ARMISelLowering.cpp revision 922ad78776a2a45fd79602475636077edfdf94fc
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/PseudoSourceValue.h"
44#include "llvm/CodeGen/SelectionDAG.h"
45#include "llvm/MC/MCSectionMachO.h"
46#include "llvm/Target/TargetOptions.h"
47#include "llvm/ADT/VectorExtras.h"
48#include "llvm/ADT/StringExtras.h"
49#include "llvm/ADT/Statistic.h"
50#include "llvm/Support/CommandLine.h"
51#include "llvm/Support/ErrorHandling.h"
52#include "llvm/Support/MathExtras.h"
53#include "llvm/Support/raw_ostream.h"
54#include <sstream>
55using namespace llvm;
56
57STATISTIC(NumTailCalls, "Number of tail calls");
58STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
59
60// This option should go away when tail calls fully work.
61static cl::opt<bool>
62EnableARMTailCalls("arm-tail-calls", cl::Hidden,
63  cl::desc("Generate tail calls (TEMPORARY OPTION)."),
64  cl::init(false));
65
66cl::opt<bool>
67EnableARMLongCalls("arm-long-calls", cl::Hidden,
68  cl::desc("Generate calls via indirect call instructions"),
69  cl::init(false));
70
71static cl::opt<bool>
72ARMInterworking("arm-interworking", cl::Hidden,
73  cl::desc("Enable / disable ARM interworking (for debugging only)"),
74  cl::init(true));
75
76namespace llvm {
77  class ARMCCState : public CCState {
78  public:
79    ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
80               const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
81               LLVMContext &C, ParmContext PC)
82        : CCState(CC, isVarArg, MF, TM, locs, C) {
83      assert(((PC == Call) || (PC == Prologue)) &&
84             "ARMCCState users must specify whether their context is call"
85             "or prologue generation.");
86      CallOrPrologue = PC;
87    }
88  };
89}
90
91// The APCS parameter registers.
92static const unsigned GPRArgRegs[] = {
93  ARM::R0, ARM::R1, ARM::R2, ARM::R3
94};
95
96void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
97                                       EVT PromotedBitwiseVT) {
98  if (VT != PromotedLdStVT) {
99    setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
100    AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
101                       PromotedLdStVT.getSimpleVT());
102
103    setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
104    AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
105                       PromotedLdStVT.getSimpleVT());
106  }
107
108  EVT ElemTy = VT.getVectorElementType();
109  if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
110    setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
111  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
112  if (ElemTy != MVT::i32) {
113    setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
114    setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
115    setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
116    setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
117  }
118  setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
119  setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
120  setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
121  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
122  setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
123  setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
124  if (VT.isInteger()) {
125    setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
126    setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
127    setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
128    setLoadExtAction(ISD::SEXTLOAD, VT.getSimpleVT(), Expand);
129    setLoadExtAction(ISD::ZEXTLOAD, VT.getSimpleVT(), Expand);
130    for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
131         InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
132      setTruncStoreAction(VT.getSimpleVT(),
133                          (MVT::SimpleValueType)InnerVT, Expand);
134  }
135  setLoadExtAction(ISD::EXTLOAD, VT.getSimpleVT(), Expand);
136
137  // Promote all bit-wise operations.
138  if (VT.isInteger() && VT != PromotedBitwiseVT) {
139    setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
140    AddPromotedToType (ISD::AND, VT.getSimpleVT(),
141                       PromotedBitwiseVT.getSimpleVT());
142    setOperationAction(ISD::OR,  VT.getSimpleVT(), Promote);
143    AddPromotedToType (ISD::OR,  VT.getSimpleVT(),
144                       PromotedBitwiseVT.getSimpleVT());
145    setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
146    AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
147                       PromotedBitwiseVT.getSimpleVT());
148  }
149
150  // Neon does not support vector divide/remainder operations.
151  setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
152  setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
153  setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
154  setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
155  setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
156  setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
157}
158
159void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
160  addRegisterClass(VT, ARM::DPRRegisterClass);
161  addTypeForNEON(VT, MVT::f64, MVT::v2i32);
162}
163
164void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
165  addRegisterClass(VT, ARM::QPRRegisterClass);
166  addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
167}
168
169static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
170  if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
171    return new TargetLoweringObjectFileMachO();
172
173  return new ARMElfTargetObjectFile();
174}
175
176ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
177    : TargetLowering(TM, createTLOF(TM)) {
178  Subtarget = &TM.getSubtarget<ARMSubtarget>();
179  RegInfo = TM.getRegisterInfo();
180  Itins = TM.getInstrItineraryData();
181
182  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
183
184  if (Subtarget->isTargetDarwin()) {
185    // Uses VFP for Thumb libfuncs if available.
186    if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
187      // Single-precision floating-point arithmetic.
188      setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
189      setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
190      setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
191      setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
192
193      // Double-precision floating-point arithmetic.
194      setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
195      setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
196      setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
197      setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
198
199      // Single-precision comparisons.
200      setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
201      setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
202      setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
203      setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
204      setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
205      setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
206      setLibcallName(RTLIB::UO_F32,  "__unordsf2vfp");
207      setLibcallName(RTLIB::O_F32,   "__unordsf2vfp");
208
209      setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
210      setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
211      setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
212      setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
213      setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
214      setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
215      setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
216      setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
217
218      // Double-precision comparisons.
219      setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
220      setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
221      setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
222      setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
223      setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
224      setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
225      setLibcallName(RTLIB::UO_F64,  "__unorddf2vfp");
226      setLibcallName(RTLIB::O_F64,   "__unorddf2vfp");
227
228      setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
229      setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
230      setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
231      setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
232      setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
233      setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
234      setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
235      setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
236
237      // Floating-point to integer conversions.
238      // i64 conversions are done via library routines even when generating VFP
239      // instructions, so use the same ones.
240      setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
241      setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
242      setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
243      setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
244
245      // Conversions between floating types.
246      setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
247      setLibcallName(RTLIB::FPEXT_F32_F64,   "__extendsfdf2vfp");
248
249      // Integer to floating-point conversions.
250      // i64 conversions are done via library routines even when generating VFP
251      // instructions, so use the same ones.
252      // FIXME: There appears to be some naming inconsistency in ARM libgcc:
253      // e.g., __floatunsidf vs. __floatunssidfvfp.
254      setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
255      setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
256      setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
257      setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
258    }
259  }
260
261  // These libcalls are not available in 32-bit.
262  setLibcallName(RTLIB::SHL_I128, 0);
263  setLibcallName(RTLIB::SRL_I128, 0);
264  setLibcallName(RTLIB::SRA_I128, 0);
265
266  if (Subtarget->isAAPCS_ABI()) {
267    // Double-precision floating-point arithmetic helper functions
268    // RTABI chapter 4.1.2, Table 2
269    setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
270    setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
271    setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
272    setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
273    setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
274    setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
275    setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
276    setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
277
278    // Double-precision floating-point comparison helper functions
279    // RTABI chapter 4.1.2, Table 3
280    setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
281    setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
282    setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
283    setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
284    setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
285    setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
286    setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
287    setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
288    setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
289    setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
290    setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
291    setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
292    setLibcallName(RTLIB::UO_F64,  "__aeabi_dcmpun");
293    setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
294    setLibcallName(RTLIB::O_F64,   "__aeabi_dcmpun");
295    setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
296    setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
297    setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
298    setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
299    setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
300    setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
301    setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
302    setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
303    setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
304
305    // Single-precision floating-point arithmetic helper functions
306    // RTABI chapter 4.1.2, Table 4
307    setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
308    setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
309    setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
310    setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
311    setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
312    setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
313    setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
314    setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
315
316    // Single-precision floating-point comparison helper functions
317    // RTABI chapter 4.1.2, Table 5
318    setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
319    setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
320    setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
321    setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
322    setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
323    setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
324    setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
325    setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
326    setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
327    setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
328    setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
329    setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
330    setLibcallName(RTLIB::UO_F32,  "__aeabi_fcmpun");
331    setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
332    setLibcallName(RTLIB::O_F32,   "__aeabi_fcmpun");
333    setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
334    setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
335    setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
336    setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
337    setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
338    setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
339    setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
340    setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
341    setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
342
343    // Floating-point to integer conversions.
344    // RTABI chapter 4.1.2, Table 6
345    setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
346    setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
347    setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
348    setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
349    setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
350    setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
351    setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
352    setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
353    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
354    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
355    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
356    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
357    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
358    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
359    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
360    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
361
362    // Conversions between floating types.
363    // RTABI chapter 4.1.2, Table 7
364    setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
365    setLibcallName(RTLIB::FPEXT_F32_F64,   "__aeabi_f2d");
366    setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
367    setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
368
369    // Integer to floating-point conversions.
370    // RTABI chapter 4.1.2, Table 8
371    setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
372    setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
373    setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
374    setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
375    setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
376    setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
377    setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
378    setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
379    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
380    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
381    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
382    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
383    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
384    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
385    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
386    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
387
388    // Long long helper functions
389    // RTABI chapter 4.2, Table 9
390    setLibcallName(RTLIB::MUL_I64,  "__aeabi_lmul");
391    setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
392    setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
393    setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
394    setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
395    setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
396    setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
397    setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
398    setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
399    setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
400    setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
401    setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
402
403    // Integer division functions
404    // RTABI chapter 4.3.1
405    setLibcallName(RTLIB::SDIV_I8,  "__aeabi_idiv");
406    setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
407    setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
408    setLibcallName(RTLIB::UDIV_I8,  "__aeabi_uidiv");
409    setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
410    setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
411    setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
412    setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
413    setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
414    setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
415    setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
416    setLibcallCallingConv(RTLIB::UDIV_I32, 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  }
424
425  // Use divmod compiler-rt calls for iOS 5.0 and later.
426  if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
427      !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
428    setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
429    setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
430  }
431
432  if (Subtarget->isThumb1Only())
433    addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
434  else
435    addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
436  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
437    addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
438    if (!Subtarget->isFPOnlySP())
439      addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
440
441    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
442  }
443
444  if (Subtarget->hasNEON()) {
445    addDRTypeForNEON(MVT::v2f32);
446    addDRTypeForNEON(MVT::v8i8);
447    addDRTypeForNEON(MVT::v4i16);
448    addDRTypeForNEON(MVT::v2i32);
449    addDRTypeForNEON(MVT::v1i64);
450
451    addQRTypeForNEON(MVT::v4f32);
452    addQRTypeForNEON(MVT::v2f64);
453    addQRTypeForNEON(MVT::v16i8);
454    addQRTypeForNEON(MVT::v8i16);
455    addQRTypeForNEON(MVT::v4i32);
456    addQRTypeForNEON(MVT::v2i64);
457
458    // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
459    // neither Neon nor VFP support any arithmetic operations on it.
460    setOperationAction(ISD::FADD, MVT::v2f64, Expand);
461    setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
462    setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
463    setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
464    setOperationAction(ISD::FREM, MVT::v2f64, Expand);
465    setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
466    setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
467    setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
468    setOperationAction(ISD::FABS, MVT::v2f64, Expand);
469    setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
470    setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
471    setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
472    setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
473    setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
474    setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
475    setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
476    setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
477    setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
478    setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
479    setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
480    setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
481    setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
482    setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
483    setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
484
485    setTruncStoreAction(MVT::v2f64, MVT::v2f32, Expand);
486
487    // Neon does not support some operations on v1i64 and v2i64 types.
488    setOperationAction(ISD::MUL, MVT::v1i64, Expand);
489    // Custom handling for some quad-vector types to detect VMULL.
490    setOperationAction(ISD::MUL, MVT::v8i16, Custom);
491    setOperationAction(ISD::MUL, MVT::v4i32, Custom);
492    setOperationAction(ISD::MUL, MVT::v2i64, Custom);
493    // Custom handling for some vector types to avoid expensive expansions
494    setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
495    setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
496    setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
497    setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
498    setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
499    setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
500    // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
501    // a destination type that is wider than the source.
502    setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
503    setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
504
505    setTargetDAGCombine(ISD::INTRINSIC_VOID);
506    setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
507    setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
508    setTargetDAGCombine(ISD::SHL);
509    setTargetDAGCombine(ISD::SRL);
510    setTargetDAGCombine(ISD::SRA);
511    setTargetDAGCombine(ISD::SIGN_EXTEND);
512    setTargetDAGCombine(ISD::ZERO_EXTEND);
513    setTargetDAGCombine(ISD::ANY_EXTEND);
514    setTargetDAGCombine(ISD::SELECT_CC);
515    setTargetDAGCombine(ISD::BUILD_VECTOR);
516    setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
517    setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
518    setTargetDAGCombine(ISD::STORE);
519    setTargetDAGCombine(ISD::FP_TO_SINT);
520    setTargetDAGCombine(ISD::FP_TO_UINT);
521    setTargetDAGCombine(ISD::FDIV);
522
523    setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
524  }
525
526  computeRegisterProperties();
527
528  // ARM does not have f32 extending load.
529  setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
530
531  // ARM does not have i1 sign extending load.
532  setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
533
534  // ARM supports all 4 flavors of integer indexed load / store.
535  if (!Subtarget->isThumb1Only()) {
536    for (unsigned im = (unsigned)ISD::PRE_INC;
537         im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
538      setIndexedLoadAction(im,  MVT::i1,  Legal);
539      setIndexedLoadAction(im,  MVT::i8,  Legal);
540      setIndexedLoadAction(im,  MVT::i16, Legal);
541      setIndexedLoadAction(im,  MVT::i32, Legal);
542      setIndexedStoreAction(im, MVT::i1,  Legal);
543      setIndexedStoreAction(im, MVT::i8,  Legal);
544      setIndexedStoreAction(im, MVT::i16, Legal);
545      setIndexedStoreAction(im, MVT::i32, Legal);
546    }
547  }
548
549  // i64 operation support.
550  setOperationAction(ISD::MUL,     MVT::i64, Expand);
551  setOperationAction(ISD::MULHU,   MVT::i32, Expand);
552  if (Subtarget->isThumb1Only()) {
553    setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
554    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
555  }
556  if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
557      || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
558    setOperationAction(ISD::MULHS, MVT::i32, Expand);
559
560  setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
561  setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
562  setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
563  setOperationAction(ISD::SRL,       MVT::i64, Custom);
564  setOperationAction(ISD::SRA,       MVT::i64, Custom);
565
566  if (!Subtarget->isThumb1Only()) {
567    // FIXME: We should do this for Thumb1 as well.
568    setOperationAction(ISD::ADDC,    MVT::i32, Custom);
569    setOperationAction(ISD::ADDE,    MVT::i32, Custom);
570    setOperationAction(ISD::SUBC,    MVT::i32, Custom);
571    setOperationAction(ISD::SUBE,    MVT::i32, Custom);
572  }
573
574  // ARM does not have ROTL.
575  setOperationAction(ISD::ROTL,  MVT::i32, Expand);
576  setOperationAction(ISD::CTTZ,  MVT::i32, Custom);
577  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
578  if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
579    setOperationAction(ISD::CTLZ, MVT::i32, Expand);
580
581  // Only ARMv6 has BSWAP.
582  if (!Subtarget->hasV6Ops())
583    setOperationAction(ISD::BSWAP, MVT::i32, Expand);
584
585  // These are expanded into libcalls.
586  if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
587    // v7M has a hardware divider
588    setOperationAction(ISD::SDIV,  MVT::i32, Expand);
589    setOperationAction(ISD::UDIV,  MVT::i32, Expand);
590  }
591  setOperationAction(ISD::SREM,  MVT::i32, Expand);
592  setOperationAction(ISD::UREM,  MVT::i32, Expand);
593  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
594  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
595
596  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
597  setOperationAction(ISD::ConstantPool,  MVT::i32,   Custom);
598  setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
599  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
600  setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
601
602  setOperationAction(ISD::TRAP, MVT::Other, Legal);
603
604  // Use the default implementation.
605  setOperationAction(ISD::VASTART,            MVT::Other, Custom);
606  setOperationAction(ISD::VAARG,              MVT::Other, Expand);
607  setOperationAction(ISD::VACOPY,             MVT::Other, Expand);
608  setOperationAction(ISD::VAEND,              MVT::Other, Expand);
609  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
610  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
611  setOperationAction(ISD::EHSELECTION,        MVT::i32,   Expand);
612  setOperationAction(ISD::EXCEPTIONADDR,      MVT::i32,   Expand);
613  setExceptionPointerRegister(ARM::R0);
614  setExceptionSelectorRegister(ARM::R1);
615
616  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
617  // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
618  // the default expansion.
619  // FIXME: This should be checking for v6k, not just v6.
620  if (Subtarget->hasDataBarrier() ||
621      (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
622    // membarrier needs custom lowering; the rest are legal and handled
623    // normally.
624    setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
625    setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
626    // Custom lowering for 64-bit ops
627    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i64, Custom);
628    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i64, Custom);
629    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i64, Custom);
630    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i64, Custom);
631    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i64, Custom);
632    setOperationAction(ISD::ATOMIC_SWAP,  MVT::i64, Custom);
633    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i64, Custom);
634    // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
635    setInsertFencesForAtomic(true);
636  } else {
637    // Set them all for expansion, which will force libcalls.
638    setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
639    setOperationAction(ISD::ATOMIC_FENCE,   MVT::Other, Expand);
640    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i32, Expand);
641    setOperationAction(ISD::ATOMIC_SWAP,      MVT::i32, Expand);
642    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i32, Expand);
643    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i32, Expand);
644    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i32, Expand);
645    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i32, Expand);
646    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i32, Expand);
647    setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
648    setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
649    setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
650    setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
651    setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
652    // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
653    // Unordered/Monotonic case.
654    setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
655    setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
656    // Since the libcalls include locking, fold in the fences
657    setShouldFoldAtomicFences(true);
658  }
659
660  setOperationAction(ISD::PREFETCH,         MVT::Other, Custom);
661
662  // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
663  if (!Subtarget->hasV6Ops()) {
664    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
665    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
666  }
667  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
668
669  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
670    // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
671    // iff target supports vfp2.
672    setOperationAction(ISD::BITCAST, MVT::i64, Custom);
673    setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
674  }
675
676  // We want to custom lower some of our intrinsics.
677  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
678  if (Subtarget->isTargetDarwin()) {
679    setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
680    setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
681    setOperationAction(ISD::EH_SJLJ_DISPATCHSETUP, MVT::Other, Custom);
682    setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
683  }
684
685  setOperationAction(ISD::SETCC,     MVT::i32, Expand);
686  setOperationAction(ISD::SETCC,     MVT::f32, Expand);
687  setOperationAction(ISD::SETCC,     MVT::f64, Expand);
688  setOperationAction(ISD::SELECT,    MVT::i32, Custom);
689  setOperationAction(ISD::SELECT,    MVT::f32, Custom);
690  setOperationAction(ISD::SELECT,    MVT::f64, Custom);
691  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
692  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
693  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
694
695  setOperationAction(ISD::BRCOND,    MVT::Other, Expand);
696  setOperationAction(ISD::BR_CC,     MVT::i32,   Custom);
697  setOperationAction(ISD::BR_CC,     MVT::f32,   Custom);
698  setOperationAction(ISD::BR_CC,     MVT::f64,   Custom);
699  setOperationAction(ISD::BR_JT,     MVT::Other, Custom);
700
701  // We don't support sin/cos/fmod/copysign/pow
702  setOperationAction(ISD::FSIN,      MVT::f64, Expand);
703  setOperationAction(ISD::FSIN,      MVT::f32, Expand);
704  setOperationAction(ISD::FCOS,      MVT::f32, Expand);
705  setOperationAction(ISD::FCOS,      MVT::f64, Expand);
706  setOperationAction(ISD::FREM,      MVT::f64, Expand);
707  setOperationAction(ISD::FREM,      MVT::f32, Expand);
708  if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) {
709    setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
710    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
711  }
712  setOperationAction(ISD::FPOW,      MVT::f64, Expand);
713  setOperationAction(ISD::FPOW,      MVT::f32, Expand);
714
715  setOperationAction(ISD::FMA, MVT::f64, Expand);
716  setOperationAction(ISD::FMA, MVT::f32, Expand);
717
718  // Various VFP goodness
719  if (!UseSoftFloat && !Subtarget->isThumb1Only()) {
720    // int <-> fp are custom expanded into bit_convert + ARMISD ops.
721    if (Subtarget->hasVFP2()) {
722      setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
723      setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
724      setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
725      setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
726    }
727    // Special handling for half-precision FP.
728    if (!Subtarget->hasFP16()) {
729      setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
730      setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
731    }
732  }
733
734  // We have target-specific dag combine patterns for the following nodes:
735  // ARMISD::VMOVRRD  - No need to call setTargetDAGCombine
736  setTargetDAGCombine(ISD::ADD);
737  setTargetDAGCombine(ISD::SUB);
738  setTargetDAGCombine(ISD::MUL);
739
740  if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON())
741    setTargetDAGCombine(ISD::OR);
742  if (Subtarget->hasNEON())
743    setTargetDAGCombine(ISD::AND);
744
745  setStackPointerRegisterToSaveRestore(ARM::SP);
746
747  if (UseSoftFloat || Subtarget->isThumb1Only() || !Subtarget->hasVFP2())
748    setSchedulingPreference(Sched::RegPressure);
749  else
750    setSchedulingPreference(Sched::Hybrid);
751
752  //// temporary - rewrite interface to use type
753  maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
754
755  // On ARM arguments smaller than 4 bytes are extended, so all arguments
756  // are at least 4 bytes aligned.
757  setMinStackArgumentAlignment(4);
758
759  benefitFromCodePlacementOpt = true;
760
761  setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
762}
763
764// FIXME: It might make sense to define the representative register class as the
765// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
766// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
767// SPR's representative would be DPR_VFP2. This should work well if register
768// pressure tracking were modified such that a register use would increment the
769// pressure of the register class's representative and all of it's super
770// classes' representatives transitively. We have not implemented this because
771// of the difficulty prior to coalescing of modeling operand register classes
772// due to the common occurrence of cross class copies and subregister insertions
773// and extractions.
774std::pair<const TargetRegisterClass*, uint8_t>
775ARMTargetLowering::findRepresentativeClass(EVT VT) const{
776  const TargetRegisterClass *RRC = 0;
777  uint8_t Cost = 1;
778  switch (VT.getSimpleVT().SimpleTy) {
779  default:
780    return TargetLowering::findRepresentativeClass(VT);
781  // Use DPR as representative register class for all floating point
782  // and vector types. Since there are 32 SPR registers and 32 DPR registers so
783  // the cost is 1 for both f32 and f64.
784  case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
785  case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
786    RRC = ARM::DPRRegisterClass;
787    // When NEON is used for SP, only half of the register file is available
788    // because operations that define both SP and DP results will be constrained
789    // to the VFP2 class (D0-D15). We currently model this constraint prior to
790    // coalescing by double-counting the SP regs. See the FIXME above.
791    if (Subtarget->useNEONForSinglePrecisionFP())
792      Cost = 2;
793    break;
794  case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
795  case MVT::v4f32: case MVT::v2f64:
796    RRC = ARM::DPRRegisterClass;
797    Cost = 2;
798    break;
799  case MVT::v4i64:
800    RRC = ARM::DPRRegisterClass;
801    Cost = 4;
802    break;
803  case MVT::v8i64:
804    RRC = ARM::DPRRegisterClass;
805    Cost = 8;
806    break;
807  }
808  return std::make_pair(RRC, Cost);
809}
810
811const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
812  switch (Opcode) {
813  default: return 0;
814  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
815  case ARMISD::WrapperDYN:    return "ARMISD::WrapperDYN";
816  case ARMISD::WrapperPIC:    return "ARMISD::WrapperPIC";
817  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
818  case ARMISD::CALL:          return "ARMISD::CALL";
819  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
820  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
821  case ARMISD::tCALL:         return "ARMISD::tCALL";
822  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
823  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
824  case ARMISD::BR2_JT:        return "ARMISD::BR2_JT";
825  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
826  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
827  case ARMISD::CMP:           return "ARMISD::CMP";
828  case ARMISD::CMPZ:          return "ARMISD::CMPZ";
829  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
830  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
831  case ARMISD::BCC_i64:       return "ARMISD::BCC_i64";
832  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";
833  case ARMISD::CMOV:          return "ARMISD::CMOV";
834
835  case ARMISD::RBIT:          return "ARMISD::RBIT";
836
837  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
838  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
839  case ARMISD::SITOF:         return "ARMISD::SITOF";
840  case ARMISD::UITOF:         return "ARMISD::UITOF";
841
842  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
843  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
844  case ARMISD::RRX:           return "ARMISD::RRX";
845
846  case ARMISD::ADDC:          return "ARMISD::ADDC";
847  case ARMISD::ADDE:          return "ARMISD::ADDE";
848  case ARMISD::SUBC:          return "ARMISD::SUBC";
849  case ARMISD::SUBE:          return "ARMISD::SUBE";
850
851  case ARMISD::VMOVRRD:       return "ARMISD::VMOVRRD";
852  case ARMISD::VMOVDRR:       return "ARMISD::VMOVDRR";
853
854  case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
855  case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
856  case ARMISD::EH_SJLJ_DISPATCHSETUP:return "ARMISD::EH_SJLJ_DISPATCHSETUP";
857
858  case ARMISD::TC_RETURN:     return "ARMISD::TC_RETURN";
859
860  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
861
862  case ARMISD::DYN_ALLOC:     return "ARMISD::DYN_ALLOC";
863
864  case ARMISD::MEMBARRIER:    return "ARMISD::MEMBARRIER";
865  case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
866
867  case ARMISD::PRELOAD:       return "ARMISD::PRELOAD";
868
869  case ARMISD::VCEQ:          return "ARMISD::VCEQ";
870  case ARMISD::VCEQZ:         return "ARMISD::VCEQZ";
871  case ARMISD::VCGE:          return "ARMISD::VCGE";
872  case ARMISD::VCGEZ:         return "ARMISD::VCGEZ";
873  case ARMISD::VCLEZ:         return "ARMISD::VCLEZ";
874  case ARMISD::VCGEU:         return "ARMISD::VCGEU";
875  case ARMISD::VCGT:          return "ARMISD::VCGT";
876  case ARMISD::VCGTZ:         return "ARMISD::VCGTZ";
877  case ARMISD::VCLTZ:         return "ARMISD::VCLTZ";
878  case ARMISD::VCGTU:         return "ARMISD::VCGTU";
879  case ARMISD::VTST:          return "ARMISD::VTST";
880
881  case ARMISD::VSHL:          return "ARMISD::VSHL";
882  case ARMISD::VSHRs:         return "ARMISD::VSHRs";
883  case ARMISD::VSHRu:         return "ARMISD::VSHRu";
884  case ARMISD::VSHLLs:        return "ARMISD::VSHLLs";
885  case ARMISD::VSHLLu:        return "ARMISD::VSHLLu";
886  case ARMISD::VSHLLi:        return "ARMISD::VSHLLi";
887  case ARMISD::VSHRN:         return "ARMISD::VSHRN";
888  case ARMISD::VRSHRs:        return "ARMISD::VRSHRs";
889  case ARMISD::VRSHRu:        return "ARMISD::VRSHRu";
890  case ARMISD::VRSHRN:        return "ARMISD::VRSHRN";
891  case ARMISD::VQSHLs:        return "ARMISD::VQSHLs";
892  case ARMISD::VQSHLu:        return "ARMISD::VQSHLu";
893  case ARMISD::VQSHLsu:       return "ARMISD::VQSHLsu";
894  case ARMISD::VQSHRNs:       return "ARMISD::VQSHRNs";
895  case ARMISD::VQSHRNu:       return "ARMISD::VQSHRNu";
896  case ARMISD::VQSHRNsu:      return "ARMISD::VQSHRNsu";
897  case ARMISD::VQRSHRNs:      return "ARMISD::VQRSHRNs";
898  case ARMISD::VQRSHRNu:      return "ARMISD::VQRSHRNu";
899  case ARMISD::VQRSHRNsu:     return "ARMISD::VQRSHRNsu";
900  case ARMISD::VGETLANEu:     return "ARMISD::VGETLANEu";
901  case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
902  case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
903  case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
904  case ARMISD::VDUP:          return "ARMISD::VDUP";
905  case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
906  case ARMISD::VEXT:          return "ARMISD::VEXT";
907  case ARMISD::VREV64:        return "ARMISD::VREV64";
908  case ARMISD::VREV32:        return "ARMISD::VREV32";
909  case ARMISD::VREV16:        return "ARMISD::VREV16";
910  case ARMISD::VZIP:          return "ARMISD::VZIP";
911  case ARMISD::VUZP:          return "ARMISD::VUZP";
912  case ARMISD::VTRN:          return "ARMISD::VTRN";
913  case ARMISD::VTBL1:         return "ARMISD::VTBL1";
914  case ARMISD::VTBL2:         return "ARMISD::VTBL2";
915  case ARMISD::VMULLs:        return "ARMISD::VMULLs";
916  case ARMISD::VMULLu:        return "ARMISD::VMULLu";
917  case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
918  case ARMISD::FMAX:          return "ARMISD::FMAX";
919  case ARMISD::FMIN:          return "ARMISD::FMIN";
920  case ARMISD::BFI:           return "ARMISD::BFI";
921  case ARMISD::VORRIMM:       return "ARMISD::VORRIMM";
922  case ARMISD::VBICIMM:       return "ARMISD::VBICIMM";
923  case ARMISD::VBSL:          return "ARMISD::VBSL";
924  case ARMISD::VLD2DUP:       return "ARMISD::VLD2DUP";
925  case ARMISD::VLD3DUP:       return "ARMISD::VLD3DUP";
926  case ARMISD::VLD4DUP:       return "ARMISD::VLD4DUP";
927  case ARMISD::VLD1_UPD:      return "ARMISD::VLD1_UPD";
928  case ARMISD::VLD2_UPD:      return "ARMISD::VLD2_UPD";
929  case ARMISD::VLD3_UPD:      return "ARMISD::VLD3_UPD";
930  case ARMISD::VLD4_UPD:      return "ARMISD::VLD4_UPD";
931  case ARMISD::VLD2LN_UPD:    return "ARMISD::VLD2LN_UPD";
932  case ARMISD::VLD3LN_UPD:    return "ARMISD::VLD3LN_UPD";
933  case ARMISD::VLD4LN_UPD:    return "ARMISD::VLD4LN_UPD";
934  case ARMISD::VLD2DUP_UPD:   return "ARMISD::VLD2DUP_UPD";
935  case ARMISD::VLD3DUP_UPD:   return "ARMISD::VLD3DUP_UPD";
936  case ARMISD::VLD4DUP_UPD:   return "ARMISD::VLD4DUP_UPD";
937  case ARMISD::VST1_UPD:      return "ARMISD::VST1_UPD";
938  case ARMISD::VST2_UPD:      return "ARMISD::VST2_UPD";
939  case ARMISD::VST3_UPD:      return "ARMISD::VST3_UPD";
940  case ARMISD::VST4_UPD:      return "ARMISD::VST4_UPD";
941  case ARMISD::VST2LN_UPD:    return "ARMISD::VST2LN_UPD";
942  case ARMISD::VST3LN_UPD:    return "ARMISD::VST3LN_UPD";
943  case ARMISD::VST4LN_UPD:    return "ARMISD::VST4LN_UPD";
944  }
945}
946
947EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
948  if (!VT.isVector()) return getPointerTy();
949  return VT.changeVectorElementTypeToInteger();
950}
951
952/// getRegClassFor - Return the register class that should be used for the
953/// specified value type.
954TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
955  // Map v4i64 to QQ registers but do not make the type legal. Similarly map
956  // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
957  // load / store 4 to 8 consecutive D registers.
958  if (Subtarget->hasNEON()) {
959    if (VT == MVT::v4i64)
960      return ARM::QQPRRegisterClass;
961    else if (VT == MVT::v8i64)
962      return ARM::QQQQPRRegisterClass;
963  }
964  return TargetLowering::getRegClassFor(VT);
965}
966
967// Create a fast isel object.
968FastISel *
969ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
970  return ARM::createFastISel(funcInfo);
971}
972
973/// getMaximalGlobalOffset - Returns the maximal possible offset which can
974/// be used for loads / stores from the global.
975unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
976  return (Subtarget->isThumb1Only() ? 127 : 4095);
977}
978
979Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
980  unsigned NumVals = N->getNumValues();
981  if (!NumVals)
982    return Sched::RegPressure;
983
984  for (unsigned i = 0; i != NumVals; ++i) {
985    EVT VT = N->getValueType(i);
986    if (VT == MVT::Glue || VT == MVT::Other)
987      continue;
988    if (VT.isFloatingPoint() || VT.isVector())
989      return Sched::Latency;
990  }
991
992  if (!N->isMachineOpcode())
993    return Sched::RegPressure;
994
995  // Load are scheduled for latency even if there instruction itinerary
996  // is not available.
997  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
998  const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
999
1000  if (MCID.getNumDefs() == 0)
1001    return Sched::RegPressure;
1002  if (!Itins->isEmpty() &&
1003      Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1004    return Sched::Latency;
1005
1006  return Sched::RegPressure;
1007}
1008
1009//===----------------------------------------------------------------------===//
1010// Lowering Code
1011//===----------------------------------------------------------------------===//
1012
1013/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1014static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1015  switch (CC) {
1016  default: llvm_unreachable("Unknown condition code!");
1017  case ISD::SETNE:  return ARMCC::NE;
1018  case ISD::SETEQ:  return ARMCC::EQ;
1019  case ISD::SETGT:  return ARMCC::GT;
1020  case ISD::SETGE:  return ARMCC::GE;
1021  case ISD::SETLT:  return ARMCC::LT;
1022  case ISD::SETLE:  return ARMCC::LE;
1023  case ISD::SETUGT: return ARMCC::HI;
1024  case ISD::SETUGE: return ARMCC::HS;
1025  case ISD::SETULT: return ARMCC::LO;
1026  case ISD::SETULE: return ARMCC::LS;
1027  }
1028}
1029
1030/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1031static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1032                        ARMCC::CondCodes &CondCode2) {
1033  CondCode2 = ARMCC::AL;
1034  switch (CC) {
1035  default: llvm_unreachable("Unknown FP condition!");
1036  case ISD::SETEQ:
1037  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1038  case ISD::SETGT:
1039  case ISD::SETOGT: CondCode = ARMCC::GT; break;
1040  case ISD::SETGE:
1041  case ISD::SETOGE: CondCode = ARMCC::GE; break;
1042  case ISD::SETOLT: CondCode = ARMCC::MI; break;
1043  case ISD::SETOLE: CondCode = ARMCC::LS; break;
1044  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1045  case ISD::SETO:   CondCode = ARMCC::VC; break;
1046  case ISD::SETUO:  CondCode = ARMCC::VS; break;
1047  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1048  case ISD::SETUGT: CondCode = ARMCC::HI; break;
1049  case ISD::SETUGE: CondCode = ARMCC::PL; break;
1050  case ISD::SETLT:
1051  case ISD::SETULT: CondCode = ARMCC::LT; break;
1052  case ISD::SETLE:
1053  case ISD::SETULE: CondCode = ARMCC::LE; break;
1054  case ISD::SETNE:
1055  case ISD::SETUNE: CondCode = ARMCC::NE; break;
1056  }
1057}
1058
1059//===----------------------------------------------------------------------===//
1060//                      Calling Convention Implementation
1061//===----------------------------------------------------------------------===//
1062
1063#include "ARMGenCallingConv.inc"
1064
1065/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1066/// given CallingConvention value.
1067CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1068                                                 bool Return,
1069                                                 bool isVarArg) const {
1070  switch (CC) {
1071  default:
1072    llvm_unreachable("Unsupported calling convention");
1073  case CallingConv::Fast:
1074    if (Subtarget->hasVFP2() && !isVarArg) {
1075      if (!Subtarget->isAAPCS_ABI())
1076        return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1077      // For AAPCS ABI targets, just use VFP variant of the calling convention.
1078      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1079    }
1080    // Fallthrough
1081  case CallingConv::C: {
1082    // Use target triple & subtarget features to do actual dispatch.
1083    if (!Subtarget->isAAPCS_ABI())
1084      return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1085    else if (Subtarget->hasVFP2() &&
1086             FloatABIType == FloatABI::Hard && !isVarArg)
1087      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1088    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1089  }
1090  case CallingConv::ARM_AAPCS_VFP:
1091    return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1092  case CallingConv::ARM_AAPCS:
1093    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1094  case CallingConv::ARM_APCS:
1095    return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1096  }
1097}
1098
1099/// LowerCallResult - Lower the result values of a call into the
1100/// appropriate copies out of appropriate physical registers.
