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) :
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::FMINNUM, VT, Expand);
351 setOperationAction(ISD::FMAXNUM, VT, Expand);
352 setOperationAction(ISD::FADD, VT, Expand);
353 setOperationAction(ISD::FCEIL, VT, Expand);
354 setOperationAction(ISD::FCOS, VT, Expand);
355 setOperationAction(ISD::FDIV, VT, Expand);
356 setOperationAction(ISD::FEXP2, VT, Expand);
357 setOperationAction(ISD::FLOG2, VT, Expand);
358 setOperationAction(ISD::FREM, VT, Expand);
359 setOperationAction(ISD::FPOW, VT, Expand);
360 setOperationAction(ISD::FFLOOR, VT, Expand);
361 setOperationAction(ISD::FTRUNC, VT, Expand);
362 setOperationAction(ISD::FMUL, VT, Expand);
363 setOperationAction(ISD::FMA, VT, Expand);
364 setOperationAction(ISD::FRINT, VT, Expand);
365 setOperationAction(ISD::FNEARBYINT, VT, Expand);
366 setOperationAction(ISD::FSQRT, VT, Expand);
367 setOperationAction(ISD::FSIN, VT, Expand);
368 setOperationAction(ISD::FSUB, VT, Expand);
369 setOperationAction(ISD::FNEG, VT, Expand);
370 setOperationAction(ISD::SELECT, VT, Expand);
371 setOperationAction(ISD::VSELECT, VT, Expand);
372 setOperationAction(ISD::SELECT_CC, VT, Expand);
373 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
374 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
377 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
378 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
380 setTargetDAGCombine(ISD::MUL);
381 setTargetDAGCombine(ISD::SELECT);
382 setTargetDAGCombine(ISD::SELECT_CC);
383 setTargetDAGCombine(ISD::STORE);
385 setBooleanContents(ZeroOrNegativeOneBooleanContent);
386 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
388 setSchedulingPreference(Sched::RegPressure);
389 setJumpIsExpensive(true);
391 // SI at least has hardware support for floating point exceptions, but no way
392 // of using or handling them is implemented. They are also optional in OpenCL
394 setHasFloatingPointExceptions(false);
396 setSelectIsExpensive(false);
397 PredictableSelectIsExpensive = false;
399 // There are no integer divide instructions, and these expand to a pretty
400 // large sequence of instructions.
401 setIntDivIsCheap(false);
402 setPow2SDivIsCheap(false);
404 // FIXME: Need to really handle these.
405 MaxStoresPerMemcpy = 4096;
406 MaxStoresPerMemmove = 4096;
407 MaxStoresPerMemset = 4096;
410 //===----------------------------------------------------------------------===//
411 // Target Information
412 //===----------------------------------------------------------------------===//
414 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
418 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
422 // The backend supports 32 and 64 bit floating point immediates.
423 // FIXME: Why are we reporting vectors of FP immediates as legal?
424 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
425 EVT ScalarVT = VT.getScalarType();
426 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
429 // We don't want to shrink f64 / f32 constants.
430 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
431 EVT ScalarVT = VT.getScalarType();
432 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
435 bool AMDGPUTargetLowering::shouldReduceLoadWidth(SDNode *N,
439 unsigned NewSize = NewVT.getStoreSizeInBits();
441 // If we are reducing to a 32-bit load, this is always better.
445 EVT OldVT = N->getValueType(0);
446 unsigned OldSize = OldVT.getStoreSizeInBits();
448 // Don't produce extloads from sub 32-bit types. SI doesn't have scalar
449 // extloads, so doing one requires using a buffer_load. In cases where we
450 // still couldn't use a scalar load, using the wider load shouldn't really
453 // If the old size already had to be an extload, there's no harm in continuing
454 // to reduce the width.
455 return (OldSize < 32);
458 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
460 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
463 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
464 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
466 return ((LScalarSize <= CastScalarSize) ||
467 (CastScalarSize >= 32) ||
471 //===---------------------------------------------------------------------===//
473 //===---------------------------------------------------------------------===//
475 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
476 assert(VT.isFloatingPoint());
477 return VT == MVT::f32 || VT == MVT::f64;
480 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
481 assert(VT.isFloatingPoint());
482 return VT == MVT::f32 || VT == MVT::f64;
485 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
486 // Truncate is just accessing a subregister.
487 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
490 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
491 // Truncate is just accessing a subregister.
492 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
493 (Dest->getPrimitiveSizeInBits() % 32 == 0);
496 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
497 const DataLayout *DL = getDataLayout();
498 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
499 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
501 return SrcSize == 32 && DestSize == 64;
504 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
505 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
506 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
507 // this will enable reducing 64-bit operations the 32-bit, which is always
509 return Src == MVT::i32 && Dest == MVT::i64;
512 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
513 return isZExtFree(Val.getValueType(), VT2);
516 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
517 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
518 // limited number of native 64-bit operations. Shrinking an operation to fit
519 // in a single 32-bit register should always be helpful. As currently used,
520 // this is much less general than the name suggests, and is only used in
521 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
522 // not profitable, and may actually be harmful.
