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 // 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 // Initialize target lowering borrowed from AMDIL
114 // We need to custom lower some of the intrinsics
115 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
117 // Library functions. These default to Expand, but we have instructions
119 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
120 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
121 setOperationAction(ISD::FPOW, MVT::f32, Legal);
122 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
123 setOperationAction(ISD::FABS, MVT::f32, Legal);
124 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
125 setOperationAction(ISD::FRINT, MVT::f32, Legal);
126 setOperationAction(ISD::FROUND, MVT::f32, Legal);
127 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
129 // The hardware supports 32-bit ROTR, but not ROTL.
130 setOperationAction(ISD::ROTL, MVT::i32, Expand);
131 setOperationAction(ISD::ROTL, MVT::i64, Expand);
132 setOperationAction(ISD::ROTR, MVT::i64, Expand);
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::v4f32, Promote);
143 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
145 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
146 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
148 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
149 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
151 setOperationAction(ISD::STORE, MVT::f64, Promote);
152 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
154 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
155 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
157 // Custom lowering of vector stores is required for local address space
159 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
160 // XXX: Native v2i32 local address space stores are possible, but not
161 // currently implemented.
162 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
164 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
165 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
166 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
168 // XXX: This can be change to Custom, once ExpandVectorStores can
169 // handle 64-bit stores.
170 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
172 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
173 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
174 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
175 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
176 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
179 setOperationAction(ISD::LOAD, MVT::f32, Promote);
180 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
182 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
183 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
185 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
186 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
188 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
189 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
191 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
192 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
194 setOperationAction(ISD::LOAD, MVT::f64, Promote);
195 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
197 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
198 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
200 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
201 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
202 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
203 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
204 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
205 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
206 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
207 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
208 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
209 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
211 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
212 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
213 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
214 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
216 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
217 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
218 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
220 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
221 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
224 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
226 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
228 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
230 setOperationAction(ISD::MUL, MVT::i64, Expand);
231 setOperationAction(ISD::SUB, MVT::i64, Expand);
233 setOperationAction(ISD::UDIV, MVT::i32, Expand);
234 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
235 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
236 setOperationAction(ISD::UREM, MVT::i32, Expand);
238 if (!Subtarget->hasBFI()) {
239 // fcopysign can be done in a single instruction with BFI.
240 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
241 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
244 if (!Subtarget->hasBCNT(32))
245 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
247 if (!Subtarget->hasBCNT(64))
248 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
250 MVT VTs[] = { MVT::i32, MVT::i64 };
252 setOperationAction(ISD::CTTZ, VT, Expand);
253 setOperationAction(ISD::CTLZ, VT, Expand);
256 static const MVT::SimpleValueType IntTypes[] = {
257 MVT::v2i32, MVT::v4i32
260 for (MVT VT : IntTypes) {
261 // Expand the following operations for the current type by default.
262 setOperationAction(ISD::ADD, VT, Expand);
263 setOperationAction(ISD::AND, VT, Expand);
264 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
265 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
266 setOperationAction(ISD::MUL, VT, Expand);
267 setOperationAction(ISD::OR, VT, Expand);
268 setOperationAction(ISD::SHL, VT, Expand);
269 setOperationAction(ISD::SRA, VT, Expand);
270 setOperationAction(ISD::SRL, VT, Expand);
271 setOperationAction(ISD::ROTL, VT, Expand);
272 setOperationAction(ISD::ROTR, VT, Expand);
273 setOperationAction(ISD::SUB, VT, Expand);
274 setOperationAction(ISD::UDIV, VT, Expand);
275 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
276 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
277 setOperationAction(ISD::UREM, VT, Expand);
278 setOperationAction(ISD::SELECT, VT, Expand);