1101SDValue
1102ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1103                                   CallingConv::ID CallConv, bool isVarArg,
1104                                   const SmallVectorImpl<ISD::InputArg> &Ins,
1105                                   DebugLoc dl, SelectionDAG &DAG,
1106                                   SmallVectorImpl<SDValue> &InVals) const {
1107
1108  // Assign locations to each value returned by this call.
1109  SmallVector<CCValAssign, 16> RVLocs;
1110  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1111                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1112  CCInfo.AnalyzeCallResult(Ins,
1113                           CCAssignFnForNode(CallConv, /* Return*/ true,
1114                                             isVarArg));
1115
1116  // Copy all of the result registers out of their specified physreg.
1117  for (unsigned i = 0; i != RVLocs.size(); ++i) {
1118    CCValAssign VA = RVLocs[i];
1119
1120    SDValue Val;
1121    if (VA.needsCustom()) {
1122      // Handle f64 or half of a v2f64.
1123      SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1124                                      InFlag);
1125      Chain = Lo.getValue(1);
1126      InFlag = Lo.getValue(2);
1127      VA = RVLocs[++i]; // skip ahead to next loc
1128      SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1129                                      InFlag);
1130      Chain = Hi.getValue(1);
1131      InFlag = Hi.getValue(2);
1132      Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1133
1134      if (VA.getLocVT() == MVT::v2f64) {
1135        SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1136        Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1137                          DAG.getConstant(0, MVT::i32));
1138
1139        VA = RVLocs[++i]; // skip ahead to next loc
1140        Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1141        Chain = Lo.getValue(1);
1142        InFlag = Lo.getValue(2);
1143        VA = RVLocs[++i]; // skip ahead to next loc
1144        Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1145        Chain = Hi.getValue(1);
1146        InFlag = Hi.getValue(2);
1147        Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1148        Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1149                          DAG.getConstant(1, MVT::i32));
1150      }
1151    } else {
1152      Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1153                               InFlag);
1154      Chain = Val.getValue(1);
1155      InFlag = Val.getValue(2);
1156    }
1157
1158    switch (VA.getLocInfo()) {
1159    default: llvm_unreachable("Unknown loc info!");
1160    case CCValAssign::Full: break;
1161    case CCValAssign::BCvt:
1162      Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1163      break;
1164    }
1165
1166    InVals.push_back(Val);
1167  }
1168
1169  return Chain;
1170}
1171
1172/// LowerMemOpCallTo - Store the argument to the stack.
1173SDValue
1174ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1175                                    SDValue StackPtr, SDValue Arg,
1176                                    DebugLoc dl, SelectionDAG &DAG,
1177                                    const CCValAssign &VA,
1178                                    ISD::ArgFlagsTy Flags) const {
1179  unsigned LocMemOffset = VA.getLocMemOffset();
1180  SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1181  PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1182  return DAG.getStore(Chain, dl, Arg, PtrOff,
1183                      MachinePointerInfo::getStack(LocMemOffset),
1184                      false, false, 0);
1185}
1186
1187void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1188                                         SDValue Chain, SDValue &Arg,
1189                                         RegsToPassVector &RegsToPass,
1190                                         CCValAssign &VA, CCValAssign &NextVA,
1191                                         SDValue &StackPtr,
1192                                         SmallVector<SDValue, 8> &MemOpChains,
1193                                         ISD::ArgFlagsTy Flags) const {
1194
1195  SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1196                              DAG.getVTList(MVT::i32, MVT::i32), Arg);
1197  RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1198
1199  if (NextVA.isRegLoc())
1200    RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1201  else {
1202    assert(NextVA.isMemLoc());
1203    if (StackPtr.getNode() == 0)
1204      StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1205
1206    MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1207                                           dl, DAG, NextVA,
1208                                           Flags));
1209  }
1210}
1211
1212/// LowerCall - Lowering a call into a callseq_start <-
1213/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1214/// nodes.
1215SDValue
1216ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1217                             CallingConv::ID CallConv, bool isVarArg,
1218                             bool &isTailCall,
1219                             const SmallVectorImpl<ISD::OutputArg> &Outs,
1220                             const SmallVectorImpl<SDValue> &OutVals,
1221                             const SmallVectorImpl<ISD::InputArg> &Ins,
1222                             DebugLoc dl, SelectionDAG &DAG,
1223                             SmallVectorImpl<SDValue> &InVals) const {
1224  MachineFunction &MF = DAG.getMachineFunction();
1225  bool IsStructRet    = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1226  bool IsSibCall = false;
1227  // Disable tail calls if they're not supported.
1228  if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1229    isTailCall = false;
1230  if (isTailCall) {
1231    // Check if it's really possible to do a tail call.
1232    isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1233                    isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1234                                                   Outs, OutVals, Ins, DAG);
1235    // We don't support GuaranteedTailCallOpt for ARM, only automatically
1236    // detected sibcalls.
1237    if (isTailCall) {
1238      ++NumTailCalls;
1239      IsSibCall = true;
1240    }
1241  }
1242
1243  // Analyze operands of the call, assigning locations to each operand.
1244  SmallVector<CCValAssign, 16> ArgLocs;
1245  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1246                 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1247  CCInfo.AnalyzeCallOperands(Outs,
1248                             CCAssignFnForNode(CallConv, /* Return*/ false,
1249                                               isVarArg));
1250
1251  // Get a count of how many bytes are to be pushed on the stack.
1252  unsigned NumBytes = CCInfo.getNextStackOffset();
1253
1254  // For tail calls, memory operands are available in our caller's stack.
1255  if (IsSibCall)
1256    NumBytes = 0;
1257
1258  // Adjust the stack pointer for the new arguments...
1259  // These operations are automatically eliminated by the prolog/epilog pass
1260  if (!IsSibCall)
1261    Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1262
1263  SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1264
1265  RegsToPassVector RegsToPass;
1266  SmallVector<SDValue, 8> MemOpChains;
1267
1268  // Walk the register/memloc assignments, inserting copies/loads.  In the case
1269  // of tail call optimization, arguments are handled later.
1270  for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1271       i != e;
1272       ++i, ++realArgIdx) {
1273    CCValAssign &VA = ArgLocs[i];
1274    SDValue Arg = OutVals[realArgIdx];
1275    ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1276    bool isByVal = Flags.isByVal();
1277
1278    // Promote the value if needed.
1279    switch (VA.getLocInfo()) {
1280    default: llvm_unreachable("Unknown loc info!");
1281    case CCValAssign::Full: break;
1282    case CCValAssign::SExt:
1283      Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1284      break;
1285    case CCValAssign::ZExt:
1286      Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1287      break;
1288    case CCValAssign::AExt:
1289      Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1290      break;
1291    case CCValAssign::BCvt:
1292      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1293      break;
1294    }
1295
1296    // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1297    if (VA.needsCustom()) {
1298      if (VA.getLocVT() == MVT::v2f64) {
1299        SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1300                                  DAG.getConstant(0, MVT::i32));
1301        SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1302                                  DAG.getConstant(1, MVT::i32));
1303
1304        PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1305                         VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1306
1307        VA = ArgLocs[++i]; // skip ahead to next loc
1308        if (VA.isRegLoc()) {
1309          PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1310                           VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1311        } else {
1312          assert(VA.isMemLoc());
1313
1314          MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1315                                                 dl, DAG, VA, Flags));
1316        }
1317      } else {
1318        PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1319                         StackPtr, MemOpChains, Flags);
1320      }
1321    } else if (VA.isRegLoc()) {
1322      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1323    } else if (isByVal) {
1324      assert(VA.isMemLoc());
1325      unsigned offset = 0;
1326
1327      // True if this byval aggregate will be split between registers
1328      // and memory.
1329      if (CCInfo.isFirstByValRegValid()) {
1330        EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1331        unsigned int i, j;
1332        for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1333          SDValue Const = DAG.getConstant(4*i, MVT::i32);
1334          SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1335          SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1336                                     MachinePointerInfo(),
1337                                     false, false, 0);
1338          MemOpChains.push_back(Load.getValue(1));
1339          RegsToPass.push_back(std::make_pair(j, Load));
1340        }
1341        offset = ARM::R4 - CCInfo.getFirstByValReg();
1342        CCInfo.clearFirstByValReg();
1343      }
1344
1345      unsigned LocMemOffset = VA.getLocMemOffset();
1346      SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1347      SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1348                                StkPtrOff);
1349      SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1350      SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1351      SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1352                                         MVT::i32);
1353      // TODO: Disable AlwaysInline when it becomes possible
1354      //       to emit a nested call sequence.
1355      MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1356                                          Flags.getByValAlign(),
1357                                          /*isVolatile=*/false,
1358                                          /*AlwaysInline=*/true,
1359                                          MachinePointerInfo(0),
1360                                          MachinePointerInfo(0)));
1361
1362    } else if (!IsSibCall) {
1363      assert(VA.isMemLoc());
1364
1365      MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1366                                             dl, DAG, VA, Flags));
1367    }
1368  }
1369
1370  if (!MemOpChains.empty())
1371    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1372                        &MemOpChains[0], MemOpChains.size());
1373
1374  // Build a sequence of copy-to-reg nodes chained together with token chain
1375  // and flag operands which copy the outgoing args into the appropriate regs.
1376  SDValue InFlag;
1377  // Tail call byval lowering might overwrite argument registers so in case of
1378  // tail call optimization the copies to registers are lowered later.
1379  if (!isTailCall)
1380    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1381      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1382                               RegsToPass[i].second, InFlag);
1383      InFlag = Chain.getValue(1);
1384    }
1385
1386  // For tail calls lower the arguments to the 'real' stack slot.
1387  if (isTailCall) {
1388    // Force all the incoming stack arguments to be loaded from the stack
1389    // before any new outgoing arguments are stored to the stack, because the
1390    // outgoing stack slots may alias the incoming argument stack slots, and
1391    // the alias isn't otherwise explicit. This is slightly more conservative
1392    // than necessary, because it means that each store effectively depends
1393    // on every argument instead of just those arguments it would clobber.
1394
1395    // Do not flag preceding copytoreg stuff together with the following stuff.
1396    InFlag = SDValue();
1397    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1398      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1399                               RegsToPass[i].second, InFlag);
1400      InFlag = Chain.getValue(1);
1401    }
1402    InFlag =SDValue();
1403  }
1404
1405  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1406  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1407  // node so that legalize doesn't hack it.
1408  bool isDirect = false;
1409  bool isARMFunc = false;
1410  bool isLocalARMFunc = false;
1411  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1412
1413  if (EnableARMLongCalls) {
1414    assert (getTargetMachine().getRelocationModel() == Reloc::Static
1415            && "long-calls with non-static relocation model!");
1416    // Handle a global address or an external symbol. If it's not one of
1417    // those, the target's already in a register, so we don't need to do
1418    // anything extra.
1419    if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1420      const GlobalValue *GV = G->getGlobal();
1421      // Create a constant pool entry for the callee address
1422      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1423      ARMConstantPoolValue *CPV =
1424        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1425
1426      // Get the address of the callee into a register
1427      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1428      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1429      Callee = DAG.getLoad(getPointerTy(), dl,
1430                           DAG.getEntryNode(), CPAddr,
1431                           MachinePointerInfo::getConstantPool(),
1432                           false, false, 0);
1433    } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1434      const char *Sym = S->getSymbol();
1435
1436      // Create a constant pool entry for the callee address
1437      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1438      ARMConstantPoolValue *CPV =
1439        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1440                                      ARMPCLabelIndex, 0);
1441      // Get the address of the callee into a register
1442      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1443      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1444      Callee = DAG.getLoad(getPointerTy(), dl,
1445                           DAG.getEntryNode(), CPAddr,
1446                           MachinePointerInfo::getConstantPool(),
1447                           false, false, 0);
1448    }
1449  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1450    const GlobalValue *GV = G->getGlobal();
1451    isDirect = true;
1452    bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1453    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1454                   getTargetMachine().getRelocationModel() != Reloc::Static;
1455    isARMFunc = !Subtarget->isThumb() || isStub;
1456    // ARM call to a local ARM function is predicable.
1457    isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1458    // tBX takes a register source operand.
1459    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1460      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1461      ARMConstantPoolValue *CPV =
1462        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1463      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1464      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1465      Callee = DAG.getLoad(getPointerTy(), dl,
1466                           DAG.getEntryNode(), CPAddr,
1467                           MachinePointerInfo::getConstantPool(),
1468                           false, false, 0);
1469      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1470      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1471                           getPointerTy(), Callee, PICLabel);
1472    } else {
1473      // On ELF targets for PIC code, direct calls should go through the PLT
1474      unsigned OpFlags = 0;
1475      if (Subtarget->isTargetELF() &&
1476                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1477        OpFlags = ARMII::MO_PLT;
1478      Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1479    }
1480  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1481    isDirect = true;
1482    bool isStub = Subtarget->isTargetDarwin() &&
1483                  getTargetMachine().getRelocationModel() != Reloc::Static;
1484    isARMFunc = !Subtarget->isThumb() || isStub;
1485    // tBX takes a register source operand.
1486    const char *Sym = S->getSymbol();
1487    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1488      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1489      ARMConstantPoolValue *CPV =
1490        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1491                                      ARMPCLabelIndex, 4);
1492      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1493      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1494      Callee = DAG.getLoad(getPointerTy(), dl,
1495                           DAG.getEntryNode(), CPAddr,
1496                           MachinePointerInfo::getConstantPool(),
1497                           false, false, 0);
1498      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1499      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1500                           getPointerTy(), Callee, PICLabel);
1501    } else {
1502      unsigned OpFlags = 0;
1503      // On ELF targets for PIC code, direct calls should go through the PLT
1504      if (Subtarget->isTargetELF() &&
1505                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
1506        OpFlags = ARMII::MO_PLT;
1507      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1508    }
1509  }
1510
1511  // FIXME: handle tail calls differently.
1512  unsigned CallOpc;
1513  if (Subtarget->isThumb()) {
1514    if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1515      CallOpc = ARMISD::CALL_NOLINK;
1516    else
1517      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1518  } else {
1519    CallOpc = (isDirect || Subtarget->hasV5TOps())
1520      ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL)
1521      : ARMISD::CALL_NOLINK;
1522  }
1523
1524  std::vector<SDValue> Ops;
1525  Ops.push_back(Chain);
1526  Ops.push_back(Callee);
1527
1528  // Add argument registers to the end of the list so that they are known live
1529  // into the call.
1530  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1531    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1532                                  RegsToPass[i].second.getValueType()));
1533
1534  if (InFlag.getNode())
1535    Ops.push_back(InFlag);
1536
1537  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1538  if (isTailCall)
1539    return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1540
1541  // Returns a chain and a flag for retval copy to use.
1542  Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1543  InFlag = Chain.getValue(1);
1544
1545  Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1546                             DAG.getIntPtrConstant(0, true), InFlag);
1547  if (!Ins.empty())
1548    InFlag = Chain.getValue(1);
1549
1550  // Handle result values, copying them out of physregs into vregs that we
1551  // return.
1552  return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1553                         dl, DAG, InVals);
1554}
1555
1556/// HandleByVal - Every parameter *after* a byval parameter is passed
1557/// on the stack.  Remember the next parameter register to allocate,
1558/// and then confiscate the rest of the parameter registers to insure
1559/// this.
1560void
1561llvm::ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1562  unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1563  assert((State->getCallOrPrologue() == Prologue ||
1564          State->getCallOrPrologue() == Call) &&
1565         "unhandled ParmContext");
1566  if ((!State->isFirstByValRegValid()) &&
1567      (ARM::R0 <= reg) && (reg <= ARM::R3)) {
1568    State->setFirstByValReg(reg);
1569    // At a call site, a byval parameter that is split between
1570    // registers and memory needs its size truncated here.  In a
1571    // function prologue, such byval parameters are reassembled in
1572    // memory, and are not truncated.
1573    if (State->getCallOrPrologue() == Call) {
1574      unsigned excess = 4 * (ARM::R4 - reg);
1575      assert(size >= excess && "expected larger existing stack allocation");
1576      size -= excess;
1577    }
1578  }
1579  // Confiscate any remaining parameter registers to preclude their
1580  // assignment to subsequent parameters.
1581  while (State->AllocateReg(GPRArgRegs, 4))
1582    ;
1583}
1584
1585/// MatchingStackOffset - Return true if the given stack call argument is
1586/// already available in the same position (relatively) of the caller's
1587/// incoming argument stack.
1588static
1589bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1590                         MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1591                         const ARMInstrInfo *TII) {
1592  unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1593  int FI = INT_MAX;
1594  if (Arg.getOpcode() == ISD::CopyFromReg) {
1595    unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1596    if (!TargetRegisterInfo::isVirtualRegister(VR))
1597      return false;
1598    MachineInstr *Def = MRI->getVRegDef(VR);
1599    if (!Def)
1600      return false;
1601    if (!Flags.isByVal()) {
1602      if (!TII->isLoadFromStackSlot(Def, FI))
1603        return false;
1604    } else {
1605      return false;
1606    }
1607  } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1608    if (Flags.isByVal())
1609      // ByVal argument is passed in as a pointer but it's now being
1610      // dereferenced. e.g.
1611      // define @foo(%struct.X* %A) {
1612      //   tail call @bar(%struct.X* byval %A)
1613      // }
1614      return false;
1615    SDValue Ptr = Ld->getBasePtr();
1616    FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1617    if (!FINode)
1618      return false;
1619    FI = FINode->getIndex();
1620  } else
1621    return false;
1622
1623  assert(FI != INT_MAX);
1624  if (!MFI->isFixedObjectIndex(FI))
1625    return false;
1626  return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1627}
1628
1629/// IsEligibleForTailCallOptimization - Check whether the call is eligible
1630/// for tail call optimization. Targets which want to do tail call
1631/// optimization should implement this function.
1632bool
1633ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1634                                                     CallingConv::ID CalleeCC,
1635                                                     bool isVarArg,
1636                                                     bool isCalleeStructRet,
1637                                                     bool isCallerStructRet,
1638                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
1639                                    const SmallVectorImpl<SDValue> &OutVals,
1640                                    const SmallVectorImpl<ISD::InputArg> &Ins,
1641                                                     SelectionDAG& DAG) const {
1642  const Function *CallerF = DAG.getMachineFunction().getFunction();
1643  CallingConv::ID CallerCC = CallerF->getCallingConv();
1644  bool CCMatch = CallerCC == CalleeCC;
1645
1646  // Look for obvious safe cases to perform tail call optimization that do not
1647  // require ABI changes. This is what gcc calls sibcall.
1648
1649  // Do not sibcall optimize vararg calls unless the call site is not passing
1650  // any arguments.
1651  if (isVarArg && !Outs.empty())
1652    return false;
1653
1654  // Also avoid sibcall optimization if either caller or callee uses struct
1655  // return semantics.
1656  if (isCalleeStructRet || isCallerStructRet)
1657    return false;
1658
1659  // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1660  // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1661  // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1662  // support in the assembler and linker to be used. This would need to be
1663  // fixed to fully support tail calls in Thumb1.
1664  //
1665  // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1666  // LR.  This means if we need to reload LR, it takes an extra instructions,
1667  // which outweighs the value of the tail call; but here we don't know yet
1668  // whether LR is going to be used.  Probably the right approach is to
1669  // generate the tail call here and turn it back into CALL/RET in
1670  // emitEpilogue if LR is used.
1671
1672  // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1673  // but we need to make sure there are enough registers; the only valid
1674  // registers are the 4 used for parameters.  We don't currently do this
1675  // case.
1676  if (Subtarget->isThumb1Only())
1677    return false;
1678
1679  // If the calling conventions do not match, then we'd better make sure the
1680  // results are returned in the same way as what the caller expects.
1681  if (!CCMatch) {
1682    SmallVector<CCValAssign, 16> RVLocs1;
1683    ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1684                       getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1685    CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1686
1687    SmallVector<CCValAssign, 16> RVLocs2;
1688    ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1689                       getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1690    CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1691
1692    if (RVLocs1.size() != RVLocs2.size())
1693      return false;
1694    for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1695      if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1696        return false;
1697      if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1698        return false;
1699      if (RVLocs1[i].isRegLoc()) {
1700        if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1701          return false;
1702      } else {
1703        if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1704          return false;
1705      }
1706    }
1707  }
1708
1709  // If the callee takes no arguments then go on to check the results of the
1710  // call.
1711  if (!Outs.empty()) {
1712    // Check if stack adjustment is needed. For now, do not do this if any
1713    // argument is passed on the stack.
1714    SmallVector<CCValAssign, 16> ArgLocs;
1715    ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1716                      getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1717    CCInfo.AnalyzeCallOperands(Outs,
1718                               CCAssignFnForNode(CalleeCC, false, isVarArg));
1719    if (CCInfo.getNextStackOffset()) {
1720      MachineFunction &MF = DAG.getMachineFunction();
1721
1722      // Check if the arguments are already laid out in the right way as
1723      // the caller's fixed stack objects.
1724      MachineFrameInfo *MFI = MF.getFrameInfo();
1725      const MachineRegisterInfo *MRI = &MF.getRegInfo();
1726      const ARMInstrInfo *TII =
1727        ((ARMTargetMachine&)getTargetMachine()).getInstrInfo();
1728      for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1729           i != e;
1730           ++i, ++realArgIdx) {
1731        CCValAssign &VA = ArgLocs[i];
1732        EVT RegVT = VA.getLocVT();
1733        SDValue Arg = OutVals[realArgIdx];
1734        ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1735        if (VA.getLocInfo() == CCValAssign::Indirect)
1736          return false;
1737        if (VA.needsCustom()) {
1738          // f64 and vector types are split into multiple registers or
1739          // register/stack-slot combinations.  The types will not match
1740          // the registers; give up on memory f64 refs until we figure
1741          // out what to do about this.
1742          if (!VA.isRegLoc())
1743            return false;
1744          if (!ArgLocs[++i].isRegLoc())
1745            return false;
1746          if (RegVT == MVT::v2f64) {
1747            if (!ArgLocs[++i].isRegLoc())
1748              return false;
1749            if (!ArgLocs[++i].isRegLoc())
1750              return false;
1751          }
1752        } else if (!VA.isRegLoc()) {
1753          if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1754                                   MFI, MRI, TII))
1755            return false;
1756        }
1757      }
1758    }
1759  }
1760
1761  return true;
1762}
1763
1764SDValue
1765ARMTargetLowering::LowerReturn(SDValue Chain,
1766                               CallingConv::ID CallConv, bool isVarArg,
1767                               const SmallVectorImpl<ISD::OutputArg> &Outs,
1768                               const SmallVectorImpl<SDValue> &OutVals,
1769                               DebugLoc dl, SelectionDAG &DAG) const {
1770
1771  // CCValAssign - represent the assignment of the return value to a location.
1772  SmallVector<CCValAssign, 16> RVLocs;
1773
1774  // CCState - Info about the registers and stack slots.
1775  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1776                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1777
1778  // Analyze outgoing return values.
1779  CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1780                                               isVarArg));
1781
1782  // If this is the first return lowered for this function, add
1783  // the regs to the liveout set for the function.
1784  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1785    for (unsigned i = 0; i != RVLocs.size(); ++i)
1786      if (RVLocs[i].isRegLoc())
1787        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1788  }
1789
1790  SDValue Flag;
1791
1792  // Copy the result values into the output registers.
1793  for (unsigned i = 0, realRVLocIdx = 0;
1794       i != RVLocs.size();
1795       ++i, ++realRVLocIdx) {
1796    CCValAssign &VA = RVLocs[i];
1797    assert(VA.isRegLoc() && "Can only return in registers!");
1798
1799    SDValue Arg = OutVals[realRVLocIdx];
1800
1801    switch (VA.getLocInfo()) {
1802    default: llvm_unreachable("Unknown loc info!");
1803    case CCValAssign::Full: break;
1804    case CCValAssign::BCvt:
1805      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1806      break;
1807    }
1808
1809    if (VA.needsCustom()) {
1810      if (VA.getLocVT() == MVT::v2f64) {
1811        // Extract the first half and return it in two registers.
1812        SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1813                                   DAG.getConstant(0, MVT::i32));
1814        SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1815                                       DAG.getVTList(MVT::i32, MVT::i32), Half);
1816
1817        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1818        Flag = Chain.getValue(1);
1819        VA = RVLocs[++i]; // skip ahead to next loc
1820        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1821                                 HalfGPRs.getValue(1), Flag);
1822        Flag = Chain.getValue(1);
1823        VA = RVLocs[++i]; // skip ahead to next loc
1824
1825        // Extract the 2nd half and fall through to handle it as an f64 value.
1826        Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1827                          DAG.getConstant(1, MVT::i32));
1828      }
1829      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
1830      // available.
1831      SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1832                                  DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1833      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1834      Flag = Chain.getValue(1);
1835      VA = RVLocs[++i]; // skip ahead to next loc
1836      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1837                               Flag);
1838    } else
1839      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1840
1841    // Guarantee that all emitted copies are
1842    // stuck together, avoiding something bad.
1843    Flag = Chain.getValue(1);
1844  }
1845
1846  SDValue result;
1847  if (Flag.getNode())
1848    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1849  else // Return Void
1850    result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1851
1852  return result;
1853}
1854
1855bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1856  if (N->getNumValues() != 1)
1857    return false;
1858  if (!N->hasNUsesOfValue(1, 0))
1859    return false;
1860
1861  unsigned NumCopies = 0;
1862  SDNode* Copies[2];
1863  SDNode *Use = *N->use_begin();
1864  if (Use->getOpcode() == ISD::CopyToReg) {
1865    Copies[NumCopies++] = Use;
1866  } else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1867    // f64 returned in a pair of GPRs.
1868    for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1869         UI != UE; ++UI) {
1870      if (UI->getOpcode() != ISD::CopyToReg)
1871        return false;
1872      Copies[UI.getUse().getResNo()] = *UI;
1873      ++NumCopies;
1874    }
1875  } else if (Use->getOpcode() == ISD::BITCAST) {
1876    // f32 returned in a single GPR.
1877    if (!Use->hasNUsesOfValue(1, 0))
1878      return false;
1879    Use = *Use->use_begin();
1880    if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1881      return false;
1882    Copies[NumCopies++] = Use;
1883  } else {
1884    return false;
1885  }
1886
1887  if (NumCopies != 1 && NumCopies != 2)
1888    return false;
1889
1890  bool HasRet = false;
1891  for (unsigned i = 0; i < NumCopies; ++i) {
1892    SDNode *Copy = Copies[i];
1893    for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1894         UI != UE; ++UI) {
1895      if (UI->getOpcode() == ISD::CopyToReg) {
1896        SDNode *Use = *UI;
1897        if (Use == Copies[0] || Use == Copies[1])
1898          continue;
1899        return false;
1900      }
1901      if (UI->getOpcode() != ARMISD::RET_FLAG)
1902        return false;
1903      HasRet = true;
1904    }
1905  }
1906
1907  return HasRet;
1908}
1909
1910bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1911  if (!EnableARMTailCalls)
1912    return false;
1913
1914  if (!CI->isTailCall())
1915    return false;
1916
1917  return !Subtarget->isThumb1Only();
1918}
1919
1920// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
1921// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
1922// one of the above mentioned nodes. It has to be wrapped because otherwise
1923// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
1924// be used to form addressing mode. These wrapped nodes will be selected
1925// into MOVi.
1926static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
1927  EVT PtrVT = Op.getValueType();
1928  // FIXME there is no actual debug info here
1929  DebugLoc dl = Op.getDebugLoc();
1930  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1931  SDValue Res;
1932  if (CP->isMachineConstantPoolEntry())
1933    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
1934                                    CP->getAlignment());
1935  else
1936    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
1937                                    CP->getAlignment());
1938  return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
1939}
1940
1941unsigned ARMTargetLowering::getJumpTableEncoding() const {
1942  return MachineJumpTableInfo::EK_Inline;
1943}
1944
1945SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
1946                                             SelectionDAG &DAG) const {
1947  MachineFunction &MF = DAG.getMachineFunction();
1948  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1949  unsigned ARMPCLabelIndex = 0;
1950  DebugLoc DL = Op.getDebugLoc();
1951  EVT PtrVT = getPointerTy();
1952  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1953  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1954  SDValue CPAddr;
1955  if (RelocM == Reloc::Static) {
1956    CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
1957  } else {
1958    unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
1959    ARMPCLabelIndex = AFI->createPICLabelUId();
1960    ARMConstantPoolValue *CPV =
1961      ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
1962                                      ARMCP::CPBlockAddress, PCAdj);
1963    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1964  }
1965  CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
1966  SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
1967                               MachinePointerInfo::getConstantPool(),
1968                               false, false, 0);
1969  if (RelocM == Reloc::Static)
1970    return Result;
1971  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1972  return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
1973}
1974
1975// Lower ISD::GlobalTLSAddress using the "general dynamic" model
1976SDValue
1977ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
1978                                                 SelectionDAG &DAG) const {
1979  DebugLoc dl = GA->getDebugLoc();
1980  EVT PtrVT = getPointerTy();
1981  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
1982  MachineFunction &MF = DAG.getMachineFunction();
1983  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1984  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1985  ARMConstantPoolValue *CPV =
1986    ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
1987                                    ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
1988  SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
1989  Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
1990  Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
1991                         MachinePointerInfo::getConstantPool(),
1992                         false, false, 0);
1993  SDValue Chain = Argument.getValue(1);
1994
1995  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1996  Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
1997
1998  // call __tls_get_addr.
1999  ArgListTy Args;
2000  ArgListEntry Entry;
2001  Entry.Node = Argument;
2002  Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2003  Args.push_back(Entry);
2004  // FIXME: is there useful debug info available here?
2005  std::pair<SDValue, SDValue> CallResult =
2006    LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2007                false, false, false, false,
2008                0, CallingConv::C, false, /*isReturnValueUsed=*/true,
2009                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2010  return CallResult.first;
2011}
2012
2013// Lower ISD::GlobalTLSAddress using the "initial exec" or
2014// "local exec" model.
2015SDValue
2016ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2017                                        SelectionDAG &DAG) const {
2018  const GlobalValue *GV = GA->getGlobal();
2019  DebugLoc dl = GA->getDebugLoc();
2020  SDValue Offset;
2021  SDValue Chain = DAG.getEntryNode();
2022  EVT PtrVT = getPointerTy();
2023  // Get the Thread Pointer
2024  SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2025
2026  if (GV->isDeclaration()) {
2027    MachineFunction &MF = DAG.getMachineFunction();
2028    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2029    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2030    // Initial exec model.
2031    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2032    ARMConstantPoolValue *CPV =
2033      ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2034                                      ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2035                                      true);
2036    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2037    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2038    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2039                         MachinePointerInfo::getConstantPool(),
2040                         false, false, 0);
2041    Chain = Offset.getValue(1);
2042
2043    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2044    Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2045
2046    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2047                         MachinePointerInfo::getConstantPool(),
2048                         false, false, 0);
2049  } else {
2050    // local exec model
2051    ARMConstantPoolValue *CPV =
2052      ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2053    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2054    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2055    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2056                         MachinePointerInfo::getConstantPool(),
2057                         false, false, 0);
2058  }
2059
2060  // The address of the thread local variable is the add of the thread
2061  // pointer with the offset of the variable.
2062  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2063}
2064
2065SDValue
2066ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2067  // TODO: implement the "local dynamic" model
2068  assert(Subtarget->isTargetELF() &&
2069         "TLS not implemented for non-ELF targets");
2070  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2071  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
2072  // otherwise use the "Local Exec" TLS Model
2073  if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2074    return LowerToTLSGeneralDynamicModel(GA, DAG);
2075  else
2076    return LowerToTLSExecModels(GA, DAG);
2077}
2078
2079SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2080                                                 SelectionDAG &DAG) const {
2081  EVT PtrVT = getPointerTy();
2082  DebugLoc dl = Op.getDebugLoc();
2083  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2084  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2085  if (RelocM == Reloc::PIC_) {
2086    bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2087    ARMConstantPoolValue *CPV =
2088      ARMConstantPoolConstant::Create(GV,
2089                                      UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2090    SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2091    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2092    SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2093                                 CPAddr,
2094                                 MachinePointerInfo::getConstantPool(),
2095                                 false, false, 0);
2096    SDValue Chain = Result.getValue(1);
2097    SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2098    Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2099    if (!UseGOTOFF)
2100      Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2101                           MachinePointerInfo::getGOT(), false, false, 0);
2102    return Result;
2103  }
2104
2105  // If we have T2 ops, we can materialize the address directly via movt/movw
2106  // pair. This is always cheaper.
2107  if (Subtarget->useMovt()) {
2108    ++NumMovwMovt;
2109    // FIXME: Once remat is capable of dealing with instructions with register
2110    // operands, expand this into two nodes.
2111    return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2112                       DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2113  } else {
2114    SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2115    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2116    return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2117                       MachinePointerInfo::getConstantPool(),
2118                       false, false, 0);
2119  }
2120}
2121
2122SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2123                                                    SelectionDAG &DAG) const {
2124  EVT PtrVT = getPointerTy();
2125  DebugLoc dl = Op.getDebugLoc();
2126  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2127  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2128  MachineFunction &MF = DAG.getMachineFunction();
2129  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2130
2131  // FIXME: Enable this for static codegen when tool issues are fixed.
2132  if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2133    ++NumMovwMovt;
2134    // FIXME: Once remat is capable of dealing with instructions with register
2135    // operands, expand this into two nodes.
2136    if (RelocM == Reloc::Static)
2137      return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2138                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2139
2140    unsigned Wrapper = (RelocM == Reloc::PIC_)
2141      ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2142    SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2143                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2144    if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2145      Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2146                           MachinePointerInfo::getGOT(), false, false, 0);
2147    return Result;
2148  }
2149
2150  unsigned ARMPCLabelIndex = 0;
2151  SDValue CPAddr;
2152  if (RelocM == Reloc::Static) {
2153    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2154  } else {
2155    ARMPCLabelIndex = AFI->createPICLabelUId();
2156    unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2157    ARMConstantPoolValue *CPV =
2158      ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2159                                      PCAdj);
2160    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2161  }
2162  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2163
2164  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2165                               MachinePointerInfo::getConstantPool(),
2166                               false, false, 0);
2167  SDValue Chain = Result.getValue(1);
2168
2169  if (RelocM == Reloc::PIC_) {
2170    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2171    Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2172  }
2173
2174  if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2175    Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2176                         false, false, 0);
2177
2178  return Result;
2179}
2180
2181SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2182                                                    SelectionDAG &DAG) const {
2183  assert(Subtarget->isTargetELF() &&
2184         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2185  MachineFunction &MF = DAG.getMachineFunction();
2186  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2187  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2188  EVT PtrVT = getPointerTy();
2189  DebugLoc dl = Op.getDebugLoc();
2190  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2191  ARMConstantPoolValue *CPV =
2192    ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2193                                  ARMPCLabelIndex, PCAdj);
2194  SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2195  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2196  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2197                               MachinePointerInfo::getConstantPool(),
2198                               false, false, 0);
2199  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2200  return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2201}
2202
2203SDValue
2204ARMTargetLowering::LowerEH_SJLJ_DISPATCHSETUP(SDValue Op, SelectionDAG &DAG)
2205  const {
2206  DebugLoc dl = Op.getDebugLoc();
2207  return DAG.getNode(ARMISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
2208                     Op.getOperand(0), Op.getOperand(1));
2209}
2210
2211SDValue
2212ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2213  DebugLoc dl = Op.getDebugLoc();
2214  SDValue Val = DAG.getConstant(0, MVT::i32);
2215  return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2216                     DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2217                     Op.getOperand(1), Val);
2218}
2219
2220SDValue
2221ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2222  DebugLoc dl = Op.getDebugLoc();
2223  return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2224                     Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2225}
2226
2227SDValue
2228ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2229                                          const ARMSubtarget *Subtarget) const {
2230  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2231  DebugLoc dl = Op.getDebugLoc();
2232  switch (IntNo) {
2233  default: return SDValue();    // Don't custom lower most intrinsics.
