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 ROTR, but not ROTL
130 setOperationAction(ISD::ROTL, MVT::i32, Expand);
132 // Lower floating point store/load to integer store/load to reduce the number
133 // of patterns in tablegen.
134 setOperationAction(ISD::STORE, MVT::f32, Promote);
135 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
137 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
138 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
140 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
141 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
143 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
144 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
146 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
147 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
149 setOperationAction(ISD::STORE, MVT::f64, Promote);
150 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
152 setOperationAction(ISD::STORE, MVT::v2f64, Promote);
153 AddPromotedToType(ISD::STORE, MVT::v2f64, MVT::v2i64);
155 // Custom lowering of vector stores is required for local address space
157 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
158 // XXX: Native v2i32 local address space stores are possible, but not
159 // currently implemented.
160 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
162 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
163 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
164 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
166 // XXX: This can be change to Custom, once ExpandVectorStores can
167 // handle 64-bit stores.
168 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
170 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
171 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
172 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
173 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
174 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
177 setOperationAction(ISD::LOAD, MVT::f32, Promote);
178 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
180 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
181 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
183 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
184 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
186 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
187 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
189 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
190 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
192 setOperationAction(ISD::LOAD, MVT::f64, Promote);
193 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
195 setOperationAction(ISD::LOAD, MVT::v2f64, Promote);
196 AddPromotedToType(ISD::LOAD, MVT::v2f64, MVT::v2i64);
198 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
199 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
200 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
201 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
202 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
203 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
204 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
205 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
206 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
207 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
209 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
210 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
211 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
212 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
213 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
214 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
215 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
216 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
217 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
218 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
219 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
222 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
224 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
226 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
228 setOperationAction(ISD::MUL, MVT::i64, Expand);
229 setOperationAction(ISD::SUB, MVT::i64, Expand);
231 setOperationAction(ISD::UDIV, MVT::i32, Expand);
232 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
233 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
234 setOperationAction(ISD::UREM, MVT::i32, Expand);
236 if (!Subtarget->hasBFI()) {
237 // fcopysign can be done in a single instruction with BFI.
238 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
239 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
242 if (!Subtarget->hasBCNT(32))
243 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
245 if (!Subtarget->hasBCNT(64))
246 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
248 MVT VTs[] = { MVT::i32, MVT::i64 };
250 setOperationAction(ISD::CTTZ, VT, Expand);
251 setOperationAction(ISD::CTLZ, VT, Expand);
254 static const MVT::SimpleValueType IntTypes[] = {
255 MVT::v2i32, MVT::v4i32
258 for (MVT VT : IntTypes) {
259 // Expand the following operations for the current type by default.
260 setOperationAction(ISD::ADD, VT, Expand);
261 setOperationAction(ISD::AND, VT, Expand);
262 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
263 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
264 setOperationAction(ISD::MUL, VT, Expand);
265 setOperationAction(ISD::OR, VT, Expand);
266 setOperationAction(ISD::SHL, VT, Expand);
267 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
268 setOperationAction(ISD::SRL, VT, Expand);
269 setOperationAction(ISD::SRA, VT, Expand);
270 setOperationAction(ISD::SUB, VT, Expand);
271 setOperationAction(ISD::UDIV, VT, Expand);
272 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
273 setOperationAction(ISD::UREM, VT, Expand);
274 setOperationAction(ISD::SELECT, VT, Expand);
275 setOperationAction(ISD::VSELECT, VT, Expand);
276 setOperationAction(ISD::XOR, VT, Expand);
277 setOperationAction(ISD::BSWAP, VT, Expand);
278 setOperationAction(ISD::CTPOP, VT, Expand);
279 setOperationAction(ISD::CTTZ, VT, Expand);
280 setOperationAction(ISD::CTLZ, VT, Expand);
283 static const MVT::SimpleValueType FloatTypes[] = {
284 MVT::v2f32, MVT::v4f32
287 for (MVT VT : FloatTypes) {
288 setOperationAction(ISD::FABS, VT, Expand);
289 setOperationAction(ISD::FADD, VT, Expand);
290 setOperationAction(ISD::FCOS, VT, Expand);
291 setOperationAction(ISD::FDIV, VT, Expand);
292 setOperationAction(ISD::FPOW, VT, Expand);
293 setOperationAction(ISD::FFLOOR, VT, Expand);
294 setOperationAction(ISD::FTRUNC, VT, Expand);
295 setOperationAction(ISD::FMUL, VT, Expand);
296 setOperationAction(ISD::FRINT, VT, Expand);
297 setOperationAction(ISD::FSQRT, VT, Expand);
298 setOperationAction(ISD::FSIN, VT, Expand);
299 setOperationAction(ISD::FSUB, VT, Expand);
300 setOperationAction(ISD::FNEG, VT, Expand);
301 setOperationAction(ISD::SELECT, VT, Expand);
302 setOperationAction(ISD::VSELECT, VT, Expand);
303 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
306 setTargetDAGCombine(ISD::MUL);
307 setTargetDAGCombine(ISD::SELECT_CC);
310 //===----------------------------------------------------------------------===//
311 // Target Information
312 //===----------------------------------------------------------------------===//
314 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
318 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
320 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
323 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
324 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
326 return ((LScalarSize <= CastScalarSize) ||
327 (CastScalarSize >= 32) ||
331 //===---------------------------------------------------------------------===//
333 //===---------------------------------------------------------------------===//
335 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
336 assert(VT.isFloatingPoint());
337 return VT == MVT::f32;
340 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
341 assert(VT.isFloatingPoint());
342 return VT == MVT::f32;
345 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
346 // Truncate is just accessing a subregister.
