ARMISelLowering.cpp revision 44d23825d61d530b8d562329ec8fc2d4f843bb8d
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::OR);
799  if (Subtarget->hasNEON())
800    setTargetDAGCombine(ISD::AND);
801
802  setStackPointerRegisterToSaveRestore(ARM::SP);
803
804  if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
805      !Subtarget->hasVFP2())
806    setSchedulingPreference(Sched::RegPressure);
807  else
808    setSchedulingPreference(Sched::Hybrid);
809
810  //// temporary - rewrite interface to use type
811  maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
812  maxStoresPerMemset = 16;
813  maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
814
815  // On ARM arguments smaller than 4 bytes are extended, so all arguments
816  // are at least 4 bytes aligned.
817  setMinStackArgumentAlignment(4);
818
819  benefitFromCodePlacementOpt = true;
820
821  setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
822}
823
824// FIXME: It might make sense to define the representative register class as the
825// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
826// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
827// SPR's representative would be DPR_VFP2. This should work well if register
828// pressure tracking were modified such that a register use would increment the
829// pressure of the register class's representative and all of it's super
830// classes' representatives transitively. We have not implemented this because
831// of the difficulty prior to coalescing of modeling operand register classes
832// due to the common occurrence of cross class copies and subregister insertions
833// and extractions.
834std::pair<const TargetRegisterClass*, uint8_t>
835ARMTargetLowering::findRepresentativeClass(EVT VT) const{
836  const TargetRegisterClass *RRC = 0;
837  uint8_t Cost = 1;
838  switch (VT.getSimpleVT().SimpleTy) {
839  default:
840    return TargetLowering::findRepresentativeClass(VT);
841  // Use DPR as representative register class for all floating point
842  // and vector types. Since there are 32 SPR registers and 32 DPR registers so
843  // the cost is 1 for both f32 and f64.
844  case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
845  case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
846    RRC = ARM::DPRRegisterClass;
847    // When NEON is used for SP, only half of the register file is available
848    // because operations that define both SP and DP results will be constrained
849    // to the VFP2 class (D0-D15). We currently model this constraint prior to
850    // coalescing by double-counting the SP regs. See the FIXME above.
851    if (Subtarget->useNEONForSinglePrecisionFP())
852      Cost = 2;
853    break;
854  case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
855  case MVT::v4f32: case MVT::v2f64:
856    RRC = ARM::DPRRegisterClass;
857    Cost = 2;
858    break;
859  case MVT::v4i64:
860    RRC = ARM::DPRRegisterClass;
861    Cost = 4;
862    break;
863  case MVT::v8i64:
864    RRC = ARM::DPRRegisterClass;
865    Cost = 8;
866    break;
867  }
868  return std::make_pair(RRC, Cost);
869}
870
871const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
872  switch (Opcode) {
873  default: return 0;
874  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
875  case ARMISD::WrapperDYN:    return "ARMISD::WrapperDYN";
876  case ARMISD::WrapperPIC:    return "ARMISD::WrapperPIC";
877  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
878  case ARMISD::CALL:          return "ARMISD::CALL";
879  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
880  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
881  case ARMISD::tCALL:         return "ARMISD::tCALL";
882  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
883  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
884  case ARMISD::BR2_JT:        return "ARMISD::BR2_JT";
885  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
886  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
887  case ARMISD::CMP:           return "ARMISD::CMP";
888  case ARMISD::CMPZ:          return "ARMISD::CMPZ";
889  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
890  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
891  case ARMISD::BCC_i64:       return "ARMISD::BCC_i64";
892  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";
893  case ARMISD::CMOV:          return "ARMISD::CMOV";
894
895  case ARMISD::RBIT:          return "ARMISD::RBIT";
896
897  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
898  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
899  case ARMISD::SITOF:         return "ARMISD::SITOF";
900  case ARMISD::UITOF:         return "ARMISD::UITOF";
901
902  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
903  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
904  case ARMISD::RRX:           return "ARMISD::RRX";
905
906  case ARMISD::ADDC:          return "ARMISD::ADDC";
907  case ARMISD::ADDE:          return "ARMISD::ADDE";
908  case ARMISD::SUBC:          return "ARMISD::SUBC";
909  case ARMISD::SUBE:          return "ARMISD::SUBE";
910
911  case ARMISD::VMOVRRD:       return "ARMISD::VMOVRRD";
912  case ARMISD::VMOVDRR:       return "ARMISD::VMOVDRR";
913
914  case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
915  case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
916
917  case ARMISD::TC_RETURN:     return "ARMISD::TC_RETURN";
918
919  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
920
921  case ARMISD::DYN_ALLOC:     return "ARMISD::DYN_ALLOC";
922
923  case ARMISD::MEMBARRIER:    return "ARMISD::MEMBARRIER";
924  case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
925
926  case ARMISD::PRELOAD:       return "ARMISD::PRELOAD";
927
928  case ARMISD::VCEQ:          return "ARMISD::VCEQ";
929  case ARMISD::VCEQZ:         return "ARMISD::VCEQZ";
930  case ARMISD::VCGE:          return "ARMISD::VCGE";
931  case ARMISD::VCGEZ:         return "ARMISD::VCGEZ";
932  case ARMISD::VCLEZ:         return "ARMISD::VCLEZ";
933  case ARMISD::VCGEU:         return "ARMISD::VCGEU";
934  case ARMISD::VCGT:          return "ARMISD::VCGT";
935  case ARMISD::VCGTZ:         return "ARMISD::VCGTZ";
936  case ARMISD::VCLTZ:         return "ARMISD::VCLTZ";
937  case ARMISD::VCGTU:         return "ARMISD::VCGTU";
938  case ARMISD::VTST:          return "ARMISD::VTST";
939
940  case ARMISD::VSHL:          return "ARMISD::VSHL";
941  case ARMISD::VSHRs:         return "ARMISD::VSHRs";
942  case ARMISD::VSHRu:         return "ARMISD::VSHRu";
943  case ARMISD::VSHLLs:        return "ARMISD::VSHLLs";
944  case ARMISD::VSHLLu:        return "ARMISD::VSHLLu";
945  case ARMISD::VSHLLi:        return "ARMISD::VSHLLi";
946  case ARMISD::VSHRN:         return "ARMISD::VSHRN";
947  case ARMISD::VRSHRs:        return "ARMISD::VRSHRs";
948  case ARMISD::VRSHRu:        return "ARMISD::VRSHRu";
949  case ARMISD::VRSHRN:        return "ARMISD::VRSHRN";
950  case ARMISD::VQSHLs:        return "ARMISD::VQSHLs";
951  case ARMISD::VQSHLu:        return "ARMISD::VQSHLu";
952  case ARMISD::VQSHLsu:       return "ARMISD::VQSHLsu";
953  case ARMISD::VQSHRNs:       return "ARMISD::VQSHRNs";
954  case ARMISD::VQSHRNu:       return "ARMISD::VQSHRNu";
955  case ARMISD::VQSHRNsu:      return "ARMISD::VQSHRNsu";
956  case ARMISD::VQRSHRNs:      return "ARMISD::VQRSHRNs";
957  case ARMISD::VQRSHRNu:      return "ARMISD::VQRSHRNu";
958  case ARMISD::VQRSHRNsu:     return "ARMISD::VQRSHRNsu";
959  case ARMISD::VGETLANEu:     return "ARMISD::VGETLANEu";
960  case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
961  case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
962  case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
963  case ARMISD::VMOVFPIMM:     return "ARMISD::VMOVFPIMM";
964  case ARMISD::VDUP:          return "ARMISD::VDUP";
965  case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
966  case ARMISD::VEXT:          return "ARMISD::VEXT";
967  case ARMISD::VREV64:        return "ARMISD::VREV64";
968  case ARMISD::VREV32:        return "ARMISD::VREV32";
969  case ARMISD::VREV16:        return "ARMISD::VREV16";
970  case ARMISD::VZIP:          return "ARMISD::VZIP";
971  case ARMISD::VUZP:          return "ARMISD::VUZP";
972  case ARMISD::VTRN:          return "ARMISD::VTRN";
973  case ARMISD::VTBL1:         return "ARMISD::VTBL1";
974  case ARMISD::VTBL2:         return "ARMISD::VTBL2";
975  case ARMISD::VMULLs:        return "ARMISD::VMULLs";
976  case ARMISD::VMULLu:        return "ARMISD::VMULLu";
977  case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
978  case ARMISD::FMAX:          return "ARMISD::FMAX";
979  case ARMISD::FMIN:          return "ARMISD::FMIN";
980  case ARMISD::BFI:           return "ARMISD::BFI";
981  case ARMISD::VORRIMM:       return "ARMISD::VORRIMM";
982  case ARMISD::VBICIMM:       return "ARMISD::VBICIMM";
983  case ARMISD::VBSL:          return "ARMISD::VBSL";
984  case ARMISD::VLD2DUP:       return "ARMISD::VLD2DUP";
985  case ARMISD::VLD3DUP:       return "ARMISD::VLD3DUP";
986  case ARMISD::VLD4DUP:       return "ARMISD::VLD4DUP";
987  case ARMISD::VLD1_UPD:      return "ARMISD::VLD1_UPD";
988  case ARMISD::VLD2_UPD:      return "ARMISD::VLD2_UPD";
989  case ARMISD::VLD3_UPD:      return "ARMISD::VLD3_UPD";
990  case ARMISD::VLD4_UPD:      return "ARMISD::VLD4_UPD";
991  case ARMISD::VLD2LN_UPD:    return "ARMISD::VLD2LN_UPD";
992  case ARMISD::VLD3LN_UPD:    return "ARMISD::VLD3LN_UPD";
993  case ARMISD::VLD4LN_UPD:    return "ARMISD::VLD4LN_UPD";
994  case ARMISD::VLD2DUP_UPD:   return "ARMISD::VLD2DUP_UPD";
995  case ARMISD::VLD3DUP_UPD:   return "ARMISD::VLD3DUP_UPD";
996  case ARMISD::VLD4DUP_UPD:   return "ARMISD::VLD4DUP_UPD";
997  case ARMISD::VST1_UPD:      return "ARMISD::VST1_UPD";
998  case ARMISD::VST2_UPD:      return "ARMISD::VST2_UPD";
999  case ARMISD::VST3_UPD:      return "ARMISD::VST3_UPD";
1000  case ARMISD::VST4_UPD:      return "ARMISD::VST4_UPD";
1001  case ARMISD::VST2LN_UPD:    return "ARMISD::VST2LN_UPD";
1002  case ARMISD::VST3LN_UPD:    return "ARMISD::VST3LN_UPD";
1003  case ARMISD::VST4LN_UPD:    return "ARMISD::VST4LN_UPD";
1004  }
1005}
1006
1007EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1008  if (!VT.isVector()) return getPointerTy();
1009  return VT.changeVectorElementTypeToInteger();
1010}
1011
1012/// getRegClassFor - Return the register class that should be used for the
1013/// specified value type.
1014const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1015  // Map v4i64 to QQ registers but do not make the type legal. Similarly map
1016  // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1017  // load / store 4 to 8 consecutive D registers.
1018  if (Subtarget->hasNEON()) {
1019    if (VT == MVT::v4i64)
1020      return ARM::QQPRRegisterClass;
1021    else if (VT == MVT::v8i64)
1022      return ARM::QQQQPRRegisterClass;
1023  }
1024  return TargetLowering::getRegClassFor(VT);
1025}
1026
1027// Create a fast isel object.
1028FastISel *
1029ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1030  return ARM::createFastISel(funcInfo);
1031}
1032
1033/// getMaximalGlobalOffset - Returns the maximal possible offset which can
1034/// be used for loads / stores from the global.
1035unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1036  return (Subtarget->isThumb1Only() ? 127 : 4095);
1037}
1038
1039Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1040  unsigned NumVals = N->getNumValues();
1041  if (!NumVals)
1042    return Sched::RegPressure;
1043
1044  for (unsigned i = 0; i != NumVals; ++i) {
1045    EVT VT = N->getValueType(i);
1046    if (VT == MVT::Glue || VT == MVT::Other)
1047      continue;
1048    if (VT.isFloatingPoint() || VT.isVector())
1049      return Sched::ILP;
1050  }
1051
1052  if (!N->isMachineOpcode())
1053    return Sched::RegPressure;
1054
1055  // Load are scheduled for latency even if there instruction itinerary
1056  // is not available.
1057  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1058  const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1059
1060  if (MCID.getNumDefs() == 0)
1061    return Sched::RegPressure;
1062  if (!Itins->isEmpty() &&
1063      Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1064    return Sched::ILP;
1065
1066  return Sched::RegPressure;
1067}
1068
1069//===----------------------------------------------------------------------===//
1070// Lowering Code
1071//===----------------------------------------------------------------------===//
1072
1073/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1074static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1075  switch (CC) {
1076  default: llvm_unreachable("Unknown condition code!");
1077  case ISD::SETNE:  return ARMCC::NE;
1078  case ISD::SETEQ:  return ARMCC::EQ;
1079  case ISD::SETGT:  return ARMCC::GT;
1080  case ISD::SETGE:  return ARMCC::GE;
1081  case ISD::SETLT:  return ARMCC::LT;
1082  case ISD::SETLE:  return ARMCC::LE;
1083  case ISD::SETUGT: return ARMCC::HI;
1084  case ISD::SETUGE: return ARMCC::HS;
1085  case ISD::SETULT: return ARMCC::LO;
1086  case ISD::SETULE: return ARMCC::LS;
1087  }
1088}
1089
1090/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1091static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1092                        ARMCC::CondCodes &CondCode2) {
1093  CondCode2 = ARMCC::AL;
1094  switch (CC) {
1095  default: llvm_unreachable("Unknown FP condition!");
1096  case ISD::SETEQ:
1097  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1098  case ISD::SETGT:
1099  case ISD::SETOGT: CondCode = ARMCC::GT; break;
1100  case ISD::SETGE:
1101  case ISD::SETOGE: CondCode = ARMCC::GE; break;
1102  case ISD::SETOLT: CondCode = ARMCC::MI; break;
1103  case ISD::SETOLE: CondCode = ARMCC::LS; break;
1104  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1105  case ISD::SETO:   CondCode = ARMCC::VC; break;
1106  case ISD::SETUO:  CondCode = ARMCC::VS; break;
1107  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1108  case ISD::SETUGT: CondCode = ARMCC::HI; break;
1109  case ISD::SETUGE: CondCode = ARMCC::PL; break;
1110  case ISD::SETLT:
1111  case ISD::SETULT: CondCode = ARMCC::LT; break;
1112  case ISD::SETLE:
1113  case ISD::SETULE: CondCode = ARMCC::LE; break;
1114  case ISD::SETNE:
1115  case ISD::SETUNE: CondCode = ARMCC::NE; break;
1116  }
1117}
1118
1119//===----------------------------------------------------------------------===//
1120//                      Calling Convention Implementation
1121//===----------------------------------------------------------------------===//
1122
1123#include "ARMGenCallingConv.inc"
1124
1125/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1126/// given CallingConvention value.
1127CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1128                                                 bool Return,
1129                                                 bool isVarArg) const {
1130  switch (CC) {
1131  default:
1132    llvm_unreachable("Unsupported calling convention");
1133  case CallingConv::Fast:
1134    if (Subtarget->hasVFP2() && !isVarArg) {
1135      if (!Subtarget->isAAPCS_ABI())
1136        return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1137      // For AAPCS ABI targets, just use VFP variant of the calling convention.
1138      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1139    }
1140    // Fallthrough
1141  case CallingConv::C: {
1142    // Use target triple & subtarget features to do actual dispatch.
1143    if (!Subtarget->isAAPCS_ABI())
1144      return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1145    else if (Subtarget->hasVFP2() &&
1146             getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1147             !isVarArg)
1148      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1149    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1150  }
1151  case CallingConv::ARM_AAPCS_VFP:
1152    if (!isVarArg)
1153      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1154    // Fallthrough
1155  case CallingConv::ARM_AAPCS:
1156    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1157  case CallingConv::ARM_APCS:
1158    return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1159  }
1160}
1161
1162/// LowerCallResult - Lower the result values of a call into the
1163/// appropriate copies out of appropriate physical registers.
1164SDValue
1165ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1166                                   CallingConv::ID CallConv, bool isVarArg,
1167                                   const SmallVectorImpl<ISD::InputArg> &Ins,
1168                                   DebugLoc dl, SelectionDAG &DAG,
1169                                   SmallVectorImpl<SDValue> &InVals) const {
1170
1171  // Assign locations to each value returned by this call.
1172  SmallVector<CCValAssign, 16> RVLocs;
1173  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1174                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1175  CCInfo.AnalyzeCallResult(Ins,
1176                           CCAssignFnForNode(CallConv, /* Return*/ true,
1177                                             isVarArg));
1178
1179  // Copy all of the result registers out of their specified physreg.
1180  for (unsigned i = 0; i != RVLocs.size(); ++i) {
1181    CCValAssign VA = RVLocs[i];
1182
1183    SDValue Val;
1184    if (VA.needsCustom()) {
1185      // Handle f64 or half of a v2f64.
1186      SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1187                                      InFlag);
1188      Chain = Lo.getValue(1);
1189      InFlag = Lo.getValue(2);
1190      VA = RVLocs[++i]; // skip ahead to next loc
1191      SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1192                                      InFlag);
1193      Chain = Hi.getValue(1);
1194      InFlag = Hi.getValue(2);
1195      Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1196
1197      if (VA.getLocVT() == MVT::v2f64) {
1198        SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1199        Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1200                          DAG.getConstant(0, MVT::i32));
1201
1202        VA = RVLocs[++i]; // skip ahead to next loc
1203        Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1204        Chain = Lo.getValue(1);
1205        InFlag = Lo.getValue(2);
1206        VA = RVLocs[++i]; // skip ahead to next loc
1207        Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1208        Chain = Hi.getValue(1);
1209        InFlag = Hi.getValue(2);
1210        Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1211        Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1212                          DAG.getConstant(1, MVT::i32));
1213      }
1214    } else {
1215      Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1216                               InFlag);
1217      Chain = Val.getValue(1);
1218      InFlag = Val.getValue(2);
1219    }
1220
1221    switch (VA.getLocInfo()) {
1222    default: llvm_unreachable("Unknown loc info!");
1223    case CCValAssign::Full: break;
1224    case CCValAssign::BCvt:
1225      Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1226      break;
1227    }
1228
1229    InVals.push_back(Val);
1230  }
1231
1232  return Chain;
1233}
1234
1235/// LowerMemOpCallTo - Store the argument to the stack.
1236SDValue
1237ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1238                                    SDValue StackPtr, SDValue Arg,
1239                                    DebugLoc dl, SelectionDAG &DAG,
1240                                    const CCValAssign &VA,
1241                                    ISD::ArgFlagsTy Flags) const {
1242  unsigned LocMemOffset = VA.getLocMemOffset();
1243  SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1244  PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1245  return DAG.getStore(Chain, dl, Arg, PtrOff,
1246                      MachinePointerInfo::getStack(LocMemOffset),
1247                      false, false, 0);
1248}
1249
1250void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1251                                         SDValue Chain, SDValue &Arg,
1252                                         RegsToPassVector &RegsToPass,
1253                                         CCValAssign &VA, CCValAssign &NextVA,
1254                                         SDValue &StackPtr,
1255                                         SmallVector<SDValue, 8> &MemOpChains,
1256                                         ISD::ArgFlagsTy Flags) const {
1257
1258  SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1259                              DAG.getVTList(MVT::i32, MVT::i32), Arg);
1260  RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1261
1262  if (NextVA.isRegLoc())
1263    RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1264  else {
1265    assert(NextVA.isMemLoc());
1266    if (StackPtr.getNode() == 0)
1267      StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1268
1269    MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1270                                           dl, DAG, NextVA,
1271                                           Flags));
1272  }
1273}
1274
1275/// LowerCall - Lowering a call into a callseq_start <-
1276/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1277/// nodes.
1278SDValue
1279ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1280                             CallingConv::ID CallConv, bool isVarArg,
1281                             bool &isTailCall,
1282                             const SmallVectorImpl<ISD::OutputArg> &Outs,
1283                             const SmallVectorImpl<SDValue> &OutVals,
1284                             const SmallVectorImpl<ISD::InputArg> &Ins,
1285                             DebugLoc dl, SelectionDAG &DAG,
1286                             SmallVectorImpl<SDValue> &InVals) const {
1287  MachineFunction &MF = DAG.getMachineFunction();
1288  bool IsStructRet    = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1289  bool IsSibCall = false;
1290  // Disable tail calls if they're not supported.
1291  if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1292    isTailCall = false;
1293  if (isTailCall) {
1294    // Check if it's really possible to do a tail call.
1295    isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1296                    isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1297                                                   Outs, OutVals, Ins, DAG);
1298    // We don't support GuaranteedTailCallOpt for ARM, only automatically
1299    // detected sibcalls.
1300    if (isTailCall) {
1301      ++NumTailCalls;
1302      IsSibCall = true;
1303    }
1304  }
1305
1306  // Analyze operands of the call, assigning locations to each operand.
1307  SmallVector<CCValAssign, 16> ArgLocs;
1308  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1309                 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1310  CCInfo.AnalyzeCallOperands(Outs,
1311                             CCAssignFnForNode(CallConv, /* Return*/ false,
1312                                               isVarArg));
1313
1314  // Get a count of how many bytes are to be pushed on the stack.
1315  unsigned NumBytes = CCInfo.getNextStackOffset();
1316
1317  // For tail calls, memory operands are available in our caller's stack.
1318  if (IsSibCall)
1319    NumBytes = 0;
1320
1321  // Adjust the stack pointer for the new arguments...
1322  // These operations are automatically eliminated by the prolog/epilog pass
1323  if (!IsSibCall)
1324    Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1325
1326  SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1327
1328  RegsToPassVector RegsToPass;
1329  SmallVector<SDValue, 8> MemOpChains;
1330
1331  // Walk the register/memloc assignments, inserting copies/loads.  In the case
1332  // of tail call optimization, arguments are handled later.
1333  for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1334       i != e;
1335       ++i, ++realArgIdx) {
1336    CCValAssign &VA = ArgLocs[i];
1337    SDValue Arg = OutVals[realArgIdx];
1338    ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1339    bool isByVal = Flags.isByVal();
1340
1341    // Promote the value if needed.
1342    switch (VA.getLocInfo()) {
1343    default: llvm_unreachable("Unknown loc info!");
1344    case CCValAssign::Full: break;
1345    case CCValAssign::SExt:
1346      Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1347      break;
1348    case CCValAssign::ZExt:
1349      Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1350      break;
1351    case CCValAssign::AExt:
1352      Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1353      break;
1354    case CCValAssign::BCvt:
1355      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1356      break;
1357    }
1358
1359    // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1360    if (VA.needsCustom()) {
1361      if (VA.getLocVT() == MVT::v2f64) {
1362        SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1363                                  DAG.getConstant(0, MVT::i32));
1364        SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1365                                  DAG.getConstant(1, MVT::i32));
1366
1367        PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1368                         VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1369
1370        VA = ArgLocs[++i]; // skip ahead to next loc
1371        if (VA.isRegLoc()) {
1372          PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1373                           VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1374        } else {
1375          assert(VA.isMemLoc());
1376
1377          MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1378                                                 dl, DAG, VA, Flags));
1379        }
1380      } else {
1381        PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1382                         StackPtr, MemOpChains, Flags);
1383      }
1384    } else if (VA.isRegLoc()) {
1385      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1386    } else if (isByVal) {
1387      assert(VA.isMemLoc());
1388      unsigned offset = 0;
1389
1390      // True if this byval aggregate will be split between registers
1391      // and memory.
1392      if (CCInfo.isFirstByValRegValid()) {
1393        EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1394        unsigned int i, j;
1395        for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1396          SDValue Const = DAG.getConstant(4*i, MVT::i32);
1397          SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1398          SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1399                                     MachinePointerInfo(),
1400                                     false, false, false, 0);
1401          MemOpChains.push_back(Load.getValue(1));
1402          RegsToPass.push_back(std::make_pair(j, Load));
1403        }
1404        offset = ARM::R4 - CCInfo.getFirstByValReg();
1405        CCInfo.clearFirstByValReg();
1406      }
1407
1408      unsigned LocMemOffset = VA.getLocMemOffset();
1409      SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1410      SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1411                                StkPtrOff);
1412      SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1413      SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1414      SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1415                                         MVT::i32);
1416      MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1417                                          Flags.getByValAlign(),
1418                                          /*isVolatile=*/false,
1419                                          /*AlwaysInline=*/false,
1420                                          MachinePointerInfo(0),
1421                                          MachinePointerInfo(0)));
1422
1423    } else if (!IsSibCall) {
1424      assert(VA.isMemLoc());
1425
1426      MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1427                                             dl, DAG, VA, Flags));
1428    }
1429  }
1430
1431  if (!MemOpChains.empty())
1432    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1433                        &MemOpChains[0], MemOpChains.size());
1434
1435  // Build a sequence of copy-to-reg nodes chained together with token chain
1436  // and flag operands which copy the outgoing args into the appropriate regs.
1437  SDValue InFlag;
1438  // Tail call byval lowering might overwrite argument registers so in case of
1439  // tail call optimization the copies to registers are lowered later.
1440  if (!isTailCall)
1441    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1442      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1443                               RegsToPass[i].second, InFlag);
1444      InFlag = Chain.getValue(1);
1445    }
1446
1447  // For tail calls lower the arguments to the 'real' stack slot.
1448  if (isTailCall) {
1449    // Force all the incoming stack arguments to be loaded from the stack
1450    // before any new outgoing arguments are stored to the stack, because the
1451    // outgoing stack slots may alias the incoming argument stack slots, and
1452    // the alias isn't otherwise explicit. This is slightly more conservative
1453    // than necessary, because it means that each store effectively depends
1454    // on every argument instead of just those arguments it would clobber.
1455
1456    // Do not flag preceding copytoreg stuff together with the following stuff.
1457    InFlag = SDValue();
1458    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1459      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1460                               RegsToPass[i].second, InFlag);
1461      InFlag = Chain.getValue(1);
1462    }
1463    InFlag =SDValue();
1464  }
1465
1466  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1467  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1468  // node so that legalize doesn't hack it.
1469  bool isDirect = false;
1470  bool isARMFunc = false;
1471  bool isLocalARMFunc = false;
1472  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1473
1474  if (EnableARMLongCalls) {
1475    assert (getTargetMachine().getRelocationModel() == Reloc::Static
1476            && "long-calls with non-static relocation model!");
1477    // Handle a global address or an external symbol. If it's not one of
1478    // those, the target's already in a register, so we don't need to do
1479    // anything extra.
1480    if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1481      const GlobalValue *GV = G->getGlobal();
1482      // Create a constant pool entry for the callee address
1483      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1484      ARMConstantPoolValue *CPV =
1485        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1486
1487      // Get the address of the callee into a register
1488      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1489      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1490      Callee = DAG.getLoad(getPointerTy(), dl,
1491                           DAG.getEntryNode(), CPAddr,
1492                           MachinePointerInfo::getConstantPool(),
1493                           false, false, false, 0);
1494    } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1495      const char *Sym = S->getSymbol();
1496
1497      // Create a constant pool entry for the callee address
1498      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1499      ARMConstantPoolValue *CPV =
1500        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1501                                      ARMPCLabelIndex, 0);
1502      // Get the address of the callee into a register
1503      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1504      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1505      Callee = DAG.getLoad(getPointerTy(), dl,
1506                           DAG.getEntryNode(), CPAddr,
1507                           MachinePointerInfo::getConstantPool(),
1508                           false, false, false, 0);
1509    }
1510  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1511    const GlobalValue *GV = G->getGlobal();
1512    isDirect = true;
1513    bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1514    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1515                   getTargetMachine().getRelocationModel() != Reloc::Static;
1516    isARMFunc = !Subtarget->isThumb() || isStub;
1517    // ARM call to a local ARM function is predicable.
1518    isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1519    // tBX takes a register source operand.
1520    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1521      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1522      ARMConstantPoolValue *CPV =
1523        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1524      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1525      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1526      Callee = DAG.getLoad(getPointerTy(), dl,
1527                           DAG.getEntryNode(), CPAddr,
1528                           MachinePointerInfo::getConstantPool(),
1529                           false, false, false, 0);
1530      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1531      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1532                           getPointerTy(), Callee, PICLabel);
1533    } else {
1534      // On ELF targets for PIC code, direct calls should go through the PLT
1535      unsigned OpFlags = 0;
1536      if (Subtarget->isTargetELF() &&
1537                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1538        OpFlags = ARMII::MO_PLT;
1539      Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1540    }
1541  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1542    isDirect = true;
1543    bool isStub = Subtarget->isTargetDarwin() &&
1544                  getTargetMachine().getRelocationModel() != Reloc::Static;
1545    isARMFunc = !Subtarget->isThumb() || isStub;
1546    // tBX takes a register source operand.
1547    const char *Sym = S->getSymbol();
1548    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1549      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1550      ARMConstantPoolValue *CPV =
1551        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1552                                      ARMPCLabelIndex, 4);
1553      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1554      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1555      Callee = DAG.getLoad(getPointerTy(), dl,
1556                           DAG.getEntryNode(), CPAddr,
1557                           MachinePointerInfo::getConstantPool(),
1558                           false, false, false, 0);
1559      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1560      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1561                           getPointerTy(), Callee, PICLabel);
1562    } else {
1563      unsigned OpFlags = 0;
1564      // On ELF targets for PIC code, direct calls should go through the PLT
1565      if (Subtarget->isTargetELF() &&
1566                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1567        OpFlags = ARMII::MO_PLT;
1568      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1569    }
1570  }
1571
1572  // FIXME: handle tail calls differently.
1573  unsigned CallOpc;
1574  if (Subtarget->isThumb()) {
1575    if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1576      CallOpc = ARMISD::CALL_NOLINK;
1577    else
1578      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1579  } else {
1580    CallOpc = (isDirect || Subtarget->hasV5TOps())
1581      ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1582      : ARMISD::CALL_NOLINK;
1583  }
1584
1585  std::vector<SDValue> Ops;
1586  Ops.push_back(Chain);
1587  Ops.push_back(Callee);
1588
1589  // Add argument registers to the end of the list so that they are known live
1590  // into the call.
1591  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1592    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1593                                  RegsToPass[i].second.getValueType()));
1594
1595  if (InFlag.getNode())
1596    Ops.push_back(InFlag);
1597
1598  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1599  if (isTailCall)
1600    return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1601
1602  // Returns a chain and a flag for retval copy to use.
1603  Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1604  InFlag = Chain.getValue(1);
1605
1606  Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1607                             DAG.getIntPtrConstant(0, true), InFlag);
1608  if (!Ins.empty())
1609    InFlag = Chain.getValue(1);
1610
1611  // Handle result values, copying them out of physregs into vregs that we
1612  // return.
1613  return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1614                         dl, DAG, InVals);
1615}
1616
1617/// HandleByVal - Every parameter *after* a byval parameter is passed
1618/// on the stack.  Remember the next parameter register to allocate,
1619/// and then confiscate the rest of the parameter registers to insure
1620/// this.
1621void
1622llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1623  unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1624  assert((State->getCallOrPrologue() == Prologue ||
1625          State->getCallOrPrologue() == Call) &&
1626         "unhandled ParmContext");
1627  if ((!State->isFirstByValRegValid()) &&
1628      (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1629    State->setFirstByValReg(reg);
1630    // At a call site, a byval parameter that is split between
1631    // registers and memory needs its size truncated here.  In a
1632    // function prologue, such byval parameters are reassembled in
1633    // memory, and are not truncated.
1634    if (State->getCallOrPrologue() == Call) {
1635      unsigned excess = 4 * (ARM::R4 - reg);
1636      assert(size >= excess && "expected larger existing stack allocation");
1637      size -= excess;
1638    }
1639  }
1640  // Confiscate any remaining parameter registers to preclude their
1641  // assignment to subsequent parameters.
1642  while (State->AllocateReg(GPRArgRegs, 4))
1643    ;
1644}
1645
1646/// MatchingStackOffset - Return true if the given stack call argument is
1647/// already available in the same position (relatively) of the caller's
1648/// incoming argument stack.
1649static
1650bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1651                         MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1652                         const ARMInstrInfo *TII) {
1653  unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1654  int FI = INT_MAX;
1655  if (Arg.getOpcode() == ISD::CopyFromReg) {
1656    unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1657    if (!TargetRegisterInfo::isVirtualRegister(VR))
1658      return false;
1659    MachineInstr *Def = MRI->getVRegDef(VR);
1660    if (!Def)
1661      return false;
1662    if (!Flags.isByVal()) {
1663      if (!TII->isLoadFromStackSlot(Def, FI))
1664        return false;
1665    } else {
1666      return false;
1667    }
1668  } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1669    if (Flags.isByVal())
1670      // ByVal argument is passed in as a pointer but it's now being
1671      // dereferenced. e.g.
1672      // define @foo(%struct.X* %A) {
1673      //   tail call @bar(%struct.X* byval %A)
1674      // }
1675      return false;
1676    SDValue Ptr = Ld->getBasePtr();
1677    FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1678    if (!FINode)
1679      return false;
1680    FI = FINode->getIndex();
1681  } else
1682    return false;
1683
1684  assert(FI != INT_MAX);
1685  if (!MFI->isFixedObjectIndex(FI))
1686    return false;
1687  return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1688}
1689
1690/// IsEligibleForTailCallOptimization - Check whether the call is eligible
1691/// for tail call optimization. Targets which want to do tail call
1692/// optimization should implement this function.