279 setOperationAction(ISD::VSELECT, VT, Expand);
280 setOperationAction(ISD::XOR, VT, Expand);
281 setOperationAction(ISD::BSWAP, VT, Expand);
282 setOperationAction(ISD::CTPOP, VT, Expand);
283 setOperationAction(ISD::CTTZ, VT, Expand);
284 setOperationAction(ISD::CTLZ, VT, Expand);
287 static const MVT::SimpleValueType FloatTypes[] = {
288 MVT::v2f32, MVT::v4f32
291 for (MVT VT : FloatTypes) {
292 setOperationAction(ISD::FABS, VT, Expand);
293 setOperationAction(ISD::FADD, VT, Expand);
294 setOperationAction(ISD::FCOS, VT, Expand);
295 setOperationAction(ISD::FDIV, VT, Expand);
296 setOperationAction(ISD::FPOW, VT, Expand);
297 setOperationAction(ISD::FFLOOR, VT, Expand);
298 setOperationAction(ISD::FTRUNC, VT, Expand);
299 setOperationAction(ISD::FMUL, VT, Expand);
300 setOperationAction(ISD::FRINT, VT, Expand);
301 setOperationAction(ISD::FSQRT, VT, Expand);
302 setOperationAction(ISD::FSIN, VT, Expand);
303 setOperationAction(ISD::FSUB, VT, Expand);
304 setOperationAction(ISD::FNEG, VT, Expand);
305 setOperationAction(ISD::SELECT, VT, Expand);
306 setOperationAction(ISD::VSELECT, VT, Expand);
307 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
310 setTargetDAGCombine(ISD::MUL);
311 setTargetDAGCombine(ISD::SELECT_CC);
314 //===----------------------------------------------------------------------===//
315 // Target Information
316 //===----------------------------------------------------------------------===//
318 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
322 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
324 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
327 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
328 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
330 return ((LScalarSize <= CastScalarSize) ||
331 (CastScalarSize >= 32) ||
335 //===---------------------------------------------------------------------===//
337 //===---------------------------------------------------------------------===//
339 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
340 assert(VT.isFloatingPoint());
341 return VT == MVT::f32;
344 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
345 assert(VT.isFloatingPoint());
346 return VT == MVT::f32;
349 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
350 // Truncate is just accessing a subregister.
351 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
354 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
355 // Truncate is just accessing a subregister.
356 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
357 (Dest->getPrimitiveSizeInBits() % 32 == 0);
360 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
361 const DataLayout *DL = getDataLayout();
362 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
363 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
365 return SrcSize == 32 && DestSize == 64;
368 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
369 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
370 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
371 // this will enable reducing 64-bit operations the 32-bit, which is always
373 return Src == MVT::i32 && Dest == MVT::i64;
376 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
377 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
378 // limited number of native 64-bit operations. Shrinking an operation to fit
379 // in a single 32-bit register should always be helpful. As currently used,
380 // this is much less general than the name suggests, and is only used in
381 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
382 // not profitable, and may actually be harmful.
383 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
386 //===---------------------------------------------------------------------===//
387 // TargetLowering Callbacks
388 //===---------------------------------------------------------------------===//
390 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
391 const SmallVectorImpl<ISD::InputArg> &Ins) const {
393 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
396 SDValue AMDGPUTargetLowering::LowerReturn(
398 CallingConv::ID CallConv,
400 const SmallVectorImpl<ISD::OutputArg> &Outs,
401 const SmallVectorImpl<SDValue> &OutVals,
402 SDLoc DL, SelectionDAG &DAG) const {
403 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
406 //===---------------------------------------------------------------------===//
407 // Target specific lowering
408 //===---------------------------------------------------------------------===//
410 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
411 SmallVectorImpl<SDValue> &InVals) const {
412 SDValue Callee = CLI.Callee;
413 SelectionDAG &DAG = CLI.DAG;
415 const Function &Fn = *DAG.getMachineFunction().getFunction();
417 StringRef FuncName("<unknown>");
419 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
420 FuncName = G->getSymbol();
421 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
422 FuncName = G->getGlobal()->getName();
424 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
425 DAG.getContext()->diagnose(NoCalls);
429 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
431 switch (Op.getOpcode()) {
433 Op.getNode()->dump();
434 llvm_unreachable("Custom lowering code for this"
435 "instruction is not implemented yet!");
437 // AMDIL DAG lowering
438 case ISD::SDIV: return LowerSDIV(Op, DAG);
439 case ISD::SREM: return LowerSREM(Op, DAG);
440 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
441 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
442 // AMDGPU DAG lowering
443 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
444 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
445 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
446 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
447 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
448 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
453 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
454 SmallVectorImpl<SDValue> &Results,
455 SelectionDAG &DAG) const {
456 switch (N->getOpcode()) {
457 case ISD::SIGN_EXTEND_INREG:
458 // Different parts of legalization seem to interpret which type of
459 // sign_extend_inreg is the one to check for custom lowering. The extended
460 // from type is what really matters, but some places check for custom
461 // lowering of the result type. This results in trying to use
462 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
463 // nothing here and let the illegal result integer be handled normally.
