1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
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 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUFrameLowering.h"
19 #include "AMDGPUIntrinsicInfo.h"
20 #include "AMDGPURegisterInfo.h"
21 #include "AMDGPUSubtarget.h"
22 #include "R600MachineFunctionInfo.h"
23 #include "SIMachineFunctionInfo.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/SelectionDAG.h"
28 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/DiagnosticPrinter.h"
37 /// Diagnostic information for unimplemented or unsupported feature reporting.
38 class DiagnosticInfoUnsupported : public DiagnosticInfo {
40 const Twine &Description;
45 static int getKindID() {
47 KindID = llvm::getNextAvailablePluginDiagnosticKind();
52 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
53 DiagnosticSeverity Severity = DS_Error)
54 : DiagnosticInfo(getKindID(), Severity),
58 const Function &getFunction() const { return Fn; }
59 const Twine &getDescription() const { return Description; }
61 void print(DiagnosticPrinter &DP) const override {
62 DP << "unsupported " << getDescription() << " in " << Fn.getName();
65 static bool classof(const DiagnosticInfo *DI) {
66 return DI->getKind() == getKindID();
70 int DiagnosticInfoUnsupported::KindID = 0;
74 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
75 CCValAssign::LocInfo LocInfo,
76 ISD::ArgFlagsTy ArgFlags, CCState &State) {
77 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
78 ArgFlags.getOrigAlign());
79 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
84 #include "AMDGPUGenCallingConv.inc"
86 // Find a larger type to do a load / store of a vector with.
87 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
88 unsigned StoreSize = VT.getStoreSizeInBits();
90 return EVT::getIntegerVT(Ctx, StoreSize);
92 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
93 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
96 // Type for a vector that will be loaded to.
97 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
98 unsigned StoreSize = VT.getStoreSizeInBits();
100 return EVT::getIntegerVT(Ctx, 32);
102 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
105 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
106 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
108 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
110 setOperationAction(ISD::Constant, MVT::i32, Legal);
111 setOperationAction(ISD::Constant, MVT::i64, Legal);
112 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
113 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
115 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
116 setOperationAction(ISD::BRIND, MVT::Other, Expand);
118 // We need to custom lower some of the intrinsics
119 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
121 // Library functions. These default to Expand, but we have instructions
123 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
124 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
125 setOperationAction(ISD::FPOW, MVT::f32, Legal);
126 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
127 setOperationAction(ISD::FABS, MVT::f32, Legal);
128 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
129 setOperationAction(ISD::FRINT, MVT::f32, Legal);
130 setOperationAction(ISD::FROUND, MVT::f32, Legal);
131 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
133 setOperationAction(ISD::FREM, MVT::f32, Custom);
134 setOperationAction(ISD::FREM, MVT::f64, Custom);
136 // Lower floating point store/load to integer store/load to reduce the number
137 // of patterns in tablegen.
138 setOperationAction(ISD::STORE, MVT::f32, Promote);
139 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
141 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
142 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
144 setOperationAction(ISD::STORE, MVT::i64, Promote);
145 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
147 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
148 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
150 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
151 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
153 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
154 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
156 setOperationAction(ISD::STORE, MVT::f64, Promote);
157 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
159 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
160 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
162 // Custom lowering of vector stores is required for local address space
164 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
165 // XXX: Native v2i32 local address space stores are possible, but not
166 // currently implemented.
167 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
169 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
170 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
171 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
173 // XXX: This can be change to Custom, once ExpandVectorStores can
174 // handle 64-bit stores.
175 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
177 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
178 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
179 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
180 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
181 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
184 setOperationAction(ISD::LOAD, MVT::f32, Promote);
185 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
187 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
188 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
190 setOperationAction(ISD::LOAD, MVT::i64, Promote);
191 AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
193 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
194 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
196 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
197 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
199 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
200 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
202 setOperationAction(ISD::LOAD, MVT::f64, Promote);
203 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
205 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
206 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
208 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
209 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
210 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
211 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
212 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
213 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
214 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
215 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
216 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
217 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
219 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
220 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
222 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
223 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
224 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
225 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
226 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
227 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
228 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
229 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
230 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
232 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
234 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
235 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
236 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
237 setOperationAction(ISD::FRINT, MVT::f64, Custom);
238 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
241 if (!Subtarget->hasBFI()) {
242 // fcopysign can be done in a single instruction with BFI.
243 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
244 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
247 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
249 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
250 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
251 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
253 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
254 for (MVT VT : ScalarIntVTs) {
255 setOperationAction(ISD::SREM, VT, Expand);
256 setOperationAction(ISD::SDIV, VT, Expand);
258 // GPU does not have divrem function for signed or unsigned.
259 setOperationAction(ISD::SDIVREM, VT, Custom);
260 setOperationAction(ISD::UDIVREM, VT, Custom);
262 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
263 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
264 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
266 setOperationAction(ISD::BSWAP, VT, Expand);
267 setOperationAction(ISD::CTTZ, VT, Expand);
268 setOperationAction(ISD::CTLZ, VT, Expand);
271 if (!Subtarget->hasBCNT(32))
272 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
274 if (!Subtarget->hasBCNT(64))
275 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
277 // The hardware supports 32-bit ROTR, but not ROTL.
278 setOperationAction(ISD::ROTL, MVT::i32, Expand);
279 setOperationAction(ISD::ROTL, MVT::i64, Expand);
280 setOperationAction(ISD::ROTR, MVT::i64, Expand);
282 setOperationAction(ISD::MUL, MVT::i64, Expand);
283 setOperationAction(ISD::MULHU, MVT::i64, Expand);
284 setOperationAction(ISD::MULHS, MVT::i64, Expand);
285 setOperationAction(ISD::UDIV, MVT::i32, Expand);
286 setOperationAction(ISD::UREM, MVT::i32, Expand);
287 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
288 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
289 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
290 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
291 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
293 if (!Subtarget->hasFFBH())
294 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
296 if (!Subtarget->hasFFBL())
297 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
299 static const MVT::SimpleValueType VectorIntTypes[] = {
300 MVT::v2i32, MVT::v4i32
303 for (MVT VT : VectorIntTypes) {
304 // Expand the following operations for the current type by default.
