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::isLoadBitCastBeneficial(EVT LoadTy,
437 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
440 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
441 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
443 return ((LScalarSize <= CastScalarSize) ||
444 (CastScalarSize >= 32) ||
448 //===---------------------------------------------------------------------===//
450 //===---------------------------------------------------------------------===//
452 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
453 assert(VT.isFloatingPoint());
454 return VT == MVT::f32 || VT == MVT::f64;
457 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
458 assert(VT.isFloatingPoint());
459 return VT == MVT::f32 || VT == MVT::f64;
462 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
463 // Truncate is just accessing a subregister.
464 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
467 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
468 // Truncate is just accessing a subregister.
469 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
470 (Dest->getPrimitiveSizeInBits() % 32 == 0);
473 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
474 const DataLayout *DL = getDataLayout();
475 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
476 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
478 return SrcSize == 32 && DestSize == 64;
481 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
482 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
483 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
484 // this will enable reducing 64-bit operations the 32-bit, which is always
486 return Src == MVT::i32 && Dest == MVT::i64;
489 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
490 return isZExtFree(Val.getValueType(), VT2);
493 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
494 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
495 // limited number of native 64-bit operations. Shrinking an operation to fit
496 // in a single 32-bit register should always be helpful. As currently used,
497 // this is much less general than the name suggests, and is only used in
498 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
499 // not profitable, and may actually be harmful.
500 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
503 //===---------------------------------------------------------------------===//
504 // TargetLowering Callbacks
505 //===---------------------------------------------------------------------===//
507 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
508 const SmallVectorImpl<ISD::InputArg> &Ins) const {
510 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
513 SDValue AMDGPUTargetLowering::LowerReturn(
515 CallingConv::ID CallConv,
517 const SmallVectorImpl<ISD::OutputArg> &Outs,
518 const SmallVectorImpl<SDValue> &OutVals,
519 SDLoc DL, SelectionDAG &DAG) const {
520 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
523 //===---------------------------------------------------------------------===//
524 // Target specific lowering
525 //===---------------------------------------------------------------------===//
527 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
528 SmallVectorImpl<SDValue> &InVals) const {
529 SDValue Callee = CLI.Callee;
530 SelectionDAG &DAG = CLI.DAG;
532 const Function &Fn = *DAG.getMachineFunction().getFunction();
534 StringRef FuncName("<unknown>");
536 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
537 FuncName = G->getSymbol();
538 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
539 FuncName = G->getGlobal()->getName();
541 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
542 DAG.getContext()->diagnose(NoCalls);
546 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
547 SelectionDAG &DAG) const {
548 switch (Op.getOpcode()) {
550 Op.getNode()->dump();
551 llvm_unreachable("Custom lowering code for this"
552 "instruction is not implemented yet!");
554 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
555 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
556 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
557 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
558 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
559 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
560 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
561 case ISD::FREM: return LowerFREM(Op, DAG);
562 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
563 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
564 case ISD::FRINT: return LowerFRINT(Op, DAG);
565 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
566 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
567 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
568 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
569 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
570 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
575 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
576 SmallVectorImpl<SDValue> &Results,
577 SelectionDAG &DAG) const {
578 switch (N->getOpcode()) {
579 case ISD::SIGN_EXTEND_INREG:
580 // Different parts of legalization seem to interpret which type of
581 // sign_extend_inreg is the one to check for custom lowering. The extended
582 // from type is what really matters, but some places check for custom
583 // lowering of the result type. This results in trying to use
584 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
585 // nothing here and let the illegal result integer be handled normally.
588 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
592 Results.push_back(SDValue(Node, 0));
593 Results.push_back(SDValue(Node, 1));
594 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
596 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
600 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
601 if (Lowered.getNode())
602 Results.push_back(Lowered);
610 // FIXME: This implements accesses to initialized globals in the constant
611 // address space by copying them to private and accessing that. It does not
612 // properly handle illegal types or vectors. The private vector loads are not
613 // scalarized, and the illegal scalars hit an assertion. This technique will not
614 // work well with large initializers, and this should eventually be
615 // removed. Initialized globals should be placed into a data section that the
616 // runtime will load into a buffer before the kernel is executed. Uses of the
617 // global need to be replaced with a pointer loaded from an implicit kernel
618 // argument into this buffer holding the copy of the data, which will remove the
619 // need for any of this.
620 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
621 const GlobalValue *GV,
622 const SDValue &InitPtr,
624 SelectionDAG &DAG) const {
625 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
627 Type *InitTy = Init->getType();
629 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
630 EVT VT = EVT::getEVT(InitTy);
631 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
632 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
633 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
634 TD->getPrefTypeAlignment(InitTy));
637 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
638 EVT VT = EVT::getEVT(CFP->getType());
639 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
640 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
641 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
642 TD->getPrefTypeAlignment(CFP->getType()));
645 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
646 const StructLayout *SL = TD->getStructLayout(ST);
648 EVT PtrVT = InitPtr.getValueType();
649 SmallVector<SDValue, 8> Chains;
651 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
652 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
653 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
655 Constant *Elt = Init->getAggregateElement(I);
656 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
659 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
662 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
663 EVT PtrVT = InitPtr.getValueType();
665 unsigned NumElements;
666 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
667 NumElements = AT->getNumElements();
668 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
669 NumElements = VT->getNumElements();
671 llvm_unreachable("Unexpected type");
673 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
674 SmallVector<SDValue, 8> Chains;
675 for (unsigned i = 0; i < NumElements; ++i) {
676 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
677 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
679 Constant *Elt = Init->getAggregateElement(i);
680 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
683 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
686 if (isa<UndefValue>(Init)) {
687 EVT VT = EVT::getEVT(InitTy);
688 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
689 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
690 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
691 TD->getPrefTypeAlignment(InitTy));
695 llvm_unreachable("Unhandled constant initializer");
698 static bool hasDefinedInitializer(const GlobalValue *GV) {
699 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
700 if (!GVar || !GVar->hasInitializer())
703 if (isa<UndefValue>(GVar->getInitializer()))
709 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
711 SelectionDAG &DAG) const {
713 const DataLayout *TD = getTargetMachine().getSubtargetImpl()->getDataLayout();
714 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
715 const GlobalValue *GV = G->getGlobal();
717 switch (G->getAddressSpace()) {
718 case AMDGPUAS::LOCAL_ADDRESS: {
719 // XXX: What does the value of G->getOffset() mean?
720 assert(G->getOffset() == 0 &&
721 "Do not know what to do with an non-zero offset");
723 // TODO: We could emit code to handle the initialization somewhere.
