1
//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This file defines the interfaces that ARM uses to lower LLVM code into a
13
//===----------------------------------------------------------------------===//
15
#define DEBUG_TYPE "arm-isel"
16
#include "ARMISelLowering.h"
18
#include "ARMCallingConv.h"
19
#include "ARMConstantPoolValue.h"
20
#include "ARMMachineFunctionInfo.h"
21
#include "ARMPerfectShuffle.h"
22
#include "ARMSubtarget.h"
23
#include "ARMTargetMachine.h"
24
#include "ARMTargetObjectFile.h"
25
#include "MCTargetDesc/ARMAddressingModes.h"
26
#include "llvm/CallingConv.h"
27
#include "llvm/Constants.h"
28
#include "llvm/Function.h"
29
#include "llvm/GlobalValue.h"
30
#include "llvm/Instruction.h"
31
#include "llvm/Instructions.h"
32
#include "llvm/Intrinsics.h"
33
#include "llvm/Type.h"
34
#include "llvm/CodeGen/CallingConvLower.h"
35
#include "llvm/CodeGen/IntrinsicLowering.h"
36
#include "llvm/CodeGen/MachineBasicBlock.h"
37
#include "llvm/CodeGen/MachineFrameInfo.h"
38
#include "llvm/CodeGen/MachineFunction.h"
39
#include "llvm/CodeGen/MachineInstrBuilder.h"
40
#include "llvm/CodeGen/MachineModuleInfo.h"
41
#include "llvm/CodeGen/MachineRegisterInfo.h"
42
#include "llvm/CodeGen/SelectionDAG.h"
43
#include "llvm/MC/MCSectionMachO.h"
44
#include "llvm/Target/TargetOptions.h"
45
#include "llvm/ADT/StringExtras.h"
46
#include "llvm/ADT/Statistic.h"
47
#include "llvm/Support/CommandLine.h"
48
#include "llvm/Support/ErrorHandling.h"
49
#include "llvm/Support/MathExtras.h"
50
#include "llvm/Support/raw_ostream.h"
53
STATISTIC(NumTailCalls, "Number of tail calls");
54
STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
56
// This option should go away when tail calls fully work.
58
EnableARMTailCalls("arm-tail-calls", cl::Hidden,
59
cl::desc("Generate tail calls (TEMPORARY OPTION)."),
63
EnableARMLongCalls("arm-long-calls", cl::Hidden,
64
cl::desc("Generate calls via indirect call instructions"),
68
ARMInterworking("arm-interworking", cl::Hidden,
69
cl::desc("Enable / disable ARM interworking (for debugging only)"),
73
class ARMCCState : public CCState {
75
ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
76
const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
77
LLVMContext &C, ParmContext PC)
78
: CCState(CC, isVarArg, MF, TM, locs, C) {
79
assert(((PC == Call) || (PC == Prologue)) &&
80
"ARMCCState users must specify whether their context is call"
81
"or prologue generation.");
87
// The APCS parameter registers.
88
static const uint16_t GPRArgRegs[] = {
89
ARM::R0, ARM::R1, ARM::R2, ARM::R3
92
void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT,
93
EVT PromotedBitwiseVT) {
94
if (VT != PromotedLdStVT) {
95
setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
96
AddPromotedToType (ISD::LOAD, VT.getSimpleVT(),
97
PromotedLdStVT.getSimpleVT());
99
setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
100
AddPromotedToType (ISD::STORE, VT.getSimpleVT(),
101
PromotedLdStVT.getSimpleVT());
104
EVT ElemTy = VT.getVectorElementType();
105
if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
106
setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
107
setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
108
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
109
if (ElemTy == MVT::i32) {
110
setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Custom);
111
setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Custom);
112
setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
113
setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);
115
setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
116
setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
117
setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
118
setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
120
setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
121
setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
122
setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);
123
setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Legal);
124
setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
125
setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
126
setOperationAction(ISD::SIGN_EXTEND_INREG, VT.getSimpleVT(), Expand);
127
if (VT.isInteger()) {
128
setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
129
setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
130
setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
133
// Promote all bit-wise operations.
134
if (VT.isInteger() && VT != PromotedBitwiseVT) {
135
setOperationAction(ISD::AND, VT.getSimpleVT(), Promote);
136
AddPromotedToType (ISD::AND, VT.getSimpleVT(),
137
PromotedBitwiseVT.getSimpleVT());
138
setOperationAction(ISD::OR, VT.getSimpleVT(), Promote);
139
AddPromotedToType (ISD::OR, VT.getSimpleVT(),
140
PromotedBitwiseVT.getSimpleVT());
141
setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote);
142
AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
143
PromotedBitwiseVT.getSimpleVT());
146
// Neon does not support vector divide/remainder operations.
147
setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
148
setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
149
setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
150
setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
151
setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
152
setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
155
void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
156
addRegisterClass(VT, ARM::DPRRegisterClass);
157
addTypeForNEON(VT, MVT::f64, MVT::v2i32);
160
void ARMTargetLowering::addQRTypeForNEON(EVT VT) {
161
addRegisterClass(VT, ARM::QPRRegisterClass);
162
addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
165
static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
166
if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
167
return new TargetLoweringObjectFileMachO();
169
return new ARMElfTargetObjectFile();
172
ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
173
: TargetLowering(TM, createTLOF(TM)) {
174
Subtarget = &TM.getSubtarget<ARMSubtarget>();
175
RegInfo = TM.getRegisterInfo();
176
Itins = TM.getInstrItineraryData();
178
setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
180
if (Subtarget->isTargetDarwin()) {
181
// Uses VFP for Thumb libfuncs if available.
182
if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
183
// Single-precision floating-point arithmetic.
184
setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
185
setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
186
setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
187
setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");
189
// Double-precision floating-point arithmetic.
190
setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
191
setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
192
setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
193
setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");
195
// Single-precision comparisons.
196
setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
197
setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
198
setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
199
setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
200
setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
201
setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
202
setLibcallName(RTLIB::UO_F32, "__unordsf2vfp");
203
setLibcallName(RTLIB::O_F32, "__unordsf2vfp");
205
setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
206
setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
207
setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
208
setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
209
setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
210
setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
211
setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
212
setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
214
// Double-precision comparisons.
215
setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
216
setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
217
setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
218
setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
219
setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
220
setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
221
setLibcallName(RTLIB::UO_F64, "__unorddf2vfp");
222
setLibcallName(RTLIB::O_F64, "__unorddf2vfp");
224
setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
225
setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
226
setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
227
setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
228
setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
229
setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
230
setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
231
setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
233
// Floating-point to integer conversions.
234
// i64 conversions are done via library routines even when generating VFP
235
// instructions, so use the same ones.
236
setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
237
setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
238
setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
239
setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");
241
// Conversions between floating types.
242
setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
243
setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp");
245
// Integer to floating-point conversions.
246
// i64 conversions are done via library routines even when generating VFP
247
// instructions, so use the same ones.
248
// FIXME: There appears to be some naming inconsistency in ARM libgcc:
249
// e.g., __floatunsidf vs. __floatunssidfvfp.
250
setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
251
setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
252
setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
253
setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
257
// These libcalls are not available in 32-bit.
258
setLibcallName(RTLIB::SHL_I128, 0);
259
setLibcallName(RTLIB::SRL_I128, 0);
260
setLibcallName(RTLIB::SRA_I128, 0);
262
if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
263
// Double-precision floating-point arithmetic helper functions
264
// RTABI chapter 4.1.2, Table 2
265
setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
266
setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
267
setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
268
setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
269
setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
270
setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
271
setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
272
setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);
274
// Double-precision floating-point comparison helper functions
275
// RTABI chapter 4.1.2, Table 3
276
setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
277
setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
278
setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
279
setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
280
setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
281
setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
282
setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
283
setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
284
setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
285
setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
286
setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
287
setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
288
setLibcallName(RTLIB::UO_F64, "__aeabi_dcmpun");
289
setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE);
290
setLibcallName(RTLIB::O_F64, "__aeabi_dcmpun");
291
setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ);
292
setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
293
setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
294
setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
295
setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
296
setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
297
setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
298
setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
299
setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);
301
// Single-precision floating-point arithmetic helper functions
302
// RTABI chapter 4.1.2, Table 4
303
setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
304
setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
305
setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
306
setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
307
setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
308
setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
309
setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
310
setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);
312
// Single-precision floating-point comparison helper functions
313
// RTABI chapter 4.1.2, Table 5
314
setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
315
setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
316
setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
317
setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
318
setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
319
setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
320
setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
321
setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
322
setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
323
setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
324
setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
325
setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
326
setLibcallName(RTLIB::UO_F32, "__aeabi_fcmpun");
327
setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE);
328
setLibcallName(RTLIB::O_F32, "__aeabi_fcmpun");
329
setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ);
330
setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
331
setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
332
setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
333
setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
334
setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
335
setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
336
setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
337
setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);
339
// Floating-point to integer conversions.
340
// RTABI chapter 4.1.2, Table 6
341
setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
342
setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
343
setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
344
setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
345
setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
346
setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
347
setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
348
setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
349
setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
350
setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
351
setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
352
setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
353
setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
354
setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
355
setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
356
setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);
358
// Conversions between floating types.
359
// RTABI chapter 4.1.2, Table 7
360
setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
361
setLibcallName(RTLIB::FPEXT_F32_F64, "__aeabi_f2d");
362
setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
363
setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);
365
// Integer to floating-point conversions.
366
// RTABI chapter 4.1.2, Table 8
367
setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
368
setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
369
setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
370
setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
371
setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
372
setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
373
setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
374
setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
375
setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
376
setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
377
setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
378
setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
379
setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
380
setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
381
setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
382
setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
384
// Long long helper functions
385
// RTABI chapter 4.2, Table 9
386
setLibcallName(RTLIB::MUL_I64, "__aeabi_lmul");
387
setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
388
setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
389
setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
390
setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
391
setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
392
setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
393
setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
394
setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
395
setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);
397
// Integer division functions
398
// RTABI chapter 4.3.1
399
setLibcallName(RTLIB::SDIV_I8, "__aeabi_idiv");
400
setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
401
setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
402
setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
403
setLibcallName(RTLIB::UDIV_I8, "__aeabi_uidiv");
404
setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
405
setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
406
setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
407
setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
408
setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
409
setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
410
setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
411
setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
412
setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
413
setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
414
setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
417
// RTABI chapter 4.3.4
418
setLibcallName(RTLIB::MEMCPY, "__aeabi_memcpy");
419
setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
420
setLibcallName(RTLIB::MEMSET, "__aeabi_memset");
421
setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
422
setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
423
setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
426
// Use divmod compiler-rt calls for iOS 5.0 and later.
427
if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
428
!Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
429
setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
430
setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
433
if (Subtarget->isThumb1Only())
434
addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
436
addRegisterClass(MVT::i32, ARM::GPRRegisterClass);
437
if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
438
!Subtarget->isThumb1Only()) {
439
addRegisterClass(MVT::f32, ARM::SPRRegisterClass);
440
if (!Subtarget->isFPOnlySP())
441
addRegisterClass(MVT::f64, ARM::DPRRegisterClass);
443
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
446
for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
447
VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
448
for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
449
InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
450
setTruncStoreAction((MVT::SimpleValueType)VT,
451
(MVT::SimpleValueType)InnerVT, Expand);
452
setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
453
setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
454
setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
457
setOperationAction(ISD::ConstantFP, MVT::f32, Custom);
459
if (Subtarget->hasNEON()) {
460
addDRTypeForNEON(MVT::v2f32);
461
addDRTypeForNEON(MVT::v8i8);
462
addDRTypeForNEON(MVT::v4i16);
463
addDRTypeForNEON(MVT::v2i32);
464
addDRTypeForNEON(MVT::v1i64);
466
addQRTypeForNEON(MVT::v4f32);
467
addQRTypeForNEON(MVT::v2f64);
468
addQRTypeForNEON(MVT::v16i8);
469
addQRTypeForNEON(MVT::v8i16);
470
addQRTypeForNEON(MVT::v4i32);
471
addQRTypeForNEON(MVT::v2i64);
473
// v2f64 is legal so that QR subregs can be extracted as f64 elements, but
474
// neither Neon nor VFP support any arithmetic operations on it.
475
// The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
476
// supported for v4f32.
477
setOperationAction(ISD::FADD, MVT::v2f64, Expand);
478
setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
479
setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
480
// FIXME: Code duplication: FDIV and FREM are expanded always, see
481
// ARMTargetLowering::addTypeForNEON method for details.
482
setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
483
setOperationAction(ISD::FREM, MVT::v2f64, Expand);
484
// FIXME: Create unittest.
485
// In another words, find a way when "copysign" appears in DAG with vector
487
setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
488
// FIXME: Code duplication: SETCC has custom operation action, see
489
// ARMTargetLowering::addTypeForNEON method for details.
490
setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
491
// FIXME: Create unittest for FNEG and for FABS.
492
setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
493
setOperationAction(ISD::FABS, MVT::v2f64, Expand);
494
setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
495
setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
496
setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
497
setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
498
setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
499
setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
500
setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
501
setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
502
setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
503
setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
504
// FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
505
setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
506
setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
507
setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
508
setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
509
setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
511
setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
512
setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
513
setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
514
setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
515
setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
516
setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
517
setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
518
setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
519
setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
520
setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
522
// Neon does not support some operations on v1i64 and v2i64 types.
523
setOperationAction(ISD::MUL, MVT::v1i64, Expand);
524
// Custom handling for some quad-vector types to detect VMULL.
525
setOperationAction(ISD::MUL, MVT::v8i16, Custom);
526
setOperationAction(ISD::MUL, MVT::v4i32, Custom);
527
setOperationAction(ISD::MUL, MVT::v2i64, Custom);
528
// Custom handling for some vector types to avoid expensive expansions
529
setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
530
setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
531
setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
532
setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
533
setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
534
setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
535
// Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
536
// a destination type that is wider than the source, and nor does
537
// it have a FP_TO_[SU]INT instruction with a narrower destination than
539
setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
540
setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
541
setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
542
setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
544
setTargetDAGCombine(ISD::INTRINSIC_VOID);
545
setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
546
setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
547
setTargetDAGCombine(ISD::SHL);
548
setTargetDAGCombine(ISD::SRL);
549
setTargetDAGCombine(ISD::SRA);
550
setTargetDAGCombine(ISD::SIGN_EXTEND);
551
setTargetDAGCombine(ISD::ZERO_EXTEND);
552
setTargetDAGCombine(ISD::ANY_EXTEND);
553
setTargetDAGCombine(ISD::SELECT_CC);
554
setTargetDAGCombine(ISD::BUILD_VECTOR);
555
setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
556
setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
557
setTargetDAGCombine(ISD::STORE);
558
setTargetDAGCombine(ISD::FP_TO_SINT);
559
setTargetDAGCombine(ISD::FP_TO_UINT);
560
setTargetDAGCombine(ISD::FDIV);
562
// It is legal to extload from v4i8 to v4i16 or v4i32.
563
MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
564
MVT::v4i16, MVT::v2i16,
566
for (unsigned i = 0; i < 6; ++i) {
567
setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
568
setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
569
setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
573
computeRegisterProperties();
575
// ARM does not have f32 extending load.
576
setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
578
// ARM does not have i1 sign extending load.
579
setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
581
// ARM supports all 4 flavors of integer indexed load / store.
582
if (!Subtarget->isThumb1Only()) {
583
for (unsigned im = (unsigned)ISD::PRE_INC;
584
im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
585
setIndexedLoadAction(im, MVT::i1, Legal);
586
setIndexedLoadAction(im, MVT::i8, Legal);
587
setIndexedLoadAction(im, MVT::i16, Legal);
588
setIndexedLoadAction(im, MVT::i32, Legal);
589
setIndexedStoreAction(im, MVT::i1, Legal);
590
setIndexedStoreAction(im, MVT::i8, Legal);
591
setIndexedStoreAction(im, MVT::i16, Legal);
592
setIndexedStoreAction(im, MVT::i32, Legal);
596
// i64 operation support.
597
setOperationAction(ISD::MUL, MVT::i64, Expand);
598
setOperationAction(ISD::MULHU, MVT::i32, Expand);
599
if (Subtarget->isThumb1Only()) {
600
setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
601
setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
603
if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
604
|| (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
605
setOperationAction(ISD::MULHS, MVT::i32, Expand);
607
setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
608
setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
609
setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
610
setOperationAction(ISD::SRL, MVT::i64, Custom);
611
setOperationAction(ISD::SRA, MVT::i64, Custom);
613
if (!Subtarget->isThumb1Only()) {
614
// FIXME: We should do this for Thumb1 as well.
615
setOperationAction(ISD::ADDC, MVT::i32, Custom);
616
setOperationAction(ISD::ADDE, MVT::i32, Custom);
617
setOperationAction(ISD::SUBC, MVT::i32, Custom);
618
setOperationAction(ISD::SUBE, MVT::i32, Custom);
621
// ARM does not have ROTL.
622
setOperationAction(ISD::ROTL, MVT::i32, Expand);
623
setOperationAction(ISD::CTTZ, MVT::i32, Custom);
624
setOperationAction(ISD::CTPOP, MVT::i32, Expand);
625
if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
626
setOperationAction(ISD::CTLZ, MVT::i32, Expand);
628
// These just redirect to CTTZ and CTLZ on ARM.
629
setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
630
setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
632
// Only ARMv6 has BSWAP.
633
if (!Subtarget->hasV6Ops())
634
setOperationAction(ISD::BSWAP, MVT::i32, Expand);
636
// These are expanded into libcalls.
637
if (!Subtarget->hasDivide() || !Subtarget->isThumb2()) {
638
// v7M has a hardware divider
639
setOperationAction(ISD::SDIV, MVT::i32, Expand);
640
setOperationAction(ISD::UDIV, MVT::i32, Expand);
642
setOperationAction(ISD::SREM, MVT::i32, Expand);
643
setOperationAction(ISD::UREM, MVT::i32, Expand);
644
setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
645
setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
647
setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
648
setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
649
setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
650
setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
651
setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
653
setOperationAction(ISD::TRAP, MVT::Other, Legal);
655
// Use the default implementation.
656
setOperationAction(ISD::VASTART, MVT::Other, Custom);
657
setOperationAction(ISD::VAARG, MVT::Other, Expand);
658
setOperationAction(ISD::VACOPY, MVT::Other, Expand);
659
setOperationAction(ISD::VAEND, MVT::Other, Expand);
660
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
661
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
663
if (!Subtarget->isTargetDarwin()) {
664
// Non-Darwin platforms may return values in these registers via the
665
// personality function.
666
setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
667
setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
668
setExceptionPointerRegister(ARM::R0);
669
setExceptionSelectorRegister(ARM::R1);
672
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
673
// ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
674
// the default expansion.
675
// FIXME: This should be checking for v6k, not just v6.
676
if (Subtarget->hasDataBarrier() ||
677
(Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
678
// membarrier needs custom lowering; the rest are legal and handled
680
setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
681
setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
682
// Custom lowering for 64-bit ops
683
setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
684
setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
685
setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
686
setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
687
setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
688
setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
689
setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
690
// Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
691
setInsertFencesForAtomic(true);
693
// Set them all for expansion, which will force libcalls.
694
setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
695
setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Expand);
696
setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Expand);
697
setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, Expand);
698
setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, Expand);
699
setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
700
setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Expand);
701
setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, Expand);
702
setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, Expand);
703
setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
704
setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
705
setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
706
setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
707
setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
708
// Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
709
// Unordered/Monotonic case.
710
setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
711
setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
712
// Since the libcalls include locking, fold in the fences
713
setShouldFoldAtomicFences(true);
716
setOperationAction(ISD::PREFETCH, MVT::Other, Custom);
718
// Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
719
if (!Subtarget->hasV6Ops()) {
720
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
721
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
723
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
725
if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
726
!Subtarget->isThumb1Only()) {
727
// Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
728
// iff target supports vfp2.
729
setOperationAction(ISD::BITCAST, MVT::i64, Custom);
730
setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
733
// We want to custom lower some of our intrinsics.
734
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
735
if (Subtarget->isTargetDarwin()) {
736
setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
737
setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
738
setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
741
setOperationAction(ISD::SETCC, MVT::i32, Expand);
742
setOperationAction(ISD::SETCC, MVT::f32, Expand);
743
setOperationAction(ISD::SETCC, MVT::f64, Expand);
744
setOperationAction(ISD::SELECT, MVT::i32, Custom);
745
setOperationAction(ISD::SELECT, MVT::f32, Custom);
746
setOperationAction(ISD::SELECT, MVT::f64, Custom);
747
setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
748
setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
749
setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
751
setOperationAction(ISD::BRCOND, MVT::Other, Expand);
752
setOperationAction(ISD::BR_CC, MVT::i32, Custom);
753
setOperationAction(ISD::BR_CC, MVT::f32, Custom);
754
setOperationAction(ISD::BR_CC, MVT::f64, Custom);
755
setOperationAction(ISD::BR_JT, MVT::Other, Custom);
757
// We don't support sin/cos/fmod/copysign/pow
758
setOperationAction(ISD::FSIN, MVT::f64, Expand);
759
setOperationAction(ISD::FSIN, MVT::f32, Expand);
760
setOperationAction(ISD::FCOS, MVT::f32, Expand);
761
setOperationAction(ISD::FCOS, MVT::f64, Expand);
762
setOperationAction(ISD::FREM, MVT::f64, Expand);
763
setOperationAction(ISD::FREM, MVT::f32, Expand);
764
if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
765
!Subtarget->isThumb1Only()) {
766
setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
767
setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
769
setOperationAction(ISD::FPOW, MVT::f64, Expand);
770
setOperationAction(ISD::FPOW, MVT::f32, Expand);
772
setOperationAction(ISD::FMA, MVT::f64, Expand);
773
setOperationAction(ISD::FMA, MVT::f32, Expand);
775
// Various VFP goodness
776
if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
777
// int <-> fp are custom expanded into bit_convert + ARMISD ops.
778
if (Subtarget->hasVFP2()) {
779
setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
780
setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
781
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
782
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
784
// Special handling for half-precision FP.
785
if (!Subtarget->hasFP16()) {
786
setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
787
setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
791
// We have target-specific dag combine patterns for the following nodes:
792
// ARMISD::VMOVRRD - No need to call setTargetDAGCombine
793
setTargetDAGCombine(ISD::ADD);
794
setTargetDAGCombine(ISD::SUB);
795
setTargetDAGCombine(ISD::MUL);
797
if (Subtarget->hasV6T2Ops() || Subtarget->hasNEON()) {
798
setTargetDAGCombine(ISD::AND);
799
setTargetDAGCombine(ISD::OR);
800
setTargetDAGCombine(ISD::XOR);
803
if (Subtarget->hasV6Ops())
804
setTargetDAGCombine(ISD::SRL);
806
setStackPointerRegisterToSaveRestore(ARM::SP);
808
if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
809
!Subtarget->hasVFP2())
810
setSchedulingPreference(Sched::RegPressure);
812
setSchedulingPreference(Sched::Hybrid);
814
//// temporary - rewrite interface to use type
815
maxStoresPerMemcpy = maxStoresPerMemcpyOptSize = 1;
816
maxStoresPerMemset = 16;
817
maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
819
// On ARM arguments smaller than 4 bytes are extended, so all arguments
820
// are at least 4 bytes aligned.
821
setMinStackArgumentAlignment(4);
823
benefitFromCodePlacementOpt = true;
825
setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
828
// FIXME: It might make sense to define the representative register class as the
829
// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
830
// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
831
// SPR's representative would be DPR_VFP2. This should work well if register
832
// pressure tracking were modified such that a register use would increment the
833
// pressure of the register class's representative and all of it's super
834
// classes' representatives transitively. We have not implemented this because
835
// of the difficulty prior to coalescing of modeling operand register classes
836
// due to the common occurrence of cross class copies and subregister insertions
838
std::pair<const TargetRegisterClass*, uint8_t>
839
ARMTargetLowering::findRepresentativeClass(EVT VT) const{
840
const TargetRegisterClass *RRC = 0;
842
switch (VT.getSimpleVT().SimpleTy) {
844
return TargetLowering::findRepresentativeClass(VT);
845
// Use DPR as representative register class for all floating point
846
// and vector types. Since there are 32 SPR registers and 32 DPR registers so
847
// the cost is 1 for both f32 and f64.
848
case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
849
case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
850
RRC = ARM::DPRRegisterClass;
851
// When NEON is used for SP, only half of the register file is available
852
// because operations that define both SP and DP results will be constrained
853
// to the VFP2 class (D0-D15). We currently model this constraint prior to
854
// coalescing by double-counting the SP regs. See the FIXME above.
855
if (Subtarget->useNEONForSinglePrecisionFP())
858
case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
859
case MVT::v4f32: case MVT::v2f64:
860
RRC = ARM::DPRRegisterClass;
864
RRC = ARM::DPRRegisterClass;
868
RRC = ARM::DPRRegisterClass;
872
return std::make_pair(RRC, Cost);
875
const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
878
case ARMISD::Wrapper: return "ARMISD::Wrapper";
879
case ARMISD::WrapperDYN: return "ARMISD::WrapperDYN";
880
case ARMISD::WrapperPIC: return "ARMISD::WrapperPIC";
881
case ARMISD::WrapperJT: return "ARMISD::WrapperJT";
882
case ARMISD::CALL: return "ARMISD::CALL";
883
case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED";
884
case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK";
885
case ARMISD::tCALL: return "ARMISD::tCALL";
886
case ARMISD::BRCOND: return "ARMISD::BRCOND";
887
case ARMISD::BR_JT: return "ARMISD::BR_JT";
888
case ARMISD::BR2_JT: return "ARMISD::BR2_JT";
889
case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG";
890
case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD";
891
case ARMISD::CMP: return "ARMISD::CMP";
892
case ARMISD::CMPZ: return "ARMISD::CMPZ";
893
case ARMISD::CMPFP: return "ARMISD::CMPFP";
894
case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0";
895
case ARMISD::BCC_i64: return "ARMISD::BCC_i64";
896
case ARMISD::FMSTAT: return "ARMISD::FMSTAT";
898
case ARMISD::CMOV: return "ARMISD::CMOV";
899
case ARMISD::CAND: return "ARMISD::CAND";
900
case ARMISD::COR: return "ARMISD::COR";
901
case ARMISD::CXOR: return "ARMISD::CXOR";
903
case ARMISD::RBIT: return "ARMISD::RBIT";
905
case ARMISD::FTOSI: return "ARMISD::FTOSI";
906
case ARMISD::FTOUI: return "ARMISD::FTOUI";
907
case ARMISD::SITOF: return "ARMISD::SITOF";
908
case ARMISD::UITOF: return "ARMISD::UITOF";
910
case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG";
911
case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
912
case ARMISD::RRX: return "ARMISD::RRX";
914
case ARMISD::ADDC: return "ARMISD::ADDC";
915
case ARMISD::ADDE: return "ARMISD::ADDE";
916
case ARMISD::SUBC: return "ARMISD::SUBC";
917
case ARMISD::SUBE: return "ARMISD::SUBE";
919
case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
920
case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
922
case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
923
case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
925
case ARMISD::TC_RETURN: return "ARMISD::TC_RETURN";
927
case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
929
case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
931
case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
932
case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";
934
case ARMISD::PRELOAD: return "ARMISD::PRELOAD";
936
case ARMISD::VCEQ: return "ARMISD::VCEQ";
937
case ARMISD::VCEQZ: return "ARMISD::VCEQZ";
938
case ARMISD::VCGE: return "ARMISD::VCGE";
939
case ARMISD::VCGEZ: return "ARMISD::VCGEZ";
940
case ARMISD::VCLEZ: return "ARMISD::VCLEZ";
941
case ARMISD::VCGEU: return "ARMISD::VCGEU";
942
case ARMISD::VCGT: return "ARMISD::VCGT";
943
case ARMISD::VCGTZ: return "ARMISD::VCGTZ";
944
case ARMISD::VCLTZ: return "ARMISD::VCLTZ";
945
case ARMISD::VCGTU: return "ARMISD::VCGTU";
946
case ARMISD::VTST: return "ARMISD::VTST";
948
case ARMISD::VSHL: return "ARMISD::VSHL";
949
case ARMISD::VSHRs: return "ARMISD::VSHRs";
950
case ARMISD::VSHRu: return "ARMISD::VSHRu";
951
case ARMISD::VSHLLs: return "ARMISD::VSHLLs";
952
case ARMISD::VSHLLu: return "ARMISD::VSHLLu";
953
case ARMISD::VSHLLi: return "ARMISD::VSHLLi";
954
case ARMISD::VSHRN: return "ARMISD::VSHRN";
955
case ARMISD::VRSHRs: return "ARMISD::VRSHRs";
956
case ARMISD::VRSHRu: return "ARMISD::VRSHRu";
957
case ARMISD::VRSHRN: return "ARMISD::VRSHRN";
958
case ARMISD::VQSHLs: return "ARMISD::VQSHLs";
959
case ARMISD::VQSHLu: return "ARMISD::VQSHLu";
960
case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu";
961
case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs";
962
case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu";
963
case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu";
964
case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs";
965
case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu";
966
case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
967
case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
968
case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
969
case ARMISD::VMOVIMM: return "ARMISD::VMOVIMM";
970
case ARMISD::VMVNIMM: return "ARMISD::VMVNIMM";
971
case ARMISD::VMOVFPIMM: return "ARMISD::VMOVFPIMM";
972
case ARMISD::VDUP: return "ARMISD::VDUP";
973
case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
974
case ARMISD::VEXT: return "ARMISD::VEXT";
975
case ARMISD::VREV64: return "ARMISD::VREV64";
976
case ARMISD::VREV32: return "ARMISD::VREV32";
977
case ARMISD::VREV16: return "ARMISD::VREV16";
978
case ARMISD::VZIP: return "ARMISD::VZIP";
979
case ARMISD::VUZP: return "ARMISD::VUZP";
980
case ARMISD::VTRN: return "ARMISD::VTRN";
981
case ARMISD::VTBL1: return "ARMISD::VTBL1";
982
case ARMISD::VTBL2: return "ARMISD::VTBL2";
983
case ARMISD::VMULLs: return "ARMISD::VMULLs";
984
case ARMISD::VMULLu: return "ARMISD::VMULLu";
985
case ARMISD::BUILD_VECTOR: return "ARMISD::BUILD_VECTOR";
986
case ARMISD::FMAX: return "ARMISD::FMAX";
987
case ARMISD::FMIN: return "ARMISD::FMIN";
988
case ARMISD::BFI: return "ARMISD::BFI";
989
case ARMISD::VORRIMM: return "ARMISD::VORRIMM";
990
case ARMISD::VBICIMM: return "ARMISD::VBICIMM";
991
case ARMISD::VBSL: return "ARMISD::VBSL";
992
case ARMISD::VLD2DUP: return "ARMISD::VLD2DUP";
993
case ARMISD::VLD3DUP: return "ARMISD::VLD3DUP";
994
case ARMISD::VLD4DUP: return "ARMISD::VLD4DUP";
995
case ARMISD::VLD1_UPD: return "ARMISD::VLD1_UPD";
996
case ARMISD::VLD2_UPD: return "ARMISD::VLD2_UPD";
997
case ARMISD::VLD3_UPD: return "ARMISD::VLD3_UPD";
998
case ARMISD::VLD4_UPD: return "ARMISD::VLD4_UPD";
999
case ARMISD::VLD2LN_UPD: return "ARMISD::VLD2LN_UPD";
1000
case ARMISD::VLD3LN_UPD: return "ARMISD::VLD3LN_UPD";
1001
case ARMISD::VLD4LN_UPD: return "ARMISD::VLD4LN_UPD";
1002
case ARMISD::VLD2DUP_UPD: return "ARMISD::VLD2DUP_UPD";
1003
case ARMISD::VLD3DUP_UPD: return "ARMISD::VLD3DUP_UPD";
1004
case ARMISD::VLD4DUP_UPD: return "ARMISD::VLD4DUP_UPD";
1005
case ARMISD::VST1_UPD: return "ARMISD::VST1_UPD";
1006
case ARMISD::VST2_UPD: return "ARMISD::VST2_UPD";
1007
case ARMISD::VST3_UPD: return "ARMISD::VST3_UPD";
1008
case ARMISD::VST4_UPD: return "ARMISD::VST4_UPD";
1009
case ARMISD::VST2LN_UPD: return "ARMISD::VST2LN_UPD";
1010
case ARMISD::VST3LN_UPD: return "ARMISD::VST3LN_UPD";
1011
case ARMISD::VST4LN_UPD: return "ARMISD::VST4LN_UPD";
1015
EVT ARMTargetLowering::getSetCCResultType(EVT VT) const {
1016
if (!VT.isVector()) return getPointerTy();
1017
return VT.changeVectorElementTypeToInteger();
1020
/// getRegClassFor - Return the register class that should be used for the
1021
/// specified value type.
1022
const TargetRegisterClass *ARMTargetLowering::getRegClassFor(EVT VT) const {
1023
// Map v4i64 to QQ registers but do not make the type legal. Similarly map
1024
// v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
1025
// load / store 4 to 8 consecutive D registers.
1026
if (Subtarget->hasNEON()) {
1027
if (VT == MVT::v4i64)
1028
return ARM::QQPRRegisterClass;
1029
else if (VT == MVT::v8i64)
1030
return ARM::QQQQPRRegisterClass;
1032
return TargetLowering::getRegClassFor(VT);
1035
// Create a fast isel object.
1037
ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
1038
return ARM::createFastISel(funcInfo);
1041
/// getMaximalGlobalOffset - Returns the maximal possible offset which can
1042
/// be used for loads / stores from the global.
1043
unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
1044
return (Subtarget->isThumb1Only() ? 127 : 4095);
1047
Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
1048
unsigned NumVals = N->getNumValues();
1050
return Sched::RegPressure;
1052
for (unsigned i = 0; i != NumVals; ++i) {
1053
EVT VT = N->getValueType(i);
1054
if (VT == MVT::Glue || VT == MVT::Other)
1056
if (VT.isFloatingPoint() || VT.isVector())
1060
if (!N->isMachineOpcode())
1061
return Sched::RegPressure;
1063
// Load are scheduled for latency even if there instruction itinerary
1064
// is not available.
1065
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1066
const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
1068
if (MCID.getNumDefs() == 0)
1069
return Sched::RegPressure;
1070
if (!Itins->isEmpty() &&
1071
Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
1074
return Sched::RegPressure;
1077
//===----------------------------------------------------------------------===//
1079
//===----------------------------------------------------------------------===//
1081
/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
1082
static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
1084
default: llvm_unreachable("Unknown condition code!");
1085
case ISD::SETNE: return ARMCC::NE;
1086
case ISD::SETEQ: return ARMCC::EQ;
1087
case ISD::SETGT: return ARMCC::GT;
1088
case ISD::SETGE: return ARMCC::GE;
1089
case ISD::SETLT: return ARMCC::LT;
1090
case ISD::SETLE: return ARMCC::LE;
1091
case ISD::SETUGT: return ARMCC::HI;
1092
case ISD::SETUGE: return ARMCC::HS;
1093
case ISD::SETULT: return ARMCC::LO;
1094
case ISD::SETULE: return ARMCC::LS;
1098
/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
1099
static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
1100
ARMCC::CondCodes &CondCode2) {
1101
CondCode2 = ARMCC::AL;
1103
default: llvm_unreachable("Unknown FP condition!");
1105
case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
1107
case ISD::SETOGT: CondCode = ARMCC::GT; break;
1109
case ISD::SETOGE: CondCode = ARMCC::GE; break;
1110
case ISD::SETOLT: CondCode = ARMCC::MI; break;
1111
case ISD::SETOLE: CondCode = ARMCC::LS; break;
1112
case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
1113
case ISD::SETO: CondCode = ARMCC::VC; break;
1114
case ISD::SETUO: CondCode = ARMCC::VS; break;
1115
case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
1116
case ISD::SETUGT: CondCode = ARMCC::HI; break;
1117
case ISD::SETUGE: CondCode = ARMCC::PL; break;
1119
case ISD::SETULT: CondCode = ARMCC::LT; break;
1121
case ISD::SETULE: CondCode = ARMCC::LE; break;
1123
case ISD::SETUNE: CondCode = ARMCC::NE; break;
1127
//===----------------------------------------------------------------------===//
1128
// Calling Convention Implementation
1129
//===----------------------------------------------------------------------===//
1131
#include "ARMGenCallingConv.inc"
1133
/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1134
/// given CallingConvention value.