523 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
526 //===---------------------------------------------------------------------===//
527 // TargetLowering Callbacks
528 //===---------------------------------------------------------------------===//
530 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
531 const SmallVectorImpl<ISD::InputArg> &Ins) const {
533 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
536 SDValue AMDGPUTargetLowering::LowerReturn(
538 CallingConv::ID CallConv,
540 const SmallVectorImpl<ISD::OutputArg> &Outs,
541 const SmallVectorImpl<SDValue> &OutVals,
542 SDLoc DL, SelectionDAG &DAG) const {
543 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
546 //===---------------------------------------------------------------------===//
547 // Target specific lowering
548 //===---------------------------------------------------------------------===//
550 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
551 SmallVectorImpl<SDValue> &InVals) const {
552 SDValue Callee = CLI.Callee;
553 SelectionDAG &DAG = CLI.DAG;
555 const Function &Fn = *DAG.getMachineFunction().getFunction();
557 StringRef FuncName("<unknown>");
559 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
560 FuncName = G->getSymbol();
561 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
562 FuncName = G->getGlobal()->getName();
564 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
565 DAG.getContext()->diagnose(NoCalls);
569 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
570 SelectionDAG &DAG) const {
571 switch (Op.getOpcode()) {
573 Op.getNode()->dump();
574 llvm_unreachable("Custom lowering code for this"
575 "instruction is not implemented yet!");
577 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
578 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
579 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
580 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
581 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
582 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
583 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
584 case ISD::FREM: return LowerFREM(Op, DAG);
585 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
586 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
587 case ISD::FRINT: return LowerFRINT(Op, DAG);
588 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
589 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
590 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
591 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
592 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
593 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
598 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
599 SmallVectorImpl<SDValue> &Results,
600 SelectionDAG &DAG) const {
601 switch (N->getOpcode()) {
602 case ISD::SIGN_EXTEND_INREG:
603 // Different parts of legalization seem to interpret which type of
604 // sign_extend_inreg is the one to check for custom lowering. The extended
605 // from type is what really matters, but some places check for custom
606 // lowering of the result type. This results in trying to use
607 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
608 // nothing here and let the illegal result integer be handled normally.
611 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
615 Results.push_back(SDValue(Node, 0));
616 Results.push_back(SDValue(Node, 1));
617 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
619 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
623 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
624 if (Lowered.getNode())
625 Results.push_back(Lowered);
633 // FIXME: This implements accesses to initialized globals in the constant
634 // address space by copying them to private and accessing that. It does not
635 // properly handle illegal types or vectors. The private vector loads are not
636 // scalarized, and the illegal scalars hit an assertion. This technique will not
637 // work well with large initializers, and this should eventually be
638 // removed. Initialized globals should be placed into a data section that the
639 // runtime will load into a buffer before the kernel is executed. Uses of the
640 // global need to be replaced with a pointer loaded from an implicit kernel
641 // argument into this buffer holding the copy of the data, which will remove the
642 // need for any of this.
643 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
644 const GlobalValue *GV,
645 const SDValue &InitPtr,
647 SelectionDAG &DAG) const {
648 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
650 Type *InitTy = Init->getType();
652 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
653 EVT VT = EVT::getEVT(InitTy);
654 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
655 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
656 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
657 TD->getPrefTypeAlignment(InitTy));
660 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
661 EVT VT = EVT::getEVT(CFP->getType());
662 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
663 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
664 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
665 TD->getPrefTypeAlignment(CFP->getType()));
668 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
669 const StructLayout *SL = TD->getStructLayout(ST);
671 EVT PtrVT = InitPtr.getValueType();
672 SmallVector<SDValue, 8> Chains;
674 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
675 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
676 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
678 Constant *Elt = Init->getAggregateElement(I);
679 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
682 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
685 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
686 EVT PtrVT = InitPtr.getValueType();
688 unsigned NumElements;
689 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
690 NumElements = AT->getNumElements();
691 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
692 NumElements = VT->getNumElements();
694 llvm_unreachable("Unexpected type");
696 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
697 SmallVector<SDValue, 8> Chains;
698 for (unsigned i = 0; i < NumElements; ++i) {
699 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
700 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
702 Constant *Elt = Init->getAggregateElement(i);
703 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
706 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
709 if (isa<UndefValue>(Init)) {
710 EVT VT = EVT::getEVT(InitTy);
711 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
712 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
713 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
714 TD->getPrefTypeAlignment(InitTy));
718 llvm_unreachable("Unhandled constant initializer");
721 static bool hasDefinedInitializer(const GlobalValue *GV) {
722 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
723 if (!GVar || !GVar->hasInitializer())
726 if (isa<UndefValue>(GVar->getInitializer()))
732 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
734 SelectionDAG &DAG) const {
736 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
737 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
738 const GlobalValue *GV = G->getGlobal();
740 switch (G->getAddressSpace()) {
741 case AMDGPUAS::LOCAL_ADDRESS: {
742 // XXX: What does the value of G->getOffset() mean?
743 assert(G->getOffset() == 0 &&
744 "Do not know what to do with an non-zero offset");
746 // TODO: We could emit code to handle the initialization somewhere.
747 if (hasDefinedInitializer(GV))
751 if (MFI->LocalMemoryObjects.count(GV) == 0) {
752 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
753 Offset = MFI->LDSSize;
754 MFI->LocalMemoryObjects[GV] = Offset;
755 // XXX: Account for alignment?
756 MFI->LDSSize += Size;
758 Offset = MFI->LocalMemoryObjects[GV];
761 return DAG.getConstant(Offset, getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
763 case AMDGPUAS::CONSTANT_ADDRESS: {
764 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
765 Type *EltType = GV->getType()->getElementType();
766 unsigned Size = TD->getTypeAllocSize(EltType);
767 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
769 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
770 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
772 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
773 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
775 const GlobalVariable *Var = cast<GlobalVariable>(GV);
776 if (!Var->hasInitializer()) {
777 // This has no use, but bugpoint will hit it.
778 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
781 const Constant *Init = Var->getInitializer();
782 SmallVector<SDNode*, 8> WorkList;
784 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
785 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
786 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
788 WorkList.push_back(*I);
790 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
791 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
792 E = WorkList.end(); I != E; ++I) {
793 SmallVector<SDValue, 8> Ops;
794 Ops.push_back(Chain);
795 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
796 Ops.push_back((*I)->getOperand(i));
798 DAG.UpdateNodeOperands(*I, Ops);
800 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
804 const Function &Fn = *DAG.getMachineFunction().getFunction();
805 DiagnosticInfoUnsupported BadInit(Fn,
806 "initializer for address space");
807 DAG.getContext()->diagnose(BadInit);
811 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
812 SelectionDAG &DAG) const {
813 SmallVector<SDValue, 8> Args;
814 SDValue A = Op.getOperand(0);
815 SDValue B = Op.getOperand(1);
817 DAG.ExtractVectorElements(A, Args);
818 DAG.ExtractVectorElements(B, Args);
820 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
823 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
824 SelectionDAG &DAG) const {
826 SmallVector<SDValue, 8> Args;
827 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
828 EVT VT = Op.getValueType();
829 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
830 VT.getVectorNumElements());
832 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
835 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
836 SelectionDAG &DAG) const {
838 MachineFunction &MF = DAG.getMachineFunction();
839 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
840 getTargetMachine().getSubtargetImpl()->getFrameLowering());
842 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
844 unsigned FrameIndex = FIN->getIndex();
845 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
846 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
850 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
851 SelectionDAG &DAG) const {
852 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
854 EVT VT = Op.getValueType();
856 switch (IntrinsicID) {
858 case AMDGPUIntrinsic::AMDGPU_abs:
859 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
860 return LowerIntrinsicIABS(Op, DAG);
861 case AMDGPUIntrinsic::AMDGPU_lrp:
862 return LowerIntrinsicLRP(Op, DAG);
863 case AMDGPUIntrinsic::AMDGPU_fract:
864 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
865 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
867 case AMDGPUIntrinsic::AMDGPU_clamp:
868 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
869 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
870 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
872 case Intrinsic::AMDGPU_div_scale: {
873 // 3rd parameter required to be a constant.
