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 "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDILIntrinsicInfo.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 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
88 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
90 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
92 // Initialize target lowering borrowed from AMDIL
95 // We need to custom lower some of the intrinsics
96 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
98 // Library functions. These default to Expand, but we have instructions
100 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
101 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
102 setOperationAction(ISD::FPOW, MVT::f32, Legal);
103 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
104 setOperationAction(ISD::FABS, MVT::f32, Legal);
105 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
106 setOperationAction(ISD::FRINT, MVT::f32, Legal);
107 setOperationAction(ISD::FROUND, MVT::f32, Legal);
108 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
110 // The hardware supports ROTR, but not ROTL
111 setOperationAction(ISD::ROTL, MVT::i32, Expand);
113 // Lower floating point store/load to integer store/load to reduce the number
114 // of patterns in tablegen.
115 setOperationAction(ISD::STORE, MVT::f32, Promote);
116 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
118 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
119 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
121 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
122 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
124 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
125 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
127 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
128 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
130 setOperationAction(ISD::STORE, MVT::f64, Promote);
131 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
133 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
134 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
136 // Custom lowering of vector stores is required for local address space
138 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
139 // XXX: Native v2i32 local address space stores are possible, but not
140 // currently implemented.
141 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
143 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
144 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
145 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
147 // XXX: This can be change to Custom, once ExpandVectorStores can
148 // handle 64-bit stores.
149 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
151 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
152 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
153 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
154 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
155 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
158 setOperationAction(ISD::LOAD, MVT::f32, Promote);
159 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
161 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
162 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
164 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
165 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
167 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
168 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
170 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
171 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
173 setOperationAction(ISD::LOAD, MVT::f64, Promote);
174 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
176 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
177 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
179 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
180 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
181 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
182 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
183 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
184 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
185 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
186 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
187 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
188 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
190 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
191 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
192 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
193 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
194 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
195 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
196 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
197 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
198 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
199 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
201 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
203 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
205 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
207 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
209 setOperationAction(ISD::MUL, MVT::i64, Expand);
210 setOperationAction(ISD::SUB, MVT::i64, Expand);
212 setOperationAction(ISD::UDIV, MVT::i32, Expand);
213 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
214 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
215 setOperationAction(ISD::UREM, MVT::i32, Expand);
217 if (!Subtarget->hasBFI()) {
218 // fcopysign can be done in a single instruction with BFI.
219 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
220 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
223 for (MVT VT : { MVT::i32, MVT::i64 }) {
224 // TODO: Evergreen has BCNT_INT for CTPOP
225 setOperationAction(ISD::CTPOP, VT, Expand);
226 setOperationAction(ISD::CTTZ, VT, Expand);
227 setOperationAction(ISD::CTLZ, VT, Expand);
230 static const MVT::SimpleValueType IntTypes[] = {
231 MVT::v2i32, MVT::v4i32
234 for (MVT VT : IntTypes) {
235 // Expand the following operations for the current type by default.
236 setOperationAction(ISD::ADD, VT, Expand);
237 setOperationAction(ISD::AND, VT, Expand);
238 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
239 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
240 setOperationAction(ISD::MUL, VT, Expand);
241 setOperationAction(ISD::OR, VT, Expand);
242 setOperationAction(ISD::SHL, VT, Expand);
243 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
244 setOperationAction(ISD::SRL, VT, Expand);
245 setOperationAction(ISD::SRA, VT, Expand);
246 setOperationAction(ISD::SUB, VT, Expand);