724 if (hasDefinedInitializer(GV))
728 if (MFI->LocalMemoryObjects.count(GV) == 0) {
729 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
730 Offset = MFI->LDSSize;
731 MFI->LocalMemoryObjects[GV] = Offset;
732 // XXX: Account for alignment?
733 MFI->LDSSize += Size;
735 Offset = MFI->LocalMemoryObjects[GV];
738 return DAG.getConstant(Offset, getPointerTy(AMDGPUAS::LOCAL_ADDRESS));
740 case AMDGPUAS::CONSTANT_ADDRESS: {
741 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
742 Type *EltType = GV->getType()->getElementType();
743 unsigned Size = TD->getTypeAllocSize(EltType);
744 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
746 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
747 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
749 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
750 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
752 const GlobalVariable *Var = cast<GlobalVariable>(GV);
753 if (!Var->hasInitializer()) {
754 // This has no use, but bugpoint will hit it.
755 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
758 const Constant *Init = Var->getInitializer();
759 SmallVector<SDNode*, 8> WorkList;
761 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
762 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
763 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
765 WorkList.push_back(*I);
767 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
768 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
769 E = WorkList.end(); I != E; ++I) {
770 SmallVector<SDValue, 8> Ops;
771 Ops.push_back(Chain);
772 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
773 Ops.push_back((*I)->getOperand(i));
775 DAG.UpdateNodeOperands(*I, Ops);
777 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
781 const Function &Fn = *DAG.getMachineFunction().getFunction();
782 DiagnosticInfoUnsupported BadInit(Fn,
783 "initializer for address space");
784 DAG.getContext()->diagnose(BadInit);
788 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
789 SelectionDAG &DAG) const {
790 SmallVector<SDValue, 8> Args;
791 SDValue A = Op.getOperand(0);
792 SDValue B = Op.getOperand(1);
794 DAG.ExtractVectorElements(A, Args);
795 DAG.ExtractVectorElements(B, Args);
797 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
800 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
801 SelectionDAG &DAG) const {
803 SmallVector<SDValue, 8> Args;
804 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
805 EVT VT = Op.getValueType();
806 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
807 VT.getVectorNumElements());
809 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
812 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
813 SelectionDAG &DAG) const {
815 MachineFunction &MF = DAG.getMachineFunction();
816 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
817 getTargetMachine().getSubtargetImpl()->getFrameLowering());
819 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
821 unsigned FrameIndex = FIN->getIndex();
822 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
823 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
827 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
828 SelectionDAG &DAG) const {
829 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
831 EVT VT = Op.getValueType();
833 switch (IntrinsicID) {
835 case AMDGPUIntrinsic::AMDGPU_abs:
836 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
837 return LowerIntrinsicIABS(Op, DAG);
838 case AMDGPUIntrinsic::AMDGPU_lrp:
839 return LowerIntrinsicLRP(Op, DAG);
840 case AMDGPUIntrinsic::AMDGPU_fract:
841 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
842 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
844 case AMDGPUIntrinsic::AMDGPU_clamp:
845 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
846 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
847 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
849 case Intrinsic::AMDGPU_div_scale: {
850 // 3rd parameter required to be a constant.
851 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
853 return DAG.getUNDEF(VT);
855 // Translate to the operands expected by the machine instruction. The
856 // first parameter must be the same as the first instruction.
857 SDValue Numerator = Op.getOperand(1);
858 SDValue Denominator = Op.getOperand(2);
860 // Note this order is opposite of the machine instruction's operations,
861 // which is s0.f = Quotient, s1.f = Denominator, s2.f = Numerator. The
862 // intrinsic has the numerator as the first operand to match a normal
863 // division operation.
865 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
867 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
868 Denominator, Numerator);
871 case Intrinsic::AMDGPU_div_fmas:
872 // FIXME: Dropping bool parameter. Work is needed to support the implicit
874 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
875 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
877 case Intrinsic::AMDGPU_div_fixup:
878 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
879 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
881 case Intrinsic::AMDGPU_trig_preop:
882 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
883 Op.getOperand(1), Op.getOperand(2));
885 case Intrinsic::AMDGPU_rcp:
886 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
888 case Intrinsic::AMDGPU_rsq:
889 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
891 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
892 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
894 case Intrinsic::AMDGPU_rsq_clamped:
895 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
897 case Intrinsic::AMDGPU_ldexp:
898 return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
901 case AMDGPUIntrinsic::AMDGPU_imax:
902 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
904 case AMDGPUIntrinsic::AMDGPU_umax:
905 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
907 case AMDGPUIntrinsic::AMDGPU_imin:
908 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
910 case AMDGPUIntrinsic::AMDGPU_umin:
911 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
914 case AMDGPUIntrinsic::AMDGPU_umul24:
915 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
916 Op.getOperand(1), Op.getOperand(2));
918 case AMDGPUIntrinsic::AMDGPU_imul24:
919 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
920 Op.getOperand(1), Op.getOperand(2));
922 case AMDGPUIntrinsic::AMDGPU_umad24:
923 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
924 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
926 case AMDGPUIntrinsic::AMDGPU_imad24:
927 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
928 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
930 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
931 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
933 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
934 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
936 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
937 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
939 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
940 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
942 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
943 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
948 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
949 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
954 case AMDGPUIntrinsic::AMDGPU_bfi:
955 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
960 case AMDGPUIntrinsic::AMDGPU_bfm:
961 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
965 case AMDGPUIntrinsic::AMDGPU_brev:
966 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
968 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
969 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
971 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
972 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
973 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
974 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
978 ///IABS(a) = SMAX(sub(0, a), a)
979 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
980 SelectionDAG &DAG) const {
982 EVT VT = Op.getValueType();
983 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
986 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
989 /// Linear Interpolation
990 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
991 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
992 SelectionDAG &DAG) const {
994 EVT VT = Op.getValueType();
995 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
996 DAG.getConstantFP(1.0f, MVT::f32),
998 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
1000 return DAG.getNode(ISD::FADD, DL, VT,
1001 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
1005 /// \brief Generate Min/Max node
1006 SDValue AMDGPUTargetLowering::CombineFMinMax(SDLoc DL,
1013 SelectionDAG &DAG) const {
1014 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1017 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1026 case ISD::SETFALSE2:
1038 // We need to permute the operands to get the correct NaN behavior. The
1039 // selected operand is the second one based on the failing compare with NaN,
1040 // so permute it based on the compare type the hardware uses.