874 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
876 return DAG.getUNDEF(VT);
878 // Translate to the operands expected by the machine instruction. The
879 // first parameter must be the same as the first instruction.
880 SDValue Numerator = Op.getOperand(1);
881 SDValue Denominator = Op.getOperand(2);
883 // Note this order is opposite of the machine instruction's operations,
884 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
885 // intrinsic has the numerator as the first operand to match a normal
886 // division operation.
888 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
890 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
891 Denominator, Numerator);
894 case Intrinsic::AMDGPU_div_fmas:
895 // FIXME: Dropping bool parameter. Work is needed to support the implicit
897 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
898 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
900 case Intrinsic::AMDGPU_div_fixup:
901 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
902 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
904 case Intrinsic::AMDGPU_trig_preop:
905 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
906 Op.getOperand(1), Op.getOperand(2));
908 case Intrinsic::AMDGPU_rcp:
909 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
911 case Intrinsic::AMDGPU_rsq:
912 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
914 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
915 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
917 case Intrinsic::AMDGPU_rsq_clamped:
918 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
919 Type *Type = VT.getTypeForEVT(*DAG.getContext());
920 APFloat Max = APFloat::getLargest(Type->getFltSemantics());
921 APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
923 SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
924 SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
925 DAG.getConstantFP(Max, VT));
926 return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
927 DAG.getConstantFP(Min, VT));
929 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
932 case Intrinsic::AMDGPU_ldexp:
933 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
936 case AMDGPUIntrinsic::AMDGPU_imax:
937 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
939 case AMDGPUIntrinsic::AMDGPU_umax:
940 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
942 case AMDGPUIntrinsic::AMDGPU_imin:
943 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
945 case AMDGPUIntrinsic::AMDGPU_umin:
946 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
949 case AMDGPUIntrinsic::AMDGPU_umul24:
950 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
951 Op.getOperand(1), Op.getOperand(2));
953 case AMDGPUIntrinsic::AMDGPU_imul24:
954 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
955 Op.getOperand(1), Op.getOperand(2));
957 case AMDGPUIntrinsic::AMDGPU_umad24:
958 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
959 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
961 case AMDGPUIntrinsic::AMDGPU_imad24:
962 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
963 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
965 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
966 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
968 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
969 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
971 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
972 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
974 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
975 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
977 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
978 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
983 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
984 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
989 case AMDGPUIntrinsic::AMDGPU_bfi:
990 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
995 case AMDGPUIntrinsic::AMDGPU_bfm:
996 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
1000 case AMDGPUIntrinsic::AMDGPU_brev:
1001 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
1003 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
1004 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
1006 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
1007 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
1008 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
1009 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
1013 ///IABS(a) = SMAX(sub(0, a), a)
1014 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
1015 SelectionDAG &DAG) const {
1017 EVT VT = Op.getValueType();
1018 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1021 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
1024 /// Linear Interpolation
1025 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
1026 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
1027 SelectionDAG &DAG) const {
1029 EVT VT = Op.getValueType();
1030 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
1031 DAG.getConstantFP(1.0f, MVT::f32),
1033 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1035 return DAG.getNode(ISD::FADD, DL, VT,
1036 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1040 /// \brief Generate Min/Max node
1041 SDValue AMDGPUTargetLowering::CombineFMinMax(SDLoc DL,
1048 SelectionDAG &DAG) const {
1049 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1052 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1061 case ISD::SETFALSE2:
1073 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1076 // We need to permute the operands to get the correct NaN behavior. The
1077 // selected operand is the second one based on the failing compare with NaN,
1078 // so permute it based on the compare type the hardware uses.
1080 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1081 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1089 if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
1093 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1094 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1096 case ISD::SETCC_INVALID:
1097 llvm_unreachable("Invalid setcc condcode!");
1102 /// \brief Generate Min/Max node
1103 SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1110 SelectionDAG &DAG) const {
1111 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1114 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1118 unsigned Opc = (LHS == True) ? AMDGPUISD::UMIN : AMDGPUISD::UMAX;
1119 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1123 unsigned Opc = (LHS == True) ? AMDGPUISD::SMIN : AMDGPUISD::SMAX;
1124 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1128 unsigned Opc = (LHS == True) ? AMDGPUISD::SMAX : AMDGPUISD::SMIN;
1129 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1133 unsigned Opc = (LHS == True) ? AMDGPUISD::UMAX : AMDGPUISD::UMIN;
1134 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1141 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1142 SelectionDAG &DAG) const {
1143 LoadSDNode *Load = cast<LoadSDNode>(Op);
1144 EVT MemVT = Load->getMemoryVT();
1145 EVT MemEltVT = MemVT.getVectorElementType();
1147 EVT LoadVT = Op.getValueType();
1148 EVT EltVT = LoadVT.getVectorElementType();
1149 EVT PtrVT = Load->getBasePtr().getValueType();
1151 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1152 SmallVector<SDValue, 8> Loads;
1153 SmallVector<SDValue, 8> Chains;
1156 unsigned MemEltSize = MemEltVT.getStoreSize();
1157 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1159 for (unsigned i = 0; i < NumElts; ++i) {
1160 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1161 DAG.getConstant(i * MemEltSize, PtrVT));
1164 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1165 Load->getChain(), Ptr,
1166 SrcValue.getWithOffset(i * MemEltSize),
1167 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1168 Load->isInvariant(), Load->getAlignment());
1169 Loads.push_back(NewLoad.getValue(0));
1170 Chains.push_back(NewLoad.getValue(1));
1174 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1175 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1178 return DAG.getMergeValues(Ops, SL);
1181 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1182 SelectionDAG &DAG) const {
1183 EVT VT = Op.getValueType();
1185 // If this is a 2 element vector, we really want to scalarize and not create
1186 // weird 1 element vectors.