305 setOperationAction(ISD::ADD, VT, Expand);
306 setOperationAction(ISD::AND, VT, Expand);
307 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
308 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
309 setOperationAction(ISD::MUL, VT, Expand);
310 setOperationAction(ISD::OR, VT, Expand);
311 setOperationAction(ISD::SHL, VT, Expand);
312 setOperationAction(ISD::SRA, VT, Expand);
313 setOperationAction(ISD::SRL, VT, Expand);
314 setOperationAction(ISD::ROTL, VT, Expand);
315 setOperationAction(ISD::ROTR, VT, Expand);
316 setOperationAction(ISD::SUB, VT, Expand);
317 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
318 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
319 setOperationAction(ISD::SDIV, VT, Expand);
320 setOperationAction(ISD::UDIV, VT, Expand);
321 setOperationAction(ISD::SREM, VT, Expand);
322 setOperationAction(ISD::UREM, VT, Expand);
323 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
324 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
325 setOperationAction(ISD::SDIVREM, VT, Custom);
326 setOperationAction(ISD::UDIVREM, VT, Custom);
327 setOperationAction(ISD::ADDC, VT, Expand);
328 setOperationAction(ISD::SUBC, VT, Expand);
329 setOperationAction(ISD::ADDE, VT, Expand);
330 setOperationAction(ISD::SUBE, VT, Expand);
331 setOperationAction(ISD::SELECT, VT, Expand);
332 setOperationAction(ISD::VSELECT, VT, Expand);
333 setOperationAction(ISD::SELECT_CC, VT, Expand);
334 setOperationAction(ISD::XOR, VT, Expand);
335 setOperationAction(ISD::BSWAP, VT, Expand);
336 setOperationAction(ISD::CTPOP, VT, Expand);
337 setOperationAction(ISD::CTTZ, VT, Expand);
338 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
339 setOperationAction(ISD::CTLZ, VT, Expand);
340 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
341 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
344 static const MVT::SimpleValueType FloatVectorTypes[] = {
345 MVT::v2f32, MVT::v4f32
348 for (MVT VT : FloatVectorTypes) {
349 setOperationAction(ISD::FABS, VT, Expand);
350 setOperationAction(ISD::FADD, VT, Expand);
351 setOperationAction(ISD::FCEIL, VT, Expand);
352 setOperationAction(ISD::FCOS, VT, Expand);
353 setOperationAction(ISD::FDIV, VT, Expand);
354 setOperationAction(ISD::FEXP2, VT, Expand);
355 setOperationAction(ISD::FLOG2, VT, Expand);
356 setOperationAction(ISD::FREM, VT, Expand);
357 setOperationAction(ISD::FPOW, VT, Expand);
358 setOperationAction(ISD::FFLOOR, VT, Expand);
359 setOperationAction(ISD::FTRUNC, VT, Expand);
360 setOperationAction(ISD::FMUL, VT, Expand);
361 setOperationAction(ISD::FMA, VT, Expand);
362 setOperationAction(ISD::FRINT, VT, Expand);
363 setOperationAction(ISD::FNEARBYINT, VT, Expand);
364 setOperationAction(ISD::FSQRT, VT, Expand);
365 setOperationAction(ISD::FSIN, VT, Expand);
366 setOperationAction(ISD::FSUB, VT, Expand);
367 setOperationAction(ISD::FNEG, VT, Expand);
368 setOperationAction(ISD::SELECT, VT, Expand);
369 setOperationAction(ISD::VSELECT, VT, Expand);
370 setOperationAction(ISD::SELECT_CC, VT, Expand);
371 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
372 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
375 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
376 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
378 setTargetDAGCombine(ISD::MUL);
379 setTargetDAGCombine(ISD::SELECT_CC);
380 setTargetDAGCombine(ISD::STORE);
382 setSchedulingPreference(Sched::RegPressure);
383 setJumpIsExpensive(true);
385 // SI at least has hardware support for floating point exceptions, but no way
386 // of using or handling them is implemented. They are also optional in OpenCL
388 setHasFloatingPointExceptions(false);
390 setSelectIsExpensive(false);
391 PredictableSelectIsExpensive = false;
393 // There are no integer divide instructions, and these expand to a pretty
394 // large sequence of instructions.
395 setIntDivIsCheap(false);
396 setPow2SDivIsCheap(false);
398 // FIXME: Need to really handle these.
399 MaxStoresPerMemcpy = 4096;
400 MaxStoresPerMemmove = 4096;
401 MaxStoresPerMemset = 4096;
404 //===----------------------------------------------------------------------===//
405 // Target Information
406 //===----------------------------------------------------------------------===//
408 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
412 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
416 // The backend supports 32 and 64 bit floating point immediates.
417 // FIXME: Why are we reporting vectors of FP immediates as legal?
418 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
419 EVT ScalarVT = VT.getScalarType();
420 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
423 // We don't want to shrink f64 / f32 constants.
424 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
425 EVT ScalarVT = VT.getScalarType();
426 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
429 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
431 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
434 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
435 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
437 return ((LScalarSize <= CastScalarSize) ||
438 (CastScalarSize >= 32) ||
442 //===---------------------------------------------------------------------===//
444 //===---------------------------------------------------------------------===//
446 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
447 assert(VT.isFloatingPoint());
448 return VT == MVT::f32 || VT == MVT::f64;
451 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
452 assert(VT.isFloatingPoint());
453 return VT == MVT::f32 || VT == MVT::f64;
456 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
457 // Truncate is just accessing a subregister.
458 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
461 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
462 // Truncate is just accessing a subregister.
463 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
464 (Dest->getPrimitiveSizeInBits() % 32 == 0);
467 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
468 const DataLayout *DL = getDataLayout();
469 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
470 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
472 return SrcSize == 32 && DestSize == 64;
475 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
476 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
477 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
478 // this will enable reducing 64-bit operations the 32-bit, which is always
480 return Src == MVT::i32 && Dest == MVT::i64;
483 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
484 return isZExtFree(Val.getValueType(), VT2);
487 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
488 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
489 // limited number of native 64-bit operations. Shrinking an operation to fit
490 // in a single 32-bit register should always be helpful. As currently used,
491 // this is much less general than the name suggests, and is only used in
492 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
493 // not profitable, and may actually be harmful.