247 setOperationAction(ISD::UDIV, VT, Expand);
248 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
249 setOperationAction(ISD::UREM, VT, Expand);
250 setOperationAction(ISD::SELECT, VT, Expand);
251 setOperationAction(ISD::VSELECT, VT, Expand);
252 setOperationAction(ISD::XOR, VT, Expand);
253 setOperationAction(ISD::BSWAP, VT, Expand);
254 setOperationAction(ISD::CTPOP, VT, Expand);
255 setOperationAction(ISD::CTTZ, VT, Expand);
256 setOperationAction(ISD::CTLZ, VT, Expand);
259 static const MVT::SimpleValueType FloatTypes[] = {
260 MVT::v2f32, MVT::v4f32
263 for (MVT VT : FloatTypes) {
264 setOperationAction(ISD::FABS, VT, Expand);
265 setOperationAction(ISD::FADD, VT, Expand);
266 setOperationAction(ISD::FCOS, VT, Expand);
267 setOperationAction(ISD::FDIV, VT, Expand);
268 setOperationAction(ISD::FPOW, VT, Expand);
269 setOperationAction(ISD::FFLOOR, VT, Expand);
270 setOperationAction(ISD::FTRUNC, VT, Expand);
271 setOperationAction(ISD::FMUL, VT, Expand);
272 setOperationAction(ISD::FRINT, VT, Expand);
273 setOperationAction(ISD::FSQRT, VT, Expand);
274 setOperationAction(ISD::FSIN, VT, Expand);
275 setOperationAction(ISD::FSUB, VT, Expand);
276 setOperationAction(ISD::FNEG, VT, Expand);
277 setOperationAction(ISD::SELECT, VT, Expand);
278 setOperationAction(ISD::VSELECT, VT, Expand);
279 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
282 setTargetDAGCombine(ISD::MUL);
283 setTargetDAGCombine(ISD::SELECT_CC);
286 //===----------------------------------------------------------------------===//
287 // Target Information
288 //===----------------------------------------------------------------------===//
290 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
294 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
296 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
299 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
300 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
302 return ((LScalarSize <= CastScalarSize) ||
303 (CastScalarSize >= 32) ||
307 //===---------------------------------------------------------------------===//
309 //===---------------------------------------------------------------------===//
311 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
312 assert(VT.isFloatingPoint());
313 return VT == MVT::f32;
316 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
317 assert(VT.isFloatingPoint());
318 return VT == MVT::f32;
321 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
322 // Truncate is just accessing a subregister.
323 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
326 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
327 // Truncate is just accessing a subregister.
328 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
329 (Dest->getPrimitiveSizeInBits() % 32 == 0);
332 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
333 const DataLayout *DL = getDataLayout();
334 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
335 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
337 return SrcSize == 32 && DestSize == 64;
340 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
341 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
342 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
343 // this will enable reducing 64-bit operations the 32-bit, which is always
345 return Src == MVT::i32 && Dest == MVT::i64;
348 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
349 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
350 // limited number of native 64-bit operations. Shrinking an operation to fit
351 // in a single 32-bit register should always be helpful. As currently used,
352 // this is much less general than the name suggests, and is only used in
353 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
354 // not profitable, and may actually be harmful.
355 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
358 //===---------------------------------------------------------------------===//
359 // TargetLowering Callbacks
360 //===---------------------------------------------------------------------===//
362 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
363 const SmallVectorImpl<ISD::InputArg> &Ins) const {
365 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
368 SDValue AMDGPUTargetLowering::LowerReturn(
370 CallingConv::ID CallConv,
372 const SmallVectorImpl<ISD::OutputArg> &Outs,
373 const SmallVectorImpl<SDValue> &OutVals,
374 SDLoc DL, SelectionDAG &DAG) const {
375 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
378 //===---------------------------------------------------------------------===//
379 // Target specific lowering
380 //===---------------------------------------------------------------------===//
382 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
383 SmallVectorImpl<SDValue> &InVals) const {
384 SDValue Callee = CLI.Callee;
385 SelectionDAG &DAG = CLI.DAG;
387 const Function &Fn = *DAG.getMachineFunction().getFunction();
389 StringRef FuncName("<unknown>");
391 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
392 FuncName = G->getSymbol();
393 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
394 FuncName = G->getGlobal()->getName();
396 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
397 DAG.getContext()->diagnose(NoCalls);
401 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
403 switch (Op.getOpcode()) {
405 Op.getNode()->dump();
406 llvm_unreachable("Custom lowering code for this"
407 "instruction is not implemented yet!");
409 // AMDIL DAG lowering
410 case ISD::SDIV: return LowerSDIV(Op, DAG);
411 case ISD::SREM: return LowerSREM(Op, DAG);
412 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
413 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
414 // AMDGPU DAG lowering
415 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
416 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
417 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
418 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
419 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
420 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
425 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
426 SmallVectorImpl<SDValue> &Results,
427 SelectionDAG &DAG) const {
428 switch (N->getOpcode()) {
429 case ISD::SIGN_EXTEND_INREG:
430 // Different parts of legalization seem to interpret which type of
431 // sign_extend_inreg is the one to check for custom lowering. The extended
432 // from type is what really matters, but some places check for custom
433 // lowering of the result type. This results in trying to use
434 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
435 // nothing here and let the illegal result integer be handled normally.