1693bool
1694ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1695                                                     CallingConv::ID CalleeCC,
1696                                                     bool isVarArg,
1697                                                     bool isCalleeStructRet,
1698                                                     bool isCallerStructRet,
1699                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
1700                                    const SmallVectorImpl<SDValue> &OutVals,
1701                                    const SmallVectorImpl<ISD::InputArg> &Ins,
1702                                                     SelectionDAG& DAG) const {
1703  const Function *CallerF = DAG.getMachineFunction().getFunction();
1704  CallingConv::ID CallerCC = CallerF->getCallingConv();
1705  bool CCMatch = CallerCC == CalleeCC;
1706
1707  // Look for obvious safe cases to perform tail call optimization that do not
1708  // require ABI changes. This is what gcc calls sibcall.
1709
1710  // Do not sibcall optimize vararg calls unless the call site is not passing
1711  // any arguments.
1712  if (isVarArg && !Outs.empty())
1713    return false;
1714
1715  // Also avoid sibcall optimization if either caller or callee uses struct
1716  // return semantics.
1717  if (isCalleeStructRet || isCallerStructRet)
1718    return false;
1719
1720  // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1721  // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1722  // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1723  // support in the assembler and linker to be used. This would need to be
1724  // fixed to fully support tail calls in Thumb1.
1725  //
1726  // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1727  // LR.  This means if we need to reload LR, it takes an extra instructions,
1728  // which outweighs the value of the tail call; but here we don't know yet
1729  // whether LR is going to be used.  Probably the right approach is to
1730  // generate the tail call here and turn it back into CALL/RET in
1731  // emitEpilogue if LR is used.
1732
1733  // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1734  // but we need to make sure there are enough registers; the only valid
1735  // registers are the 4 used for parameters.  We don't currently do this
1736  // case.
1737  if (Subtarget->isThumb1Only())
1738    return false;
1739
1740  // If the calling conventions do not match, then we'd better make sure the
1741  // results are returned in the same way as what the caller expects.
1742  if (!CCMatch) {
1743    SmallVector<CCValAssign, 16> RVLocs1;
1744    ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1745                       getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1746    CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1747
1748    SmallVector<CCValAssign, 16> RVLocs2;
1749    ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1750                       getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1751    CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1752
1753    if (RVLocs1.size() != RVLocs2.size())
1754      return false;
1755    for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1756      if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1757        return false;
1758      if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1759        return false;
1760      if (RVLocs1[i].isRegLoc()) {
1761        if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1762          return false;
1763      } else {
1764        if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1765          return false;
1766      }
1767    }
1768  }
1769
1770  // If the callee takes no arguments then go on to check the results of the
1771  // call.
1772  if (!Outs.empty()) {
1773    // Check if stack adjustment is needed. For now, do not do this if any
1774    // argument is passed on the stack.
1775    SmallVector<CCValAssign, 16> ArgLocs;
1776    ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1777                      getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1778    CCInfo.AnalyzeCallOperands(Outs,
1779                               CCAssignFnForNode(CalleeCC, false, isVarArg));
1780    if (CCInfo.getNextStackOffset()) {
1781      MachineFunction &MF = DAG.getMachineFunction();
1782
1783      // Check if the arguments are already laid out in the right way as
1784      // the caller's fixed stack objects.
1785      MachineFrameInfo *MFI = MF.getFrameInfo();
1786      const MachineRegisterInfo *MRI = &MF.getRegInfo();
1787      const ARMInstrInfo *TII =
1788        ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1789      for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1790           i != e;
1791           ++i, ++realArgIdx) {
1792        CCValAssign &VA = ArgLocs[i];
1793        EVT RegVT = VA.getLocVT();
1794        SDValue Arg = OutVals[realArgIdx];
1795        ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1796        if (VA.getLocInfo() == CCValAssign::Indirect)
1797          return false;
1798        if (VA.needsCustom()) {
1799          // f64 and vector types are split into multiple registers or
1800          // register/stack-slot combinations.  The types will not match
1801          // the registers; give up on memory f64 refs until we figure
1802          // out what to do about this.
1803          if (!VA.isRegLoc())
1804            return false;
1805          if (!ArgLocs[++i].isRegLoc())
1806            return false;
1807          if (RegVT == MVT::v2f64) {
1808            if (!ArgLocs[++i].isRegLoc())
1809              return false;
1810            if (!ArgLocs[++i].isRegLoc())
1811              return false;
1812          }
1813        } else if (!VA.isRegLoc()) {
1814          if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1815                                   MFI, MRI, TII))
1816            return false;
1817        }
1818      }
1819    }
1820  }
1821
1822  return true;
1823}
1824
1825SDValue
1826ARMTargetLowering::LowerReturn(SDValue Chain,
1827                               CallingConv::ID CallConv, bool isVarArg,
1828                               const SmallVectorImpl<ISD::OutputArg> &Outs,
1829                               const SmallVectorImpl<SDValue> &OutVals,
1830                               DebugLoc dl, SelectionDAG &DAG) const {
1831
1832  // CCValAssign - represent the assignment of the return value to a location.
1833  SmallVector<CCValAssign, 16> RVLocs;
1834
1835  // CCState - Info about the registers and stack slots.
1836  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1837                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1838
1839  // Analyze outgoing return values.
1840  CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1841                                               isVarArg));
1842
1843  // If this is the first return lowered for this function, add
1844  // the regs to the liveout set for the function.
1845  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1846    for (unsigned i = 0; i != RVLocs.size(); ++i)
1847      if (RVLocs[i].isRegLoc())
1848        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1849  }
1850
1851  SDValue Flag;
1852
1853  // Copy the result values into the output registers.
1854  for (unsigned i = 0, realRVLocIdx = 0;
1855       i != RVLocs.size();
1856       ++i, ++realRVLocIdx) {
1857    CCValAssign &VA = RVLocs[i];
1858    assert(VA.isRegLoc() && "Can only return in registers!");
1859
1860    SDValue Arg = OutVals[realRVLocIdx];
1861
1862    switch (VA.getLocInfo()) {
1863    default: llvm_unreachable("Unknown loc info!");
1864    case CCValAssign::Full: break;
1865    case CCValAssign::BCvt:
1866      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1867      break;
1868    }
1869
1870    if (VA.needsCustom()) {
1871      if (VA.getLocVT() == MVT::v2f64) {
1872        // Extract the first half and return it in two registers.
1873        SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1874                                   DAG.getConstant(0, MVT::i32));
1875        SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1876                                       DAG.getVTList(MVT::i32, MVT::i32), Half);
1877
1878        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1879        Flag = Chain.getValue(1);
1880        VA = RVLocs[++i]; // skip ahead to next loc
1881        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1882                                 HalfGPRs.getValue(1), Flag);
1883        Flag = Chain.getValue(1);
1884        VA = RVLocs[++i]; // skip ahead to next loc
1885
1886        // Extract the 2nd half and fall through to handle it as an f64 value.
1887        Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1888                          DAG.getConstant(1, MVT::i32));
1889      }
1890      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
1891      // available.
1892      SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1893                                  DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1894      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1895      Flag = Chain.getValue(1);
1896      VA = RVLocs[++i]; // skip ahead to next loc
1897      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1898                               Flag);
1899    } else
1900      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1901
1902    // Guarantee that all emitted copies are
1903    // stuck together, avoiding something bad.
1904    Flag = Chain.getValue(1);
1905  }
1906
1907  SDValue result;
1908  if (Flag.getNode())
1909    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1910  else // Return Void
1911    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1912
1913  return result;
1914}
1915
1916bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1917  if (N->getNumValues() != 1)
1918    return false;
1919  if (!N->hasNUsesOfValue(1, 0))
1920    return false;
1921
1922  unsigned NumCopies = 0;
1923  SDNode* Copies[2] = { 0, 0 };
1924  SDNode *Use = *N->use_begin();
1925  if (Use->getOpcode() == ISD::CopyToReg) {
1926    Copies[NumCopies++] = Use;
1927  } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1928    // f64 returned in a pair of GPRs.
1929    for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1930         UI != UE; ++UI) {
1931      if (UI->getOpcode() != ISD::CopyToReg)
1932        return false;
1933      Copies[UI.getUse().getResNo()] = *UI;
1934      ++NumCopies;
1935    }
1936  } else if (Use->getOpcode() == ISD::BITCAST) {
1937    // f32 returned in a single GPR.
1938    if (!Use->hasNUsesOfValue(1, 0))
1939      return false;
1940    Use = *Use->use_begin();
1941    if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1942      return false;
1943    Copies[NumCopies++] = Use;
1944  } else {
1945    return false;
1946  }
1947
1948  if (NumCopies != 1 && NumCopies != 2)
1949    return false;
1950
1951  bool HasRet = false;
1952  for (unsigned i = 0; i < NumCopies; ++i) {
1953    SDNode *Copy = Copies[i];
1954    for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1955         UI != UE; ++UI) {
1956      if (UI->getOpcode() == ISD::CopyToReg) {
1957        SDNode *Use = *UI;
1958        if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1959          continue;
1960        return false;
1961      }
1962      if (UI->getOpcode() != ARMISD::RET_FLAG)
1963        return false;
1964      HasRet = true;
1965    }
1966  }
1967
1968  return HasRet;
1969}
1970
1971bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1972  if (!EnableARMTailCalls)
1973    return false;
1974
1975  if (!CI->isTailCall())
1976    return false;
1977
1978  return !Subtarget->isThumb1Only();
1979}
1980
1981// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1982// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1983// one of the above mentioned nodes. It has to be wrapped because otherwise
1984// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1985// be used to form addressing mode. These wrapped nodes will be selected
1986// into MOVi.
1987static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1988  EVT PtrVT = Op.getValueType();
1989  // FIXME there is no actual debug info here
1990  DebugLoc dl = Op.getDebugLoc();
1991  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1992  SDValue Res;
1993  if (CP->isMachineConstantPoolEntry())
1994    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1995                                    CP->getAlignment());
1996  else
1997    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1998                                    CP->getAlignment());
1999  return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2000}
2001
2002unsigned ARMTargetLowering::getJumpTableEncoding() const {
2003  return MachineJumpTableInfo::EK_Inline;
2004}
2005
2006SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2007                                             SelectionDAG &DAG) const {
2008  MachineFunction &MF = DAG.getMachineFunction();
2009  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2010  unsigned ARMPCLabelIndex = 0;
2011  DebugLoc DL = Op.getDebugLoc();
2012  EVT PtrVT = getPointerTy();
2013  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2014  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2015  SDValue CPAddr;
2016  if (RelocM == Reloc::Static) {
2017    CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2018  } else {
2019    unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2020    ARMPCLabelIndex = AFI->createPICLabelUId();
2021    ARMConstantPoolValue *CPV =
2022      ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2023                                      ARMCP::CPBlockAddress, PCAdj);
2024    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2025  }
2026  CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2027  SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2028                               MachinePointerInfo::getConstantPool(),
2029                               false, false, false, 0);
2030  if (RelocM == Reloc::Static)
2031    return Result;
2032  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2033  return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2034}
2035
2036// Lower ISD::GlobalTLSAddress using the "general dynamic" model
2037SDValue
2038ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2039                                                 SelectionDAG &DAG) const {
2040  DebugLoc dl = GA->getDebugLoc();
2041  EVT PtrVT = getPointerTy();
2042  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2043  MachineFunction &MF = DAG.getMachineFunction();
2044  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2045  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2046  ARMConstantPoolValue *CPV =
2047    ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2048                                    ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2049  SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2050  Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2051  Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2052                         MachinePointerInfo::getConstantPool(),
2053                         false, false, false, 0);
2054  SDValue Chain = Argument.getValue(1);
2055
2056  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2057  Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2058
2059  // call __tls_get_addr.
2060  ArgListTy Args;
2061  ArgListEntry Entry;
2062  Entry.Node = Argument;
2063  Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2064  Args.push_back(Entry);
2065  // FIXME: is there useful debug info available here?
2066  std::pair<SDValue, SDValue> CallResult =
2067    LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2068                false, false, false, false,
2069                0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2070                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2071  return CallResult.first;
2072}
2073
2074// Lower ISD::GlobalTLSAddress using the "initial exec" or
2075// "local exec" model.
2076SDValue
2077ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2078                                        SelectionDAG &DAG) const {
2079  const GlobalValue *GV = GA->getGlobal();
2080  DebugLoc dl = GA->getDebugLoc();
2081  SDValue Offset;
2082  SDValue Chain = DAG.getEntryNode();
2083  EVT PtrVT = getPointerTy();
2084  // Get the Thread Pointer
2085  SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2086
2087  if (GV->isDeclaration()) {
2088    MachineFunction &MF = DAG.getMachineFunction();
2089    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2090    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2091    // Initial exec model.
2092    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2093    ARMConstantPoolValue *CPV =
2094      ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2095                                      ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2096                                      true);
2097    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2098    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2099    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2100                         MachinePointerInfo::getConstantPool(),
2101                         false, false, false, 0);
2102    Chain = Offset.getValue(1);
2103
2104    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2105    Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2106
2107    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2108                         MachinePointerInfo::getConstantPool(),
2109                         false, false, false, 0);
2110  } else {
2111    // local exec model
2112    ARMConstantPoolValue *CPV =
2113      ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2114    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2115    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2116    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2117                         MachinePointerInfo::getConstantPool(),
2118                         false, false, false, 0);
2119  }
2120
2121  // The address of the thread local variable is the add of the thread
2122  // pointer with the offset of the variable.
2123  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2124}
2125
2126SDValue
2127ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2128  // TODO: implement the "local dynamic" model
2129  assert(Subtarget->isTargetELF() &&
2130         "TLS not implemented for non-ELF targets");
2131  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2132  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2133  // otherwise use the "Local Exec" TLS Model
2134  if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2135    return LowerToTLSGeneralDynamicModel(GA, DAG);
2136  else
2137    return LowerToTLSExecModels(GA, DAG);
2138}
2139
2140SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2141                                                 SelectionDAG &DAG) const {
2142  EVT PtrVT = getPointerTy();
2143  DebugLoc dl = Op.getDebugLoc();
2144  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2145  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2146  if (RelocM == Reloc::PIC_) {
2147    bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2148    ARMConstantPoolValue *CPV =
2149      ARMConstantPoolConstant::Create(GV,
2150                                      UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2151    SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2152    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2153    SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2154                                 CPAddr,
2155                                 MachinePointerInfo::getConstantPool(),
2156                                 false, false, false, 0);
2157    SDValue Chain = Result.getValue(1);
2158    SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2159    Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2160    if (!UseGOTOFF)
2161      Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2162                           MachinePointerInfo::getGOT(),
2163                           false, false, false, 0);
2164    return Result;
2165  }
2166
2167  // If we have T2 ops, we can materialize the address directly via movt/movw
2168  // pair. This is always cheaper.
2169  if (Subtarget->useMovt()) {
2170    ++NumMovwMovt;
2171    // FIXME: Once remat is capable of dealing with instructions with register
2172    // operands, expand this into two nodes.
2173    return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2174                       DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2175  } else {
2176    SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2177    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2178    return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2179                       MachinePointerInfo::getConstantPool(),
2180                       false, false, false, 0);
2181  }
2182}
2183
2184SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2185                                                    SelectionDAG &DAG) const {
2186  EVT PtrVT = getPointerTy();
2187  DebugLoc dl = Op.getDebugLoc();
2188  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2189  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2190  MachineFunction &MF = DAG.getMachineFunction();
2191  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2192
2193  // FIXME: Enable this for static codegen when tool issues are fixed.  Also
2194  // update ARMFastISel::ARMMaterializeGV.
2195  if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2196    ++NumMovwMovt;
2197    // FIXME: Once remat is capable of dealing with instructions with register
2198    // operands, expand this into two nodes.
2199    if (RelocM == Reloc::Static)
2200      return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2201                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2202
2203    unsigned Wrapper = (RelocM == Reloc::PIC_)
2204      ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2205    SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2206                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2207    if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2208      Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2209                           MachinePointerInfo::getGOT(),
2210                           false, false, false, 0);
2211    return Result;
2212  }
2213
2214  unsigned ARMPCLabelIndex = 0;
2215  SDValue CPAddr;
2216  if (RelocM == Reloc::Static) {
2217    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2218  } else {
2219    ARMPCLabelIndex = AFI->createPICLabelUId();
2220    unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2221    ARMConstantPoolValue *CPV =
2222      ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2223                                      PCAdj);
2224    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2225  }
2226  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2227
2228  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2229                               MachinePointerInfo::getConstantPool(),
2230                               false, false, false, 0);
2231  SDValue Chain = Result.getValue(1);
2232
2233  if (RelocM == Reloc::PIC_) {
2234    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2235    Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2236  }
2237
2238  if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2239    Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2240                         false, false, false, 0);
2241
2242  return Result;
2243}
2244
2245SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2246                                                    SelectionDAG &DAG) const {
2247  assert(Subtarget->isTargetELF() &&
2248         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2249  MachineFunction &MF = DAG.getMachineFunction();
2250  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2251  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2252  EVT PtrVT = getPointerTy();
2253  DebugLoc dl = Op.getDebugLoc();
2254  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2255  ARMConstantPoolValue *CPV =
2256    ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2257                                  ARMPCLabelIndex, PCAdj);
2258  SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2259  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2260  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2261                               MachinePointerInfo::getConstantPool(),
2262                               false, false, false, 0);
2263  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2264  return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2265}
2266
2267SDValue
2268ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2269  DebugLoc dl = Op.getDebugLoc();
2270  SDValue Val = DAG.getConstant(0, MVT::i32);
2271  return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2272                     DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2273                     Op.getOperand(1), Val);
2274}
2275
2276SDValue
2277ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2278  DebugLoc dl = Op.getDebugLoc();
2279  return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2280                     Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2281}
2282
2283SDValue
2284ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2285                                          const ARMSubtarget *Subtarget) const {
2286  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2287  DebugLoc dl = Op.getDebugLoc();
2288  switch (IntNo) {
2289  default: return SDValue();    // Don't custom lower most intrinsics.
2290  case Intrinsic::arm_thread_pointer: {
2291    EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2292    return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2293  }
2294  case Intrinsic::eh_sjlj_lsda: {
2295    MachineFunction &MF = DAG.getMachineFunction();
2296    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2297    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2298    EVT PtrVT = getPointerTy();
2299    DebugLoc dl = Op.getDebugLoc();
2300    Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2301    SDValue CPAddr;
2302    unsigned PCAdj = (RelocM != Reloc::PIC_)
2303      ? 0 : (Subtarget->isThumb() ? 4 : 8);
2304    ARMConstantPoolValue *CPV =
2305      ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2306                                      ARMCP::CPLSDA, PCAdj);
2307    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2308    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2309    SDValue Result =
2310      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2311                  MachinePointerInfo::getConstantPool(),
2312                  false, false, false, 0);
2313
2314    if (RelocM == Reloc::PIC_) {
2315      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2316      Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2317    }
2318    return Result;
2319  }
2320  case Intrinsic::arm_neon_vmulls:
2321  case Intrinsic::arm_neon_vmullu: {
2322    unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2323      ? ARMISD::VMULLs : ARMISD::VMULLu;
2324    return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2325                       Op.getOperand(1), Op.getOperand(2));
2326  }
2327  }
2328}
2329
2330static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2331                               const ARMSubtarget *Subtarget) {
2332  DebugLoc dl = Op.getDebugLoc();
2333  if (!Subtarget->hasDataBarrier()) {
2334    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2335    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2336    // here.
2337    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2338           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2339    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2340                       DAG.getConstant(0, MVT::i32));
2341  }
2342
2343  SDValue Op5 = Op.getOperand(5);
2344  bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2345  unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2346  unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2347  bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2348
2349  ARM_MB::MemBOpt DMBOpt;
2350  if (isDeviceBarrier)
2351    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2352  else
2353    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2354  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2355                     DAG.getConstant(DMBOpt, MVT::i32));
2356}
2357
2358
2359static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2360                                 const ARMSubtarget *Subtarget) {
2361  // FIXME: handle "fence singlethread" more efficiently.
2362  DebugLoc dl = Op.getDebugLoc();
2363  if (!Subtarget->hasDataBarrier()) {
2364    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2365    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2366    // here.
2367    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2368           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2369    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2370                       DAG.getConstant(0, MVT::i32));
2371  }
2372
2373  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2374                     DAG.getConstant(ARM_MB::ISH, MVT::i32));
2375}
2376
2377static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2378                             const ARMSubtarget *Subtarget) {
2379  // ARM pre v5TE and Thumb1 does not have preload instructions.
2380  if (!(Subtarget->isThumb2() ||
2381        (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2382    // Just preserve the chain.
2383    return Op.getOperand(0);
2384
2385  DebugLoc dl = Op.getDebugLoc();
2386  unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2387  if (!isRead &&
2388      (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2389    // ARMv7 with MP extension has PLDW.
2390    return Op.getOperand(0);
2391
2392  unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2393  if (Subtarget->isThumb()) {
2394    // Invert the bits.
2395    isRead = ~isRead & 1;
2396    isData = ~isData & 1;
2397  }
2398
2399  return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2400                     Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2401                     DAG.getConstant(isData, MVT::i32));
2402}
2403
2404static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2405  MachineFunction &MF = DAG.getMachineFunction();
2406  ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2407
2408  // vastart just stores the address of the VarArgsFrameIndex slot into the
2409  // memory location argument.
2410  DebugLoc dl = Op.getDebugLoc();
2411  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2412  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2413  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2414  return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2415                      MachinePointerInfo(SV), false, false, 0);
2416}
2417
2418SDValue
2419ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2420                                        SDValue &Root, SelectionDAG &DAG,
2421                                        DebugLoc dl) const {
2422  MachineFunction &MF = DAG.getMachineFunction();
2423  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2424
2425  const TargetRegisterClass *RC;
2426  if (AFI->isThumb1OnlyFunction())
2427    RC = ARM::tGPRRegisterClass;
2428  else
2429    RC = ARM::GPRRegisterClass;
2430
2431  // Transform the arguments stored in physical registers into virtual ones.
2432  unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2433  SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2434
2435  SDValue ArgValue2;
2436  if (NextVA.isMemLoc()) {
2437    MachineFrameInfo *MFI = MF.getFrameInfo();
2438    int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2439
2440    // Create load node to retrieve arguments from the stack.
2441    SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2442    ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2443                            MachinePointerInfo::getFixedStack(FI),
2444                            false, false, false, 0);
2445  } else {
2446    Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2447    ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2448  }
2449
2450  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2451}
2452
2453void
2454ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2455                                  unsigned &VARegSize, unsigned &VARegSaveSize)
2456  const {
2457  unsigned NumGPRs;
2458  if (CCInfo.isFirstByValRegValid())
2459    NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2460  else {
2461    unsigned int firstUnalloced;
2462    firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2463                                                sizeof(GPRArgRegs) /
2464                                                sizeof(GPRArgRegs[0]));
2465    NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2466  }
2467
2468  unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2469  VARegSize = NumGPRs * 4;
2470  VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2471}
2472
2473// The remaining GPRs hold either the beginning of variable-argument
2474// data, or the beginning of an aggregate passed by value (usuall
2475// byval).  Either way, we allocate stack slots adjacent to the data
2476// provided by our caller, and store the unallocated registers there.
2477// If this is a variadic function, the va_list pointer will begin with
2478// these values; otherwise, this reassembles a (byval) structure that
2479// was split between registers and memory.
2480void
2481ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2482                                        DebugLoc dl, SDValue &Chain,
2483                                        unsigned ArgOffset) const {
2484  MachineFunction &MF = DAG.getMachineFunction();
2485  MachineFrameInfo *MFI = MF.getFrameInfo();
2486  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2487  unsigned firstRegToSaveIndex;
2488  if (CCInfo.isFirstByValRegValid())
2489    firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2490  else {
2491    firstRegToSaveIndex = CCInfo.getFirstUnallocated
2492      (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2493  }
2494
2495  unsigned VARegSize, VARegSaveSize;
2496  computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2497  if (VARegSaveSize) {
2498    // If this function is vararg, store any remaining integer argument regs
2499    // to their spots on the stack so that they may be loaded by deferencing
2500    // the result of va_next.
2501    AFI->setVarArgsRegSaveSize(VARegSaveSize);
2502    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2503                                                     ArgOffset + VARegSaveSize
2504                                                     - VARegSize,
2505                                                     false));
2506    SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2507                                    getPointerTy());
2508
2509    SmallVector<SDValue, 4> MemOps;
2510    for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2511      const TargetRegisterClass *RC;
2512      if (AFI->isThumb1OnlyFunction())
2513        RC = ARM::tGPRRegisterClass;
2514      else
2515        RC = ARM::GPRRegisterClass;
2516
2517      unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2518      SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2519      SDValue Store =
2520        DAG.getStore(Val.getValue(1), dl, Val, FIN,
2521                 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2522                     false, false, 0);
2523      MemOps.push_back(Store);
2524      FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2525                        DAG.getConstant(4, getPointerTy()));
2526    }
2527    if (!MemOps.empty())
2528      Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2529                          &MemOps[0], MemOps.size());
2530  } else
2531    // This will point to the next argument passed via stack.
2532    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2533}
2534
2535SDValue
2536ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2537                                        CallingConv::ID CallConv, bool isVarArg,
2538                                        const SmallVectorImpl<ISD::InputArg>
2539                                          &Ins,
2540                                        DebugLoc dl, SelectionDAG &DAG,
2541                                        SmallVectorImpl<SDValue> &InVals)
2542                                          const {
2543  MachineFunction &MF = DAG.getMachineFunction();
2544  MachineFrameInfo *MFI = MF.getFrameInfo();
2545
2546  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2547
2548  // Assign locations to all of the incoming arguments.
2549  SmallVector<CCValAssign, 16> ArgLocs;
2550  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2551                    getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2552  CCInfo.AnalyzeFormalArguments(Ins,
2553                                CCAssignFnForNode(CallConv, /* Return*/ false,
2554                                                  isVarArg));
2555
2556  SmallVector<SDValue, 16> ArgValues;
2557  int lastInsIndex = -1;
2558
2559  SDValue ArgValue;
2560  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2561    CCValAssign &VA = ArgLocs[i];
2562
2563    // Arguments stored in registers.
2564    if (VA.isRegLoc()) {
2565      EVT RegVT = VA.getLocVT();
2566
2567      if (VA.needsCustom()) {
2568        // f64 and vector types are split up into multiple registers or
2569        // combinations of registers and stack slots.
2570        if (VA.getLocVT() == MVT::v2f64) {
2571          SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2572                                                   Chain, DAG, dl);
2573          VA = ArgLocs[++i]; // skip ahead to next loc
2574          SDValue ArgValue2;
2575          if (VA.isMemLoc()) {
2576            int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2577            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2578            ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2579                                    MachinePointerInfo::getFixedStack(FI),
2580                                    false, false, false, 0);
2581          } else {
2582            ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2583                                             Chain, DAG, dl);
2584          }
2585          ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2586          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2587                                 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2588          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2589                                 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2590        } else
2591          ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2592
2593      } else {
2594        const TargetRegisterClass *RC;
2595
2596        if (RegVT == MVT::f32)
2597          RC = ARM::SPRRegisterClass;
2598        else if (RegVT == MVT::f64)
2599          RC = ARM::DPRRegisterClass;
2600        else if (RegVT == MVT::v2f64)
2601          RC = ARM::QPRRegisterClass;
2602        else if (RegVT == MVT::i32)
2603          RC = (AFI->isThumb1OnlyFunction() ?
2604                ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2605        else
2606          llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2607
2608        // Transform the arguments in physical registers into virtual ones.
2609        unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2610        ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2611      }
2612
2613      // If this is an 8 or 16-bit value, it is really passed promoted
2614      // to 32 bits.  Insert an assert[sz]ext to capture this, then
2615      // truncate to the right size.
2616      switch (VA.getLocInfo()) {
2617      default: llvm_unreachable("Unknown loc info!");
2618      case CCValAssign::Full: break;
2619      case CCValAssign::BCvt:
2620        ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2621        break;
2622      case CCValAssign::SExt:
2623        ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2624                               DAG.getValueType(VA.getValVT()));
2625        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2626        break;
2627      case CCValAssign::ZExt:
2628        ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2629                               DAG.getValueType(VA.getValVT()));
2630        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2631        break;
2632      }
2633
2634      InVals.push_back(ArgValue);
2635
2636    } else { // VA.isRegLoc()
2637
2638      // sanity check
2639      assert(VA.isMemLoc());
2640      assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2641
2642      int index = ArgLocs[i].getValNo();
2643
2644      // Some Ins[] entries become multiple ArgLoc[] entries.
2645      // Process them only once.
2646      if (index != lastInsIndex)
2647        {
2648          ISD::ArgFlagsTy Flags = Ins[index].Flags;
2649          // FIXME: For now, all byval parameter objects are marked mutable.
2650          // This can be changed with more analysis.
2651          // In case of tail call optimization mark all arguments mutable.
2652          // Since they could be overwritten by lowering of arguments in case of
2653          // a tail call.
2654          if (Flags.isByVal()) {
2655            unsigned VARegSize, VARegSaveSize;
2656            computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2657            VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2658            unsigned Bytes = Flags.getByValSize() - VARegSize;
2659            if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2660            int FI = MFI->CreateFixedObject(Bytes,
2661                                            VA.getLocMemOffset(), false);
2662            InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2663          } else {
2664            int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2665                                            VA.getLocMemOffset(), true);
2666
2667            // Create load nodes to retrieve arguments from the stack.
2668            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2669            InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2670                                         MachinePointerInfo::getFixedStack(FI),
2671                                         false, false, false, 0));
2672          }
2673          lastInsIndex = index;
2674        }
2675    }
2676  }
2677
2678  // varargs
2679  if (isVarArg)
2680    VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2681
2682  return Chain;
2683}
2684
2685/// isFloatingPointZero - Return true if this is +0.0.
2686static bool isFloatingPointZero(SDValue Op) {
2687  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2688    return CFP->getValueAPF().isPosZero();
2689  else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2690    // Maybe this has already been legalized into the constant pool?
2691    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2692      SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2693      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2694        if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2695          return CFP->getValueAPF().isPosZero();
2696    }
2697  }
2698  return false;
2699}
2700
2701/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2702/// the given operands.
2703SDValue
2704ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2705                             SDValue &ARMcc, SelectionDAG &DAG,
2706                             DebugLoc dl) const {
2707  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2708    unsigned C = RHSC->getZExtValue();
2709    if (!isLegalICmpImmediate(C)) {
2710      // Constant does not fit, try adjusting it by one?
2711      switch (CC) {
2712      default: break;
2713      case ISD::SETLT:
2714      case ISD::SETGE:
2715        if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2716          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2717          RHS = DAG.getConstant(C-1, MVT::i32);
2718        }
2719        break;
2720      case ISD::SETULT:
2721      case ISD::SETUGE:
2722        if (C != 0 && isLegalICmpImmediate(C-1)) {
2723          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2724          RHS = DAG.getConstant(C-1, MVT::i32);
2725        }
2726        break;
2727      case ISD::SETLE:
2728      case ISD::SETGT:
2729        if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2730          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2731          RHS = DAG.getConstant(C+1, MVT::i32);
2732        }
2733        break;
2734      case ISD::SETULE:
2735      case ISD::SETUGT:
2736        if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2737          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2738          RHS = DAG.getConstant(C+1, MVT::i32);
2739        }
2740        break;
2741      }
2742    }
2743  }
2744
2745  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2746  ARMISD::NodeType CompareType;
2747  switch (CondCode) {
2748  default:
2749    CompareType = ARMISD::CMP;
2750    break;
2751  case ARMCC::EQ:
2752  case ARMCC::NE:
2753    // Uses only Z Flag
2754    CompareType = ARMISD::CMPZ;
2755    break;
2756  }
2757  ARMcc = DAG.getConstant(CondCode, MVT::i32);
2758  return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2759}
2760
2761/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2762SDValue
2763ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2764                             DebugLoc dl) const {
2765  SDValue Cmp;
2766  if (!isFloatingPointZero(RHS))
2767    Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2768  else
2769    Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2770  return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2771}
2772
2773/// duplicateCmp - Glue values can have only one use, so this function
2774/// duplicates a comparison node.
2775SDValue
2776ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2777  unsigned Opc = Cmp.getOpcode();
2778  DebugLoc DL = Cmp.getDebugLoc();
2779  if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2780    return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2781
2782  assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2783  Cmp = Cmp.getOperand(0);
2784  Opc = Cmp.getOpcode();
2785  if (Opc == ARMISD::CMPFP)
2786    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2787  else {
2788    assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2789    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2790  }
2791  return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2792}
2793
2794SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2795  SDValue Cond = Op.getOperand(0);
2796  SDValue SelectTrue = Op.getOperand(1);
2797  SDValue SelectFalse = Op.getOperand(2);
2798  DebugLoc dl = Op.getDebugLoc();
2799
2800  // Convert:
2801  //
2802  //   (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2803  //   (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2804  //
2805  if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2806    const ConstantSDNode *CMOVTrue =
2807      dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2808    const ConstantSDNode *CMOVFalse =
2809      dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2810
2811    if (CMOVTrue && CMOVFalse) {
2812      unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2813      unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2814
2815      SDValue True;
2816      SDValue False;
2817      if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2818        True = SelectTrue;
2819        False = SelectFalse;
2820      } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2821        True = SelectFalse;
2822        False = SelectTrue;
2823      }
2824
2825      if (True.getNode() && False.getNode()) {
2826        EVT VT = Op.getValueType();
2827        SDValue ARMcc = Cond.getOperand(2);
2828        SDValue CCR = Cond.getOperand(3);
2829        SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2830        assert(True.getValueType() == VT);
2831        return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2832      }
2833    }
2834  }
2835
2836  return DAG.getSelectCC(dl, Cond,
2837                         DAG.getConstant(0, Cond.getValueType()),
2838                         SelectTrue, SelectFalse, ISD::SETNE);
2839}
2840
2841SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2842  EVT VT = Op.getValueType();
2843  SDValue LHS = Op.getOperand(0);
2844  SDValue RHS = Op.getOperand(1);
2845  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2846  SDValue TrueVal = Op.getOperand(2);
2847  SDValue FalseVal = Op.getOperand(3);
2848  DebugLoc dl = Op.getDebugLoc();
2849
2850  if (LHS.getValueType() == MVT::i32) {
2851    SDValue ARMcc;
2852    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2853    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2854    return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2855  }
2856
2857  ARMCC::CondCodes CondCode, CondCode2;
2858  FPCCToARMCC(CC, CondCode, CondCode2);
2859
2860  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2861  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2862  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2863  SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2864                               ARMcc, CCR, Cmp);
2865  if (CondCode2 != ARMCC::AL) {
2866    SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2867    // FIXME: Needs another CMP because flag can have but one use.