1187 if (VT.getVectorNumElements() == 2)
1188 return ScalarizeVectorLoad(Op, DAG);
1190 LoadSDNode *Load = cast<LoadSDNode>(Op);
1191 SDValue BasePtr = Load->getBasePtr();
1192 EVT PtrVT = BasePtr.getValueType();
1193 EVT MemVT = Load->getMemoryVT();
1195 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1198 EVT LoMemVT, HiMemVT;
1201 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1202 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1203 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1205 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1206 Load->getChain(), BasePtr,
1208 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1209 Load->isInvariant(), Load->getAlignment());
1211 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1212 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1215 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1216 Load->getChain(), HiPtr,
1217 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1218 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1219 Load->isInvariant(), Load->getAlignment());
1222 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1223 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1224 LoLoad.getValue(1), HiLoad.getValue(1))
1227 return DAG.getMergeValues(Ops, SL);
1230 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1231 SelectionDAG &DAG) const {
1232 StoreSDNode *Store = cast<StoreSDNode>(Op);
1233 EVT MemVT = Store->getMemoryVT();
1234 unsigned MemBits = MemVT.getSizeInBits();
1236 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1237 // truncating store into an i32 store.
1238 // XXX: We could also handle optimize other vector bitwidths.
1239 if (!MemVT.isVector() || MemBits > 32) {
1244 SDValue Value = Store->getValue();
1245 EVT VT = Value.getValueType();
1246 EVT ElemVT = VT.getVectorElementType();
1247 SDValue Ptr = Store->getBasePtr();
1248 EVT MemEltVT = MemVT.getVectorElementType();
1249 unsigned MemEltBits = MemEltVT.getSizeInBits();
1250 unsigned MemNumElements = MemVT.getVectorNumElements();
1251 unsigned PackedSize = MemVT.getStoreSizeInBits();
1252 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1254 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1256 SDValue PackedValue;
1257 for (unsigned i = 0; i < MemNumElements; ++i) {
1258 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1259 DAG.getConstant(i, MVT::i32));
1260 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1261 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1263 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1264 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1269 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1273 if (PackedSize < 32) {
1274 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1275 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1276 Store->getMemOperand()->getPointerInfo(),
1278 Store->isNonTemporal(), Store->isVolatile(),
1279 Store->getAlignment());
1282 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1283 Store->getMemOperand()->getPointerInfo(),
1284 Store->isVolatile(), Store->isNonTemporal(),
1285 Store->getAlignment());
1288 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1289 SelectionDAG &DAG) const {
1290 StoreSDNode *Store = cast<StoreSDNode>(Op);
1291 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1292 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1293 EVT PtrVT = Store->getBasePtr().getValueType();
1294 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1297 SmallVector<SDValue, 8> Chains;
1299 unsigned EltSize = MemEltVT.getStoreSize();
1300 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1302 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1303 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1305 DAG.getConstant(i, MVT::i32));
1307 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), PtrVT);
1308 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1310 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1311 SrcValue.getWithOffset(i * EltSize),
1312 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1313 Store->getAlignment());
1314 Chains.push_back(NewStore);
1317 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1320 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1321 SelectionDAG &DAG) const {
1322 StoreSDNode *Store = cast<StoreSDNode>(Op);
1323 SDValue Val = Store->getValue();
1324 EVT VT = Val.getValueType();
1326 // If this is a 2 element vector, we really want to scalarize and not create
1327 // weird 1 element vectors.
1328 if (VT.getVectorNumElements() == 2)
1329 return ScalarizeVectorStore(Op, DAG);
1331 EVT MemVT = Store->getMemoryVT();
1332 SDValue Chain = Store->getChain();
1333 SDValue BasePtr = Store->getBasePtr();
1337 EVT LoMemVT, HiMemVT;
1340 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1341 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1342 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1344 EVT PtrVT = BasePtr.getValueType();
1345 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1346 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1348 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1350 = DAG.getTruncStore(Chain, SL, Lo,
1354 Store->isNonTemporal(),
1355 Store->isVolatile(),
1356 Store->getAlignment());
1358 = DAG.getTruncStore(Chain, SL, Hi,
1360 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1362 Store->isNonTemporal(),
1363 Store->isVolatile(),
1364 Store->getAlignment());
1366 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1370 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1372 LoadSDNode *Load = cast<LoadSDNode>(Op);
1373 ISD::LoadExtType ExtType = Load->getExtensionType();
1374 EVT VT = Op.getValueType();
1375 EVT MemVT = Load->getMemoryVT();
1377 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1378 // We can do the extload to 32-bits, and then need to separately extend to
1381 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1385 Load->getMemOperand());
1388 DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
1389 ExtLoad32.getValue(1)
1392 return DAG.getMergeValues(Ops, DL);
1395 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1396 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1397 // FIXME: Copied from PPC
1398 // First, load into 32 bits, then truncate to 1 bit.