438 SDValue Op = SDValue(N, 0);
440 EVT VT = Op.getValueType();
441 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
442 N->getOperand(0), N->getOperand(1));
443 Results.push_back(UDIVREM);
447 SDValue Op = SDValue(N, 0);
449 EVT VT = Op.getValueType();
450 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
451 N->getOperand(0), N->getOperand(1));
452 Results.push_back(UDIVREM.getValue(1));
456 SDValue Op = SDValue(N, 0);
458 EVT VT = Op.getValueType();
459 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
461 SDValue one = DAG.getConstant(1, HalfVT);
462 SDValue zero = DAG.getConstant(0, HalfVT);
465 SDValue LHS = N->getOperand(0);
466 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
467 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
469 SDValue RHS = N->getOperand(1);
470 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
471 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
473 // Get Speculative values
474 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
475 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
477 SDValue REM_Hi = zero;
478 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
480 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
481 SDValue DIV_Lo = zero;
483 const unsigned halfBitWidth = HalfVT.getSizeInBits();
485 for (unsigned i = 0; i < halfBitWidth; ++i) {
486 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
487 // Get Value of high bit
489 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
490 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
492 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
493 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
496 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
497 DAG.getConstant(halfBitWidth - 1, HalfVT));
498 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
499 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
501 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
502 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
505 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
507 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
508 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
510 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
514 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
516 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
517 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
518 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
521 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
522 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
523 Results.push_back(DIV);
524 Results.push_back(REM);
532 // FIXME: This implements accesses to initialized globals in the constant
533 // address space by copying them to private and accessing that. It does not
534 // properly handle illegal types or vectors. The private vector loads are not
535 // scalarized, and the illegal scalars hit an assertion. This technique will not
536 // work well with large initializers, and this should eventually be
537 // removed. Initialized globals should be placed into a data section that the
538 // runtime will load into a buffer before the kernel is executed. Uses of the
539 // global need to be replaced with a pointer loaded from an implicit kernel
540 // argument into this buffer holding the copy of the data, which will remove the
541 // need for any of this.
542 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
543 const GlobalValue *GV,
544 const SDValue &InitPtr,
546 SelectionDAG &DAG) const {
547 const DataLayout *TD = getTargetMachine().getDataLayout();
549 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
550 EVT VT = EVT::getEVT(CI->getType());
551 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
552 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
553 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
554 TD->getPrefTypeAlignment(CI->getType()));
557 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
558 EVT VT = EVT::getEVT(CFP->getType());
559 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
560 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
561 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
562 TD->getPrefTypeAlignment(CFP->getType()));
565 Type *InitTy = Init->getType();
566 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
567 const StructLayout *SL = TD->getStructLayout(ST);
569 EVT PtrVT = InitPtr.getValueType();
570 SmallVector<SDValue, 8> Chains;
572 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
573 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
574 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
576 Constant *Elt = Init->getAggregateElement(I);
577 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
580 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
583 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
584 EVT PtrVT = InitPtr.getValueType();
586 unsigned NumElements;
587 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
588 NumElements = AT->getNumElements();
589 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
590 NumElements = VT->getNumElements();
592 llvm_unreachable("Unexpected type");
594 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
595 SmallVector<SDValue, 8> Chains;
596 for (unsigned i = 0; i < NumElements; ++i) {
597 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
598 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
600 Constant *Elt = Init->getAggregateElement(i);
601 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
604 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
608 llvm_unreachable("Unhandled constant initializer");
611 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
613 SelectionDAG &DAG) const {
615 const DataLayout *TD = getTargetMachine().getDataLayout();
616 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
617 const GlobalValue *GV = G->getGlobal();
619 switch (G->getAddressSpace()) {
620 default: llvm_unreachable("Global Address lowering not implemented for this "
622 case AMDGPUAS::LOCAL_ADDRESS: {
623 // XXX: What does the value of G->getOffset() mean?
624 assert(G->getOffset() == 0 &&
625 "Do not know what to do with an non-zero offset");
628 if (MFI->LocalMemoryObjects.count(GV) == 0) {
629 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
630 Offset = MFI->LDSSize;
631 MFI->LocalMemoryObjects[GV] = Offset;
632 // XXX: Account for alignment?
