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/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
38 /// Diagnostic information for unimplemented or unsupported feature reporting.
39 class DiagnosticInfoUnsupported : public DiagnosticInfo {
41 const Twine &Description;
46 static int getKindID() {
48 KindID = llvm::getNextAvailablePluginDiagnosticKind();
53 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
54 DiagnosticSeverity Severity = DS_Error)
55 : DiagnosticInfo(getKindID(), Severity),
59 const Function &getFunction() const { return Fn; }
60 const Twine &getDescription() const { return Description; }
62 void print(DiagnosticPrinter &DP) const override {
63 DP << "unsupported " << getDescription() << " in " << Fn.getName();
66 static bool classof(const DiagnosticInfo *DI) {
67 return DI->getKind() == getKindID();
71 int DiagnosticInfoUnsupported::KindID = 0;
75 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
76 CCValAssign::LocInfo LocInfo,
77 ISD::ArgFlagsTy ArgFlags, CCState &State) {
78 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
79 ArgFlags.getOrigAlign());
80 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
85 #include "AMDGPUGenCallingConv.inc"
87 // Find a larger type to do a load / store of a vector with.
88 EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
89 unsigned StoreSize = VT.getStoreSizeInBits();
91 return EVT::getIntegerVT(Ctx, StoreSize);
93 assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
94 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
97 // Type for a vector that will be loaded to.
98 EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
99 unsigned StoreSize = VT.getStoreSizeInBits();
101 return EVT::getIntegerVT(Ctx, 32);
103 return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
106 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
107 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
109 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
111 setOperationAction(ISD::Constant, MVT::i32, Legal);
112 setOperationAction(ISD::Constant, MVT::i64, Legal);
113 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
114 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
116 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
117 setOperationAction(ISD::BRIND, MVT::Other, Expand);
119 // We need to custom lower some of the intrinsics
120 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
122 // Library functions. These default to Expand, but we have instructions
124 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
125 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
126 setOperationAction(ISD::FPOW, MVT::f32, Legal);
127 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
128 setOperationAction(ISD::FABS, MVT::f32, Legal);
129 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
130 setOperationAction(ISD::FRINT, MVT::f32, Legal);
131 setOperationAction(ISD::FROUND, MVT::f32, Legal);
132 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
134 // Lower floating point store/load to integer store/load to reduce the number
135 // of patterns in tablegen.
136 setOperationAction(ISD::STORE, MVT::f32, Promote);
137 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
139 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
140 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
142 setOperationAction(ISD::STORE, MVT::i64, Promote);
143 AddPromotedToType(ISD::STORE, MVT::i64, MVT::v2i32);
145 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
146 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
148 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
149 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
151 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
152 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
154 setOperationAction(ISD::STORE, MVT::f64, Promote);
155 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
157 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
158 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
160 // Custom lowering of vector stores is required for local address space
162 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
163 // XXX: Native v2i32 local address space stores are possible, but not
164 // currently implemented.
165 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
167 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
168 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
169 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
171 // XXX: This can be change to Custom, once ExpandVectorStores can
172 // handle 64-bit stores.
173 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
175 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
176 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
177 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
178 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
179 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
182 setOperationAction(ISD::LOAD, MVT::f32, Promote);
183 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
185 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
186 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
188 setOperationAction(ISD::LOAD, MVT::i64, Promote);
189 AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
191 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
192 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
194 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
195 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
197 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
198 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
200 setOperationAction(ISD::LOAD, MVT::f64, Promote);
201 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
203 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
204 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
206 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
207 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
208 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
209 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
210 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
211 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
212 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
213 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
214 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
215 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
217 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
218 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
220 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
221 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
223 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
224 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
225 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
226 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
227 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
228 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
230 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
232 if (Subtarget->getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
233 setOperationAction(ISD::FCEIL, MVT::f64, Custom);
234 setOperationAction(ISD::FTRUNC, MVT::f64, Custom);
235 setOperationAction(ISD::FRINT, MVT::f64, Custom);
236 setOperationAction(ISD::FFLOOR, MVT::f64, Custom);
239 if (!Subtarget->hasBFI()) {
240 // fcopysign can be done in a single instruction with BFI.
241 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
242 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
245 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
247 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
248 for (MVT VT : ScalarIntVTs) {
249 setOperationAction(ISD::SREM, VT, Expand);
250 setOperationAction(ISD::SDIV, VT, Expand);
252 // GPU does not have divrem function for signed or unsigned.
253 setOperationAction(ISD::SDIVREM, VT, Custom);
254 setOperationAction(ISD::UDIVREM, VT, Custom);
256 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
257 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
258 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
260 setOperationAction(ISD::BSWAP, VT, Expand);
261 setOperationAction(ISD::CTTZ, VT, Expand);
262 setOperationAction(ISD::CTLZ, VT, Expand);
265 if (!Subtarget->hasBCNT(32))
266 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
268 if (!Subtarget->hasBCNT(64))
269 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
271 // The hardware supports 32-bit ROTR, but not ROTL.
272 setOperationAction(ISD::ROTL, MVT::i32, Expand);
273 setOperationAction(ISD::ROTL, MVT::i64, Expand);
274 setOperationAction(ISD::ROTR, MVT::i64, Expand);
276 setOperationAction(ISD::MUL, MVT::i64, Expand);
277 setOperationAction(ISD::MULHU, MVT::i64, Expand);
278 setOperationAction(ISD::MULHS, MVT::i64, Expand);
279 setOperationAction(ISD::UDIV, MVT::i32, Expand);
280 setOperationAction(ISD::UREM, MVT::i32, Expand);
281 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
282 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
284 if (!Subtarget->hasFFBH())
285 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
287 if (!Subtarget->hasFFBL())
288 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
290 static const MVT::SimpleValueType VectorIntTypes[] = {
291 MVT::v2i32, MVT::v4i32
294 for (MVT VT : VectorIntTypes) {
295 // Expand the following operations for the current type by default.
296 setOperationAction(ISD::ADD, VT, Expand);
297 setOperationAction(ISD::AND, VT, Expand);
298 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
299 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
300 setOperationAction(ISD::MUL, VT, Expand);
301 setOperationAction(ISD::OR, VT, Expand);
302 setOperationAction(ISD::SHL, VT, Expand);
303 setOperationAction(ISD::SRA, VT, Expand);
304 setOperationAction(ISD::SRL, VT, Expand);
305 setOperationAction(ISD::ROTL, VT, Expand);
306 setOperationAction(ISD::ROTR, VT, Expand);
307 setOperationAction(ISD::SUB, VT, Expand);
308 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
309 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
310 // TODO: Implement custom UREM / SREM routines.