1400 SDValue Chain = Load->getChain();
1401 SDValue BasePtr = Load->getBasePtr();
1402 MachineMemOperand *MMO = Load->getMemOperand();
1404 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1405 BasePtr, MVT::i8, MMO);
1408 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1412 return DAG.getMergeValues(Ops, DL);
1415 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1416 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1417 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1421 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1422 DAG.getConstant(2, MVT::i32));
1423 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1424 Load->getChain(), Ptr,
1425 DAG.getTargetConstant(0, MVT::i32),
1427 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1429 DAG.getConstant(0x3, MVT::i32));
1430 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1431 DAG.getConstant(3, MVT::i32));
1433 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1435 EVT MemEltVT = MemVT.getScalarType();
1436 if (ExtType == ISD::SEXTLOAD) {
1437 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1440 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1444 return DAG.getMergeValues(Ops, DL);
1448 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1452 return DAG.getMergeValues(Ops, DL);
1455 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1457 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1458 if (Result.getNode()) {
1462 StoreSDNode *Store = cast<StoreSDNode>(Op);
1463 SDValue Chain = Store->getChain();
1464 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1465 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1466 Store->getValue().getValueType().isVector()) {
1467 return ScalarizeVectorStore(Op, DAG);
1470 EVT MemVT = Store->getMemoryVT();
1471 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1472 MemVT.bitsLT(MVT::i32)) {
1474 if (Store->getMemoryVT() == MVT::i8) {
1476 } else if (Store->getMemoryVT() == MVT::i16) {
1479 SDValue BasePtr = Store->getBasePtr();
1480 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1481 DAG.getConstant(2, MVT::i32));
1482 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1483 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1485 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1486 DAG.getConstant(0x3, MVT::i32));
1488 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1489 DAG.getConstant(3, MVT::i32));
1491 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1494 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1496 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1497 MaskedValue, ShiftAmt);
1499 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1501 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1502 DAG.getConstant(0xffffffff, MVT::i32));
1503 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1505 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1506 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1507 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1512 // This is a shortcut for integer division because we have fast i32<->f32
1513 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1514 // float is enough to accurately represent up to a 24-bit integer.
1515 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1517 EVT VT = Op.getValueType();
1518 SDValue LHS = Op.getOperand(0);
1519 SDValue RHS = Op.getOperand(1);
1520 MVT IntVT = MVT::i32;
1521 MVT FltVT = MVT::f32;
1523 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1524 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1526 if (VT.isVector()) {
1527 unsigned NElts = VT.getVectorNumElements();
1528 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1529 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1532 unsigned BitSize = VT.getScalarType().getSizeInBits();
1534 SDValue jq = DAG.getConstant(1, IntVT);
1537 // char|short jq = ia ^ ib;
1538 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1540 // jq = jq >> (bitsize - 2)
1541 jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
1544 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
1547 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1550 // int ia = (int)LHS;
1552 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1554 // int ib, (int)RHS;
1556 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1558 // float fa = (float)ia;
1559 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1561 // float fb = (float)ib;
1562 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1564 // float fq = native_divide(fa, fb);
1565 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1566 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1569 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1571 // float fqneg = -fq;
1572 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1574 // float fr = mad(fqneg, fb, fa);
1575 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1576 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1578 // int iq = (int)fq;
1579 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1582 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1585 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1587 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1589 // int cv = fr >= fb;
1590 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1592 // jq = (cv ? jq : 0);
1593 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
1595 // dst = trunc/extend to legal type
1596 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1599 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1601 // Rem needs compensation, it's easier to recompute it
1602 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1603 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1609 return DAG.getMergeValues(Res, DL);
1612 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1614 SmallVectorImpl<SDValue> &Results) const {
1615 assert(Op.getValueType() == MVT::i64);
1618 EVT VT = Op.getValueType();
1619 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1621 SDValue one = DAG.getConstant(1, HalfVT);
1622 SDValue zero = DAG.getConstant(0, HalfVT);
1625 SDValue LHS = Op.getOperand(0);
1626 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1627 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1629 SDValue RHS = Op.getOperand(1);
1630 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1631 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1633 // Get Speculative values
1634 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1635 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1637 SDValue REM_Hi = zero;
1638 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1640 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1641 SDValue DIV_Lo = zero;
1643 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1645 for (unsigned i = 0; i < halfBitWidth; ++i) {
1646 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
1647 // Get Value of high bit
1649 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
1650 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
1652 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1653 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1656 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
1657 DAG.getConstant(halfBitWidth - 1, HalfVT));
1658 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
1659 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
1661 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
1662 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
1665 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1667 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
1668 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1670 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1674 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1676 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1677 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
1678 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
1681 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1682 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1683 Results.push_back(DIV);
1684 Results.push_back(REM);
1687 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1688 SelectionDAG &DAG) const {
1690 EVT VT = Op.getValueType();
1692 if (VT == MVT::i64) {
1693 SmallVector<SDValue, 2> Results;
1694 LowerUDIVREM64(Op, DAG, Results);
1695 return DAG.getMergeValues(Results, DL);
1698 SDValue Num = Op.getOperand(0);
1699 SDValue Den = Op.getOperand(1);
1701 if (VT == MVT::i32) {
1702 if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
1703 DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
1704 // TODO: We technically could do this for i64, but shouldn't that just be
1705 // handled by something generally reducing 64-bit division on 32-bit
1706 // values to 32-bit?
1707 return LowerDIVREM24(Op, DAG, false);
1711 // RCP = URECIP(Den) = 2^32 / Den + e
1712 // e is rounding error.
1713 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1715 // RCP_LO = mul(RCP, Den) */
1716 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1718 // RCP_HI = mulhu (RCP, Den) */
1719 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1721 // NEG_RCP_LO = -RCP_LO
1722 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1725 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1726 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1729 // Calculate the rounding error from the URECIP instruction
1730 // E = mulhu(ABS_RCP_LO, RCP)
1731 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1733 // RCP_A_E = RCP + E
1734 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1736 // RCP_S_E = RCP - E
1737 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1739 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1740 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1743 // Quotient = mulhu(Tmp0, Num)
1744 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1746 // Num_S_Remainder = Quotient * Den
1747 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1749 // Remainder = Num - Num_S_Remainder
1750 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1752 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1753 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1754 DAG.getConstant(-1, VT),
1755 DAG.getConstant(0, VT),
1757 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1758 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1760 DAG.getConstant(-1, VT),
1761 DAG.getConstant(0, VT),
1763 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1764 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1767 // Calculate Division result:
1769 // Quotient_A_One = Quotient + 1
1770 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1771 DAG.getConstant(1, VT));
1773 // Quotient_S_One = Quotient - 1
1774 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1775 DAG.getConstant(1, VT));
1777 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1778 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1779 Quotient, Quotient_A_One, ISD::SETEQ);
1781 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1782 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1783 Quotient_S_One, Div, ISD::SETEQ);
1785 // Calculate Rem result:
1787 // Remainder_S_Den = Remainder - Den
1788 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1790 // Remainder_A_Den = Remainder + Den
1791 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1793 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1794 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1795 Remainder, Remainder_S_Den, ISD::SETEQ);
1797 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1798 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1799 Remainder_A_Den, Rem, ISD::SETEQ);
1804 return DAG.getMergeValues(Ops, DL);
1807 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1808 SelectionDAG &DAG) const {
1810 EVT VT = Op.getValueType();
1812 SDValue LHS = Op.getOperand(0);
1813 SDValue RHS = Op.getOperand(1);
1815 if (VT == MVT::i32) {
1816 if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
1817 DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
1818 // TODO: We technically could do this for i64, but shouldn't that just be
1819 // handled by something generally reducing 64-bit division on 32-bit
1820 // values to 32-bit?