1042 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, RHS, LHS);
1043 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, LHS, RHS);
1052 return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
1053 return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
1055 case ISD::SETCC_INVALID:
1056 llvm_unreachable("Invalid setcc condcode!");
1061 /// \brief Generate Min/Max node
1062 SDValue AMDGPUTargetLowering::CombineIMinMax(SDLoc DL,
1069 SelectionDAG &DAG) const {
1070 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
1073 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
1077 unsigned Opc = (LHS == True) ? AMDGPUISD::UMIN : AMDGPUISD::UMAX;
1078 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1082 unsigned Opc = (LHS == True) ? AMDGPUISD::SMIN : AMDGPUISD::SMAX;
1083 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1087 unsigned Opc = (LHS == True) ? AMDGPUISD::SMAX : AMDGPUISD::SMIN;
1088 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1092 unsigned Opc = (LHS == True) ? AMDGPUISD::UMAX : AMDGPUISD::UMIN;
1093 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1100 SDValue AMDGPUTargetLowering::ScalarizeVectorLoad(const SDValue Op,
1101 SelectionDAG &DAG) const {
1102 LoadSDNode *Load = cast<LoadSDNode>(Op);
1103 EVT MemVT = Load->getMemoryVT();
1104 EVT MemEltVT = MemVT.getVectorElementType();
1106 EVT LoadVT = Op.getValueType();
1107 EVT EltVT = LoadVT.getVectorElementType();
1108 EVT PtrVT = Load->getBasePtr().getValueType();
1110 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1111 SmallVector<SDValue, 8> Loads;
1112 SmallVector<SDValue, 8> Chains;
1115 unsigned MemEltSize = MemEltVT.getStoreSize();
1116 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1118 for (unsigned i = 0; i < NumElts; ++i) {
1119 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1120 DAG.getConstant(i * MemEltSize, PtrVT));
1123 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1124 Load->getChain(), Ptr,
1125 SrcValue.getWithOffset(i * MemEltSize),
1126 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1127 Load->isInvariant(), Load->getAlignment());
1128 Loads.push_back(NewLoad.getValue(0));
1129 Chains.push_back(NewLoad.getValue(1));
1133 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1134 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1137 return DAG.getMergeValues(Ops, SL);
1140 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue Op,
1141 SelectionDAG &DAG) const {
1142 EVT VT = Op.getValueType();
1144 // If this is a 2 element vector, we really want to scalarize and not create
1145 // weird 1 element vectors.
1146 if (VT.getVectorNumElements() == 2)
1147 return ScalarizeVectorLoad(Op, DAG);
1149 LoadSDNode *Load = cast<LoadSDNode>(Op);
1150 SDValue BasePtr = Load->getBasePtr();
1151 EVT PtrVT = BasePtr.getValueType();
1152 EVT MemVT = Load->getMemoryVT();
1154 MachinePointerInfo SrcValue(Load->getMemOperand()->getValue());
1157 EVT LoMemVT, HiMemVT;
1160 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1161 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1162 std::tie(Lo, Hi) = DAG.SplitVector(Op, SL, LoVT, HiVT);
1164 = DAG.getExtLoad(Load->getExtensionType(), SL, LoVT,
1165 Load->getChain(), BasePtr,
1167 LoMemVT, Load->isVolatile(), Load->isNonTemporal(),
1168 Load->isInvariant(), Load->getAlignment());
1170 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1171 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1174 = DAG.getExtLoad(Load->getExtensionType(), SL, HiVT,
1175 Load->getChain(), HiPtr,
1176 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1177 HiMemVT, Load->isVolatile(), Load->isNonTemporal(),
1178 Load->isInvariant(), Load->getAlignment());
1181 DAG.getNode(ISD::CONCAT_VECTORS, SL, VT, LoLoad, HiLoad),
1182 DAG.getNode(ISD::TokenFactor, SL, MVT::Other,
1183 LoLoad.getValue(1), HiLoad.getValue(1))
1186 return DAG.getMergeValues(Ops, SL);
1189 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1190 SelectionDAG &DAG) const {
1191 StoreSDNode *Store = cast<StoreSDNode>(Op);
1192 EVT MemVT = Store->getMemoryVT();
1193 unsigned MemBits = MemVT.getSizeInBits();
1195 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1196 // truncating store into an i32 store.
1197 // XXX: We could also handle optimize other vector bitwidths.
1198 if (!MemVT.isVector() || MemBits > 32) {
1203 SDValue Value = Store->getValue();
1204 EVT VT = Value.getValueType();
1205 EVT ElemVT = VT.getVectorElementType();
1206 SDValue Ptr = Store->getBasePtr();
1207 EVT MemEltVT = MemVT.getVectorElementType();
1208 unsigned MemEltBits = MemEltVT.getSizeInBits();
1209 unsigned MemNumElements = MemVT.getVectorNumElements();
1210 unsigned PackedSize = MemVT.getStoreSizeInBits();
1211 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1213 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1215 SDValue PackedValue;
1216 for (unsigned i = 0; i < MemNumElements; ++i) {
1217 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1218 DAG.getConstant(i, MVT::i32));
1219 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1220 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1222 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1223 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1228 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1232 if (PackedSize < 32) {
1233 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1234 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1235 Store->getMemOperand()->getPointerInfo(),
1237 Store->isNonTemporal(), Store->isVolatile(),
1238 Store->getAlignment());
1241 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1242 Store->getMemOperand()->getPointerInfo(),
1243 Store->isVolatile(), Store->isNonTemporal(),
1244 Store->getAlignment());
1247 SDValue AMDGPUTargetLowering::ScalarizeVectorStore(SDValue Op,
1248 SelectionDAG &DAG) const {
1249 StoreSDNode *Store = cast<StoreSDNode>(Op);
1250 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1251 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1252 EVT PtrVT = Store->getBasePtr().getValueType();
1253 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1256 SmallVector<SDValue, 8> Chains;
1258 unsigned EltSize = MemEltVT.getStoreSize();
1259 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1261 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1262 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1264 DAG.getConstant(i, MVT::i32));
1266 SDValue Offset = DAG.getConstant(i * MemEltVT.getStoreSize(), PtrVT);
1267 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Store->getBasePtr(), Offset);
1269 DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1270 SrcValue.getWithOffset(i * EltSize),
1271 MemEltVT, Store->isNonTemporal(), Store->isVolatile(),
1272 Store->getAlignment());
1273 Chains.push_back(NewStore);
1276 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1279 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1280 SelectionDAG &DAG) const {
1281 StoreSDNode *Store = cast<StoreSDNode>(Op);
1282 SDValue Val = Store->getValue();
1283 EVT VT = Val.getValueType();
1285 // If this is a 2 element vector, we really want to scalarize and not create
1286 // weird 1 element vectors.