2234  case Intrinsic::arm_thread_pointer: {
2235    EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2236    return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2237  }
2238  case Intrinsic::eh_sjlj_lsda: {
2239    MachineFunction &MF = DAG.getMachineFunction();
2240    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2241    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2242    EVT PtrVT = getPointerTy();
2243    DebugLoc dl = Op.getDebugLoc();
2244    Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2245    SDValue CPAddr;
2246    unsigned PCAdj = (RelocM != Reloc::PIC_)
2247      ? 0 : (Subtarget->isThumb() ? 4 : 8);
2248    ARMConstantPoolValue *CPV =
2249      ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2250                                      ARMCP::CPLSDA, PCAdj);
2251    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2252    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2253    SDValue Result =
2254      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2255                  MachinePointerInfo::getConstantPool(),
2256                  false, false, 0);
2257
2258    if (RelocM == Reloc::PIC_) {
2259      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2260      Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2261    }
2262    return Result;
2263  }
2264  case Intrinsic::arm_neon_vmulls:
2265  case Intrinsic::arm_neon_vmullu: {
2266    unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2267      ? ARMISD::VMULLs : ARMISD::VMULLu;
2268    return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2269                       Op.getOperand(1), Op.getOperand(2));
2270  }
2271  }
2272}
2273
2274static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2275                               const ARMSubtarget *Subtarget) {
2276  DebugLoc dl = Op.getDebugLoc();
2277  if (!Subtarget->hasDataBarrier()) {
2278    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2279    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2280    // here.
2281    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2282           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2283    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2284                       DAG.getConstant(0, MVT::i32));
2285  }
2286
2287  SDValue Op5 = Op.getOperand(5);
2288  bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2289  unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2290  unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2291  bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2292
2293  ARM_MB::MemBOpt DMBOpt;
2294  if (isDeviceBarrier)
2295    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2296  else
2297    DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2298  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2299                     DAG.getConstant(DMBOpt, MVT::i32));
2300}
2301
2302
2303static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2304                                 const ARMSubtarget *Subtarget) {
2305  // FIXME: handle "fence singlethread" more efficiently.
2306  DebugLoc dl = Op.getDebugLoc();
2307  if (!Subtarget->hasDataBarrier()) {
2308    // Some ARMv6 cpus can support data barriers with an mcr instruction.
2309    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2310    // here.
2311    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2312           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2313    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2314                       DAG.getConstant(0, MVT::i32));
2315  }
2316
2317  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2318                     DAG.getConstant(ARM_MB::ISH, MVT::i32));
2319}
2320
2321static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2322                             const ARMSubtarget *Subtarget) {
2323  // ARM pre v5TE and Thumb1 does not have preload instructions.
2324  if (!(Subtarget->isThumb2() ||
2325        (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2326    // Just preserve the chain.
2327    return Op.getOperand(0);
2328
2329  DebugLoc dl = Op.getDebugLoc();
2330  unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2331  if (!isRead &&
2332      (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2333    // ARMv7 with MP extension has PLDW.
2334    return Op.getOperand(0);
2335
2336  unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2337  if (Subtarget->isThumb()) {
2338    // Invert the bits.
2339    isRead = ~isRead & 1;
2340    isData = ~isData & 1;
2341  }
2342
2343  return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2344                     Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2345                     DAG.getConstant(isData, MVT::i32));
2346}
2347
2348static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2349  MachineFunction &MF = DAG.getMachineFunction();
2350  ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2351
2352  // vastart just stores the address of the VarArgsFrameIndex slot into the
2353  // memory location argument.
2354  DebugLoc dl = Op.getDebugLoc();
2355  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2356  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2357  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2358  return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2359                      MachinePointerInfo(SV), false, false, 0);
2360}
2361
2362SDValue
2363ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2364                                        SDValue &Root, SelectionDAG &DAG,
2365                                        DebugLoc dl) const {
2366  MachineFunction &MF = DAG.getMachineFunction();
2367  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2368
2369  TargetRegisterClass *RC;
2370  if (AFI->isThumb1OnlyFunction())
2371    RC = ARM::tGPRRegisterClass;
2372  else
2373    RC = ARM::GPRRegisterClass;
2374
2375  // Transform the arguments stored in physical registers into virtual ones.
2376  unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2377  SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2378
2379  SDValue ArgValue2;
2380  if (NextVA.isMemLoc()) {
2381    MachineFrameInfo *MFI = MF.getFrameInfo();
2382    int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2383
2384    // Create load node to retrieve arguments from the stack.
2385    SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2386    ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2387                            MachinePointerInfo::getFixedStack(FI),
2388                            false, false, 0);
2389  } else {
2390    Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2391    ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2392  }
2393
2394  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2395}
2396
2397void
2398ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2399                                  unsigned &VARegSize, unsigned &VARegSaveSize)
2400  const {
2401  unsigned NumGPRs;
2402  if (CCInfo.isFirstByValRegValid())
2403    NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2404  else {
2405    unsigned int firstUnalloced;
2406    firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2407                                                sizeof(GPRArgRegs) /
2408                                                sizeof(GPRArgRegs[0]));
2409    NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2410  }
2411
2412  unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2413  VARegSize = NumGPRs * 4;
2414  VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2415}
2416
2417// The remaining GPRs hold either the beginning of variable-argument
2418// data, or the beginning of an aggregate passed by value (usuall
2419// byval).  Either way, we allocate stack slots adjacent to the data
2420// provided by our caller, and store the unallocated registers there.
2421// If this is a variadic function, the va_list pointer will begin with
2422// these values; otherwise, this reassembles a (byval) structure that
2423// was split between registers and memory.
2424void
2425ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2426                                        DebugLoc dl, SDValue &Chain,
2427                                        unsigned ArgOffset) const {
2428  MachineFunction &MF = DAG.getMachineFunction();
2429  MachineFrameInfo *MFI = MF.getFrameInfo();
2430  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2431  unsigned firstRegToSaveIndex;
2432  if (CCInfo.isFirstByValRegValid())
2433    firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2434  else {
2435    firstRegToSaveIndex = CCInfo.getFirstUnallocated
2436      (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2437  }
2438
2439  unsigned VARegSize, VARegSaveSize;
2440  computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2441  if (VARegSaveSize) {
2442    // If this function is vararg, store any remaining integer argument regs
2443    // to their spots on the stack so that they may be loaded by deferencing
2444    // the result of va_next.
2445    AFI->setVarArgsRegSaveSize(VARegSaveSize);
2446    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2447                                                     ArgOffset + VARegSaveSize
2448                                                     - VARegSize,
2449                                                     false));
2450    SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2451                                    getPointerTy());
2452
2453    SmallVector<SDValue, 4> MemOps;
2454    for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2455      TargetRegisterClass *RC;
2456      if (AFI->isThumb1OnlyFunction())
2457        RC = ARM::tGPRRegisterClass;
2458      else
2459        RC = ARM::GPRRegisterClass;
2460
2461      unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2462      SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2463      SDValue Store =
2464        DAG.getStore(Val.getValue(1), dl, Val, FIN,
2465                 MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2466                     false, false, 0);
2467      MemOps.push_back(Store);
2468      FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2469                        DAG.getConstant(4, getPointerTy()));
2470    }
2471    if (!MemOps.empty())
2472      Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2473                          &MemOps[0], MemOps.size());
2474  } else
2475    // This will point to the next argument passed via stack.
2476    AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2477}
2478
2479SDValue
2480ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2481                                        CallingConv::ID CallConv, bool isVarArg,
2482                                        const SmallVectorImpl<ISD::InputArg>
2483                                          &Ins,
2484                                        DebugLoc dl, SelectionDAG &DAG,
2485                                        SmallVectorImpl<SDValue> &InVals)
2486                                          const {
2487  MachineFunction &MF = DAG.getMachineFunction();
2488  MachineFrameInfo *MFI = MF.getFrameInfo();
2489
2490  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2491
2492  // Assign locations to all of the incoming arguments.
2493  SmallVector<CCValAssign, 16> ArgLocs;
2494  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2495                    getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2496  CCInfo.AnalyzeFormalArguments(Ins,
2497                                CCAssignFnForNode(CallConv, /* Return*/ false,
2498                                                  isVarArg));
2499
2500  SmallVector<SDValue, 16> ArgValues;
2501  int lastInsIndex = -1;
2502
2503  SDValue ArgValue;
2504  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2505    CCValAssign &VA = ArgLocs[i];
2506
2507    // Arguments stored in registers.
2508    if (VA.isRegLoc()) {
2509      EVT RegVT = VA.getLocVT();
2510
2511      if (VA.needsCustom()) {
2512        // f64 and vector types are split up into multiple registers or
2513        // combinations of registers and stack slots.
2514        if (VA.getLocVT() == MVT::v2f64) {
2515          SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2516                                                   Chain, DAG, dl);
2517          VA = ArgLocs[++i]; // skip ahead to next loc
2518          SDValue ArgValue2;
2519          if (VA.isMemLoc()) {
2520            int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2521            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2522            ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2523                                    MachinePointerInfo::getFixedStack(FI),
2524                                    false, false, 0);
2525          } else {
2526            ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2527                                             Chain, DAG, dl);
2528          }
2529          ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2530          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2531                                 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2532          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2533                                 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2534        } else
2535          ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2536
2537      } else {
2538        TargetRegisterClass *RC;
2539
2540        if (RegVT == MVT::f32)
2541          RC = ARM::SPRRegisterClass;
2542        else if (RegVT == MVT::f64)
2543          RC = ARM::DPRRegisterClass;
2544        else if (RegVT == MVT::v2f64)
2545          RC = ARM::QPRRegisterClass;
2546        else if (RegVT == MVT::i32)
2547          RC = (AFI->isThumb1OnlyFunction() ?
2548                ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2549        else
2550          llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2551
2552        // Transform the arguments in physical registers into virtual ones.
2553        unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2554        ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2555      }
2556
2557      // If this is an 8 or 16-bit value, it is really passed promoted
2558      // to 32 bits.  Insert an assert[sz]ext to capture this, then
2559      // truncate to the right size.
2560      switch (VA.getLocInfo()) {
2561      default: llvm_unreachable("Unknown loc info!");
2562      case CCValAssign::Full: break;
2563      case CCValAssign::BCvt:
2564        ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2565        break;
2566      case CCValAssign::SExt:
2567        ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2568                               DAG.getValueType(VA.getValVT()));
2569        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2570        break;
2571      case CCValAssign::ZExt:
2572        ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2573                               DAG.getValueType(VA.getValVT()));
2574        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2575        break;
2576      }
2577
2578      InVals.push_back(ArgValue);
2579
2580    } else { // VA.isRegLoc()
2581
2582      // sanity check
2583      assert(VA.isMemLoc());
2584      assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2585
2586      int index = ArgLocs[i].getValNo();
2587
2588      // Some Ins[] entries become multiple ArgLoc[] entries.
2589      // Process them only once.
2590      if (index != lastInsIndex)
2591        {
2592          ISD::ArgFlagsTy Flags = Ins[index].Flags;
2593          // FIXME: For now, all byval parameter objects are marked mutable.
2594          // This can be changed with more analysis.
2595          // In case of tail call optimization mark all arguments mutable.
2596          // Since they could be overwritten by lowering of arguments in case of
2597          // a tail call.
2598          if (Flags.isByVal()) {
2599            unsigned VARegSize, VARegSaveSize;
2600            computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2601            VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2602            unsigned Bytes = Flags.getByValSize() - VARegSize;
2603            if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2604            int FI = MFI->CreateFixedObject(Bytes,
2605                                            VA.getLocMemOffset(), false);
2606            InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2607          } else {
2608            int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2609                                            VA.getLocMemOffset(), true);
2610
2611            // Create load nodes to retrieve arguments from the stack.
2612            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2613            InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2614                                         MachinePointerInfo::getFixedStack(FI),
2615                                         false, false, 0));
2616          }
2617          lastInsIndex = index;
2618        }
2619    }
2620  }
2621
2622  // varargs
2623  if (isVarArg)
2624    VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2625
2626  return Chain;
2627}
2628
2629/// isFloatingPointZero - Return true if this is +0.0.
2630static bool isFloatingPointZero(SDValue Op) {
2631  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2632    return CFP->getValueAPF().isPosZero();
2633  else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2634    // Maybe this has already been legalized into the constant pool?
2635    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2636      SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2637      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2638        if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2639          return CFP->getValueAPF().isPosZero();
2640    }
2641  }
2642  return false;
2643}
2644
2645/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2646/// the given operands.
2647SDValue
2648ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2649                             SDValue &ARMcc, SelectionDAG &DAG,
2650                             DebugLoc dl) const {
2651  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2652    unsigned C = RHSC->getZExtValue();
2653    if (!isLegalICmpImmediate(C)) {
2654      // Constant does not fit, try adjusting it by one?
2655      switch (CC) {
2656      default: break;
2657      case ISD::SETLT:
2658      case ISD::SETGE:
2659        if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2660          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2661          RHS = DAG.getConstant(C-1, MVT::i32);
2662        }
2663        break;
2664      case ISD::SETULT:
2665      case ISD::SETUGE:
2666        if (C != 0 && isLegalICmpImmediate(C-1)) {
2667          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2668          RHS = DAG.getConstant(C-1, MVT::i32);
2669        }
2670        break;
2671      case ISD::SETLE:
2672      case ISD::SETGT:
2673        if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2674          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2675          RHS = DAG.getConstant(C+1, MVT::i32);
2676        }
2677        break;
2678      case ISD::SETULE:
2679      case ISD::SETUGT:
2680        if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2681          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2682          RHS = DAG.getConstant(C+1, MVT::i32);
2683        }
2684        break;
2685      }
2686    }
2687  }
2688
2689  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2690  ARMISD::NodeType CompareType;
2691  switch (CondCode) {
2692  default:
2693    CompareType = ARMISD::CMP;
2694    break;
2695  case ARMCC::EQ:
2696  case ARMCC::NE:
2697    // Uses only Z Flag
2698    CompareType = ARMISD::CMPZ;
2699    break;
2700  }
2701  ARMcc = DAG.getConstant(CondCode, MVT::i32);
2702  return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2703}
2704
2705/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2706SDValue
2707ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2708                             DebugLoc dl) const {
2709  SDValue Cmp;
2710  if (!isFloatingPointZero(RHS))
2711    Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2712  else
2713    Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2714  return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2715}
2716
2717/// duplicateCmp - Glue values can have only one use, so this function
2718/// duplicates a comparison node.
2719SDValue
2720ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2721  unsigned Opc = Cmp.getOpcode();
2722  DebugLoc DL = Cmp.getDebugLoc();
2723  if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2724    return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2725
2726  assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2727  Cmp = Cmp.getOperand(0);
2728  Opc = Cmp.getOpcode();
2729  if (Opc == ARMISD::CMPFP)
2730    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2731  else {
2732    assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2733    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2734  }
2735  return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2736}
2737
2738SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2739  SDValue Cond = Op.getOperand(0);
2740  SDValue SelectTrue = Op.getOperand(1);
2741  SDValue SelectFalse = Op.getOperand(2);
2742  DebugLoc dl = Op.getDebugLoc();
2743
2744  // Convert:
2745  //
2746  //   (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2747  //   (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2748  //
2749  if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2750    const ConstantSDNode *CMOVTrue =
2751      dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2752    const ConstantSDNode *CMOVFalse =
2753      dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2754
2755    if (CMOVTrue && CMOVFalse) {
2756      unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2757      unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2758
2759      SDValue True;
2760      SDValue False;
2761      if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2762        True = SelectTrue;
2763        False = SelectFalse;
2764      } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2765        True = SelectFalse;
2766        False = SelectTrue;
2767      }
2768
2769      if (True.getNode() && False.getNode()) {
2770        EVT VT = Op.getValueType();
2771        SDValue ARMcc = Cond.getOperand(2);
2772        SDValue CCR = Cond.getOperand(3);
2773        SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2774        assert(True.getValueType() == VT);
2775        return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2776      }
2777    }
2778  }
2779
2780  return DAG.getSelectCC(dl, Cond,
2781                         DAG.getConstant(0, Cond.getValueType()),
2782                         SelectTrue, SelectFalse, ISD::SETNE);
2783}
2784
2785SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2786  EVT VT = Op.getValueType();
2787  SDValue LHS = Op.getOperand(0);
2788  SDValue RHS = Op.getOperand(1);
2789  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2790  SDValue TrueVal = Op.getOperand(2);
2791  SDValue FalseVal = Op.getOperand(3);
2792  DebugLoc dl = Op.getDebugLoc();
2793
2794  if (LHS.getValueType() == MVT::i32) {
2795    SDValue ARMcc;
2796    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2797    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2798    return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2799  }
2800
2801  ARMCC::CondCodes CondCode, CondCode2;
2802  FPCCToARMCC(CC, CondCode, CondCode2);
2803
2804  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2805  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2806  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2807  SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2808                               ARMcc, CCR, Cmp);
2809  if (CondCode2 != ARMCC::AL) {
2810    SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2811    // FIXME: Needs another CMP because flag can have but one use.
2812    SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2813    Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2814                         Result, TrueVal, ARMcc2, CCR, Cmp2);
2815  }
2816  return Result;
2817}
2818
2819/// canChangeToInt - Given the fp compare operand, return true if it is suitable
2820/// to morph to an integer compare sequence.
2821static bool canChangeToInt(SDValue Op, bool &SeenZero,
2822                           const ARMSubtarget *Subtarget) {
2823  SDNode *N = Op.getNode();
2824  if (!N->hasOneUse())
2825    // Otherwise it requires moving the value from fp to integer registers.
2826    return false;
2827  if (!N->getNumValues())
2828    return false;
2829  EVT VT = Op.getValueType();
2830  if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2831    // f32 case is generally profitable. f64 case only makes sense when vcmpe +
2832    // vmrs are very slow, e.g. cortex-a8.
2833    return false;
2834
2835  if (isFloatingPointZero(Op)) {
2836    SeenZero = true;
2837    return true;
2838  }
2839  return ISD::isNormalLoad(N);
2840}
2841
2842static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2843  if (isFloatingPointZero(Op))
2844    return DAG.getConstant(0, MVT::i32);
2845
2846  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2847    return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2848                       Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2849                       Ld->isVolatile(), Ld->isNonTemporal(),
2850                       Ld->getAlignment());
2851
2852  llvm_unreachable("Unknown VFP cmp argument!");
2853}
2854
2855static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2856                           SDValue &RetVal1, SDValue &RetVal2) {
2857  if (isFloatingPointZero(Op)) {
2858    RetVal1 = DAG.getConstant(0, MVT::i32);
2859    RetVal2 = DAG.getConstant(0, MVT::i32);
2860    return;
2861  }
2862
2863  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2864    SDValue Ptr = Ld->getBasePtr();
2865    RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2866                          Ld->getChain(), Ptr,
2867                          Ld->getPointerInfo(),
2868                          Ld->isVolatile(), Ld->isNonTemporal(),
2869                          Ld->getAlignment());
2870
2871    EVT PtrType = Ptr.getValueType();
2872    unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2873    SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2874                                 PtrType, Ptr, DAG.getConstant(4, PtrType));
2875    RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2876                          Ld->getChain(), NewPtr,
2877                          Ld->getPointerInfo().getWithOffset(4),
2878                          Ld->isVolatile(), Ld->isNonTemporal(),
2879                          NewAlign);
2880    return;
2881  }
2882
2883  llvm_unreachable("Unknown VFP cmp argument!");
2884}
2885
2886/// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2887/// f32 and even f64 comparisons to integer ones.
2888SDValue
2889ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2890  SDValue Chain = Op.getOperand(0);
2891  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2892  SDValue LHS = Op.getOperand(2);
2893  SDValue RHS = Op.getOperand(3);
2894  SDValue Dest = Op.getOperand(4);
2895  DebugLoc dl = Op.getDebugLoc();
2896
2897  bool SeenZero = false;
2898  if (canChangeToInt(LHS, SeenZero, Subtarget) &&
2899      canChangeToInt(RHS, SeenZero, Subtarget) &&
2900      // If one of the operand is zero, it's safe to ignore the NaN case since
2901      // we only care about equality comparisons.
2902      (SeenZero || (DAG.isKnownNeverNaN(LHS) && DAG.isKnownNeverNaN(RHS)))) {
2903    // If unsafe fp math optimization is enabled and there are no other uses of
2904    // the CMP operands, and the condition code is EQ or NE, we can optimize it
2905    // to an integer comparison.
2906    if (CC == ISD::SETOEQ)
2907      CC = ISD::SETEQ;
2908    else if (CC == ISD::SETUNE)
2909      CC = ISD::SETNE;
2910
2911    SDValue ARMcc;
2912    if (LHS.getValueType() == MVT::f32) {
2913      LHS = bitcastf32Toi32(LHS, DAG);
2914      RHS = bitcastf32Toi32(RHS, DAG);
2915      SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2916      SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2917      return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2918                         Chain, Dest, ARMcc, CCR, Cmp);
2919    }
2920
2921    SDValue LHS1, LHS2;
2922    SDValue RHS1, RHS2;
2923    expandf64Toi32(LHS, DAG, LHS1, LHS2);
2924    expandf64Toi32(RHS, DAG, RHS1, RHS2);
2925    ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2926    ARMcc = DAG.getConstant(CondCode, MVT::i32);
2927    SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2928    SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
2929    return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
2930  }
2931
2932  return SDValue();
2933}
2934
2935SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2936  SDValue Chain = Op.getOperand(0);
2937  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2938  SDValue LHS = Op.getOperand(2);
2939  SDValue RHS = Op.getOperand(3);
2940  SDValue Dest = Op.getOperand(4);
2941  DebugLoc dl = Op.getDebugLoc();
2942
2943  if (LHS.getValueType() == MVT::i32) {
2944    SDValue ARMcc;
2945    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2946    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2947    return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
2948                       Chain, Dest, ARMcc, CCR, Cmp);
2949  }
2950
2951  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
2952
2953  if (UnsafeFPMath &&
2954      (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
2955       CC == ISD::SETNE || CC == ISD::SETUNE)) {
2956    SDValue Result = OptimizeVFPBrcond(Op, DAG);
2957    if (Result.getNode())
2958      return Result;
2959  }
2960
2961  ARMCC::CondCodes CondCode, CondCode2;
2962  FPCCToARMCC(CC, CondCode, CondCode2);
2963
2964  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2965  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2966  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2967  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
2968  SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
2969  SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2970  if (CondCode2 != ARMCC::AL) {
2971    ARMcc = DAG.getConstant(CondCode2, MVT::i32);
2972    SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
2973    Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
2974  }
2975  return Res;
2976}
2977
2978SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
2979  SDValue Chain = Op.getOperand(0);
2980  SDValue Table = Op.getOperand(1);
2981  SDValue Index = Op.getOperand(2);
2982  DebugLoc dl = Op.getDebugLoc();
2983
2984  EVT PTy = getPointerTy();
2985  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
2986  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
2987  SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
2988  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
2989  Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
2990  Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
2991  SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
2992  if (Subtarget->isThumb2()) {
2993    // Thumb2 uses a two-level jump. That is, it jumps into the jump table
2994    // which does another jump to the destination. This also makes it easier
2995    // to translate it to TBB / TBH later.
2996    // FIXME: This might not work if the function is extremely large.
2997    return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
2998                       Addr, Op.getOperand(2), JTI, UId);
2999  }
3000  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3001    Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3002                       MachinePointerInfo::getJumpTable(),
3003                       false, false, 0);
3004    Chain = Addr.getValue(1);
3005    Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3006    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3007  } else {
3008    Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3009                       MachinePointerInfo::getJumpTable(), false, false, 0);
3010    Chain = Addr.getValue(1);
3011    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3012  }
3013}
3014
3015static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3016  DebugLoc dl = Op.getDebugLoc();
3017  unsigned Opc;
3018
3019  switch (Op.getOpcode()) {
3020  default:
3021    assert(0 && "Invalid opcode!");
3022  case ISD::FP_TO_SINT:
3023    Opc = ARMISD::FTOSI;
3024    break;
3025  case ISD::FP_TO_UINT:
3026    Opc = ARMISD::FTOUI;
3027    break;
3028  }
3029  Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3030  return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3031}
3032
3033static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3034  EVT VT = Op.getValueType();
3035  DebugLoc dl = Op.getDebugLoc();
3036
3037  assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3038         "Invalid type for custom lowering!");
3039  if (VT != MVT::v4f32)
3040    return DAG.UnrollVectorOp(Op.getNode());
3041
3042  unsigned CastOpc;
3043  unsigned Opc;
3044  switch (Op.getOpcode()) {
3045  default:
3046    assert(0 && "Invalid opcode!");
3047  case ISD::SINT_TO_FP:
3048    CastOpc = ISD::SIGN_EXTEND;
3049    Opc = ISD::SINT_TO_FP;
3050    break;
3051  case ISD::UINT_TO_FP:
3052    CastOpc = ISD::ZERO_EXTEND;
3053    Opc = ISD::UINT_TO_FP;
3054    break;
3055  }
3056
3057  Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3058  return DAG.getNode(Opc, dl, VT, Op);
3059}
3060
3061static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3062  EVT VT = Op.getValueType();
3063  if (VT.isVector())
3064    return LowerVectorINT_TO_FP(Op, DAG);
3065
3066  DebugLoc dl = Op.getDebugLoc();
3067  unsigned Opc;
3068
3069  switch (Op.getOpcode()) {
3070  default:
3071    assert(0 && "Invalid opcode!");
3072  case ISD::SINT_TO_FP:
3073    Opc = ARMISD::SITOF;
3074    break;
3075  case ISD::UINT_TO_FP:
3076    Opc = ARMISD::UITOF;
3077    break;
3078  }
3079
3080  Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3081  return DAG.getNode(Opc, dl, VT, Op);
3082}
3083
3084SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3085  // Implement fcopysign with a fabs and a conditional fneg.
3086  SDValue Tmp0 = Op.getOperand(0);
3087  SDValue Tmp1 = Op.getOperand(1);
3088  DebugLoc dl = Op.getDebugLoc();
3089  EVT VT = Op.getValueType();
3090  EVT SrcVT = Tmp1.getValueType();
3091  bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3092    Tmp0.getOpcode() == ARMISD::VMOVDRR;
3093  bool UseNEON = !InGPR && Subtarget->hasNEON();
3094
3095  if (UseNEON) {
3096    // Use VBSL to copy the sign bit.
3097    unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3098    SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3099                               DAG.getTargetConstant(EncodedVal, MVT::i32));
3100    EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3101    if (VT == MVT::f64)
3102      Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3103                         DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3104                         DAG.getConstant(32, MVT::i32));
3105    else /*if (VT == MVT::f32)*/
3106      Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3107    if (SrcVT == MVT::f32) {
3108      Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3109      if (VT == MVT::f64)
3110        Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3111                           DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3112                           DAG.getConstant(32, MVT::i32));
3113    } else if (VT == MVT::f32)
3114      Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3115                         DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3116                         DAG.getConstant(32, MVT::i32));
3117    Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3118    Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3119
3120    SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3121                                            MVT::i32);
3122    AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3123    SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3124                                  DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3125
3126    SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3127                              DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3128                              DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3129    if (VT == MVT::f32) {
3130      Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3131      Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3132                        DAG.getConstant(0, MVT::i32));
3133    } else {
3134      Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3135    }
3136
3137    return Res;
3138  }
3139
3140  // Bitcast operand 1 to i32.
3141  if (SrcVT == MVT::f64)
3142    Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3143                       &Tmp1, 1).getValue(1);
3144  Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3145
3146  // Or in the signbit with integer operations.
3147  SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3148  SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3149  Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3150  if (VT == MVT::f32) {
3151    Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3152                       DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3153    return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3154                       DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3155  }
3156
3157  // f64: Or the high part with signbit and then combine two parts.
3158  Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3159                     &Tmp0, 1);
3160  SDValue Lo = Tmp0.getValue(0);
3161  SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3162  Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3163  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3164}
3165
3166SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3167  MachineFunction &MF = DAG.getMachineFunction();
3168  MachineFrameInfo *MFI = MF.getFrameInfo();
3169  MFI->setReturnAddressIsTaken(true);
3170
3171  EVT VT = Op.getValueType();
3172  DebugLoc dl = Op.getDebugLoc();
3173  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3174  if (Depth) {
3175    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3176    SDValue Offset = DAG.getConstant(4, MVT::i32);
3177    return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3178                       DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3179                       MachinePointerInfo(), false, false, 0);
3180  }
3181
3182  // Return LR, which contains the return address. Mark it an implicit live-in.
3183  unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3184  return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3185}
3186
3187SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3188  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3189  MFI->setFrameAddressIsTaken(true);
3190
3191  EVT VT = Op.getValueType();
3192  DebugLoc dl = Op.getDebugLoc();  // FIXME probably not meaningful
3193  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3194  unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3195    ? ARM::R7 : ARM::R11;
3196  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3197  while (Depth--)
3198    FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3199                            MachinePointerInfo(),
3200                            false, false, 0);
3201  return FrameAddr;
3202}
3203
3204/// ExpandBITCAST - If the target supports VFP, this function is called to
3205/// expand a bit convert where either the source or destination type is i64 to
3206/// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
3207/// operand type is illegal (e.g., v2f32 for a target that doesn't support
3208/// vectors), since the legalizer won't know what to do with that.
3209static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3210  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3211  DebugLoc dl = N->getDebugLoc();
3212  SDValue Op = N->getOperand(0);
3213
3214  // This function is only supposed to be called for i64 types, either as the
3215  // source or destination of the bit convert.
3216  EVT SrcVT = Op.getValueType();
3217  EVT DstVT = N->getValueType(0);
3218  assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3219         "ExpandBITCAST called for non-i64 type");
3220
3221  // Turn i64->f64 into VMOVDRR.
3222  if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3223    SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3224                             DAG.getConstant(0, MVT::i32));
3225    SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3226                             DAG.getConstant(1, MVT::i32));
3227    return DAG.getNode(ISD::BITCAST, dl, DstVT,
3228                       DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3229  }
3230
3231  // Turn f64->i64 into VMOVRRD.
3232  if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3233    SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3234                              DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3235    // Merge the pieces into a single i64 value.
3236    return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3237  }
3238
3239  return SDValue();
3240}
3241
3242/// getZeroVector - Returns a vector of specified type with all zero elements.
3243/// Zero vectors are used to represent vector negation and in those cases
3244/// will be implemented with the NEON VNEG instruction.  However, VNEG does
3245/// not support i64 elements, so sometimes the zero vectors will need to be
3246/// explicitly constructed.  Regardless, use a canonical VMOV to create the
3247/// zero vector.
3248static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3249  assert(VT.isVector() && "Expected a vector type");
3250  // The canonical modified immediate encoding of a zero vector is....0!
3251  SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3252  EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3253  SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3254  return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3255}
3256
3257/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3258/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3259SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3260                                                SelectionDAG &DAG) const {
3261  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3262  EVT VT = Op.getValueType();
3263  unsigned VTBits = VT.getSizeInBits();
3264  DebugLoc dl = Op.getDebugLoc();
3265  SDValue ShOpLo = Op.getOperand(0);
3266  SDValue ShOpHi = Op.getOperand(1);
3267  SDValue ShAmt  = Op.getOperand(2);
3268  SDValue ARMcc;
3269  unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3270
3271  assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3272
3273  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3274                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3275  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3276  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3277                                   DAG.getConstant(VTBits, MVT::i32));
3278  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3279  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3280  SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3281
3282  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3283  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3284                          ARMcc, DAG, dl);
3285  SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3286  SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3287                           CCR, Cmp);
3288
3289  SDValue Ops[2] = { Lo, Hi };
3290  return DAG.getMergeValues(Ops, 2, dl);
3291}
3292
3293/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3294/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3295SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3296                                               SelectionDAG &DAG) const {
3297  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3298  EVT VT = Op.getValueType();
3299  unsigned VTBits = VT.getSizeInBits();
3300  DebugLoc dl = Op.getDebugLoc();
3301  SDValue ShOpLo = Op.getOperand(0);
3302  SDValue ShOpHi = Op.getOperand(1);
3303  SDValue ShAmt  = Op.getOperand(2);
3304  SDValue ARMcc;
3305
3306  assert(Op.getOpcode() == ISD::SHL_PARTS);
3307  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3308                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
3309  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3310  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3311                                   DAG.getConstant(VTBits, MVT::i32));
3312  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3313  SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3314
3315  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3316  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3317  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3318                          ARMcc, DAG, dl);
3319  SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3320  SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3321                           CCR, Cmp);
3322
3323  SDValue Ops[2] = { Lo, Hi };
3324  return DAG.getMergeValues(Ops, 2, dl);
3325}
3326
3327SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3328                                            SelectionDAG &DAG) const {
3329  // The rounding mode is in bits 23:22 of the FPSCR.
3330  // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3331  // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3332  // so that the shift + and get folded into a bitfield extract.
3333  DebugLoc dl = Op.getDebugLoc();
3334  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3335                              DAG.getConstant(Intrinsic::arm_get_fpscr,
3336                                              MVT::i32));
3337  SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3338                                  DAG.getConstant(1U << 22, MVT::i32));
3339  SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3340                              DAG.getConstant(22, MVT::i32));
3341  return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3342                     DAG.getConstant(3, MVT::i32));
3343}
3344
3345static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3346                         const ARMSubtarget *ST) {
3347  EVT VT = N->getValueType(0);
3348  DebugLoc dl = N->getDebugLoc();
3349
3350  if (!ST->hasV6T2Ops())
3351    return SDValue();
3352
3353  SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3354  return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3355}
3356
3357static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3358                          const ARMSubtarget *ST) {
3359  EVT VT = N->getValueType(0);
3360  DebugLoc dl = N->getDebugLoc();
3361
3362  if (!VT.isVector())
3363    return SDValue();
3364
3365  // Lower vector shifts on NEON to use VSHL.
3366  assert(ST->hasNEON() && "unexpected vector shift");
3367
3368  // Left shifts translate directly to the vshiftu intrinsic.
3369  if (N->getOpcode() == ISD::SHL)
3370    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3371                       DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3372                       N->getOperand(0), N->getOperand(1));
3373
3374  assert((N->getOpcode() == ISD::SRA ||
3375          N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3376
3377  // NEON uses the same intrinsics for both left and right shifts.  For
3378  // right shifts, the shift amounts are negative, so negate the vector of
3379  // shift amounts.
3380  EVT ShiftVT = N->getOperand(1).getValueType();
3381  SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3382                                     getZeroVector(ShiftVT, DAG, dl),
3383                                     N->getOperand(1));
3384  Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3385                             Intrinsic::arm_neon_vshifts :
3386                             Intrinsic::arm_neon_vshiftu);
3387  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3388                     DAG.getConstant(vshiftInt, MVT::i32),
3389                     N->getOperand(0), NegatedCount);
3390}
3391
3392static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3393                                const ARMSubtarget *ST) {
3394  EVT VT = N->getValueType(0);
3395  DebugLoc dl = N->getDebugLoc();
3396
3397  // We can get here for a node like i32 = ISD::SHL i32, i64
3398  if (VT != MVT::i64)
3399    return SDValue();
3400
3401  assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3402         "Unknown shift to lower!");
3403
3404  // We only lower SRA, SRL of 1 here, all others use generic lowering.
3405  if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3406      cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3407    return SDValue();
3408
3409  // If we are in thumb mode, we don't have RRX.
3410  if (ST->isThumb1Only()) return SDValue();
3411
3412  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
3413  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3414                           DAG.getConstant(0, MVT::i32));
3415  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3416                           DAG.getConstant(1, MVT::i32));
3417
3418  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3419  // captures the result into a carry flag.
3420  unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3421  Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3422
3423  // The low part is an ARMISD::RRX operand, which shifts the carry in.
3424  Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3425
3426  // Merge the pieces into a single i64 value.
3427 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3428}
3429
3430static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3431  SDValue TmpOp0, TmpOp1;
3432  bool Invert = false;
3433  bool Swap = false;
3434  unsigned Opc = 0;
3435
3436  SDValue Op0 = Op.getOperand(0);
3437  SDValue Op1 = Op.getOperand(1);
3438  SDValue CC = Op.getOperand(2);
3439  EVT VT = Op.getValueType();
3440  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3441  DebugLoc dl = Op.getDebugLoc();
3442
3443  if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3444    switch (SetCCOpcode) {
3445    default: llvm_unreachable("Illegal FP comparison"); break;
3446    case ISD::SETUNE:
3447    case ISD::SETNE:  Invert = true; // Fallthrough
3448    case ISD::SETOEQ:
3449    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3450    case ISD::SETOLT:
3451    case ISD::SETLT: Swap = true; // Fallthrough
3452    case ISD::SETOGT:
3453    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3454    case ISD::SETOLE:
3455    case ISD::SETLE:  Swap = true; // Fallthrough
3456    case ISD::SETOGE:
3457    case ISD::SETGE: Opc = ARMISD::VCGE; break;
3458    case ISD::SETUGE: Swap = true; // Fallthrough
3459    case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3460    case ISD::SETUGT: Swap = true; // Fallthrough
3461    case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3462    case ISD::SETUEQ: Invert = true; // Fallthrough
3463    case ISD::SETONE:
3464      // Expand this to (OLT | OGT).