311 setOperationAction(ISD::SDIV, VT, Expand);
312 setOperationAction(ISD::UDIV, VT, Expand);
313 setOperationAction(ISD::SREM, VT, Expand);
314 setOperationAction(ISD::UREM, VT, Expand);
315 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
316 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
317 setOperationAction(ISD::SDIVREM, VT, Custom);
318 setOperationAction(ISD::UDIVREM, VT, Custom);
319 setOperationAction(ISD::ADDC, VT, Expand);
320 setOperationAction(ISD::SUBC, VT, Expand);
321 setOperationAction(ISD::ADDE, VT, Expand);
322 setOperationAction(ISD::SUBE, VT, Expand);
323 setOperationAction(ISD::SELECT, VT, Expand);
324 setOperationAction(ISD::VSELECT, VT, Expand);
325 setOperationAction(ISD::SELECT_CC, VT, Expand);
326 setOperationAction(ISD::XOR, VT, Expand);
327 setOperationAction(ISD::BSWAP, VT, Expand);
328 setOperationAction(ISD::CTPOP, VT, Expand);
329 setOperationAction(ISD::CTTZ, VT, Expand);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
331 setOperationAction(ISD::CTLZ, VT, Expand);
332 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
333 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
336 static const MVT::SimpleValueType FloatVectorTypes[] = {
337 MVT::v2f32, MVT::v4f32
340 for (MVT VT : FloatVectorTypes) {
341 setOperationAction(ISD::FABS, VT, Expand);
342 setOperationAction(ISD::FADD, VT, Expand);
343 setOperationAction(ISD::FCEIL, VT, Expand);
344 setOperationAction(ISD::FCOS, VT, Expand);
345 setOperationAction(ISD::FDIV, VT, Expand);
346 setOperationAction(ISD::FEXP2, VT, Expand);
347 setOperationAction(ISD::FLOG2, VT, Expand);
348 setOperationAction(ISD::FPOW, VT, Expand);
349 setOperationAction(ISD::FFLOOR, VT, Expand);
350 setOperationAction(ISD::FTRUNC, VT, Expand);
351 setOperationAction(ISD::FMUL, VT, Expand);
352 setOperationAction(ISD::FMA, VT, Expand);
353 setOperationAction(ISD::FRINT, VT, Expand);
354 setOperationAction(ISD::FNEARBYINT, VT, Expand);
355 setOperationAction(ISD::FSQRT, VT, Expand);
356 setOperationAction(ISD::FSIN, VT, Expand);
357 setOperationAction(ISD::FSUB, VT, Expand);
358 setOperationAction(ISD::FNEG, VT, Expand);
359 setOperationAction(ISD::SELECT, VT, Expand);
360 setOperationAction(ISD::VSELECT, VT, Expand);
361 setOperationAction(ISD::SELECT_CC, VT, Expand);
362 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
363 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
366 setOperationAction(ISD::FNEARBYINT, MVT::f32, Custom);
367 setOperationAction(ISD::FNEARBYINT, MVT::f64, Custom);
369 setTargetDAGCombine(ISD::MUL);
370 setTargetDAGCombine(ISD::SELECT_CC);
371 setTargetDAGCombine(ISD::STORE);
373 setSchedulingPreference(Sched::RegPressure);
374 setJumpIsExpensive(true);
376 setSelectIsExpensive(false);
377 PredictableSelectIsExpensive = false;
379 // There are no integer divide instructions, and these expand to a pretty
380 // large sequence of instructions.
381 setIntDivIsCheap(false);
382 setPow2DivIsCheap(false);
384 // TODO: Investigate this when 64-bit divides are implemented.
385 addBypassSlowDiv(64, 32);
387 // FIXME: Need to really handle these.
388 MaxStoresPerMemcpy = 4096;
389 MaxStoresPerMemmove = 4096;
390 MaxStoresPerMemset = 4096;
393 //===----------------------------------------------------------------------===//
394 // Target Information
395 //===----------------------------------------------------------------------===//
397 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
401 bool AMDGPUTargetLowering::isSelectSupported(SelectSupportKind SelType) const {
405 // The backend supports 32 and 64 bit floating point immediates.
406 // FIXME: Why are we reporting vectors of FP immediates as legal?
407 bool AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
408 EVT ScalarVT = VT.getScalarType();
409 return (ScalarVT == MVT::f32 || ScalarVT == MVT::f64);
412 // We don't want to shrink f64 / f32 constants.
413 bool AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
414 EVT ScalarVT = VT.getScalarType();
415 return (ScalarVT != MVT::f32 && ScalarVT != MVT::f64);
418 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
420 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
423 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
424 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
426 return ((LScalarSize <= CastScalarSize) ||
427 (CastScalarSize >= 32) ||
431 //===---------------------------------------------------------------------===//
433 //===---------------------------------------------------------------------===//
435 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
436 assert(VT.isFloatingPoint());
437 return VT == MVT::f32;
440 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
441 assert(VT.isFloatingPoint());
442 return VT == MVT::f32;
445 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
446 // Truncate is just accessing a subregister.
447 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
450 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
451 // Truncate is just accessing a subregister.
452 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
453 (Dest->getPrimitiveSizeInBits() % 32 == 0);
456 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
457 const DataLayout *DL = getDataLayout();
458 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
459 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
461 return SrcSize == 32 && DestSize == 64;
464 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
465 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
466 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
467 // this will enable reducing 64-bit operations the 32-bit, which is always
469 return Src == MVT::i32 && Dest == MVT::i64;
472 bool AMDGPUTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
473 return isZExtFree(Val.getValueType(), VT2);
476 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
477 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
478 // limited number of native 64-bit operations. Shrinking an operation to fit
479 // in a single 32-bit register should always be helpful. As currently used,
480 // this is much less general than the name suggests, and is only used in
481 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
482 // not profitable, and may actually be harmful.
483 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
486 //===---------------------------------------------------------------------===//
487 // TargetLowering Callbacks
488 //===---------------------------------------------------------------------===//
490 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
491 const SmallVectorImpl<ISD::InputArg> &Ins) const {
493 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
496 SDValue AMDGPUTargetLowering::LowerReturn(
498 CallingConv::ID CallConv,
500 const SmallVectorImpl<ISD::OutputArg> &Outs,
501 const SmallVectorImpl<SDValue> &OutVals,
502 SDLoc DL, SelectionDAG &DAG) const {
503 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
506 //===---------------------------------------------------------------------===//
507 // Target specific lowering
508 //===---------------------------------------------------------------------===//
510 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
511 SmallVectorImpl<SDValue> &InVals) const {
512 SDValue Callee = CLI.Callee;
513 SelectionDAG &DAG = CLI.DAG;
515 const Function &Fn = *DAG.getMachineFunction().getFunction();
517 StringRef FuncName("<unknown>");
519 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
520 FuncName = G->getSymbol();
521 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
522 FuncName = G->getGlobal()->getName();
524 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
525 DAG.getContext()->diagnose(NoCalls);
529 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
530 SelectionDAG &DAG) const {
531 switch (Op.getOpcode()) {
533 Op.getNode()->dump();
534 llvm_unreachable("Custom lowering code for this"
535 "instruction is not implemented yet!");
537 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
538 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
539 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
540 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
541 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
542 case ISD::SDIV: return LowerSDIV(Op, DAG);
543 case ISD::SREM: return LowerSREM(Op, DAG);
544 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
545 case ISD::SDIVREM: return LowerSDIVREM(Op, DAG);
546 case ISD::FCEIL: return LowerFCEIL(Op, DAG);
547 case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
548 case ISD::FRINT: return LowerFRINT(Op, DAG);
549 case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
550 case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
551 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
556 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
557 SmallVectorImpl<SDValue> &Results,
558 SelectionDAG &DAG) const {
559 switch (N->getOpcode()) {
560 case ISD::SIGN_EXTEND_INREG:
561 // Different parts of legalization seem to interpret which type of
562 // sign_extend_inreg is the one to check for custom lowering. The extended
563 // from type is what really matters, but some places check for custom
564 // lowering of the result type. This results in trying to use
565 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
566 // nothing here and let the illegal result integer be handled normally.