1135
CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
1137
bool isVarArg) const {
1140
llvm_unreachable("Unsupported calling convention");
1141
case CallingConv::Fast:
1142
if (Subtarget->hasVFP2() && !isVarArg) {
1143
if (!Subtarget->isAAPCS_ABI())
1144
return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
1145
// For AAPCS ABI targets, just use VFP variant of the calling convention.
1146
return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1149
case CallingConv::C: {
1150
// Use target triple & subtarget features to do actual dispatch.
1151
if (!Subtarget->isAAPCS_ABI())
1152
return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1153
else if (Subtarget->hasVFP2() &&
1154
getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
1156
return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1157
return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1159
case CallingConv::ARM_AAPCS_VFP:
1161
return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
1163
case CallingConv::ARM_AAPCS:
1164
return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
1165
case CallingConv::ARM_APCS:
1166
return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
1170
/// LowerCallResult - Lower the result values of a call into the
1171
/// appropriate copies out of appropriate physical registers.
1173
ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1174
CallingConv::ID CallConv, bool isVarArg,
1175
const SmallVectorImpl<ISD::InputArg> &Ins,
1176
DebugLoc dl, SelectionDAG &DAG,
1177
SmallVectorImpl<SDValue> &InVals) const {
1179
// Assign locations to each value returned by this call.
1180
SmallVector<CCValAssign, 16> RVLocs;
1181
ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1182
getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1183
CCInfo.AnalyzeCallResult(Ins,
1184
CCAssignFnForNode(CallConv, /* Return*/ true,
1187
// Copy all of the result registers out of their specified physreg.
1188
for (unsigned i = 0; i != RVLocs.size(); ++i) {
1189
CCValAssign VA = RVLocs[i];
1192
if (VA.needsCustom()) {
1193
// Handle f64 or half of a v2f64.
1194
SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1196
Chain = Lo.getValue(1);
1197
InFlag = Lo.getValue(2);
1198
VA = RVLocs[++i]; // skip ahead to next loc
1199
SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
1201
Chain = Hi.getValue(1);
1202
InFlag = Hi.getValue(2);
1203
Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1205
if (VA.getLocVT() == MVT::v2f64) {
1206
SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
1207
Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1208
DAG.getConstant(0, MVT::i32));
1210
VA = RVLocs[++i]; // skip ahead to next loc
1211
Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1212
Chain = Lo.getValue(1);
1213
InFlag = Lo.getValue(2);
1214
VA = RVLocs[++i]; // skip ahead to next loc
1215
Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
1216
Chain = Hi.getValue(1);
1217
InFlag = Hi.getValue(2);
1218
Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
1219
Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
1220
DAG.getConstant(1, MVT::i32));
1223
Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1225
Chain = Val.getValue(1);
1226
InFlag = Val.getValue(2);
1229
switch (VA.getLocInfo()) {
1230
default: llvm_unreachable("Unknown loc info!");
1231
case CCValAssign::Full: break;
1232
case CCValAssign::BCvt:
1233
Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1237
InVals.push_back(Val);
1243
/// LowerMemOpCallTo - Store the argument to the stack.
1245
ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
1246
SDValue StackPtr, SDValue Arg,
1247
DebugLoc dl, SelectionDAG &DAG,
1248
const CCValAssign &VA,
1249
ISD::ArgFlagsTy Flags) const {
1250
unsigned LocMemOffset = VA.getLocMemOffset();
1251
SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1252
PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1253
return DAG.getStore(Chain, dl, Arg, PtrOff,
1254
MachinePointerInfo::getStack(LocMemOffset),
1258
void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
1259
SDValue Chain, SDValue &Arg,
1260
RegsToPassVector &RegsToPass,
1261
CCValAssign &VA, CCValAssign &NextVA,
1263
SmallVector<SDValue, 8> &MemOpChains,
1264
ISD::ArgFlagsTy Flags) const {
1266
SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1267
DAG.getVTList(MVT::i32, MVT::i32), Arg);
1268
RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
1270
if (NextVA.isRegLoc())
1271
RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
1273
assert(NextVA.isMemLoc());
1274
if (StackPtr.getNode() == 0)
1275
StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1277
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
1283
/// LowerCall - Lowering a call into a callseq_start <-
1284
/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
1287
ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1288
CallingConv::ID CallConv, bool isVarArg,
1289
bool doesNotRet, bool &isTailCall,
1290
const SmallVectorImpl<ISD::OutputArg> &Outs,
1291
const SmallVectorImpl<SDValue> &OutVals,
1292
const SmallVectorImpl<ISD::InputArg> &Ins,
1293
DebugLoc dl, SelectionDAG &DAG,
1294
SmallVectorImpl<SDValue> &InVals) const {
1295
MachineFunction &MF = DAG.getMachineFunction();
1296
bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
1297
bool IsSibCall = false;
1298
// Disable tail calls if they're not supported.
1299
if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
1302
// Check if it's really possible to do a tail call.
1303
isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1304
isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1305
Outs, OutVals, Ins, DAG);
1306
// We don't support GuaranteedTailCallOpt for ARM, only automatically
1307
// detected sibcalls.
1314
// Analyze operands of the call, assigning locations to each operand.
1315
SmallVector<CCValAssign, 16> ArgLocs;
1316
ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1317
getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1318
CCInfo.AnalyzeCallOperands(Outs,
1319
CCAssignFnForNode(CallConv, /* Return*/ false,
1322
// Get a count of how many bytes are to be pushed on the stack.
1323
unsigned NumBytes = CCInfo.getNextStackOffset();
1325
// For tail calls, memory operands are available in our caller's stack.
1329
// Adjust the stack pointer for the new arguments...
1330
// These operations are automatically eliminated by the prolog/epilog pass
1332
Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1334
SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());
1336
RegsToPassVector RegsToPass;
1337
SmallVector<SDValue, 8> MemOpChains;
1339
// Walk the register/memloc assignments, inserting copies/loads. In the case
1340
// of tail call optimization, arguments are handled later.
1341
for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1343
++i, ++realArgIdx) {
1344
CCValAssign &VA = ArgLocs[i];
1345
SDValue Arg = OutVals[realArgIdx];
1346
ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1347
bool isByVal = Flags.isByVal();
1349
// Promote the value if needed.
1350
switch (VA.getLocInfo()) {
1351
default: llvm_unreachable("Unknown loc info!");
1352
case CCValAssign::Full: break;
1353
case CCValAssign::SExt:
1354
Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1356
case CCValAssign::ZExt:
1357
Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1359
case CCValAssign::AExt:
1360
Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1362
case CCValAssign::BCvt:
1363
Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1367
// f64 and v2f64 might be passed in i32 pairs and must be split into pieces
1368
if (VA.needsCustom()) {
1369
if (VA.getLocVT() == MVT::v2f64) {
1370
SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1371
DAG.getConstant(0, MVT::i32));
1372
SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1373
DAG.getConstant(1, MVT::i32));
1375
PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
1376
VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1378
VA = ArgLocs[++i]; // skip ahead to next loc
1379
if (VA.isRegLoc()) {
1380
PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
1381
VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
1383
assert(VA.isMemLoc());
1385
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
1386
dl, DAG, VA, Flags));
1389
PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
1390
StackPtr, MemOpChains, Flags);
1392
} else if (VA.isRegLoc()) {
1393
RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1394
} else if (isByVal) {
1395
assert(VA.isMemLoc());
1396
unsigned offset = 0;
1398
// True if this byval aggregate will be split between registers
1400
if (CCInfo.isFirstByValRegValid()) {
1401
EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
1403
for (i = 0, j = CCInfo.getFirstByValReg(); j < ARM::R4; i++, j++) {
1404
SDValue Const = DAG.getConstant(4*i, MVT::i32);
1405
SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
1406
SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
1407
MachinePointerInfo(),
1408
false, false, false, 0);
1409
MemOpChains.push_back(Load.getValue(1));
1410
RegsToPass.push_back(std::make_pair(j, Load));
1412
offset = ARM::R4 - CCInfo.getFirstByValReg();
1413
CCInfo.clearFirstByValReg();
1416
unsigned LocMemOffset = VA.getLocMemOffset();
1417
SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
1418
SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
1420
SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
1421
SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
1422
SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
1424
MemOpChains.push_back(DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode,
1425
Flags.getByValAlign(),
1426
/*isVolatile=*/false,
1427
/*AlwaysInline=*/false,
1428
MachinePointerInfo(0),
1429
MachinePointerInfo(0)));
1431
} else if (!IsSibCall) {
1432
assert(VA.isMemLoc());
1434
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1435
dl, DAG, VA, Flags));
1439
if (!MemOpChains.empty())
1440
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1441
&MemOpChains[0], MemOpChains.size());
1443
// Build a sequence of copy-to-reg nodes chained together with token chain
1444
// and flag operands which copy the outgoing args into the appropriate regs.
1446
// Tail call byval lowering might overwrite argument registers so in case of
1447
// tail call optimization the copies to registers are lowered later.
1449
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1450
Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1451
RegsToPass[i].second, InFlag);
1452
InFlag = Chain.getValue(1);
1455
// For tail calls lower the arguments to the 'real' stack slot.
1457
// Force all the incoming stack arguments to be loaded from the stack
1458
// before any new outgoing arguments are stored to the stack, because the
1459
// outgoing stack slots may alias the incoming argument stack slots, and
1460
// the alias isn't otherwise explicit. This is slightly more conservative
1461
// than necessary, because it means that each store effectively depends
1462
// on every argument instead of just those arguments it would clobber.
1464
// Do not flag preceding copytoreg stuff together with the following stuff.
1466
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1467
Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1468
RegsToPass[i].second, InFlag);
1469
InFlag = Chain.getValue(1);
1474
// If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1475
// direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1476
// node so that legalize doesn't hack it.
1477
bool isDirect = false;
1478
bool isARMFunc = false;
1479
bool isLocalARMFunc = false;
1480
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
1482
if (EnableARMLongCalls) {
1483
assert (getTargetMachine().getRelocationModel() == Reloc::Static
1484
&& "long-calls with non-static relocation model!");
1485
// Handle a global address or an external symbol. If it's not one of
1486
// those, the target's already in a register, so we don't need to do
1488
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1489
const GlobalValue *GV = G->getGlobal();
1490
// Create a constant pool entry for the callee address
1491
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1492
ARMConstantPoolValue *CPV =
1493
ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);
1495
// Get the address of the callee into a register
1496
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1497
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1498
Callee = DAG.getLoad(getPointerTy(), dl,
1499
DAG.getEntryNode(), CPAddr,
1500
MachinePointerInfo::getConstantPool(),
1501
false, false, false, 0);
1502
} else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
1503
const char *Sym = S->getSymbol();
1505
// Create a constant pool entry for the callee address
1506
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1507
ARMConstantPoolValue *CPV =
1508
ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1509
ARMPCLabelIndex, 0);
1510
// Get the address of the callee into a register
1511
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1512
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1513
Callee = DAG.getLoad(getPointerTy(), dl,
1514
DAG.getEntryNode(), CPAddr,
1515
MachinePointerInfo::getConstantPool(),
1516
false, false, false, 0);
1518
} else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1519
const GlobalValue *GV = G->getGlobal();
1521
bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
1522
bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
1523
getTargetMachine().getRelocationModel() != Reloc::Static;
1524
isARMFunc = !Subtarget->isThumb() || isStub;
1525
// ARM call to a local ARM function is predicable.
1526
isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
1527
// tBX takes a register source operand.
1528
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1529
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1530
ARMConstantPoolValue *CPV =
1531
ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
1532
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1533
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1534
Callee = DAG.getLoad(getPointerTy(), dl,
1535
DAG.getEntryNode(), CPAddr,
1536
MachinePointerInfo::getConstantPool(),
1537
false, false, false, 0);
1538
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1539
Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1540
getPointerTy(), Callee, PICLabel);
1542
// On ELF targets for PIC code, direct calls should go through the PLT
1543
unsigned OpFlags = 0;
1544
if (Subtarget->isTargetELF() &&
1545
getTargetMachine().getRelocationModel() == Reloc::PIC_)
1546
OpFlags = ARMII::MO_PLT;
1547
Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
1549
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1551
bool isStub = Subtarget->isTargetDarwin() &&
1552
getTargetMachine().getRelocationModel() != Reloc::Static;
1553
isARMFunc = !Subtarget->isThumb() || isStub;
1554
// tBX takes a register source operand.
1555
const char *Sym = S->getSymbol();
1556
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
1557
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
1558
ARMConstantPoolValue *CPV =
1559
ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
1560
ARMPCLabelIndex, 4);
1561
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
1562
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
1563
Callee = DAG.getLoad(getPointerTy(), dl,
1564
DAG.getEntryNode(), CPAddr,
1565
MachinePointerInfo::getConstantPool(),
1566
false, false, false, 0);
1567
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
1568
Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
1569
getPointerTy(), Callee, PICLabel);
1571
unsigned OpFlags = 0;
1572
// On ELF targets for PIC code, direct calls should go through the PLT
1573
if (Subtarget->isTargetELF() &&
1574
getTargetMachine().getRelocationModel() == Reloc::PIC_)
1575
OpFlags = ARMII::MO_PLT;
1576
Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
1580
// FIXME: handle tail calls differently.
1582
if (Subtarget->isThumb()) {
1583
if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
1584
CallOpc = ARMISD::CALL_NOLINK;
1585
else if (doesNotRet && isDirect && !isARMFunc &&
1586
Subtarget->hasRAS() && !Subtarget->isThumb1Only())
1587
// "mov lr, pc; b _foo" to avoid confusing the RSP
1588
CallOpc = ARMISD::CALL_NOLINK;
1590
CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
1592
if (!isDirect && !Subtarget->hasV5TOps()) {
1593
CallOpc = ARMISD::CALL_NOLINK;
1594
} else if (doesNotRet && isDirect && Subtarget->hasRAS())
1595
// "mov lr, pc; b _foo" to avoid confusing the RSP
1596
CallOpc = ARMISD::CALL_NOLINK;
1598
CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
1601
std::vector<SDValue> Ops;
1602
Ops.push_back(Chain);
1603
Ops.push_back(Callee);
1605
// Add argument registers to the end of the list so that they are known live
1607
for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1608
Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1609
RegsToPass[i].second.getValueType()));
1611
// Add a register mask operand representing the call-preserved registers.
1612
const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1613
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1614
assert(Mask && "Missing call preserved mask for calling convention");
1615
Ops.push_back(DAG.getRegisterMask(Mask));
1617
if (InFlag.getNode())
1618
Ops.push_back(InFlag);
1620
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1622
return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1624
// Returns a chain and a flag for retval copy to use.
1625
Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
1626
InFlag = Chain.getValue(1);
1628
Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1629
DAG.getIntPtrConstant(0, true), InFlag);
1631
InFlag = Chain.getValue(1);
1633
// Handle result values, copying them out of physregs into vregs that we
1635
return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins,
1639
/// HandleByVal - Every parameter *after* a byval parameter is passed
1640
/// on the stack. Remember the next parameter register to allocate,
1641
/// and then confiscate the rest of the parameter registers to insure
1644
ARMTargetLowering::HandleByVal(CCState *State, unsigned &size) const {
1645
unsigned reg = State->AllocateReg(GPRArgRegs, 4);
1646
assert((State->getCallOrPrologue() == Prologue ||
1647
State->getCallOrPrologue() == Call) &&
1648
"unhandled ParmContext");
1649
if ((!State->isFirstByValRegValid()) &&
1650
(ARM::R0 <= reg) && (reg <= ARM::R3)) {
1651
State->setFirstByValReg(reg);
1652
// At a call site, a byval parameter that is split between
1653
// registers and memory needs its size truncated here. In a
1654
// function prologue, such byval parameters are reassembled in
1655
// memory, and are not truncated.
1656
if (State->getCallOrPrologue() == Call) {
1657
unsigned excess = 4 * (ARM::R4 - reg);
1658
assert(size >= excess && "expected larger existing stack allocation");
1662
// Confiscate any remaining parameter registers to preclude their
1663
// assignment to subsequent parameters.
1664
while (State->AllocateReg(GPRArgRegs, 4))
1668
/// MatchingStackOffset - Return true if the given stack call argument is
1669
/// already available in the same position (relatively) of the caller's
1670
/// incoming argument stack.
1672
bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
1673
MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
1674
const TargetInstrInfo *TII) {
1675
unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
1677
if (Arg.getOpcode() == ISD::CopyFromReg) {
1678
unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
1679
if (!TargetRegisterInfo::isVirtualRegister(VR))
1681
MachineInstr *Def = MRI->getVRegDef(VR);
1684
if (!Flags.isByVal()) {
1685
if (!TII->isLoadFromStackSlot(Def, FI))
1690
} else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
1691
if (Flags.isByVal())
1692
// ByVal argument is passed in as a pointer but it's now being
1693
// dereferenced. e.g.
1694
// define @foo(%struct.X* %A) {
1695
// tail call @bar(%struct.X* byval %A)
1698
SDValue Ptr = Ld->getBasePtr();
1699
FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
1702
FI = FINode->getIndex();
1706
assert(FI != INT_MAX);
1707
if (!MFI->isFixedObjectIndex(FI))
1709
return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
1712
/// IsEligibleForTailCallOptimization - Check whether the call is eligible
1713
/// for tail call optimization. Targets which want to do tail call
1714
/// optimization should implement this function.
1716
ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1717
CallingConv::ID CalleeCC,
1719
bool isCalleeStructRet,
1720
bool isCallerStructRet,
1721
const SmallVectorImpl<ISD::OutputArg> &Outs,
1722
const SmallVectorImpl<SDValue> &OutVals,
1723
const SmallVectorImpl<ISD::InputArg> &Ins,
1724
SelectionDAG& DAG) const {
1725
const Function *CallerF = DAG.getMachineFunction().getFunction();
1726
CallingConv::ID CallerCC = CallerF->getCallingConv();
1727
bool CCMatch = CallerCC == CalleeCC;
1729
// Look for obvious safe cases to perform tail call optimization that do not
1730
// require ABI changes. This is what gcc calls sibcall.
1732
// Do not sibcall optimize vararg calls unless the call site is not passing
1734
if (isVarArg && !Outs.empty())
1737
// Also avoid sibcall optimization if either caller or callee uses struct
1738
// return semantics.
1739
if (isCalleeStructRet || isCallerStructRet)
1742
// FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
1743
// emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
1744
// the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
1745
// support in the assembler and linker to be used. This would need to be
1746
// fixed to fully support tail calls in Thumb1.
1748
// Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
1749
// LR. This means if we need to reload LR, it takes an extra instructions,
1750
// which outweighs the value of the tail call; but here we don't know yet
1751
// whether LR is going to be used. Probably the right approach is to
1752
// generate the tail call here and turn it back into CALL/RET in
1753
// emitEpilogue if LR is used.
1755
// Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
1756
// but we need to make sure there are enough registers; the only valid
1757
// registers are the 4 used for parameters. We don't currently do this
1759
if (Subtarget->isThumb1Only())
1762
// If the calling conventions do not match, then we'd better make sure the
1763
// results are returned in the same way as what the caller expects.
1765
SmallVector<CCValAssign, 16> RVLocs1;
1766
ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1767
getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
1768
CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));
1770
SmallVector<CCValAssign, 16> RVLocs2;
1771
ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1772
getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
1773
CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));
1775
if (RVLocs1.size() != RVLocs2.size())
1777
for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1778
if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1780
if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1782
if (RVLocs1[i].isRegLoc()) {
1783
if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1786
if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1792
// If the callee takes no arguments then go on to check the results of the
1794
if (!Outs.empty()) {
1795
// Check if stack adjustment is needed. For now, do not do this if any
1796
// argument is passed on the stack.
1797
SmallVector<CCValAssign, 16> ArgLocs;
1798
ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
1799
getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
1800
CCInfo.AnalyzeCallOperands(Outs,
1801
CCAssignFnForNode(CalleeCC, false, isVarArg));
1802
if (CCInfo.getNextStackOffset()) {
1803
MachineFunction &MF = DAG.getMachineFunction();
1805
// Check if the arguments are already laid out in the right way as
1806
// the caller's fixed stack objects.
1807
MachineFrameInfo *MFI = MF.getFrameInfo();
1808
const MachineRegisterInfo *MRI = &MF.getRegInfo();
1809
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
1810
for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
1812
++i, ++realArgIdx) {
1813
CCValAssign &VA = ArgLocs[i];
1814
EVT RegVT = VA.getLocVT();
1815
SDValue Arg = OutVals[realArgIdx];
1816
ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
1817
if (VA.getLocInfo() == CCValAssign::Indirect)
1819
if (VA.needsCustom()) {
1820
// f64 and vector types are split into multiple registers or
1821
// register/stack-slot combinations. The types will not match
1822
// the registers; give up on memory f64 refs until we figure
1823
// out what to do about this.
1826
if (!ArgLocs[++i].isRegLoc())
1828
if (RegVT == MVT::v2f64) {
1829
if (!ArgLocs[++i].isRegLoc())
1831
if (!ArgLocs[++i].isRegLoc())
1834
} else if (!VA.isRegLoc()) {
1835
if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
1847
ARMTargetLowering::LowerReturn(SDValue Chain,
1848
CallingConv::ID CallConv, bool isVarArg,
1849
const SmallVectorImpl<ISD::OutputArg> &Outs,
1850
const SmallVectorImpl<SDValue> &OutVals,
1851
DebugLoc dl, SelectionDAG &DAG) const {
1853
// CCValAssign - represent the assignment of the return value to a location.
1854
SmallVector<CCValAssign, 16> RVLocs;
1856
// CCState - Info about the registers and stack slots.
1857
ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1858
getTargetMachine(), RVLocs, *DAG.getContext(), Call);
1860
// Analyze outgoing return values.
1861
CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
1864
// If this is the first return lowered for this function, add
1865
// the regs to the liveout set for the function.
1866
if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1867
for (unsigned i = 0; i != RVLocs.size(); ++i)
1868
if (RVLocs[i].isRegLoc())
1869
DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1874
// Copy the result values into the output registers.
1875
for (unsigned i = 0, realRVLocIdx = 0;
1877
++i, ++realRVLocIdx) {
1878
CCValAssign &VA = RVLocs[i];
1879
assert(VA.isRegLoc() && "Can only return in registers!");
1881
SDValue Arg = OutVals[realRVLocIdx];
1883
switch (VA.getLocInfo()) {
1884
default: llvm_unreachable("Unknown loc info!");
1885
case CCValAssign::Full: break;
1886
case CCValAssign::BCvt:
1887
Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1891
if (VA.needsCustom()) {
1892
if (VA.getLocVT() == MVT::v2f64) {
1893
// Extract the first half and return it in two registers.
1894
SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1895
DAG.getConstant(0, MVT::i32));
1896
SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
1897
DAG.getVTList(MVT::i32, MVT::i32), Half);
1899
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
1900
Flag = Chain.getValue(1);
1901
VA = RVLocs[++i]; // skip ahead to next loc
1902
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
1903
HalfGPRs.getValue(1), Flag);
1904
Flag = Chain.getValue(1);
1905
VA = RVLocs[++i]; // skip ahead to next loc
1907
// Extract the 2nd half and fall through to handle it as an f64 value.
1908
Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
1909
DAG.getConstant(1, MVT::i32));
1911
// Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
1913
SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
1914
DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
1915
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
1916
Flag = Chain.getValue(1);
1917
VA = RVLocs[++i]; // skip ahead to next loc
1918
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
1921
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1923
// Guarantee that all emitted copies are
1924
// stuck together, avoiding something bad.
1925
Flag = Chain.getValue(1);
1930
result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
1932
result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain);
1937
bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N) const {
1938
if (N->getNumValues() != 1)
1940
if (!N->hasNUsesOfValue(1, 0))
1943
unsigned NumCopies = 0;
1944
SDNode* Copies[2] = { 0, 0 };
1945
SDNode *Use = *N->use_begin();
1946
if (Use->getOpcode() == ISD::CopyToReg) {
1947
Copies[NumCopies++] = Use;
1948
} else if (Use->getOpcode() == ARMISD::VMOVRRD) {
1949
// f64 returned in a pair of GPRs.
1950
for (SDNode::use_iterator UI = Use->use_begin(), UE = Use->use_end();
1952
if (UI->getOpcode() != ISD::CopyToReg)
1954
Copies[UI.getUse().getResNo()] = *UI;
1957
} else if (Use->getOpcode() == ISD::BITCAST) {
1958
// f32 returned in a single GPR.
1959
if (!Use->hasNUsesOfValue(1, 0))
1961
Use = *Use->use_begin();
1962
if (Use->getOpcode() != ISD::CopyToReg || !Use->hasNUsesOfValue(1, 0))
1964
Copies[NumCopies++] = Use;
1969
if (NumCopies != 1 && NumCopies != 2)
1972
bool HasRet = false;
1973
for (unsigned i = 0; i < NumCopies; ++i) {
1974
SDNode *Copy = Copies[i];
1975
for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1977
if (UI->getOpcode() == ISD::CopyToReg) {
1979
if (Use == Copies[0] || ((NumCopies == 2) && (Use == Copies[1])))
1983
if (UI->getOpcode() != ARMISD::RET_FLAG)
1992
bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1993
if (!EnableARMTailCalls)
1996
if (!CI->isTailCall())
1999
return !Subtarget->isThumb1Only();
2002
// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
2003
// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
2004
// one of the above mentioned nodes. It has to be wrapped because otherwise
2005
// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
2006
// be used to form addressing mode. These wrapped nodes will be selected
2008
static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
2009
EVT PtrVT = Op.getValueType();
2010
// FIXME there is no actual debug info here
2011
DebugLoc dl = Op.getDebugLoc();
2012
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
2014
if (CP->isMachineConstantPoolEntry())
2015
Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
2016
CP->getAlignment());
2018
Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
2019
CP->getAlignment());
2020
return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
2023
unsigned ARMTargetLowering::getJumpTableEncoding() const {
2024
return MachineJumpTableInfo::EK_Inline;
2027
SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
2028
SelectionDAG &DAG) const {
2029
MachineFunction &MF = DAG.getMachineFunction();
2030
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2031
unsigned ARMPCLabelIndex = 0;
2032
DebugLoc DL = Op.getDebugLoc();
2033
EVT PtrVT = getPointerTy();
2034
const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2035
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2037
if (RelocM == Reloc::Static) {
2038
CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
2040
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2041
ARMPCLabelIndex = AFI->createPICLabelUId();
2042
ARMConstantPoolValue *CPV =
2043
ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
2044
ARMCP::CPBlockAddress, PCAdj);
2045
CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2047
CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
2048
SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
2049
MachinePointerInfo::getConstantPool(),
2050
false, false, false, 0);
2051
if (RelocM == Reloc::Static)
2053
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2054
return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
2057
// Lower ISD::GlobalTLSAddress using the "general dynamic" model
2059
ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
2060
SelectionDAG &DAG) const {
2061
DebugLoc dl = GA->getDebugLoc();
2062
EVT PtrVT = getPointerTy();
2063
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2064
MachineFunction &MF = DAG.getMachineFunction();
2065
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2066
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2067
ARMConstantPoolValue *CPV =
2068
ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2069
ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
2070
SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2071
Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
2072
Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
2073
MachinePointerInfo::getConstantPool(),
2074
false, false, false, 0);
2075
SDValue Chain = Argument.getValue(1);
2077
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2078
Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
2080
// call __tls_get_addr.
2083
Entry.Node = Argument;
2084
Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
2085
Args.push_back(Entry);
2086
// FIXME: is there useful debug info available here?
2087
std::pair<SDValue, SDValue> CallResult =
2088
LowerCallTo(Chain, (Type *) Type::getInt32Ty(*DAG.getContext()),
2089
false, false, false, false,
2090
0, CallingConv::C, /*isTailCall=*/false,
2091
/*doesNotRet=*/false, /*isReturnValueUsed=*/true,
2092
DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
2093
return CallResult.first;
2096
// Lower ISD::GlobalTLSAddress using the "initial exec" or
2097
// "local exec" model.
2099
ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
2100
SelectionDAG &DAG) const {
2101
const GlobalValue *GV = GA->getGlobal();
2102
DebugLoc dl = GA->getDebugLoc();
2104
SDValue Chain = DAG.getEntryNode();
2105
EVT PtrVT = getPointerTy();
2106
// Get the Thread Pointer
2107
SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2109
if (GV->isDeclaration()) {
2110
MachineFunction &MF = DAG.getMachineFunction();
2111
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2112
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2113
// Initial exec model.
2114
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
2115
ARMConstantPoolValue *CPV =
2116
ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
2117
ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
2119
Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2120
Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2121
Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2122
MachinePointerInfo::getConstantPool(),
2123
false, false, false, 0);
2124
Chain = Offset.getValue(1);
2126
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2127
Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
2129
Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2130
MachinePointerInfo::getConstantPool(),
2131
false, false, false, 0);
2134
ARMConstantPoolValue *CPV =
2135
ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
2136
Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2137
Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
2138
Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
2139
MachinePointerInfo::getConstantPool(),
2140
false, false, false, 0);
2143
// The address of the thread local variable is the add of the thread
2144
// pointer with the offset of the variable.
2145
return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
2149
ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
2150
// TODO: implement the "local dynamic" model
2151
assert(Subtarget->isTargetELF() &&
2152
"TLS not implemented for non-ELF targets");
2153
GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2154
// If the relocation model is PIC, use the "General Dynamic" TLS Model,
2155
// otherwise use the "Local Exec" TLS Model
2156
if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
2157
return LowerToTLSGeneralDynamicModel(GA, DAG);
2159
return LowerToTLSExecModels(GA, DAG);
2162
SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
2163
SelectionDAG &DAG) const {
2164
EVT PtrVT = getPointerTy();
2165
DebugLoc dl = Op.getDebugLoc();
2166
const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2167
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2168
if (RelocM == Reloc::PIC_) {
2169
bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
2170
ARMConstantPoolValue *CPV =
2171
ARMConstantPoolConstant::Create(GV,
2172
UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
2173
SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2174
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2175
SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
2177
MachinePointerInfo::getConstantPool(),
2178
false, false, false, 0);
2179
SDValue Chain = Result.getValue(1);
2180
SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
2181
Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
2183
Result = DAG.getLoad(PtrVT, dl, Chain, Result,
2184
MachinePointerInfo::getGOT(),
2185
false, false, false, 0);
2189
// If we have T2 ops, we can materialize the address directly via movt/movw
2190
// pair. This is always cheaper.
2191
if (Subtarget->useMovt()) {
2193
// FIXME: Once remat is capable of dealing with instructions with register
2194
// operands, expand this into two nodes.
2195
return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2196
DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2198
SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2199
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2200
return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2201
MachinePointerInfo::getConstantPool(),
2202
false, false, false, 0);
2206
SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
2207
SelectionDAG &DAG) const {
2208
EVT PtrVT = getPointerTy();
2209
DebugLoc dl = Op.getDebugLoc();
2210
const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
2211
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2212
MachineFunction &MF = DAG.getMachineFunction();
2213
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2215
// FIXME: Enable this for static codegen when tool issues are fixed. Also
2216
// update ARMFastISel::ARMMaterializeGV.
2217
if (Subtarget->useMovt() && RelocM != Reloc::Static) {
2219
// FIXME: Once remat is capable of dealing with instructions with register
2220
// operands, expand this into two nodes.
2221
if (RelocM == Reloc::Static)
2222
return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
2223
DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2225
unsigned Wrapper = (RelocM == Reloc::PIC_)
2226
? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
2227
SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
2228
DAG.getTargetGlobalAddress(GV, dl, PtrVT));
2229
if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2230
Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
2231
MachinePointerInfo::getGOT(),
2232
false, false, false, 0);
2236
unsigned ARMPCLabelIndex = 0;
2238
if (RelocM == Reloc::Static) {
2239
CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
2241
ARMPCLabelIndex = AFI->createPICLabelUId();
2242
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
2243
ARMConstantPoolValue *CPV =
2244
ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
2246
CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2248
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2250
SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2251
MachinePointerInfo::getConstantPool(),
2252
false, false, false, 0);
2253
SDValue Chain = Result.getValue(1);
2255
if (RelocM == Reloc::PIC_) {
2256
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2257
Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2260
if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
2261
Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
2262
false, false, false, 0);
2267
SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
2268
SelectionDAG &DAG) const {
2269
assert(Subtarget->isTargetELF() &&
2270
"GLOBAL OFFSET TABLE not implemented for non-ELF targets");
2271
MachineFunction &MF = DAG.getMachineFunction();
2272
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2273
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2274
EVT PtrVT = getPointerTy();
2275
DebugLoc dl = Op.getDebugLoc();
2276
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
2277
ARMConstantPoolValue *CPV =
2278
ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
2279
ARMPCLabelIndex, PCAdj);
2280
SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2281
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2282
SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2283
MachinePointerInfo::getConstantPool(),
2284
false, false, false, 0);
2285
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2286
return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2290
ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
2291
DebugLoc dl = Op.getDebugLoc();
2292
SDValue Val = DAG.getConstant(0, MVT::i32);
2293
return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
2294
DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
2295
Op.getOperand(1), Val);
2299
ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
2300
DebugLoc dl = Op.getDebugLoc();
2301
return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
2302
Op.getOperand(1), DAG.getConstant(0, MVT::i32));
2306
ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
2307
const ARMSubtarget *Subtarget) const {
2308
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2309
DebugLoc dl = Op.getDebugLoc();
2311
default: return SDValue(); // Don't custom lower most intrinsics.
2312
case Intrinsic::arm_thread_pointer: {
2313
EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2314
return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
2316
case Intrinsic::eh_sjlj_lsda: {
2317
MachineFunction &MF = DAG.getMachineFunction();
2318
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2319
unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
2320
EVT PtrVT = getPointerTy();
2321
DebugLoc dl = Op.getDebugLoc();
2322
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2324
unsigned PCAdj = (RelocM != Reloc::PIC_)
2325
? 0 : (Subtarget->isThumb() ? 4 : 8);
2326
ARMConstantPoolValue *CPV =
2327
ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
2328
ARMCP::CPLSDA, PCAdj);
2329
CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
2330
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
2332
DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
2333
MachinePointerInfo::getConstantPool(),
2334
false, false, false, 0);
2336
if (RelocM == Reloc::PIC_) {
2337
SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
2338
Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
2342
case Intrinsic::arm_neon_vmulls:
2343
case Intrinsic::arm_neon_vmullu: {
2344
unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
2345
? ARMISD::VMULLs : ARMISD::VMULLu;
2346
return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
2347
Op.getOperand(1), Op.getOperand(2));
2352
static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
2353
const ARMSubtarget *Subtarget) {
2354
DebugLoc dl = Op.getDebugLoc();
2355
if (!Subtarget->hasDataBarrier()) {
2356
// Some ARMv6 cpus can support data barriers with an mcr instruction.