3465      TmpOp0 = Op0;
3466      TmpOp1 = Op1;
3467      Opc = ISD::OR;
3468      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3469      Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3470      break;
3471    case ISD::SETUO: Invert = true; // Fallthrough
3472    case ISD::SETO:
3473      // Expand this to (OLT | OGE).
3474      TmpOp0 = Op0;
3475      TmpOp1 = Op1;
3476      Opc = ISD::OR;
3477      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3478      Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3479      break;
3480    }
3481  } else {
3482    // Integer comparisons.
3483    switch (SetCCOpcode) {
3484    default: llvm_unreachable("Illegal integer comparison"); break;
3485    case ISD::SETNE:  Invert = true;
3486    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
3487    case ISD::SETLT:  Swap = true;
3488    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
3489    case ISD::SETLE:  Swap = true;
3490    case ISD::SETGE:  Opc = ARMISD::VCGE; break;
3491    case ISD::SETULT: Swap = true;
3492    case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3493    case ISD::SETULE: Swap = true;
3494    case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3495    }
3496
3497    // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3498    if (Opc == ARMISD::VCEQ) {
3499
3500      SDValue AndOp;
3501      if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3502        AndOp = Op0;
3503      else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3504        AndOp = Op1;
3505
3506      // Ignore bitconvert.
3507      if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3508        AndOp = AndOp.getOperand(0);
3509
3510      if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3511        Opc = ARMISD::VTST;
3512        Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3513        Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3514        Invert = !Invert;
3515      }
3516    }
3517  }
3518
3519  if (Swap)
3520    std::swap(Op0, Op1);
3521
3522  // If one of the operands is a constant vector zero, attempt to fold the
3523  // comparison to a specialized compare-against-zero form.
3524  SDValue SingleOp;
3525  if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3526    SingleOp = Op0;
3527  else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3528    if (Opc == ARMISD::VCGE)
3529      Opc = ARMISD::VCLEZ;
3530    else if (Opc == ARMISD::VCGT)
3531      Opc = ARMISD::VCLTZ;
3532    SingleOp = Op1;
3533  }
3534
3535  SDValue Result;
3536  if (SingleOp.getNode()) {
3537    switch (Opc) {
3538    case ARMISD::VCEQ:
3539      Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3540    case ARMISD::VCGE:
3541      Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3542    case ARMISD::VCLEZ:
3543      Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3544    case ARMISD::VCGT:
3545      Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3546    case ARMISD::VCLTZ:
3547      Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3548    default:
3549      Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3550    }
3551  } else {
3552     Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3553  }
3554
3555  if (Invert)
3556    Result = DAG.getNOT(dl, Result, VT);
3557
3558  return Result;
3559}
3560
3561/// isNEONModifiedImm - Check if the specified splat value corresponds to a
3562/// valid vector constant for a NEON instruction with a "modified immediate"
3563/// operand (e.g., VMOV).  If so, return the encoded value.
3564static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3565                                 unsigned SplatBitSize, SelectionDAG &DAG,
3566                                 EVT &VT, bool is128Bits, NEONModImmType type) {
3567  unsigned OpCmode, Imm;
3568
3569  // SplatBitSize is set to the smallest size that splats the vector, so a
3570  // zero vector will always have SplatBitSize == 8.  However, NEON modified
3571  // immediate instructions others than VMOV do not support the 8-bit encoding
3572  // of a zero vector, and the default encoding of zero is supposed to be the
3573  // 32-bit version.
3574  if (SplatBits == 0)
3575    SplatBitSize = 32;
3576
3577  switch (SplatBitSize) {
3578  case 8:
3579    if (type != VMOVModImm)
3580      return SDValue();
3581    // Any 1-byte value is OK.  Op=0, Cmode=1110.
3582    assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3583    OpCmode = 0xe;
3584    Imm = SplatBits;
3585    VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3586    break;
3587
3588  case 16:
3589    // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3590    VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3591    if ((SplatBits & ~0xff) == 0) {
3592      // Value = 0x00nn: Op=x, Cmode=100x.
3593      OpCmode = 0x8;
3594      Imm = SplatBits;
3595      break;
3596    }
3597    if ((SplatBits & ~0xff00) == 0) {
3598      // Value = 0xnn00: Op=x, Cmode=101x.
3599      OpCmode = 0xa;
3600      Imm = SplatBits >> 8;
3601      break;
3602    }
3603    return SDValue();
3604
3605  case 32:
3606    // NEON's 32-bit VMOV supports splat values where:
3607    // * only one byte is nonzero, or
3608    // * the least significant byte is 0xff and the second byte is nonzero, or
3609    // * the least significant 2 bytes are 0xff and the third is nonzero.
3610    VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3611    if ((SplatBits & ~0xff) == 0) {
3612      // Value = 0x000000nn: Op=x, Cmode=000x.
3613      OpCmode = 0;
3614      Imm = SplatBits;
3615      break;
3616    }
3617    if ((SplatBits & ~0xff00) == 0) {
3618      // Value = 0x0000nn00: Op=x, Cmode=001x.
3619      OpCmode = 0x2;
3620      Imm = SplatBits >> 8;
3621      break;
3622    }
3623    if ((SplatBits & ~0xff0000) == 0) {
3624      // Value = 0x00nn0000: Op=x, Cmode=010x.
3625      OpCmode = 0x4;
3626      Imm = SplatBits >> 16;
3627      break;
3628    }
3629    if ((SplatBits & ~0xff000000) == 0) {
3630      // Value = 0xnn000000: Op=x, Cmode=011x.
3631      OpCmode = 0x6;
3632      Imm = SplatBits >> 24;
3633      break;
3634    }
3635
3636    // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3637    if (type == OtherModImm) return SDValue();
3638
3639    if ((SplatBits & ~0xffff) == 0 &&
3640        ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3641      // Value = 0x0000nnff: Op=x, Cmode=1100.
3642      OpCmode = 0xc;
3643      Imm = SplatBits >> 8;
3644      SplatBits |= 0xff;
3645      break;
3646    }
3647
3648    if ((SplatBits & ~0xffffff) == 0 &&
3649        ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3650      // Value = 0x00nnffff: Op=x, Cmode=1101.
3651      OpCmode = 0xd;
3652      Imm = SplatBits >> 16;
3653      SplatBits |= 0xffff;
3654      break;
3655    }
3656
3657    // Note: there are a few 32-bit splat values (specifically: 00ffff00,
3658    // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3659    // VMOV.I32.  A (very) minor optimization would be to replicate the value
3660    // and fall through here to test for a valid 64-bit splat.  But, then the
3661    // caller would also need to check and handle the change in size.
3662    return SDValue();
3663
3664  case 64: {
3665    if (type != VMOVModImm)
3666      return SDValue();
3667    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3668    uint64_t BitMask = 0xff;
3669    uint64_t Val = 0;
3670    unsigned ImmMask = 1;
3671    Imm = 0;
3672    for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3673      if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3674        Val |= BitMask;
3675        Imm |= ImmMask;
3676      } else if ((SplatBits & BitMask) != 0) {
3677        return SDValue();
3678      }
3679      BitMask <<= 8;
3680      ImmMask <<= 1;
3681    }
3682    // Op=1, Cmode=1110.
3683    OpCmode = 0x1e;
3684    SplatBits = Val;
3685    VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3686    break;
3687  }
3688
3689  default:
3690    llvm_unreachable("unexpected size for isNEONModifiedImm");
3691    return SDValue();
3692  }
3693
3694  unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3695  return DAG.getTargetConstant(EncodedVal, MVT::i32);
3696}
3697
3698static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
3699                       bool &ReverseVEXT, unsigned &Imm) {
3700  unsigned NumElts = VT.getVectorNumElements();
3701  ReverseVEXT = false;
3702
3703  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
3704  if (M[0] < 0)
3705    return false;
3706
3707  Imm = M[0];
3708
3709  // If this is a VEXT shuffle, the immediate value is the index of the first
3710  // element.  The other shuffle indices must be the successive elements after
3711  // the first one.
3712  unsigned ExpectedElt = Imm;
3713  for (unsigned i = 1; i < NumElts; ++i) {
3714    // Increment the expected index.  If it wraps around, it may still be
3715    // a VEXT but the source vectors must be swapped.
3716    ExpectedElt += 1;
3717    if (ExpectedElt == NumElts * 2) {
3718      ExpectedElt = 0;
3719      ReverseVEXT = true;
3720    }
3721
3722    if (M[i] < 0) continue; // ignore UNDEF indices
3723    if (ExpectedElt != static_cast<unsigned>(M[i]))
3724      return false;
3725  }
3726
3727  // Adjust the index value if the source operands will be swapped.
3728  if (ReverseVEXT)
3729    Imm -= NumElts;
3730
3731  return true;
3732}
3733
3734/// isVREVMask - Check if a vector shuffle corresponds to a VREV
3735/// instruction with the specified blocksize.  (The order of the elements
3736/// within each block of the vector is reversed.)
3737static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
3738                       unsigned BlockSize) {
3739  assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3740         "Only possible block sizes for VREV are: 16, 32, 64");
3741
3742  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3743  if (EltSz == 64)
3744    return false;
3745
3746  unsigned NumElts = VT.getVectorNumElements();
3747  unsigned BlockElts = M[0] + 1;
3748  // If the first shuffle index is UNDEF, be optimistic.
3749  if (M[0] < 0)
3750    BlockElts = BlockSize / EltSz;
3751
3752  if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3753    return false;
3754
3755  for (unsigned i = 0; i < NumElts; ++i) {
3756    if (M[i] < 0) continue; // ignore UNDEF indices
3757    if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3758      return false;
3759  }
3760
3761  return true;
3762}
3763
3764static bool isVTBLMask(const SmallVectorImpl<int> &M, EVT VT) {
3765  // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3766  // range, then 0 is placed into the resulting vector. So pretty much any mask
3767  // of 8 elements can work here.
3768  return VT == MVT::v8i8 && M.size() == 8;
3769}
3770
3771static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
3772                       unsigned &WhichResult) {
3773  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3774  if (EltSz == 64)
3775    return false;
3776
3777  unsigned NumElts = VT.getVectorNumElements();
3778  WhichResult = (M[0] == 0 ? 0 : 1);
3779  for (unsigned i = 0; i < NumElts; i += 2) {
3780    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3781        (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3782      return false;
3783  }
3784  return true;
3785}
3786
3787/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3788/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3789/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3790static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3791                                unsigned &WhichResult) {
3792  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3793  if (EltSz == 64)
3794    return false;
3795
3796  unsigned NumElts = VT.getVectorNumElements();
3797  WhichResult = (M[0] == 0 ? 0 : 1);
3798  for (unsigned i = 0; i < NumElts; i += 2) {
3799    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3800        (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3801      return false;
3802  }
3803  return true;
3804}
3805
3806static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
3807                       unsigned &WhichResult) {
3808  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3809  if (EltSz == 64)
3810    return false;
3811
3812  unsigned NumElts = VT.getVectorNumElements();
3813  WhichResult = (M[0] == 0 ? 0 : 1);
3814  for (unsigned i = 0; i != NumElts; ++i) {
3815    if (M[i] < 0) continue; // ignore UNDEF indices
3816    if ((unsigned) M[i] != 2 * i + WhichResult)
3817      return false;
3818  }
3819
3820  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3821  if (VT.is64BitVector() && EltSz == 32)
3822    return false;
3823
3824  return true;
3825}
3826
3827/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3828/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3829/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3830static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3831                                unsigned &WhichResult) {
3832  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3833  if (EltSz == 64)
3834    return false;
3835
3836  unsigned Half = VT.getVectorNumElements() / 2;
3837  WhichResult = (M[0] == 0 ? 0 : 1);
3838  for (unsigned j = 0; j != 2; ++j) {
3839    unsigned Idx = WhichResult;
3840    for (unsigned i = 0; i != Half; ++i) {
3841      int MIdx = M[i + j * Half];
3842      if (MIdx >= 0 && (unsigned) MIdx != Idx)
3843        return false;
3844      Idx += 2;
3845    }
3846  }
3847
3848  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3849  if (VT.is64BitVector() && EltSz == 32)
3850    return false;
3851
3852  return true;
3853}
3854
3855static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
3856                       unsigned &WhichResult) {
3857  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3858  if (EltSz == 64)
3859    return false;
3860
3861  unsigned NumElts = VT.getVectorNumElements();
3862  WhichResult = (M[0] == 0 ? 0 : 1);
3863  unsigned Idx = WhichResult * NumElts / 2;
3864  for (unsigned i = 0; i != NumElts; i += 2) {
3865    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3866        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3867      return false;
3868    Idx += 1;
3869  }
3870
3871  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3872  if (VT.is64BitVector() && EltSz == 32)
3873    return false;
3874
3875  return true;
3876}
3877
3878/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
3879/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3880/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
3881static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
3882                                unsigned &WhichResult) {
3883  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3884  if (EltSz == 64)
3885    return false;
3886
3887  unsigned NumElts = VT.getVectorNumElements();
3888  WhichResult = (M[0] == 0 ? 0 : 1);
3889  unsigned Idx = WhichResult * NumElts / 2;
3890  for (unsigned i = 0; i != NumElts; i += 2) {
3891    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3892        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
3893      return false;
3894    Idx += 1;
3895  }
3896
3897  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3898  if (VT.is64BitVector() && EltSz == 32)
3899    return false;
3900
3901  return true;
3902}
3903
3904// If N is an integer constant that can be moved into a register in one
3905// instruction, return an SDValue of such a constant (will become a MOV
3906// instruction).  Otherwise return null.
3907static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
3908                                     const ARMSubtarget *ST, DebugLoc dl) {
3909  uint64_t Val;
3910  if (!isa<ConstantSDNode>(N))
3911    return SDValue();
3912  Val = cast<ConstantSDNode>(N)->getZExtValue();
3913
3914  if (ST->isThumb1Only()) {
3915    if (Val <= 255 || ~Val <= 255)
3916      return DAG.getConstant(Val, MVT::i32);
3917  } else {
3918    if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
3919      return DAG.getConstant(Val, MVT::i32);
3920  }
3921  return SDValue();
3922}
3923
3924// If this is a case we can't handle, return null and let the default
3925// expansion code take care of it.
3926SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
3927                                             const ARMSubtarget *ST) const {
3928  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
3929  DebugLoc dl = Op.getDebugLoc();
3930  EVT VT = Op.getValueType();
3931
3932  APInt SplatBits, SplatUndef;
3933  unsigned SplatBitSize;
3934  bool HasAnyUndefs;
3935  if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
3936    if (SplatBitSize <= 64) {
3937      // Check if an immediate VMOV works.
3938      EVT VmovVT;
3939      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
3940                                      SplatUndef.getZExtValue(), SplatBitSize,
3941                                      DAG, VmovVT, VT.is128BitVector(),
3942                                      VMOVModImm);
3943      if (Val.getNode()) {
3944        SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
3945        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3946      }
3947
3948      // Try an immediate VMVN.
3949      uint64_t NegatedImm = (~SplatBits).getZExtValue();
3950      Val = isNEONModifiedImm(NegatedImm,
3951                                      SplatUndef.getZExtValue(), SplatBitSize,
3952                                      DAG, VmovVT, VT.is128BitVector(),
3953                                      VMVNModImm);
3954      if (Val.getNode()) {
3955        SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
3956        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3957      }
3958    }
3959  }
3960
3961  // Scan through the operands to see if only one value is used.
3962  unsigned NumElts = VT.getVectorNumElements();
3963  bool isOnlyLowElement = true;
3964  bool usesOnlyOneValue = true;
3965  bool isConstant = true;
3966  SDValue Value;
3967  for (unsigned i = 0; i < NumElts; ++i) {
3968    SDValue V = Op.getOperand(i);
3969    if (V.getOpcode() == ISD::UNDEF)
3970      continue;
3971    if (i > 0)
3972      isOnlyLowElement = false;
3973    if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
3974      isConstant = false;
3975
3976    if (!Value.getNode())
3977      Value = V;
3978    else if (V != Value)
3979      usesOnlyOneValue = false;
3980  }
3981
3982  if (!Value.getNode())
3983    return DAG.getUNDEF(VT);
3984
3985  if (isOnlyLowElement)
3986    return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
3987
3988  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
3989
3990  // Use VDUP for non-constant splats.  For f32 constant splats, reduce to
3991  // i32 and try again.
3992  if (usesOnlyOneValue && EltSize <= 32) {
3993    if (!isConstant)
3994      return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
3995    if (VT.getVectorElementType().isFloatingPoint()) {
3996      SmallVector<SDValue, 8> Ops;
3997      for (unsigned i = 0; i < NumElts; ++i)
3998        Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
3999                                  Op.getOperand(i)));
4000      EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4001      SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4002      Val = LowerBUILD_VECTOR(Val, DAG, ST);
4003      if (Val.getNode())
4004        return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4005    }
4006    SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4007    if (Val.getNode())
4008      return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4009  }
4010
4011  // If all elements are constants and the case above didn't get hit, fall back
4012  // to the default expansion, which will generate a load from the constant
4013  // pool.
4014  if (isConstant)
4015    return SDValue();
4016
4017  // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4018  if (NumElts >= 4) {
4019    SDValue shuffle = ReconstructShuffle(Op, DAG);
4020    if (shuffle != SDValue())
4021      return shuffle;
4022  }
4023
4024  // Vectors with 32- or 64-bit elements can be built by directly assigning
4025  // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
4026  // will be legalized.
4027  if (EltSize >= 32) {
4028    // Do the expansion with floating-point types, since that is what the VFP
4029    // registers are defined to use, and since i64 is not legal.
4030    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4031    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4032    SmallVector<SDValue, 8> Ops;
4033    for (unsigned i = 0; i < NumElts; ++i)
4034      Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4035    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4036    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4037  }
4038
4039  return SDValue();
4040}
4041
4042// Gather data to see if the operation can be modelled as a
4043// shuffle in combination with VEXTs.
4044SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4045                                              SelectionDAG &DAG) const {
4046  DebugLoc dl = Op.getDebugLoc();
4047  EVT VT = Op.getValueType();
4048  unsigned NumElts = VT.getVectorNumElements();
4049
4050  SmallVector<SDValue, 2> SourceVecs;
4051  SmallVector<unsigned, 2> MinElts;
4052  SmallVector<unsigned, 2> MaxElts;
4053
4054  for (unsigned i = 0; i < NumElts; ++i) {
4055    SDValue V = Op.getOperand(i);
4056    if (V.getOpcode() == ISD::UNDEF)
4057      continue;
4058    else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4059      // A shuffle can only come from building a vector from various
4060      // elements of other vectors.
4061      return SDValue();
4062    } else if (V.getOperand(0).getValueType().getVectorElementType() !=
4063               VT.getVectorElementType()) {
4064      // This code doesn't know how to handle shuffles where the vector
4065      // element types do not match (this happens because type legalization
4066      // promotes the return type of EXTRACT_VECTOR_ELT).
4067      // FIXME: It might be appropriate to extend this code to handle
4068      // mismatched types.
4069      return SDValue();
4070    }
4071
4072    // Record this extraction against the appropriate vector if possible...
4073    SDValue SourceVec = V.getOperand(0);
4074    unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4075    bool FoundSource = false;
4076    for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4077      if (SourceVecs[j] == SourceVec) {
4078        if (MinElts[j] > EltNo)
4079          MinElts[j] = EltNo;
4080        if (MaxElts[j] < EltNo)
4081          MaxElts[j] = EltNo;
4082        FoundSource = true;
4083        break;
4084      }
4085    }
4086
4087    // Or record a new source if not...
4088    if (!FoundSource) {
4089      SourceVecs.push_back(SourceVec);
4090      MinElts.push_back(EltNo);
4091      MaxElts.push_back(EltNo);
4092    }
4093  }
4094
4095  // Currently only do something sane when at most two source vectors
4096  // involved.
4097  if (SourceVecs.size() > 2)
4098    return SDValue();
4099
4100  SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4101  int VEXTOffsets[2] = {0, 0};
4102
4103  // This loop extracts the usage patterns of the source vectors
4104  // and prepares appropriate SDValues for a shuffle if possible.
4105  for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4106    if (SourceVecs[i].getValueType() == VT) {
4107      // No VEXT necessary
4108      ShuffleSrcs[i] = SourceVecs[i];
4109      VEXTOffsets[i] = 0;
4110      continue;
4111    } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4112      // It probably isn't worth padding out a smaller vector just to
4113      // break it down again in a shuffle.
4114      return SDValue();
4115    }
4116
4117    // Since only 64-bit and 128-bit vectors are legal on ARM and
4118    // we've eliminated the other cases...
4119    assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4120           "unexpected vector sizes in ReconstructShuffle");
4121
4122    if (MaxElts[i] - MinElts[i] >= NumElts) {
4123      // Span too large for a VEXT to cope
4124      return SDValue();
4125    }
4126
4127    if (MinElts[i] >= NumElts) {
4128      // The extraction can just take the second half
4129      VEXTOffsets[i] = NumElts;
4130      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4131                                   SourceVecs[i],
4132                                   DAG.getIntPtrConstant(NumElts));
4133    } else if (MaxElts[i] < NumElts) {
4134      // The extraction can just take the first half
4135      VEXTOffsets[i] = 0;
4136      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4137                                   SourceVecs[i],
4138                                   DAG.getIntPtrConstant(0));
4139    } else {
4140      // An actual VEXT is needed
4141      VEXTOffsets[i] = MinElts[i];
4142      SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4143                                     SourceVecs[i],
4144                                     DAG.getIntPtrConstant(0));
4145      SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4146                                     SourceVecs[i],
4147                                     DAG.getIntPtrConstant(NumElts));
4148      ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4149                                   DAG.getConstant(VEXTOffsets[i], MVT::i32));
4150    }
4151  }
4152
4153  SmallVector<int, 8> Mask;
4154
4155  for (unsigned i = 0; i < NumElts; ++i) {
4156    SDValue Entry = Op.getOperand(i);
4157    if (Entry.getOpcode() == ISD::UNDEF) {
4158      Mask.push_back(-1);
4159      continue;
4160    }
4161
4162    SDValue ExtractVec = Entry.getOperand(0);
4163    int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4164                                          .getOperand(1))->getSExtValue();
4165    if (ExtractVec == SourceVecs[0]) {
4166      Mask.push_back(ExtractElt - VEXTOffsets[0]);
4167    } else {
4168      Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4169    }
4170  }
4171
4172  // Final check before we try to produce nonsense...
4173  if (isShuffleMaskLegal(Mask, VT))
4174    return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4175                                &Mask[0]);
4176
4177  return SDValue();
4178}
4179
4180/// isShuffleMaskLegal - Targets can use this to indicate that they only
4181/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4182/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4183/// are assumed to be legal.
4184bool
4185ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4186                                      EVT VT) const {
4187  if (VT.getVectorNumElements() == 4 &&
4188      (VT.is128BitVector() || VT.is64BitVector())) {
4189    unsigned PFIndexes[4];
4190    for (unsigned i = 0; i != 4; ++i) {
4191      if (M[i] < 0)
4192        PFIndexes[i] = 8;
4193      else
4194        PFIndexes[i] = M[i];
4195    }
4196
4197    // Compute the index in the perfect shuffle table.
4198    unsigned PFTableIndex =
4199      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4200    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4201    unsigned Cost = (PFEntry >> 30);
4202
4203    if (Cost <= 4)
4204      return true;
4205  }
4206
4207  bool ReverseVEXT;
4208  unsigned Imm, WhichResult;
4209
4210  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4211  return (EltSize >= 32 ||
4212          ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4213          isVREVMask(M, VT, 64) ||
4214          isVREVMask(M, VT, 32) ||
4215          isVREVMask(M, VT, 16) ||
4216          isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4217          isVTBLMask(M, VT) ||
4218          isVTRNMask(M, VT, WhichResult) ||
4219          isVUZPMask(M, VT, WhichResult) ||
4220          isVZIPMask(M, VT, WhichResult) ||
4221          isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4222          isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4223          isVZIP_v_undef_Mask(M, VT, WhichResult));
4224}
4225
4226/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4227/// the specified operations to build the shuffle.
4228static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4229                                      SDValue RHS, SelectionDAG &DAG,
4230                                      DebugLoc dl) {
4231  unsigned OpNum = (PFEntry >> 26) & 0x0F;
4232  unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4233  unsigned RHSID = (PFEntry >>  0) & ((1 << 13)-1);
4234
4235  enum {
4236    OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4237    OP_VREV,
4238    OP_VDUP0,
4239    OP_VDUP1,
4240    OP_VDUP2,
4241    OP_VDUP3,
4242    OP_VEXT1,
4243    OP_VEXT2,
4244    OP_VEXT3,
4245    OP_VUZPL, // VUZP, left result
4246    OP_VUZPR, // VUZP, right result
4247    OP_VZIPL, // VZIP, left result
4248    OP_VZIPR, // VZIP, right result
4249    OP_VTRNL, // VTRN, left result
4250    OP_VTRNR  // VTRN, right result
4251  };
4252
4253  if (OpNum == OP_COPY) {
4254    if (LHSID == (1*9+2)*9+3) return LHS;
4255    assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4256    return RHS;
4257  }
4258
4259  SDValue OpLHS, OpRHS;
4260  OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4261  OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4262  EVT VT = OpLHS.getValueType();
4263
4264  switch (OpNum) {
4265  default: llvm_unreachable("Unknown shuffle opcode!");
4266  case OP_VREV:
4267    // VREV divides the vector in half and swaps within the half.
4268    if (VT.getVectorElementType() == MVT::i32 ||
4269        VT.getVectorElementType() == MVT::f32)
4270      return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4271    // vrev <4 x i16> -> VREV32
4272    if (VT.getVectorElementType() == MVT::i16)
4273      return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4274    // vrev <4 x i8> -> VREV16
4275    assert(VT.getVectorElementType() == MVT::i8);
4276    return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4277  case OP_VDUP0:
4278  case OP_VDUP1:
4279  case OP_VDUP2:
4280  case OP_VDUP3:
4281    return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4282                       OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4283  case OP_VEXT1:
4284  case OP_VEXT2:
4285  case OP_VEXT3:
4286    return DAG.getNode(ARMISD::VEXT, dl, VT,
4287                       OpLHS, OpRHS,
4288                       DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4289  case OP_VUZPL:
4290  case OP_VUZPR:
4291    return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4292                       OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4293  case OP_VZIPL:
4294  case OP_VZIPR:
4295    return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4296                       OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4297  case OP_VTRNL:
4298  case OP_VTRNR:
4299    return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4300                       OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4301  }
4302}
4303
4304static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4305                                       SmallVectorImpl<int> &ShuffleMask,
4306                                       SelectionDAG &DAG) {
4307  // Check to see if we can use the VTBL instruction.
4308  SDValue V1 = Op.getOperand(0);
4309  SDValue V2 = Op.getOperand(1);
4310  DebugLoc DL = Op.getDebugLoc();
4311
4312  SmallVector<SDValue, 8> VTBLMask;
4313  for (SmallVectorImpl<int>::iterator
4314         I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4315    VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4316
4317  if (V2.getNode()->getOpcode() == ISD::UNDEF)
4318    return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4319                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4320                                   &VTBLMask[0], 8));
4321
4322  return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4323                     DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4324                                 &VTBLMask[0], 8));
4325}
4326
4327static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4328  SDValue V1 = Op.getOperand(0);
4329  SDValue V2 = Op.getOperand(1);
4330  DebugLoc dl = Op.getDebugLoc();
4331  EVT VT = Op.getValueType();
4332  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4333  SmallVector<int, 8> ShuffleMask;
4334
4335  // Convert shuffles that are directly supported on NEON to target-specific
4336  // DAG nodes, instead of keeping them as shuffles and matching them again
4337  // during code selection.  This is more efficient and avoids the possibility
4338  // of inconsistencies between legalization and selection.
4339  // FIXME: floating-point vectors should be canonicalized to integer vectors
4340  // of the same time so that they get CSEd properly.
4341  SVN->getMask(ShuffleMask);
4342
4343  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4344  if (EltSize <= 32) {
4345    if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4346      int Lane = SVN->getSplatIndex();
4347      // If this is undef splat, generate it via "just" vdup, if possible.
4348      if (Lane == -1) Lane = 0;
4349
4350      if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4351        return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4352      }
4353      return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4354                         DAG.getConstant(Lane, MVT::i32));
4355    }
4356
4357    bool ReverseVEXT;
4358    unsigned Imm;
4359    if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4360      if (ReverseVEXT)
4361        std::swap(V1, V2);
4362      return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4363                         DAG.getConstant(Imm, MVT::i32));
4364    }
4365
4366    if (isVREVMask(ShuffleMask, VT, 64))
4367      return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4368    if (isVREVMask(ShuffleMask, VT, 32))
4369      return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4370    if (isVREVMask(ShuffleMask, VT, 16))
4371      return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4372
4373    // Check for Neon shuffles that modify both input vectors in place.
4374    // If both results are used, i.e., if there are two shuffles with the same
4375    // source operands and with masks corresponding to both results of one of
4376    // these operations, DAG memoization will ensure that a single node is
4377    // used for both shuffles.
4378    unsigned WhichResult;
4379    if (isVTRNMask(ShuffleMask, VT, WhichResult))
4380      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4381                         V1, V2).getValue(WhichResult);
4382    if (isVUZPMask(ShuffleMask, VT, WhichResult))
4383      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4384                         V1, V2).getValue(WhichResult);
4385    if (isVZIPMask(ShuffleMask, VT, WhichResult))
4386      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4387                         V1, V2).getValue(WhichResult);
4388
4389    if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4390      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4391                         V1, V1).getValue(WhichResult);
4392    if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4393      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4394                         V1, V1).getValue(WhichResult);
4395    if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4396      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4397                         V1, V1).getValue(WhichResult);
4398  }
4399
4400  // If the shuffle is not directly supported and it has 4 elements, use
4401  // the PerfectShuffle-generated table to synthesize it from other shuffles.
4402  unsigned NumElts = VT.getVectorNumElements();
4403  if (NumElts == 4) {
4404    unsigned PFIndexes[4];
4405    for (unsigned i = 0; i != 4; ++i) {
4406      if (ShuffleMask[i] < 0)
4407        PFIndexes[i] = 8;
4408      else
4409        PFIndexes[i] = ShuffleMask[i];
4410    }
4411
4412    // Compute the index in the perfect shuffle table.
4413    unsigned PFTableIndex =
4414      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4415    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4416    unsigned Cost = (PFEntry >> 30);
4417
4418    if (Cost <= 4)
4419      return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4420  }
4421
4422  // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4423  if (EltSize >= 32) {
4424    // Do the expansion with floating-point types, since that is what the VFP
4425    // registers are defined to use, and since i64 is not legal.
4426    EVT EltVT = EVT::getFloatingPointVT(EltSize);
4427    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4428    V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4429    V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4430    SmallVector<SDValue, 8> Ops;
4431    for (unsigned i = 0; i < NumElts; ++i) {
4432      if (ShuffleMask[i] < 0)
4433        Ops.push_back(DAG.getUNDEF(EltVT));
4434      else
4435        Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4436                                  ShuffleMask[i] < (int)NumElts ? V1 : V2,
4437                                  DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4438                                                  MVT::i32)));
4439    }
4440    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4441    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4442  }
4443
4444  if (VT == MVT::v8i8) {
4445    SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4446    if (NewOp.getNode())
4447      return NewOp;
4448  }
4449
4450  return SDValue();
4451}
4452
4453static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4454  // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4455  SDValue Lane = Op.getOperand(1);
4456  if (!isa<ConstantSDNode>(Lane))
4457    return SDValue();
4458
4459  SDValue Vec = Op.getOperand(0);
4460  if (Op.getValueType() == MVT::i32 &&
4461      Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4462    DebugLoc dl = Op.getDebugLoc();
4463    return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4464  }
4465
4466  return Op;
4467}
4468
4469static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4470  // The only time a CONCAT_VECTORS operation can have legal types is when
4471  // two 64-bit vectors are concatenated to a 128-bit vector.
4472  assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4473         "unexpected CONCAT_VECTORS");
4474  DebugLoc dl = Op.getDebugLoc();
4475  SDValue Val = DAG.getUNDEF(MVT::v2f64);
4476  SDValue Op0 = Op.getOperand(0);
4477  SDValue Op1 = Op.getOperand(1);
4478  if (Op0.getOpcode() != ISD::UNDEF)
4479    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4480                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4481                      DAG.getIntPtrConstant(0));
4482  if (Op1.getOpcode() != ISD::UNDEF)
4483    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4484                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4485                      DAG.getIntPtrConstant(1));
4486  return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4487}
4488
4489/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4490/// element has been zero/sign-extended, depending on the isSigned parameter,
4491/// from an integer type half its size.
4492static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4493                                   bool isSigned) {
4494  // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4495  EVT VT = N->getValueType(0);
4496  if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4497    SDNode *BVN = N->getOperand(0).getNode();
4498    if (BVN->getValueType(0) != MVT::v4i32 ||
4499        BVN->getOpcode() != ISD::BUILD_VECTOR)
4500      return false;
4501    unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4502    unsigned HiElt = 1 - LoElt;
4503    ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4504    ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4505    ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4506    ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4507    if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4508      return false;
4509    if (isSigned) {
4510      if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4511          Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4512        return true;
4513    } else {
4514      if (Hi0->isNullValue() && Hi1->isNullValue())
4515        return true;
4516    }
4517    return false;
4518  }
4519
4520  if (N->getOpcode() != ISD::BUILD_VECTOR)
4521    return false;
4522
4523  for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4524    SDNode *Elt = N->getOperand(i).getNode();
4525    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4526      unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4527      unsigned HalfSize = EltSize / 2;
4528      if (isSigned) {
4529        if (!isIntN(HalfSize, C->getSExtValue()))
4530          return false;
4531      } else {
4532        if (!isUIntN(HalfSize, C->getZExtValue()))
4533          return false;
4534      }
4535      continue;
4536    }
4537    return false;
4538  }
4539
4540  return true;
4541}
4542
4543/// isSignExtended - Check if a node is a vector value that is sign-extended
4544/// or a constant BUILD_VECTOR with sign-extended elements.
4545static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4546  if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4547    return true;
4548  if (isExtendedBUILD_VECTOR(N, DAG, true))
4549    return true;
4550  return false;
4551}
4552
4553/// isZeroExtended - Check if a node is a vector value that is zero-extended
4554/// or a constant BUILD_VECTOR with zero-extended elements.
4555static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4556  if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4557    return true;
4558  if (isExtendedBUILD_VECTOR(N, DAG, false))
4559    return true;
4560  return false;
4561}
4562
4563/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4564/// load, or BUILD_VECTOR with extended elements, return the unextended value.
4565static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4566  if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4567    return N->getOperand(0);
4568  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4569    return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4570                       LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4571                       LD->isNonTemporal(), LD->getAlignment());
4572  // Otherwise, the value must be a BUILD_VECTOR.  For v2i64, it will
4573  // have been legalized as a BITCAST from v4i32.
4574  if (N->getOpcode() == ISD::BITCAST) {
4575    SDNode *BVN = N->getOperand(0).getNode();
4576    assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4577           BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4578    unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4579    return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4580                       BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4581  }
4582  // Construct a new BUILD_VECTOR with elements truncated to half the size.
4583  assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4584  EVT VT = N->getValueType(0);
4585  unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4586  unsigned NumElts = VT.getVectorNumElements();
4587  MVT TruncVT = MVT::getIntegerVT(EltSize);
4588  SmallVector<SDValue, 8> Ops;
4589  for (unsigned i = 0; i != NumElts; ++i) {
4590    ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4591    const APInt &CInt = C->getAPIntValue();
4592    Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4593  }
4594  return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4595                     MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4596}
4597
4598static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4599  unsigned Opcode = N->getOpcode();
4600  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4601    SDNode *N0 = N->getOperand(0).getNode();
4602    SDNode *N1 = N->getOperand(1).getNode();
4603    return N0->hasOneUse() && N1->hasOneUse() &&
4604      isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4605  }
4606  return false;
4607}
4608
4609static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4610  unsigned Opcode = N->getOpcode();
4611  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4612    SDNode *N0 = N->getOperand(0).getNode();
4613    SDNode *N1 = N->getOperand(1).getNode();
4614    return N0->hasOneUse() && N1->hasOneUse() &&
4615      isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4616  }
4617  return false;
4618}
4619
4620static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4621  // Multiplications are only custom-lowered for 128-bit vectors so that
4622  // VMULL can be detected.  Otherwise v2i64 multiplications are not legal.