569 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
573 Results.push_back(SDValue(Node, 0));
574 Results.push_back(SDValue(Node, 1));
575 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
577 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
581 SDValue Lowered = LowerSTORE(SDValue(N, 0), DAG);
582 if (Lowered.getNode())
583 Results.push_back(Lowered);
591 // FIXME: This implements accesses to initialized globals in the constant
592 // address space by copying them to private and accessing that. It does not
593 // properly handle illegal types or vectors. The private vector loads are not
594 // scalarized, and the illegal scalars hit an assertion. This technique will not
595 // work well with large initializers, and this should eventually be
596 // removed. Initialized globals should be placed into a data section that the
597 // runtime will load into a buffer before the kernel is executed. Uses of the
598 // global need to be replaced with a pointer loaded from an implicit kernel
599 // argument into this buffer holding the copy of the data, which will remove the
600 // need for any of this.
601 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
602 const GlobalValue *GV,
603 const SDValue &InitPtr,
605 SelectionDAG &DAG) const {
606 const DataLayout *TD = getTargetMachine().getDataLayout();
608 Type *InitTy = Init->getType();
610 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
611 EVT VT = EVT::getEVT(InitTy);
612 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
613 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
614 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
615 TD->getPrefTypeAlignment(InitTy));
618 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
619 EVT VT = EVT::getEVT(CFP->getType());
620 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
621 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
622 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
623 TD->getPrefTypeAlignment(CFP->getType()));
626 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
627 const StructLayout *SL = TD->getStructLayout(ST);
629 EVT PtrVT = InitPtr.getValueType();
630 SmallVector<SDValue, 8> Chains;
632 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
633 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
634 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
636 Constant *Elt = Init->getAggregateElement(I);
637 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
640 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
643 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
644 EVT PtrVT = InitPtr.getValueType();
646 unsigned NumElements;
647 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
648 NumElements = AT->getNumElements();
649 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
650 NumElements = VT->getNumElements();
652 llvm_unreachable("Unexpected type");
654 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
655 SmallVector<SDValue, 8> Chains;
656 for (unsigned i = 0; i < NumElements; ++i) {
657 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
658 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
660 Constant *Elt = Init->getAggregateElement(i);
661 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
664 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
667 if (isa<UndefValue>(Init)) {
668 EVT VT = EVT::getEVT(InitTy);
669 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
670 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
671 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
672 TD->getPrefTypeAlignment(InitTy));
676 llvm_unreachable("Unhandled constant initializer");
679 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
681 SelectionDAG &DAG) const {
683 const DataLayout *TD = getTargetMachine().getDataLayout();
684 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
685 const GlobalValue *GV = G->getGlobal();
687 switch (G->getAddressSpace()) {
688 default: llvm_unreachable("Global Address lowering not implemented for this "
690 case AMDGPUAS::LOCAL_ADDRESS: {
691 // XXX: What does the value of G->getOffset() mean?
692 assert(G->getOffset() == 0 &&
693 "Do not know what to do with an non-zero offset");
696 if (MFI->LocalMemoryObjects.count(GV) == 0) {
697 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
698 Offset = MFI->LDSSize;
699 MFI->LocalMemoryObjects[GV] = Offset;
700 // XXX: Account for alignment?
701 MFI->LDSSize += Size;
703 Offset = MFI->LocalMemoryObjects[GV];
706 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
708 case AMDGPUAS::CONSTANT_ADDRESS: {
709 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
710 Type *EltType = GV->getType()->getElementType();
711 unsigned Size = TD->getTypeAllocSize(EltType);
712 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
714 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
715 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
717 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
718 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
720 const GlobalVariable *Var = cast<GlobalVariable>(GV);
721 if (!Var->hasInitializer()) {
722 // This has no use, but bugpoint will hit it.
723 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
726 const Constant *Init = Var->getInitializer();
727 SmallVector<SDNode*, 8> WorkList;
729 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
730 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
731 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
733 WorkList.push_back(*I);
735 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
736 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
737 E = WorkList.end(); I != E; ++I) {
738 SmallVector<SDValue, 8> Ops;
739 Ops.push_back(Chain);
740 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
741 Ops.push_back((*I)->getOperand(i));
743 DAG.UpdateNodeOperands(*I, Ops);
745 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
750 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
751 SelectionDAG &DAG) const {
752 SmallVector<SDValue, 8> Args;
753 SDValue A = Op.getOperand(0);
754 SDValue B = Op.getOperand(1);
756 DAG.ExtractVectorElements(A, Args);
757 DAG.ExtractVectorElements(B, Args);
759 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
762 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
763 SelectionDAG &DAG) const {
765 SmallVector<SDValue, 8> Args;
766 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
767 EVT VT = Op.getValueType();
768 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
769 VT.getVectorNumElements());
771 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
774 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
775 SelectionDAG &DAG) const {
777 MachineFunction &MF = DAG.getMachineFunction();
778 const AMDGPUFrameLowering *TFL =
779 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
781 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
783 unsigned FrameIndex = FIN->getIndex();
784 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
785 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
789 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
790 SelectionDAG &DAG) const {
791 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
793 EVT VT = Op.getValueType();
795 switch (IntrinsicID) {
797 case AMDGPUIntrinsic::AMDGPU_abs:
798 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
799 return LowerIntrinsicIABS(Op, DAG);
800 case AMDGPUIntrinsic::AMDGPU_lrp:
801 return LowerIntrinsicLRP(Op, DAG);
802 case AMDGPUIntrinsic::AMDGPU_fract:
803 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
804 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
806 case AMDGPUIntrinsic::AMDGPU_clamp:
807 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
808 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
809 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
811 case Intrinsic::AMDGPU_div_scale: {
812 // 3rd parameter required to be a constant.
813 const ConstantSDNode *Param = dyn_cast<ConstantSDNode>(Op.getOperand(3));
815 return DAG.getUNDEF(VT);
817 // Translate to the operands expected by the machine instruction. The
818 // first parameter must be the same as the first instruction.
819 SDValue Numerator = Op.getOperand(1);
820 SDValue Denominator = Op.getOperand(2);
821 SDValue Src0 = Param->isAllOnesValue() ? Numerator : Denominator;
823 return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
824 Src0, Denominator, Numerator);
827 case Intrinsic::AMDGPU_div_fmas:
828 return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
829 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
831 case Intrinsic::AMDGPU_div_fixup:
832 return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
833 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
835 case Intrinsic::AMDGPU_trig_preop:
836 return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
837 Op.getOperand(1), Op.getOperand(2));
839 case Intrinsic::AMDGPU_rcp:
840 return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
842 case Intrinsic::AMDGPU_rsq:
843 return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
845 case AMDGPUIntrinsic::AMDGPU_legacy_rsq:
846 return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
848 case Intrinsic::AMDGPU_rsq_clamped:
849 return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
851 case AMDGPUIntrinsic::AMDGPU_imax:
852 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
854 case AMDGPUIntrinsic::AMDGPU_umax:
855 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
857 case AMDGPUIntrinsic::AMDGPU_imin:
858 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
860 case AMDGPUIntrinsic::AMDGPU_umin:
861 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
864 case AMDGPUIntrinsic::AMDGPU_umul24:
865 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
866 Op.getOperand(1), Op.getOperand(2));
868 case AMDGPUIntrinsic::AMDGPU_imul24:
869 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
870 Op.getOperand(1), Op.getOperand(2));
872 case AMDGPUIntrinsic::AMDGPU_umad24:
873 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
874 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
876 case AMDGPUIntrinsic::AMDGPU_imad24:
877 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
878 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
880 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
881 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
883 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
884 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
886 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
887 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
889 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
890 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
892 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
893 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
898 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
899 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
904 case AMDGPUIntrinsic::AMDGPU_bfi:
905 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
910 case AMDGPUIntrinsic::AMDGPU_bfm:
911 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
915 case AMDGPUIntrinsic::AMDGPU_brev:
916 return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
918 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
919 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
921 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
922 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
923 case AMDGPUIntrinsic::AMDGPU_trunc: // Legacy name.