494 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
497 //===---------------------------------------------------------------------===//
498 // TargetLowering Callbacks
499 //===---------------------------------------------------------------------===//
501 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
502 const SmallVectorImpl<ISD::InputArg> &Ins) const {
504 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
507 SDValue AMDGPUTargetLowering::LowerReturn(
509 CallingConv::ID CallConv,
511 const SmallVectorImpl<ISD::OutputArg> &Outs,
512 const SmallVectorImpl<SDValue> &OutVals,
513 SDLoc DL, SelectionDAG &DAG) const {
514 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
517 //===---------------------------------------------------------------------===//
518 // Target specific lowering
519 //===---------------------------------------------------------------------===//
521 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
522 SmallVectorImpl<SDValue> &InVals) const {
523 SDValue Callee = CLI.Callee;
524 SelectionDAG &DAG = CLI.DAG;
526 const Function &Fn = *DAG.getMachineFunction().getFunction();
528 StringRef FuncName("<unknown>");
530 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
531 FuncName = G->getSymbol();
532 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
533 FuncName = G->getGlobal()->getName();
535 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
536 DAG.getContext()->diagnose(NoCalls);
540 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
541 SelectionDAG &DAG) const {
542 switch (Op.getOpcode()) {
544 Op.getNode()->dump();
545 llvm_unreachable("Custom lowering code for this"
546 "instruction is not implemented yet!");
548 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
549 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
550 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
551 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
552 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
553 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
554 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
555 case ISD::FREM: return LowerFREM(Op, DAG);
556 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
557 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
558 case ISD::FRINT: return LowerFRINT(Op, DAG);
559 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
560 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
561 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
562 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
563 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
564 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
569 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
570 SmallVectorImpl<SDValue> &Results,
571 SelectionDAG &DAG) const {
572 switch (N->getOpcode()) {
573 case ISD::SIGN_EXTEND_INREG:
574 // Different parts of legalization seem to interpret which type of
575 // sign_extend_inreg is the one to check for custom lowering. The extended
576 // from type is what really matters, but some places check for custom
577 // lowering of the result type. This results in trying to use
578 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
579 // nothing here and let the illegal result integer be handled normally.
582 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
586 Results.push_back(SDValue(Node, 0));
587 Results.push_back(SDValue(Node, 1));
588 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
590 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
594 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
595 if (Lowered.getNode())
596 Results.push_back(Lowered);
604 // FIXME: This implements accesses to initialized globals in the constant
605 // address space by copying them to private and accessing that. It does not
606 // properly handle illegal types or vectors. The private vector loads are not
607 // scalarized, and the illegal scalars hit an assertion. This technique will not
608 // work well with large initializers, and this should eventually be
609 // removed. Initialized globals should be placed into a data section that the
610 // runtime will load into a buffer before the kernel is executed. Uses of the
611 // global need to be replaced with a pointer loaded from an implicit kernel
612 // argument into this buffer holding the copy of the data, which will remove the
613 // need for any of this.
614 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
615 const GlobalValue *GV,
616 const SDValue &InitPtr,
618 SelectionDAG &DAG) const {
619 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
621 Type *InitTy = Init->getType();
623 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
624 EVT VT = EVT::getEVT(InitTy);
625 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
626 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
627 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
628 TD->getPrefTypeAlignment(InitTy));
631 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
632 EVT VT = EVT::getEVT(CFP->getType());
633 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
634 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
635 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
636 TD->getPrefTypeAlignment(CFP->getType()));
639 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
640 const StructLayout *SL = TD->getStructLayout(ST);
642 EVT PtrVT = InitPtr.getValueType();
643 SmallVector<SDValue, 8> Chains;
645 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
646 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
647 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
649 Constant *Elt = Init->getAggregateElement(I);
650 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
653 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
656 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
657 EVT PtrVT = InitPtr.getValueType();
659 unsigned NumElements;
660 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
661 NumElements = AT->getNumElements();
662 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
663 NumElements = VT->getNumElements();
665 llvm_unreachable("Unexpected type");
667 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
668 SmallVector<SDValue, 8> Chains;
669 for (unsigned i = 0; i < NumElements; ++i) {
670 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
671 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
673 Constant *Elt = Init->getAggregateElement(i);
674 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
677 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
680 if (isa<UndefValue>(Init)) {
681 EVT VT = EVT::getEVT(InitTy);
682 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
683 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
684 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
685 TD->getPrefTypeAlignment(InitTy));
689 llvm_unreachable("Unhandled constant initializer");
692 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
694 SelectionDAG &DAG) const {
696 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
697 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
698 const GlobalValue *GV = G->getGlobal();
700 switch (G->getAddressSpace()) {
701 default: llvm_unreachable("Global Address lowering not implemented for this "
703 case AMDGPUAS::LOCAL_ADDRESS: {
704 // XXX: What does the value of G->getOffset() mean?
705 assert(G->getOffset() == 0 &&
706 "Do not know what to do with an non-zero offset");
709 if (MFI->LocalMemoryObjects.count(GV) == 0) {
710 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
711 Offset = MFI->LDSSize;
712 MFI->LocalMemoryObjects[GV] = Offset;
713 // XXX: Account for alignment?
714 MFI->LDSSize += Size;
716 Offset = MFI->LocalMemoryObjects[GV];
719 return DAG.getConstant(Offset, getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
721 case AMDGPUAS::CONSTANT_ADDRESS: {
722 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
723 Type *EltType = GV->getType()->getElementType();
724 unsigned Size = TD->getTypeAllocSize(EltType);
725 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
727 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
728 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
730 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
731 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
733 const GlobalVariable *Var = cast<GlobalVariable>(GV);
734 if (!Var->hasInitializer()) {
735 // This has no use, but bugpoint will hit it.
736 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
739 const Constant *Init = Var->getInitializer();
740 SmallVector<SDNode*, 8> WorkList;
742 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
743 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
744 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
746 WorkList.push_back(*I);
748 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
749 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
750 E = WorkList.end(); I != E; ++I) {
751 SmallVector<SDValue, 8> Ops;
752 Ops.push_back(Chain);
753 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
754 Ops.push_back((*I)->getOperand(i));
756 DAG.UpdateNodeOperands(*I, Ops);
758 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
763 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
764 SelectionDAG &DAG) const {
765 SmallVector<SDValue, 8> Args;
766 SDValue A = Op.getOperand(0);
767 SDValue B = Op.getOperand(1);
769 DAG.ExtractVectorElements(A, Args);
770 DAG.ExtractVectorElements(B, Args);
772 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
775 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
776 SelectionDAG &DAG) const {
778 SmallVector<SDValue, 8> Args;
779 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
780 EVT VT = Op.getValueType();
781 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
782 VT.getVectorNumElements());
784 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
787 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
788 SelectionDAG &DAG) const {
790 MachineFunction &MF = DAG.getMachineFunction();
791 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
792 getTargetMachine().getSubtargetImpl()->getFrameLowering());
794 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
796 unsigned FrameIndex = FIN->getIndex();
797 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
798 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
802 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
803 SelectionDAG &DAG) const {
804 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
806 EVT VT = Op.getValueType();
808 switch (IntrinsicID) {
810 case AMDGPUIntrinsic::AMDGPU_abs:
811 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
812 return LowerIntrinsicIABS(Op, DAG);
813 case AMDGPUIntrinsic::AMDGPU_lrp:
814 return LowerIntrinsicLRP(Op, DAG);
815 case AMDGPUIntrinsic::AMDGPU_fract:
816 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
817 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
819 case AMDGPUIntrinsic::AMDGPU_clamp:
820 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
821 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
822 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
824 case Intrinsic::AMDGPU_div_scale: {
825 // 3rd parameter required to be a constant.