2357
// Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2359
assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2360
"Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2361
return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2362
DAG.getConstant(0, MVT::i32));
2365
SDValue Op5 = Op.getOperand(5);
2366
bool isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue() != 0;
2367
unsigned isLL = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
2368
unsigned isLS = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
2369
bool isOnlyStoreBarrier = (isLL == 0 && isLS == 0);
2371
ARM_MB::MemBOpt DMBOpt;
2372
if (isDeviceBarrier)
2373
DMBOpt = isOnlyStoreBarrier ? ARM_MB::ST : ARM_MB::SY;
2375
DMBOpt = isOnlyStoreBarrier ? ARM_MB::ISHST : ARM_MB::ISH;
2376
return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2377
DAG.getConstant(DMBOpt, MVT::i32));
2381
static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
2382
const ARMSubtarget *Subtarget) {
2383
// FIXME: handle "fence singlethread" more efficiently.
2384
DebugLoc dl = Op.getDebugLoc();
2385
if (!Subtarget->hasDataBarrier()) {
2386
// Some ARMv6 cpus can support data barriers with an mcr instruction.
2387
// Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
2389
assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
2390
"Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
2391
return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
2392
DAG.getConstant(0, MVT::i32));
2395
return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
2396
DAG.getConstant(ARM_MB::ISH, MVT::i32));
2399
static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
2400
const ARMSubtarget *Subtarget) {
2401
// ARM pre v5TE and Thumb1 does not have preload instructions.
2402
if (!(Subtarget->isThumb2() ||
2403
(!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
2404
// Just preserve the chain.
2405
return Op.getOperand(0);
2407
DebugLoc dl = Op.getDebugLoc();
2408
unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
2410
(!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
2411
// ARMv7 with MP extension has PLDW.
2412
return Op.getOperand(0);
2414
unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
2415
if (Subtarget->isThumb()) {
2417
isRead = ~isRead & 1;
2418
isData = ~isData & 1;
2421
return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
2422
Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
2423
DAG.getConstant(isData, MVT::i32));
2426
static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
2427
MachineFunction &MF = DAG.getMachineFunction();
2428
ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();
2430
// vastart just stores the address of the VarArgsFrameIndex slot into the
2431
// memory location argument.
2432
DebugLoc dl = Op.getDebugLoc();
2433
EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
2434
SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
2435
const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2436
return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
2437
MachinePointerInfo(SV), false, false, 0);
2441
ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
2442
SDValue &Root, SelectionDAG &DAG,
2443
DebugLoc dl) const {
2444
MachineFunction &MF = DAG.getMachineFunction();
2445
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2447
const TargetRegisterClass *RC;
2448
if (AFI->isThumb1OnlyFunction())
2449
RC = ARM::tGPRRegisterClass;
2451
RC = ARM::GPRRegisterClass;
2453
// Transform the arguments stored in physical registers into virtual ones.
2454
unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2455
SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2458
if (NextVA.isMemLoc()) {
2459
MachineFrameInfo *MFI = MF.getFrameInfo();
2460
int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);
2462
// Create load node to retrieve arguments from the stack.
2463
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2464
ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
2465
MachinePointerInfo::getFixedStack(FI),
2466
false, false, false, 0);
2468
Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
2469
ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
2472
return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
2476
ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
2477
unsigned &VARegSize, unsigned &VARegSaveSize)
2480
if (CCInfo.isFirstByValRegValid())
2481
NumGPRs = ARM::R4 - CCInfo.getFirstByValReg();
2483
unsigned int firstUnalloced;
2484
firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
2485
sizeof(GPRArgRegs) /
2486
sizeof(GPRArgRegs[0]));
2487
NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
2490
unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
2491
VARegSize = NumGPRs * 4;
2492
VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
2495
// The remaining GPRs hold either the beginning of variable-argument
2496
// data, or the beginning of an aggregate passed by value (usuall
2497
// byval). Either way, we allocate stack slots adjacent to the data
2498
// provided by our caller, and store the unallocated registers there.
2499
// If this is a variadic function, the va_list pointer will begin with
2500
// these values; otherwise, this reassembles a (byval) structure that
2501
// was split between registers and memory.
2503
ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
2504
DebugLoc dl, SDValue &Chain,
2505
unsigned ArgOffset) const {
2506
MachineFunction &MF = DAG.getMachineFunction();
2507
MachineFrameInfo *MFI = MF.getFrameInfo();
2508
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2509
unsigned firstRegToSaveIndex;
2510
if (CCInfo.isFirstByValRegValid())
2511
firstRegToSaveIndex = CCInfo.getFirstByValReg() - ARM::R0;
2513
firstRegToSaveIndex = CCInfo.getFirstUnallocated
2514
(GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
2517
unsigned VARegSize, VARegSaveSize;
2518
computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2519
if (VARegSaveSize) {
2520
// If this function is vararg, store any remaining integer argument regs
2521
// to their spots on the stack so that they may be loaded by deferencing
2522
// the result of va_next.
2523
AFI->setVarArgsRegSaveSize(VARegSaveSize);
2524
AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(VARegSaveSize,
2525
ArgOffset + VARegSaveSize
2528
SDValue FIN = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(),
2531
SmallVector<SDValue, 4> MemOps;
2532
for (; firstRegToSaveIndex < 4; ++firstRegToSaveIndex) {
2533
const TargetRegisterClass *RC;
2534
if (AFI->isThumb1OnlyFunction())
2535
RC = ARM::tGPRRegisterClass;
2537
RC = ARM::GPRRegisterClass;
2539
unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
2540
SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
2542
DAG.getStore(Val.getValue(1), dl, Val, FIN,
2543
MachinePointerInfo::getFixedStack(AFI->getVarArgsFrameIndex()),
2545
MemOps.push_back(Store);
2546
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
2547
DAG.getConstant(4, getPointerTy()));
2549
if (!MemOps.empty())
2550
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2551
&MemOps[0], MemOps.size());
2553
// This will point to the next argument passed via stack.
2554
AFI->setVarArgsFrameIndex(MFI->CreateFixedObject(4, ArgOffset, true));
2558
ARMTargetLowering::LowerFormalArguments(SDValue Chain,
2559
CallingConv::ID CallConv, bool isVarArg,
2560
const SmallVectorImpl<ISD::InputArg>
2562
DebugLoc dl, SelectionDAG &DAG,
2563
SmallVectorImpl<SDValue> &InVals)
2565
MachineFunction &MF = DAG.getMachineFunction();
2566
MachineFrameInfo *MFI = MF.getFrameInfo();
2568
ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
2570
// Assign locations to all of the incoming arguments.
2571
SmallVector<CCValAssign, 16> ArgLocs;
2572
ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
2573
getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
2574
CCInfo.AnalyzeFormalArguments(Ins,
2575
CCAssignFnForNode(CallConv, /* Return*/ false,
2578
SmallVector<SDValue, 16> ArgValues;
2579
int lastInsIndex = -1;
2582
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2583
CCValAssign &VA = ArgLocs[i];
2585
// Arguments stored in registers.
2586
if (VA.isRegLoc()) {
2587
EVT RegVT = VA.getLocVT();
2589
if (VA.needsCustom()) {
2590
// f64 and vector types are split up into multiple registers or
2591
// combinations of registers and stack slots.
2592
if (VA.getLocVT() == MVT::v2f64) {
2593
SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
2595
VA = ArgLocs[++i]; // skip ahead to next loc
2597
if (VA.isMemLoc()) {
2598
int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
2599
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2600
ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
2601
MachinePointerInfo::getFixedStack(FI),
2602
false, false, false, 0);
2604
ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
2607
ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
2608
ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2609
ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
2610
ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
2611
ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
2613
ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);
2616
const TargetRegisterClass *RC;
2618
if (RegVT == MVT::f32)
2619
RC = ARM::SPRRegisterClass;
2620
else if (RegVT == MVT::f64)
2621
RC = ARM::DPRRegisterClass;
2622
else if (RegVT == MVT::v2f64)
2623
RC = ARM::QPRRegisterClass;
2624
else if (RegVT == MVT::i32)
2625
RC = (AFI->isThumb1OnlyFunction() ?
2626
ARM::tGPRRegisterClass : ARM::GPRRegisterClass);
2628
llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");
2630
// Transform the arguments in physical registers into virtual ones.
2631
unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2632
ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2635
// If this is an 8 or 16-bit value, it is really passed promoted
2636
// to 32 bits. Insert an assert[sz]ext to capture this, then
2637
// truncate to the right size.
2638
switch (VA.getLocInfo()) {
2639
default: llvm_unreachable("Unknown loc info!");
2640
case CCValAssign::Full: break;
2641
case CCValAssign::BCvt:
2642
ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2644
case CCValAssign::SExt:
2645
ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2646
DAG.getValueType(VA.getValVT()));
2647
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2649
case CCValAssign::ZExt:
2650
ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2651
DAG.getValueType(VA.getValVT()));
2652
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2656
InVals.push_back(ArgValue);
2658
} else { // VA.isRegLoc()
2661
assert(VA.isMemLoc());
2662
assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
2664
int index = ArgLocs[i].getValNo();
2666
// Some Ins[] entries become multiple ArgLoc[] entries.
2667
// Process them only once.
2668
if (index != lastInsIndex)
2670
ISD::ArgFlagsTy Flags = Ins[index].Flags;
2671
// FIXME: For now, all byval parameter objects are marked mutable.
2672
// This can be changed with more analysis.
2673
// In case of tail call optimization mark all arguments mutable.
2674
// Since they could be overwritten by lowering of arguments in case of
2676
if (Flags.isByVal()) {
2677
unsigned VARegSize, VARegSaveSize;
2678
computeRegArea(CCInfo, MF, VARegSize, VARegSaveSize);
2679
VarArgStyleRegisters(CCInfo, DAG, dl, Chain, 0);
2680
unsigned Bytes = Flags.getByValSize() - VARegSize;
2681
if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2682
int FI = MFI->CreateFixedObject(Bytes,
2683
VA.getLocMemOffset(), false);
2684
InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
2686
int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
2687
VA.getLocMemOffset(), true);
2689
// Create load nodes to retrieve arguments from the stack.
2690
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2691
InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
2692
MachinePointerInfo::getFixedStack(FI),
2693
false, false, false, 0));
2695
lastInsIndex = index;
2702
VarArgStyleRegisters(CCInfo, DAG, dl, Chain, CCInfo.getNextStackOffset());
2707
/// isFloatingPointZero - Return true if this is +0.0.
2708
static bool isFloatingPointZero(SDValue Op) {
2709
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
2710
return CFP->getValueAPF().isPosZero();
2711
else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
2712
// Maybe this has already been legalized into the constant pool?
2713
if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
2714
SDValue WrapperOp = Op.getOperand(1).getOperand(0);
2715
if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
2716
if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
2717
return CFP->getValueAPF().isPosZero();
2723
/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
2724
/// the given operands.
2726
ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
2727
SDValue &ARMcc, SelectionDAG &DAG,
2728
DebugLoc dl) const {
2729
if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2730
unsigned C = RHSC->getZExtValue();
2731
if (!isLegalICmpImmediate(C)) {
2732
// Constant does not fit, try adjusting it by one?
2737
if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
2738
CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2739
RHS = DAG.getConstant(C-1, MVT::i32);
2744
if (C != 0 && isLegalICmpImmediate(C-1)) {
2745
CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2746
RHS = DAG.getConstant(C-1, MVT::i32);
2751
if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
2752
CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2753
RHS = DAG.getConstant(C+1, MVT::i32);
2758
if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
2759
CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2760
RHS = DAG.getConstant(C+1, MVT::i32);
2767
ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
2768
ARMISD::NodeType CompareType;
2771
CompareType = ARMISD::CMP;
2776
CompareType = ARMISD::CMPZ;
2779
ARMcc = DAG.getConstant(CondCode, MVT::i32);
2780
return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
2783
/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
2785
ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
2786
DebugLoc dl) const {
2788
if (!isFloatingPointZero(RHS))
2789
Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
2791
Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
2792
return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
2795
/// duplicateCmp - Glue values can have only one use, so this function
2796
/// duplicates a comparison node.
2798
ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
2799
unsigned Opc = Cmp.getOpcode();
2800
DebugLoc DL = Cmp.getDebugLoc();
2801
if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
2802
return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2804
assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
2805
Cmp = Cmp.getOperand(0);
2806
Opc = Cmp.getOpcode();
2807
if (Opc == ARMISD::CMPFP)
2808
Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
2810
assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
2811
Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
2813
return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
2816
SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2817
SDValue Cond = Op.getOperand(0);
2818
SDValue SelectTrue = Op.getOperand(1);
2819
SDValue SelectFalse = Op.getOperand(2);
2820
DebugLoc dl = Op.getDebugLoc();
2824
// (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
2825
// (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
2827
if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
2828
const ConstantSDNode *CMOVTrue =
2829
dyn_cast<ConstantSDNode>(Cond.getOperand(0));
2830
const ConstantSDNode *CMOVFalse =
2831
dyn_cast<ConstantSDNode>(Cond.getOperand(1));
2833
if (CMOVTrue && CMOVFalse) {
2834
unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
2835
unsigned CMOVFalseVal = CMOVFalse->getZExtValue();
2839
if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
2841
False = SelectFalse;
2842
} else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
2847
if (True.getNode() && False.getNode()) {
2848
EVT VT = Op.getValueType();
2849
SDValue ARMcc = Cond.getOperand(2);
2850
SDValue CCR = Cond.getOperand(3);
2851
SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
2852
assert(True.getValueType() == VT);
2853
return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
2858
// ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
2859
// undefined bits before doing a full-word comparison with zero.
2860
Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
2861
DAG.getConstant(1, Cond.getValueType()));
2863
return DAG.getSelectCC(dl, Cond,
2864
DAG.getConstant(0, Cond.getValueType()),
2865
SelectTrue, SelectFalse, ISD::SETNE);
2868
SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2869
EVT VT = Op.getValueType();
2870
SDValue LHS = Op.getOperand(0);
2871
SDValue RHS = Op.getOperand(1);
2872
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2873
SDValue TrueVal = Op.getOperand(2);
2874
SDValue FalseVal = Op.getOperand(3);
2875
DebugLoc dl = Op.getDebugLoc();
2877
if (LHS.getValueType() == MVT::i32) {
2879
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2880
SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
2881
return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
2884
ARMCC::CondCodes CondCode, CondCode2;
2885
FPCCToARMCC(CC, CondCode, CondCode2);
2887
SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
2888
SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
2889
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
2890
SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
2892
if (CondCode2 != ARMCC::AL) {
2893
SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
2894
// FIXME: Needs another CMP because flag can have but one use.
2895
SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
2896
Result = DAG.getNode(ARMISD::CMOV, dl, VT,
2897
Result, TrueVal, ARMcc2, CCR, Cmp2);
2902
/// canChangeToInt - Given the fp compare operand, return true if it is suitable
2903
/// to morph to an integer compare sequence.
2904
static bool canChangeToInt(SDValue Op, bool &SeenZero,
2905
const ARMSubtarget *Subtarget) {
2906
SDNode *N = Op.getNode();
2907
if (!N->hasOneUse())
2908
// Otherwise it requires moving the value from fp to integer registers.
2910
if (!N->getNumValues())
2912
EVT VT = Op.getValueType();
2913
if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
2914
// f32 case is generally profitable. f64 case only makes sense when vcmpe +
2915
// vmrs are very slow, e.g. cortex-a8.
2918
if (isFloatingPointZero(Op)) {
2922
return ISD::isNormalLoad(N);
2925
static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
2926
if (isFloatingPointZero(Op))
2927
return DAG.getConstant(0, MVT::i32);
2929
if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
2930
return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2931
Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
2932
Ld->isVolatile(), Ld->isNonTemporal(),
2933
Ld->isInvariant(), Ld->getAlignment());
2935
llvm_unreachable("Unknown VFP cmp argument!");
2938
static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
2939
SDValue &RetVal1, SDValue &RetVal2) {
2940
if (isFloatingPointZero(Op)) {
2941
RetVal1 = DAG.getConstant(0, MVT::i32);
2942
RetVal2 = DAG.getConstant(0, MVT::i32);
2946
if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
2947
SDValue Ptr = Ld->getBasePtr();
2948
RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2949
Ld->getChain(), Ptr,
2950
Ld->getPointerInfo(),
2951
Ld->isVolatile(), Ld->isNonTemporal(),
2952
Ld->isInvariant(), Ld->getAlignment());
2954
EVT PtrType = Ptr.getValueType();
2955
unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
2956
SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
2957
PtrType, Ptr, DAG.getConstant(4, PtrType));
2958
RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
2959
Ld->getChain(), NewPtr,
2960
Ld->getPointerInfo().getWithOffset(4),
2961
Ld->isVolatile(), Ld->isNonTemporal(),
2962
Ld->isInvariant(), NewAlign);
2966
llvm_unreachable("Unknown VFP cmp argument!");
2969
/// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
2970
/// f32 and even f64 comparisons to integer ones.
2972
ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
2973
SDValue Chain = Op.getOperand(0);
2974
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2975
SDValue LHS = Op.getOperand(2);
2976
SDValue RHS = Op.getOperand(3);
2977
SDValue Dest = Op.getOperand(4);
2978
DebugLoc dl = Op.getDebugLoc();
2980
bool LHSSeenZero = false;
2981
bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
2982
bool RHSSeenZero = false;
2983
bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
2984
if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
2985
// If unsafe fp math optimization is enabled and there are no other uses of
2986
// the CMP operands, and the condition code is EQ or NE, we can optimize it
2987
// to an integer comparison.
2988
if (CC == ISD::SETOEQ)
2990
else if (CC == ISD::SETUNE)
2993
SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
2995
if (LHS.getValueType() == MVT::f32) {
2996
LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2997
bitcastf32Toi32(LHS, DAG), Mask);
2998
RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
2999
bitcastf32Toi32(RHS, DAG), Mask);
3000
SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3001
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3002
return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3003
Chain, Dest, ARMcc, CCR, Cmp);
3008
expandf64Toi32(LHS, DAG, LHS1, LHS2);
3009
expandf64Toi32(RHS, DAG, RHS1, RHS2);
3010
LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
3011
RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
3012
ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
3013
ARMcc = DAG.getConstant(CondCode, MVT::i32);
3014
SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3015
SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
3016
return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
3022
SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
3023
SDValue Chain = Op.getOperand(0);
3024
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
3025
SDValue LHS = Op.getOperand(2);
3026
SDValue RHS = Op.getOperand(3);
3027
SDValue Dest = Op.getOperand(4);
3028
DebugLoc dl = Op.getDebugLoc();
3030
if (LHS.getValueType() == MVT::i32) {
3032
SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
3033
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3034
return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
3035
Chain, Dest, ARMcc, CCR, Cmp);
3038
assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
3040
if (getTargetMachine().Options.UnsafeFPMath &&
3041
(CC == ISD::SETEQ || CC == ISD::SETOEQ ||
3042
CC == ISD::SETNE || CC == ISD::SETUNE)) {
3043
SDValue Result = OptimizeVFPBrcond(Op, DAG);
3044
if (Result.getNode())
3048
ARMCC::CondCodes CondCode, CondCode2;
3049
FPCCToARMCC(CC, CondCode, CondCode2);
3051
SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
3052
SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
3053
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3054
SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
3055
SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
3056
SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3057
if (CondCode2 != ARMCC::AL) {
3058
ARMcc = DAG.getConstant(CondCode2, MVT::i32);
3059
SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
3060
Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
3065
SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
3066
SDValue Chain = Op.getOperand(0);
3067
SDValue Table = Op.getOperand(1);
3068
SDValue Index = Op.getOperand(2);
3069
DebugLoc dl = Op.getDebugLoc();
3071
EVT PTy = getPointerTy();
3072
JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
3073
ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
3074
SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
3075
SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
3076
Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
3077
Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
3078
SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
3079
if (Subtarget->isThumb2()) {
3080
// Thumb2 uses a two-level jump. That is, it jumps into the jump table
3081
// which does another jump to the destination. This also makes it easier
3082
// to translate it to TBB / TBH later.
3083
// FIXME: This might not work if the function is extremely large.
3084
return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
3085
Addr, Op.getOperand(2), JTI, UId);
3087
if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
3088
Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
3089
MachinePointerInfo::getJumpTable(),
3090
false, false, false, 0);
3091
Chain = Addr.getValue(1);
3092
Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
3093
return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3095
Addr = DAG.getLoad(PTy, dl, Chain, Addr,
3096
MachinePointerInfo::getJumpTable(),
3097
false, false, false, 0);
3098
Chain = Addr.getValue(1);
3099
return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
3103
static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3104
EVT VT = Op.getValueType();
3105
DebugLoc dl = Op.getDebugLoc();
3107
if (Op.getValueType().getVectorElementType() == MVT::i32) {
3108
if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
3110
return DAG.UnrollVectorOp(Op.getNode());
3113
assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
3114
"Invalid type for custom lowering!");
3115
if (VT != MVT::v4i16)
3116
return DAG.UnrollVectorOp(Op.getNode());
3118
Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
3119
return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
3122
static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
3123
EVT VT = Op.getValueType();
3125
return LowerVectorFP_TO_INT(Op, DAG);
3127
DebugLoc dl = Op.getDebugLoc();
3130
switch (Op.getOpcode()) {
3131
default: llvm_unreachable("Invalid opcode!");
3132
case ISD::FP_TO_SINT:
3133
Opc = ARMISD::FTOSI;
3135
case ISD::FP_TO_UINT:
3136
Opc = ARMISD::FTOUI;
3139
Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
3140
return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
3143
static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3144
EVT VT = Op.getValueType();
3145
DebugLoc dl = Op.getDebugLoc();
3147
if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
3148
if (VT.getVectorElementType() == MVT::f32)
3150
return DAG.UnrollVectorOp(Op.getNode());
3153
assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
3154
"Invalid type for custom lowering!");
3155
if (VT != MVT::v4f32)
3156
return DAG.UnrollVectorOp(Op.getNode());
3160
switch (Op.getOpcode()) {
3161
default: llvm_unreachable("Invalid opcode!");
3162
case ISD::SINT_TO_FP:
3163
CastOpc = ISD::SIGN_EXTEND;
3164
Opc = ISD::SINT_TO_FP;
3166
case ISD::UINT_TO_FP:
3167
CastOpc = ISD::ZERO_EXTEND;
3168
Opc = ISD::UINT_TO_FP;
3172
Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
3173
return DAG.getNode(Opc, dl, VT, Op);
3176
static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
3177
EVT VT = Op.getValueType();
3179
return LowerVectorINT_TO_FP(Op, DAG);
3181
DebugLoc dl = Op.getDebugLoc();
3184
switch (Op.getOpcode()) {
3185
default: llvm_unreachable("Invalid opcode!");
3186
case ISD::SINT_TO_FP:
3187
Opc = ARMISD::SITOF;
3189
case ISD::UINT_TO_FP:
3190
Opc = ARMISD::UITOF;
3194
Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
3195
return DAG.getNode(Opc, dl, VT, Op);
3198
SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
3199
// Implement fcopysign with a fabs and a conditional fneg.
3200
SDValue Tmp0 = Op.getOperand(0);
3201
SDValue Tmp1 = Op.getOperand(1);
3202
DebugLoc dl = Op.getDebugLoc();
3203
EVT VT = Op.getValueType();
3204
EVT SrcVT = Tmp1.getValueType();
3205
bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
3206
Tmp0.getOpcode() == ARMISD::VMOVDRR;
3207
bool UseNEON = !InGPR && Subtarget->hasNEON();
3210
// Use VBSL to copy the sign bit.
3211
unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
3212
SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
3213
DAG.getTargetConstant(EncodedVal, MVT::i32));
3214
EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
3216
Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3217
DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
3218
DAG.getConstant(32, MVT::i32));
3219
else /*if (VT == MVT::f32)*/
3220
Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
3221
if (SrcVT == MVT::f32) {
3222
Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
3224
Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
3225
DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
3226
DAG.getConstant(32, MVT::i32));
3227
} else if (VT == MVT::f32)
3228
Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
3229
DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
3230
DAG.getConstant(32, MVT::i32));
3231
Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
3232
Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);
3234
SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
3236
AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
3237
SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
3238
DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));
3240
SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
3241
DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
3242
DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
3243
if (VT == MVT::f32) {
3244
Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
3245
Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
3246
DAG.getConstant(0, MVT::i32));
3248
Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
3254
// Bitcast operand 1 to i32.
3255
if (SrcVT == MVT::f64)
3256
Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3257
&Tmp1, 1).getValue(1);
3258
Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);
3260
// Or in the signbit with integer operations.
3261
SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
3262
SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
3263
Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
3264
if (VT == MVT::f32) {
3265
Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
3266
DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
3267
return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
3268
DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
3271
// f64: Or the high part with signbit and then combine two parts.
3272
Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
3274
SDValue Lo = Tmp0.getValue(0);
3275
SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
3276
Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
3277
return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
3280
SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
3281
MachineFunction &MF = DAG.getMachineFunction();
3282
MachineFrameInfo *MFI = MF.getFrameInfo();
3283
MFI->setReturnAddressIsTaken(true);
3285
EVT VT = Op.getValueType();
3286
DebugLoc dl = Op.getDebugLoc();
3287
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3289
SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
3290
SDValue Offset = DAG.getConstant(4, MVT::i32);
3291
return DAG.getLoad(VT, dl, DAG.getEntryNode(),
3292
DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
3293
MachinePointerInfo(), false, false, false, 0);
3296
// Return LR, which contains the return address. Mark it an implicit live-in.
3297
unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
3298
return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
3301
SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
3302
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3303
MFI->setFrameAddressIsTaken(true);
3305
EVT VT = Op.getValueType();
3306
DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
3307
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
3308
unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
3309
? ARM::R7 : ARM::R11;
3310
SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
3312
FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
3313
MachinePointerInfo(),
3314
false, false, false, 0);
3318
/// ExpandBITCAST - If the target supports VFP, this function is called to
3319
/// expand a bit convert where either the source or destination type is i64 to
3320
/// use a VMOVDRR or VMOVRRD node. This should not be done when the non-i64
3321
/// operand type is illegal (e.g., v2f32 for a target that doesn't support
3322
/// vectors), since the legalizer won't know what to do with that.
3323
static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
3324
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3325
DebugLoc dl = N->getDebugLoc();
3326
SDValue Op = N->getOperand(0);
3328
// This function is only supposed to be called for i64 types, either as the
3329
// source or destination of the bit convert.
3330
EVT SrcVT = Op.getValueType();
3331
EVT DstVT = N->getValueType(0);
3332
assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
3333
"ExpandBITCAST called for non-i64 type");
3335
// Turn i64->f64 into VMOVDRR.
3336
if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
3337
SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3338
DAG.getConstant(0, MVT::i32));
3339
SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
3340
DAG.getConstant(1, MVT::i32));
3341
return DAG.getNode(ISD::BITCAST, dl, DstVT,
3342
DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
3345
// Turn f64->i64 into VMOVRRD.
3346
if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
3347
SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
3348
DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
3349
// Merge the pieces into a single i64 value.
3350
return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
3356
/// getZeroVector - Returns a vector of specified type with all zero elements.
3357
/// Zero vectors are used to represent vector negation and in those cases
3358
/// will be implemented with the NEON VNEG instruction. However, VNEG does
3359
/// not support i64 elements, so sometimes the zero vectors will need to be
3360
/// explicitly constructed. Regardless, use a canonical VMOV to create the
3362
static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3363
assert(VT.isVector() && "Expected a vector type");
3364
// The canonical modified immediate encoding of a zero vector is....0!
3365
SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
3366
EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
3367
SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
3368
return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
3371
/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
3372
/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3373
SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
3374
SelectionDAG &DAG) const {
3375
assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3376
EVT VT = Op.getValueType();
3377
unsigned VTBits = VT.getSizeInBits();
3378
DebugLoc dl = Op.getDebugLoc();
3379
SDValue ShOpLo = Op.getOperand(0);
3380
SDValue ShOpHi = Op.getOperand(1);
3381
SDValue ShAmt = Op.getOperand(2);
3383
unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
3385
assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
3387
SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3388
DAG.getConstant(VTBits, MVT::i32), ShAmt);
3389
SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
3390
SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3391
DAG.getConstant(VTBits, MVT::i32));
3392
SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
3393
SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3394
SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
3396
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3397
SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3399
SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
3400
SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
3403
SDValue Ops[2] = { Lo, Hi };
3404
return DAG.getMergeValues(Ops, 2, dl);
3407
/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
3408
/// i32 values and take a 2 x i32 value to shift plus a shift amount.
3409
SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
3410
SelectionDAG &DAG) const {
3411
assert(Op.getNumOperands() == 3 && "Not a double-shift!");
3412
EVT VT = Op.getValueType();
3413
unsigned VTBits = VT.getSizeInBits();
3414
DebugLoc dl = Op.getDebugLoc();
3415
SDValue ShOpLo = Op.getOperand(0);
3416
SDValue ShOpHi = Op.getOperand(1);
3417
SDValue ShAmt = Op.getOperand(2);
3420
assert(Op.getOpcode() == ISD::SHL_PARTS);
3421
SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
3422
DAG.getConstant(VTBits, MVT::i32), ShAmt);
3423
SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
3424
SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
3425
DAG.getConstant(VTBits, MVT::i32));
3426
SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
3427
SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
3429
SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
3430
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
3431
SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
3433
SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
3434
SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
3437
SDValue Ops[2] = { Lo, Hi };
3438
return DAG.getMergeValues(Ops, 2, dl);
3441
SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
3442
SelectionDAG &DAG) const {
3443
// The rounding mode is in bits 23:22 of the FPSCR.
3444
// The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
3445
// The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
3446
// so that the shift + and get folded into a bitfield extract.
3447
DebugLoc dl = Op.getDebugLoc();
3448
SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
3449
DAG.getConstant(Intrinsic::arm_get_fpscr,
3451
SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
3452
DAG.getConstant(1U << 22, MVT::i32));
3453
SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
3454
DAG.getConstant(22, MVT::i32));
3455
return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
3456
DAG.getConstant(3, MVT::i32));
3459
static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
3460
const ARMSubtarget *ST) {
3461
EVT VT = N->getValueType(0);
3462
DebugLoc dl = N->getDebugLoc();
3464
if (!ST->hasV6T2Ops())
3467
SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
3468
return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
3471
static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
3472
const ARMSubtarget *ST) {
3473
EVT VT = N->getValueType(0);
3474
DebugLoc dl = N->getDebugLoc();
3479
// Lower vector shifts on NEON to use VSHL.
3480
assert(ST->hasNEON() && "unexpected vector shift");
3482
// Left shifts translate directly to the vshiftu intrinsic.
3483
if (N->getOpcode() == ISD::SHL)
3484
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3485
DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
3486
N->getOperand(0), N->getOperand(1));
3488
assert((N->getOpcode() == ISD::SRA ||
3489
N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");
3491
// NEON uses the same intrinsics for both left and right shifts. For
3492
// right shifts, the shift amounts are negative, so negate the vector of
3494
EVT ShiftVT = N->getOperand(1).getValueType();
3495
SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
3496
getZeroVector(ShiftVT, DAG, dl),
3498
Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
3499
Intrinsic::arm_neon_vshifts :
3500
Intrinsic::arm_neon_vshiftu);
3501
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
3502
DAG.getConstant(vshiftInt, MVT::i32),
3503
N->getOperand(0), NegatedCount);
3506
static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
3507
const ARMSubtarget *ST) {
3508
EVT VT = N->getValueType(0);
3509
DebugLoc dl = N->getDebugLoc();
3511
// We can get here for a node like i32 = ISD::SHL i32, i64
3515
assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
3516
"Unknown shift to lower!");
3518
// We only lower SRA, SRL of 1 here, all others use generic lowering.
3519
if (!isa<ConstantSDNode>(N->getOperand(1)) ||
3520
cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
3523
// If we are in thumb mode, we don't have RRX.
3524
if (ST->isThumb1Only()) return SDValue();
3526
// Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr.
3527
SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3528
DAG.getConstant(0, MVT::i32));
3529
SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
3530
DAG.getConstant(1, MVT::i32));
3532
// First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
3533
// captures the result into a carry flag.
3534
unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
3535
Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);
3537
// The low part is an ARMISD::RRX operand, which shifts the carry in.
3538
Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));
3540
// Merge the pieces into a single i64 value.
3541
return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
3544
static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
3545
SDValue TmpOp0, TmpOp1;
3546
bool Invert = false;
3550
SDValue Op0 = Op.getOperand(0);
3551
SDValue Op1 = Op.getOperand(1);
3552
SDValue CC = Op.getOperand(2);
3553
EVT VT = Op.getValueType();
3554
ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
3555
DebugLoc dl = Op.getDebugLoc();
3557
if (Op.getOperand(1).getValueType().isFloatingPoint()) {
3558
switch (SetCCOpcode) {
3559
default: llvm_unreachable("Illegal FP comparison");
3561
case ISD::SETNE: Invert = true; // Fallthrough
3563
case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3565
case ISD::SETLT: Swap = true; // Fallthrough
3567
case ISD::SETGT: Opc = ARMISD::VCGT; break;
3569
case ISD::SETLE: Swap = true; // Fallthrough
3571
case ISD::SETGE: Opc = ARMISD::VCGE; break;
3572
case ISD::SETUGE: Swap = true; // Fallthrough
3573
case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
3574
case ISD::SETUGT: Swap = true; // Fallthrough
3575
case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
3576
case ISD::SETUEQ: Invert = true; // Fallthrough
3578
// Expand this to (OLT | OGT).
3582
Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3583
Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
3585
case ISD::SETUO: Invert = true; // Fallthrough
3587
// Expand this to (OLT | OGE).
3591
Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
3592
Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
3596
// Integer comparisons.
3597
switch (SetCCOpcode) {
3598
default: llvm_unreachable("Illegal integer comparison");
3599
case ISD::SETNE: Invert = true;
3600
case ISD::SETEQ: Opc = ARMISD::VCEQ; break;
3601
case ISD::SETLT: Swap = true;
3602
case ISD::SETGT: Opc = ARMISD::VCGT; break;
3603
case ISD::SETLE: Swap = true;
3604
case ISD::SETGE: Opc = ARMISD::VCGE; break;
3605
case ISD::SETULT: Swap = true;
3606
case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
3607
case ISD::SETULE: Swap = true;
3608
case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
3611
// Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
3612
if (Opc == ARMISD::VCEQ) {
3615
if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3617
else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
3620
// Ignore bitconvert.
3621
if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
3622
AndOp = AndOp.getOperand(0);
3624
if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
3626
Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
3627
Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
3634
std::swap(Op0, Op1);
3636
// If one of the operands is a constant vector zero, attempt to fold the
3637
// comparison to a specialized compare-against-zero form.
3639
if (ISD::isBuildVectorAllZeros(Op1.getNode()))
3641
else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
3642
if (Opc == ARMISD::VCGE)
3643
Opc = ARMISD::VCLEZ;
3644
else if (Opc == ARMISD::VCGT)
3645
Opc = ARMISD::VCLTZ;
3650
if (SingleOp.getNode()) {
3653
Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
3655
Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
3657
Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
3659
Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
3661
Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
3663
Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3666
Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
3670
Result = DAG.getNOT(dl, Result, VT);
3675
SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
3676
const ARMSubtarget *ST) const {
3677
if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
3680
ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
3681
assert(Op.getValueType() == MVT::f32 &&
3682
"ConstantFP custom lowering should only occur for f32.");
3684
APFloat FPVal = CFP->getValueAPF();
3685
int ImmVal = ARM_AM::getFP32Imm(FPVal);
3689
DebugLoc DL = Op.getDebugLoc();
3690
SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
3691
SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32, NewVal);
3692
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
3693
DAG.getConstant(0, MVT::i32));
3696
/// isNEONModifiedImm - Check if the specified splat value corresponds to a
3697
/// valid vector constant for a NEON instruction with a "modified immediate"
3698
/// operand (e.g., VMOV). If so, return the encoded value.
3699
static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3700
unsigned SplatBitSize, SelectionDAG &DAG,
3701
EVT &VT, bool is128Bits, NEONModImmType type) {
3702
unsigned OpCmode, Imm;
3704
// SplatBitSize is set to the smallest size that splats the vector, so a
3705
// zero vector will always have SplatBitSize == 8. However, NEON modified
3706
// immediate instructions others than VMOV do not support the 8-bit encoding
3707
// of a zero vector, and the default encoding of zero is supposed to be the
3712
switch (SplatBitSize) {
3714
if (type != VMOVModImm)
3716
// Any 1-byte value is OK. Op=0, Cmode=1110.
