ARMISelLowering.cpp revision fc501a3ec9d97e372ecb1bd9cf32d861da46b2c9
1//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the interfaces that ARM uses to lower LLVM code into a
11// selection DAG.
12//
13//===----------------------------------------------------------------------===//
14
15#define DEBUG_TYPE "arm-isel"
16#include "ARM.h"
17#include "ARMCallingConv.h"
18#include "ARMConstantPoolValue.h"
19#include "ARMISelLowering.h"
20#include "ARMMachineFunctionInfo.h"
21#include "ARMPerfectShuffle.h"
22#include "ARMRegisterInfo.h"
23#include "ARMSubtarget.h"
24#include "ARMTargetMachine.h"
25#include "ARMTargetObjectFile.h"
26#include "MCTargetDesc/ARMAddressingModes.h"
27#include "llvm/CallingConv.h"
28#include "llvm/Constants.h"
29#include "llvm/Function.h"
30#include "llvm/GlobalValue.h"
31#include "llvm/Instruction.h"
32#include "llvm/Instructions.h"
33#include "llvm/Intrinsics.h"
34#include "llvm/Type.h"
35#include "llvm/CodeGen/CallingConvLower.h"
36#include "llvm/CodeGen/IntrinsicLowering.h"
37#include "llvm/CodeGen/MachineBasicBlock.h"
38#include "llvm/CodeGen/MachineFrameInfo.h"
39#include "llvm/CodeGen/MachineFunction.h"
40#include "llvm/CodeGen/MachineInstrBuilder.h"
41#include "llvm/CodeGen/MachineModuleInfo.h"
42#include "llvm/CodeGen/MachineRegisterInfo.h"
43#include "llvm/CodeGen/SelectionDAG.h"
44#include "llvm/MC/MCSectionMachO.h"
45#include "llvm/Target/TargetOptions.h"
46#include "llvm/ADT/StringExtras.h"
47#include "llvm/ADT/Statistic.h"
48#include "llvm/Support/CommandLine.h"
49#include "llvm/Support/ErrorHandling.h"
50#include "llvm/Support/MathExtras.h"
51#include "llvm/Support/raw_ostream.h"
52#include <sstream>
53using namespace llvm;
54
55STATISTIC(NumTailCalls, "Number of tail calls");
56STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
57
58// This option should go away when tail calls fully work.
59static cl::opt<bool>
60EnableARMTailCalls("arm-tail-calls", cl::Hidden,
61  cl::desc("Generate tail calls (TEMPORARY OPTION)."),
62  cl::init(false));
63
64cl::opt<bool>
65EnableARMLongCalls("arm-long-calls", cl::Hidden,
66  cl::desc("Generate calls via indirect call instructions"),
67  cl::init(false));
68
69static cl::opt<bool>
70ARMInterworking("arm-interworking", cl::Hidden,
71  cl::desc("Enable / disable ARM interworking (for debugging only)"),
72  cl::init(true));
73
74namespace {
75  class ARMCCState : public CCState {
76  public:
77    ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
78               const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
79               LLVMContext &C, ParmContext PC)
80        : CCState(CC, isVarArg, MF, TM, locs, C) {
81      assert(((PC == Call) || (PC == Prologue)) &&
82             "ARMCCState users must specify whether their context is call"
83             "or prologue generation.");
84      CallOrPrologue = PC;
85    }
86  };
87}
88
89// The APCS parameter registers.
90static const unsigned GPRArgRegs[] = {
91  ARM::R0, ARM::R1, ARM::R2, ARM::R3
92};
93
94void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
95                                       EVT PromotedBitwiseVT) {
96  if (VT != PromotedLdStVT) {
97    setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
98    AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
99                       PromotedLdStVT.getSimpleVT());
100
101    setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
102    AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
103                       PromotedLdStVT.getSimpleVT());
104  }
105
106  EVT ElemTy = VT.getVectorElementType();
107  if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
108    setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
109  setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
110  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
111  if (ElemTy == MVT::i32) {
112    setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
113    setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
114    setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
115    setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
116  } else {
117    setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
118    setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
119    setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
120    setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
121  }
122  setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
123  setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
124  setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
125  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
126  setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
127  setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
128  setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
129  if (VT.isInteger()) {
130    setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
131    setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
132    setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
133  }
134
135  // Promote all bit-wise operations.
136  if (VT.isInteger() && VT != PromotedBitwiseVT) {
137    setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
138    AddPromotedToType (ISD::AND, VT.getSimpleVT(),
139                       PromotedBitwiseVT.getSimpleVT());
140    setOperationAction(ISD::OR,  VT.getSimpleVT(), Promote);
141    AddPromotedToType (ISD::OR,  VT.getSimpleVT(),
142                       PromotedBitwiseVT.getSimpleVT());
143    setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
144    AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
145                       PromotedBitwiseVT.getSimpleVT());
146  }
147
148  // Neon does not support vector divide/remainder operations.
149  setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
150  setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
151  setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
152  setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
153  setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
154  setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
155}
156
157void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
158  addRegisterClass(VT, ARM::DPRRegisterClass);
159  addTypeForNEON(VT, MVT::f64, MVT::v2i32);
160}
161
162void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
163  addRegisterClass(VT, ARM::QPRRegisterClass);
164  addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
165}
166
167static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
168  if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
169    return new TargetLoweringObjectFileMachO();
170
171  return new ARMElfTargetObjectFile();
172}
173
174ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
175    : TargetLowering(TM, createTLOF(TM)) {
176  Subtarget = &TM.getSubtarget<ARMSubtarget>();
177  RegInfo = TM.getRegisterInfo();
178  Itins = TM.getInstrItineraryData();
179
180  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
181
182  if (Subtarget->isTargetDarwin()) {
183    // Uses VFP for Thumb libfuncs if available.
184    if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
185      // Single-precision floating-point arithmetic.
186      setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
187      setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
188      setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
189      setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
190
191      // Double-precision floating-point arithmetic.
192      setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
193      setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
194      setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
195      setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
196
197      // Single-precision comparisons.
198      setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
199      setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
200      setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
201      setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
202      setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
203      setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
204      setLibcallName(RTLIB::UO_F32,  "__unordsf2vfp");
205      setLibcallName(RTLIB::O_F32,   "__unordsf2vfp");
206
207      setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
208      setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
209      setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
210      setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
211      setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
212      setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
213      setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
214      setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
215
216      // Double-precision comparisons.
217      setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
218      setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
219      setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
220      setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
221      setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
222      setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
223      setLibcallName(RTLIB::UO_F64,  "__unorddf2vfp");
224      setLibcallName(RTLIB::O_F64,   "__unorddf2vfp");
225
226      setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
227      setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
228      setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
229      setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
230      setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
231      setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
232      setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
233      setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
234
235      // Floating-point to integer conversions.
236      // i64 conversions are done via library routines even when generating VFP
237      // instructions, so use the same ones.
238      setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
239      setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
240      setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
241      setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
242
243      // Conversions between floating types.
244      setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
245      setLibcallName(RTLIB::FPEXT_F32_F64,   "__extendsfdf2vfp");
246
247      // Integer to floating-point conversions.
248      // i64 conversions are done via library routines even when generating VFP
249      // instructions, so use the same ones.
250      // FIXME: There appears to be some naming inconsistency in ARM libgcc:
251      // e.g., __floatunsidf vs. __floatunssidfvfp.
252      setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
253      setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
254      setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
255      setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
256    }
257  }
258
259  // These libcalls are not available in 32-bit.
260  setLibcallName(RTLIB::SHL_I128, 0);
261  setLibcallName(RTLIB::SRL_I128, 0);
262  setLibcallName(RTLIB::SRA_I128, 0);
263
264  if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
265    // Double-precision floating-point arithmetic helper functions
266    // RTABI chapter 4.1.2, Table 2
267    setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
268    setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
269    setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
270    setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
271    setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
272    setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
273    setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
274    setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
275
276    // Double-precision floating-point comparison helper functions
277    // RTABI chapter 4.1.2, Table 3
278    setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
279    setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
280    setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
281    setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
282    setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
283    setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
284    setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
285    setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
286    setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
287    setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
288    setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
289    setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
290    setLibcallName(RTLIB::UO_F64,  "__aeabi_dcmpun");
291    setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
292    setLibcallName(RTLIB::O_F64,   "__aeabi_dcmpun");
293    setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
294    setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
295    setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
296    setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
297    setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
298    setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
299    setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
300    setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
301    setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
302
303    // Single-precision floating-point arithmetic helper functions
304    // RTABI chapter 4.1.2, Table 4
305    setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
306    setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
307    setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
308    setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
309    setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
310    setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
311    setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
312    setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
313
314    // Single-precision floating-point comparison helper functions
315    // RTABI chapter 4.1.2, Table 5
316    setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
317    setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
318    setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
319    setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
320    setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
321    setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
322    setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
323    setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
324    setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
325    setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
326    setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
327    setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
328    setLibcallName(RTLIB::UO_F32,  "__aeabi_fcmpun");
329    setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
330    setLibcallName(RTLIB::O_F32,   "__aeabi_fcmpun");
331    setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
332    setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
333    setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
334    setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
335    setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
336    setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
337    setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
338    setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
339    setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
340
341    // Floating-point to integer conversions.
342    // RTABI chapter 4.1.2, Table 6
343    setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
344    setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
345    setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
346    setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
347    setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
348    setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
349    setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
350    setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
351    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
352    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
353    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
354    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
355    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
356    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
357    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
358    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
359
360    // Conversions between floating types.
361    // RTABI chapter 4.1.2, Table 7
362    setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
363    setLibcallName(RTLIB::FPEXT_F32_F64,   "__aeabi_f2d");
364    setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
365    setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
366
367    // Integer to floating-point conversions.
368    // RTABI chapter 4.1.2, Table 8
369    setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
370    setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
371    setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
372    setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
373    setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
374    setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
375    setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
376    setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
377    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
378    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
379    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
380    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
381    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
382    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
383    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
384    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
385
386    // Long long helper functions
387    // RTABI chapter 4.2, Table 9
388    setLibcallName(RTLIB::MUL_I64,  "__aeabi_lmul");
389    setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
390    setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
391    setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
392    setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
393    setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
394    setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
395    setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
396    setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
397    setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
398
399    // Integer division functions
400    // RTABI chapter 4.3.1
401    setLibcallName(RTLIB::SDIV_I8,  "__aeabi_idiv");
402    setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
403    setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
404    setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
405    setLibcallName(RTLIB::UDIV_I8,  "__aeabi_uidiv");
406    setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
407    setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
408    setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
409    setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
410    setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
411    setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
412    setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
413    setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
414    setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
415    setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
416    setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
417
418    // Memory operations
419    // RTABI chapter 4.3.4
420    setLibcallName(RTLIB::MEMCPY,  "__aeabi_memcpy");
421    setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
422    setLibcallName(RTLIB::MEMSET,  "__aeabi_memset");
423    setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
424    setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
425    setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
426  }
427
428  // Use divmod compiler-rt calls for iOS 5.0 and later.
429  if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
430      !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
431    setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
432    setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
433  }
434
435  if (Subtarget->isThumb1Only())
436    addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
437  else
438    addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
439  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
440      !Subtarget->isThumb1Only()) {
441    addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
442    if (!Subtarget->isFPOnlySP())
443      addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
444
445    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
446  }
447
448  for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
449       VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
450    for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
451         InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
452      setTruncStoreAction((MVT::SimpleValueType)VT,
453                          (MVT::SimpleValueType)InnerVT, Expand);
454    setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
455    setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
456    setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
457  }
458
459  if (Subtarget->hasNEON()) {
460    addDRTypeForNEON(MVT::v2f32);
461    addDRTypeForNEON(MVT::v8i8);
462    addDRTypeForNEON(MVT::v4i16);
463    addDRTypeForNEON(MVT::v2i32);
464    addDRTypeForNEON(MVT::v1i64);
465
466    addQRTypeForNEON(MVT::v4f32);
467    addQRTypeForNEON(MVT::v2f64);
468    addQRTypeForNEON(MVT::v16i8);
469    addQRTypeForNEON(MVT::v8i16);
470    addQRTypeForNEON(MVT::v4i32);
471    addQRTypeForNEON(MVT::v2i64);
472
473    // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
474    // neither Neon nor VFP support any arithmetic operations on it.
475    // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
476    // supported for v4f32.
477    setOperationAction(ISD::FADD, MVT::v2f64, Expand);
478    setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
479    setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
480    // FIXME: Code duplication: FDIV and FREM are expanded always, see
481    // ARMTargetLowering::addTypeForNEON method for details.
482    setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
483    setOperationAction(ISD::FREM, MVT::v2f64, Expand);
484    // FIXME: Create unittest.
485    // In another words, find a way when "copysign" appears in DAG with vector
486    // operands.
487    setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
488    // FIXME: Code duplication: SETCC has custom operation action, see
489    // ARMTargetLowering::addTypeForNEON method for details.
490    setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
491    // FIXME: Create unittest for FNEG and for FABS.
492    setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
493    setOperationAction(ISD::FABS, MVT::v2f64, Expand);
494    setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
495    setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
496    setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
497    setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
498    setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
499    setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
500    setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
501    setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
502    setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
503    setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
504    // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
505    setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
506    setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
507    setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
508    setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
509    setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
510
511    setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
512    setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
513    setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
514    setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
515    setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
516    setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
517    setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
518    setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
519    setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
520    setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
521
522    // Neon does not support some operations on v1i64 and v2i64 types.
523    setOperationAction(ISD::MUL, MVT::v1i64, Expand);
524    // Custom handling for some quad-vector types to detect VMULL.
525    setOperationAction(ISD::MUL, MVT::v8i16, Custom);
526    setOperationAction(ISD::MUL, MVT::v4i32, Custom);
527    setOperationAction(ISD::MUL, MVT::v2i64, Custom);
528    // Custom handling for some vector types to avoid expensive expansions
529    setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
530    setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
531    setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
532    setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
533    setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
534    setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
535    // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
536    // a destination type that is wider than the source, and nor does
537    // it have a FP_TO_[SU]INT instruction with a narrower destination than
538    // source.
539    setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
540    setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
541    setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
542    setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
543
544    setTargetDAGCombine(ISD::INTRINSIC_VOID);
545    setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
546    setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
547    setTargetDAGCombine(ISD::SHL);
548    setTargetDAGCombine(ISD::SRL);
549    setTargetDAGCombine(ISD::SRA);
550    setTargetDAGCombine(ISD::SIGN_EXTEND);
551    setTargetDAGCombine(ISD::ZERO_EXTEND);
552    setTargetDAGCombine(ISD::ANY_EXTEND);
553    setTargetDAGCombine(ISD::SELECT_CC);
554    setTargetDAGCombine(ISD::BUILD_VECTOR);
555    setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
556    setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
557    setTargetDAGCombine(ISD::STORE);
558    setTargetDAGCombine(ISD::FP_TO_SINT);
559    setTargetDAGCombine(ISD::FP_TO_UINT);
560    setTargetDAGCombine(ISD::FDIV);
561
562    // It is legal to extload from v4i8 to v4i16 or v4i32.
563    MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
564                  MVT::v4i16, MVT::v2i16,
565                  MVT::v2i32};
566    for (unsigned i = 0; i < 6; ++i) {
567      setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
568      setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
569      setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
570    }
571  }
572
573  computeRegisterProperties();
574
575  // ARM does not have f32 extending load.
576  setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
577
578  // ARM does not have i1 sign extending load.
579  setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
580
581  // ARM supports all 4 flavors of integer indexed load / store.
582  if (!Subtarget->isThumb1Only()) {
583    for (unsigned im = (unsigned)ISD::PRE_INC;
584         im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
585      setIndexedLoadAction(im,  MVT::i1,  Legal);
586      setIndexedLoadAction(im,  MVT::i8,  Legal);
587      setIndexedLoadAction(im,  MVT::i16, Legal);
588      setIndexedLoadAction(im,  MVT::i32, Legal);
589      setIndexedStoreAction(im, MVT::i1,  Legal);
590      setIndexedStoreAction(im, MVT::i8,  Legal);
591      setIndexedStoreAction(im, MVT::i16, Legal);
592      setIndexedStoreAction(im, MVT::i32, Legal);
593    }
594  }
595
596  // i64 operation support.
597  setOperationAction(ISD::MUL,     MVT::i64, Expand);
598  setOperationAction(ISD::MULHU,   MVT::i32, Expand);
599  if (Subtarget->isThumb1Only()) {
600    setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
601    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
602  }
603  if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
604      || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
605    setOperationAction(ISD::MULHS, MVT::i32, Expand);
606
607  setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
608  setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
609  setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
610  setOperationAction(ISD::SRL,       MVT::i64, Custom);
611  setOperationAction(ISD::SRA,       MVT::i64, Custom);
612
613  if (!Subtarget->isThumb1Only()) {
614    // FIXME: We should do this for Thumb1 as well.
615    setOperationAction(ISD::ADDC,    MVT::i32, Custom);
616    setOperationAction(ISD::ADDE,    MVT::i32, Custom);
617    setOperationAction(ISD::SUBC,    MVT::i32, Custom);
618    setOperationAction(ISD::SUBE,    MVT::i32, Custom);
619  }
620
621  // ARM does not have ROTL.
622  setOperationAction(ISD::ROTL,  MVT::i32, Expand);
623  setOperationAction(ISD::CTTZ,  MVT::i32, Custom);
624  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
625  if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
626    setOperationAction(ISD::CTLZ, MVT::i32, Expand);
627
628  // These just redirect to CTTZ and CTLZ on ARM.
629  setOperationAction(ISD::CTTZ_ZERO_UNDEF  , MVT::i32  , Expand);
630  setOperationAction(ISD::CTLZ_ZERO_UNDEF  , MVT::i32  , Expand);
631
632  // Only ARMv6 has BSWAP.
633  if (!Subtarget->hasV6Ops())
634    setOperationAction(ISD::BSWAP, MVT::i32, Expand);
635
636  // These are expanded into libcalls.
637  if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
638    // v7M has a hardware divider
639    setOperationAction(ISD::SDIV,  MVT::i32, Expand);
640    setOperationAction(ISD::UDIV,  MVT::i32, Expand);
641  }
642  setOperationAction(ISD::SREM,  MVT::i32, Expand);
643  setOperationAction(ISD::UREM,  MVT::i32, Expand);
644  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
645  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
646
647  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
648  setOperationAction(ISD::ConstantPool,  MVT::i32,   Custom);
649  setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
650  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
651  setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
652
653  setOperationAction(ISD::TRAP, MVT::Other, Legal);
654
655  // Use the default implementation.
656  setOperationAction(ISD::VASTART,            MVT::Other, Custom);
657  setOperationAction(ISD::VAARG,              MVT::Other, Expand);
658  setOperationAction(ISD::VACOPY,             MVT::Other, Expand);
659  setOperationAction(ISD::VAEND,              MVT::Other, Expand);
660  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
661  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
662
663  if (!Subtarget->isTargetDarwin()) {
664    // Non-Darwin platforms may return values in these registers via the
665    // personality function.
666    setOperationAction(ISD::EHSELECTION,      MVT::i32,   Expand);
667    setOperationAction(ISD::EXCEPTIONADDR,    MVT::i32,   Expand);
668    setExceptionPointerRegister(ARM::R0);
669    setExceptionSelectorRegister(ARM::R1);
670  }
671
672  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
673  // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
674  // the default expansion.
675  // FIXME: This should be checking for v6k, not just v6.
676  if (Subtarget->hasDataBarrier() ||
677      (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
678    // membarrier needs custom lowering; the rest are legal and handled
679    // normally.
680    setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
681    setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
682    // Custom lowering for 64-bit ops
683    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i64, Custom);
684    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i64, Custom);
685    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i64, Custom);
686    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i64, Custom);
687    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i64, Custom);
688    setOperationAction(ISD::ATOMIC_SWAP,  MVT::i64, Custom);
689    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i64, Custom);
690    // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
691    setInsertFencesForAtomic(true);
692  } else {
693    // Set them all for expansion, which will force libcalls.
694    setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
695    setOperationAction(ISD::ATOMIC_FENCE,   MVT::Other, Expand);
696    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i32, Expand);
697    setOperationAction(ISD::ATOMIC_SWAP,      MVT::i32, Expand);
698    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i32, Expand);
699    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i32, Expand);
700    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i32, Expand);
701    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i32, Expand);
702    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i32, Expand);
703    setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
704    setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
705    setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
706    setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
707    setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
708    // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
709    // Unordered/Monotonic case.
710    setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
711    setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
712    // Since the libcalls include locking, fold in the fences
713    setShouldFoldAtomicFences(true);
714  }
715
716  setOperationAction(ISD::PREFETCH,         MVT::Other, Custom);
717
718  // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
719  if (!Subtarget->hasV6Ops()) {
720    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
721    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
722  }
723  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
724
725  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
726      !Subtarget->isThumb1Only()) {
727    // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
728    // iff target supports vfp2.
729    setOperationAction(ISD::BITCAST, MVT::i64, Custom);
730    setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
731  }
732
733  // We want to custom lower some of our intrinsics.
734  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
735  if (Subtarget->isTargetDarwin()) {
736    setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
737    setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
738    setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
739  }
740
741  setOperationAction(ISD::SETCC,     MVT::i32, Expand);
742  setOperationAction(ISD::SETCC,     MVT::f32, Expand);
743  setOperationAction(ISD::SETCC,     MVT::f64, Expand);
744  setOperationAction(ISD::SELECT,    MVT::i32, Custom);
745  setOperationAction(ISD::SELECT,    MVT::f32, Custom);
746  setOperationAction(ISD::SELECT,    MVT::f64, Custom);
747  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
748  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
749  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
750
751  setOperationAction(ISD::BRCOND,    MVT::Other, Expand);
752  setOperationAction(ISD::BR_CC,     MVT::i32,   Custom);
753  setOperationAction(ISD::BR_CC,     MVT::f32,   Custom);
754  setOperationAction(ISD::BR_CC,     MVT::f64,   Custom);
755  setOperationAction(ISD::BR_JT,     MVT::Other, Custom);
756
757  // We don't support sin/cos/fmod/copysign/pow
758  setOperationAction(ISD::FSIN,      MVT::f64, Expand);
759  setOperationAction(ISD::FSIN,      MVT::f32, Expand);
760  setOperationAction(ISD::FCOS,      MVT::f32, Expand);
761  setOperationAction(ISD::FCOS,      MVT::f64, Expand);
762  setOperationAction(ISD::FREM,      MVT::f64, Expand);
763  setOperationAction(ISD::FREM,      MVT::f32, Expand);
764  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
765      !Subtarget->isThumb1Only()) {
766    setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
767    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
768  }
769  setOperationAction(ISD::FPOW,      MVT::f64, Expand);
770  setOperationAction(ISD::FPOW,      MVT::f32, Expand);
771
772  setOperationAction(ISD::FMA, MVT::f64, Expand);
773  setOperationAction(ISD::FMA, MVT::f32, Expand);
774
775  // Various VFP goodness
776  if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
777    // int <-> fp are custom expanded into bit_convert + ARMISD ops.
778    if (Subtarget->hasVFP2()) {
779      setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
780      setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
781      setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
782      setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
783    }
784    // Special handling for half-precision FP.
785    if (!Subtarget->hasFP16()) {
786      setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
787      setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
788    }
789  }
790
791  // We have target-specific dag combine patterns for the following nodes:
792  // ARMISD::VMOVRRD  - No need to call setTargetDAGCombine
793  setTargetDAGCombine(ISD::ADD);
794  setTargetDAGCombine(ISD::SUB);
795  setTargetDAGCombine(ISD::MUL);
796
797  if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
798    setTargetDAGCombine(ISD::AND);
799    setTargetDAGCombine(ISD::OR);
800    setTargetDAGCombine(ISD::XOR);
801  }
802
803  if (Subtarget->hasV6Ops())
804    setTargetDAGCombine(ISD::SRL);
805
806  setStackPointerRegisterToSaveRestore(ARM::SP);
807
808  if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
809      !Subtarget->hasVFP2())
810    setSchedulingPreference(Sched::RegPressure);
811  else
812    setSchedulingPreference(Sched::Hybrid);
813
814  //// temporary - rewrite interface to use type
815  maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
816  maxStoresPerMemset = 16;
817  maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
818
819  // On ARM arguments smaller than 4 bytes are extended, so all arguments
820  // are at least 4 bytes aligned.
821  setMinStackArgumentAlignment(4);
822
823  benefitFromCodePlacementOpt = true;
824
825  setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
826}
827
828// FIXME: It might make sense to define the representative register class as the
829// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
830// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
831// SPR's representative would be DPR_VFP2. This should work well if register
832// pressure tracking were modified such that a register use would increment the
833// pressure of the register class's representative and all of it's super
834// classes' representatives transitively. We have not implemented this because
835// of the difficulty prior to coalescing of modeling operand register classes
836// due to the common occurrence of cross class copies and subregister insertions
837// and extractions.
838std::pair<const TargetRegisterClass*, uint8_t>
839ARMTargetLowering::findRepresentativeClass(EVT VT) const{
840  const TargetRegisterClass *RRC = 0;
841  uint8_t Cost = 1;
842  switch (VT.getSimpleVT().SimpleTy) {
843  default:
844    return TargetLowering::findRepresentativeClass(VT);
845  // Use DPR as representative register class for all floating point
846  // and vector types. Since there are 32 SPR registers and 32 DPR registers so
847  // the cost is 1 for both f32 and f64.
848  case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
849  case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
850    RRC = ARM::DPRRegisterClass;
851    // When NEON is used for SP, only half of the register file is available
852    // because operations that define both SP and DP results will be constrained
853    // to the VFP2 class (D0-D15). We currently model this constraint prior to
854    // coalescing by double-counting the SP regs. See the FIXME above.
855    if (Subtarget->useNEONForSinglePrecisionFP())
856      Cost = 2;
857    break;
858  case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
859  case MVT::v4f32: case MVT::v2f64:
860    RRC = ARM::DPRRegisterClass;
861    Cost = 2;
862    break;
863  case MVT::v4i64:
864    RRC = ARM::DPRRegisterClass;
865    Cost = 4;
866    break;
867  case MVT::v8i64:
868    RRC = ARM::DPRRegisterClass;
869    Cost = 8;
870    break;
871  }
872  return std::make_pair(RRC, Cost);
873}
874
875const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
876  switch (Opcode) {
877  default: return 0;
878  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
879  case ARMISD::WrapperDYN:    return "ARMISD::WrapperDYN";
880  case ARMISD::WrapperPIC:    return "ARMISD::WrapperPIC";
881  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
882  case ARMISD::CALL:          return "ARMISD::CALL";
883  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
884  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
885  case ARMISD::tCALL:         return "ARMISD::tCALL";
886  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
887  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
888  case ARMISD::BR2_JT:        return "ARMISD::BR2_JT";
889  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
890  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
891  case ARMISD::CMP:           return "ARMISD::CMP";
892  case ARMISD::CMPZ:          return "ARMISD::CMPZ";
893  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
894  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
895  case ARMISD::BCC_i64:       return "ARMISD::BCC_i64";
896  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";
897
898  case ARMISD::CMOV:          return "ARMISD::CMOV";
899  case ARMISD::CAND:          return "ARMISD::CAND";
900  case ARMISD::COR:           return "ARMISD::COR";
901  case ARMISD::CXOR:          return "ARMISD::CXOR";
902
903  case ARMISD::RBIT:          return "ARMISD::RBIT";
904
905  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
906  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
907  case ARMISD::SITOF:         return "ARMISD::SITOF";
908  case ARMISD::UITOF:         return "ARMISD::UITOF";
909
910  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
911  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
912  case ARMISD::RRX:           return "ARMISD::RRX";
913
914  case ARMISD::ADDC:          return "ARMISD::ADDC";
915  case ARMISD::ADDE:          return "ARMISD::ADDE";
916  case ARMISD::SUBC:          return "ARMISD::SUBC";
917  case ARMISD::SUBE:          return "ARMISD::SUBE";
918
919  case ARMISD::VMOVRRD:       return "ARMISD::VMOVRRD";
920  case ARMISD::VMOVDRR:       return "ARMISD::VMOVDRR";
921
922  case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
923  case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
924
925  case ARMISD::TC_RETURN:     return "ARMISD::TC_RETURN";
926
927  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
928
929  case ARMISD::DYN_ALLOC:     return "ARMISD::DYN_ALLOC";
930
931  case ARMISD::MEMBARRIER:    return "ARMISD::MEMBARRIER";
932  case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
933
934  case ARMISD::PRELOAD:       return "ARMISD::PRELOAD";
935
936  case ARMISD::VCEQ:          return "ARMISD::VCEQ";
937  case ARMISD::VCEQZ:         return "ARMISD::VCEQZ";
938  case ARMISD::VCGE:          return "ARMISD::VCGE";
939  case ARMISD::VCGEZ:         return "ARMISD::VCGEZ";
940  case ARMISD::VCLEZ:         return "ARMISD::VCLEZ";
941  case ARMISD::VCGEU:         return "ARMISD::VCGEU";
942  case ARMISD::VCGT:          return "ARMISD::VCGT";
943  case ARMISD::VCGTZ:         return "ARMISD::VCGTZ";
944  case ARMISD::VCLTZ:         return "ARMISD::VCLTZ";
945  case ARMISD::VCGTU:         return "ARMISD::VCGTU";
946  case ARMISD::VTST:          return "ARMISD::VTST";
947
948  case ARMISD::VSHL:          return "ARMISD::VSHL";
949  case ARMISD::VSHRs:         return "ARMISD::VSHRs";
950  case ARMISD::VSHRu:         return "ARMISD::VSHRu";
951  case ARMISD::VSHLLs:        return "ARMISD::VSHLLs";
952  case ARMISD::VSHLLu:        return "ARMISD::VSHLLu";
953  case ARMISD::VSHLLi:        return "ARMISD::VSHLLi";
954  case ARMISD::VSHRN:         return "ARMISD::VSHRN";
955  case ARMISD::VRSHRs:        return "ARMISD::VRSHRs";
956  case ARMISD::VRSHRu:        return "ARMISD::VRSHRu";
957  case ARMISD::VRSHRN:        return "ARMISD::VRSHRN";
958  case ARMISD::VQSHLs:        return "ARMISD::VQSHLs";
959  case ARMISD::VQSHLu:        return "ARMISD::VQSHLu";
960  case ARMISD::VQSHLsu:       return "ARMISD::VQSHLsu";
961  case ARMISD::VQSHRNs:       return "ARMISD::VQSHRNs";
962  case ARMISD::VQSHRNu:       return "ARMISD::VQSHRNu";
963  case ARMISD::VQSHRNsu:      return "ARMISD::VQSHRNsu";
964  case ARMISD::VQRSHRNs:      return "ARMISD::VQRSHRNs";
965  case ARMISD::VQRSHRNu:      return "ARMISD::VQRSHRNu";
966  case ARMISD::VQRSHRNsu:     return "ARMISD::VQRSHRNsu";
967  case ARMISD::VGETLANEu:     return "ARMISD::VGETLANEu";
968  case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
969  case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
970  case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
971  case ARMISD::VMOVFPIMM:     return "ARMISD::VMOVFPIMM";
972  case ARMISD::VDUP:          return "ARMISD::VDUP";
973  case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
974  case ARMISD::VEXT:          return "ARMISD::VEXT";
975  case ARMISD::VREV64:        return "ARMISD::VREV64";
976  case ARMISD::VREV32:        return "ARMISD::VREV32";
977  case ARMISD::VREV16:        return "ARMISD::VREV16";
978  case ARMISD::VZIP:          return "ARMISD::VZIP";
979  case ARMISD::VUZP:          return "ARMISD::VUZP";
980  case ARMISD::VTRN:          return "ARMISD::VTRN";
981  case ARMISD::VTBL1:         return "ARMISD::VTBL1";
982  case ARMISD::VTBL2:         return "ARMISD::VTBL2";
983  case ARMISD::VMULLs:        return "ARMISD::VMULLs";
984  case ARMISD::VMULLu:        return "ARMISD::VMULLu";
985  case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
986  case ARMISD::FMAX:          return "ARMISD::FMAX";
987  case ARMISD::FMIN:          return "ARMISD::FMIN";
988  case ARMISD::BFI:           return "ARMISD::BFI";
989  case ARMISD::VORRIMM:       return "ARMISD::VORRIMM";
990  case ARMISD::VBICIMM:       return "ARMISD::VBICIMM";
991  case ARMISD::VBSL:          return "ARMISD::VBSL";
992  case ARMISD::VLD2DUP:       return "ARMISD::VLD2DUP";
993  case ARMISD::VLD3DUP:       return "ARMISD::VLD3DUP";
994  case ARMISD::VLD4DUP:       return "ARMISD::VLD4DUP";
995  case ARMISD::VLD1_UPD:      return "ARMISD::VLD1_UPD";
996  case ARMISD::VLD2_UPD:      return "ARMISD::VLD2_UPD";
997  case ARMISD::VLD3_UPD:      return "ARMISD::VLD3_UPD";
998  case ARMISD::VLD4_UPD:      return "ARMISD::VLD4_UPD";
999  case ARMISD::VLD2LN_UPD:    return "ARMISD::VLD2LN_UPD";
1000  case ARMISD::VLD3LN_UPD:    return "ARMISD::VLD3LN_UPD";
1001  case ARMISD::VLD4LN_UPD:    return "ARMISD::VLD4LN_UPD";
1002  case ARMISD::VLD2DUP_UPD:   return "ARMISD::VLD2DUP_UPD";
1003  case ARMISD::VLD3DUP_UPD:   return "ARMISD::VLD3DUP_UPD";
1004  case ARMISD::VLD4DUP_UPD:   return "ARMISD::VLD4DUP_UPD";
1005  case ARMISD::VST1_UPD:      return "ARMISD::VST1_UPD";
1006  case ARMISD::VST2_UPD:      return "ARMISD::VST2_UPD";
1007  case ARMISD::VST3_UPD:      return "ARMISD::VST3_UPD";
1008  case ARMISD::VST4_UPD:      return "ARMISD::VST4_UPD";
1009  case ARMISD::VST2LN_UPD:    return "ARMISD::VST2LN_UPD";
1010  case ARMISD::VST3LN_UPD:    return "ARMISD::VST3LN_UPD";
1011  case ARMISD::VST4LN_UPD:    return "ARMISD::VST4LN_UPD";
1012  }
1013}
1014
1015EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1016  if (!VT.isVector()) return getPointerTy();
1017  return VT.changeVectorElementTypeToInteger();
1018}
1019
1020/// getRegClassFor - Return the register class that should be used for the
1021/// specified value type.
1022const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1023  // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1024  // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1025  // load / store 4 to 8 consecutive D registers.
1026  if (Subtarget->hasNEON()) {
1027    if (VT == MVT::v4i64)
1028      return ARM::QQPRRegisterClass;
1029    else if (VT == MVT::v8i64)
1030      return ARM::QQQQPRRegisterClass;
1031  }
1032  return TargetLowering::getRegClassFor(VT);
1033}
1034
1035// Create a fast isel object.
1036FastISel *
1037ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1038  return ARM::createFastISel(funcInfo);
1039}
1040
1041/// getMaximalGlobalOffset - Returns the maximal possible offset which can
1042/// be used for loads / stores from the global.
1043unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1044  return (Subtarget->isThumb1Only() ? 127 : 4095);
1045}
1046
1047Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1048  unsigned NumVals = N->getNumValues();
1049  if (!NumVals)
1050    return Sched::RegPressure;
1051
1052  for (unsigned i = 0; i != NumVals; ++i) {
1053    EVT VT = N->getValueType(i);
1054    if (VT == MVT::Glue || VT == MVT::Other)
1055      continue;
1056    if (VT.isFloatingPoint() || VT.isVector())
1057      return Sched::ILP;
1058  }
1059
1060  if (!N->isMachineOpcode())
1061    return Sched::RegPressure;
1062
1063  // Load are scheduled for latency even if there instruction itinerary
1064  // is not available.