633 MFI->LDSSize += Size;
635 Offset = MFI->LocalMemoryObjects[GV];
638 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
640 case AMDGPUAS::CONSTANT_ADDRESS: {
641 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
642 Type *EltType = GV->getType()->getElementType();
643 unsigned Size = TD->getTypeAllocSize(EltType);
644 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
646 const GlobalVariable *Var = cast<GlobalVariable>(GV);
647 const Constant *Init = Var->getInitializer();
648 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
649 SDValue InitPtr = DAG.getFrameIndex(FI,
650 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
651 SmallVector<SDNode*, 8> WorkList;
653 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
654 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
655 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
657 WorkList.push_back(*I);
659 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
660 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
661 E = WorkList.end(); I != E; ++I) {
662 SmallVector<SDValue, 8> Ops;
663 Ops.push_back(Chain);
664 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
665 Ops.push_back((*I)->getOperand(i));
667 DAG.UpdateNodeOperands(*I, Ops);
669 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
670 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
675 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
676 SelectionDAG &DAG) const {
677 SmallVector<SDValue, 8> Args;
678 SDValue A = Op.getOperand(0);
679 SDValue B = Op.getOperand(1);
681 DAG.ExtractVectorElements(A, Args);
682 DAG.ExtractVectorElements(B, Args);
684 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
687 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
688 SelectionDAG &DAG) const {
690 SmallVector<SDValue, 8> Args;
691 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
692 EVT VT = Op.getValueType();
693 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
694 VT.getVectorNumElements());
696 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
699 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
700 SelectionDAG &DAG) const {
702 MachineFunction &MF = DAG.getMachineFunction();
703 const AMDGPUFrameLowering *TFL =
704 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
706 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
709 unsigned FrameIndex = FIN->getIndex();
710 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
711 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
715 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
716 SelectionDAG &DAG) const {
717 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
719 EVT VT = Op.getValueType();
721 switch (IntrinsicID) {
723 case AMDGPUIntrinsic::AMDIL_abs:
724 return LowerIntrinsicIABS(Op, DAG);
725 case AMDGPUIntrinsic::AMDIL_exp:
726 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
727 case AMDGPUIntrinsic::AMDGPU_lrp:
728 return LowerIntrinsicLRP(Op, DAG);
729 case AMDGPUIntrinsic::AMDIL_fraction:
730 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
731 case AMDGPUIntrinsic::AMDIL_max:
732 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
734 case AMDGPUIntrinsic::AMDGPU_imax:
735 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
737 case AMDGPUIntrinsic::AMDGPU_umax:
738 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
740 case AMDGPUIntrinsic::AMDIL_min:
741 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
743 case AMDGPUIntrinsic::AMDGPU_imin:
744 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
746 case AMDGPUIntrinsic::AMDGPU_umin:
747 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
750 case AMDGPUIntrinsic::AMDGPU_umul24:
751 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
752 Op.getOperand(1), Op.getOperand(2));
754 case AMDGPUIntrinsic::AMDGPU_imul24:
755 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
756 Op.getOperand(1), Op.getOperand(2));
758 case AMDGPUIntrinsic::AMDGPU_umad24:
759 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
760 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
762 case AMDGPUIntrinsic::AMDGPU_imad24:
763 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
764 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
766 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
767 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
772 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
773 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
778 case AMDGPUIntrinsic::AMDGPU_bfi:
779 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
784 case AMDGPUIntrinsic::AMDGPU_bfm:
785 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
789 case AMDGPUIntrinsic::AMDIL_round_nearest:
790 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
794 ///IABS(a) = SMAX(sub(0, a), a)
795 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
796 SelectionDAG &DAG) const {
798 EVT VT = Op.getValueType();
799 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
802 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
805 /// Linear Interpolation
806 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
807 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
808 SelectionDAG &DAG) const {
810 EVT VT = Op.getValueType();
811 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
812 DAG.getConstantFP(1.0f, MVT::f32),
814 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
816 return DAG.getNode(ISD::FADD, DL, VT,
817 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
821 /// \brief Generate Min/Max node
822 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
823 SelectionDAG &DAG) const {
825 EVT VT = N->getValueType(0);
827 SDValue LHS = N->getOperand(0);
828 SDValue RHS = N->getOperand(1);
829 SDValue True = N->getOperand(2);
830 SDValue False = N->getOperand(3);
831 SDValue CC = N->getOperand(4);
833 if (VT != MVT::f32 ||
834 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
838 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
852 llvm_unreachable("Operation should already be optimised!");
859 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
860 return DAG.getNode(Opc, DL, VT, LHS, RHS);
868 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
869 return DAG.getNode(Opc, DL, VT, LHS, RHS);
871 case ISD::SETCC_INVALID:
872 llvm_unreachable("Invalid setcc condcode!");
877 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
878 SelectionDAG &DAG) const {
879 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
880 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
881 EVT EltVT = Op.getValueType().getVectorElementType();
882 EVT PtrVT = Load->getBasePtr().getValueType();
883 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
884 SmallVector<SDValue, 8> Loads;
887 for (unsigned i = 0, e = NumElts; i != e; ++i) {
888 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
889 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
890 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
891 Load->getChain(), Ptr,
892 MachinePointerInfo(Load->getMemOperand()->getValue()),
893 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
894 Load->getAlignment()));
896 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
899 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
900 SelectionDAG &DAG) const {
901 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
902 EVT MemVT = Store->getMemoryVT();
903 unsigned MemBits = MemVT.getSizeInBits();
905 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
906 // truncating store into an i32 store.