3717
assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3720
VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3724
// NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
3725
VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3726
if ((SplatBits & ~0xff) == 0) {
3727
// Value = 0x00nn: Op=x, Cmode=100x.
3732
if ((SplatBits & ~0xff00) == 0) {
3733
// Value = 0xnn00: Op=x, Cmode=101x.
3735
Imm = SplatBits >> 8;
3741
// NEON's 32-bit VMOV supports splat values where:
3742
// * only one byte is nonzero, or
3743
// * the least significant byte is 0xff and the second byte is nonzero, or
3744
// * the least significant 2 bytes are 0xff and the third is nonzero.
3745
VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3746
if ((SplatBits & ~0xff) == 0) {
3747
// Value = 0x000000nn: Op=x, Cmode=000x.
3752
if ((SplatBits & ~0xff00) == 0) {
3753
// Value = 0x0000nn00: Op=x, Cmode=001x.
3755
Imm = SplatBits >> 8;
3758
if ((SplatBits & ~0xff0000) == 0) {
3759
// Value = 0x00nn0000: Op=x, Cmode=010x.
3761
Imm = SplatBits >> 16;
3764
if ((SplatBits & ~0xff000000) == 0) {
3765
// Value = 0xnn000000: Op=x, Cmode=011x.
3767
Imm = SplatBits >> 24;
3771
// cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
3772
if (type == OtherModImm) return SDValue();
3774
if ((SplatBits & ~0xffff) == 0 &&
3775
((SplatBits | SplatUndef) & 0xff) == 0xff) {
3776
// Value = 0x0000nnff: Op=x, Cmode=1100.
3778
Imm = SplatBits >> 8;
3783
if ((SplatBits & ~0xffffff) == 0 &&
3784
((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3785
// Value = 0x00nnffff: Op=x, Cmode=1101.
3787
Imm = SplatBits >> 16;
3788
SplatBits |= 0xffff;
3792
// Note: there are a few 32-bit splat values (specifically: 00ffff00,
3793
// ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
3794
// VMOV.I32. A (very) minor optimization would be to replicate the value
3795
// and fall through here to test for a valid 64-bit splat. But, then the
3796
// caller would also need to check and handle the change in size.
3800
if (type != VMOVModImm)
3802
// NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
3803
uint64_t BitMask = 0xff;
3805
unsigned ImmMask = 1;
3807
for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3808
if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3811
} else if ((SplatBits & BitMask) != 0) {
3817
// Op=1, Cmode=1110.
3820
VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3825
llvm_unreachable("unexpected size for isNEONModifiedImm");
3828
unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
3829
return DAG.getTargetConstant(EncodedVal, MVT::i32);
3832
static bool isVEXTMask(ArrayRef<int> M, EVT VT,
3833
bool &ReverseVEXT, unsigned &Imm) {
3834
unsigned NumElts = VT.getVectorNumElements();
3835
ReverseVEXT = false;
3837
// Assume that the first shuffle index is not UNDEF. Fail if it is.
3843
// If this is a VEXT shuffle, the immediate value is the index of the first
3844
// element. The other shuffle indices must be the successive elements after
3846
unsigned ExpectedElt = Imm;
3847
for (unsigned i = 1; i < NumElts; ++i) {
3848
// Increment the expected index. If it wraps around, it may still be
3849
// a VEXT but the source vectors must be swapped.
3851
if (ExpectedElt == NumElts * 2) {
3856
if (M[i] < 0) continue; // ignore UNDEF indices
3857
if (ExpectedElt != static_cast<unsigned>(M[i]))
3861
// Adjust the index value if the source operands will be swapped.
3868
/// isVREVMask - Check if a vector shuffle corresponds to a VREV
3869
/// instruction with the specified blocksize. (The order of the elements
3870
/// within each block of the vector is reversed.)
3871
static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
3872
assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
3873
"Only possible block sizes for VREV are: 16, 32, 64");
3875
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3879
unsigned NumElts = VT.getVectorNumElements();
3880
unsigned BlockElts = M[0] + 1;
3881
// If the first shuffle index is UNDEF, be optimistic.
3883
BlockElts = BlockSize / EltSz;
3885
if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
3888
for (unsigned i = 0; i < NumElts; ++i) {
3889
if (M[i] < 0) continue; // ignore UNDEF indices
3890
if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
3897
static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
3898
// We can handle <8 x i8> vector shuffles. If the index in the mask is out of
3899
// range, then 0 is placed into the resulting vector. So pretty much any mask
3900
// of 8 elements can work here.
3901
return VT == MVT::v8i8 && M.size() == 8;
3904
static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3905
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3909
unsigned NumElts = VT.getVectorNumElements();
3910
WhichResult = (M[0] == 0 ? 0 : 1);
3911
for (unsigned i = 0; i < NumElts; i += 2) {
3912
if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3913
(M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
3919
/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
3920
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3921
/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
3922
static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3923
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3927
unsigned NumElts = VT.getVectorNumElements();
3928
WhichResult = (M[0] == 0 ? 0 : 1);
3929
for (unsigned i = 0; i < NumElts; i += 2) {
3930
if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
3931
(M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
3937
static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3938
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3942
unsigned NumElts = VT.getVectorNumElements();
3943
WhichResult = (M[0] == 0 ? 0 : 1);
3944
for (unsigned i = 0; i != NumElts; ++i) {
3945
if (M[i] < 0) continue; // ignore UNDEF indices
3946
if ((unsigned) M[i] != 2 * i + WhichResult)
3950
// VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3951
if (VT.is64BitVector() && EltSz == 32)
3957
/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
3958
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
3959
/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
3960
static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
3961
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3965
unsigned Half = VT.getVectorNumElements() / 2;
3966
WhichResult = (M[0] == 0 ? 0 : 1);
3967
for (unsigned j = 0; j != 2; ++j) {
3968
unsigned Idx = WhichResult;
3969
for (unsigned i = 0; i != Half; ++i) {
3970
int MIdx = M[i + j * Half];
3971
if (MIdx >= 0 && (unsigned) MIdx != Idx)
3977
// VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
3978
if (VT.is64BitVector() && EltSz == 32)
3984
static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
3985
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
3989
unsigned NumElts = VT.getVectorNumElements();
3990
WhichResult = (M[0] == 0 ? 0 : 1);
3991
unsigned Idx = WhichResult * NumElts / 2;
3992
for (unsigned i = 0; i != NumElts; i += 2) {
3993
if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
3994
(M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
3999
// VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4000
if (VT.is64BitVector() && EltSz == 32)
4006
/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
4007
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
4008
/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
4009
static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
4010
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4014
unsigned NumElts = VT.getVectorNumElements();
4015
WhichResult = (M[0] == 0 ? 0 : 1);
4016
unsigned Idx = WhichResult * NumElts / 2;
4017
for (unsigned i = 0; i != NumElts; i += 2) {
4018
if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
4019
(M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
4024
// VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
4025
if (VT.is64BitVector() && EltSz == 32)
4031
// If N is an integer constant that can be moved into a register in one
4032
// instruction, return an SDValue of such a constant (will become a MOV
4033
// instruction). Otherwise return null.
4034
static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
4035
const ARMSubtarget *ST, DebugLoc dl) {
4037
if (!isa<ConstantSDNode>(N))
4039
Val = cast<ConstantSDNode>(N)->getZExtValue();
4041
if (ST->isThumb1Only()) {
4042
if (Val <= 255 || ~Val <= 255)
4043
return DAG.getConstant(Val, MVT::i32);
4045
if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
4046
return DAG.getConstant(Val, MVT::i32);
4051
// If this is a case we can't handle, return null and let the default
4052
// expansion code take care of it.
4053
SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4054
const ARMSubtarget *ST) const {
4055
BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4056
DebugLoc dl = Op.getDebugLoc();
4057
EVT VT = Op.getValueType();
4059
APInt SplatBits, SplatUndef;
4060
unsigned SplatBitSize;
4062
if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4063
if (SplatBitSize <= 64) {
4064
// Check if an immediate VMOV works.
4066
SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
4067
SplatUndef.getZExtValue(), SplatBitSize,
4068
DAG, VmovVT, VT.is128BitVector(),
4070
if (Val.getNode()) {
4071
SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
4072
return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4075
// Try an immediate VMVN.
4076
uint64_t NegatedImm = (~SplatBits).getZExtValue();
4077
Val = isNEONModifiedImm(NegatedImm,
4078
SplatUndef.getZExtValue(), SplatBitSize,
4079
DAG, VmovVT, VT.is128BitVector(),
4081
if (Val.getNode()) {
4082
SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
4083
return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
4086
// Use vmov.f32 to materialize other v2f32 and v4f32 splats.
4087
if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
4088
int ImmVal = ARM_AM::getFP32Imm(SplatBits);
4090
SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4091
return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
4097
// Scan through the operands to see if only one value is used.
4098
unsigned NumElts = VT.getVectorNumElements();
4099
bool isOnlyLowElement = true;
4100
bool usesOnlyOneValue = true;
4101
bool isConstant = true;
4103
for (unsigned i = 0; i < NumElts; ++i) {
4104
SDValue V = Op.getOperand(i);
4105
if (V.getOpcode() == ISD::UNDEF)
4108
isOnlyLowElement = false;
4109
if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4112
if (!Value.getNode())
4114
else if (V != Value)
4115
usesOnlyOneValue = false;
4118
if (!Value.getNode())
4119
return DAG.getUNDEF(VT);
4121
if (isOnlyLowElement)
4122
return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);
4124
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4126
// Use VDUP for non-constant splats. For f32 constant splats, reduce to
4127
// i32 and try again.
4128
if (usesOnlyOneValue && EltSize <= 32) {
4130
return DAG.getNode(ARMISD::VDUP, dl, VT, Value);
4131
if (VT.getVectorElementType().isFloatingPoint()) {
4132
SmallVector<SDValue, 8> Ops;
4133
for (unsigned i = 0; i < NumElts; ++i)
4134
Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
4136
EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
4137
SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
4138
Val = LowerBUILD_VECTOR(Val, DAG, ST);
4140
return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4142
SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
4144
return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
4147
// If all elements are constants and the case above didn't get hit, fall back
4148
// to the default expansion, which will generate a load from the constant
4153
// Empirical tests suggest this is rarely worth it for vectors of length <= 2.
4155
SDValue shuffle = ReconstructShuffle(Op, DAG);
4156
if (shuffle != SDValue())
4160
// Vectors with 32- or 64-bit elements can be built by directly assigning
4161
// the subregisters. Lower it to an ARMISD::BUILD_VECTOR so the operands
4162
// will be legalized.
4163
if (EltSize >= 32) {
4164
// Do the expansion with floating-point types, since that is what the VFP
4165
// registers are defined to use, and since i64 is not legal.
4166
EVT EltVT = EVT::getFloatingPointVT(EltSize);
4167
EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4168
SmallVector<SDValue, 8> Ops;
4169
for (unsigned i = 0; i < NumElts; ++i)
4170
Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
4171
SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4172
return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4178
// Gather data to see if the operation can be modelled as a
4179
// shuffle in combination with VEXTs.
4180
SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
4181
SelectionDAG &DAG) const {
4182
DebugLoc dl = Op.getDebugLoc();
4183
EVT VT = Op.getValueType();
4184
unsigned NumElts = VT.getVectorNumElements();
4186
SmallVector<SDValue, 2> SourceVecs;
4187
SmallVector<unsigned, 2> MinElts;
4188
SmallVector<unsigned, 2> MaxElts;
4190
for (unsigned i = 0; i < NumElts; ++i) {
4191
SDValue V = Op.getOperand(i);
4192
if (V.getOpcode() == ISD::UNDEF)
4194
else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
4195
// A shuffle can only come from building a vector from various
4196
// elements of other vectors.
4198
} else if (V.getOperand(0).getValueType().getVectorElementType() !=
4199
VT.getVectorElementType()) {
4200
// This code doesn't know how to handle shuffles where the vector
4201
// element types do not match (this happens because type legalization
4202
// promotes the return type of EXTRACT_VECTOR_ELT).
4203
// FIXME: It might be appropriate to extend this code to handle
4204
// mismatched types.
4208
// Record this extraction against the appropriate vector if possible...
4209
SDValue SourceVec = V.getOperand(0);
4210
unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
4211
bool FoundSource = false;
4212
for (unsigned j = 0; j < SourceVecs.size(); ++j) {
4213
if (SourceVecs[j] == SourceVec) {
4214
if (MinElts[j] > EltNo)
4216
if (MaxElts[j] < EltNo)
4223
// Or record a new source if not...
4225
SourceVecs.push_back(SourceVec);
4226
MinElts.push_back(EltNo);
4227
MaxElts.push_back(EltNo);
4231
// Currently only do something sane when at most two source vectors
4233
if (SourceVecs.size() > 2)
4236
SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
4237
int VEXTOffsets[2] = {0, 0};
4239
// This loop extracts the usage patterns of the source vectors
4240
// and prepares appropriate SDValues for a shuffle if possible.
4241
for (unsigned i = 0; i < SourceVecs.size(); ++i) {
4242
if (SourceVecs[i].getValueType() == VT) {
4243
// No VEXT necessary
4244
ShuffleSrcs[i] = SourceVecs[i];
4247
} else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
4248
// It probably isn't worth padding out a smaller vector just to
4249
// break it down again in a shuffle.
4253
// Since only 64-bit and 128-bit vectors are legal on ARM and
4254
// we've eliminated the other cases...
4255
assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
4256
"unexpected vector sizes in ReconstructShuffle");
4258
if (MaxElts[i] - MinElts[i] >= NumElts) {
4259
// Span too large for a VEXT to cope
4263
if (MinElts[i] >= NumElts) {
4264
// The extraction can just take the second half
4265
VEXTOffsets[i] = NumElts;
4266
ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4268
DAG.getIntPtrConstant(NumElts));
4269
} else if (MaxElts[i] < NumElts) {
4270
// The extraction can just take the first half
4272
ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4274
DAG.getIntPtrConstant(0));
4276
// An actual VEXT is needed
4277
VEXTOffsets[i] = MinElts[i];
4278
SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4280
DAG.getIntPtrConstant(0));
4281
SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
4283
DAG.getIntPtrConstant(NumElts));
4284
ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
4285
DAG.getConstant(VEXTOffsets[i], MVT::i32));
4289
SmallVector<int, 8> Mask;
4291
for (unsigned i = 0; i < NumElts; ++i) {
4292
SDValue Entry = Op.getOperand(i);
4293
if (Entry.getOpcode() == ISD::UNDEF) {
4298
SDValue ExtractVec = Entry.getOperand(0);
4299
int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
4300
.getOperand(1))->getSExtValue();
4301
if (ExtractVec == SourceVecs[0]) {
4302
Mask.push_back(ExtractElt - VEXTOffsets[0]);
4304
Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
4308
// Final check before we try to produce nonsense...
4309
if (isShuffleMaskLegal(Mask, VT))
4310
return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
4316
/// isShuffleMaskLegal - Targets can use this to indicate that they only
4317
/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
4318
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
4319
/// are assumed to be legal.
4321
ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
4323
if (VT.getVectorNumElements() == 4 &&
4324
(VT.is128BitVector() || VT.is64BitVector())) {
4325
unsigned PFIndexes[4];
4326
for (unsigned i = 0; i != 4; ++i) {
4330
PFIndexes[i] = M[i];
4333
// Compute the index in the perfect shuffle table.
4334
unsigned PFTableIndex =
4335
PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4336
unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4337
unsigned Cost = (PFEntry >> 30);
4344
unsigned Imm, WhichResult;
4346
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4347
return (EltSize >= 32 ||
4348
ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
4349
isVREVMask(M, VT, 64) ||
4350
isVREVMask(M, VT, 32) ||
4351
isVREVMask(M, VT, 16) ||
4352
isVEXTMask(M, VT, ReverseVEXT, Imm) ||
4353
isVTBLMask(M, VT) ||
4354
isVTRNMask(M, VT, WhichResult) ||
4355
isVUZPMask(M, VT, WhichResult) ||
4356
isVZIPMask(M, VT, WhichResult) ||
4357
isVTRN_v_undef_Mask(M, VT, WhichResult) ||
4358
isVUZP_v_undef_Mask(M, VT, WhichResult) ||
4359
isVZIP_v_undef_Mask(M, VT, WhichResult));
4362
/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
4363
/// the specified operations to build the shuffle.
4364
static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
4365
SDValue RHS, SelectionDAG &DAG,
4367
unsigned OpNum = (PFEntry >> 26) & 0x0F;
4368
unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
4369
unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
4372
OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
4381
OP_VUZPL, // VUZP, left result
4382
OP_VUZPR, // VUZP, right result
4383
OP_VZIPL, // VZIP, left result
4384
OP_VZIPR, // VZIP, right result
4385
OP_VTRNL, // VTRN, left result
4386
OP_VTRNR // VTRN, right result
4389
if (OpNum == OP_COPY) {
4390
if (LHSID == (1*9+2)*9+3) return LHS;
4391
assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
4395
SDValue OpLHS, OpRHS;
4396
OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
4397
OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
4398
EVT VT = OpLHS.getValueType();
4401
default: llvm_unreachable("Unknown shuffle opcode!");
4403
// VREV divides the vector in half and swaps within the half.
4404
if (VT.getVectorElementType() == MVT::i32 ||
4405
VT.getVectorElementType() == MVT::f32)
4406
return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
4407
// vrev <4 x i16> -> VREV32
4408
if (VT.getVectorElementType() == MVT::i16)
4409
return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
4410
// vrev <4 x i8> -> VREV16
4411
assert(VT.getVectorElementType() == MVT::i8);
4412
return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
4417
return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
4418
OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
4422
return DAG.getNode(ARMISD::VEXT, dl, VT,
4424
DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
4427
return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4428
OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
4431
return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4432
OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
4435
return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4436
OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
4440
static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
4441
ArrayRef<int> ShuffleMask,
4442
SelectionDAG &DAG) {
4443
// Check to see if we can use the VTBL instruction.
4444
SDValue V1 = Op.getOperand(0);
4445
SDValue V2 = Op.getOperand(1);
4446
DebugLoc DL = Op.getDebugLoc();
4448
SmallVector<SDValue, 8> VTBLMask;
4449
for (ArrayRef<int>::iterator
4450
I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
4451
VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));
4453
if (V2.getNode()->getOpcode() == ISD::UNDEF)
4454
return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
4455
DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4458
return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
4459
DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
4463
static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4464
SDValue V1 = Op.getOperand(0);
4465
SDValue V2 = Op.getOperand(1);
4466
DebugLoc dl = Op.getDebugLoc();
4467
EVT VT = Op.getValueType();
4468
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4470
// Convert shuffles that are directly supported on NEON to target-specific
4471
// DAG nodes, instead of keeping them as shuffles and matching them again
4472
// during code selection. This is more efficient and avoids the possibility
4473
// of inconsistencies between legalization and selection.
4474
// FIXME: floating-point vectors should be canonicalized to integer vectors
4475
// of the same time so that they get CSEd properly.
4476
ArrayRef<int> ShuffleMask = SVN->getMask();
4478
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4479
if (EltSize <= 32) {
4480
if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4481
int Lane = SVN->getSplatIndex();
4482
// If this is undef splat, generate it via "just" vdup, if possible.
4483
if (Lane == -1) Lane = 0;
4485
// Test if V1 is a SCALAR_TO_VECTOR.
4486
if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4487
return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4489
// Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
4490
// (and probably will turn into a SCALAR_TO_VECTOR once legalization
4492
if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
4493
!isa<ConstantSDNode>(V1.getOperand(0))) {
4494
bool IsScalarToVector = true;
4495
for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
4496
if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
4497
IsScalarToVector = false;
4500
if (IsScalarToVector)
4501
return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
4503
return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
4504
DAG.getConstant(Lane, MVT::i32));
4509
if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
4512
return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
4513
DAG.getConstant(Imm, MVT::i32));
4516
if (isVREVMask(ShuffleMask, VT, 64))
4517
return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
4518
if (isVREVMask(ShuffleMask, VT, 32))
4519
return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
4520
if (isVREVMask(ShuffleMask, VT, 16))
4521
return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
4523
// Check for Neon shuffles that modify both input vectors in place.
4524
// If both results are used, i.e., if there are two shuffles with the same
4525
// source operands and with masks corresponding to both results of one of
4526
// these operations, DAG memoization will ensure that a single node is
4527
// used for both shuffles.
4528
unsigned WhichResult;
4529
if (isVTRNMask(ShuffleMask, VT, WhichResult))
4530
return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4531
V1, V2).getValue(WhichResult);
4532
if (isVUZPMask(ShuffleMask, VT, WhichResult))
4533
return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4534
V1, V2).getValue(WhichResult);
4535
if (isVZIPMask(ShuffleMask, VT, WhichResult))
4536
return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4537
V1, V2).getValue(WhichResult);
4539
if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
4540
return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
4541
V1, V1).getValue(WhichResult);
4542
if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4543
return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
4544
V1, V1).getValue(WhichResult);
4545
if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
4546
return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
4547
V1, V1).getValue(WhichResult);
4550
// If the shuffle is not directly supported and it has 4 elements, use
4551
// the PerfectShuffle-generated table to synthesize it from other shuffles.
4552
unsigned NumElts = VT.getVectorNumElements();
4554
unsigned PFIndexes[4];
4555
for (unsigned i = 0; i != 4; ++i) {
4556
if (ShuffleMask[i] < 0)
4559
PFIndexes[i] = ShuffleMask[i];
4562
// Compute the index in the perfect shuffle table.
4563
unsigned PFTableIndex =
4564
PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
4565
unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
4566
unsigned Cost = (PFEntry >> 30);
4569
return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
4572
// Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
4573
if (EltSize >= 32) {
4574
// Do the expansion with floating-point types, since that is what the VFP
4575
// registers are defined to use, and since i64 is not legal.
4576
EVT EltVT = EVT::getFloatingPointVT(EltSize);
4577
EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
4578
V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
4579
V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
4580
SmallVector<SDValue, 8> Ops;
4581
for (unsigned i = 0; i < NumElts; ++i) {
4582
if (ShuffleMask[i] < 0)
4583
Ops.push_back(DAG.getUNDEF(EltVT));
4585
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
4586
ShuffleMask[i] < (int)NumElts ? V1 : V2,
4587
DAG.getConstant(ShuffleMask[i] & (NumElts-1),
4590
SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
4591
return DAG.getNode(ISD::BITCAST, dl, VT, Val);
4594
if (VT == MVT::v8i8) {
4595
SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
4596
if (NewOp.getNode())
4603
static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4604
// INSERT_VECTOR_ELT is legal only for immediate indexes.
4605
SDValue Lane = Op.getOperand(2);
4606
if (!isa<ConstantSDNode>(Lane))
4612
static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4613
// EXTRACT_VECTOR_ELT is legal only for immediate indexes.
4614
SDValue Lane = Op.getOperand(1);
4615
if (!isa<ConstantSDNode>(Lane))
4618
SDValue Vec = Op.getOperand(0);
4619
if (Op.getValueType() == MVT::i32 &&
4620
Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
4621
DebugLoc dl = Op.getDebugLoc();
4622
return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
4628
static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
4629
// The only time a CONCAT_VECTORS operation can have legal types is when
4630
// two 64-bit vectors are concatenated to a 128-bit vector.
4631
assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
4632
"unexpected CONCAT_VECTORS");
4633
DebugLoc dl = Op.getDebugLoc();
4634
SDValue Val = DAG.getUNDEF(MVT::v2f64);
4635
SDValue Op0 = Op.getOperand(0);
4636
SDValue Op1 = Op.getOperand(1);
4637
if (Op0.getOpcode() != ISD::UNDEF)
4638
Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4639
DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
4640
DAG.getIntPtrConstant(0));
4641
if (Op1.getOpcode() != ISD::UNDEF)
4642
Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
4643
DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
4644
DAG.getIntPtrConstant(1));
4645
return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
4648
/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
4649
/// element has been zero/sign-extended, depending on the isSigned parameter,
4650
/// from an integer type half its size.
4651
static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
4653
// A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
4654
EVT VT = N->getValueType(0);
4655
if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
4656
SDNode *BVN = N->getOperand(0).getNode();
4657
if (BVN->getValueType(0) != MVT::v4i32 ||
4658
BVN->getOpcode() != ISD::BUILD_VECTOR)
4660
unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4661
unsigned HiElt = 1 - LoElt;
4662
ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
4663
ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
4664
ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
4665
ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
4666
if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
4669
if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
4670
Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
4673
if (Hi0->isNullValue() && Hi1->isNullValue())
4679
if (N->getOpcode() != ISD::BUILD_VECTOR)
4682
for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
4683
SDNode *Elt = N->getOperand(i).getNode();
4684
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
4685
unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4686
unsigned HalfSize = EltSize / 2;
4688
if (!isIntN(HalfSize, C->getSExtValue()))
4691
if (!isUIntN(HalfSize, C->getZExtValue()))
4702
/// isSignExtended - Check if a node is a vector value that is sign-extended
4703
/// or a constant BUILD_VECTOR with sign-extended elements.
4704
static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
4705
if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
4707
if (isExtendedBUILD_VECTOR(N, DAG, true))
4712
/// isZeroExtended - Check if a node is a vector value that is zero-extended
4713
/// or a constant BUILD_VECTOR with zero-extended elements.
4714
static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
4715
if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
4717
if (isExtendedBUILD_VECTOR(N, DAG, false))
4722
/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
4723
/// load, or BUILD_VECTOR with extended elements, return the unextended value.
4724
static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
4725
if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
4726
return N->getOperand(0);
4727
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
4728
return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
4729
LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
4730
LD->isNonTemporal(), LD->isInvariant(),
4731
LD->getAlignment());
4732
// Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
4733
// have been legalized as a BITCAST from v4i32.
4734
if (N->getOpcode() == ISD::BITCAST) {
4735
SDNode *BVN = N->getOperand(0).getNode();
4736
assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
4737
BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
4738
unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
4739
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
4740
BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
4742
// Construct a new BUILD_VECTOR with elements truncated to half the size.
4743
assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
4744
EVT VT = N->getValueType(0);
4745
unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
4746
unsigned NumElts = VT.getVectorNumElements();
4747
MVT TruncVT = MVT::getIntegerVT(EltSize);
4748
SmallVector<SDValue, 8> Ops;
4749
for (unsigned i = 0; i != NumElts; ++i) {
4750
ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
4751
const APInt &CInt = C->getAPIntValue();
4752
Ops.push_back(DAG.getConstant(CInt.trunc(EltSize), TruncVT));
4754
return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
4755
MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
4758
static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
4759
unsigned Opcode = N->getOpcode();
4760
if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4761
SDNode *N0 = N->getOperand(0).getNode();
4762
SDNode *N1 = N->getOperand(1).getNode();
4763
return N0->hasOneUse() && N1->hasOneUse() &&
4764
isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
4769
static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
4770
unsigned Opcode = N->getOpcode();
4771
if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
4772
SDNode *N0 = N->getOperand(0).getNode();
4773
SDNode *N1 = N->getOperand(1).getNode();
4774
return N0->hasOneUse() && N1->hasOneUse() &&
4775
isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
4780
static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
4781
// Multiplications are only custom-lowered for 128-bit vectors so that
4782
// VMULL can be detected. Otherwise v2i64 multiplications are not legal.
4783
EVT VT = Op.getValueType();
4784
assert(VT.is128BitVector() && "unexpected type for custom-lowering ISD::MUL");
4785
SDNode *N0 = Op.getOperand(0).getNode();
4786
SDNode *N1 = Op.getOperand(1).getNode();
4787
unsigned NewOpc = 0;
4789
bool isN0SExt = isSignExtended(N0, DAG);
4790
bool isN1SExt = isSignExtended(N1, DAG);
4791
if (isN0SExt && isN1SExt)
4792
NewOpc = ARMISD::VMULLs;
4794
bool isN0ZExt = isZeroExtended(N0, DAG);
4795
bool isN1ZExt = isZeroExtended(N1, DAG);
4796
if (isN0ZExt && isN1ZExt)
4797
NewOpc = ARMISD::VMULLu;
4798
else if (isN1SExt || isN1ZExt) {
4799
// Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
4800
// into (s/zext A * s/zext C) + (s/zext B * s/zext C)
4801
if (isN1SExt && isAddSubSExt(N0, DAG)) {
4802
NewOpc = ARMISD::VMULLs;
4804
} else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
4805
NewOpc = ARMISD::VMULLu;
4807
} else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
4809
NewOpc = ARMISD::VMULLu;
4815
if (VT == MVT::v2i64)
4816
// Fall through to expand this. It is not legal.
4819
// Other vector multiplications are legal.
4824
// Legalize to a VMULL instruction.
4825
DebugLoc DL = Op.getDebugLoc();
4827
SDValue Op1 = SkipExtension(N1, DAG);
4829
Op0 = SkipExtension(N0, DAG);
4830
assert(Op0.getValueType().is64BitVector() &&
4831
Op1.getValueType().is64BitVector() &&
4832
"unexpected types for extended operands to VMULL");
4833
return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
4836
// Optimizing (zext A + zext B) * C, to (VMULL A, C) + (VMULL B, C) during
4837
// isel lowering to take advantage of no-stall back to back vmul + vmla.
4844
SDValue N00 = SkipExtension(N0->getOperand(0).getNode(), DAG);
4845
SDValue N01 = SkipExtension(N0->getOperand(1).getNode(), DAG);
4846
EVT Op1VT = Op1.getValueType();
4847
return DAG.getNode(N0->getOpcode(), DL, VT,
4848
DAG.getNode(NewOpc, DL, VT,
4849
DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
4850
DAG.getNode(NewOpc, DL, VT,
4851
DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
4855
LowerSDIV_v4i8(SDValue X, SDValue Y, DebugLoc dl, SelectionDAG &DAG) {
4857
// float4 xf = vcvt_f32_s32(vmovl_s16(a.lo));
4858
// float4 yf = vcvt_f32_s32(vmovl_s16(b.lo));
4859
X = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, X);
4860
Y = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Y);
4861
X = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, X);
4862
Y = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, Y);
4863
// Get reciprocal estimate.
4864
// float4 recip = vrecpeq_f32(yf);
4865
Y = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4866
DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), Y);
4867
// Because char has a smaller range than uchar, we can actually get away
4868
// without any newton steps. This requires that we use a weird bias
4869
// of 0xb000, however (again, this has been exhaustively tested).
4870
// float4 result = as_float4(as_int4(xf*recip) + 0xb000);
4871
X = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, X, Y);
4872
X = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, X);
4873
Y = DAG.getConstant(0xb000, MVT::i32);
4874
Y = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Y, Y, Y, Y);
4875
X = DAG.getNode(ISD::ADD, dl, MVT::v4i32, X, Y);
4876
X = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, X);
4877
// Convert back to short.
4878
X = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, X);
4879
X = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, X);
4884
LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
4886
// Convert to float.
4887
// float4 yf = vcvt_f32_s32(vmovl_s16(y));
4888
// float4 xf = vcvt_f32_s32(vmovl_s16(x));
4889
N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N0);
4890
N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, N1);
4891
N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4892
N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4894
// Use reciprocal estimate and one refinement step.
4895
// float4 recip = vrecpeq_f32(yf);
4896
// recip *= vrecpsq_f32(yf, recip);
4897
N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4898
DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
4899
N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
4900
DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
4902
N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
4903
// Because short has a smaller range than ushort, we can actually get away
4904
// with only a single newton step. This requires that we use a weird bias
4905
// of 89, however (again, this has been exhaustively tested).
4906
// float4 result = as_float4(as_int4(xf*recip) + 0x89);
4907
N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
4908
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
4909
N1 = DAG.getConstant(0x89, MVT::i32);
4910
N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
4911
N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
4912
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
4913
// Convert back to integer and return.
4914
// return vmovn_s32(vcvt_s32_f32(result));
4915
N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
4916
N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
4920
static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
4921
EVT VT = Op.getValueType();
4922
assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4923
"unexpected type for custom-lowering ISD::SDIV");
4925
DebugLoc dl = Op.getDebugLoc();
4926
SDValue N0 = Op.getOperand(0);
4927
SDValue N1 = Op.getOperand(1);
4930
if (VT == MVT::v8i8) {
4931
N0 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N0);
4932
N1 = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i16, N1);
4934
N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4935
DAG.getIntPtrConstant(4));
4936
N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4937
DAG.getIntPtrConstant(4));
4938
N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4939
DAG.getIntPtrConstant(0));
4940
N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4941
DAG.getIntPtrConstant(0));
4943
N0 = LowerSDIV_v4i8(N0, N1, dl, DAG); // v4i16
4944
N2 = LowerSDIV_v4i8(N2, N3, dl, DAG); // v4i16
4946
N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4947
N0 = LowerCONCAT_VECTORS(N0, DAG);
4949
N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v8i8, N0);
4952
return LowerSDIV_v4i16(N0, N1, dl, DAG);
4955
static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
4956
EVT VT = Op.getValueType();
4957
assert((VT == MVT::v4i16 || VT == MVT::v8i8) &&
4958
"unexpected type for custom-lowering ISD::UDIV");
4960
DebugLoc dl = Op.getDebugLoc();
4961
SDValue N0 = Op.getOperand(0);
4962
SDValue N1 = Op.getOperand(1);
4965
if (VT == MVT::v8i8) {
4966
N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N0);
4967
N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v8i16, N1);
4969
N2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4970
DAG.getIntPtrConstant(4));
4971
N3 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4972
DAG.getIntPtrConstant(4));
4973
N0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N0,
4974
DAG.getIntPtrConstant(0));
4975
N1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i16, N1,
4976
DAG.getIntPtrConstant(0));
4978
N0 = LowerSDIV_v4i16(N0, N1, dl, DAG); // v4i16
4979
N2 = LowerSDIV_v4i16(N2, N3, dl, DAG); // v4i16
4981
N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i16, N0, N2);
4982
N0 = LowerCONCAT_VECTORS(N0, DAG);
4984
N0 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v8i8,
4985
DAG.getConstant(Intrinsic::arm_neon_vqmovnsu, MVT::i32),
4990
// v4i16 sdiv ... Convert to float.
4991
// float4 yf = vcvt_f32_s32(vmovl_u16(y));
4992
// float4 xf = vcvt_f32_s32(vmovl_u16(x));
4993
N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
4994
N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
4995
N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
4996
SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
4998
// Use reciprocal estimate and two refinement steps.
4999
// float4 recip = vrecpeq_f32(yf);
5000
// recip *= vrecpsq_f32(yf, recip);
5001
// recip *= vrecpsq_f32(yf, recip);
5002
N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5003
DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
5004
N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5005
DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5007
N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5008
N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
5009
DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
5011
N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
5012
// Simply multiplying by the reciprocal estimate can leave us a few ulps
5013
// too low, so we add 2 ulps (exhaustive testing shows that this is enough,
5014
// and that it will never cause us to return an answer too large).
5015
// float4 result = as_float4(as_int4(xf*recip) + 2);
5016
N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
5017
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
5018
N1 = DAG.getConstant(2, MVT::i32);
5019
N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
5020
N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
5021
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
5022
// Convert back to integer and return.