1065  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1066  const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1067
1068  if (MCID.getNumDefs() == 0)
1069    return Sched::RegPressure;
1070  if (!Itins->isEmpty() &&
1071      Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1072    return Sched::ILP;
1073
1074  return Sched::RegPressure;
1075}
1076
1077//===----------------------------------------------------------------------===//
1078// Lowering Code
1079//===----------------------------------------------------------------------===//
1080
1081/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1082static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1083  switch (CC) {
1084  default: llvm_unreachable("Unknown condition code!");
1085  case ISD::SETNE:  return ARMCC::NE;
1086  case ISD::SETEQ:  return ARMCC::EQ;
1087  case ISD::SETGT:  return ARMCC::GT;
1088  case ISD::SETGE:  return ARMCC::GE;
1089  case ISD::SETLT:  return ARMCC::LT;
1090  case ISD::SETLE:  return ARMCC::LE;
1091  case ISD::SETUGT: return ARMCC::HI;
1092  case ISD::SETUGE: return ARMCC::HS;
1093  case ISD::SETULT: return ARMCC::LO;
1094  case ISD::SETULE: return ARMCC::LS;
1095  }
1096}
1097
1098/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1099static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1100                        ARMCC::CondCodes &CondCode2) {
1101  CondCode2 = ARMCC::AL;
1102  switch (CC) {
1103  default: llvm_unreachable("Unknown FP condition!");
1104  case ISD::SETEQ:
1105  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1106  case ISD::SETGT:
1107  case ISD::SETOGT: CondCode = ARMCC::GT; break;
1108  case ISD::SETGE:
1109  case ISD::SETOGE: CondCode = ARMCC::GE; break;
1110  case ISD::SETOLT: CondCode = ARMCC::MI; break;
1111  case ISD::SETOLE: CondCode = ARMCC::LS; break;
1112  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1113  case ISD::SETO:   CondCode = ARMCC::VC; break;
1114  case ISD::SETUO:  CondCode = ARMCC::VS; break;
1115  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1116  case ISD::SETUGT: CondCode = ARMCC::HI; break;
1117  case ISD::SETUGE: CondCode = ARMCC::PL; break;
1118  case ISD::SETLT:
1119  case ISD::SETULT: CondCode = ARMCC::LT; break;
1120  case ISD::SETLE:
1121  case ISD::SETULE: CondCode = ARMCC::LE; break;
1122  case ISD::SETNE:
1123  case ISD::SETUNE: CondCode = ARMCC::NE; break;
1124  }
1125}
1126
1127//===----------------------------------------------------------------------===//
1128//                      Calling Convention Implementation
1129//===----------------------------------------------------------------------===//
1130
1131#include "ARMGenCallingConv.inc"
1132
1133/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1134/// given CallingConvention value.
1135CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1136                                                 bool Return,
1137                                                 bool isVarArg) const {
1138  switch (CC) {
1139  default:
1140    llvm_unreachable("Unsupported calling convention");
1141  case CallingConv::Fast:
1142    if (Subtarget->hasVFP2() && !isVarArg) {
1143      if (!Subtarget->isAAPCS_ABI())
1144        return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1145      // For AAPCS ABI targets, just use VFP variant of the calling convention.
1146      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1147    }
1148    // Fallthrough
1149  case CallingConv::C: {
1150    // Use target triple & subtarget features to do actual dispatch.
1151    if (!Subtarget->isAAPCS_ABI())
1152      return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1153    else if (Subtarget->hasVFP2() &&
1154             getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1155             !isVarArg)
1156      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1157    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1158  }
1159  case CallingConv::ARM_AAPCS_VFP:
1160    if (!isVarArg)
1161      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1162    // Fallthrough
1163  case CallingConv::ARM_AAPCS:
1164    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165  case CallingConv::ARM_APCS:
1166    return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1167  }
1168}
1169
1170/// LowerCallResult - Lower the result values of a call into the
1171/// appropriate copies out of appropriate physical registers.
1172SDValue
1173ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1174                                   CallingConv::ID CallConv, bool isVarArg,
1175                                   const SmallVectorImpl<ISD::InputArg> &Ins,
1176                                   DebugLoc dl, SelectionDAG &DAG,
1177                                   SmallVectorImpl<SDValue> &InVals) const {
1178
1179  // Assign locations to each value returned by this call.
1180  SmallVector<CCValAssign, 16> RVLocs;
1181  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1182                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1183  CCInfo.AnalyzeCallResult(Ins,
1184                           CCAssignFnForNode(CallConv, /* Return*/ true,
1185                                             isVarArg));
1186
1187  // Copy all of the result registers out of their specified physreg.
1188  for (unsigned i = 0; i != RVLocs.size(); ++i) {
1189    CCValAssign VA = RVLocs[i];
1190
1191    SDValue Val;
1192    if (VA.needsCustom()) {
1193      // Handle f64 or half of a v2f64.
1194      SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1195                                      InFlag);
1196      Chain = Lo.getValue(1);
1197      InFlag = Lo.getValue(2);
1198      VA = RVLocs[++i]; // skip ahead to next loc
1199      SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1200                                      InFlag);
1201      Chain = Hi.getValue(1);
1202      InFlag = Hi.getValue(2);
1203      Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1204
1205      if (VA.getLocVT() == MVT::v2f64) {
1206        SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1207        Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1208                          DAG.getConstant(0, MVT::i32));
1209
1210        VA = RVLocs[++i]; // skip ahead to next loc
1211        Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1212        Chain = Lo.getValue(1);
1213        InFlag = Lo.getValue(2);
1214        VA = RVLocs[++i]; // skip ahead to next loc
1215        Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1216        Chain = Hi.getValue(1);
1217        InFlag = Hi.getValue(2);
1218        Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1219        Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1220                          DAG.getConstant(1, MVT::i32));
1221      }
1222    } else {
1223      Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1224                               InFlag);
1225      Chain = Val.getValue(1);
1226      InFlag = Val.getValue(2);
1227    }
1228
1229    switch (VA.getLocInfo()) {
1230    default: llvm_unreachable("Unknown loc info!");
1231    case CCValAssign::Full: break;
1232    case CCValAssign::BCvt:
1233      Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1234      break;
1235    }
1236
1237    InVals.push_back(Val);
1238  }
1239
1240  return Chain;
1241}
1242
1243/// LowerMemOpCallTo - Store the argument to the stack.
1244SDValue
1245ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1246                                    SDValue StackPtr, SDValue Arg,
1247                                    DebugLoc dl, SelectionDAG &DAG,
1248                                    const CCValAssign &VA,
1249                                    ISD::ArgFlagsTy Flags) const {
1250  unsigned LocMemOffset = VA.getLocMemOffset();
1251  SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1252  PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1253  return DAG.getStore(Chain, dl, Arg, PtrOff,
1254                      MachinePointerInfo::getStack(LocMemOffset),
1255                      false, false, 0);
1256}
1257
1258void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1259                                         SDValue Chain, SDValue &Arg,
1260                                         RegsToPassVector &RegsToPass,
1261                                         CCValAssign &VA, CCValAssign &NextVA,
1262                                         SDValue &StackPtr,
1263                                         SmallVector<SDValue, 8> &MemOpChains,
1264                                         ISD::ArgFlagsTy Flags) const {
1265
1266  SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1267                              DAG.getVTList(MVT::i32, MVT::i32), Arg);
1268  RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1269
1270  if (NextVA.isRegLoc())
1271    RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1272  else {
1273    assert(NextVA.isMemLoc());
1274    if (StackPtr.getNode() == 0)
1275      StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1276
1277    MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1278                                           dl, DAG, NextVA,
1279                                           Flags));
1280  }
1281}
1282
1283/// LowerCall - Lowering a call into a callseq_start <-
1284/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1285/// nodes.
1286SDValue
1287ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1288                             CallingConv::ID CallConv, bool isVarArg,
1289                             bool doesNotRet, bool &isTailCall,
1290                             const SmallVectorImpl<ISD::OutputArg> &Outs,
1291                             const SmallVectorImpl<SDValue> &OutVals,
1292                             const SmallVectorImpl<ISD::InputArg> &Ins,
1293                             DebugLoc dl, SelectionDAG &DAG,
1294                             SmallVectorImpl<SDValue> &InVals) const {
1295  MachineFunction &MF = DAG.getMachineFunction();
1296  bool IsStructRet    = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1297  bool IsSibCall = false;
1298  // Disable tail calls if they're not supported.
1299  if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1300    isTailCall = false;
1301  if (isTailCall) {
1302    // Check if it's really possible to do a tail call.
1303    isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1304                    isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1305                                                   Outs, OutVals, Ins, DAG);
1306    // We don't support GuaranteedTailCallOpt for ARM, only automatically
1307    // detected sibcalls.
1308    if (isTailCall) {
1309      ++NumTailCalls;
1310      IsSibCall = true;
1311    }
1312  }
1313
1314  // Analyze operands of the call, assigning locations to each operand.
1315  SmallVector<CCValAssign, 16> ArgLocs;
1316  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1317                 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1318  CCInfo.AnalyzeCallOperands(Outs,
1319                             CCAssignFnForNode(CallConv, /* Return*/ false,
1320                                               isVarArg));
1321
1322  // Get a count of how many bytes are to be pushed on the stack.
1323  unsigned NumBytes = CCInfo.getNextStackOffset();
1324
1325  // For tail calls, memory operands are available in our caller's stack.
1326  if (IsSibCall)
1327    NumBytes = 0;
1328
1329  // Adjust the stack pointer for the new arguments...
1330  // These operations are automatically eliminated by the prolog/epilog pass
1331  if (!IsSibCall)
1332    Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1333
1334  SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1335
1336  RegsToPassVector RegsToPass;
1337  SmallVector<SDValue, 8> MemOpChains;
1338
1339  // Walk the register/memloc assignments, inserting copies/loads.  In the case
1340  // of tail call optimization, arguments are handled later.
1341  for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1342       i != e;
1343       ++i, ++realArgIdx) {
1344    CCValAssign &VA = ArgLocs[i];
1345    SDValue Arg = OutVals[realArgIdx];
1346    ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1347    bool isByVal = Flags.isByVal();
1348
1349    // Promote the value if needed.
1350    switch (VA.getLocInfo()) {
1351    default: llvm_unreachable("Unknown loc info!");
1352    case CCValAssign::Full: break;
1353    case CCValAssign::SExt:
1354      Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1355      break;
1356    case CCValAssign::ZExt:
1357      Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1358      break;
1359    case CCValAssign::AExt:
1360      Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1361      break;
1362    case CCValAssign::BCvt:
1363      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1364      break;
1365    }
1366
1367    // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1368    if (VA.needsCustom()) {
1369      if (VA.getLocVT() == MVT::v2f64) {
1370        SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1371                                  DAG.getConstant(0, MVT::i32));
1372        SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1373                                  DAG.getConstant(1, MVT::i32));
1374
1375        PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1376                         VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1377
1378        VA = ArgLocs[++i]; // skip ahead to next loc
1379        if (VA.isRegLoc()) {
1380          PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1381                           VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1382        } else {
1383          assert(VA.isMemLoc());
1384
1385          MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1386                                                 dl, DAG, VA, Flags));
1387        }
1388      } else {
1389        PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1390                         StackPtr, MemOpChains, Flags);
1391      }
1392    } else if (VA.isRegLoc()) {
1393      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1394    } else if (isByVal) {
1395      assert(VA.isMemLoc());
1396      unsigned offset = 0;
1397
1398      // True if this byval aggregate will be split between registers
1399      // and memory.
1400      if (CCInfo.isFirstByValRegValid()) {
1401        EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1402        unsigned int i, j;
1403        for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1404          SDValue Const = DAG.getConstant(4*i, MVT::i32);
1405          SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1406          SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1407                                     MachinePointerInfo(),
1408                                     false, false, false, 0);
1409          MemOpChains.push_back(Load.getValue(1));
1410          RegsToPass.push_back(std::make_pair(j, Load));
1411        }
1412        offset = ARM::R4 - CCInfo.getFirstByValReg();
1413        CCInfo.clearFirstByValReg();
1414      }
1415
1416      unsigned LocMemOffset = VA.getLocMemOffset();
1417      SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418      SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1419                                StkPtrOff);
1420      SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1421      SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1422      SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1423                                         MVT::i32);
1424      MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1425                                          Flags.getByValAlign(),
1426                                          /*isVolatile=*/false,
1427                                          /*AlwaysInline=*/false,
1428                                          MachinePointerInfo(0),
1429                                          MachinePointerInfo(0)));
1430
1431    } else if (!IsSibCall) {
1432      assert(VA.isMemLoc());
1433
1434      MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1435                                             dl, DAG, VA, Flags));
1436    }
1437  }
1438
1439  if (!MemOpChains.empty())
1440    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1441                        &MemOpChains[0], MemOpChains.size());
1442
1443  // Build a sequence of copy-to-reg nodes chained together with token chain
1444  // and flag operands which copy the outgoing args into the appropriate regs.
1445  SDValue InFlag;
1446  // Tail call byval lowering might overwrite argument registers so in case of
1447  // tail call optimization the copies to registers are lowered later.
1448  if (!isTailCall)
1449    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1450      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1451                               RegsToPass[i].second, InFlag);
1452      InFlag = Chain.getValue(1);
1453    }
1454
1455  // For tail calls lower the arguments to the 'real' stack slot.
1456  if (isTailCall) {
1457    // Force all the incoming stack arguments to be loaded from the stack
1458    // before any new outgoing arguments are stored to the stack, because the
1459    // outgoing stack slots may alias the incoming argument stack slots, and
1460    // the alias isn't otherwise explicit. This is slightly more conservative
1461    // than necessary, because it means that each store effectively depends
1462    // on every argument instead of just those arguments it would clobber.
1463
1464    // Do not flag preceding copytoreg stuff together with the following stuff.
1465    InFlag = SDValue();
1466    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1467      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1468                               RegsToPass[i].second, InFlag);
1469      InFlag = Chain.getValue(1);
1470    }
1471    InFlag =SDValue();
1472  }
1473
1474  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1475  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1476  // node so that legalize doesn't hack it.
1477  bool isDirect = false;
1478  bool isARMFunc = false;
1479  bool isLocalARMFunc = false;
1480  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1481
1482  if (EnableARMLongCalls) {
1483    assert (getTargetMachine().getRelocationModel() == Reloc::Static
1484            && "long-calls with non-static relocation model!");
1485    // Handle a global address or an external symbol. If it's not one of
1486    // those, the target's already in a register, so we don't need to do
1487    // anything extra.
1488    if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1489      const GlobalValue *GV = G->getGlobal();
1490      // Create a constant pool entry for the callee address
1491      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1492      ARMConstantPoolValue *CPV =
1493        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1494
1495      // Get the address of the callee into a register
1496      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1497      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1498      Callee = DAG.getLoad(getPointerTy(), dl,
1499                           DAG.getEntryNode(), CPAddr,
1500                           MachinePointerInfo::getConstantPool(),
1501                           false, false, false, 0);
1502    } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1503      const char *Sym = S->getSymbol();
1504
1505      // Create a constant pool entry for the callee address
1506      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507      ARMConstantPoolValue *CPV =
1508        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1509                                      ARMPCLabelIndex, 0);
1510      // Get the address of the callee into a register
1511      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513      Callee = DAG.getLoad(getPointerTy(), dl,
1514                           DAG.getEntryNode(), CPAddr,
1515                           MachinePointerInfo::getConstantPool(),
1516                           false, false, false, 0);
1517    }
1518  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1519    const GlobalValue *GV = G->getGlobal();
1520    isDirect = true;
1521    bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1522    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1523                   getTargetMachine().getRelocationModel() != Reloc::Static;
1524    isARMFunc = !Subtarget->isThumb() || isStub;
1525    // ARM call to a local ARM function is predicable.
1526    isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1527    // tBX takes a register source operand.
1528    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1529      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1530      ARMConstantPoolValue *CPV =
1531        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1532      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1533      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1534      Callee = DAG.getLoad(getPointerTy(), dl,
1535                           DAG.getEntryNode(), CPAddr,
1536                           MachinePointerInfo::getConstantPool(),
1537                           false, false, false, 0);
1538      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1539      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1540                           getPointerTy(), Callee, PICLabel);
1541    } else {
1542      // On ELF targets for PIC code, direct calls should go through the PLT
1543      unsigned OpFlags = 0;
1544      if (Subtarget->isTargetELF() &&
1545                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1546        OpFlags = ARMII::MO_PLT;
1547      Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1548    }
1549  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1550    isDirect = true;
1551    bool isStub = Subtarget->isTargetDarwin() &&
1552                  getTargetMachine().getRelocationModel() != Reloc::Static;
1553    isARMFunc = !Subtarget->isThumb() || isStub;
1554    // tBX takes a register source operand.
1555    const char *Sym = S->getSymbol();
1556    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1557      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1558      ARMConstantPoolValue *CPV =
1559        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1560                                      ARMPCLabelIndex, 4);
1561      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1562      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1563      Callee = DAG.getLoad(getPointerTy(), dl,
1564                           DAG.getEntryNode(), CPAddr,
1565                           MachinePointerInfo::getConstantPool(),
1566                           false, false, false, 0);
1567      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1568      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1569                           getPointerTy(), Callee, PICLabel);
1570    } else {
1571      unsigned OpFlags = 0;
1572      // On ELF targets for PIC code, direct calls should go through the PLT
1573      if (Subtarget->isTargetELF() &&
1574                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1575        OpFlags = ARMII::MO_PLT;
1576      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1577    }
1578  }
1579
1580  // FIXME: handle tail calls differently.
1581  unsigned CallOpc;
1582  if (Subtarget->isThumb()) {
1583    if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1584      CallOpc = ARMISD::CALL_NOLINK;
1585    else if (doesNotRet && isDirect && !isARMFunc &&
1586             Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1587      // "mov lr, pc; b _foo" to avoid confusing the RSP
1588      CallOpc = ARMISD::CALL_NOLINK;
1589    else
1590      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1591  } else {
1592    if (!isDirect && !Subtarget->hasV5TOps()) {
1593      CallOpc = ARMISD::CALL_NOLINK;
1594    } else if (doesNotRet && isDirect && Subtarget->hasRAS())
1595      // "mov lr, pc; b _foo" to avoid confusing the RSP
1596      CallOpc = ARMISD::CALL_NOLINK;
1597    else
1598      CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1599  }
1600
1601  std::vector<SDValue> Ops;
1602  Ops.push_back(Chain);
1603  Ops.push_back(Callee);
1604
1605  // Add argument registers to the end of the list so that they are known live
1606  // into the call.
1607  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1608    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1609                                  RegsToPass[i].second.getValueType()));
1610
1611  // Add a register mask operand representing the call-preserved registers.
1612  const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1613  const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1614  assert(Mask && "Missing call preserved mask for calling convention");
1615  Ops.push_back(DAG.getRegisterMask(Mask));
1616
1617  if (InFlag.getNode())
1618    Ops.push_back(InFlag);
1619
1620  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1621  if (isTailCall)
1622    return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1623
1624  // Returns a chain and a flag for retval copy to use.
1625  Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1626  InFlag = Chain.getValue(1);
1627
1628  Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1629                             DAG.getIntPtrConstant(0, true), InFlag);
1630  if (!Ins.empty())
1631    InFlag = Chain.getValue(1);
1632
1633  // Handle result values, copying them out of physregs into vregs that we
1634  // return.
1635  return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1636                         dl, DAG, InVals);
1637}
1638
1639/// HandleByVal - Every parameter *after* a byval parameter is passed
1640/// on the stack.  Remember the next parameter register to allocate,
1641/// and then confiscate the rest of the parameter registers to insure
1642/// this.
1643void
1644llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1645  unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1646  assert((State->getCallOrPrologue() == Prologue ||
1647          State->getCallOrPrologue() == Call) &&
1648         "unhandled ParmContext");
1649  if ((!State->isFirstByValRegValid()) &&
1650      (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1651    State->setFirstByValReg(reg);
1652    // At a call site, a byval parameter that is split between
1653    // registers and memory needs its size truncated here.  In a
1654    // function prologue, such byval parameters are reassembled in
1655    // memory, and are not truncated.
1656    if (State->getCallOrPrologue() == Call) {
1657      unsigned excess = 4 * (ARM::R4 - reg);
1658      assert(size >= excess && "expected larger existing stack allocation");
1659      size -= excess;
1660    }
1661  }
1662  // Confiscate any remaining parameter registers to preclude their
1663  // assignment to subsequent parameters.
1664  while (State->AllocateReg(GPRArgRegs, 4))
1665    ;
1666}
1667
1668/// MatchingStackOffset - Return true if the given stack call argument is
1669/// already available in the same position (relatively) of the caller's
1670/// incoming argument stack.
1671static
1672bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1673                         MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1674                         const ARMInstrInfo *TII) {
1675  unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1676  int FI = INT_MAX;
1677  if (Arg.getOpcode() == ISD::CopyFromReg) {
1678    unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1679    if (!TargetRegisterInfo::isVirtualRegister(VR))
1680      return false;
1681    MachineInstr *Def = MRI->getVRegDef(VR);
1682    if (!Def)
1683      return false;
1684    if (!Flags.isByVal()) {
1685      if (!TII->isLoadFromStackSlot(Def, FI))
1686        return false;
1687    } else {
1688      return false;
1689    }
1690  } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1691    if (Flags.isByVal())
1692      // ByVal argument is passed in as a pointer but it's now being
1693      // dereferenced. e.g.
1694      // define @foo(%struct.X* %A) {
1695      //   tail call @bar(%struct.X* byval %A)
1696      // }
1697      return false;
1698    SDValue Ptr = Ld->getBasePtr();
1699    FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1700    if (!FINode)
1701      return false;
1702    FI = FINode->getIndex();
1703  } else
1704    return false;
1705
1706  assert(FI != INT_MAX);
1707  if (!MFI->isFixedObjectIndex(FI))
1708    return false;
1709  return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1710}
1711
1712/// IsEligibleForTailCallOptimization - Check whether the call is eligible
1713/// for tail call optimization. Targets which want to do tail call
1714/// optimization should implement this function.
1715bool
1716ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1717                                                     CallingConv::ID CalleeCC,
1718                                                     bool isVarArg,
1719                                                     bool isCalleeStructRet,
1720                                                     bool isCallerStructRet,
1721                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
1722                                    const SmallVectorImpl<SDValue> &OutVals,
1723                                    const SmallVectorImpl<ISD::InputArg> &Ins,
1724                                                     SelectionDAG& DAG) const {
1725  const Function *CallerF = DAG.getMachineFunction().getFunction();
1726  CallingConv::ID CallerCC = CallerF->getCallingConv();
1727  bool CCMatch = CallerCC == CalleeCC;
1728
1729  // Look for obvious safe cases to perform tail call optimization that do not
1730  // require ABI changes. This is what gcc calls sibcall.
1731
1732  // Do not sibcall optimize vararg calls unless the call site is not passing
1733  // any arguments.
1734  if (isVarArg && !Outs.empty())
1735    return false;
1736
1737  // Also avoid sibcall optimization if either caller or callee uses struct
1738  // return semantics.
1739  if (isCalleeStructRet || isCallerStructRet)
1740    return false;
1741
1742  // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1743  // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1744  // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1745  // support in the assembler and linker to be used. This would need to be
1746  // fixed to fully support tail calls in Thumb1.
1747  //
1748  // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1749  // LR.  This means if we need to reload LR, it takes an extra instructions,
1750  // which outweighs the value of the tail call; but here we don't know yet
1751  // whether LR is going to be used.  Probably the right approach is to
1752  // generate the tail call here and turn it back into CALL/RET in
1753  // emitEpilogue if LR is used.
1754
1755  // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1756  // but we need to make sure there are enough registers; the only valid
1757  // registers are the 4 used for parameters.  We don't currently do this
1758  // case.
1759  if (Subtarget->isThumb1Only())
1760    return false;
1761
1762  // If the calling conventions do not match, then we'd better make sure the
1763  // results are returned in the same way as what the caller expects.
1764  if (!CCMatch) {
1765    SmallVector<CCValAssign, 16> RVLocs1;
1766    ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1767                       getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1768    CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1769
1770    SmallVector<CCValAssign, 16> RVLocs2;
1771    ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1772                       getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1773    CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1774
1775    if (RVLocs1.size() != RVLocs2.size())
1776      return false;
1777    for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1778      if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1779        return false;
1780      if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1781        return false;
1782      if (RVLocs1[i].isRegLoc()) {
1783        if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1784          return false;
1785      } else {
1786        if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1787          return false;
1788      }
1789    }
1790  }
1791
1792  // If the callee takes no arguments then go on to check the results of the
1793  // call.
1794  if (!Outs.empty()) {
1795    // Check if stack adjustment is needed. For now, do not do this if any
1796    // argument is passed on the stack.
1797    SmallVector<CCValAssign, 16> ArgLocs;
1798    ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1799                      getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1800    CCInfo.AnalyzeCallOperands(Outs,
1801                               CCAssignFnForNode(CalleeCC, false, isVarArg));
1802    if (CCInfo.getNextStackOffset()) {
1803      MachineFunction &MF = DAG.getMachineFunction();
1804
1805      // Check if the arguments are already laid out in the right way as
1806      // the caller's fixed stack objects.
1807      MachineFrameInfo *MFI = MF.getFrameInfo();
1808      const MachineRegisterInfo *MRI = &MF.getRegInfo();
1809      const ARMInstrInfo *TII =
1810        ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1811      for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1812           i != e;
1813           ++i, ++realArgIdx) {
1814        CCValAssign &VA = ArgLocs[i];
1815        EVT RegVT = VA.getLocVT();
1816        SDValue Arg = OutVals[realArgIdx];
1817        ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1818        if (VA.getLocInfo() == CCValAssign::Indirect)
1819          return false;
1820        if (VA.needsCustom()) {
1821          // f64 and vector types are split into multiple registers or
1822          // register/stack-slot combinations.  The types will not match
1823          // the registers; give up on memory f64 refs until we figure
1824          // out what to do about this.
1825          if (!VA.isRegLoc())
1826            return false;
1827          if (!ArgLocs[++i].isRegLoc())
1828            return false;
1829          if (RegVT == MVT::v2f64) {
1830            if (!ArgLocs[++i].isRegLoc())
1831              return false;
1832            if (!ArgLocs[++i].isRegLoc())
1833              return false;
1834          }
1835        } else if (!VA.isRegLoc()) {
1836          if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1837                                   MFI, MRI, TII))
1838            return false;
1839        }
1840      }
1841    }
1842  }
1843
1844  return true;
1845}
1846
1847SDValue
1848ARMTargetLowering::LowerReturn(SDValue Chain,
1849                               CallingConv::ID CallConv, bool isVarArg,
1850                               const SmallVectorImpl<ISD::OutputArg> &Outs,
1851                               const SmallVectorImpl<SDValue> &OutVals,
1852                               DebugLoc dl, SelectionDAG &DAG) const {
1853
1854  // CCValAssign - represent the assignment of the return value to a location.
1855  SmallVector<CCValAssign, 16> RVLocs;
1856
1857  // CCState - Info about the registers and stack slots.
1858  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1859                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1860
1861  // Analyze outgoing return values.
1862  CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1863                                               isVarArg));
1864
1865  // If this is the first return lowered for this function, add
1866  // the regs to the liveout set for the function.
1867  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1868    for (unsigned i = 0; i != RVLocs.size(); ++i)
1869      if (RVLocs[i].isRegLoc())
1870        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1871  }
1872
1873  SDValue Flag;
1874
1875  // Copy the result values into the output registers.
1876  for (unsigned i = 0, realRVLocIdx = 0;
1877       i != RVLocs.size();
1878       ++i, ++realRVLocIdx) {
1879    CCValAssign &VA = RVLocs[i];
1880    assert(VA.isRegLoc() && "Can only return in registers!");
1881
1882    SDValue Arg = OutVals[realRVLocIdx];
1883
1884    switch (VA.getLocInfo()) {
1885    default: llvm_unreachable("Unknown loc info!");
1886    case CCValAssign::Full: break;
1887    case CCValAssign::BCvt:
1888      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1889      break;
1890    }
1891
1892    if (VA.needsCustom()) {
1893      if (VA.getLocVT() == MVT::v2f64) {
1894        // Extract the first half and return it in two registers.
1895        SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1896                                   DAG.getConstant(0, MVT::i32));
1897        SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1898                                       DAG.getVTList(MVT::i32, MVT::i32), Half);
1899
1900        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1901        Flag = Chain.getValue(1);
1902        VA = RVLocs[++i]; // skip ahead to next loc
1903        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1904                                 HalfGPRs.getValue(1), Flag);
1905        Flag = Chain.getValue(1);
1906        VA = RVLocs[++i]; // skip ahead to next loc
1907
1908        // Extract the 2nd half and fall through to handle it as an f64 value.
1909        Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1910                          DAG.getConstant(1, MVT::i32));
1911      }
1912      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
1913      // available.
1914      SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1915                                  DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1916      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1917      Flag = Chain.getValue(1);
1918      VA = RVLocs[++i]; // skip ahead to next loc
1919      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1920                               Flag);
1921    } else
1922      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1923
1924    // Guarantee that all emitted copies are
1925    // stuck together, avoiding something bad.
1926    Flag = Chain.getValue(1);
1927  }
1928
1929  SDValue result;
1930  if (Flag.getNode())
1931    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1932  else // Return Void
1933    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1934
1935  return result;
1936}
1937
1938bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1939  if (N->getNumValues() != 1)
1940    return false;
1941  if (!N->hasNUsesOfValue(1, 0))
1942    return false;
1943
1944  unsigned NumCopies = 0;
1945  SDNode* Copies[2] = { 0, 0 };
1946  SDNode *Use = *N->use_begin();
1947  if (Use->getOpcode() == ISD::CopyToReg) {
1948    Copies[NumCopies++] = Use;
1949  } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1950    // f64 returned in a pair of GPRs.
1951    for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1952         UI != UE; ++UI) {
1953      if (UI->getOpcode() != ISD::CopyToReg)
1954        return false;
1955      Copies[UI.getUse().getResNo()] = *UI;
1956      ++NumCopies;
1957    }
1958  } else if (Use->getOpcode() == ISD::BITCAST) {
1959    // f32 returned in a single GPR.
1960    if (!Use->hasNUsesOfValue(1, 0))
1961      return false;
1962    Use = *Use->use_begin();
1963    if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1964      return false;
1965    Copies[NumCopies++] = Use;
1966  } else {
1967    return false;
1968  }
1969
1970  if (NumCopies != 1 && NumCopies != 2)
1971    return false;
1972
1973  bool HasRet = false;
1974  for (unsigned i = 0; i < NumCopies; ++i) {
1975    SDNode *Copy = Copies[i];
1976    for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1977         UI != UE; ++UI) {
1978      if (UI->getOpcode() == ISD::CopyToReg) {
1979        SDNode *Use = *UI;
1980        if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1981          continue;
1982        return false;
1983      }
1984      if (UI->getOpcode() != ARMISD::RET_FLAG)
1985        return false;
1986      HasRet = true;
1987    }
1988  }
1989
1990  return HasRet;
1991}
1992
1993bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1994  if (!EnableARMTailCalls)
1995    return false;
1996
1997  if (!CI->isTailCall())
1998    return false;
1999
2000  return !Subtarget->isThumb1Only();
2001}
2002
2003// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2004// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2005// one of the above mentioned nodes. It has to be wrapped because otherwise
2006// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2007// be used to form addressing mode. These wrapped nodes will be selected
2008// into MOVi.
2009static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2010  EVT PtrVT = Op.getValueType();
2011  // FIXME there is no actual debug info here
2012  DebugLoc dl = Op.getDebugLoc();
2013  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2014  SDValue Res;
2015  if (CP->isMachineConstantPoolEntry())
2016    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2017                                    CP->getAlignment());
2018  else
2019    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2020                                    CP->getAlignment());
2021  return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2022}
2023
2024unsigned ARMTargetLowering::getJumpTableEncoding() const {
2025  return MachineJumpTableInfo::EK_Inline;
2026}
2027
2028SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2029                                             SelectionDAG &DAG) const {
2030  MachineFunction &MF = DAG.getMachineFunction();
2031  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2032  unsigned ARMPCLabelIndex = 0;
2033  DebugLoc DL = Op.getDebugLoc();
2034  EVT PtrVT = getPointerTy();
2035  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2036  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2037  SDValue CPAddr;
2038  if (RelocM == Reloc::Static) {
2039    CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2040  } else {
2041    unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2042    ARMPCLabelIndex = AFI->createPICLabelUId();
2043    ARMConstantPoolValue *CPV =
2044      ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2045                                      ARMCP::CPBlockAddress, PCAdj);
2046    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2047  }
2048  CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2049  SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2050                               MachinePointerInfo::getConstantPool(),
2051                               false, false, false, 0);
2052  if (RelocM == Reloc::Static)
2053    return Result;
2054  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2055  return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2056}
2057
2058// Lower ISD::GlobalTLSAddress using the "general dynamic" model
2059SDValue
2060ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2061                                                 SelectionDAG &DAG) const {
2062  DebugLoc dl = GA->getDebugLoc();
2063  EVT PtrVT = getPointerTy();
2064  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2065  MachineFunction &MF = DAG.getMachineFunction();
2066  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2067  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2068  ARMConstantPoolValue *CPV =
2069    ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2070                                    ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2071  SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2072  Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2073  Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2074                         MachinePointerInfo::getConstantPool(),
2075                         false, false, false, 0);
2076  SDValue Chain = Argument.getValue(1);
2077
2078  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2079  Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2080
2081  // call __tls_get_addr.
2082  ArgListTy Args;
2083  ArgListEntry Entry;
2084  Entry.Node = Argument;
2085  Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2086  Args.push_back(Entry);
2087  // FIXME: is there useful debug info available here?
2088  std::pair<SDValue, SDValue> CallResult =
2089    LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2090                false, false, false, false,
2091                0, CallingConv::C, /*isTailCall=*/false,
2092                /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2093                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2094  return CallResult.first;
2095}
2096
2097// Lower ISD::GlobalTLSAddress using the "initial exec" or
2098// "local exec" model.
2099SDValue
2100ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2101                                        SelectionDAG &DAG) const {
2102  const GlobalValue *GV = GA->getGlobal();
2103  DebugLoc dl = GA->getDebugLoc();
2104  SDValue Offset;
2105  SDValue Chain = DAG.getEntryNode();
2106  EVT PtrVT = getPointerTy();
2107  // Get the Thread Pointer
2108  SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2109
2110  if (GV->isDeclaration()) {
2111    MachineFunction &MF = DAG.getMachineFunction();
2112    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2113    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2114    // Initial exec model.
2115    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2116    ARMConstantPoolValue *CPV =
2117      ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2118                                      ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2119                                      true);
2120    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2121    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2122    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2123                         MachinePointerInfo::getConstantPool(),
2124                         false, false, false, 0);
2125    Chain = Offset.getValue(1);
2126
2127    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2128    Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2129
2130    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2131                         MachinePointerInfo::getConstantPool(),
2132                         false, false, false, 0);
2133  } else {
2134    // local exec model
2135    ARMConstantPoolValue *CPV =
2136      ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2137    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2138    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2139    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2140                         MachinePointerInfo::getConstantPool(),
2141                         false, false, false, 0);
2142  }
2143
2144  // The address of the thread local variable is the add of the thread
2145  // pointer with the offset of the variable.
2146  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2147}
2148
2149SDValue
2150ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2151  // TODO: implement the "local dynamic" model
2152  assert(Subtarget->isTargetELF() &&
2153         "TLS not implemented for non-ELF targets");
2154  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2155  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2156  // otherwise use the "Local Exec" TLS Model
2157  if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2158    return LowerToTLSGeneralDynamicModel(GA, DAG);
2159  else
2160    return LowerToTLSExecModels(GA, DAG);
2161}
2162
2163SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2164                                                 SelectionDAG &DAG) const {
2165  EVT PtrVT = getPointerTy();
2166  DebugLoc dl = Op.getDebugLoc();
2167  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2168  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2169  if (RelocM == Reloc::PIC_) {
2170    bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2171    ARMConstantPoolValue *CPV =
2172      ARMConstantPoolConstant::Create(GV,
2173                                      UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2174    SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2175    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2176    SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2177                                 CPAddr,
2178                                 MachinePointerInfo::getConstantPool(),
2179                                 false, false, false, 0);
2180    SDValue Chain = Result.getValue(1);
2181    SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2182    Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2183    if (!UseGOTOFF)
2184      Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2185                           MachinePointerInfo::getGOT(),
2186                           false, false, false, 0);
2187    return Result;
2188  }
2189
2190  // If we have T2 ops, we can materialize the address directly via movt/movw
2191  // pair. This is always cheaper.
2192  if (Subtarget->useMovt()) {
2193    ++NumMovwMovt;
2194    // FIXME: Once remat is capable of dealing with instructions with register
2195    // operands, expand this into two nodes.
2196    return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2197                       DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2198  } else {
2199    SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2200    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2201    return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2202                       MachinePointerInfo::getConstantPool(),
2203                       false, false, false, 0);
2204  }
2205}
2206
2207SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2208                                                    SelectionDAG &DAG) const {
2209  EVT PtrVT = getPointerTy();
2210  DebugLoc dl = Op.getDebugLoc();
2211  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2212  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2213  MachineFunction &MF = DAG.getMachineFunction();
2214  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2215
2216  // FIXME: Enable this for static codegen when tool issues are fixed.  Also
2217  // update ARMFastISel::ARMMaterializeGV.
2218  if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2219    ++NumMovwMovt;
2220    // FIXME: Once remat is capable of dealing with instructions with register
2221    // operands, expand this into two nodes.