4623  EVT VT = Op.getValueType();
4624  assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4625  SDNode *N0 = Op.getOperand(0).getNode();
4626  SDNode *N1 = Op.getOperand(1).getNode();
4627  unsigned NewOpc = 0;
4628  bool isMLA = false;
4629  bool isN0SExt = isSignExtended(N0, DAG);
4630  bool isN1SExt = isSignExtended(N1, DAG);
4631  if (isN0SExt && isN1SExt)
4632    NewOpc = ARMISD::VMULLs;
4633  else {
4634    bool isN0ZExt = isZeroExtended(N0, DAG);
4635    bool isN1ZExt = isZeroExtended(N1, DAG);
4636    if (isN0ZExt && isN1ZExt)
4637      NewOpc = ARMISD::VMULLu;
4638    else if (isN1SExt || isN1ZExt) {
4639      // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4640      // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4641      if (isN1SExt && isAddSubSExt(N0, DAG)) {
4642        NewOpc = ARMISD::VMULLs;
4643        isMLA = true;
4644      } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4645        NewOpc = ARMISD::VMULLu;
4646        isMLA = true;
4647      } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4648        std::swap(N0, N1);
4649        NewOpc = ARMISD::VMULLu;
4650        isMLA = true;
4651      }
4652    }
4653
4654    if (!NewOpc) {
4655      if (VT == MVT::v2i64)
4656        // Fall through to expand this.  It is not legal.
4657        return SDValue();
4658      else
4659        // Other vector multiplications are legal.
4660        return Op;
4661    }
4662  }
4663
4664  // Legalize to a VMULL instruction.
4665  DebugLoc DL = Op.getDebugLoc();
4666  SDValue Op0;
4667  SDValue Op1 = SkipExtension(N1, DAG);
4668  if (!isMLA) {
4669    Op0 = SkipExtension(N0, DAG);
4670    assert(Op0.getValueType().is64BitVector() &&
4671           Op1.getValueType().is64BitVector() &&
4672           "unexpected types for extended operands to VMULL");
4673    return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4674  }
4675
4676  // Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4677  // isel lowering to take advantage of no-stall back to back vmul + vmla.
4678  //   vmull q0, d4, d6
4679  //   vmlal q0, d5, d6
4680  // is faster than
4681  //   vaddl q0, d4, d5
4682  //   vmovl q1, d6
4683  //   vmul  q0, q0, q1
4684  SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4685  SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4686  EVT Op1VT = Op1.getValueType();
4687  return DAG.getNode(N0->getOpcode(), DL, VT,
4688                     DAG.getNode(NewOpc, DL, VT,
4689                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4690                     DAG.getNode(NewOpc, DL, VT,
4691                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4692}
4693
4694static SDValue
4695LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4696  // Convert to float
4697  // float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4698  // float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4699  X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4700  Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4701  X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4702  Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4703  // Get reciprocal estimate.
4704  // float4 recip = vrecpeq_f32(yf);
4705  Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4706                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4707  // Because char has a smaller range than uchar, we can actually get away
4708  // without any newton steps.  This requires that we use a weird bias
4709  // of 0xb000, however (again, this has been exhaustively tested).
4710  // float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4711  X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4712  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4713  Y = DAG.getConstant(0xb000, MVT::i32);
4714  Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4715  X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4716  X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4717  // Convert back to short.
4718  X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4719  X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4720  return X;
4721}
4722
4723static SDValue
4724LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4725  SDValue N2;
4726  // Convert to float.
4727  // float4 yf = vcvt_f32_s32(vmovl_s16(y));
4728  // float4 xf = vcvt_f32_s32(vmovl_s16(x));
4729  N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4730  N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4731  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4732  N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4733
4734  // Use reciprocal estimate and one refinement step.
4735  // float4 recip = vrecpeq_f32(yf);
4736  // recip *= vrecpsq_f32(yf, recip);
4737  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4738                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4739  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4740                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4741                   N1, N2);
4742  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4743  // Because short has a smaller range than ushort, we can actually get away
4744  // with only a single newton step.  This requires that we use a weird bias
4745  // of 89, however (again, this has been exhaustively tested).
4746  // float4 result = as_float4(as_int4(xf*recip) + 0x89);
4747  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4748  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4749  N1 = DAG.getConstant(0x89, MVT::i32);
4750  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4751  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4752  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4753  // Convert back to integer and return.
4754  // return vmovn_s32(vcvt_s32_f32(result));
4755  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4756  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4757  return N0;
4758}
4759
4760static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4761  EVT VT = Op.getValueType();
4762  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4763         "unexpected type for custom-lowering ISD::SDIV");
4764
4765  DebugLoc dl = Op.getDebugLoc();
4766  SDValue N0 = Op.getOperand(0);
4767  SDValue N1 = Op.getOperand(1);
4768  SDValue N2, N3;
4769
4770  if (VT == MVT::v8i8) {
4771    N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4772    N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4773
4774    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4775                     DAG.getIntPtrConstant(4));
4776    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4777                     DAG.getIntPtrConstant(4));
4778    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4779                     DAG.getIntPtrConstant(0));
4780    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4781                     DAG.getIntPtrConstant(0));
4782
4783    N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4784    N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4785
4786    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4787    N0 = LowerCONCAT_VECTORS(N0, DAG);
4788
4789    N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4790    return N0;
4791  }
4792  return LowerSDIV_v4i16(N0, N1, dl, DAG);
4793}
4794
4795static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4796  EVT VT = Op.getValueType();
4797  assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4798         "unexpected type for custom-lowering ISD::UDIV");
4799
4800  DebugLoc dl = Op.getDebugLoc();
4801  SDValue N0 = Op.getOperand(0);
4802  SDValue N1 = Op.getOperand(1);
4803  SDValue N2, N3;
4804
4805  if (VT == MVT::v8i8) {
4806    N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4807    N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4808
4809    N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4810                     DAG.getIntPtrConstant(4));
4811    N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4812                     DAG.getIntPtrConstant(4));
4813    N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4814                     DAG.getIntPtrConstant(0));
4815    N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4816                     DAG.getIntPtrConstant(0));
4817
4818    N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4819    N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4820
4821    N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4822    N0 = LowerCONCAT_VECTORS(N0, DAG);
4823
4824    N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4825                     DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4826                     N0);
4827    return N0;
4828  }
4829
4830  // v4i16 sdiv ... Convert to float.
4831  // float4 yf = vcvt_f32_s32(vmovl_u16(y));
4832  // float4 xf = vcvt_f32_s32(vmovl_u16(x));
4833  N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4834  N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4835  N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4836  SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4837
4838  // Use reciprocal estimate and two refinement steps.
4839  // float4 recip = vrecpeq_f32(yf);
4840  // recip *= vrecpsq_f32(yf, recip);
4841  // recip *= vrecpsq_f32(yf, recip);
4842  N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4843                   DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
4844  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4845                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4846                   BN1, N2);
4847  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4848  N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4849                   DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4850                   BN1, N2);
4851  N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4852  // Simply multiplying by the reciprocal estimate can leave us a few ulps
4853  // too low, so we add 2 ulps (exhaustive testing shows that this is enough,
4854  // and that it will never cause us to return an answer too large).
4855  // float4 result = as_float4(as_int4(xf*recip) + 2);
4856  N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4857  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4858  N1 = DAG.getConstant(2, MVT::i32);
4859  N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4860  N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4861  N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4862  // Convert back to integer and return.
4863  // return vmovn_u32(vcvt_s32_f32(result));
4864  N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4865  N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4866  return N0;
4867}
4868
4869static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
4870  EVT VT = Op.getNode()->getValueType(0);
4871  SDVTList VTs = DAG.getVTList(VT, MVT::i32);
4872
4873  unsigned Opc;
4874  bool ExtraOp = false;
4875  switch (Op.getOpcode()) {
4876  default: assert(0 && "Invalid code");
4877  case ISD::ADDC: Opc = ARMISD::ADDC; break;
4878  case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
4879  case ISD::SUBC: Opc = ARMISD::SUBC; break;
4880  case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
4881  }
4882
4883  if (!ExtraOp)
4884    return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4885                       Op.getOperand(1));
4886  return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
4887                     Op.getOperand(1), Op.getOperand(2));
4888}
4889
4890static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
4891  // Monotonic load/store is legal for all targets
4892  if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
4893    return Op;
4894
4895  // Aquire/Release load/store is not legal for targets without a
4896  // dmb or equivalent available.
4897  return SDValue();
4898}
4899
4900
4901static void
4902ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
4903                    SelectionDAG &DAG, unsigned NewOp) {
4904  DebugLoc dl = Node->getDebugLoc();
4905  assert (Node->getValueType(0) == MVT::i64 &&
4906          "Only know how to expand i64 atomics");
4907
4908  SmallVector<SDValue, 6> Ops;
4909  Ops.push_back(Node->getOperand(0)); // Chain
4910  Ops.push_back(Node->getOperand(1)); // Ptr
4911  // Low part of Val1
4912  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4913                            Node->getOperand(2), DAG.getIntPtrConstant(0)));
4914  // High part of Val1
4915  Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4916                            Node->getOperand(2), DAG.getIntPtrConstant(1)));
4917  if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
4918    // High part of Val1
4919    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4920                              Node->getOperand(3), DAG.getIntPtrConstant(0)));
4921    // High part of Val2
4922    Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4923                              Node->getOperand(3), DAG.getIntPtrConstant(1)));
4924  }
4925  SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
4926  SDValue Result =
4927    DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
4928                            cast<MemSDNode>(Node)->getMemOperand());
4929  SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
4930  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
4931  Results.push_back(Result.getValue(2));
4932}
4933
4934SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
4935  switch (Op.getOpcode()) {
4936  default: llvm_unreachable("Don't know how to custom lower this!");
4937  case ISD::ConstantPool:  return LowerConstantPool(Op, DAG);
4938  case ISD::BlockAddress:  return LowerBlockAddress(Op, DAG);
4939  case ISD::GlobalAddress:
4940    return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
4941      LowerGlobalAddressELF(Op, DAG);
4942  case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
4943  case ISD::SELECT:        return LowerSELECT(Op, DAG);
4944  case ISD::SELECT_CC:     return LowerSELECT_CC(Op, DAG);
4945  case ISD::BR_CC:         return LowerBR_CC(Op, DAG);
4946  case ISD::BR_JT:         return LowerBR_JT(Op, DAG);
4947  case ISD::VASTART:       return LowerVASTART(Op, DAG);
4948  case ISD::MEMBARRIER:    return LowerMEMBARRIER(Op, DAG, Subtarget);
4949  case ISD::ATOMIC_FENCE:  return LowerATOMIC_FENCE(Op, DAG, Subtarget);
4950  case ISD::PREFETCH:      return LowerPREFETCH(Op, DAG, Subtarget);
4951  case ISD::SINT_TO_FP:
4952  case ISD::UINT_TO_FP:    return LowerINT_TO_FP(Op, DAG);
4953  case ISD::FP_TO_SINT:
4954  case ISD::FP_TO_UINT:    return LowerFP_TO_INT(Op, DAG);
4955  case ISD::FCOPYSIGN:     return LowerFCOPYSIGN(Op, DAG);
4956  case ISD::RETURNADDR:    return LowerRETURNADDR(Op, DAG);
4957  case ISD::FRAMEADDR:     return LowerFRAMEADDR(Op, DAG);
4958  case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
4959  case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
4960  case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
4961  case ISD::EH_SJLJ_DISPATCHSETUP: return LowerEH_SJLJ_DISPATCHSETUP(Op, DAG);
4962  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
4963                                                               Subtarget);
4964  case ISD::BITCAST:       return ExpandBITCAST(Op.getNode(), DAG);
4965  case ISD::SHL:
4966  case ISD::SRL:
4967  case ISD::SRA:           return LowerShift(Op.getNode(), DAG, Subtarget);
4968  case ISD::SHL_PARTS:     return LowerShiftLeftParts(Op, DAG);
4969  case ISD::SRL_PARTS:
4970  case ISD::SRA_PARTS:     return LowerShiftRightParts(Op, DAG);
4971  case ISD::CTTZ:          return LowerCTTZ(Op.getNode(), DAG, Subtarget);
4972  case ISD::SETCC:         return LowerVSETCC(Op, DAG);
4973  case ISD::BUILD_VECTOR:  return LowerBUILD_VECTOR(Op, DAG, Subtarget);
4974  case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
4975  case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
4976  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
4977  case ISD::FLT_ROUNDS_:   return LowerFLT_ROUNDS_(Op, DAG);
4978  case ISD::MUL:           return LowerMUL(Op, DAG);
4979  case ISD::SDIV:          return LowerSDIV(Op, DAG);
4980  case ISD::UDIV:          return LowerUDIV(Op, DAG);
4981  case ISD::ADDC:
4982  case ISD::ADDE:
4983  case ISD::SUBC:
4984  case ISD::SUBE:          return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
4985  case ISD::ATOMIC_LOAD:
4986  case ISD::ATOMIC_STORE:  return LowerAtomicLoadStore(Op, DAG);
4987  }
4988  return SDValue();
4989}
4990
4991/// ReplaceNodeResults - Replace the results of node with an illegal result
4992/// type with new values built out of custom code.
4993void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
4994                                           SmallVectorImpl<SDValue>&Results,
4995                                           SelectionDAG &DAG) const {
4996  SDValue Res;
4997  switch (N->getOpcode()) {
4998  default:
4999    llvm_unreachable("Don't know how to custom expand this!");
5000    break;
5001  case ISD::BITCAST:
5002    Res = ExpandBITCAST(N, DAG);
5003    break;
5004  case ISD::SRL:
5005  case ISD::SRA:
5006    Res = Expand64BitShift(N, DAG, Subtarget);
5007    break;
5008  case ISD::ATOMIC_LOAD_ADD:
5009    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5010    return;
5011  case ISD::ATOMIC_LOAD_AND:
5012    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5013    return;
5014  case ISD::ATOMIC_LOAD_NAND:
5015    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5016    return;
5017  case ISD::ATOMIC_LOAD_OR:
5018    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5019    return;
5020  case ISD::ATOMIC_LOAD_SUB:
5021    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5022    return;
5023  case ISD::ATOMIC_LOAD_XOR:
5024    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5025    return;
5026  case ISD::ATOMIC_SWAP:
5027    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5028    return;
5029  case ISD::ATOMIC_CMP_SWAP:
5030    ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5031    return;
5032  }
5033  if (Res.getNode())
5034    Results.push_back(Res);
5035}
5036
5037//===----------------------------------------------------------------------===//
5038//                           ARM Scheduler Hooks
5039//===----------------------------------------------------------------------===//
5040
5041MachineBasicBlock *
5042ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5043                                     MachineBasicBlock *BB,
5044                                     unsigned Size) const {
5045  unsigned dest    = MI->getOperand(0).getReg();
5046  unsigned ptr     = MI->getOperand(1).getReg();
5047  unsigned oldval  = MI->getOperand(2).getReg();
5048  unsigned newval  = MI->getOperand(3).getReg();
5049  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5050  DebugLoc dl = MI->getDebugLoc();
5051  bool isThumb2 = Subtarget->isThumb2();
5052
5053  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5054  unsigned scratch =
5055    MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5056                                       : ARM::GPRRegisterClass);
5057
5058  if (isThumb2) {
5059    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5060    MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5061    MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5062  }
5063
5064  unsigned ldrOpc, strOpc;
5065  switch (Size) {
5066  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5067  case 1:
5068    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5069    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5070    break;
5071  case 2:
5072    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5073    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5074    break;
5075  case 4:
5076    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5077    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5078    break;
5079  }
5080
5081  MachineFunction *MF = BB->getParent();
5082  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5083  MachineFunction::iterator It = BB;
5084  ++It; // insert the new blocks after the current block
5085
5086  MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5087  MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5088  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5089  MF->insert(It, loop1MBB);
5090  MF->insert(It, loop2MBB);
5091  MF->insert(It, exitMBB);
5092
5093  // Transfer the remainder of BB and its successor edges to exitMBB.
5094  exitMBB->splice(exitMBB->begin(), BB,
5095                  llvm::next(MachineBasicBlock::iterator(MI)),
5096                  BB->end());
5097  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5098
5099  //  thisMBB:
5100  //   ...
5101  //   fallthrough --> loop1MBB
5102  BB->addSuccessor(loop1MBB);
5103
5104  // loop1MBB:
5105  //   ldrex dest, [ptr]
5106  //   cmp dest, oldval
5107  //   bne exitMBB
5108  BB = loop1MBB;
5109  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5110  if (ldrOpc == ARM::t2LDREX)
5111    MIB.addImm(0);
5112  AddDefaultPred(MIB);
5113  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5114                 .addReg(dest).addReg(oldval));
5115  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5116    .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5117  BB->addSuccessor(loop2MBB);
5118  BB->addSuccessor(exitMBB);
5119
5120  // loop2MBB:
5121  //   strex scratch, newval, [ptr]
5122  //   cmp scratch, #0
5123  //   bne loop1MBB
5124  BB = loop2MBB;
5125  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5126  if (strOpc == ARM::t2STREX)
5127    MIB.addImm(0);
5128  AddDefaultPred(MIB);
5129  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5130                 .addReg(scratch).addImm(0));
5131  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5132    .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5133  BB->addSuccessor(loop1MBB);
5134  BB->addSuccessor(exitMBB);
5135
5136  //  exitMBB:
5137  //   ...
5138  BB = exitMBB;
5139
5140  MI->eraseFromParent();   // The instruction is gone now.
5141
5142  return BB;
5143}
5144
5145MachineBasicBlock *
5146ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5147                                    unsigned Size, unsigned BinOpcode) const {
5148  // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5149  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5150
5151  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5152  MachineFunction *MF = BB->getParent();
5153  MachineFunction::iterator It = BB;
5154  ++It;
5155
5156  unsigned dest = MI->getOperand(0).getReg();
5157  unsigned ptr = MI->getOperand(1).getReg();
5158  unsigned incr = MI->getOperand(2).getReg();
5159  DebugLoc dl = MI->getDebugLoc();
5160  bool isThumb2 = Subtarget->isThumb2();
5161
5162  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5163  if (isThumb2) {
5164    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5165    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5166  }
5167
5168  unsigned ldrOpc, strOpc;
5169  switch (Size) {
5170  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5171  case 1:
5172    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5173    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5174    break;
5175  case 2:
5176    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5177    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5178    break;
5179  case 4:
5180    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5181    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5182    break;
5183  }
5184
5185  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5186  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5187  MF->insert(It, loopMBB);
5188  MF->insert(It, exitMBB);
5189
5190  // Transfer the remainder of BB and its successor edges to exitMBB.
5191  exitMBB->splice(exitMBB->begin(), BB,
5192                  llvm::next(MachineBasicBlock::iterator(MI)),
5193                  BB->end());
5194  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5195
5196  TargetRegisterClass *TRC =
5197    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5198  unsigned scratch = MRI.createVirtualRegister(TRC);
5199  unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5200
5201  //  thisMBB:
5202  //   ...
5203  //   fallthrough --> loopMBB
5204  BB->addSuccessor(loopMBB);
5205
5206  //  loopMBB:
5207  //   ldrex dest, ptr
5208  //   <binop> scratch2, dest, incr
5209  //   strex scratch, scratch2, ptr
5210  //   cmp scratch, #0
5211  //   bne- loopMBB
5212  //   fallthrough --> exitMBB
5213  BB = loopMBB;
5214  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5215  if (ldrOpc == ARM::t2LDREX)
5216    MIB.addImm(0);
5217  AddDefaultPred(MIB);
5218  if (BinOpcode) {
5219    // operand order needs to go the other way for NAND
5220    if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5221      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5222                     addReg(incr).addReg(dest)).addReg(0);
5223    else
5224      AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5225                     addReg(dest).addReg(incr)).addReg(0);
5226  }
5227
5228  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5229  if (strOpc == ARM::t2STREX)
5230    MIB.addImm(0);
5231  AddDefaultPred(MIB);
5232  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5233                 .addReg(scratch).addImm(0));
5234  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5235    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5236
5237  BB->addSuccessor(loopMBB);
5238  BB->addSuccessor(exitMBB);
5239
5240  //  exitMBB:
5241  //   ...
5242  BB = exitMBB;
5243
5244  MI->eraseFromParent();   // The instruction is gone now.
5245
5246  return BB;
5247}
5248
5249MachineBasicBlock *
5250ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5251                                          MachineBasicBlock *BB,
5252                                          unsigned Size,
5253                                          bool signExtend,
5254                                          ARMCC::CondCodes Cond) const {
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  unsigned oldval = dest;
5266  DebugLoc dl = MI->getDebugLoc();
5267  bool isThumb2 = Subtarget->isThumb2();
5268
5269  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5270  if (isThumb2) {
5271    MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5272    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5273  }
5274
5275  unsigned ldrOpc, strOpc, extendOpc;
5276  switch (Size) {
5277  default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5278  case 1:
5279    ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5280    strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5281    extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5282    break;
5283  case 2:
5284    ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5285    strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5286    extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5287    break;
5288  case 4:
5289    ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5290    strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5291    extendOpc = 0;
5292    break;
5293  }
5294
5295  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5296  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5297  MF->insert(It, loopMBB);
5298  MF->insert(It, exitMBB);
5299
5300  // Transfer the remainder of BB and its successor edges to exitMBB.
5301  exitMBB->splice(exitMBB->begin(), BB,
5302                  llvm::next(MachineBasicBlock::iterator(MI)),
5303                  BB->end());
5304  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5305
5306  TargetRegisterClass *TRC =
5307    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5308  unsigned scratch = MRI.createVirtualRegister(TRC);
5309  unsigned scratch2 = MRI.createVirtualRegister(TRC);
5310
5311  //  thisMBB:
5312  //   ...
5313  //   fallthrough --> loopMBB
5314  BB->addSuccessor(loopMBB);
5315
5316  //  loopMBB:
5317  //   ldrex dest, ptr
5318  //   (sign extend dest, if required)
5319  //   cmp dest, incr
5320  //   cmov.cond scratch2, dest, incr
5321  //   strex scratch, scratch2, ptr
5322  //   cmp scratch, #0
5323  //   bne- loopMBB
5324  //   fallthrough --> exitMBB
5325  BB = loopMBB;
5326  MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5327  if (ldrOpc == ARM::t2LDREX)
5328    MIB.addImm(0);
5329  AddDefaultPred(MIB);
5330
5331  // Sign extend the value, if necessary.
5332  if (signExtend && extendOpc) {
5333    oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5334    AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5335                     .addReg(dest)
5336                     .addImm(0));
5337  }
5338
5339  // Build compare and cmov instructions.
5340  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5341                 .addReg(oldval).addReg(incr));
5342  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5343         .addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5344
5345  MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5346  if (strOpc == ARM::t2STREX)
5347    MIB.addImm(0);
5348  AddDefaultPred(MIB);
5349  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5350                 .addReg(scratch).addImm(0));
5351  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5352    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5353
5354  BB->addSuccessor(loopMBB);
5355  BB->addSuccessor(exitMBB);
5356
5357  //  exitMBB:
5358  //   ...
5359  BB = exitMBB;
5360
5361  MI->eraseFromParent();   // The instruction is gone now.
5362
5363  return BB;
5364}
5365
5366MachineBasicBlock *
5367ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5368                                      unsigned Op1, unsigned Op2,
5369                                      bool NeedsCarry, bool IsCmpxchg) const {
5370  // This also handles ATOMIC_SWAP, indicated by Op1==0.
5371  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5372
5373  const BasicBlock *LLVM_BB = BB->getBasicBlock();
5374  MachineFunction *MF = BB->getParent();
5375  MachineFunction::iterator It = BB;
5376  ++It;
5377
5378  unsigned destlo = MI->getOperand(0).getReg();
5379  unsigned desthi = MI->getOperand(1).getReg();
5380  unsigned ptr = MI->getOperand(2).getReg();
5381  unsigned vallo = MI->getOperand(3).getReg();
5382  unsigned valhi = MI->getOperand(4).getReg();
5383  DebugLoc dl = MI->getDebugLoc();
5384  bool isThumb2 = Subtarget->isThumb2();
5385
5386  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5387  if (isThumb2) {
5388    MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5389    MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5390    MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5391  }
5392
5393  unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5394  unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5395
5396  MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5397  MachineBasicBlock *contBB = 0, *cont2BB = 0;
5398  if (IsCmpxchg) {
5399    contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5400    cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5401  }
5402  MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5403  MF->insert(It, loopMBB);
5404  if (IsCmpxchg) {
5405    MF->insert(It, contBB);
5406    MF->insert(It, cont2BB);
5407  }
5408  MF->insert(It, exitMBB);
5409
5410  // Transfer the remainder of BB and its successor edges to exitMBB.
5411  exitMBB->splice(exitMBB->begin(), BB,
5412                  llvm::next(MachineBasicBlock::iterator(MI)),
5413                  BB->end());
5414  exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5415
5416  TargetRegisterClass *TRC =
5417    isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5418  unsigned storesuccess = MRI.createVirtualRegister(TRC);
5419
5420  //  thisMBB:
5421  //   ...
5422  //   fallthrough --> loopMBB
5423  BB->addSuccessor(loopMBB);
5424
5425  //  loopMBB:
5426  //   ldrexd r2, r3, ptr
5427  //   <binopa> r0, r2, incr
5428  //   <binopb> r1, r3, incr
5429  //   strexd storesuccess, r0, r1, ptr
5430  //   cmp storesuccess, #0
5431  //   bne- loopMBB
5432  //   fallthrough --> exitMBB
5433  //
5434  // Note that the registers are explicitly specified because there is not any
5435  // way to force the register allocator to allocate a register pair.
5436  //
5437  // FIXME: The hardcoded registers are not necessary for Thumb2, but we
5438  // need to properly enforce the restriction that the two output registers
5439  // for ldrexd must be different.
5440  BB = loopMBB;
5441  // Load
5442  AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5443                 .addReg(ARM::R2, RegState::Define)
5444                 .addReg(ARM::R3, RegState::Define).addReg(ptr));
5445  // Copy r2/r3 into dest.  (This copy will normally be coalesced.)
5446  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5447  BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5448
5449  if (IsCmpxchg) {
5450    // Add early exit
5451    for (unsigned i = 0; i < 2; i++) {
5452      AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5453                                                         ARM::CMPrr))
5454                     .addReg(i == 0 ? destlo : desthi)
5455                     .addReg(i == 0 ? vallo : valhi));
5456      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5457        .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5458      BB->addSuccessor(exitMBB);
5459      BB->addSuccessor(i == 0 ? contBB : cont2BB);
5460      BB = (i == 0 ? contBB : cont2BB);
5461    }
5462
5463    // Copy to physregs for strexd
5464    unsigned setlo = MI->getOperand(5).getReg();
5465    unsigned sethi = MI->getOperand(6).getReg();
5466    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5467    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5468  } else if (Op1) {
5469    // Perform binary operation
5470    AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5471                   .addReg(destlo).addReg(vallo))
5472        .addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5473    AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5474                   .addReg(desthi).addReg(valhi)).addReg(0);
5475  } else {
5476    // Copy to physregs for strexd
5477    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5478    BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5479  }
5480
5481  // Store
5482  AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5483                 .addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5484  // Cmp+jump
5485  AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5486                 .addReg(storesuccess).addImm(0));
5487  BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5488    .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5489
5490  BB->addSuccessor(loopMBB);
5491  BB->addSuccessor(exitMBB);
5492
5493  //  exitMBB:
5494  //   ...
5495  BB = exitMBB;
5496
5497  MI->eraseFromParent();   // The instruction is gone now.
5498
5499  return BB;
5500}
5501
5502/// EmitBasePointerRecalculation - For functions using a base pointer, we
5503/// rematerialize it (via the frame pointer).
5504void ARMTargetLowering::
5505EmitBasePointerRecalculation(MachineInstr *MI, MachineBasicBlock *MBB,
5506                             MachineBasicBlock *DispatchBB) const {
5507  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5508  const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
5509  MachineFunction &MF = *MI->getParent()->getParent();
5510  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
5511  const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
5512
5513  if (!RI.hasBasePointer(MF)) return;
5514
5515  MachineBasicBlock::iterator MBBI = MI;
5516
5517  int32_t NumBytes = AFI->getFramePtrSpillOffset();
5518  unsigned FramePtr = RI.getFrameRegister(MF);
5519  assert(MF.getTarget().getFrameLowering()->hasFP(MF) &&
5520         "Base pointer without frame pointer?");
5521
5522  if (AFI->isThumb2Function())
5523    llvm::emitT2RegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5524                                 FramePtr, -NumBytes, ARMCC::AL, 0, *AII);
5525  else if (AFI->isThumbFunction())
5526    llvm::emitThumbRegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5527                                    FramePtr, -NumBytes, *AII, RI);
5528  else
5529    llvm::emitARMRegPlusImmediate(*MBB, MBBI, MI->getDebugLoc(), ARM::R6,
5530                                  FramePtr, -NumBytes, ARMCC::AL, 0, *AII);
5531
5532  if (!RI.needsStackRealignment(MF)) return;
5533
5534  // If there's dynamic realignment, adjust for it.
5535  MachineFrameInfo *MFI = MF.getFrameInfo();
5536  unsigned MaxAlign = MFI->getMaxAlignment();
5537  assert(!AFI->isThumb1OnlyFunction());
5538
5539  // Emit bic r6, r6, MaxAlign
5540  unsigned bicOpc = AFI->isThumbFunction() ? ARM::t2BICri : ARM::BICri;
5541  AddDefaultCC(
5542    AddDefaultPred(
5543      BuildMI(*MBB, MBBI, MI->getDebugLoc(), TII->get(bicOpc), ARM::R6)
5544      .addReg(ARM::R6, RegState::Kill)
5545      .addImm(MaxAlign - 1)));
5546}
5547
5548/// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5549/// registers the function context.
5550void ARMTargetLowering::
5551SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5552                       MachineBasicBlock *DispatchBB, int FI) const {
5553  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5554  DebugLoc dl = MI->getDebugLoc();
5555  MachineFunction *MF = MBB->getParent();
5556  MachineRegisterInfo *MRI = &MF->getRegInfo();
5557  MachineConstantPool *MCP = MF->getConstantPool();
5558  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5559  const Function *F = MF->getFunction();
5560
5561  bool isThumb = Subtarget->isThumb();
5562  bool isThumb2 = Subtarget->isThumb2();
5563
5564  unsigned PCLabelId = AFI->createPICLabelUId();
5565  unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5566  ARMConstantPoolValue *CPV =
5567    ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5568  unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5569
5570  const TargetRegisterClass *TRC =
5571    isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5572
5573  // Grab constant pool and fixed stack memory operands.
5574  MachineMemOperand *CPMMO =
5575    MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5576                             MachineMemOperand::MOLoad, 4, 4);
5577
5578  MachineMemOperand *FIMMOSt =
5579    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5580                             MachineMemOperand::MOStore, 4, 4);
5581
5582  EmitBasePointerRecalculation(MI, MBB, DispatchBB);
5583
5584  // Load the address of the dispatch MBB into the jump buffer.
5585  if (isThumb2) {
5586    // Incoming value: jbuf
5587    //   ldr.n  r5, LCPI1_1
5588    //   orr    r5, r5, #1
5589    //   add    r5, pc
5590    //   str    r5, [$jbuf, #+4] ; &jbuf[1]
5591    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5592    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5593                   .addConstantPoolIndex(CPI)
5594                   .addMemOperand(CPMMO));
5595    // Set the low bit because of thumb mode.
5596    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5597    AddDefaultCC(
5598      AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5599                     .addReg(NewVReg1, RegState::Kill)
5600                     .addImm(0x01)));
5601    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5602    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5603      .addReg(NewVReg2, RegState::Kill)
5604      .addImm(PCLabelId);
5605    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5606                   .addReg(NewVReg3, RegState::Kill)
5607                   .addFrameIndex(FI)
5608                   .addImm(36)  // &jbuf[1] :: pc
5609                   .addMemOperand(FIMMOSt));
5610  } else if (isThumb) {
5611    // Incoming value: jbuf
5612    //   ldr.n  r1, LCPI1_4
5613    //   add    r1, pc
5614    //   mov    r2, #1
5615    //   orrs   r1, r2
5616    //   add    r2, $jbuf, #+4 ; &jbuf[1]
5617    //   str    r1, [r2]
5618    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5619    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5620                   .addConstantPoolIndex(CPI)
5621                   .addMemOperand(CPMMO));
5622    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5623    BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5624      .addReg(NewVReg1, RegState::Kill)
5625      .addImm(PCLabelId);
5626    // Set the low bit because of thumb mode.
5627    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5628    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5629                   .addReg(ARM::CPSR, RegState::Define)
5630                   .addImm(1));
5631    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5632    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5633                   .addReg(ARM::CPSR, RegState::Define)
5634                   .addReg(NewVReg2, RegState::Kill)
5635                   .addReg(NewVReg3, RegState::Kill));
5636    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5637    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5638                   .addFrameIndex(FI)
5639                   .addImm(36)); // &jbuf[1] :: pc
5640    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5641                   .addReg(NewVReg4, RegState::Kill)
5642                   .addReg(NewVReg5, RegState::Kill)
5643                   .addImm(0)
5644                   .addMemOperand(FIMMOSt));
5645  } else {
5646    // Incoming value: jbuf
5647    //   ldr  r1, LCPI1_1
5648    //   add  r1, pc, r1
5649    //   str  r1, [$jbuf, #+4] ; &jbuf[1]
5650    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5651    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12),  NewVReg1)
5652                   .addConstantPoolIndex(CPI)
5653                   .addImm(0)
5654                   .addMemOperand(CPMMO));
5655    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5656    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5657                   .addReg(NewVReg1, RegState::Kill)
5658                   .addImm(PCLabelId));
5659    AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5660                   .addReg(NewVReg2, RegState::Kill)
5661                   .addFrameIndex(FI)
5662                   .addImm(36)  // &jbuf[1] :: pc
5663                   .addMemOperand(FIMMOSt));
5664  }
5665}
5666
5667MachineBasicBlock *ARMTargetLowering::
5668EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5669  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5670  DebugLoc dl = MI->getDebugLoc();
5671  MachineFunction *MF = MBB->getParent();
5672  MachineRegisterInfo *MRI = &MF->getRegInfo();
5673  ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5674  MachineFrameInfo *MFI = MF->getFrameInfo();
5675  int FI = MFI->getFunctionContextIndex();
5676
5677  const TargetRegisterClass *TRC =
5678    Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5679
5680  // Get a mapping of the call site numbers to all of the landing pads they're
5681  // associated with.
5682  DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5683  unsigned MaxCSNum = 0;
5684  MachineModuleInfo &MMI = MF->getMMI();
5685  for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5686    if (!BB->isLandingPad()) continue;
5687
5688    // FIXME: We should assert that the EH_LABEL is the first MI in the landing
5689    // pad.
5690    for (MachineBasicBlock::iterator
5691           II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5692      if (!II->isEHLabel()) continue;
5693
5694      MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5695      if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5696
5697      SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5698      for (SmallVectorImpl<unsigned>::iterator
5699             CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5700           CSI != CSE; ++CSI) {
5701        CallSiteNumToLPad[*CSI].push_back(BB);
5702        MaxCSNum = std::max(MaxCSNum, *CSI);
5703      }
5704      break;
5705    }
5706  }
5707
5708  // Get an ordered list of the machine basic blocks for the jump table.
5709  std::vector<MachineBasicBlock*> LPadList;
5710  SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5711  LPadList.reserve(CallSiteNumToLPad.size());
5712  for (unsigned I = 1; I <= MaxCSNum; ++I) {
5713    SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5714    for (SmallVectorImpl<MachineBasicBlock*>::iterator
5715           II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5716      LPadList.push_back(*II);
5717      InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5718    }
5719  }
5720
5721  assert(!LPadList.empty() &&
5722         "No landing pad destinations for the dispatch jump table!");
5723
5724  // Create the jump table and associated information.
5725  MachineJumpTableInfo *JTI =
5726    MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5727  unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5728  unsigned UId = AFI->createJumpTableUId();
5729
5730  // Create the MBBs for the dispatch code.
5731
5732  // Shove the dispatch's address into the return slot in the function context.
5733  MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5734  DispatchBB->setIsLandingPad();
5735
5736  MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5737  BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5738  DispatchBB->addSuccessor(TrapBB);
5739
5740  MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5741  DispatchBB->addSuccessor(DispContBB);
5742
5743  // Insert and MBBs.
5744  MF->insert(MF->end(), DispatchBB);
5745  MF->insert(MF->end(), DispContBB);
5746  MF->insert(MF->end(), TrapBB);
5747
5748  // Insert code into the entry block that creates and registers the function
5749  // context.