924 return DAG.getNode(ISD::FTRUNC, DL, VT, Op.getOperand(1));
928 ///IABS(a) = SMAX(sub(0, a), a)
929 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
930 SelectionDAG &DAG) const {
932 EVT VT = Op.getValueType();
933 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
936 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
939 /// Linear Interpolation
940 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
941 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
942 SelectionDAG &DAG) const {
944 EVT VT = Op.getValueType();
945 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
946 DAG.getConstantFP(1.0f, MVT::f32),
948 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
950 return DAG.getNode(ISD::FADD, DL, VT,
951 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
955 /// \brief Generate Min/Max node
956 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
957 SelectionDAG &DAG) const {
959 EVT VT = N->getValueType(0);
961 SDValue LHS = N->getOperand(0);
962 SDValue RHS = N->getOperand(1);
963 SDValue True = N->getOperand(2);
964 SDValue False = N->getOperand(3);
965 SDValue CC = N->getOperand(4);
967 if (VT != MVT::f32 ||
968 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
972 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
986 llvm_unreachable("Operation should already be optimised!");
993 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
994 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1002 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
1003 return DAG.getNode(Opc, DL, VT, LHS, RHS);
1005 case ISD::SETCC_INVALID:
1006 llvm_unreachable("Invalid setcc condcode!");
1011 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
1012 SelectionDAG &DAG) const {
1013 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
1014 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
1015 EVT LoadVT = Op.getValueType();
1016 EVT EltVT = Op.getValueType().getVectorElementType();
1017 EVT PtrVT = Load->getBasePtr().getValueType();
1019 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
1020 SmallVector<SDValue, 8> Loads;
1021 SmallVector<SDValue, 8> Chains;
1025 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1026 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
1027 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
1030 = DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
1031 Load->getChain(), Ptr,
1032 MachinePointerInfo(Load->getMemOperand()->getValue()),
1033 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
1034 Load->getAlignment());
1035 Loads.push_back(NewLoad.getValue(0));
1036 Chains.push_back(NewLoad.getValue(1));
1040 DAG.getNode(ISD::BUILD_VECTOR, SL, LoadVT, Loads),
1041 DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains)
1044 return DAG.getMergeValues(Ops, SL);
1047 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
1048 SelectionDAG &DAG) const {
1049 StoreSDNode *Store = cast<StoreSDNode>(Op);
1050 EVT MemVT = Store->getMemoryVT();
1051 unsigned MemBits = MemVT.getSizeInBits();
1053 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
1054 // truncating store into an i32 store.
1055 // XXX: We could also handle optimize other vector bitwidths.
1056 if (!MemVT.isVector() || MemBits > 32) {
1061 SDValue Value = Store->getValue();
1062 EVT VT = Value.getValueType();
1063 EVT ElemVT = VT.getVectorElementType();
1064 SDValue Ptr = Store->getBasePtr();
1065 EVT MemEltVT = MemVT.getVectorElementType();
1066 unsigned MemEltBits = MemEltVT.getSizeInBits();
1067 unsigned MemNumElements = MemVT.getVectorNumElements();
1068 unsigned PackedSize = MemVT.getStoreSizeInBits();
1069 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1071 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1073 SDValue PackedValue;
1074 for (unsigned i = 0; i < MemNumElements; ++i) {
1075 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1076 DAG.getConstant(i, MVT::i32));
1077 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1078 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1080 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1081 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1086 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1090 if (PackedSize < 32) {
1091 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1092 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1093 Store->getMemOperand()->getPointerInfo(),
1095 Store->isNonTemporal(), Store->isVolatile(),
1096 Store->getAlignment());
1099 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1100 Store->getMemOperand()->getPointerInfo(),
1101 Store->isVolatile(), Store->isNonTemporal(),
1102 Store->getAlignment());
1105 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1106 SelectionDAG &DAG) const {
1107 StoreSDNode *Store = cast<StoreSDNode>(Op);
1108 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1109 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1110 EVT PtrVT = Store->getBasePtr().getValueType();
1111 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1114 SmallVector<SDValue, 8> Chains;
1116 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1117 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1118 Store->getValue(), DAG.getConstant(i, MVT::i32));
1119 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1120 Store->getBasePtr(),
1121 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1123 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1124 MachinePointerInfo(Store->getMemOperand()->getValue()),
1125 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1126 Store->getAlignment()));
1128 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1131 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1133 LoadSDNode *Load = cast<LoadSDNode>(Op);
1134 ISD::LoadExtType ExtType = Load->getExtensionType();
1135 EVT VT = Op.getValueType();
1136 EVT MemVT = Load->getMemoryVT();
1138 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1139 // We can do the extload to 32-bits, and then need to separately extend to
1142 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1146 Load->getMemOperand());
1149 DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32),
1150 ExtLoad32.getValue(1)
1153 return DAG.getMergeValues(Ops, DL);
1156 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1157 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1158 // FIXME: Copied from PPC
1159 // First, load into 32 bits, then truncate to 1 bit.