907 // XXX: We could also handle optimize other vector bitwidths.
908 if (!MemVT.isVector() || MemBits > 32) {
913 SDValue Value = Store->getValue();
914 EVT VT = Value.getValueType();
915 EVT ElemVT = VT.getVectorElementType();
916 SDValue Ptr = Store->getBasePtr();
917 EVT MemEltVT = MemVT.getVectorElementType();
918 unsigned MemEltBits = MemEltVT.getSizeInBits();
919 unsigned MemNumElements = MemVT.getVectorNumElements();
920 unsigned PackedSize = MemVT.getStoreSizeInBits();
921 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
923 assert(Value.getValueType().getScalarSizeInBits() >= 32);
926 for (unsigned i = 0; i < MemNumElements; ++i) {
927 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
928 DAG.getConstant(i, MVT::i32));
929 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
930 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
932 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
933 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
938 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
942 if (PackedSize < 32) {
943 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
944 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
945 Store->getMemOperand()->getPointerInfo(),
947 Store->isNonTemporal(), Store->isVolatile(),
948 Store->getAlignment());
951 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
952 Store->getMemOperand()->getPointerInfo(),
953 Store->isVolatile(), Store->isNonTemporal(),
954 Store->getAlignment());
957 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
958 SelectionDAG &DAG) const {
959 StoreSDNode *Store = cast<StoreSDNode>(Op);
960 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
961 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
962 EVT PtrVT = Store->getBasePtr().getValueType();
963 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
966 SmallVector<SDValue, 8> Chains;
968 for (unsigned i = 0, e = NumElts; i != e; ++i) {
969 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
970 Store->getValue(), DAG.getConstant(i, MVT::i32));
971 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
973 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
975 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
976 MachinePointerInfo(Store->getMemOperand()->getValue()),
977 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
978 Store->getAlignment()));
980 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
983 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
985 LoadSDNode *Load = cast<LoadSDNode>(Op);
986 ISD::LoadExtType ExtType = Load->getExtensionType();
987 EVT VT = Op.getValueType();
988 EVT MemVT = Load->getMemoryVT();
990 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
991 // We can do the extload to 32-bits, and then need to separately extend to
994 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
998 Load->getMemOperand());
999 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1002 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1003 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1004 // FIXME: Copied from PPC
1005 // First, load into 32 bits, then truncate to 1 bit.
1007 SDValue Chain = Load->getChain();
1008 SDValue BasePtr = Load->getBasePtr();
1009 MachineMemOperand *MMO = Load->getMemOperand();
1011 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1012 BasePtr, MVT::i8, MMO);
1013 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1016 // Lower loads constant address space global variable loads
1017 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1018 isa<GlobalVariable>(
1019 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1021 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1022 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1023 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1024 DAG.getConstant(2, MVT::i32));
1025 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1026 Load->getChain(), Ptr,
1027 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1030 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1031 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1035 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1036 DAG.getConstant(2, MVT::i32));
1037 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1038 Load->getChain(), Ptr,
1039 DAG.getTargetConstant(0, MVT::i32),
1041 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1043 DAG.getConstant(0x3, MVT::i32));
1044 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1045 DAG.getConstant(3, MVT::i32));
1047 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1049 EVT MemEltVT = MemVT.getScalarType();
1050 if (ExtType == ISD::SEXTLOAD) {
1051 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1052 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1055 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1058 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1060 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1061 if (Result.getNode()) {
1065 StoreSDNode *Store = cast<StoreSDNode>(Op);
1066 SDValue Chain = Store->getChain();
1067 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1068 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1069 Store->getValue().getValueType().isVector()) {
1070 return SplitVectorStore(Op, DAG);
1073 EVT MemVT = Store->getMemoryVT();
1074 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1075 MemVT.bitsLT(MVT::i32)) {
1077 if (Store->getMemoryVT() == MVT::i8) {
1079 } else if (Store->getMemoryVT() == MVT::i16) {
1082 SDValue BasePtr = Store->getBasePtr();
1083 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1084 DAG.getConstant(2, MVT::i32));
1085 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1086 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1088 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1089 DAG.getConstant(0x3, MVT::i32));
1091 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1092 DAG.getConstant(3, MVT::i32));
1094 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1097 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1099 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1100 MaskedValue, ShiftAmt);
1102 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1104 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1105 DAG.getConstant(0xffffffff, MVT::i32));
1106 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1108 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1109 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1110 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1115 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1116 SelectionDAG &DAG) const {
1118 EVT VT = Op.getValueType();
1120 SDValue Num = Op.getOperand(0);
1121 SDValue Den = Op.getOperand(1);
1123 // RCP = URECIP(Den) = 2^32 / Den + e
1124 // e is rounding error.