1821 return LowerDIVREM24(Op, DAG, true);
1825 SDValue Zero = DAG.getConstant(0, VT);
1826 SDValue NegOne = DAG.getConstant(-1, VT);
1828 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1829 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1830 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1831 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1833 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1834 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1836 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1837 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1839 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1840 SDValue Rem = Div.getValue(1);
1842 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1843 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1845 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1846 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1852 return DAG.getMergeValues(Res, DL);
1855 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1856 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1858 EVT VT = Op.getValueType();
1859 SDValue X = Op.getOperand(0);
1860 SDValue Y = Op.getOperand(1);
1862 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1863 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1864 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1866 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1869 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1871 SDValue Src = Op.getOperand(0);
1873 // result = trunc(src)
1874 // if (src > 0.0 && src != result)
1877 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1879 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1880 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1882 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1884 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1885 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1886 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1888 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1889 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1892 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1894 SDValue Src = Op.getOperand(0);
1896 assert(Op.getValueType() == MVT::f64);
1898 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1899 const SDValue One = DAG.getConstant(1, MVT::i32);
1901 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1903 // Extract the upper half, since this is where we will find the sign and
1905 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1907 const unsigned FractBits = 52;
1908 const unsigned ExpBits = 11;
1910 // Extract the exponent.
1911 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1913 DAG.getConstant(FractBits - 32, MVT::i32),
1914 DAG.getConstant(ExpBits, MVT::i32));
1915 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1916 DAG.getConstant(1023, MVT::i32));
1918 // Extract the sign bit.
1919 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1920 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1922 // Extend back to to 64-bits.
1923 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1925 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1927 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1928 const SDValue FractMask
1929 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1931 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1932 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1933 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1935 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1937 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1939 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1940 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1942 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1943 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1945 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1948 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1950 SDValue Src = Op.getOperand(0);
1952 assert(Op.getValueType() == MVT::f64);
1954 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1955 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1956 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1958 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1959 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1961 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1963 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1964 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1966 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1967 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1969 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1972 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1973 // FNEARBYINT and FRINT are the same, except in their handling of FP
1974 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1975 // rint, so just treat them as equivalent.
1976 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1979 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1981 SDValue Src = Op.getOperand(0);
1983 // result = trunc(src);
1984 // if (src < 0.0 && src != result)
1987 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1989 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1990 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1992 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1994 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1995 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1996 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1998 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1999 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
2002 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
2003 bool Signed) const {
2005 SDValue Src = Op.getOperand(0);
2007 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
2009 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2010 DAG.getConstant(0, MVT::i32));
2011 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
2012 DAG.getConstant(1, MVT::i32));
2014 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
2017 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
2019 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
2020 DAG.getConstant(32, MVT::i32));
2022 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
2025 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
2026 SelectionDAG &DAG) const {
2027 SDValue S0 = Op.getOperand(0);
2028 if (S0.getValueType() != MVT::i64)
2031 EVT DestVT = Op.getValueType();
2032 if (DestVT == MVT::f64)
2033 return LowerINT_TO_FP64(Op, DAG, false);
2035 assert(DestVT == MVT::f32);
2039 // f32 uint_to_fp i64
2040 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2041 DAG.getConstant(0, MVT::i32));
2042 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2043 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2044 DAG.getConstant(1, MVT::i32));
2045 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2046 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2047 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
2048 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2051 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2052 SelectionDAG &DAG) const {
2053 SDValue Src = Op.getOperand(0);
2054 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2055 return LowerINT_TO_FP64(Op, DAG, true);
2060 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2061 bool Signed) const {
2064 SDValue Src = Op.getOperand(0);
2066 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2069 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), MVT::f64);
2071 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), MVT::f64);
2073 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2075 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2078 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2080 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2081 MVT::i32, FloorMul);
2082 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2084 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2086 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2089 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2090 SelectionDAG &DAG) const {
2091 SDValue Src = Op.getOperand(0);
2093 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2094 return LowerFP64_TO_INT(Op, DAG, true);
2099 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2100 SelectionDAG &DAG) const {
2101 SDValue Src = Op.getOperand(0);
2103 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2104 return LowerFP64_TO_INT(Op, DAG, false);
2109 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2110 SelectionDAG &DAG) const {
2111 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2112 MVT VT = Op.getSimpleValueType();
2113 MVT ScalarVT = VT.getScalarType();
2118 SDValue Src = Op.getOperand(0);
2121 // TODO: Don't scalarize on Evergreen?
2122 unsigned NElts = VT.getVectorNumElements();
2123 SmallVector<SDValue, 8> Args;
2124 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2126 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2127 for (unsigned I = 0; I < NElts; ++I)
2128 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2130 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2133 //===----------------------------------------------------------------------===//
2134 // Custom DAG optimizations
2135 //===----------------------------------------------------------------------===//
2137 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2138 APInt KnownZero, KnownOne;
2139 EVT VT = Op.getValueType();
2140 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2142 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2145 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2146 EVT VT = Op.getValueType();
2148 // In order for this to be a signed 24-bit value, bit 23, must
2150 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2151 // as unsigned 24-bit values.