466 SDValue Op = SDValue(N, 0);
468 EVT VT = Op.getValueType();
469 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
470 N->getOperand(0), N->getOperand(1));
471 Results.push_back(UDIVREM);
475 SDValue Op = SDValue(N, 0);
477 EVT VT = Op.getValueType();
478 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
479 N->getOperand(0), N->getOperand(1));
480 Results.push_back(UDIVREM.getValue(1));
484 SDValue Op = SDValue(N, 0);
486 EVT VT = Op.getValueType();
487 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
489 SDValue one = DAG.getConstant(1, HalfVT);
490 SDValue zero = DAG.getConstant(0, HalfVT);
493 SDValue LHS = N->getOperand(0);
494 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
495 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
497 SDValue RHS = N->getOperand(1);
498 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
499 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
501 // Get Speculative values
502 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
503 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
505 SDValue REM_Hi = zero;
506 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
508 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
509 SDValue DIV_Lo = zero;
511 const unsigned halfBitWidth = HalfVT.getSizeInBits();
513 for (unsigned i = 0; i < halfBitWidth; ++i) {
514 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
515 // Get Value of high bit
517 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
518 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
520 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
521 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
524 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
525 DAG.getConstant(halfBitWidth - 1, HalfVT));
526 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
527 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
529 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
530 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
533 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
535 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
536 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
538 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
542 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
544 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
545 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
546 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
549 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
550 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
551 Results.push_back(DIV);
552 Results.push_back(REM);
560 // FIXME: This implements accesses to initialized globals in the constant
561 // address space by copying them to private and accessing that. It does not
562 // properly handle illegal types or vectors. The private vector loads are not
563 // scalarized, and the illegal scalars hit an assertion. This technique will not
564 // work well with large initializers, and this should eventually be
565 // removed. Initialized globals should be placed into a data section that the
566 // runtime will load into a buffer before the kernel is executed. Uses of the
567 // global need to be replaced with a pointer loaded from an implicit kernel
568 // argument into this buffer holding the copy of the data, which will remove the
569 // need for any of this.
570 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
571 const GlobalValue *GV,
572 const SDValue &InitPtr,
574 SelectionDAG &DAG) const {
575 const DataLayout *TD = getTargetMachine().getDataLayout();
577 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
578 EVT VT = EVT::getEVT(CI->getType());
579 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
580 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
581 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
582 TD->getPrefTypeAlignment(CI->getType()));
585 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
586 EVT VT = EVT::getEVT(CFP->getType());
587 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
588 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
589 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
590 TD->getPrefTypeAlignment(CFP->getType()));
593 Type *InitTy = Init->getType();
594 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
595 const StructLayout *SL = TD->getStructLayout(ST);
597 EVT PtrVT = InitPtr.getValueType();
598 SmallVector<SDValue, 8> Chains;
600 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
601 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
602 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
604 Constant *Elt = Init->getAggregateElement(I);
605 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
608 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
611 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
612 EVT PtrVT = InitPtr.getValueType();
614 unsigned NumElements;
615 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
616 NumElements = AT->getNumElements();
617 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
618 NumElements = VT->getNumElements();
620 llvm_unreachable("Unexpected type");
622 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
623 SmallVector<SDValue, 8> Chains;
624 for (unsigned i = 0; i < NumElements; ++i) {
625 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
626 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
628 Constant *Elt = Init->getAggregateElement(i);
629 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
632 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
636 llvm_unreachable("Unhandled constant initializer");
639 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
641 SelectionDAG &DAG) const {
643 const DataLayout *TD = getTargetMachine().getDataLayout();
644 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
645 const GlobalValue *GV = G->getGlobal();
647 switch (G->getAddressSpace()) {
648 default: llvm_unreachable("Global Address lowering not implemented for this "
650 case AMDGPUAS::LOCAL_ADDRESS: {
651 // XXX: What does the value of G->getOffset() mean?
652 assert(G->getOffset() == 0 &&
653 "Do not know what to do with an non-zero offset");
656 if (MFI->LocalMemoryObjects.count(GV) == 0) {
657 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
658 Offset = MFI->LDSSize;
659 MFI->LocalMemoryObjects[GV] = Offset;
660 // XXX: Account for alignment?