1287 if (VT.getVectorNumElements() == 2)
1288 return ScalarizeVectorStore(Op, DAG);
1290 EVT MemVT = Store->getMemoryVT();
1291 SDValue Chain = Store->getChain();
1292 SDValue BasePtr = Store->getBasePtr();
1296 EVT LoMemVT, HiMemVT;
1299 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
1300 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemVT);
1301 std::tie(Lo, Hi) = DAG.SplitVector(Val, SL, LoVT, HiVT);
1303 EVT PtrVT = BasePtr.getValueType();
1304 SDValue HiPtr = DAG.getNode(ISD::ADD, SL, PtrVT, BasePtr,
1305 DAG.getConstant(LoMemVT.getStoreSize(), PtrVT));
1307 MachinePointerInfo SrcValue(Store->getMemOperand()->getValue());
1309 = DAG.getTruncStore(Chain, SL, Lo,
1313 Store->isNonTemporal(),
1314 Store->isVolatile(),
1315 Store->getAlignment());
1317 = DAG.getTruncStore(Chain, SL, Hi,
1319 SrcValue.getWithOffset(LoMemVT.getStoreSize()),
1321 Store->isNonTemporal(),
1322 Store->isVolatile(),
1323 Store->getAlignment());
1325 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, LoStore, HiStore);
1329 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1331 LoadSDNode *Load = cast<LoadSDNode>(Op);
1332 ISD::LoadExtType ExtType = Load->getExtensionType();
1333 EVT VT = Op.getValueType();
1334 EVT MemVT = Load->getMemoryVT();
1336 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1337 // We can do the extload to 32-bits, and then need to separately extend to
1340 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1344 Load->getMemOperand());
1347 DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
1348 ExtLoad32.getValue(1)
1351 return DAG.getMergeValues(Ops, DL);
1354 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1355 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1356 // FIXME: Copied from PPC
1357 // First, load into 32 bits, then truncate to 1 bit.
1359 SDValue Chain = Load->getChain();
1360 SDValue BasePtr = Load->getBasePtr();
1361 MachineMemOperand *MMO = Load->getMemOperand();
1363 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1364 BasePtr, MVT::i8, MMO);
1367 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1371 return DAG.getMergeValues(Ops, DL);
1374 if (Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS ||
1375 Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1376 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1380 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1381 DAG.getConstant(2, MVT::i32));
1382 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1383 Load->getChain(), Ptr,
1384 DAG.getTargetConstant(0, MVT::i32),
1386 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1388 DAG.getConstant(0x3, MVT::i32));
1389 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1390 DAG.getConstant(3, MVT::i32));
1392 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1394 EVT MemEltVT = MemVT.getScalarType();
1395 if (ExtType == ISD::SEXTLOAD) {
1396 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1399 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1403 return DAG.getMergeValues(Ops, DL);
1407 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1411 return DAG.getMergeValues(Ops, DL);
1414 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1416 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1417 if (Result.getNode()) {
1421 StoreSDNode *Store = cast<StoreSDNode>(Op);
1422 SDValue Chain = Store->getChain();
1423 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1424 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1425 Store->getValue().getValueType().isVector()) {
1426 return ScalarizeVectorStore(Op, DAG);
1429 EVT MemVT = Store->getMemoryVT();
1430 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1431 MemVT.bitsLT(MVT::i32)) {
1433 if (Store->getMemoryVT() == MVT::i8) {
1435 } else if (Store->getMemoryVT() == MVT::i16) {
1438 SDValue BasePtr = Store->getBasePtr();
1439 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1440 DAG.getConstant(2, MVT::i32));
1441 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1442 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1444 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1445 DAG.getConstant(0x3, MVT::i32));
1447 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1448 DAG.getConstant(3, MVT::i32));
1450 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1453 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1455 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1456 MaskedValue, ShiftAmt);
1458 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1460 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1461 DAG.getConstant(0xffffffff, MVT::i32));
1462 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1464 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1465 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1466 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1471 // This is a shortcut for integer division because we have fast i32<->f32
1472 // conversions, and fast f32 reciprocal instructions. The fractional part of a
1473 // float is enough to accurately represent up to a 24-bit integer.
1474 SDValue AMDGPUTargetLowering::LowerDIVREM24(SDValue Op, SelectionDAG &DAG, bool sign) const {
1476 EVT VT = Op.getValueType();
1477 SDValue LHS = Op.getOperand(0);
1478 SDValue RHS = Op.getOperand(1);
1479 MVT IntVT = MVT::i32;
1480 MVT FltVT = MVT::f32;
1482 ISD::NodeType ToFp = sign ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
1483 ISD::NodeType ToInt = sign ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
1485 if (VT.isVector()) {
1486 unsigned NElts = VT.getVectorNumElements();
1487 IntVT = MVT::getVectorVT(MVT::i32, NElts);
1488 FltVT = MVT::getVectorVT(MVT::f32, NElts);
1491 unsigned BitSize = VT.getScalarType().getSizeInBits();
1493 SDValue jq = DAG.getConstant(1, IntVT);
1496 // char|short jq = ia ^ ib;
1497 jq = DAG.getNode(ISD::XOR, DL, VT, LHS, RHS);
1499 // jq = jq >> (bitsize - 2)
1500 jq = DAG.getNode(ISD::SRA, DL, VT, jq, DAG.getConstant(BitSize - 2, VT));
1503 jq = DAG.getNode(ISD::OR, DL, VT, jq, DAG.getConstant(1, VT));
1506 jq = DAG.getSExtOrTrunc(jq, DL, IntVT);
1509 // int ia = (int)LHS;
1511 DAG.getSExtOrTrunc(LHS, DL, IntVT) : DAG.getZExtOrTrunc(LHS, DL, IntVT);
1513 // int ib, (int)RHS;
1515 DAG.getSExtOrTrunc(RHS, DL, IntVT) : DAG.getZExtOrTrunc(RHS, DL, IntVT);
1517 // float fa = (float)ia;
1518 SDValue fa = DAG.getNode(ToFp, DL, FltVT, ia);
1520 // float fb = (float)ib;
1521 SDValue fb = DAG.getNode(ToFp, DL, FltVT, ib);
1523 // float fq = native_divide(fa, fb);
1524 SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
1525 fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
1528 fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
1530 // float fqneg = -fq;
1531 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
1533 // float fr = mad(fqneg, fb, fa);
1534 SDValue fr = DAG.getNode(ISD::FADD, DL, FltVT,
1535 DAG.getNode(ISD::FMUL, DL, FltVT, fqneg, fb), fa);
1537 // int iq = (int)fq;
1538 SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
1541 fr = DAG.getNode(ISD::FABS, DL, FltVT, fr);
1544 fb = DAG.getNode(ISD::FABS, DL, FltVT, fb);
1546 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), VT);