2868    SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2869    Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2870                         Result, TrueVal, ARMcc2, CCR, Cmp2);
2871  }
2872  return Result;
2873}
2874
2875/// canChangeToInt - Given the fp compare operand, return true if it is suitable
2876/// to morph to an integer compare sequence.
2877static bool canChangeToInt(SDValue Op, bool &SeenZero,
2878                           const ARMSubtarget *Subtarget) {
2879  SDNode *N = Op.getNode();
2880  if (!N->hasOneUse())
2881    // Otherwise it requires moving the value from fp to integer registers.
2882    return false;
2883  if (!N->getNumValues())
2884    return false;
2885  EVT VT = Op.getValueType();
2886  if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2887    // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2888    // vmrs are very slow, e.g. cortex-a8.
2889    return false;
2890
2891  if (isFloatingPointZero(Op)) {
2892    SeenZero = true;
2893    return true;
2894  }
2895  return ISD::isNormalLoad(N);
2896}
2897
2898static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2899  if (isFloatingPointZero(Op))
2900    return DAG.getConstant(0, MVT::i32);
2901
2902  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2903    return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2904                       Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2905                       Ld->isVolatile(), Ld->isNonTemporal(),
2906                       Ld->isInvariant(), Ld->getAlignment());
2907
2908  llvm_unreachable("Unknown VFP cmp argument!");
2909}
2910
2911static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2912                           SDValue &RetVal1, SDValue &RetVal2) {
2913  if (isFloatingPointZero(Op)) {
2914    RetVal1 = DAG.getConstant(0, MVT::i32);
2915    RetVal2 = DAG.getConstant(0, MVT::i32);
2916    return;
2917  }
2918
2919  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2920    SDValue Ptr = Ld->getBasePtr();
2921    RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2922                          Ld->getChain(), Ptr,
2923                          Ld->getPointerInfo(),
2924                          Ld->isVolatile(), Ld->isNonTemporal(),
2925                          Ld->isInvariant(), Ld->getAlignment());
2926
2927    EVT PtrType = Ptr.getValueType();
2928    unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2929    SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2930                                 PtrType, Ptr, DAG.getConstant(4, PtrType));
2931    RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2932                          Ld->getChain(), NewPtr,
2933                          Ld->getPointerInfo().getWithOffset(4),
2934                          Ld->isVolatile(), Ld->isNonTemporal(),
2935                          Ld->isInvariant(), NewAlign);
2936    return;
2937  }
2938
2939  llvm_unreachable("Unknown VFP cmp argument!");
2940}
2941
2942/// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2943/// f32 and even f64 comparisons to integer ones.
2944SDValue
2945ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2946  SDValue Chain = Op.getOperand(0);
2947  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2948  SDValue LHS = Op.getOperand(2);
2949  SDValue RHS = Op.getOperand(3);
2950  SDValue Dest = Op.getOperand(4);
2951  DebugLoc dl = Op.getDebugLoc();
2952
2953  bool SeenZero = false;
2954  if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2955      canChangeToInt(RHS, SeenZero, Subtarget) &&
2956      // If one of the operand is zero, it's safe to ignore the NaN case since
2957      // we only care about equality comparisons.
2958      (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2959    // If unsafe fp math optimization is enabled and there are no other uses of
2960    // the CMP operands, and the condition code is EQ or NE, we can optimize it
2961    // to an integer comparison.
2962    if (CC == ISD::SETOEQ)
2963      CC = ISD::SETEQ;
2964    else if (CC == ISD::SETUNE)
2965      CC = ISD::SETNE;
2966
2967    SDValue ARMcc;
2968    if (LHS.getValueType() == MVT::f32) {
2969      LHS = bitcastf32Toi32(LHS, DAG);
2970      RHS = bitcastf32Toi32(RHS, DAG);
2971      SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2972      SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2973      return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2974                         Chain, Dest, ARMcc, CCR, Cmp);
2975    }
2976
2977    SDValue LHS1, LHS2;
2978    SDValue RHS1, RHS2;
2979    expandf64Toi32(LHS, DAG, LHS1, LHS2);
2980    expandf64Toi32(RHS, DAG, RHS1, RHS2);
2981    ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2982    ARMcc = DAG.getConstant(CondCode, MVT::i32);
2983    SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2984    SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2985    return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2986  }
2987
2988  return SDValue();
2989}
2990
2991SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2992  SDValue Chain = Op.getOperand(0);
2993  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2994  SDValue LHS = Op.getOperand(2);
2995  SDValue RHS = Op.getOperand(3);
2996  SDValue Dest = Op.getOperand(4);
2997  DebugLoc dl = Op.getDebugLoc();
2998
2999  if (LHS.getValueType() == MVT::i32) {
3000    SDValue ARMcc;
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  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3008
3009  if (getTargetMachine().Options.UnsafeFPMath &&
3010      (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3011       CC == ISD::SETNE || CC == ISD::SETUNE)) {
3012    SDValue Result = OptimizeVFPBrcond(Op, DAG);
3013    if (Result.getNode())
3014      return Result;
3015  }
3016
3017  ARMCC::CondCodes CondCode, CondCode2;
3018  FPCCToARMCC(CC, CondCode, CondCode2);
3019
3020  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3021  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3022  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3023  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3024  SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3025  SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3026  if (CondCode2 != ARMCC::AL) {
3027    ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3028    SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3029    Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3030  }
3031  return Res;
3032}
3033
3034SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3035  SDValue Chain = Op.getOperand(0);
3036  SDValue Table = Op.getOperand(1);
3037  SDValue Index = Op.getOperand(2);
3038  DebugLoc dl = Op.getDebugLoc();
3039
3040  EVT PTy = getPointerTy();
3041  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3042  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3043  SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3044  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3045  Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3046  Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3047  SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3048  if (Subtarget->isThumb2()) {
3049    // Thumb2 uses a two-level jump. That is, it jumps into the jump table
3050    // which does another jump to the destination. This also makes it easier
3051    // to translate it to TBB / TBH later.
3052    // FIXME: This might not work if the function is extremely large.
3053    return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3054                       Addr, Op.getOperand(2), JTI, UId);
3055  }
3056  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3057    Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3058                       MachinePointerInfo::getJumpTable(),
3059                       false, false, false, 0);
3060    Chain = Addr.getValue(1);
3061    Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3062    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3063  } else {
3064    Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3065                       MachinePointerInfo::getJumpTable(),
3066                       false, false, false, 0);
3067    Chain = Addr.getValue(1);
3068    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3069  }
3070}
3071
3072static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3073  EVT VT = Op.getValueType();
3074  DebugLoc dl = Op.getDebugLoc();
3075
3076  if (Op.getValueType().getVectorElementType() == MVT::i32) {
3077    if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3078      return Op;
3079    return DAG.UnrollVectorOp(Op.getNode());
3080  }
3081
3082  assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3083         "Invalid type for custom lowering!");
3084  if (VT != MVT::v4i16)
3085    return DAG.UnrollVectorOp(Op.getNode());
3086
3087  Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3088  return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3089}
3090
3091static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3092  EVT VT = Op.getValueType();
3093  if (VT.isVector())
3094    return LowerVectorFP_TO_INT(Op, DAG);
3095
3096  DebugLoc dl = Op.getDebugLoc();
3097  unsigned Opc;
3098
3099  switch (Op.getOpcode()) {
3100  default: llvm_unreachable("Invalid opcode!");
3101  case ISD::FP_TO_SINT:
3102    Opc = ARMISD::FTOSI;
3103    break;
3104  case ISD::FP_TO_UINT:
3105    Opc = ARMISD::FTOUI;
3106    break;
3107  }
3108  Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3109  return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3110}
3111
3112static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3113  EVT VT = Op.getValueType();
3114  DebugLoc dl = Op.getDebugLoc();
3115
3116  if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3117    if (VT.getVectorElementType() == MVT::f32)
3118      return Op;
3119    return DAG.UnrollVectorOp(Op.getNode());
3120  }
3121
3122  assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3123         "Invalid type for custom lowering!");
3124  if (VT != MVT::v4f32)
3125    return DAG.UnrollVectorOp(Op.getNode());
3126
3127  unsigned CastOpc;
3128  unsigned Opc;
3129  switch (Op.getOpcode()) {
3130  default: llvm_unreachable("Invalid opcode!");
3131  case ISD::SINT_TO_FP:
3132    CastOpc = ISD::SIGN_EXTEND;
3133    Opc = ISD::SINT_TO_FP;
3134    break;
3135  case ISD::UINT_TO_FP:
3136    CastOpc = ISD::ZERO_EXTEND;
3137    Opc = ISD::UINT_TO_FP;
3138    break;
3139  }
3140
3141  Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3142  return DAG.getNode(Opc, dl, VT, Op);
3143}
3144
3145static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3146  EVT VT = Op.getValueType();
3147  if (VT.isVector())
3148    return LowerVectorINT_TO_FP(Op, DAG);
3149
3150  DebugLoc dl = Op.getDebugLoc();
3151  unsigned Opc;
3152
3153  switch (Op.getOpcode()) {
3154  default: llvm_unreachable("Invalid opcode!");
3155  case ISD::SINT_TO_FP:
3156    Opc = ARMISD::SITOF;
3157    break;
3158  case ISD::UINT_TO_FP:
3159    Opc = ARMISD::UITOF;
3160    break;
3161  }
3162
3163  Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3164  return DAG.getNode(Opc, dl, VT, Op);
3165}
3166
3167SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3168  // Implement fcopysign with a fabs and a conditional fneg.
3169  SDValue Tmp0 = Op.getOperand(0);
3170  SDValue Tmp1 = Op.getOperand(1);
3171  DebugLoc dl = Op.getDebugLoc();
3172  EVT VT = Op.getValueType();
3173  EVT SrcVT = Tmp1.getValueType();
3174  bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3175    Tmp0.getOpcode() == ARMISD::VMOVDRR;
3176  bool UseNEON = !InGPR && Subtarget->hasNEON();
3177
3178  if (UseNEON) {
3179    // Use VBSL to copy the sign bit.
3180    unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3181    SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3182                               DAG.getTargetConstant(EncodedVal, MVT::i32));
3183    EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3184    if (VT == MVT::f64)
3185      Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3186                         DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3187                         DAG.getConstant(32, MVT::i32));
3188    else /*if (VT == MVT::f32)*/
3189      Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3190    if (SrcVT == MVT::f32) {
3191      Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3192      if (VT == MVT::f64)
3193        Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3194                           DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3195                           DAG.getConstant(32, MVT::i32));
3196    } else if (VT == MVT::f32)
3197      Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3198                         DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3199                         DAG.getConstant(32, MVT::i32));
3200    Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3201    Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3202
3203    SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3204                                            MVT::i32);
3205    AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3206    SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3207                                  DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3208
3209    SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3210                              DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3211                              DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3212    if (VT == MVT::f32) {
3213      Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3214      Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3215                        DAG.getConstant(0, MVT::i32));
3216    } else {
3217      Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3218    }
3219
3220    return Res;
3221  }
3222
3223  // Bitcast operand 1 to i32.
3224  if (SrcVT == MVT::f64)
3225    Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3226                       &Tmp1, 1).getValue(1);
3227  Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3228
3229  // Or in the signbit with integer operations.
3230  SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3231  SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3232  Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3233  if (VT == MVT::f32) {
3234    Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3235                       DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3236    return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3237                       DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3238  }
3239
3240  // f64: Or the high part with signbit and then combine two parts.
3241  Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3242                     &Tmp0, 1);
3243  SDValue Lo = Tmp0.getValue(0);
3244  SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3245  Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3246  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3247}
3248
3249SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3250  MachineFunction &MF = DAG.getMachineFunction();
3251  MachineFrameInfo *MFI = MF.getFrameInfo();
3252  MFI->setReturnAddressIsTaken(true);
3253
3254  EVT VT = Op.getValueType();
3255  DebugLoc dl = Op.getDebugLoc();
3256  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3257  if (Depth) {
3258    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3259    SDValue Offset = DAG.getConstant(4, MVT::i32);
3260    return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3261                       DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3262                       MachinePointerInfo(), false, false, false, 0);
3263  }
3264
3265  // Return LR, which contains the return address. Mark it an implicit live-in.
3266  unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3267  return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3268}
3269
3270SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3271  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3272  MFI->setFrameAddressIsTaken(true);
3273
3274  EVT VT = Op.getValueType();
3275  DebugLoc dl = Op.getDebugLoc();  // FIXME probably not meaningful
3276  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3277  unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3278    ? ARM::R7 : ARM::R11;
3279  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3280  while (Depth--)
3281    FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3282                            MachinePointerInfo(),
3283                            false, false, false, 0);
3284  return FrameAddr;
3285}
3286
3287/// ExpandBITCAST - If the target supports VFP, this function is called to
3288/// expand a bit convert where either the source or destination type is i64 to
3289/// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
3290/// operand type is illegal (e.g., v2f32 for a target that doesn't support
3291/// vectors), since the legalizer won't know what to do with that.
3292static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3293  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3294  DebugLoc dl = N->getDebugLoc();
3295  SDValue Op = N->getOperand(0);
3296
3297  // This function is only supposed to be called for i64 types, either as the
3298  // source or destination of the bit convert.
3299  EVT SrcVT = Op.getValueType();
3300  EVT DstVT = N->getValueType(0);
3301  assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3302         "ExpandBITCAST called for non-i64 type");
3303
3304  // Turn i64->f64 into VMOVDRR.
3305  if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3306    SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3307                             DAG.getConstant(0, MVT::i32));
3308    SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3309                             DAG.getConstant(1, MVT::i32));
3310    return DAG.getNode(ISD::BITCAST, dl, DstVT,
3311                       DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3312  }
3313
3314  // Turn f64->i64 into VMOVRRD.
3315  if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3316    SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3317                              DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3318    // Merge the pieces into a single i64 value.
3319    return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3320  }
3321
3322  return SDValue();
3323}
3324
3325/// getZeroVector - Returns a vector of specified type with all zero elements.
3326/// Zero vectors are used to represent vector negation and in those cases
3327/// will be implemented with the NEON VNEG instruction.  However, VNEG does
3328/// not support i64 elements, so sometimes the zero vectors will need to be
3329/// explicitly constructed.  Regardless, use a canonical VMOV to create the
3330/// zero vector.
3331static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3332  assert(VT.isVector() && "Expected a vector type");
3333  // The canonical modified immediate encoding of a zero vector is....0!
3334  SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3335  EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3336  SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3337  return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3338}
3339
3340/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3341/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3342SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3343                                                SelectionDAG &DAG) const {
3344  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3345  EVT VT = Op.getValueType();
3346  unsigned VTBits = VT.getSizeInBits();
3347  DebugLoc dl = Op.getDebugLoc();
3348  SDValue ShOpLo = Op.getOperand(0);
3349  SDValue ShOpHi = Op.getOperand(1);
3350  SDValue ShAmt  = Op.getOperand(2);
3351  SDValue ARMcc;
3352  unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3353
3354  assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3355
3356  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3357                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3358  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3359  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3360                                   DAG.getConstant(VTBits, MVT::i32));
3361  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3362  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3363  SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3364
3365  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3366  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3367                          ARMcc, DAG, dl);
3368  SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3369  SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3370                           CCR, Cmp);
3371
3372  SDValue Ops[2] = { Lo, Hi };
3373  return DAG.getMergeValues(Ops, 2, dl);
3374}
3375
3376/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3377/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3378SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3379                                               SelectionDAG &DAG) const {
3380  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3381  EVT VT = Op.getValueType();
3382  unsigned VTBits = VT.getSizeInBits();
3383  DebugLoc dl = Op.getDebugLoc();
3384  SDValue ShOpLo = Op.getOperand(0);
3385  SDValue ShOpHi = Op.getOperand(1);
3386  SDValue ShAmt  = Op.getOperand(2);
3387  SDValue ARMcc;
3388
3389  assert(Op.getOpcode() == ISD::SHL_PARTS);
3390  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3391                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3392  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3393  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3394                                   DAG.getConstant(VTBits, MVT::i32));
3395  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3396  SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3397
3398  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3399  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3400  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3401                          ARMcc, DAG, dl);
3402  SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3403  SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3404                           CCR, Cmp);
3405
3406  SDValue Ops[2] = { Lo, Hi };
3407  return DAG.getMergeValues(Ops, 2, dl);
3408}
3409
3410SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3411                                            SelectionDAG &DAG) const {
3412  // The rounding mode is in bits 23:22 of the FPSCR.
3413  // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3414  // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3415  // so that the shift + and get folded into a bitfield extract.
3416  DebugLoc dl = Op.getDebugLoc();
3417  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3418                              DAG.getConstant(Intrinsic::arm_get_fpscr,
3419                                              MVT::i32));
3420  SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3421                                  DAG.getConstant(1U << 22, MVT::i32));
3422  SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3423                              DAG.getConstant(22, MVT::i32));
3424  return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3425                     DAG.getConstant(3, MVT::i32));
3426}
3427
3428static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3429                         const ARMSubtarget *ST) {
3430  EVT VT = N->getValueType(0);
3431  DebugLoc dl = N->getDebugLoc();
3432
3433  if (!ST->hasV6T2Ops())
3434    return SDValue();
3435
3436  SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3437  return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3438}
3439
3440static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3441                          const ARMSubtarget *ST) {
3442  EVT VT = N->getValueType(0);
3443  DebugLoc dl = N->getDebugLoc();
3444
3445  if (!VT.isVector())
3446    return SDValue();
3447
3448  // Lower vector shifts on NEON to use VSHL.
3449  assert(ST->hasNEON() && "unexpected vector shift");
3450
3451  // Left shifts translate directly to the vshiftu intrinsic.
3452  if (N->getOpcode() == ISD::SHL)
3453    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3454                       DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3455                       N->getOperand(0), N->getOperand(1));
3456
3457  assert((N->getOpcode() == ISD::SRA ||
3458          N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3459
3460  // NEON uses the same intrinsics for both left and right shifts.  For
3461  // right shifts, the shift amounts are negative, so negate the vector of
3462  // shift amounts.
3463  EVT ShiftVT = N->getOperand(1).getValueType();
3464  SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3465                                     getZeroVector(ShiftVT, DAG, dl),
3466                                     N->getOperand(1));
3467  Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3468                             Intrinsic::arm_neon_vshifts :
3469                             Intrinsic::arm_neon_vshiftu);
3470  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3471                     DAG.getConstant(vshiftInt, MVT::i32),
3472                     N->getOperand(0), NegatedCount);
3473}
3474
3475static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3476                                const ARMSubtarget *ST) {
3477  EVT VT = N->getValueType(0);
3478  DebugLoc dl = N->getDebugLoc();
3479
3480  // We can get here for a node like i32 = ISD::SHL i32, i64
3481  if (VT != MVT::i64)
3482    return SDValue();
3483
3484  assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3485         "Unknown shift to lower!");
3486
3487  // We only lower SRA, SRL of 1 here, all others use generic lowering.
3488  if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3489      cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3490    return SDValue();
3491
3492  // If we are in thumb mode, we don't have RRX.
3493  if (ST->isThumb1Only()) return SDValue();
3494
3495  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
3496  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3497                           DAG.getConstant(0, MVT::i32));
3498  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3499                           DAG.getConstant(1, MVT::i32));
3500
3501  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3502  // captures the result into a carry flag.
3503  unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3504  Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3505
3506  // The low part is an ARMISD::RRX operand, which shifts the carry in.
3507  Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3508
3509  // Merge the pieces into a single i64 value.
3510 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3511}
3512
3513static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3514  SDValue TmpOp0, TmpOp1;
3515  bool Invert = false;
3516  bool Swap = false;
3517  unsigned Opc = 0;
3518
3519  SDValue Op0 = Op.getOperand(0);
3520  SDValue Op1 = Op.getOperand(1);
3521  SDValue CC = Op.getOperand(2);
3522  EVT VT = Op.getValueType();
3523  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3524  DebugLoc dl = Op.getDebugLoc();
3525
3526  if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3527    switch (SetCCOpcode) {
3528    default: llvm_unreachable("Illegal FP comparison");
3529    case ISD::SETUNE:
3530    case ISD::SETNE:  Invert = true; // Fallthrough
3531    case ISD::SETOEQ:
3532    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3533    case ISD::SETOLT:
3534    case ISD::SETLT: Swap = true; // Fallthrough
3535    case ISD::SETOGT:
3536    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3537    case ISD::SETOLE:
3538    case ISD::SETLE:  Swap = true; // Fallthrough
3539    case ISD::SETOGE:
3540    case ISD::SETGE: Opc = ARMISD::VCGE; break;
3541    case ISD::SETUGE: Swap = true; // Fallthrough
3542    case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3543    case ISD::SETUGT: Swap = true; // Fallthrough
3544    case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3545    case ISD::SETUEQ: Invert = true; // Fallthrough
3546    case ISD::SETONE:
3547      // Expand this to (OLT | OGT).
3548      TmpOp0 = Op0;
3549      TmpOp1 = Op1;
3550      Opc = ISD::OR;
3551      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3552      Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3553      break;
3554    case ISD::SETUO: Invert = true; // Fallthrough
3555    case ISD::SETO:
3556      // Expand this to (OLT | OGE).
3557      TmpOp0 = Op0;
3558      TmpOp1 = Op1;
3559      Opc = ISD::OR;
3560      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3561      Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3562      break;
3563    }
3564  } else {
3565    // Integer comparisons.
3566    switch (SetCCOpcode) {
3567    default: llvm_unreachable("Illegal integer comparison");
3568    case ISD::SETNE:  Invert = true;
3569    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3570    case ISD::SETLT:  Swap = true;
3571    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3572    case ISD::SETLE:  Swap = true;
3573    case ISD::SETGE:  Opc = ARMISD::VCGE; break;
3574    case ISD::SETULT: Swap = true;
3575    case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3576    case ISD::SETULE: Swap = true;
3577    case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3578    }
3579
3580    // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3581    if (Opc == ARMISD::VCEQ) {
3582
3583      SDValue AndOp;
3584      if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3585        AndOp = Op0;
3586      else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3587        AndOp = Op1;
3588
3589      // Ignore bitconvert.
3590      if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3591        AndOp = AndOp.getOperand(0);
3592
3593      if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3594        Opc = ARMISD::VTST;
3595        Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3596        Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3597        Invert = !Invert;
3598      }
3599    }
3600  }
3601
3602  if (Swap)
3603    std::swap(Op0, Op1);
3604
3605  // If one of the operands is a constant vector zero, attempt to fold the
3606  // comparison to a specialized compare-against-zero form.
3607  SDValue SingleOp;
3608  if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3609    SingleOp = Op0;
3610  else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3611    if (Opc == ARMISD::VCGE)
3612      Opc = ARMISD::VCLEZ;
3613    else if (Opc == ARMISD::VCGT)
3614      Opc = ARMISD::VCLTZ;
3615    SingleOp = Op1;
3616  }
3617
3618  SDValue Result;
3619  if (SingleOp.getNode()) {
3620    switch (Opc) {
3621    case ARMISD::VCEQ:
3622      Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3623    case ARMISD::VCGE:
3624      Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3625    case ARMISD::VCLEZ:
3626      Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3627    case ARMISD::VCGT:
3628      Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3629    case ARMISD::VCLTZ:
3630      Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3631    default:
3632      Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3633    }
3634  } else {
3635     Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3636  }
3637
3638  if (Invert)
3639    Result = DAG.getNOT(dl, Result, VT);
3640
3641  return Result;
3642}
3643
3644/// isNEONModifiedImm - Check if the specified splat value corresponds to a
3645/// valid vector constant for a NEON instruction with a "modified immediate"
3646/// operand (e.g., VMOV).  If so, return the encoded value.
3647static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3648                                 unsigned SplatBitSize, SelectionDAG &DAG,
3649                                 EVT &VT, bool is128Bits, NEONModImmType type) {
3650  unsigned OpCmode, Imm;
3651
3652  // SplatBitSize is set to the smallest size that splats the vector, so a
3653  // zero vector will always have SplatBitSize == 8.  However, NEON modified
3654  // immediate instructions others than VMOV do not support the 8-bit encoding
3655  // of a zero vector, and the default encoding of zero is supposed to be the
3656  // 32-bit version.
3657  if (SplatBits == 0)
3658    SplatBitSize = 32;
3659
3660  switch (SplatBitSize) {
3661  case 8:
3662    if (type != VMOVModImm)
3663      return SDValue();
3664    // Any 1-byte value is OK.  Op=0, Cmode=1110.
3665    assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3666    OpCmode = 0xe;
3667    Imm = SplatBits;
3668    VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3669    break;
3670
3671  case 16:
3672    // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3673    VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3674    if ((SplatBits & ~0xff) == 0) {
3675      // Value = 0x00nn: Op=x, Cmode=100x.
3676      OpCmode = 0x8;
3677      Imm = SplatBits;
3678      break;
3679    }
3680    if ((SplatBits & ~0xff00) == 0) {
3681      // Value = 0xnn00: Op=x, Cmode=101x.
3682      OpCmode = 0xa;
3683      Imm = SplatBits >> 8;
3684      break;
3685    }
3686    return SDValue();
3687
3688  case 32:
3689    // NEON's 32-bit VMOV supports splat values where:
3690    // * only one byte is nonzero, or
3691    // * the least significant byte is 0xff and the second byte is nonzero, or
3692    // * the least significant 2 bytes are 0xff and the third is nonzero.
3693    VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3694    if ((SplatBits & ~0xff) == 0) {
3695      // Value = 0x000000nn: Op=x, Cmode=000x.
3696      OpCmode = 0;
3697      Imm = SplatBits;
3698      break;
3699    }
3700    if ((SplatBits & ~0xff00) == 0) {
3701      // Value = 0x0000nn00: Op=x, Cmode=001x.
3702      OpCmode = 0x2;
3703      Imm = SplatBits >> 8;
3704      break;
3705    }
3706    if ((SplatBits & ~0xff0000) == 0) {
3707      // Value = 0x00nn0000: Op=x, Cmode=010x.
3708      OpCmode = 0x4;
3709      Imm = SplatBits >> 16;
3710      break;
3711    }
3712    if ((SplatBits & ~0xff000000) == 0) {
3713      // Value = 0xnn000000: Op=x, Cmode=011x.
3714      OpCmode = 0x6;
3715      Imm = SplatBits >> 24;
3716      break;
3717    }
3718
3719    // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3720    if (type == OtherModImm) return SDValue();
3721
3722    if ((SplatBits & ~0xffff) == 0 &&
3723        ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3724      // Value = 0x0000nnff: Op=x, Cmode=1100.
3725      OpCmode = 0xc;
3726      Imm = SplatBits >> 8;
3727      SplatBits |= 0xff;
3728      break;
3729    }
3730
3731    if ((SplatBits & ~0xffffff) == 0 &&
3732        ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3733      // Value = 0x00nnffff: Op=x, Cmode=1101.
3734      OpCmode = 0xd;
3735      Imm = SplatBits >> 16;
3736      SplatBits |= 0xffff;
3737      break;
3738    }
3739
3740    // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3741    // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3742    // VMOV.I32.  A (very) minor optimization would be to replicate the value
3743    // and fall through here to test for a valid 64-bit splat.  But, then the
3744    // caller would also need to check and handle the change in size.
3745    return SDValue();
3746
3747  case 64: {
3748    if (type != VMOVModImm)
3749      return SDValue();
3750    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3751    uint64_t BitMask = 0xff;
3752    uint64_t Val = 0;
3753    unsigned ImmMask = 1;
3754    Imm = 0;
3755    for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3756      if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3757        Val |= BitMask;
3758        Imm |= ImmMask;
3759      } else if ((SplatBits & BitMask) != 0) {
3760        return SDValue();
3761      }
3762      BitMask <<= 8;
3763      ImmMask <<= 1;
3764    }
3765    // Op=1, Cmode=1110.
3766    OpCmode = 0x1e;
3767    SplatBits = Val;
3768    VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3769    break;
3770  }
3771
3772  default:
3773    llvm_unreachable("unexpected size for isNEONModifiedImm");
3774  }
3775
3776  unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3777  return DAG.getTargetConstant(EncodedVal, MVT::i32);
3778}
3779
3780static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3781                       bool &ReverseVEXT, unsigned &Imm) {
3782  unsigned NumElts = VT.getVectorNumElements();
3783  ReverseVEXT = false;
3784
3785  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
3786  if (M[0] < 0)
3787    return false;
3788
3789  Imm = M[0];
3790
3791  // If this is a VEXT shuffle, the immediate value is the index of the first
3792  // element.  The other shuffle indices must be the successive elements after
3793  // the first one.
3794  unsigned ExpectedElt = Imm;
3795  for (unsigned i = 1; i < NumElts; ++i) {
3796    // Increment the expected index.  If it wraps around, it may still be
3797    // a VEXT but the source vectors must be swapped.
3798    ExpectedElt += 1;
3799    if (ExpectedElt == NumElts * 2) {
3800      ExpectedElt = 0;
3801      ReverseVEXT = true;
3802    }
3803
3804    if (M[i] < 0) continue; // ignore UNDEF indices
3805    if (ExpectedElt != static_cast<unsigned>(M[i]))
3806      return false;
3807  }
3808
3809  // Adjust the index value if the source operands will be swapped.
3810  if (ReverseVEXT)
3811    Imm -= NumElts;
3812
3813  return true;
3814}
3815
3816/// isVREVMask - Check if a vector shuffle corresponds to a VREV
3817/// instruction with the specified blocksize.  (The order of the elements
3818/// within each block of the vector is reversed.)
3819static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3820  assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3821         "Only possible block sizes for VREV are: 16, 32, 64");
3822
3823  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3824  if (EltSz == 64)
3825    return false;
3826
3827  unsigned NumElts = VT.getVectorNumElements();
3828  unsigned BlockElts = M[0] + 1;
3829  // If the first shuffle index is UNDEF, be optimistic.
3830  if (M[0] < 0)
3831    BlockElts = BlockSize / EltSz;
3832
3833  if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3834    return false;
3835
3836  for (unsigned i = 0; i < NumElts; ++i) {
3837    if (M[i] < 0) continue; // ignore UNDEF indices
3838    if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3839      return false;
3840  }
3841
3842  return true;
3843}
3844
3845static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3846  // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3847  // range, then 0 is placed into the resulting vector. So pretty much any mask
3848  // of 8 elements can work here.
3849  return VT == MVT::v8i8 && M.size() == 8;
3850}
3851
3852static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3853  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3854  if (EltSz == 64)
3855    return false;
3856
3857  unsigned NumElts = VT.getVectorNumElements();
3858  WhichResult = (M[0] == 0 ? 0 : 1);
3859  for (unsigned i = 0; i < NumElts; i += 2) {
3860    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3861        (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3862      return false;
3863  }
3864  return true;
3865}
3866
3867/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3868/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3869/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3870static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3871  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3872  if (EltSz == 64)
3873    return false;
3874
3875  unsigned NumElts = VT.getVectorNumElements();
3876  WhichResult = (M[0] == 0 ? 0 : 1);
3877  for (unsigned i = 0; i < NumElts; i += 2) {
3878    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3879        (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3880      return false;
3881  }
3882  return true;
3883}
3884
3885static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3886  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3887  if (EltSz == 64)
3888    return false;
3889
3890  unsigned NumElts = VT.getVectorNumElements();
3891  WhichResult = (M[0] == 0 ? 0 : 1);
3892  for (unsigned i = 0; i != NumElts; ++i) {
3893    if (M[i] < 0) continue; // ignore UNDEF indices
3894    if ((unsigned) M[i] != 2 * i + WhichResult)
3895      return false;
3896  }
3897
3898  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3899  if (VT.is64BitVector() && EltSz == 32)
3900    return false;
3901
3902  return true;
3903}
3904
3905/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3906/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3907/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3908static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3909  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3910  if (EltSz == 64)
3911    return false;
3912
3913  unsigned Half = VT.getVectorNumElements() / 2;
3914  WhichResult = (M[0] == 0 ? 0 : 1);
3915  for (unsigned j = 0; j != 2; ++j) {
3916    unsigned Idx = WhichResult;
3917    for (unsigned i = 0; i != Half; ++i) {
3918      int MIdx = M[i + j * Half];
3919      if (MIdx >= 0 && (unsigned) MIdx != Idx)
3920        return false;
3921      Idx += 2;
3922    }
3923  }
3924
3925  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3926  if (VT.is64BitVector() && EltSz == 32)
3927    return false;
3928
3929  return true;
3930}
3931
3932static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3933  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3934  if (EltSz == 64)
3935    return false;
3936
3937  unsigned NumElts = VT.getVectorNumElements();
3938  WhichResult = (M[0] == 0 ? 0 : 1);
3939  unsigned Idx = WhichResult * NumElts / 2;
3940  for (unsigned i = 0; i != NumElts; i += 2) {
3941    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3942        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3943      return false;
3944    Idx += 1;
3945  }
3946
3947  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3948  if (VT.is64BitVector() && EltSz == 32)
3949    return false;
3950
3951  return true;
3952}
3953
3954/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3955/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3956/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3957static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3958  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3959  if (EltSz == 64)
3960    return false;
3961
3962  unsigned NumElts = VT.getVectorNumElements();
3963  WhichResult = (M[0] == 0 ? 0 : 1);
3964  unsigned Idx = WhichResult * NumElts / 2;
3965  for (unsigned i = 0; i != NumElts; i += 2) {
3966    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3967        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3968      return false;
3969    Idx += 1;
3970  }
3971
3972  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3973  if (VT.is64BitVector() && EltSz == 32)
3974    return false;
3975
3976  return true;
3977}
3978
3979// If N is an integer constant that can be moved into a register in one
3980// instruction, return an SDValue of such a constant (will become a MOV
3981// instruction).  Otherwise return null.