1125 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1127 // RCP_LO = umulo(RCP, Den) */
1128 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1130 // RCP_HI = mulhu (RCP, Den) */
1131 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1133 // NEG_RCP_LO = -RCP_LO
1134 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1137 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1138 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1141 // Calculate the rounding error from the URECIP instruction
1142 // E = mulhu(ABS_RCP_LO, RCP)
1143 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1145 // RCP_A_E = RCP + E
1146 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1148 // RCP_S_E = RCP - E
1149 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1151 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1152 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1155 // Quotient = mulhu(Tmp0, Num)
1156 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1158 // Num_S_Remainder = Quotient * Den
1159 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1161 // Remainder = Num - Num_S_Remainder
1162 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1164 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1165 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1166 DAG.getConstant(-1, VT),
1167 DAG.getConstant(0, VT),
1169 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1170 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1172 DAG.getConstant(-1, VT),
1173 DAG.getConstant(0, VT),
1175 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1176 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1179 // Calculate Division result:
1181 // Quotient_A_One = Quotient + 1
1182 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1183 DAG.getConstant(1, VT));
1185 // Quotient_S_One = Quotient - 1
1186 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1187 DAG.getConstant(1, VT));
1189 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1190 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1191 Quotient, Quotient_A_One, ISD::SETEQ);
1193 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1194 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1195 Quotient_S_One, Div, ISD::SETEQ);
1197 // Calculate Rem result:
1199 // Remainder_S_Den = Remainder - Den
1200 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1202 // Remainder_A_Den = Remainder + Den
1203 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1205 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1206 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1207 Remainder, Remainder_S_Den, ISD::SETEQ);
1209 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1210 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1211 Remainder_A_Den, Rem, ISD::SETEQ);
1216 return DAG.getMergeValues(Ops, DL);
1219 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1220 SelectionDAG &DAG) const {
1221 SDValue S0 = Op.getOperand(0);
1223 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1226 // f32 uint_to_fp i64
1227 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1228 DAG.getConstant(0, MVT::i32));
1229 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1230 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1231 DAG.getConstant(1, MVT::i32));
1232 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1233 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1234 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1235 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1239 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1241 SelectionDAG &DAG) const {
1242 MVT VT = Op.getSimpleValueType();
1244 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1245 // Shift left by 'Shift' bits.
1246 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1247 // Signed shift Right by 'Shift' bits.
1248 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1251 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1252 SelectionDAG &DAG) const {
1253 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1254 MVT VT = Op.getSimpleValueType();
1255 MVT ScalarVT = VT.getScalarType();
1260 SDValue Src = Op.getOperand(0);
1263 // TODO: Don't scalarize on Evergreen?
1264 unsigned NElts = VT.getVectorNumElements();
1265 SmallVector<SDValue, 8> Args;
1266 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1268 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1269 for (unsigned I = 0; I < NElts; ++I)
1270 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1272 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1275 //===----------------------------------------------------------------------===//
1276 // Custom DAG optimizations
1277 //===----------------------------------------------------------------------===//
1279 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1280 APInt KnownZero, KnownOne;
1281 EVT VT = Op.getValueType();
1282 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1284 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1287 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1288 EVT VT = Op.getValueType();
1290 // In order for this to be a signed 24-bit value, bit 23, must
1292 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1293 // as unsigned 24-bit values.