1548 // int cv = fr >= fb;
1549 SDValue cv = DAG.getSetCC(DL, SetCCVT, fr, fb, ISD::SETOGE);
1551 // jq = (cv ? jq : 0);
1552 jq = DAG.getNode(ISD::SELECT, DL, VT, cv, jq, DAG.getConstant(0, VT));
1554 // dst = trunc/extend to legal type
1555 iq = sign ? DAG.getSExtOrTrunc(iq, DL, VT) : DAG.getZExtOrTrunc(iq, DL, VT);
1558 SDValue Div = DAG.getNode(ISD::ADD, DL, VT, iq, jq);
1560 // Rem needs compensation, it's easier to recompute it
1561 SDValue Rem = DAG.getNode(ISD::MUL, DL, VT, Div, RHS);
1562 Rem = DAG.getNode(ISD::SUB, DL, VT, LHS, Rem);
1568 return DAG.getMergeValues(Res, DL);
1571 void AMDGPUTargetLowering::LowerUDIVREM64(SDValue Op,
1573 SmallVectorImpl<SDValue> &Results) const {
1574 assert(Op.getValueType() == MVT::i64);
1577 EVT VT = Op.getValueType();
1578 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
1580 SDValue one = DAG.getConstant(1, HalfVT);
1581 SDValue zero = DAG.getConstant(0, HalfVT);
1584 SDValue LHS = Op.getOperand(0);
1585 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
1586 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
1588 SDValue RHS = Op.getOperand(1);
1589 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
1590 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
1592 // Get Speculative values
1593 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
1594 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
1596 SDValue REM_Hi = zero;
1597 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
1599 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
1600 SDValue DIV_Lo = zero;
1602 const unsigned halfBitWidth = HalfVT.getSizeInBits();
1604 for (unsigned i = 0; i < halfBitWidth; ++i) {
1605 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
1606 // Get Value of high bit
1608 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
1609 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
1611 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
1612 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
1615 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
1616 DAG.getConstant(halfBitWidth - 1, HalfVT));
1617 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
1618 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
1620 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
1621 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
1624 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1626 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
1627 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETUGE);
1629 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
1633 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
1635 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETUGE);
1636 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
1637 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
1640 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
1641 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
1642 Results.push_back(DIV);
1643 Results.push_back(REM);
1646 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1647 SelectionDAG &DAG) const {
1649 EVT VT = Op.getValueType();
1651 if (VT == MVT::i64) {
1652 SmallVector<SDValue, 2> Results;
1653 LowerUDIVREM64(Op, DAG, Results);
1654 return DAG.getMergeValues(Results, DL);
1657 SDValue Num = Op.getOperand(0);
1658 SDValue Den = Op.getOperand(1);
1660 if (VT == MVT::i32) {
1661 if (DAG.MaskedValueIsZero(Op.getOperand(0), APInt(32, 0xff << 24)) &&
1662 DAG.MaskedValueIsZero(Op.getOperand(1), APInt(32, 0xff << 24))) {
1663 // TODO: We technically could do this for i64, but shouldn't that just be
1664 // handled by something generally reducing 64-bit division on 32-bit
1665 // values to 32-bit?
1666 return LowerDIVREM24(Op, DAG, false);
1670 // RCP = URECIP(Den) = 2^32 / Den + e
1671 // e is rounding error.
1672 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1674 // RCP_LO = mul(RCP, Den) */
1675 SDValue RCP_LO = DAG.getNode(ISD::MUL, DL, VT, RCP, Den);
1677 // RCP_HI = mulhu (RCP, Den) */
1678 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1680 // NEG_RCP_LO = -RCP_LO
1681 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1684 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1685 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1688 // Calculate the rounding error from the URECIP instruction
1689 // E = mulhu(ABS_RCP_LO, RCP)
1690 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1692 // RCP_A_E = RCP + E
1693 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1695 // RCP_S_E = RCP - E
1696 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1698 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1699 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1702 // Quotient = mulhu(Tmp0, Num)
1703 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1705 // Num_S_Remainder = Quotient * Den
1706 SDValue Num_S_Remainder = DAG.getNode(ISD::MUL, DL, VT, Quotient, Den);
1708 // Remainder = Num - Num_S_Remainder
1709 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1711 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1712 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1713 DAG.getConstant(-1, VT),
1714 DAG.getConstant(0, VT),
1716 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1717 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1719 DAG.getConstant(-1, VT),
1720 DAG.getConstant(0, VT),
1722 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1723 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1726 // Calculate Division result:
1728 // Quotient_A_One = Quotient + 1
1729 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1730 DAG.getConstant(1, VT));
1732 // Quotient_S_One = Quotient - 1
1733 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1734 DAG.getConstant(1, VT));
1736 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1737 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1738 Quotient, Quotient_A_One, ISD::SETEQ);
1740 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1741 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1742 Quotient_S_One, Div, ISD::SETEQ);
1744 // Calculate Rem result:
1746 // Remainder_S_Den = Remainder - Den
1747 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1749 // Remainder_A_Den = Remainder + Den
1750 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1752 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1753 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1754 Remainder, Remainder_S_Den, ISD::SETEQ);
1756 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1757 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1758 Remainder_A_Den, Rem, ISD::SETEQ);
1763 return DAG.getMergeValues(Ops, DL);
1766 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1767 SelectionDAG &DAG) const {
1769 EVT VT = Op.getValueType();
1771 SDValue LHS = Op.getOperand(0);
1772 SDValue RHS = Op.getOperand(1);
1774 if (VT == MVT::i32) {
1775 if (DAG.ComputeNumSignBits(Op.getOperand(0)) > 8 &&
1776 DAG.ComputeNumSignBits(Op.getOperand(1)) > 8) {
1777 // TODO: We technically could do this for i64, but shouldn't that just be
1778 // handled by something generally reducing 64-bit division on 32-bit
1779 // values to 32-bit?