2222    if (RelocM == Reloc::Static)
2223      return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2224                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2225
2226    unsigned Wrapper = (RelocM == Reloc::PIC_)
2227      ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2228    SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2229                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2230    if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2231      Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2232                           MachinePointerInfo::getGOT(),
2233                           false, false, false, 0);
2234    return Result;
2235  }
2236
2237  unsigned ARMPCLabelIndex = 0;
2238  SDValue CPAddr;
2239  if (RelocM == Reloc::Static) {
2240    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2241  } else {
2242    ARMPCLabelIndex = AFI->createPICLabelUId();
2243    unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2244    ARMConstantPoolValue *CPV =
2245      ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2246                                      PCAdj);
2247    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2248  }
2249  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2250
2251  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2252                               MachinePointerInfo::getConstantPool(),
2253                               false, false, false, 0);
2254  SDValue Chain = Result.getValue(1);
2255
2256  if (RelocM == Reloc::PIC_) {
2257    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2258    Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2259  }
2260
2261  if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2262    Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2263                         false, false, false, 0);
2264
2265  return Result;
2266}
2267
2268SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2269                                                    SelectionDAG &DAG) const {
2270  assert(Subtarget->isTargetELF() &&
2271         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2272  MachineFunction &MF = DAG.getMachineFunction();
2273  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2274  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2275  EVT PtrVT = getPointerTy();
2276  DebugLoc dl = Op.getDebugLoc();
2277  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2278  ARMConstantPoolValue *CPV =
2279    ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2280                                  ARMPCLabelIndex, PCAdj);
2281  SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2282  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2283  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2284                               MachinePointerInfo::getConstantPool(),
2285                               false, false, false, 0);
2286  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2287  return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2288}
2289
2290SDValue
2291ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2292  DebugLoc dl = Op.getDebugLoc();
2293  SDValue Val = DAG.getConstant(0, MVT::i32);
2294  return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2295                     DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2296                     Op.getOperand(1), Val);
2297}
2298
2299SDValue
2300ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2301  DebugLoc dl = Op.getDebugLoc();
2302  return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2303                     Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2304}
2305
2306SDValue
2307ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2308                                          const ARMSubtarget *Subtarget) const {
2309  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2310  DebugLoc dl = Op.getDebugLoc();
2311  switch (IntNo) {
2312  default: return SDValue();    // Don't custom lower most intrinsics.
2313  case Intrinsic::arm_thread_pointer: {
2314    EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2315    return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2316  }
2317  case Intrinsic::eh_sjlj_lsda: {
2318    MachineFunction &MF = DAG.getMachineFunction();
2319    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2320    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2321    EVT PtrVT = getPointerTy();
2322    DebugLoc dl = Op.getDebugLoc();
2323    Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2324    SDValue CPAddr;
2325    unsigned PCAdj = (RelocM != Reloc::PIC_)
2326      ? 0 : (Subtarget->isThumb() ? 4 : 8);
2327    ARMConstantPoolValue *CPV =
2328      ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2329                                      ARMCP::CPLSDA, PCAdj);
2330    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2331    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2332    SDValue Result =
2333      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2334                  MachinePointerInfo::getConstantPool(),
2335                  false, false, false, 0);
2336
2337    if (RelocM == Reloc::PIC_) {
2338      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2339      Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2340    }
2341    return Result;
2342  }
2343  case Intrinsic::arm_neon_vmulls:
2344  case Intrinsic::arm_neon_vmullu: {
2345    unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2346      ? ARMISD::VMULLs : ARMISD::VMULLu;
2347    return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2348                       Op.getOperand(1), Op.getOperand(2));
2349  }
2350  }
2351}
2352
2353static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2354                               const ARMSubtarget *Subtarget) {
2355  DebugLoc dl = Op.getDebugLoc();
2356  if (!Subtarget->hasDataBarrier()) {
2357    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2358    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2359    // here.
2360    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2361           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2362    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2363                       DAG.getConstant(0, MVT::i32));
2364  }
2365
2366  SDValue Op5 = Op.getOperand(5);
2367  bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2368  unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2369  unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2370  bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2371
2372  ARM_MB::MemBOpt DMBOpt;
2373  if (isDeviceBarrier)
2374    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2375  else
2376    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2377  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2378                     DAG.getConstant(DMBOpt, MVT::i32));
2379}
2380
2381
2382static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2383                                 const ARMSubtarget *Subtarget) {
2384  // FIXME: handle "fence singlethread" more efficiently.
2385  DebugLoc dl = Op.getDebugLoc();
2386  if (!Subtarget->hasDataBarrier()) {
2387    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2388    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2389    // here.
2390    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2391           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2392    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2393                       DAG.getConstant(0, MVT::i32));
2394  }
2395
2396  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2397                     DAG.getConstant(ARM_MB::ISH, MVT::i32));
2398}
2399
2400static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2401                             const ARMSubtarget *Subtarget) {
2402  // ARM pre v5TE and Thumb1 does not have preload instructions.
2403  if (!(Subtarget->isThumb2() ||
2404        (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2405    // Just preserve the chain.
2406    return Op.getOperand(0);
2407
2408  DebugLoc dl = Op.getDebugLoc();
2409  unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2410  if (!isRead &&
2411      (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2412    // ARMv7 with MP extension has PLDW.
2413    return Op.getOperand(0);
2414
2415  unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2416  if (Subtarget->isThumb()) {
2417    // Invert the bits.
2418    isRead = ~isRead & 1;
2419    isData = ~isData & 1;
2420  }
2421
2422  return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2423                     Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2424                     DAG.getConstant(isData, MVT::i32));
2425}
2426
2427static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2428  MachineFunction &MF = DAG.getMachineFunction();
2429  ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2430
2431  // vastart just stores the address of the VarArgsFrameIndex slot into the
2432  // memory location argument.
2433  DebugLoc dl = Op.getDebugLoc();
2434  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2435  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2436  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2437  return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2438                      MachinePointerInfo(SV), false, false, 0);
2439}
2440
2441SDValue
2442ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2443                                        SDValue &Root, SelectionDAG &DAG,
2444                                        DebugLoc dl) const {
2445  MachineFunction &MF = DAG.getMachineFunction();
2446  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2447
2448  const TargetRegisterClass *RC;
2449  if (AFI->isThumb1OnlyFunction())
2450    RC = ARM::tGPRRegisterClass;
2451  else
2452    RC = ARM::GPRRegisterClass;
2453
2454  // Transform the arguments stored in physical registers into virtual ones.
2455  unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2456  SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2457
2458  SDValue ArgValue2;
2459  if (NextVA.isMemLoc()) {
2460    MachineFrameInfo *MFI = MF.getFrameInfo();
2461    int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2462
2463    // Create load node to retrieve arguments from the stack.
2464    SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2465    ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2466                            MachinePointerInfo::getFixedStack(FI),
2467                            false, false, false, 0);
2468  } else {
2469    Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2470    ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2471  }
2472
2473  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2474}
2475
2476void
2477ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2478                                  unsigned &VARegSize, unsigned &VARegSaveSize)
2479  const {
2480  unsigned NumGPRs;
2481  if (CCInfo.isFirstByValRegValid())
2482    NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2483  else {
2484    unsigned int firstUnalloced;
2485    firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2486                                                sizeof(GPRArgRegs) /
2487                                                sizeof(GPRArgRegs[0]));
2488    NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2489  }
2490
2491  unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2492  VARegSize = NumGPRs * 4;
2493  VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2494}
2495
2496// The remaining GPRs hold either the beginning of variable-argument
2497// data, or the beginning of an aggregate passed by value (usuall
2498// byval).  Either way, we allocate stack slots adjacent to the data
2499// provided by our caller, and store the unallocated registers there.
2500// If this is a variadic function, the va_list pointer will begin with
2501// these values; otherwise, this reassembles a (byval) structure that
2502// was split between registers and memory.
2503void
2504ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2505                                        DebugLoc dl, SDValue &Chain,
2506                                        unsigned ArgOffset) const {
2507  MachineFunction &MF = DAG.getMachineFunction();
2508  MachineFrameInfo *MFI = MF.getFrameInfo();
2509  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2510  unsigned firstRegToSaveIndex;
2511  if (CCInfo.isFirstByValRegValid())
2512    firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2513  else {
2514    firstRegToSaveIndex = CCInfo.getFirstUnallocated
2515      (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2516  }
2517
2518  unsigned VARegSize, VARegSaveSize;
2519  computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2520  if (VARegSaveSize) {
2521    // If this function is vararg, store any remaining integer argument regs
2522    // to their spots on the stack so that they may be loaded by deferencing
2523    // the result of va_next.
2524    AFI->setVarArgsRegSaveSize(VARegSaveSize);
2525    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2526                                                     ArgOffset + VARegSaveSize
2527                                                     - VARegSize,
2528                                                     false));
2529    SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2530                                    getPointerTy());
2531
2532    SmallVector<SDValue, 4> MemOps;
2533    for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2534      const TargetRegisterClass *RC;
2535      if (AFI->isThumb1OnlyFunction())
2536        RC = ARM::tGPRRegisterClass;
2537      else
2538        RC = ARM::GPRRegisterClass;
2539
2540      unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2541      SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2542      SDValue Store =
2543        DAG.getStore(Val.getValue(1), dl, Val, FIN,
2544                 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2545                     false, false, 0);
2546      MemOps.push_back(Store);
2547      FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2548                        DAG.getConstant(4, getPointerTy()));
2549    }
2550    if (!MemOps.empty())
2551      Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2552                          &MemOps[0], MemOps.size());
2553  } else
2554    // This will point to the next argument passed via stack.
2555    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2556}
2557
2558SDValue
2559ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2560                                        CallingConv::ID CallConv, bool isVarArg,
2561                                        const SmallVectorImpl<ISD::InputArg>
2562                                          &Ins,
2563                                        DebugLoc dl, SelectionDAG &DAG,
2564                                        SmallVectorImpl<SDValue> &InVals)
2565                                          const {
2566  MachineFunction &MF = DAG.getMachineFunction();
2567  MachineFrameInfo *MFI = MF.getFrameInfo();
2568
2569  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2570
2571  // Assign locations to all of the incoming arguments.
2572  SmallVector<CCValAssign, 16> ArgLocs;
2573  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2574                    getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2575  CCInfo.AnalyzeFormalArguments(Ins,
2576                                CCAssignFnForNode(CallConv, /* Return*/ false,
2577                                                  isVarArg));
2578
2579  SmallVector<SDValue, 16> ArgValues;
2580  int lastInsIndex = -1;
2581
2582  SDValue ArgValue;
2583  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2584    CCValAssign &VA = ArgLocs[i];
2585
2586    // Arguments stored in registers.
2587    if (VA.isRegLoc()) {
2588      EVT RegVT = VA.getLocVT();
2589
2590      if (VA.needsCustom()) {
2591        // f64 and vector types are split up into multiple registers or
2592        // combinations of registers and stack slots.
2593        if (VA.getLocVT() == MVT::v2f64) {
2594          SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2595                                                   Chain, DAG, dl);
2596          VA = ArgLocs[++i]; // skip ahead to next loc
2597          SDValue ArgValue2;
2598          if (VA.isMemLoc()) {
2599            int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2600            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2601            ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2602                                    MachinePointerInfo::getFixedStack(FI),
2603                                    false, false, false, 0);
2604          } else {
2605            ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2606                                             Chain, DAG, dl);
2607          }
2608          ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2609          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2610                                 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2611          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2612                                 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2613        } else
2614          ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2615
2616      } else {
2617        const TargetRegisterClass *RC;
2618
2619        if (RegVT == MVT::f32)
2620          RC = ARM::SPRRegisterClass;
2621        else if (RegVT == MVT::f64)
2622          RC = ARM::DPRRegisterClass;
2623        else if (RegVT == MVT::v2f64)
2624          RC = ARM::QPRRegisterClass;
2625        else if (RegVT == MVT::i32)
2626          RC = (AFI->isThumb1OnlyFunction() ?
2627                ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2628        else
2629          llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2630
2631        // Transform the arguments in physical registers into virtual ones.
2632        unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2633        ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2634      }
2635
2636      // If this is an 8 or 16-bit value, it is really passed promoted
2637      // to 32 bits.  Insert an assert[sz]ext to capture this, then
2638      // truncate to the right size.
2639      switch (VA.getLocInfo()) {
2640      default: llvm_unreachable("Unknown loc info!");
2641      case CCValAssign::Full: break;
2642      case CCValAssign::BCvt:
2643        ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2644        break;
2645      case CCValAssign::SExt:
2646        ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2647                               DAG.getValueType(VA.getValVT()));
2648        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2649        break;
2650      case CCValAssign::ZExt:
2651        ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2652                               DAG.getValueType(VA.getValVT()));
2653        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2654        break;
2655      }
2656
2657      InVals.push_back(ArgValue);
2658
2659    } else { // VA.isRegLoc()
2660
2661      // sanity check
2662      assert(VA.isMemLoc());
2663      assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2664
2665      int index = ArgLocs[i].getValNo();
2666
2667      // Some Ins[] entries become multiple ArgLoc[] entries.
2668      // Process them only once.
2669      if (index != lastInsIndex)
2670        {
2671          ISD::ArgFlagsTy Flags = Ins[index].Flags;
2672          // FIXME: For now, all byval parameter objects are marked mutable.
2673          // This can be changed with more analysis.
2674          // In case of tail call optimization mark all arguments mutable.
2675          // Since they could be overwritten by lowering of arguments in case of
2676          // a tail call.
2677          if (Flags.isByVal()) {
2678            unsigned VARegSize, VARegSaveSize;
2679            computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2680            VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2681            unsigned Bytes = Flags.getByValSize() - VARegSize;
2682            if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2683            int FI = MFI->CreateFixedObject(Bytes,
2684                                            VA.getLocMemOffset(), false);
2685            InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2686          } else {
2687            int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2688                                            VA.getLocMemOffset(), true);
2689
2690            // Create load nodes to retrieve arguments from the stack.
2691            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2692            InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2693                                         MachinePointerInfo::getFixedStack(FI),
2694                                         false, false, false, 0));
2695          }
2696          lastInsIndex = index;
2697        }
2698    }
2699  }
2700
2701  // varargs
2702  if (isVarArg)
2703    VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2704
2705  return Chain;
2706}
2707
2708/// isFloatingPointZero - Return true if this is +0.0.
2709static bool isFloatingPointZero(SDValue Op) {
2710  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2711    return CFP->getValueAPF().isPosZero();
2712  else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2713    // Maybe this has already been legalized into the constant pool?
2714    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2715      SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2716      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2717        if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2718          return CFP->getValueAPF().isPosZero();
2719    }
2720  }
2721  return false;
2722}
2723
2724/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2725/// the given operands.
2726SDValue
2727ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2728                             SDValue &ARMcc, SelectionDAG &DAG,
2729                             DebugLoc dl) const {
2730  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2731    unsigned C = RHSC->getZExtValue();
2732    if (!isLegalICmpImmediate(C)) {
2733      // Constant does not fit, try adjusting it by one?
2734      switch (CC) {
2735      default: break;
2736      case ISD::SETLT:
2737      case ISD::SETGE:
2738        if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2739          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2740          RHS = DAG.getConstant(C-1, MVT::i32);
2741        }
2742        break;
2743      case ISD::SETULT:
2744      case ISD::SETUGE:
2745        if (C != 0 && isLegalICmpImmediate(C-1)) {
2746          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2747          RHS = DAG.getConstant(C-1, MVT::i32);
2748        }
2749        break;
2750      case ISD::SETLE:
2751      case ISD::SETGT:
2752        if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2753          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2754          RHS = DAG.getConstant(C+1, MVT::i32);
2755        }
2756        break;
2757      case ISD::SETULE:
2758      case ISD::SETUGT:
2759        if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2760          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2761          RHS = DAG.getConstant(C+1, MVT::i32);
2762        }
2763        break;
2764      }
2765    }
2766  }
2767
2768  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2769  ARMISD::NodeType CompareType;
2770  switch (CondCode) {
2771  default:
2772    CompareType = ARMISD::CMP;
2773    break;
2774  case ARMCC::EQ:
2775  case ARMCC::NE:
2776    // Uses only Z Flag
2777    CompareType = ARMISD::CMPZ;
2778    break;
2779  }
2780  ARMcc = DAG.getConstant(CondCode, MVT::i32);
2781  return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2782}
2783
2784/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2785SDValue
2786ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2787                             DebugLoc dl) const {
2788  SDValue Cmp;
2789  if (!isFloatingPointZero(RHS))
2790    Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2791  else
2792    Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2793  return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2794}
2795
2796/// duplicateCmp - Glue values can have only one use, so this function
2797/// duplicates a comparison node.
2798SDValue
2799ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2800  unsigned Opc = Cmp.getOpcode();
2801  DebugLoc DL = Cmp.getDebugLoc();
2802  if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2803    return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2804
2805  assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2806  Cmp = Cmp.getOperand(0);
2807  Opc = Cmp.getOpcode();
2808  if (Opc == ARMISD::CMPFP)
2809    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2810  else {
2811    assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2812    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2813  }
2814  return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2815}
2816
2817SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2818  SDValue Cond = Op.getOperand(0);
2819  SDValue SelectTrue = Op.getOperand(1);
2820  SDValue SelectFalse = Op.getOperand(2);
2821  DebugLoc dl = Op.getDebugLoc();
2822
2823  // Convert:
2824  //
2825  //   (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2826  //   (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2827  //
2828  if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2829    const ConstantSDNode *CMOVTrue =
2830      dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2831    const ConstantSDNode *CMOVFalse =
2832      dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2833
2834    if (CMOVTrue && CMOVFalse) {
2835      unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2836      unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2837
2838      SDValue True;
2839      SDValue False;
2840      if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2841        True = SelectTrue;
2842        False = SelectFalse;
2843      } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2844        True = SelectFalse;
2845        False = SelectTrue;
2846      }
2847
2848      if (True.getNode() && False.getNode()) {
2849        EVT VT = Op.getValueType();
2850        SDValue ARMcc = Cond.getOperand(2);
2851        SDValue CCR = Cond.getOperand(3);
2852        SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2853        assert(True.getValueType() == VT);
2854        return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2855      }
2856    }
2857  }
2858
2859  // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2860  // undefined bits before doing a full-word comparison with zero.
2861  Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2862                     DAG.getConstant(1, Cond.getValueType()));
2863
2864  return DAG.getSelectCC(dl, Cond,
2865                         DAG.getConstant(0, Cond.getValueType()),
2866                         SelectTrue, SelectFalse, ISD::SETNE);
2867}
2868
2869SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2870  EVT VT = Op.getValueType();
2871  SDValue LHS = Op.getOperand(0);
2872  SDValue RHS = Op.getOperand(1);
2873  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2874  SDValue TrueVal = Op.getOperand(2);
2875  SDValue FalseVal = Op.getOperand(3);
2876  DebugLoc dl = Op.getDebugLoc();
2877
2878  if (LHS.getValueType() == MVT::i32) {
2879    SDValue ARMcc;
2880    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2881    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2882    return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2883  }
2884
2885  ARMCC::CondCodes CondCode, CondCode2;
2886  FPCCToARMCC(CC, CondCode, CondCode2);
2887
2888  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2889  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2890  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2891  SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2892                               ARMcc, CCR, Cmp);
2893  if (CondCode2 != ARMCC::AL) {
2894    SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2895    // FIXME: Needs another CMP because flag can have but one use.
2896    SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2897    Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2898                         Result, TrueVal, ARMcc2, CCR, Cmp2);
2899  }
2900  return Result;
2901}
2902
2903/// canChangeToInt - Given the fp compare operand, return true if it is suitable
2904/// to morph to an integer compare sequence.
2905static bool canChangeToInt(SDValue Op, bool &SeenZero,
2906                           const ARMSubtarget *Subtarget) {
2907  SDNode *N = Op.getNode();
2908  if (!N->hasOneUse())
2909    // Otherwise it requires moving the value from fp to integer registers.
2910    return false;
2911  if (!N->getNumValues())
2912    return false;
2913  EVT VT = Op.getValueType();
2914  if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2915    // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2916    // vmrs are very slow, e.g. cortex-a8.
2917    return false;
2918
2919  if (isFloatingPointZero(Op)) {
2920    SeenZero = true;
2921    return true;
2922  }
2923  return ISD::isNormalLoad(N);
2924}
2925
2926static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2927  if (isFloatingPointZero(Op))
2928    return DAG.getConstant(0, MVT::i32);
2929
2930  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2931    return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2932                       Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2933                       Ld->isVolatile(), Ld->isNonTemporal(),
2934                       Ld->isInvariant(), Ld->getAlignment());
2935
2936  llvm_unreachable("Unknown VFP cmp argument!");
2937}
2938
2939static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2940                           SDValue &RetVal1, SDValue &RetVal2) {
2941  if (isFloatingPointZero(Op)) {
2942    RetVal1 = DAG.getConstant(0, MVT::i32);
2943    RetVal2 = DAG.getConstant(0, MVT::i32);
2944    return;
2945  }
2946
2947  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2948    SDValue Ptr = Ld->getBasePtr();
2949    RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2950                          Ld->getChain(), Ptr,
2951                          Ld->getPointerInfo(),
2952                          Ld->isVolatile(), Ld->isNonTemporal(),
2953                          Ld->isInvariant(), Ld->getAlignment());
2954
2955    EVT PtrType = Ptr.getValueType();
2956    unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2957    SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2958                                 PtrType, Ptr, DAG.getConstant(4, PtrType));
2959    RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2960                          Ld->getChain(), NewPtr,
2961                          Ld->getPointerInfo().getWithOffset(4),
2962                          Ld->isVolatile(), Ld->isNonTemporal(),
2963                          Ld->isInvariant(), NewAlign);
2964    return;
2965  }
2966
2967  llvm_unreachable("Unknown VFP cmp argument!");
2968}
2969
2970/// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2971/// f32 and even f64 comparisons to integer ones.
2972SDValue
2973ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2974  SDValue Chain = Op.getOperand(0);
2975  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2976  SDValue LHS = Op.getOperand(2);
2977  SDValue RHS = Op.getOperand(3);
2978  SDValue Dest = Op.getOperand(4);
2979  DebugLoc dl = Op.getDebugLoc();
2980
2981  bool LHSSeenZero = false;
2982  bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
2983  bool RHSSeenZero = false;
2984  bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
2985  if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
2986    // If unsafe fp math optimization is enabled and there are no other uses of
2987    // the CMP operands, and the condition code is EQ or NE, we can optimize it
2988    // to an integer comparison.
2989    if (CC == ISD::SETOEQ)
2990      CC = ISD::SETEQ;
2991    else if (CC == ISD::SETUNE)
2992      CC = ISD::SETNE;
2993
2994    SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
2995    SDValue ARMcc;
2996    if (LHS.getValueType() == MVT::f32) {
2997      LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2998                        bitcastf32Toi32(LHS, DAG), Mask);
2999      RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
3000                        bitcastf32Toi32(RHS, DAG), Mask);
3001      SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3002      SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3003      return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3004                         Chain, Dest, ARMcc, CCR, Cmp);
3005    }
3006
3007    SDValue LHS1, LHS2;
3008    SDValue RHS1, RHS2;
3009    expandf64Toi32(LHS, DAG, LHS1, LHS2);
3010    expandf64Toi32(RHS, DAG, RHS1, RHS2);
3011    LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3012    RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3013    ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3014    ARMcc = DAG.getConstant(CondCode, MVT::i32);
3015    SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3016    SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3017    return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3018  }
3019
3020  return SDValue();
3021}
3022
3023SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3024  SDValue Chain = Op.getOperand(0);
3025  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3026  SDValue LHS = Op.getOperand(2);
3027  SDValue RHS = Op.getOperand(3);
3028  SDValue Dest = Op.getOperand(4);
3029  DebugLoc dl = Op.getDebugLoc();
3030
3031  if (LHS.getValueType() == MVT::i32) {
3032    SDValue ARMcc;
3033    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3034    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3035    return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3036                       Chain, Dest, ARMcc, CCR, Cmp);
3037  }
3038
3039  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3040
3041  if (getTargetMachine().Options.UnsafeFPMath &&
3042      (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3043       CC == ISD::SETNE || CC == ISD::SETUNE)) {
3044    SDValue Result = OptimizeVFPBrcond(Op, DAG);
3045    if (Result.getNode())
3046      return Result;
3047  }
3048
3049  ARMCC::CondCodes CondCode, CondCode2;
3050  FPCCToARMCC(CC, CondCode, CondCode2);
3051
3052  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3053  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3054  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3055  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3056  SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3057  SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3058  if (CondCode2 != ARMCC::AL) {
3059    ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3060    SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3061    Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3062  }
3063  return Res;
3064}
3065
3066SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3067  SDValue Chain = Op.getOperand(0);
3068  SDValue Table = Op.getOperand(1);
3069  SDValue Index = Op.getOperand(2);
3070  DebugLoc dl = Op.getDebugLoc();
3071
3072  EVT PTy = getPointerTy();
3073  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3074  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3075  SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3076  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3077  Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3078  Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3079  SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3080  if (Subtarget->isThumb2()) {
3081    // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3082    // which does another jump to the destination. This also makes it easier
3083    // to translate it to TBB / TBH later.
3084    // FIXME: This might not work if the function is extremely large.
3085    return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3086                       Addr, Op.getOperand(2), JTI, UId);
3087  }
3088  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3089    Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3090                       MachinePointerInfo::getJumpTable(),
3091                       false, false, false, 0);
3092    Chain = Addr.getValue(1);
3093    Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3094    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3095  } else {
3096    Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3097                       MachinePointerInfo::getJumpTable(),
3098                       false, false, false, 0);
3099    Chain = Addr.getValue(1);
3100    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3101  }
3102}
3103
3104static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3105  EVT VT = Op.getValueType();
3106  DebugLoc dl = Op.getDebugLoc();
3107
3108  if (Op.getValueType().getVectorElementType() == MVT::i32) {
3109    if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3110      return Op;
3111    return DAG.UnrollVectorOp(Op.getNode());
3112  }
3113
3114  assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3115         "Invalid type for custom lowering!");
3116  if (VT != MVT::v4i16)
3117    return DAG.UnrollVectorOp(Op.getNode());
3118
3119  Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3120  return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3121}
3122
3123static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3124  EVT VT = Op.getValueType();
3125  if (VT.isVector())
3126    return LowerVectorFP_TO_INT(Op, DAG);
3127
3128  DebugLoc dl = Op.getDebugLoc();
3129  unsigned Opc;
3130
3131  switch (Op.getOpcode()) {
3132  default: llvm_unreachable("Invalid opcode!");
3133  case ISD::FP_TO_SINT:
3134    Opc = ARMISD::FTOSI;
3135    break;
3136  case ISD::FP_TO_UINT:
3137    Opc = ARMISD::FTOUI;
3138    break;
3139  }
3140  Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3141  return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3142}
3143
3144static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3145  EVT VT = Op.getValueType();
3146  DebugLoc dl = Op.getDebugLoc();
3147
3148  if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3149    if (VT.getVectorElementType() == MVT::f32)
3150      return Op;
3151    return DAG.UnrollVectorOp(Op.getNode());
3152  }
3153
3154  assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3155         "Invalid type for custom lowering!");
3156  if (VT != MVT::v4f32)
3157    return DAG.UnrollVectorOp(Op.getNode());
3158
3159  unsigned CastOpc;
3160  unsigned Opc;
3161  switch (Op.getOpcode()) {
3162  default: llvm_unreachable("Invalid opcode!");
3163  case ISD::SINT_TO_FP:
3164    CastOpc = ISD::SIGN_EXTEND;
3165    Opc = ISD::SINT_TO_FP;
3166    break;
3167  case ISD::UINT_TO_FP:
3168    CastOpc = ISD::ZERO_EXTEND;
3169    Opc = ISD::UINT_TO_FP;
3170    break;
3171  }
3172
3173  Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3174  return DAG.getNode(Opc, dl, VT, Op);
3175}
3176
3177static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3178  EVT VT = Op.getValueType();
3179  if (VT.isVector())
3180    return LowerVectorINT_TO_FP(Op, DAG);
3181
3182  DebugLoc dl = Op.getDebugLoc();
3183  unsigned Opc;
3184
3185  switch (Op.getOpcode()) {
3186  default: llvm_unreachable("Invalid opcode!");
3187  case ISD::SINT_TO_FP:
3188    Opc = ARMISD::SITOF;
3189    break;
3190  case ISD::UINT_TO_FP:
3191    Opc = ARMISD::UITOF;
3192    break;
3193  }
3194
3195  Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3196  return DAG.getNode(Opc, dl, VT, Op);
3197}
3198
3199SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3200  // Implement fcopysign with a fabs and a conditional fneg.
3201  SDValue Tmp0 = Op.getOperand(0);
3202  SDValue Tmp1 = Op.getOperand(1);
3203  DebugLoc dl = Op.getDebugLoc();
3204  EVT VT = Op.getValueType();
3205  EVT SrcVT = Tmp1.getValueType();
3206  bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3207    Tmp0.getOpcode() == ARMISD::VMOVDRR;
3208  bool UseNEON = !InGPR && Subtarget->hasNEON();
3209
3210  if (UseNEON) {
3211    // Use VBSL to copy the sign bit.
3212    unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3213    SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3214                               DAG.getTargetConstant(EncodedVal, MVT::i32));
3215    EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3216    if (VT == MVT::f64)
3217      Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3218                         DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3219                         DAG.getConstant(32, MVT::i32));
3220    else /*if (VT == MVT::f32)*/
3221      Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3222    if (SrcVT == MVT::f32) {
3223      Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3224      if (VT == MVT::f64)
3225        Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3226                           DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3227                           DAG.getConstant(32, MVT::i32));
3228    } else if (VT == MVT::f32)
3229      Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3230                         DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3231                         DAG.getConstant(32, MVT::i32));
3232    Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3233    Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3234
3235    SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3236                                            MVT::i32);
3237    AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3238    SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3239                                  DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3240
3241    SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3242                              DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3243                              DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3244    if (VT == MVT::f32) {
3245      Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3246      Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3247                        DAG.getConstant(0, MVT::i32));
3248    } else {
3249      Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3250    }
3251
3252    return Res;
3253  }
3254
3255  // Bitcast operand 1 to i32.
3256  if (SrcVT == MVT::f64)
3257    Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3258                       &Tmp1, 1).getValue(1);
3259  Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3260
3261  // Or in the signbit with integer operations.
3262  SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3263  SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3264  Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3265  if (VT == MVT::f32) {
3266    Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3267                       DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3268    return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3269                       DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3270  }
3271
3272  // f64: Or the high part with signbit and then combine two parts.
3273  Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3274                     &Tmp0, 1);
3275  SDValue Lo = Tmp0.getValue(0);
3276  SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3277  Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3278  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3279}
3280
3281SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3282  MachineFunction &MF = DAG.getMachineFunction();
3283  MachineFrameInfo *MFI = MF.getFrameInfo();
3284  MFI->setReturnAddressIsTaken(true);
3285
3286  EVT VT = Op.getValueType();
3287  DebugLoc dl = Op.getDebugLoc();
3288  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3289  if (Depth) {
3290    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3291    SDValue Offset = DAG.getConstant(4, MVT::i32);
3292    return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3293                       DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3294                       MachinePointerInfo(), false, false, false, 0);
3295  }
3296
3297  // Return LR, which contains the return address. Mark it an implicit live-in.
3298  unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3299  return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3300}
3301
3302SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3303  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3304  MFI->setFrameAddressIsTaken(true);
3305
3306  EVT VT = Op.getValueType();
3307  DebugLoc dl = Op.getDebugLoc();  // FIXME probably not meaningful
3308  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3309  unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3310    ? ARM::R7 : ARM::R11;
3311  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3312  while (Depth--)
3313    FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3314                            MachinePointerInfo(),
3315                            false, false, false, 0);
3316  return FrameAddr;
3317}
3318
3319/// ExpandBITCAST - If the target supports VFP, this function is called to
3320/// expand a bit convert where either the source or destination type is i64 to
3321/// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
3322/// operand type is illegal (e.g., v2f32 for a target that doesn't support
3323/// vectors), since the legalizer won't know what to do with that.
3324static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3325  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3326  DebugLoc dl = N->getDebugLoc();
3327  SDValue Op = N->getOperand(0);
3328
3329  // This function is only supposed to be called for i64 types, either as the
3330  // source or destination of the bit convert.
3331  EVT SrcVT = Op.getValueType();
3332  EVT DstVT = N->getValueType(0);
3333  assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3334         "ExpandBITCAST called for non-i64 type");
3335
3336  // Turn i64->f64 into VMOVDRR.
3337  if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3338    SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3339                             DAG.getConstant(0, MVT::i32));
3340    SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3341                             DAG.getConstant(1, MVT::i32));
3342    return DAG.getNode(ISD::BITCAST, dl, DstVT,
3343                       DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3344  }
3345
3346  // Turn f64->i64 into VMOVRRD.
3347  if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3348    SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3349                              DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3350    // Merge the pieces into a single i64 value.
3351    return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3352  }
3353
3354  return SDValue();
3355}
3356
3357/// getZeroVector - Returns a vector of specified type with all zero elements.
3358/// Zero vectors are used to represent vector negation and in those cases
3359/// will be implemented with the NEON VNEG instruction.  However, VNEG does
3360/// not support i64 elements, so sometimes the zero vectors will need to be
3361/// explicitly constructed.  Regardless, use a canonical VMOV to create the
3362/// zero vector.
3363static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3364  assert(VT.isVector() && "Expected a vector type");
3365  // The canonical modified immediate encoding of a zero vector is....0!
3366  SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3367  EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3368  SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3369  return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3370}
3371
3372/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3373/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3374SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3375                                                SelectionDAG &DAG) const {
3376  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3377  EVT VT = Op.getValueType();
3378  unsigned VTBits = VT.getSizeInBits();
3379  DebugLoc dl = Op.getDebugLoc();
3380  SDValue ShOpLo = Op.getOperand(0);
3381  SDValue ShOpHi = Op.getOperand(1);
3382  SDValue ShAmt  = Op.getOperand(2);
3383  SDValue ARMcc;
3384  unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3385
3386  assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3387
3388  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3389                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3390  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3391  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3392                                   DAG.getConstant(VTBits, MVT::i32));
3393  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3394  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3395  SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3396
3397  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3398  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3399                          ARMcc, DAG, dl);
3400  SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3401  SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3402                           CCR, Cmp);
3403
3404  SDValue Ops[2] = { Lo, Hi };
3405  return DAG.getMergeValues(Ops, 2, dl);
3406}
3407
3408/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3409/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3410SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3411                                               SelectionDAG &DAG) const {
3412  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3413  EVT VT = Op.getValueType();
3414  unsigned VTBits = VT.getSizeInBits();
3415  DebugLoc dl = Op.getDebugLoc();
3416  SDValue ShOpLo = Op.getOperand(0);
3417  SDValue ShOpHi = Op.getOperand(1);
3418  SDValue ShAmt  = Op.getOperand(2);
3419  SDValue ARMcc;
3420
3421  assert(Op.getOpcode() == ISD::SHL_PARTS);
3422  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3423                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3424  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3425  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3426                                   DAG.getConstant(VTBits, MVT::i32));
3427  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3428  SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3429
3430  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3431  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3432  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3433                          ARMcc, DAG, dl);
3434  SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3435  SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3436                           CCR, Cmp);
3437
3438  SDValue Ops[2] = { Lo, Hi };
3439  return DAG.getMergeValues(Ops, 2, dl);
3440}
3441
3442SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3443                                            SelectionDAG &DAG) const {
3444  // The rounding mode is in bits 23:22 of the FPSCR.
3445  // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3446  // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3447  // so that the shift + and get folded into a bitfield extract.
3448  DebugLoc dl = Op.getDebugLoc();
3449  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3450                              DAG.getConstant(Intrinsic::arm_get_fpscr,
3451                                              MVT::i32));
3452  SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3453                                  DAG.getConstant(1U << 22, MVT::i32));
3454  SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3455                              DAG.getConstant(22, MVT::i32));
3456  return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3457                     DAG.getConstant(3, MVT::i32));
3458}
3459
3460static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3461                         const ARMSubtarget *ST) {
3462  EVT VT = N->getValueType(0);
3463  DebugLoc dl = N->getDebugLoc();
3464
3465  if (!ST->hasV6T2Ops())
3466    return SDValue();
3467
3468  SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3469  return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3470}
3471
3472static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3473                          const ARMSubtarget *ST) {
3474  EVT VT = N->getValueType(0);
3475  DebugLoc dl = N->getDebugLoc();
3476
3477  if (!VT.isVector())
3478    return SDValue();
3479
3480  // Lower vector shifts on NEON to use VSHL.
3481  assert(ST->hasNEON() && "unexpected vector shift");
3482
3483  // Left shifts translate directly to the vshiftu intrinsic.
3484  if (N->getOpcode() == ISD::SHL)
3485    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3486                       DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3487                       N->getOperand(0), N->getOperand(1));
3488
3489  assert((N->getOpcode() == ISD::SRA ||
3490          N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3491
3492  // NEON uses the same intrinsics for both left and right shifts.  For
3493  // right shifts, the shift amounts are negative, so negate the vector of
3494  // shift amounts.