5023
// return vmovn_u32(vcvt_s32_f32(result));
5024
N0 = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::v4i32, N0);
5025
N0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::v4i16, N0);
5029
static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
5030
EVT VT = Op.getNode()->getValueType(0);
5031
SDVTList VTs = DAG.getVTList(VT, MVT::i32);
5034
bool ExtraOp = false;
5035
switch (Op.getOpcode()) {
5036
default: llvm_unreachable("Invalid code");
5037
case ISD::ADDC: Opc = ARMISD::ADDC; break;
5038
case ISD::ADDE: Opc = ARMISD::ADDE; ExtraOp = true; break;
5039
case ISD::SUBC: Opc = ARMISD::SUBC; break;
5040
case ISD::SUBE: Opc = ARMISD::SUBE; ExtraOp = true; break;
5044
return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5046
return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
5047
Op.getOperand(1), Op.getOperand(2));
5050
static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
5051
// Monotonic load/store is legal for all targets
5052
if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
5055
// Aquire/Release load/store is not legal for targets without a
5056
// dmb or equivalent available.
5062
ReplaceATOMIC_OP_64(SDNode *Node, SmallVectorImpl<SDValue>& Results,
5063
SelectionDAG &DAG, unsigned NewOp) {
5064
DebugLoc dl = Node->getDebugLoc();
5065
assert (Node->getValueType(0) == MVT::i64 &&
5066
"Only know how to expand i64 atomics");
5068
SmallVector<SDValue, 6> Ops;
5069
Ops.push_back(Node->getOperand(0)); // Chain
5070
Ops.push_back(Node->getOperand(1)); // Ptr
5072
Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5073
Node->getOperand(2), DAG.getIntPtrConstant(0)));
5074
// High part of Val1
5075
Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5076
Node->getOperand(2), DAG.getIntPtrConstant(1)));
5077
if (NewOp == ARMISD::ATOMCMPXCHG64_DAG) {
5078
// High part of Val1
5079
Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5080
Node->getOperand(3), DAG.getIntPtrConstant(0)));
5081
// High part of Val2
5082
Ops.push_back(DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5083
Node->getOperand(3), DAG.getIntPtrConstant(1)));
5085
SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
5087
DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops.data(), Ops.size(), MVT::i64,
5088
cast<MemSDNode>(Node)->getMemOperand());
5089
SDValue OpsF[] = { Result.getValue(0), Result.getValue(1) };
5090
Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
5091
Results.push_back(Result.getValue(2));
5094
SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
5095
switch (Op.getOpcode()) {
5096
default: llvm_unreachable("Don't know how to custom lower this!");
5097
case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
5098
case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
5099
case ISD::GlobalAddress:
5100
return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
5101
LowerGlobalAddressELF(Op, DAG);
5102
case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
5103
case ISD::SELECT: return LowerSELECT(Op, DAG);
5104
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
5105
case ISD::BR_CC: return LowerBR_CC(Op, DAG);
5106
case ISD::BR_JT: return LowerBR_JT(Op, DAG);
5107
case ISD::VASTART: return LowerVASTART(Op, DAG);
5108
case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
5109
case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG, Subtarget);
5110
case ISD::PREFETCH: return LowerPREFETCH(Op, DAG, Subtarget);
5111
case ISD::SINT_TO_FP:
5112
case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
5113
case ISD::FP_TO_SINT:
5114
case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG);
5115
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
5116
case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
5117
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
5118
case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
5119
case ISD::EH_SJLJ_SETJMP: return LowerEH_SJLJ_SETJMP(Op, DAG);
5120
case ISD::EH_SJLJ_LONGJMP: return LowerEH_SJLJ_LONGJMP(Op, DAG);
5121
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG,
5123
case ISD::BITCAST: return ExpandBITCAST(Op.getNode(), DAG);
5126
case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
5127
case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
5128
case ISD::SRL_PARTS:
5129
case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
5130
case ISD::CTTZ: return LowerCTTZ(Op.getNode(), DAG, Subtarget);
5131
case ISD::SETCC: return LowerVSETCC(Op, DAG);
5132
case ISD::ConstantFP: return LowerConstantFP(Op, DAG, Subtarget);
5133
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG, Subtarget);
5134
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
5135
case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
5136
case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
5137
case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
5138
case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
5139
case ISD::MUL: return LowerMUL(Op, DAG);
5140
case ISD::SDIV: return LowerSDIV(Op, DAG);
5141
case ISD::UDIV: return LowerUDIV(Op, DAG);
5145
case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
5146
case ISD::ATOMIC_LOAD:
5147
case ISD::ATOMIC_STORE: return LowerAtomicLoadStore(Op, DAG);
5151
/// ReplaceNodeResults - Replace the results of node with an illegal result
5152
/// type with new values built out of custom code.
5153
void ARMTargetLowering::ReplaceNodeResults(SDNode *N,
5154
SmallVectorImpl<SDValue>&Results,
5155
SelectionDAG &DAG) const {
5157
switch (N->getOpcode()) {
5159
llvm_unreachable("Don't know how to custom expand this!");
5161
Res = ExpandBITCAST(N, DAG);
5165
Res = Expand64BitShift(N, DAG, Subtarget);
5167
case ISD::ATOMIC_LOAD_ADD:
5168
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMADD64_DAG);
5170
case ISD::ATOMIC_LOAD_AND:
5171
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMAND64_DAG);
5173
case ISD::ATOMIC_LOAD_NAND:
5174
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMNAND64_DAG);
5176
case ISD::ATOMIC_LOAD_OR:
5177
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMOR64_DAG);
5179
case ISD::ATOMIC_LOAD_SUB:
5180
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSUB64_DAG);
5182
case ISD::ATOMIC_LOAD_XOR:
5183
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMXOR64_DAG);
5185
case ISD::ATOMIC_SWAP:
5186
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMSWAP64_DAG);
5188
case ISD::ATOMIC_CMP_SWAP:
5189
ReplaceATOMIC_OP_64(N, Results, DAG, ARMISD::ATOMCMPXCHG64_DAG);
5193
Results.push_back(Res);
5196
//===----------------------------------------------------------------------===//
5197
// ARM Scheduler Hooks
5198
//===----------------------------------------------------------------------===//
5201
ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
5202
MachineBasicBlock *BB,
5203
unsigned Size) const {
5204
unsigned dest = MI->getOperand(0).getReg();
5205
unsigned ptr = MI->getOperand(1).getReg();
5206
unsigned oldval = MI->getOperand(2).getReg();
5207
unsigned newval = MI->getOperand(3).getReg();
5208
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5209
DebugLoc dl = MI->getDebugLoc();
5210
bool isThumb2 = Subtarget->isThumb2();
5212
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5214
MRI.createVirtualRegister(isThumb2 ? ARM::rGPRRegisterClass
5215
: ARM::GPRRegisterClass);
5218
MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5219
MRI.constrainRegClass(oldval, ARM::rGPRRegisterClass);
5220
MRI.constrainRegClass(newval, ARM::rGPRRegisterClass);
5223
unsigned ldrOpc, strOpc;
5225
default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5227
ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5228
strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5231
ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5232
strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5235
ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5236
strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5240
MachineFunction *MF = BB->getParent();
5241
const BasicBlock *LLVM_BB = BB->getBasicBlock();
5242
MachineFunction::iterator It = BB;
5243
++It; // insert the new blocks after the current block
5245
MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5246
MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
5247
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5248
MF->insert(It, loop1MBB);
5249
MF->insert(It, loop2MBB);
5250
MF->insert(It, exitMBB);
5252
// Transfer the remainder of BB and its successor edges to exitMBB.
5253
exitMBB->splice(exitMBB->begin(), BB,
5254
llvm::next(MachineBasicBlock::iterator(MI)),
5256
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5260
// fallthrough --> loop1MBB
5261
BB->addSuccessor(loop1MBB);
5264
// ldrex dest, [ptr]
5268
MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5269
if (ldrOpc == ARM::t2LDREX)
5271
AddDefaultPred(MIB);
5272
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5273
.addReg(dest).addReg(oldval));
5274
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5275
.addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5276
BB->addSuccessor(loop2MBB);
5277
BB->addSuccessor(exitMBB);
5280
// strex scratch, newval, [ptr]
5284
MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval).addReg(ptr);
5285
if (strOpc == ARM::t2STREX)
5287
AddDefaultPred(MIB);
5288
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5289
.addReg(scratch).addImm(0));
5290
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5291
.addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5292
BB->addSuccessor(loop1MBB);
5293
BB->addSuccessor(exitMBB);
5299
MI->eraseFromParent(); // The instruction is gone now.
5305
ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
5306
unsigned Size, unsigned BinOpcode) const {
5307
// This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
5308
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5310
const BasicBlock *LLVM_BB = BB->getBasicBlock();
5311
MachineFunction *MF = BB->getParent();
5312
MachineFunction::iterator It = BB;
5315
unsigned dest = MI->getOperand(0).getReg();
5316
unsigned ptr = MI->getOperand(1).getReg();
5317
unsigned incr = MI->getOperand(2).getReg();
5318
DebugLoc dl = MI->getDebugLoc();
5319
bool isThumb2 = Subtarget->isThumb2();
5321
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5323
MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5324
MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5327
unsigned ldrOpc, strOpc;
5329
default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5331
ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5332
strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5335
ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5336
strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5339
ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5340
strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5344
MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5345
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5346
MF->insert(It, loopMBB);
5347
MF->insert(It, exitMBB);
5349
// Transfer the remainder of BB and its successor edges to exitMBB.
5350
exitMBB->splice(exitMBB->begin(), BB,
5351
llvm::next(MachineBasicBlock::iterator(MI)),
5353
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5355
const TargetRegisterClass *TRC =
5356
isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5357
unsigned scratch = MRI.createVirtualRegister(TRC);
5358
unsigned scratch2 = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
5362
// fallthrough --> loopMBB
5363
BB->addSuccessor(loopMBB);
5367
// <binop> scratch2, dest, incr
5368
// strex scratch, scratch2, ptr
5371
// fallthrough --> exitMBB
5373
MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5374
if (ldrOpc == ARM::t2LDREX)
5376
AddDefaultPred(MIB);
5378
// operand order needs to go the other way for NAND
5379
if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
5380
AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5381
addReg(incr).addReg(dest)).addReg(0);
5383
AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
5384
addReg(dest).addReg(incr)).addReg(0);
5387
MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5388
if (strOpc == ARM::t2STREX)
5390
AddDefaultPred(MIB);
5391
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5392
.addReg(scratch).addImm(0));
5393
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5394
.addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5396
BB->addSuccessor(loopMBB);
5397
BB->addSuccessor(exitMBB);
5403
MI->eraseFromParent(); // The instruction is gone now.
5409
ARMTargetLowering::EmitAtomicBinaryMinMax(MachineInstr *MI,
5410
MachineBasicBlock *BB,
5413
ARMCC::CondCodes Cond) const {
5414
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5416
const BasicBlock *LLVM_BB = BB->getBasicBlock();
5417
MachineFunction *MF = BB->getParent();
5418
MachineFunction::iterator It = BB;
5421
unsigned dest = MI->getOperand(0).getReg();
5422
unsigned ptr = MI->getOperand(1).getReg();
5423
unsigned incr = MI->getOperand(2).getReg();
5424
unsigned oldval = dest;
5425
DebugLoc dl = MI->getDebugLoc();
5426
bool isThumb2 = Subtarget->isThumb2();
5428
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5430
MRI.constrainRegClass(dest, ARM::rGPRRegisterClass);
5431
MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5434
unsigned ldrOpc, strOpc, extendOpc;
5436
default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
5438
ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
5439
strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
5440
extendOpc = isThumb2 ? ARM::t2SXTB : ARM::SXTB;
5443
ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
5444
strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
5445
extendOpc = isThumb2 ? ARM::t2SXTH : ARM::SXTH;
5448
ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
5449
strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
5454
MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5455
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5456
MF->insert(It, loopMBB);
5457
MF->insert(It, exitMBB);
5459
// Transfer the remainder of BB and its successor edges to exitMBB.
5460
exitMBB->splice(exitMBB->begin(), BB,
5461
llvm::next(MachineBasicBlock::iterator(MI)),
5463
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5465
const TargetRegisterClass *TRC =
5466
isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5467
unsigned scratch = MRI.createVirtualRegister(TRC);
5468
unsigned scratch2 = MRI.createVirtualRegister(TRC);
5472
// fallthrough --> loopMBB
5473
BB->addSuccessor(loopMBB);
5477
// (sign extend dest, if required)
5479
// cmov.cond scratch2, dest, incr
5480
// strex scratch, scratch2, ptr
5483
// fallthrough --> exitMBB
5485
MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
5486
if (ldrOpc == ARM::t2LDREX)
5488
AddDefaultPred(MIB);
5490
// Sign extend the value, if necessary.
5491
if (signExtend && extendOpc) {
5492
oldval = MRI.createVirtualRegister(ARM::GPRRegisterClass);
5493
AddDefaultPred(BuildMI(BB, dl, TII->get(extendOpc), oldval)
5498
// Build compare and cmov instructions.
5499
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
5500
.addReg(oldval).addReg(incr));
5501
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr), scratch2)
5502
.addReg(oldval).addReg(incr).addImm(Cond).addReg(ARM::CPSR);
5504
MIB = BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2).addReg(ptr);
5505
if (strOpc == ARM::t2STREX)
5507
AddDefaultPred(MIB);
5508
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5509
.addReg(scratch).addImm(0));
5510
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5511
.addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5513
BB->addSuccessor(loopMBB);
5514
BB->addSuccessor(exitMBB);
5520
MI->eraseFromParent(); // The instruction is gone now.
5526
ARMTargetLowering::EmitAtomicBinary64(MachineInstr *MI, MachineBasicBlock *BB,
5527
unsigned Op1, unsigned Op2,
5528
bool NeedsCarry, bool IsCmpxchg) const {
5529
// This also handles ATOMIC_SWAP, indicated by Op1==0.
5530
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5532
const BasicBlock *LLVM_BB = BB->getBasicBlock();
5533
MachineFunction *MF = BB->getParent();
5534
MachineFunction::iterator It = BB;
5537
unsigned destlo = MI->getOperand(0).getReg();
5538
unsigned desthi = MI->getOperand(1).getReg();
5539
unsigned ptr = MI->getOperand(2).getReg();
5540
unsigned vallo = MI->getOperand(3).getReg();
5541
unsigned valhi = MI->getOperand(4).getReg();
5542
DebugLoc dl = MI->getDebugLoc();
5543
bool isThumb2 = Subtarget->isThumb2();
5545
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
5547
MRI.constrainRegClass(destlo, ARM::rGPRRegisterClass);
5548
MRI.constrainRegClass(desthi, ARM::rGPRRegisterClass);
5549
MRI.constrainRegClass(ptr, ARM::rGPRRegisterClass);
5552
unsigned ldrOpc = isThumb2 ? ARM::t2LDREXD : ARM::LDREXD;
5553
unsigned strOpc = isThumb2 ? ARM::t2STREXD : ARM::STREXD;
5555
MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5556
MachineBasicBlock *contBB = 0, *cont2BB = 0;
5558
contBB = MF->CreateMachineBasicBlock(LLVM_BB);
5559
cont2BB = MF->CreateMachineBasicBlock(LLVM_BB);
5561
MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
5562
MF->insert(It, loopMBB);
5564
MF->insert(It, contBB);
5565
MF->insert(It, cont2BB);
5567
MF->insert(It, exitMBB);
5569
// Transfer the remainder of BB and its successor edges to exitMBB.
5570
exitMBB->splice(exitMBB->begin(), BB,
5571
llvm::next(MachineBasicBlock::iterator(MI)),
5573
exitMBB->transferSuccessorsAndUpdatePHIs(BB);
5575
const TargetRegisterClass *TRC =
5576
isThumb2 ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5577
unsigned storesuccess = MRI.createVirtualRegister(TRC);
5581
// fallthrough --> loopMBB
5582
BB->addSuccessor(loopMBB);
5585
// ldrexd r2, r3, ptr
5586
// <binopa> r0, r2, incr
5587
// <binopb> r1, r3, incr
5588
// strexd storesuccess, r0, r1, ptr
5589
// cmp storesuccess, #0
5591
// fallthrough --> exitMBB
5593
// Note that the registers are explicitly specified because there is not any
5594
// way to force the register allocator to allocate a register pair.
5596
// FIXME: The hardcoded registers are not necessary for Thumb2, but we
5597
// need to properly enforce the restriction that the two output registers
5598
// for ldrexd must be different.
5601
AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc))
5602
.addReg(ARM::R2, RegState::Define)
5603
.addReg(ARM::R3, RegState::Define).addReg(ptr));
5604
// Copy r2/r3 into dest. (This copy will normally be coalesced.)
5605
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), destlo).addReg(ARM::R2);
5606
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), desthi).addReg(ARM::R3);
5610
for (unsigned i = 0; i < 2; i++) {
5611
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr :
5613
.addReg(i == 0 ? destlo : desthi)
5614
.addReg(i == 0 ? vallo : valhi));
5615
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5616
.addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5617
BB->addSuccessor(exitMBB);
5618
BB->addSuccessor(i == 0 ? contBB : cont2BB);
5619
BB = (i == 0 ? contBB : cont2BB);
5622
// Copy to physregs for strexd
5623
unsigned setlo = MI->getOperand(5).getReg();
5624
unsigned sethi = MI->getOperand(6).getReg();
5625
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(setlo);
5626
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(sethi);
5628
// Perform binary operation
5629
AddDefaultPred(BuildMI(BB, dl, TII->get(Op1), ARM::R0)
5630
.addReg(destlo).addReg(vallo))
5631
.addReg(NeedsCarry ? ARM::CPSR : 0, getDefRegState(NeedsCarry));
5632
AddDefaultPred(BuildMI(BB, dl, TII->get(Op2), ARM::R1)
5633
.addReg(desthi).addReg(valhi)).addReg(0);
5635
// Copy to physregs for strexd
5636
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R0).addReg(vallo);
5637
BuildMI(BB, dl, TII->get(TargetOpcode::COPY), ARM::R1).addReg(valhi);
5641
AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), storesuccess)
5642
.addReg(ARM::R0).addReg(ARM::R1).addReg(ptr));
5644
AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
5645
.addReg(storesuccess).addImm(0));
5646
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
5647
.addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
5649
BB->addSuccessor(loopMBB);
5650
BB->addSuccessor(exitMBB);
5656
MI->eraseFromParent(); // The instruction is gone now.
5661
/// SetupEntryBlockForSjLj - Insert code into the entry block that creates and
5662
/// registers the function context.
5663
void ARMTargetLowering::
5664
SetupEntryBlockForSjLj(MachineInstr *MI, MachineBasicBlock *MBB,
5665
MachineBasicBlock *DispatchBB, int FI) const {
5666
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5667
DebugLoc dl = MI->getDebugLoc();
5668
MachineFunction *MF = MBB->getParent();
5669
MachineRegisterInfo *MRI = &MF->getRegInfo();
5670
MachineConstantPool *MCP = MF->getConstantPool();
5671
ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5672
const Function *F = MF->getFunction();
5674
bool isThumb = Subtarget->isThumb();
5675
bool isThumb2 = Subtarget->isThumb2();
5677
unsigned PCLabelId = AFI->createPICLabelUId();
5678
unsigned PCAdj = (isThumb || isThumb2) ? 4 : 8;
5679
ARMConstantPoolValue *CPV =
5680
ARMConstantPoolMBB::Create(F->getContext(), DispatchBB, PCLabelId, PCAdj);
5681
unsigned CPI = MCP->getConstantPoolIndex(CPV, 4);
5683
const TargetRegisterClass *TRC =
5684
isThumb ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5686
// Grab constant pool and fixed stack memory operands.
5687
MachineMemOperand *CPMMO =
5688
MF->getMachineMemOperand(MachinePointerInfo::getConstantPool(),
5689
MachineMemOperand::MOLoad, 4, 4);
5691
MachineMemOperand *FIMMOSt =
5692
MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5693
MachineMemOperand::MOStore, 4, 4);
5695
// Load the address of the dispatch MBB into the jump buffer.
5697
// Incoming value: jbuf
5698
// ldr.n r5, LCPI1_1
5701
// str r5, [$jbuf, #+4] ; &jbuf[1]
5702
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5703
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2LDRpci), NewVReg1)
5704
.addConstantPoolIndex(CPI)
5705
.addMemOperand(CPMMO));
5706
// Set the low bit because of thumb mode.
5707
unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5709
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2ORRri), NewVReg2)
5710
.addReg(NewVReg1, RegState::Kill)
5712
unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5713
BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg3)
5714
.addReg(NewVReg2, RegState::Kill)
5716
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::t2STRi12))
5717
.addReg(NewVReg3, RegState::Kill)
5719
.addImm(36) // &jbuf[1] :: pc
5720
.addMemOperand(FIMMOSt));
5721
} else if (isThumb) {
5722
// Incoming value: jbuf
5723
// ldr.n r1, LCPI1_4
5727
// add r2, $jbuf, #+4 ; &jbuf[1]
5729
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5730
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tLDRpci), NewVReg1)
5731
.addConstantPoolIndex(CPI)
5732
.addMemOperand(CPMMO));
5733
unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5734
BuildMI(*MBB, MI, dl, TII->get(ARM::tPICADD), NewVReg2)
5735
.addReg(NewVReg1, RegState::Kill)
5737
// Set the low bit because of thumb mode.
5738
unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5739
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tMOVi8), NewVReg3)
5740
.addReg(ARM::CPSR, RegState::Define)
5742
unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5743
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tORR), NewVReg4)
5744
.addReg(ARM::CPSR, RegState::Define)
5745
.addReg(NewVReg2, RegState::Kill)
5746
.addReg(NewVReg3, RegState::Kill));
5747
unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5748
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tADDrSPi), NewVReg5)
5750
.addImm(36)); // &jbuf[1] :: pc
5751
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::tSTRi))
5752
.addReg(NewVReg4, RegState::Kill)
5753
.addReg(NewVReg5, RegState::Kill)
5755
.addMemOperand(FIMMOSt));
5757
// Incoming value: jbuf
5760
// str r1, [$jbuf, #+4] ; &jbuf[1]
5761
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5762
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::LDRi12), NewVReg1)
5763
.addConstantPoolIndex(CPI)
5765
.addMemOperand(CPMMO));
5766
unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5767
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::PICADD), NewVReg2)
5768
.addReg(NewVReg1, RegState::Kill)
5769
.addImm(PCLabelId));
5770
AddDefaultPred(BuildMI(*MBB, MI, dl, TII->get(ARM::STRi12))
5771
.addReg(NewVReg2, RegState::Kill)
5773
.addImm(36) // &jbuf[1] :: pc
5774
.addMemOperand(FIMMOSt));
5778
MachineBasicBlock *ARMTargetLowering::
5779
EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
5780
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
5781
DebugLoc dl = MI->getDebugLoc();
5782
MachineFunction *MF = MBB->getParent();
5783
MachineRegisterInfo *MRI = &MF->getRegInfo();
5784
ARMFunctionInfo *AFI = MF->getInfo<ARMFunctionInfo>();
5785
MachineFrameInfo *MFI = MF->getFrameInfo();
5786
int FI = MFI->getFunctionContextIndex();
5788
const TargetRegisterClass *TRC =
5789
Subtarget->isThumb() ? ARM::tGPRRegisterClass : ARM::GPRRegisterClass;
5791
// Get a mapping of the call site numbers to all of the landing pads they're
5793
DenseMap<unsigned, SmallVector<MachineBasicBlock*, 2> > CallSiteNumToLPad;
5794
unsigned MaxCSNum = 0;
5795
MachineModuleInfo &MMI = MF->getMMI();
5796
for (MachineFunction::iterator BB = MF->begin(), E = MF->end(); BB != E; ++BB) {
5797
if (!BB->isLandingPad()) continue;
5799
// FIXME: We should assert that the EH_LABEL is the first MI in the landing
5801
for (MachineBasicBlock::iterator
5802
II = BB->begin(), IE = BB->end(); II != IE; ++II) {
5803
if (!II->isEHLabel()) continue;
5805
MCSymbol *Sym = II->getOperand(0).getMCSymbol();
5806
if (!MMI.hasCallSiteLandingPad(Sym)) continue;
5808
SmallVectorImpl<unsigned> &CallSiteIdxs = MMI.getCallSiteLandingPad(Sym);
5809
for (SmallVectorImpl<unsigned>::iterator
5810
CSI = CallSiteIdxs.begin(), CSE = CallSiteIdxs.end();
5811
CSI != CSE; ++CSI) {
5812
CallSiteNumToLPad[*CSI].push_back(BB);
5813
MaxCSNum = std::max(MaxCSNum, *CSI);
5819
// Get an ordered list of the machine basic blocks for the jump table.
5820
std::vector<MachineBasicBlock*> LPadList;
5821
SmallPtrSet<MachineBasicBlock*, 64> InvokeBBs;
5822
LPadList.reserve(CallSiteNumToLPad.size());
5823
for (unsigned I = 1; I <= MaxCSNum; ++I) {
5824
SmallVectorImpl<MachineBasicBlock*> &MBBList = CallSiteNumToLPad[I];
5825
for (SmallVectorImpl<MachineBasicBlock*>::iterator
5826
II = MBBList.begin(), IE = MBBList.end(); II != IE; ++II) {
5827
LPadList.push_back(*II);
5828
InvokeBBs.insert((*II)->pred_begin(), (*II)->pred_end());
5832
assert(!LPadList.empty() &&
5833
"No landing pad destinations for the dispatch jump table!");
5835
// Create the jump table and associated information.
5836
MachineJumpTableInfo *JTI =
5837
MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
5838
unsigned MJTI = JTI->createJumpTableIndex(LPadList);
5839
unsigned UId = AFI->createJumpTableUId();
5841
// Create the MBBs for the dispatch code.
5843
// Shove the dispatch's address into the return slot in the function context.
5844
MachineBasicBlock *DispatchBB = MF->CreateMachineBasicBlock();
5845
DispatchBB->setIsLandingPad();
5847
MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
5848
BuildMI(TrapBB, dl, TII->get(Subtarget->isThumb() ? ARM::tTRAP : ARM::TRAP));
5849
DispatchBB->addSuccessor(TrapBB);
5851
MachineBasicBlock *DispContBB = MF->CreateMachineBasicBlock();
5852
DispatchBB->addSuccessor(DispContBB);
5855
MF->insert(MF->end(), DispatchBB);
5856
MF->insert(MF->end(), DispContBB);
5857
MF->insert(MF->end(), TrapBB);
5859
// Insert code into the entry block that creates and registers the function
5861
SetupEntryBlockForSjLj(MI, MBB, DispatchBB, FI);
5863
MachineMemOperand *FIMMOLd =
5864
MF->getMachineMemOperand(MachinePointerInfo::getFixedStack(FI),
5865
MachineMemOperand::MOLoad |
5866
MachineMemOperand::MOVolatile, 4, 4);
5868
if (AFI->isThumb1OnlyFunction())
5869
BuildMI(DispatchBB, dl, TII->get(ARM::tInt_eh_sjlj_dispatchsetup));
5870
else if (!Subtarget->hasVFP2())
5871
BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup_nofp));
5873
BuildMI(DispatchBB, dl, TII->get(ARM::Int_eh_sjlj_dispatchsetup));
5875
unsigned NumLPads = LPadList.size();
5876
if (Subtarget->isThumb2()) {
5877
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5878
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2LDRi12), NewVReg1)
5881
.addMemOperand(FIMMOLd));
5883
if (NumLPads < 256) {
5884
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPri))
5886
.addImm(LPadList.size()));
5888
unsigned VReg1 = MRI->createVirtualRegister(TRC);
5889
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVi16), VReg1)
5890
.addImm(NumLPads & 0xFFFF));
5892
unsigned VReg2 = VReg1;
5893
if ((NumLPads & 0xFFFF0000) != 0) {
5894
VReg2 = MRI->createVirtualRegister(TRC);
5895
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2MOVTi16), VReg2)
5897
.addImm(NumLPads >> 16));
5900
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::t2CMPrr))
5905
BuildMI(DispatchBB, dl, TII->get(ARM::t2Bcc))
5910
unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5911
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::t2LEApcrelJT),NewVReg3)
5912
.addJumpTableIndex(MJTI)
5915
unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5918
BuildMI(DispContBB, dl, TII->get(ARM::t2ADDrs), NewVReg4)
5919
.addReg(NewVReg3, RegState::Kill)
5921
.addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
5923
BuildMI(DispContBB, dl, TII->get(ARM::t2BR_JT))
5924
.addReg(NewVReg4, RegState::Kill)
5926
.addJumpTableIndex(MJTI)
5928
} else if (Subtarget->isThumb()) {
5929
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
5930
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRspi), NewVReg1)
5933
.addMemOperand(FIMMOLd));
5935
if (NumLPads < 256) {
5936
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPi8))
5940
MachineConstantPool *ConstantPool = MF->getConstantPool();
5941
Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
5942
const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
5944
// MachineConstantPool wants an explicit alignment.
5945
unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
5947
Align = getTargetData()->getTypeAllocSize(C->getType());
5948
unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
5950
unsigned VReg1 = MRI->createVirtualRegister(TRC);
5951
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tLDRpci))
5952
.addReg(VReg1, RegState::Define)
5953
.addConstantPoolIndex(Idx));
5954
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::tCMPr))
5959
BuildMI(DispatchBB, dl, TII->get(ARM::tBcc))
5964
unsigned NewVReg2 = MRI->createVirtualRegister(TRC);
5965
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLSLri), NewVReg2)
5966
.addReg(ARM::CPSR, RegState::Define)
5970
unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
5971
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLEApcrelJT), NewVReg3)
5972
.addJumpTableIndex(MJTI)
5975
unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
5976
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg4)
5977
.addReg(ARM::CPSR, RegState::Define)
5978
.addReg(NewVReg2, RegState::Kill)
5981
MachineMemOperand *JTMMOLd =
5982
MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
5983
MachineMemOperand::MOLoad, 4, 4);
5985
unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
5986
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tLDRi), NewVReg5)
5987
.addReg(NewVReg4, RegState::Kill)
5989
.addMemOperand(JTMMOLd));
5991
unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
5992
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
5993
.addReg(ARM::CPSR, RegState::Define)
5994
.addReg(NewVReg5, RegState::Kill)
5997
BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
5998
.addReg(NewVReg6, RegState::Kill)
5999
.addJumpTableIndex(MJTI)
6002
unsigned NewVReg1 = MRI->createVirtualRegister(TRC);
6003
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRi12), NewVReg1)
6006
.addMemOperand(FIMMOLd));
6008
if (NumLPads < 256) {
6009
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPri))
6012
} else if (Subtarget->hasV6T2Ops() && isUInt<16>(NumLPads)) {
6013
unsigned VReg1 = MRI->createVirtualRegister(TRC);
6014
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVi16), VReg1)
6015
.addImm(NumLPads & 0xFFFF));
6017
unsigned VReg2 = VReg1;
6018
if ((NumLPads & 0xFFFF0000) != 0) {
6019
VReg2 = MRI->createVirtualRegister(TRC);
6020
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::MOVTi16), VReg2)
6022
.addImm(NumLPads >> 16));
6025
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6029
MachineConstantPool *ConstantPool = MF->getConstantPool();
6030
Type *Int32Ty = Type::getInt32Ty(MF->getFunction()->getContext());
6031
const Constant *C = ConstantInt::get(Int32Ty, NumLPads);
6033
// MachineConstantPool wants an explicit alignment.
6034
unsigned Align = getTargetData()->getPrefTypeAlignment(Int32Ty);
6036
Align = getTargetData()->getTypeAllocSize(C->getType());
6037
unsigned Idx = ConstantPool->getConstantPoolIndex(C, Align);
6039
unsigned VReg1 = MRI->createVirtualRegister(TRC);
6040
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::LDRcp))
6041
.addReg(VReg1, RegState::Define)
6042
.addConstantPoolIndex(Idx)
6044
AddDefaultPred(BuildMI(DispatchBB, dl, TII->get(ARM::CMPrr))
6046
.addReg(VReg1, RegState::Kill));
6049
BuildMI(DispatchBB, dl, TII->get(ARM::Bcc))
6054
unsigned NewVReg3 = MRI->createVirtualRegister(TRC);
6056
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::MOVsi), NewVReg3)
6058
.addImm(ARM_AM::getSORegOpc(ARM_AM::lsl, 2))));
6059
unsigned NewVReg4 = MRI->createVirtualRegister(TRC);
6060
AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::LEApcrelJT), NewVReg4)
6061
.addJumpTableIndex(MJTI)
6064
MachineMemOperand *JTMMOLd =
6065
MF->getMachineMemOperand(MachinePointerInfo::getJumpTable(),
6066
MachineMemOperand::MOLoad, 4, 4);
6067
unsigned NewVReg5 = MRI->createVirtualRegister(TRC);
6069
BuildMI(DispContBB, dl, TII->get(ARM::LDRrs), NewVReg5)
6070
.addReg(NewVReg3, RegState::Kill)
6073
.addMemOperand(JTMMOLd));
6075
BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
6076
.addReg(NewVReg5, RegState::Kill)
6078
.addJumpTableIndex(MJTI)
6082
// Add the jump table entries as successors to the MBB.
6083
MachineBasicBlock *PrevMBB = 0;
6084
for (std::vector<MachineBasicBlock*>::iterator
6085
I = LPadList.begin(), E = LPadList.end(); I != E; ++I) {
6086
MachineBasicBlock *CurMBB = *I;
6087
if (PrevMBB != CurMBB)
6088
DispContBB->addSuccessor(CurMBB);
6092
// N.B. the order the invoke BBs are processed in doesn't matter here.
6093
const ARMBaseInstrInfo *AII = static_cast<const ARMBaseInstrInfo*>(TII);
6094
const ARMBaseRegisterInfo &RI = AII->getRegisterInfo();
6095
const uint16_t *SavedRegs = RI.getCalleeSavedRegs(MF);
6096
SmallVector<MachineBasicBlock*, 64> MBBLPads;
6097
for (SmallPtrSet<MachineBasicBlock*, 64>::iterator
6098
I = InvokeBBs.begin(), E = InvokeBBs.end(); I != E; ++I) {
6099
MachineBasicBlock *BB = *I;
6101
// Remove the landing pad successor from the invoke block and replace it
6102
// with the new dispatch block.
6103
SmallVector<MachineBasicBlock*, 4> Successors(BB->succ_begin(),
6105
while (!Successors.empty()) {
6106
MachineBasicBlock *SMBB = Successors.pop_back_val();
6107
if (SMBB->isLandingPad()) {
6108
BB->removeSuccessor(SMBB);
6109
MBBLPads.push_back(SMBB);
6113
BB->addSuccessor(DispatchBB);
6115
// Find the invoke call and mark all of the callee-saved registers as
6116
// 'implicit defined' so that they're spilled. This prevents code from
6117
// moving instructions to before the EH block, where they will never be
6119
for (MachineBasicBlock::reverse_iterator
6120
II = BB->rbegin(), IE = BB->rend(); II != IE; ++II) {
6121
if (!II->isCall()) continue;
6123
DenseMap<unsigned, bool> DefRegs;
6124
for (MachineInstr::mop_iterator
6125
OI = II->operands_begin(), OE = II->operands_end();
6127
if (!OI->isReg()) continue;
6128
DefRegs[OI->getReg()] = true;
6131
MachineInstrBuilder MIB(&*II);
6133
for (unsigned i = 0; SavedRegs[i] != 0; ++i) {
6134
unsigned Reg = SavedRegs[i];
6135
if (Subtarget->isThumb2() &&
6136
!ARM::tGPRRegisterClass->contains(Reg) &&
6137
!ARM::hGPRRegisterClass->contains(Reg))
6139
else if (Subtarget->isThumb1Only() &&
6140
!ARM::tGPRRegisterClass->contains(Reg))
6142
else if (!Subtarget->isThumb() &&
6143
!ARM::GPRRegisterClass->contains(Reg))
6146
MIB.addReg(Reg, RegState::ImplicitDefine | RegState::Dead);
6153
// Mark all former landing pads as non-landing pads. The dispatch is the only
6155
for (SmallVectorImpl<MachineBasicBlock*>::iterator
6156
I = MBBLPads.begin(), E = MBBLPads.end(); I != E; ++I)
6157
(*I)->setIsLandingPad(false);
6159
// The instruction is gone now.