5750  SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5751
5752  MachineMemOperand *FIMMOLd =
5753    MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5754                             MachineMemOperand::MOLoad |
5755                             MachineMemOperand::MOVolatile, 4, 4);
5756
5757  unsigned NumLPads = LPadList.size();
5758  if (Subtarget->isThumb2()) {
5759    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5760    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5761                   .addFrameIndex(FI)
5762                   .addImm(4)
5763                   .addMemOperand(FIMMOLd));
5764
5765    if (NumLPads < 256) {
5766      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5767                     .addReg(NewVReg1)
5768                     .addImm(LPadList.size()));
5769    } else {
5770      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5771      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5772                     .addImm(NumLPads & 0xFFFF));
5773
5774      unsigned VReg2 = VReg1;
5775      if ((NumLPads & 0xFFFF0000) != 0) {
5776        VReg2 = MRI->createVirtualRegister(TRC);
5777        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5778                       .addReg(VReg1)
5779                       .addImm(NumLPads >> 16));
5780      }
5781
5782      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5783                     .addReg(NewVReg1)
5784                     .addReg(VReg2));
5785    }
5786
5787    BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5788      .addMBB(TrapBB)
5789      .addImm(ARMCC::HI)
5790      .addReg(ARM::CPSR);
5791
5792    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5793    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5794                   .addJumpTableIndex(MJTI)
5795                   .addImm(UId));
5796
5797    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5798    AddDefaultCC(
5799      AddDefaultPred(
5800        BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5801        .addReg(NewVReg3, RegState::Kill)
5802        .addReg(NewVReg1)
5803        .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5804
5805    BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5806      .addReg(NewVReg4, RegState::Kill)
5807      .addReg(NewVReg1)
5808      .addJumpTableIndex(MJTI)
5809      .addImm(UId);
5810  } else if (Subtarget->isThumb()) {
5811    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5812    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5813                   .addFrameIndex(FI)
5814                   .addImm(1)
5815                   .addMemOperand(FIMMOLd));
5816
5817    if (NumLPads < 256) {
5818      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5819                     .addReg(NewVReg1)
5820                     .addImm(NumLPads));
5821    } else {
5822      MachineConstantPool *ConstantPool = MF->getConstantPool();
5823      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5824      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5825
5826      // MachineConstantPool wants an explicit alignment.
5827      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5828      if (Align == 0)
5829        Align = getTargetData()->getTypeAllocSize(C->getType());
5830      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5831
5832      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5833      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5834                     .addReg(VReg1, RegState::Define)
5835                     .addConstantPoolIndex(Idx));
5836      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5837                     .addReg(NewVReg1)
5838                     .addReg(VReg1));
5839    }
5840
5841    BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5842      .addMBB(TrapBB)
5843      .addImm(ARMCC::HI)
5844      .addReg(ARM::CPSR);
5845
5846    unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5847    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5848                   .addReg(ARM::CPSR, RegState::Define)
5849                   .addReg(NewVReg1)
5850                   .addImm(2));
5851
5852    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5853    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5854                   .addJumpTableIndex(MJTI)
5855                   .addImm(UId));
5856
5857    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5858    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5859                   .addReg(ARM::CPSR, RegState::Define)
5860                   .addReg(NewVReg2, RegState::Kill)
5861                   .addReg(NewVReg3));
5862
5863    MachineMemOperand *JTMMOLd =
5864      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5865                               MachineMemOperand::MOLoad, 4, 4);
5866
5867    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5868    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5869                   .addReg(NewVReg4, RegState::Kill)
5870                   .addImm(0)
5871                   .addMemOperand(JTMMOLd));
5872
5873    unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5874    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5875                   .addReg(ARM::CPSR, RegState::Define)
5876                   .addReg(NewVReg5, RegState::Kill)
5877                   .addReg(NewVReg3));
5878
5879    BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5880      .addReg(NewVReg6, RegState::Kill)
5881      .addJumpTableIndex(MJTI)
5882      .addImm(UId);
5883  } else {
5884    unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5885    AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
5886                   .addFrameIndex(FI)
5887                   .addImm(4)
5888                   .addMemOperand(FIMMOLd));
5889
5890    if (NumLPads < 256) {
5891      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
5892                     .addReg(NewVReg1)
5893                     .addImm(NumLPads));
5894    } else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
5895      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5896      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
5897                     .addImm(NumLPads & 0xFFFF));
5898
5899      unsigned VReg2 = VReg1;
5900      if ((NumLPads & 0xFFFF0000) != 0) {
5901        VReg2 = MRI->createVirtualRegister(TRC);
5902        AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
5903                       .addReg(VReg1)
5904                       .addImm(NumLPads >> 16));
5905      }
5906
5907      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5908                     .addReg(NewVReg1)
5909                     .addReg(VReg2));
5910    } else {
5911      MachineConstantPool *ConstantPool = MF->getConstantPool();
5912      Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5913      const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5914
5915      // MachineConstantPool wants an explicit alignment.
5916      unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5917      if (Align == 0)
5918        Align = getTargetData()->getTypeAllocSize(C->getType());
5919      unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5920
5921      unsigned VReg1 = MRI->createVirtualRegister(TRC);
5922      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
5923                     .addReg(VReg1, RegState::Define)
5924                     .addConstantPoolIndex(Idx));
5925      AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
5926                     .addReg(NewVReg1)
5927                     .addReg(VReg1, RegState::Kill));
5928    }
5929
5930    BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
5931      .addMBB(TrapBB)
5932      .addImm(ARMCC::HI)
5933      .addReg(ARM::CPSR);
5934
5935    unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5936    AddDefaultCC(
5937      AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
5938                     .addReg(NewVReg1)
5939                     .addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5940    unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5941    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
5942                   .addJumpTableIndex(MJTI)
5943                   .addImm(UId));
5944
5945    MachineMemOperand *JTMMOLd =
5946      MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5947                               MachineMemOperand::MOLoad, 4, 4);
5948    unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5949    AddDefaultPred(
5950      BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
5951      .addReg(NewVReg3, RegState::Kill)
5952      .addReg(NewVReg4)
5953      .addImm(0)
5954      .addMemOperand(JTMMOLd));
5955
5956    BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
5957      .addReg(NewVReg5, RegState::Kill)
5958      .addReg(NewVReg4)
5959      .addJumpTableIndex(MJTI)
5960      .addImm(UId);
5961  }
5962
5963  // Add the jump table entries as successors to the MBB.
5964  MachineBasicBlock *PrevMBB = 0;
5965  for (std::vector<MachineBasicBlock*>::iterator
5966         I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
5967    MachineBasicBlock *CurMBB = *I;
5968    if (PrevMBB != CurMBB)
5969      DispContBB->addSuccessor(CurMBB);
5970    PrevMBB = CurMBB;
5971  }
5972
5973  // N.B. the order the invoke BBs are processed in doesn't matter here.
5974  const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
5975  const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
5976  const unsigned *SavedRegs = RI.getCalleeSavedRegs(MF);
5977  SmallVector<MachineBasicBlock*, 64> MBBLPads;
5978  for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
5979         I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
5980    MachineBasicBlock *BB = *I;
5981
5982    // Remove the landing pad successor from the invoke block and replace it
5983    // with the new dispatch block.
5984    for (MachineBasicBlock::succ_iterator
5985           SI = BB->succ_begin(), SE = BB->succ_end(); SI != SE; ++SI) {
5986      MachineBasicBlock *SMBB = *SI;
5987      if (SMBB->isLandingPad()) {
5988        BB->removeSuccessor(SMBB);
5989        MBBLPads.push_back(SMBB);
5990      }
5991    }
5992
5993    BB->addSuccessor(DispatchBB);
5994
5995    // Find the invoke call and mark all of the callee-saved registers as
5996    // 'implicit defined' so that they're spilled. This prevents code from
5997    // moving instructions to before the EH block, where they will never be
5998    // executed.
5999    for (MachineBasicBlock::reverse_iterator
6000           II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6001      if (!II->getDesc().isCall()) continue;
6002
6003      DenseMap<unsigned, bool> DefRegs;
6004      for (MachineInstr::mop_iterator
6005             OI = II->operands_begin(), OE = II->operands_end();
6006           OI != OE; ++OI) {
6007        if (!OI->isReg()) continue;
6008        DefRegs[OI->getReg()] = true;
6009      }
6010
6011      MachineInstrBuilder MIB(&*II);
6012
6013      for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6014        if (!TRC->contains(SavedRegs[i])) continue;
6015        if (!DefRegs[SavedRegs[i]])
6016          MIB.addReg(SavedRegs[i], RegState::ImplicitDefine | RegState::Dead);
6017      }
6018
6019      break;
6020    }
6021  }
6022
6023  // Mark all former landing pads as non-landing pads. The dispatch is the only
6024  // landing pad now.
6025  for (SmallVectorImpl<MachineBasicBlock*>::iterator
6026         I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6027    (*I)->setIsLandingPad(false);
6028
6029  // The instruction is gone now.
6030  MI->eraseFromParent();
6031
6032  return MBB;
6033}
6034
6035static
6036MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6037  for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6038       E = MBB->succ_end(); I != E; ++I)
6039    if (*I != Succ)
6040      return *I;
6041  llvm_unreachable("Expecting a BB with two successors!");
6042}
6043
6044MachineBasicBlock *
6045ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6046                                               MachineBasicBlock *BB) const {
6047  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6048  DebugLoc dl = MI->getDebugLoc();
6049  bool isThumb2 = Subtarget->isThumb2();
6050  switch (MI->getOpcode()) {
6051  default: {
6052    MI->dump();
6053    llvm_unreachable("Unexpected instr type to insert");
6054  }
6055  // The Thumb2 pre-indexed stores have the same MI operands, they just
6056  // define them differently in the .td files from the isel patterns, so
6057  // they need pseudos.
6058  case ARM::t2STR_preidx:
6059    MI->setDesc(TII->get(ARM::t2STR_PRE));
6060    return BB;
6061  case ARM::t2STRB_preidx:
6062    MI->setDesc(TII->get(ARM::t2STRB_PRE));
6063    return BB;
6064  case ARM::t2STRH_preidx:
6065    MI->setDesc(TII->get(ARM::t2STRH_PRE));
6066    return BB;
6067
6068  case ARM::STRi_preidx:
6069  case ARM::STRBi_preidx: {
6070    unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6071      ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6072    // Decode the offset.
6073    unsigned Offset = MI->getOperand(4).getImm();
6074    bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6075    Offset = ARM_AM::getAM2Offset(Offset);
6076    if (isSub)
6077      Offset = -Offset;
6078
6079    MachineMemOperand *MMO = *MI->memoperands_begin();
6080    BuildMI(*BB, MI, dl, TII->get(NewOpc))
6081      .addOperand(MI->getOperand(0))  // Rn_wb
6082      .addOperand(MI->getOperand(1))  // Rt
6083      .addOperand(MI->getOperand(2))  // Rn
6084      .addImm(Offset)                 // offset (skip GPR==zero_reg)
6085      .addOperand(MI->getOperand(5))  // pred
6086      .addOperand(MI->getOperand(6))
6087      .addMemOperand(MMO);
6088    MI->eraseFromParent();
6089    return BB;
6090  }
6091  case ARM::STRr_preidx:
6092  case ARM::STRBr_preidx:
6093  case ARM::STRH_preidx: {
6094    unsigned NewOpc;
6095    switch (MI->getOpcode()) {
6096    default: llvm_unreachable("unexpected opcode!");
6097    case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6098    case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6099    case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6100    }
6101    MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6102    for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6103      MIB.addOperand(MI->getOperand(i));
6104    MI->eraseFromParent();
6105    return BB;
6106  }
6107  case ARM::ATOMIC_LOAD_ADD_I8:
6108     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6109  case ARM::ATOMIC_LOAD_ADD_I16:
6110     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6111  case ARM::ATOMIC_LOAD_ADD_I32:
6112     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6113
6114  case ARM::ATOMIC_LOAD_AND_I8:
6115     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6116  case ARM::ATOMIC_LOAD_AND_I16:
6117     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6118  case ARM::ATOMIC_LOAD_AND_I32:
6119     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6120
6121  case ARM::ATOMIC_LOAD_OR_I8:
6122     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6123  case ARM::ATOMIC_LOAD_OR_I16:
6124     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6125  case ARM::ATOMIC_LOAD_OR_I32:
6126     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6127
6128  case ARM::ATOMIC_LOAD_XOR_I8:
6129     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6130  case ARM::ATOMIC_LOAD_XOR_I16:
6131     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6132  case ARM::ATOMIC_LOAD_XOR_I32:
6133     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6134
6135  case ARM::ATOMIC_LOAD_NAND_I8:
6136     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6137  case ARM::ATOMIC_LOAD_NAND_I16:
6138     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6139  case ARM::ATOMIC_LOAD_NAND_I32:
6140     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6141
6142  case ARM::ATOMIC_LOAD_SUB_I8:
6143     return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6144  case ARM::ATOMIC_LOAD_SUB_I16:
6145     return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6146  case ARM::ATOMIC_LOAD_SUB_I32:
6147     return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6148
6149  case ARM::ATOMIC_LOAD_MIN_I8:
6150     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6151  case ARM::ATOMIC_LOAD_MIN_I16:
6152     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6153  case ARM::ATOMIC_LOAD_MIN_I32:
6154     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6155
6156  case ARM::ATOMIC_LOAD_MAX_I8:
6157     return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6158  case ARM::ATOMIC_LOAD_MAX_I16:
6159     return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6160  case ARM::ATOMIC_LOAD_MAX_I32:
6161     return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6162
6163  case ARM::ATOMIC_LOAD_UMIN_I8:
6164     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6165  case ARM::ATOMIC_LOAD_UMIN_I16:
6166     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6167  case ARM::ATOMIC_LOAD_UMIN_I32:
6168     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6169
6170  case ARM::ATOMIC_LOAD_UMAX_I8:
6171     return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6172  case ARM::ATOMIC_LOAD_UMAX_I16:
6173     return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6174  case ARM::ATOMIC_LOAD_UMAX_I32:
6175     return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6176
6177  case ARM::ATOMIC_SWAP_I8:  return EmitAtomicBinary(MI, BB, 1, 0);
6178  case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6179  case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6180
6181  case ARM::ATOMIC_CMP_SWAP_I8:  return EmitAtomicCmpSwap(MI, BB, 1);
6182  case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6183  case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6184
6185
6186  case ARM::ATOMADD6432:
6187    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6188                              isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6189                              /*NeedsCarry*/ true);
6190  case ARM::ATOMSUB6432:
6191    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6192                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6193                              /*NeedsCarry*/ true);
6194  case ARM::ATOMOR6432:
6195    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6196                              isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6197  case ARM::ATOMXOR6432:
6198    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6199                              isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6200  case ARM::ATOMAND6432:
6201    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6202                              isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6203  case ARM::ATOMSWAP6432:
6204    return EmitAtomicBinary64(MI, BB, 0, 0, false);
6205  case ARM::ATOMCMPXCHG6432:
6206    return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6207                              isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6208                              /*NeedsCarry*/ false, /*IsCmpxchg*/true);
6209
6210  case ARM::tMOVCCr_pseudo: {
6211    // To "insert" a SELECT_CC instruction, we actually have to insert the
6212    // diamond control-flow pattern.  The incoming instruction knows the
6213    // destination vreg to set, the condition code register to branch on, the
6214    // true/false values to select between, and a branch opcode to use.
6215    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6216    MachineFunction::iterator It = BB;
6217    ++It;
6218
6219    //  thisMBB:
6220    //  ...
6221    //   TrueVal = ...
6222    //   cmpTY ccX, r1, r2
6223    //   bCC copy1MBB
6224    //   fallthrough --> copy0MBB
6225    MachineBasicBlock *thisMBB  = BB;
6226    MachineFunction *F = BB->getParent();
6227    MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6228    MachineBasicBlock *sinkMBB  = F->CreateMachineBasicBlock(LLVM_BB);
6229    F->insert(It, copy0MBB);
6230    F->insert(It, sinkMBB);
6231
6232    // Transfer the remainder of BB and its successor edges to sinkMBB.
6233    sinkMBB->splice(sinkMBB->begin(), BB,
6234                    llvm::next(MachineBasicBlock::iterator(MI)),
6235                    BB->end());
6236    sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6237
6238    BB->addSuccessor(copy0MBB);
6239    BB->addSuccessor(sinkMBB);
6240
6241    BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6242      .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6243
6244    //  copy0MBB:
6245    //   %FalseValue = ...
6246    //   # fallthrough to sinkMBB
6247    BB = copy0MBB;
6248
6249    // Update machine-CFG edges
6250    BB->addSuccessor(sinkMBB);
6251
6252    //  sinkMBB:
6253    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6254    //  ...
6255    BB = sinkMBB;
6256    BuildMI(*BB, BB->begin(), dl,
6257            TII->get(ARM::PHI), MI->getOperand(0).getReg())
6258      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6259      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6260
6261    MI->eraseFromParent();   // The pseudo instruction is gone now.
6262    return BB;
6263  }
6264
6265  case ARM::BCCi64:
6266  case ARM::BCCZi64: {
6267    // If there is an unconditional branch to the other successor, remove it.
6268    BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6269
6270    // Compare both parts that make up the double comparison separately for
6271    // equality.
6272    bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6273
6274    unsigned LHS1 = MI->getOperand(1).getReg();
6275    unsigned LHS2 = MI->getOperand(2).getReg();
6276    if (RHSisZero) {
6277      AddDefaultPred(BuildMI(BB, dl,
6278                             TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6279                     .addReg(LHS1).addImm(0));
6280      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6281        .addReg(LHS2).addImm(0)
6282        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6283    } else {
6284      unsigned RHS1 = MI->getOperand(3).getReg();
6285      unsigned RHS2 = MI->getOperand(4).getReg();
6286      AddDefaultPred(BuildMI(BB, dl,
6287                             TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6288                     .addReg(LHS1).addReg(RHS1));
6289      BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6290        .addReg(LHS2).addReg(RHS2)
6291        .addImm(ARMCC::EQ).addReg(ARM::CPSR);
6292    }
6293
6294    MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6295    MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6296    if (MI->getOperand(0).getImm() == ARMCC::NE)
6297      std::swap(destMBB, exitMBB);
6298
6299    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6300      .addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6301    if (isThumb2)
6302      AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6303    else
6304      BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6305
6306    MI->eraseFromParent();   // The pseudo instruction is gone now.
6307    return BB;
6308  }
6309
6310  case ARM::Int_eh_sjlj_setjmp:
6311  case ARM::Int_eh_sjlj_setjmp_nofp:
6312  case ARM::tInt_eh_sjlj_setjmp:
6313  case ARM::t2Int_eh_sjlj_setjmp:
6314  case ARM::t2Int_eh_sjlj_setjmp_nofp:
6315    EmitSjLjDispatchBlock(MI, BB);
6316    return BB;
6317
6318  case ARM::ABS:
6319  case ARM::t2ABS: {
6320    // To insert an ABS instruction, we have to insert the
6321    // diamond control-flow pattern.  The incoming instruction knows the
6322    // source vreg to test against 0, the destination vreg to set,
6323    // the condition code register to branch on, the
6324    // true/false values to select between, and a branch opcode to use.
6325    // It transforms
6326    //     V1 = ABS V0
6327    // into
6328    //     V2 = MOVS V0
6329    //     BCC                      (branch to SinkBB if V0 >= 0)
6330    //     RSBBB: V3 = RSBri V2, 0  (compute ABS if V2 < 0)
6331    //     SinkBB: V1 = PHI(V2, V3)
6332    const BasicBlock *LLVM_BB = BB->getBasicBlock();
6333    MachineFunction::iterator BBI = BB;
6334    ++BBI;
6335    MachineFunction *Fn = BB->getParent();
6336    MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6337    MachineBasicBlock *SinkBB  = Fn->CreateMachineBasicBlock(LLVM_BB);
6338    Fn->insert(BBI, RSBBB);
6339    Fn->insert(BBI, SinkBB);
6340
6341    unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6342    unsigned int ABSDstReg = MI->getOperand(0).getReg();
6343    bool isThumb2 = Subtarget->isThumb2();
6344    MachineRegisterInfo &MRI = Fn->getRegInfo();
6345    // In Thumb mode S must not be specified if source register is the SP or
6346    // PC and if destination register is the SP, so restrict register class
6347    unsigned NewMovDstReg = MRI.createVirtualRegister(
6348      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6349    unsigned NewRsbDstReg = MRI.createVirtualRegister(
6350      isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6351
6352    // Transfer the remainder of BB and its successor edges to sinkMBB.
6353    SinkBB->splice(SinkBB->begin(), BB,
6354      llvm::next(MachineBasicBlock::iterator(MI)),
6355      BB->end());
6356    SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6357
6358    BB->addSuccessor(RSBBB);
6359    BB->addSuccessor(SinkBB);
6360
6361    // fall through to SinkMBB
6362    RSBBB->addSuccessor(SinkBB);
6363
6364    // insert a movs at the end of BB
6365    BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6366      NewMovDstReg)
6367      .addReg(ABSSrcReg, RegState::Kill)
6368      .addImm((unsigned)ARMCC::AL).addReg(0)
6369      .addReg(ARM::CPSR, RegState::Define);
6370
6371    // insert a bcc with opposite CC to ARMCC::MI at the end of BB
6372    BuildMI(BB, dl,
6373      TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6374      .addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6375
6376    // insert rsbri in RSBBB
6377    // Note: BCC and rsbri will be converted into predicated rsbmi
6378    // by if-conversion pass
6379    BuildMI(*RSBBB, RSBBB->begin(), dl,
6380      TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6381      .addReg(NewMovDstReg, RegState::Kill)
6382      .addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6383
6384    // insert PHI in SinkBB,
6385    // reuse ABSDstReg to not change uses of ABS instruction
6386    BuildMI(*SinkBB, SinkBB->begin(), dl,
6387      TII->get(ARM::PHI), ABSDstReg)
6388      .addReg(NewRsbDstReg).addMBB(RSBBB)
6389      .addReg(NewMovDstReg).addMBB(BB);
6390
6391    // remove ABS instruction
6392    MI->eraseFromParent();
6393
6394    // return last added BB
6395    return SinkBB;
6396  }
6397  }
6398}
6399
6400void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6401                                                      SDNode *Node) const {
6402  const MCInstrDesc *MCID = &MI->getDesc();
6403  if (!MCID->hasPostISelHook()) {
6404    assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6405           "Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6406    return;
6407  }
6408
6409  // Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6410  // RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6411  // operand is still set to noreg. If needed, set the optional operand's
6412  // register to CPSR, and remove the redundant implicit def.
6413  //
6414  // e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6415
6416  // Rename pseudo opcodes.
6417  unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6418  if (NewOpc) {
6419    const ARMBaseInstrInfo *TII =
6420      static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6421    MCID = &TII->get(NewOpc);
6422
6423    assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6424           "converted opcode should be the same except for cc_out");
6425
6426    MI->setDesc(*MCID);
6427
6428    // Add the optional cc_out operand
6429    MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6430  }
6431  unsigned ccOutIdx = MCID->getNumOperands() - 1;
6432
6433  // Any ARM instruction that sets the 's' bit should specify an optional
6434  // "cc_out" operand in the last operand position.
6435  if (!MCID->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6436    assert(!NewOpc && "Optional cc_out operand required");
6437    return;
6438  }
6439  // Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6440  // since we already have an optional CPSR def.
6441  bool definesCPSR = false;
6442  bool deadCPSR = false;
6443  for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6444       i != e; ++i) {
6445    const MachineOperand &MO = MI->getOperand(i);
6446    if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6447      definesCPSR = true;
6448      if (MO.isDead())
6449        deadCPSR = true;
6450      MI->RemoveOperand(i);
6451      break;
6452    }
6453  }
6454  if (!definesCPSR) {
6455    assert(!NewOpc && "Optional cc_out operand required");
6456    return;
6457  }
6458  assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6459  if (deadCPSR) {
6460    assert(!MI->getOperand(ccOutIdx).getReg() &&
6461           "expect uninitialized optional cc_out operand");
6462    return;
6463  }
6464
6465  // If this instruction was defined with an optional CPSR def and its dag node
6466  // had a live implicit CPSR def, then activate the optional CPSR def.
6467  MachineOperand &MO = MI->getOperand(ccOutIdx);
6468  MO.setReg(ARM::CPSR);
6469  MO.setIsDef(true);
6470}
6471
6472//===----------------------------------------------------------------------===//
6473//                           ARM Optimization Hooks
6474//===----------------------------------------------------------------------===//
6475
6476static
6477SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6478                            TargetLowering::DAGCombinerInfo &DCI) {
6479  SelectionDAG &DAG = DCI.DAG;
6480  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6481  EVT VT = N->getValueType(0);
6482  unsigned Opc = N->getOpcode();
6483  bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6484  SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6485  SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6486  ISD::CondCode CC = ISD::SETCC_INVALID;
6487
6488  if (isSlctCC) {
6489    CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6490  } else {
6491    SDValue CCOp = Slct.getOperand(0);
6492    if (CCOp.getOpcode() == ISD::SETCC)
6493      CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6494  }
6495
6496  bool DoXform = false;
6497  bool InvCC = false;
6498  assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6499          "Bad input!");
6500
6501  if (LHS.getOpcode() == ISD::Constant &&
6502      cast<ConstantSDNode>(LHS)->isNullValue()) {
6503    DoXform = true;
6504  } else if (CC != ISD::SETCC_INVALID &&
6505             RHS.getOpcode() == ISD::Constant &&
6506             cast<ConstantSDNode>(RHS)->isNullValue()) {
6507    std::swap(LHS, RHS);
6508    SDValue Op0 = Slct.getOperand(0);
6509    EVT OpVT = isSlctCC ? Op0.getValueType() :
6510                          Op0.getOperand(0).getValueType();
6511    bool isInt = OpVT.isInteger();
6512    CC = ISD::getSetCCInverse(CC, isInt);
6513
6514    if (!TLI.isCondCodeLegal(CC, OpVT))
6515      return SDValue();         // Inverse operator isn't legal.
6516
6517    DoXform = true;
6518    InvCC = true;
6519  }
6520
6521  if (DoXform) {
6522    SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6523    if (isSlctCC)
6524      return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6525                             Slct.getOperand(0), Slct.getOperand(1), CC);
6526    SDValue CCOp = Slct.getOperand(0);
6527    if (InvCC)
6528      CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6529                          CCOp.getOperand(0), CCOp.getOperand(1), CC);
6530    return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6531                       CCOp, OtherOp, Result);
6532  }
6533  return SDValue();
6534}
6535
6536// AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6537// (only after legalization).
6538static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6539                                 TargetLowering::DAGCombinerInfo &DCI,
6540                                 const ARMSubtarget *Subtarget) {
6541
6542  // Only perform optimization if after legalize, and if NEON is available. We
6543  // also expected both operands to be BUILD_VECTORs.
6544  if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6545      || N0.getOpcode() != ISD::BUILD_VECTOR
6546      || N1.getOpcode() != ISD::BUILD_VECTOR)
6547    return SDValue();
6548
6549  // Check output type since VPADDL operand elements can only be 8, 16, or 32.
6550  EVT VT = N->getValueType(0);
6551  if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6552    return SDValue();
6553
6554  // Check that the vector operands are of the right form.
6555  // N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6556  // operands, where N is the size of the formed vector.
6557  // Each EXTRACT_VECTOR should have the same input vector and odd or even
6558  // index such that we have a pair wise add pattern.
6559
6560  // Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6561  if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6562    return SDValue();
6563  SDValue Vec = N0->getOperand(0)->getOperand(0);
6564  SDNode *V = Vec.getNode();
6565  unsigned nextIndex = 0;
6566
6567  // For each operands to the ADD which are BUILD_VECTORs,
6568  // check to see if each of their operands are an EXTRACT_VECTOR with
6569  // the same vector and appropriate index.
6570  for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6571    if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6572        && N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6573
6574      SDValue ExtVec0 = N0->getOperand(i);
6575      SDValue ExtVec1 = N1->getOperand(i);
6576
6577      // First operand is the vector, verify its the same.
6578      if (V != ExtVec0->getOperand(0).getNode() ||
6579          V != ExtVec1->getOperand(0).getNode())
6580        return SDValue();
6581
6582      // Second is the constant, verify its correct.
6583      ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6584      ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6585
6586      // For the constant, we want to see all the even or all the odd.
6587      if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6588          || C1->getZExtValue() != nextIndex+1)
6589        return SDValue();
6590
6591      // Increment index.
6592      nextIndex+=2;
6593    } else
6594      return SDValue();
6595  }
6596
6597  // Create VPADDL node.
6598  SelectionDAG &DAG = DCI.DAG;
6599  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6600
6601  // Build operand list.
6602  SmallVector<SDValue, 8> Ops;
6603  Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6604                                TLI.getPointerTy()));
6605
6606  // Input is the vector.
6607  Ops.push_back(Vec);
6608
6609  // Get widened type and narrowed type.
6610  MVT widenType;
6611  unsigned numElem = VT.getVectorNumElements();
6612  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6613    case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6614    case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6615    case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6616    default:
6617      assert(0 && "Invalid vector element type for padd optimization.");
6618  }
6619
6620  SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6621                            widenType, &Ops[0], Ops.size());
6622  return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6623}
6624
6625/// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6626/// operands N0 and N1.  This is a helper for PerformADDCombine that is
6627/// called with the default operands, and if that fails, with commuted
6628/// operands.
6629static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6630                                          TargetLowering::DAGCombinerInfo &DCI,
6631                                          const ARMSubtarget *Subtarget){
6632
6633  // Attempt to create vpaddl for this add.
6634  SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6635  if (Result.getNode())
6636    return Result;
6637
6638  // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6639  if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6640    SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6641    if (Result.getNode()) return Result;
6642  }
6643  return SDValue();
6644}
6645
6646/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6647///
6648static SDValue PerformADDCombine(SDNode *N,
6649                                 TargetLowering::DAGCombinerInfo &DCI,
6650                                 const ARMSubtarget *Subtarget) {
6651  SDValue N0 = N->getOperand(0);
6652  SDValue N1 = N->getOperand(1);
6653
6654  // First try with the default operand order.
6655  SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6656  if (Result.getNode())
6657    return Result;
6658
6659  // If that didn't work, try again with the operands commuted.
6660  return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6661}
6662
6663/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6664///
6665static SDValue PerformSUBCombine(SDNode *N,
6666                                 TargetLowering::DAGCombinerInfo &DCI) {
6667  SDValue N0 = N->getOperand(0);
6668  SDValue N1 = N->getOperand(1);
6669
6670  // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6671  if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6672    SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6673    if (Result.getNode()) return Result;
6674  }
6675
6676  return SDValue();
6677}
6678
6679/// PerformVMULCombine
6680/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6681/// special multiplier accumulator forwarding.
6682///   vmul d3, d0, d2
6683///   vmla d3, d1, d2
6684/// is faster than
6685///   vadd d3, d0, d1
6686///   vmul d3, d3, d2
6687static SDValue PerformVMULCombine(SDNode *N,
6688                                  TargetLowering::DAGCombinerInfo &DCI,
6689                                  const ARMSubtarget *Subtarget) {
6690  if (!Subtarget->hasVMLxForwarding())
6691    return SDValue();
6692
6693  SelectionDAG &DAG = DCI.DAG;
6694  SDValue N0 = N->getOperand(0);
6695  SDValue N1 = N->getOperand(1);
6696  unsigned Opcode = N0.getOpcode();
6697  if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6698      Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6699    Opcode = N1.getOpcode();
6700    if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6701        Opcode != ISD::FADD && Opcode != ISD::FSUB)
6702      return SDValue();
6703    std::swap(N0, N1);
6704  }
6705
6706  EVT VT = N->getValueType(0);
6707  DebugLoc DL = N->getDebugLoc();
6708  SDValue N00 = N0->getOperand(0);
6709  SDValue N01 = N0->getOperand(1);
6710  return DAG.getNode(Opcode, DL, VT,
6711                     DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6712                     DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6713}
6714
6715static SDValue PerformMULCombine(SDNode *N,
6716                                 TargetLowering::DAGCombinerInfo &DCI,
6717                                 const ARMSubtarget *Subtarget) {
6718  SelectionDAG &DAG = DCI.DAG;
6719
6720  if (Subtarget->isThumb1Only())
6721    return SDValue();
6722
6723  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6724    return SDValue();
6725
6726  EVT VT = N->getValueType(0);
6727  if (VT.is64BitVector() || VT.is128BitVector())
6728    return PerformVMULCombine(N, DCI, Subtarget);
6729  if (VT != MVT::i32)
6730    return SDValue();
6731
6732  ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6733  if (!C)
6734    return SDValue();
6735
6736  uint64_t MulAmt = C->getZExtValue();
6737  unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6738  ShiftAmt = ShiftAmt & (32 - 1);
6739  SDValue V = N->getOperand(0);
6740  DebugLoc DL = N->getDebugLoc();
6741
6742  SDValue Res;
6743  MulAmt >>= ShiftAmt;
6744  if (isPowerOf2_32(MulAmt - 1)) {
6745    // (mul x, 2^N + 1) => (add (shl x, N), x)
6746    Res = DAG.getNode(ISD::ADD, DL, VT,
6747                      V, DAG.getNode(ISD::SHL, DL, VT,
6748                                     V, DAG.getConstant(Log2_32(MulAmt-1),
6749                                                        MVT::i32)));
6750  } else if (isPowerOf2_32(MulAmt + 1)) {
6751    // (mul x, 2^N - 1) => (sub (shl x, N), x)
6752    Res = DAG.getNode(ISD::SUB, DL, VT,
6753                      DAG.getNode(ISD::SHL, DL, VT,
6754                                  V, DAG.getConstant(Log2_32(MulAmt+1),
6755                                                     MVT::i32)),
6756                                                     V);
6757  } else
6758    return SDValue();
6759
6760  if (ShiftAmt != 0)
6761    Res = DAG.getNode(ISD::SHL, DL, VT, Res,
6762                      DAG.getConstant(ShiftAmt, MVT::i32));
6763
6764  // Do not add new nodes to DAG combiner worklist.
6765  DCI.CombineTo(N, Res, false);
6766  return SDValue();
6767}
6768
6769static SDValue PerformANDCombine(SDNode *N,
6770                                TargetLowering::DAGCombinerInfo &DCI) {
6771
6772  // Attempt to use immediate-form VBIC
6773  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6774  DebugLoc dl = N->getDebugLoc();
6775  EVT VT = N->getValueType(0);
6776  SelectionDAG &DAG = DCI.DAG;
6777
6778  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6779    return SDValue();
6780
6781  APInt SplatBits, SplatUndef;
6782  unsigned SplatBitSize;
6783  bool HasAnyUndefs;
6784  if (BVN &&
6785      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6786    if (SplatBitSize <= 64) {
6787      EVT VbicVT;
6788      SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6789                                      SplatUndef.getZExtValue(), SplatBitSize,
6790                                      DAG, VbicVT, VT.is128BitVector(),
6791                                      OtherModImm);
6792      if (Val.getNode()) {
6793        SDValue Input =
6794          DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6795        SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
6796        return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
6797      }
6798    }
6799  }
6800
6801  return SDValue();
6802}
6803
6804/// PerformORCombine - Target-specific dag combine xforms for ISD::OR
6805static SDValue PerformORCombine(SDNode *N,
6806                                TargetLowering::DAGCombinerInfo &DCI,
6807                                const ARMSubtarget *Subtarget) {
6808  // Attempt to use immediate-form VORR
6809  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6810  DebugLoc dl = N->getDebugLoc();
6811  EVT VT = N->getValueType(0);
6812  SelectionDAG &DAG = DCI.DAG;
6813
6814  if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6815    return SDValue();
6816
6817  APInt SplatBits, SplatUndef;
6818  unsigned SplatBitSize;
6819  bool HasAnyUndefs;
6820  if (BVN && Subtarget->hasNEON() &&
6821      BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6822    if (SplatBitSize <= 64) {
6823      EVT VorrVT;
6824      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
6825                                      SplatUndef.getZExtValue(), SplatBitSize,
6826                                      DAG, VorrVT, VT.is128BitVector(),
6827                                      OtherModImm);
6828      if (Val.getNode()) {
6829        SDValue Input =
6830          DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
6831        SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
6832        return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
6833      }
6834    }
6835  }
6836
6837  SDValue N0 = N->getOperand(0);
6838  if (N0.getOpcode() != ISD::AND)
6839    return SDValue();
6840  SDValue N1 = N->getOperand(1);
6841
6842  // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
6843  if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
6844      DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
6845    APInt SplatUndef;
6846    unsigned SplatBitSize;
6847    bool HasAnyUndefs;
6848
6849    BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
6850    APInt SplatBits0;
6851    if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
6852                                  HasAnyUndefs) && !HasAnyUndefs) {
6853      BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
6854      APInt SplatBits1;
6855      if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
6856                                    HasAnyUndefs) && !HasAnyUndefs &&
6857          SplatBits0 == ~SplatBits1) {
6858        // Canonicalize the vector type to make instruction selection simpler.