3495  EVT ShiftVT = N->getOperand(1).getValueType();
3496  SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3497                                     getZeroVector(ShiftVT, DAG, dl),
3498                                     N->getOperand(1));
3499  Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3500                             Intrinsic::arm_neon_vshifts :
3501                             Intrinsic::arm_neon_vshiftu);
3502  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3503                     DAG.getConstant(vshiftInt, MVT::i32),
3504                     N->getOperand(0), NegatedCount);
3505}
3506
3507static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3508                                const ARMSubtarget *ST) {
3509  EVT VT = N->getValueType(0);
3510  DebugLoc dl = N->getDebugLoc();
3511
3512  // We can get here for a node like i32 = ISD::SHL i32, i64
3513  if (VT != MVT::i64)
3514    return SDValue();
3515
3516  assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3517         "Unknown shift to lower!");
3518
3519  // We only lower SRA, SRL of 1 here, all others use generic lowering.
3520  if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3521      cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3522    return SDValue();
3523
3524  // If we are in thumb mode, we don't have RRX.
3525  if (ST->isThumb1Only()) return SDValue();
3526
3527  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
3528  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3529                           DAG.getConstant(0, MVT::i32));
3530  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3531                           DAG.getConstant(1, MVT::i32));
3532
3533  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3534  // captures the result into a carry flag.
3535  unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3536  Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3537
3538  // The low part is an ARMISD::RRX operand, which shifts the carry in.
3539  Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3540
3541  // Merge the pieces into a single i64 value.
3542 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3543}
3544
3545static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3546  SDValue TmpOp0, TmpOp1;
3547  bool Invert = false;
3548  bool Swap = false;
3549  unsigned Opc = 0;
3550
3551  SDValue Op0 = Op.getOperand(0);
3552  SDValue Op1 = Op.getOperand(1);
3553  SDValue CC = Op.getOperand(2);
3554  EVT VT = Op.getValueType();
3555  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3556  DebugLoc dl = Op.getDebugLoc();
3557
3558  if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3559    switch (SetCCOpcode) {
3560    default: llvm_unreachable("Illegal FP comparison");
3561    case ISD::SETUNE:
3562    case ISD::SETNE:  Invert = true; // Fallthrough
3563    case ISD::SETOEQ:
3564    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3565    case ISD::SETOLT:
3566    case ISD::SETLT: Swap = true; // Fallthrough
3567    case ISD::SETOGT:
3568    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3569    case ISD::SETOLE:
3570    case ISD::SETLE:  Swap = true; // Fallthrough
3571    case ISD::SETOGE:
3572    case ISD::SETGE: Opc = ARMISD::VCGE; break;
3573    case ISD::SETUGE: Swap = true; // Fallthrough
3574    case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3575    case ISD::SETUGT: Swap = true; // Fallthrough
3576    case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3577    case ISD::SETUEQ: Invert = true; // Fallthrough
3578    case ISD::SETONE:
3579      // Expand this to (OLT | OGT).
3580      TmpOp0 = Op0;
3581      TmpOp1 = Op1;
3582      Opc = ISD::OR;
3583      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3584      Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3585      break;
3586    case ISD::SETUO: Invert = true; // Fallthrough
3587    case ISD::SETO:
3588      // Expand this to (OLT | OGE).
3589      TmpOp0 = Op0;
3590      TmpOp1 = Op1;
3591      Opc = ISD::OR;
3592      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3593      Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3594      break;
3595    }
3596  } else {
3597    // Integer comparisons.
3598    switch (SetCCOpcode) {
3599    default: llvm_unreachable("Illegal integer comparison");
3600    case ISD::SETNE:  Invert = true;
3601    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3602    case ISD::SETLT:  Swap = true;
3603    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3604    case ISD::SETLE:  Swap = true;
3605    case ISD::SETGE:  Opc = ARMISD::VCGE; break;
3606    case ISD::SETULT: Swap = true;
3607    case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3608    case ISD::SETULE: Swap = true;
3609    case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3610    }
3611
3612    // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3613    if (Opc == ARMISD::VCEQ) {
3614
3615      SDValue AndOp;
3616      if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3617        AndOp = Op0;
3618      else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3619        AndOp = Op1;
3620
3621      // Ignore bitconvert.
3622      if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3623        AndOp = AndOp.getOperand(0);
3624
3625      if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3626        Opc = ARMISD::VTST;
3627        Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3628        Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3629        Invert = !Invert;
3630      }
3631    }
3632  }
3633
3634  if (Swap)
3635    std::swap(Op0, Op1);
3636
3637  // If one of the operands is a constant vector zero, attempt to fold the
3638  // comparison to a specialized compare-against-zero form.
3639  SDValue SingleOp;
3640  if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3641    SingleOp = Op0;
3642  else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3643    if (Opc == ARMISD::VCGE)
3644      Opc = ARMISD::VCLEZ;
3645    else if (Opc == ARMISD::VCGT)
3646      Opc = ARMISD::VCLTZ;
3647    SingleOp = Op1;
3648  }
3649
3650  SDValue Result;
3651  if (SingleOp.getNode()) {
3652    switch (Opc) {
3653    case ARMISD::VCEQ:
3654      Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3655    case ARMISD::VCGE:
3656      Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3657    case ARMISD::VCLEZ:
3658      Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3659    case ARMISD::VCGT:
3660      Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3661    case ARMISD::VCLTZ:
3662      Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3663    default:
3664      Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3665    }
3666  } else {
3667     Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3668  }
3669
3670  if (Invert)
3671    Result = DAG.getNOT(dl, Result, VT);
3672
3673  return Result;
3674}
3675
3676/// isNEONModifiedImm - Check if the specified splat value corresponds to a
3677/// valid vector constant for a NEON instruction with a "modified immediate"
3678/// operand (e.g., VMOV).  If so, return the encoded value.
3679static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3680                                 unsigned SplatBitSize, SelectionDAG &DAG,
3681                                 EVT &VT, bool is128Bits, NEONModImmType type) {
3682  unsigned OpCmode, Imm;
3683
3684  // SplatBitSize is set to the smallest size that splats the vector, so a
3685  // zero vector will always have SplatBitSize == 8.  However, NEON modified
3686  // immediate instructions others than VMOV do not support the 8-bit encoding
3687  // of a zero vector, and the default encoding of zero is supposed to be the
3688  // 32-bit version.
3689  if (SplatBits == 0)
3690    SplatBitSize = 32;
3691
3692  switch (SplatBitSize) {
3693  case 8:
3694    if (type != VMOVModImm)
3695      return SDValue();
3696    // Any 1-byte value is OK.  Op=0, Cmode=1110.
3697    assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3698    OpCmode = 0xe;
3699    Imm = SplatBits;
3700    VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3701    break;
3702
3703  case 16:
3704    // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3705    VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3706    if ((SplatBits & ~0xff) == 0) {
3707      // Value = 0x00nn: Op=x, Cmode=100x.
3708      OpCmode = 0x8;
3709      Imm = SplatBits;
3710      break;
3711    }
3712    if ((SplatBits & ~0xff00) == 0) {
3713      // Value = 0xnn00: Op=x, Cmode=101x.
3714      OpCmode = 0xa;
3715      Imm = SplatBits >> 8;
3716      break;
3717    }
3718    return SDValue();
3719
3720  case 32:
3721    // NEON's 32-bit VMOV supports splat values where:
3722    // * only one byte is nonzero, or
3723    // * the least significant byte is 0xff and the second byte is nonzero, or
3724    // * the least significant 2 bytes are 0xff and the third is nonzero.
3725    VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3726    if ((SplatBits & ~0xff) == 0) {
3727      // Value = 0x000000nn: Op=x, Cmode=000x.
3728      OpCmode = 0;
3729      Imm = SplatBits;
3730      break;
3731    }
3732    if ((SplatBits & ~0xff00) == 0) {
3733      // Value = 0x0000nn00: Op=x, Cmode=001x.
3734      OpCmode = 0x2;
3735      Imm = SplatBits >> 8;
3736      break;
3737    }
3738    if ((SplatBits & ~0xff0000) == 0) {
3739      // Value = 0x00nn0000: Op=x, Cmode=010x.
3740      OpCmode = 0x4;
3741      Imm = SplatBits >> 16;
3742      break;
3743    }
3744    if ((SplatBits & ~0xff000000) == 0) {
3745      // Value = 0xnn000000: Op=x, Cmode=011x.
3746      OpCmode = 0x6;
3747      Imm = SplatBits >> 24;
3748      break;
3749    }
3750
3751    // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3752    if (type == OtherModImm) return SDValue();
3753
3754    if ((SplatBits & ~0xffff) == 0 &&
3755        ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3756      // Value = 0x0000nnff: Op=x, Cmode=1100.
3757      OpCmode = 0xc;
3758      Imm = SplatBits >> 8;
3759      SplatBits |= 0xff;
3760      break;
3761    }
3762
3763    if ((SplatBits & ~0xffffff) == 0 &&
3764        ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3765      // Value = 0x00nnffff: Op=x, Cmode=1101.
3766      OpCmode = 0xd;
3767      Imm = SplatBits >> 16;
3768      SplatBits |= 0xffff;
3769      break;
3770    }
3771
3772    // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3773    // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3774    // VMOV.I32.  A (very) minor optimization would be to replicate the value
3775    // and fall through here to test for a valid 64-bit splat.  But, then the
3776    // caller would also need to check and handle the change in size.
3777    return SDValue();
3778
3779  case 64: {
3780    if (type != VMOVModImm)
3781      return SDValue();
3782    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3783    uint64_t BitMask = 0xff;
3784    uint64_t Val = 0;
3785    unsigned ImmMask = 1;
3786    Imm = 0;
3787    for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3788      if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3789        Val |= BitMask;
3790        Imm |= ImmMask;
3791      } else if ((SplatBits & BitMask) != 0) {
3792        return SDValue();
3793      }
3794      BitMask <<= 8;
3795      ImmMask <<= 1;
3796    }
3797    // Op=1, Cmode=1110.
3798    OpCmode = 0x1e;
3799    SplatBits = Val;
3800    VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3801    break;
3802  }
3803
3804  default:
3805    llvm_unreachable("unexpected size for isNEONModifiedImm");
3806  }
3807
3808  unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3809  return DAG.getTargetConstant(EncodedVal, MVT::i32);
3810}
3811
3812static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3813                       bool &ReverseVEXT, unsigned &Imm) {
3814  unsigned NumElts = VT.getVectorNumElements();
3815  ReverseVEXT = false;
3816
3817  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
3818  if (M[0] < 0)
3819    return false;
3820
3821  Imm = M[0];
3822
3823  // If this is a VEXT shuffle, the immediate value is the index of the first
3824  // element.  The other shuffle indices must be the successive elements after
3825  // the first one.
3826  unsigned ExpectedElt = Imm;
3827  for (unsigned i = 1; i < NumElts; ++i) {
3828    // Increment the expected index.  If it wraps around, it may still be
3829    // a VEXT but the source vectors must be swapped.
3830    ExpectedElt += 1;
3831    if (ExpectedElt == NumElts * 2) {
3832      ExpectedElt = 0;
3833      ReverseVEXT = true;
3834    }
3835
3836    if (M[i] < 0) continue; // ignore UNDEF indices
3837    if (ExpectedElt != static_cast<unsigned>(M[i]))
3838      return false;
3839  }
3840
3841  // Adjust the index value if the source operands will be swapped.
3842  if (ReverseVEXT)
3843    Imm -= NumElts;
3844
3845  return true;
3846}
3847
3848/// isVREVMask - Check if a vector shuffle corresponds to a VREV
3849/// instruction with the specified blocksize.  (The order of the elements
3850/// within each block of the vector is reversed.)
3851static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3852  assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3853         "Only possible block sizes for VREV are: 16, 32, 64");
3854
3855  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3856  if (EltSz == 64)
3857    return false;
3858
3859  unsigned NumElts = VT.getVectorNumElements();
3860  unsigned BlockElts = M[0] + 1;
3861  // If the first shuffle index is UNDEF, be optimistic.
3862  if (M[0] < 0)
3863    BlockElts = BlockSize / EltSz;
3864
3865  if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3866    return false;
3867
3868  for (unsigned i = 0; i < NumElts; ++i) {
3869    if (M[i] < 0) continue; // ignore UNDEF indices
3870    if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3871      return false;
3872  }
3873
3874  return true;
3875}
3876
3877static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3878  // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3879  // range, then 0 is placed into the resulting vector. So pretty much any mask
3880  // of 8 elements can work here.
3881  return VT == MVT::v8i8 && M.size() == 8;
3882}
3883
3884static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3885  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3886  if (EltSz == 64)
3887    return false;
3888
3889  unsigned NumElts = VT.getVectorNumElements();
3890  WhichResult = (M[0] == 0 ? 0 : 1);
3891  for (unsigned i = 0; i < NumElts; i += 2) {
3892    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3893        (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3894      return false;
3895  }
3896  return true;
3897}
3898
3899/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3900/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3901/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3902static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3903  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3904  if (EltSz == 64)
3905    return false;
3906
3907  unsigned NumElts = VT.getVectorNumElements();
3908  WhichResult = (M[0] == 0 ? 0 : 1);
3909  for (unsigned i = 0; i < NumElts; i += 2) {
3910    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3911        (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3912      return false;
3913  }
3914  return true;
3915}
3916
3917static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3918  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3919  if (EltSz == 64)
3920    return false;
3921
3922  unsigned NumElts = VT.getVectorNumElements();
3923  WhichResult = (M[0] == 0 ? 0 : 1);
3924  for (unsigned i = 0; i != NumElts; ++i) {
3925    if (M[i] < 0) continue; // ignore UNDEF indices
3926    if ((unsigned) M[i] != 2 * i + WhichResult)
3927      return false;
3928  }
3929
3930  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3931  if (VT.is64BitVector() && EltSz == 32)
3932    return false;
3933
3934  return true;
3935}
3936
3937/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3938/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3939/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3940static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3941  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3942  if (EltSz == 64)
3943    return false;
3944
3945  unsigned Half = VT.getVectorNumElements() / 2;
3946  WhichResult = (M[0] == 0 ? 0 : 1);
3947  for (unsigned j = 0; j != 2; ++j) {
3948    unsigned Idx = WhichResult;
3949    for (unsigned i = 0; i != Half; ++i) {
3950      int MIdx = M[i + j * Half];
3951      if (MIdx >= 0 && (unsigned) MIdx != Idx)
3952        return false;
3953      Idx += 2;
3954    }
3955  }
3956
3957  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3958  if (VT.is64BitVector() && EltSz == 32)
3959    return false;
3960
3961  return true;
3962}
3963
3964static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3965  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3966  if (EltSz == 64)
3967    return false;
3968
3969  unsigned NumElts = VT.getVectorNumElements();
3970  WhichResult = (M[0] == 0 ? 0 : 1);
3971  unsigned Idx = WhichResult * NumElts / 2;
3972  for (unsigned i = 0; i != NumElts; i += 2) {
3973    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3974        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3975      return false;
3976    Idx += 1;
3977  }
3978
3979  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3980  if (VT.is64BitVector() && EltSz == 32)
3981    return false;
3982
3983  return true;
3984}
3985
3986/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3987/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3988/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3989static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3990  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3991  if (EltSz == 64)
3992    return false;
3993
3994  unsigned NumElts = VT.getVectorNumElements();
3995  WhichResult = (M[0] == 0 ? 0 : 1);
3996  unsigned Idx = WhichResult * NumElts / 2;
3997  for (unsigned i = 0; i != NumElts; i += 2) {
3998    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3999        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4000      return false;
4001    Idx += 1;
4002  }
4003
4004  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4005  if (VT.is64BitVector() && EltSz == 32)
4006    return false;
4007
4008  return true;
4009}
4010
4011// If N is an integer constant that can be moved into a register in one
4012// instruction, return an SDValue of such a constant (will become a MOV
4013// instruction).  Otherwise return null.
4014static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4015                                     const ARMSubtarget *ST, DebugLoc dl) {
4016  uint64_t Val;
4017  if (!isa<ConstantSDNode>(N))
4018    return SDValue();
4019  Val = cast<ConstantSDNode>(N)->getZExtValue();
4020
4021  if (ST->isThumb1Only()) {
4022    if (Val <= 255 || ~Val <= 255)
4023      return DAG.getConstant(Val, MVT::i32);
4024  } else {
4025    if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4026      return DAG.getConstant(Val, MVT::i32);
4027  }
4028  return SDValue();
4029}
4030
4031// If this is a case we can't handle, return null and let the default
4032// expansion code take care of it.
4033SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4034                                             const ARMSubtarget *ST) const {
4035  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4036  DebugLoc dl = Op.getDebugLoc();
4037  EVT VT = Op.getValueType();
4038
4039  APInt SplatBits, SplatUndef;
4040  unsigned SplatBitSize;
4041  bool HasAnyUndefs;
4042  if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4043    if (SplatBitSize <= 64) {
4044      // Check if an immediate VMOV works.
4045      EVT VmovVT;
4046      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4047                                      SplatUndef.getZExtValue(), SplatBitSize,
4048                                      DAG, VmovVT, VT.is128BitVector(),
4049                                      VMOVModImm);
4050      if (Val.getNode()) {
4051        SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4052        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4053      }
4054
4055      // Try an immediate VMVN.
4056      uint64_t NegatedImm = (~SplatBits).getZExtValue();
4057      Val = isNEONModifiedImm(NegatedImm,
4058                                      SplatUndef.getZExtValue(), SplatBitSize,
4059                                      DAG, VmovVT, VT.is128BitVector(),
4060                                      VMVNModImm);
4061      if (Val.getNode()) {
4062        SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4063        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4064      }
4065
4066      // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4067      if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4068        int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4069        if (ImmVal != -1) {
4070          SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4071          return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4072        }
4073      }
4074    }
4075  }
4076
4077  // Scan through the operands to see if only one value is used.
4078  unsigned NumElts = VT.getVectorNumElements();
4079  bool isOnlyLowElement = true;
4080  bool usesOnlyOneValue = true;
4081  bool isConstant = true;
4082  SDValue Value;
4083  for (unsigned i = 0; i < NumElts; ++i) {
4084    SDValue V = Op.getOperand(i);
4085    if (V.getOpcode() == ISD::UNDEF)
4086      continue;
4087    if (i > 0)
4088      isOnlyLowElement = false;
4089    if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4090      isConstant = false;
4091
4092    if (!Value.getNode())
4093      Value = V;
4094    else if (V != Value)
4095      usesOnlyOneValue = false;
4096  }
4097
4098  if (!Value.getNode())
4099    return DAG.getUNDEF(VT);
4100
4101  if (isOnlyLowElement)
4102    return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4103
4104  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4105
4106  // Use VDUP for non-constant splats.  For f32 constant splats, reduce to
4107  // i32 and try again.
4108  if (usesOnlyOneValue && EltSize <= 32) {
4109    if (!isConstant)
4110      return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4111    if (VT.getVectorElementType().isFloatingPoint()) {
4112      SmallVector<SDValue, 8> Ops;
4113      for (unsigned i = 0; i < NumElts; ++i)
4114        Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4115                                  Op.getOperand(i)));
4116      EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4117      SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4118      Val = LowerBUILD_VECTOR(Val, DAG, ST);
4119      if (Val.getNode())
4120        return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4121    }
4122    SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4123    if (Val.getNode())
4124      return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4125  }
4126
4127  // If all elements are constants and the case above didn't get hit, fall back
4128  // to the default expansion, which will generate a load from the constant
4129  // pool.
4130  if (isConstant)
4131    return SDValue();
4132
4133  // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4134  if (NumElts >= 4) {
4135    SDValue shuffle = ReconstructShuffle(Op, DAG);
4136    if (shuffle != SDValue())
4137      return shuffle;
4138  }
4139
4140  // Vectors with 32- or 64-bit elements can be built by directly assigning
4141  // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
4142  // will be legalized.
4143  if (EltSize >= 32) {
4144    // Do the expansion with floating-point types, since that is what the VFP
4145    // registers are defined to use, and since i64 is not legal.
4146    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4147    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4148    SmallVector<SDValue, 8> Ops;
4149    for (unsigned i = 0; i < NumElts; ++i)
4150      Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4151    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4152    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4153  }
4154
4155  return SDValue();
4156}
4157
4158// Gather data to see if the operation can be modelled as a
4159// shuffle in combination with VEXTs.
4160SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4161                                              SelectionDAG &DAG) const {
4162  DebugLoc dl = Op.getDebugLoc();
4163  EVT VT = Op.getValueType();
4164  unsigned NumElts = VT.getVectorNumElements();
4165
4166  SmallVector<SDValue, 2> SourceVecs;
4167  SmallVector<unsigned, 2> MinElts;
4168  SmallVector<unsigned, 2> MaxElts;
4169
4170  for (unsigned i = 0; i < NumElts; ++i) {
4171    SDValue V = Op.getOperand(i);
4172    if (V.getOpcode() == ISD::UNDEF)
4173      continue;
4174    else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4175      // A shuffle can only come from building a vector from various
4176      // elements of other vectors.
4177      return SDValue();
4178    } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4179               VT.getVectorElementType()) {
4180      // This code doesn't know how to handle shuffles where the vector
4181      // element types do not match (this happens because type legalization
4182      // promotes the return type of EXTRACT_VECTOR_ELT).
4183      // FIXME: It might be appropriate to extend this code to handle
4184      // mismatched types.
4185      return SDValue();
4186    }
4187
4188    // Record this extraction against the appropriate vector if possible...
4189    SDValue SourceVec = V.getOperand(0);
4190    unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4191    bool FoundSource = false;
4192    for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4193      if (SourceVecs[j] == SourceVec) {
4194        if (MinElts[j] > EltNo)
4195          MinElts[j] = EltNo;
4196        if (MaxElts[j] < EltNo)
4197          MaxElts[j] = EltNo;
4198        FoundSource = true;
4199        break;
4200      }
4201    }
4202
4203    // Or record a new source if not...
4204    if (!FoundSource) {
4205      SourceVecs.push_back(SourceVec);
4206      MinElts.push_back(EltNo);
4207      MaxElts.push_back(EltNo);
4208    }
4209  }
4210
4211  // Currently only do something sane when at most two source vectors
4212  // involved.
4213  if (SourceVecs.size() > 2)
4214    return SDValue();
4215
4216  SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4217  int VEXTOffsets[2] = {0, 0};
4218
4219  // This loop extracts the usage patterns of the source vectors
4220  // and prepares appropriate SDValues for a shuffle if possible.
4221  for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4222    if (SourceVecs[i].getValueType() == VT) {
4223      // No VEXT necessary
4224      ShuffleSrcs[i] = SourceVecs[i];
4225      VEXTOffsets[i] = 0;
4226      continue;
4227    } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4228      // It probably isn't worth padding out a smaller vector just to
4229      // break it down again in a shuffle.
4230      return SDValue();
4231    }
4232
4233    // Since only 64-bit and 128-bit vectors are legal on ARM and
4234    // we've eliminated the other cases...
4235    assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4236           "unexpected vector sizes in ReconstructShuffle");
4237
4238    if (MaxElts[i] - MinElts[i] >= NumElts) {
4239      // Span too large for a VEXT to cope
4240      return SDValue();
4241    }
4242
4243    if (MinElts[i] >= NumElts) {
4244      // The extraction can just take the second half
4245      VEXTOffsets[i] = NumElts;
4246      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4247                                   SourceVecs[i],
4248                                   DAG.getIntPtrConstant(NumElts));
4249    } else if (MaxElts[i] < NumElts) {
4250      // The extraction can just take the first half
4251      VEXTOffsets[i] = 0;
4252      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4253                                   SourceVecs[i],
4254                                   DAG.getIntPtrConstant(0));
4255    } else {
4256      // An actual VEXT is needed
4257      VEXTOffsets[i] = MinElts[i];
4258      SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4259                                     SourceVecs[i],
4260                                     DAG.getIntPtrConstant(0));
4261      SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4262                                     SourceVecs[i],
4263                                     DAG.getIntPtrConstant(NumElts));
4264      ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4265                                   DAG.getConstant(VEXTOffsets[i], MVT::i32));
4266    }
4267  }
4268
4269  SmallVector<int, 8> Mask;
4270
4271  for (unsigned i = 0; i < NumElts; ++i) {
4272    SDValue Entry = Op.getOperand(i);
4273    if (Entry.getOpcode() == ISD::UNDEF) {
4274      Mask.push_back(-1);
4275      continue;
4276    }
4277
4278    SDValue ExtractVec = Entry.getOperand(0);
4279    int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4280                                          .getOperand(1))->getSExtValue();
4281    if (ExtractVec == SourceVecs[0]) {
4282      Mask.push_back(ExtractElt - VEXTOffsets[0]);
4283    } else {
4284      Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4285    }
4286  }
4287
4288  // Final check before we try to produce nonsense...
4289  if (isShuffleMaskLegal(Mask, VT))
4290    return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4291                                &Mask[0]);
4292
4293  return SDValue();
4294}
4295
4296/// isShuffleMaskLegal - Targets can use this to indicate that they only
4297/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4298/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4299/// are assumed to be legal.
4300bool
4301ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4302                                      EVT VT) const {
4303  if (VT.getVectorNumElements() == 4 &&
4304      (VT.is128BitVector() || VT.is64BitVector())) {
4305    unsigned PFIndexes[4];
4306    for (unsigned i = 0; i != 4; ++i) {
4307      if (M[i] < 0)
4308        PFIndexes[i] = 8;
4309      else
4310        PFIndexes[i] = M[i];
4311    }
4312
4313    // Compute the index in the perfect shuffle table.
4314    unsigned PFTableIndex =
4315      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4316    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4317    unsigned Cost = (PFEntry >> 30);
4318
4319    if (Cost <= 4)
4320      return true;
4321  }
4322
4323  bool ReverseVEXT;
4324  unsigned Imm, WhichResult;
4325
4326  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4327  return (EltSize >= 32 ||
4328          ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4329          isVREVMask(M, VT, 64) ||
4330          isVREVMask(M, VT, 32) ||
4331          isVREVMask(M, VT, 16) ||
4332          isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4333          isVTBLMask(M, VT) ||
4334          isVTRNMask(M, VT, WhichResult) ||
4335          isVUZPMask(M, VT, WhichResult) ||
4336          isVZIPMask(M, VT, WhichResult) ||
4337          isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4338          isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4339          isVZIP_v_undef_Mask(M, VT, WhichResult));
4340}
4341
4342/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4343/// the specified operations to build the shuffle.
4344static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4345                                      SDValue RHS, SelectionDAG &DAG,
4346                                      DebugLoc dl) {
4347  unsigned OpNum = (PFEntry >> 26) & 0x0F;
4348  unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4349  unsigned RHSID = (PFEntry >>  0) & ((1 << 13)-1);
4350
4351  enum {
4352    OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4353    OP_VREV,
4354    OP_VDUP0,
4355    OP_VDUP1,
4356    OP_VDUP2,
4357    OP_VDUP3,
4358    OP_VEXT1,
4359    OP_VEXT2,
4360    OP_VEXT3,
4361    OP_VUZPL, // VUZP, left result
4362    OP_VUZPR, // VUZP, right result
4363    OP_VZIPL, // VZIP, left result
4364    OP_VZIPR, // VZIP, right result
4365    OP_VTRNL, // VTRN, left result
4366    OP_VTRNR  // VTRN, right result
4367  };
4368
4369  if (OpNum == OP_COPY) {
4370    if (LHSID == (1*9+2)*9+3) return LHS;
4371    assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4372    return RHS;
4373  }
4374
4375  SDValue OpLHS, OpRHS;
4376  OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4377  OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4378  EVT VT = OpLHS.getValueType();
4379
4380  switch (OpNum) {
4381  default: llvm_unreachable("Unknown shuffle opcode!");
4382  case OP_VREV:
4383    // VREV divides the vector in half and swaps within the half.
4384    if (VT.getVectorElementType() == MVT::i32 ||
4385        VT.getVectorElementType() == MVT::f32)
4386      return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4387    // vrev <4 x i16> -> VREV32
4388    if (VT.getVectorElementType() == MVT::i16)
4389      return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4390    // vrev <4 x i8> -> VREV16
4391    assert(VT.getVectorElementType() == MVT::i8);
4392    return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4393  case OP_VDUP0:
4394  case OP_VDUP1:
4395  case OP_VDUP2:
4396  case OP_VDUP3:
4397    return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4398                       OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4399  case OP_VEXT1:
4400  case OP_VEXT2:
4401  case OP_VEXT3:
4402    return DAG.getNode(ARMISD::VEXT, dl, VT,
4403                       OpLHS, OpRHS,
4404                       DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4405  case OP_VUZPL:
4406  case OP_VUZPR:
4407    return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4408                       OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4409  case OP_VZIPL:
4410  case OP_VZIPR:
4411    return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4412                       OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4413  case OP_VTRNL:
4414  case OP_VTRNR:
4415    return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4416                       OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4417  }
4418}
4419
4420static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4421                                       ArrayRef<int> ShuffleMask,
4422                                       SelectionDAG &DAG) {
4423  // Check to see if we can use the VTBL instruction.
4424  SDValue V1 = Op.getOperand(0);
4425  SDValue V2 = Op.getOperand(1);
4426  DebugLoc DL = Op.getDebugLoc();
4427
4428  SmallVector<SDValue, 8> VTBLMask;
4429  for (ArrayRef<int>::iterator
4430         I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4431    VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4432
4433  if (V2.getNode()->getOpcode() == ISD::UNDEF)
4434    return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4435                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4436                                   &VTBLMask[0], 8));
4437
4438  return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4439                     DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4440                                 &VTBLMask[0], 8));
4441}
4442
4443static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4444  SDValue V1 = Op.getOperand(0);
4445  SDValue V2 = Op.getOperand(1);
4446  DebugLoc dl = Op.getDebugLoc();
4447  EVT VT = Op.getValueType();
4448  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4449
4450  // Convert shuffles that are directly supported on NEON to target-specific
4451  // DAG nodes, instead of keeping them as shuffles and matching them again
4452  // during code selection.  This is more efficient and avoids the possibility
4453  // of inconsistencies between legalization and selection.
4454  // FIXME: floating-point vectors should be canonicalized to integer vectors
4455  // of the same time so that they get CSEd properly.
4456  ArrayRef<int> ShuffleMask = SVN->getMask();
4457
4458  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4459  if (EltSize <= 32) {
4460    if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4461      int Lane = SVN->getSplatIndex();
4462      // If this is undef splat, generate it via "just" vdup, if possible.
4463      if (Lane == -1) Lane = 0;
4464
4465      // Test if V1 is a SCALAR_TO_VECTOR.
4466      if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4467        return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4468      }
4469      // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4470      // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4471      // reaches it).
4472      if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4473          !isa<ConstantSDNode>(V1.getOperand(0))) {
4474        bool IsScalarToVector = true;
4475        for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4476          if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4477            IsScalarToVector = false;
4478            break;
4479          }
4480        if (IsScalarToVector)
4481          return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4482      }
4483      return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4484                         DAG.getConstant(Lane, MVT::i32));
4485    }
4486
4487    bool ReverseVEXT;
4488    unsigned Imm;
4489    if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4490      if (ReverseVEXT)
4491        std::swap(V1, V2);
4492      return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4493                         DAG.getConstant(Imm, MVT::i32));
4494    }
4495
4496    if (isVREVMask(ShuffleMask, VT, 64))
4497      return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4498    if (isVREVMask(ShuffleMask, VT, 32))
4499      return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4500    if (isVREVMask(ShuffleMask, VT, 16))
4501      return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4502
4503    // Check for Neon shuffles that modify both input vectors in place.
4504    // If both results are used, i.e., if there are two shuffles with the same
4505    // source operands and with masks corresponding to both results of one of
4506    // these operations, DAG memoization will ensure that a single node is
4507    // used for both shuffles.
4508    unsigned WhichResult;
4509    if (isVTRNMask(ShuffleMask, VT, WhichResult))
4510      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4511                         V1, V2).getValue(WhichResult);
4512    if (isVUZPMask(ShuffleMask, VT, WhichResult))
4513      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4514                         V1, V2).getValue(WhichResult);
4515    if (isVZIPMask(ShuffleMask, VT, WhichResult))
4516      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4517                         V1, V2).getValue(WhichResult);
4518
4519    if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4520      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4521                         V1, V1).getValue(WhichResult);
4522    if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4523      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4524                         V1, V1).getValue(WhichResult);
4525    if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4526      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4527                         V1, V1).getValue(WhichResult);
4528  }
4529
4530  // If the shuffle is not directly supported and it has 4 elements, use
4531  // the PerfectShuffle-generated table to synthesize it from other shuffles.
4532  unsigned NumElts = VT.getVectorNumElements();
4533  if (NumElts == 4) {
4534    unsigned PFIndexes[4];
4535    for (unsigned i = 0; i != 4; ++i) {
4536      if (ShuffleMask[i] < 0)
4537        PFIndexes[i] = 8;
4538      else
4539        PFIndexes[i] = ShuffleMask[i];
4540    }
4541
4542    // Compute the index in the perfect shuffle table.
4543    unsigned PFTableIndex =
4544      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4545    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4546    unsigned Cost = (PFEntry >> 30);
4547
4548    if (Cost <= 4)
4549      return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4550  }
4551
4552  // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4553  if (EltSize >= 32) {
4554    // Do the expansion with floating-point types, since that is what the VFP
4555    // registers are defined to use, and since i64 is not legal.
4556    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4557    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4558    V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4559    V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4560    SmallVector<SDValue, 8> Ops;
4561    for (unsigned i = 0; i < NumElts; ++i) {
4562      if (ShuffleMask[i] < 0)
4563        Ops.push_back(DAG.getUNDEF(EltVT));
4564      else
4565        Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4566                                  ShuffleMask[i] < (int)NumElts ? V1 : V2,
4567                                  DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4568                                                  MVT::i32)));
4569    }
4570    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4571    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4572  }
4573
4574  if (VT == MVT::v8i8) {
4575    SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4576    if (NewOp.getNode())
4577      return NewOp;
4578  }
4579
4580  return SDValue();
4581}
4582
4583static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4584  // INSERT_VECTOR_ELT is legal only for immediate indexes.
4585  SDValue Lane = Op.getOperand(2);
4586  if (!isa<ConstantSDNode>(Lane))
4587    return SDValue();
4588
4589  return Op;
4590}
4591
4592static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4593  // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4594  SDValue Lane = Op.getOperand(1);
4595  if (!isa<ConstantSDNode>(Lane))
4596    return SDValue();
4597
4598  SDValue Vec = Op.getOperand(0);
4599  if (Op.getValueType() == MVT::i32 &&
4600      Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4601    DebugLoc dl = Op.getDebugLoc();
4602    return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4603  }
4604
4605  return Op;
4606}
4607
4608static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4609  // The only time a CONCAT_VECTORS operation can have legal types is when
4610  // two 64-bit vectors are concatenated to a 128-bit vector.
4611  assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4612         "unexpected CONCAT_VECTORS");
4613  DebugLoc dl = Op.getDebugLoc();
4614  SDValue Val = DAG.getUNDEF(MVT::v2f64);
4615  SDValue Op0 = Op.getOperand(0);
4616  SDValue Op1 = Op.getOperand(1);
4617  if (Op0.getOpcode() != ISD::UNDEF)
4618    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4619                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4620                      DAG.getIntPtrConstant(0));
4621  if (Op1.getOpcode() != ISD::UNDEF)
4622    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4623                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4624                      DAG.getIntPtrConstant(1));
4625  return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4626}
4627
4628/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4629/// element has been zero/sign-extended, depending on the isSigned parameter,
4630/// from an integer type half its size.
4631static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4632                                   bool isSigned) {
4633  // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4634  EVT VT = N->getValueType(0);
4635  if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4636    SDNode *BVN = N->getOperand(0).getNode();
4637    if (BVN->getValueType(0) != MVT::v4i32 ||
4638        BVN->getOpcode() != ISD::BUILD_VECTOR)
4639      return false;
4640    unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4641    unsigned HiElt = 1 - LoElt;
4642    ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4643    ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4644    ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4645    ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4646    if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4647      return false;
4648    if (isSigned) {
4649      if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4650          Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4651        return true;
4652    } else {
4653      if (Hi0->isNullValue() && Hi1->isNullValue())
4654        return true;
4655    }
4656    return false;
4657  }
4658
4659  if (N->getOpcode() != ISD::BUILD_VECTOR)
4660    return false;
4661
4662  for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4663    SDNode *Elt = N->getOperand(i).getNode();
4664    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4665      unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4666      unsigned HalfSize = EltSize / 2;
4667      if (isSigned) {
4668        if (!isIntN(HalfSize, C->getSExtValue()))
4669          return false;
4670      } else {
4671        if (!isUIntN(HalfSize, C->getZExtValue()))
4672          return false;
4673      }
4674      continue;
4675    }
4676    return false;
4677  }
4678
4679  return true;
4680}
4681
4682/// isSignExtended - Check if a node is a vector value that is sign-extended
4683/// or a constant BUILD_VECTOR with sign-extended elements.
4684static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4685  if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4686    return true;
4687  if (isExtendedBUILD_VECTOR(N, DAG, true))
4688    return true;
4689  return false;
4690}
4691
4692/// isZeroExtended - Check if a node is a vector value that is zero-extended
4693/// or a constant BUILD_VECTOR with zero-extended elements.