1780 return LowerDIVREM24(Op, DAG, true);
1784 SDValue Zero = DAG.getConstant(0, VT);
1785 SDValue NegOne = DAG.getConstant(-1, VT);
1787 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1788 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1789 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1790 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1792 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1793 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1795 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1796 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1798 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1799 SDValue Rem = Div.getValue(1);
1801 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1802 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1804 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1805 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1811 return DAG.getMergeValues(Res, DL);
1814 // (frem x, y) -> (fsub x, (fmul (ftrunc (fdiv x, y)), y))
1815 SDValue AMDGPUTargetLowering::LowerFREM(SDValue Op, SelectionDAG &DAG) const {
1817 EVT VT = Op.getValueType();
1818 SDValue X = Op.getOperand(0);
1819 SDValue Y = Op.getOperand(1);
1821 SDValue Div = DAG.getNode(ISD::FDIV, SL, VT, X, Y);
1822 SDValue Floor = DAG.getNode(ISD::FTRUNC, SL, VT, Div);
1823 SDValue Mul = DAG.getNode(ISD::FMUL, SL, VT, Floor, Y);
1825 return DAG.getNode(ISD::FSUB, SL, VT, X, Mul);
1828 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1830 SDValue Src = Op.getOperand(0);
1832 // result = trunc(src)
1833 // if (src > 0.0 && src != result)
1836 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1838 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1839 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1841 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1843 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1844 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1845 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1847 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1848 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1851 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1853 SDValue Src = Op.getOperand(0);
1855 assert(Op.getValueType() == MVT::f64);
1857 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1858 const SDValue One = DAG.getConstant(1, MVT::i32);
1860 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1862 // Extract the upper half, since this is where we will find the sign and
1864 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1866 const unsigned FractBits = 52;
1867 const unsigned ExpBits = 11;
1869 // Extract the exponent.
1870 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_U32, SL, MVT::i32,
1872 DAG.getConstant(FractBits - 32, MVT::i32),
1873 DAG.getConstant(ExpBits, MVT::i32));
1874 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1875 DAG.getConstant(1023, MVT::i32));
1877 // Extract the sign bit.
1878 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1879 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1881 // Extend back to to 64-bits.
1882 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1884 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1886 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1887 const SDValue FractMask
1888 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1890 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1891 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1892 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1894 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1896 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1898 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1899 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1901 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1902 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1904 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1907 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1909 SDValue Src = Op.getOperand(0);
1911 assert(Op.getValueType() == MVT::f64);
1913 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1914 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1915 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1917 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1918 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1920 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1922 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1923 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1925 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1926 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1928 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1931 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1932 // FNEARBYINT and FRINT are the same, except in their handling of FP
1933 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1934 // rint, so just treat them as equivalent.
1935 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1938 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1940 SDValue Src = Op.getOperand(0);
1942 // result = trunc(src);
1943 // if (src < 0.0 && src != result)
1946 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1948 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1949 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1951 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1953 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1954 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1955 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1957 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1958 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1961 SDValue AMDGPUTargetLowering::LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG,
1962 bool Signed) const {
1964 SDValue Src = Op.getOperand(0);
1966 SDValue BC = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1968 SDValue Lo = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
1969 DAG.getConstant(0, MVT::i32));
1970 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, BC,
1971 DAG.getConstant(1, MVT::i32));
1973 SDValue CvtHi = DAG.getNode(Signed ? ISD::SINT_TO_FP : ISD::UINT_TO_FP,
1976 SDValue CvtLo = DAG.getNode(ISD::UINT_TO_FP, SL, MVT::f64, Lo);
1978 SDValue LdExp = DAG.getNode(AMDGPUISD::LDEXP, SL, MVT::f64, CvtHi,
1979 DAG.getConstant(32, MVT::i32));
1981 return DAG.getNode(ISD::FADD, SL, MVT::f64, LdExp, CvtLo);
1984 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1985 SelectionDAG &DAG) const {
1986 SDValue S0 = Op.getOperand(0);
1987 if (S0.getValueType() != MVT::i64)
1990 EVT DestVT = Op.getValueType();
1991 if (DestVT == MVT::f64)
1992 return LowerINT_TO_FP64(Op, DAG, false);
1994 assert(DestVT == MVT::f32);
1998 // f32 uint_to_fp i64
1999 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2000 DAG.getConstant(0, MVT::i32));
2001 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
2002 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
2003 DAG.getConstant(1, MVT::i32));
2004 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
2005 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
2006 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
2007 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
2010 SDValue AMDGPUTargetLowering::LowerSINT_TO_FP(SDValue Op,
2011 SelectionDAG &DAG) const {
2012 SDValue Src = Op.getOperand(0);
2013 if (Src.getValueType() == MVT::i64 && Op.getValueType() == MVT::f64)
2014 return LowerINT_TO_FP64(Op, DAG, true);
2019 SDValue AMDGPUTargetLowering::LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG,
2020 bool Signed) const {
2023 SDValue Src = Op.getOperand(0);
2025 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
2028 = DAG.getConstantFP(BitsToDouble(UINT64_C(0x3df0000000000000)), MVT::f64);
2030 = DAG.getConstantFP(BitsToDouble(UINT64_C(0xc1f0000000000000)), MVT::f64);
2032 SDValue Mul = DAG.getNode(ISD::FMUL, SL, MVT::f64, Trunc, K0);
2034 SDValue FloorMul = DAG.getNode(ISD::FFLOOR, SL, MVT::f64, Mul);
2037 SDValue Fma = DAG.getNode(ISD::FMA, SL, MVT::f64, FloorMul, K1, Trunc);
2039 SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
2040 MVT::i32, FloorMul);
2041 SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
2043 SDValue Result = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, Lo, Hi);
2045 return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
2048 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
2049 SelectionDAG &DAG) const {
2050 SDValue Src = Op.getOperand(0);
2052 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2053 return LowerFP64_TO_INT(Op, DAG, true);
2058 SDValue AMDGPUTargetLowering::LowerFP_TO_UINT(SDValue Op,
2059 SelectionDAG &DAG) const {
2060 SDValue Src = Op.getOperand(0);
2062 if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
2063 return LowerFP64_TO_INT(Op, DAG, false);
2068 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2069 SelectionDAG &DAG) const {
2070 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2071 MVT VT = Op.getSimpleValueType();
2072 MVT ScalarVT = VT.getScalarType();
2077 SDValue Src = Op.getOperand(0);
2080 // TODO: Don't scalarize on Evergreen?
2081 unsigned NElts = VT.getVectorNumElements();
2082 SmallVector<SDValue, 8> Args;
2083 DAG.ExtractVectorElements(Src, Args, 0, NElts);
2085 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
2086 for (unsigned I = 0; I < NElts; ++I)
2087 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
2089 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
2092 //===----------------------------------------------------------------------===//
2093 // Custom DAG optimizations
2094 //===----------------------------------------------------------------------===//
2096 static bool isU24(SDValue Op, SelectionDAG &DAG) {
2097 APInt KnownZero, KnownOne;
2098 EVT VT = Op.getValueType();
2099 DAG.computeKnownBits(Op, KnownZero, KnownOne);
2101 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
2104 static bool isI24(SDValue Op, SelectionDAG &DAG) {
2105 EVT VT = Op.getValueType();
2107 // In order for this to be a signed 24-bit value, bit 23, must
2109 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
2110 // as unsigned 24-bit values.