3982static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3983                                     const ARMSubtarget *ST, DebugLoc dl) {
3984  uint64_t Val;
3985  if (!isa<ConstantSDNode>(N))
3986    return SDValue();
3987  Val = cast<ConstantSDNode>(N)->getZExtValue();
3988
3989  if (ST->isThumb1Only()) {
3990    if (Val <= 255 || ~Val <= 255)
3991      return DAG.getConstant(Val, MVT::i32);
3992  } else {
3993    if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3994      return DAG.getConstant(Val, MVT::i32);
3995  }
3996  return SDValue();
3997}
3998
3999// If this is a case we can't handle, return null and let the default
4000// expansion code take care of it.
4001SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4002                                             const ARMSubtarget *ST) const {
4003  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4004  DebugLoc dl = Op.getDebugLoc();
4005  EVT VT = Op.getValueType();
4006
4007  APInt SplatBits, SplatUndef;
4008  unsigned SplatBitSize;
4009  bool HasAnyUndefs;
4010  if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4011    if (SplatBitSize <= 64) {
4012      // Check if an immediate VMOV works.
4013      EVT VmovVT;
4014      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4015                                      SplatUndef.getZExtValue(), SplatBitSize,
4016                                      DAG, VmovVT, VT.is128BitVector(),
4017                                      VMOVModImm);
4018      if (Val.getNode()) {
4019        SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4020        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4021      }
4022
4023      // Try an immediate VMVN.
4024      uint64_t NegatedImm = (~SplatBits).getZExtValue();
4025      Val = isNEONModifiedImm(NegatedImm,
4026                                      SplatUndef.getZExtValue(), SplatBitSize,
4027                                      DAG, VmovVT, VT.is128BitVector(),
4028                                      VMVNModImm);
4029      if (Val.getNode()) {
4030        SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4031        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4032      }
4033
4034      // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4035      if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4036        int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4037        if (ImmVal != -1) {
4038          SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4039          return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4040        }
4041      }
4042    }
4043  }
4044
4045  // Scan through the operands to see if only one value is used.
4046  unsigned NumElts = VT.getVectorNumElements();
4047  bool isOnlyLowElement = true;
4048  bool usesOnlyOneValue = true;
4049  bool isConstant = true;
4050  SDValue Value;
4051  for (unsigned i = 0; i < NumElts; ++i) {
4052    SDValue V = Op.getOperand(i);
4053    if (V.getOpcode() == ISD::UNDEF)
4054      continue;
4055    if (i > 0)
4056      isOnlyLowElement = false;
4057    if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4058      isConstant = false;
4059
4060    if (!Value.getNode())
4061      Value = V;
4062    else if (V != Value)
4063      usesOnlyOneValue = false;
4064  }
4065
4066  if (!Value.getNode())
4067    return DAG.getUNDEF(VT);
4068
4069  if (isOnlyLowElement)
4070    return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4071
4072  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4073
4074  // Use VDUP for non-constant splats.  For f32 constant splats, reduce to
4075  // i32 and try again.
4076  if (usesOnlyOneValue && EltSize <= 32) {
4077    if (!isConstant)
4078      return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4079    if (VT.getVectorElementType().isFloatingPoint()) {
4080      SmallVector<SDValue, 8> Ops;
4081      for (unsigned i = 0; i < NumElts; ++i)
4082        Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4083                                  Op.getOperand(i)));
4084      EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4085      SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4086      Val = LowerBUILD_VECTOR(Val, DAG, ST);
4087      if (Val.getNode())
4088        return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4089    }
4090    SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4091    if (Val.getNode())
4092      return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4093  }
4094
4095  // If all elements are constants and the case above didn't get hit, fall back
4096  // to the default expansion, which will generate a load from the constant
4097  // pool.
4098  if (isConstant)
4099    return SDValue();
4100
4101  // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4102  if (NumElts >= 4) {
4103    SDValue shuffle = ReconstructShuffle(Op, DAG);
4104    if (shuffle != SDValue())
4105      return shuffle;
4106  }
4107
4108  // Vectors with 32- or 64-bit elements can be built by directly assigning
4109  // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
4110  // will be legalized.
4111  if (EltSize >= 32) {
4112    // Do the expansion with floating-point types, since that is what the VFP
4113    // registers are defined to use, and since i64 is not legal.
4114    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4115    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4116    SmallVector<SDValue, 8> Ops;
4117    for (unsigned i = 0; i < NumElts; ++i)
4118      Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4119    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4120    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4121  }
4122
4123  return SDValue();
4124}
4125
4126// Gather data to see if the operation can be modelled as a
4127// shuffle in combination with VEXTs.
4128SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4129                                              SelectionDAG &DAG) const {
4130  DebugLoc dl = Op.getDebugLoc();
4131  EVT VT = Op.getValueType();
4132  unsigned NumElts = VT.getVectorNumElements();
4133
4134  SmallVector<SDValue, 2> SourceVecs;
4135  SmallVector<unsigned, 2> MinElts;
4136  SmallVector<unsigned, 2> MaxElts;
4137
4138  for (unsigned i = 0; i < NumElts; ++i) {
4139    SDValue V = Op.getOperand(i);
4140    if (V.getOpcode() == ISD::UNDEF)
4141      continue;
4142    else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4143      // A shuffle can only come from building a vector from various
4144      // elements of other vectors.
4145      return SDValue();
4146    } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4147               VT.getVectorElementType()) {
4148      // This code doesn't know how to handle shuffles where the vector
4149      // element types do not match (this happens because type legalization
4150      // promotes the return type of EXTRACT_VECTOR_ELT).
4151      // FIXME: It might be appropriate to extend this code to handle
4152      // mismatched types.
4153      return SDValue();
4154    }
4155
4156    // Record this extraction against the appropriate vector if possible...
4157    SDValue SourceVec = V.getOperand(0);
4158    unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4159    bool FoundSource = false;
4160    for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4161      if (SourceVecs[j] == SourceVec) {
4162        if (MinElts[j] > EltNo)
4163          MinElts[j] = EltNo;
4164        if (MaxElts[j] < EltNo)
4165          MaxElts[j] = EltNo;
4166        FoundSource = true;
4167        break;
4168      }
4169    }
4170
4171    // Or record a new source if not...
4172    if (!FoundSource) {
4173      SourceVecs.push_back(SourceVec);
4174      MinElts.push_back(EltNo);
4175      MaxElts.push_back(EltNo);
4176    }
4177  }
4178
4179  // Currently only do something sane when at most two source vectors
4180  // involved.
4181  if (SourceVecs.size() > 2)
4182    return SDValue();
4183
4184  SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4185  int VEXTOffsets[2] = {0, 0};
4186
4187  // This loop extracts the usage patterns of the source vectors
4188  // and prepares appropriate SDValues for a shuffle if possible.
4189  for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4190    if (SourceVecs[i].getValueType() == VT) {
4191      // No VEXT necessary
4192      ShuffleSrcs[i] = SourceVecs[i];
4193      VEXTOffsets[i] = 0;
4194      continue;
4195    } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4196      // It probably isn't worth padding out a smaller vector just to
4197      // break it down again in a shuffle.
4198      return SDValue();
4199    }
4200
4201    // Since only 64-bit and 128-bit vectors are legal on ARM and
4202    // we've eliminated the other cases...
4203    assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4204           "unexpected vector sizes in ReconstructShuffle");
4205
4206    if (MaxElts[i] - MinElts[i] >= NumElts) {
4207      // Span too large for a VEXT to cope
4208      return SDValue();
4209    }
4210
4211    if (MinElts[i] >= NumElts) {
4212      // The extraction can just take the second half
4213      VEXTOffsets[i] = NumElts;
4214      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4215                                   SourceVecs[i],
4216                                   DAG.getIntPtrConstant(NumElts));
4217    } else if (MaxElts[i] < NumElts) {
4218      // The extraction can just take the first half
4219      VEXTOffsets[i] = 0;
4220      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4221                                   SourceVecs[i],
4222                                   DAG.getIntPtrConstant(0));
4223    } else {
4224      // An actual VEXT is needed
4225      VEXTOffsets[i] = MinElts[i];
4226      SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4227                                     SourceVecs[i],
4228                                     DAG.getIntPtrConstant(0));
4229      SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4230                                     SourceVecs[i],
4231                                     DAG.getIntPtrConstant(NumElts));
4232      ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4233                                   DAG.getConstant(VEXTOffsets[i], MVT::i32));
4234    }
4235  }
4236
4237  SmallVector<int, 8> Mask;
4238
4239  for (unsigned i = 0; i < NumElts; ++i) {
4240    SDValue Entry = Op.getOperand(i);
4241    if (Entry.getOpcode() == ISD::UNDEF) {
4242      Mask.push_back(-1);
4243      continue;
4244    }
4245
4246    SDValue ExtractVec = Entry.getOperand(0);
4247    int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4248                                          .getOperand(1))->getSExtValue();
4249    if (ExtractVec == SourceVecs[0]) {
4250      Mask.push_back(ExtractElt - VEXTOffsets[0]);
4251    } else {
4252      Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4253    }
4254  }
4255
4256  // Final check before we try to produce nonsense...
4257  if (isShuffleMaskLegal(Mask, VT))
4258    return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4259                                &Mask[0]);
4260
4261  return SDValue();
4262}
4263
4264/// isShuffleMaskLegal - Targets can use this to indicate that they only
4265/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4266/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4267/// are assumed to be legal.
4268bool
4269ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4270                                      EVT VT) const {
4271  if (VT.getVectorNumElements() == 4 &&
4272      (VT.is128BitVector() || VT.is64BitVector())) {
4273    unsigned PFIndexes[4];
4274    for (unsigned i = 0; i != 4; ++i) {
4275      if (M[i] < 0)
4276        PFIndexes[i] = 8;
4277      else
4278        PFIndexes[i] = M[i];
4279    }
4280
4281    // Compute the index in the perfect shuffle table.
4282    unsigned PFTableIndex =
4283      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4284    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4285    unsigned Cost = (PFEntry >> 30);
4286
4287    if (Cost <= 4)
4288      return true;
4289  }
4290
4291  bool ReverseVEXT;
4292  unsigned Imm, WhichResult;
4293
4294  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4295  return (EltSize >= 32 ||
4296          ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4297          isVREVMask(M, VT, 64) ||
4298          isVREVMask(M, VT, 32) ||
4299          isVREVMask(M, VT, 16) ||
4300          isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4301          isVTBLMask(M, VT) ||
4302          isVTRNMask(M, VT, WhichResult) ||
4303          isVUZPMask(M, VT, WhichResult) ||
4304          isVZIPMask(M, VT, WhichResult) ||
4305          isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4306          isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4307          isVZIP_v_undef_Mask(M, VT, WhichResult));
4308}
4309
4310/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4311/// the specified operations to build the shuffle.
4312static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4313                                      SDValue RHS, SelectionDAG &DAG,
4314                                      DebugLoc dl) {
4315  unsigned OpNum = (PFEntry >> 26) & 0x0F;
4316  unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4317  unsigned RHSID = (PFEntry >>  0) & ((1 << 13)-1);
4318
4319  enum {
4320    OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4321    OP_VREV,
4322    OP_VDUP0,
4323    OP_VDUP1,
4324    OP_VDUP2,
4325    OP_VDUP3,
4326    OP_VEXT1,
4327    OP_VEXT2,
4328    OP_VEXT3,
4329    OP_VUZPL, // VUZP, left result
4330    OP_VUZPR, // VUZP, right result
4331    OP_VZIPL, // VZIP, left result
4332    OP_VZIPR, // VZIP, right result
4333    OP_VTRNL, // VTRN, left result
4334    OP_VTRNR  // VTRN, right result
4335  };
4336
4337  if (OpNum == OP_COPY) {
4338    if (LHSID == (1*9+2)*9+3) return LHS;
4339    assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4340    return RHS;
4341  }
4342
4343  SDValue OpLHS, OpRHS;
4344  OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4345  OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4346  EVT VT = OpLHS.getValueType();
4347
4348  switch (OpNum) {
4349  default: llvm_unreachable("Unknown shuffle opcode!");
4350  case OP_VREV:
4351    // VREV divides the vector in half and swaps within the half.
4352    if (VT.getVectorElementType() == MVT::i32 ||
4353        VT.getVectorElementType() == MVT::f32)
4354      return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4355    // vrev <4 x i16> -> VREV32
4356    if (VT.getVectorElementType() == MVT::i16)
4357      return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4358    // vrev <4 x i8> -> VREV16
4359    assert(VT.getVectorElementType() == MVT::i8);
4360    return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4361  case OP_VDUP0:
4362  case OP_VDUP1:
4363  case OP_VDUP2:
4364  case OP_VDUP3:
4365    return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4366                       OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4367  case OP_VEXT1:
4368  case OP_VEXT2:
4369  case OP_VEXT3:
4370    return DAG.getNode(ARMISD::VEXT, dl, VT,
4371                       OpLHS, OpRHS,
4372                       DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4373  case OP_VUZPL:
4374  case OP_VUZPR:
4375    return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4376                       OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4377  case OP_VZIPL:
4378  case OP_VZIPR:
4379    return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4380                       OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4381  case OP_VTRNL:
4382  case OP_VTRNR:
4383    return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4384                       OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4385  }
4386}
4387
4388static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4389                                       ArrayRef<int> ShuffleMask,
4390                                       SelectionDAG &DAG) {
4391  // Check to see if we can use the VTBL instruction.
4392  SDValue V1 = Op.getOperand(0);
4393  SDValue V2 = Op.getOperand(1);
4394  DebugLoc DL = Op.getDebugLoc();
4395
4396  SmallVector<SDValue, 8> VTBLMask;
4397  for (ArrayRef<int>::iterator
4398         I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4399    VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4400
4401  if (V2.getNode()->getOpcode() == ISD::UNDEF)
4402    return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4403                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4404                                   &VTBLMask[0], 8));
4405
4406  return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4407                     DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4408                                 &VTBLMask[0], 8));
4409}
4410
4411static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4412  SDValue V1 = Op.getOperand(0);
4413  SDValue V2 = Op.getOperand(1);
4414  DebugLoc dl = Op.getDebugLoc();
4415  EVT VT = Op.getValueType();
4416  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4417
4418  // Convert shuffles that are directly supported on NEON to target-specific
4419  // DAG nodes, instead of keeping them as shuffles and matching them again
4420  // during code selection.  This is more efficient and avoids the possibility
4421  // of inconsistencies between legalization and selection.
4422  // FIXME: floating-point vectors should be canonicalized to integer vectors
4423  // of the same time so that they get CSEd properly.
4424  ArrayRef<int> ShuffleMask = SVN->getMask();
4425
4426  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4427  if (EltSize <= 32) {
4428    if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4429      int Lane = SVN->getSplatIndex();
4430      // If this is undef splat, generate it via "just" vdup, if possible.
4431      if (Lane == -1) Lane = 0;
4432
4433      // Test if V1 is a SCALAR_TO_VECTOR.
4434      if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4435        return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4436      }
4437      // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4438      // (and probably will turn into a SCALAR_TO_VECTOR once legalization
4439      // reaches it).
4440      if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4441          !isa<ConstantSDNode>(V1.getOperand(0))) {
4442        bool IsScalarToVector = true;
4443        for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4444          if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4445            IsScalarToVector = false;
4446            break;
4447          }
4448        if (IsScalarToVector)
4449          return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4450      }
4451      return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4452                         DAG.getConstant(Lane, MVT::i32));
4453    }
4454
4455    bool ReverseVEXT;
4456    unsigned Imm;
4457    if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4458      if (ReverseVEXT)
4459        std::swap(V1, V2);
4460      return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4461                         DAG.getConstant(Imm, MVT::i32));
4462    }
4463
4464    if (isVREVMask(ShuffleMask, VT, 64))
4465      return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4466    if (isVREVMask(ShuffleMask, VT, 32))
4467      return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4468    if (isVREVMask(ShuffleMask, VT, 16))
4469      return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4470
4471    // Check for Neon shuffles that modify both input vectors in place.
4472    // If both results are used, i.e., if there are two shuffles with the same
4473    // source operands and with masks corresponding to both results of one of
4474    // these operations, DAG memoization will ensure that a single node is
4475    // used for both shuffles.
4476    unsigned WhichResult;
4477    if (isVTRNMask(ShuffleMask, VT, WhichResult))
4478      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4479                         V1, V2).getValue(WhichResult);
4480    if (isVUZPMask(ShuffleMask, VT, WhichResult))
4481      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4482                         V1, V2).getValue(WhichResult);
4483    if (isVZIPMask(ShuffleMask, VT, WhichResult))
4484      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4485                         V1, V2).getValue(WhichResult);
4486
4487    if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4488      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4489                         V1, V1).getValue(WhichResult);
4490    if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4491      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4492                         V1, V1).getValue(WhichResult);
4493    if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4494      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4495                         V1, V1).getValue(WhichResult);
4496  }
4497
4498  // If the shuffle is not directly supported and it has 4 elements, use
4499  // the PerfectShuffle-generated table to synthesize it from other shuffles.
4500  unsigned NumElts = VT.getVectorNumElements();
4501  if (NumElts == 4) {
4502    unsigned PFIndexes[4];
4503    for (unsigned i = 0; i != 4; ++i) {
4504      if (ShuffleMask[i] < 0)
4505        PFIndexes[i] = 8;
4506      else
4507        PFIndexes[i] = ShuffleMask[i];
4508    }
4509
4510    // Compute the index in the perfect shuffle table.
4511    unsigned PFTableIndex =
4512      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4513    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4514    unsigned Cost = (PFEntry >> 30);
4515
4516    if (Cost <= 4)
4517      return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4518  }
4519
4520  // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4521  if (EltSize >= 32) {
4522    // Do the expansion with floating-point types, since that is what the VFP
4523    // registers are defined to use, and since i64 is not legal.
4524    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4525    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4526    V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4527    V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4528    SmallVector<SDValue, 8> Ops;
4529    for (unsigned i = 0; i < NumElts; ++i) {
4530      if (ShuffleMask[i] < 0)
4531        Ops.push_back(DAG.getUNDEF(EltVT));
4532      else
4533        Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4534                                  ShuffleMask[i] < (int)NumElts ? V1 : V2,
4535                                  DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4536                                                  MVT::i32)));
4537    }
4538    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4539    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4540  }
4541
4542  if (VT == MVT::v8i8) {
4543    SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4544    if (NewOp.getNode())
4545      return NewOp;
4546  }
4547
4548  return SDValue();
4549}
4550
4551static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4552  // INSERT_VECTOR_ELT is legal only for immediate indexes.
4553  SDValue Lane = Op.getOperand(2);
4554  if (!isa<ConstantSDNode>(Lane))
4555    return SDValue();
4556
4557  return Op;
4558}
4559
4560static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4561  // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4562  SDValue Lane = Op.getOperand(1);
4563  if (!isa<ConstantSDNode>(Lane))
4564    return SDValue();
4565
4566  SDValue Vec = Op.getOperand(0);
4567  if (Op.getValueType() == MVT::i32 &&
4568      Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4569    DebugLoc dl = Op.getDebugLoc();
4570    return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4571  }
4572
4573  return Op;
4574}
4575
4576static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4577  // The only time a CONCAT_VECTORS operation can have legal types is when
4578  // two 64-bit vectors are concatenated to a 128-bit vector.
4579  assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4580         "unexpected CONCAT_VECTORS");
4581  DebugLoc dl = Op.getDebugLoc();
4582  SDValue Val = DAG.getUNDEF(MVT::v2f64);
4583  SDValue Op0 = Op.getOperand(0);
4584  SDValue Op1 = Op.getOperand(1);
4585  if (Op0.getOpcode() != ISD::UNDEF)
4586    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4587                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4588                      DAG.getIntPtrConstant(0));
4589  if (Op1.getOpcode() != ISD::UNDEF)
4590    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4591                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4592                      DAG.getIntPtrConstant(1));
4593  return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4594}
4595
4596/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4597/// element has been zero/sign-extended, depending on the isSigned parameter,
4598/// from an integer type half its size.
4599static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4600                                   bool isSigned) {
4601  // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4602  EVT VT = N->getValueType(0);
4603  if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4604    SDNode *BVN = N->getOperand(0).getNode();
4605    if (BVN->getValueType(0) != MVT::v4i32 ||
4606        BVN->getOpcode() != ISD::BUILD_VECTOR)
4607      return false;
4608    unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4609    unsigned HiElt = 1 - LoElt;
4610    ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4611    ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4612    ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4613    ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4614    if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4615      return false;
4616    if (isSigned) {
4617      if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4618          Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4619        return true;
4620    } else {
4621      if (Hi0->isNullValue() && Hi1->isNullValue())
4622        return true;
4623    }
4624    return false;
4625  }
4626
4627  if (N->getOpcode() != ISD::BUILD_VECTOR)
4628    return false;
4629
4630  for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4631    SDNode *Elt = N->getOperand(i).getNode();
4632    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4633      unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4634      unsigned HalfSize = EltSize / 2;
4635      if (isSigned) {
4636        if (!isIntN(HalfSize, C->getSExtValue()))
4637          return false;
4638      } else {
4639        if (!isUIntN(HalfSize, C->getZExtValue()))
4640          return false;
4641      }
4642      continue;
4643    }
4644    return false;
4645  }
4646
4647  return true;
4648}
4649
4650/// isSignExtended - Check if a node is a vector value that is sign-extended
4651/// or a constant BUILD_VECTOR with sign-extended elements.
4652static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4653  if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4654    return true;
4655  if (isExtendedBUILD_VECTOR(N, DAG, true))
4656    return true;
4657  return false;
4658}
4659
4660/// isZeroExtended - Check if a node is a vector value that is zero-extended
4661/// or a constant BUILD_VECTOR with zero-extended elements.
4662static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4663  if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4664    return true;
4665  if (isExtendedBUILD_VECTOR(N, DAG, false))
4666    return true;
4667  return false;
4668}
4669
4670/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4671/// load, or BUILD_VECTOR with extended elements, return the unextended value.
4672static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4673  if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4674    return N->getOperand(0);
4675  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4676    return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4677                       LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4678                       LD->isNonTemporal(), LD->isInvariant(),
4679                       LD->getAlignment());
4680  // Otherwise, the value must be a BUILD_VECTOR.  For v2i64, it will
4681  // have been legalized as a BITCAST from v4i32.
4682  if (N->getOpcode() == ISD::BITCAST) {
4683    SDNode *BVN = N->getOperand(0).getNode();
4684    assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4685           BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4686    unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4687    return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4688                       BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4689  }
4690  // Construct a new BUILD_VECTOR with elements truncated to half the size.
4691  assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4692  EVT VT = N->getValueType(0);
4693  unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4694  unsigned NumElts = VT.getVectorNumElements();
4695  MVT TruncVT = MVT::getIntegerVT(EltSize);
4696  SmallVector<SDValue, 8> Ops;
4697  for (unsigned i = 0; i != NumElts; ++i) {
4698    ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4699    const APInt &CInt = C->getAPIntValue();
4700    Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4701  }
4702  return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4703                     MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4704}
4705
4706static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4707  unsigned Opcode = N->getOpcode();
4708  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4709    SDNode *N0 = N->getOperand(0).getNode();
4710    SDNode *N1 = N->getOperand(1).getNode();
4711    return N0->hasOneUse() && N1->hasOneUse() &&
4712      isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4713  }
4714  return false;
4715}
4716
4717static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4718  unsigned Opcode = N->getOpcode();
4719  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4720    SDNode *N0 = N->getOperand(0).getNode();
4721    SDNode *N1 = N->getOperand(1).getNode();
4722    return N0->hasOneUse() && N1->hasOneUse() &&
4723      isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4724  }
4725  return false;
4726}
4727
4728static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4729  // Multiplications are only custom-lowered for 128-bit vectors so that
4730  // VMULL can be detected.  Otherwise v2i64 multiplications are not legal.
4731  EVT VT = Op.getValueType();
4732  assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4733  SDNode *N0 = Op.getOperand(0).getNode();
4734  SDNode *N1 = Op.getOperand(1).getNode();
4735  unsigned NewOpc = 0;
4736  bool isMLA = false;
4737  bool isN0SExt = isSignExtended(N0, DAG);
4738  bool isN1SExt = isSignExtended(N1, DAG);
4739  if (isN0SExt && isN1SExt)
4740    NewOpc = ARMISD::VMULLs;
4741  else {
4742    bool isN0ZExt = isZeroExtended(N0, DAG);
4743    bool isN1ZExt = isZeroExtended(N1, DAG);
4744    if (isN0ZExt && isN1ZExt)
4745      NewOpc = ARMISD::VMULLu;
4746    else if (isN1SExt || isN1ZExt) {
4747      // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4748      // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4749      if (isN1SExt && isAddSubSExt(N0, DAG)) {
4750        NewOpc = ARMISD::VMULLs;
4751        isMLA = true;
4752      } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4753        NewOpc = ARMISD::VMULLu;
4754        isMLA = true;
4755      } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4756        std::swap(N0, N1);
4757        NewOpc = ARMISD::VMULLu;
4758        isMLA = true;
4759      }
4760    }
4761
4762    if (!NewOpc) {
4763      if (VT == MVT::v2i64)
4764        // Fall through to expand this.  It is not legal.
4765        return SDValue();
4766      else
4767        // Other vector multiplications are legal.
4768        return Op;
4769    }
4770  }
4771
4772  // Legalize to a VMULL instruction.
4773  DebugLoc DL = Op.getDebugLoc();
4774  SDValue Op0;
4775  SDValue Op1 = SkipExtension(N1, DAG);
4776  if (!isMLA) {
4777    Op0 = SkipExtension(N0, DAG);
4778    assert(Op0.getValueType().is64BitVector() &&
4779           Op1.getValueType().is64BitVector() &&
4780           "unexpected types for extended operands to VMULL");
4781    return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4782  }
4783
4784  // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4785  // isel lowering to take advantage of no-stall back to back vmul + vmla.
4786  //   vmull q0, d4, d6
4787  //   vmlal q0, d5, d6
4788  // is faster than
4789  //   vaddl q0, d4, d5
4790  //   vmovl q1, d6
4791  //   vmul  q0, q0, q1
4792  SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4793  SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4794  EVT Op1VT = Op1.getValueType();
4795  return DAG.getNode(N0->getOpcode(), DL, VT,
4796                     DAG.getNode(NewOpc, DL, VT,
4797                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4798                     DAG.getNode(NewOpc, DL, VT,
4799                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4800}
4801
4802static SDValue
4803LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4804  // Convert to float
4805  // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4806  // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4807  X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4808  Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4809  X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4810  Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4811  // Get reciprocal estimate.
4812  // float4 recip = vrecpeq_f32(yf);
4813  Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4814                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4815  // Because char has a smaller range than uchar, we can actually get away
4816  // without any newton steps.  This requires that we use a weird bias
4817  // of 0xb000, however (again, this has been exhaustively tested).
4818  // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4819  X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4820  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4821  Y = DAG.getConstant(0xb000, MVT::i32);
4822  Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4823  X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4824  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4825  // Convert back to short.
4826  X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4827  X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4828  return X;
4829}
4830
4831static SDValue
4832LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4833  SDValue N2;
4834  // Convert to float.
4835  // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4836  // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4837  N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4838  N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4839  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4840  N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4841
4842  // Use reciprocal estimate and one refinement step.
4843  // float4 recip = vrecpeq_f32(yf);
4844  // recip *= vrecpsq_f32(yf, recip);
4845  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4846                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4847  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4848                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4849                   N1, N2);
4850  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4851  // Because short has a smaller range than ushort, we can actually get away
4852  // with only a single newton step.  This requires that we use a weird bias
4853  // of 89, however (again, this has been exhaustively tested).
4854  // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4855  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4856  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4857  N1 = DAG.getConstant(0x89, MVT::i32);
4858  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4859  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4860  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4861  // Convert back to integer and return.
4862  // return vmovn_s32(vcvt_s32_f32(result));
4863  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4864  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4865  return N0;
4866}
4867
4868static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4869  EVT VT = Op.getValueType();
4870  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4871         "unexpected type for custom-lowering ISD::SDIV");
4872
4873  DebugLoc dl = Op.getDebugLoc();
4874  SDValue N0 = Op.getOperand(0);
4875  SDValue N1 = Op.getOperand(1);
4876  SDValue N2, N3;
4877
4878  if (VT == MVT::v8i8) {
4879    N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4880    N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4881
4882    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4883                     DAG.getIntPtrConstant(4));
4884    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4885                     DAG.getIntPtrConstant(4));
4886    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4887                     DAG.getIntPtrConstant(0));
4888    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4889                     DAG.getIntPtrConstant(0));
4890
4891    N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4892    N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4893
4894    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4895    N0 = LowerCONCAT_VECTORS(N0, DAG);
4896
4897    N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4898    return N0;
4899  }
4900  return LowerSDIV_v4i16(N0, N1, dl, DAG);
4901}
4902
4903static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4904  EVT VT = Op.getValueType();
4905  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4906         "unexpected type for custom-lowering ISD::UDIV");
4907
4908  DebugLoc dl = Op.getDebugLoc();
4909  SDValue N0 = Op.getOperand(0);
4910  SDValue N1 = Op.getOperand(1);
4911  SDValue N2, N3;
4912
4913  if (VT == MVT::v8i8) {
4914    N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4915    N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4916
4917    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4918                     DAG.getIntPtrConstant(4));
4919    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4920                     DAG.getIntPtrConstant(4));
4921    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4922                     DAG.getIntPtrConstant(0));
4923    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4924                     DAG.getIntPtrConstant(0));
4925
4926    N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4927    N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4928
4929    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4930    N0 = LowerCONCAT_VECTORS(N0, DAG);
4931
4932    N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4933                     DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4934                     N0);
4935    return N0;
4936  }
4937
4938  // v4i16 sdiv ... Convert to float.
4939  // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4940  // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4941  N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4942  N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4943  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4944  SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4945
4946  // Use reciprocal estimate and two refinement steps.
4947  // float4 recip = vrecpeq_f32(yf);
4948  // recip *= vrecpsq_f32(yf, recip);
4949  // recip *= vrecpsq_f32(yf, recip);
4950  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4951                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4952  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4953                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4954                   BN1, N2);
4955  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4956  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4957                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4958                   BN1, N2);
4959  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4960  // Simply multiplying by the reciprocal estimate can leave us a few ulps
4961  // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4962  // and that it will never cause us to return an answer too large).
4963  // float4 result = as_float4(as_int4(xf*recip) + 2);
4964  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4965  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4966  N1 = DAG.getConstant(2, MVT::i32);
4967  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4968  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4969  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4970  // Convert back to integer and return.
4971  // return vmovn_u32(vcvt_s32_f32(result));
4972  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4973  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4974  return N0;
4975}
4976
4977static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4978  EVT VT = Op.getNode()->getValueType(0);
4979  SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4980
4981  unsigned Opc;
4982  bool ExtraOp = false;
4983  switch (Op.getOpcode()) {
4984  default: llvm_unreachable("Invalid code");
4985  case ISD::ADDC: Opc = ARMISD::ADDC; break;
4986  case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4987  case ISD::SUBC: Opc = ARMISD::SUBC; break;
4988  case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4989  }
4990
4991  if (!ExtraOp)
4992    return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4993                       Op.getOperand(1));
4994  return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4995                     Op.getOperand(1), Op.getOperand(2));
4996}
4997
4998static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
4999  // Monotonic load/store is legal for all targets
5000  if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5001    return Op;
5002
5003  // Aquire/Release load/store is not legal for targets without a
5004  // dmb or equivalent available.