4694static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4695  if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4696    return true;
4697  if (isExtendedBUILD_VECTOR(N, DAG, false))
4698    return true;
4699  return false;
4700}
4701
4702/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4703/// load, or BUILD_VECTOR with extended elements, return the unextended value.
4704static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4705  if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4706    return N->getOperand(0);
4707  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4708    return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4709                       LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4710                       LD->isNonTemporal(), LD->isInvariant(),
4711                       LD->getAlignment());
4712  // Otherwise, the value must be a BUILD_VECTOR.  For v2i64, it will
4713  // have been legalized as a BITCAST from v4i32.
4714  if (N->getOpcode() == ISD::BITCAST) {
4715    SDNode *BVN = N->getOperand(0).getNode();
4716    assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4717           BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4718    unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4719    return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4720                       BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4721  }
4722  // Construct a new BUILD_VECTOR with elements truncated to half the size.
4723  assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4724  EVT VT = N->getValueType(0);
4725  unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4726  unsigned NumElts = VT.getVectorNumElements();
4727  MVT TruncVT = MVT::getIntegerVT(EltSize);
4728  SmallVector<SDValue, 8> Ops;
4729  for (unsigned i = 0; i != NumElts; ++i) {
4730    ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4731    const APInt &CInt = C->getAPIntValue();
4732    Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4733  }
4734  return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4735                     MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4736}
4737
4738static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4739  unsigned Opcode = N->getOpcode();
4740  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4741    SDNode *N0 = N->getOperand(0).getNode();
4742    SDNode *N1 = N->getOperand(1).getNode();
4743    return N0->hasOneUse() && N1->hasOneUse() &&
4744      isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4745  }
4746  return false;
4747}
4748
4749static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4750  unsigned Opcode = N->getOpcode();
4751  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4752    SDNode *N0 = N->getOperand(0).getNode();
4753    SDNode *N1 = N->getOperand(1).getNode();
4754    return N0->hasOneUse() && N1->hasOneUse() &&
4755      isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4756  }
4757  return false;
4758}
4759
4760static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4761  // Multiplications are only custom-lowered for 128-bit vectors so that
4762  // VMULL can be detected.  Otherwise v2i64 multiplications are not legal.
4763  EVT VT = Op.getValueType();
4764  assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4765  SDNode *N0 = Op.getOperand(0).getNode();
4766  SDNode *N1 = Op.getOperand(1).getNode();
4767  unsigned NewOpc = 0;
4768  bool isMLA = false;
4769  bool isN0SExt = isSignExtended(N0, DAG);
4770  bool isN1SExt = isSignExtended(N1, DAG);
4771  if (isN0SExt && isN1SExt)
4772    NewOpc = ARMISD::VMULLs;
4773  else {
4774    bool isN0ZExt = isZeroExtended(N0, DAG);
4775    bool isN1ZExt = isZeroExtended(N1, DAG);
4776    if (isN0ZExt && isN1ZExt)
4777      NewOpc = ARMISD::VMULLu;
4778    else if (isN1SExt || isN1ZExt) {
4779      // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4780      // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4781      if (isN1SExt && isAddSubSExt(N0, DAG)) {
4782        NewOpc = ARMISD::VMULLs;
4783        isMLA = true;
4784      } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4785        NewOpc = ARMISD::VMULLu;
4786        isMLA = true;
4787      } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4788        std::swap(N0, N1);
4789        NewOpc = ARMISD::VMULLu;
4790        isMLA = true;
4791      }
4792    }
4793
4794    if (!NewOpc) {
4795      if (VT == MVT::v2i64)
4796        // Fall through to expand this.  It is not legal.
4797        return SDValue();
4798      else
4799        // Other vector multiplications are legal.
4800        return Op;
4801    }
4802  }
4803
4804  // Legalize to a VMULL instruction.
4805  DebugLoc DL = Op.getDebugLoc();
4806  SDValue Op0;
4807  SDValue Op1 = SkipExtension(N1, DAG);
4808  if (!isMLA) {
4809    Op0 = SkipExtension(N0, DAG);
4810    assert(Op0.getValueType().is64BitVector() &&
4811           Op1.getValueType().is64BitVector() &&
4812           "unexpected types for extended operands to VMULL");
4813    return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4814  }
4815
4816  // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4817  // isel lowering to take advantage of no-stall back to back vmul + vmla.
4818  //   vmull q0, d4, d6
4819  //   vmlal q0, d5, d6
4820  // is faster than
4821  //   vaddl q0, d4, d5
4822  //   vmovl q1, d6
4823  //   vmul  q0, q0, q1
4824  SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4825  SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4826  EVT Op1VT = Op1.getValueType();
4827  return DAG.getNode(N0->getOpcode(), DL, VT,
4828                     DAG.getNode(NewOpc, DL, VT,
4829                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4830                     DAG.getNode(NewOpc, DL, VT,
4831                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4832}
4833
4834static SDValue
4835LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4836  // Convert to float
4837  // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4838  // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4839  X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4840  Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4841  X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4842  Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4843  // Get reciprocal estimate.
4844  // float4 recip = vrecpeq_f32(yf);
4845  Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4846                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4847  // Because char has a smaller range than uchar, we can actually get away
4848  // without any newton steps.  This requires that we use a weird bias
4849  // of 0xb000, however (again, this has been exhaustively tested).
4850  // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4851  X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4852  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4853  Y = DAG.getConstant(0xb000, MVT::i32);
4854  Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4855  X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4856  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4857  // Convert back to short.
4858  X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4859  X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4860  return X;
4861}
4862
4863static SDValue
4864LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4865  SDValue N2;
4866  // Convert to float.
4867  // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4868  // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4869  N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4870  N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4871  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4872  N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4873
4874  // Use reciprocal estimate and one refinement step.
4875  // float4 recip = vrecpeq_f32(yf);
4876  // recip *= vrecpsq_f32(yf, recip);
4877  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4878                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4879  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4880                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4881                   N1, N2);
4882  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4883  // Because short has a smaller range than ushort, we can actually get away
4884  // with only a single newton step.  This requires that we use a weird bias
4885  // of 89, however (again, this has been exhaustively tested).
4886  // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4887  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4888  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4889  N1 = DAG.getConstant(0x89, MVT::i32);
4890  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4891  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4892  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4893  // Convert back to integer and return.
4894  // return vmovn_s32(vcvt_s32_f32(result));
4895  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4896  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4897  return N0;
4898}
4899
4900static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4901  EVT VT = Op.getValueType();
4902  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4903         "unexpected type for custom-lowering ISD::SDIV");
4904
4905  DebugLoc dl = Op.getDebugLoc();
4906  SDValue N0 = Op.getOperand(0);
4907  SDValue N1 = Op.getOperand(1);
4908  SDValue N2, N3;
4909
4910  if (VT == MVT::v8i8) {
4911    N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4912    N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4913
4914    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4915                     DAG.getIntPtrConstant(4));
4916    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4917                     DAG.getIntPtrConstant(4));
4918    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4919                     DAG.getIntPtrConstant(0));
4920    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4921                     DAG.getIntPtrConstant(0));
4922
4923    N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4924    N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4925
4926    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4927    N0 = LowerCONCAT_VECTORS(N0, DAG);
4928
4929    N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4930    return N0;
4931  }
4932  return LowerSDIV_v4i16(N0, N1, dl, DAG);
4933}
4934
4935static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4936  EVT VT = Op.getValueType();
4937  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4938         "unexpected type for custom-lowering ISD::UDIV");
4939
4940  DebugLoc dl = Op.getDebugLoc();
4941  SDValue N0 = Op.getOperand(0);
4942  SDValue N1 = Op.getOperand(1);
4943  SDValue N2, N3;
4944
4945  if (VT == MVT::v8i8) {
4946    N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4947    N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4948
4949    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4950                     DAG.getIntPtrConstant(4));
4951    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4952                     DAG.getIntPtrConstant(4));
4953    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4954                     DAG.getIntPtrConstant(0));
4955    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4956                     DAG.getIntPtrConstant(0));
4957
4958    N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4959    N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4960
4961    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4962    N0 = LowerCONCAT_VECTORS(N0, DAG);
4963
4964    N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4965                     DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4966                     N0);
4967    return N0;
4968  }
4969
4970  // v4i16 sdiv ... Convert to float.
4971  // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4972  // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4973  N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4974  N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4975  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4976  SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4977
4978  // Use reciprocal estimate and two refinement steps.
4979  // float4 recip = vrecpeq_f32(yf);
4980  // recip *= vrecpsq_f32(yf, recip);
4981  // recip *= vrecpsq_f32(yf, recip);
4982  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4983                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4984  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4985                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4986                   BN1, N2);
4987  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4988  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4989                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4990                   BN1, N2);
4991  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4992  // Simply multiplying by the reciprocal estimate can leave us a few ulps
4993  // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4994  // and that it will never cause us to return an answer too large).
4995  // float4 result = as_float4(as_int4(xf*recip) + 2);
4996  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4997  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4998  N1 = DAG.getConstant(2, MVT::i32);
4999  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5000  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5001  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5002  // Convert back to integer and return.
5003  // return vmovn_u32(vcvt_s32_f32(result));
5004  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5005  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5006  return N0;
5007}
5008
5009static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5010  EVT VT = Op.getNode()->getValueType(0);
5011  SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5012
5013  unsigned Opc;
5014  bool ExtraOp = false;
5015  switch (Op.getOpcode()) {
5016  default: llvm_unreachable("Invalid code");
5017  case ISD::ADDC: Opc = ARMISD::ADDC; break;
5018  case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5019  case ISD::SUBC: Opc = ARMISD::SUBC; break;
5020  case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5021  }
5022
5023  if (!ExtraOp)
5024    return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5025                       Op.getOperand(1));
5026  return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5027                     Op.getOperand(1), Op.getOperand(2));
5028}
5029
5030static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5031  // Monotonic load/store is legal for all targets
5032  if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5033    return Op;
5034
5035  // Aquire/Release load/store is not legal for targets without a
5036  // dmb or equivalent available.
5037  return SDValue();
5038}
5039
5040
5041static void
5042ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5043                    SelectionDAG &DAG, unsigned NewOp) {
5044  DebugLoc dl = Node->getDebugLoc();
5045  assert (Node->getValueType(0) == MVT::i64 &&
5046          "Only know how to expand i64 atomics");
5047
5048  SmallVector<SDValue, 6> Ops;
5049  Ops.push_back(Node->getOperand(0)); // Chain
5050  Ops.push_back(Node->getOperand(1)); // Ptr
5051  // Low part of Val1
5052  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5053                            Node->getOperand(2), DAG.getIntPtrConstant(0)));
5054  // High part of Val1
5055  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5056                            Node->getOperand(2), DAG.getIntPtrConstant(1)));
5057  if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5058    // High part of Val1
5059    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5060                              Node->getOperand(3), DAG.getIntPtrConstant(0)));
5061    // High part of Val2
5062    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5063                              Node->getOperand(3), DAG.getIntPtrConstant(1)));
5064  }
5065  SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5066  SDValue Result =
5067    DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5068                            cast<MemSDNode>(Node)->getMemOperand());
5069  SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5070  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5071  Results.push_back(Result.getValue(2));
5072}
5073
5074SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5075  switch (Op.getOpcode()) {
5076  default: llvm_unreachable("Don't know how to custom lower this!");
5077  case ISD::ConstantPool:  return LowerConstantPool(Op, DAG);
5078  case ISD::BlockAddress:  return LowerBlockAddress(Op, DAG);
5079  case ISD::GlobalAddress:
5080    return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5081      LowerGlobalAddressELF(Op, DAG);
5082  case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5083  case ISD::SELECT:        return LowerSELECT(Op, DAG);
5084  case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG);
5085  case ISD::BR_CC:         return LowerBR_CC(Op, DAG);
5086  case ISD::BR_JT:         return LowerBR_JT(Op, DAG);
5087  case ISD::VASTART:       return LowerVASTART(Op, DAG);
5088  case ISD::MEMBARRIER:    return LowerMEMBARRIER(Op, DAG, Subtarget);
5089  case ISD::ATOMIC_FENCE:  return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5090  case ISD::PREFETCH:      return LowerPREFETCH(Op, DAG, Subtarget);
5091  case ISD::SINT_TO_FP:
5092  case ISD::UINT_TO_FP:    return LowerINT_TO_FP(Op, DAG);
5093  case ISD::FP_TO_SINT:
5094  case ISD::FP_TO_UINT:    return LowerFP_TO_INT(Op, DAG);
5095  case ISD::FCOPYSIGN:     return LowerFCOPYSIGN(Op, DAG);
5096  case ISD::RETURNADDR:    return LowerRETURNADDR(Op, DAG);
5097  case ISD::FRAMEADDR:     return LowerFRAMEADDR(Op, DAG);
5098  case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5099  case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5100  case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5101  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5102                                                               Subtarget);
5103  case ISD::BITCAST:       return ExpandBITCAST(Op.getNode(), DAG);
5104  case ISD::SHL:
5105  case ISD::SRL:
5106  case ISD::SRA:           return LowerShift(Op.getNode(), DAG, Subtarget);
5107  case ISD::SHL_PARTS:     return LowerShiftLeftParts(Op, DAG);
5108  case ISD::SRL_PARTS:
5109  case ISD::SRA_PARTS:     return LowerShiftRightParts(Op, DAG);
5110  case ISD::CTTZ:          return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5111  case ISD::SETCC:         return LowerVSETCC(Op, DAG);
5112  case ISD::BUILD_VECTOR:  return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5113  case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5114  case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5115  case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5116  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5117  case ISD::FLT_ROUNDS_:   return LowerFLT_ROUNDS_(Op, DAG);
5118  case ISD::MUL:           return LowerMUL(Op, DAG);
5119  case ISD::SDIV:          return LowerSDIV(Op, DAG);
5120  case ISD::UDIV:          return LowerUDIV(Op, DAG);
5121  case ISD::ADDC:
5122  case ISD::ADDE:
5123  case ISD::SUBC:
5124  case ISD::SUBE:          return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5125  case ISD::ATOMIC_LOAD:
5126  case ISD::ATOMIC_STORE:  return LowerAtomicLoadStore(Op, DAG);
5127  }
5128}
5129
5130/// ReplaceNodeResults - Replace the results of node with an illegal result
5131/// type with new values built out of custom code.
5132void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5133                                           SmallVectorImpl<SDValue>&Results,
5134                                           SelectionDAG &DAG) const {
5135  SDValue Res;
5136  switch (N->getOpcode()) {
5137  default:
5138    llvm_unreachable("Don't know how to custom expand this!");
5139  case ISD::BITCAST:
5140    Res = ExpandBITCAST(N, DAG);
5141    break;
5142  case ISD::SRL:
5143  case ISD::SRA:
5144    Res = Expand64BitShift(N, DAG, Subtarget);
5145    break;
5146  case ISD::ATOMIC_LOAD_ADD:
5147    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5148    return;
5149  case ISD::ATOMIC_LOAD_AND:
5150    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5151    return;
5152  case ISD::ATOMIC_LOAD_NAND:
5153    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5154    return;
5155  case ISD::ATOMIC_LOAD_OR:
5156    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5157    return;
5158  case ISD::ATOMIC_LOAD_SUB:
5159    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5160    return;
5161  case ISD::ATOMIC_LOAD_XOR:
5162    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5163    return;
5164  case ISD::ATOMIC_SWAP:
5165    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5166    return;
5167  case ISD::ATOMIC_CMP_SWAP:
5168    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5169    return;
5170  }
5171  if (Res.getNode())
5172    Results.push_back(Res);
5173}
5174
5175//===----------------------------------------------------------------------===//
5176//                           ARM Scheduler Hooks
5177//===----------------------------------------------------------------------===//
5178
5179MachineBasicBlock *
5180ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5181                                     MachineBasicBlock *BB,
5182                                     unsigned Size) const {
5183  unsigned dest    = MI->getOperand(0).getReg();
5184  unsigned ptr     = MI->getOperand(1).getReg();
5185  unsigned oldval  = MI->getOperand(2).getReg();
5186  unsigned newval  = MI->getOperand(3).getReg();
5187  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5188  DebugLoc dl = MI->getDebugLoc();
5189  bool isThumb2 = Subtarget->isThumb2();
5190
5191  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5192  unsigned scratch =
5193    MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5194                                       : ARM::GPRRegisterClass);
5195
5196  if (isThumb2) {
5197    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5198    MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5199    MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5200  }
5201
5202  unsigned ldrOpc, strOpc;
5203  switch (Size) {
5204  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5205  case 1:
5206    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5207    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5208    break;
5209  case 2:
5210    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5211    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5212    break;
5213  case 4:
5214    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5215    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5216    break;
5217  }
5218
5219  MachineFunction *MF = BB->getParent();
5220  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5221  MachineFunction::iterator It = BB;
5222  ++It; // insert the new blocks after the current block
5223
5224  MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5225  MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5226  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5227  MF->insert(It, loop1MBB);
5228  MF->insert(It, loop2MBB);
5229  MF->insert(It, exitMBB);
5230
5231  // Transfer the remainder of BB and its successor edges to exitMBB.
5232  exitMBB->splice(exitMBB->begin(), BB,
5233                  llvm::next(MachineBasicBlock::iterator(MI)),
5234                  BB->end());
5235  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5236
5237  //  thisMBB:
5238  //   ...
5239  //   fallthrough --> loop1MBB
5240  BB->addSuccessor(loop1MBB);
5241
5242  // loop1MBB:
5243  //   ldrex dest, [ptr]
5244  //   cmp dest, oldval
5245  //   bne exitMBB
5246  BB = loop1MBB;
5247  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5248  if (ldrOpc == ARM::t2LDREX)
5249    MIB.addImm(0);
5250  AddDefaultPred(MIB);
5251  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5252                 .addReg(dest).addReg(oldval));
5253  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5254    .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5255  BB->addSuccessor(loop2MBB);
5256  BB->addSuccessor(exitMBB);
5257
5258  // loop2MBB:
5259  //   strex scratch, newval, [ptr]
5260  //   cmp scratch, #0
5261  //   bne loop1MBB
5262  BB = loop2MBB;
5263  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5264  if (strOpc == ARM::t2STREX)
5265    MIB.addImm(0);
5266  AddDefaultPred(MIB);
5267  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5268                 .addReg(scratch).addImm(0));
5269  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5270    .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5271  BB->addSuccessor(loop1MBB);
5272  BB->addSuccessor(exitMBB);
5273
5274  //  exitMBB:
5275  //   ...
5276  BB = exitMBB;
5277
5278  MI->eraseFromParent();   // The instruction is gone now.
5279
5280  return BB;
5281}
5282
5283MachineBasicBlock *
5284ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5285                                    unsigned Size, unsigned BinOpcode) const {
5286  // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5287  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5288
5289  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5290  MachineFunction *MF = BB->getParent();
5291  MachineFunction::iterator It = BB;
5292  ++It;
5293
5294  unsigned dest = MI->getOperand(0).getReg();
5295  unsigned ptr = MI->getOperand(1).getReg();
5296  unsigned incr = MI->getOperand(2).getReg();
5297  DebugLoc dl = MI->getDebugLoc();
5298  bool isThumb2 = Subtarget->isThumb2();
5299
5300  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5301  if (isThumb2) {
5302    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5303    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5304  }
5305
5306  unsigned ldrOpc, strOpc;
5307  switch (Size) {
5308  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5309  case 1:
5310    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5311    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5312    break;
5313  case 2:
5314    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5315    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5316    break;
5317  case 4:
5318    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5319    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5320    break;
5321  }
5322
5323  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5324  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5325  MF->insert(It, loopMBB);
5326  MF->insert(It, exitMBB);
5327
5328  // Transfer the remainder of BB and its successor edges to exitMBB.
5329  exitMBB->splice(exitMBB->begin(), BB,
5330                  llvm::next(MachineBasicBlock::iterator(MI)),
5331                  BB->end());
5332  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5333
5334  const TargetRegisterClass *TRC =
5335    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5336  unsigned scratch = MRI.createVirtualRegister(TRC);
5337  unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5338
5339  //  thisMBB:
5340  //   ...
5341  //   fallthrough --> loopMBB
5342  BB->addSuccessor(loopMBB);
5343
5344  //  loopMBB:
5345  //   ldrex dest, ptr
5346  //   <binop> scratch2, dest, incr
5347  //   strex scratch, scratch2, ptr
5348  //   cmp scratch, #0
5349  //   bne- loopMBB
5350  //   fallthrough --> exitMBB
5351  BB = loopMBB;
5352  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5353  if (ldrOpc == ARM::t2LDREX)
5354    MIB.addImm(0);
5355  AddDefaultPred(MIB);
5356  if (BinOpcode) {
5357    // operand order needs to go the other way for NAND
5358    if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5359      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5360                     addReg(incr).addReg(dest)).addReg(0);
5361    else
5362      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5363                     addReg(dest).addReg(incr)).addReg(0);
5364  }
5365
5366  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5367  if (strOpc == ARM::t2STREX)
5368    MIB.addImm(0);
5369  AddDefaultPred(MIB);
5370  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5371                 .addReg(scratch).addImm(0));
5372  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5373    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5374
5375  BB->addSuccessor(loopMBB);
5376  BB->addSuccessor(exitMBB);
5377
5378  //  exitMBB:
5379  //   ...
5380  BB = exitMBB;
5381
5382  MI->eraseFromParent();   // The instruction is gone now.
5383
5384  return BB;
5385}
5386
5387MachineBasicBlock *
5388ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5389                                          MachineBasicBlock *BB,
5390                                          unsigned Size,
5391                                          bool signExtend,
5392                                          ARMCC::CondCodes Cond) const {
5393  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5394
5395  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5396  MachineFunction *MF = BB->getParent();
5397  MachineFunction::iterator It = BB;
5398  ++It;
5399
5400  unsigned dest = MI->getOperand(0).getReg();
5401  unsigned ptr = MI->getOperand(1).getReg();
5402  unsigned incr = MI->getOperand(2).getReg();
5403  unsigned oldval = dest;
5404  DebugLoc dl = MI->getDebugLoc();
5405  bool isThumb2 = Subtarget->isThumb2();
5406
5407  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5408  if (isThumb2) {
5409    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5410    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5411  }
5412
5413  unsigned ldrOpc, strOpc, extendOpc;
5414  switch (Size) {
5415  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5416  case 1:
5417    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5418    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5419    extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5420    break;
5421  case 2:
5422    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5423    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5424    extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5425    break;
5426  case 4:
5427    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5428    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5429    extendOpc = 0;
5430    break;
5431  }
5432
5433  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5434  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5435  MF->insert(It, loopMBB);
5436  MF->insert(It, exitMBB);
5437
5438  // Transfer the remainder of BB and its successor edges to exitMBB.
5439  exitMBB->splice(exitMBB->begin(), BB,
5440                  llvm::next(MachineBasicBlock::iterator(MI)),
5441                  BB->end());
5442  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5443
5444  const TargetRegisterClass *TRC =
5445    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5446  unsigned scratch = MRI.createVirtualRegister(TRC);
5447  unsigned scratch2 = MRI.createVirtualRegister(TRC);
5448
5449  //  thisMBB:
5450  //   ...
5451  //   fallthrough --> loopMBB
5452  BB->addSuccessor(loopMBB);
5453
5454  //  loopMBB:
5455  //   ldrex dest, ptr
5456  //   (sign extend dest, if required)
5457  //   cmp dest, incr
5458  //   cmov.cond scratch2, dest, incr
5459  //   strex scratch, scratch2, ptr
5460  //   cmp scratch, #0
5461  //   bne- loopMBB
5462  //   fallthrough --> exitMBB
5463  BB = loopMBB;
5464  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5465  if (ldrOpc == ARM::t2LDREX)
5466    MIB.addImm(0);
5467  AddDefaultPred(MIB);
5468
5469  // Sign extend the value, if necessary.
5470  if (signExtend && extendOpc) {
5471    oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5472    AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5473                     .addReg(dest)
5474                     .addImm(0));
5475  }
5476
5477  // Build compare and cmov instructions.
5478  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5479                 .addReg(oldval).addReg(incr));
5480  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5481         .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5482
5483  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5484  if (strOpc == ARM::t2STREX)
5485    MIB.addImm(0);
5486  AddDefaultPred(MIB);
5487  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5488                 .addReg(scratch).addImm(0));
5489  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5490    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5491
5492  BB->addSuccessor(loopMBB);
5493  BB->addSuccessor(exitMBB);
5494
5495  //  exitMBB:
5496  //   ...
5497  BB = exitMBB;
5498
5499  MI->eraseFromParent();   // The instruction is gone now.
5500
5501  return BB;
5502}
5503
5504MachineBasicBlock *
5505ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5506                                      unsigned Op1, unsigned Op2,
5507                                      bool NeedsCarry, bool IsCmpxchg) const {
5508  // This also handles ATOMIC_SWAP, indicated by Op1==0.
5509  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5510
5511  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5512  MachineFunction *MF = BB->getParent();
5513  MachineFunction::iterator It = BB;
5514  ++It;
5515
5516  unsigned destlo = MI->getOperand(0).getReg();
5517  unsigned desthi = MI->getOperand(1).getReg();
5518  unsigned ptr = MI->getOperand(2).getReg();
5519  unsigned vallo = MI->getOperand(3).getReg();
5520  unsigned valhi = MI->getOperand(4).getReg();
5521  DebugLoc dl = MI->getDebugLoc();
5522  bool isThumb2 = Subtarget->isThumb2();
5523
5524  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5525  if (isThumb2) {
5526    MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5527    MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5528    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5529  }
5530
5531  unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5532  unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5533
5534  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5535  MachineBasicBlock *contBB = 0, *cont2BB = 0;
5536  if (IsCmpxchg) {
5537    contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5538    cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5539  }
5540  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5541  MF->insert(It, loopMBB);
5542  if (IsCmpxchg) {
5543    MF->insert(It, contBB);
5544    MF->insert(It, cont2BB);
5545  }
5546  MF->insert(It, exitMBB);
5547
5548  // Transfer the remainder of BB and its successor edges to exitMBB.
5549  exitMBB->splice(exitMBB->begin(), BB,
5550                  llvm::next(MachineBasicBlock::iterator(MI)),
5551                  BB->end());
5552  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5553
5554  const TargetRegisterClass *TRC =
5555    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5556  unsigned storesuccess = MRI.createVirtualRegister(TRC);
5557
5558  //  thisMBB:
5559  //   ...
5560  //   fallthrough --> loopMBB
5561  BB->addSuccessor(loopMBB);
5562
5563  //  loopMBB:
5564  //   ldrexd r2, r3, ptr
5565  //   <binopa> r0, r2, incr
5566  //   <binopb> r1, r3, incr
5567  //   strexd storesuccess, r0, r1, ptr
5568  //   cmp storesuccess, #0
5569  //   bne- loopMBB
5570  //   fallthrough --> exitMBB
5571  //
5572  // Note that the registers are explicitly specified because there is not any
5573  // way to force the register allocator to allocate a register pair.
5574  //
5575  // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5576  // need to properly enforce the restriction that the two output registers
5577  // for ldrexd must be different.
5578  BB = loopMBB;
5579  // Load
5580  AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5581                 .addReg(ARM::R2, RegState::Define)
5582                 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5583  // Copy r2/r3 into dest.  (This copy will normally be coalesced.)
5584  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5585  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5586
5587  if (IsCmpxchg) {
5588    // Add early exit
5589    for (unsigned i = 0; i < 2; i++) {
5590      AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5591                                                         ARM::CMPrr))
5592                     .addReg(i == 0 ? destlo : desthi)
5593                     .addReg(i == 0 ? vallo : valhi));
5594      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5595        .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5596      BB->addSuccessor(exitMBB);
5597      BB->addSuccessor(i == 0 ? contBB : cont2BB);
5598      BB = (i == 0 ? contBB : cont2BB);
5599    }
5600
5601    // Copy to physregs for strexd
5602    unsigned setlo = MI->getOperand(5).getReg();
5603    unsigned sethi = MI->getOperand(6).getReg();
5604    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5605    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5606  } else if (Op1) {
5607    // Perform binary operation
5608    AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5609                   .addReg(destlo).addReg(vallo))
5610        .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5611    AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5612                   .addReg(desthi).addReg(valhi)).addReg(0);
5613  } else {
5614    // Copy to physregs for strexd
5615    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5616    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5617  }
5618
5619  // Store
5620  AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5621                 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5622  // Cmp+jump
5623  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5624                 .addReg(storesuccess).addImm(0));
5625  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5626    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5627
5628  BB->addSuccessor(loopMBB);
5629  BB->addSuccessor(exitMBB);
5630
5631  //  exitMBB:
5632  //   ...
5633  BB = exitMBB;
5634
5635  MI->eraseFromParent();   // The instruction is gone now.
5636
5637  return BB;
5638}
5639
5640/// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5641/// registers the function context.
5642void ARMTargetLowering::
5643SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5644                       MachineBasicBlock *DispatchBB, int FI) const {
5645  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5646  DebugLoc dl = MI->getDebugLoc();
5647  MachineFunction *MF = MBB->getParent();
5648  MachineRegisterInfo *MRI = &MF->getRegInfo();
5649  MachineConstantPool *MCP = MF->getConstantPool();
5650  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5651  const Function *F = MF->getFunction();
5652
5653  bool isThumb = Subtarget->isThumb();
5654  bool isThumb2 = Subtarget->isThumb2();
5655
5656  unsigned PCLabelId = AFI->createPICLabelUId();
5657  unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5658  ARMConstantPoolValue *CPV =
5659    ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5660  unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5661
5662  const TargetRegisterClass *TRC =
5663    isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5664
5665  // Grab constant pool and fixed stack memory operands.
5666  MachineMemOperand *CPMMO =
5667    MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5668                             MachineMemOperand::MOLoad, 4, 4);
5669
5670  MachineMemOperand *FIMMOSt =
5671    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5672                             MachineMemOperand::MOStore, 4, 4);
5673
5674  // Load the address of the dispatch MBB into the jump buffer.
5675  if (isThumb2) {
5676    // Incoming value: jbuf
5677    //   ldr.n  r5, LCPI1_1
5678    //   orr    r5, r5, #1
5679    //   add    r5, pc
5680    //   str    r5, [$jbuf, #+4] ; &jbuf[1]
5681    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5682    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5683                   .addConstantPoolIndex(CPI)
5684                   .addMemOperand(CPMMO));
5685    // Set the low bit because of thumb mode.
5686    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5687    AddDefaultCC(
5688      AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5689                     .addReg(NewVReg1, RegState::Kill)
5690                     .addImm(0x01)));
5691    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5692    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5693      .addReg(NewVReg2, RegState::Kill)
5694      .addImm(PCLabelId);
5695    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5696                   .addReg(NewVReg3, RegState::Kill)
5697                   .addFrameIndex(FI)
5698                   .addImm(36)  // &jbuf[1] :: pc
5699                   .addMemOperand(FIMMOSt));
5700  } else if (isThumb) {
5701    // Incoming value: jbuf
5702    //   ldr.n  r1, LCPI1_4
5703    //   add    r1, pc
5704    //   mov    r2, #1
5705    //   orrs   r1, r2
5706    //   add    r2, $jbuf, #+4 ; &jbuf[1]
5707    //   str    r1, [r2]
5708    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5709    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5710                   .addConstantPoolIndex(CPI)
5711                   .addMemOperand(CPMMO));
5712    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5713    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5714      .addReg(NewVReg1, RegState::Kill)
5715      .addImm(PCLabelId);
5716    // Set the low bit because of thumb mode.
5717    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5718    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5719                   .addReg(ARM::CPSR, RegState::Define)
5720                   .addImm(1));
5721    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5722    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5723                   .addReg(ARM::CPSR, RegState::Define)
5724                   .addReg(NewVReg2, RegState::Kill)
5725                   .addReg(NewVReg3, RegState::Kill));
5726    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5727    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5728                   .addFrameIndex(FI)
5729                   .addImm(36)); // &jbuf[1] :: pc
5730    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5731                   .addReg(NewVReg4, RegState::Kill)
5732                   .addReg(NewVReg5, RegState::Kill)
5733                   .addImm(0)
5734                   .addMemOperand(FIMMOSt));
5735  } else {
5736    // Incoming value: jbuf
5737    //   ldr  r1, LCPI1_1
5738    //   add  r1, pc, r1
5739    //   str  r1, [$jbuf, #+4] ; &jbuf[1]
5740    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5741    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12),  NewVReg1)
5742                   .addConstantPoolIndex(CPI)
5743                   .addImm(0)
5744                   .addMemOperand(CPMMO));
5745    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5746    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5747                   .addReg(NewVReg1, RegState::Kill)
5748                   .addImm(PCLabelId));
5749    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5750                   .addReg(NewVReg2, RegState::Kill)
5751                   .addFrameIndex(FI)
5752                   .addImm(36)  // &jbuf[1] :: pc
5753                   .addMemOperand(FIMMOSt));
5754  }
5755}
5756
5757MachineBasicBlock *ARMTargetLowering::
5758EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5759  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5760  DebugLoc dl = MI->getDebugLoc();
5761  MachineFunction *MF = MBB->getParent();
5762  MachineRegisterInfo *MRI = &MF->getRegInfo();
5763  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5764  MachineFrameInfo *MFI = MF->getFrameInfo();
5765  int FI = MFI->getFunctionContextIndex();
5766
5767  const TargetRegisterClass *TRC =
5768    Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5769
5770  // Get a mapping of the call site numbers to all of the landing pads they're
5771  // associated with.
5772  DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5773  unsigned MaxCSNum = 0;
5774  MachineModuleInfo &MMI = MF->getMMI();
5775  for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5776    if (!BB->isLandingPad()) continue;
5777
5778    // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5779    // pad.
5780    for (MachineBasicBlock::iterator
5781           II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5782      if (!II->isEHLabel()) continue;
5783
5784      MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5785      if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5786
5787      SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5788      for (SmallVectorImpl<unsigned>::iterator
5789             CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5790           CSI != CSE; ++CSI) {
5791        CallSiteNumToLPad[*CSI].push_back(BB);
5792        MaxCSNum = std::max(MaxCSNum, *CSI);
5793      }
5794      break;
5795    }
5796  }
5797
5798  // Get an ordered list of the machine basic blocks for the jump table.
5799  std::vector<MachineBasicBlock*> LPadList;
5800  SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5801  LPadList.reserve(CallSiteNumToLPad.size());
5802  for (unsigned I = 1; I <= MaxCSNum; ++I) {
5803    SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5804    for (SmallVectorImpl<MachineBasicBlock*>::iterator
5805           II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5806      LPadList.push_back(*II);
5807      InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5808    }
5809  }
5810
5811  assert(!LPadList.empty() &&
5812         "No landing pad destinations for the dispatch jump table!");
5813
5814  // Create the jump table and associated information.
5815  MachineJumpTableInfo *JTI =
5816    MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5817  unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5818  unsigned UId = AFI->createJumpTableUId();
5819
5820  // Create the MBBs for the dispatch code.
5821
5822  // Shove the dispatch's address into the return slot in the function context.
5823  MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5824  DispatchBB->setIsLandingPad();
5825
5826  MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5827  BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5828  DispatchBB->addSuccessor(TrapBB);
5829
5830  MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5831  DispatchBB->addSuccessor(DispContBB);
5832
5833  // Insert and MBBs.
5834  MF->insert(MF->end(), DispatchBB);
5835  MF->insert(MF->end(), DispContBB);
5836  MF->insert(MF->end(), TrapBB);
5837
5838  // Insert code into the entry block that creates and registers the function
5839  // context.
5840  SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5841
5842  MachineMemOperand *FIMMOLd =
5843    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5844                             MachineMemOperand::MOLoad |
5845                             MachineMemOperand::MOVolatile, 4, 4);
5846
5847  if (AFI->isThumb1OnlyFunction())
5848    BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5849  else if (!Subtarget->hasVFP2())
5850    BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5851  else
5852    BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5853
5854  unsigned NumLPads = LPadList.size();
5855  if (Subtarget->isThumb2()) {
5856    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5857    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5858                   .addFrameIndex(FI)
5859                   .addImm(4)
5860                   .addMemOperand(FIMMOLd));
5861
5862    if (NumLPads < 256) {
5863      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5864                     .addReg(NewVReg1)
5865                     .addImm(LPadList.size()));
5866    } else {
5867      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5868      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5869                     .addImm(NumLPads & 0xFFFF));
5870
5871      unsigned VReg2 = VReg1;
5872      if ((NumLPads & 0xFFFF0000) != 0) {
5873        VReg2 = MRI->createVirtualRegister(TRC);
5874        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5875                       .addReg(VReg1)
5876                       .addImm(NumLPads >> 16));
5877      }
5878
5879      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5880                     .addReg(NewVReg1)
5881                     .addReg(VReg2));
5882    }
5883
5884    BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5885      .addMBB(TrapBB)
5886      .addImm(ARMCC::HI)
5887      .addReg(ARM::CPSR);
5888
5889    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5890    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5891                   .addJumpTableIndex(MJTI)
5892                   .addImm(UId));
5893
5894    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5895    AddDefaultCC(
5896      AddDefaultPred(
5897        BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5898        .addReg(NewVReg3, RegState::Kill)
5899        .addReg(NewVReg1)
5900        .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5901
5902    BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5903      .addReg(NewVReg4, RegState::Kill)
5904      .addReg(NewVReg1)
5905      .addJumpTableIndex(MJTI)
5906      .addImm(UId);
5907  } else if (Subtarget->isThumb()) {
5908    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5909    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5910                   .addFrameIndex(FI)
5911                   .addImm(1)
5912                   .addMemOperand(FIMMOLd));
5913
5914    if (NumLPads < 256) {
5915      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5916                     .addReg(NewVReg1)
5917                     .addImm(NumLPads));
5918    } else {
5919      MachineConstantPool *ConstantPool = MF->getConstantPool();
5920      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5921      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5922
5923      // MachineConstantPool wants an explicit alignment.