6160
MI->eraseFromParent();
6166
MachineBasicBlock *OtherSucc(MachineBasicBlock *MBB, MachineBasicBlock *Succ) {
6167
for (MachineBasicBlock::succ_iterator I = MBB->succ_begin(),
6168
E = MBB->succ_end(); I != E; ++I)
6171
llvm_unreachable("Expecting a BB with two successors!");
6175
ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
6176
MachineBasicBlock *BB) const {
6177
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
6178
DebugLoc dl = MI->getDebugLoc();
6179
bool isThumb2 = Subtarget->isThumb2();
6180
switch (MI->getOpcode()) {
6183
llvm_unreachable("Unexpected instr type to insert");
6185
// The Thumb2 pre-indexed stores have the same MI operands, they just
6186
// define them differently in the .td files from the isel patterns, so
6187
// they need pseudos.
6188
case ARM::t2STR_preidx:
6189
MI->setDesc(TII->get(ARM::t2STR_PRE));
6191
case ARM::t2STRB_preidx:
6192
MI->setDesc(TII->get(ARM::t2STRB_PRE));
6194
case ARM::t2STRH_preidx:
6195
MI->setDesc(TII->get(ARM::t2STRH_PRE));
6198
case ARM::STRi_preidx:
6199
case ARM::STRBi_preidx: {
6200
unsigned NewOpc = MI->getOpcode() == ARM::STRi_preidx ?
6201
ARM::STR_PRE_IMM : ARM::STRB_PRE_IMM;
6202
// Decode the offset.
6203
unsigned Offset = MI->getOperand(4).getImm();
6204
bool isSub = ARM_AM::getAM2Op(Offset) == ARM_AM::sub;
6205
Offset = ARM_AM::getAM2Offset(Offset);
6209
MachineMemOperand *MMO = *MI->memoperands_begin();
6210
BuildMI(*BB, MI, dl, TII->get(NewOpc))
6211
.addOperand(MI->getOperand(0)) // Rn_wb
6212
.addOperand(MI->getOperand(1)) // Rt
6213
.addOperand(MI->getOperand(2)) // Rn
6214
.addImm(Offset) // offset (skip GPR==zero_reg)
6215
.addOperand(MI->getOperand(5)) // pred
6216
.addOperand(MI->getOperand(6))
6217
.addMemOperand(MMO);
6218
MI->eraseFromParent();
6221
case ARM::STRr_preidx:
6222
case ARM::STRBr_preidx:
6223
case ARM::STRH_preidx: {
6225
switch (MI->getOpcode()) {
6226
default: llvm_unreachable("unexpected opcode!");
6227
case ARM::STRr_preidx: NewOpc = ARM::STR_PRE_REG; break;
6228
case ARM::STRBr_preidx: NewOpc = ARM::STRB_PRE_REG; break;
6229
case ARM::STRH_preidx: NewOpc = ARM::STRH_PRE; break;
6231
MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(NewOpc));
6232
for (unsigned i = 0; i < MI->getNumOperands(); ++i)
6233
MIB.addOperand(MI->getOperand(i));
6234
MI->eraseFromParent();
6237
case ARM::ATOMIC_LOAD_ADD_I8:
6238
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6239
case ARM::ATOMIC_LOAD_ADD_I16:
6240
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6241
case ARM::ATOMIC_LOAD_ADD_I32:
6242
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
6244
case ARM::ATOMIC_LOAD_AND_I8:
6245
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6246
case ARM::ATOMIC_LOAD_AND_I16:
6247
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6248
case ARM::ATOMIC_LOAD_AND_I32:
6249
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6251
case ARM::ATOMIC_LOAD_OR_I8:
6252
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6253
case ARM::ATOMIC_LOAD_OR_I16:
6254
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6255
case ARM::ATOMIC_LOAD_OR_I32:
6256
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6258
case ARM::ATOMIC_LOAD_XOR_I8:
6259
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6260
case ARM::ATOMIC_LOAD_XOR_I16:
6261
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6262
case ARM::ATOMIC_LOAD_XOR_I32:
6263
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6265
case ARM::ATOMIC_LOAD_NAND_I8:
6266
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6267
case ARM::ATOMIC_LOAD_NAND_I16:
6268
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6269
case ARM::ATOMIC_LOAD_NAND_I32:
6270
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
6272
case ARM::ATOMIC_LOAD_SUB_I8:
6273
return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6274
case ARM::ATOMIC_LOAD_SUB_I16:
6275
return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6276
case ARM::ATOMIC_LOAD_SUB_I32:
6277
return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
6279
case ARM::ATOMIC_LOAD_MIN_I8:
6280
return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::LT);
6281
case ARM::ATOMIC_LOAD_MIN_I16:
6282
return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::LT);
6283
case ARM::ATOMIC_LOAD_MIN_I32:
6284
return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::LT);
6286
case ARM::ATOMIC_LOAD_MAX_I8:
6287
return EmitAtomicBinaryMinMax(MI, BB, 1, true, ARMCC::GT);
6288
case ARM::ATOMIC_LOAD_MAX_I16:
6289
return EmitAtomicBinaryMinMax(MI, BB, 2, true, ARMCC::GT);
6290
case ARM::ATOMIC_LOAD_MAX_I32:
6291
return EmitAtomicBinaryMinMax(MI, BB, 4, true, ARMCC::GT);
6293
case ARM::ATOMIC_LOAD_UMIN_I8:
6294
return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::LO);
6295
case ARM::ATOMIC_LOAD_UMIN_I16:
6296
return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::LO);
6297
case ARM::ATOMIC_LOAD_UMIN_I32:
6298
return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::LO);
6300
case ARM::ATOMIC_LOAD_UMAX_I8:
6301
return EmitAtomicBinaryMinMax(MI, BB, 1, false, ARMCC::HI);
6302
case ARM::ATOMIC_LOAD_UMAX_I16:
6303
return EmitAtomicBinaryMinMax(MI, BB, 2, false, ARMCC::HI);
6304
case ARM::ATOMIC_LOAD_UMAX_I32:
6305
return EmitAtomicBinaryMinMax(MI, BB, 4, false, ARMCC::HI);
6307
case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
6308
case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
6309
case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
6311
case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
6312
case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
6313
case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
6316
case ARM::ATOMADD6432:
6317
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr,
6318
isThumb2 ? ARM::t2ADCrr : ARM::ADCrr,
6319
/*NeedsCarry*/ true);
6320
case ARM::ATOMSUB6432:
6321
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6322
isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6323
/*NeedsCarry*/ true);
6324
case ARM::ATOMOR6432:
6325
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr,
6326
isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
6327
case ARM::ATOMXOR6432:
6328
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2EORrr : ARM::EORrr,
6329
isThumb2 ? ARM::t2EORrr : ARM::EORrr);
6330
case ARM::ATOMAND6432:
6331
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr,
6332
isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
6333
case ARM::ATOMSWAP6432:
6334
return EmitAtomicBinary64(MI, BB, 0, 0, false);
6335
case ARM::ATOMCMPXCHG6432:
6336
return EmitAtomicBinary64(MI, BB, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr,
6337
isThumb2 ? ARM::t2SBCrr : ARM::SBCrr,
6338
/*NeedsCarry*/ false, /*IsCmpxchg*/true);
6340
case ARM::tMOVCCr_pseudo: {
6341
// To "insert" a SELECT_CC instruction, we actually have to insert the
6342
// diamond control-flow pattern. The incoming instruction knows the
6343
// destination vreg to set, the condition code register to branch on, the
6344
// true/false values to select between, and a branch opcode to use.
6345
const BasicBlock *LLVM_BB = BB->getBasicBlock();
6346
MachineFunction::iterator It = BB;
6352
// cmpTY ccX, r1, r2
6354
// fallthrough --> copy0MBB
6355
MachineBasicBlock *thisMBB = BB;
6356
MachineFunction *F = BB->getParent();
6357
MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
6358
MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
6359
F->insert(It, copy0MBB);
6360
F->insert(It, sinkMBB);
6362
// Transfer the remainder of BB and its successor edges to sinkMBB.
6363
sinkMBB->splice(sinkMBB->begin(), BB,
6364
llvm::next(MachineBasicBlock::iterator(MI)),
6366
sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
6368
BB->addSuccessor(copy0MBB);
6369
BB->addSuccessor(sinkMBB);
6371
BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB)
6372
.addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg());
6375
// %FalseValue = ...
6376
// # fallthrough to sinkMBB
6379
// Update machine-CFG edges
6380
BB->addSuccessor(sinkMBB);
6383
// %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
6386
BuildMI(*BB, BB->begin(), dl,
6387
TII->get(ARM::PHI), MI->getOperand(0).getReg())
6388
.addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
6389
.addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
6391
MI->eraseFromParent(); // The pseudo instruction is gone now.
6396
case ARM::BCCZi64: {
6397
// If there is an unconditional branch to the other successor, remove it.
6398
BB->erase(llvm::next(MachineBasicBlock::iterator(MI)), BB->end());
6400
// Compare both parts that make up the double comparison separately for
6402
bool RHSisZero = MI->getOpcode() == ARM::BCCZi64;
6404
unsigned LHS1 = MI->getOperand(1).getReg();
6405
unsigned LHS2 = MI->getOperand(2).getReg();
6407
AddDefaultPred(BuildMI(BB, dl,
6408
TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6409
.addReg(LHS1).addImm(0));
6410
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
6411
.addReg(LHS2).addImm(0)
6412
.addImm(ARMCC::EQ).addReg(ARM::CPSR);
6414
unsigned RHS1 = MI->getOperand(3).getReg();
6415
unsigned RHS2 = MI->getOperand(4).getReg();
6416
AddDefaultPred(BuildMI(BB, dl,
6417
TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6418
.addReg(LHS1).addReg(RHS1));
6419
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
6420
.addReg(LHS2).addReg(RHS2)
6421
.addImm(ARMCC::EQ).addReg(ARM::CPSR);
6424
MachineBasicBlock *destMBB = MI->getOperand(RHSisZero ? 3 : 5).getMBB();
6425
MachineBasicBlock *exitMBB = OtherSucc(BB, destMBB);
6426
if (MI->getOperand(0).getImm() == ARMCC::NE)
6427
std::swap(destMBB, exitMBB);
6429
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
6430
.addMBB(destMBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
6432
AddDefaultPred(BuildMI(BB, dl, TII->get(ARM::t2B)).addMBB(exitMBB));
6434
BuildMI(BB, dl, TII->get(ARM::B)) .addMBB(exitMBB);
6436
MI->eraseFromParent(); // The pseudo instruction is gone now.
6440
case ARM::Int_eh_sjlj_setjmp:
6441
case ARM::Int_eh_sjlj_setjmp_nofp:
6442
case ARM::tInt_eh_sjlj_setjmp:
6443
case ARM::t2Int_eh_sjlj_setjmp:
6444
case ARM::t2Int_eh_sjlj_setjmp_nofp:
6445
EmitSjLjDispatchBlock(MI, BB);
6450
// To insert an ABS instruction, we have to insert the
6451
// diamond control-flow pattern. The incoming instruction knows the
6452
// source vreg to test against 0, the destination vreg to set,
6453
// the condition code register to branch on, the
6454
// true/false values to select between, and a branch opcode to use.
6459
// BCC (branch to SinkBB if V0 >= 0)
6460
// RSBBB: V3 = RSBri V2, 0 (compute ABS if V2 < 0)
6461
// SinkBB: V1 = PHI(V2, V3)
6462
const BasicBlock *LLVM_BB = BB->getBasicBlock();
6463
MachineFunction::iterator BBI = BB;
6465
MachineFunction *Fn = BB->getParent();
6466
MachineBasicBlock *RSBBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6467
MachineBasicBlock *SinkBB = Fn->CreateMachineBasicBlock(LLVM_BB);
6468
Fn->insert(BBI, RSBBB);
6469
Fn->insert(BBI, SinkBB);
6471
unsigned int ABSSrcReg = MI->getOperand(1).getReg();
6472
unsigned int ABSDstReg = MI->getOperand(0).getReg();
6473
bool isThumb2 = Subtarget->isThumb2();
6474
MachineRegisterInfo &MRI = Fn->getRegInfo();
6475
// In Thumb mode S must not be specified if source register is the SP or
6476
// PC and if destination register is the SP, so restrict register class
6477
unsigned NewMovDstReg = MRI.createVirtualRegister(
6478
isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6479
unsigned NewRsbDstReg = MRI.createVirtualRegister(
6480
isThumb2 ? ARM::rGPRRegisterClass : ARM::GPRRegisterClass);
6482
// Transfer the remainder of BB and its successor edges to sinkMBB.
6483
SinkBB->splice(SinkBB->begin(), BB,
6484
llvm::next(MachineBasicBlock::iterator(MI)),
6486
SinkBB->transferSuccessorsAndUpdatePHIs(BB);
6488
BB->addSuccessor(RSBBB);
6489
BB->addSuccessor(SinkBB);
6491
// fall through to SinkMBB
6492
RSBBB->addSuccessor(SinkBB);
6494
// insert a movs at the end of BB
6495
BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2MOVr : ARM::MOVr),
6497
.addReg(ABSSrcReg, RegState::Kill)
6498
.addImm((unsigned)ARMCC::AL).addReg(0)
6499
.addReg(ARM::CPSR, RegState::Define);
6501
// insert a bcc with opposite CC to ARMCC::MI at the end of BB
6503
TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc)).addMBB(SinkBB)
6504
.addImm(ARMCC::getOppositeCondition(ARMCC::MI)).addReg(ARM::CPSR);
6506
// insert rsbri in RSBBB
6507
// Note: BCC and rsbri will be converted into predicated rsbmi
6508
// by if-conversion pass
6509
BuildMI(*RSBBB, RSBBB->begin(), dl,
6510
TII->get(isThumb2 ? ARM::t2RSBri : ARM::RSBri), NewRsbDstReg)
6511
.addReg(NewMovDstReg, RegState::Kill)
6512
.addImm(0).addImm((unsigned)ARMCC::AL).addReg(0).addReg(0);
6514
// insert PHI in SinkBB,
6515
// reuse ABSDstReg to not change uses of ABS instruction
6516
BuildMI(*SinkBB, SinkBB->begin(), dl,
6517
TII->get(ARM::PHI), ABSDstReg)
6518
.addReg(NewRsbDstReg).addMBB(RSBBB)
6519
.addReg(NewMovDstReg).addMBB(BB);
6521
// remove ABS instruction
6522
MI->eraseFromParent();
6524
// return last added BB
6530
void ARMTargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
6531
SDNode *Node) const {
6532
if (!MI->hasPostISelHook()) {
6533
assert(!convertAddSubFlagsOpcode(MI->getOpcode()) &&
6534
"Pseudo flag-setting opcodes must be marked with 'hasPostISelHook'");
6538
const MCInstrDesc *MCID = &MI->getDesc();
6539
// Adjust potentially 's' setting instructions after isel, i.e. ADC, SBC, RSB,
6540
// RSC. Coming out of isel, they have an implicit CPSR def, but the optional
6541
// operand is still set to noreg. If needed, set the optional operand's
6542
// register to CPSR, and remove the redundant implicit def.
6544
// e.g. ADCS (..., CPSR<imp-def>) -> ADC (... opt:CPSR<def>).
6546
// Rename pseudo opcodes.
6547
unsigned NewOpc = convertAddSubFlagsOpcode(MI->getOpcode());
6549
const ARMBaseInstrInfo *TII =
6550
static_cast<const ARMBaseInstrInfo*>(getTargetMachine().getInstrInfo());
6551
MCID = &TII->get(NewOpc);
6553
assert(MCID->getNumOperands() == MI->getDesc().getNumOperands() + 1 &&
6554
"converted opcode should be the same except for cc_out");
6558
// Add the optional cc_out operand
6559
MI->addOperand(MachineOperand::CreateReg(0, /*isDef=*/true));
6561
unsigned ccOutIdx = MCID->getNumOperands() - 1;
6563
// Any ARM instruction that sets the 's' bit should specify an optional
6564
// "cc_out" operand in the last operand position.
6565
if (!MI->hasOptionalDef() || !MCID->OpInfo[ccOutIdx].isOptionalDef()) {
6566
assert(!NewOpc && "Optional cc_out operand required");
6569
// Look for an implicit def of CPSR added by MachineInstr ctor. Remove it
6570
// since we already have an optional CPSR def.
6571
bool definesCPSR = false;
6572
bool deadCPSR = false;
6573
for (unsigned i = MCID->getNumOperands(), e = MI->getNumOperands();
6575
const MachineOperand &MO = MI->getOperand(i);
6576
if (MO.isReg() && MO.isDef() && MO.getReg() == ARM::CPSR) {
6580
MI->RemoveOperand(i);
6585
assert(!NewOpc && "Optional cc_out operand required");
6588
assert(deadCPSR == !Node->hasAnyUseOfValue(1) && "inconsistent dead flag");
6590
assert(!MI->getOperand(ccOutIdx).getReg() &&
6591
"expect uninitialized optional cc_out operand");
6595
// If this instruction was defined with an optional CPSR def and its dag node
6596
// had a live implicit CPSR def, then activate the optional CPSR def.
6597
MachineOperand &MO = MI->getOperand(ccOutIdx);
6598
MO.setReg(ARM::CPSR);
6602
//===----------------------------------------------------------------------===//
6603
// ARM Optimization Hooks
6604
//===----------------------------------------------------------------------===//
6607
SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
6608
TargetLowering::DAGCombinerInfo &DCI) {
6609
SelectionDAG &DAG = DCI.DAG;
6610
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6611
EVT VT = N->getValueType(0);
6612
unsigned Opc = N->getOpcode();
6613
bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC;
6614
SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1);
6615
SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2);
6616
ISD::CondCode CC = ISD::SETCC_INVALID;
6619
CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get();
6621
SDValue CCOp = Slct.getOperand(0);
6622
if (CCOp.getOpcode() == ISD::SETCC)
6623
CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get();
6626
bool DoXform = false;
6628
assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) &&
6631
if (LHS.getOpcode() == ISD::Constant &&
6632
cast<ConstantSDNode>(LHS)->isNullValue()) {
6634
} else if (CC != ISD::SETCC_INVALID &&
6635
RHS.getOpcode() == ISD::Constant &&
6636
cast<ConstantSDNode>(RHS)->isNullValue()) {
6637
std::swap(LHS, RHS);
6638
SDValue Op0 = Slct.getOperand(0);
6639
EVT OpVT = isSlctCC ? Op0.getValueType() :
6640
Op0.getOperand(0).getValueType();
6641
bool isInt = OpVT.isInteger();
6642
CC = ISD::getSetCCInverse(CC, isInt);
6644
if (!TLI.isCondCodeLegal(CC, OpVT))
6645
return SDValue(); // Inverse operator isn't legal.
6652
SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS);
6654
return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result,
6655
Slct.getOperand(0), Slct.getOperand(1), CC);
6656
SDValue CCOp = Slct.getOperand(0);
6658
CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(),
6659
CCOp.getOperand(0), CCOp.getOperand(1), CC);
6660
return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
6661
CCOp, OtherOp, Result);
6666
// AddCombineToVPADDL- For pair-wise add on neon, use the vpaddl instruction
6667
// (only after legalization).
6668
static SDValue AddCombineToVPADDL(SDNode *N, SDValue N0, SDValue N1,
6669
TargetLowering::DAGCombinerInfo &DCI,
6670
const ARMSubtarget *Subtarget) {
6672
// Only perform optimization if after legalize, and if NEON is available. We
6673
// also expected both operands to be BUILD_VECTORs.
6674
if (DCI.isBeforeLegalize() || !Subtarget->hasNEON()
6675
|| N0.getOpcode() != ISD::BUILD_VECTOR
6676
|| N1.getOpcode() != ISD::BUILD_VECTOR)
6679
// Check output type since VPADDL operand elements can only be 8, 16, or 32.
6680
EVT VT = N->getValueType(0);
6681
if (!VT.isInteger() || VT.getVectorElementType() == MVT::i64)
6684
// Check that the vector operands are of the right form.
6685
// N0 and N1 are BUILD_VECTOR nodes with N number of EXTRACT_VECTOR
6686
// operands, where N is the size of the formed vector.
6687
// Each EXTRACT_VECTOR should have the same input vector and odd or even
6688
// index such that we have a pair wise add pattern.
6690
// Grab the vector that all EXTRACT_VECTOR nodes should be referencing.
6691
if (N0->getOperand(0)->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6693
SDValue Vec = N0->getOperand(0)->getOperand(0);
6694
SDNode *V = Vec.getNode();
6695
unsigned nextIndex = 0;
6697
// For each operands to the ADD which are BUILD_VECTORs,
6698
// check to see if each of their operands are an EXTRACT_VECTOR with
6699
// the same vector and appropriate index.
6700
for (unsigned i = 0, e = N0->getNumOperands(); i != e; ++i) {
6701
if (N0->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT
6702
&& N1->getOperand(i)->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
6704
SDValue ExtVec0 = N0->getOperand(i);
6705
SDValue ExtVec1 = N1->getOperand(i);
6707
// First operand is the vector, verify its the same.
6708
if (V != ExtVec0->getOperand(0).getNode() ||
6709
V != ExtVec1->getOperand(0).getNode())
6712
// Second is the constant, verify its correct.
6713
ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(ExtVec0->getOperand(1));
6714
ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(ExtVec1->getOperand(1));
6716
// For the constant, we want to see all the even or all the odd.
6717
if (!C0 || !C1 || C0->getZExtValue() != nextIndex
6718
|| C1->getZExtValue() != nextIndex+1)
6727
// Create VPADDL node.
6728
SelectionDAG &DAG = DCI.DAG;
6729
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6731
// Build operand list.
6732
SmallVector<SDValue, 8> Ops;
6733
Ops.push_back(DAG.getConstant(Intrinsic::arm_neon_vpaddls,
6734
TLI.getPointerTy()));
6736
// Input is the vector.
6739
// Get widened type and narrowed type.
6741
unsigned numElem = VT.getVectorNumElements();
6742
switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
6743
case MVT::i8: widenType = MVT::getVectorVT(MVT::i16, numElem); break;
6744
case MVT::i16: widenType = MVT::getVectorVT(MVT::i32, numElem); break;
6745
case MVT::i32: widenType = MVT::getVectorVT(MVT::i64, numElem); break;
6747
llvm_unreachable("Invalid vector element type for padd optimization.");
6750
SDValue tmp = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
6751
widenType, &Ops[0], Ops.size());
6752
return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, tmp);
6755
/// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
6756
/// operands N0 and N1. This is a helper for PerformADDCombine that is
6757
/// called with the default operands, and if that fails, with commuted
6759
static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
6760
TargetLowering::DAGCombinerInfo &DCI,
6761
const ARMSubtarget *Subtarget){
6763
// Attempt to create vpaddl for this add.
6764
SDValue Result = AddCombineToVPADDL(N, N0, N1, DCI, Subtarget);
6765
if (Result.getNode())
6768
// fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
6769
if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) {
6770
SDValue Result = combineSelectAndUse(N, N0, N1, DCI);
6771
if (Result.getNode()) return Result;
6776
/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
6778
static SDValue PerformADDCombine(SDNode *N,
6779
TargetLowering::DAGCombinerInfo &DCI,
6780
const ARMSubtarget *Subtarget) {
6781
SDValue N0 = N->getOperand(0);
6782
SDValue N1 = N->getOperand(1);
6784
// First try with the default operand order.
6785
SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget);
6786
if (Result.getNode())
6789
// If that didn't work, try again with the operands commuted.
6790
return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget);
6793
/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
6795
static SDValue PerformSUBCombine(SDNode *N,
6796
TargetLowering::DAGCombinerInfo &DCI) {
6797
SDValue N0 = N->getOperand(0);
6798
SDValue N1 = N->getOperand(1);
6800
// fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
6801
if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) {
6802
SDValue Result = combineSelectAndUse(N, N1, N0, DCI);
6803
if (Result.getNode()) return Result;
6809
/// PerformVMULCombine
6810
/// Distribute (A + B) * C to (A * C) + (B * C) to take advantage of the
6811
/// special multiplier accumulator forwarding.
6817
static SDValue PerformVMULCombine(SDNode *N,
6818
TargetLowering::DAGCombinerInfo &DCI,
6819
const ARMSubtarget *Subtarget) {
6820
if (!Subtarget->hasVMLxForwarding())
6823
SelectionDAG &DAG = DCI.DAG;
6824
SDValue N0 = N->getOperand(0);
6825
SDValue N1 = N->getOperand(1);
6826
unsigned Opcode = N0.getOpcode();
6827
if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6828
Opcode != ISD::FADD && Opcode != ISD::FSUB) {
6829
Opcode = N1.getOpcode();
6830
if (Opcode != ISD::ADD && Opcode != ISD::SUB &&
6831
Opcode != ISD::FADD && Opcode != ISD::FSUB)
6836
EVT VT = N->getValueType(0);
6837
DebugLoc DL = N->getDebugLoc();
6838
SDValue N00 = N0->getOperand(0);
6839
SDValue N01 = N0->getOperand(1);
6840
return DAG.getNode(Opcode, DL, VT,
6841
DAG.getNode(ISD::MUL, DL, VT, N00, N1),
6842
DAG.getNode(ISD::MUL, DL, VT, N01, N1));
6845
static SDValue PerformMULCombine(SDNode *N,
6846
TargetLowering::DAGCombinerInfo &DCI,
6847
const ARMSubtarget *Subtarget) {
6848
SelectionDAG &DAG = DCI.DAG;
6850
if (Subtarget->isThumb1Only())
6853
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
6856
EVT VT = N->getValueType(0);
6857
if (VT.is64BitVector() || VT.is128BitVector())
6858
return PerformVMULCombine(N, DCI, Subtarget);
6862
ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
6866
int64_t MulAmt = C->getSExtValue();
6867
unsigned ShiftAmt = CountTrailingZeros_64(MulAmt);
6869
ShiftAmt = ShiftAmt & (32 - 1);
6870
SDValue V = N->getOperand(0);
6871
DebugLoc DL = N->getDebugLoc();
6874
MulAmt >>= ShiftAmt;
6877
if (isPowerOf2_32(MulAmt - 1)) {
6878
// (mul x, 2^N + 1) => (add (shl x, N), x)
6879
Res = DAG.getNode(ISD::ADD, DL, VT,
6881
DAG.getNode(ISD::SHL, DL, VT,
6883
DAG.getConstant(Log2_32(MulAmt - 1),
6885
} else if (isPowerOf2_32(MulAmt + 1)) {
6886
// (mul x, 2^N - 1) => (sub (shl x, N), x)
6887
Res = DAG.getNode(ISD::SUB, DL, VT,
6888
DAG.getNode(ISD::SHL, DL, VT,
6890
DAG.getConstant(Log2_32(MulAmt + 1),
6896
uint64_t MulAmtAbs = -MulAmt;
6897
if (isPowerOf2_32(MulAmtAbs + 1)) {
6898
// (mul x, -(2^N - 1)) => (sub x, (shl x, N))
6899
Res = DAG.getNode(ISD::SUB, DL, VT,
6901
DAG.getNode(ISD::SHL, DL, VT,
6903
DAG.getConstant(Log2_32(MulAmtAbs + 1),
6905
} else if (isPowerOf2_32(MulAmtAbs - 1)) {
6906
// (mul x, -(2^N + 1)) => - (add (shl x, N), x)
6907
Res = DAG.getNode(ISD::ADD, DL, VT,
6909
DAG.getNode(ISD::SHL, DL, VT,
6911
DAG.getConstant(Log2_32(MulAmtAbs-1),
6913
Res = DAG.getNode(ISD::SUB, DL, VT,
6914
DAG.getConstant(0, MVT::i32),Res);
6921
Res = DAG.getNode(ISD::SHL, DL, VT,
6922
Res, DAG.getConstant(ShiftAmt, MVT::i32));
6924
// Do not add new nodes to DAG combiner worklist.
6925
DCI.CombineTo(N, Res, false);
6929
static bool isCMOVWithZeroOrAllOnesLHS(SDValue N, bool AllOnes) {
6930
if (N.getOpcode() != ARMISD::CMOV || !N.getNode()->hasOneUse())
6933
SDValue FalseVal = N.getOperand(0);
6934
ConstantSDNode *C = dyn_cast<ConstantSDNode>(FalseVal);
6938
return C->isAllOnesValue();
6939
return C->isNullValue();
6942
/// formConditionalOp - Combine an operation with a conditional move operand
6943
/// to form a conditional op. e.g. (or x, (cmov 0, y, cond)) => (or.cond x, y)
6944
/// (and x, (cmov -1, y, cond)) => (and.cond, x, y)
6945
static SDValue formConditionalOp(SDNode *N, SelectionDAG &DAG,
6947
SDValue N0 = N->getOperand(0);
6948
SDValue N1 = N->getOperand(1);
6950
bool isAND = N->getOpcode() == ISD::AND;
6951
bool isCand = isCMOVWithZeroOrAllOnesLHS(N1, isAND);
6952
if (!isCand && Commutable) {
6953
isCand = isCMOVWithZeroOrAllOnesLHS(N0, isAND);
6961
switch (N->getOpcode()) {
6962
default: llvm_unreachable("Unexpected node");
6963
case ISD::AND: Opc = ARMISD::CAND; break;
6964
case ISD::OR: Opc = ARMISD::COR; break;
6965
case ISD::XOR: Opc = ARMISD::CXOR; break;
6967
return DAG.getNode(Opc, N->getDebugLoc(), N->getValueType(0), N0,
6968
N1.getOperand(1), N1.getOperand(2), N1.getOperand(3),
6972
static SDValue PerformANDCombine(SDNode *N,
6973
TargetLowering::DAGCombinerInfo &DCI,
6974
const ARMSubtarget *Subtarget) {
6976
// Attempt to use immediate-form VBIC
6977
BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
6978
DebugLoc dl = N->getDebugLoc();
6979
EVT VT = N->getValueType(0);
6980
SelectionDAG &DAG = DCI.DAG;
6982
if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
6985
APInt SplatBits, SplatUndef;
6986
unsigned SplatBitSize;
6989
BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
6990
if (SplatBitSize <= 64) {
6992
SDValue Val = isNEONModifiedImm((~SplatBits).getZExtValue(),
6993
SplatUndef.getZExtValue(), SplatBitSize,
6994
DAG, VbicVT, VT.is128BitVector(),
6996
if (Val.getNode()) {
6998
DAG.getNode(ISD::BITCAST, dl, VbicVT, N->getOperand(0));
6999
SDValue Vbic = DAG.getNode(ARMISD::VBICIMM, dl, VbicVT, Input, Val);
7000
return DAG.getNode(ISD::BITCAST, dl, VT, Vbic);
7005
if (!Subtarget->isThumb1Only()) {
7006
// (and x, (cmov -1, y, cond)) => (and.cond x, y)
7007
SDValue CAND = formConditionalOp(N, DAG, true);
7015
/// PerformORCombine - Target-specific dag combine xforms for ISD::OR
7016
static SDValue PerformORCombine(SDNode *N,
7017
TargetLowering::DAGCombinerInfo &DCI,
7018
const ARMSubtarget *Subtarget) {
7019
// Attempt to use immediate-form VORR
7020
BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(1));
7021
DebugLoc dl = N->getDebugLoc();
7022
EVT VT = N->getValueType(0);
7023
SelectionDAG &DAG = DCI.DAG;
7025
if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7028
APInt SplatBits, SplatUndef;
7029
unsigned SplatBitSize;
7031
if (BVN && Subtarget->hasNEON() &&
7032
BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
7033
if (SplatBitSize <= 64) {
7035
SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
7036
SplatUndef.getZExtValue(), SplatBitSize,
7037
DAG, VorrVT, VT.is128BitVector(),
7039
if (Val.getNode()) {
7041
DAG.getNode(ISD::BITCAST, dl, VorrVT, N->getOperand(0));
7042
SDValue Vorr = DAG.getNode(ARMISD::VORRIMM, dl, VorrVT, Input, Val);
7043
return DAG.getNode(ISD::BITCAST, dl, VT, Vorr);
7048
if (!Subtarget->isThumb1Only()) {
7049
// (or x, (cmov 0, y, cond)) => (or.cond x, y)
7050
SDValue COR = formConditionalOp(N, DAG, true);
7055
SDValue N0 = N->getOperand(0);
7056
if (N0.getOpcode() != ISD::AND)
7058
SDValue N1 = N->getOperand(1);
7060
// (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
7061
if (Subtarget->hasNEON() && N1.getOpcode() == ISD::AND && VT.isVector() &&
7062
DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
7064
unsigned SplatBitSize;
7067
BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
7069
if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
7070
HasAnyUndefs) && !HasAnyUndefs) {
7071
BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
7073
if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
7074
HasAnyUndefs) && !HasAnyUndefs &&
7075
SplatBits0 == ~SplatBits1) {
7076
// Canonicalize the vector type to make instruction selection simpler.
7077
EVT CanonicalVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
7078
SDValue Result = DAG.getNode(ARMISD::VBSL, dl, CanonicalVT,
7079
N0->getOperand(1), N0->getOperand(0),
7081
return DAG.getNode(ISD::BITCAST, dl, VT, Result);
7086
// Try to use the ARM/Thumb2 BFI (bitfield insert) instruction when
7089
// BFI is only available on V6T2+
7090
if (Subtarget->isThumb1Only() || !Subtarget->hasV6T2Ops())
7093
DebugLoc DL = N->getDebugLoc();
7094
// 1) or (and A, mask), val => ARMbfi A, val, mask
7095
// iff (val & mask) == val
7097
// 2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7098
// 2a) iff isBitFieldInvertedMask(mask) && isBitFieldInvertedMask(~mask2)
7099
// && mask == ~mask2
7100
// 2b) iff isBitFieldInvertedMask(~mask) && isBitFieldInvertedMask(mask2)
7101
// && ~mask == mask2
7102
// (i.e., copy a bitfield value into another bitfield of the same width)
7107
SDValue N00 = N0.getOperand(0);
7109
// The value and the mask need to be constants so we can verify this is
7110
// actually a bitfield set. If the mask is 0xffff, we can do better
7111
// via a movt instruction, so don't use BFI in that case.
7112
SDValue MaskOp = N0.getOperand(1);
7113
ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(MaskOp);
7116
unsigned Mask = MaskC->getZExtValue();
7120
// Case (1): or (and A, mask), val => ARMbfi A, val, mask
7121
ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
7123
unsigned Val = N1C->getZExtValue();
7124
if ((Val & ~Mask) != Val)
7127
if (ARM::isBitFieldInvertedMask(Mask)) {
7128
Val >>= CountTrailingZeros_32(~Mask);
7130
Res = DAG.getNode(ARMISD::BFI, DL, VT, N00,
7131
DAG.getConstant(Val, MVT::i32),
7132
DAG.getConstant(Mask, MVT::i32));
7134
// Do not add new nodes to DAG combiner worklist.
7135
DCI.CombineTo(N, Res, false);
7138
} else if (N1.getOpcode() == ISD::AND) {
7139
// case (2) or (and A, mask), (and B, mask2) => ARMbfi A, (lsr B, amt), mask
7140
ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7143
unsigned Mask2 = N11C->getZExtValue();
7145
// Mask and ~Mask2 (or reverse) must be equivalent for the BFI pattern
7147
if (ARM::isBitFieldInvertedMask(Mask) &&
7149
// The pack halfword instruction works better for masks that fit it,
7150
// so use that when it's available.