6859        EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
6860        SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
6861                                     N0->getOperand(1), N0->getOperand(0),
6862                                     N1->getOperand(0));
6863        return DAG.getNode(ISD::BITCAST, dl, VT, Result);
6864      }
6865    }
6866  }
6867
6868  // Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
6869  // reasonable.
6870
6871  // BFI is only available on V6T2+
6872  if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
6873    return SDValue();
6874
6875  DebugLoc DL = N->getDebugLoc();
6876  // 1) or (and A, mask), val => ARMbfi A, val, mask
6877  //      iff (val & mask) == val
6878  //
6879  // 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6880  //  2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
6881  //          && mask == ~mask2
6882  //  2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
6883  //          && ~mask == mask2
6884  //  (i.e., copy a bitfield value into another bitfield of the same width)
6885
6886  if (VT != MVT::i32)
6887    return SDValue();
6888
6889  SDValue N00 = N0.getOperand(0);
6890
6891  // The value and the mask need to be constants so we can verify this is
6892  // actually a bitfield set. If the mask is 0xffff, we can do better
6893  // via a movt instruction, so don't use BFI in that case.
6894  SDValue MaskOp = N0.getOperand(1);
6895  ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
6896  if (!MaskC)
6897    return SDValue();
6898  unsigned Mask = MaskC->getZExtValue();
6899  if (Mask == 0xffff)
6900    return SDValue();
6901  SDValue Res;
6902  // Case (1): or (and A, mask), val => ARMbfi A, val, mask
6903  ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
6904  if (N1C) {
6905    unsigned Val = N1C->getZExtValue();
6906    if ((Val & ~Mask) != Val)
6907      return SDValue();
6908
6909    if (ARM::isBitFieldInvertedMask(Mask)) {
6910      Val >>= CountTrailingZeros_32(~Mask);
6911
6912      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
6913                        DAG.getConstant(Val, MVT::i32),
6914                        DAG.getConstant(Mask, MVT::i32));
6915
6916      // Do not add new nodes to DAG combiner worklist.
6917      DCI.CombineTo(N, Res, false);
6918      return SDValue();
6919    }
6920  } else if (N1.getOpcode() == ISD::AND) {
6921    // case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
6922    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6923    if (!N11C)
6924      return SDValue();
6925    unsigned Mask2 = N11C->getZExtValue();
6926
6927    // Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
6928    // as is to match.
6929    if (ARM::isBitFieldInvertedMask(Mask) &&
6930        (Mask == ~Mask2)) {
6931      // The pack halfword instruction works better for masks that fit it,
6932      // so use that when it's available.
6933      if (Subtarget->hasT2ExtractPack() &&
6934          (Mask == 0xffff || Mask == 0xffff0000))
6935        return SDValue();
6936      // 2a
6937      unsigned amt = CountTrailingZeros_32(Mask2);
6938      Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
6939                        DAG.getConstant(amt, MVT::i32));
6940      Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
6941                        DAG.getConstant(Mask, MVT::i32));
6942      // Do not add new nodes to DAG combiner worklist.
6943      DCI.CombineTo(N, Res, false);
6944      return SDValue();
6945    } else if (ARM::isBitFieldInvertedMask(~Mask) &&
6946               (~Mask == Mask2)) {
6947      // The pack halfword instruction works better for masks that fit it,
6948      // so use that when it's available.
6949      if (Subtarget->hasT2ExtractPack() &&
6950          (Mask2 == 0xffff || Mask2 == 0xffff0000))
6951        return SDValue();
6952      // 2b
6953      unsigned lsb = CountTrailingZeros_32(Mask);
6954      Res = DAG.getNode(ISD::SRL, DL, VT, N00,
6955                        DAG.getConstant(lsb, MVT::i32));
6956      Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
6957                        DAG.getConstant(Mask2, MVT::i32));
6958      // Do not add new nodes to DAG combiner worklist.
6959      DCI.CombineTo(N, Res, false);
6960      return SDValue();
6961    }
6962  }
6963
6964  if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
6965      N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
6966      ARM::isBitFieldInvertedMask(~Mask)) {
6967    // Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
6968    // where lsb(mask) == #shamt and masked bits of B are known zero.
6969    SDValue ShAmt = N00.getOperand(1);
6970    unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6971    unsigned LSB = CountTrailingZeros_32(Mask);
6972    if (ShAmtC != LSB)
6973      return SDValue();
6974
6975    Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
6976                      DAG.getConstant(~Mask, MVT::i32));
6977
6978    // Do not add new nodes to DAG combiner worklist.
6979    DCI.CombineTo(N, Res, false);
6980  }
6981
6982  return SDValue();
6983}
6984
6985/// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
6986/// the bits being cleared by the AND are not demanded by the BFI.
6987static SDValue PerformBFICombine(SDNode *N,
6988                                 TargetLowering::DAGCombinerInfo &DCI) {
6989  SDValue N1 = N->getOperand(1);
6990  if (N1.getOpcode() == ISD::AND) {
6991    ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
6992    if (!N11C)
6993      return SDValue();
6994    unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
6995    unsigned LSB = CountTrailingZeros_32(~InvMask);
6996    unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
6997    unsigned Mask = (1 << Width)-1;
6998    unsigned Mask2 = N11C->getZExtValue();
6999    if ((Mask & (~Mask2)) == 0)
7000      return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7001                             N->getOperand(0), N1.getOperand(0),
7002                             N->getOperand(2));
7003  }
7004  return SDValue();
7005}
7006
7007/// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7008/// ARMISD::VMOVRRD.
7009static SDValue PerformVMOVRRDCombine(SDNode *N,
7010                                     TargetLowering::DAGCombinerInfo &DCI) {
7011  // vmovrrd(vmovdrr x, y) -> x,y
7012  SDValue InDouble = N->getOperand(0);
7013  if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7014    return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7015
7016  // vmovrrd(load f64) -> (load i32), (load i32)
7017  SDNode *InNode = InDouble.getNode();
7018  if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7019      InNode->getValueType(0) == MVT::f64 &&
7020      InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7021      !cast<LoadSDNode>(InNode)->isVolatile()) {
7022    // TODO: Should this be done for non-FrameIndex operands?
7023    LoadSDNode *LD = cast<LoadSDNode>(InNode);
7024
7025    SelectionDAG &DAG = DCI.DAG;
7026    DebugLoc DL = LD->getDebugLoc();
7027    SDValue BasePtr = LD->getBasePtr();
7028    SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7029                                 LD->getPointerInfo(), LD->isVolatile(),
7030                                 LD->isNonTemporal(), LD->getAlignment());
7031
7032    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7033                                    DAG.getConstant(4, MVT::i32));
7034    SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7035                                 LD->getPointerInfo(), LD->isVolatile(),
7036                                 LD->isNonTemporal(),
7037                                 std::min(4U, LD->getAlignment() / 2));
7038
7039    DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7040    SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7041    DCI.RemoveFromWorklist(LD);
7042    DAG.DeleteNode(LD);
7043    return Result;
7044  }
7045
7046  return SDValue();
7047}
7048
7049/// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7050/// ARMISD::VMOVDRR.  This is also used for BUILD_VECTORs with 2 operands.
7051static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7052  // N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7053  SDValue Op0 = N->getOperand(0);
7054  SDValue Op1 = N->getOperand(1);
7055  if (Op0.getOpcode() == ISD::BITCAST)
7056    Op0 = Op0.getOperand(0);
7057  if (Op1.getOpcode() == ISD::BITCAST)
7058    Op1 = Op1.getOperand(0);
7059  if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7060      Op0.getNode() == Op1.getNode() &&
7061      Op0.getResNo() == 0 && Op1.getResNo() == 1)
7062    return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7063                       N->getValueType(0), Op0.getOperand(0));
7064  return SDValue();
7065}
7066
7067/// PerformSTORECombine - Target-specific dag combine xforms for
7068/// ISD::STORE.
7069static SDValue PerformSTORECombine(SDNode *N,
7070                                   TargetLowering::DAGCombinerInfo &DCI) {
7071  // Bitcast an i64 store extracted from a vector to f64.
7072  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7073  StoreSDNode *St = cast<StoreSDNode>(N);
7074  SDValue StVal = St->getValue();
7075  if (!ISD::isNormalStore(St) || St->isVolatile())
7076    return SDValue();
7077
7078  if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7079      StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7080    SelectionDAG  &DAG = DCI.DAG;
7081    DebugLoc DL = St->getDebugLoc();
7082    SDValue BasePtr = St->getBasePtr();
7083    SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7084                                  StVal.getNode()->getOperand(0), BasePtr,
7085                                  St->getPointerInfo(), St->isVolatile(),
7086                                  St->isNonTemporal(), St->getAlignment());
7087
7088    SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7089                                    DAG.getConstant(4, MVT::i32));
7090    return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7091                        OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7092                        St->isNonTemporal(),
7093                        std::min(4U, St->getAlignment() / 2));
7094  }
7095
7096  if (StVal.getValueType() != MVT::i64 ||
7097      StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7098    return SDValue();
7099
7100  SelectionDAG &DAG = DCI.DAG;
7101  DebugLoc dl = StVal.getDebugLoc();
7102  SDValue IntVec = StVal.getOperand(0);
7103  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7104                                 IntVec.getValueType().getVectorNumElements());
7105  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7106  SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7107                               Vec, StVal.getOperand(1));
7108  dl = N->getDebugLoc();
7109  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7110  // Make the DAGCombiner fold the bitcasts.
7111  DCI.AddToWorklist(Vec.getNode());
7112  DCI.AddToWorklist(ExtElt.getNode());
7113  DCI.AddToWorklist(V.getNode());
7114  return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7115                      St->getPointerInfo(), St->isVolatile(),
7116                      St->isNonTemporal(), St->getAlignment(),
7117                      St->getTBAAInfo());
7118}
7119
7120/// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7121/// are normal, non-volatile loads.  If so, it is profitable to bitcast an
7122/// i64 vector to have f64 elements, since the value can then be loaded
7123/// directly into a VFP register.
7124static bool hasNormalLoadOperand(SDNode *N) {
7125  unsigned NumElts = N->getValueType(0).getVectorNumElements();
7126  for (unsigned i = 0; i < NumElts; ++i) {
7127    SDNode *Elt = N->getOperand(i).getNode();
7128    if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7129      return true;
7130  }
7131  return false;
7132}
7133
7134/// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7135/// ISD::BUILD_VECTOR.
7136static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7137                                          TargetLowering::DAGCombinerInfo &DCI){
7138  // build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7139  // VMOVRRD is introduced when legalizing i64 types.  It forces the i64 value
7140  // into a pair of GPRs, which is fine when the value is used as a scalar,
7141  // but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7142  SelectionDAG &DAG = DCI.DAG;
7143  if (N->getNumOperands() == 2) {
7144    SDValue RV = PerformVMOVDRRCombine(N, DAG);
7145    if (RV.getNode())
7146      return RV;
7147  }
7148
7149  // Load i64 elements as f64 values so that type legalization does not split
7150  // them up into i32 values.
7151  EVT VT = N->getValueType(0);
7152  if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7153    return SDValue();
7154  DebugLoc dl = N->getDebugLoc();
7155  SmallVector<SDValue, 8> Ops;
7156  unsigned NumElts = VT.getVectorNumElements();
7157  for (unsigned i = 0; i < NumElts; ++i) {
7158    SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7159    Ops.push_back(V);
7160    // Make the DAGCombiner fold the bitcast.
7161    DCI.AddToWorklist(V.getNode());
7162  }
7163  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7164  SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7165  return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7166}
7167
7168/// PerformInsertEltCombine - Target-specific dag combine xforms for
7169/// ISD::INSERT_VECTOR_ELT.
7170static SDValue PerformInsertEltCombine(SDNode *N,
7171                                       TargetLowering::DAGCombinerInfo &DCI) {
7172  // Bitcast an i64 load inserted into a vector to f64.
7173  // Otherwise, the i64 value will be legalized to a pair of i32 values.
7174  EVT VT = N->getValueType(0);
7175  SDNode *Elt = N->getOperand(1).getNode();
7176  if (VT.getVectorElementType() != MVT::i64 ||
7177      !ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7178    return SDValue();
7179
7180  SelectionDAG &DAG = DCI.DAG;
7181  DebugLoc dl = N->getDebugLoc();
7182  EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7183                                 VT.getVectorNumElements());
7184  SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7185  SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7186  // Make the DAGCombiner fold the bitcasts.
7187  DCI.AddToWorklist(Vec.getNode());
7188  DCI.AddToWorklist(V.getNode());
7189  SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7190                               Vec, V, N->getOperand(2));
7191  return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7192}
7193
7194/// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7195/// ISD::VECTOR_SHUFFLE.
7196static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7197  // The LLVM shufflevector instruction does not require the shuffle mask
7198  // length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7199  // have that requirement.  When translating to ISD::VECTOR_SHUFFLE, if the
7200  // operands do not match the mask length, they are extended by concatenating
7201  // them with undef vectors.  That is probably the right thing for other
7202  // targets, but for NEON it is better to concatenate two double-register
7203  // size vector operands into a single quad-register size vector.  Do that
7204  // transformation here:
7205  //   shuffle(concat(v1, undef), concat(v2, undef)) ->
7206  //   shuffle(concat(v1, v2), undef)
7207  SDValue Op0 = N->getOperand(0);
7208  SDValue Op1 = N->getOperand(1);
7209  if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7210      Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7211      Op0.getNumOperands() != 2 ||
7212      Op1.getNumOperands() != 2)
7213    return SDValue();
7214  SDValue Concat0Op1 = Op0.getOperand(1);
7215  SDValue Concat1Op1 = Op1.getOperand(1);
7216  if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7217      Concat1Op1.getOpcode() != ISD::UNDEF)
7218    return SDValue();
7219  // Skip the transformation if any of the types are illegal.
7220  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7221  EVT VT = N->getValueType(0);
7222  if (!TLI.isTypeLegal(VT) ||
7223      !TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7224      !TLI.isTypeLegal(Concat1Op1.getValueType()))
7225    return SDValue();
7226
7227  SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7228                                  Op0.getOperand(0), Op1.getOperand(0));
7229  // Translate the shuffle mask.
7230  SmallVector<int, 16> NewMask;
7231  unsigned NumElts = VT.getVectorNumElements();
7232  unsigned HalfElts = NumElts/2;
7233  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7234  for (unsigned n = 0; n < NumElts; ++n) {
7235    int MaskElt = SVN->getMaskElt(n);
7236    int NewElt = -1;
7237    if (MaskElt < (int)HalfElts)
7238      NewElt = MaskElt;
7239    else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7240      NewElt = HalfElts + MaskElt - NumElts;
7241    NewMask.push_back(NewElt);
7242  }
7243  return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7244                              DAG.getUNDEF(VT), NewMask.data());
7245}
7246
7247/// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7248/// NEON load/store intrinsics to merge base address updates.
7249static SDValue CombineBaseUpdate(SDNode *N,
7250                                 TargetLowering::DAGCombinerInfo &DCI) {
7251  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7252    return SDValue();
7253
7254  SelectionDAG &DAG = DCI.DAG;
7255  bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7256                      N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7257  unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7258  SDValue Addr = N->getOperand(AddrOpIdx);
7259
7260  // Search for a use of the address operand that is an increment.
7261  for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7262         UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7263    SDNode *User = *UI;
7264    if (User->getOpcode() != ISD::ADD ||
7265        UI.getUse().getResNo() != Addr.getResNo())
7266      continue;
7267
7268    // Check that the add is independent of the load/store.  Otherwise, folding
7269    // it would create a cycle.
7270    if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7271      continue;
7272
7273    // Find the new opcode for the updating load/store.
7274    bool isLoad = true;
7275    bool isLaneOp = false;
7276    unsigned NewOpc = 0;
7277    unsigned NumVecs = 0;
7278    if (isIntrinsic) {
7279      unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7280      switch (IntNo) {
7281      default: assert(0 && "unexpected intrinsic for Neon base update");
7282      case Intrinsic::arm_neon_vld1:     NewOpc = ARMISD::VLD1_UPD;
7283        NumVecs = 1; break;
7284      case Intrinsic::arm_neon_vld2:     NewOpc = ARMISD::VLD2_UPD;
7285        NumVecs = 2; break;
7286      case Intrinsic::arm_neon_vld3:     NewOpc = ARMISD::VLD3_UPD;
7287        NumVecs = 3; break;
7288      case Intrinsic::arm_neon_vld4:     NewOpc = ARMISD::VLD4_UPD;
7289        NumVecs = 4; break;
7290      case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7291        NumVecs = 2; isLaneOp = true; break;
7292      case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7293        NumVecs = 3; isLaneOp = true; break;
7294      case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7295        NumVecs = 4; isLaneOp = true; break;
7296      case Intrinsic::arm_neon_vst1:     NewOpc = ARMISD::VST1_UPD;
7297        NumVecs = 1; isLoad = false; break;
7298      case Intrinsic::arm_neon_vst2:     NewOpc = ARMISD::VST2_UPD;
7299        NumVecs = 2; isLoad = false; break;
7300      case Intrinsic::arm_neon_vst3:     NewOpc = ARMISD::VST3_UPD;
7301        NumVecs = 3; isLoad = false; break;
7302      case Intrinsic::arm_neon_vst4:     NewOpc = ARMISD::VST4_UPD;
7303        NumVecs = 4; isLoad = false; break;
7304      case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7305        NumVecs = 2; isLoad = false; isLaneOp = true; break;
7306      case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7307        NumVecs = 3; isLoad = false; isLaneOp = true; break;
7308      case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7309        NumVecs = 4; isLoad = false; isLaneOp = true; break;
7310      }
7311    } else {
7312      isLaneOp = true;
7313      switch (N->getOpcode()) {
7314      default: assert(0 && "unexpected opcode for Neon base update");
7315      case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7316      case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7317      case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7318      }
7319    }
7320
7321    // Find the size of memory referenced by the load/store.
7322    EVT VecTy;
7323    if (isLoad)
7324      VecTy = N->getValueType(0);
7325    else
7326      VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7327    unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7328    if (isLaneOp)
7329      NumBytes /= VecTy.getVectorNumElements();
7330
7331    // If the increment is a constant, it must match the memory ref size.
7332    SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7333    if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7334      uint64_t IncVal = CInc->getZExtValue();
7335      if (IncVal != NumBytes)
7336        continue;
7337    } else if (NumBytes >= 3 * 16) {
7338      // VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7339      // separate instructions that make it harder to use a non-constant update.
7340      continue;
7341    }
7342
7343    // Create the new updating load/store node.
7344    EVT Tys[6];
7345    unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7346    unsigned n;
7347    for (n = 0; n < NumResultVecs; ++n)
7348      Tys[n] = VecTy;
7349    Tys[n++] = MVT::i32;
7350    Tys[n] = MVT::Other;
7351    SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7352    SmallVector<SDValue, 8> Ops;
7353    Ops.push_back(N->getOperand(0)); // incoming chain
7354    Ops.push_back(N->getOperand(AddrOpIdx));
7355    Ops.push_back(Inc);
7356    for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7357      Ops.push_back(N->getOperand(i));
7358    }
7359    MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7360    SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7361                                           Ops.data(), Ops.size(),
7362                                           MemInt->getMemoryVT(),
7363                                           MemInt->getMemOperand());
7364
7365    // Update the uses.
7366    std::vector<SDValue> NewResults;
7367    for (unsigned i = 0; i < NumResultVecs; ++i) {
7368      NewResults.push_back(SDValue(UpdN.getNode(), i));
7369    }
7370    NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7371    DCI.CombineTo(N, NewResults);
7372    DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7373
7374    break;
7375  }
7376  return SDValue();
7377}
7378
7379/// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7380/// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7381/// are also VDUPLANEs.  If so, combine them to a vldN-dup operation and
7382/// return true.
7383static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7384  SelectionDAG &DAG = DCI.DAG;
7385  EVT VT = N->getValueType(0);
7386  // vldN-dup instructions only support 64-bit vectors for N > 1.
7387  if (!VT.is64BitVector())
7388    return false;
7389
7390  // Check if the VDUPLANE operand is a vldN-dup intrinsic.
7391  SDNode *VLD = N->getOperand(0).getNode();
7392  if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7393    return false;
7394  unsigned NumVecs = 0;
7395  unsigned NewOpc = 0;
7396  unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7397  if (IntNo == Intrinsic::arm_neon_vld2lane) {
7398    NumVecs = 2;
7399    NewOpc = ARMISD::VLD2DUP;
7400  } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7401    NumVecs = 3;
7402    NewOpc = ARMISD::VLD3DUP;
7403  } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7404    NumVecs = 4;
7405    NewOpc = ARMISD::VLD4DUP;
7406  } else {
7407    return false;
7408  }
7409
7410  // First check that all the vldN-lane uses are VDUPLANEs and that the lane
7411  // numbers match the load.
7412  unsigned VLDLaneNo =
7413    cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7414  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7415       UI != UE; ++UI) {
7416    // Ignore uses of the chain result.
7417    if (UI.getUse().getResNo() == NumVecs)
7418      continue;
7419    SDNode *User = *UI;
7420    if (User->getOpcode() != ARMISD::VDUPLANE ||
7421        VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7422      return false;
7423  }
7424
7425  // Create the vldN-dup node.
7426  EVT Tys[5];
7427  unsigned n;
7428  for (n = 0; n < NumVecs; ++n)
7429    Tys[n] = VT;
7430  Tys[n] = MVT::Other;
7431  SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7432  SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7433  MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7434  SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7435                                           Ops, 2, VLDMemInt->getMemoryVT(),
7436                                           VLDMemInt->getMemOperand());
7437
7438  // Update the uses.
7439  for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7440       UI != UE; ++UI) {
7441    unsigned ResNo = UI.getUse().getResNo();
7442    // Ignore uses of the chain result.
7443    if (ResNo == NumVecs)
7444      continue;
7445    SDNode *User = *UI;
7446    DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7447  }
7448
7449  // Now the vldN-lane intrinsic is dead except for its chain result.
7450  // Update uses of the chain.
7451  std::vector<SDValue> VLDDupResults;
7452  for (unsigned n = 0; n < NumVecs; ++n)
7453    VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7454  VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7455  DCI.CombineTo(VLD, VLDDupResults);
7456
7457  return true;
7458}
7459
7460/// PerformVDUPLANECombine - Target-specific dag combine xforms for
7461/// ARMISD::VDUPLANE.
7462static SDValue PerformVDUPLANECombine(SDNode *N,
7463                                      TargetLowering::DAGCombinerInfo &DCI) {
7464  SDValue Op = N->getOperand(0);
7465
7466  // If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7467  // of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7468  if (CombineVLDDUP(N, DCI))
7469    return SDValue(N, 0);
7470
7471  // If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7472  // redundant.  Ignore bit_converts for now; element sizes are checked below.
7473  while (Op.getOpcode() == ISD::BITCAST)
7474    Op = Op.getOperand(0);
7475  if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7476    return SDValue();
7477
7478  // Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7479  unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7480  // The canonical VMOV for a zero vector uses a 32-bit element size.
7481  unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7482  unsigned EltBits;
7483  if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7484    EltSize = 8;
7485  EVT VT = N->getValueType(0);
7486  if (EltSize > VT.getVectorElementType().getSizeInBits())
7487    return SDValue();
7488
7489  return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7490}
7491
7492// isConstVecPow2 - Return true if each vector element is a power of 2, all
7493// elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7494static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7495{
7496  integerPart cN;
7497  integerPart c0 = 0;
7498  for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7499       I != E; I++) {
7500    ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7501    if (!C)
7502      return false;
7503
7504    bool isExact;
7505    APFloat APF = C->getValueAPF();
7506    if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7507        != APFloat::opOK || !isExact)
7508      return false;
7509
7510    c0 = (I == 0) ? cN : c0;
7511    if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7512      return false;
7513  }
7514  C = c0;
7515  return true;
7516}
7517
7518/// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7519/// can replace combinations of VMUL and VCVT (floating-point to integer)
7520/// when the VMUL has a constant operand that is a power of 2.
7521///
7522/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7523///  vmul.f32        d16, d17, d16
7524///  vcvt.s32.f32    d16, d16
7525/// becomes:
7526///  vcvt.s32.f32    d16, d16, #3
7527static SDValue PerformVCVTCombine(SDNode *N,
7528                                  TargetLowering::DAGCombinerInfo &DCI,
7529                                  const ARMSubtarget *Subtarget) {
7530  SelectionDAG &DAG = DCI.DAG;
7531  SDValue Op = N->getOperand(0);
7532
7533  if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7534      Op.getOpcode() != ISD::FMUL)
7535    return SDValue();
7536
7537  uint64_t C;
7538  SDValue N0 = Op->getOperand(0);
7539  SDValue ConstVec = Op->getOperand(1);
7540  bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7541
7542  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7543      !isConstVecPow2(ConstVec, isSigned, C))
7544    return SDValue();
7545
7546  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7547    Intrinsic::arm_neon_vcvtfp2fxu;
7548  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7549                     N->getValueType(0),
7550                     DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7551                     DAG.getConstant(Log2_64(C), MVT::i32));
7552}
7553
7554/// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7555/// can replace combinations of VCVT (integer to floating-point) and VDIV
7556/// when the VDIV has a constant operand that is a power of 2.
7557///
7558/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7559///  vcvt.f32.s32    d16, d16
7560///  vdiv.f32        d16, d17, d16
7561/// becomes:
7562///  vcvt.f32.s32    d16, d16, #3
7563static SDValue PerformVDIVCombine(SDNode *N,
7564                                  TargetLowering::DAGCombinerInfo &DCI,
7565                                  const ARMSubtarget *Subtarget) {
7566  SelectionDAG &DAG = DCI.DAG;
7567  SDValue Op = N->getOperand(0);
7568  unsigned OpOpcode = Op.getNode()->getOpcode();
7569
7570  if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7571      (OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7572    return SDValue();
7573
7574  uint64_t C;
7575  SDValue ConstVec = N->getOperand(1);
7576  bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7577
7578  if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7579      !isConstVecPow2(ConstVec, isSigned, C))
7580    return SDValue();
7581
7582  unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7583    Intrinsic::arm_neon_vcvtfxu2fp;
7584  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7585                     Op.getValueType(),
7586                     DAG.getConstant(IntrinsicOpcode, MVT::i32),
7587                     Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7588}
7589
7590/// Getvshiftimm - Check if this is a valid build_vector for the immediate
7591/// operand of a vector shift operation, where all the elements of the
7592/// build_vector must have the same constant integer value.
7593static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7594  // Ignore bit_converts.
7595  while (Op.getOpcode() == ISD::BITCAST)
7596    Op = Op.getOperand(0);
7597  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7598  APInt SplatBits, SplatUndef;
7599  unsigned SplatBitSize;
7600  bool HasAnyUndefs;
7601  if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7602                                      HasAnyUndefs, ElementBits) ||
7603      SplatBitSize > ElementBits)
7604    return false;
7605  Cnt = SplatBits.getSExtValue();
7606  return true;
7607}
7608
7609/// isVShiftLImm - Check if this is a valid build_vector for the immediate
7610/// operand of a vector shift left operation.  That value must be in the range:
7611///   0 <= Value < ElementBits for a left shift; or
7612///   0 <= Value <= ElementBits for a long left shift.
7613static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7614  assert(VT.isVector() && "vector shift count is not a vector type");
7615  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7616  if (! getVShiftImm(Op, ElementBits, Cnt))
7617    return false;
7618  return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7619}
7620
7621/// isVShiftRImm - Check if this is a valid build_vector for the immediate
7622/// operand of a vector shift right operation.  For a shift opcode, the value
7623/// is positive, but for an intrinsic the value count must be negative. The
7624/// absolute value must be in the range:
7625///   1 <= |Value| <= ElementBits for a right shift; or
7626///   1 <= |Value| <= ElementBits/2 for a narrow right shift.
7627static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7628                         int64_t &Cnt) {
7629  assert(VT.isVector() && "vector shift count is not a vector type");
7630  unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7631  if (! getVShiftImm(Op, ElementBits, Cnt))
7632    return false;
7633  if (isIntrinsic)
7634    Cnt = -Cnt;
7635  return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7636}
7637
7638/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7639static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7640  unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7641  switch (IntNo) {
7642  default:
7643    // Don't do anything for most intrinsics.
7644    break;
7645
7646  // Vector shifts: check for immediate versions and lower them.
7647  // Note: This is done during DAG combining instead of DAG legalizing because
7648  // the build_vectors for 64-bit vector element shift counts are generally
7649  // not legal, and it is hard to see their values after they get legalized to
7650  // loads from a constant pool.
7651  case Intrinsic::arm_neon_vshifts:
7652  case Intrinsic::arm_neon_vshiftu:
7653  case Intrinsic::arm_neon_vshiftls:
7654  case Intrinsic::arm_neon_vshiftlu:
7655  case Intrinsic::arm_neon_vshiftn:
7656  case Intrinsic::arm_neon_vrshifts:
7657  case Intrinsic::arm_neon_vrshiftu:
7658  case Intrinsic::arm_neon_vrshiftn:
7659  case Intrinsic::arm_neon_vqshifts:
7660  case Intrinsic::arm_neon_vqshiftu:
7661  case Intrinsic::arm_neon_vqshiftsu:
7662  case Intrinsic::arm_neon_vqshiftns:
7663  case Intrinsic::arm_neon_vqshiftnu:
7664  case Intrinsic::arm_neon_vqshiftnsu:
7665  case Intrinsic::arm_neon_vqrshiftns:
7666  case Intrinsic::arm_neon_vqrshiftnu:
7667  case Intrinsic::arm_neon_vqrshiftnsu: {
7668    EVT VT = N->getOperand(1).getValueType();
7669    int64_t Cnt;
7670    unsigned VShiftOpc = 0;
7671
7672    switch (IntNo) {
7673    case Intrinsic::arm_neon_vshifts:
7674    case Intrinsic::arm_neon_vshiftu:
7675      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7676        VShiftOpc = ARMISD::VSHL;
7677        break;
7678      }
7679      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7680        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7681                     ARMISD::VSHRs : ARMISD::VSHRu);
7682        break;
7683      }
7684      return SDValue();
7685
7686    case Intrinsic::arm_neon_vshiftls:
7687    case Intrinsic::arm_neon_vshiftlu:
7688      if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7689        break;
7690      llvm_unreachable("invalid shift count for vshll intrinsic");
7691
7692    case Intrinsic::arm_neon_vrshifts:
7693    case Intrinsic::arm_neon_vrshiftu:
7694      if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7695        break;
7696      return SDValue();
7697
7698    case Intrinsic::arm_neon_vqshifts:
7699    case Intrinsic::arm_neon_vqshiftu:
7700      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7701        break;
7702      return SDValue();
7703
7704    case Intrinsic::arm_neon_vqshiftsu:
7705      if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7706        break;
7707      llvm_unreachable("invalid shift count for vqshlu intrinsic");
7708
7709    case Intrinsic::arm_neon_vshiftn:
7710    case Intrinsic::arm_neon_vrshiftn:
7711    case Intrinsic::arm_neon_vqshiftns:
7712    case Intrinsic::arm_neon_vqshiftnu:
7713    case Intrinsic::arm_neon_vqshiftnsu:
7714    case Intrinsic::arm_neon_vqrshiftns:
7715    case Intrinsic::arm_neon_vqrshiftnu:
7716    case Intrinsic::arm_neon_vqrshiftnsu:
7717      // Narrowing shifts require an immediate right shift.
7718      if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7719        break;
7720      llvm_unreachable("invalid shift count for narrowing vector shift "
7721                       "intrinsic");
7722
7723    default:
7724      llvm_unreachable("unhandled vector shift");
7725    }
7726
7727    switch (IntNo) {
7728    case Intrinsic::arm_neon_vshifts:
7729    case Intrinsic::arm_neon_vshiftu:
7730      // Opcode already set above.
7731      break;
7732    case Intrinsic::arm_neon_vshiftls:
7733    case Intrinsic::arm_neon_vshiftlu:
7734      if (Cnt == VT.getVectorElementType().getSizeInBits())
7735        VShiftOpc = ARMISD::VSHLLi;
7736      else
7737        VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7738                     ARMISD::VSHLLs : ARMISD::VSHLLu);
7739      break;
7740    case Intrinsic::arm_neon_vshiftn:
7741      VShiftOpc = ARMISD::VSHRN; break;
7742    case Intrinsic::arm_neon_vrshifts:
7743      VShiftOpc = ARMISD::VRSHRs; break;
7744    case Intrinsic::arm_neon_vrshiftu:
7745      VShiftOpc = ARMISD::VRSHRu; break;
7746    case Intrinsic::arm_neon_vrshiftn:
7747      VShiftOpc = ARMISD::VRSHRN; break;
7748    case Intrinsic::arm_neon_vqshifts:
7749      VShiftOpc = ARMISD::VQSHLs; break;
7750    case Intrinsic::arm_neon_vqshiftu:
7751      VShiftOpc = ARMISD::VQSHLu; break;
7752    case Intrinsic::arm_neon_vqshiftsu:
7753      VShiftOpc = ARMISD::VQSHLsu; break;
7754    case Intrinsic::arm_neon_vqshiftns:
7755      VShiftOpc = ARMISD::VQSHRNs; break;
7756    case Intrinsic::arm_neon_vqshiftnu:
7757      VShiftOpc = ARMISD::VQSHRNu; break;
7758    case Intrinsic::arm_neon_vqshiftnsu:
7759      VShiftOpc = ARMISD::VQSHRNsu; break;
7760    case Intrinsic::arm_neon_vqrshiftns:
7761      VShiftOpc = ARMISD::VQRSHRNs; break;
7762    case Intrinsic::arm_neon_vqrshiftnu:
7763      VShiftOpc = ARMISD::VQRSHRNu; break;
7764    case Intrinsic::arm_neon_vqrshiftnsu:
7765      VShiftOpc = ARMISD::VQRSHRNsu; break;
7766    }
7767
7768    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7769                       N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
7770  }
7771
7772  case Intrinsic::arm_neon_vshiftins: {
7773    EVT VT = N->getOperand(1).getValueType();
7774    int64_t Cnt;
7775    unsigned VShiftOpc = 0;
7776
7777    if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
7778      VShiftOpc = ARMISD::VSLI;
7779    else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
7780      VShiftOpc = ARMISD::VSRI;
7781    else {
7782      llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
7783    }
7784
7785    return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
7786                       N->getOperand(1), N->getOperand(2),
7787                       DAG.getConstant(Cnt, MVT::i32));
7788  }
7789
7790  case Intrinsic::arm_neon_vqrshifts:
7791  case Intrinsic::arm_neon_vqrshiftu:
7792    // No immediate versions of these to check for.
7793    break;
7794  }
7795
7796  return SDValue();
7797}
7798
7799/// PerformShiftCombine - Checks for immediate versions of vector shifts and
7800/// lowers them.  As with the vector shift intrinsics, this is done during DAG
7801/// combining instead of DAG legalizing because the build_vectors for 64-bit
7802/// vector element shift counts are generally not legal, and it is hard to see
7803/// their values after they get legalized to loads from a constant pool.
7804static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
7805                                   const ARMSubtarget *ST) {
7806  EVT VT = N->getValueType(0);
7807
7808  // Nothing to be done for scalar shifts.
7809  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7810  if (!VT.isVector() || !TLI.isTypeLegal(VT))
7811    return SDValue();
7812
7813  assert(ST->hasNEON() && "unexpected vector shift");
7814  int64_t Cnt;
7815
7816  switch (N->getOpcode()) {
7817  default: llvm_unreachable("unexpected shift opcode");
7818
7819  case ISD::SHL:
7820    if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
7821      return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
7822                         DAG.getConstant(Cnt, MVT::i32));
7823    break;
7824
7825  case ISD::SRA:
7826  case ISD::SRL:
7827    if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
7828      unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
7829                            ARMISD::VSHRs : ARMISD::VSHRu);
7830      return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
7831                         DAG.getConstant(Cnt, MVT::i32));
7832    }
7833  }
7834  return SDValue();
7835}
7836
7837/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
7838/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
7839static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
7840                                    const ARMSubtarget *ST) {
7841  SDValue N0 = N->getOperand(0);
7842
7843  // Check for sign- and zero-extensions of vector extract operations of 8-
7844  // and 16-bit vector elements.  NEON supports these directly.  They are
7845  // handled during DAG combining because type legalization will promote them
7846  // to 32-bit types and it is messy to recognize the operations after that.