5005  return SDValue();
5006}
5007
5008
5009static void
5010ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5011                    SelectionDAG &DAG, unsigned NewOp) {
5012  DebugLoc dl = Node->getDebugLoc();
5013  assert (Node->getValueType(0) == MVT::i64 &&
5014          "Only know how to expand i64 atomics");
5015
5016  SmallVector<SDValue, 6> Ops;
5017  Ops.push_back(Node->getOperand(0)); // Chain
5018  Ops.push_back(Node->getOperand(1)); // Ptr
5019  // Low part of Val1
5020  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5021                            Node->getOperand(2), DAG.getIntPtrConstant(0)));
5022  // High part of Val1
5023  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5024                            Node->getOperand(2), DAG.getIntPtrConstant(1)));
5025  if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5026    // High part of Val1
5027    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5028                              Node->getOperand(3), DAG.getIntPtrConstant(0)));
5029    // High part of Val2
5030    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5031                              Node->getOperand(3), DAG.getIntPtrConstant(1)));
5032  }
5033  SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5034  SDValue Result =
5035    DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5036                            cast<MemSDNode>(Node)->getMemOperand());
5037  SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5038  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5039  Results.push_back(Result.getValue(2));
5040}
5041
5042SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5043  switch (Op.getOpcode()) {
5044  default: llvm_unreachable("Don't know how to custom lower this!");
5045  case ISD::ConstantPool:  return LowerConstantPool(Op, DAG);
5046  case ISD::BlockAddress:  return LowerBlockAddress(Op, DAG);
5047  case ISD::GlobalAddress:
5048    return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5049      LowerGlobalAddressELF(Op, DAG);
5050  case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5051  case ISD::SELECT:        return LowerSELECT(Op, DAG);
5052  case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG);
5053  case ISD::BR_CC:         return LowerBR_CC(Op, DAG);
5054  case ISD::BR_JT:         return LowerBR_JT(Op, DAG);
5055  case ISD::VASTART:       return LowerVASTART(Op, DAG);
5056  case ISD::MEMBARRIER:    return LowerMEMBARRIER(Op, DAG, Subtarget);
5057  case ISD::ATOMIC_FENCE:  return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5058  case ISD::PREFETCH:      return LowerPREFETCH(Op, DAG, Subtarget);
5059  case ISD::SINT_TO_FP:
5060  case ISD::UINT_TO_FP:    return LowerINT_TO_FP(Op, DAG);
5061  case ISD::FP_TO_SINT:
5062  case ISD::FP_TO_UINT:    return LowerFP_TO_INT(Op, DAG);
5063  case ISD::FCOPYSIGN:     return LowerFCOPYSIGN(Op, DAG);
5064  case ISD::RETURNADDR:    return LowerRETURNADDR(Op, DAG);
5065  case ISD::FRAMEADDR:     return LowerFRAMEADDR(Op, DAG);
5066  case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5067  case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5068  case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5069  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5070                                                               Subtarget);
5071  case ISD::BITCAST:       return ExpandBITCAST(Op.getNode(), DAG);
5072  case ISD::SHL:
5073  case ISD::SRL:
5074  case ISD::SRA:           return LowerShift(Op.getNode(), DAG, Subtarget);
5075  case ISD::SHL_PARTS:     return LowerShiftLeftParts(Op, DAG);
5076  case ISD::SRL_PARTS:
5077  case ISD::SRA_PARTS:     return LowerShiftRightParts(Op, DAG);
5078  case ISD::CTTZ:          return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5079  case ISD::SETCC:         return LowerVSETCC(Op, DAG);
5080  case ISD::BUILD_VECTOR:  return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5081  case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5082  case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5083  case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5084  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5085  case ISD::FLT_ROUNDS_:   return LowerFLT_ROUNDS_(Op, DAG);
5086  case ISD::MUL:           return LowerMUL(Op, DAG);
5087  case ISD::SDIV:          return LowerSDIV(Op, DAG);
5088  case ISD::UDIV:          return LowerUDIV(Op, DAG);
5089  case ISD::ADDC:
5090  case ISD::ADDE:
5091  case ISD::SUBC:
5092  case ISD::SUBE:          return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5093  case ISD::ATOMIC_LOAD:
5094  case ISD::ATOMIC_STORE:  return LowerAtomicLoadStore(Op, DAG);
5095  }
5096}
5097
5098/// ReplaceNodeResults - Replace the results of node with an illegal result
5099/// type with new values built out of custom code.
5100void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5101                                           SmallVectorImpl<SDValue>&Results,
5102                                           SelectionDAG &DAG) const {
5103  SDValue Res;
5104  switch (N->getOpcode()) {
5105  default:
5106    llvm_unreachable("Don't know how to custom expand this!");
5107  case ISD::BITCAST:
5108    Res = ExpandBITCAST(N, DAG);
5109    break;
5110  case ISD::SRL:
5111  case ISD::SRA:
5112    Res = Expand64BitShift(N, DAG, Subtarget);
5113    break;
5114  case ISD::ATOMIC_LOAD_ADD:
5115    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5116    return;
5117  case ISD::ATOMIC_LOAD_AND:
5118    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5119    return;
5120  case ISD::ATOMIC_LOAD_NAND:
5121    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5122    return;
5123  case ISD::ATOMIC_LOAD_OR:
5124    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5125    return;
5126  case ISD::ATOMIC_LOAD_SUB:
5127    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5128    return;
5129  case ISD::ATOMIC_LOAD_XOR:
5130    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5131    return;
5132  case ISD::ATOMIC_SWAP:
5133    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5134    return;
5135  case ISD::ATOMIC_CMP_SWAP:
5136    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5137    return;
5138  }
5139  if (Res.getNode())
5140    Results.push_back(Res);
5141}
5142
5143//===----------------------------------------------------------------------===//
5144//                           ARM Scheduler Hooks
5145//===----------------------------------------------------------------------===//
5146
5147MachineBasicBlock *
5148ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5149                                     MachineBasicBlock *BB,
5150                                     unsigned Size) const {
5151  unsigned dest    = MI->getOperand(0).getReg();
5152  unsigned ptr     = MI->getOperand(1).getReg();
5153  unsigned oldval  = MI->getOperand(2).getReg();
5154  unsigned newval  = MI->getOperand(3).getReg();
5155  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5156  DebugLoc dl = MI->getDebugLoc();
5157  bool isThumb2 = Subtarget->isThumb2();
5158
5159  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5160  unsigned scratch =
5161    MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5162                                       : ARM::GPRRegisterClass);
5163
5164  if (isThumb2) {
5165    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5166    MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5167    MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5168  }
5169
5170  unsigned ldrOpc, strOpc;
5171  switch (Size) {
5172  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5173  case 1:
5174    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5175    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5176    break;
5177  case 2:
5178    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5179    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5180    break;
5181  case 4:
5182    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5183    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5184    break;
5185  }
5186
5187  MachineFunction *MF = BB->getParent();
5188  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5189  MachineFunction::iterator It = BB;
5190  ++It; // insert the new blocks after the current block
5191
5192  MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5193  MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5194  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5195  MF->insert(It, loop1MBB);
5196  MF->insert(It, loop2MBB);
5197  MF->insert(It, exitMBB);
5198
5199  // Transfer the remainder of BB and its successor edges to exitMBB.
5200  exitMBB->splice(exitMBB->begin(), BB,
5201                  llvm::next(MachineBasicBlock::iterator(MI)),
5202                  BB->end());
5203  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5204
5205  //  thisMBB:
5206  //   ...
5207  //   fallthrough --> loop1MBB
5208  BB->addSuccessor(loop1MBB);
5209
5210  // loop1MBB:
5211  //   ldrex dest, [ptr]
5212  //   cmp dest, oldval
5213  //   bne exitMBB
5214  BB = loop1MBB;
5215  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5216  if (ldrOpc == ARM::t2LDREX)
5217    MIB.addImm(0);
5218  AddDefaultPred(MIB);
5219  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5220                 .addReg(dest).addReg(oldval));
5221  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5222    .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5223  BB->addSuccessor(loop2MBB);
5224  BB->addSuccessor(exitMBB);
5225
5226  // loop2MBB:
5227  //   strex scratch, newval, [ptr]
5228  //   cmp scratch, #0
5229  //   bne loop1MBB
5230  BB = loop2MBB;
5231  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5232  if (strOpc == ARM::t2STREX)
5233    MIB.addImm(0);
5234  AddDefaultPred(MIB);
5235  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5236                 .addReg(scratch).addImm(0));
5237  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5238    .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5239  BB->addSuccessor(loop1MBB);
5240  BB->addSuccessor(exitMBB);
5241
5242  //  exitMBB:
5243  //   ...
5244  BB = exitMBB;
5245
5246  MI->eraseFromParent();   // The instruction is gone now.
5247
5248  return BB;
5249}
5250
5251MachineBasicBlock *
5252ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5253                                    unsigned Size, unsigned BinOpcode) const {
5254  // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5255  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5256
5257  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5258  MachineFunction *MF = BB->getParent();
5259  MachineFunction::iterator It = BB;
5260  ++It;
5261
5262  unsigned dest = MI->getOperand(0).getReg();
5263  unsigned ptr = MI->getOperand(1).getReg();
5264  unsigned incr = MI->getOperand(2).getReg();
5265  DebugLoc dl = MI->getDebugLoc();
5266  bool isThumb2 = Subtarget->isThumb2();
5267
5268  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5269  if (isThumb2) {
5270    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5271    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5272  }
5273
5274  unsigned ldrOpc, strOpc;
5275  switch (Size) {
5276  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5277  case 1:
5278    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5279    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5280    break;
5281  case 2:
5282    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5283    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5284    break;
5285  case 4:
5286    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5287    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5288    break;
5289  }
5290
5291  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5292  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5293  MF->insert(It, loopMBB);
5294  MF->insert(It, exitMBB);
5295
5296  // Transfer the remainder of BB and its successor edges to exitMBB.
5297  exitMBB->splice(exitMBB->begin(), BB,
5298                  llvm::next(MachineBasicBlock::iterator(MI)),
5299                  BB->end());
5300  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5301
5302  const TargetRegisterClass *TRC =
5303    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5304  unsigned scratch = MRI.createVirtualRegister(TRC);
5305  unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5306
5307  //  thisMBB:
5308  //   ...
5309  //   fallthrough --> loopMBB
5310  BB->addSuccessor(loopMBB);
5311
5312  //  loopMBB:
5313  //   ldrex dest, ptr
5314  //   <binop> scratch2, dest, incr
5315  //   strex scratch, scratch2, ptr
5316  //   cmp scratch, #0
5317  //   bne- loopMBB
5318  //   fallthrough --> exitMBB
5319  BB = loopMBB;
5320  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5321  if (ldrOpc == ARM::t2LDREX)
5322    MIB.addImm(0);
5323  AddDefaultPred(MIB);
5324  if (BinOpcode) {
5325    // operand order needs to go the other way for NAND
5326    if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5327      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5328                     addReg(incr).addReg(dest)).addReg(0);
5329    else
5330      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5331                     addReg(dest).addReg(incr)).addReg(0);
5332  }
5333
5334  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5335  if (strOpc == ARM::t2STREX)
5336    MIB.addImm(0);
5337  AddDefaultPred(MIB);
5338  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5339                 .addReg(scratch).addImm(0));
5340  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5341    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5342
5343  BB->addSuccessor(loopMBB);
5344  BB->addSuccessor(exitMBB);
5345
5346  //  exitMBB:
5347  //   ...
5348  BB = exitMBB;
5349
5350  MI->eraseFromParent();   // The instruction is gone now.
5351
5352  return BB;
5353}
5354
5355MachineBasicBlock *
5356ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5357                                          MachineBasicBlock *BB,
5358                                          unsigned Size,
5359                                          bool signExtend,
5360                                          ARMCC::CondCodes Cond) const {
5361  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5362
5363  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5364  MachineFunction *MF = BB->getParent();
5365  MachineFunction::iterator It = BB;
5366  ++It;
5367
5368  unsigned dest = MI->getOperand(0).getReg();
5369  unsigned ptr = MI->getOperand(1).getReg();
5370  unsigned incr = MI->getOperand(2).getReg();
5371  unsigned oldval = dest;
5372  DebugLoc dl = MI->getDebugLoc();
5373  bool isThumb2 = Subtarget->isThumb2();
5374
5375  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5376  if (isThumb2) {
5377    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5378    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5379  }
5380
5381  unsigned ldrOpc, strOpc, extendOpc;
5382  switch (Size) {
5383  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5384  case 1:
5385    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5386    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5387    extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5388    break;
5389  case 2:
5390    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5391    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5392    extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5393    break;
5394  case 4:
5395    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5396    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5397    extendOpc = 0;
5398    break;
5399  }
5400
5401  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5402  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5403  MF->insert(It, loopMBB);
5404  MF->insert(It, exitMBB);
5405
5406  // Transfer the remainder of BB and its successor edges to exitMBB.
5407  exitMBB->splice(exitMBB->begin(), BB,
5408                  llvm::next(MachineBasicBlock::iterator(MI)),
5409                  BB->end());
5410  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5411
5412  const TargetRegisterClass *TRC =
5413    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5414  unsigned scratch = MRI.createVirtualRegister(TRC);
5415  unsigned scratch2 = MRI.createVirtualRegister(TRC);
5416
5417  //  thisMBB:
5418  //   ...
5419  //   fallthrough --> loopMBB
5420  BB->addSuccessor(loopMBB);
5421
5422  //  loopMBB:
5423  //   ldrex dest, ptr
5424  //   (sign extend dest, if required)
5425  //   cmp dest, incr
5426  //   cmov.cond scratch2, dest, incr
5427  //   strex scratch, scratch2, ptr
5428  //   cmp scratch, #0
5429  //   bne- loopMBB
5430  //   fallthrough --> exitMBB
5431  BB = loopMBB;
5432  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5433  if (ldrOpc == ARM::t2LDREX)
5434    MIB.addImm(0);
5435  AddDefaultPred(MIB);
5436
5437  // Sign extend the value, if necessary.
5438  if (signExtend && extendOpc) {
5439    oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5440    AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5441                     .addReg(dest)
5442                     .addImm(0));
5443  }
5444
5445  // Build compare and cmov instructions.
5446  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5447                 .addReg(oldval).addReg(incr));
5448  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5449         .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5450
5451  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5452  if (strOpc == ARM::t2STREX)
5453    MIB.addImm(0);
5454  AddDefaultPred(MIB);
5455  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5456                 .addReg(scratch).addImm(0));
5457  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5458    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5459
5460  BB->addSuccessor(loopMBB);
5461  BB->addSuccessor(exitMBB);
5462
5463  //  exitMBB:
5464  //   ...
5465  BB = exitMBB;
5466
5467  MI->eraseFromParent();   // The instruction is gone now.
5468
5469  return BB;
5470}
5471
5472MachineBasicBlock *
5473ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5474                                      unsigned Op1, unsigned Op2,
5475                                      bool NeedsCarry, bool IsCmpxchg) const {
5476  // This also handles ATOMIC_SWAP, indicated by Op1==0.
5477  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5478
5479  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5480  MachineFunction *MF = BB->getParent();
5481  MachineFunction::iterator It = BB;
5482  ++It;
5483
5484  unsigned destlo = MI->getOperand(0).getReg();
5485  unsigned desthi = MI->getOperand(1).getReg();
5486  unsigned ptr = MI->getOperand(2).getReg();
5487  unsigned vallo = MI->getOperand(3).getReg();
5488  unsigned valhi = MI->getOperand(4).getReg();
5489  DebugLoc dl = MI->getDebugLoc();
5490  bool isThumb2 = Subtarget->isThumb2();
5491
5492  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5493  if (isThumb2) {
5494    MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5495    MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5496    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5497  }
5498
5499  unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5500  unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5501
5502  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5503  MachineBasicBlock *contBB = 0, *cont2BB = 0;
5504  if (IsCmpxchg) {
5505    contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5506    cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5507  }
5508  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5509  MF->insert(It, loopMBB);
5510  if (IsCmpxchg) {
5511    MF->insert(It, contBB);
5512    MF->insert(It, cont2BB);
5513  }
5514  MF->insert(It, exitMBB);
5515
5516  // Transfer the remainder of BB and its successor edges to exitMBB.
5517  exitMBB->splice(exitMBB->begin(), BB,
5518                  llvm::next(MachineBasicBlock::iterator(MI)),
5519                  BB->end());
5520  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5521
5522  const TargetRegisterClass *TRC =
5523    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5524  unsigned storesuccess = MRI.createVirtualRegister(TRC);
5525
5526  //  thisMBB:
5527  //   ...
5528  //   fallthrough --> loopMBB
5529  BB->addSuccessor(loopMBB);
5530
5531  //  loopMBB:
5532  //   ldrexd r2, r3, ptr
5533  //   <binopa> r0, r2, incr
5534  //   <binopb> r1, r3, incr
5535  //   strexd storesuccess, r0, r1, ptr
5536  //   cmp storesuccess, #0
5537  //   bne- loopMBB
5538  //   fallthrough --> exitMBB
5539  //
5540  // Note that the registers are explicitly specified because there is not any
5541  // way to force the register allocator to allocate a register pair.
5542  //
5543  // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5544  // need to properly enforce the restriction that the two output registers
5545  // for ldrexd must be different.
5546  BB = loopMBB;
5547  // Load
5548  AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5549                 .addReg(ARM::R2, RegState::Define)
5550                 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5551  // Copy r2/r3 into dest.  (This copy will normally be coalesced.)
5552  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5553  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5554
5555  if (IsCmpxchg) {
5556    // Add early exit
5557    for (unsigned i = 0; i < 2; i++) {
5558      AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5559                                                         ARM::CMPrr))
5560                     .addReg(i == 0 ? destlo : desthi)
5561                     .addReg(i == 0 ? vallo : valhi));
5562      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5563        .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5564      BB->addSuccessor(exitMBB);
5565      BB->addSuccessor(i == 0 ? contBB : cont2BB);
5566      BB = (i == 0 ? contBB : cont2BB);
5567    }
5568
5569    // Copy to physregs for strexd
5570    unsigned setlo = MI->getOperand(5).getReg();
5571    unsigned sethi = MI->getOperand(6).getReg();
5572    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5573    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5574  } else if (Op1) {
5575    // Perform binary operation
5576    AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5577                   .addReg(destlo).addReg(vallo))
5578        .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5579    AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5580                   .addReg(desthi).addReg(valhi)).addReg(0);
5581  } else {
5582    // Copy to physregs for strexd
5583    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5584    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5585  }
5586
5587  // Store
5588  AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5589                 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5590  // Cmp+jump
5591  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5592                 .addReg(storesuccess).addImm(0));
5593  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5594    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5595
5596  BB->addSuccessor(loopMBB);
5597  BB->addSuccessor(exitMBB);
5598
5599  //  exitMBB:
5600  //   ...
5601  BB = exitMBB;
5602
5603  MI->eraseFromParent();   // The instruction is gone now.
5604
5605  return BB;
5606}
5607
5608/// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5609/// registers the function context.
5610void ARMTargetLowering::
5611SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5612                       MachineBasicBlock *DispatchBB, int FI) const {
5613  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5614  DebugLoc dl = MI->getDebugLoc();
5615  MachineFunction *MF = MBB->getParent();
5616  MachineRegisterInfo *MRI = &MF->getRegInfo();
5617  MachineConstantPool *MCP = MF->getConstantPool();
5618  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5619  const Function *F = MF->getFunction();
5620
5621  bool isThumb = Subtarget->isThumb();
5622  bool isThumb2 = Subtarget->isThumb2();
5623
5624  unsigned PCLabelId = AFI->createPICLabelUId();
5625  unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5626  ARMConstantPoolValue *CPV =
5627    ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5628  unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5629
5630  const TargetRegisterClass *TRC =
5631    isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5632
5633  // Grab constant pool and fixed stack memory operands.
5634  MachineMemOperand *CPMMO =
5635    MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5636                             MachineMemOperand::MOLoad, 4, 4);
5637
5638  MachineMemOperand *FIMMOSt =
5639    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5640                             MachineMemOperand::MOStore, 4, 4);
5641
5642  // Load the address of the dispatch MBB into the jump buffer.
5643  if (isThumb2) {
5644    // Incoming value: jbuf
5645    //   ldr.n  r5, LCPI1_1
5646    //   orr    r5, r5, #1
5647    //   add    r5, pc
5648    //   str    r5, [$jbuf, #+4] ; &jbuf[1]
5649    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5650    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5651                   .addConstantPoolIndex(CPI)
5652                   .addMemOperand(CPMMO));
5653    // Set the low bit because of thumb mode.
5654    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5655    AddDefaultCC(
5656      AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5657                     .addReg(NewVReg1, RegState::Kill)
5658                     .addImm(0x01)));
5659    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5660    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5661      .addReg(NewVReg2, RegState::Kill)
5662      .addImm(PCLabelId);
5663    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5664                   .addReg(NewVReg3, RegState::Kill)
5665                   .addFrameIndex(FI)
5666                   .addImm(36)  // &jbuf[1] :: pc
5667                   .addMemOperand(FIMMOSt));
5668  } else if (isThumb) {
5669    // Incoming value: jbuf
5670    //   ldr.n  r1, LCPI1_4
5671    //   add    r1, pc
5672    //   mov    r2, #1
5673    //   orrs   r1, r2
5674    //   add    r2, $jbuf, #+4 ; &jbuf[1]
5675    //   str    r1, [r2]
5676    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5677    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5678                   .addConstantPoolIndex(CPI)
5679                   .addMemOperand(CPMMO));
5680    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5681    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5682      .addReg(NewVReg1, RegState::Kill)
5683      .addImm(PCLabelId);
5684    // Set the low bit because of thumb mode.
5685    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5686    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5687                   .addReg(ARM::CPSR, RegState::Define)
5688                   .addImm(1));
5689    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5690    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5691                   .addReg(ARM::CPSR, RegState::Define)
5692                   .addReg(NewVReg2, RegState::Kill)
5693                   .addReg(NewVReg3, RegState::Kill));
5694    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5695    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5696                   .addFrameIndex(FI)
5697                   .addImm(36)); // &jbuf[1] :: pc
5698    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5699                   .addReg(NewVReg4, RegState::Kill)
5700                   .addReg(NewVReg5, RegState::Kill)
5701                   .addImm(0)
5702                   .addMemOperand(FIMMOSt));
5703  } else {
5704    // Incoming value: jbuf
5705    //   ldr  r1, LCPI1_1
5706    //   add  r1, pc, r1
5707    //   str  r1, [$jbuf, #+4] ; &jbuf[1]
5708    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5709    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12),  NewVReg1)
5710                   .addConstantPoolIndex(CPI)
5711                   .addImm(0)
5712                   .addMemOperand(CPMMO));
5713    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5714    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5715                   .addReg(NewVReg1, RegState::Kill)
5716                   .addImm(PCLabelId));
5717    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5718                   .addReg(NewVReg2, RegState::Kill)
5719                   .addFrameIndex(FI)
5720                   .addImm(36)  // &jbuf[1] :: pc
5721                   .addMemOperand(FIMMOSt));
5722  }
5723}
5724
5725MachineBasicBlock *ARMTargetLowering::
5726EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5727  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5728  DebugLoc dl = MI->getDebugLoc();
5729  MachineFunction *MF = MBB->getParent();
5730  MachineRegisterInfo *MRI = &MF->getRegInfo();
5731  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5732  MachineFrameInfo *MFI = MF->getFrameInfo();
5733  int FI = MFI->getFunctionContextIndex();
5734
5735  const TargetRegisterClass *TRC =
5736    Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5737
5738  // Get a mapping of the call site numbers to all of the landing pads they're
5739  // associated with.
5740  DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5741  unsigned MaxCSNum = 0;
5742  MachineModuleInfo &MMI = MF->getMMI();
5743  for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5744    if (!BB->isLandingPad()) continue;
5745
5746    // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5747    // pad.
5748    for (MachineBasicBlock::iterator
5749           II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5750      if (!II->isEHLabel()) continue;
5751
5752      MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5753      if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5754
5755      SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5756      for (SmallVectorImpl<unsigned>::iterator
5757             CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5758           CSI != CSE; ++CSI) {
5759        CallSiteNumToLPad[*CSI].push_back(BB);
5760        MaxCSNum = std::max(MaxCSNum, *CSI);
5761      }
5762      break;
5763    }
5764  }
5765
5766  // Get an ordered list of the machine basic blocks for the jump table.
5767  std::vector<MachineBasicBlock*> LPadList;
5768  SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5769  LPadList.reserve(CallSiteNumToLPad.size());
5770  for (unsigned I = 1; I <= MaxCSNum; ++I) {
5771    SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5772    for (SmallVectorImpl<MachineBasicBlock*>::iterator
5773           II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5774      LPadList.push_back(*II);
5775      InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5776    }
5777  }
5778
5779  assert(!LPadList.empty() &&
5780         "No landing pad destinations for the dispatch jump table!");
5781
5782  // Create the jump table and associated information.
5783  MachineJumpTableInfo *JTI =
5784    MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5785  unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5786  unsigned UId = AFI->createJumpTableUId();
5787
5788  // Create the MBBs for the dispatch code.
5789
5790  // Shove the dispatch's address into the return slot in the function context.
5791  MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5792  DispatchBB->setIsLandingPad();
5793
5794  MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5795  BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5796  DispatchBB->addSuccessor(TrapBB);
5797
5798  MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5799  DispatchBB->addSuccessor(DispContBB);
5800
5801  // Insert and MBBs.
5802  MF->insert(MF->end(), DispatchBB);
5803  MF->insert(MF->end(), DispContBB);
5804  MF->insert(MF->end(), TrapBB);
5805
5806  // Insert code into the entry block that creates and registers the function
5807  // context.
5808  SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5809
5810  MachineMemOperand *FIMMOLd =
5811    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5812                             MachineMemOperand::MOLoad |
5813                             MachineMemOperand::MOVolatile, 4, 4);
5814
5815  if (AFI->isThumb1OnlyFunction())
5816    BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5817  else if (!Subtarget->hasVFP2())
5818    BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5819  else
5820    BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5821
5822  unsigned NumLPads = LPadList.size();
5823  if (Subtarget->isThumb2()) {
5824    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5825    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5826                   .addFrameIndex(FI)
5827                   .addImm(4)
5828                   .addMemOperand(FIMMOLd));
5829
5830    if (NumLPads < 256) {
5831      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5832                     .addReg(NewVReg1)
5833                     .addImm(LPadList.size()));
5834    } else {
5835      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5836      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5837                     .addImm(NumLPads & 0xFFFF));
5838
5839      unsigned VReg2 = VReg1;
5840      if ((NumLPads & 0xFFFF0000) != 0) {
5841        VReg2 = MRI->createVirtualRegister(TRC);
5842        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5843                       .addReg(VReg1)
5844                       .addImm(NumLPads >> 16));
5845      }
5846
5847      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5848                     .addReg(NewVReg1)
5849                     .addReg(VReg2));
5850    }
5851
5852    BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5853      .addMBB(TrapBB)
5854      .addImm(ARMCC::HI)
5855      .addReg(ARM::CPSR);
5856
5857    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5858    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5859                   .addJumpTableIndex(MJTI)
5860                   .addImm(UId));
5861
5862    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5863    AddDefaultCC(
5864      AddDefaultPred(
5865        BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5866        .addReg(NewVReg3, RegState::Kill)
5867        .addReg(NewVReg1)
5868        .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5869
5870    BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5871      .addReg(NewVReg4, RegState::Kill)
5872      .addReg(NewVReg1)
5873      .addJumpTableIndex(MJTI)
5874      .addImm(UId);
5875  } else if (Subtarget->isThumb()) {
5876    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5877    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5878                   .addFrameIndex(FI)
5879                   .addImm(1)
5880                   .addMemOperand(FIMMOLd));
5881
5882    if (NumLPads < 256) {
5883      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5884                     .addReg(NewVReg1)
5885                     .addImm(NumLPads));
5886    } else {
5887      MachineConstantPool *ConstantPool = MF->getConstantPool();
5888      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5889      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5890
5891      // MachineConstantPool wants an explicit alignment.
5892      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5893      if (Align == 0)
5894        Align = getTargetData()->getTypeAllocSize(C->getType());
5895      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5896
5897      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5898      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5899                     .addReg(VReg1, RegState::Define)
5900                     .addConstantPoolIndex(Idx));
5901      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5902                     .addReg(NewVReg1)
5903                     .addReg(VReg1));
5904    }
5905
5906    BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5907      .addMBB(TrapBB)
5908      .addImm(ARMCC::HI)
5909      .addReg(ARM::CPSR);
5910
5911    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5912    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5913                   .addReg(ARM::CPSR, RegState::Define)
5914                   .addReg(NewVReg1)
5915                   .addImm(2));
5916
5917    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5918    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5919                   .addJumpTableIndex(MJTI)
5920                   .addImm(UId));
5921
5922    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5923    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5924                   .addReg(ARM::CPSR, RegState::Define)
5925                   .addReg(NewVReg2, RegState::Kill)
5926                   .addReg(NewVReg3));
5927
5928    MachineMemOperand *JTMMOLd =
5929      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5930                               MachineMemOperand::MOLoad, 4, 4);
5931
5932    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5933    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5934                   .addReg(NewVReg4, RegState::Kill)
5935                   .addImm(0)
5936                   .addMemOperand(JTMMOLd));
5937
5938    unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5939    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5940                   .addReg(ARM::CPSR, RegState::Define)
5941                   .addReg(NewVReg5, RegState::Kill)
5942                   .addReg(NewVReg3));
5943
5944    BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5945      .addReg(NewVReg6, RegState::Kill)
5946      .addJumpTableIndex(MJTI)
5947      .addImm(UId);
5948  } else {
5949    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5950    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5951                   .addFrameIndex(FI)
5952                   .addImm(4)
5953                   .addMemOperand(FIMMOLd));
5954
5955    if (NumLPads < 256) {
5956      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5957                     .addReg(NewVReg1)
5958                     .addImm(NumLPads));
5959    } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5960      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5961      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5962                     .addImm(NumLPads & 0xFFFF));
5963
5964      unsigned VReg2 = VReg1;
5965      if ((NumLPads & 0xFFFF0000) != 0) {
5966        VReg2 = MRI->createVirtualRegister(TRC);
5967        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5968                       .addReg(VReg1)
5969                       .addImm(NumLPads >> 16));
5970      }
5971
5972      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5973                     .addReg(NewVReg1)
5974                     .addReg(VReg2));
5975    } else {
5976      MachineConstantPool *ConstantPool = MF->getConstantPool();
5977      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5978      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5979
5980      // MachineConstantPool wants an explicit alignment.
5981      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5982      if (Align == 0)
5983        Align = getTargetData()->getTypeAllocSize(C->getType());
5984      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5985
5986      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5987      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
5988                     .addReg(VReg1, RegState::Define)
5989                     .addConstantPoolIndex(Idx)
5990                     .addImm(0));
5991      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5992                     .addReg(NewVReg1)
5993                     .addReg(VReg1, RegState::Kill));
5994    }
5995
5996    BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
5997      .addMBB(TrapBB)
5998      .addImm(ARMCC::HI)
5999      .addReg(ARM::CPSR);
6000
6001    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6002    AddDefaultCC(
6003      AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6004                     .addReg(NewVReg1)
6005                     .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6006    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6007    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6008                   .addJumpTableIndex(MJTI)
6009                   .addImm(UId));
6010
6011    MachineMemOperand *JTMMOLd =
6012      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6013                               MachineMemOperand::MOLoad, 4, 4);
6014    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6015    AddDefaultPred(
6016      BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6017      .addReg(NewVReg3, RegState::Kill)
6018      .addReg(NewVReg4)
6019      .addImm(0)
6020      .addMemOperand(JTMMOLd));
6021
6022    BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6023      .addReg(NewVReg5, RegState::Kill)
6024      .addReg(NewVReg4)
6025      .addJumpTableIndex(MJTI)
6026      .addImm(UId);
6027  }
6028
6029  // Add the jump table entries as successors to the MBB.
6030  MachineBasicBlock *PrevMBB = 0;
6031  for (std::vector<MachineBasicBlock*>::iterator
6032         I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6033    MachineBasicBlock *CurMBB = *I;
6034    if (PrevMBB != CurMBB)
6035      DispContBB->addSuccessor(CurMBB);
6036    PrevMBB = CurMBB;
6037  }
6038
6039  // N.B. the order the invoke BBs are processed in doesn't matter here.
6040  const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6041  const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6042  const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
6043  SmallVector<MachineBasicBlock*, 64> MBBLPads;
6044  for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6045         I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6046    MachineBasicBlock *BB = *I;
6047
6048    // Remove the landing pad successor from the invoke block and replace it
6049    // with the new dispatch block.
6050    SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6051                                                  BB->succ_end());
6052    while (!Successors.empty()) {
6053      MachineBasicBlock *SMBB = Successors.pop_back_val();
6054      if (SMBB->isLandingPad()) {
6055        BB->removeSuccessor(SMBB);
6056        MBBLPads.push_back(SMBB);
6057      }
6058    }
6059
6060    BB->addSuccessor(DispatchBB);
6061
6062    // Find the invoke call and mark all of the callee-saved registers as
6063    // 'implicit defined' so that they're spilled. This prevents code from
6064    // moving instructions to before the EH block, where they will never be
6065    // executed.
6066    for (MachineBasicBlock::reverse_iterator
6067           II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6068      if (!II->isCall()) continue;
6069
6070      DenseMap<unsigned, bool> DefRegs;
6071      for (MachineInstr::mop_iterator
6072             OI = II->operands_begin(), OE = II->operands_end();
6073           OI != OE; ++OI) {
6074        if (!OI->isReg()) continue;
6075        DefRegs[OI->getReg()] = true;
6076      }
6077
6078      MachineInstrBuilder MIB(&*II);
6079
6080      for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6081        unsigned Reg = SavedRegs[i];
6082        if (Subtarget->isThumb2() &&
6083            !ARM::tGPRRegisterClass->contains(Reg) &&
6084            !ARM::hGPRRegisterClass->contains(Reg))
6085          continue;
6086        else if (Subtarget->isThumb1Only() &&
6087                 !ARM::tGPRRegisterClass->contains(Reg))
6088          continue;
6089        else if (!Subtarget->isThumb() &&
6090                 !ARM::GPRRegisterClass->contains(Reg))
6091          continue;
6092        if (!DefRegs[Reg])
6093          MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6094      }
6095
6096      break;
6097    }
6098  }
6099
6100  // Mark all former landing pads as non-landing pads. The dispatch is the only
6101  // landing pad now.