661 MFI->LDSSize += Size;
663 Offset = MFI->LocalMemoryObjects[GV];
666 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
668 case AMDGPUAS::CONSTANT_ADDRESS: {
669 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
670 Type *EltType = GV->getType()->getElementType();
671 unsigned Size = TD->getTypeAllocSize(EltType);
672 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
674 const GlobalVariable *Var = cast<GlobalVariable>(GV);
675 const Constant *Init = Var->getInitializer();
676 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
677 SDValue InitPtr = DAG.getFrameIndex(FI,
678 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
679 SmallVector<SDNode*, 8> WorkList;
681 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
682 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
683 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
685 WorkList.push_back(*I);
687 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
688 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
689 E = WorkList.end(); I != E; ++I) {
690 SmallVector<SDValue, 8> Ops;
691 Ops.push_back(Chain);
692 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
693 Ops.push_back((*I)->getOperand(i));
695 DAG.UpdateNodeOperands(*I, Ops);
697 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
698 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
703 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
704 SelectionDAG &DAG) const {
705 SmallVector<SDValue, 8> Args;
706 SDValue A = Op.getOperand(0);
707 SDValue B = Op.getOperand(1);
709 DAG.ExtractVectorElements(A, Args);
710 DAG.ExtractVectorElements(B, Args);
712 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
715 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
716 SelectionDAG &DAG) const {
718 SmallVector<SDValue, 8> Args;
719 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
720 EVT VT = Op.getValueType();
721 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
722 VT.getVectorNumElements());
724 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
727 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
728 SelectionDAG &DAG) const {
730 MachineFunction &MF = DAG.getMachineFunction();
731 const AMDGPUFrameLowering *TFL =
732 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
734 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
736 unsigned FrameIndex = FIN->getIndex();
737 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
738 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
742 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
743 SelectionDAG &DAG) const {
744 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
746 EVT VT = Op.getValueType();
748 switch (IntrinsicID) {
750 case AMDGPUIntrinsic::AMDGPU_abs:
751 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
752 return LowerIntrinsicIABS(Op, DAG);
753 case AMDGPUIntrinsic::AMDGPU_lrp:
754 return LowerIntrinsicLRP(Op, DAG);
755 case AMDGPUIntrinsic::AMDGPU_fract:
756 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
757 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
759 case AMDGPUIntrinsic::AMDGPU_clamp:
760 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
761 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
762 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
764 case AMDGPUIntrinsic::AMDGPU_imax:
765 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
767 case AMDGPUIntrinsic::AMDGPU_umax:
768 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
770 case AMDGPUIntrinsic::AMDGPU_imin:
771 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
773 case AMDGPUIntrinsic::AMDGPU_umin:
774 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
777 case AMDGPUIntrinsic::AMDGPU_umul24:
778 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
779 Op.getOperand(1), Op.getOperand(2));
781 case AMDGPUIntrinsic::AMDGPU_imul24:
782 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
783 Op.getOperand(1), Op.getOperand(2));
785 case AMDGPUIntrinsic::AMDGPU_umad24:
786 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
787 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
789 case AMDGPUIntrinsic::AMDGPU_imad24:
790 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
791 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
793 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
794 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
796 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
797 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
799 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
800 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
802 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
803 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
805 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
806 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
811 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
812 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
817 case AMDGPUIntrinsic::AMDGPU_bfi:
818 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
823 case AMDGPUIntrinsic::AMDGPU_bfm:
824 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
828 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
829 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
831 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
832 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
836 ///IABS(a) = SMAX(sub(0, a), a)
837 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
838 SelectionDAG &DAG) const {
840 EVT VT = Op.getValueType();
841 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
844 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
847 /// Linear Interpolation
848 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
849 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
850 SelectionDAG &DAG) const {
852 EVT VT = Op.getValueType();
853 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
854 DAG.getConstantFP(1.0f, MVT::f32),
856 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
858 return DAG.getNode(ISD::FADD, DL, VT,
859 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
863 /// \brief Generate Min/Max node
864 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
865 SelectionDAG &DAG) const {
867 EVT VT = N->getValueType(0);
869 SDValue LHS = N->getOperand(0);
870 SDValue RHS = N->getOperand(1);
871 SDValue True = N->getOperand(2);
872 SDValue False = N->getOperand(3);
873 SDValue CC = N->getOperand(4);
875 if (VT != MVT::f32 ||
876 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
880 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
894 llvm_unreachable("Operation should already be optimised!");
901 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
902 return DAG.getNode(Opc, DL, VT, LHS, RHS);
910 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
911 return DAG.getNode(Opc, DL, VT, LHS, RHS);
913 case ISD::SETCC_INVALID:
914 llvm_unreachable("Invalid setcc condcode!");
919 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
920 SelectionDAG &DAG) const {
921 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
922 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
923 EVT EltVT = Op.getValueType().getVectorElementType();
924 EVT PtrVT = Load->getBasePtr().getValueType();
925 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
926 SmallVector<SDValue, 8> Loads;
929 for (unsigned i = 0, e = NumElts; i != e; ++i) {
930 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
931 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
932 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
933 Load->getChain(), Ptr,
934 MachinePointerInfo(Load->getMemOperand()->getValue()),
935 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
936 Load->getAlignment()));
938 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
941 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
942 SelectionDAG &DAG) const {
943 StoreSDNode *Store = cast<StoreSDNode>(Op);
944 EVT MemVT = Store->getMemoryVT();
945 unsigned MemBits = MemVT.getSizeInBits();
947 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
948 // truncating store into an i32 store.