826 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
828 return DAG.getUNDEF(VT);
830 // Translate to the operands expected by the machine instruction. The
831 // first parameter must be the same as the first instruction.
832 SDValue Numerator = Op.getOperand(1);
833 SDValue Denominator = Op.getOperand(2);
835 // Note this order is opposite of the machine instruction's operations,
836 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
837 // intrinsic has the numerator as the first operand to match a normal
838 // division operation.
840 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
842 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
843 Denominator, Numerator);
846 case Intrinsic::AMDGPU_div_fmas:
847 // FIXME: Dropping bool parameter. Work is needed to support the implicit
849 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
850 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
852 case Intrinsic::AMDGPU_div_fixup:
853 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
854 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
856 case Intrinsic::AMDGPU_trig_preop:
857 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
858 Op.getOperand(1), Op.getOperand(2));
860 case Intrinsic::AMDGPU_rcp:
861 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
863 case Intrinsic::AMDGPU_rsq:
864 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
866 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
867 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
869 case Intrinsic::AMDGPU_rsq_clamped:
870 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
872 case Intrinsic::AMDGPU_ldexp:
873 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
876 case AMDGPUIntrinsic::AMDGPU_imax:
877 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
879 case AMDGPUIntrinsic::AMDGPU_umax:
880 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
882 case AMDGPUIntrinsic::AMDGPU_imin:
883 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
885 case AMDGPUIntrinsic::AMDGPU_umin:
886 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
889 case AMDGPUIntrinsic::AMDGPU_umul24:
890 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
891 Op.getOperand(1), Op.getOperand(2));
893 case AMDGPUIntrinsic::AMDGPU_imul24:
894 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
895 Op.getOperand(1), Op.getOperand(2));
897 case AMDGPUIntrinsic::AMDGPU_umad24:
898 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
899 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
901 case AMDGPUIntrinsic::AMDGPU_imad24:
902 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
903 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
905 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
906 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
908 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
909 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
911 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
912 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
914 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
915 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
917 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
918 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
923 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
924 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
929 case AMDGPUIntrinsic::AMDGPU_bfi:
930 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
935 case AMDGPUIntrinsic::AMDGPU_bfm:
936 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
940 case AMDGPUIntrinsic::AMDGPU_brev:
941 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
943 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
944 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
946 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
947 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
948 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
949 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
953 ///IABS(a) = SMAX(sub(0, a), a)
954 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
955 SelectionDAG &DAG) const {
957 EVT VT = Op.getValueType();
958 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
961 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
964 /// Linear Interpolation
965 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
966 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
967 SelectionDAG &DAG) const {
969 EVT VT = Op.getValueType();
970 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
971 DAG.getConstantFP(1.0f, MVT::f32),
973 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
975 return DAG.getNode(ISD::FADD, DL, VT,
976 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
980 /// \brief Generate Min/Max node
981 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
982 SelectionDAG &DAG) const {
984 EVT VT = N->getValueType(0);
986 SDValue LHS = N->getOperand(0);
987 SDValue RHS = N->getOperand(1);
988 SDValue True = N->getOperand(2);
989 SDValue False = N->getOperand(3);
990 SDValue CC = N->getOperand(4);
992 if (VT != MVT::f32 ||
993 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
997 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1006 case ISD::SETFALSE2:
1011 llvm_unreachable("Operation should already be optimised!");
1018 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
1019 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1027 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
1028 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1030 case ISD::SETCC_INVALID:
1031 llvm_unreachable("Invalid setcc condcode!");
1036 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1037 SelectionDAG &DAG) const {
1038 LoadSDNode *Load = cast<LoadSDNode>(Op);
1039 EVT MemVT = Load->getMemoryVT();
1040 EVT MemEltVT = MemVT.getVectorElementType();
1042 EVT LoadVT = Op.getValueType();
1043 EVT EltVT = LoadVT.getVectorElementType();
1044 EVT PtrVT = Load->getBasePtr().getValueType();
1046 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1047 SmallVector<SDValue, 8> Loads;
1048 SmallVector<SDValue, 8> Chains;
1051 unsigned MemEltSize = MemEltVT.getStoreSize();
1052 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1054 for (unsigned i = 0; i < NumElts; ++i) {
1055 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1056 DAG.getConstant(i * MemEltSize, PtrVT));
1059 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1060 Load->getChain(), Ptr,
1061 SrcValue.getWithOffset(i * MemEltSize),
1062 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1063 Load->isInvariant(), Load->getAlignment());
1064 Loads.push_back(NewLoad.getValue(0));
1065 Chains.push_back(NewLoad.getValue(1));
1069 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1070 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1073 return DAG.getMergeValues(Ops, SL);
1076 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1077 SelectionDAG &DAG) const {
1078 EVT VT = Op.getValueType();
1080 // If this is a 2 element vector, we really want to scalarize and not create
1081 // weird 1 element vectors.
1082 if (VT.getVectorNumElements() == 2)
1083 return ScalarizeVectorLoad(Op, DAG);
1085 LoadSDNode *Load = cast<LoadSDNode>(Op);
1086 SDValue BasePtr = Load->getBasePtr();
1087 EVT PtrVT = BasePtr.getValueType();
1088 EVT MemVT = Load->getMemoryVT();
1090 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1093 EVT LoMemVT, HiMemVT;
1096 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1097 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1098 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1100 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1101 Load->getChain(), BasePtr,
1103 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1104 Load->isInvariant(), Load->getAlignment());
1106 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1107 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1110 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1111 Load->getChain(), HiPtr,
1112 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1113 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1114 Load->isInvariant(), Load->getAlignment());
1117 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1118 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1119 LoLoad.getValue(1), HiLoad.getValue(1))
1122 return DAG.getMergeValues(Ops, SL);
1125 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1126 SelectionDAG &DAG) const {
1127 StoreSDNode *Store = cast<StoreSDNode>(Op);
1128 EVT MemVT = Store->getMemoryVT();
1129 unsigned MemBits = MemVT.getSizeInBits();
1131 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1132 // truncating store into an i32 store.