5924      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5925      if (Align == 0)
5926        Align = getTargetData()->getTypeAllocSize(C->getType());
5927      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5928
5929      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5930      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5931                     .addReg(VReg1, RegState::Define)
5932                     .addConstantPoolIndex(Idx));
5933      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5934                     .addReg(NewVReg1)
5935                     .addReg(VReg1));
5936    }
5937
5938    BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5939      .addMBB(TrapBB)
5940      .addImm(ARMCC::HI)
5941      .addReg(ARM::CPSR);
5942
5943    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5944    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5945                   .addReg(ARM::CPSR, RegState::Define)
5946                   .addReg(NewVReg1)
5947                   .addImm(2));
5948
5949    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5950    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5951                   .addJumpTableIndex(MJTI)
5952                   .addImm(UId));
5953
5954    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5955    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5956                   .addReg(ARM::CPSR, RegState::Define)
5957                   .addReg(NewVReg2, RegState::Kill)
5958                   .addReg(NewVReg3));
5959
5960    MachineMemOperand *JTMMOLd =
5961      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5962                               MachineMemOperand::MOLoad, 4, 4);
5963
5964    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5965    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5966                   .addReg(NewVReg4, RegState::Kill)
5967                   .addImm(0)
5968                   .addMemOperand(JTMMOLd));
5969
5970    unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5971    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5972                   .addReg(ARM::CPSR, RegState::Define)
5973                   .addReg(NewVReg5, RegState::Kill)
5974                   .addReg(NewVReg3));
5975
5976    BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5977      .addReg(NewVReg6, RegState::Kill)
5978      .addJumpTableIndex(MJTI)
5979      .addImm(UId);
5980  } else {
5981    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5982    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5983                   .addFrameIndex(FI)
5984                   .addImm(4)
5985                   .addMemOperand(FIMMOLd));
5986
5987    if (NumLPads < 256) {
5988      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5989                     .addReg(NewVReg1)
5990                     .addImm(NumLPads));
5991    } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5992      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5993      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5994                     .addImm(NumLPads & 0xFFFF));
5995
5996      unsigned VReg2 = VReg1;
5997      if ((NumLPads & 0xFFFF0000) != 0) {
5998        VReg2 = MRI->createVirtualRegister(TRC);
5999        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6000                       .addReg(VReg1)
6001                       .addImm(NumLPads >> 16));
6002      }
6003
6004      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6005                     .addReg(NewVReg1)
6006                     .addReg(VReg2));
6007    } else {
6008      MachineConstantPool *ConstantPool = MF->getConstantPool();
6009      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6010      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6011
6012      // MachineConstantPool wants an explicit alignment.
6013      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6014      if (Align == 0)
6015        Align = getTargetData()->getTypeAllocSize(C->getType());
6016      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6017
6018      unsigned VReg1 = MRI->createVirtualRegister(TRC);
6019      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6020                     .addReg(VReg1, RegState::Define)
6021                     .addConstantPoolIndex(Idx)
6022                     .addImm(0));
6023      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6024                     .addReg(NewVReg1)
6025                     .addReg(VReg1, RegState::Kill));
6026    }
6027
6028    BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6029      .addMBB(TrapBB)
6030      .addImm(ARMCC::HI)
6031      .addReg(ARM::CPSR);
6032
6033    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6034    AddDefaultCC(
6035      AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6036                     .addReg(NewVReg1)
6037                     .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6038    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6039    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6040                   .addJumpTableIndex(MJTI)
6041                   .addImm(UId));
6042
6043    MachineMemOperand *JTMMOLd =
6044      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6045                               MachineMemOperand::MOLoad, 4, 4);
6046    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6047    AddDefaultPred(
6048      BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6049      .addReg(NewVReg3, RegState::Kill)
6050      .addReg(NewVReg4)
6051      .addImm(0)
6052      .addMemOperand(JTMMOLd));
6053
6054    BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6055      .addReg(NewVReg5, RegState::Kill)
6056      .addReg(NewVReg4)
6057      .addJumpTableIndex(MJTI)
6058      .addImm(UId);
6059  }
6060
6061  // Add the jump table entries as successors to the MBB.
6062  MachineBasicBlock *PrevMBB = 0;
6063  for (std::vector<MachineBasicBlock*>::iterator
6064         I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6065    MachineBasicBlock *CurMBB = *I;
6066    if (PrevMBB != CurMBB)
6067      DispContBB->addSuccessor(CurMBB);
6068    PrevMBB = CurMBB;
6069  }
6070
6071  // N.B. the order the invoke BBs are processed in doesn't matter here.
6072  const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6073  const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6074  const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
6075  SmallVector<MachineBasicBlock*, 64> MBBLPads;
6076  for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6077         I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6078    MachineBasicBlock *BB = *I;
6079
6080    // Remove the landing pad successor from the invoke block and replace it
6081    // with the new dispatch block.
6082    SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6083                                                  BB->succ_end());
6084    while (!Successors.empty()) {
6085      MachineBasicBlock *SMBB = Successors.pop_back_val();
6086      if (SMBB->isLandingPad()) {
6087        BB->removeSuccessor(SMBB);
6088        MBBLPads.push_back(SMBB);
6089      }
6090    }
6091
6092    BB->addSuccessor(DispatchBB);
6093
6094    // Find the invoke call and mark all of the callee-saved registers as
6095    // 'implicit defined' so that they're spilled. This prevents code from
6096    // moving instructions to before the EH block, where they will never be
6097    // executed.
6098    for (MachineBasicBlock::reverse_iterator
6099           II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6100      if (!II->isCall()) continue;
6101
6102      DenseMap<unsigned, bool> DefRegs;
6103      for (MachineInstr::mop_iterator
6104             OI = II->operands_begin(), OE = II->operands_end();
6105           OI != OE; ++OI) {
6106        if (!OI->isReg()) continue;
6107        DefRegs[OI->getReg()] = true;
6108      }
6109
6110      MachineInstrBuilder MIB(&*II);
6111
6112      for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6113        unsigned Reg = SavedRegs[i];
6114        if (Subtarget->isThumb2() &&
6115            !ARM::tGPRRegisterClass->contains(Reg) &&
6116            !ARM::hGPRRegisterClass->contains(Reg))
6117          continue;
6118        else if (Subtarget->isThumb1Only() &&
6119                 !ARM::tGPRRegisterClass->contains(Reg))
6120          continue;
6121        else if (!Subtarget->isThumb() &&
6122                 !ARM::GPRRegisterClass->contains(Reg))
6123          continue;
6124        if (!DefRegs[Reg])
6125          MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6126      }
6127
6128      break;
6129    }
6130  }
6131
6132  // Mark all former landing pads as non-landing pads. The dispatch is the only
6133  // landing pad now.
6134  for (SmallVectorImpl<MachineBasicBlock*>::iterator
6135         I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6136    (*I)->setIsLandingPad(false);
6137
6138  // The instruction is gone now.
6139  MI->eraseFromParent();
6140
6141  return MBB;
6142}
6143
6144static
6145MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6146  for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6147       E = MBB->succ_end(); I != E; ++I)
6148    if (*I != Succ)
6149      return *I;
6150  llvm_unreachable("Expecting a BB with two successors!");
6151}
6152
6153MachineBasicBlock *
6154ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6155                                               MachineBasicBlock *BB) const {
6156  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6157  DebugLoc dl = MI->getDebugLoc();
6158  bool isThumb2 = Subtarget->isThumb2();
6159  switch (MI->getOpcode()) {
6160  default: {
6161    MI->dump();
6162    llvm_unreachable("Unexpected instr type to insert");
6163  }
6164  // The Thumb2 pre-indexed stores have the same MI operands, they just
6165  // define them differently in the .td files from the isel patterns, so
6166  // they need pseudos.
6167  case ARM::t2STR_preidx:
6168    MI->setDesc(TII->get(ARM::t2STR_PRE));
6169    return BB;
6170  case ARM::t2STRB_preidx:
6171    MI->setDesc(TII->get(ARM::t2STRB_PRE));
6172    return BB;
6173  case ARM::t2STRH_preidx:
6174    MI->setDesc(TII->get(ARM::t2STRH_PRE));
6175    return BB;
6176
6177  case ARM::STRi_preidx:
6178  case ARM::STRBi_preidx: {
6179    unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6180      ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6181    // Decode the offset.
6182    unsigned Offset = MI->getOperand(4).getImm();
6183    bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6184    Offset = ARM_AM::getAM2Offset(Offset);
6185    if (isSub)
6186      Offset = -Offset;
6187
6188    MachineMemOperand *MMO = *MI->memoperands_begin();
6189    BuildMI(*BB, MI, dl, TII->get(NewOpc))
6190      .addOperand(MI->getOperand(0))  // Rn_wb
6191      .addOperand(MI->getOperand(1))  // Rt
6192      .addOperand(MI->getOperand(2))  // Rn
6193      .addImm(Offset)                 // offset (skip GPR==zero_reg)
6194      .addOperand(MI->getOperand(5))  // pred
6195      .addOperand(MI->getOperand(6))
6196      .addMemOperand(MMO);
6197    MI->eraseFromParent();
6198    return BB;
6199  }
6200  case ARM::STRr_preidx:
6201  case ARM::STRBr_preidx:
6202  case ARM::STRH_preidx: {
6203    unsigned NewOpc;
6204    switch (MI->getOpcode()) {
6205    default: llvm_unreachable("unexpected opcode!");
6206    case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6207    case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6208    case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6209    }
6210    MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6211    for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6212      MIB.addOperand(MI->getOperand(i));
6213    MI->eraseFromParent();
6214    return BB;
6215  }
6216  case ARM::ATOMIC_LOAD_ADD_I8:
6217     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6218  case ARM::ATOMIC_LOAD_ADD_I16:
6219     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6220  case ARM::ATOMIC_LOAD_ADD_I32:
6221     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6222
6223  case ARM::ATOMIC_LOAD_AND_I8:
6224     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6225  case ARM::ATOMIC_LOAD_AND_I16:
6226     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6227  case ARM::ATOMIC_LOAD_AND_I32:
6228     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6229
6230  case ARM::ATOMIC_LOAD_OR_I8:
6231     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6232  case ARM::ATOMIC_LOAD_OR_I16:
6233     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6234  case ARM::ATOMIC_LOAD_OR_I32:
6235     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6236
6237  case ARM::ATOMIC_LOAD_XOR_I8:
6238     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6239  case ARM::ATOMIC_LOAD_XOR_I16:
6240     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6241  case ARM::ATOMIC_LOAD_XOR_I32:
6242     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6243
6244  case ARM::ATOMIC_LOAD_NAND_I8:
6245     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6246  case ARM::ATOMIC_LOAD_NAND_I16:
6247     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6248  case ARM::ATOMIC_LOAD_NAND_I32:
6249     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6250
6251  case ARM::ATOMIC_LOAD_SUB_I8:
6252     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6253  case ARM::ATOMIC_LOAD_SUB_I16:
6254     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6255  case ARM::ATOMIC_LOAD_SUB_I32:
6256     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6257
6258  case ARM::ATOMIC_LOAD_MIN_I8:
6259     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6260  case ARM::ATOMIC_LOAD_MIN_I16:
6261     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6262  case ARM::ATOMIC_LOAD_MIN_I32:
6263     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6264
6265  case ARM::ATOMIC_LOAD_MAX_I8:
6266     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6267  case ARM::ATOMIC_LOAD_MAX_I16:
6268     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6269  case ARM::ATOMIC_LOAD_MAX_I32:
6270     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6271
6272  case ARM::ATOMIC_LOAD_UMIN_I8:
6273     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6274  case ARM::ATOMIC_LOAD_UMIN_I16:
6275     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6276  case ARM::ATOMIC_LOAD_UMIN_I32:
6277     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6278
6279  case ARM::ATOMIC_LOAD_UMAX_I8:
6280     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6281  case ARM::ATOMIC_LOAD_UMAX_I16:
6282     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6283  case ARM::ATOMIC_LOAD_UMAX_I32:
6284     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6285
6286  case ARM::ATOMIC_SWAP_I8:  return EmitAtomicBinary(MI, BB, 1, 0);
6287  case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6288  case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6289
6290  case ARM::ATOMIC_CMP_SWAP_I8:  return EmitAtomicCmpSwap(MI, BB, 1);
6291  case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6292  case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6293
6294
6295  case ARM::ATOMADD6432:
6296    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6297                              isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6298                              /*NeedsCarry*/ true);
6299  case ARM::ATOMSUB6432:
6300    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6301                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6302                              /*NeedsCarry*/ true);
6303  case ARM::ATOMOR6432:
6304    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6305                              isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6306  case ARM::ATOMXOR6432:
6307    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6308                              isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6309  case ARM::ATOMAND6432:
6310    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6311                              isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6312  case ARM::ATOMSWAP6432:
6313    return EmitAtomicBinary64(MI, BB, 0, 0, false);
6314  case ARM::ATOMCMPXCHG6432:
6315    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6316                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6317                              /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6318
6319  case ARM::tMOVCCr_pseudo: {
6320    // To "insert" a SELECT_CC instruction, we actually have to insert the
6321    // diamond control-flow pattern.  The incoming instruction knows the
6322    // destination vreg to set, the condition code register to branch on, the
6323    // true/false values to select between, and a branch opcode to use.
6324    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6325    MachineFunction::iterator It = BB;
6326    ++It;
6327
6328    //  thisMBB:
6329    //  ...
6330    //   TrueVal = ...
6331    //   cmpTY ccX, r1, r2
6332    //   bCC copy1MBB
6333    //   fallthrough --> copy0MBB
6334    MachineBasicBlock *thisMBB  = BB;
6335    MachineFunction *F = BB->getParent();
6336    MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6337    MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
6338    F->insert(It, copy0MBB);
6339    F->insert(It, sinkMBB);
6340
6341    // Transfer the remainder of BB and its successor edges to sinkMBB.
6342    sinkMBB->splice(sinkMBB->begin(), BB,
6343                    llvm::next(MachineBasicBlock::iterator(MI)),
6344                    BB->end());
6345    sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6346
6347    BB->addSuccessor(copy0MBB);
6348    BB->addSuccessor(sinkMBB);
6349
6350    BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6351      .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6352
6353    //  copy0MBB:
6354    //   %FalseValue = ...
6355    //   # fallthrough to sinkMBB
6356    BB = copy0MBB;
6357
6358    // Update machine-CFG edges
6359    BB->addSuccessor(sinkMBB);
6360
6361    //  sinkMBB:
6362    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6363    //  ...
6364    BB = sinkMBB;
6365    BuildMI(*BB, BB->begin(), dl,
6366            TII->get(ARM::PHI), MI->getOperand(0).getReg())
6367      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6368      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6369
6370    MI->eraseFromParent();   // The pseudo instruction is gone now.
6371    return BB;
6372  }
6373
6374  case ARM::BCCi64:
6375  case ARM::BCCZi64: {
6376    // If there is an unconditional branch to the other successor, remove it.
6377    BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6378
6379    // Compare both parts that make up the double comparison separately for
6380    // equality.
6381    bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6382
6383    unsigned LHS1 = MI->getOperand(1).getReg();
6384    unsigned LHS2 = MI->getOperand(2).getReg();
6385    if (RHSisZero) {
6386      AddDefaultPred(BuildMI(BB, dl,
6387                             TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6388                     .addReg(LHS1).addImm(0));
6389      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6390        .addReg(LHS2).addImm(0)
6391        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6392    } else {
6393      unsigned RHS1 = MI->getOperand(3).getReg();
6394      unsigned RHS2 = MI->getOperand(4).getReg();
6395      AddDefaultPred(BuildMI(BB, dl,
6396                             TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6397                     .addReg(LHS1).addReg(RHS1));
6398      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6399        .addReg(LHS2).addReg(RHS2)
6400        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6401    }
6402
6403    MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6404    MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6405    if (MI->getOperand(0).getImm() == ARMCC::NE)
6406      std::swap(destMBB, exitMBB);
6407
6408    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6409      .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6410    if (isThumb2)
6411      AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6412    else
6413      BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6414
6415    MI->eraseFromParent();   // The pseudo instruction is gone now.
6416    return BB;
6417  }
6418
6419  case ARM::Int_eh_sjlj_setjmp:
6420  case ARM::Int_eh_sjlj_setjmp_nofp:
6421  case ARM::tInt_eh_sjlj_setjmp:
6422  case ARM::t2Int_eh_sjlj_setjmp:
6423  case ARM::t2Int_eh_sjlj_setjmp_nofp:
6424    EmitSjLjDispatchBlock(MI, BB);
6425    return BB;
6426
6427  case ARM::ABS:
6428  case ARM::t2ABS: {
6429    // To insert an ABS instruction, we have to insert the
6430    // diamond control-flow pattern.  The incoming instruction knows the
6431    // source vreg to test against 0, the destination vreg to set,
6432    // the condition code register to branch on, the
6433    // true/false values to select between, and a branch opcode to use.
6434    // It transforms
6435    //     V1 = ABS V0
6436    // into
6437    //     V2 = MOVS V0
6438    //     BCC                      (branch to SinkBB if V0 >= 0)
6439    //     RSBBB: V3 = RSBri V2, 0  (compute ABS if V2 < 0)
6440    //     SinkBB: V1 = PHI(V2, V3)
6441    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6442    MachineFunction::iterator BBI = BB;
6443    ++BBI;
6444    MachineFunction *Fn = BB->getParent();
6445    MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6446    MachineBasicBlock *SinkBB  = Fn->CreateMachineBasicBlock(LLVM_BB);
6447    Fn->insert(BBI, RSBBB);
6448    Fn->insert(BBI, SinkBB);
6449
6450    unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6451    unsigned int ABSDstReg = MI->getOperand(0).getReg();
6452    bool isThumb2 = Subtarget->isThumb2();
6453    MachineRegisterInfo &MRI = Fn->getRegInfo();
6454    // In Thumb mode S must not be specified if source register is the SP or
6455    // PC and if destination register is the SP, so restrict register class
6456    unsigned NewMovDstReg = MRI.createVirtualRegister(
6457      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6458    unsigned NewRsbDstReg = MRI.createVirtualRegister(
6459      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6460
6461    // Transfer the remainder of BB and its successor edges to sinkMBB.
6462    SinkBB->splice(SinkBB->begin(), BB,
6463      llvm::next(MachineBasicBlock::iterator(MI)),
6464      BB->end());
6465    SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6466
6467    BB->addSuccessor(RSBBB);
6468    BB->addSuccessor(SinkBB);
6469
6470    // fall through to SinkMBB
6471    RSBBB->addSuccessor(SinkBB);
6472
6473    // insert a movs at the end of BB
6474    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6475      NewMovDstReg)
6476      .addReg(ABSSrcReg, RegState::Kill)
6477      .addImm((unsigned)ARMCC::AL).addReg(0)
6478      .addReg(ARM::CPSR, RegState::Define);
6479
6480    // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6481    BuildMI(BB, dl,
6482      TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6483      .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6484
6485    // insert rsbri in RSBBB
6486    // Note: BCC and rsbri will be converted into predicated rsbmi
6487    // by if-conversion pass
6488    BuildMI(*RSBBB, RSBBB->begin(), dl,
6489      TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6490      .addReg(NewMovDstReg, RegState::Kill)
6491      .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6492
6493    // insert PHI in SinkBB,
6494    // reuse ABSDstReg to not change uses of ABS instruction
6495    BuildMI(*SinkBB, SinkBB->begin(), dl,
6496      TII->get(ARM::PHI), ABSDstReg)
6497      .addReg(NewRsbDstReg).addMBB(RSBBB)
6498      .addReg(NewMovDstReg).addMBB(BB);
6499
6500    // remove ABS instruction
6501    MI->eraseFromParent();
6502
6503    // return last added BB
6504    return SinkBB;
6505  }
6506  }
6507}
6508
6509void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6510                                                      SDNode *Node) const {
6511  if (!MI->hasPostISelHook()) {
6512    assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6513           "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6514    return;
6515  }
6516
6517  const MCInstrDesc *MCID = &MI->getDesc();
6518  // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6519  // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6520  // operand is still set to noreg. If needed, set the optional operand's
6521  // register to CPSR, and remove the redundant implicit def.
6522  //
6523  // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6524
6525  // Rename pseudo opcodes.
6526  unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6527  if (NewOpc) {
6528    const ARMBaseInstrInfo *TII =
6529      static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6530    MCID = &TII->get(NewOpc);
6531
6532    assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6533           "converted opcode should be the same except for cc_out");
6534
6535    MI->setDesc(*MCID);
6536
6537    // Add the optional cc_out operand
6538    MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6539  }
6540  unsigned ccOutIdx = MCID->getNumOperands() - 1;
6541
6542  // Any ARM instruction that sets the 's' bit should specify an optional
6543  // "cc_out" operand in the last operand position.
6544  if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6545    assert(!NewOpc && "Optional cc_out operand required");
6546    return;
6547  }
6548  // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6549  // since we already have an optional CPSR def.
6550  bool definesCPSR = false;
6551  bool deadCPSR = false;
6552  for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6553       i != e; ++i) {
6554    const MachineOperand &MO = MI->getOperand(i);
6555    if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6556      definesCPSR = true;
6557      if (MO.isDead())
6558        deadCPSR = true;
6559      MI->RemoveOperand(i);
6560      break;
6561    }
6562  }
6563  if (!definesCPSR) {
6564    assert(!NewOpc && "Optional cc_out operand required");
6565    return;
6566  }
6567  assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6568  if (deadCPSR) {
6569    assert(!MI->getOperand(ccOutIdx).getReg() &&
6570           "expect uninitialized optional cc_out operand");
6571    return;
6572  }
6573
6574  // If this instruction was defined with an optional CPSR def and its dag node
6575  // had a live implicit CPSR def, then activate the optional CPSR def.
6576  MachineOperand &MO = MI->getOperand(ccOutIdx);
6577  MO.setReg(ARM::CPSR);
6578  MO.setIsDef(true);
6579}
6580
6581//===----------------------------------------------------------------------===//
6582//                           ARM Optimization Hooks
6583//===----------------------------------------------------------------------===//
6584
6585static
6586SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6587                            TargetLowering::DAGCombinerInfo &DCI) {
6588  SelectionDAG &DAG = DCI.DAG;
6589  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6590  EVT VT = N->getValueType(0);
6591  unsigned Opc = N->getOpcode();
6592  bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6593  SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6594  SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6595  ISD::CondCode CC = ISD::SETCC_INVALID;
6596
6597  if (isSlctCC) {
6598    CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6599  } else {
6600    SDValue CCOp = Slct.getOperand(0);
6601    if (CCOp.getOpcode() == ISD::SETCC)
6602      CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6603  }
6604
6605  bool DoXform = false;
6606  bool InvCC = false;
6607  assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6608          "Bad input!");
6609
6610  if (LHS.getOpcode() == ISD::Constant &&
6611      cast<ConstantSDNode>(LHS)->isNullValue()) {
6612    DoXform = true;
6613  } else if (CC != ISD::SETCC_INVALID &&
6614             RHS.getOpcode() == ISD::Constant &&
6615             cast<ConstantSDNode>(RHS)->isNullValue()) {
6616    std::swap(LHS, RHS);
6617    SDValue Op0 = Slct.getOperand(0);
6618    EVT OpVT = isSlctCC ? Op0.getValueType() :
6619                          Op0.getOperand(0).getValueType();
6620    bool isInt = OpVT.isInteger();
6621    CC = ISD::getSetCCInverse(CC, isInt);
6622
6623    if (!TLI.isCondCodeLegal(CC, OpVT))
6624      return SDValue();         // Inverse operator isn't legal.
6625
6626    DoXform = true;
6627    InvCC = true;
6628  }
6629
6630  if (DoXform) {
6631    SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6632    if (isSlctCC)
6633      return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6634                             Slct.getOperand(0), Slct.getOperand(1), CC);
6635    SDValue CCOp = Slct.getOperand(0);
6636    if (InvCC)
6637      CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6638                          CCOp.getOperand(0), CCOp.getOperand(1), CC);
6639    return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6640                       CCOp, OtherOp, Result);
6641  }
6642  return SDValue();
6643}
6644
6645// AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6646// (only after legalization).
6647static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6648                                 TargetLowering::DAGCombinerInfo &DCI,
6649                                 const ARMSubtarget *Subtarget) {
6650
6651  // Only perform optimization if after legalize, and if NEON is available. We
6652  // also expected both operands to be BUILD_VECTORs.
6653  if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6654      || N0.getOpcode() != ISD::BUILD_VECTOR
6655      || N1.getOpcode() != ISD::BUILD_VECTOR)
6656    return SDValue();
6657
6658  // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6659  EVT VT = N->getValueType(0);
6660  if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6661    return SDValue();
6662
6663  // Check that the vector operands are of the right form.
6664  // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6665  // operands, where N is the size of the formed vector.
6666  // Each EXTRACT_VECTOR should have the same input vector and odd or even
6667  // index such that we have a pair wise add pattern.
6668
6669  // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6670  if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6671    return SDValue();
6672  SDValue Vec = N0->getOperand(0)->getOperand(0);
6673  SDNode *V = Vec.getNode();
6674  unsigned nextIndex = 0;
6675
6676  // For each operands to the ADD which are BUILD_VECTORs,
6677  // check to see if each of their operands are an EXTRACT_VECTOR with
6678  // the same vector and appropriate index.
6679  for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6680    if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6681        && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6682
6683      SDValue ExtVec0 = N0->getOperand(i);
6684      SDValue ExtVec1 = N1->getOperand(i);
6685
6686      // First operand is the vector, verify its the same.
6687      if (V != ExtVec0->getOperand(0).getNode() ||
6688          V != ExtVec1->getOperand(0).getNode())
6689        return SDValue();
6690
6691      // Second is the constant, verify its correct.
6692      ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6693      ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6694
6695      // For the constant, we want to see all the even or all the odd.
6696      if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6697          || C1->getZExtValue() != nextIndex+1)
6698        return SDValue();
6699
6700      // Increment index.
6701      nextIndex+=2;
6702    } else
6703      return SDValue();
6704  }
6705
6706  // Create VPADDL node.
6707  SelectionDAG &DAG = DCI.DAG;
6708  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6709
6710  // Build operand list.
6711  SmallVector<SDValue, 8> Ops;
6712  Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6713                                TLI.getPointerTy()));
6714
6715  // Input is the vector.
6716  Ops.push_back(Vec);
6717
6718  // Get widened type and narrowed type.
6719  MVT widenType;
6720  unsigned numElem = VT.getVectorNumElements();
6721  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6722    case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6723    case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6724    case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6725    default:
6726      llvm_unreachable("Invalid vector element type for padd optimization.");
6727  }
6728
6729  SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6730                            widenType, &Ops[0], Ops.size());
6731  return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6732}
6733
6734/// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6735/// operands N0 and N1.  This is a helper for PerformADDCombine that is
6736/// called with the default operands, and if that fails, with commuted
6737/// operands.
6738static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6739                                          TargetLowering::DAGCombinerInfo &DCI,
6740                                          const ARMSubtarget *Subtarget){
6741
6742  // Attempt to create vpaddl for this add.
6743  SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6744  if (Result.getNode())
6745    return Result;
6746
6747  // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6748  if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6749    SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6750    if (Result.getNode()) return Result;
6751  }
6752  return SDValue();
6753}
6754
6755/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6756///
6757static SDValue PerformADDCombine(SDNode *N,
6758                                 TargetLowering::DAGCombinerInfo &DCI,
6759                                 const ARMSubtarget *Subtarget) {
6760  SDValue N0 = N->getOperand(0);
6761  SDValue N1 = N->getOperand(1);
6762
6763  // First try with the default operand order.
6764  SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6765  if (Result.getNode())
6766    return Result;
6767
6768  // If that didn't work, try again with the operands commuted.
6769  return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6770}
6771
6772/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6773///
6774static SDValue PerformSUBCombine(SDNode *N,
6775                                 TargetLowering::DAGCombinerInfo &DCI) {
6776  SDValue N0 = N->getOperand(0);
6777  SDValue N1 = N->getOperand(1);
6778
6779  // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6780  if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6781    SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6782    if (Result.getNode()) return Result;
6783  }
6784
6785  return SDValue();
6786}
6787
6788/// PerformVMULCombine
6789/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6790/// special multiplier accumulator forwarding.
6791///   vmul d3, d0, d2
6792///   vmla d3, d1, d2
6793/// is faster than
6794///   vadd d3, d0, d1
6795///   vmul d3, d3, d2
6796static SDValue PerformVMULCombine(SDNode *N,
6797                                  TargetLowering::DAGCombinerInfo &DCI,
6798                                  const ARMSubtarget *Subtarget) {
6799  if (!Subtarget->hasVMLxForwarding())
6800    return SDValue();
6801
6802  SelectionDAG &DAG = DCI.DAG;
6803  SDValue N0 = N->getOperand(0);
6804  SDValue N1 = N->getOperand(1);
6805  unsigned Opcode = N0.getOpcode();
6806  if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6807      Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6808    Opcode = N1.getOpcode();
6809    if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6810        Opcode != ISD::FADD && Opcode != ISD::FSUB)
6811      return SDValue();
6812    std::swap(N0, N1);
6813  }
6814
6815  EVT VT = N->getValueType(0);
6816  DebugLoc DL = N->getDebugLoc();
6817  SDValue N00 = N0->getOperand(0);
6818  SDValue N01 = N0->getOperand(1);
6819  return DAG.getNode(Opcode, DL, VT,
6820                     DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6821                     DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6822}
6823
6824static SDValue PerformMULCombine(SDNode *N,
6825                                 TargetLowering::DAGCombinerInfo &DCI,
6826                                 const ARMSubtarget *Subtarget) {
6827  SelectionDAG &DAG = DCI.DAG;
6828
6829  if (Subtarget->isThumb1Only())
6830    return SDValue();
6831
6832  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6833    return SDValue();
6834
6835  EVT VT = N->getValueType(0);
6836  if (VT.is64BitVector() || VT.is128BitVector())
6837    return PerformVMULCombine(N, DCI, Subtarget);
6838  if (VT != MVT::i32)
6839    return SDValue();
6840
6841  ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6842  if (!C)
6843    return SDValue();
6844
6845  uint64_t MulAmt = C->getZExtValue();
6846  unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6847  ShiftAmt = ShiftAmt & (32 - 1);
6848  SDValue V = N->getOperand(0);
6849  DebugLoc DL = N->getDebugLoc();
6850
6851  SDValue Res;
6852  MulAmt >>= ShiftAmt;
6853  if (isPowerOf2_32(MulAmt - 1)) {
6854    // (mul x, 2^N + 1) => (add (shl x, N), x)
6855    Res = DAG.getNode(ISD::ADD, DL, VT,
6856                      V, DAG.getNode(ISD::SHL, DL, VT,
6857                                     V, DAG.getConstant(Log2_32(MulAmt-1),
6858                                                        MVT::i32)));
6859  } else if (isPowerOf2_32(MulAmt + 1)) {
6860    // (mul x, 2^N - 1) => (sub (shl x, N), x)
6861    Res = DAG.getNode(ISD::SUB, DL, VT,
6862                      DAG.getNode(ISD::SHL, DL, VT,
6863                                  V, DAG.getConstant(Log2_32(MulAmt+1),
6864                                                     MVT::i32)),
6865                                                     V);
6866  } else
6867    return SDValue();
6868
6869  if (ShiftAmt != 0)
6870    Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6871                      DAG.getConstant(ShiftAmt, MVT::i32));
6872
6873  // Do not add new nodes to DAG combiner worklist.
6874  DCI.CombineTo(N, Res, false);
6875  return SDValue();
6876}
6877
6878static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6879  if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6880    return false;
6881
6882  SDValue FalseVal = N.getOperand(0);
6883  ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6884  if (!C)
6885    return false;
6886  if (AllOnes)
6887    return C->isAllOnesValue();
6888  return C->isNullValue();
6889}
6890
6891/// formConditionalOp - Combine an operation with a conditional move operand
6892/// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
6893/// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
6894static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
6895                                 bool Commutable) {
6896  SDValue N0 = N->getOperand(0);
6897  SDValue N1 = N->getOperand(1);
6898
6899  bool isAND = N->getOpcode() == ISD::AND;
6900  bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
6901  if (!isCand && Commutable) {
6902    isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
6903    if (isCand)
6904      std::swap(N0, N1);
6905  }
6906  if (!isCand)
6907    return SDValue();
6908
6909  unsigned Opc = 0;
6910  switch (N->getOpcode()) {
6911  default: llvm_unreachable("Unexpected node");
6912  case ISD::AND: Opc = ARMISD::CAND; break;
6913  case ISD::OR:  Opc = ARMISD::COR; break;
6914  case ISD::XOR: Opc = ARMISD::CXOR; break;
6915  }
6916  return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
6917                     N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
6918                     N1.getOperand(4));
6919}
6920
6921static SDValue PerformANDCombine(SDNode *N,
6922                                 TargetLowering::DAGCombinerInfo &DCI,
6923                                 const ARMSubtarget *Subtarget) {
6924
6925  // Attempt to use immediate-form VBIC
6926  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6927  DebugLoc dl = N->getDebugLoc();
6928  EVT VT = N->getValueType(0);
6929  SelectionDAG &DAG = DCI.DAG;
6930
6931  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6932    return SDValue();
6933
6934  APInt SplatBits, SplatUndef;
6935  unsigned SplatBitSize;
6936  bool HasAnyUndefs;
6937  if (BVN &&
6938      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6939    if (SplatBitSize <= 64) {
6940      EVT VbicVT;
6941      SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6942                                      SplatUndef.getZExtValue(), SplatBitSize,
6943                                      DAG, VbicVT, VT.is128BitVector(),
6944                                      OtherModImm);
6945      if (Val.getNode()) {
6946        SDValue Input =
6947          DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6948        SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6949        return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6950      }
6951    }
6952  }
6953
6954  if (!Subtarget->isThumb1Only()) {
6955    // (and x, (cmov -1, y, cond)) => (and.cond x, y)
6956    SDValue CAND = formConditionalOp(N, DAG, true);
6957    if (CAND.getNode())
6958      return CAND;
6959  }
6960
6961  return SDValue();
6962}
6963
6964/// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6965static SDValue PerformORCombine(SDNode *N,
6966                                TargetLowering::DAGCombinerInfo &DCI,
6967                                const ARMSubtarget *Subtarget) {
6968  // Attempt to use immediate-form VORR
6969  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6970  DebugLoc dl = N->getDebugLoc();
6971  EVT VT = N->getValueType(0);
6972  SelectionDAG &DAG = DCI.DAG;
6973
6974  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6975    return SDValue();
6976
6977  APInt SplatBits, SplatUndef;
6978  unsigned SplatBitSize;
6979  bool HasAnyUndefs;
6980  if (BVN && Subtarget->hasNEON() &&
6981      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6982    if (SplatBitSize <= 64) {
6983      EVT VorrVT;
6984      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6985                                      SplatUndef.getZExtValue(), SplatBitSize,
6986                                      DAG, VorrVT, VT.is128BitVector(),
6987                                      OtherModImm);
6988      if (Val.getNode()) {
6989        SDValue Input =
6990          DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6991        SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6992        return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6993      }
6994    }
6995  }
6996
6997  if (!Subtarget->isThumb1Only()) {
6998    // (or x, (cmov 0, y, cond)) => (or.cond x, y)
6999    SDValue COR = formConditionalOp(N, DAG, true);
7000    if (COR.getNode())
7001      return COR;
7002  }
7003
7004  SDValue N0 = N->getOperand(0);
7005  if (N0.getOpcode() != ISD::AND)
7006    return SDValue();
7007  SDValue N1 = N->getOperand(1);
7008
7009  // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7010  if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7011      DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7012    APInt SplatUndef;
7013    unsigned SplatBitSize;
7014    bool HasAnyUndefs;
7015
7016    BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7017    APInt SplatBits0;
7018    if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7019                                  HasAnyUndefs) && !HasAnyUndefs) {
7020      BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7021      APInt SplatBits1;
7022      if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7023                                    HasAnyUndefs) && !HasAnyUndefs &&
7024          SplatBits0 == ~SplatBits1) {
7025        // Canonicalize the vector type to make instruction selection simpler.