7151
if (Subtarget->hasT2ExtractPack() &&
7152
(Mask == 0xffff || Mask == 0xffff0000))
7155
unsigned amt = CountTrailingZeros_32(Mask2);
7156
Res = DAG.getNode(ISD::SRL, DL, VT, N1.getOperand(0),
7157
DAG.getConstant(amt, MVT::i32));
7158
Res = DAG.getNode(ARMISD::BFI, DL, VT, N00, Res,
7159
DAG.getConstant(Mask, MVT::i32));
7160
// Do not add new nodes to DAG combiner worklist.
7161
DCI.CombineTo(N, Res, false);
7163
} else if (ARM::isBitFieldInvertedMask(~Mask) &&
7165
// The pack halfword instruction works better for masks that fit it,
7166
// so use that when it's available.
7167
if (Subtarget->hasT2ExtractPack() &&
7168
(Mask2 == 0xffff || Mask2 == 0xffff0000))
7171
unsigned lsb = CountTrailingZeros_32(Mask);
7172
Res = DAG.getNode(ISD::SRL, DL, VT, N00,
7173
DAG.getConstant(lsb, MVT::i32));
7174
Res = DAG.getNode(ARMISD::BFI, DL, VT, N1.getOperand(0), Res,
7175
DAG.getConstant(Mask2, MVT::i32));
7176
// Do not add new nodes to DAG combiner worklist.
7177
DCI.CombineTo(N, Res, false);
7182
if (DAG.MaskedValueIsZero(N1, MaskC->getAPIntValue()) &&
7183
N00.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N00.getOperand(1)) &&
7184
ARM::isBitFieldInvertedMask(~Mask)) {
7185
// Case (3): or (and (shl A, #shamt), mask), B => ARMbfi B, A, ~mask
7186
// where lsb(mask) == #shamt and masked bits of B are known zero.
7187
SDValue ShAmt = N00.getOperand(1);
7188
unsigned ShAmtC = cast<ConstantSDNode>(ShAmt)->getZExtValue();
7189
unsigned LSB = CountTrailingZeros_32(Mask);
7193
Res = DAG.getNode(ARMISD::BFI, DL, VT, N1, N00.getOperand(0),
7194
DAG.getConstant(~Mask, MVT::i32));
7196
// Do not add new nodes to DAG combiner worklist.
7197
DCI.CombineTo(N, Res, false);
7203
static SDValue PerformXORCombine(SDNode *N,
7204
TargetLowering::DAGCombinerInfo &DCI,
7205
const ARMSubtarget *Subtarget) {
7206
EVT VT = N->getValueType(0);
7207
SelectionDAG &DAG = DCI.DAG;
7209
if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
7212
if (!Subtarget->isThumb1Only()) {
7213
// (xor x, (cmov 0, y, cond)) => (xor.cond x, y)
7214
SDValue CXOR = formConditionalOp(N, DAG, true);
7222
/// PerformBFICombine - (bfi A, (and B, Mask1), Mask2) -> (bfi A, B, Mask2) iff
7223
/// the bits being cleared by the AND are not demanded by the BFI.
7224
static SDValue PerformBFICombine(SDNode *N,
7225
TargetLowering::DAGCombinerInfo &DCI) {
7226
SDValue N1 = N->getOperand(1);
7227
if (N1.getOpcode() == ISD::AND) {
7228
ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
7231
unsigned InvMask = cast<ConstantSDNode>(N->getOperand(2))->getZExtValue();
7232
unsigned LSB = CountTrailingZeros_32(~InvMask);
7233
unsigned Width = (32 - CountLeadingZeros_32(~InvMask)) - LSB;
7234
unsigned Mask = (1 << Width)-1;
7235
unsigned Mask2 = N11C->getZExtValue();
7236
if ((Mask & (~Mask2)) == 0)
7237
return DCI.DAG.getNode(ARMISD::BFI, N->getDebugLoc(), N->getValueType(0),
7238
N->getOperand(0), N1.getOperand(0),
7244
/// PerformVMOVRRDCombine - Target-specific dag combine xforms for
7245
/// ARMISD::VMOVRRD.
7246
static SDValue PerformVMOVRRDCombine(SDNode *N,
7247
TargetLowering::DAGCombinerInfo &DCI) {
7248
// vmovrrd(vmovdrr x, y) -> x,y
7249
SDValue InDouble = N->getOperand(0);
7250
if (InDouble.getOpcode() == ARMISD::VMOVDRR)
7251
return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
7253
// vmovrrd(load f64) -> (load i32), (load i32)
7254
SDNode *InNode = InDouble.getNode();
7255
if (ISD::isNormalLoad(InNode) && InNode->hasOneUse() &&
7256
InNode->getValueType(0) == MVT::f64 &&
7257
InNode->getOperand(1).getOpcode() == ISD::FrameIndex &&
7258
!cast<LoadSDNode>(InNode)->isVolatile()) {
7259
// TODO: Should this be done for non-FrameIndex operands?
7260
LoadSDNode *LD = cast<LoadSDNode>(InNode);
7262
SelectionDAG &DAG = DCI.DAG;
7263
DebugLoc DL = LD->getDebugLoc();
7264
SDValue BasePtr = LD->getBasePtr();
7265
SDValue NewLD1 = DAG.getLoad(MVT::i32, DL, LD->getChain(), BasePtr,
7266
LD->getPointerInfo(), LD->isVolatile(),
7267
LD->isNonTemporal(), LD->isInvariant(),
7268
LD->getAlignment());
7270
SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7271
DAG.getConstant(4, MVT::i32));
7272
SDValue NewLD2 = DAG.getLoad(MVT::i32, DL, NewLD1.getValue(1), OffsetPtr,
7273
LD->getPointerInfo(), LD->isVolatile(),
7274
LD->isNonTemporal(), LD->isInvariant(),
7275
std::min(4U, LD->getAlignment() / 2));
7277
DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLD2.getValue(1));
7278
SDValue Result = DCI.CombineTo(N, NewLD1, NewLD2);
7279
DCI.RemoveFromWorklist(LD);
7287
/// PerformVMOVDRRCombine - Target-specific dag combine xforms for
7288
/// ARMISD::VMOVDRR. This is also used for BUILD_VECTORs with 2 operands.
7289
static SDValue PerformVMOVDRRCombine(SDNode *N, SelectionDAG &DAG) {
7290
// N=vmovrrd(X); vmovdrr(N:0, N:1) -> bit_convert(X)
7291
SDValue Op0 = N->getOperand(0);
7292
SDValue Op1 = N->getOperand(1);
7293
if (Op0.getOpcode() == ISD::BITCAST)
7294
Op0 = Op0.getOperand(0);
7295
if (Op1.getOpcode() == ISD::BITCAST)
7296
Op1 = Op1.getOperand(0);
7297
if (Op0.getOpcode() == ARMISD::VMOVRRD &&
7298
Op0.getNode() == Op1.getNode() &&
7299
Op0.getResNo() == 0 && Op1.getResNo() == 1)
7300
return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
7301
N->getValueType(0), Op0.getOperand(0));
7305
/// PerformSTORECombine - Target-specific dag combine xforms for
7307
static SDValue PerformSTORECombine(SDNode *N,
7308
TargetLowering::DAGCombinerInfo &DCI) {
7309
// Bitcast an i64 store extracted from a vector to f64.
7310
// Otherwise, the i64 value will be legalized to a pair of i32 values.
7311
StoreSDNode *St = cast<StoreSDNode>(N);
7312
SDValue StVal = St->getValue();
7313
if (!ISD::isNormalStore(St) || St->isVolatile())
7316
if (StVal.getNode()->getOpcode() == ARMISD::VMOVDRR &&
7317
StVal.getNode()->hasOneUse() && !St->isVolatile()) {
7318
SelectionDAG &DAG = DCI.DAG;
7319
DebugLoc DL = St->getDebugLoc();
7320
SDValue BasePtr = St->getBasePtr();
7321
SDValue NewST1 = DAG.getStore(St->getChain(), DL,
7322
StVal.getNode()->getOperand(0), BasePtr,
7323
St->getPointerInfo(), St->isVolatile(),
7324
St->isNonTemporal(), St->getAlignment());
7326
SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i32, BasePtr,
7327
DAG.getConstant(4, MVT::i32));
7328
return DAG.getStore(NewST1.getValue(0), DL, StVal.getNode()->getOperand(1),
7329
OffsetPtr, St->getPointerInfo(), St->isVolatile(),
7330
St->isNonTemporal(),
7331
std::min(4U, St->getAlignment() / 2));
7334
if (StVal.getValueType() != MVT::i64 ||
7335
StVal.getNode()->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
7338
SelectionDAG &DAG = DCI.DAG;
7339
DebugLoc dl = StVal.getDebugLoc();
7340
SDValue IntVec = StVal.getOperand(0);
7341
EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7342
IntVec.getValueType().getVectorNumElements());
7343
SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, IntVec);
7344
SDValue ExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7345
Vec, StVal.getOperand(1));
7346
dl = N->getDebugLoc();
7347
SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ExtElt);
7348
// Make the DAGCombiner fold the bitcasts.
7349
DCI.AddToWorklist(Vec.getNode());
7350
DCI.AddToWorklist(ExtElt.getNode());
7351
DCI.AddToWorklist(V.getNode());
7352
return DAG.getStore(St->getChain(), dl, V, St->getBasePtr(),
7353
St->getPointerInfo(), St->isVolatile(),
7354
St->isNonTemporal(), St->getAlignment(),
7358
/// hasNormalLoadOperand - Check if any of the operands of a BUILD_VECTOR node
7359
/// are normal, non-volatile loads. If so, it is profitable to bitcast an
7360
/// i64 vector to have f64 elements, since the value can then be loaded
7361
/// directly into a VFP register.
7362
static bool hasNormalLoadOperand(SDNode *N) {
7363
unsigned NumElts = N->getValueType(0).getVectorNumElements();
7364
for (unsigned i = 0; i < NumElts; ++i) {
7365
SDNode *Elt = N->getOperand(i).getNode();
7366
if (ISD::isNormalLoad(Elt) && !cast<LoadSDNode>(Elt)->isVolatile())
7372
/// PerformBUILD_VECTORCombine - Target-specific dag combine xforms for
7373
/// ISD::BUILD_VECTOR.
7374
static SDValue PerformBUILD_VECTORCombine(SDNode *N,
7375
TargetLowering::DAGCombinerInfo &DCI){
7376
// build_vector(N=ARMISD::VMOVRRD(X), N:1) -> bit_convert(X):
7377
// VMOVRRD is introduced when legalizing i64 types. It forces the i64 value
7378
// into a pair of GPRs, which is fine when the value is used as a scalar,
7379
// but if the i64 value is converted to a vector, we need to undo the VMOVRRD.
7380
SelectionDAG &DAG = DCI.DAG;
7381
if (N->getNumOperands() == 2) {
7382
SDValue RV = PerformVMOVDRRCombine(N, DAG);
7387
// Load i64 elements as f64 values so that type legalization does not split
7388
// them up into i32 values.
7389
EVT VT = N->getValueType(0);
7390
if (VT.getVectorElementType() != MVT::i64 || !hasNormalLoadOperand(N))
7392
DebugLoc dl = N->getDebugLoc();
7393
SmallVector<SDValue, 8> Ops;
7394
unsigned NumElts = VT.getVectorNumElements();
7395
for (unsigned i = 0; i < NumElts; ++i) {
7396
SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(i));
7398
// Make the DAGCombiner fold the bitcast.
7399
DCI.AddToWorklist(V.getNode());
7401
EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, NumElts);
7402
SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, FloatVT, Ops.data(), NumElts);
7403
return DAG.getNode(ISD::BITCAST, dl, VT, BV);
7406
/// PerformInsertEltCombine - Target-specific dag combine xforms for
7407
/// ISD::INSERT_VECTOR_ELT.
7408
static SDValue PerformInsertEltCombine(SDNode *N,
7409
TargetLowering::DAGCombinerInfo &DCI) {
7410
// Bitcast an i64 load inserted into a vector to f64.
7411
// Otherwise, the i64 value will be legalized to a pair of i32 values.
7412
EVT VT = N->getValueType(0);
7413
SDNode *Elt = N->getOperand(1).getNode();
7414
if (VT.getVectorElementType() != MVT::i64 ||
7415
!ISD::isNormalLoad(Elt) || cast<LoadSDNode>(Elt)->isVolatile())
7418
SelectionDAG &DAG = DCI.DAG;
7419
DebugLoc dl = N->getDebugLoc();
7420
EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64,
7421
VT.getVectorNumElements());
7422
SDValue Vec = DAG.getNode(ISD::BITCAST, dl, FloatVT, N->getOperand(0));
7423
SDValue V = DAG.getNode(ISD::BITCAST, dl, MVT::f64, N->getOperand(1));
7424
// Make the DAGCombiner fold the bitcasts.
7425
DCI.AddToWorklist(Vec.getNode());
7426
DCI.AddToWorklist(V.getNode());
7427
SDValue InsElt = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, FloatVT,
7428
Vec, V, N->getOperand(2));
7429
return DAG.getNode(ISD::BITCAST, dl, VT, InsElt);
7432
/// PerformVECTOR_SHUFFLECombine - Target-specific dag combine xforms for
7433
/// ISD::VECTOR_SHUFFLE.
7434
static SDValue PerformVECTOR_SHUFFLECombine(SDNode *N, SelectionDAG &DAG) {
7435
// The LLVM shufflevector instruction does not require the shuffle mask
7436
// length to match the operand vector length, but ISD::VECTOR_SHUFFLE does
7437
// have that requirement. When translating to ISD::VECTOR_SHUFFLE, if the
7438
// operands do not match the mask length, they are extended by concatenating
7439
// them with undef vectors. That is probably the right thing for other
7440
// targets, but for NEON it is better to concatenate two double-register
7441
// size vector operands into a single quad-register size vector. Do that
7442
// transformation here:
7443
// shuffle(concat(v1, undef), concat(v2, undef)) ->
7444
// shuffle(concat(v1, v2), undef)
7445
SDValue Op0 = N->getOperand(0);
7446
SDValue Op1 = N->getOperand(1);
7447
if (Op0.getOpcode() != ISD::CONCAT_VECTORS ||
7448
Op1.getOpcode() != ISD::CONCAT_VECTORS ||
7449
Op0.getNumOperands() != 2 ||
7450
Op1.getNumOperands() != 2)
7452
SDValue Concat0Op1 = Op0.getOperand(1);
7453
SDValue Concat1Op1 = Op1.getOperand(1);
7454
if (Concat0Op1.getOpcode() != ISD::UNDEF ||
7455
Concat1Op1.getOpcode() != ISD::UNDEF)
7457
// Skip the transformation if any of the types are illegal.
7458
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
7459
EVT VT = N->getValueType(0);
7460
if (!TLI.isTypeLegal(VT) ||
7461
!TLI.isTypeLegal(Concat0Op1.getValueType()) ||
7462
!TLI.isTypeLegal(Concat1Op1.getValueType()))
7465
SDValue NewConcat = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7466
Op0.getOperand(0), Op1.getOperand(0));
7467
// Translate the shuffle mask.
7468
SmallVector<int, 16> NewMask;
7469
unsigned NumElts = VT.getVectorNumElements();
7470
unsigned HalfElts = NumElts/2;
7471
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7472
for (unsigned n = 0; n < NumElts; ++n) {
7473
int MaskElt = SVN->getMaskElt(n);
7475
if (MaskElt < (int)HalfElts)
7477
else if (MaskElt >= (int)NumElts && MaskElt < (int)(NumElts + HalfElts))
7478
NewElt = HalfElts + MaskElt - NumElts;
7479
NewMask.push_back(NewElt);
7481
return DAG.getVectorShuffle(VT, N->getDebugLoc(), NewConcat,
7482
DAG.getUNDEF(VT), NewMask.data());
7485
/// CombineBaseUpdate - Target-specific DAG combine function for VLDDUP and
7486
/// NEON load/store intrinsics to merge base address updates.
7487
static SDValue CombineBaseUpdate(SDNode *N,
7488
TargetLowering::DAGCombinerInfo &DCI) {
7489
if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
7492
SelectionDAG &DAG = DCI.DAG;
7493
bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
7494
N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
7495
unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
7496
SDValue Addr = N->getOperand(AddrOpIdx);
7498
// Search for a use of the address operand that is an increment.
7499
for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
7500
UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
7502
if (User->getOpcode() != ISD::ADD ||
7503
UI.getUse().getResNo() != Addr.getResNo())
7506
// Check that the add is independent of the load/store. Otherwise, folding
7507
// it would create a cycle.
7508
if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
7511
// Find the new opcode for the updating load/store.
7513
bool isLaneOp = false;
7514
unsigned NewOpc = 0;
7515
unsigned NumVecs = 0;
7517
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
7519
default: llvm_unreachable("unexpected intrinsic for Neon base update");
7520
case Intrinsic::arm_neon_vld1: NewOpc = ARMISD::VLD1_UPD;
7522
case Intrinsic::arm_neon_vld2: NewOpc = ARMISD::VLD2_UPD;
7524
case Intrinsic::arm_neon_vld3: NewOpc = ARMISD::VLD3_UPD;
7526
case Intrinsic::arm_neon_vld4: NewOpc = ARMISD::VLD4_UPD;
7528
case Intrinsic::arm_neon_vld2lane: NewOpc = ARMISD::VLD2LN_UPD;
7529
NumVecs = 2; isLaneOp = true; break;
7530
case Intrinsic::arm_neon_vld3lane: NewOpc = ARMISD::VLD3LN_UPD;
7531
NumVecs = 3; isLaneOp = true; break;
7532
case Intrinsic::arm_neon_vld4lane: NewOpc = ARMISD::VLD4LN_UPD;
7533
NumVecs = 4; isLaneOp = true; break;
7534
case Intrinsic::arm_neon_vst1: NewOpc = ARMISD::VST1_UPD;
7535
NumVecs = 1; isLoad = false; break;
7536
case Intrinsic::arm_neon_vst2: NewOpc = ARMISD::VST2_UPD;
7537
NumVecs = 2; isLoad = false; break;
7538
case Intrinsic::arm_neon_vst3: NewOpc = ARMISD::VST3_UPD;
7539
NumVecs = 3; isLoad = false; break;
7540
case Intrinsic::arm_neon_vst4: NewOpc = ARMISD::VST4_UPD;
7541
NumVecs = 4; isLoad = false; break;
7542
case Intrinsic::arm_neon_vst2lane: NewOpc = ARMISD::VST2LN_UPD;
7543
NumVecs = 2; isLoad = false; isLaneOp = true; break;
7544
case Intrinsic::arm_neon_vst3lane: NewOpc = ARMISD::VST3LN_UPD;
7545
NumVecs = 3; isLoad = false; isLaneOp = true; break;
7546
case Intrinsic::arm_neon_vst4lane: NewOpc = ARMISD::VST4LN_UPD;
7547
NumVecs = 4; isLoad = false; isLaneOp = true; break;
7551
switch (N->getOpcode()) {
7552
default: llvm_unreachable("unexpected opcode for Neon base update");
7553
case ARMISD::VLD2DUP: NewOpc = ARMISD::VLD2DUP_UPD; NumVecs = 2; break;
7554
case ARMISD::VLD3DUP: NewOpc = ARMISD::VLD3DUP_UPD; NumVecs = 3; break;
7555
case ARMISD::VLD4DUP: NewOpc = ARMISD::VLD4DUP_UPD; NumVecs = 4; break;
7559
// Find the size of memory referenced by the load/store.
7562
VecTy = N->getValueType(0);
7564
VecTy = N->getOperand(AddrOpIdx+1).getValueType();
7565
unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
7567
NumBytes /= VecTy.getVectorNumElements();
7569
// If the increment is a constant, it must match the memory ref size.
7570
SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
7571
if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
7572
uint64_t IncVal = CInc->getZExtValue();
7573
if (IncVal != NumBytes)
7575
} else if (NumBytes >= 3 * 16) {
7576
// VLD3/4 and VST3/4 for 128-bit vectors are implemented with two
7577
// separate instructions that make it harder to use a non-constant update.
7581
// Create the new updating load/store node.
7583
unsigned NumResultVecs = (isLoad ? NumVecs : 0);
7585
for (n = 0; n < NumResultVecs; ++n)
7587
Tys[n++] = MVT::i32;
7588
Tys[n] = MVT::Other;
7589
SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs+2);
7590
SmallVector<SDValue, 8> Ops;
7591
Ops.push_back(N->getOperand(0)); // incoming chain
7592
Ops.push_back(N->getOperand(AddrOpIdx));
7594
for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
7595
Ops.push_back(N->getOperand(i));
7597
MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
7598
SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, N->getDebugLoc(), SDTys,
7599
Ops.data(), Ops.size(),
7600
MemInt->getMemoryVT(),
7601
MemInt->getMemOperand());
7604
std::vector<SDValue> NewResults;
7605
for (unsigned i = 0; i < NumResultVecs; ++i) {
7606
NewResults.push_back(SDValue(UpdN.getNode(), i));
7608
NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs+1)); // chain
7609
DCI.CombineTo(N, NewResults);
7610
DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
7617
/// CombineVLDDUP - For a VDUPLANE node N, check if its source operand is a
7618
/// vldN-lane (N > 1) intrinsic, and if all the other uses of that intrinsic
7619
/// are also VDUPLANEs. If so, combine them to a vldN-dup operation and
7621
static bool CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
7622
SelectionDAG &DAG = DCI.DAG;
7623
EVT VT = N->getValueType(0);
7624
// vldN-dup instructions only support 64-bit vectors for N > 1.
7625
if (!VT.is64BitVector())
7628
// Check if the VDUPLANE operand is a vldN-dup intrinsic.
7629
SDNode *VLD = N->getOperand(0).getNode();
7630
if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
7632
unsigned NumVecs = 0;
7633
unsigned NewOpc = 0;
7634
unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
7635
if (IntNo == Intrinsic::arm_neon_vld2lane) {
7637
NewOpc = ARMISD::VLD2DUP;
7638
} else if (IntNo == Intrinsic::arm_neon_vld3lane) {
7640
NewOpc = ARMISD::VLD3DUP;
7641
} else if (IntNo == Intrinsic::arm_neon_vld4lane) {
7643
NewOpc = ARMISD::VLD4DUP;
7648
// First check that all the vldN-lane uses are VDUPLANEs and that the lane
7649
// numbers match the load.
7650
unsigned VLDLaneNo =
7651
cast<ConstantSDNode>(VLD->getOperand(NumVecs+3))->getZExtValue();
7652
for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7654
// Ignore uses of the chain result.
7655
if (UI.getUse().getResNo() == NumVecs)
7658
if (User->getOpcode() != ARMISD::VDUPLANE ||
7659
VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
7663
// Create the vldN-dup node.
7666
for (n = 0; n < NumVecs; ++n)
7668
Tys[n] = MVT::Other;
7669
SDVTList SDTys = DAG.getVTList(Tys, NumVecs+1);
7670
SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
7671
MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
7672
SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, VLD->getDebugLoc(), SDTys,
7673
Ops, 2, VLDMemInt->getMemoryVT(),
7674
VLDMemInt->getMemOperand());
7677
for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
7679
unsigned ResNo = UI.getUse().getResNo();
7680
// Ignore uses of the chain result.
7681
if (ResNo == NumVecs)
7684
DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
7687
// Now the vldN-lane intrinsic is dead except for its chain result.
7688
// Update uses of the chain.
7689
std::vector<SDValue> VLDDupResults;
7690
for (unsigned n = 0; n < NumVecs; ++n)
7691
VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
7692
VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
7693
DCI.CombineTo(VLD, VLDDupResults);
7698
/// PerformVDUPLANECombine - Target-specific dag combine xforms for
7699
/// ARMISD::VDUPLANE.
7700
static SDValue PerformVDUPLANECombine(SDNode *N,
7701
TargetLowering::DAGCombinerInfo &DCI) {
7702
SDValue Op = N->getOperand(0);
7704
// If the source is a vldN-lane (N > 1) intrinsic, and all the other uses
7705
// of that intrinsic are also VDUPLANEs, combine them to a vldN-dup operation.
7706
if (CombineVLDDUP(N, DCI))
7707
return SDValue(N, 0);
7709
// If the source is already a VMOVIMM or VMVNIMM splat, the VDUPLANE is
7710
// redundant. Ignore bit_converts for now; element sizes are checked below.
7711
while (Op.getOpcode() == ISD::BITCAST)
7712
Op = Op.getOperand(0);
7713
if (Op.getOpcode() != ARMISD::VMOVIMM && Op.getOpcode() != ARMISD::VMVNIMM)
7716
// Make sure the VMOV element size is not bigger than the VDUPLANE elements.
7717
unsigned EltSize = Op.getValueType().getVectorElementType().getSizeInBits();
7718
// The canonical VMOV for a zero vector uses a 32-bit element size.
7719
unsigned Imm = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7721
if (ARM_AM::decodeNEONModImm(Imm, EltBits) == 0)
7723
EVT VT = N->getValueType(0);
7724
if (EltSize > VT.getVectorElementType().getSizeInBits())
7727
return DCI.DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
7730
// isConstVecPow2 - Return true if each vector element is a power of 2, all
7731
// elements are the same constant, C, and Log2(C) ranges from 1 to 32.
7732
static bool isConstVecPow2(SDValue ConstVec, bool isSigned, uint64_t &C)
7736
for (unsigned I = 0, E = ConstVec.getValueType().getVectorNumElements();
7738
ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(ConstVec.getOperand(I));
7743
APFloat APF = C->getValueAPF();
7744
if (APF.convertToInteger(&cN, 64, isSigned, APFloat::rmTowardZero, &isExact)
7745
!= APFloat::opOK || !isExact)
7748
c0 = (I == 0) ? cN : c0;
7749
if (!isPowerOf2_64(cN) || c0 != cN || Log2_64(c0) < 1 || Log2_64(c0) > 32)
7756
/// PerformVCVTCombine - VCVT (floating-point to fixed-point, Advanced SIMD)
7757
/// can replace combinations of VMUL and VCVT (floating-point to integer)
7758
/// when the VMUL has a constant operand that is a power of 2.
7760
/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7761
/// vmul.f32 d16, d17, d16
7762
/// vcvt.s32.f32 d16, d16
7764
/// vcvt.s32.f32 d16, d16, #3
7765
static SDValue PerformVCVTCombine(SDNode *N,
7766
TargetLowering::DAGCombinerInfo &DCI,
7767
const ARMSubtarget *Subtarget) {
7768
SelectionDAG &DAG = DCI.DAG;
7769
SDValue Op = N->getOperand(0);
7771
if (!Subtarget->hasNEON() || !Op.getValueType().isVector() ||
7772
Op.getOpcode() != ISD::FMUL)
7776
SDValue N0 = Op->getOperand(0);
7777
SDValue ConstVec = Op->getOperand(1);
7778
bool isSigned = N->getOpcode() == ISD::FP_TO_SINT;
7780
if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7781
!isConstVecPow2(ConstVec, isSigned, C))
7784
unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfp2fxs :
7785
Intrinsic::arm_neon_vcvtfp2fxu;
7786
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7788
DAG.getConstant(IntrinsicOpcode, MVT::i32), N0,
7789
DAG.getConstant(Log2_64(C), MVT::i32));
7792
/// PerformVDIVCombine - VCVT (fixed-point to floating-point, Advanced SIMD)
7793
/// can replace combinations of VCVT (integer to floating-point) and VDIV
7794
/// when the VDIV has a constant operand that is a power of 2.
7796
/// Example (assume d17 = <float 8.000000e+00, float 8.000000e+00>):
7797
/// vcvt.f32.s32 d16, d16
7798
/// vdiv.f32 d16, d17, d16
7800
/// vcvt.f32.s32 d16, d16, #3
7801
static SDValue PerformVDIVCombine(SDNode *N,
7802
TargetLowering::DAGCombinerInfo &DCI,
7803
const ARMSubtarget *Subtarget) {
7804
SelectionDAG &DAG = DCI.DAG;
7805
SDValue Op = N->getOperand(0);
7806
unsigned OpOpcode = Op.getNode()->getOpcode();
7808
if (!Subtarget->hasNEON() || !N->getValueType(0).isVector() ||
7809
(OpOpcode != ISD::SINT_TO_FP && OpOpcode != ISD::UINT_TO_FP))
7813
SDValue ConstVec = N->getOperand(1);
7814
bool isSigned = OpOpcode == ISD::SINT_TO_FP;
7816
if (ConstVec.getOpcode() != ISD::BUILD_VECTOR ||
7817
!isConstVecPow2(ConstVec, isSigned, C))
7820
unsigned IntrinsicOpcode = isSigned ? Intrinsic::arm_neon_vcvtfxs2fp :
7821
Intrinsic::arm_neon_vcvtfxu2fp;
7822
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, N->getDebugLoc(),
7824
DAG.getConstant(IntrinsicOpcode, MVT::i32),
7825
Op.getOperand(0), DAG.getConstant(Log2_64(C), MVT::i32));
7828
/// Getvshiftimm - Check if this is a valid build_vector for the immediate
7829
/// operand of a vector shift operation, where all the elements of the
7830
/// build_vector must have the same constant integer value.
7831
static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
7832
// Ignore bit_converts.
7833
while (Op.getOpcode() == ISD::BITCAST)
7834
Op = Op.getOperand(0);
7835
BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
7836
APInt SplatBits, SplatUndef;
7837
unsigned SplatBitSize;
7839
if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
7840
HasAnyUndefs, ElementBits) ||
7841
SplatBitSize > ElementBits)
7843
Cnt = SplatBits.getSExtValue();
7847
/// isVShiftLImm - Check if this is a valid build_vector for the immediate
7848
/// operand of a vector shift left operation. That value must be in the range:
7849
/// 0 <= Value < ElementBits for a left shift; or
7850
/// 0 <= Value <= ElementBits for a long left shift.
7851
static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
7852
assert(VT.isVector() && "vector shift count is not a vector type");
7853
unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7854
if (! getVShiftImm(Op, ElementBits, Cnt))
7856
return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits);
7859
/// isVShiftRImm - Check if this is a valid build_vector for the immediate
7860
/// operand of a vector shift right operation. For a shift opcode, the value
7861
/// is positive, but for an intrinsic the value count must be negative. The
7862
/// absolute value must be in the range:
7863
/// 1 <= |Value| <= ElementBits for a right shift; or
7864
/// 1 <= |Value| <= ElementBits/2 for a narrow right shift.
7865
static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic,
7867
assert(VT.isVector() && "vector shift count is not a vector type");
7868
unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
7869
if (! getVShiftImm(Op, ElementBits, Cnt))
7873
return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits));
7876
/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics.
7877
static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
7878
unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
7881
// Don't do anything for most intrinsics.
7884
// Vector shifts: check for immediate versions and lower them.
7885
// Note: This is done during DAG combining instead of DAG legalizing because
7886
// the build_vectors for 64-bit vector element shift counts are generally
7887
// not legal, and it is hard to see their values after they get legalized to
7888
// loads from a constant pool.
7889
case Intrinsic::arm_neon_vshifts:
7890
case Intrinsic::arm_neon_vshiftu:
7891
case Intrinsic::arm_neon_vshiftls:
7892
case Intrinsic::arm_neon_vshiftlu:
7893
case Intrinsic::arm_neon_vshiftn:
7894
case Intrinsic::arm_neon_vrshifts:
7895
case Intrinsic::arm_neon_vrshiftu:
7896
case Intrinsic::arm_neon_vrshiftn:
7897
case Intrinsic::arm_neon_vqshifts:
7898
case Intrinsic::arm_neon_vqshiftu:
7899
case Intrinsic::arm_neon_vqshiftsu:
7900
case Intrinsic::arm_neon_vqshiftns:
7901
case Intrinsic::arm_neon_vqshiftnu:
7902
case Intrinsic::arm_neon_vqshiftnsu:
7903
case Intrinsic::arm_neon_vqrshiftns:
7904
case Intrinsic::arm_neon_vqrshiftnu:
7905
case Intrinsic::arm_neon_vqrshiftnsu: {
7906
EVT VT = N->getOperand(1).getValueType();
7908
unsigned VShiftOpc = 0;
7911
case Intrinsic::arm_neon_vshifts:
7912
case Intrinsic::arm_neon_vshiftu:
7913
if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) {
7914
VShiftOpc = ARMISD::VSHL;
7917
if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) {
7918
VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ?
7919
ARMISD::VSHRs : ARMISD::VSHRu);
7924
case Intrinsic::arm_neon_vshiftls:
7925
case Intrinsic::arm_neon_vshiftlu:
7926
if (isVShiftLImm(N->getOperand(2), VT, true, Cnt))
7928
llvm_unreachable("invalid shift count for vshll intrinsic");
7930
case Intrinsic::arm_neon_vrshifts:
7931
case Intrinsic::arm_neon_vrshiftu:
7932
if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt))
7936
case Intrinsic::arm_neon_vqshifts:
7937
case Intrinsic::arm_neon_vqshiftu:
7938
if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7942
case Intrinsic::arm_neon_vqshiftsu:
7943
if (isVShiftLImm(N->getOperand(2), VT, false, Cnt))
7945
llvm_unreachable("invalid shift count for vqshlu intrinsic");
7947
case Intrinsic::arm_neon_vshiftn:
7948
case Intrinsic::arm_neon_vrshiftn:
7949
case Intrinsic::arm_neon_vqshiftns:
7950
case Intrinsic::arm_neon_vqshiftnu:
7951
case Intrinsic::arm_neon_vqshiftnsu:
7952
case Intrinsic::arm_neon_vqrshiftns:
7953
case Intrinsic::arm_neon_vqrshiftnu:
7954
case Intrinsic::arm_neon_vqrshiftnsu:
7955
// Narrowing shifts require an immediate right shift.
7956
if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt))
7958
llvm_unreachable("invalid shift count for narrowing vector shift "
7962
llvm_unreachable("unhandled vector shift");
7966
case Intrinsic::arm_neon_vshifts:
7967
case Intrinsic::arm_neon_vshiftu:
7968
// Opcode already set above.
7970
case Intrinsic::arm_neon_vshiftls:
7971
case Intrinsic::arm_neon_vshiftlu:
7972
if (Cnt == VT.getVectorElementType().getSizeInBits())
7973
VShiftOpc = ARMISD::VSHLLi;
7975
VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ?
7976
ARMISD::VSHLLs : ARMISD::VSHLLu);
7978
case Intrinsic::arm_neon_vshiftn:
7979
VShiftOpc = ARMISD::VSHRN; break;
7980
case Intrinsic::arm_neon_vrshifts:
7981
VShiftOpc = ARMISD::VRSHRs; break;
7982
case Intrinsic::arm_neon_vrshiftu:
7983
VShiftOpc = ARMISD::VRSHRu; break;
7984
case Intrinsic::arm_neon_vrshiftn:
7985
VShiftOpc = ARMISD::VRSHRN; break;
7986
case Intrinsic::arm_neon_vqshifts:
7987
VShiftOpc = ARMISD::VQSHLs; break;
7988
case Intrinsic::arm_neon_vqshiftu:
7989
VShiftOpc = ARMISD::VQSHLu; break;
7990
case Intrinsic::arm_neon_vqshiftsu:
7991
VShiftOpc = ARMISD::VQSHLsu; break;
7992
case Intrinsic::arm_neon_vqshiftns:
7993
VShiftOpc = ARMISD::VQSHRNs; break;
7994
case Intrinsic::arm_neon_vqshiftnu:
7995
VShiftOpc = ARMISD::VQSHRNu; break;
7996
case Intrinsic::arm_neon_vqshiftnsu:
7997
VShiftOpc = ARMISD::VQSHRNsu; break;
7998
case Intrinsic::arm_neon_vqrshiftns:
7999
VShiftOpc = ARMISD::VQRSHRNs; break;
8000
case Intrinsic::arm_neon_vqrshiftnu:
8001
VShiftOpc = ARMISD::VQRSHRNu; break;
8002
case Intrinsic::arm_neon_vqrshiftnsu:
8003
VShiftOpc = ARMISD::VQRSHRNsu; break;
8006
return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8007
N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
8010
case Intrinsic::arm_neon_vshiftins: {
8011
EVT VT = N->getOperand(1).getValueType();
8013
unsigned VShiftOpc = 0;
8015
if (isVShiftLImm(N->getOperand(3), VT, false, Cnt))
8016
VShiftOpc = ARMISD::VSLI;
8017
else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt))
8018
VShiftOpc = ARMISD::VSRI;
8020
llvm_unreachable("invalid shift count for vsli/vsri intrinsic");
8023
return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0),
8024
N->getOperand(1), N->getOperand(2),
8025
DAG.getConstant(Cnt, MVT::i32));
8028
case Intrinsic::arm_neon_vqrshifts:
8029
case Intrinsic::arm_neon_vqrshiftu:
8030
// No immediate versions of these to check for.