1133 // XXX: We could also handle optimize other vector bitwidths.
1134 if (!MemVT.isVector() || MemBits > 32) {
1139 SDValue Value = Store->getValue();
1140 EVT VT = Value.getValueType();
1141 EVT ElemVT = VT.getVectorElementType();
1142 SDValue Ptr = Store->getBasePtr();
1143 EVT MemEltVT = MemVT.getVectorElementType();
1144 unsigned MemEltBits = MemEltVT.getSizeInBits();
1145 unsigned MemNumElements = MemVT.getVectorNumElements();
1146 unsigned PackedSize = MemVT.getStoreSizeInBits();
1147 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1149 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1151 SDValue PackedValue;
1152 for (unsigned i = 0; i < MemNumElements; ++i) {
1153 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1154 DAG.getConstant(i, MVT::i32));
1155 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1156 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1158 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1159 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1164 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1168 if (PackedSize < 32) {
1169 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1170 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1171 Store->getMemOperand()->getPointerInfo(),
1173 Store->isNonTemporal(), Store->isVolatile(),
1174 Store->getAlignment());
1177 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1178 Store->getMemOperand()->getPointerInfo(),
1179 Store->isVolatile(), Store->isNonTemporal(),
1180 Store->getAlignment());
1183 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1184 SelectionDAG &DAG) const {
1185 StoreSDNode *Store = cast<StoreSDNode>(Op);
1186 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1187 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1188 EVT PtrVT = Store->getBasePtr().getValueType();
1189 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1192 SmallVector<SDValue, 8> Chains;
1194 unsigned EltSize = MemEltVT.getStoreSize();
1195 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1197 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1198 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1200 DAG.getConstant(i, MVT::i32));
1202 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), PtrVT);
1203 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1205 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1206 SrcValue.getWithOffset(i * EltSize),
1207 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1208 Store->getAlignment());
1209 Chains.push_back(NewStore);
1212 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1215 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1216 SelectionDAG &DAG) const {
1217 StoreSDNode *Store = cast<StoreSDNode>(Op);
1218 SDValue Val = Store->getValue();
1219 EVT VT = Val.getValueType();
1221 // If this is a 2 element vector, we really want to scalarize and not create
1222 // weird 1 element vectors.
1223 if (VT.getVectorNumElements() == 2)
1224 return ScalarizeVectorStore(Op, DAG);
1226 EVT MemVT = Store->getMemoryVT();
1227 SDValue Chain = Store->getChain();
1228 SDValue BasePtr = Store->getBasePtr();
1232 EVT LoMemVT, HiMemVT;
1235 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1236 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1237 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1239 EVT PtrVT = BasePtr.getValueType();
1240 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1241 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1243 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1245 = DAG.getTruncStore(Chain, SL, Lo,
1249 Store->isNonTemporal(),
1250 Store->isVolatile(),
1251 Store->getAlignment());
1253 = DAG.getTruncStore(Chain, SL, Hi,
1255 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1257 Store->isNonTemporal(),
1258 Store->isVolatile(),
1259 Store->getAlignment());
1261 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1265 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1267 LoadSDNode *Load = cast<LoadSDNode>(Op);
1268 ISD::LoadExtType ExtType = Load->getExtensionType();
1269 EVT VT = Op.getValueType();
1270 EVT MemVT = Load->getMemoryVT();
1272 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1273 // We can do the extload to 32-bits, and then need to separately extend to
1276 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1280 Load->getMemOperand());
1283 DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
1284 ExtLoad32.getValue(1)
1287 return DAG.getMergeValues(Ops, DL);
1290 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1291 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1292 // FIXME: Copied from PPC
1293 // First, load into 32 bits, then truncate to 1 bit.
1295 SDValue Chain = Load->getChain();
1296 SDValue BasePtr = Load->getBasePtr();
1297 MachineMemOperand *MMO = Load->getMemOperand();
1299 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1300 BasePtr, MVT::i8, MMO);
1303 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1307 return DAG.getMergeValues(Ops, DL);
1310 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1311 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1312 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1316 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1317 DAG.getConstant(2, MVT::i32));
1318 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1319 Load->getChain(), Ptr,
1320 DAG.getTargetConstant(0, MVT::i32),
1322 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1324 DAG.getConstant(0x3, MVT::i32));
1325 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1326 DAG.getConstant(3, MVT::i32));
1328 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1330 EVT MemEltVT = MemVT.getScalarType();
1331 if (ExtType == ISD::SEXTLOAD) {
1332 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1335 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1339 return DAG.getMergeValues(Ops, DL);
1343 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1347 return DAG.getMergeValues(Ops, DL);
1350 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1352 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1353 if (Result.getNode()) {
1357 StoreSDNode *Store = cast<StoreSDNode>(Op);
1358 SDValue Chain = Store->getChain();
1359 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1360 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1361 Store->getValue().getValueType().isVector()) {
1362 return ScalarizeVectorStore(Op, DAG);
1365 EVT MemVT = Store->getMemoryVT();
1366 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1367 MemVT.bitsLT(MVT::i32)) {
1369 if (Store->getMemoryVT() == MVT::i8) {
1371 } else if (Store->getMemoryVT() == MVT::i16) {
1374 SDValue BasePtr = Store->getBasePtr();
1375 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1376 DAG.getConstant(2, MVT::i32));
1377 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1378 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1380 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1381 DAG.getConstant(0x3, MVT::i32));
1383 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1384 DAG.getConstant(3, MVT::i32));
1386 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1389 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1391 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1392 MaskedValue, ShiftAmt);
1394 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1396 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1397 DAG.getConstant(0xffffffff, MVT::i32));
1398 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1400 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1401 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1402 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1407 // This is a shortcut for integer division because we have fast i32<->f32
1408 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1409 // float is enough to accurately represent up to a 24-bit integer.