7026        EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7027        SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7028                                     N0->getOperand(1), N0->getOperand(0),
7029                                     N1->getOperand(0));
7030        return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7031      }
7032    }
7033  }
7034
7035  // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7036  // reasonable.
7037
7038  // BFI is only available on V6T2+
7039  if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7040    return SDValue();
7041
7042  DebugLoc DL = N->getDebugLoc();
7043  // 1) or (and A, mask), val => ARMbfi A, val, mask
7044  //      iff (val & mask) == val
7045  //
7046  // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7047  //  2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7048  //          && mask == ~mask2
7049  //  2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7050  //          && ~mask == mask2
7051  //  (i.e., copy a bitfield value into another bitfield of the same width)
7052
7053  if (VT != MVT::i32)
7054    return SDValue();
7055
7056  SDValue N00 = N0.getOperand(0);
7057
7058  // The value and the mask need to be constants so we can verify this is
7059  // actually a bitfield set. If the mask is 0xffff, we can do better
7060  // via a movt instruction, so don't use BFI in that case.
7061  SDValue MaskOp = N0.getOperand(1);
7062  ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7063  if (!MaskC)
7064    return SDValue();
7065  unsigned Mask = MaskC->getZExtValue();
7066  if (Mask == 0xffff)
7067    return SDValue();
7068  SDValue Res;
7069  // Case (1): or (and A, mask), val => ARMbfi A, val, mask
7070  ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7071  if (N1C) {
7072    unsigned Val = N1C->getZExtValue();
7073    if ((Val & ~Mask) != Val)
7074      return SDValue();
7075
7076    if (ARM::isBitFieldInvertedMask(Mask)) {
7077      Val >>= CountTrailingZeros_32(~Mask);
7078
7079      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7080                        DAG.getConstant(Val, MVT::i32),
7081                        DAG.getConstant(Mask, MVT::i32));
7082
7083      // Do not add new nodes to DAG combiner worklist.
7084      DCI.CombineTo(N, Res, false);
7085      return SDValue();
7086    }
7087  } else if (N1.getOpcode() == ISD::AND) {
7088    // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7089    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7090    if (!N11C)
7091      return SDValue();
7092    unsigned Mask2 = N11C->getZExtValue();
7093
7094    // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7095    // as is to match.
7096    if (ARM::isBitFieldInvertedMask(Mask) &&
7097        (Mask == ~Mask2)) {
7098      // The pack halfword instruction works better for masks that fit it,
7099      // so use that when it's available.
7100      if (Subtarget->hasT2ExtractPack() &&
7101          (Mask == 0xffff || Mask == 0xffff0000))
7102        return SDValue();
7103      // 2a
7104      unsigned amt = CountTrailingZeros_32(Mask2);
7105      Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7106                        DAG.getConstant(amt, MVT::i32));
7107      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7108                        DAG.getConstant(Mask, MVT::i32));
7109      // Do not add new nodes to DAG combiner worklist.
7110      DCI.CombineTo(N, Res, false);
7111      return SDValue();
7112    } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7113               (~Mask == Mask2)) {
7114      // The pack halfword instruction works better for masks that fit it,
7115      // so use that when it's available.
7116      if (Subtarget->hasT2ExtractPack() &&
7117          (Mask2 == 0xffff || Mask2 == 0xffff0000))
7118        return SDValue();
7119      // 2b
7120      unsigned lsb = CountTrailingZeros_32(Mask);
7121      Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7122                        DAG.getConstant(lsb, MVT::i32));
7123      Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7124                        DAG.getConstant(Mask2, MVT::i32));
7125      // Do not add new nodes to DAG combiner worklist.
7126      DCI.CombineTo(N, Res, false);
7127      return SDValue();
7128    }
7129  }
7130
7131  if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7132      N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7133      ARM::isBitFieldInvertedMask(~Mask)) {
7134    // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7135    // where lsb(mask) == #shamt and masked bits of B are known zero.
7136    SDValue ShAmt = N00.getOperand(1);
7137    unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7138    unsigned LSB = CountTrailingZeros_32(Mask);
7139    if (ShAmtC != LSB)
7140      return SDValue();
7141
7142    Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7143                      DAG.getConstant(~Mask, MVT::i32));
7144
7145    // Do not add new nodes to DAG combiner worklist.
7146    DCI.CombineTo(N, Res, false);
7147  }
7148
7149  return SDValue();
7150}
7151
7152static SDValue PerformXORCombine(SDNode *N,
7153                                 TargetLowering::DAGCombinerInfo &DCI,
7154                                 const ARMSubtarget *Subtarget) {
7155  EVT VT = N->getValueType(0);
7156  SelectionDAG &DAG = DCI.DAG;
7157
7158  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7159    return SDValue();
7160
7161  if (!Subtarget->isThumb1Only()) {
7162    // (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7163    SDValue CXOR = formConditionalOp(N, DAG, true);
7164    if (CXOR.getNode())
7165      return CXOR;
7166  }
7167
7168  return SDValue();
7169}
7170
7171/// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7172/// the bits being cleared by the AND are not demanded by the BFI.
7173static SDValue PerformBFICombine(SDNode *N,
7174                                 TargetLowering::DAGCombinerInfo &DCI) {
7175  SDValue N1 = N->getOperand(1);
7176  if (N1.getOpcode() == ISD::AND) {
7177    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7178    if (!N11C)
7179      return SDValue();
7180    unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7181    unsigned LSB = CountTrailingZeros_32(~InvMask);
7182    unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7183    unsigned Mask = (1 << Width)-1;
7184    unsigned Mask2 = N11C->getZExtValue();
7185    if ((Mask & (~Mask2)) == 0)
7186      return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7187                             N->getOperand(0), N1.getOperand(0),
7188                             N->getOperand(2));
7189  }
7190  return SDValue();
7191}
7192
7193/// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7194/// ARMISD::VMOVRRD.
7195static SDValue PerformVMOVRRDCombine(SDNode *N,
7196                                     TargetLowering::DAGCombinerInfo &DCI) {
7197  // vmovrrd(vmovdrr x, y) -> x,y
7198  SDValue InDouble = N->getOperand(0);
7199  if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7200    return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7201
7202  // vmovrrd(load f64) -> (load i32), (load i32)
7203  SDNode *InNode = InDouble.getNode();
7204  if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7205      InNode->getValueType(0) == MVT::f64 &&
7206      InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7207      !cast<LoadSDNode>(InNode)->isVolatile()) {
7208    // TODO: Should this be done for non-FrameIndex operands?
7209    LoadSDNode *LD = cast<LoadSDNode>(InNode);
7210
7211    SelectionDAG &DAG = DCI.DAG;
7212    DebugLoc DL = LD->getDebugLoc();
7213    SDValue BasePtr = LD->getBasePtr();
7214    SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7215                                 LD->getPointerInfo(), LD->isVolatile(),
7216                                 LD->isNonTemporal(), LD->isInvariant(),
7217                                 LD->getAlignment());
7218
7219    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7220                                    DAG.getConstant(4, MVT::i32));
7221    SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7222                                 LD->getPointerInfo(), LD->isVolatile(),
7223                                 LD->isNonTemporal(), LD->isInvariant(),
7224                                 std::min(4U, LD->getAlignment() / 2));
7225
7226    DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7227    SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7228    DCI.RemoveFromWorklist(LD);
7229    DAG.DeleteNode(LD);
7230    return Result;
7231  }
7232
7233  return SDValue();
7234}
7235
7236/// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7237/// ARMISD::VMOVDRR.  This is also used for BUILD_VECTORs with 2 operands.
7238static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7239  // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7240  SDValue Op0 = N->getOperand(0);
7241  SDValue Op1 = N->getOperand(1);
7242  if (Op0.getOpcode() == ISD::BITCAST)
7243    Op0 = Op0.getOperand(0);
7244  if (Op1.getOpcode() == ISD::BITCAST)
7245    Op1 = Op1.getOperand(0);
7246  if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7247      Op0.getNode() == Op1.getNode() &&
7248      Op0.getResNo() == 0 && Op1.getResNo() == 1)
7249    return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7250                       N->getValueType(0), Op0.getOperand(0));
7251  return SDValue();
7252}
7253
7254/// PerformSTORECombine - Target-specific dag combine xforms for
7255/// ISD::STORE.
7256static SDValue PerformSTORECombine(SDNode *N,
7257                                   TargetLowering::DAGCombinerInfo &DCI) {
7258  // Bitcast an i64 store extracted from a vector to f64.
7259  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7260  StoreSDNode *St = cast<StoreSDNode>(N);
7261  SDValue StVal = St->getValue();
7262  if (!ISD::isNormalStore(St) || St->isVolatile())
7263    return SDValue();
7264
7265  if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7266      StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7267    SelectionDAG  &DAG = DCI.DAG;
7268    DebugLoc DL = St->getDebugLoc();
7269    SDValue BasePtr = St->getBasePtr();
7270    SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7271                                  StVal.getNode()->getOperand(0), BasePtr,
7272                                  St->getPointerInfo(), St->isVolatile(),
7273                                  St->isNonTemporal(), St->getAlignment());
7274
7275    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7276                                    DAG.getConstant(4, MVT::i32));
7277    return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7278                        OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7279                        St->isNonTemporal(),
7280                        std::min(4U, St->getAlignment() / 2));
7281  }
7282
7283  if (StVal.getValueType() != MVT::i64 ||
7284      StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7285    return SDValue();
7286
7287  SelectionDAG &DAG = DCI.DAG;
7288  DebugLoc dl = StVal.getDebugLoc();
7289  SDValue IntVec = StVal.getOperand(0);
7290  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7291                                 IntVec.getValueType().getVectorNumElements());
7292  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7293  SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7294                               Vec, StVal.getOperand(1));
7295  dl = N->getDebugLoc();
7296  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7297  // Make the DAGCombiner fold the bitcasts.
7298  DCI.AddToWorklist(Vec.getNode());
7299  DCI.AddToWorklist(ExtElt.getNode());
7300  DCI.AddToWorklist(V.getNode());
7301  return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7302                      St->getPointerInfo(), St->isVolatile(),
7303                      St->isNonTemporal(), St->getAlignment(),
7304                      St->getTBAAInfo());
7305}
7306
7307/// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7308/// are normal, non-volatile loads.  If so, it is profitable to bitcast an
7309/// i64 vector to have f64 elements, since the value can then be loaded
7310/// directly into a VFP register.
7311static bool hasNormalLoadOperand(SDNode *N) {
7312  unsigned NumElts = N->getValueType(0).getVectorNumElements();
7313  for (unsigned i = 0; i < NumElts; ++i) {
7314    SDNode *Elt = N->getOperand(i).getNode();
7315    if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7316      return true;
7317  }
7318  return false;
7319}
7320
7321/// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7322/// ISD::BUILD_VECTOR.
7323static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7324                                          TargetLowering::DAGCombinerInfo &DCI){
7325  // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7326  // VMOVRRD is introduced when legalizing i64 types.  It forces the i64 value
7327  // into a pair of GPRs, which is fine when the value is used as a scalar,
7328  // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7329  SelectionDAG &DAG = DCI.DAG;
7330  if (N->getNumOperands() == 2) {
7331    SDValue RV = PerformVMOVDRRCombine(N, DAG);
7332    if (RV.getNode())
7333      return RV;
7334  }
7335
7336  // Load i64 elements as f64 values so that type legalization does not split
7337  // them up into i32 values.
7338  EVT VT = N->getValueType(0);
7339  if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7340    return SDValue();
7341  DebugLoc dl = N->getDebugLoc();
7342  SmallVector<SDValue, 8> Ops;
7343  unsigned NumElts = VT.getVectorNumElements();
7344  for (unsigned i = 0; i < NumElts; ++i) {
7345    SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7346    Ops.push_back(V);
7347    // Make the DAGCombiner fold the bitcast.
7348    DCI.AddToWorklist(V.getNode());
7349  }
7350  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7351  SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7352  return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7353}
7354
7355/// PerformInsertEltCombine - Target-specific dag combine xforms for
7356/// ISD::INSERT_VECTOR_ELT.
7357static SDValue PerformInsertEltCombine(SDNode *N,
7358                                       TargetLowering::DAGCombinerInfo &DCI) {
7359  // Bitcast an i64 load inserted into a vector to f64.
7360  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7361  EVT VT = N->getValueType(0);
7362  SDNode *Elt = N->getOperand(1).getNode();
7363  if (VT.getVectorElementType() != MVT::i64 ||
7364      !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7365    return SDValue();
7366
7367  SelectionDAG &DAG = DCI.DAG;
7368  DebugLoc dl = N->getDebugLoc();
7369  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7370                                 VT.getVectorNumElements());
7371  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7372  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7373  // Make the DAGCombiner fold the bitcasts.
7374  DCI.AddToWorklist(Vec.getNode());
7375  DCI.AddToWorklist(V.getNode());
7376  SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7377                               Vec, V, N->getOperand(2));
7378  return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7379}
7380
7381/// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7382/// ISD::VECTOR_SHUFFLE.
7383static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7384  // The LLVM shufflevector instruction does not require the shuffle mask
7385  // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7386  // have that requirement.  When translating to ISD::VECTOR_SHUFFLE, if the
7387  // operands do not match the mask length, they are extended by concatenating
7388  // them with undef vectors.  That is probably the right thing for other
7389  // targets, but for NEON it is better to concatenate two double-register
7390  // size vector operands into a single quad-register size vector.  Do that
7391  // transformation here:
7392  //   shuffle(concat(v1, undef), concat(v2, undef)) ->
7393  //   shuffle(concat(v1, v2), undef)
7394  SDValue Op0 = N->getOperand(0);
7395  SDValue Op1 = N->getOperand(1);
7396  if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7397      Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7398      Op0.getNumOperands() != 2 ||
7399      Op1.getNumOperands() != 2)
7400    return SDValue();
7401  SDValue Concat0Op1 = Op0.getOperand(1);
7402  SDValue Concat1Op1 = Op1.getOperand(1);
7403  if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7404      Concat1Op1.getOpcode() != ISD::UNDEF)
7405    return SDValue();
7406  // Skip the transformation if any of the types are illegal.
7407  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7408  EVT VT = N->getValueType(0);
7409  if (!TLI.isTypeLegal(VT) ||
7410      !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7411      !TLI.isTypeLegal(Concat1Op1.getValueType()))
7412    return SDValue();
7413
7414  SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7415                                  Op0.getOperand(0), Op1.getOperand(0));
7416  // Translate the shuffle mask.
7417  SmallVector<int, 16> NewMask;
7418  unsigned NumElts = VT.getVectorNumElements();
7419  unsigned HalfElts = NumElts/2;
7420  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7421  for (unsigned n = 0; n < NumElts; ++n) {
7422    int MaskElt = SVN->getMaskElt(n);
7423    int NewElt = -1;
7424    if (MaskElt < (int)HalfElts)
7425      NewElt = MaskElt;
7426    else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7427      NewElt = HalfElts + MaskElt - NumElts;
7428    NewMask.push_back(NewElt);
7429  }
7430  return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7431                              DAG.getUNDEF(VT), NewMask.data());
7432}
7433
7434/// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7435/// NEON load/store intrinsics to merge base address updates.
7436static SDValue CombineBaseUpdate(SDNode *N,
7437                                 TargetLowering::DAGCombinerInfo &DCI) {
7438  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7439    return SDValue();
7440
7441  SelectionDAG &DAG = DCI.DAG;
7442  bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7443                      N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7444  unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7445  SDValue Addr = N->getOperand(AddrOpIdx);
7446
7447  // Search for a use of the address operand that is an increment.
7448  for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7449         UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7450    SDNode *User = *UI;
7451    if (User->getOpcode() != ISD::ADD ||
7452        UI.getUse().getResNo() != Addr.getResNo())
7453      continue;
7454
7455    // Check that the add is independent of the load/store.  Otherwise, folding
7456    // it would create a cycle.
7457    if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7458      continue;
7459
7460    // Find the new opcode for the updating load/store.
7461    bool isLoad = true;
7462    bool isLaneOp = false;
7463    unsigned NewOpc = 0;
7464    unsigned NumVecs = 0;
7465    if (isIntrinsic) {
7466      unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7467      switch (IntNo) {
7468      default: llvm_unreachable("unexpected intrinsic for Neon base update");
7469      case Intrinsic::arm_neon_vld1:     NewOpc = ARMISD::VLD1_UPD;
7470        NumVecs = 1; break;
7471      case Intrinsic::arm_neon_vld2:     NewOpc = ARMISD::VLD2_UPD;
7472        NumVecs = 2; break;
7473      case Intrinsic::arm_neon_vld3:     NewOpc = ARMISD::VLD3_UPD;
7474        NumVecs = 3; break;
7475      case Intrinsic::arm_neon_vld4:     NewOpc = ARMISD::VLD4_UPD;
7476        NumVecs = 4; break;
7477      case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7478        NumVecs = 2; isLaneOp = true; break;
7479      case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7480        NumVecs = 3; isLaneOp = true; break;
7481      case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7482        NumVecs = 4; isLaneOp = true; break;
7483      case Intrinsic::arm_neon_vst1:     NewOpc = ARMISD::VST1_UPD;
7484        NumVecs = 1; isLoad = false; break;
7485      case Intrinsic::arm_neon_vst2:     NewOpc = ARMISD::VST2_UPD;
7486        NumVecs = 2; isLoad = false; break;
7487      case Intrinsic::arm_neon_vst3:     NewOpc = ARMISD::VST3_UPD;
7488        NumVecs = 3; isLoad = false; break;
7489      case Intrinsic::arm_neon_vst4:     NewOpc = ARMISD::VST4_UPD;
7490        NumVecs = 4; isLoad = false; break;
7491      case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7492        NumVecs = 2; isLoad = false; isLaneOp = true; break;
7493      case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7494        NumVecs = 3; isLoad = false; isLaneOp = true; break;
7495      case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7496        NumVecs = 4; isLoad = false; isLaneOp = true; break;
7497      }
7498    } else {
7499      isLaneOp = true;
7500      switch (N->getOpcode()) {
7501      default: llvm_unreachable("unexpected opcode for Neon base update");
7502      case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7503      case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7504      case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7505      }
7506    }
7507
7508    // Find the size of memory referenced by the load/store.
7509    EVT VecTy;
7510    if (isLoad)
7511      VecTy = N->getValueType(0);
7512    else
7513      VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7514    unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7515    if (isLaneOp)
7516      NumBytes /= VecTy.getVectorNumElements();
7517
7518    // If the increment is a constant, it must match the memory ref size.
7519    SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7520    if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7521      uint64_t IncVal = CInc->getZExtValue();
7522      if (IncVal != NumBytes)
7523        continue;
7524    } else if (NumBytes >= 3 * 16) {
7525      // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7526      // separate instructions that make it harder to use a non-constant update.
7527      continue;
7528    }
7529
7530    // Create the new updating load/store node.
7531    EVT Tys[6];
7532    unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7533    unsigned n;
7534    for (n = 0; n < NumResultVecs; ++n)
7535      Tys[n] = VecTy;
7536    Tys[n++] = MVT::i32;
7537    Tys[n] = MVT::Other;
7538    SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7539    SmallVector<SDValue, 8> Ops;
7540    Ops.push_back(N->getOperand(0)); // incoming chain
7541    Ops.push_back(N->getOperand(AddrOpIdx));
7542    Ops.push_back(Inc);
7543    for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7544      Ops.push_back(N->getOperand(i));
7545    }
7546    MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7547    SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7548                                           Ops.data(), Ops.size(),
7549                                           MemInt->getMemoryVT(),
7550                                           MemInt->getMemOperand());
7551
7552    // Update the uses.
7553    std::vector<SDValue> NewResults;
7554    for (unsigned i = 0; i < NumResultVecs; ++i) {
7555      NewResults.push_back(SDValue(UpdN.getNode(), i));
7556    }
7557    NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7558    DCI.CombineTo(N, NewResults);
7559    DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7560
7561    break;
7562  }
7563  return SDValue();
7564}
7565
7566/// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7567/// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7568/// are also VDUPLANEs.  If so, combine them to a vldN-dup operation and
7569/// return true.
7570static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7571  SelectionDAG &DAG = DCI.DAG;
7572  EVT VT = N->getValueType(0);
7573  // vldN-dup instructions only support 64-bit vectors for N > 1.
7574  if (!VT.is64BitVector())
7575    return false;
7576
7577  // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7578  SDNode *VLD = N->getOperand(0).getNode();
7579  if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7580    return false;
7581  unsigned NumVecs = 0;
7582  unsigned NewOpc = 0;
7583  unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7584  if (IntNo == Intrinsic::arm_neon_vld2lane) {
7585    NumVecs = 2;
7586    NewOpc = ARMISD::VLD2DUP;
7587  } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7588    NumVecs = 3;
7589    NewOpc = ARMISD::VLD3DUP;
7590  } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7591    NumVecs = 4;
7592    NewOpc = ARMISD::VLD4DUP;
7593  } else {
7594    return false;
7595  }
7596
7597  // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7598  // numbers match the load.
7599  unsigned VLDLaneNo =
7600    cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7601  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7602       UI != UE; ++UI) {
7603    // Ignore uses of the chain result.
7604    if (UI.getUse().getResNo() == NumVecs)
7605      continue;
7606    SDNode *User = *UI;
7607    if (User->getOpcode() != ARMISD::VDUPLANE ||
7608        VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7609      return false;
7610  }
7611
7612  // Create the vldN-dup node.
7613  EVT Tys[5];
7614  unsigned n;
7615  for (n = 0; n < NumVecs; ++n)
7616    Tys[n] = VT;
7617  Tys[n] = MVT::Other;
7618  SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7619  SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7620  MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7621  SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7622                                           Ops, 2, VLDMemInt->getMemoryVT(),
7623                                           VLDMemInt->getMemOperand());
7624
7625  // Update the uses.
7626  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7627       UI != UE; ++UI) {
7628    unsigned ResNo = UI.getUse().getResNo();
7629    // Ignore uses of the chain result.
7630    if (ResNo == NumVecs)
7631      continue;
7632    SDNode *User = *UI;
7633    DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7634  }
7635
7636  // Now the vldN-lane intrinsic is dead except for its chain result.
7637  // Update uses of the chain.
7638  std::vector<SDValue> VLDDupResults;
7639  for (unsigned n = 0; n < NumVecs; ++n)
7640    VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7641  VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7642  DCI.CombineTo(VLD, VLDDupResults);
7643
7644  return true;
7645}
7646
7647/// PerformVDUPLANECombine - Target-specific dag combine xforms for
7648/// ARMISD::VDUPLANE.
7649static SDValue PerformVDUPLANECombine(SDNode *N,
7650                                      TargetLowering::DAGCombinerInfo &DCI) {
7651  SDValue Op = N->getOperand(0);
7652
7653  // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7654  // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7655  if (CombineVLDDUP(N, DCI))
7656    return SDValue(N, 0);
7657
7658  // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7659  // redundant.  Ignore bit_converts for now; element sizes are checked below.
7660  while (Op.getOpcode() == ISD::BITCAST)
7661    Op = Op.getOperand(0);
7662  if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7663    return SDValue();
7664
7665  // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7666  unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7667  // The canonical VMOV for a zero vector uses a 32-bit element size.
7668  unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7669  unsigned EltBits;
7670  if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7671    EltSize = 8;
7672  EVT VT = N->getValueType(0);
7673  if (EltSize > VT.getVectorElementType().getSizeInBits())
7674    return SDValue();
7675
7676  return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7677}
7678
7679// isConstVecPow2 - Return true if each vector element is a power of 2, all
7680// elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7681static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7682{
7683  integerPart cN;
7684  integerPart c0 = 0;
7685  for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7686       I != E; I++) {
7687    ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7688    if (!C)
7689      return false;
7690
7691    bool isExact;
7692    APFloat APF = C->getValueAPF();
7693    if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7694        != APFloat::opOK || !isExact)
7695      return false;
7696
7697    c0 = (I == 0) ? cN : c0;
7698    if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7699      return false;
7700  }
7701  C = c0;
7702  return true;
7703}
7704
7705/// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7706/// can replace combinations of VMUL and VCVT (floating-point to integer)
7707/// when the VMUL has a constant operand that is a power of 2.
7708///
7709/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7710///  vmul.f32        d16, d17, d16
7711///  vcvt.s32.f32    d16, d16
7712/// becomes:
7713///  vcvt.s32.f32    d16, d16, #3
7714static SDValue PerformVCVTCombine(SDNode *N,
7715                                  TargetLowering::DAGCombinerInfo &DCI,
7716                                  const ARMSubtarget *Subtarget) {
7717  SelectionDAG &DAG = DCI.DAG;
7718  SDValue Op = N->getOperand(0);
7719
7720  if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7721      Op.getOpcode() != ISD::FMUL)
7722    return SDValue();
7723
7724  uint64_t C;
7725  SDValue N0 = Op->getOperand(0);
7726  SDValue ConstVec = Op->getOperand(1);
7727  bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7728
7729  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7730      !isConstVecPow2(ConstVec, isSigned, C))
7731    return SDValue();
7732
7733  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7734    Intrinsic::arm_neon_vcvtfp2fxu;
7735  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7736                     N->getValueType(0),
7737                     DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7738                     DAG.getConstant(Log2_64(C), MVT::i32));
7739}
7740
7741/// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7742/// can replace combinations of VCVT (integer to floating-point) and VDIV
7743/// when the VDIV has a constant operand that is a power of 2.
7744///
7745/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7746///  vcvt.f32.s32    d16, d16
7747///  vdiv.f32        d16, d17, d16
7748/// becomes:
7749///  vcvt.f32.s32    d16, d16, #3
7750static SDValue PerformVDIVCombine(SDNode *N,
7751                                  TargetLowering::DAGCombinerInfo &DCI,
7752                                  const ARMSubtarget *Subtarget) {
7753  SelectionDAG &DAG = DCI.DAG;
7754  SDValue Op = N->getOperand(0);
7755  unsigned OpOpcode = Op.getNode()->getOpcode();
7756
7757  if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7758      (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7759    return SDValue();
7760
7761  uint64_t C;
7762  SDValue ConstVec = N->getOperand(1);
7763  bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7764
7765  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7766      !isConstVecPow2(ConstVec, isSigned, C))
7767    return SDValue();
7768
7769  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7770    Intrinsic::arm_neon_vcvtfxu2fp;
7771  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7772                     Op.getValueType(),
7773                     DAG.getConstant(IntrinsicOpcode, MVT::i32),
7774                     Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7775}
7776
7777/// Getvshiftimm - Check if this is a valid build_vector for the immediate
7778/// operand of a vector shift operation, where all the elements of the
7779/// build_vector must have the same constant integer value.
7780static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7781  // Ignore bit_converts.
7782  while (Op.getOpcode() == ISD::BITCAST)
7783    Op = Op.getOperand(0);
7784  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7785  APInt SplatBits, SplatUndef;
7786  unsigned SplatBitSize;
7787  bool HasAnyUndefs;
7788  if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7789                                      HasAnyUndefs, ElementBits) ||
7790      SplatBitSize > ElementBits)
7791    return false;
7792  Cnt = SplatBits.getSExtValue();
7793  return true;
7794}
7795
7796/// isVShiftLImm - Check if this is a valid build_vector for the immediate
7797/// operand of a vector shift left operation.  That value must be in the range:
7798///   0 <= Value < ElementBits for a left shift; or
7799///   0 <= Value <= ElementBits for a long left shift.
7800static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7801  assert(VT.isVector() && "vector shift count is not a vector type");
7802  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7803  if (! getVShiftImm(Op, ElementBits, Cnt))
7804    return false;
7805  return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7806}
7807
7808/// isVShiftRImm - Check if this is a valid build_vector for the immediate
7809/// operand of a vector shift right operation.  For a shift opcode, the value
7810/// is positive, but for an intrinsic the value count must be negative. The
7811/// absolute value must be in the range:
7812///   1 <= |Value| <= ElementBits for a right shift; or
7813///   1 <= |Value| <= ElementBits/2 for a narrow right shift.
7814static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7815                         int64_t &Cnt) {
7816  assert(VT.isVector() && "vector shift count is not a vector type");
7817  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7818  if (! getVShiftImm(Op, ElementBits, Cnt))
7819    return false;
7820  if (isIntrinsic)
7821    Cnt = -Cnt;
7822  return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7823}
7824
7825/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7826static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7827  unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7828  switch (IntNo) {
7829  default:
7830    // Don't do anything for most intrinsics.
7831    break;
7832
7833  // Vector shifts: check for immediate versions and lower them.
7834  // Note: This is done during DAG combining instead of DAG legalizing because
7835  // the build_vectors for 64-bit vector element shift counts are generally
7836  // not legal, and it is hard to see their values after they get legalized to
7837  // loads from a constant pool.
7838  case Intrinsic::arm_neon_vshifts:
7839  case Intrinsic::arm_neon_vshiftu:
7840  case Intrinsic::arm_neon_vshiftls:
7841  case Intrinsic::arm_neon_vshiftlu:
7842  case Intrinsic::arm_neon_vshiftn:
7843  case Intrinsic::arm_neon_vrshifts:
7844  case Intrinsic::arm_neon_vrshiftu:
7845  case Intrinsic::arm_neon_vrshiftn:
7846  case Intrinsic::arm_neon_vqshifts:
7847  case Intrinsic::arm_neon_vqshiftu:
7848  case Intrinsic::arm_neon_vqshiftsu:
7849  case Intrinsic::arm_neon_vqshiftns:
7850  case Intrinsic::arm_neon_vqshiftnu:
7851  case Intrinsic::arm_neon_vqshiftnsu:
7852  case Intrinsic::arm_neon_vqrshiftns:
7853  case Intrinsic::arm_neon_vqrshiftnu:
7854  case Intrinsic::arm_neon_vqrshiftnsu: {
7855    EVT VT = N->getOperand(1).getValueType();
7856    int64_t Cnt;
7857    unsigned VShiftOpc = 0;
7858
7859    switch (IntNo) {
7860    case Intrinsic::arm_neon_vshifts:
7861    case Intrinsic::arm_neon_vshiftu:
7862      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7863        VShiftOpc = ARMISD::VSHL;
7864        break;
7865      }
7866      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7867        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7868                     ARMISD::VSHRs : ARMISD::VSHRu);
7869        break;
7870      }
7871      return SDValue();
7872
7873    case Intrinsic::arm_neon_vshiftls:
7874    case Intrinsic::arm_neon_vshiftlu:
7875      if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7876        break;
7877      llvm_unreachable("invalid shift count for vshll intrinsic");
7878
7879    case Intrinsic::arm_neon_vrshifts:
7880    case Intrinsic::arm_neon_vrshiftu:
7881      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7882        break;
7883      return SDValue();
7884
7885    case Intrinsic::arm_neon_vqshifts:
7886    case Intrinsic::arm_neon_vqshiftu:
7887      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7888        break;
7889      return SDValue();
7890
7891    case Intrinsic::arm_neon_vqshiftsu:
7892      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7893        break;
7894      llvm_unreachable("invalid shift count for vqshlu intrinsic");
7895
7896    case Intrinsic::arm_neon_vshiftn:
7897    case Intrinsic::arm_neon_vrshiftn:
7898    case Intrinsic::arm_neon_vqshiftns:
7899    case Intrinsic::arm_neon_vqshiftnu:
7900    case Intrinsic::arm_neon_vqshiftnsu:
7901    case Intrinsic::arm_neon_vqrshiftns:
7902    case Intrinsic::arm_neon_vqrshiftnu:
7903    case Intrinsic::arm_neon_vqrshiftnsu:
7904      // Narrowing shifts require an immediate right shift.
7905      if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7906        break;
7907      llvm_unreachable("invalid shift count for narrowing vector shift "
7908                       "intrinsic");
7909
7910    default:
7911      llvm_unreachable("unhandled vector shift");
7912    }
7913
7914    switch (IntNo) {
7915    case Intrinsic::arm_neon_vshifts:
7916    case Intrinsic::arm_neon_vshiftu:
7917      // Opcode already set above.
7918      break;
7919    case Intrinsic::arm_neon_vshiftls:
7920    case Intrinsic::arm_neon_vshiftlu:
7921      if (Cnt == VT.getVectorElementType().getSizeInBits())
7922        VShiftOpc = ARMISD::VSHLLi;
7923      else
7924        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7925                     ARMISD::VSHLLs : ARMISD::VSHLLu);
7926      break;
7927    case Intrinsic::arm_neon_vshiftn:
7928      VShiftOpc = ARMISD::VSHRN; break;
7929    case Intrinsic::arm_neon_vrshifts:
7930      VShiftOpc = ARMISD::VRSHRs; break;
7931    case Intrinsic::arm_neon_vrshiftu:
7932      VShiftOpc = ARMISD::VRSHRu; break;
7933    case Intrinsic::arm_neon_vrshiftn:
7934      VShiftOpc = ARMISD::VRSHRN; break;
7935    case Intrinsic::arm_neon_vqshifts:
7936      VShiftOpc = ARMISD::VQSHLs; break;
7937    case Intrinsic::arm_neon_vqshiftu:
7938      VShiftOpc = ARMISD::VQSHLu; break;
7939    case Intrinsic::arm_neon_vqshiftsu:
7940      VShiftOpc = ARMISD::VQSHLsu; break;
7941    case Intrinsic::arm_neon_vqshiftns:
7942      VShiftOpc = ARMISD::VQSHRNs; break;
7943    case Intrinsic::arm_neon_vqshiftnu:
7944      VShiftOpc = ARMISD::VQSHRNu; break;
7945    case Intrinsic::arm_neon_vqshiftnsu:
7946      VShiftOpc = ARMISD::VQSHRNsu; break;
7947    case Intrinsic::arm_neon_vqrshiftns:
7948      VShiftOpc = ARMISD::VQRSHRNs; break;
7949    case Intrinsic::arm_neon_vqrshiftnu:
7950      VShiftOpc = ARMISD::VQRSHRNu; break;
7951    case Intrinsic::arm_neon_vqrshiftnsu:
7952      VShiftOpc = ARMISD::VQRSHRNsu; break;
7953    }
7954
7955    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7956                       N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7957  }
7958
7959  case Intrinsic::arm_neon_vshiftins: {
7960    EVT VT = N->getOperand(1).getValueType();
7961    int64_t Cnt;
7962    unsigned VShiftOpc = 0;
7963
7964    if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7965      VShiftOpc = ARMISD::VSLI;
7966    else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7967      VShiftOpc = ARMISD::VSRI;
7968    else {
7969      llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7970    }
7971
7972    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7973                       N->getOperand(1), N->getOperand(2),
7974                       DAG.getConstant(Cnt, MVT::i32));
7975  }
7976
7977  case Intrinsic::arm_neon_vqrshifts:
7978  case Intrinsic::arm_neon_vqrshiftu:
7979    // No immediate versions of these to check for.
7980    break;
7981  }
7982
7983  return SDValue();
7984}
7985
7986/// PerformShiftCombine - Checks for immediate versions of vector shifts and
7987/// lowers them.  As with the vector shift intrinsics, this is done during DAG
7988/// combining instead of DAG legalizing because the build_vectors for 64-bit
7989/// vector element shift counts are generally not legal, and it is hard to see
7990/// their values after they get legalized to loads from a constant pool.
7991static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7992                                   const ARMSubtarget *ST) {
7993  EVT VT = N->getValueType(0);
7994  if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
7995    // Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
7996    // 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
7997    SDValue N1 = N->getOperand(1);
7998    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
7999      SDValue N0 = N->getOperand(0);
8000      if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8001          DAG.MaskedValueIsZero(N0.getOperand(0),
8002                                APInt::getHighBitsSet(32, 16)))
8003        return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8004    }
8005  }
8006
8007  // Nothing to be done for scalar shifts.
8008  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8009  if (!VT.isVector() || !TLI.isTypeLegal(VT))
8010    return SDValue();
8011
8012  assert(ST->hasNEON() && "unexpected vector shift");
8013  int64_t Cnt;
8014
8015  switch (N->getOpcode()) {
8016  default: llvm_unreachable("unexpected shift opcode");
8017
8018  case ISD::SHL:
8019    if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8020      return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8021                         DAG.getConstant(Cnt, MVT::i32));
8022    break;
8023
8024  case ISD::SRA:
8025  case ISD::SRL:
8026    if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8027      unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8028                            ARMISD::VSHRs : ARMISD::VSHRu);
8029      return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8030                         DAG.getConstant(Cnt, MVT::i32));
8031    }
8032  }
8033  return SDValue();
8034}
8035
8036/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8037/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8038static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8039                                    const ARMSubtarget *ST) {
8040  SDValue N0 = N->getOperand(0);
8041
8042  // Check for sign- and zero-extensions of vector extract operations of 8-
8043  // and 16-bit vector elements.  NEON supports these directly.  They are
8044  // handled during DAG combining because type legalization will promote them
8045  // to 32-bit types and it is messy to recognize the operations after that.