2152 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2155 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2157 SelectionDAG &DAG = DCI.DAG;
2158 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2159 EVT VT = Op.getValueType();
2161 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2162 APInt KnownZero, KnownOne;
2163 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2164 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2165 DCI.CommitTargetLoweringOpt(TLO);
2168 template <typename IntTy>
2169 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2170 uint32_t Offset, uint32_t Width) {
2171 if (Width + Offset < 32) {
2172 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2173 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2174 return DAG.getConstant(Result, MVT::i32);
2177 return DAG.getConstant(Src0 >> Offset, MVT::i32);
2180 static bool usesAllNormalStores(SDNode *LoadVal) {
2181 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2182 if (!ISD::isNormalStore(*I))
2189 // If we have a copy of an illegal type, replace it with a load / store of an
2190 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2191 // instructions emitted when handling extloads and truncstores. Ideally we could
2192 // recognize the pack / unpack pattern to eliminate it.
2193 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2194 DAGCombinerInfo &DCI) const {
2195 if (!DCI.isBeforeLegalize())
2198 StoreSDNode *SN = cast<StoreSDNode>(N);
2199 SDValue Value = SN->getValue();
2200 EVT VT = Value.getValueType();
2202 if (isTypeLegal(VT) || SN->isVolatile() ||
2203 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2206 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2207 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2210 EVT MemVT = LoadVal->getMemoryVT();
2213 SelectionDAG &DAG = DCI.DAG;
2214 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2216 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2218 LoadVal->getChain(),
2219 LoadVal->getBasePtr(),
2220 LoadVal->getOffset(),
2222 LoadVal->getMemOperand());
2224 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2225 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2227 return DAG.getStore(SN->getChain(), SL, NewLoad,
2228 SN->getBasePtr(), SN->getMemOperand());
2231 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2232 DAGCombinerInfo &DCI) const {
2233 EVT VT = N->getValueType(0);
2235 if (VT.isVector() || VT.getSizeInBits() > 32)
2238 SelectionDAG &DAG = DCI.DAG;
2241 SDValue N0 = N->getOperand(0);
2242 SDValue N1 = N->getOperand(1);
2245 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2246 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2247 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2248 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2249 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2250 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2251 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2252 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2257 // We need to use sext even for MUL_U24, because MUL_U24 is used
2258 // for signed multiply of 8 and 16-bit types.
2259 return DAG.getSExtOrTrunc(Mul, DL, VT);
2262 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2263 DAGCombinerInfo &DCI) const {
2264 SelectionDAG &DAG = DCI.DAG;
2267 switch(N->getOpcode()) {
2270 return performMulCombine(N, DCI);
2271 case AMDGPUISD::MUL_I24:
2272 case AMDGPUISD::MUL_U24: {
2273 SDValue N0 = N->getOperand(0);
2274 SDValue N1 = N->getOperand(1);
2275 simplifyI24(N0, DCI);
2276 simplifyI24(N1, DCI);
2279 case ISD::SELECT_CC: {
2281 EVT VT = N->getValueType(0);
2283 if (VT == MVT::f32 ||
2285 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
2286 SDValue LHS = N->getOperand(0);
2287 SDValue RHS = N->getOperand(1);
2288 SDValue True = N->getOperand(2);
2289 SDValue False = N->getOperand(3);
2290 SDValue CC = N->getOperand(4);
2292 return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2298 SDValue Cond = N->getOperand(0);
2299 if (Cond.getOpcode() == ISD::SETCC) {
2301 EVT VT = N->getValueType(0);
2302 SDValue LHS = Cond.getOperand(0);
2303 SDValue RHS = Cond.getOperand(1);
2304 SDValue CC = Cond.getOperand(2);
2306 SDValue True = N->getOperand(1);
2307 SDValue False = N->getOperand(2);
2309 if (VT == MVT::f32 ||
2311 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
2312 return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2315 // TODO: Implement min / max Evergreen instructions.
2316 if (VT == MVT::i32 &&
2317 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2318 return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2324 case AMDGPUISD::BFE_I32:
2325 case AMDGPUISD::BFE_U32: {
2326 assert(!N->getValueType(0).isVector() &&
2327 "Vector handling of BFE not implemented");
2328 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2332 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2334 return DAG.getConstant(0, MVT::i32);
2336 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2340 SDValue BitsFrom = N->getOperand(0);
2341 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2343 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2345 if (OffsetVal == 0) {
2346 // This is already sign / zero extended, so try to fold away extra BFEs.
2347 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2349 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2350 if (OpSignBits >= SignBits)
2353 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2355 // This is a sign_extend_inreg. Replace it to take advantage of existing
2356 // DAG Combines. If not eliminated, we will match back to BFE during
2359 // TODO: The sext_inreg of extended types ends, although we can could
2360 // handle them in a single BFE.
2361 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2362 DAG.getValueType(SmallVT));
2365 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2368 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2370 return constantFoldBFE<int32_t>(DAG,
2371 CVal->getSExtValue(),
2376 return constantFoldBFE<uint32_t>(DAG,
2377 CVal->getZExtValue(),
2382 if ((OffsetVal + WidthVal) >= 32) {
2383 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2384 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2385 BitsFrom, ShiftVal);
2388 if (BitsFrom.hasOneUse()) {
2389 APInt Demanded = APInt::getBitsSet(32,
2391 OffsetVal + WidthVal);
2393 APInt KnownZero, KnownOne;
2394 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2395 !DCI.isBeforeLegalizeOps());
2396 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2397 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2398 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2399 KnownZero, KnownOne, TLO)) {
2400 DCI.CommitTargetLoweringOpt(TLO);
2408 return performStoreCombine(N, DCI);
2413 //===----------------------------------------------------------------------===//
2415 //===----------------------------------------------------------------------===//
2417 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2420 const SmallVectorImpl<ISD::InputArg> &Ins,
2421 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2423 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2424 if (Ins[i].ArgVT == Ins[i].VT) {
2425 OrigIns.push_back(Ins[i]);
2430 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2431 // Vector has been split into scalars.