1410 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1412 EVT VT = Op.getValueType();
1413 SDValue LHS = Op.getOperand(0);
1414 SDValue RHS = Op.getOperand(1);
1415 MVT IntVT = MVT::i32;
1416 MVT FltVT = MVT::f32;
1418 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1419 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1421 if (VT.isVector()) {
1422 unsigned NElts = VT.getVectorNumElements();
1423 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1424 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1427 unsigned BitSize = VT.getScalarType().getSizeInBits();
1429 SDValue jq = DAG.getConstant(1, IntVT);
1432 // char|short jq = ia ^ ib;
1433 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1435 // jq = jq >> (bitsize - 2)
1436 jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
1439 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
1442 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1445 // int ia = (int)LHS;
1447 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1449 // int ib, (int)RHS;
1451 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1453 // float fa = (float)ia;
1454 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1456 // float fb = (float)ib;
1457 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1459 // float fq = native_divide(fa, fb);
1460 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1461 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1464 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1466 // float fqneg = -fq;
1467 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1469 // float fr = mad(fqneg, fb, fa);
1470 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1471 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1473 // int iq = (int)fq;
1474 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1477 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1480 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1482 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1484 // int cv = fr >= fb;
1485 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1487 // jq = (cv ? jq : 0);
1488 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
1490 // dst = trunc/extend to legal type
1491 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1494 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1496 // Rem needs compensation, it's easier to recompute it
1497 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1498 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1504 return DAG.getMergeValues(Res, DL);
1507 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1508 SelectionDAG &DAG) const {
1510 EVT VT = Op.getValueType();
1512 SDValue Num = Op.getOperand(0);
1513 SDValue Den = Op.getOperand(1);
1515 if (VT == MVT::i32) {
1516 if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
1517 DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
1518 // TODO: We technically could do this for i64, but shouldn't that just be
1519 // handled by something generally reducing 64-bit division on 32-bit
1520 // values to 32-bit?
1521 return LowerDIVREM24(Op, DAG, false);
1525 // RCP = URECIP(Den) = 2^32 / Den + e
1526 // e is rounding error.
1527 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1529 // RCP_LO = mul(RCP, Den) */
1530 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1532 // RCP_HI = mulhu (RCP, Den) */
1533 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1535 // NEG_RCP_LO = -RCP_LO
1536 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1539 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1540 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1543 // Calculate the rounding error from the URECIP instruction
1544 // E = mulhu(ABS_RCP_LO, RCP)
1545 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1547 // RCP_A_E = RCP + E
1548 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1550 // RCP_S_E = RCP - E
1551 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1553 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1554 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1557 // Quotient = mulhu(Tmp0, Num)
1558 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1560 // Num_S_Remainder = Quotient * Den
1561 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1563 // Remainder = Num - Num_S_Remainder
1564 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1566 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1567 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1568 DAG.getConstant(-1, VT),
1569 DAG.getConstant(0, VT),
1571 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1572 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1574 DAG.getConstant(-1, VT),
1575 DAG.getConstant(0, VT),
1577 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1578 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1581 // Calculate Division result:
1583 // Quotient_A_One = Quotient + 1
1584 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1585 DAG.getConstant(1, VT));
1587 // Quotient_S_One = Quotient - 1
1588 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1589 DAG.getConstant(1, VT));
1591 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1592 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1593 Quotient, Quotient_A_One, ISD::SETEQ);
1595 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1596 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1597 Quotient_S_One, Div, ISD::SETEQ);
1599 // Calculate Rem result:
1601 // Remainder_S_Den = Remainder - Den
1602 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1604 // Remainder_A_Den = Remainder + Den
1605 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1607 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1608 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1609 Remainder, Remainder_S_Den, ISD::SETEQ);
1611 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1612 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1613 Remainder_A_Den, Rem, ISD::SETEQ);
1618 return DAG.getMergeValues(Ops, DL);
1621 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1622 SelectionDAG &DAG) const {
1624 EVT VT = Op.getValueType();
1626 SDValue LHS = Op.getOperand(0);
1627 SDValue RHS = Op.getOperand(1);
1629 if (VT == MVT::i32) {
1630 if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
1631 DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
1632 // TODO: We technically could do this for i64, but shouldn't that just be
1633 // handled by something generally reducing 64-bit division on 32-bit
1634 // values to 32-bit?
1635 return LowerDIVREM24(Op, DAG, true);
1639 SDValue Zero = DAG.getConstant(0, VT);
1640 SDValue NegOne = DAG.getConstant(-1, VT);
1642 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1643 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1644 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1645 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1647 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1648 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1650 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1651 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1653 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1654 SDValue Rem = Div.getValue(1);
1656 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1657 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1659 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1660 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1666 return DAG.getMergeValues(Res, DL);
1669 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1670 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1672 EVT VT = Op.getValueType();
1673 SDValue X = Op.getOperand(0);
1674 SDValue Y = Op.getOperand(1);
1676 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1677 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1678 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1680 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1683 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1685 SDValue Src = Op.getOperand(0);
1687 // result = trunc(src)
1688 // if (src > 0.0 && src != result)
1691 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1693 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1694 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1696 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1698 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1699 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1700 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1702 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1703 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1706 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1708 SDValue Src = Op.getOperand(0);
1710 assert(Op.getValueType() == MVT::f64);
1712 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1713 const SDValue One = DAG.getConstant(1, MVT::i32);
1715 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1717 // Extract the upper half, since this is where we will find the sign and
1719 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1721 const unsigned FractBits = 52;
1722 const unsigned ExpBits = 11;
1724 // Extract the exponent.
1725 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1727 DAG.getConstant(FractBits - 32, MVT::i32),
1728 DAG.getConstant(ExpBits, MVT::i32));
1729 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1730 DAG.getConstant(1023, MVT::i32));
1732 // Extract the sign bit.
1733 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1734 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1736 // Extend back to to 64-bits.
1737 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1739 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1741 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1742 const SDValue FractMask
1743 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1745 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1746 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1747 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1749 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1751 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1753 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1754 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1756 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1757 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1759 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1762 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1764 SDValue Src = Op.getOperand(0);
1766 assert(Op.getValueType() == MVT::f64);
1768 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1769 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1770 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1772 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1773 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1775 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1777 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1778 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1780 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1781 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1783 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1786 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1787 // FNEARBYINT and FRINT are the same, except in their handling of FP
1788 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1789 // rint, so just treat them as equivalent.