949 // XXX: We could also handle optimize other vector bitwidths.
950 if (!MemVT.isVector() || MemBits > 32) {
955 SDValue Value = Store->getValue();
956 EVT VT = Value.getValueType();
957 EVT ElemVT = VT.getVectorElementType();
958 SDValue Ptr = Store->getBasePtr();
959 EVT MemEltVT = MemVT.getVectorElementType();
960 unsigned MemEltBits = MemEltVT.getSizeInBits();
961 unsigned MemNumElements = MemVT.getVectorNumElements();
962 unsigned PackedSize = MemVT.getStoreSizeInBits();
963 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
965 assert(Value.getValueType().getScalarSizeInBits() >= 32);
968 for (unsigned i = 0; i < MemNumElements; ++i) {
969 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
970 DAG.getConstant(i, MVT::i32));
971 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
972 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
974 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
975 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
980 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
984 if (PackedSize < 32) {
985 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
986 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
987 Store->getMemOperand()->getPointerInfo(),
989 Store->isNonTemporal(), Store->isVolatile(),
990 Store->getAlignment());
993 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
994 Store->getMemOperand()->getPointerInfo(),
995 Store->isVolatile(), Store->isNonTemporal(),
996 Store->getAlignment());
999 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1000 SelectionDAG &DAG) const {
1001 StoreSDNode *Store = cast<StoreSDNode>(Op);
1002 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1003 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1004 EVT PtrVT = Store->getBasePtr().getValueType();
1005 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1008 SmallVector<SDValue, 8> Chains;
1010 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1011 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1012 Store->getValue(), DAG.getConstant(i, MVT::i32));
1013 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1014 Store->getBasePtr(),
1015 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1017 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1018 MachinePointerInfo(Store->getMemOperand()->getValue()),
1019 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1020 Store->getAlignment()));
1022 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1025 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1027 LoadSDNode *Load = cast<LoadSDNode>(Op);
1028 ISD::LoadExtType ExtType = Load->getExtensionType();
1029 EVT VT = Op.getValueType();
1030 EVT MemVT = Load->getMemoryVT();
1032 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1033 // We can do the extload to 32-bits, and then need to separately extend to
1036 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1040 Load->getMemOperand());
1041 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1044 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1045 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1046 // FIXME: Copied from PPC
1047 // First, load into 32 bits, then truncate to 1 bit.
1049 SDValue Chain = Load->getChain();
1050 SDValue BasePtr = Load->getBasePtr();
1051 MachineMemOperand *MMO = Load->getMemOperand();
1053 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1054 BasePtr, MVT::i8, MMO);
1055 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1058 // Lower loads constant address space global variable loads
1059 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1060 isa<GlobalVariable>(
1061 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1063 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1064 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1065 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1066 DAG.getConstant(2, MVT::i32));
1067 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1068 Load->getChain(), Ptr,
1069 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1072 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1073 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1077 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1078 DAG.getConstant(2, MVT::i32));
1079 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1080 Load->getChain(), Ptr,
1081 DAG.getTargetConstant(0, MVT::i32),
1083 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1085 DAG.getConstant(0x3, MVT::i32));
1086 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1087 DAG.getConstant(3, MVT::i32));
1089 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1091 EVT MemEltVT = MemVT.getScalarType();
1092 if (ExtType == ISD::SEXTLOAD) {
1093 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1094 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1097 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1100 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1102 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1103 if (Result.getNode()) {
1107 StoreSDNode *Store = cast<StoreSDNode>(Op);
1108 SDValue Chain = Store->getChain();
1109 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1110 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1111 Store->getValue().getValueType().isVector()) {
1112 return SplitVectorStore(Op, DAG);
1115 EVT MemVT = Store->getMemoryVT();
1116 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1117 MemVT.bitsLT(MVT::i32)) {
1119 if (Store->getMemoryVT() == MVT::i8) {
1121 } else if (Store->getMemoryVT() == MVT::i16) {
1124 SDValue BasePtr = Store->getBasePtr();
1125 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1126 DAG.getConstant(2, MVT::i32));
1127 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1128 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1130 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1131 DAG.getConstant(0x3, MVT::i32));
1133 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1134 DAG.getConstant(3, MVT::i32));
1136 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1139 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1141 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1142 MaskedValue, ShiftAmt);
1144 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1146 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1147 DAG.getConstant(0xffffffff, MVT::i32));
1148 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1150 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1151 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1152 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1157 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1158 SelectionDAG &DAG) const {
1160 EVT VT = Op.getValueType();
1162 SDValue Num = Op.getOperand(0);
1163 SDValue Den = Op.getOperand(1);
1165 // RCP = URECIP(Den) = 2^32 / Den + e
1166 // e is rounding error.