1161 SDValue Chain = Load->getChain();
1162 SDValue BasePtr = Load->getBasePtr();
1163 MachineMemOperand *MMO = Load->getMemOperand();
1165 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1166 BasePtr, MVT::i8, MMO);
1169 DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD),
1173 return DAG.getMergeValues(Ops, DL);
1176 // Lower loads constant address space global variable loads
1177 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1178 isa<GlobalVariable>(
1179 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1182 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1183 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1184 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1185 DAG.getConstant(2, MVT::i32));
1186 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op->getVTList(),
1187 Load->getChain(), Ptr,
1188 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1191 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1192 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1196 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1197 DAG.getConstant(2, MVT::i32));
1198 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1199 Load->getChain(), Ptr,
1200 DAG.getTargetConstant(0, MVT::i32),
1202 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1204 DAG.getConstant(0x3, MVT::i32));
1205 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1206 DAG.getConstant(3, MVT::i32));
1208 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1210 EVT MemEltVT = MemVT.getScalarType();
1211 if (ExtType == ISD::SEXTLOAD) {
1212 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1215 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode),
1219 return DAG.getMergeValues(Ops, DL);
1223 DAG.getZeroExtendInReg(Ret, DL, MemEltVT),
1227 return DAG.getMergeValues(Ops, DL);
1230 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1232 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1233 if (Result.getNode()) {
1237 StoreSDNode *Store = cast<StoreSDNode>(Op);
1238 SDValue Chain = Store->getChain();
1239 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1240 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1241 Store->getValue().getValueType().isVector()) {
1242 return SplitVectorStore(Op, DAG);
1245 EVT MemVT = Store->getMemoryVT();
1246 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1247 MemVT.bitsLT(MVT::i32)) {
1249 if (Store->getMemoryVT() == MVT::i8) {
1251 } else if (Store->getMemoryVT() == MVT::i16) {
1254 SDValue BasePtr = Store->getBasePtr();
1255 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1256 DAG.getConstant(2, MVT::i32));
1257 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1258 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1260 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1261 DAG.getConstant(0x3, MVT::i32));
1263 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1264 DAG.getConstant(3, MVT::i32));
1266 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1269 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1271 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1272 MaskedValue, ShiftAmt);
1274 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1276 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1277 DAG.getConstant(0xffffffff, MVT::i32));
1278 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1280 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1281 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1282 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1287 SDValue AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
1289 EVT OVT = Op.getValueType();
1290 SDValue LHS = Op.getOperand(0);
1291 SDValue RHS = Op.getOperand(1);
1294 if (!OVT.isVector()) {
1297 } else if (OVT.getVectorNumElements() == 2) {
1300 } else if (OVT.getVectorNumElements() == 4) {
1304 unsigned bitsize = OVT.getScalarType().getSizeInBits();
1305 // char|short jq = ia ^ ib;
1306 SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
1308 // jq = jq >> (bitsize - 2)
1309 jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT));
1312 jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
1315 jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
1317 // int ia = (int)LHS;
1318 SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
1320 // int ib, (int)RHS;
1321 SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
1323 // float fa = (float)ia;
1324 SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
1326 // float fb = (float)ib;
1327 SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
1329 // float fq = native_divide(fa, fb);
1330 SDValue fq = DAG.getNode(ISD::FMUL, DL, FLTTY,
1331 fa, DAG.getNode(AMDGPUISD::RCP, DL, FLTTY, fb));
1334 fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
1336 // float fqneg = -fq;
1337 SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
1339 // float fr = mad(fqneg, fb, fa);
1340 SDValue fr = DAG.getNode(ISD::FADD, DL, FLTTY,
1341 DAG.getNode(ISD::MUL, DL, FLTTY, fqneg, fb), fa);
1343 // int iq = (int)fq;
1344 SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
1347 fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
1350 fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
1352 // int cv = fr >= fb;
1354 if (INTTY == MVT::i32) {
1355 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1357 cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
1359 // jq = (cv ? jq : 0);
1360 jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq,
1361 DAG.getConstant(0, OVT));
1363 iq = DAG.getSExtOrTrunc(iq, DL, OVT);
1364 iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
1368 SDValue AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
1370 EVT OVT = Op.getValueType();
1371 SDValue LHS = Op.getOperand(0);
1372 SDValue RHS = Op.getOperand(1);
1373 // The LowerSDIV32 function generates equivalent to the following IL.
1383 // ixor r10, r10, r11
1385 // ixor DST, r0, r10
1394 SDValue r10 = DAG.getSelectCC(DL,
1395 r0, DAG.getConstant(0, OVT),
1396 DAG.getConstant(-1, OVT),
1397 DAG.getConstant(0, OVT),
1401 SDValue r11 = DAG.getSelectCC(DL,
1402 r1, DAG.getConstant(0, OVT),
1403 DAG.getConstant(-1, OVT),
1404 DAG.getConstant(0, OVT),
1408 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1411 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1414 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1417 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1420 r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
1422 // ixor r10, r10, r11
1423 r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
1426 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1428 // ixor DST, r0, r10
1429 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1433 SDValue AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
1434 return SDValue(Op.getNode(), 0);
1437 SDValue AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
1438 EVT OVT = Op.getValueType().getScalarType();
1440 if (OVT == MVT::i64)
1441 return LowerSDIV64(Op, DAG);
1443 if (OVT.getScalarType() == MVT::i32)
1444 return LowerSDIV32(Op, DAG);
1446 if (OVT == MVT::i16 || OVT == MVT::i8) {
1447 // FIXME: We should be checking for the masked bits. This isn't reached
1448 // because i8 and i16 are not legal types.
1449 return LowerSDIV24(Op, DAG);
1452 return SDValue(Op.getNode(), 0);
1455 SDValue AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
1457 EVT OVT = Op.getValueType();
1458 SDValue LHS = Op.getOperand(0);
1459 SDValue RHS = Op.getOperand(1);
1460 // The LowerSREM32 function generates equivalent to the following IL.
1470 // umul r20, r20, r1
1473 // ixor DST, r0, r10
1482 SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
1485 SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
1488 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1491 r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
1494 r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1497 r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
1500 SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
1502 // umul r20, r20, r1
1503 r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
1506 r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
1509 r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
1511 // ixor DST, r0, r10
1512 SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
1516 SDValue AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
1517 return SDValue(Op.getNode(), 0);
1520 SDValue AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
1521 EVT OVT = Op.getValueType();
1523 if (OVT.getScalarType() == MVT::i64)
1524 return LowerSREM64(Op, DAG);
1526 if (OVT.getScalarType() == MVT::i32)
1527 return LowerSREM32(Op, DAG);
1529 return SDValue(Op.getNode(), 0);
1532 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1533 SelectionDAG &DAG) const {
1535 EVT VT = Op.getValueType();
1537 SDValue Num = Op.getOperand(0);
1538 SDValue Den = Op.getOperand(1);
1540 // RCP = URECIP(Den) = 2^32 / Den + e
1541 // e is rounding error.