1294 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1297 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1299 SelectionDAG &DAG = DCI.DAG;
1300 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1301 EVT VT = Op.getValueType();
1303 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1304 APInt KnownZero, KnownOne;
1305 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1306 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1307 DCI.CommitTargetLoweringOpt(TLO);
1310 template <typename IntTy>
1311 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1312 uint32_t Offset, uint32_t Width) {
1313 if (Width + Offset < 32) {
1314 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1315 return DAG.getConstant(Result, MVT::i32);
1318 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1321 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1322 DAGCombinerInfo &DCI) const {
1323 SelectionDAG &DAG = DCI.DAG;
1326 switch(N->getOpcode()) {
1329 EVT VT = N->getValueType(0);
1330 SDValue N0 = N->getOperand(0);
1331 SDValue N1 = N->getOperand(1);
1334 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1335 if (VT.isVector() || VT.getSizeInBits() > 32)
1338 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1339 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1340 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1341 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1342 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1343 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1344 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1345 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1350 // We need to use sext even for MUL_U24, because MUL_U24 is used
1351 // for signed multiply of 8 and 16-bit types.
1352 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1356 case AMDGPUISD::MUL_I24:
1357 case AMDGPUISD::MUL_U24: {
1358 SDValue N0 = N->getOperand(0);
1359 SDValue N1 = N->getOperand(1);
1360 simplifyI24(N0, DCI);
1361 simplifyI24(N1, DCI);
1364 case ISD::SELECT_CC: {
1365 return CombineMinMax(N, DAG);
1367 case AMDGPUISD::BFE_I32:
1368 case AMDGPUISD::BFE_U32: {
1369 assert(!N->getValueType(0).isVector() &&
1370 "Vector handling of BFE not implemented");
1371 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1375 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1377 return DAG.getConstant(0, MVT::i32);
1379 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1383 SDValue BitsFrom = N->getOperand(0);
1384 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1386 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1388 if (OffsetVal == 0) {
1389 // This is already sign / zero extended, so try to fold away extra BFEs.
1390 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1392 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1393 if (OpSignBits >= SignBits)
1396 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1398 // This is a sign_extend_inreg. Replace it to take advantage of existing
1399 // DAG Combines. If not eliminated, we will match back to BFE during
1402 // TODO: The sext_inreg of extended types ends, although we can could
1403 // handle them in a single BFE.
1404 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1405 DAG.getValueType(SmallVT));
1408 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1411 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1413 return constantFoldBFE<int32_t>(DAG,
1414 Val->getSExtValue(),
1419 return constantFoldBFE<uint32_t>(DAG,
1420 Val->getZExtValue(),
1425 APInt Demanded = APInt::getBitsSet(32,
1427 OffsetVal + WidthVal);
1429 if ((OffsetVal + WidthVal) >= 32) {
1430 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1431 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1432 BitsFrom, ShiftVal);
1435 APInt KnownZero, KnownOne;
1436 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1437 !DCI.isBeforeLegalizeOps());
1438 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1439 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1440 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1441 DCI.CommitTargetLoweringOpt(TLO);
1450 //===----------------------------------------------------------------------===//
1452 //===----------------------------------------------------------------------===//
1454 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1457 const SmallVectorImpl<ISD::InputArg> &Ins,
1458 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1460 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1461 if (Ins[i].ArgVT == Ins[i].VT) {
1462 OrigIns.push_back(Ins[i]);
1467 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1468 // Vector has been split into scalars.
1469 VT = Ins[i].ArgVT.getVectorElementType();
1470 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1471 Ins[i].ArgVT.getVectorElementType() !=
1472 Ins[i].VT.getVectorElementType()) {
1473 // Vector elements have been promoted
1476 // Vector has been spilt into smaller vectors.