1167 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1169 // RCP_LO = umulo(RCP, Den) */
1170 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1172 // RCP_HI = mulhu (RCP, Den) */
1173 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1175 // NEG_RCP_LO = -RCP_LO
1176 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1179 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1180 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1183 // Calculate the rounding error from the URECIP instruction
1184 // E = mulhu(ABS_RCP_LO, RCP)
1185 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1187 // RCP_A_E = RCP + E
1188 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1190 // RCP_S_E = RCP - E
1191 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1193 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1194 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1197 // Quotient = mulhu(Tmp0, Num)
1198 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1200 // Num_S_Remainder = Quotient * Den
1201 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1203 // Remainder = Num - Num_S_Remainder
1204 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1206 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1207 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1208 DAG.getConstant(-1, VT),
1209 DAG.getConstant(0, VT),
1211 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1212 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1214 DAG.getConstant(-1, VT),
1215 DAG.getConstant(0, VT),
1217 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1218 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1221 // Calculate Division result:
1223 // Quotient_A_One = Quotient + 1
1224 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1225 DAG.getConstant(1, VT));
1227 // Quotient_S_One = Quotient - 1
1228 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1229 DAG.getConstant(1, VT));
1231 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1232 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1233 Quotient, Quotient_A_One, ISD::SETEQ);
1235 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1236 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1237 Quotient_S_One, Div, ISD::SETEQ);
1239 // Calculate Rem result:
1241 // Remainder_S_Den = Remainder - Den
1242 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1244 // Remainder_A_Den = Remainder + Den
1245 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1247 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1248 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1249 Remainder, Remainder_S_Den, ISD::SETEQ);
1251 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1252 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1253 Remainder_A_Den, Rem, ISD::SETEQ);
1258 return DAG.getMergeValues(Ops, DL);
1261 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1262 SelectionDAG &DAG) const {
1263 SDValue S0 = Op.getOperand(0);
1265 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1268 // f32 uint_to_fp i64
1269 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1270 DAG.getConstant(0, MVT::i32));
1271 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1272 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1273 DAG.getConstant(1, MVT::i32));
1274 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1275 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1276 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1277 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1280 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1282 SelectionDAG &DAG) const {
1283 MVT VT = Op.getSimpleValueType();
1285 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1286 // Shift left by 'Shift' bits.
1287 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1288 // Signed shift Right by 'Shift' bits.
1289 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1292 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1293 SelectionDAG &DAG) const {
1294 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1295 MVT VT = Op.getSimpleValueType();
1296 MVT ScalarVT = VT.getScalarType();
1301 SDValue Src = Op.getOperand(0);
1304 // TODO: Don't scalarize on Evergreen?
1305 unsigned NElts = VT.getVectorNumElements();
1306 SmallVector<SDValue, 8> Args;
1307 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1309 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1310 for (unsigned I = 0; I < NElts; ++I)
1311 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1313 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1316 //===----------------------------------------------------------------------===//
1317 // Custom DAG optimizations
1318 //===----------------------------------------------------------------------===//
1320 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1321 APInt KnownZero, KnownOne;
1322 EVT VT = Op.getValueType();
1323 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1325 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1328 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1329 EVT VT = Op.getValueType();
1331 // In order for this to be a signed 24-bit value, bit 23, must
1333 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1334 // as unsigned 24-bit values.