2432 VT = Ins[i].ArgVT.getVectorElementType();
2433 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2434 Ins[i].ArgVT.getVectorElementType() !=
2435 Ins[i].VT.getVectorElementType()) {
2436 // Vector elements have been promoted
2439 // Vector has been spilt into smaller vectors.
2443 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2444 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2445 OrigIns.push_back(Arg);
2449 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2450 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2451 return CFP->isExactlyValue(1.0);
2453 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2454 return C->isAllOnesValue();
2459 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2460 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2461 return CFP->getValueAPF().isZero();
2463 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2464 return C->isNullValue();
2469 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2470 const TargetRegisterClass *RC,
2471 unsigned Reg, EVT VT) const {
2472 MachineFunction &MF = DAG.getMachineFunction();
2473 MachineRegisterInfo &MRI = MF.getRegInfo();
2474 unsigned VirtualRegister;
2475 if (!MRI.isLiveIn(Reg)) {
2476 VirtualRegister = MRI.createVirtualRegister(RC);
2477 MRI.addLiveIn(Reg, VirtualRegister);
2479 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2481 return DAG.getRegister(VirtualRegister, VT);
2484 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2486 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2488 default: return nullptr;
2490 NODE_NAME_CASE(CALL);
2491 NODE_NAME_CASE(UMUL);
2492 NODE_NAME_CASE(RET_FLAG);
2493 NODE_NAME_CASE(BRANCH_COND);
2496 NODE_NAME_CASE(DWORDADDR)
2497 NODE_NAME_CASE(FRACT)
2498 NODE_NAME_CASE(CLAMP)
2500 NODE_NAME_CASE(FMAX_LEGACY)
2501 NODE_NAME_CASE(SMAX)
2502 NODE_NAME_CASE(UMAX)
2503 NODE_NAME_CASE(FMIN_LEGACY)
2504 NODE_NAME_CASE(SMIN)
2505 NODE_NAME_CASE(UMIN)
2506 NODE_NAME_CASE(FMAX3)
2507 NODE_NAME_CASE(SMAX3)
2508 NODE_NAME_CASE(UMAX3)
2509 NODE_NAME_CASE(FMIN3)
2510 NODE_NAME_CASE(SMIN3)
2511 NODE_NAME_CASE(UMIN3)
2512 NODE_NAME_CASE(URECIP)
2513 NODE_NAME_CASE(DIV_SCALE)
2514 NODE_NAME_CASE(DIV_FMAS)
2515 NODE_NAME_CASE(DIV_FIXUP)
2516 NODE_NAME_CASE(TRIG_PREOP)
2519 NODE_NAME_CASE(RSQ_LEGACY)
2520 NODE_NAME_CASE(RSQ_CLAMPED)
2521 NODE_NAME_CASE(LDEXP)
2522 NODE_NAME_CASE(DOT4)
2523 NODE_NAME_CASE(BFE_U32)
2524 NODE_NAME_CASE(BFE_I32)
2527 NODE_NAME_CASE(BREV)
2528 NODE_NAME_CASE(MUL_U24)
2529 NODE_NAME_CASE(MUL_I24)
2530 NODE_NAME_CASE(MAD_U24)
2531 NODE_NAME_CASE(MAD_I24)
2532 NODE_NAME_CASE(EXPORT)
2533 NODE_NAME_CASE(CONST_ADDRESS)
2534 NODE_NAME_CASE(REGISTER_LOAD)
2535 NODE_NAME_CASE(REGISTER_STORE)
2536 NODE_NAME_CASE(LOAD_CONSTANT)
2537 NODE_NAME_CASE(LOAD_INPUT)
2538 NODE_NAME_CASE(SAMPLE)
2539 NODE_NAME_CASE(SAMPLEB)
2540 NODE_NAME_CASE(SAMPLED)
2541 NODE_NAME_CASE(SAMPLEL)
2542 NODE_NAME_CASE(CVT_F32_UBYTE0)
2543 NODE_NAME_CASE(CVT_F32_UBYTE1)
2544 NODE_NAME_CASE(CVT_F32_UBYTE2)
2545 NODE_NAME_CASE(CVT_F32_UBYTE3)
2546 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2547 NODE_NAME_CASE(CONST_DATA_PTR)
2548 NODE_NAME_CASE(STORE_MSKOR)
2549 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2553 static void computeKnownBitsForMinMax(const SDValue Op0,
2557 const SelectionDAG &DAG,
2559 APInt Op0Zero, Op0One;
2560 APInt Op1Zero, Op1One;
2561 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2562 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2564 KnownZero = Op0Zero & Op1Zero;
2565 KnownOne = Op0One & Op1One;
2568 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2572 const SelectionDAG &DAG,
2573 unsigned Depth) const {
2575 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2579 unsigned Opc = Op.getOpcode();
2584 case ISD::INTRINSIC_WO_CHAIN: {
2585 // FIXME: The intrinsic should just use the node.
2586 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2587 case AMDGPUIntrinsic::AMDGPU_imax:
2588 case AMDGPUIntrinsic::AMDGPU_umax:
2589 case AMDGPUIntrinsic::AMDGPU_imin:
2590 case AMDGPUIntrinsic::AMDGPU_umin:
2591 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2592 KnownZero, KnownOne, DAG, Depth);
2600 case AMDGPUISD::SMAX:
2601 case AMDGPUISD::UMAX:
2602 case AMDGPUISD::SMIN:
2603 case AMDGPUISD::UMIN:
2604 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2605 KnownZero, KnownOne, DAG, Depth);
2608 case AMDGPUISD::BFE_I32:
2609 case AMDGPUISD::BFE_U32: {
2610 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2614 unsigned BitWidth = 32;
2615 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2617 if (Opc == AMDGPUISD::BFE_U32)
2618 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2625 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2627 const SelectionDAG &DAG,
2628 unsigned Depth) const {
2629 switch (Op.getOpcode()) {
2630 case AMDGPUISD::BFE_I32: {
2631 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2635 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2636 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2637 if (!Offset || !Offset->isNullValue())
2640 // TODO: Could probably figure something out with non-0 offsets.
2641 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2642 return std::max(SignBits, Op0SignBits);
2645 case AMDGPUISD::BFE_U32: {
2646 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2647 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;