2111 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
2114 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
2116 SelectionDAG &DAG = DCI.DAG;
2117 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2118 EVT VT = Op.getValueType();
2120 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
2121 APInt KnownZero, KnownOne;
2122 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
2123 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
2124 DCI.CommitTargetLoweringOpt(TLO);
2127 template <typename IntTy>
2128 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
2129 uint32_t Offset, uint32_t Width) {
2130 if (Width + Offset < 32) {
2131 uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
2132 IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
2133 return DAG.getConstant(Result, MVT::i32);
2136 return DAG.getConstant(Src0 >> Offset, MVT::i32);
2139 static bool usesAllNormalStores(SDNode *LoadVal) {
2140 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
2141 if (!ISD::isNormalStore(*I))
2148 // If we have a copy of an illegal type, replace it with a load / store of an
2149 // equivalently sized legal type. This avoids intermediate bit pack / unpack
2150 // instructions emitted when handling extloads and truncstores. Ideally we could
2151 // recognize the pack / unpack pattern to eliminate it.
2152 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
2153 DAGCombinerInfo &DCI) const {
2154 if (!DCI.isBeforeLegalize())
2157 StoreSDNode *SN = cast<StoreSDNode>(N);
2158 SDValue Value = SN->getValue();
2159 EVT VT = Value.getValueType();
2161 if (isTypeLegal(VT) || SN->isVolatile() ||
2162 !ISD::isNormalLoad(Value.getNode()) || VT.getSizeInBits() < 8)
2165 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
2166 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
2169 EVT MemVT = LoadVal->getMemoryVT();
2172 SelectionDAG &DAG = DCI.DAG;
2173 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
2175 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
2177 LoadVal->getChain(),
2178 LoadVal->getBasePtr(),
2179 LoadVal->getOffset(),
2181 LoadVal->getMemOperand());
2183 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
2184 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
2186 return DAG.getStore(SN->getChain(), SL, NewLoad,
2187 SN->getBasePtr(), SN->getMemOperand());
2190 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
2191 DAGCombinerInfo &DCI) const {
2192 EVT VT = N->getValueType(0);
2194 if (VT.isVector() || VT.getSizeInBits() > 32)
2197 SelectionDAG &DAG = DCI.DAG;
2200 SDValue N0 = N->getOperand(0);
2201 SDValue N1 = N->getOperand(1);
2204 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
2205 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
2206 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
2207 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
2208 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
2209 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
2210 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
2211 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
2216 // We need to use sext even for MUL_U24, because MUL_U24 is used
2217 // for signed multiply of 8 and 16-bit types.
2218 return DAG.getSExtOrTrunc(Mul, DL, VT);
2221 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
2222 DAGCombinerInfo &DCI) const {
2223 SelectionDAG &DAG = DCI.DAG;
2226 switch(N->getOpcode()) {
2229 return performMulCombine(N, DCI);
2230 case AMDGPUISD::MUL_I24:
2231 case AMDGPUISD::MUL_U24: {
2232 SDValue N0 = N->getOperand(0);
2233 SDValue N1 = N->getOperand(1);
2234 simplifyI24(N0, DCI);
2235 simplifyI24(N1, DCI);
2238 case ISD::SELECT_CC: {
2240 EVT VT = N->getValueType(0);
2242 if (VT == MVT::f32 ||
2244 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
2245 SDValue LHS = N->getOperand(0);
2246 SDValue RHS = N->getOperand(1);
2247 SDValue True = N->getOperand(2);
2248 SDValue False = N->getOperand(3);
2249 SDValue CC = N->getOperand(4);
2251 return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2257 SDValue Cond = N->getOperand(0);
2258 if (Cond.getOpcode() == ISD::SETCC) {
2260 EVT VT = N->getValueType(0);
2261 SDValue LHS = Cond.getOperand(0);
2262 SDValue RHS = Cond.getOperand(1);
2263 SDValue CC = Cond.getOperand(2);
2265 SDValue True = N->getOperand(1);
2266 SDValue False = N->getOperand(2);
2268 if (VT == MVT::f32 ||
2270 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)) {
2271 return CombineFMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2274 // TODO: Implement min / max Evergreen instructions.
2275 if (VT == MVT::i32 &&
2276 Subtarget->getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
2277 return CombineIMinMax(DL, VT, LHS, RHS, True, False, CC, DAG);
2283 case AMDGPUISD::BFE_I32:
2284 case AMDGPUISD::BFE_U32: {
2285 assert(!N->getValueType(0).isVector() &&
2286 "Vector handling of BFE not implemented");
2287 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2291 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2293 return DAG.getConstant(0, MVT::i32);
2295 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2299 SDValue BitsFrom = N->getOperand(0);
2300 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2302 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2304 if (OffsetVal == 0) {
2305 // This is already sign / zero extended, so try to fold away extra BFEs.
2306 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2308 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2309 if (OpSignBits >= SignBits)
2312 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2314 // This is a sign_extend_inreg. Replace it to take advantage of existing
2315 // DAG Combines. If not eliminated, we will match back to BFE during
2318 // TODO: The sext_inreg of extended types ends, although we can could
2319 // handle them in a single BFE.
2320 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2321 DAG.getValueType(SmallVT));
2324 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2327 if (ConstantSDNode *CVal = dyn_cast<ConstantSDNode>(BitsFrom)) {
2329 return constantFoldBFE<int32_t>(DAG,
2330 CVal->getSExtValue(),
2335 return constantFoldBFE<uint32_t>(DAG,
2336 CVal->getZExtValue(),
2341 if ((OffsetVal + WidthVal) >= 32) {
2342 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2343 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2344 BitsFrom, ShiftVal);
2347 if (BitsFrom.hasOneUse()) {
2348 APInt Demanded = APInt::getBitsSet(32,
2350 OffsetVal + WidthVal);
2352 APInt KnownZero, KnownOne;
2353 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2354 !DCI.isBeforeLegalizeOps());
2355 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2356 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2357 TLI.SimplifyDemandedBits(BitsFrom, Demanded,
2358 KnownZero, KnownOne, TLO)) {
2359 DCI.CommitTargetLoweringOpt(TLO);
2367 return performStoreCombine(N, DCI);
2372 //===----------------------------------------------------------------------===//
2374 //===----------------------------------------------------------------------===//
2376 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2379 const SmallVectorImpl<ISD::InputArg> &Ins,
2380 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2382 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2383 if (Ins[i].ArgVT == Ins[i].VT) {
2384 OrigIns.push_back(Ins[i]);
2389 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2390 // Vector has been split into scalars.