8037
/// PerformShiftCombine - Checks for immediate versions of vector shifts and
8038
/// lowers them. As with the vector shift intrinsics, this is done during DAG
8039
/// combining instead of DAG legalizing because the build_vectors for 64-bit
8040
/// vector element shift counts are generally not legal, and it is hard to see
8041
/// their values after they get legalized to loads from a constant pool.
8042
static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG,
8043
const ARMSubtarget *ST) {
8044
EVT VT = N->getValueType(0);
8045
if (N->getOpcode() == ISD::SRL && VT == MVT::i32 && ST->hasV6Ops()) {
8046
// Canonicalize (srl (bswap x), 16) to (rotr (bswap x), 16) if the high
8047
// 16-bits of x is zero. This optimizes rev + lsr 16 to rev16.
8048
SDValue N1 = N->getOperand(1);
8049
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
8050
SDValue N0 = N->getOperand(0);
8051
if (C->getZExtValue() == 16 && N0.getOpcode() == ISD::BSWAP &&
8052
DAG.MaskedValueIsZero(N0.getOperand(0),
8053
APInt::getHighBitsSet(32, 16)))
8054
return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, N0, N1);
8058
// Nothing to be done for scalar shifts.
8059
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8060
if (!VT.isVector() || !TLI.isTypeLegal(VT))
8063
assert(ST->hasNEON() && "unexpected vector shift");
8066
switch (N->getOpcode()) {
8067
default: llvm_unreachable("unexpected shift opcode");
8070
if (isVShiftLImm(N->getOperand(1), VT, false, Cnt))
8071
return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0),
8072
DAG.getConstant(Cnt, MVT::i32));
8077
if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) {
8078
unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ?
8079
ARMISD::VSHRs : ARMISD::VSHRu);
8080
return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0),
8081
DAG.getConstant(Cnt, MVT::i32));
8087
/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND,
8088
/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND.
8089
static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG,
8090
const ARMSubtarget *ST) {
8091
SDValue N0 = N->getOperand(0);
8093
// Check for sign- and zero-extensions of vector extract operations of 8-
8094
// and 16-bit vector elements. NEON supports these directly. They are
8095
// handled during DAG combining because type legalization will promote them
8096
// to 32-bit types and it is messy to recognize the operations after that.
8097
if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
8098
SDValue Vec = N0.getOperand(0);
8099
SDValue Lane = N0.getOperand(1);
8100
EVT VT = N->getValueType(0);
8101
EVT EltVT = N0.getValueType();
8102
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
8104
if (VT == MVT::i32 &&
8105
(EltVT == MVT::i8 || EltVT == MVT::i16) &&
8106
TLI.isTypeLegal(Vec.getValueType()) &&
8107
isa<ConstantSDNode>(Lane)) {
8110
switch (N->getOpcode()) {
8111
default: llvm_unreachable("unexpected opcode");
8112
case ISD::SIGN_EXTEND:
8113
Opc = ARMISD::VGETLANEs;
8115
case ISD::ZERO_EXTEND:
8116
case ISD::ANY_EXTEND:
8117
Opc = ARMISD::VGETLANEu;
8120
return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane);
8127
/// PerformSELECT_CCCombine - Target-specific DAG combining for ISD::SELECT_CC
8128
/// to match f32 max/min patterns to use NEON vmax/vmin instructions.
8129
static SDValue PerformSELECT_CCCombine(SDNode *N, SelectionDAG &DAG,
8130
const ARMSubtarget *ST) {
8131
// If the target supports NEON, try to use vmax/vmin instructions for f32
8132
// selects like "x < y ? x : y". Unless the NoNaNsFPMath option is set,
8133
// be careful about NaNs: NEON's vmax/vmin return NaN if either operand is
8134
// a NaN; only do the transformation when it matches that behavior.
8136
// For now only do this when using NEON for FP operations; if using VFP, it
8137
// is not obvious that the benefit outweighs the cost of switching to the
8139
if (!ST->hasNEON() || !ST->useNEONForSinglePrecisionFP() ||
8140
N->getValueType(0) != MVT::f32)
8143
SDValue CondLHS = N->getOperand(0);
8144
SDValue CondRHS = N->getOperand(1);
8145
SDValue LHS = N->getOperand(2);
8146
SDValue RHS = N->getOperand(3);
8147
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
8149
unsigned Opcode = 0;
8151
if (DAG.isEqualTo(LHS, CondLHS) && DAG.isEqualTo(RHS, CondRHS)) {
8152
IsReversed = false; // x CC y ? x : y
8153
} else if (DAG.isEqualTo(LHS, CondRHS) && DAG.isEqualTo(RHS, CondLHS)) {
8154
IsReversed = true ; // x CC y ? y : x
8168
// If LHS is NaN, an ordered comparison will be false and the result will
8169
// be the RHS, but vmin(NaN, RHS) = NaN. Avoid this by checking that LHS
8170
// != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8171
IsUnordered = (CC == ISD::SETULT || CC == ISD::SETULE);
8172
if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8174
// For less-than-or-equal comparisons, "+0 <= -0" will be true but vmin
8175
// will return -0, so vmin can only be used for unsafe math or if one of
8176
// the operands is known to be nonzero.
8177
if ((CC == ISD::SETLE || CC == ISD::SETOLE || CC == ISD::SETULE) &&
8178
!DAG.getTarget().Options.UnsafeFPMath &&
8179
!(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8181
Opcode = IsReversed ? ARMISD::FMAX : ARMISD::FMIN;
8190
// If LHS is NaN, an ordered comparison will be false and the result will
8191
// be the RHS, but vmax(NaN, RHS) = NaN. Avoid this by checking that LHS
8192
// != NaN. Likewise, for unordered comparisons, check for RHS != NaN.
8193
IsUnordered = (CC == ISD::SETUGT || CC == ISD::SETUGE);
8194
if (!DAG.isKnownNeverNaN(IsUnordered ? RHS : LHS))
8196
// For greater-than-or-equal comparisons, "-0 >= +0" will be true but vmax
8197
// will return +0, so vmax can only be used for unsafe math or if one of
8198
// the operands is known to be nonzero.
8199
if ((CC == ISD::SETGE || CC == ISD::SETOGE || CC == ISD::SETUGE) &&
8200
!DAG.getTarget().Options.UnsafeFPMath &&
8201
!(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
8203
Opcode = IsReversed ? ARMISD::FMIN : ARMISD::FMAX;
8209
return DAG.getNode(Opcode, N->getDebugLoc(), N->getValueType(0), LHS, RHS);
8212
/// PerformCMOVCombine - Target-specific DAG combining for ARMISD::CMOV.
8214
ARMTargetLowering::PerformCMOVCombine(SDNode *N, SelectionDAG &DAG) const {
8215
SDValue Cmp = N->getOperand(4);
8216
if (Cmp.getOpcode() != ARMISD::CMPZ)
8217
// Only looking at EQ and NE cases.
8220
EVT VT = N->getValueType(0);
8221
DebugLoc dl = N->getDebugLoc();
8222
SDValue LHS = Cmp.getOperand(0);
8223
SDValue RHS = Cmp.getOperand(1);
8224
SDValue FalseVal = N->getOperand(0);
8225
SDValue TrueVal = N->getOperand(1);
8226
SDValue ARMcc = N->getOperand(2);
8227
ARMCC::CondCodes CC =
8228
(ARMCC::CondCodes)cast<ConstantSDNode>(ARMcc)->getZExtValue();
8246
/// FIXME: Turn this into a target neutral optimization?
8248
if (CC == ARMCC::NE && FalseVal == RHS && FalseVal != LHS) {
8249
Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, TrueVal, ARMcc,
8250
N->getOperand(3), Cmp);
8251
} else if (CC == ARMCC::EQ && TrueVal == RHS) {
8253
SDValue NewCmp = getARMCmp(LHS, RHS, ISD::SETNE, ARMcc, DAG, dl);
8254
Res = DAG.getNode(ARMISD::CMOV, dl, VT, LHS, FalseVal, ARMcc,
8255
N->getOperand(3), NewCmp);
8258
if (Res.getNode()) {
8259
APInt KnownZero, KnownOne;
8260
APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits());
8261
DAG.ComputeMaskedBits(SDValue(N,0), Mask, KnownZero, KnownOne);
8262
// Capture demanded bits information that would be otherwise lost.
8263
if (KnownZero == 0xfffffffe)
8264
Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8265
DAG.getValueType(MVT::i1));
8266
else if (KnownZero == 0xffffff00)
8267
Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8268
DAG.getValueType(MVT::i8));
8269
else if (KnownZero == 0xffff0000)
8270
Res = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Res,
8271
DAG.getValueType(MVT::i16));
8277
SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N,
8278
DAGCombinerInfo &DCI) const {
8279
switch (N->getOpcode()) {
8281
case ISD::ADD: return PerformADDCombine(N, DCI, Subtarget);
8282
case ISD::SUB: return PerformSUBCombine(N, DCI);
8283
case ISD::MUL: return PerformMULCombine(N, DCI, Subtarget);
8284
case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
8285
case ISD::XOR: return PerformXORCombine(N, DCI, Subtarget);
8286
case ISD::AND: return PerformANDCombine(N, DCI, Subtarget);
8287
case ARMISD::BFI: return PerformBFICombine(N, DCI);
8288
case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
8289
case ARMISD::VMOVDRR: return PerformVMOVDRRCombine(N, DCI.DAG);
8290
case ISD::STORE: return PerformSTORECombine(N, DCI);
8291
case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DCI);
8292
case ISD::INSERT_VECTOR_ELT: return PerformInsertEltCombine(N, DCI);
8293
case ISD::VECTOR_SHUFFLE: return PerformVECTOR_SHUFFLECombine(N, DCI.DAG);
8294
case ARMISD::VDUPLANE: return PerformVDUPLANECombine(N, DCI);
8295
case ISD::FP_TO_SINT:
8296
case ISD::FP_TO_UINT: return PerformVCVTCombine(N, DCI, Subtarget);
8297
case ISD::FDIV: return PerformVDIVCombine(N, DCI, Subtarget);
8298
case ISD::INTRINSIC_WO_CHAIN: return PerformIntrinsicCombine(N, DCI.DAG);
8301
case ISD::SRL: return PerformShiftCombine(N, DCI.DAG, Subtarget);
8302
case ISD::SIGN_EXTEND:
8303
case ISD::ZERO_EXTEND:
8304
case ISD::ANY_EXTEND: return PerformExtendCombine(N, DCI.DAG, Subtarget);
8305
case ISD::SELECT_CC: return PerformSELECT_CCCombine(N, DCI.DAG, Subtarget);
8306
case ARMISD::CMOV: return PerformCMOVCombine(N, DCI.DAG);
8307
case ARMISD::VLD2DUP:
8308
case ARMISD::VLD3DUP:
8309
case ARMISD::VLD4DUP:
8310
return CombineBaseUpdate(N, DCI);
8311
case ISD::INTRINSIC_VOID:
8312
case ISD::INTRINSIC_W_CHAIN:
8313
switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
8314
case Intrinsic::arm_neon_vld1:
8315
case Intrinsic::arm_neon_vld2:
8316
case Intrinsic::arm_neon_vld3:
8317
case Intrinsic::arm_neon_vld4:
8318
case Intrinsic::arm_neon_vld2lane:
8319
case Intrinsic::arm_neon_vld3lane:
8320
case Intrinsic::arm_neon_vld4lane:
8321
case Intrinsic::arm_neon_vst1:
8322
case Intrinsic::arm_neon_vst2:
8323
case Intrinsic::arm_neon_vst3:
8324
case Intrinsic::arm_neon_vst4:
8325
case Intrinsic::arm_neon_vst2lane:
8326
case Intrinsic::arm_neon_vst3lane:
8327
case Intrinsic::arm_neon_vst4lane:
8328
return CombineBaseUpdate(N, DCI);
8336
bool ARMTargetLowering::isDesirableToTransformToIntegerOp(unsigned Opc,
8338
return (VT == MVT::f32) && (Opc == ISD::LOAD || Opc == ISD::STORE);
8341
bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
8342
if (!Subtarget->allowsUnalignedMem())
8345
switch (VT.getSimpleVT().SimpleTy) {
8352
// FIXME: VLD1 etc with standard alignment is legal.
8356
static bool memOpAlign(unsigned DstAlign, unsigned SrcAlign,
8357
unsigned AlignCheck) {
8358
return ((SrcAlign == 0 || SrcAlign % AlignCheck == 0) &&
8359
(DstAlign == 0 || DstAlign % AlignCheck == 0));
8362
EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
8363
unsigned DstAlign, unsigned SrcAlign,
8366
MachineFunction &MF) const {
8367
const Function *F = MF.getFunction();
8369
// See if we can use NEON instructions for this...
8371
!F->hasFnAttr(Attribute::NoImplicitFloat) &&
8372
Subtarget->hasNEON()) {
8373
if (memOpAlign(SrcAlign, DstAlign, 16) && Size >= 16) {
8375
} else if (memOpAlign(SrcAlign, DstAlign, 8) && Size >= 8) {
8380
// Lowering to i32/i16 if the size permits.
8383
} else if (Size >= 2) {
8387
// Let the target-independent logic figure it out.
8391
static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
8396
switch (VT.getSimpleVT().SimpleTy) {
8397
default: return false;
8412
if ((V & (Scale - 1)) != 0)
8415
return V == (V & ((1LL << 5) - 1));
8418
static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
8419
const ARMSubtarget *Subtarget) {
8426
switch (VT.getSimpleVT().SimpleTy) {
8427
default: return false;
8432
// + imm12 or - imm8
8434
return V == (V & ((1LL << 8) - 1));
8435
return V == (V & ((1LL << 12) - 1));
8438
// Same as ARM mode. FIXME: NEON?
8439
if (!Subtarget->hasVFP2())
8444
return V == (V & ((1LL << 8) - 1));
8448
/// isLegalAddressImmediate - Return true if the integer value can be used
8449
/// as the offset of the target addressing mode for load / store of the
8451
static bool isLegalAddressImmediate(int64_t V, EVT VT,
8452
const ARMSubtarget *Subtarget) {
8459
if (Subtarget->isThumb1Only())
8460
return isLegalT1AddressImmediate(V, VT);
8461
else if (Subtarget->isThumb2())
8462
return isLegalT2AddressImmediate(V, VT, Subtarget);
8467
switch (VT.getSimpleVT().SimpleTy) {
8468
default: return false;
8473
return V == (V & ((1LL << 12) - 1));
8476
return V == (V & ((1LL << 8) - 1));
8479
if (!Subtarget->hasVFP2()) // FIXME: NEON?
8484
return V == (V & ((1LL << 8) - 1));
8488
bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
8490
int Scale = AM.Scale;
8494
switch (VT.getSimpleVT().SimpleTy) {
8495
default: return false;
8504
return Scale == 2 || Scale == 4 || Scale == 8;
8507
if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8511
// Note, we allow "void" uses (basically, uses that aren't loads or
8512
// stores), because arm allows folding a scale into many arithmetic
8513
// operations. This should be made more precise and revisited later.
8515
// Allow r << imm, but the imm has to be a multiple of two.
8516
if (Scale & 1) return false;
8517
return isPowerOf2_32(Scale);
8521
/// isLegalAddressingMode - Return true if the addressing mode represented
8522
/// by AM is legal for this target, for a load/store of the specified type.
8523
bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
8525
EVT VT = getValueType(Ty, true);
8526
if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget))
8529
// Can never fold addr of global into load/store.
8534
case 0: // no scale reg, must be "r+i" or "r", or "i".
8537
if (Subtarget->isThumb1Only())
8541
// ARM doesn't support any R+R*scale+imm addr modes.
8548
if (Subtarget->isThumb2())
8549
return isLegalT2ScaledAddressingMode(AM, VT);
8551
int Scale = AM.Scale;
8552
switch (VT.getSimpleVT().SimpleTy) {
8553
default: return false;
8557
if (Scale < 0) Scale = -Scale;
8561
return isPowerOf2_32(Scale & ~1);
8565
if (((unsigned)AM.HasBaseReg + Scale) <= 2)
8570
// Note, we allow "void" uses (basically, uses that aren't loads or
8571
// stores), because arm allows folding a scale into many arithmetic
8572
// operations. This should be made more precise and revisited later.
8574
// Allow r << imm, but the imm has to be a multiple of two.
8575
if (Scale & 1) return false;
8576
return isPowerOf2_32(Scale);
8582
/// isLegalICmpImmediate - Return true if the specified immediate is legal
8583
/// icmp immediate, that is the target has icmp instructions which can compare
8584
/// a register against the immediate without having to materialize the
8585
/// immediate into a register.
8586
bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
8587
if (!Subtarget->isThumb())
8588
return ARM_AM::getSOImmVal(Imm) != -1;
8589
if (Subtarget->isThumb2())
8590
return ARM_AM::getT2SOImmVal(Imm) != -1;
8591
return Imm >= 0 && Imm <= 255;
8594
/// isLegalAddImmediate - Return true if the specified immediate is legal
8595
/// add immediate, that is the target has add instructions which can add
8596
/// a register with the immediate without having to materialize the
8597
/// immediate into a register.
8598
bool ARMTargetLowering::isLegalAddImmediate(int64_t Imm) const {
8599
return ARM_AM::getSOImmVal(Imm) != -1;
8602
static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
8603
bool isSEXTLoad, SDValue &Base,
8604
SDValue &Offset, bool &isInc,
8605
SelectionDAG &DAG) {
8606
if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8609
if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) {
8611
Base = Ptr->getOperand(0);
8612
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8613
int RHSC = (int)RHS->getZExtValue();
8614
if (RHSC < 0 && RHSC > -256) {
8615
assert(Ptr->getOpcode() == ISD::ADD);
8617
Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8621
isInc = (Ptr->getOpcode() == ISD::ADD);
8622
Offset = Ptr->getOperand(1);
8624
} else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) {
8626
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8627
int RHSC = (int)RHS->getZExtValue();
8628
if (RHSC < 0 && RHSC > -0x1000) {
8629
assert(Ptr->getOpcode() == ISD::ADD);
8631
Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8632
Base = Ptr->getOperand(0);
8637
if (Ptr->getOpcode() == ISD::ADD) {
8639
ARM_AM::ShiftOpc ShOpcVal=
8640
ARM_AM::getShiftOpcForNode(Ptr->getOperand(0).getOpcode());
8641
if (ShOpcVal != ARM_AM::no_shift) {
8642
Base = Ptr->getOperand(1);
8643
Offset = Ptr->getOperand(0);
8645
Base = Ptr->getOperand(0);
8646
Offset = Ptr->getOperand(1);
8651
isInc = (Ptr->getOpcode() == ISD::ADD);
8652
Base = Ptr->getOperand(0);
8653
Offset = Ptr->getOperand(1);
8657
// FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
8661
static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT,
8662
bool isSEXTLoad, SDValue &Base,
8663
SDValue &Offset, bool &isInc,
8664
SelectionDAG &DAG) {
8665
if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB)
8668
Base = Ptr->getOperand(0);
8669
if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) {
8670
int RHSC = (int)RHS->getZExtValue();
8671
if (RHSC < 0 && RHSC > -0x100) { // 8 bits.
8672
assert(Ptr->getOpcode() == ISD::ADD);
8674
Offset = DAG.getConstant(-RHSC, RHS->getValueType(0));
8676
} else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero.
8677
isInc = Ptr->getOpcode() == ISD::ADD;
8678
Offset = DAG.getConstant(RHSC, RHS->getValueType(0));
8686
/// getPreIndexedAddressParts - returns true by value, base pointer and
8687
/// offset pointer and addressing mode by reference if the node's address
8688
/// can be legally represented as pre-indexed load / store address.
8690
ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
8692
ISD::MemIndexedMode &AM,
8693
SelectionDAG &DAG) const {
8694
if (Subtarget->isThumb1Only())
8699
bool isSEXTLoad = false;
8700
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8701
Ptr = LD->getBasePtr();
8702
VT = LD->getMemoryVT();
8703
isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8704
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8705
Ptr = ST->getBasePtr();
8706
VT = ST->getMemoryVT();
8711
bool isLegal = false;
8712
if (Subtarget->isThumb2())
8713
isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8714
Offset, isInc, DAG);
8716
isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
8717
Offset, isInc, DAG);
8721
AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC;
8725
/// getPostIndexedAddressParts - returns true by value, base pointer and
8726
/// offset pointer and addressing mode by reference if this node can be
8727
/// combined with a load / store to form a post-indexed load / store.
8728
bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
8731
ISD::MemIndexedMode &AM,
8732
SelectionDAG &DAG) const {
8733
if (Subtarget->isThumb1Only())
8738
bool isSEXTLoad = false;
8739
if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8740
VT = LD->getMemoryVT();
8741
Ptr = LD->getBasePtr();
8742
isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
8743
} else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8744
VT = ST->getMemoryVT();
8745
Ptr = ST->getBasePtr();
8750
bool isLegal = false;
8751
if (Subtarget->isThumb2())
8752
isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8755
isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
8761
// Swap base ptr and offset to catch more post-index load / store when
8762
// it's legal. In Thumb2 mode, offset must be an immediate.
8763
if (Ptr == Offset && Op->getOpcode() == ISD::ADD &&
8764
!Subtarget->isThumb2())
8765
std::swap(Base, Offset);
8767
// Post-indexed load / store update the base pointer.
8772
AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
8776
void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8780
const SelectionDAG &DAG,
8781
unsigned Depth) const {
8782
KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0);
8783
switch (Op.getOpcode()) {
8785
case ARMISD::CMOV: {
8786
// Bits are known zero/one if known on the LHS and RHS.
8787
DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
8788
if (KnownZero == 0 && KnownOne == 0) return;
8790
APInt KnownZeroRHS, KnownOneRHS;
8791
DAG.ComputeMaskedBits(Op.getOperand(1), Mask,
8792
KnownZeroRHS, KnownOneRHS, Depth+1);
8793
KnownZero &= KnownZeroRHS;
8794
KnownOne &= KnownOneRHS;
8800
//===----------------------------------------------------------------------===//
8801
// ARM Inline Assembly Support
8802
//===----------------------------------------------------------------------===//
8804
bool ARMTargetLowering::ExpandInlineAsm(CallInst *CI) const {
8805
// Looking for "rev" which is V6+.
8806
if (!Subtarget->hasV6Ops())
8809
InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8810
std::string AsmStr = IA->getAsmString();
8811
SmallVector<StringRef, 4> AsmPieces;
8812
SplitString(AsmStr, AsmPieces, ";\n");
8814
switch (AsmPieces.size()) {
8815
default: return false;
8817
AsmStr = AsmPieces[0];
8819
SplitString(AsmStr, AsmPieces, " \t,");
8822
if (AsmPieces.size() == 3 &&
8823
AsmPieces[0] == "rev" && AsmPieces[1] == "$0" && AsmPieces[2] == "$1" &&
8824
IA->getConstraintString().compare(0, 4, "=l,l") == 0) {
8825
IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8826
if (Ty && Ty->getBitWidth() == 32)
8827
return IntrinsicLowering::LowerToByteSwap(CI);
8835
/// getConstraintType - Given a constraint letter, return the type of
8836
/// constraint it is for this target.
8837
ARMTargetLowering::ConstraintType
8838
ARMTargetLowering::getConstraintType(const std::string &Constraint) const {
8839
if (Constraint.size() == 1) {
8840
switch (Constraint[0]) {
8842
case 'l': return C_RegisterClass;
8843
case 'w': return C_RegisterClass;
8844
case 'h': return C_RegisterClass;
8845
case 'x': return C_RegisterClass;
8846
case 't': return C_RegisterClass;
8847
case 'j': return C_Other; // Constant for movw.
8848
// An address with a single base register. Due to the way we
8849
// currently handle addresses it is the same as an 'r' memory constraint.
8850
case 'Q': return C_Memory;
8852
} else if (Constraint.size() == 2) {
8853
switch (Constraint[0]) {
8855
// All 'U+' constraints are addresses.
8856
case 'U': return C_Memory;
8859
return TargetLowering::getConstraintType(Constraint);
8862
/// Examine constraint type and operand type and determine a weight value.
8863
/// This object must already have been set up with the operand type
8864
/// and the current alternative constraint selected.
8865
TargetLowering::ConstraintWeight
8866
ARMTargetLowering::getSingleConstraintMatchWeight(
8867
AsmOperandInfo &info, const char *constraint) const {
8868
ConstraintWeight weight = CW_Invalid;
8869
Value *CallOperandVal = info.CallOperandVal;
8870
// If we don't have a value, we can't do a match,
8871
// but allow it at the lowest weight.
8872
if (CallOperandVal == NULL)
8874
Type *type = CallOperandVal->getType();
8875
// Look at the constraint type.
8876
switch (*constraint) {
8878
weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
8881
if (type->isIntegerTy()) {
8882
if (Subtarget->isThumb())
8883
weight = CW_SpecificReg;
8885
weight = CW_Register;
8889
if (type->isFloatingPointTy())
8890
weight = CW_Register;
8896
typedef std::pair<unsigned, const TargetRegisterClass*> RCPair;
8898
ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
8900
if (Constraint.size() == 1) {
8901
// GCC ARM Constraint Letters
8902
switch (Constraint[0]) {
8903
case 'l': // Low regs or general regs.
8904
if (Subtarget->isThumb())
8905
return RCPair(0U, ARM::tGPRRegisterClass);
8907
return RCPair(0U, ARM::GPRRegisterClass);
8908
case 'h': // High regs or no regs.
8909
if (Subtarget->isThumb())
8910
return RCPair(0U, ARM::hGPRRegisterClass);
8913
return RCPair(0U, ARM::GPRRegisterClass);
8916
return RCPair(0U, ARM::SPRRegisterClass);
8917
if (VT.getSizeInBits() == 64)
8918
return RCPair(0U, ARM::DPRRegisterClass);
8919
if (VT.getSizeInBits() == 128)
8920
return RCPair(0U, ARM::QPRRegisterClass);
8924
return RCPair(0U, ARM::SPR_8RegisterClass);
8925
if (VT.getSizeInBits() == 64)
8926
return RCPair(0U, ARM::DPR_8RegisterClass);
8927
if (VT.getSizeInBits() == 128)
8928
return RCPair(0U, ARM::QPR_8RegisterClass);
8932
return RCPair(0U, ARM::SPRRegisterClass);
8936
if (StringRef("{cc}").equals_lower(Constraint))
8937
return std::make_pair(unsigned(ARM::CPSR), ARM::CCRRegisterClass);
8939
return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
8942
/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8943
/// vector. If it is invalid, don't add anything to Ops.
8944
void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8945
std::string &Constraint,
8946
std::vector<SDValue>&Ops,
8947
SelectionDAG &DAG) const {
8948
SDValue Result(0, 0);
8950
// Currently only support length 1 constraints.
8951
if (Constraint.length() != 1) return;
8953
char ConstraintLetter = Constraint[0];
8954
switch (ConstraintLetter) {
8957
case 'I': case 'J': case 'K': case 'L':
8958
case 'M': case 'N': case 'O':
8959
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
8963
int64_t CVal64 = C->getSExtValue();
8964
int CVal = (int) CVal64;
8965
// None of these constraints allow values larger than 32 bits. Check
8966
// that the value fits in an int.
8970
switch (ConstraintLetter) {
8972
// Constant suitable for movw, must be between 0 and
8974
if (Subtarget->hasV6T2Ops())
8975
if (CVal >= 0 && CVal <= 65535)
8979
if (Subtarget->isThumb1Only()) {
8980
// This must be a constant between 0 and 255, for ADD
8982
if (CVal >= 0 && CVal <= 255)
8984
} else if (Subtarget->isThumb2()) {
8985
// A constant that can be used as an immediate value in a
8986
// data-processing instruction.
8987
if (ARM_AM::getT2SOImmVal(CVal) != -1)
8990
// A constant that can be used as an immediate value in a
8991
// data-processing instruction.
8992
if (ARM_AM::getSOImmVal(CVal) != -1)
8998
if (Subtarget->isThumb()) { // FIXME thumb2
8999
// This must be a constant between -255 and -1, for negated ADD
9000
// immediates. This can be used in GCC with an "n" modifier that
9001
// prints the negated value, for use with SUB instructions. It is
9002
// not useful otherwise but is implemented for compatibility.
9003
if (CVal >= -255 && CVal <= -1)
9006
// This must be a constant between -4095 and 4095. It is not clear
9007
// what this constraint is intended for. Implemented for
9008
// compatibility with GCC.
9009
if (CVal >= -4095 && CVal <= 4095)
9015
if (Subtarget->isThumb1Only()) {
9016
// A 32-bit value where only one byte has a nonzero value. Exclude
9017
// zero to match GCC. This constraint is used by GCC internally for
9018
// constants that can be loaded with a move/shift combination.
9019
// It is not useful otherwise but is implemented for compatibility.
9020
if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal))
9022
} else if (Subtarget->isThumb2()) {
9023
// A constant whose bitwise inverse can be used as an immediate
9024
// value in a data-processing instruction. This can be used in GCC
9025
// with a "B" modifier that prints the inverted value, for use with
9026
// BIC and MVN instructions. It is not useful otherwise but is
9027
// implemented for compatibility.
9028
if (ARM_AM::getT2SOImmVal(~CVal) != -1)
9031
// A constant whose bitwise inverse can be used as an immediate
9032
// value in a data-processing instruction. This can be used in GCC
9033
// with a "B" modifier that prints the inverted value, for use with
9034
// BIC and MVN instructions. It is not useful otherwise but is
9035
// implemented for compatibility.
9036
if (ARM_AM::getSOImmVal(~CVal) != -1)
9042
if (Subtarget->isThumb1Only()) {
9043
// This must be a constant between -7 and 7,
9044
// for 3-operand ADD/SUB immediate instructions.
9045
if (CVal >= -7 && CVal < 7)
9047
} else if (Subtarget->isThumb2()) {
9048
// A constant whose negation can be used as an immediate value in a
9049
// data-processing instruction. This can be used in GCC with an "n"
9050
// modifier that prints the negated value, for use with SUB
9051
// instructions. It is not useful otherwise but is implemented for
9053
if (ARM_AM::getT2SOImmVal(-CVal) != -1)
9056
// A constant whose negation can be used as an immediate value in a
9057
// data-processing instruction. This can be used in GCC with an "n"
9058
// modifier that prints the negated value, for use with SUB
9059
// instructions. It is not useful otherwise but is implemented for
9061
if (ARM_AM::getSOImmVal(-CVal) != -1)
9067
if (Subtarget->isThumb()) { // FIXME thumb2
9068
// This must be a multiple of 4 between 0 and 1020, for
9069
// ADD sp + immediate.
9070
if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0))
9073
// A power of two or a constant between 0 and 32. This is used in
9074
// GCC for the shift amount on shifted register operands, but it is
9075
// useful in general for any shift amounts.
9076
if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0))
9082
if (Subtarget->isThumb()) { // FIXME thumb2
9083
// This must be a constant between 0 and 31, for shift amounts.
9084
if (CVal >= 0 && CVal <= 31)
9090
if (Subtarget->isThumb()) { // FIXME thumb2
9091
// This must be a multiple of 4 between -508 and 508, for
9092
// ADD/SUB sp = sp + immediate.
9093
if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0))
9098
Result = DAG.getTargetConstant(CVal, Op.getValueType());
9102
if (Result.getNode()) {
9103
Ops.push_back(Result);
9106
return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
9110
ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
9111
// The ARM target isn't yet aware of offsets.
9115
bool ARM::isBitFieldInvertedMask(unsigned v) {
9116
if (v == 0xffffffff)
9118
// there can be 1's on either or both "outsides", all the "inside"
9120
unsigned int lsb = 0, msb = 31;
9121
while (v & (1 << msb)) --msb;
9122
while (v & (1 << lsb)) ++lsb;
9123
for (unsigned int i = lsb; i <= msb; ++i) {
9130
/// isFPImmLegal - Returns true if the target can instruction select the
9131
/// specified FP immediate natively. If false, the legalizer will
9132
/// materialize the FP immediate as a load from a constant pool.
9133
bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
9134
if (!Subtarget->hasVFP3())
9137
return ARM_AM::getFP32Imm(Imm) != -1;
9139
return ARM_AM::getFP64Imm(Imm) != -1;
9143
/// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
9144
/// MemIntrinsicNodes. The associated MachineMemOperands record the alignment
9145
/// specified in the intrinsic calls.
9146
bool ARMTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
9148
unsigned Intrinsic) const {
9149
switch (Intrinsic) {
9150
case Intrinsic::arm_neon_vld1:
9151
case Intrinsic::arm_neon_vld2:
9152
case Intrinsic::arm_neon_vld3:
9153
case Intrinsic::arm_neon_vld4:
9154
case Intrinsic::arm_neon_vld2lane:
9155
case Intrinsic::arm_neon_vld3lane:
9156
case Intrinsic::arm_neon_vld4lane: {
9157
Info.opc = ISD::INTRINSIC_W_CHAIN;
9158
// Conservatively set memVT to the entire set of vectors loaded.
9159
uint64_t NumElts = getTargetData()->getTypeAllocSize(I.getType()) / 8;
9160
Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9161
Info.ptrVal = I.getArgOperand(0);
9163
Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9164
Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9165
Info.vol = false; // volatile loads with NEON intrinsics not supported
9166
Info.readMem = true;
9167
Info.writeMem = false;
9170
case Intrinsic::arm_neon_vst1:
9171
case Intrinsic::arm_neon_vst2:
9172
case Intrinsic::arm_neon_vst3:
9173
case Intrinsic::arm_neon_vst4:
9174
case Intrinsic::arm_neon_vst2lane:
9175
case Intrinsic::arm_neon_vst3lane:
9176
case Intrinsic::arm_neon_vst4lane: {
9177
Info.opc = ISD::INTRINSIC_VOID;
9178
// Conservatively set memVT to the entire set of vectors stored.
9179
unsigned NumElts = 0;
9180
for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
9181
Type *ArgTy = I.getArgOperand(ArgI)->getType();
9182
if (!ArgTy->isVectorTy())
9184
NumElts += getTargetData()->getTypeAllocSize(ArgTy) / 8;
9186
Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
9187
Info.ptrVal = I.getArgOperand(0);
9189
Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
9190
Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
9191
Info.vol = false; // volatile stores with NEON intrinsics not supported
9192
Info.readMem = false;
9193
Info.writeMem = true;
9196
case Intrinsic::arm_strexd: {
9197
Info.opc = ISD::INTRINSIC_W_CHAIN;
9198
Info.memVT = MVT::i64;
9199
Info.ptrVal = I.getArgOperand(2);
9203
Info.readMem = false;
9204
Info.writeMem = true;
9207
case Intrinsic::arm_ldrexd: {
9208
Info.opc = ISD::INTRINSIC_W_CHAIN;
9209
Info.memVT = MVT::i64;
9210
Info.ptrVal = I.getArgOperand(0);
9214
Info.readMem = true;
9215
Info.writeMem = false;