6102  for (SmallVectorImpl<MachineBasicBlock*>::iterator
6103         I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6104    (*I)->setIsLandingPad(false);
6105
6106  // The instruction is gone now.
6107  MI->eraseFromParent();
6108
6109  return MBB;
6110}
6111
6112static
6113MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6114  for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6115       E = MBB->succ_end(); I != E; ++I)
6116    if (*I != Succ)
6117      return *I;
6118  llvm_unreachable("Expecting a BB with two successors!");
6119}
6120
6121MachineBasicBlock *
6122ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6123                                               MachineBasicBlock *BB) const {
6124  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6125  DebugLoc dl = MI->getDebugLoc();
6126  bool isThumb2 = Subtarget->isThumb2();
6127  switch (MI->getOpcode()) {
6128  default: {
6129    MI->dump();
6130    llvm_unreachable("Unexpected instr type to insert");
6131  }
6132  // The Thumb2 pre-indexed stores have the same MI operands, they just
6133  // define them differently in the .td files from the isel patterns, so
6134  // they need pseudos.
6135  case ARM::t2STR_preidx:
6136    MI->setDesc(TII->get(ARM::t2STR_PRE));
6137    return BB;
6138  case ARM::t2STRB_preidx:
6139    MI->setDesc(TII->get(ARM::t2STRB_PRE));
6140    return BB;
6141  case ARM::t2STRH_preidx:
6142    MI->setDesc(TII->get(ARM::t2STRH_PRE));
6143    return BB;
6144
6145  case ARM::STRi_preidx:
6146  case ARM::STRBi_preidx: {
6147    unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6148      ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6149    // Decode the offset.
6150    unsigned Offset = MI->getOperand(4).getImm();
6151    bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6152    Offset = ARM_AM::getAM2Offset(Offset);
6153    if (isSub)
6154      Offset = -Offset;
6155
6156    MachineMemOperand *MMO = *MI->memoperands_begin();
6157    BuildMI(*BB, MI, dl, TII->get(NewOpc))
6158      .addOperand(MI->getOperand(0))  // Rn_wb
6159      .addOperand(MI->getOperand(1))  // Rt
6160      .addOperand(MI->getOperand(2))  // Rn
6161      .addImm(Offset)                 // offset (skip GPR==zero_reg)
6162      .addOperand(MI->getOperand(5))  // pred
6163      .addOperand(MI->getOperand(6))
6164      .addMemOperand(MMO);
6165    MI->eraseFromParent();
6166    return BB;
6167  }
6168  case ARM::STRr_preidx:
6169  case ARM::STRBr_preidx:
6170  case ARM::STRH_preidx: {
6171    unsigned NewOpc;
6172    switch (MI->getOpcode()) {
6173    default: llvm_unreachable("unexpected opcode!");
6174    case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6175    case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6176    case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6177    }
6178    MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6179    for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6180      MIB.addOperand(MI->getOperand(i));
6181    MI->eraseFromParent();
6182    return BB;
6183  }
6184  case ARM::ATOMIC_LOAD_ADD_I8:
6185     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6186  case ARM::ATOMIC_LOAD_ADD_I16:
6187     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6188  case ARM::ATOMIC_LOAD_ADD_I32:
6189     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6190
6191  case ARM::ATOMIC_LOAD_AND_I8:
6192     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6193  case ARM::ATOMIC_LOAD_AND_I16:
6194     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6195  case ARM::ATOMIC_LOAD_AND_I32:
6196     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6197
6198  case ARM::ATOMIC_LOAD_OR_I8:
6199     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6200  case ARM::ATOMIC_LOAD_OR_I16:
6201     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6202  case ARM::ATOMIC_LOAD_OR_I32:
6203     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6204
6205  case ARM::ATOMIC_LOAD_XOR_I8:
6206     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6207  case ARM::ATOMIC_LOAD_XOR_I16:
6208     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6209  case ARM::ATOMIC_LOAD_XOR_I32:
6210     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6211
6212  case ARM::ATOMIC_LOAD_NAND_I8:
6213     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6214  case ARM::ATOMIC_LOAD_NAND_I16:
6215     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6216  case ARM::ATOMIC_LOAD_NAND_I32:
6217     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6218
6219  case ARM::ATOMIC_LOAD_SUB_I8:
6220     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6221  case ARM::ATOMIC_LOAD_SUB_I16:
6222     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6223  case ARM::ATOMIC_LOAD_SUB_I32:
6224     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6225
6226  case ARM::ATOMIC_LOAD_MIN_I8:
6227     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6228  case ARM::ATOMIC_LOAD_MIN_I16:
6229     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6230  case ARM::ATOMIC_LOAD_MIN_I32:
6231     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6232
6233  case ARM::ATOMIC_LOAD_MAX_I8:
6234     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6235  case ARM::ATOMIC_LOAD_MAX_I16:
6236     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6237  case ARM::ATOMIC_LOAD_MAX_I32:
6238     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6239
6240  case ARM::ATOMIC_LOAD_UMIN_I8:
6241     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6242  case ARM::ATOMIC_LOAD_UMIN_I16:
6243     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6244  case ARM::ATOMIC_LOAD_UMIN_I32:
6245     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6246
6247  case ARM::ATOMIC_LOAD_UMAX_I8:
6248     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6249  case ARM::ATOMIC_LOAD_UMAX_I16:
6250     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6251  case ARM::ATOMIC_LOAD_UMAX_I32:
6252     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6253
6254  case ARM::ATOMIC_SWAP_I8:  return EmitAtomicBinary(MI, BB, 1, 0);
6255  case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6256  case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6257
6258  case ARM::ATOMIC_CMP_SWAP_I8:  return EmitAtomicCmpSwap(MI, BB, 1);
6259  case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6260  case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6261
6262
6263  case ARM::ATOMADD6432:
6264    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6265                              isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6266                              /*NeedsCarry*/ true);
6267  case ARM::ATOMSUB6432:
6268    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6269                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6270                              /*NeedsCarry*/ true);
6271  case ARM::ATOMOR6432:
6272    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6273                              isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6274  case ARM::ATOMXOR6432:
6275    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6276                              isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6277  case ARM::ATOMAND6432:
6278    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6279                              isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6280  case ARM::ATOMSWAP6432:
6281    return EmitAtomicBinary64(MI, BB, 0, 0, false);
6282  case ARM::ATOMCMPXCHG6432:
6283    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6284                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6285                              /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6286
6287  case ARM::tMOVCCr_pseudo: {
6288    // To "insert" a SELECT_CC instruction, we actually have to insert the
6289    // diamond control-flow pattern.  The incoming instruction knows the
6290    // destination vreg to set, the condition code register to branch on, the
6291    // true/false values to select between, and a branch opcode to use.
6292    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6293    MachineFunction::iterator It = BB;
6294    ++It;
6295
6296    //  thisMBB:
6297    //  ...
6298    //   TrueVal = ...
6299    //   cmpTY ccX, r1, r2
6300    //   bCC copy1MBB
6301    //   fallthrough --> copy0MBB
6302    MachineBasicBlock *thisMBB  = BB;
6303    MachineFunction *F = BB->getParent();
6304    MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6305    MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
6306    F->insert(It, copy0MBB);
6307    F->insert(It, sinkMBB);
6308
6309    // Transfer the remainder of BB and its successor edges to sinkMBB.
6310    sinkMBB->splice(sinkMBB->begin(), BB,
6311                    llvm::next(MachineBasicBlock::iterator(MI)),
6312                    BB->end());
6313    sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6314
6315    BB->addSuccessor(copy0MBB);
6316    BB->addSuccessor(sinkMBB);
6317
6318    BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6319      .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6320
6321    //  copy0MBB:
6322    //   %FalseValue = ...
6323    //   # fallthrough to sinkMBB
6324    BB = copy0MBB;
6325
6326    // Update machine-CFG edges
6327    BB->addSuccessor(sinkMBB);
6328
6329    //  sinkMBB:
6330    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6331    //  ...
6332    BB = sinkMBB;
6333    BuildMI(*BB, BB->begin(), dl,
6334            TII->get(ARM::PHI), MI->getOperand(0).getReg())
6335      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6336      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6337
6338    MI->eraseFromParent();   // The pseudo instruction is gone now.
6339    return BB;
6340  }
6341
6342  case ARM::BCCi64:
6343  case ARM::BCCZi64: {
6344    // If there is an unconditional branch to the other successor, remove it.
6345    BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6346
6347    // Compare both parts that make up the double comparison separately for
6348    // equality.
6349    bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6350
6351    unsigned LHS1 = MI->getOperand(1).getReg();
6352    unsigned LHS2 = MI->getOperand(2).getReg();
6353    if (RHSisZero) {
6354      AddDefaultPred(BuildMI(BB, dl,
6355                             TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6356                     .addReg(LHS1).addImm(0));
6357      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6358        .addReg(LHS2).addImm(0)
6359        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6360    } else {
6361      unsigned RHS1 = MI->getOperand(3).getReg();
6362      unsigned RHS2 = MI->getOperand(4).getReg();
6363      AddDefaultPred(BuildMI(BB, dl,
6364                             TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6365                     .addReg(LHS1).addReg(RHS1));
6366      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6367        .addReg(LHS2).addReg(RHS2)
6368        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6369    }
6370
6371    MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6372    MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6373    if (MI->getOperand(0).getImm() == ARMCC::NE)
6374      std::swap(destMBB, exitMBB);
6375
6376    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6377      .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6378    if (isThumb2)
6379      AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6380    else
6381      BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6382
6383    MI->eraseFromParent();   // The pseudo instruction is gone now.
6384    return BB;
6385  }
6386
6387  case ARM::Int_eh_sjlj_setjmp:
6388  case ARM::Int_eh_sjlj_setjmp_nofp:
6389  case ARM::tInt_eh_sjlj_setjmp:
6390  case ARM::t2Int_eh_sjlj_setjmp:
6391  case ARM::t2Int_eh_sjlj_setjmp_nofp:
6392    EmitSjLjDispatchBlock(MI, BB);
6393    return BB;
6394
6395  case ARM::ABS:
6396  case ARM::t2ABS: {
6397    // To insert an ABS instruction, we have to insert the
6398    // diamond control-flow pattern.  The incoming instruction knows the
6399    // source vreg to test against 0, the destination vreg to set,
6400    // the condition code register to branch on, the
6401    // true/false values to select between, and a branch opcode to use.
6402    // It transforms
6403    //     V1 = ABS V0
6404    // into
6405    //     V2 = MOVS V0
6406    //     BCC                      (branch to SinkBB if V0 >= 0)
6407    //     RSBBB: V3 = RSBri V2, 0  (compute ABS if V2 < 0)
6408    //     SinkBB: V1 = PHI(V2, V3)
6409    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6410    MachineFunction::iterator BBI = BB;
6411    ++BBI;
6412    MachineFunction *Fn = BB->getParent();
6413    MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6414    MachineBasicBlock *SinkBB  = Fn->CreateMachineBasicBlock(LLVM_BB);
6415    Fn->insert(BBI, RSBBB);
6416    Fn->insert(BBI, SinkBB);
6417
6418    unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6419    unsigned int ABSDstReg = MI->getOperand(0).getReg();
6420    bool isThumb2 = Subtarget->isThumb2();
6421    MachineRegisterInfo &MRI = Fn->getRegInfo();
6422    // In Thumb mode S must not be specified if source register is the SP or
6423    // PC and if destination register is the SP, so restrict register class
6424    unsigned NewMovDstReg = MRI.createVirtualRegister(
6425      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6426    unsigned NewRsbDstReg = MRI.createVirtualRegister(
6427      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6428
6429    // Transfer the remainder of BB and its successor edges to sinkMBB.
6430    SinkBB->splice(SinkBB->begin(), BB,
6431      llvm::next(MachineBasicBlock::iterator(MI)),
6432      BB->end());
6433    SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6434
6435    BB->addSuccessor(RSBBB);
6436    BB->addSuccessor(SinkBB);
6437
6438    // fall through to SinkMBB
6439    RSBBB->addSuccessor(SinkBB);
6440
6441    // insert a movs at the end of BB
6442    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6443      NewMovDstReg)
6444      .addReg(ABSSrcReg, RegState::Kill)
6445      .addImm((unsigned)ARMCC::AL).addReg(0)
6446      .addReg(ARM::CPSR, RegState::Define);
6447
6448    // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6449    BuildMI(BB, dl,
6450      TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6451      .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6452
6453    // insert rsbri in RSBBB
6454    // Note: BCC and rsbri will be converted into predicated rsbmi
6455    // by if-conversion pass
6456    BuildMI(*RSBBB, RSBBB->begin(), dl,
6457      TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6458      .addReg(NewMovDstReg, RegState::Kill)
6459      .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6460
6461    // insert PHI in SinkBB,
6462    // reuse ABSDstReg to not change uses of ABS instruction
6463    BuildMI(*SinkBB, SinkBB->begin(), dl,
6464      TII->get(ARM::PHI), ABSDstReg)
6465      .addReg(NewRsbDstReg).addMBB(RSBBB)
6466      .addReg(NewMovDstReg).addMBB(BB);
6467
6468    // remove ABS instruction
6469    MI->eraseFromParent();
6470
6471    // return last added BB
6472    return SinkBB;
6473  }
6474  }
6475}
6476
6477void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6478                                                      SDNode *Node) const {
6479  if (!MI->hasPostISelHook()) {
6480    assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6481           "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6482    return;
6483  }
6484
6485  const MCInstrDesc *MCID = &MI->getDesc();
6486  // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6487  // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6488  // operand is still set to noreg. If needed, set the optional operand's
6489  // register to CPSR, and remove the redundant implicit def.
6490  //
6491  // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6492
6493  // Rename pseudo opcodes.
6494  unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6495  if (NewOpc) {
6496    const ARMBaseInstrInfo *TII =
6497      static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6498    MCID = &TII->get(NewOpc);
6499
6500    assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6501           "converted opcode should be the same except for cc_out");
6502
6503    MI->setDesc(*MCID);
6504
6505    // Add the optional cc_out operand
6506    MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6507  }
6508  unsigned ccOutIdx = MCID->getNumOperands() - 1;
6509
6510  // Any ARM instruction that sets the 's' bit should specify an optional
6511  // "cc_out" operand in the last operand position.
6512  if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6513    assert(!NewOpc && "Optional cc_out operand required");
6514    return;
6515  }
6516  // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6517  // since we already have an optional CPSR def.
6518  bool definesCPSR = false;
6519  bool deadCPSR = false;
6520  for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6521       i != e; ++i) {
6522    const MachineOperand &MO = MI->getOperand(i);
6523    if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6524      definesCPSR = true;
6525      if (MO.isDead())
6526        deadCPSR = true;
6527      MI->RemoveOperand(i);
6528      break;
6529    }
6530  }
6531  if (!definesCPSR) {
6532    assert(!NewOpc && "Optional cc_out operand required");
6533    return;
6534  }
6535  assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6536  if (deadCPSR) {
6537    assert(!MI->getOperand(ccOutIdx).getReg() &&
6538           "expect uninitialized optional cc_out operand");
6539    return;
6540  }
6541
6542  // If this instruction was defined with an optional CPSR def and its dag node
6543  // had a live implicit CPSR def, then activate the optional CPSR def.
6544  MachineOperand &MO = MI->getOperand(ccOutIdx);
6545  MO.setReg(ARM::CPSR);
6546  MO.setIsDef(true);
6547}
6548
6549//===----------------------------------------------------------------------===//
6550//                           ARM Optimization Hooks
6551//===----------------------------------------------------------------------===//
6552
6553static
6554SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6555                            TargetLowering::DAGCombinerInfo &DCI) {
6556  SelectionDAG &DAG = DCI.DAG;
6557  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6558  EVT VT = N->getValueType(0);
6559  unsigned Opc = N->getOpcode();
6560  bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6561  SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6562  SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6563  ISD::CondCode CC = ISD::SETCC_INVALID;
6564
6565  if (isSlctCC) {
6566    CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6567  } else {
6568    SDValue CCOp = Slct.getOperand(0);
6569    if (CCOp.getOpcode() == ISD::SETCC)
6570      CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6571  }
6572
6573  bool DoXform = false;
6574  bool InvCC = false;
6575  assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6576          "Bad input!");
6577
6578  if (LHS.getOpcode() == ISD::Constant &&
6579      cast<ConstantSDNode>(LHS)->isNullValue()) {
6580    DoXform = true;
6581  } else if (CC != ISD::SETCC_INVALID &&
6582             RHS.getOpcode() == ISD::Constant &&
6583             cast<ConstantSDNode>(RHS)->isNullValue()) {
6584    std::swap(LHS, RHS);
6585    SDValue Op0 = Slct.getOperand(0);
6586    EVT OpVT = isSlctCC ? Op0.getValueType() :
6587                          Op0.getOperand(0).getValueType();
6588    bool isInt = OpVT.isInteger();
6589    CC = ISD::getSetCCInverse(CC, isInt);
6590
6591    if (!TLI.isCondCodeLegal(CC, OpVT))
6592      return SDValue();         // Inverse operator isn't legal.
6593
6594    DoXform = true;
6595    InvCC = true;
6596  }
6597
6598  if (DoXform) {
6599    SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6600    if (isSlctCC)
6601      return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6602                             Slct.getOperand(0), Slct.getOperand(1), CC);
6603    SDValue CCOp = Slct.getOperand(0);
6604    if (InvCC)
6605      CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6606                          CCOp.getOperand(0), CCOp.getOperand(1), CC);
6607    return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6608                       CCOp, OtherOp, Result);
6609  }
6610  return SDValue();
6611}
6612
6613// AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6614// (only after legalization).
6615static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6616                                 TargetLowering::DAGCombinerInfo &DCI,
6617                                 const ARMSubtarget *Subtarget) {
6618
6619  // Only perform optimization if after legalize, and if NEON is available. We
6620  // also expected both operands to be BUILD_VECTORs.
6621  if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6622      || N0.getOpcode() != ISD::BUILD_VECTOR
6623      || N1.getOpcode() != ISD::BUILD_VECTOR)
6624    return SDValue();
6625
6626  // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6627  EVT VT = N->getValueType(0);
6628  if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6629    return SDValue();
6630
6631  // Check that the vector operands are of the right form.
6632  // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6633  // operands, where N is the size of the formed vector.
6634  // Each EXTRACT_VECTOR should have the same input vector and odd or even
6635  // index such that we have a pair wise add pattern.
6636
6637  // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6638  if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6639    return SDValue();
6640  SDValue Vec = N0->getOperand(0)->getOperand(0);
6641  SDNode *V = Vec.getNode();
6642  unsigned nextIndex = 0;
6643
6644  // For each operands to the ADD which are BUILD_VECTORs,
6645  // check to see if each of their operands are an EXTRACT_VECTOR with
6646  // the same vector and appropriate index.
6647  for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6648    if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6649        && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6650
6651      SDValue ExtVec0 = N0->getOperand(i);
6652      SDValue ExtVec1 = N1->getOperand(i);
6653
6654      // First operand is the vector, verify its the same.
6655      if (V != ExtVec0->getOperand(0).getNode() ||
6656          V != ExtVec1->getOperand(0).getNode())
6657        return SDValue();
6658
6659      // Second is the constant, verify its correct.
6660      ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6661      ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6662
6663      // For the constant, we want to see all the even or all the odd.
6664      if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6665          || C1->getZExtValue() != nextIndex+1)
6666        return SDValue();
6667
6668      // Increment index.
6669      nextIndex+=2;
6670    } else
6671      return SDValue();
6672  }
6673
6674  // Create VPADDL node.
6675  SelectionDAG &DAG = DCI.DAG;
6676  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6677
6678  // Build operand list.
6679  SmallVector<SDValue, 8> Ops;
6680  Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6681                                TLI.getPointerTy()));
6682
6683  // Input is the vector.
6684  Ops.push_back(Vec);
6685
6686  // Get widened type and narrowed type.
6687  MVT widenType;
6688  unsigned numElem = VT.getVectorNumElements();
6689  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6690    case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6691    case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6692    case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6693    default:
6694      llvm_unreachable("Invalid vector element type for padd optimization.");
6695  }
6696
6697  SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6698                            widenType, &Ops[0], Ops.size());
6699  return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6700}
6701
6702/// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6703/// operands N0 and N1.  This is a helper for PerformADDCombine that is
6704/// called with the default operands, and if that fails, with commuted
6705/// operands.
6706static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6707                                          TargetLowering::DAGCombinerInfo &DCI,
6708                                          const ARMSubtarget *Subtarget){
6709
6710  // Attempt to create vpaddl for this add.
6711  SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6712  if (Result.getNode())
6713    return Result;
6714
6715  // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6716  if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6717    SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6718    if (Result.getNode()) return Result;
6719  }
6720  return SDValue();
6721}
6722
6723/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6724///
6725static SDValue PerformADDCombine(SDNode *N,
6726                                 TargetLowering::DAGCombinerInfo &DCI,
6727                                 const ARMSubtarget *Subtarget) {
6728  SDValue N0 = N->getOperand(0);
6729  SDValue N1 = N->getOperand(1);
6730
6731  // First try with the default operand order.
6732  SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6733  if (Result.getNode())
6734    return Result;
6735
6736  // If that didn't work, try again with the operands commuted.
6737  return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6738}
6739
6740/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6741///
6742static SDValue PerformSUBCombine(SDNode *N,
6743                                 TargetLowering::DAGCombinerInfo &DCI) {
6744  SDValue N0 = N->getOperand(0);
6745  SDValue N1 = N->getOperand(1);
6746
6747  // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6748  if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6749    SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6750    if (Result.getNode()) return Result;
6751  }
6752
6753  return SDValue();
6754}
6755
6756/// PerformVMULCombine
6757/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6758/// special multiplier accumulator forwarding.
6759///   vmul d3, d0, d2
6760///   vmla d3, d1, d2
6761/// is faster than
6762///   vadd d3, d0, d1
6763///   vmul d3, d3, d2
6764static SDValue PerformVMULCombine(SDNode *N,
6765                                  TargetLowering::DAGCombinerInfo &DCI,
6766                                  const ARMSubtarget *Subtarget) {
6767  if (!Subtarget->hasVMLxForwarding())
6768    return SDValue();
6769
6770  SelectionDAG &DAG = DCI.DAG;
6771  SDValue N0 = N->getOperand(0);
6772  SDValue N1 = N->getOperand(1);
6773  unsigned Opcode = N0.getOpcode();
6774  if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6775      Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6776    Opcode = N1.getOpcode();
6777    if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6778        Opcode != ISD::FADD && Opcode != ISD::FSUB)
6779      return SDValue();
6780    std::swap(N0, N1);
6781  }
6782
6783  EVT VT = N->getValueType(0);
6784  DebugLoc DL = N->getDebugLoc();
6785  SDValue N00 = N0->getOperand(0);
6786  SDValue N01 = N0->getOperand(1);
6787  return DAG.getNode(Opcode, DL, VT,
6788                     DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6789                     DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6790}
6791
6792static SDValue PerformMULCombine(SDNode *N,
6793                                 TargetLowering::DAGCombinerInfo &DCI,
6794                                 const ARMSubtarget *Subtarget) {
6795  SelectionDAG &DAG = DCI.DAG;
6796
6797  if (Subtarget->isThumb1Only())
6798    return SDValue();
6799
6800  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6801    return SDValue();
6802
6803  EVT VT = N->getValueType(0);
6804  if (VT.is64BitVector() || VT.is128BitVector())
6805    return PerformVMULCombine(N, DCI, Subtarget);
6806  if (VT != MVT::i32)
6807    return SDValue();
6808
6809  ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6810  if (!C)
6811    return SDValue();
6812
6813  uint64_t MulAmt = C->getZExtValue();
6814  unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6815  ShiftAmt = ShiftAmt & (32 - 1);
6816  SDValue V = N->getOperand(0);
6817  DebugLoc DL = N->getDebugLoc();
6818
6819  SDValue Res;
6820  MulAmt >>= ShiftAmt;
6821  if (isPowerOf2_32(MulAmt - 1)) {
6822    // (mul x, 2^N + 1) => (add (shl x, N), x)
6823    Res = DAG.getNode(ISD::ADD, DL, VT,
6824                      V, DAG.getNode(ISD::SHL, DL, VT,
6825                                     V, DAG.getConstant(Log2_32(MulAmt-1),
6826                                                        MVT::i32)));
6827  } else if (isPowerOf2_32(MulAmt + 1)) {
6828    // (mul x, 2^N - 1) => (sub (shl x, N), x)
6829    Res = DAG.getNode(ISD::SUB, DL, VT,
6830                      DAG.getNode(ISD::SHL, DL, VT,
6831                                  V, DAG.getConstant(Log2_32(MulAmt+1),
6832                                                     MVT::i32)),
6833                                                     V);
6834  } else
6835    return SDValue();
6836
6837  if (ShiftAmt != 0)
6838    Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6839                      DAG.getConstant(ShiftAmt, MVT::i32));
6840
6841  // Do not add new nodes to DAG combiner worklist.
6842  DCI.CombineTo(N, Res, false);
6843  return SDValue();
6844}
6845
6846static SDValue PerformANDCombine(SDNode *N,
6847                                TargetLowering::DAGCombinerInfo &DCI) {
6848
6849  // Attempt to use immediate-form VBIC
6850  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6851  DebugLoc dl = N->getDebugLoc();
6852  EVT VT = N->getValueType(0);
6853  SelectionDAG &DAG = DCI.DAG;
6854
6855  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6856    return SDValue();
6857
6858  APInt SplatBits, SplatUndef;
6859  unsigned SplatBitSize;
6860  bool HasAnyUndefs;
6861  if (BVN &&
6862      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6863    if (SplatBitSize <= 64) {
6864      EVT VbicVT;
6865      SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6866                                      SplatUndef.getZExtValue(), SplatBitSize,
6867                                      DAG, VbicVT, VT.is128BitVector(),
6868                                      OtherModImm);
6869      if (Val.getNode()) {
6870        SDValue Input =
6871          DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6872        SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6873        return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6874      }
6875    }
6876  }
6877
6878  return SDValue();
6879}
6880
6881/// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6882static SDValue PerformORCombine(SDNode *N,
6883                                TargetLowering::DAGCombinerInfo &DCI,
6884                                const ARMSubtarget *Subtarget) {
6885  // Attempt to use immediate-form VORR
6886  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6887  DebugLoc dl = N->getDebugLoc();
6888  EVT VT = N->getValueType(0);
6889  SelectionDAG &DAG = DCI.DAG;
6890
6891  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6892    return SDValue();
6893
6894  APInt SplatBits, SplatUndef;
6895  unsigned SplatBitSize;
6896  bool HasAnyUndefs;
6897  if (BVN && Subtarget->hasNEON() &&
6898      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6899    if (SplatBitSize <= 64) {
6900      EVT VorrVT;
6901      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6902                                      SplatUndef.getZExtValue(), SplatBitSize,
6903                                      DAG, VorrVT, VT.is128BitVector(),
6904                                      OtherModImm);
6905      if (Val.getNode()) {
6906        SDValue Input =
6907          DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6908        SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6909        return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6910      }
6911    }
6912  }
6913
6914  SDValue N0 = N->getOperand(0);
6915  if (N0.getOpcode() != ISD::AND)
6916    return SDValue();
6917  SDValue N1 = N->getOperand(1);
6918
6919  // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6920  if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6921      DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6922    APInt SplatUndef;
6923    unsigned SplatBitSize;
6924    bool HasAnyUndefs;
6925
6926    BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6927    APInt SplatBits0;
6928    if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6929                                  HasAnyUndefs) && !HasAnyUndefs) {
6930      BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6931      APInt SplatBits1;
6932      if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6933                                    HasAnyUndefs) && !HasAnyUndefs &&
6934          SplatBits0 == ~SplatBits1) {
6935        // Canonicalize the vector type to make instruction selection simpler.
6936        EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6937        SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6938                                     N0->getOperand(1), N0->getOperand(0),
6939                                     N1->getOperand(0));
6940        return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6941      }
6942    }
6943  }
6944
6945  // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6946  // reasonable.
6947
6948  // BFI is only available on V6T2+
6949  if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6950    return SDValue();
6951
6952  DebugLoc DL = N->getDebugLoc();
6953  // 1) or (and A, mask), val => ARMbfi A, val, mask
6954  //      iff (val & mask) == val
6955  //
6956  // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6957  //  2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6958  //          && mask == ~mask2
6959  //  2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6960  //          && ~mask == mask2
6961  //  (i.e., copy a bitfield value into another bitfield of the same width)
6962
6963  if (VT != MVT::i32)
6964    return SDValue();
6965
6966  SDValue N00 = N0.getOperand(0);
6967
6968  // The value and the mask need to be constants so we can verify this is
6969  // actually a bitfield set. If the mask is 0xffff, we can do better
6970  // via a movt instruction, so don't use BFI in that case.
6971  SDValue MaskOp = N0.getOperand(1);
6972  ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6973  if (!MaskC)
6974    return SDValue();
6975  unsigned Mask = MaskC->getZExtValue();
6976  if (Mask == 0xffff)
6977    return SDValue();
6978  SDValue Res;
6979  // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6980  ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6981  if (N1C) {
6982    unsigned Val = N1C->getZExtValue();
6983    if ((Val & ~Mask) != Val)
6984      return SDValue();
6985
6986    if (ARM::isBitFieldInvertedMask(Mask)) {
6987      Val >>= CountTrailingZeros_32(~Mask);
6988
6989      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6990                        DAG.getConstant(Val, MVT::i32),
6991                        DAG.getConstant(Mask, MVT::i32));
6992
6993      // Do not add new nodes to DAG combiner worklist.
6994      DCI.CombineTo(N, Res, false);
6995      return SDValue();
6996    }
6997  } else if (N1.getOpcode() == ISD::AND) {
6998    // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6999    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7000    if (!N11C)
7001      return SDValue();
7002    unsigned Mask2 = N11C->getZExtValue();
7003
7004    // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7005    // as is to match.
7006    if (ARM::isBitFieldInvertedMask(Mask) &&
7007        (Mask == ~Mask2)) {
7008      // The pack halfword instruction works better for masks that fit it,
7009      // so use that when it's available.
7010      if (Subtarget->hasT2ExtractPack() &&
7011          (Mask == 0xffff || Mask == 0xffff0000))
7012        return SDValue();
7013      // 2a
7014      unsigned amt = CountTrailingZeros_32(Mask2);
7015      Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7016                        DAG.getConstant(amt, MVT::i32));
7017      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7018                        DAG.getConstant(Mask, MVT::i32));
7019      // Do not add new nodes to DAG combiner worklist.
7020      DCI.CombineTo(N, Res, false);
7021      return SDValue();
7022    } else if (ARM::isBitFieldInvertedMask(~Mask) &&
7023               (~Mask == Mask2)) {
7024      // The pack halfword instruction works better for masks that fit it,
7025      // so use that when it's available.
7026      if (Subtarget->hasT2ExtractPack() &&
7027          (Mask2 == 0xffff || Mask2 == 0xffff0000))
7028        return SDValue();
7029      // 2b
7030      unsigned lsb = CountTrailingZeros_32(Mask);
7031      Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7032                        DAG.getConstant(lsb, MVT::i32));
7033      Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7034                        DAG.getConstant(Mask2, MVT::i32));
7035      // Do not add new nodes to DAG combiner worklist.
7036      DCI.CombineTo(N, Res, false);
7037      return SDValue();
7038    }
7039  }
7040
7041  if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7042      N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7043      ARM::isBitFieldInvertedMask(~Mask)) {
7044    // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7045    // where lsb(mask) == #shamt and masked bits of B are known zero.
7046    SDValue ShAmt = N00.getOperand(1);
7047    unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7048    unsigned LSB = CountTrailingZeros_32(Mask);
7049    if (ShAmtC != LSB)
7050      return SDValue();
7051
7052    Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7053                      DAG.getConstant(~Mask, MVT::i32));
7054
7055    // Do not add new nodes to DAG combiner worklist.
7056    DCI.CombineTo(N, Res, false);
7057  }
7058
7059  return SDValue();
7060}
7061
7062/// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7063/// the bits being cleared by the AND are not demanded by the BFI.
7064static SDValue PerformBFICombine(SDNode *N,
7065                                 TargetLowering::DAGCombinerInfo &DCI) {
7066  SDValue N1 = N->getOperand(1);
7067  if (N1.getOpcode() == ISD::AND) {
7068    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7069    if (!N11C)
7070      return SDValue();
7071    unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7072    unsigned LSB = CountTrailingZeros_32(~InvMask);
7073    unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7074    unsigned Mask = (1 << Width)-1;
7075    unsigned Mask2 = N11C->getZExtValue();
7076    if ((Mask & (~Mask2)) == 0)
7077      return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7078                             N->getOperand(0), N1.getOperand(0),
7079                             N->getOperand(2));
7080  }
7081  return SDValue();
7082}
7083
7084/// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7085/// ARMISD::VMOVRRD.
7086static SDValue PerformVMOVRRDCombine(SDNode *N,
7087                                     TargetLowering::DAGCombinerInfo &DCI) {
7088  // vmovrrd(vmovdrr x, y) -> x,y
7089  SDValue InDouble = N->getOperand(0);
7090  if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7091    return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7092
7093  // vmovrrd(load f64) -> (load i32), (load i32)
7094  SDNode *InNode = InDouble.getNode();
7095  if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7096      InNode->getValueType(0) == MVT::f64 &&
7097      InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7098      !cast<LoadSDNode>(InNode)->isVolatile()) {
7099    // TODO: Should this be done for non-FrameIndex operands?