1335 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1338 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1340 SelectionDAG &DAG = DCI.DAG;
1341 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1342 EVT VT = Op.getValueType();
1344 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1345 APInt KnownZero, KnownOne;
1346 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1347 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1348 DCI.CommitTargetLoweringOpt(TLO);
1351 template <typename IntTy>
1352 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1353 uint32_t Offset, uint32_t Width) {
1354 if (Width + Offset < 32) {
1355 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1356 return DAG.getConstant(Result, MVT::i32);
1359 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1362 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1363 DAGCombinerInfo &DCI) const {
1364 SelectionDAG &DAG = DCI.DAG;
1367 switch(N->getOpcode()) {
1370 EVT VT = N->getValueType(0);
1371 SDValue N0 = N->getOperand(0);
1372 SDValue N1 = N->getOperand(1);
1375 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1376 if (VT.isVector() || VT.getSizeInBits() > 32)
1379 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1380 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1381 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1382 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1383 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1384 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1385 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1386 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1391 // We need to use sext even for MUL_U24, because MUL_U24 is used
1392 // for signed multiply of 8 and 16-bit types.
1393 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1397 case AMDGPUISD::MUL_I24:
1398 case AMDGPUISD::MUL_U24: {
1399 SDValue N0 = N->getOperand(0);
1400 SDValue N1 = N->getOperand(1);
1401 simplifyI24(N0, DCI);
1402 simplifyI24(N1, DCI);
1405 case ISD::SELECT_CC: {
1406 return CombineMinMax(N, DAG);
1408 case AMDGPUISD::BFE_I32:
1409 case AMDGPUISD::BFE_U32: {
1410 assert(!N->getValueType(0).isVector() &&
1411 "Vector handling of BFE not implemented");
1412 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1416 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1418 return DAG.getConstant(0, MVT::i32);
1420 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1424 SDValue BitsFrom = N->getOperand(0);
1425 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1427 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1429 if (OffsetVal == 0) {
1430 // This is already sign / zero extended, so try to fold away extra BFEs.
1431 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1433 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1434 if (OpSignBits >= SignBits)
1437 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1439 // This is a sign_extend_inreg. Replace it to take advantage of existing
1440 // DAG Combines. If not eliminated, we will match back to BFE during
1443 // TODO: The sext_inreg of extended types ends, although we can could
1444 // handle them in a single BFE.
1445 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1446 DAG.getValueType(SmallVT));
1449 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1452 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1454 return constantFoldBFE<int32_t>(DAG,
1455 Val->getSExtValue(),
1460 return constantFoldBFE<uint32_t>(DAG,
1461 Val->getZExtValue(),
1466 APInt Demanded = APInt::getBitsSet(32,
1468 OffsetVal + WidthVal);
1470 if ((OffsetVal + WidthVal) >= 32) {
1471 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1472 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1473 BitsFrom, ShiftVal);
1476 APInt KnownZero, KnownOne;
1477 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1478 !DCI.isBeforeLegalizeOps());
1479 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1480 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1481 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1482 DCI.CommitTargetLoweringOpt(TLO);
1491 //===----------------------------------------------------------------------===//
1493 //===----------------------------------------------------------------------===//
1495 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1498 const SmallVectorImpl<ISD::InputArg> &Ins,
1499 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1501 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1502 if (Ins[i].ArgVT == Ins[i].VT) {
1503 OrigIns.push_back(Ins[i]);
1508 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1509 // Vector has been split into scalars.
1510 VT = Ins[i].ArgVT.getVectorElementType();
1511 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1512 Ins[i].ArgVT.getVectorElementType() !=
1513 Ins[i].VT.getVectorElementType()) {
1514 // Vector elements have been promoted
1517 // Vector has been spilt into smaller vectors.