1542 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1544 // RCP_LO = umulo(RCP, Den) */
1545 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1547 // RCP_HI = mulhu (RCP, Den) */
1548 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1550 // NEG_RCP_LO = -RCP_LO
1551 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1554 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1555 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1558 // Calculate the rounding error from the URECIP instruction
1559 // E = mulhu(ABS_RCP_LO, RCP)
1560 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1562 // RCP_A_E = RCP + E
1563 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1565 // RCP_S_E = RCP - E
1566 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1568 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1569 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1572 // Quotient = mulhu(Tmp0, Num)
1573 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1575 // Num_S_Remainder = Quotient * Den
1576 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1578 // Remainder = Num - Num_S_Remainder
1579 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1581 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1582 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1583 DAG.getConstant(-1, VT),
1584 DAG.getConstant(0, VT),
1586 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1587 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1589 DAG.getConstant(-1, VT),
1590 DAG.getConstant(0, VT),
1592 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1593 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1596 // Calculate Division result:
1598 // Quotient_A_One = Quotient + 1
1599 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1600 DAG.getConstant(1, VT));
1602 // Quotient_S_One = Quotient - 1
1603 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1604 DAG.getConstant(1, VT));
1606 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1607 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1608 Quotient, Quotient_A_One, ISD::SETEQ);
1610 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1611 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1612 Quotient_S_One, Div, ISD::SETEQ);
1614 // Calculate Rem result:
1616 // Remainder_S_Den = Remainder - Den
1617 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1619 // Remainder_A_Den = Remainder + Den
1620 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1622 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1623 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1624 Remainder, Remainder_S_Den, ISD::SETEQ);
1626 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1627 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1628 Remainder_A_Den, Rem, ISD::SETEQ);
1633 return DAG.getMergeValues(Ops, DL);
1636 SDValue AMDGPUTargetLowering::LowerSDIVREM(SDValue Op,
1637 SelectionDAG &DAG) const {
1639 EVT VT = Op.getValueType();
1641 SDValue Zero = DAG.getConstant(0, VT);
1642 SDValue NegOne = DAG.getConstant(-1, VT);
1644 SDValue LHS = Op.getOperand(0);
1645 SDValue RHS = Op.getOperand(1);
1647 SDValue LHSign = DAG.getSelectCC(DL, LHS, Zero, NegOne, Zero, ISD::SETLT);
1648 SDValue RHSign = DAG.getSelectCC(DL, RHS, Zero, NegOne, Zero, ISD::SETLT);
1649 SDValue DSign = DAG.getNode(ISD::XOR, DL, VT, LHSign, RHSign);
1650 SDValue RSign = LHSign; // Remainder sign is the same as LHS
1652 LHS = DAG.getNode(ISD::ADD, DL, VT, LHS, LHSign);
1653 RHS = DAG.getNode(ISD::ADD, DL, VT, RHS, RHSign);
1655 LHS = DAG.getNode(ISD::XOR, DL, VT, LHS, LHSign);
1656 RHS = DAG.getNode(ISD::XOR, DL, VT, RHS, RHSign);
1658 SDValue Div = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT), LHS, RHS);
1659 SDValue Rem = Div.getValue(1);
1661 Div = DAG.getNode(ISD::XOR, DL, VT, Div, DSign);
1662 Rem = DAG.getNode(ISD::XOR, DL, VT, Rem, RSign);
1664 Div = DAG.getNode(ISD::SUB, DL, VT, Div, DSign);
1665 Rem = DAG.getNode(ISD::SUB, DL, VT, Rem, RSign);
1671 return DAG.getMergeValues(Res, DL);
1674 SDValue AMDGPUTargetLowering::LowerFCEIL(SDValue Op, SelectionDAG &DAG) const {
1676 SDValue Src = Op.getOperand(0);
1678 // result = trunc(src)
1679 // if (src > 0.0 && src != result)
1682 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1684 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1685 const SDValue One = DAG.getConstantFP(1.0, MVT::f64);
1687 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1689 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOGT);
1690 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1691 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1693 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, One, Zero);
1694 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1697 SDValue AMDGPUTargetLowering::LowerFTRUNC(SDValue Op, SelectionDAG &DAG) const {
1699 SDValue Src = Op.getOperand(0);
1701 assert(Op.getValueType() == MVT::f64);
1703 const SDValue Zero = DAG.getConstant(0, MVT::i32);
1704 const SDValue One = DAG.getConstant(1, MVT::i32);
1706 SDValue VecSrc = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Src);
1708 // Extract the upper half, since this is where we will find the sign and
1710 SDValue Hi = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, VecSrc, One);
1712 const unsigned FractBits = 52;
1713 const unsigned ExpBits = 11;
1715 // Extract the exponent.
1716 SDValue ExpPart = DAG.getNode(AMDGPUISD::BFE_I32, SL, MVT::i32,
1718 DAG.getConstant(FractBits - 32, MVT::i32),
1719 DAG.getConstant(ExpBits, MVT::i32));
1720 SDValue Exp = DAG.getNode(ISD::SUB, SL, MVT::i32, ExpPart,
1721 DAG.getConstant(1023, MVT::i32));
1723 // Extract the sign bit.
1724 const SDValue SignBitMask = DAG.getConstant(UINT32_C(1) << 31, MVT::i32);
1725 SDValue SignBit = DAG.getNode(ISD::AND, SL, MVT::i32, Hi, SignBitMask);
1727 // Extend back to to 64-bits.
1728 SDValue SignBit64 = DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32,
1730 SignBit64 = DAG.getNode(ISD::BITCAST, SL, MVT::i64, SignBit64);
1732 SDValue BcInt = DAG.getNode(ISD::BITCAST, SL, MVT::i64, Src);
1733 const SDValue FractMask
1734 = DAG.getConstant((UINT64_C(1) << FractBits) - 1, MVT::i64);
1736 SDValue Shr = DAG.getNode(ISD::SRA, SL, MVT::i64, FractMask, Exp);
1737 SDValue Not = DAG.getNOT(SL, Shr, MVT::i64);
1738 SDValue Tmp0 = DAG.getNode(ISD::AND, SL, MVT::i64, BcInt, Not);
1740 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::i32);
1742 const SDValue FiftyOne = DAG.getConstant(FractBits - 1, MVT::i32);
1744 SDValue ExpLt0 = DAG.getSetCC(SL, SetCCVT, Exp, Zero, ISD::SETLT);
1745 SDValue ExpGt51 = DAG.getSetCC(SL, SetCCVT, Exp, FiftyOne, ISD::SETGT);
1747 SDValue Tmp1 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpLt0, SignBit64, Tmp0);
1748 SDValue Tmp2 = DAG.getNode(ISD::SELECT, SL, MVT::i64, ExpGt51, BcInt, Tmp1);
1750 return DAG.getNode(ISD::BITCAST, SL, MVT::f64, Tmp2);
1753 SDValue AMDGPUTargetLowering::LowerFRINT(SDValue Op, SelectionDAG &DAG) const {
1755 SDValue Src = Op.getOperand(0);
1757 assert(Op.getValueType() == MVT::f64);
1759 APFloat C1Val(APFloat::IEEEdouble, "0x1.0p+52");
1760 SDValue C1 = DAG.getConstantFP(C1Val, MVT::f64);
1761 SDValue CopySign = DAG.getNode(ISD::FCOPYSIGN, SL, MVT::f64, C1, Src);
1763 SDValue Tmp1 = DAG.getNode(ISD::FADD, SL, MVT::f64, Src, CopySign);
1764 SDValue Tmp2 = DAG.getNode(ISD::FSUB, SL, MVT::f64, Tmp1, CopySign);
1766 SDValue Fabs = DAG.getNode(ISD::FABS, SL, MVT::f64, Src);
1768 APFloat C2Val(APFloat::IEEEdouble, "0x1.fffffffffffffp+51");
1769 SDValue C2 = DAG.getConstantFP(C2Val, MVT::f64);
1771 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1772 SDValue Cond = DAG.getSetCC(SL, SetCCVT, Fabs, C2, ISD::SETOGT);
1774 return DAG.getSelect(SL, MVT::f64, Cond, Src, Tmp2);
1777 SDValue AMDGPUTargetLowering::LowerFNEARBYINT(SDValue Op, SelectionDAG &DAG) const {
1778 // FNEARBYINT and FRINT are the same, except in their handling of FP
1779 // exceptions. Those aren't really meaningful for us, and OpenCL only has
1780 // rint, so just treat them as equivalent.