1480 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1481 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1482 OrigIns.push_back(Arg);
1486 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1487 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1488 return CFP->isExactlyValue(1.0);
1490 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1491 return C->isAllOnesValue();
1496 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1497 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1498 return CFP->getValueAPF().isZero();
1500 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1501 return C->isNullValue();
1506 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1507 const TargetRegisterClass *RC,
1508 unsigned Reg, EVT VT) const {
1509 MachineFunction &MF = DAG.getMachineFunction();
1510 MachineRegisterInfo &MRI = MF.getRegInfo();
1511 unsigned VirtualRegister;
1512 if (!MRI.isLiveIn(Reg)) {
1513 VirtualRegister = MRI.createVirtualRegister(RC);
1514 MRI.addLiveIn(Reg, VirtualRegister);
1516 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1518 return DAG.getRegister(VirtualRegister, VT);
1521 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1523 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1525 default: return nullptr;
1527 NODE_NAME_CASE(CALL);
1528 NODE_NAME_CASE(UMUL);
1529 NODE_NAME_CASE(DIV_INF);
1530 NODE_NAME_CASE(RET_FLAG);
1531 NODE_NAME_CASE(BRANCH_COND);
1534 NODE_NAME_CASE(DWORDADDR)
1535 NODE_NAME_CASE(FRACT)
1536 NODE_NAME_CASE(FMAX)
1537 NODE_NAME_CASE(SMAX)
1538 NODE_NAME_CASE(UMAX)
1539 NODE_NAME_CASE(FMIN)
1540 NODE_NAME_CASE(SMIN)
1541 NODE_NAME_CASE(UMIN)
1542 NODE_NAME_CASE(BFE_U32)
1543 NODE_NAME_CASE(BFE_I32)
1546 NODE_NAME_CASE(MUL_U24)
1547 NODE_NAME_CASE(MUL_I24)
1548 NODE_NAME_CASE(MAD_U24)
1549 NODE_NAME_CASE(MAD_I24)
1550 NODE_NAME_CASE(URECIP)
1551 NODE_NAME_CASE(DOT4)
1552 NODE_NAME_CASE(EXPORT)
1553 NODE_NAME_CASE(CONST_ADDRESS)
1554 NODE_NAME_CASE(REGISTER_LOAD)
1555 NODE_NAME_CASE(REGISTER_STORE)
1556 NODE_NAME_CASE(LOAD_CONSTANT)
1557 NODE_NAME_CASE(LOAD_INPUT)
1558 NODE_NAME_CASE(SAMPLE)
1559 NODE_NAME_CASE(SAMPLEB)
1560 NODE_NAME_CASE(SAMPLED)
1561 NODE_NAME_CASE(SAMPLEL)
1562 NODE_NAME_CASE(STORE_MSKOR)
1563 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1567 static void computeKnownBitsForMinMax(const SDValue Op0,
1571 const SelectionDAG &DAG,
1573 APInt Op0Zero, Op0One;
1574 APInt Op1Zero, Op1One;
1575 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1576 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1578 KnownZero = Op0Zero & Op1Zero;
1579 KnownOne = Op0One & Op1One;
1582 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1586 const SelectionDAG &DAG,
1587 unsigned Depth) const {
1589 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1593 unsigned Opc = Op.getOpcode();
1598 case ISD::INTRINSIC_WO_CHAIN: {
1599 // FIXME: The intrinsic should just use the node.
1600 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1601 case AMDGPUIntrinsic::AMDGPU_imax:
1602 case AMDGPUIntrinsic::AMDGPU_umax:
1603 case AMDGPUIntrinsic::AMDGPU_imin:
1604 case AMDGPUIntrinsic::AMDGPU_umin:
1605 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1606 KnownZero, KnownOne, DAG, Depth);
1614 case AMDGPUISD::SMAX:
1615 case AMDGPUISD::UMAX:
1616 case AMDGPUISD::SMIN:
1617 case AMDGPUISD::UMIN:
1618 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1619 KnownZero, KnownOne, DAG, Depth);
1622 case AMDGPUISD::BFE_I32:
1623 case AMDGPUISD::BFE_U32: {
1624 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1628 unsigned BitWidth = 32;
1629 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1631 KnownZero = APInt::getAllOnesValue(BitWidth);
1632 KnownOne = APInt::getNullValue(BitWidth);
1636 // FIXME: This could do a lot more. If offset is 0, should be the same as
1637 // sign_extend_inreg implementation, but that involves duplicating it.
1638 if (Opc == AMDGPUISD::BFE_I32)
1639 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1641 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1648 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1650 const SelectionDAG &DAG,
1651 unsigned Depth) const {
1652 switch (Op.getOpcode()) {
1653 case AMDGPUISD::BFE_I32: {
1654 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1658 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1659 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1660 if (!Offset || !Offset->isNullValue())
1663 // TODO: Could probably figure something out with non-0 offsets.
1664 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1665 return std::max(SignBits, Op0SignBits);
1668 case AMDGPUISD::BFE_U32: {
1669 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1670 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;