1790 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1793 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1795 SDValue Src = Op.getOperand(0);
1797 // result = trunc(src);
1798 // if (src < 0.0 && src != result)
1801 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1803 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1804 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1806 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1808 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1809 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1810 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1812 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1813 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1816 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
1817 bool Signed) const {
1819 SDValue Src = Op.getOperand(0);
1821 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1823 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
1824 DAG.getConstant(0, MVT::i32));
1825 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
1826 DAG.getConstant(1, MVT::i32));
1828 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
1831 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
1833 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
1834 DAG.getConstant(32, MVT::i32));
1836 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
1839 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1840 SelectionDAG &DAG) const {
1841 SDValue S0 = Op.getOperand(0);
1842 if (S0.getValueType() != MVT::i64)
1845 EVT DestVT = Op.getValueType();
1846 if (DestVT == MVT::f64)
1847 return LowerINT_TO_FP64(Op, DAG, false);
1849 assert(DestVT == MVT::f32);
1853 // f32 uint_to_fp i64
1854 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1855 DAG.getConstant(0, MVT::i32));
1856 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1857 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1858 DAG.getConstant(1, MVT::i32));
1859 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1860 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1861 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1862 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1865 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
1866 SelectionDAG &DAG) const {
1867 SDValue Src = Op.getOperand(0);
1868 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
1869 return LowerINT_TO_FP64(Op, DAG, true);
1874 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
1875 bool Signed) const {
1878 SDValue Src = Op.getOperand(0);
1880 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1883 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), MVT::f64);
1885 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), MVT::f64);
1887 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
1889 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
1892 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
1894 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
1895 MVT::i32, FloorMul);
1896 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
1898 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
1900 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
1903 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
1904 SelectionDAG &DAG) const {
1905 SDValue Src = Op.getOperand(0);
1907 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
1908 return LowerFP64_TO_INT(Op, DAG, true);
1913 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
1914 SelectionDAG &DAG) const {
1915 SDValue Src = Op.getOperand(0);
1917 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
1918 return LowerFP64_TO_INT(Op, DAG, false);
1923 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1924 SelectionDAG &DAG) const {
1925 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1926 MVT VT = Op.getSimpleValueType();
1927 MVT ScalarVT = VT.getScalarType();
1932 SDValue Src = Op.getOperand(0);
1935 // TODO: Don't scalarize on Evergreen?
1936 unsigned NElts = VT.getVectorNumElements();
1937 SmallVector<SDValue, 8> Args;
1938 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1940 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1941 for (unsigned I = 0; I < NElts; ++I)
1942 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1944 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1947 //===----------------------------------------------------------------------===//
1948 // Custom DAG optimizations
1949 //===----------------------------------------------------------------------===//
1951 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1952 APInt KnownZero, KnownOne;
1953 EVT VT = Op.getValueType();
1954 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1956 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1959 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1960 EVT VT = Op.getValueType();
1962 // In order for this to be a signed 24-bit value, bit 23, must
1964 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1965 // as unsigned 24-bit values.
1966 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1969 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1971 SelectionDAG &DAG = DCI.DAG;
1972 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1973 EVT VT = Op.getValueType();
1975 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1976 APInt KnownZero, KnownOne;
1977 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1978 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1979 DCI.CommitTargetLoweringOpt(TLO);
1982 template <typename IntTy>
1983 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1984 uint32_t Offset, uint32_t Width) {
1985 if (Width + Offset < 32) {
1986 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
1987 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
1988 return DAG.getConstant(Result, MVT::i32);
1991 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1994 static bool usesAllNormalStores(SDNode *LoadVal) {
1995 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
1996 if (!ISD::isNormalStore(*I))
2003 // If we have a copy of an illegal type, replace it with a load / store of an
2004 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2005 // instructions emitted when handling extloads and truncstores. Ideally we could
2006 // recognize the pack / unpack pattern to eliminate it.
2007 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2008 DAGCombinerInfo &DCI) const {
2009 if (!DCI.isBeforeLegalize())
2012 StoreSDNode *SN = cast<StoreSDNode>(N);
2013 SDValue Value = SN->getValue();
2014 EVT VT = Value.getValueType();
2016 if (isTypeLegal(VT) || SN->isVolatile() || !ISD::isNormalLoad(Value.getNode()))
2019 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2020 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2023 EVT MemVT = LoadVal->getMemoryVT();
2026 SelectionDAG &DAG = DCI.DAG;
2027 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2029 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2031 LoadVal->getChain(),
2032 LoadVal->getBasePtr(),
2033 LoadVal->getOffset(),
2035 LoadVal->getMemOperand());
2037 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2038 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2040 return DAG.getStore(SN->getChain(), SL, NewLoad,
2041 SN->getBasePtr(), SN->getMemOperand());
2044 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2045 DAGCombinerInfo &DCI) const {
2046 EVT VT = N->getValueType(0);
2048 if (VT.isVector() || VT.getSizeInBits() > 32)
2051 SelectionDAG &DAG = DCI.DAG;
2054 SDValue N0 = N->getOperand(0);
2055 SDValue N1 = N->getOperand(1);
2058 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2059 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2060 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2061 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2062 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2063 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2064 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2065 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2070 // We need to use sext even for MUL_U24, because MUL_U24 is used
2071 // for signed multiply of 8 and 16-bit types.
2072 return DAG.getSExtOrTrunc(Mul, DL, VT);
2075 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2076 DAGCombinerInfo &DCI) const {
2077 SelectionDAG &DAG = DCI.DAG;
2080 switch(N->getOpcode()) {
2083 return performMulCombine(N, DCI);
2084 case AMDGPUISD::MUL_I24:
2085 case AMDGPUISD::MUL_U24: {
2086 SDValue N0 = N->getOperand(0);
2087 SDValue N1 = N->getOperand(1);
2088 simplifyI24(N0, DCI);
2089 simplifyI24(N1, DCI);
2092 case ISD::SELECT_CC: {
2093 return CombineMinMax(N, DAG);
2095 case AMDGPUISD::BFE_I32:
2096 case AMDGPUISD::BFE_U32: {
2097 assert(!N->getValueType(0).isVector() &&
2098 "Vector handling of BFE not implemented");
2099 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2103 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2105 return DAG.getConstant(0, MVT::i32);
2107 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2111 SDValue BitsFrom = N->getOperand(0);
2112 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2114 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2116 if (OffsetVal == 0) {
2117 // This is already sign / zero extended, so try to fold away extra BFEs.
2118 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2120 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2121 if (OpSignBits >= SignBits)
2124 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2126 // This is a sign_extend_inreg. Replace it to take advantage of existing
2127 // DAG Combines. If not eliminated, we will match back to BFE during
2130 // TODO: The sext_inreg of extended types ends, although we can could
2131 // handle them in a single BFE.