8046  if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8047    SDValue Vec = N0.getOperand(0);
8048    SDValue Lane = N0.getOperand(1);
8049    EVT VT = N->getValueType(0);
8050    EVT EltVT = N0.getValueType();
8051    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8052
8053    if (VT == MVT::i32 &&
8054        (EltVT == MVT::i8 || EltVT == MVT::i16) &&
8055        TLI.isTypeLegal(Vec.getValueType()) &&
8056        isa<ConstantSDNode>(Lane)) {
8057
8058      unsigned Opc = 0;
8059      switch (N->getOpcode()) {
8060      default: llvm_unreachable("unexpected opcode");
8061      case ISD::SIGN_EXTEND:
8062        Opc = ARMISD::VGETLANEs;
8063        break;
8064      case ISD::ZERO_EXTEND:
8065      case ISD::ANY_EXTEND:
8066        Opc = ARMISD::VGETLANEu;
8067        break;
8068      }
8069      return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8070    }
8071  }
8072
8073  return SDValue();
8074}
8075
8076/// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8077/// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8078static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8079                                       const ARMSubtarget *ST) {
8080  // If the target supports NEON, try to use vmax/vmin instructions for f32
8081  // selects like "x < y ? x : y".  Unless the NoNaNsFPMath option is set,
8082  // be careful about NaNs:  NEON's vmax/vmin return NaN if either operand is
8083  // a NaN; only do the transformation when it matches that behavior.
8084
8085  // For now only do this when using NEON for FP operations; if using VFP, it
8086  // is not obvious that the benefit outweighs the cost of switching to the
8087  // NEON pipeline.
8088  if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8089      N->getValueType(0) != MVT::f32)
8090    return SDValue();
8091
8092  SDValue CondLHS = N->getOperand(0);
8093  SDValue CondRHS = N->getOperand(1);
8094  SDValue LHS = N->getOperand(2);
8095  SDValue RHS = N->getOperand(3);
8096  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8097
8098  unsigned Opcode = 0;
8099  bool IsReversed;
8100  if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8101    IsReversed = false; // x CC y ? x : y
8102  } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8103    IsReversed = true ; // x CC y ? y : x
8104  } else {
8105    return SDValue();
8106  }
8107
8108  bool IsUnordered;
8109  switch (CC) {
8110  default: break;
8111  case ISD::SETOLT:
8112  case ISD::SETOLE:
8113  case ISD::SETLT:
8114  case ISD::SETLE:
8115  case ISD::SETULT:
8116  case ISD::SETULE:
8117    // If LHS is NaN, an ordered comparison will be false and the result will
8118    // be the RHS, but vmin(NaN, RHS) = NaN.  Avoid this by checking that LHS
8119    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
8120    IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8121    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8122      break;
8123    // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8124    // will return -0, so vmin can only be used for unsafe math or if one of
8125    // the operands is known to be nonzero.
8126    if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8127        !DAG.getTarget().Options.UnsafeFPMath &&
8128        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8129      break;
8130    Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8131    break;
8132
8133  case ISD::SETOGT:
8134  case ISD::SETOGE:
8135  case ISD::SETGT:
8136  case ISD::SETGE:
8137  case ISD::SETUGT:
8138  case ISD::SETUGE:
8139    // If LHS is NaN, an ordered comparison will be false and the result will
8140    // be the RHS, but vmax(NaN, RHS) = NaN.  Avoid this by checking that LHS
8141    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
8142    IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8143    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8144      break;
8145    // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8146    // will return +0, so vmax can only be used for unsafe math or if one of
8147    // the operands is known to be nonzero.
8148    if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8149        !DAG.getTarget().Options.UnsafeFPMath &&
8150        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8151      break;
8152    Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8153    break;
8154  }
8155
8156  if (!Opcode)
8157    return SDValue();
8158  return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8159}
8160
8161/// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8162SDValue
8163ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8164  SDValue Cmp = N->getOperand(4);
8165  if (Cmp.getOpcode() != ARMISD::CMPZ)
8166    // Only looking at EQ and NE cases.
8167    return SDValue();
8168
8169  EVT VT = N->getValueType(0);
8170  DebugLoc dl = N->getDebugLoc();
8171  SDValue LHS = Cmp.getOperand(0);
8172  SDValue RHS = Cmp.getOperand(1);
8173  SDValue FalseVal = N->getOperand(0);
8174  SDValue TrueVal = N->getOperand(1);
8175  SDValue ARMcc = N->getOperand(2);
8176  ARMCC::CondCodes CC =
8177    (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8178
8179  // Simplify
8180  //   mov     r1, r0
8181  //   cmp     r1, x
8182  //   mov     r0, y
8183  //   moveq   r0, x
8184  // to
8185  //   cmp     r0, x
8186  //   movne   r0, y
8187  //
8188  //   mov     r1, r0
8189  //   cmp     r1, x
8190  //   mov     r0, x
8191  //   movne   r0, y
8192  // to
8193  //   cmp     r0, x
8194  //   movne   r0, y
8195  /// FIXME: Turn this into a target neutral optimization?
8196  SDValue Res;
8197  if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8198    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8199                      N->getOperand(3), Cmp);
8200  } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8201    SDValue ARMcc;
8202    SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8203    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8204                      N->getOperand(3), NewCmp);
8205  }
8206
8207  if (Res.getNode()) {
8208    APInt KnownZero, KnownOne;
8209    APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8210    DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8211    // Capture demanded bits information that would be otherwise lost.
8212    if (KnownZero == 0xfffffffe)
8213      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8214                        DAG.getValueType(MVT::i1));
8215    else if (KnownZero == 0xffffff00)
8216      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8217                        DAG.getValueType(MVT::i8));
8218    else if (KnownZero == 0xffff0000)
8219      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8220                        DAG.getValueType(MVT::i16));
8221  }
8222
8223  return Res;
8224}
8225
8226SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8227                                             DAGCombinerInfo &DCI) const {
8228  switch (N->getOpcode()) {
8229  default: break;
8230  case ISD::ADD:        return PerformADDCombine(N, DCI, Subtarget);
8231  case ISD::SUB:        return PerformSUBCombine(N, DCI);
8232  case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
8233  case ISD::OR:         return PerformORCombine(N, DCI, Subtarget);
8234  case ISD::XOR:        return PerformXORCombine(N, DCI, Subtarget);
8235  case ISD::AND:        return PerformANDCombine(N, DCI, Subtarget);
8236  case ARMISD::BFI:     return PerformBFICombine(N, DCI);
8237  case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8238  case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8239  case ISD::STORE:      return PerformSTORECombine(N, DCI);
8240  case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8241  case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8242  case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8243  case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8244  case ISD::FP_TO_SINT:
8245  case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8246  case ISD::FDIV:       return PerformVDIVCombine(N, DCI, Subtarget);
8247  case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8248  case ISD::SHL:
8249  case ISD::SRA:
8250  case ISD::SRL:        return PerformShiftCombine(N, DCI.DAG, Subtarget);
8251  case ISD::SIGN_EXTEND:
8252  case ISD::ZERO_EXTEND:
8253  case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8254  case ISD::SELECT_CC:  return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8255  case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8256  case ARMISD::VLD2DUP:
8257  case ARMISD::VLD3DUP:
8258  case ARMISD::VLD4DUP:
8259    return CombineBaseUpdate(N, DCI);
8260  case ISD::INTRINSIC_VOID:
8261  case ISD::INTRINSIC_W_CHAIN:
8262    switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8263    case Intrinsic::arm_neon_vld1:
8264    case Intrinsic::arm_neon_vld2:
8265    case Intrinsic::arm_neon_vld3:
8266    case Intrinsic::arm_neon_vld4:
8267    case Intrinsic::arm_neon_vld2lane:
8268    case Intrinsic::arm_neon_vld3lane:
8269    case Intrinsic::arm_neon_vld4lane:
8270    case Intrinsic::arm_neon_vst1:
8271    case Intrinsic::arm_neon_vst2:
8272    case Intrinsic::arm_neon_vst3:
8273    case Intrinsic::arm_neon_vst4:
8274    case Intrinsic::arm_neon_vst2lane:
8275    case Intrinsic::arm_neon_vst3lane:
8276    case Intrinsic::arm_neon_vst4lane:
8277      return CombineBaseUpdate(N, DCI);
8278    default: break;
8279    }
8280    break;
8281  }
8282  return SDValue();
8283}
8284
8285bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8286                                                          EVT VT) const {
8287  return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8288}
8289
8290bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8291  if (!Subtarget->allowsUnalignedMem())
8292    return false;
8293
8294  switch (VT.getSimpleVT().SimpleTy) {
8295  default:
8296    return false;
8297  case MVT::i8:
8298  case MVT::i16:
8299  case MVT::i32:
8300    return true;
8301  // FIXME: VLD1 etc with standard alignment is legal.
8302  }
8303}
8304
8305static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8306                       unsigned AlignCheck) {
8307  return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8308          (DstAlign == 0 || DstAlign % AlignCheck == 0));
8309}
8310
8311EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8312                                           unsigned DstAlign, unsigned SrcAlign,
8313                                           bool IsZeroVal,
8314                                           bool MemcpyStrSrc,
8315                                           MachineFunction &MF) const {
8316  const Function *F = MF.getFunction();
8317
8318  // See if we can use NEON instructions for this...
8319  if (IsZeroVal &&
8320      !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8321      Subtarget->hasNEON()) {
8322    if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8323      return MVT::v4i32;
8324    } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8325      return MVT::v2i32;
8326    }
8327  }
8328
8329  // Lowering to i32/i16 if the size permits.
8330  if (Size >= 4) {
8331    return MVT::i32;
8332  } else if (Size >= 2) {
8333    return MVT::i16;
8334  }
8335
8336  // Let the target-independent logic figure it out.
8337  return MVT::Other;
8338}
8339
8340static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8341  if (V < 0)
8342    return false;
8343
8344  unsigned Scale = 1;
8345  switch (VT.getSimpleVT().SimpleTy) {
8346  default: return false;
8347  case MVT::i1:
8348  case MVT::i8:
8349    // Scale == 1;
8350    break;
8351  case MVT::i16:
8352    // Scale == 2;
8353    Scale = 2;
8354    break;
8355  case MVT::i32:
8356    // Scale == 4;
8357    Scale = 4;
8358    break;
8359  }
8360
8361  if ((V & (Scale - 1)) != 0)
8362    return false;
8363  V /= Scale;
8364  return V == (V & ((1LL << 5) - 1));
8365}
8366
8367static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8368                                      const ARMSubtarget *Subtarget) {
8369  bool isNeg = false;
8370  if (V < 0) {
8371    isNeg = true;
8372    V = - V;
8373  }
8374
8375  switch (VT.getSimpleVT().SimpleTy) {
8376  default: return false;
8377  case MVT::i1:
8378  case MVT::i8:
8379  case MVT::i16:
8380  case MVT::i32:
8381    // + imm12 or - imm8
8382    if (isNeg)
8383      return V == (V & ((1LL << 8) - 1));
8384    return V == (V & ((1LL << 12) - 1));
8385  case MVT::f32:
8386  case MVT::f64:
8387    // Same as ARM mode. FIXME: NEON?
8388    if (!Subtarget->hasVFP2())
8389      return false;
8390    if ((V & 3) != 0)
8391      return false;
8392    V >>= 2;
8393    return V == (V & ((1LL << 8) - 1));
8394  }
8395}
8396
8397/// isLegalAddressImmediate - Return true if the integer value can be used
8398/// as the offset of the target addressing mode for load / store of the
8399/// given type.
8400static bool isLegalAddressImmediate(int64_t V, EVT VT,
8401                                    const ARMSubtarget *Subtarget) {
8402  if (V == 0)
8403    return true;
8404
8405  if (!VT.isSimple())
8406    return false;
8407
8408  if (Subtarget->isThumb1Only())
8409    return isLegalT1AddressImmediate(V, VT);
8410  else if (Subtarget->isThumb2())
8411    return isLegalT2AddressImmediate(V, VT, Subtarget);
8412
8413  // ARM mode.
8414  if (V < 0)
8415    V = - V;
8416  switch (VT.getSimpleVT().SimpleTy) {
8417  default: return false;
8418  case MVT::i1:
8419  case MVT::i8:
8420  case MVT::i32:
8421    // +- imm12
8422    return V == (V & ((1LL << 12) - 1));
8423  case MVT::i16:
8424    // +- imm8
8425    return V == (V & ((1LL << 8) - 1));
8426  case MVT::f32:
8427  case MVT::f64:
8428    if (!Subtarget->hasVFP2()) // FIXME: NEON?
8429      return false;
8430    if ((V & 3) != 0)
8431      return false;
8432    V >>= 2;
8433    return V == (V & ((1LL << 8) - 1));
8434  }
8435}
8436
8437bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8438                                                      EVT VT) const {
8439  int Scale = AM.Scale;
8440  if (Scale < 0)
8441    return false;
8442
8443  switch (VT.getSimpleVT().SimpleTy) {
8444  default: return false;
8445  case MVT::i1:
8446  case MVT::i8:
8447  case MVT::i16:
8448  case MVT::i32:
8449    if (Scale == 1)
8450      return true;
8451    // r + r << imm
8452    Scale = Scale & ~1;
8453    return Scale == 2 || Scale == 4 || Scale == 8;
8454  case MVT::i64:
8455    // r + r
8456    if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8457      return true;
8458    return false;
8459  case MVT::isVoid:
8460    // Note, we allow "void" uses (basically, uses that aren't loads or
8461    // stores), because arm allows folding a scale into many arithmetic
8462    // operations.  This should be made more precise and revisited later.
8463
8464    // Allow r << imm, but the imm has to be a multiple of two.
8465    if (Scale & 1) return false;
8466    return isPowerOf2_32(Scale);
8467  }
8468}
8469
8470/// isLegalAddressingMode - Return true if the addressing mode represented
8471/// by AM is legal for this target, for a load/store of the specified type.
8472bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8473                                              Type *Ty) const {
8474  EVT VT = getValueType(Ty, true);
8475  if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8476    return false;
8477
8478  // Can never fold addr of global into load/store.
8479  if (AM.BaseGV)
8480    return false;
8481
8482  switch (AM.Scale) {
8483  case 0:  // no scale reg, must be "r+i" or "r", or "i".
8484    break;
8485  case 1:
8486    if (Subtarget->isThumb1Only())
8487      return false;
8488    // FALL THROUGH.
8489  default:
8490    // ARM doesn't support any R+R*scale+imm addr modes.
8491    if (AM.BaseOffs)
8492      return false;
8493
8494    if (!VT.isSimple())
8495      return false;
8496
8497    if (Subtarget->isThumb2())
8498      return isLegalT2ScaledAddressingMode(AM, VT);
8499
8500    int Scale = AM.Scale;
8501    switch (VT.getSimpleVT().SimpleTy) {
8502    default: return false;
8503    case MVT::i1:
8504    case MVT::i8:
8505    case MVT::i32:
8506      if (Scale < 0) Scale = -Scale;
8507      if (Scale == 1)
8508        return true;
8509      // r + r << imm
8510      return isPowerOf2_32(Scale & ~1);
8511    case MVT::i16:
8512    case MVT::i64:
8513      // r + r
8514      if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8515        return true;
8516      return false;
8517
8518    case MVT::isVoid:
8519      // Note, we allow "void" uses (basically, uses that aren't loads or
8520      // stores), because arm allows folding a scale into many arithmetic
8521      // operations.  This should be made more precise and revisited later.
8522
8523      // Allow r << imm, but the imm has to be a multiple of two.
8524      if (Scale & 1) return false;
8525      return isPowerOf2_32(Scale);
8526    }
8527  }
8528  return true;
8529}
8530
8531/// isLegalICmpImmediate - Return true if the specified immediate is legal
8532/// icmp immediate, that is the target has icmp instructions which can compare
8533/// a register against the immediate without having to materialize the
8534/// immediate into a register.
8535bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8536  if (!Subtarget->isThumb())
8537    return ARM_AM::getSOImmVal(Imm) != -1;
8538  if (Subtarget->isThumb2())
8539    return ARM_AM::getT2SOImmVal(Imm) != -1;
8540  return Imm >= 0 && Imm <= 255;
8541}
8542
8543/// isLegalAddImmediate - Return true if the specified immediate is legal
8544/// add immediate, that is the target has add instructions which can add
8545/// a register with the immediate without having to materialize the
8546/// immediate into a register.
8547bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8548  return ARM_AM::getSOImmVal(Imm) != -1;
8549}
8550
8551static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8552                                      bool isSEXTLoad, SDValue &Base,
8553                                      SDValue &Offset, bool &isInc,
8554                                      SelectionDAG &DAG) {
8555  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8556    return false;
8557
8558  if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8559    // AddressingMode 3
8560    Base = Ptr->getOperand(0);
8561    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8562      int RHSC = (int)RHS->getZExtValue();
8563      if (RHSC < 0 && RHSC > -256) {
8564        assert(Ptr->getOpcode() == ISD::ADD);
8565        isInc = false;
8566        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8567        return true;
8568      }
8569    }
8570    isInc = (Ptr->getOpcode() == ISD::ADD);
8571    Offset = Ptr->getOperand(1);
8572    return true;
8573  } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8574    // AddressingMode 2
8575    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8576      int RHSC = (int)RHS->getZExtValue();
8577      if (RHSC < 0 && RHSC > -0x1000) {
8578        assert(Ptr->getOpcode() == ISD::ADD);
8579        isInc = false;
8580        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8581        Base = Ptr->getOperand(0);
8582        return true;
8583      }
8584    }
8585
8586    if (Ptr->getOpcode() == ISD::ADD) {
8587      isInc = true;
8588      ARM_AM::ShiftOpc ShOpcVal=
8589        ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8590      if (ShOpcVal != ARM_AM::no_shift) {
8591        Base = Ptr->getOperand(1);
8592        Offset = Ptr->getOperand(0);
8593      } else {
8594        Base = Ptr->getOperand(0);
8595        Offset = Ptr->getOperand(1);
8596      }
8597      return true;
8598    }
8599
8600    isInc = (Ptr->getOpcode() == ISD::ADD);
8601    Base = Ptr->getOperand(0);
8602    Offset = Ptr->getOperand(1);
8603    return true;
8604  }
8605
8606  // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8607  return false;
8608}
8609
8610static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8611                                     bool isSEXTLoad, SDValue &Base,
8612                                     SDValue &Offset, bool &isInc,
8613                                     SelectionDAG &DAG) {
8614  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8615    return false;
8616
8617  Base = Ptr->getOperand(0);
8618  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8619    int RHSC = (int)RHS->getZExtValue();
8620    if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8621      assert(Ptr->getOpcode() == ISD::ADD);
8622      isInc = false;
8623      Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8624      return true;
8625    } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8626      isInc = Ptr->getOpcode() == ISD::ADD;
8627      Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8628      return true;
8629    }
8630  }
8631
8632  return false;
8633}
8634
8635/// getPreIndexedAddressParts - returns true by value, base pointer and
8636/// offset pointer and addressing mode by reference if the node's address
8637/// can be legally represented as pre-indexed load / store address.
8638bool
8639ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8640                                             SDValue &Offset,
8641                                             ISD::MemIndexedMode &AM,
8642                                             SelectionDAG &DAG) const {
8643  if (Subtarget->isThumb1Only())
8644    return false;
8645
8646  EVT VT;
8647  SDValue Ptr;
8648  bool isSEXTLoad = false;
8649  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8650    Ptr = LD->getBasePtr();
8651    VT  = LD->getMemoryVT();
8652    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8653  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8654    Ptr = ST->getBasePtr();
8655    VT  = ST->getMemoryVT();
8656  } else
8657    return false;
8658
8659  bool isInc;
8660  bool isLegal = false;
8661  if (Subtarget->isThumb2())
8662    isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8663                                       Offset, isInc, DAG);
8664  else
8665    isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8666                                        Offset, isInc, DAG);
8667  if (!isLegal)
8668    return false;
8669
8670  AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8671  return true;
8672}
8673
8674/// getPostIndexedAddressParts - returns true by value, base pointer and
8675/// offset pointer and addressing mode by reference if this node can be
8676/// combined with a load / store to form a post-indexed load / store.
8677bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8678                                                   SDValue &Base,
8679                                                   SDValue &Offset,
8680                                                   ISD::MemIndexedMode &AM,
8681                                                   SelectionDAG &DAG) const {
8682  if (Subtarget->isThumb1Only())
8683    return false;
8684
8685  EVT VT;
8686  SDValue Ptr;
8687  bool isSEXTLoad = false;
8688  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8689    VT  = LD->getMemoryVT();
8690    Ptr = LD->getBasePtr();
8691    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8692  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8693    VT  = ST->getMemoryVT();
8694    Ptr = ST->getBasePtr();
8695  } else
8696    return false;
8697
8698  bool isInc;
8699  bool isLegal = false;
8700  if (Subtarget->isThumb2())
8701    isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8702                                       isInc, DAG);
8703  else
8704    isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8705                                        isInc, DAG);
8706  if (!isLegal)
8707    return false;
8708
8709  if (Ptr != Base) {
8710    // Swap base ptr and offset to catch more post-index load / store when
8711    // it's legal. In Thumb2 mode, offset must be an immediate.
8712    if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8713        !Subtarget->isThumb2())
8714      std::swap(Base, Offset);
8715
8716    // Post-indexed load / store update the base pointer.
8717    if (Ptr != Base)
8718      return false;
8719  }
8720
8721  AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8722  return true;
8723}
8724
8725void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8726                                                       const APInt &Mask,
8727                                                       APInt &KnownZero,
8728                                                       APInt &KnownOne,
8729                                                       const SelectionDAG &DAG,
8730                                                       unsigned Depth) const {
8731  KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8732  switch (Op.getOpcode()) {
8733  default: break;
8734  case ARMISD::CMOV: {
8735    // Bits are known zero/one if known on the LHS and RHS.
8736    DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8737    if (KnownZero == 0 && KnownOne == 0) return;
8738
8739    APInt KnownZeroRHS, KnownOneRHS;
8740    DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8741                          KnownZeroRHS, KnownOneRHS, Depth+1);
8742    KnownZero &= KnownZeroRHS;
8743    KnownOne  &= KnownOneRHS;
8744    return;
8745  }
8746  }
8747}
8748
8749//===----------------------------------------------------------------------===//
8750//                           ARM Inline Assembly Support
8751//===----------------------------------------------------------------------===//
8752
8753bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8754  // Looking for "rev" which is V6+.
8755  if (!Subtarget->hasV6Ops())
8756    return false;
8757
8758  InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8759  std::string AsmStr = IA->getAsmString();
8760  SmallVector<StringRef, 4> AsmPieces;
8761  SplitString(AsmStr, AsmPieces, ";\n");
8762
8763  switch (AsmPieces.size()) {
8764  default: return false;
8765  case 1:
8766    AsmStr = AsmPieces[0];
8767    AsmPieces.clear();
8768    SplitString(AsmStr, AsmPieces, " \t,");
8769
8770    // rev $0, $1
8771    if (AsmPieces.size() == 3 &&
8772        AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8773        IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8774      IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8775      if (Ty && Ty->getBitWidth() == 32)
8776        return IntrinsicLowering::LowerToByteSwap(CI);
8777    }
8778    break;
8779  }
8780
8781  return false;
8782}
8783
8784/// getConstraintType - Given a constraint letter, return the type of
8785/// constraint it is for this target.
8786ARMTargetLowering::ConstraintType
8787ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8788  if (Constraint.size() == 1) {
8789    switch (Constraint[0]) {
8790    default:  break;
8791    case 'l': return C_RegisterClass;
8792    case 'w': return C_RegisterClass;
8793    case 'h': return C_RegisterClass;
8794    case 'x': return C_RegisterClass;
8795    case 't': return C_RegisterClass;
8796    case 'j': return C_Other; // Constant for movw.
8797      // An address with a single base register. Due to the way we
8798      // currently handle addresses it is the same as an 'r' memory constraint.
8799    case 'Q': return C_Memory;
8800    }
8801  } else if (Constraint.size() == 2) {
8802    switch (Constraint[0]) {
8803    default: break;
8804    // All 'U+' constraints are addresses.
8805    case 'U': return C_Memory;
8806    }
8807  }
8808  return TargetLowering::getConstraintType(Constraint);
8809}
8810
8811/// Examine constraint type and operand type and determine a weight value.
8812/// This object must already have been set up with the operand type
8813/// and the current alternative constraint selected.
8814TargetLowering::ConstraintWeight
8815ARMTargetLowering::getSingleConstraintMatchWeight(
8816    AsmOperandInfo &info, const char *constraint) const {
8817  ConstraintWeight weight = CW_Invalid;
8818  Value *CallOperandVal = info.CallOperandVal;
8819    // If we don't have a value, we can't do a match,
8820    // but allow it at the lowest weight.
8821  if (CallOperandVal == NULL)
8822    return CW_Default;
8823  Type *type = CallOperandVal->getType();
8824  // Look at the constraint type.
8825  switch (*constraint) {
8826  default:
8827    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8828    break;
8829  case 'l':
8830    if (type->isIntegerTy()) {
8831      if (Subtarget->isThumb())
8832        weight = CW_SpecificReg;
8833      else
8834        weight = CW_Register;
8835    }
8836    break;
8837  case 'w':
8838    if (type->isFloatingPointTy())
8839      weight = CW_Register;
8840    break;
8841  }
8842  return weight;
8843}
8844
8845typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8846RCPair
8847ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8848                                                EVT VT) const {
8849  if (Constraint.size() == 1) {
8850    // GCC ARM Constraint Letters
8851    switch (Constraint[0]) {
8852    case 'l': // Low regs or general regs.
8853      if (Subtarget->isThumb())
8854        return RCPair(0U, ARM::tGPRRegisterClass);
8855      else
8856        return RCPair(0U, ARM::GPRRegisterClass);
8857    case 'h': // High regs or no regs.
8858      if (Subtarget->isThumb())
8859        return RCPair(0U, ARM::hGPRRegisterClass);
8860      break;
8861    case 'r':
8862      return RCPair(0U, ARM::GPRRegisterClass);
8863    case 'w':
8864      if (VT == MVT::f32)
8865        return RCPair(0U, ARM::SPRRegisterClass);
8866      if (VT.getSizeInBits() == 64)
8867        return RCPair(0U, ARM::DPRRegisterClass);
8868      if (VT.getSizeInBits() == 128)
8869        return RCPair(0U, ARM::QPRRegisterClass);
8870      break;
8871    case 'x':
8872      if (VT == MVT::f32)
8873        return RCPair(0U, ARM::SPR_8RegisterClass);
8874      if (VT.getSizeInBits() == 64)
8875        return RCPair(0U, ARM::DPR_8RegisterClass);
8876      if (VT.getSizeInBits() == 128)
8877        return RCPair(0U, ARM::QPR_8RegisterClass);
8878      break;
8879    case 't':
8880      if (VT == MVT::f32)
8881        return RCPair(0U, ARM::SPRRegisterClass);
8882      break;
8883    }
8884  }
8885  if (StringRef("{cc}").equals_lower(Constraint))
8886    return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8887
8888  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8889}
8890
8891/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8892/// vector.  If it is invalid, don't add anything to Ops.
8893void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8894                                                     std::string &Constraint,
8895                                                     std::vector<SDValue>&Ops,
8896                                                     SelectionDAG &DAG) const {
8897  SDValue Result(0, 0);
8898
8899  // Currently only support length 1 constraints.
8900  if (Constraint.length() != 1) return;
8901
8902  char ConstraintLetter = Constraint[0];
8903  switch (ConstraintLetter) {
8904  default: break;
8905  case 'j':
8906  case 'I': case 'J': case 'K': case 'L':
8907  case 'M': case 'N': case 'O':
8908    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8909    if (!C)
8910      return;
8911
8912    int64_t CVal64 = C->getSExtValue();
8913    int CVal = (int) CVal64;
8914    // None of these constraints allow values larger than 32 bits.  Check
8915    // that the value fits in an int.
8916    if (CVal != CVal64)
8917      return;
8918
8919    switch (ConstraintLetter) {
8920      case 'j':
8921        // Constant suitable for movw, must be between 0 and
8922        // 65535.
8923        if (Subtarget->hasV6T2Ops())
8924          if (CVal >= 0 && CVal <= 65535)
8925            break;
8926        return;
8927      case 'I':
8928        if (Subtarget->isThumb1Only()) {
8929          // This must be a constant between 0 and 255, for ADD
8930          // immediates.
8931          if (CVal >= 0 && CVal <= 255)
8932            break;
8933        } else if (Subtarget->isThumb2()) {
8934          // A constant that can be used as an immediate value in a
8935          // data-processing instruction.
8936          if (ARM_AM::getT2SOImmVal(CVal) != -1)
8937            break;
8938        } else {
8939          // A constant that can be used as an immediate value in a
8940          // data-processing instruction.
8941          if (ARM_AM::getSOImmVal(CVal) != -1)
8942            break;
8943        }
8944        return;
8945
8946      case 'J':
8947        if (Subtarget->isThumb()) {  // FIXME thumb2
8948          // This must be a constant between -255 and -1, for negated ADD
8949          // immediates. This can be used in GCC with an "n" modifier that
8950          // prints the negated value, for use with SUB instructions. It is
8951          // not useful otherwise but is implemented for compatibility.
8952          if (CVal >= -255 && CVal <= -1)
8953            break;
8954        } else {
8955          // This must be a constant between -4095 and 4095. It is not clear
8956          // what this constraint is intended for. Implemented for
8957          // compatibility with GCC.
8958          if (CVal >= -4095 && CVal <= 4095)
8959            break;
8960        }
8961        return;
8962
8963      case 'K':
8964        if (Subtarget->isThumb1Only()) {
8965          // A 32-bit value where only one byte has a nonzero value. Exclude
8966          // zero to match GCC. This constraint is used by GCC internally for
8967          // constants that can be loaded with a move/shift combination.
8968          // It is not useful otherwise but is implemented for compatibility.
8969          if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8970            break;
8971        } else if (Subtarget->isThumb2()) {
8972          // A constant whose bitwise inverse can be used as an immediate
8973          // value in a data-processing instruction. This can be used in GCC
8974          // with a "B" modifier that prints the inverted value, for use with
8975          // BIC and MVN instructions. It is not useful otherwise but is
8976          // implemented for compatibility.
8977          if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8978            break;
8979        } else {
8980          // A constant whose bitwise inverse can be used as an immediate
8981          // value in a data-processing instruction. This can be used in GCC
8982          // with a "B" modifier that prints the inverted value, for use with
8983          // BIC and MVN instructions. It is not useful otherwise but is
8984          // implemented for compatibility.
8985          if (ARM_AM::getSOImmVal(~CVal) != -1)
8986            break;
8987        }
8988        return;
8989
8990      case 'L':
8991        if (Subtarget->isThumb1Only()) {
8992          // This must be a constant between -7 and 7,
8993          // for 3-operand ADD/SUB immediate instructions.
8994          if (CVal >= -7 && CVal < 7)
8995            break;
8996        } else if (Subtarget->isThumb2()) {
8997          // A constant whose negation can be used as an immediate value in a
8998          // data-processing instruction. This can be used in GCC with an "n"
8999          // modifier that prints the negated value, for use with SUB
9000          // instructions. It is not useful otherwise but is implemented for
9001          // compatibility.
9002          if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9003            break;
9004        } else {
9005          // A constant whose negation can be used as an immediate value in a
9006          // data-processing instruction. This can be used in GCC with an "n"
9007          // modifier that prints the negated value, for use with SUB
9008          // instructions. It is not useful otherwise but is implemented for
9009          // compatibility.
9010          if (ARM_AM::getSOImmVal(-CVal) != -1)
9011            break;
9012        }
9013        return;
9014
9015      case 'M':
9016        if (Subtarget->isThumb()) { // FIXME thumb2
9017          // This must be a multiple of 4 between 0 and 1020, for
9018          // ADD sp + immediate.
9019          if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9020            break;
9021        } else {
9022          // A power of two or a constant between 0 and 32.  This is used in
9023          // GCC for the shift amount on shifted register operands, but it is
9024          // useful in general for any shift amounts.
9025          if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9026            break;
9027        }
9028        return;
9029
9030      case 'N':
9031        if (Subtarget->isThumb()) {  // FIXME thumb2
9032          // This must be a constant between 0 and 31, for shift amounts.
9033          if (CVal >= 0 && CVal <= 31)
9034            break;
9035        }
9036        return;
9037
9038      case 'O':
9039        if (Subtarget->isThumb()) {  // FIXME thumb2
9040          // This must be a multiple of 4 between -508 and 508, for
9041          // ADD/SUB sp = sp + immediate.
9042          if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9043            break;
9044        }
9045        return;
9046    }
9047    Result = DAG.getTargetConstant(CVal, Op.getValueType());
9048    break;
9049  }
9050
9051  if (Result.getNode()) {
9052    Ops.push_back(Result);
9053    return;
9054  }
9055  return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9056}
9057
9058bool
9059ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9060  // The ARM target isn't yet aware of offsets.
9061  return false;
9062}
9063
9064bool ARM::isBitFieldInvertedMask(unsigned v) {
9065  if (v == 0xffffffff)
9066    return 0;
9067  // there can be 1's on either or both "outsides", all the "inside"
9068  // bits must be 0's
9069  unsigned int lsb = 0, msb = 31;
9070  while (v & (1 << msb)) --msb;
9071  while (v & (1 << lsb)) ++lsb;
9072  for (unsigned int i = lsb; i <= msb; ++i) {
9073    if (v & (1 << i))
9074      return 0;
9075  }
9076  return 1;
9077}
9078
9079/// isFPImmLegal - Returns true if the target can instruction select the
9080/// specified FP immediate natively. If false, the legalizer will
9081/// materialize the FP immediate as a load from a constant pool.
9082bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9083  if (!Subtarget->hasVFP3())
9084    return false;
9085  if (VT == MVT::f32)
9086    return ARM_AM::getFP32Imm(Imm) != -1;
9087  if (VT == MVT::f64)
9088    return ARM_AM::getFP64Imm(Imm) != -1;
9089  return false;
9090}
9091
9092/// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9093/// MemIntrinsicNodes.  The associated MachineMemOperands record the alignment
9094/// specified in the intrinsic calls.
9095bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9096                                           const CallInst &I,
9097                                           unsigned Intrinsic) const {
9098  switch (Intrinsic) {
9099  case Intrinsic::arm_neon_vld1:
9100  case Intrinsic::arm_neon_vld2:
9101  case Intrinsic::arm_neon_vld3:
9102  case Intrinsic::arm_neon_vld4:
9103  case Intrinsic::arm_neon_vld2lane:
9104  case Intrinsic::arm_neon_vld3lane:
9105  case Intrinsic::arm_neon_vld4lane: {
9106    Info.opc = ISD::INTRINSIC_W_CHAIN;
9107    // Conservatively set memVT to the entire set of vectors loaded.
9108    uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9109    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9110    Info.ptrVal = I.getArgOperand(0);
9111    Info.offset = 0;
9112    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9113    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9114    Info.vol = false; // volatile loads with NEON intrinsics not supported
9115    Info.readMem = true;
9116    Info.writeMem = false;
9117    return true;
9118  }
9119  case Intrinsic::arm_neon_vst1:
9120  case Intrinsic::arm_neon_vst2:
9121  case Intrinsic::arm_neon_vst3:
9122  case Intrinsic::arm_neon_vst4:
9123  case Intrinsic::arm_neon_vst2lane:
9124  case Intrinsic::arm_neon_vst3lane:
9125  case Intrinsic::arm_neon_vst4lane: {
9126    Info.opc = ISD::INTRINSIC_VOID;
9127    // Conservatively set memVT to the entire set of vectors stored.
9128    unsigned NumElts = 0;
9129    for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9130      Type *ArgTy = I.getArgOperand(ArgI)->getType();
9131      if (!ArgTy->isVectorTy())
9132        break;
9133      NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9134    }
9135    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9136    Info.ptrVal = I.getArgOperand(0);
9137    Info.offset = 0;
9138    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9139    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9140    Info.vol = false; // volatile stores with NEON intrinsics not supported
9141    Info.readMem = false;
9142    Info.writeMem = true;
9143    return true;
9144  }
9145  case Intrinsic::arm_strexd: {
9146    Info.opc = ISD::INTRINSIC_W_CHAIN;
9147    Info.memVT = MVT::i64;
9148    Info.ptrVal = I.getArgOperand(2);
9149    Info.offset = 0;
9150    Info.align = 8;
9151    Info.vol = true;
9152    Info.readMem = false;
9153    Info.writeMem = true;
9154    return true;
9155  }
9156  case Intrinsic::arm_ldrexd: {
9157    Info.opc = ISD::INTRINSIC_W_CHAIN;
9158    Info.memVT = MVT::i64;
9159    Info.ptrVal = I.getArgOperand(0);
9160    Info.offset = 0;
9161    Info.align = 8;
9162    Info.vol = true;
9163    Info.readMem = true;
9164    Info.writeMem = false;
9165    return true;
9166  }
9167  default:
9168    break;
9169  }
9170
9171  return false;
9172}
9173