1781 return DAG.getNode(ISD::FRINT, SDLoc(Op), Op.getValueType(), Op.getOperand(0));
1784 SDValue AMDGPUTargetLowering::LowerFFLOOR(SDValue Op, SelectionDAG &DAG) const {
1786 SDValue Src = Op.getOperand(0);
1788 // result = trunc(src);
1789 // if (src < 0.0 && src != result)
1792 SDValue Trunc = DAG.getNode(ISD::FTRUNC, SL, MVT::f64, Src);
1794 const SDValue Zero = DAG.getConstantFP(0.0, MVT::f64);
1795 const SDValue NegOne = DAG.getConstantFP(-1.0, MVT::f64);
1797 EVT SetCCVT = getSetCCResultType(*DAG.getContext(), MVT::f64);
1799 SDValue Lt0 = DAG.getSetCC(SL, SetCCVT, Src, Zero, ISD::SETOLT);
1800 SDValue NeTrunc = DAG.getSetCC(SL, SetCCVT, Src, Trunc, ISD::SETONE);
1801 SDValue And = DAG.getNode(ISD::AND, SL, SetCCVT, Lt0, NeTrunc);
1803 SDValue Add = DAG.getNode(ISD::SELECT, SL, MVT::f64, And, NegOne, Zero);
1804 return DAG.getNode(ISD::FADD, SL, MVT::f64, Trunc, Add);
1807 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1808 SelectionDAG &DAG) const {
1809 SDValue S0 = Op.getOperand(0);
1811 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1814 // f32 uint_to_fp i64
1815 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1816 DAG.getConstant(0, MVT::i32));
1817 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1818 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1819 DAG.getConstant(1, MVT::i32));
1820 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1821 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1822 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1823 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1826 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1828 SelectionDAG &DAG) const {
1829 MVT VT = Op.getSimpleValueType();
1831 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1832 // Shift left by 'Shift' bits.
1833 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1834 // Signed shift Right by 'Shift' bits.
1835 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1838 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1839 SelectionDAG &DAG) const {
1840 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1841 MVT VT = Op.getSimpleValueType();
1842 MVT ScalarVT = VT.getScalarType();
1847 SDValue Src = Op.getOperand(0);
1850 // TODO: Don't scalarize on Evergreen?
1851 unsigned NElts = VT.getVectorNumElements();
1852 SmallVector<SDValue, 8> Args;
1853 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1855 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1856 for (unsigned I = 0; I < NElts; ++I)
1857 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1859 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1862 //===----------------------------------------------------------------------===//
1863 // Custom DAG optimizations
1864 //===----------------------------------------------------------------------===//
1866 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1867 APInt KnownZero, KnownOne;
1868 EVT VT = Op.getValueType();
1869 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1871 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1874 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1875 EVT VT = Op.getValueType();
1877 // In order for this to be a signed 24-bit value, bit 23, must
1879 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1880 // as unsigned 24-bit values.
1881 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1884 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1886 SelectionDAG &DAG = DCI.DAG;
1887 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1888 EVT VT = Op.getValueType();
1890 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1891 APInt KnownZero, KnownOne;
1892 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1893 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1894 DCI.CommitTargetLoweringOpt(TLO);
1897 template <typename IntTy>
1898 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1899 uint32_t Offset, uint32_t Width) {
1900 if (Width + Offset < 32) {
1901 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1902 return DAG.getConstant(Result, MVT::i32);
1905 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1908 static bool usesAllNormalStores(SDNode *LoadVal) {
1909 for (SDNode::use_iterator I = LoadVal->use_begin(); !I.atEnd(); ++I) {
1910 if (!ISD::isNormalStore(*I))
1917 // If we have a copy of an illegal type, replace it with a load / store of an
1918 // equivalently sized legal type. This avoids intermediate bit pack / unpack
1919 // instructions emitted when handling extloads and truncstores. Ideally we could
1920 // recognize the pack / unpack pattern to eliminate it.
1921 SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
1922 DAGCombinerInfo &DCI) const {
1923 if (!DCI.isBeforeLegalize())
1926 StoreSDNode *SN = cast<StoreSDNode>(N);
1927 SDValue Value = SN->getValue();
1928 EVT VT = Value.getValueType();
1930 if (isTypeLegal(VT) || SN->isVolatile() || !ISD::isNormalLoad(Value.getNode()))
1933 LoadSDNode *LoadVal = cast<LoadSDNode>(Value);
1934 if (LoadVal->isVolatile() || !usesAllNormalStores(LoadVal))
1937 EVT MemVT = LoadVal->getMemoryVT();
1940 SelectionDAG &DAG = DCI.DAG;
1941 EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
1943 SDValue NewLoad = DAG.getLoad(ISD::UNINDEXED, ISD::NON_EXTLOAD,
1945 LoadVal->getChain(),
1946 LoadVal->getBasePtr(),
1947 LoadVal->getOffset(),
1949 LoadVal->getMemOperand());
1951 SDValue CastLoad = DAG.getNode(ISD::BITCAST, SL, VT, NewLoad.getValue(0));
1952 DCI.CombineTo(LoadVal, CastLoad, NewLoad.getValue(1), false);
1954 return DAG.getStore(SN->getChain(), SL, NewLoad,
1955 SN->getBasePtr(), SN->getMemOperand());
1958 SDValue AMDGPUTargetLowering::performMulCombine(SDNode *N,
1959 DAGCombinerInfo &DCI) const {
1960 EVT VT = N->getValueType(0);
1962 if (VT.isVector() || VT.getSizeInBits() > 32)
1965 SelectionDAG &DAG = DCI.DAG;
1968 SDValue N0 = N->getOperand(0);
1969 SDValue N1 = N->getOperand(1);
1972 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1973 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1974 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1975 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1976 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1977 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1978 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1979 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1984 // We need to use sext even for MUL_U24, because MUL_U24 is used
1985 // for signed multiply of 8 and 16-bit types.
1986 return DAG.getSExtOrTrunc(Mul, DL, VT);
1989 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1990 DAGCombinerInfo &DCI) const {
1991 SelectionDAG &DAG = DCI.DAG;
1994 switch(N->getOpcode()) {
1997 return performMulCombine(N, DCI);
1998 case AMDGPUISD::MUL_I24:
1999 case AMDGPUISD::MUL_U24: {
2000 SDValue N0 = N->getOperand(0);
2001 SDValue N1 = N->getOperand(1);
2002 simplifyI24(N0, DCI);
2003 simplifyI24(N1, DCI);
2006 case ISD::SELECT_CC: {
2007 return CombineMinMax(N, DAG);
2009 case AMDGPUISD::BFE_I32:
2010 case AMDGPUISD::BFE_U32: {
2011 assert(!N->getValueType(0).isVector() &&
2012 "Vector handling of BFE not implemented");
2013 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
2017 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
2019 return DAG.getConstant(0, MVT::i32);
2021 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
2025 SDValue BitsFrom = N->getOperand(0);
2026 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
2028 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
2030 if (OffsetVal == 0) {
2031 // This is already sign / zero extended, so try to fold away extra BFEs.
2032 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
2034 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
2035 if (OpSignBits >= SignBits)
2038 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
2040 // This is a sign_extend_inreg. Replace it to take advantage of existing
2041 // DAG Combines. If not eliminated, we will match back to BFE during
2044 // TODO: The sext_inreg of extended types ends, although we can could
2045 // handle them in a single BFE.