2132 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2133 DAG.getValueType(SmallVT));
2136 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2139 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2141 return constantFoldBFE<int32_t>(DAG,
2142 CVal->getSExtValue(),
2147 return constantFoldBFE<uint32_t>(DAG,
2148 CVal->getZExtValue(),
2153 if ((OffsetVal + WidthVal) >= 32) {
2154 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2155 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2156 BitsFrom, ShiftVal);
2159 if (BitsFrom.hasOneUse()) {
2160 APInt Demanded = APInt::getBitsSet(32,
2162 OffsetVal + WidthVal);
2164 APInt KnownZero, KnownOne;
2165 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2166 !DCI.isBeforeLegalizeOps());
2167 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2168 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2169 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2170 KnownZero, KnownOne, TLO)) {
2171 DCI.CommitTargetLoweringOpt(TLO);
2179 return performStoreCombine(N, DCI);
2184 //===----------------------------------------------------------------------===//
2186 //===----------------------------------------------------------------------===//
2188 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2191 const SmallVectorImpl<ISD::InputArg> &Ins,
2192 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2194 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2195 if (Ins[i].ArgVT == Ins[i].VT) {
2196 OrigIns.push_back(Ins[i]);
2201 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2202 // Vector has been split into scalars.
2203 VT = Ins[i].ArgVT.getVectorElementType();
2204 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2205 Ins[i].ArgVT.getVectorElementType() !=
2206 Ins[i].VT.getVectorElementType()) {
2207 // Vector elements have been promoted
2210 // Vector has been spilt into smaller vectors.
2214 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2215 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2216 OrigIns.push_back(Arg);
2220 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2221 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2222 return CFP->isExactlyValue(1.0);
2224 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2225 return C->isAllOnesValue();
2230 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2231 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2232 return CFP->getValueAPF().isZero();
2234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2235 return C->isNullValue();
2240 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2241 const TargetRegisterClass *RC,
2242 unsigned Reg, EVT VT) const {
2243 MachineFunction &MF = DAG.getMachineFunction();
2244 MachineRegisterInfo &MRI = MF.getRegInfo();
2245 unsigned VirtualRegister;
2246 if (!MRI.isLiveIn(Reg)) {
2247 VirtualRegister = MRI.createVirtualRegister(RC);
2248 MRI.addLiveIn(Reg, VirtualRegister);
2250 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2252 return DAG.getRegister(VirtualRegister, VT);
2255 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2257 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2259 default: return nullptr;
2261 NODE_NAME_CASE(CALL);
2262 NODE_NAME_CASE(UMUL);
2263 NODE_NAME_CASE(RET_FLAG);
2264 NODE_NAME_CASE(BRANCH_COND);
2267 NODE_NAME_CASE(DWORDADDR)
2268 NODE_NAME_CASE(FRACT)
2269 NODE_NAME_CASE(CLAMP)
2271 NODE_NAME_CASE(FMAX)
2272 NODE_NAME_CASE(SMAX)
2273 NODE_NAME_CASE(UMAX)
2274 NODE_NAME_CASE(FMIN)
2275 NODE_NAME_CASE(SMIN)
2276 NODE_NAME_CASE(UMIN)
2277 NODE_NAME_CASE(URECIP)
2278 NODE_NAME_CASE(DIV_SCALE)
2279 NODE_NAME_CASE(DIV_FMAS)
2280 NODE_NAME_CASE(DIV_FIXUP)
2281 NODE_NAME_CASE(TRIG_PREOP)
2284 NODE_NAME_CASE(RSQ_LEGACY)
2285 NODE_NAME_CASE(RSQ_CLAMPED)
2286 NODE_NAME_CASE(LDEXP)
2287 NODE_NAME_CASE(DOT4)
2288 NODE_NAME_CASE(BFE_U32)
2289 NODE_NAME_CASE(BFE_I32)
2292 NODE_NAME_CASE(BREV)
2293 NODE_NAME_CASE(MUL_U24)
2294 NODE_NAME_CASE(MUL_I24)
2295 NODE_NAME_CASE(MAD_U24)
2296 NODE_NAME_CASE(MAD_I24)
2297 NODE_NAME_CASE(EXPORT)
2298 NODE_NAME_CASE(CONST_ADDRESS)
2299 NODE_NAME_CASE(REGISTER_LOAD)
2300 NODE_NAME_CASE(REGISTER_STORE)
2301 NODE_NAME_CASE(LOAD_CONSTANT)
2302 NODE_NAME_CASE(LOAD_INPUT)
2303 NODE_NAME_CASE(SAMPLE)
2304 NODE_NAME_CASE(SAMPLEB)
2305 NODE_NAME_CASE(SAMPLED)
2306 NODE_NAME_CASE(SAMPLEL)
2307 NODE_NAME_CASE(CVT_F32_UBYTE0)
2308 NODE_NAME_CASE(CVT_F32_UBYTE1)
2309 NODE_NAME_CASE(CVT_F32_UBYTE2)
2310 NODE_NAME_CASE(CVT_F32_UBYTE3)
2311 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2312 NODE_NAME_CASE(CONST_DATA_PTR)
2313 NODE_NAME_CASE(STORE_MSKOR)
2314 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2318 static void computeKnownBitsForMinMax(const SDValue Op0,
2322 const SelectionDAG &DAG,
2324 APInt Op0Zero, Op0One;
2325 APInt Op1Zero, Op1One;
2326 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2327 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2329 KnownZero = Op0Zero & Op1Zero;
2330 KnownOne = Op0One & Op1One;
2333 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2337 const SelectionDAG &DAG,
2338 unsigned Depth) const {
2340 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2344 unsigned Opc = Op.getOpcode();
2349 case ISD::INTRINSIC_WO_CHAIN: {
2350 // FIXME: The intrinsic should just use the node.
2351 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2352 case AMDGPUIntrinsic::AMDGPU_imax:
2353 case AMDGPUIntrinsic::AMDGPU_umax:
2354 case AMDGPUIntrinsic::AMDGPU_imin:
2355 case AMDGPUIntrinsic::AMDGPU_umin:
2356 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2357 KnownZero, KnownOne, DAG, Depth);
2365 case AMDGPUISD::SMAX:
2366 case AMDGPUISD::UMAX:
2367 case AMDGPUISD::SMIN:
2368 case AMDGPUISD::UMIN:
2369 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2370 KnownZero, KnownOne, DAG, Depth);
2373 case AMDGPUISD::BFE_I32:
2374 case AMDGPUISD::BFE_U32: {
2375 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2379 unsigned BitWidth = 32;
2380 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2382 if (Opc == AMDGPUISD::BFE_U32)
2383 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2390 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2392 const SelectionDAG &DAG,
2393 unsigned Depth) const {
2394 switch (Op.getOpcode()) {
2395 case AMDGPUISD::BFE_I32: {
2396 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2400 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2401 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2402 if (!Offset || !Offset->isNullValue())
2405 // TODO: Could probably figure something out with non-0 offsets.
2406 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2407 return std::max(SignBits, Op0SignBits);
2410 case AMDGPUISD::BFE_U32: {
2411 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2412 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;