1521 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1522 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1523 OrigIns.push_back(Arg);
1527 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1528 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1529 return CFP->isExactlyValue(1.0);
1531 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1532 return C->isAllOnesValue();
1537 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1538 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1539 return CFP->getValueAPF().isZero();
1541 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1542 return C->isNullValue();
1547 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1548 const TargetRegisterClass *RC,
1549 unsigned Reg, EVT VT) const {
1550 MachineFunction &MF = DAG.getMachineFunction();
1551 MachineRegisterInfo &MRI = MF.getRegInfo();
1552 unsigned VirtualRegister;
1553 if (!MRI.isLiveIn(Reg)) {
1554 VirtualRegister = MRI.createVirtualRegister(RC);
1555 MRI.addLiveIn(Reg, VirtualRegister);
1557 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1559 return DAG.getRegister(VirtualRegister, VT);
1562 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1564 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1566 default: return nullptr;
1568 NODE_NAME_CASE(CALL);
1569 NODE_NAME_CASE(UMUL);
1570 NODE_NAME_CASE(DIV_INF);
1571 NODE_NAME_CASE(RET_FLAG);
1572 NODE_NAME_CASE(BRANCH_COND);
1575 NODE_NAME_CASE(DWORDADDR)
1576 NODE_NAME_CASE(FRACT)
1577 NODE_NAME_CASE(CLAMP)
1578 NODE_NAME_CASE(FMAX)
1579 NODE_NAME_CASE(SMAX)
1580 NODE_NAME_CASE(UMAX)
1581 NODE_NAME_CASE(FMIN)
1582 NODE_NAME_CASE(SMIN)
1583 NODE_NAME_CASE(UMIN)
1584 NODE_NAME_CASE(BFE_U32)
1585 NODE_NAME_CASE(BFE_I32)
1588 NODE_NAME_CASE(MUL_U24)
1589 NODE_NAME_CASE(MUL_I24)
1590 NODE_NAME_CASE(MAD_U24)
1591 NODE_NAME_CASE(MAD_I24)
1592 NODE_NAME_CASE(URECIP)
1593 NODE_NAME_CASE(DOT4)
1594 NODE_NAME_CASE(EXPORT)
1595 NODE_NAME_CASE(CONST_ADDRESS)
1596 NODE_NAME_CASE(REGISTER_LOAD)
1597 NODE_NAME_CASE(REGISTER_STORE)
1598 NODE_NAME_CASE(LOAD_CONSTANT)
1599 NODE_NAME_CASE(LOAD_INPUT)
1600 NODE_NAME_CASE(SAMPLE)
1601 NODE_NAME_CASE(SAMPLEB)
1602 NODE_NAME_CASE(SAMPLED)
1603 NODE_NAME_CASE(SAMPLEL)
1604 NODE_NAME_CASE(CVT_F32_UBYTE0)
1605 NODE_NAME_CASE(CVT_F32_UBYTE1)
1606 NODE_NAME_CASE(CVT_F32_UBYTE2)
1607 NODE_NAME_CASE(CVT_F32_UBYTE3)
1608 NODE_NAME_CASE(STORE_MSKOR)
1609 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1613 static void computeKnownBitsForMinMax(const SDValue Op0,
1617 const SelectionDAG &DAG,
1619 APInt Op0Zero, Op0One;
1620 APInt Op1Zero, Op1One;
1621 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1622 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1624 KnownZero = Op0Zero & Op1Zero;
1625 KnownOne = Op0One & Op1One;
1628 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1632 const SelectionDAG &DAG,
1633 unsigned Depth) const {
1635 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1639 unsigned Opc = Op.getOpcode();
1644 case ISD::INTRINSIC_WO_CHAIN: {
1645 // FIXME: The intrinsic should just use the node.
1646 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1647 case AMDGPUIntrinsic::AMDGPU_imax:
1648 case AMDGPUIntrinsic::AMDGPU_umax:
1649 case AMDGPUIntrinsic::AMDGPU_imin:
1650 case AMDGPUIntrinsic::AMDGPU_umin:
1651 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1652 KnownZero, KnownOne, DAG, Depth);
1660 case AMDGPUISD::SMAX:
1661 case AMDGPUISD::UMAX:
1662 case AMDGPUISD::SMIN:
1663 case AMDGPUISD::UMIN:
1664 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1665 KnownZero, KnownOne, DAG, Depth);
1668 case AMDGPUISD::BFE_I32:
1669 case AMDGPUISD::BFE_U32: {
1670 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1674 unsigned BitWidth = 32;
1675 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1677 KnownZero = APInt::getAllOnesValue(BitWidth);
1678 KnownOne = APInt::getNullValue(BitWidth);
1682 // FIXME: This could do a lot more. If offset is 0, should be the same as
1683 // sign_extend_inreg implementation, but that involves duplicating it.
1684 if (Opc == AMDGPUISD::BFE_I32)
1685 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1687 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1694 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1696 const SelectionDAG &DAG,
1697 unsigned Depth) const {
1698 switch (Op.getOpcode()) {
1699 case AMDGPUISD::BFE_I32: {
1700 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1704 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1705 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1706 if (!Offset || !Offset->isNullValue())
1709 // TODO: Could probably figure something out with non-0 offsets.
1710 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1711 return std::max(SignBits, Op0SignBits);
1714 case AMDGPUISD::BFE_U32: {
1715 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1716 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;