2046 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
2047 DAG.getValueType(SmallVT));
2050 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
2053 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
2055 return constantFoldBFE<int32_t>(DAG,
2056 Val->getSExtValue(),
2061 return constantFoldBFE<uint32_t>(DAG,
2062 Val->getZExtValue(),
2067 APInt Demanded = APInt::getBitsSet(32,
2069 OffsetVal + WidthVal);
2071 if ((OffsetVal + WidthVal) >= 32) {
2072 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
2073 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
2074 BitsFrom, ShiftVal);
2077 APInt KnownZero, KnownOne;
2078 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
2079 !DCI.isBeforeLegalizeOps());
2080 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2081 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
2082 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
2083 DCI.CommitTargetLoweringOpt(TLO);
2090 return performStoreCombine(N, DCI);
2095 //===----------------------------------------------------------------------===//
2097 //===----------------------------------------------------------------------===//
2099 void AMDGPUTargetLowering::getOriginalFunctionArgs(
2102 const SmallVectorImpl<ISD::InputArg> &Ins,
2103 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
2105 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
2106 if (Ins[i].ArgVT == Ins[i].VT) {
2107 OrigIns.push_back(Ins[i]);
2112 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
2113 // Vector has been split into scalars.
2114 VT = Ins[i].ArgVT.getVectorElementType();
2115 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
2116 Ins[i].ArgVT.getVectorElementType() !=
2117 Ins[i].VT.getVectorElementType()) {
2118 // Vector elements have been promoted
2121 // Vector has been spilt into smaller vectors.
2125 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
2126 Ins[i].OrigArgIndex, Ins[i].PartOffset);
2127 OrigIns.push_back(Arg);
2131 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
2132 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2133 return CFP->isExactlyValue(1.0);
2135 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2136 return C->isAllOnesValue();
2141 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
2142 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2143 return CFP->getValueAPF().isZero();
2145 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
2146 return C->isNullValue();
2151 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
2152 const TargetRegisterClass *RC,
2153 unsigned Reg, EVT VT) const {
2154 MachineFunction &MF = DAG.getMachineFunction();
2155 MachineRegisterInfo &MRI = MF.getRegInfo();
2156 unsigned VirtualRegister;
2157 if (!MRI.isLiveIn(Reg)) {
2158 VirtualRegister = MRI.createVirtualRegister(RC);
2159 MRI.addLiveIn(Reg, VirtualRegister);
2161 VirtualRegister = MRI.getLiveInVirtReg(Reg);
2163 return DAG.getRegister(VirtualRegister, VT);
2166 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
2168 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
2170 default: return nullptr;
2172 NODE_NAME_CASE(CALL);
2173 NODE_NAME_CASE(UMUL);
2174 NODE_NAME_CASE(RET_FLAG);
2175 NODE_NAME_CASE(BRANCH_COND);
2178 NODE_NAME_CASE(DWORDADDR)
2179 NODE_NAME_CASE(FRACT)
2180 NODE_NAME_CASE(CLAMP)
2181 NODE_NAME_CASE(FMAX)
2182 NODE_NAME_CASE(SMAX)
2183 NODE_NAME_CASE(UMAX)
2184 NODE_NAME_CASE(FMIN)
2185 NODE_NAME_CASE(SMIN)
2186 NODE_NAME_CASE(UMIN)
2187 NODE_NAME_CASE(URECIP)
2188 NODE_NAME_CASE(DIV_SCALE)
2189 NODE_NAME_CASE(DIV_FMAS)
2190 NODE_NAME_CASE(DIV_FIXUP)
2191 NODE_NAME_CASE(TRIG_PREOP)
2194 NODE_NAME_CASE(RSQ_LEGACY)
2195 NODE_NAME_CASE(RSQ_CLAMPED)
2196 NODE_NAME_CASE(DOT4)
2197 NODE_NAME_CASE(BFE_U32)
2198 NODE_NAME_CASE(BFE_I32)
2201 NODE_NAME_CASE(BREV)
2202 NODE_NAME_CASE(MUL_U24)
2203 NODE_NAME_CASE(MUL_I24)
2204 NODE_NAME_CASE(MAD_U24)
2205 NODE_NAME_CASE(MAD_I24)
2206 NODE_NAME_CASE(EXPORT)
2207 NODE_NAME_CASE(CONST_ADDRESS)
2208 NODE_NAME_CASE(REGISTER_LOAD)
2209 NODE_NAME_CASE(REGISTER_STORE)
2210 NODE_NAME_CASE(LOAD_CONSTANT)
2211 NODE_NAME_CASE(LOAD_INPUT)
2212 NODE_NAME_CASE(SAMPLE)
2213 NODE_NAME_CASE(SAMPLEB)
2214 NODE_NAME_CASE(SAMPLED)
2215 NODE_NAME_CASE(SAMPLEL)
2216 NODE_NAME_CASE(CVT_F32_UBYTE0)
2217 NODE_NAME_CASE(CVT_F32_UBYTE1)
2218 NODE_NAME_CASE(CVT_F32_UBYTE2)
2219 NODE_NAME_CASE(CVT_F32_UBYTE3)
2220 NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
2221 NODE_NAME_CASE(STORE_MSKOR)
2222 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
2226 static void computeKnownBitsForMinMax(const SDValue Op0,
2230 const SelectionDAG &DAG,
2232 APInt Op0Zero, Op0One;
2233 APInt Op1Zero, Op1One;
2234 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
2235 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
2237 KnownZero = Op0Zero & Op1Zero;
2238 KnownOne = Op0One & Op1One;
2241 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
2245 const SelectionDAG &DAG,
2246 unsigned Depth) const {
2248 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
2252 unsigned Opc = Op.getOpcode();
2257 case ISD::INTRINSIC_WO_CHAIN: {
2258 // FIXME: The intrinsic should just use the node.
2259 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
2260 case AMDGPUIntrinsic::AMDGPU_imax:
2261 case AMDGPUIntrinsic::AMDGPU_umax:
2262 case AMDGPUIntrinsic::AMDGPU_imin:
2263 case AMDGPUIntrinsic::AMDGPU_umin:
2264 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
2265 KnownZero, KnownOne, DAG, Depth);
2273 case AMDGPUISD::SMAX:
2274 case AMDGPUISD::UMAX:
2275 case AMDGPUISD::SMIN:
2276 case AMDGPUISD::UMIN:
2277 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
2278 KnownZero, KnownOne, DAG, Depth);
2281 case AMDGPUISD::BFE_I32:
2282 case AMDGPUISD::BFE_U32: {
2283 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2287 unsigned BitWidth = 32;
2288 uint32_t Width = CWidth->getZExtValue() & 0x1f;
2290 KnownZero = APInt::getAllOnesValue(BitWidth);
2291 KnownOne = APInt::getNullValue(BitWidth);
2295 // FIXME: This could do a lot more. If offset is 0, should be the same as
2296 // sign_extend_inreg implementation, but that involves duplicating it.
2297 if (Opc == AMDGPUISD::BFE_I32)
2298 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2300 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
2307 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
2309 const SelectionDAG &DAG,
2310 unsigned Depth) const {
2311 switch (Op.getOpcode()) {
2312 case AMDGPUISD::BFE_I32: {
2313 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2317 unsigned SignBits = 32 - Width->getZExtValue() + 1;
2318 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
2319 if (!Offset || !Offset->isNullValue())
2322 // TODO: Could probably figure something out with non-0 offsets.
2323 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
2324 return std::max(SignBits, Op0SignBits);
2327 case AMDGPUISD::BFE_U32: {
2328 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
2329 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;