7100    LoadSDNode *LD = cast<LoadSDNode>(InNode);
7101
7102    SelectionDAG &DAG = DCI.DAG;
7103    DebugLoc DL = LD->getDebugLoc();
7104    SDValue BasePtr = LD->getBasePtr();
7105    SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7106                                 LD->getPointerInfo(), LD->isVolatile(),
7107                                 LD->isNonTemporal(), LD->isInvariant(),
7108                                 LD->getAlignment());
7109
7110    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7111                                    DAG.getConstant(4, MVT::i32));
7112    SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7113                                 LD->getPointerInfo(), LD->isVolatile(),
7114                                 LD->isNonTemporal(), LD->isInvariant(),
7115                                 std::min(4U, LD->getAlignment() / 2));
7116
7117    DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7118    SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7119    DCI.RemoveFromWorklist(LD);
7120    DAG.DeleteNode(LD);
7121    return Result;
7122  }
7123
7124  return SDValue();
7125}
7126
7127/// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7128/// ARMISD::VMOVDRR.  This is also used for BUILD_VECTORs with 2 operands.
7129static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7130  // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7131  SDValue Op0 = N->getOperand(0);
7132  SDValue Op1 = N->getOperand(1);
7133  if (Op0.getOpcode() == ISD::BITCAST)
7134    Op0 = Op0.getOperand(0);
7135  if (Op1.getOpcode() == ISD::BITCAST)
7136    Op1 = Op1.getOperand(0);
7137  if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7138      Op0.getNode() == Op1.getNode() &&
7139      Op0.getResNo() == 0 && Op1.getResNo() == 1)
7140    return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7141                       N->getValueType(0), Op0.getOperand(0));
7142  return SDValue();
7143}
7144
7145/// PerformSTORECombine - Target-specific dag combine xforms for
7146/// ISD::STORE.
7147static SDValue PerformSTORECombine(SDNode *N,
7148                                   TargetLowering::DAGCombinerInfo &DCI) {
7149  // Bitcast an i64 store extracted from a vector to f64.
7150  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7151  StoreSDNode *St = cast<StoreSDNode>(N);
7152  SDValue StVal = St->getValue();
7153  if (!ISD::isNormalStore(St) || St->isVolatile())
7154    return SDValue();
7155
7156  if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7157      StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7158    SelectionDAG  &DAG = DCI.DAG;
7159    DebugLoc DL = St->getDebugLoc();
7160    SDValue BasePtr = St->getBasePtr();
7161    SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7162                                  StVal.getNode()->getOperand(0), BasePtr,
7163                                  St->getPointerInfo(), St->isVolatile(),
7164                                  St->isNonTemporal(), St->getAlignment());
7165
7166    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7167                                    DAG.getConstant(4, MVT::i32));
7168    return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7169                        OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7170                        St->isNonTemporal(),
7171                        std::min(4U, St->getAlignment() / 2));
7172  }
7173
7174  if (StVal.getValueType() != MVT::i64 ||
7175      StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7176    return SDValue();
7177
7178  SelectionDAG &DAG = DCI.DAG;
7179  DebugLoc dl = StVal.getDebugLoc();
7180  SDValue IntVec = StVal.getOperand(0);
7181  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7182                                 IntVec.getValueType().getVectorNumElements());
7183  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7184  SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7185                               Vec, StVal.getOperand(1));
7186  dl = N->getDebugLoc();
7187  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7188  // Make the DAGCombiner fold the bitcasts.
7189  DCI.AddToWorklist(Vec.getNode());
7190  DCI.AddToWorklist(ExtElt.getNode());
7191  DCI.AddToWorklist(V.getNode());
7192  return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7193                      St->getPointerInfo(), St->isVolatile(),
7194                      St->isNonTemporal(), St->getAlignment(),
7195                      St->getTBAAInfo());
7196}
7197
7198/// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7199/// are normal, non-volatile loads.  If so, it is profitable to bitcast an
7200/// i64 vector to have f64 elements, since the value can then be loaded
7201/// directly into a VFP register.
7202static bool hasNormalLoadOperand(SDNode *N) {
7203  unsigned NumElts = N->getValueType(0).getVectorNumElements();
7204  for (unsigned i = 0; i < NumElts; ++i) {
7205    SDNode *Elt = N->getOperand(i).getNode();
7206    if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7207      return true;
7208  }
7209  return false;
7210}
7211
7212/// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7213/// ISD::BUILD_VECTOR.
7214static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7215                                          TargetLowering::DAGCombinerInfo &DCI){
7216  // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7217  // VMOVRRD is introduced when legalizing i64 types.  It forces the i64 value
7218  // into a pair of GPRs, which is fine when the value is used as a scalar,
7219  // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7220  SelectionDAG &DAG = DCI.DAG;
7221  if (N->getNumOperands() == 2) {
7222    SDValue RV = PerformVMOVDRRCombine(N, DAG);
7223    if (RV.getNode())
7224      return RV;
7225  }
7226
7227  // Load i64 elements as f64 values so that type legalization does not split
7228  // them up into i32 values.
7229  EVT VT = N->getValueType(0);
7230  if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7231    return SDValue();
7232  DebugLoc dl = N->getDebugLoc();
7233  SmallVector<SDValue, 8> Ops;
7234  unsigned NumElts = VT.getVectorNumElements();
7235  for (unsigned i = 0; i < NumElts; ++i) {
7236    SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7237    Ops.push_back(V);
7238    // Make the DAGCombiner fold the bitcast.
7239    DCI.AddToWorklist(V.getNode());
7240  }
7241  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7242  SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7243  return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7244}
7245
7246/// PerformInsertEltCombine - Target-specific dag combine xforms for
7247/// ISD::INSERT_VECTOR_ELT.
7248static SDValue PerformInsertEltCombine(SDNode *N,
7249                                       TargetLowering::DAGCombinerInfo &DCI) {
7250  // Bitcast an i64 load inserted into a vector to f64.
7251  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7252  EVT VT = N->getValueType(0);
7253  SDNode *Elt = N->getOperand(1).getNode();
7254  if (VT.getVectorElementType() != MVT::i64 ||
7255      !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7256    return SDValue();
7257
7258  SelectionDAG &DAG = DCI.DAG;
7259  DebugLoc dl = N->getDebugLoc();
7260  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7261                                 VT.getVectorNumElements());
7262  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7263  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7264  // Make the DAGCombiner fold the bitcasts.
7265  DCI.AddToWorklist(Vec.getNode());
7266  DCI.AddToWorklist(V.getNode());
7267  SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7268                               Vec, V, N->getOperand(2));
7269  return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7270}
7271
7272/// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7273/// ISD::VECTOR_SHUFFLE.
7274static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7275  // The LLVM shufflevector instruction does not require the shuffle mask
7276  // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7277  // have that requirement.  When translating to ISD::VECTOR_SHUFFLE, if the
7278  // operands do not match the mask length, they are extended by concatenating
7279  // them with undef vectors.  That is probably the right thing for other
7280  // targets, but for NEON it is better to concatenate two double-register
7281  // size vector operands into a single quad-register size vector.  Do that
7282  // transformation here:
7283  //   shuffle(concat(v1, undef), concat(v2, undef)) ->
7284  //   shuffle(concat(v1, v2), undef)
7285  SDValue Op0 = N->getOperand(0);
7286  SDValue Op1 = N->getOperand(1);
7287  if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7288      Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7289      Op0.getNumOperands() != 2 ||
7290      Op1.getNumOperands() != 2)
7291    return SDValue();
7292  SDValue Concat0Op1 = Op0.getOperand(1);
7293  SDValue Concat1Op1 = Op1.getOperand(1);
7294  if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7295      Concat1Op1.getOpcode() != ISD::UNDEF)
7296    return SDValue();
7297  // Skip the transformation if any of the types are illegal.
7298  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7299  EVT VT = N->getValueType(0);
7300  if (!TLI.isTypeLegal(VT) ||
7301      !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7302      !TLI.isTypeLegal(Concat1Op1.getValueType()))
7303    return SDValue();
7304
7305  SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7306                                  Op0.getOperand(0), Op1.getOperand(0));
7307  // Translate the shuffle mask.
7308  SmallVector<int, 16> NewMask;
7309  unsigned NumElts = VT.getVectorNumElements();
7310  unsigned HalfElts = NumElts/2;
7311  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7312  for (unsigned n = 0; n < NumElts; ++n) {
7313    int MaskElt = SVN->getMaskElt(n);
7314    int NewElt = -1;
7315    if (MaskElt < (int)HalfElts)
7316      NewElt = MaskElt;
7317    else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7318      NewElt = HalfElts + MaskElt - NumElts;
7319    NewMask.push_back(NewElt);
7320  }
7321  return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7322                              DAG.getUNDEF(VT), NewMask.data());
7323}
7324
7325/// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7326/// NEON load/store intrinsics to merge base address updates.
7327static SDValue CombineBaseUpdate(SDNode *N,
7328                                 TargetLowering::DAGCombinerInfo &DCI) {
7329  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7330    return SDValue();
7331
7332  SelectionDAG &DAG = DCI.DAG;
7333  bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7334                      N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7335  unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7336  SDValue Addr = N->getOperand(AddrOpIdx);
7337
7338  // Search for a use of the address operand that is an increment.
7339  for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7340         UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7341    SDNode *User = *UI;
7342    if (User->getOpcode() != ISD::ADD ||
7343        UI.getUse().getResNo() != Addr.getResNo())
7344      continue;
7345
7346    // Check that the add is independent of the load/store.  Otherwise, folding
7347    // it would create a cycle.
7348    if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7349      continue;
7350
7351    // Find the new opcode for the updating load/store.
7352    bool isLoad = true;
7353    bool isLaneOp = false;
7354    unsigned NewOpc = 0;
7355    unsigned NumVecs = 0;
7356    if (isIntrinsic) {
7357      unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7358      switch (IntNo) {
7359      default: llvm_unreachable("unexpected intrinsic for Neon base update");
7360      case Intrinsic::arm_neon_vld1:     NewOpc = ARMISD::VLD1_UPD;
7361        NumVecs = 1; break;
7362      case Intrinsic::arm_neon_vld2:     NewOpc = ARMISD::VLD2_UPD;
7363        NumVecs = 2; break;
7364      case Intrinsic::arm_neon_vld3:     NewOpc = ARMISD::VLD3_UPD;
7365        NumVecs = 3; break;
7366      case Intrinsic::arm_neon_vld4:     NewOpc = ARMISD::VLD4_UPD;
7367        NumVecs = 4; break;
7368      case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7369        NumVecs = 2; isLaneOp = true; break;
7370      case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7371        NumVecs = 3; isLaneOp = true; break;
7372      case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7373        NumVecs = 4; isLaneOp = true; break;
7374      case Intrinsic::arm_neon_vst1:     NewOpc = ARMISD::VST1_UPD;
7375        NumVecs = 1; isLoad = false; break;
7376      case Intrinsic::arm_neon_vst2:     NewOpc = ARMISD::VST2_UPD;
7377        NumVecs = 2; isLoad = false; break;
7378      case Intrinsic::arm_neon_vst3:     NewOpc = ARMISD::VST3_UPD;
7379        NumVecs = 3; isLoad = false; break;
7380      case Intrinsic::arm_neon_vst4:     NewOpc = ARMISD::VST4_UPD;
7381        NumVecs = 4; isLoad = false; break;
7382      case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7383        NumVecs = 2; isLoad = false; isLaneOp = true; break;
7384      case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7385        NumVecs = 3; isLoad = false; isLaneOp = true; break;
7386      case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7387        NumVecs = 4; isLoad = false; isLaneOp = true; break;
7388      }
7389    } else {
7390      isLaneOp = true;
7391      switch (N->getOpcode()) {
7392      default: llvm_unreachable("unexpected opcode for Neon base update");
7393      case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7394      case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7395      case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7396      }
7397    }
7398
7399    // Find the size of memory referenced by the load/store.
7400    EVT VecTy;
7401    if (isLoad)
7402      VecTy = N->getValueType(0);
7403    else
7404      VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7405    unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7406    if (isLaneOp)
7407      NumBytes /= VecTy.getVectorNumElements();
7408
7409    // If the increment is a constant, it must match the memory ref size.
7410    SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7411    if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7412      uint64_t IncVal = CInc->getZExtValue();
7413      if (IncVal != NumBytes)
7414        continue;
7415    } else if (NumBytes >= 3 * 16) {
7416      // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7417      // separate instructions that make it harder to use a non-constant update.
7418      continue;
7419    }
7420
7421    // Create the new updating load/store node.
7422    EVT Tys[6];
7423    unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7424    unsigned n;
7425    for (n = 0; n < NumResultVecs; ++n)
7426      Tys[n] = VecTy;
7427    Tys[n++] = MVT::i32;
7428    Tys[n] = MVT::Other;
7429    SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7430    SmallVector<SDValue, 8> Ops;
7431    Ops.push_back(N->getOperand(0)); // incoming chain
7432    Ops.push_back(N->getOperand(AddrOpIdx));
7433    Ops.push_back(Inc);
7434    for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7435      Ops.push_back(N->getOperand(i));
7436    }
7437    MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7438    SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7439                                           Ops.data(), Ops.size(),
7440                                           MemInt->getMemoryVT(),
7441                                           MemInt->getMemOperand());
7442
7443    // Update the uses.
7444    std::vector<SDValue> NewResults;
7445    for (unsigned i = 0; i < NumResultVecs; ++i) {
7446      NewResults.push_back(SDValue(UpdN.getNode(), i));
7447    }
7448    NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7449    DCI.CombineTo(N, NewResults);
7450    DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7451
7452    break;
7453  }
7454  return SDValue();
7455}
7456
7457/// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7458/// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7459/// are also VDUPLANEs.  If so, combine them to a vldN-dup operation and
7460/// return true.
7461static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7462  SelectionDAG &DAG = DCI.DAG;
7463  EVT VT = N->getValueType(0);
7464  // vldN-dup instructions only support 64-bit vectors for N > 1.
7465  if (!VT.is64BitVector())
7466    return false;
7467
7468  // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7469  SDNode *VLD = N->getOperand(0).getNode();
7470  if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7471    return false;
7472  unsigned NumVecs = 0;
7473  unsigned NewOpc = 0;
7474  unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7475  if (IntNo == Intrinsic::arm_neon_vld2lane) {
7476    NumVecs = 2;
7477    NewOpc = ARMISD::VLD2DUP;
7478  } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7479    NumVecs = 3;
7480    NewOpc = ARMISD::VLD3DUP;
7481  } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7482    NumVecs = 4;
7483    NewOpc = ARMISD::VLD4DUP;
7484  } else {
7485    return false;
7486  }
7487
7488  // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7489  // numbers match the load.
7490  unsigned VLDLaneNo =
7491    cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7492  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7493       UI != UE; ++UI) {
7494    // Ignore uses of the chain result.
7495    if (UI.getUse().getResNo() == NumVecs)
7496      continue;
7497    SDNode *User = *UI;
7498    if (User->getOpcode() != ARMISD::VDUPLANE ||
7499        VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7500      return false;
7501  }
7502
7503  // Create the vldN-dup node.
7504  EVT Tys[5];
7505  unsigned n;
7506  for (n = 0; n < NumVecs; ++n)
7507    Tys[n] = VT;
7508  Tys[n] = MVT::Other;
7509  SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7510  SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7511  MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7512  SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7513                                           Ops, 2, VLDMemInt->getMemoryVT(),
7514                                           VLDMemInt->getMemOperand());
7515
7516  // Update the uses.
7517  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7518       UI != UE; ++UI) {
7519    unsigned ResNo = UI.getUse().getResNo();
7520    // Ignore uses of the chain result.
7521    if (ResNo == NumVecs)
7522      continue;
7523    SDNode *User = *UI;
7524    DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7525  }
7526
7527  // Now the vldN-lane intrinsic is dead except for its chain result.
7528  // Update uses of the chain.
7529  std::vector<SDValue> VLDDupResults;
7530  for (unsigned n = 0; n < NumVecs; ++n)
7531    VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7532  VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7533  DCI.CombineTo(VLD, VLDDupResults);
7534
7535  return true;
7536}
7537
7538/// PerformVDUPLANECombine - Target-specific dag combine xforms for
7539/// ARMISD::VDUPLANE.
7540static SDValue PerformVDUPLANECombine(SDNode *N,
7541                                      TargetLowering::DAGCombinerInfo &DCI) {
7542  SDValue Op = N->getOperand(0);
7543
7544  // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7545  // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7546  if (CombineVLDDUP(N, DCI))
7547    return SDValue(N, 0);
7548
7549  // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7550  // redundant.  Ignore bit_converts for now; element sizes are checked below.
7551  while (Op.getOpcode() == ISD::BITCAST)
7552    Op = Op.getOperand(0);
7553  if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7554    return SDValue();
7555
7556  // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7557  unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7558  // The canonical VMOV for a zero vector uses a 32-bit element size.
7559  unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7560  unsigned EltBits;
7561  if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7562    EltSize = 8;
7563  EVT VT = N->getValueType(0);
7564  if (EltSize > VT.getVectorElementType().getSizeInBits())
7565    return SDValue();
7566
7567  return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7568}
7569
7570// isConstVecPow2 - Return true if each vector element is a power of 2, all
7571// elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7572static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7573{
7574  integerPart cN;
7575  integerPart c0 = 0;
7576  for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7577       I != E; I++) {
7578    ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7579    if (!C)
7580      return false;
7581
7582    bool isExact;
7583    APFloat APF = C->getValueAPF();
7584    if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7585        != APFloat::opOK || !isExact)
7586      return false;
7587
7588    c0 = (I == 0) ? cN : c0;
7589    if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7590      return false;
7591  }
7592  C = c0;
7593  return true;
7594}
7595
7596/// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7597/// can replace combinations of VMUL and VCVT (floating-point to integer)
7598/// when the VMUL has a constant operand that is a power of 2.
7599///
7600/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7601///  vmul.f32        d16, d17, d16
7602///  vcvt.s32.f32    d16, d16
7603/// becomes:
7604///  vcvt.s32.f32    d16, d16, #3
7605static SDValue PerformVCVTCombine(SDNode *N,
7606                                  TargetLowering::DAGCombinerInfo &DCI,
7607                                  const ARMSubtarget *Subtarget) {
7608  SelectionDAG &DAG = DCI.DAG;
7609  SDValue Op = N->getOperand(0);
7610
7611  if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7612      Op.getOpcode() != ISD::FMUL)
7613    return SDValue();
7614
7615  uint64_t C;
7616  SDValue N0 = Op->getOperand(0);
7617  SDValue ConstVec = Op->getOperand(1);
7618  bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7619
7620  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7621      !isConstVecPow2(ConstVec, isSigned, C))
7622    return SDValue();
7623
7624  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7625    Intrinsic::arm_neon_vcvtfp2fxu;
7626  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7627                     N->getValueType(0),
7628                     DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7629                     DAG.getConstant(Log2_64(C), MVT::i32));
7630}
7631
7632/// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7633/// can replace combinations of VCVT (integer to floating-point) and VDIV
7634/// when the VDIV has a constant operand that is a power of 2.
7635///
7636/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7637///  vcvt.f32.s32    d16, d16
7638///  vdiv.f32        d16, d17, d16
7639/// becomes:
7640///  vcvt.f32.s32    d16, d16, #3
7641static SDValue PerformVDIVCombine(SDNode *N,
7642                                  TargetLowering::DAGCombinerInfo &DCI,
7643                                  const ARMSubtarget *Subtarget) {
7644  SelectionDAG &DAG = DCI.DAG;
7645  SDValue Op = N->getOperand(0);
7646  unsigned OpOpcode = Op.getNode()->getOpcode();
7647
7648  if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7649      (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7650    return SDValue();
7651
7652  uint64_t C;
7653  SDValue ConstVec = N->getOperand(1);
7654  bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7655
7656  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7657      !isConstVecPow2(ConstVec, isSigned, C))
7658    return SDValue();
7659
7660  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7661    Intrinsic::arm_neon_vcvtfxu2fp;
7662  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7663                     Op.getValueType(),
7664                     DAG.getConstant(IntrinsicOpcode, MVT::i32),
7665                     Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7666}
7667
7668/// Getvshiftimm - Check if this is a valid build_vector for the immediate
7669/// operand of a vector shift operation, where all the elements of the
7670/// build_vector must have the same constant integer value.
7671static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7672  // Ignore bit_converts.
7673  while (Op.getOpcode() == ISD::BITCAST)
7674    Op = Op.getOperand(0);
7675  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7676  APInt SplatBits, SplatUndef;
7677  unsigned SplatBitSize;
7678  bool HasAnyUndefs;
7679  if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7680                                      HasAnyUndefs, ElementBits) ||
7681      SplatBitSize > ElementBits)
7682    return false;
7683  Cnt = SplatBits.getSExtValue();
7684  return true;
7685}
7686
7687/// isVShiftLImm - Check if this is a valid build_vector for the immediate
7688/// operand of a vector shift left operation.  That value must be in the range:
7689///   0 <= Value < ElementBits for a left shift; or
7690///   0 <= Value <= ElementBits for a long left shift.
7691static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7692  assert(VT.isVector() && "vector shift count is not a vector type");
7693  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7694  if (! getVShiftImm(Op, ElementBits, Cnt))
7695    return false;
7696  return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7697}
7698
7699/// isVShiftRImm - Check if this is a valid build_vector for the immediate
7700/// operand of a vector shift right operation.  For a shift opcode, the value
7701/// is positive, but for an intrinsic the value count must be negative. The
7702/// absolute value must be in the range:
7703///   1 <= |Value| <= ElementBits for a right shift; or
7704///   1 <= |Value| <= ElementBits/2 for a narrow right shift.
7705static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7706                         int64_t &Cnt) {
7707  assert(VT.isVector() && "vector shift count is not a vector type");
7708  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7709  if (! getVShiftImm(Op, ElementBits, Cnt))
7710    return false;
7711  if (isIntrinsic)
7712    Cnt = -Cnt;
7713  return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7714}
7715
7716/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7717static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7718  unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7719  switch (IntNo) {
7720  default:
7721    // Don't do anything for most intrinsics.
7722    break;
7723
7724  // Vector shifts: check for immediate versions and lower them.
7725  // Note: This is done during DAG combining instead of DAG legalizing because
7726  // the build_vectors for 64-bit vector element shift counts are generally
7727  // not legal, and it is hard to see their values after they get legalized to
7728  // loads from a constant pool.
7729  case Intrinsic::arm_neon_vshifts:
7730  case Intrinsic::arm_neon_vshiftu:
7731  case Intrinsic::arm_neon_vshiftls:
7732  case Intrinsic::arm_neon_vshiftlu:
7733  case Intrinsic::arm_neon_vshiftn:
7734  case Intrinsic::arm_neon_vrshifts:
7735  case Intrinsic::arm_neon_vrshiftu:
7736  case Intrinsic::arm_neon_vrshiftn:
7737  case Intrinsic::arm_neon_vqshifts:
7738  case Intrinsic::arm_neon_vqshiftu:
7739  case Intrinsic::arm_neon_vqshiftsu:
7740  case Intrinsic::arm_neon_vqshiftns:
7741  case Intrinsic::arm_neon_vqshiftnu:
7742  case Intrinsic::arm_neon_vqshiftnsu:
7743  case Intrinsic::arm_neon_vqrshiftns:
7744  case Intrinsic::arm_neon_vqrshiftnu:
7745  case Intrinsic::arm_neon_vqrshiftnsu: {
7746    EVT VT = N->getOperand(1).getValueType();
7747    int64_t Cnt;
7748    unsigned VShiftOpc = 0;
7749
7750    switch (IntNo) {
7751    case Intrinsic::arm_neon_vshifts:
7752    case Intrinsic::arm_neon_vshiftu:
7753      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7754        VShiftOpc = ARMISD::VSHL;
7755        break;
7756      }
7757      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7758        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7759                     ARMISD::VSHRs : ARMISD::VSHRu);
7760        break;
7761      }
7762      return SDValue();
7763
7764    case Intrinsic::arm_neon_vshiftls:
7765    case Intrinsic::arm_neon_vshiftlu:
7766      if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7767        break;
7768      llvm_unreachable("invalid shift count for vshll intrinsic");
7769
7770    case Intrinsic::arm_neon_vrshifts:
7771    case Intrinsic::arm_neon_vrshiftu:
7772      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7773        break;
7774      return SDValue();
7775
7776    case Intrinsic::arm_neon_vqshifts:
7777    case Intrinsic::arm_neon_vqshiftu:
7778      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7779        break;
7780      return SDValue();
7781
7782    case Intrinsic::arm_neon_vqshiftsu:
7783      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7784        break;
7785      llvm_unreachable("invalid shift count for vqshlu intrinsic");
7786
7787    case Intrinsic::arm_neon_vshiftn:
7788    case Intrinsic::arm_neon_vrshiftn:
7789    case Intrinsic::arm_neon_vqshiftns:
7790    case Intrinsic::arm_neon_vqshiftnu:
7791    case Intrinsic::arm_neon_vqshiftnsu:
7792    case Intrinsic::arm_neon_vqrshiftns:
7793    case Intrinsic::arm_neon_vqrshiftnu:
7794    case Intrinsic::arm_neon_vqrshiftnsu:
7795      // Narrowing shifts require an immediate right shift.
7796      if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7797        break;
7798      llvm_unreachable("invalid shift count for narrowing vector shift "
7799                       "intrinsic");
7800
7801    default:
7802      llvm_unreachable("unhandled vector shift");
7803    }
7804
7805    switch (IntNo) {
7806    case Intrinsic::arm_neon_vshifts:
7807    case Intrinsic::arm_neon_vshiftu:
7808      // Opcode already set above.
7809      break;
7810    case Intrinsic::arm_neon_vshiftls:
7811    case Intrinsic::arm_neon_vshiftlu:
7812      if (Cnt == VT.getVectorElementType().getSizeInBits())
7813        VShiftOpc = ARMISD::VSHLLi;
7814      else
7815        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7816                     ARMISD::VSHLLs : ARMISD::VSHLLu);
7817      break;
7818    case Intrinsic::arm_neon_vshiftn:
7819      VShiftOpc = ARMISD::VSHRN; break;
7820    case Intrinsic::arm_neon_vrshifts:
7821      VShiftOpc = ARMISD::VRSHRs; break;
7822    case Intrinsic::arm_neon_vrshiftu:
7823      VShiftOpc = ARMISD::VRSHRu; break;
7824    case Intrinsic::arm_neon_vrshiftn:
7825      VShiftOpc = ARMISD::VRSHRN; break;
7826    case Intrinsic::arm_neon_vqshifts:
7827      VShiftOpc = ARMISD::VQSHLs; break;
7828    case Intrinsic::arm_neon_vqshiftu:
7829      VShiftOpc = ARMISD::VQSHLu; break;
7830    case Intrinsic::arm_neon_vqshiftsu:
7831      VShiftOpc = ARMISD::VQSHLsu; break;
7832    case Intrinsic::arm_neon_vqshiftns:
7833      VShiftOpc = ARMISD::VQSHRNs; break;
7834    case Intrinsic::arm_neon_vqshiftnu:
7835      VShiftOpc = ARMISD::VQSHRNu; break;
7836    case Intrinsic::arm_neon_vqshiftnsu:
7837      VShiftOpc = ARMISD::VQSHRNsu; break;
7838    case Intrinsic::arm_neon_vqrshiftns:
7839      VShiftOpc = ARMISD::VQRSHRNs; break;
7840    case Intrinsic::arm_neon_vqrshiftnu:
7841      VShiftOpc = ARMISD::VQRSHRNu; break;
7842    case Intrinsic::arm_neon_vqrshiftnsu:
7843      VShiftOpc = ARMISD::VQRSHRNsu; break;
7844    }
7845
7846    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7847                       N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7848  }
7849
7850  case Intrinsic::arm_neon_vshiftins: {
7851    EVT VT = N->getOperand(1).getValueType();
7852    int64_t Cnt;
7853    unsigned VShiftOpc = 0;
7854
7855    if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7856      VShiftOpc = ARMISD::VSLI;
7857    else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7858      VShiftOpc = ARMISD::VSRI;
7859    else {
7860      llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7861    }
7862
7863    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7864                       N->getOperand(1), N->getOperand(2),
7865                       DAG.getConstant(Cnt, MVT::i32));
7866  }
7867
7868  case Intrinsic::arm_neon_vqrshifts:
7869  case Intrinsic::arm_neon_vqrshiftu:
7870    // No immediate versions of these to check for.
7871    break;
7872  }
7873
7874  return SDValue();
7875}
7876
7877/// PerformShiftCombine - Checks for immediate versions of vector shifts and
7878/// lowers them.  As with the vector shift intrinsics, this is done during DAG
7879/// combining instead of DAG legalizing because the build_vectors for 64-bit
7880/// vector element shift counts are generally not legal, and it is hard to see
7881/// their values after they get legalized to loads from a constant pool.
7882static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7883                                   const ARMSubtarget *ST) {
7884  EVT VT = N->getValueType(0);
7885
7886  // Nothing to be done for scalar shifts.
7887  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7888  if (!VT.isVector() || !TLI.isTypeLegal(VT))
7889    return SDValue();
7890
7891  assert(ST->hasNEON() && "unexpected vector shift");
7892  int64_t Cnt;
7893
7894  switch (N->getOpcode()) {
7895  default: llvm_unreachable("unexpected shift opcode");
7896
7897  case ISD::SHL:
7898    if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7899      return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7900                         DAG.getConstant(Cnt, MVT::i32));
7901    break;
7902
7903  case ISD::SRA:
7904  case ISD::SRL:
7905    if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7906      unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7907                            ARMISD::VSHRs : ARMISD::VSHRu);
7908      return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7909                         DAG.getConstant(Cnt, MVT::i32));
7910    }
7911  }
7912  return SDValue();
7913}
7914
7915/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7916/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7917static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7918                                    const ARMSubtarget *ST) {
7919  SDValue N0 = N->getOperand(0);
7920
7921  // Check for sign- and zero-extensions of vector extract operations of 8-
7922  // and 16-bit vector elements.  NEON supports these directly.  They are
7923  // handled during DAG combining because type legalization will promote them
7924  // to 32-bit types and it is messy to recognize the operations after that.
7925  if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7926    SDValue Vec = N0.getOperand(0);
7927    SDValue Lane = N0.getOperand(1);
7928    EVT VT = N->getValueType(0);
7929    EVT EltVT = N0.getValueType();
7930    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7931
7932    if (VT == MVT::i32 &&
7933        (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7934        TLI.isTypeLegal(Vec.getValueType()) &&
7935        isa<ConstantSDNode>(Lane)) {
7936
7937      unsigned Opc = 0;
7938      switch (N->getOpcode()) {
7939      default: llvm_unreachable("unexpected opcode");
7940      case ISD::SIGN_EXTEND:
7941        Opc = ARMISD::VGETLANEs;
7942        break;
7943      case ISD::ZERO_EXTEND:
7944      case ISD::ANY_EXTEND:
7945        Opc = ARMISD::VGETLANEu;
7946        break;
7947      }
7948      return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7949    }
7950  }
7951
7952  return SDValue();
7953}
7954
7955/// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7956/// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7957static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7958                                       const ARMSubtarget *ST) {
7959  // If the target supports NEON, try to use vmax/vmin instructions for f32
7960  // selects like "x < y ? x : y".  Unless the NoNaNsFPMath option is set,
7961  // be careful about NaNs:  NEON's vmax/vmin return NaN if either operand is
7962  // a NaN; only do the transformation when it matches that behavior.
7963
7964  // For now only do this when using NEON for FP operations; if using VFP, it
7965  // is not obvious that the benefit outweighs the cost of switching to the
7966  // NEON pipeline.
7967  if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7968      N->getValueType(0) != MVT::f32)
7969    return SDValue();
7970
7971  SDValue CondLHS = N->getOperand(0);
7972  SDValue CondRHS = N->getOperand(1);
7973  SDValue LHS = N->getOperand(2);
7974  SDValue RHS = N->getOperand(3);
7975  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7976
7977  unsigned Opcode = 0;
7978  bool IsReversed;
7979  if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7980    IsReversed = false; // x CC y ? x : y
7981  } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7982    IsReversed = true ; // x CC y ? y : x
7983  } else {
7984    return SDValue();
7985  }
7986
7987  bool IsUnordered;
7988  switch (CC) {
7989  default: break;
7990  case ISD::SETOLT:
7991  case ISD::SETOLE:
7992  case ISD::SETLT:
7993  case ISD::SETLE:
7994  case ISD::SETULT:
7995  case ISD::SETULE:
7996    // If LHS is NaN, an ordered comparison will be false and the result will
7997    // be the RHS, but vmin(NaN, RHS) = NaN.  Avoid this by checking that LHS
7998    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
7999    IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8000    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8001      break;
8002    // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8003    // will return -0, so vmin can only be used for unsafe math or if one of
8004    // the operands is known to be nonzero.
8005    if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8006        !DAG.getTarget().Options.UnsafeFPMath &&
8007        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8008      break;
8009    Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8010    break;
8011
8012  case ISD::SETOGT:
8013  case ISD::SETOGE:
8014  case ISD::SETGT:
8015  case ISD::SETGE:
8016  case ISD::SETUGT:
8017  case ISD::SETUGE:
8018    // If LHS is NaN, an ordered comparison will be false and the result will
8019    // be the RHS, but vmax(NaN, RHS) = NaN.  Avoid this by checking that LHS
8020    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
8021    IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8022    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8023      break;
8024    // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8025    // will return +0, so vmax can only be used for unsafe math or if one of
8026    // the operands is known to be nonzero.