7847  if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
7848    SDValue Vec = N0.getOperand(0);
7849    SDValue Lane = N0.getOperand(1);
7850    EVT VT = N->getValueType(0);
7851    EVT EltVT = N0.getValueType();
7852    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7853
7854    if (VT == MVT::i32 &&
7855        (EltVT == MVT::i8 || EltVT == MVT::i16) &&
7856        TLI.isTypeLegal(Vec.getValueType()) &&
7857        isa<ConstantSDNode>(Lane)) {
7858
7859      unsigned Opc = 0;
7860      switch (N->getOpcode()) {
7861      default: llvm_unreachable("unexpected opcode");
7862      case ISD::SIGN_EXTEND:
7863        Opc = ARMISD::VGETLANEs;
7864        break;
7865      case ISD::ZERO_EXTEND:
7866      case ISD::ANY_EXTEND:
7867        Opc = ARMISD::VGETLANEu;
7868        break;
7869      }
7870      return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
7871    }
7872  }
7873
7874  return SDValue();
7875}
7876
7877/// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
7878/// to match f32 max/min patterns to use NEON vmax/vmin instructions.
7879static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
7880                                       const ARMSubtarget *ST) {
7881  // If the target supports NEON, try to use vmax/vmin instructions for f32
7882  // selects like "x < y ? x : y".  Unless the NoNaNsFPMath option is set,
7883  // be careful about NaNs:  NEON's vmax/vmin return NaN if either operand is
7884  // a NaN; only do the transformation when it matches that behavior.
7885
7886  // For now only do this when using NEON for FP operations; if using VFP, it
7887  // is not obvious that the benefit outweighs the cost of switching to the
7888  // NEON pipeline.
7889  if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
7890      N->getValueType(0) != MVT::f32)
7891    return SDValue();
7892
7893  SDValue CondLHS = N->getOperand(0);
7894  SDValue CondRHS = N->getOperand(1);
7895  SDValue LHS = N->getOperand(2);
7896  SDValue RHS = N->getOperand(3);
7897  ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
7898
7899  unsigned Opcode = 0;
7900  bool IsReversed;
7901  if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
7902    IsReversed = false; // x CC y ? x : y
7903  } else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
7904    IsReversed = true ; // x CC y ? y : x
7905  } else {
7906    return SDValue();
7907  }
7908
7909  bool IsUnordered;
7910  switch (CC) {
7911  default: break;
7912  case ISD::SETOLT:
7913  case ISD::SETOLE:
7914  case ISD::SETLT:
7915  case ISD::SETLE:
7916  case ISD::SETULT:
7917  case ISD::SETULE:
7918    // If LHS is NaN, an ordered comparison will be false and the result will
7919    // be the RHS, but vmin(NaN, RHS) = NaN.  Avoid this by checking that LHS
7920    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
7921    IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
7922    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7923      break;
7924    // For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
7925    // will return -0, so vmin can only be used for unsafe math or if one of
7926    // the operands is known to be nonzero.
7927    if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
7928        !UnsafeFPMath &&
7929        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7930      break;
7931    Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
7932    break;
7933
7934  case ISD::SETOGT:
7935  case ISD::SETOGE:
7936  case ISD::SETGT:
7937  case ISD::SETGE:
7938  case ISD::SETUGT:
7939  case ISD::SETUGE:
7940    // If LHS is NaN, an ordered comparison will be false and the result will
7941    // be the RHS, but vmax(NaN, RHS) = NaN.  Avoid this by checking that LHS
7942    // != NaN.  Likewise, for unordered comparisons, check for RHS != NaN.
7943    IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
7944    if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
7945      break;
7946    // For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
7947    // will return +0, so vmax can only be used for unsafe math or if one of
7948    // the operands is known to be nonzero.
7949    if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
7950        !UnsafeFPMath &&
7951        !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
7952      break;
7953    Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
7954    break;
7955  }
7956
7957  if (!Opcode)
7958    return SDValue();
7959  return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
7960}
7961
7962/// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
7963SDValue
7964ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
7965  SDValue Cmp = N->getOperand(4);
7966  if (Cmp.getOpcode() != ARMISD::CMPZ)
7967    // Only looking at EQ and NE cases.
7968    return SDValue();
7969
7970  EVT VT = N->getValueType(0);
7971  DebugLoc dl = N->getDebugLoc();
7972  SDValue LHS = Cmp.getOperand(0);
7973  SDValue RHS = Cmp.getOperand(1);
7974  SDValue FalseVal = N->getOperand(0);
7975  SDValue TrueVal = N->getOperand(1);
7976  SDValue ARMcc = N->getOperand(2);
7977  ARMCC::CondCodes CC =
7978    (ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
7979
7980  // Simplify
7981  //   mov     r1, r0
7982  //   cmp     r1, x
7983  //   mov     r0, y
7984  //   moveq   r0, x
7985  // to
7986  //   cmp     r0, x
7987  //   movne   r0, y
7988  //
7989  //   mov     r1, r0
7990  //   cmp     r1, x
7991  //   mov     r0, x
7992  //   movne   r0, y
7993  // to
7994  //   cmp     r0, x
7995  //   movne   r0, y
7996  /// FIXME: Turn this into a target neutral optimization?
7997  SDValue Res;
7998  if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
7999    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8000                      N->getOperand(3), Cmp);
8001  } else if (CC == ARMCC::EQ && TrueVal == RHS) {
8002    SDValue ARMcc;
8003    SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8004    Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8005                      N->getOperand(3), NewCmp);
8006  }
8007
8008  if (Res.getNode()) {
8009    APInt KnownZero, KnownOne;
8010    APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8011    DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8012    // Capture demanded bits information that would be otherwise lost.
8013    if (KnownZero == 0xfffffffe)
8014      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8015                        DAG.getValueType(MVT::i1));
8016    else if (KnownZero == 0xffffff00)
8017      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8018                        DAG.getValueType(MVT::i8));
8019    else if (KnownZero == 0xffff0000)
8020      Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8021                        DAG.getValueType(MVT::i16));
8022  }
8023
8024  return Res;
8025}
8026
8027SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8028                                             DAGCombinerInfo &DCI) const {
8029  switch (N->getOpcode()) {
8030  default: break;
8031  case ISD::ADD:        return PerformADDCombine(N, DCI, Subtarget);
8032  case ISD::SUB:        return PerformSUBCombine(N, DCI);
8033  case ISD::MUL:        return PerformMULCombine(N, DCI, Subtarget);
8034  case ISD::OR:         return PerformORCombine(N, DCI, Subtarget);
8035  case ISD::AND:        return PerformANDCombine(N, DCI);
8036  case ARMISD::BFI:     return PerformBFICombine(N, DCI);
8037  case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8038  case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8039  case ISD::STORE:      return PerformSTORECombine(N, DCI);
8040  case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8041  case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8042  case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8043  case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8044  case ISD::FP_TO_SINT:
8045  case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8046  case ISD::FDIV:       return PerformVDIVCombine(N, DCI, Subtarget);
8047  case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8048  case ISD::SHL:
8049  case ISD::SRA:
8050  case ISD::SRL:        return PerformShiftCombine(N, DCI.DAG, Subtarget);
8051  case ISD::SIGN_EXTEND:
8052  case ISD::ZERO_EXTEND:
8053  case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8054  case ISD::SELECT_CC:  return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8055  case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8056  case ARMISD::VLD2DUP:
8057  case ARMISD::VLD3DUP:
8058  case ARMISD::VLD4DUP:
8059    return CombineBaseUpdate(N, DCI);
8060  case ISD::INTRINSIC_VOID:
8061  case ISD::INTRINSIC_W_CHAIN:
8062    switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8063    case Intrinsic::arm_neon_vld1:
8064    case Intrinsic::arm_neon_vld2:
8065    case Intrinsic::arm_neon_vld3:
8066    case Intrinsic::arm_neon_vld4:
8067    case Intrinsic::arm_neon_vld2lane:
8068    case Intrinsic::arm_neon_vld3lane:
8069    case Intrinsic::arm_neon_vld4lane:
8070    case Intrinsic::arm_neon_vst1:
8071    case Intrinsic::arm_neon_vst2:
8072    case Intrinsic::arm_neon_vst3:
8073    case Intrinsic::arm_neon_vst4:
8074    case Intrinsic::arm_neon_vst2lane:
8075    case Intrinsic::arm_neon_vst3lane:
8076    case Intrinsic::arm_neon_vst4lane:
8077      return CombineBaseUpdate(N, DCI);
8078    default: break;
8079    }
8080    break;
8081  }
8082  return SDValue();
8083}
8084
8085bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8086                                                          EVT VT) const {
8087  return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8088}
8089
8090bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8091  if (!Subtarget->allowsUnalignedMem())
8092    return false;
8093
8094  switch (VT.getSimpleVT().SimpleTy) {
8095  default:
8096    return false;
8097  case MVT::i8:
8098  case MVT::i16:
8099  case MVT::i32:
8100    return true;
8101  // FIXME: VLD1 etc with standard alignment is legal.
8102  }
8103}
8104
8105static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8106  if (V < 0)
8107    return false;
8108
8109  unsigned Scale = 1;
8110  switch (VT.getSimpleVT().SimpleTy) {
8111  default: return false;
8112  case MVT::i1:
8113  case MVT::i8:
8114    // Scale == 1;
8115    break;
8116  case MVT::i16:
8117    // Scale == 2;
8118    Scale = 2;
8119    break;
8120  case MVT::i32:
8121    // Scale == 4;
8122    Scale = 4;
8123    break;
8124  }
8125
8126  if ((V & (Scale - 1)) != 0)
8127    return false;
8128  V /= Scale;
8129  return V == (V & ((1LL << 5) - 1));
8130}
8131
8132static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8133                                      const ARMSubtarget *Subtarget) {
8134  bool isNeg = false;
8135  if (V < 0) {
8136    isNeg = true;
8137    V = - V;
8138  }
8139
8140  switch (VT.getSimpleVT().SimpleTy) {
8141  default: return false;
8142  case MVT::i1:
8143  case MVT::i8:
8144  case MVT::i16:
8145  case MVT::i32:
8146    // + imm12 or - imm8
8147    if (isNeg)
8148      return V == (V & ((1LL << 8) - 1));
8149    return V == (V & ((1LL << 12) - 1));
8150  case MVT::f32:
8151  case MVT::f64:
8152    // Same as ARM mode. FIXME: NEON?
8153    if (!Subtarget->hasVFP2())
8154      return false;
8155    if ((V & 3) != 0)
8156      return false;
8157    V >>= 2;
8158    return V == (V & ((1LL << 8) - 1));
8159  }
8160}
8161
8162/// isLegalAddressImmediate - Return true if the integer value can be used
8163/// as the offset of the target addressing mode for load / store of the
8164/// given type.
8165static bool isLegalAddressImmediate(int64_t V, EVT VT,
8166                                    const ARMSubtarget *Subtarget) {
8167  if (V == 0)
8168    return true;
8169
8170  if (!VT.isSimple())
8171    return false;
8172
8173  if (Subtarget->isThumb1Only())
8174    return isLegalT1AddressImmediate(V, VT);
8175  else if (Subtarget->isThumb2())
8176    return isLegalT2AddressImmediate(V, VT, Subtarget);
8177
8178  // ARM mode.
8179  if (V < 0)
8180    V = - V;
8181  switch (VT.getSimpleVT().SimpleTy) {
8182  default: return false;
8183  case MVT::i1:
8184  case MVT::i8:
8185  case MVT::i32:
8186    // +- imm12
8187    return V == (V & ((1LL << 12) - 1));
8188  case MVT::i16:
8189    // +- imm8
8190    return V == (V & ((1LL << 8) - 1));
8191  case MVT::f32:
8192  case MVT::f64:
8193    if (!Subtarget->hasVFP2()) // FIXME: NEON?
8194      return false;
8195    if ((V & 3) != 0)
8196      return false;
8197    V >>= 2;
8198    return V == (V & ((1LL << 8) - 1));
8199  }
8200}
8201
8202bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8203                                                      EVT VT) const {
8204  int Scale = AM.Scale;
8205  if (Scale < 0)
8206    return false;
8207
8208  switch (VT.getSimpleVT().SimpleTy) {
8209  default: return false;
8210  case MVT::i1:
8211  case MVT::i8:
8212  case MVT::i16:
8213  case MVT::i32:
8214    if (Scale == 1)
8215      return true;
8216    // r + r << imm
8217    Scale = Scale & ~1;
8218    return Scale == 2 || Scale == 4 || Scale == 8;
8219  case MVT::i64:
8220    // r + r
8221    if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8222      return true;
8223    return false;
8224  case MVT::isVoid:
8225    // Note, we allow "void" uses (basically, uses that aren't loads or
8226    // stores), because arm allows folding a scale into many arithmetic
8227    // operations.  This should be made more precise and revisited later.
8228
8229    // Allow r << imm, but the imm has to be a multiple of two.
8230    if (Scale & 1) return false;
8231    return isPowerOf2_32(Scale);
8232  }
8233}
8234
8235/// isLegalAddressingMode - Return true if the addressing mode represented
8236/// by AM is legal for this target, for a load/store of the specified type.
8237bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8238                                              Type *Ty) const {
8239  EVT VT = getValueType(Ty, true);
8240  if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8241    return false;
8242
8243  // Can never fold addr of global into load/store.
8244  if (AM.BaseGV)
8245    return false;
8246
8247  switch (AM.Scale) {
8248  case 0:  // no scale reg, must be "r+i" or "r", or "i".
8249    break;
8250  case 1:
8251    if (Subtarget->isThumb1Only())
8252      return false;
8253    // FALL THROUGH.
8254  default:
8255    // ARM doesn't support any R+R*scale+imm addr modes.
8256    if (AM.BaseOffs)
8257      return false;
8258
8259    if (!VT.isSimple())
8260      return false;
8261
8262    if (Subtarget->isThumb2())
8263      return isLegalT2ScaledAddressingMode(AM, VT);
8264
8265    int Scale = AM.Scale;
8266    switch (VT.getSimpleVT().SimpleTy) {
8267    default: return false;
8268    case MVT::i1:
8269    case MVT::i8:
8270    case MVT::i32:
8271      if (Scale < 0) Scale = -Scale;
8272      if (Scale == 1)
8273        return true;
8274      // r + r << imm
8275      return isPowerOf2_32(Scale & ~1);
8276    case MVT::i16:
8277    case MVT::i64:
8278      // r + r
8279      if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8280        return true;
8281      return false;
8282
8283    case MVT::isVoid:
8284      // Note, we allow "void" uses (basically, uses that aren't loads or
8285      // stores), because arm allows folding a scale into many arithmetic
8286      // operations.  This should be made more precise and revisited later.
8287
8288      // Allow r << imm, but the imm has to be a multiple of two.
8289      if (Scale & 1) return false;
8290      return isPowerOf2_32(Scale);
8291    }
8292    break;
8293  }
8294  return true;
8295}
8296
8297/// isLegalICmpImmediate - Return true if the specified immediate is legal
8298/// icmp immediate, that is the target has icmp instructions which can compare
8299/// a register against the immediate without having to materialize the
8300/// immediate into a register.
8301bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8302  if (!Subtarget->isThumb())
8303    return ARM_AM::getSOImmVal(Imm) != -1;
8304  if (Subtarget->isThumb2())
8305    return ARM_AM::getT2SOImmVal(Imm) != -1;
8306  return Imm >= 0 && Imm <= 255;
8307}
8308
8309/// isLegalAddImmediate - Return true if the specified immediate is legal
8310/// add immediate, that is the target has add instructions which can add
8311/// a register with the immediate without having to materialize the
8312/// immediate into a register.
8313bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8314  return ARM_AM::getSOImmVal(Imm) != -1;
8315}
8316
8317static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8318                                      bool isSEXTLoad, SDValue &Base,
8319                                      SDValue &Offset, bool &isInc,
8320                                      SelectionDAG &DAG) {
8321  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8322    return false;
8323
8324  if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8325    // AddressingMode 3
8326    Base = Ptr->getOperand(0);
8327    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8328      int RHSC = (int)RHS->getZExtValue();
8329      if (RHSC < 0 && RHSC > -256) {
8330        assert(Ptr->getOpcode() == ISD::ADD);
8331        isInc = false;
8332        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8333        return true;
8334      }
8335    }
8336    isInc = (Ptr->getOpcode() == ISD::ADD);
8337    Offset = Ptr->getOperand(1);
8338    return true;
8339  } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8340    // AddressingMode 2
8341    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8342      int RHSC = (int)RHS->getZExtValue();
8343      if (RHSC < 0 && RHSC > -0x1000) {
8344        assert(Ptr->getOpcode() == ISD::ADD);
8345        isInc = false;
8346        Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8347        Base = Ptr->getOperand(0);
8348        return true;
8349      }
8350    }
8351
8352    if (Ptr->getOpcode() == ISD::ADD) {
8353      isInc = true;
8354      ARM_AM::ShiftOpc ShOpcVal=
8355        ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8356      if (ShOpcVal != ARM_AM::no_shift) {
8357        Base = Ptr->getOperand(1);
8358        Offset = Ptr->getOperand(0);
8359      } else {
8360        Base = Ptr->getOperand(0);
8361        Offset = Ptr->getOperand(1);
8362      }
8363      return true;
8364    }
8365
8366    isInc = (Ptr->getOpcode() == ISD::ADD);
8367    Base = Ptr->getOperand(0);
8368    Offset = Ptr->getOperand(1);
8369    return true;
8370  }
8371
8372  // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8373  return false;
8374}
8375
8376static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8377                                     bool isSEXTLoad, SDValue &Base,
8378                                     SDValue &Offset, bool &isInc,
8379                                     SelectionDAG &DAG) {
8380  if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8381    return false;
8382
8383  Base = Ptr->getOperand(0);
8384  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8385    int RHSC = (int)RHS->getZExtValue();
8386    if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8387      assert(Ptr->getOpcode() == ISD::ADD);
8388      isInc = false;
8389      Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8390      return true;
8391    } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8392      isInc = Ptr->getOpcode() == ISD::ADD;
8393      Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8394      return true;
8395    }
8396  }
8397
8398  return false;
8399}
8400
8401/// getPreIndexedAddressParts - returns true by value, base pointer and
8402/// offset pointer and addressing mode by reference if the node's address
8403/// can be legally represented as pre-indexed load / store address.
8404bool
8405ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8406                                             SDValue &Offset,
8407                                             ISD::MemIndexedMode &AM,
8408                                             SelectionDAG &DAG) const {
8409  if (Subtarget->isThumb1Only())
8410    return false;
8411
8412  EVT VT;
8413  SDValue Ptr;
8414  bool isSEXTLoad = false;
8415  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8416    Ptr = LD->getBasePtr();
8417    VT  = LD->getMemoryVT();
8418    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8419  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8420    Ptr = ST->getBasePtr();
8421    VT  = ST->getMemoryVT();
8422  } else
8423    return false;
8424
8425  bool isInc;
8426  bool isLegal = false;
8427  if (Subtarget->isThumb2())
8428    isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8429                                       Offset, isInc, DAG);
8430  else
8431    isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8432                                        Offset, isInc, DAG);
8433  if (!isLegal)
8434    return false;
8435
8436  AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8437  return true;
8438}
8439
8440/// getPostIndexedAddressParts - returns true by value, base pointer and
8441/// offset pointer and addressing mode by reference if this node can be
8442/// combined with a load / store to form a post-indexed load / store.
8443bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8444                                                   SDValue &Base,
8445                                                   SDValue &Offset,
8446                                                   ISD::MemIndexedMode &AM,
8447                                                   SelectionDAG &DAG) const {
8448  if (Subtarget->isThumb1Only())
8449    return false;
8450
8451  EVT VT;
8452  SDValue Ptr;
8453  bool isSEXTLoad = false;
8454  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8455    VT  = LD->getMemoryVT();
8456    Ptr = LD->getBasePtr();
8457    isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8458  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8459    VT  = ST->getMemoryVT();
8460    Ptr = ST->getBasePtr();
8461  } else
8462    return false;
8463
8464  bool isInc;
8465  bool isLegal = false;
8466  if (Subtarget->isThumb2())
8467    isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8468                                       isInc, DAG);
8469  else
8470    isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8471                                        isInc, DAG);
8472  if (!isLegal)
8473    return false;
8474
8475  if (Ptr != Base) {
8476    // Swap base ptr and offset to catch more post-index load / store when
8477    // it's legal. In Thumb2 mode, offset must be an immediate.
8478    if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8479        !Subtarget->isThumb2())
8480      std::swap(Base, Offset);
8481
8482    // Post-indexed load / store update the base pointer.
8483    if (Ptr != Base)
8484      return false;
8485  }
8486
8487  AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8488  return true;
8489}
8490
8491void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8492                                                       const APInt &Mask,
8493                                                       APInt &KnownZero,
8494                                                       APInt &KnownOne,
8495                                                       const SelectionDAG &DAG,
8496                                                       unsigned Depth) const {
8497  KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8498  switch (Op.getOpcode()) {
8499  default: break;
8500  case ARMISD::CMOV: {
8501    // Bits are known zero/one if known on the LHS and RHS.
8502    DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8503    if (KnownZero == 0 && KnownOne == 0) return;
8504
8505    APInt KnownZeroRHS, KnownOneRHS;
8506    DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8507                          KnownZeroRHS, KnownOneRHS, Depth+1);
8508    KnownZero &= KnownZeroRHS;
8509    KnownOne  &= KnownOneRHS;
8510    return;
8511  }
8512  }
8513}
8514
8515//===----------------------------------------------------------------------===//
8516//                           ARM Inline Assembly Support
8517//===----------------------------------------------------------------------===//
8518
8519bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8520  // Looking for "rev" which is V6+.
8521  if (!Subtarget->hasV6Ops())
8522    return false;
8523
8524  InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8525  std::string AsmStr = IA->getAsmString();
8526  SmallVector<StringRef, 4> AsmPieces;
8527  SplitString(AsmStr, AsmPieces, ";\n");
8528
8529  switch (AsmPieces.size()) {
8530  default: return false;
8531  case 1:
8532    AsmStr = AsmPieces[0];
8533    AsmPieces.clear();
8534    SplitString(AsmStr, AsmPieces, " \t,");
8535
8536    // rev $0, $1
8537    if (AsmPieces.size() == 3 &&
8538        AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8539        IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8540      IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8541      if (Ty && Ty->getBitWidth() == 32)
8542        return IntrinsicLowering::LowerToByteSwap(CI);
8543    }
8544    break;
8545  }
8546
8547  return false;
8548}
8549
8550/// getConstraintType - Given a constraint letter, return the type of
8551/// constraint it is for this target.
8552ARMTargetLowering::ConstraintType
8553ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8554  if (Constraint.size() == 1) {
8555    switch (Constraint[0]) {
8556    default:  break;
8557    case 'l': return C_RegisterClass;
8558    case 'w': return C_RegisterClass;
8559    case 'h': return C_RegisterClass;
8560    case 'x': return C_RegisterClass;
8561    case 't': return C_RegisterClass;
8562    case 'j': return C_Other; // Constant for movw.
8563      // An address with a single base register. Due to the way we
8564      // currently handle addresses it is the same as an 'r' memory constraint.
8565    case 'Q': return C_Memory;
8566    }
8567  } else if (Constraint.size() == 2) {
8568    switch (Constraint[0]) {
8569    default: break;
8570    // All 'U+' constraints are addresses.
8571    case 'U': return C_Memory;
8572    }
8573  }
8574  return TargetLowering::getConstraintType(Constraint);
8575}
8576
8577/// Examine constraint type and operand type and determine a weight value.
8578/// This object must already have been set up with the operand type
8579/// and the current alternative constraint selected.
8580TargetLowering::ConstraintWeight
8581ARMTargetLowering::getSingleConstraintMatchWeight(
8582    AsmOperandInfo &info, const char *constraint) const {
8583  ConstraintWeight weight = CW_Invalid;
8584  Value *CallOperandVal = info.CallOperandVal;
8585    // If we don't have a value, we can't do a match,
8586    // but allow it at the lowest weight.
8587  if (CallOperandVal == NULL)
8588    return CW_Default;
8589  Type *type = CallOperandVal->getType();
8590  // Look at the constraint type.
8591  switch (*constraint) {
8592  default:
8593    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8594    break;
8595  case 'l':
8596    if (type->isIntegerTy()) {
8597      if (Subtarget->isThumb())
8598        weight = CW_SpecificReg;
8599      else
8600        weight = CW_Register;
8601    }
8602    break;
8603  case 'w':
8604    if (type->isFloatingPointTy())
8605      weight = CW_Register;
8606    break;
8607  }
8608  return weight;
8609}
8610
8611typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8612RCPair
8613ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8614                                                EVT VT) const {
8615  if (Constraint.size() == 1) {
8616    // GCC ARM Constraint Letters
8617    switch (Constraint[0]) {
8618    case 'l': // Low regs or general regs.
8619      if (Subtarget->isThumb())
8620        return RCPair(0U, ARM::tGPRRegisterClass);
8621      else
8622        return RCPair(0U, ARM::GPRRegisterClass);
8623    case 'h': // High regs or no regs.
8624      if (Subtarget->isThumb())
8625        return RCPair(0U, ARM::hGPRRegisterClass);
8626      break;
8627    case 'r':
8628      return RCPair(0U, ARM::GPRRegisterClass);
8629    case 'w':
8630      if (VT == MVT::f32)
8631        return RCPair(0U, ARM::SPRRegisterClass);
8632      if (VT.getSizeInBits() == 64)
8633        return RCPair(0U, ARM::DPRRegisterClass);
8634      if (VT.getSizeInBits() == 128)
8635        return RCPair(0U, ARM::QPRRegisterClass);
8636      break;
8637    case 'x':
8638      if (VT == MVT::f32)
8639        return RCPair(0U, ARM::SPR_8RegisterClass);
8640      if (VT.getSizeInBits() == 64)
8641        return RCPair(0U, ARM::DPR_8RegisterClass);
8642      if (VT.getSizeInBits() == 128)
8643        return RCPair(0U, ARM::QPR_8RegisterClass);
8644      break;
8645    case 't':
8646      if (VT == MVT::f32)
8647        return RCPair(0U, ARM::SPRRegisterClass);
8648      break;
8649    }
8650  }
8651  if (StringRef("{cc}").equals_lower(Constraint))
8652    return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8653
8654  return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8655}
8656
8657/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8658/// vector.  If it is invalid, don't add anything to Ops.
8659void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8660                                                     std::string &Constraint,
8661                                                     std::vector<SDValue>&Ops,
8662                                                     SelectionDAG &DAG) const {
8663  SDValue Result(0, 0);
8664
8665  // Currently only support length 1 constraints.
8666  if (Constraint.length() != 1) return;
8667
8668  char ConstraintLetter = Constraint[0];
8669  switch (ConstraintLetter) {
8670  default: break;
8671  case 'j':
8672  case 'I': case 'J': case 'K': case 'L':
8673  case 'M': case 'N': case 'O':
8674    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8675    if (!C)
8676      return;
8677
8678    int64_t CVal64 = C->getSExtValue();
8679    int CVal = (int) CVal64;
8680    // None of these constraints allow values larger than 32 bits.  Check
8681    // that the value fits in an int.
8682    if (CVal != CVal64)
8683      return;
8684
8685    switch (ConstraintLetter) {
8686      case 'j':
8687        // Constant suitable for movw, must be between 0 and
8688        // 65535.
8689        if (Subtarget->hasV6T2Ops())
8690          if (CVal >= 0 && CVal <= 65535)
8691            break;
8692        return;
8693      case 'I':
8694        if (Subtarget->isThumb1Only()) {
8695          // This must be a constant between 0 and 255, for ADD
8696          // immediates.
8697          if (CVal >= 0 && CVal <= 255)
8698            break;
8699        } else if (Subtarget->isThumb2()) {
8700          // A constant that can be used as an immediate value in a
8701          // data-processing instruction.
8702          if (ARM_AM::getT2SOImmVal(CVal) != -1)
8703            break;
8704        } else {
8705          // A constant that can be used as an immediate value in a
8706          // data-processing instruction.
8707          if (ARM_AM::getSOImmVal(CVal) != -1)
8708            break;
8709        }
8710        return;
8711
8712      case 'J':
8713        if (Subtarget->isThumb()) {  // FIXME thumb2
8714          // This must be a constant between -255 and -1, for negated ADD
8715          // immediates. This can be used in GCC with an "n" modifier that
8716          // prints the negated value, for use with SUB instructions. It is
8717          // not useful otherwise but is implemented for compatibility.
8718          if (CVal >= -255 && CVal <= -1)
8719            break;
8720        } else {
8721          // This must be a constant between -4095 and 4095. It is not clear
8722          // what this constraint is intended for. Implemented for
8723          // compatibility with GCC.
8724          if (CVal >= -4095 && CVal <= 4095)
8725            break;
8726        }
8727        return;
8728
8729      case 'K':
8730        if (Subtarget->isThumb1Only()) {
8731          // A 32-bit value where only one byte has a nonzero value. Exclude
8732          // zero to match GCC. This constraint is used by GCC internally for
8733          // constants that can be loaded with a move/shift combination.
8734          // It is not useful otherwise but is implemented for compatibility.
8735          if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
8736            break;
8737        } else if (Subtarget->isThumb2()) {
8738          // A constant whose bitwise inverse can be used as an immediate
8739          // value in a data-processing instruction. This can be used in GCC
8740          // with a "B" modifier that prints the inverted value, for use with
8741          // BIC and MVN instructions. It is not useful otherwise but is
8742          // implemented for compatibility.
8743          if (ARM_AM::getT2SOImmVal(~CVal) != -1)
8744            break;
8745        } else {
8746          // A constant whose bitwise inverse can be used as an immediate
8747          // value in a data-processing instruction. This can be used in GCC
8748          // with a "B" modifier that prints the inverted value, for use with
8749          // BIC and MVN instructions. It is not useful otherwise but is
8750          // implemented for compatibility.
8751          if (ARM_AM::getSOImmVal(~CVal) != -1)
8752            break;
8753        }
8754        return;
8755
8756      case 'L':
8757        if (Subtarget->isThumb1Only()) {
8758          // This must be a constant between -7 and 7,
8759          // for 3-operand ADD/SUB immediate instructions.
8760          if (CVal >= -7 && CVal < 7)
8761            break;
8762        } else if (Subtarget->isThumb2()) {
8763          // A constant whose negation can be used as an immediate value in a
8764          // data-processing instruction. This can be used in GCC with an "n"
8765          // modifier that prints the negated value, for use with SUB
8766          // instructions. It is not useful otherwise but is implemented for
8767          // compatibility.
8768          if (ARM_AM::getT2SOImmVal(-CVal) != -1)
8769            break;
8770        } else {
8771          // A constant whose negation can be used as an immediate value in a
8772          // data-processing instruction. This can be used in GCC with an "n"
8773          // modifier that prints the negated value, for use with SUB
8774          // instructions. It is not useful otherwise but is implemented for
8775          // compatibility.
8776          if (ARM_AM::getSOImmVal(-CVal) != -1)
8777            break;
8778        }
8779        return;
8780
8781      case 'M':
8782        if (Subtarget->isThumb()) { // FIXME thumb2
8783          // This must be a multiple of 4 between 0 and 1020, for
8784          // ADD sp + immediate.
8785          if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
8786            break;
8787        } else {
8788          // A power of two or a constant between 0 and 32.  This is used in
8789          // GCC for the shift amount on shifted register operands, but it is
8790          // useful in general for any shift amounts.
8791          if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
8792            break;
8793        }
8794        return;
8795
8796      case 'N':
8797        if (Subtarget->isThumb()) {  // FIXME thumb2
8798          // This must be a constant between 0 and 31, for shift amounts.
8799          if (CVal >= 0 && CVal <= 31)
8800            break;
8801        }
8802        return;
8803
8804      case 'O':
8805        if (Subtarget->isThumb()) {  // FIXME thumb2
8806          // This must be a multiple of 4 between -508 and 508, for
8807          // ADD/SUB sp = sp + immediate.
8808          if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
8809            break;
8810        }
8811        return;
8812    }
8813    Result = DAG.getTargetConstant(CVal, Op.getValueType());
8814    break;
8815  }
8816
8817  if (Result.getNode()) {
8818    Ops.push_back(Result);
8819    return;
8820  }
8821  return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
8822}
8823
8824bool
8825ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
8826  // The ARM target isn't yet aware of offsets.
8827  return false;
8828}
8829
8830bool ARM::isBitFieldInvertedMask(unsigned v) {
8831  if (v == 0xffffffff)
8832    return 0;
8833  // there can be 1's on either or both "outsides", all the "inside"
8834  // bits must be 0's
8835  unsigned int lsb = 0, msb = 31;
8836  while (v & (1 << msb)) --msb;
8837  while (v & (1 << lsb)) ++lsb;
8838  for (unsigned int i = lsb; i <= msb; ++i) {
8839    if (v & (1 << i))
8840      return 0;
8841  }
8842  return 1;
8843}
8844
8845/// isFPImmLegal - Returns true if the target can instruction select the
8846/// specified FP immediate natively. If false, the legalizer will
8847/// materialize the FP immediate as a load from a constant pool.
8848bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
8849  if (!Subtarget->hasVFP3())
8850    return false;
8851  if (VT == MVT::f32)
8852    return ARM_AM::getFP32Imm(Imm) != -1;
8853  if (VT == MVT::f64)
8854    return ARM_AM::getFP64Imm(Imm) != -1;
8855  return false;
8856}
8857
8858/// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
8859/// MemIntrinsicNodes.  The associated MachineMemOperands record the alignment
8860/// specified in the intrinsic calls.
8861bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
8862                                           const CallInst &I,
8863                                           unsigned Intrinsic) const {
8864  switch (Intrinsic) {
8865  case Intrinsic::arm_neon_vld1:
8866  case Intrinsic::arm_neon_vld2:
8867  case Intrinsic::arm_neon_vld3:
8868  case Intrinsic::arm_neon_vld4:
8869  case Intrinsic::arm_neon_vld2lane:
8870  case Intrinsic::arm_neon_vld3lane:
8871  case Intrinsic::arm_neon_vld4lane: {
8872    Info.opc = ISD::INTRINSIC_W_CHAIN;
8873    // Conservatively set memVT to the entire set of vectors loaded.
8874    uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
8875    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8876    Info.ptrVal = I.getArgOperand(0);
8877    Info.offset = 0;
8878    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8879    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8880    Info.vol = false; // volatile loads with NEON intrinsics not supported
8881    Info.readMem = true;
8882    Info.writeMem = false;
8883    return true;
8884  }
8885  case Intrinsic::arm_neon_vst1:
8886  case Intrinsic::arm_neon_vst2:
8887  case Intrinsic::arm_neon_vst3:
8888  case Intrinsic::arm_neon_vst4:
8889  case Intrinsic::arm_neon_vst2lane:
8890  case Intrinsic::arm_neon_vst3lane:
8891  case Intrinsic::arm_neon_vst4lane: {
8892    Info.opc = ISD::INTRINSIC_VOID;
8893    // Conservatively set memVT to the entire set of vectors stored.
8894    unsigned NumElts = 0;
8895    for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
8896      Type *ArgTy = I.getArgOperand(ArgI)->getType();
8897      if (!ArgTy->isVectorTy())
8898        break;
8899      NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
8900    }
8901    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
8902    Info.ptrVal = I.getArgOperand(0);
8903    Info.offset = 0;
8904    Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
8905    Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
8906    Info.vol = false; // volatile stores with NEON intrinsics not supported
8907    Info.readMem = false;
8908    Info.writeMem = true;
8909    return true;
8910  }
8911  case Intrinsic::arm_strexd: {
8912    Info.opc = ISD::INTRINSIC_W_CHAIN;
8913    Info.memVT = MVT::i64;
8914    Info.ptrVal = I.getArgOperand(2);
8915    Info.offset = 0;
8916    Info.align = 8;
8917    Info.vol = true;
8918    Info.readMem = false;
8919    Info.writeMem = true;
8920    return true;
8921  }
8922  case Intrinsic::arm_ldrexd: {
8923    Info.opc = ISD::INTRINSIC_W_CHAIN;
8924    Info.memVT = MVT::i64;
8925    Info.ptrVal = I.getArgOperand(0);
8926    Info.offset = 0;
8927    Info.align = 8;
8928    Info.vol = true;
8929    Info.readMem = true;
8930    Info.writeMem = false;
8931    return true;
8932  }
8933  default:
8934    break;
8935  }
8936
8937  return false;
8938}
8939