2391 VT = Ins[i].ArgVT.getVectorElementType();
2392 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2393 Ins[i].ArgVT.getVectorElementType() !=
2394 Ins[i].VT.getVectorElementType()) {
2395 // Vector elements have been promoted
2398 // Vector has been spilt into smaller vectors.
2402 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2403 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2404 OrigIns.push_back(Arg);
2408 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2409 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2410 return CFP->isExactlyValue(1.0);
2412 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2413 return C->isAllOnesValue();
2418 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2419 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2420 return CFP->getValueAPF().isZero();
2422 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2423 return C->isNullValue();
2428 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2429 const TargetRegisterClass *RC,
2430 unsigned Reg, EVT VT) const {
2431 MachineFunction &MF = DAG.getMachineFunction();
2432 MachineRegisterInfo &MRI = MF.getRegInfo();
2433 unsigned VirtualRegister;
2434 if (!MRI.isLiveIn(Reg)) {
2435 VirtualRegister = MRI.createVirtualRegister(RC);
2436 MRI.addLiveIn(Reg, VirtualRegister);
2438 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2440 return DAG.getRegister(VirtualRegister, VT);
2443 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2445 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2447 default: return nullptr;
2449 NODE_NAME_CASE(CALL);
2450 NODE_NAME_CASE(UMUL);
2451 NODE_NAME_CASE(RET_FLAG);
2452 NODE_NAME_CASE(BRANCH_COND);
2455 NODE_NAME_CASE(DWORDADDR)
2456 NODE_NAME_CASE(FRACT)
2457 NODE_NAME_CASE(CLAMP)
2459 NODE_NAME_CASE(FMAX_LEGACY)
2460 NODE_NAME_CASE(SMAX)
2461 NODE_NAME_CASE(UMAX)
2462 NODE_NAME_CASE(FMIN_LEGACY)
2463 NODE_NAME_CASE(SMIN)
2464 NODE_NAME_CASE(UMIN)
2465 NODE_NAME_CASE(FMAX3)
2466 NODE_NAME_CASE(SMAX3)
2467 NODE_NAME_CASE(UMAX3)
2468 NODE_NAME_CASE(FMIN3)
2469 NODE_NAME_CASE(SMIN3)
2470 NODE_NAME_CASE(UMIN3)
2471 NODE_NAME_CASE(URECIP)
2472 NODE_NAME_CASE(DIV_SCALE)
2473 NODE_NAME_CASE(DIV_FMAS)
2474 NODE_NAME_CASE(DIV_FIXUP)
2475 NODE_NAME_CASE(TRIG_PREOP)
2478 NODE_NAME_CASE(RSQ_LEGACY)
2479 NODE_NAME_CASE(RSQ_CLAMPED)
2480 NODE_NAME_CASE(LDEXP)
2481 NODE_NAME_CASE(DOT4)
2482 NODE_NAME_CASE(BFE_U32)
2483 NODE_NAME_CASE(BFE_I32)
2486 NODE_NAME_CASE(BREV)
2487 NODE_NAME_CASE(MUL_U24)
2488 NODE_NAME_CASE(MUL_I24)
2489 NODE_NAME_CASE(MAD_U24)
2490 NODE_NAME_CASE(MAD_I24)
2491 NODE_NAME_CASE(EXPORT)
2492 NODE_NAME_CASE(CONST_ADDRESS)
2493 NODE_NAME_CASE(REGISTER_LOAD)
2494 NODE_NAME_CASE(REGISTER_STORE)
2495 NODE_NAME_CASE(LOAD_CONSTANT)
2496 NODE_NAME_CASE(LOAD_INPUT)
2497 NODE_NAME_CASE(SAMPLE)
2498 NODE_NAME_CASE(SAMPLEB)
2499 NODE_NAME_CASE(SAMPLED)
2500 NODE_NAME_CASE(SAMPLEL)
2501 NODE_NAME_CASE(CVT_F32_UBYTE0)
2502 NODE_NAME_CASE(CVT_F32_UBYTE1)
2503 NODE_NAME_CASE(CVT_F32_UBYTE2)
2504 NODE_NAME_CASE(CVT_F32_UBYTE3)
2505 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2506 NODE_NAME_CASE(CONST_DATA_PTR)
2507 NODE_NAME_CASE(STORE_MSKOR)
2508 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2512 static void computeKnownBitsForMinMax(const SDValue Op0,
2516 const SelectionDAG &DAG,
2518 APInt Op0Zero, Op0One;
2519 APInt Op1Zero, Op1One;
2520 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2521 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2523 KnownZero = Op0Zero & Op1Zero;
2524 KnownOne = Op0One & Op1One;
2527 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2531 const SelectionDAG &DAG,
2532 unsigned Depth) const {
2534 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2538 unsigned Opc = Op.getOpcode();
2543 case ISD::INTRINSIC_WO_CHAIN: {
2544 // FIXME: The intrinsic should just use the node.
2545 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2546 case AMDGPUIntrinsic::AMDGPU_imax:
2547 case AMDGPUIntrinsic::AMDGPU_umax:
2548 case AMDGPUIntrinsic::AMDGPU_imin:
2549 case AMDGPUIntrinsic::AMDGPU_umin:
2550 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2551 KnownZero, KnownOne, DAG, Depth);
2559 case AMDGPUISD::SMAX:
2560 case AMDGPUISD::UMAX:
2561 case AMDGPUISD::SMIN:
2562 case AMDGPUISD::UMIN:
2563 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2564 KnownZero, KnownOne, DAG, Depth);
2567 case AMDGPUISD::BFE_I32:
2568 case AMDGPUISD::BFE_U32: {
2569 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2573 unsigned BitWidth = 32;
2574 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2576 if (Opc == AMDGPUISD::BFE_U32)
2577 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2584 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2586 const SelectionDAG &DAG,
2587 unsigned Depth) const {
2588 switch (Op.getOpcode()) {
2589 case AMDGPUISD::BFE_I32: {
2590 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2594 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2595 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2596 if (!Offset || !Offset->isNullValue())
2599 // TODO: Could probably figure something out with non-0 offsets.
2600 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2601 return std::max(SignBits, Op0SignBits);
2604 case AMDGPUISD::BFE_U32: {
2605 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2606 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;