8027    if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8028        !DAG.getTarget().Options.UnsafeFPMath &&
8029        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8030      break;
8031    Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8032    break;
8033  }
8034
8035  if (!Opcode)
8036    return SDValue();
8037  return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8038}
8039
8040/// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8041SDValue
8042ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8043  SDValue Cmp = N->getOperand(4);
8044  if (Cmp.getOpcode() != ARMISD::CMPZ)
8045    // Only looking at EQ and NE cases.
8046    return SDValue();
8047
8048  EVT VT = N->getValueType(0);
8049  DebugLoc dl = N->getDebugLoc();
8050  SDValue LHS = Cmp.getOperand(0);
8051  SDValue RHS = Cmp.getOperand(1);
8052  SDValue FalseVal = N->getOperand(0);
8053  SDValue TrueVal = N->getOperand(1);
8054  SDValue ARMcc = N->getOperand(2);
8055  ARMCC::CondCodes CC =
8056    (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8057
8058  // Simplify
8059  //   mov     r1, r0
8060  //   cmp     r1, x
8061  //   mov     r0, y
8062  //   moveq   r0, x
8063  // to
8064  //   cmp     r0, x
8065  //   movne   r0, y
8066  //
8067  //   mov     r1, r0
8068  //   cmp     r1, x
8069  //   mov     r0, x
8070  //   movne   r0, y
8071  // to
8072  //   cmp     r0, x
8073  //   movne   r0, y
8074  /// FIXME: Turn this into a target neutral optimization?
8075  SDValue Res;
8076  if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8077    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8078                      N->getOperand(3), Cmp);
8079  } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8080    SDValue ARMcc;
8081    SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8082    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8083                      N->getOperand(3), NewCmp);
8084  }
8085
8086  if (Res.getNode()) {
8087    APInt KnownZero, KnownOne;
8088    APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8089    DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8090    // Capture demanded bits information that would be otherwise lost.
8091    if (KnownZero == 0xfffffffe)
8092      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8093                        DAG.getValueType(MVT::i1));
8094    else if (KnownZero == 0xffffff00)
8095      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8096                        DAG.getValueType(MVT::i8));
8097    else if (KnownZero == 0xffff0000)
8098      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8099                        DAG.getValueType(MVT::i16));
8100  }
8101
8102  return Res;
8103}
8104
8105SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8106                                             DAGCombinerInfo &DCI) const {
8107  switch (N->getOpcode()) {
8108  default: break;
8109  case ISD::ADD:        return PerformADDCombine(N, DCI, Subtarget);
8110  case ISD::SUB:        return PerformSUBCombine(N, DCI);
8111  case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
8112  case ISD::OR:         return PerformORCombine(N, DCI, Subtarget);
8113  case ISD::AND:        return PerformANDCombine(N, DCI);
8114  case ARMISD::BFI:     return PerformBFICombine(N, DCI);
8115  case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8116  case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8117  case ISD::STORE:      return PerformSTORECombine(N, DCI);
8118  case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8119  case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8120  case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8121  case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8122  case ISD::FP_TO_SINT:
8123  case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8124  case ISD::FDIV:       return PerformVDIVCombine(N, DCI, Subtarget);
8125  case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8126  case ISD::SHL:
8127  case ISD::SRA:
8128  case ISD::SRL:        return PerformShiftCombine(N, DCI.DAG, Subtarget);
8129  case ISD::SIGN_EXTEND:
8130  case ISD::ZERO_EXTEND:
8131  case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8132  case ISD::SELECT_CC:  return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8133  case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8134  case ARMISD::VLD2DUP:
8135  case ARMISD::VLD3DUP:
8136  case ARMISD::VLD4DUP:
8137    return CombineBaseUpdate(N, DCI);
8138  case ISD::INTRINSIC_VOID:
8139  case ISD::INTRINSIC_W_CHAIN:
8140    switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8141    case Intrinsic::arm_neon_vld1:
8142    case Intrinsic::arm_neon_vld2:
8143    case Intrinsic::arm_neon_vld3:
8144    case Intrinsic::arm_neon_vld4:
8145    case Intrinsic::arm_neon_vld2lane:
8146    case Intrinsic::arm_neon_vld3lane:
8147    case Intrinsic::arm_neon_vld4lane:
8148    case Intrinsic::arm_neon_vst1:
8149    case Intrinsic::arm_neon_vst2:
8150    case Intrinsic::arm_neon_vst3:
8151    case Intrinsic::arm_neon_vst4:
8152    case Intrinsic::arm_neon_vst2lane:
8153    case Intrinsic::arm_neon_vst3lane:
8154    case Intrinsic::arm_neon_vst4lane:
8155      return CombineBaseUpdate(N, DCI);
8156    default: break;
8157    }
8158    break;
8159  }
8160  return SDValue();
8161}
8162
8163bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8164                                                          EVT VT) const {
8165  return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8166}
8167
8168bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8169  if (!Subtarget->allowsUnalignedMem())
8170    return false;
8171
8172  switch (VT.getSimpleVT().SimpleTy) {
8173  default:
8174    return false;
8175  case MVT::i8:
8176  case MVT::i16:
8177  case MVT::i32:
8178    return true;
8179  // FIXME: VLD1 etc with standard alignment is legal.
8180  }
8181}
8182
8183static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8184                       unsigned AlignCheck) {
8185  return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8186          (DstAlign == 0 || DstAlign % AlignCheck == 0));
8187}
8188
8189EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8190                                           unsigned DstAlign, unsigned SrcAlign,
8191                                           bool IsZeroVal,
8192                                           bool MemcpyStrSrc,
8193                                           MachineFunction &MF) const {
8194  const Function *F = MF.getFunction();
8195
8196  // See if we can use NEON instructions for this...
8197  if (IsZeroVal &&
8198      !F->hasFnAttr(Attribute::NoImplicitFloat) &&
8199      Subtarget->hasNEON()) {
8200    if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8201      return MVT::v4i32;
8202    } else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8203      return MVT::v2i32;
8204    }
8205  }
8206
8207  // Lowering to i32/i16 if the size permits.
8208  if (Size >= 4) {
8209    return MVT::i32;
8210  } else if (Size >= 2) {
8211    return MVT::i16;
8212  }
8213
8214  // Let the target-independent logic figure it out.
8215  return MVT::Other;
8216}
8217
8218static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8219  if (V < 0)
8220    return false;
8221
8222  unsigned Scale = 1;
8223  switch (VT.getSimpleVT().SimpleTy) {
8224  default: return false;
8225  case MVT::i1:
8226  case MVT::i8:
8227    // Scale == 1;
8228    break;
8229  case MVT::i16:
8230    // Scale == 2;
8231    Scale = 2;
8232    break;
8233  case MVT::i32:
8234    // Scale == 4;
8235    Scale = 4;
8236    break;
8237  }
8238
8239  if ((V & (Scale - 1)) != 0)
8240    return false;
8241  V /= Scale;
8242  return V == (V & ((1LL << 5) - 1));
8243}
8244
8245static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8246                                      const ARMSubtarget *Subtarget) {
8247  bool isNeg = false;
8248  if (V < 0) {
8249    isNeg = true;
8250    V = - V;
8251  }
8252
8253  switch (VT.getSimpleVT().SimpleTy) {
8254  default: return false;
8255  case MVT::i1:
8256  case MVT::i8:
8257  case MVT::i16:
8258  case MVT::i32:
8259    // + imm12 or - imm8
8260    if (isNeg)
8261      return V == (V & ((1LL << 8) - 1));
8262    return V == (V & ((1LL << 12) - 1));
8263  case MVT::f32:
8264  case MVT::f64:
8265    // Same as ARM mode. FIXME: NEON?
8266    if (!Subtarget->hasVFP2())
8267      return false;
8268    if ((V & 3) != 0)
8269      return false;
8270    V >>= 2;
8271    return V == (V & ((1LL << 8) - 1));
8272  }
8273}
8274
8275/// isLegalAddressImmediate - Return true if the integer value can be used
8276/// as the offset of the target addressing mode for load / store of the
8277/// given type.
8278static bool isLegalAddressImmediate(int64_t V, EVT VT,
8279                                    const ARMSubtarget *Subtarget) {
8280  if (V == 0)
8281    return true;
8282
8283  if (!VT.isSimple())
8284    return false;
8285
8286  if (Subtarget->isThumb1Only())
8287    return isLegalT1AddressImmediate(V, VT);
8288  else if (Subtarget->isThumb2())
8289    return isLegalT2AddressImmediate(V, VT, Subtarget);
8290
8291  // ARM mode.
8292  if (V < 0)
8293    V = - V;
8294  switch (VT.getSimpleVT().SimpleTy) {
8295  default: return false;
8296  case MVT::i1:
8297  case MVT::i8:
8298  case MVT::i32:
8299    // +- imm12
8300    return V == (V & ((1LL << 12) - 1));
8301  case MVT::i16:
8302    // +- imm8
8303    return V == (V & ((1LL << 8) - 1));
8304  case MVT::f32:
8305  case MVT::f64:
8306    if (!Subtarget->hasVFP2()) // FIXME: NEON?
8307      return false;
8308    if ((V & 3) != 0)
8309      return false;
8310    V >>= 2;
8311    return V == (V & ((1LL << 8) - 1));
8312  }
8313}
8314
8315bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8316                                                      EVT VT) const {
8317  int Scale = AM.Scale;
8318  if (Scale < 0)
8319    return false;
8320
8321  switch (VT.getSimpleVT().SimpleTy) {
8322  default: return false;
8323  case MVT::i1:
8324  case MVT::i8:
8325  case MVT::i16:
8326  case MVT::i32:
8327    if (Scale == 1)
8328      return true;
8329    // r + r << imm
8330    Scale = Scale & ~1;
8331    return Scale == 2 || Scale == 4 || Scale == 8;
8332  case MVT::i64:
8333    // r + r
8334    if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8335      return true;
8336    return false;
8337  case MVT::isVoid:
8338    // Note, we allow "void" uses (basically, uses that aren't loads or
8339    // stores), because arm allows folding a scale into many arithmetic
8340    // operations.  This should be made more precise and revisited later.
8341
8342    // Allow r << imm, but the imm has to be a multiple of two.
8343    if (Scale & 1) return false;
8344    return isPowerOf2_32(Scale);
8345  }
8346}
8347
8348/// isLegalAddressingMode - Return true if the addressing mode represented
8349/// by AM is legal for this target, for a load/store of the specified type.
8350bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8351                                              Type *Ty) const {
8352  EVT VT = getValueType(Ty, true);
8353  if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8354    return false;
8355
8356  // Can never fold addr of global into load/store.
8357  if (AM.BaseGV)
8358    return false;
8359
8360  switch (AM.Scale) {
8361  case 0:  // no scale reg, must be "r+i" or "r", or "i".
8362    break;
8363  case 1:
8364    if (Subtarget->isThumb1Only())
8365      return false;
8366    // FALL THROUGH.
8367  default:
8368    // ARM doesn't support any R+R*scale+imm addr modes.
8369    if (AM.BaseOffs)
8370      return false;
8371
8372    if (!VT.isSimple())
8373      return false;
8374
8375    if (Subtarget->isThumb2())
8376      return isLegalT2ScaledAddressingMode(AM, VT);
8377
8378    int Scale = AM.Scale;
8379    switch (VT.getSimpleVT().SimpleTy) {
8380    default: return false;
8381    case MVT::i1:
8382    case MVT::i8:
8383    case MVT::i32:
8384      if (Scale < 0) Scale = -Scale;
8385      if (Scale == 1)
8386        return true;
8387      // r + r << imm
8388      return isPowerOf2_32(Scale & ~1);
8389    case MVT::i16:
8390    case MVT::i64:
8391      // r + r
8392      if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8393        return true;
8394      return false;
8395
8396    case MVT::isVoid:
8397      // Note, we allow "void" uses (basically, uses that aren't loads or
8398      // stores), because arm allows folding a scale into many arithmetic
8399      // operations.  This should be made more precise and revisited later.
8400
8401      // Allow r << imm, but the imm has to be a multiple of two.
8402      if (Scale & 1) return false;
8403      return isPowerOf2_32(Scale);
8404    }
8405  }
8406  return true;
8407}
8408
8409/// isLegalICmpImmediate - Return true if the specified immediate is legal
8410/// icmp immediate, that is the target has icmp instructions which can compare
8411/// a register against the immediate without having to materialize the
8412/// immediate into a register.
8413bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8414  if (!Subtarget->isThumb())
8415    return ARM_AM::getSOImmVal(Imm) != -1;
8416  if (Subtarget->isThumb2())
8417    return ARM_AM::getT2SOImmVal(Imm) != -1;
8418  return Imm >= 0 && Imm <= 255;
8419}
8420
8421/// isLegalAddImmediate - Return true if the specified immediate is legal
8422/// add immediate, that is the target has add instructions which can add
8423/// a register with the immediate without having to materialize the
8424/// immediate into a register.
8425bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8426  return ARM_AM::getSOImmVal(Imm) != -1;
8427}
8428
8429static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8430                                      bool isSEXTLoad, SDValue &Base,
8431                                      SDValue &Offset, bool &isInc,
8432                                      SelectionDAG &DAG) {
8433  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8434    return false;
8435
8436  if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8437    // AddressingMode 3
8438    Base = Ptr->getOperand(0);
8439    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8440      int RHSC = (int)RHS->getZExtValue();
8441      if (RHSC < 0 && RHSC > -256) {
8442        assert(Ptr->getOpcode() == ISD::ADD);
8443        isInc = false;
8444        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8445        return true;
8446      }
8447    }
8448    isInc = (Ptr->getOpcode() == ISD::ADD);
8449    Offset = Ptr->getOperand(1);
8450    return true;
8451  } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8452    // AddressingMode 2
8453    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8454      int RHSC = (int)RHS->getZExtValue();
8455      if (RHSC < 0 && RHSC > -0x1000) {
8456        assert(Ptr->getOpcode() == ISD::ADD);
8457        isInc = false;
8458        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8459        Base = Ptr->getOperand(0);
8460        return true;
8461      }
8462    }
8463
8464    if (Ptr->getOpcode() == ISD::ADD) {
8465      isInc = true;
8466      ARM_AM::ShiftOpc ShOpcVal=
8467        ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8468      if (ShOpcVal != ARM_AM::no_shift) {
8469        Base = Ptr->getOperand(1);
8470        Offset = Ptr->getOperand(0);
8471      } else {
8472        Base = Ptr->getOperand(0);
8473        Offset = Ptr->getOperand(1);
8474      }
8475      return true;
8476    }
8477
8478    isInc = (Ptr->getOpcode() == ISD::ADD);
8479    Base = Ptr->getOperand(0);
8480    Offset = Ptr->getOperand(1);
8481    return true;
8482  }
8483
8484  // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8485  return false;
8486}
8487
8488static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8489                                     bool isSEXTLoad, SDValue &Base,
8490                                     SDValue &Offset, bool &isInc,
8491                                     SelectionDAG &DAG) {
8492  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8493    return false;
8494
8495  Base = Ptr->getOperand(0);
8496  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8497    int RHSC = (int)RHS->getZExtValue();
8498    if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8499      assert(Ptr->getOpcode() == ISD::ADD);
8500      isInc = false;
8501      Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8502      return true;
8503    } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8504      isInc = Ptr->getOpcode() == ISD::ADD;
8505      Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8506      return true;
8507    }
8508  }
8509
8510  return false;
8511}
8512
8513/// getPreIndexedAddressParts - returns true by value, base pointer and
8514/// offset pointer and addressing mode by reference if the node's address
8515/// can be legally represented as pre-indexed load / store address.
8516bool
8517ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8518                                             SDValue &Offset,
8519                                             ISD::MemIndexedMode &AM,
8520                                             SelectionDAG &DAG) const {
8521  if (Subtarget->isThumb1Only())
8522    return false;
8523
8524  EVT VT;
8525  SDValue Ptr;
8526  bool isSEXTLoad = false;
8527  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8528    Ptr = LD->getBasePtr();
8529    VT  = LD->getMemoryVT();
8530    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8531  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8532    Ptr = ST->getBasePtr();
8533    VT  = ST->getMemoryVT();
8534  } else
8535    return false;
8536
8537  bool isInc;
8538  bool isLegal = false;
8539  if (Subtarget->isThumb2())
8540    isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8541                                       Offset, isInc, DAG);
8542  else
8543    isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8544                                        Offset, isInc, DAG);
8545  if (!isLegal)
8546    return false;
8547
8548  AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8549  return true;
8550}
8551
8552/// getPostIndexedAddressParts - returns true by value, base pointer and
8553/// offset pointer and addressing mode by reference if this node can be
8554/// combined with a load / store to form a post-indexed load / store.
8555bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8556                                                   SDValue &Base,
8557                                                   SDValue &Offset,
8558                                                   ISD::MemIndexedMode &AM,
8559                                                   SelectionDAG &DAG) const {
8560  if (Subtarget->isThumb1Only())
8561    return false;
8562
8563  EVT VT;
8564  SDValue Ptr;
8565  bool isSEXTLoad = false;
8566  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8567    VT  = LD->getMemoryVT();
8568    Ptr = LD->getBasePtr();
8569    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8570  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8571    VT  = ST->getMemoryVT();
8572    Ptr = ST->getBasePtr();
8573  } else
8574    return false;
8575
8576  bool isInc;
8577  bool isLegal = false;
8578  if (Subtarget->isThumb2())
8579    isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8580                                       isInc, DAG);
8581  else
8582    isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8583                                        isInc, DAG);
8584  if (!isLegal)
8585    return false;
8586
8587  if (Ptr != Base) {
8588    // Swap base ptr and offset to catch more post-index load / store when
8589    // it's legal. In Thumb2 mode, offset must be an immediate.
8590    if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8591        !Subtarget->isThumb2())
8592      std::swap(Base, Offset);
8593
8594    // Post-indexed load / store update the base pointer.
8595    if (Ptr != Base)
8596      return false;
8597  }
8598
8599  AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8600  return true;
8601}
8602
8603void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8604                                                       const APInt &Mask,
8605                                                       APInt &KnownZero,
8606                                                       APInt &KnownOne,
8607                                                       const SelectionDAG &DAG,
8608                                                       unsigned Depth) const {
8609  KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8610  switch (Op.getOpcode()) {
8611  default: break;
8612  case ARMISD::CMOV: {
8613    // Bits are known zero/one if known on the LHS and RHS.
8614    DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8615    if (KnownZero == 0 && KnownOne == 0) return;
8616
8617    APInt KnownZeroRHS, KnownOneRHS;
8618    DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8619                          KnownZeroRHS, KnownOneRHS, Depth+1);
8620    KnownZero &= KnownZeroRHS;
8621    KnownOne  &= KnownOneRHS;
8622    return;
8623  }
8624  }
8625}
8626
8627//===----------------------------------------------------------------------===//
8628//                           ARM Inline Assembly Support
8629//===----------------------------------------------------------------------===//
8630
8631bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8632  // Looking for "rev" which is V6+.
8633  if (!Subtarget->hasV6Ops())
8634    return false;
8635
8636  InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8637  std::string AsmStr = IA->getAsmString();
8638  SmallVector<StringRef, 4> AsmPieces;
8639  SplitString(AsmStr, AsmPieces, ";\n");
8640
8641  switch (AsmPieces.size()) {
8642  default: return false;
8643  case 1:
8644    AsmStr = AsmPieces[0];
8645    AsmPieces.clear();
8646    SplitString(AsmStr, AsmPieces, " \t,");
8647
8648    // rev $0, $1
8649    if (AsmPieces.size() == 3 &&
8650        AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8651        IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8652      IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8653      if (Ty && Ty->getBitWidth() == 32)
8654        return IntrinsicLowering::LowerToByteSwap(CI);
8655    }
8656    break;
8657  }
8658
8659  return false;
8660}
8661
8662/// getConstraintType - Given a constraint letter, return the type of
8663/// constraint it is for this target.
8664ARMTargetLowering::ConstraintType
8665ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8666  if (Constraint.size() == 1) {
8667    switch (Constraint[0]) {
8668    default:  break;
8669    case 'l': return C_RegisterClass;
8670    case 'w': return C_RegisterClass;
8671    case 'h': return C_RegisterClass;
8672    case 'x': return C_RegisterClass;
8673    case 't': return C_RegisterClass;
8674    case 'j': return C_Other; // Constant for movw.
8675      // An address with a single base register. Due to the way we
8676      // currently handle addresses it is the same as an 'r' memory constraint.
8677    case 'Q': return C_Memory;
8678    }
8679  } else if (Constraint.size() == 2) {
8680    switch (Constraint[0]) {
8681    default: break;
8682    // All 'U+' constraints are addresses.
8683    case 'U': return C_Memory;
8684    }
8685  }
8686  return TargetLowering::getConstraintType(Constraint);
8687}
8688
8689/// Examine constraint type and operand type and determine a weight value.
8690/// This object must already have been set up with the operand type
8691/// and the current alternative constraint selected.
8692TargetLowering::ConstraintWeight
8693ARMTargetLowering::getSingleConstraintMatchWeight(
8694    AsmOperandInfo &info, const char *constraint) const {
8695  ConstraintWeight weight = CW_Invalid;
8696  Value *CallOperandVal = info.CallOperandVal;
8697    // If we don't have a value, we can't do a match,
8698    // but allow it at the lowest weight.
8699  if (CallOperandVal == NULL)
8700    return CW_Default;
8701  Type *type = CallOperandVal->getType();
8702  // Look at the constraint type.
8703  switch (*constraint) {
8704  default:
8705    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8706    break;
8707  case 'l':
8708    if (type->isIntegerTy()) {
8709      if (Subtarget->isThumb())
8710        weight = CW_SpecificReg;
8711      else
8712        weight = CW_Register;
8713    }
8714    break;
8715  case 'w':
8716    if (type->isFloatingPointTy())
8717      weight = CW_Register;
8718    break;
8719  }
8720  return weight;
8721}
8722
8723typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8724RCPair
8725ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8726                                                EVT VT) const {
8727  if (Constraint.size() == 1) {
8728    // GCC ARM Constraint Letters
8729    switch (Constraint[0]) {
8730    case 'l': // Low regs or general regs.
8731      if (Subtarget->isThumb())
8732        return RCPair(0U, ARM::tGPRRegisterClass);
8733      else
8734        return RCPair(0U, ARM::GPRRegisterClass);
8735    case 'h': // High regs or no regs.
8736      if (Subtarget->isThumb())
8737        return RCPair(0U, ARM::hGPRRegisterClass);
8738      break;
8739    case 'r':
8740      return RCPair(0U, ARM::GPRRegisterClass);
8741    case 'w':
8742      if (VT == MVT::f32)
8743        return RCPair(0U, ARM::SPRRegisterClass);
8744      if (VT.getSizeInBits() == 64)
8745        return RCPair(0U, ARM::DPRRegisterClass);
8746      if (VT.getSizeInBits() == 128)
8747        return RCPair(0U, ARM::QPRRegisterClass);
8748      break;
8749    case 'x':
8750      if (VT == MVT::f32)
8751        return RCPair(0U, ARM::SPR_8RegisterClass);
8752      if (VT.getSizeInBits() == 64)
8753        return RCPair(0U, ARM::DPR_8RegisterClass);
8754      if (VT.getSizeInBits() == 128)
8755        return RCPair(0U, ARM::QPR_8RegisterClass);
8756      break;
8757    case 't':
8758      if (VT == MVT::f32)
8759        return RCPair(0U, ARM::SPRRegisterClass);
8760      break;
8761    }
8762  }
8763  if (StringRef("{cc}").equals_lower(Constraint))
8764    return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8765
8766  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8767}
8768
8769/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8770/// vector.  If it is invalid, don't add anything to Ops.
8771void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8772                                                     std::string &Constraint,
8773                                                     std::vector<SDValue>&Ops,
8774                                                     SelectionDAG &DAG) const {
8775  SDValue Result(0, 0);
8776
8777  // Currently only support length 1 constraints.
8778  if (Constraint.length() != 1) return;
8779
8780  char ConstraintLetter = Constraint[0];
8781  switch (ConstraintLetter) {
8782  default: break;
8783  case 'j':
8784  case 'I': case 'J': case 'K': case 'L':
8785  case 'M': case 'N': case 'O':
8786    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8787    if (!C)
8788      return;
8789
8790    int64_t CVal64 = C->getSExtValue();
8791    int CVal = (int) CVal64;
8792    // None of these constraints allow values larger than 32 bits.  Check
8793    // that the value fits in an int.
8794    if (CVal != CVal64)
8795      return;
8796
8797    switch (ConstraintLetter) {
8798      case 'j':
8799        // Constant suitable for movw, must be between 0 and
8800        // 65535.
8801        if (Subtarget->hasV6T2Ops())
8802          if (CVal >= 0 && CVal <= 65535)
8803            break;
8804        return;
8805      case 'I':
8806        if (Subtarget->isThumb1Only()) {
8807          // This must be a constant between 0 and 255, for ADD
8808          // immediates.
8809          if (CVal >= 0 && CVal <= 255)
8810            break;
8811        } else if (Subtarget->isThumb2()) {
8812          // A constant that can be used as an immediate value in a
8813          // data-processing instruction.
8814          if (ARM_AM::getT2SOImmVal(CVal) != -1)
8815            break;
8816        } else {
8817          // A constant that can be used as an immediate value in a
8818          // data-processing instruction.
8819          if (ARM_AM::getSOImmVal(CVal) != -1)
8820            break;
8821        }
8822        return;
8823
8824      case 'J':
8825        if (Subtarget->isThumb()) {  // FIXME thumb2
8826          // This must be a constant between -255 and -1, for negated ADD
8827          // immediates. This can be used in GCC with an "n" modifier that
8828          // prints the negated value, for use with SUB instructions. It is
8829          // not useful otherwise but is implemented for compatibility.
8830          if (CVal >= -255 && CVal <= -1)
8831            break;
8832        } else {
8833          // This must be a constant between -4095 and 4095. It is not clear
8834          // what this constraint is intended for. Implemented for
8835          // compatibility with GCC.
8836          if (CVal >= -4095 && CVal <= 4095)
8837            break;
8838        }
8839        return;
8840
8841      case 'K':
8842        if (Subtarget->isThumb1Only()) {
8843          // A 32-bit value where only one byte has a nonzero value. Exclude
8844          // zero to match GCC. This constraint is used by GCC internally for
8845          // constants that can be loaded with a move/shift combination.
8846          // It is not useful otherwise but is implemented for compatibility.
8847          if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8848            break;
8849        } else if (Subtarget->isThumb2()) {
8850          // A constant whose bitwise inverse can be used as an immediate
8851          // value in a data-processing instruction. This can be used in GCC
8852          // with a "B" modifier that prints the inverted value, for use with
8853          // BIC and MVN instructions. It is not useful otherwise but is
8854          // implemented for compatibility.
8855          if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8856            break;
8857        } else {
8858          // A constant whose bitwise inverse can be used as an immediate
8859          // value in a data-processing instruction. This can be used in GCC
8860          // with a "B" modifier that prints the inverted value, for use with
8861          // BIC and MVN instructions. It is not useful otherwise but is
8862          // implemented for compatibility.
8863          if (ARM_AM::getSOImmVal(~CVal) != -1)
8864            break;
8865        }
8866        return;
8867
8868      case 'L':
8869        if (Subtarget->isThumb1Only()) {
8870          // This must be a constant between -7 and 7,
8871          // for 3-operand ADD/SUB immediate instructions.
8872          if (CVal >= -7 && CVal < 7)
8873            break;
8874        } else if (Subtarget->isThumb2()) {
8875          // A constant whose negation can be used as an immediate value in a
8876          // data-processing instruction. This can be used in GCC with an "n"
8877          // modifier that prints the negated value, for use with SUB
8878          // instructions. It is not useful otherwise but is implemented for
8879          // compatibility.
8880          if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8881            break;
8882        } else {
8883          // A constant whose negation can be used as an immediate value in a
8884          // data-processing instruction. This can be used in GCC with an "n"
8885          // modifier that prints the negated value, for use with SUB
8886          // instructions. It is not useful otherwise but is implemented for
8887          // compatibility.
8888          if (ARM_AM::getSOImmVal(-CVal) != -1)
8889            break;
8890        }
8891        return;
8892
8893      case 'M':
8894        if (Subtarget->isThumb()) { // FIXME thumb2
8895          // This must be a multiple of 4 between 0 and 1020, for
8896          // ADD sp + immediate.
8897          if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
8898            break;
8899        } else {
8900          // A power of two or a constant between 0 and 32.  This is used in
8901          // GCC for the shift amount on shifted register operands, but it is
8902          // useful in general for any shift amounts.
8903          if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8904            break;
8905        }
8906        return;
8907
8908      case 'N':
8909        if (Subtarget->isThumb()) {  // FIXME thumb2
8910          // This must be a constant between 0 and 31, for shift amounts.
8911          if (CVal >= 0 && CVal <= 31)
8912            break;
8913        }
8914        return;
8915
8916      case 'O':
8917        if (Subtarget->isThumb()) {  // FIXME thumb2
8918          // This must be a multiple of 4 between -508 and 508, for
8919          // ADD/SUB sp = sp + immediate.
8920          if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8921            break;
8922        }
8923        return;
8924    }
8925    Result = DAG.getTargetConstant(CVal, Op.getValueType());
8926    break;
8927  }
8928
8929  if (Result.getNode()) {
8930    Ops.push_back(Result);
8931    return;
8932  }
8933  return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8934}
8935
8936bool
8937ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8938  // The ARM target isn't yet aware of offsets.
8939  return false;
8940}
8941
8942bool ARM::isBitFieldInvertedMask(unsigned v) {
8943  if (v == 0xffffffff)
8944    return 0;
8945  // there can be 1's on either or both "outsides", all the "inside"
8946  // bits must be 0's
8947  unsigned int lsb = 0, msb = 31;
8948  while (v & (1 << msb)) --msb;
8949  while (v & (1 << lsb)) ++lsb;
8950  for (unsigned int i = lsb; i <= msb; ++i) {
8951    if (v & (1 << i))
8952      return 0;
8953  }
8954  return 1;
8955}
8956
8957/// isFPImmLegal - Returns true if the target can instruction select the
8958/// specified FP immediate natively. If false, the legalizer will
8959/// materialize the FP immediate as a load from a constant pool.
8960bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8961  if (!Subtarget->hasVFP3())
8962    return false;
8963  if (VT == MVT::f32)
8964    return ARM_AM::getFP32Imm(Imm) != -1;
8965  if (VT == MVT::f64)
8966    return ARM_AM::getFP64Imm(Imm) != -1;
8967  return false;
8968}
8969
8970/// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8971/// MemIntrinsicNodes.  The associated MachineMemOperands record the alignment
8972/// specified in the intrinsic calls.
8973bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8974                                           const CallInst &I,
8975                                           unsigned Intrinsic) const {
8976  switch (Intrinsic) {
8977  case Intrinsic::arm_neon_vld1:
8978  case Intrinsic::arm_neon_vld2:
8979  case Intrinsic::arm_neon_vld3:
8980  case Intrinsic::arm_neon_vld4:
8981  case Intrinsic::arm_neon_vld2lane:
8982  case Intrinsic::arm_neon_vld3lane:
8983  case Intrinsic::arm_neon_vld4lane: {
8984    Info.opc = ISD::INTRINSIC_W_CHAIN;
8985    // Conservatively set memVT to the entire set of vectors loaded.
8986    uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8987    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8988    Info.ptrVal = I.getArgOperand(0);
8989    Info.offset = 0;
8990    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8991    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8992    Info.vol = false; // volatile loads with NEON intrinsics not supported
8993    Info.readMem = true;
8994    Info.writeMem = false;
8995    return true;
8996  }
8997  case Intrinsic::arm_neon_vst1:
8998  case Intrinsic::arm_neon_vst2:
8999  case Intrinsic::arm_neon_vst3:
9000  case Intrinsic::arm_neon_vst4:
9001  case Intrinsic::arm_neon_vst2lane:
9002  case Intrinsic::arm_neon_vst3lane:
9003  case Intrinsic::arm_neon_vst4lane: {
9004    Info.opc = ISD::INTRINSIC_VOID;
9005    // Conservatively set memVT to the entire set of vectors stored.
9006    unsigned NumElts = 0;
9007    for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9008      Type *ArgTy = I.getArgOperand(ArgI)->getType();
9009      if (!ArgTy->isVectorTy())
9010        break;
9011      NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9012    }
9013    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9014    Info.ptrVal = I.getArgOperand(0);
9015    Info.offset = 0;
9016    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9017    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9018    Info.vol = false; // volatile stores with NEON intrinsics not supported
9019    Info.readMem = false;
9020    Info.writeMem = true;
9021    return true;
9022  }
9023  case Intrinsic::arm_strexd: {
9024    Info.opc = ISD::INTRINSIC_W_CHAIN;
9025    Info.memVT = MVT::i64;
9026    Info.ptrVal = I.getArgOperand(2);
9027    Info.offset = 0;
9028    Info.align = 8;
9029    Info.vol = true;
9030    Info.readMem = false;
9031    Info.writeMem = true;
9032    return true;
9033  }
9034  case Intrinsic::arm_ldrexd: {
9035    Info.opc = ISD::INTRINSIC_W_CHAIN;
9036    Info.memVT = MVT::i64;
9037    Info.ptrVal = I.getArgOperand(0);
9038    Info.offset = 0;
9039    Info.align = 8;
9040    Info.vol = true;
9041    Info.readMem = true;
9042    Info.writeMem = false;
9043    return true;
9044  }
9045  default:
9046    break;
9047  }
9048
9049  return false;
9050}
9051