347 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
350 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
351 // Truncate is just accessing a subregister.
352 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
353 (Dest->getPrimitiveSizeInBits() % 32 == 0);
356 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
357 const DataLayout *DL = getDataLayout();
358 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
359 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
361 return SrcSize == 32 && DestSize == 64;
364 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
365 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
366 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
367 // this will enable reducing 64-bit operations the 32-bit, which is always
369 return Src == MVT::i32 && Dest == MVT::i64;
372 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
373 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
374 // limited number of native 64-bit operations. Shrinking an operation to fit
375 // in a single 32-bit register should always be helpful. As currently used,
376 // this is much less general than the name suggests, and is only used in
377 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
378 // not profitable, and may actually be harmful.
379 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
382 //===---------------------------------------------------------------------===//
383 // TargetLowering Callbacks
384 //===---------------------------------------------------------------------===//
386 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
387 const SmallVectorImpl<ISD::InputArg> &Ins) const {
389 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
392 SDValue AMDGPUTargetLowering::LowerReturn(
394 CallingConv::ID CallConv,
396 const SmallVectorImpl<ISD::OutputArg> &Outs,
397 const SmallVectorImpl<SDValue> &OutVals,
398 SDLoc DL, SelectionDAG &DAG) const {
399 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
402 //===---------------------------------------------------------------------===//
403 // Target specific lowering
404 //===---------------------------------------------------------------------===//
406 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
407 SmallVectorImpl<SDValue> &InVals) const {
408 SDValue Callee = CLI.Callee;
409 SelectionDAG &DAG = CLI.DAG;
411 const Function &Fn = *DAG.getMachineFunction().getFunction();
413 StringRef FuncName("<unknown>");
415 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
416 FuncName = G->getSymbol();
417 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
418 FuncName = G->getGlobal()->getName();
420 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
421 DAG.getContext()->diagnose(NoCalls);
425 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
427 switch (Op.getOpcode()) {
429 Op.getNode()->dump();
430 llvm_unreachable("Custom lowering code for this"
431 "instruction is not implemented yet!");
433 // AMDIL DAG lowering
434 case ISD::SDIV: return LowerSDIV(Op, DAG);
435 case ISD::SREM: return LowerSREM(Op, DAG);
436 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
437 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
438 // AMDGPU DAG lowering
439 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
440 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
441 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
442 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
443 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
444 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
449 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
450 SmallVectorImpl<SDValue> &Results,
451 SelectionDAG &DAG) const {
452 switch (N->getOpcode()) {
453 case ISD::SIGN_EXTEND_INREG:
454 // Different parts of legalization seem to interpret which type of
455 // sign_extend_inreg is the one to check for custom lowering. The extended
456 // from type is what really matters, but some places check for custom
457 // lowering of the result type. This results in trying to use
458 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
459 // nothing here and let the illegal result integer be handled normally.
462 SDValue Op = SDValue(N, 0);
464 EVT VT = Op.getValueType();
465 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
466 N->getOperand(0), N->getOperand(1));
467 Results.push_back(UDIVREM);
471 SDValue Op = SDValue(N, 0);
473 EVT VT = Op.getValueType();
474 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
475 N->getOperand(0), N->getOperand(1));
476 Results.push_back(UDIVREM.getValue(1));
480 SDValue Op = SDValue(N, 0);
482 EVT VT = Op.getValueType();
483 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
485 SDValue one = DAG.getConstant(1, HalfVT);
486 SDValue zero = DAG.getConstant(0, HalfVT);
489 SDValue LHS = N->getOperand(0);
490 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
491 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
493 SDValue RHS = N->getOperand(1);
494 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
495 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
497 // Get Speculative values
498 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
499 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
501 SDValue REM_Hi = zero;
502 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
504 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
505 SDValue DIV_Lo = zero;
507 const unsigned halfBitWidth = HalfVT.getSizeInBits();
509 for (unsigned i = 0; i < halfBitWidth; ++i) {
510 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
511 // Get Value of high bit
513 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
514 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
516 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
517 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
520 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
521 DAG.getConstant(halfBitWidth - 1, HalfVT));
522 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
523 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
525 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
526 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
529 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
531 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
532 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
534 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
538 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
540 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
541 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
542 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
545 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
546 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
547 Results.push_back(DIV);
548 Results.push_back(REM);
556 // FIXME: This implements accesses to initialized globals in the constant
557 // address space by copying them to private and accessing that. It does not
558 // properly handle illegal types or vectors. The private vector loads are not
559 // scalarized, and the illegal scalars hit an assertion. This technique will not
560 // work well with large initializers, and this should eventually be
561 // removed. Initialized globals should be placed into a data section that the
562 // runtime will load into a buffer before the kernel is executed. Uses of the
563 // global need to be replaced with a pointer loaded from an implicit kernel
564 // argument into this buffer holding the copy of the data, which will remove the
565 // need for any of this.
566 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
567 const GlobalValue *GV,
568 const SDValue &InitPtr,
570 SelectionDAG &DAG) const {
571 const DataLayout *TD = getTargetMachine().getDataLayout();
573 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
574 EVT VT = EVT::getEVT(CI->getType());
575 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
576 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
577 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
578 TD->getPrefTypeAlignment(CI->getType()));
581 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
582 EVT VT = EVT::getEVT(CFP->getType());
583 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
584 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
585 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
586 TD->getPrefTypeAlignment(CFP->getType()));
589 Type *InitTy = Init->getType();
590 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
591 const StructLayout *SL = TD->getStructLayout(ST);
593 EVT PtrVT = InitPtr.getValueType();
594 SmallVector<SDValue, 8> Chains;
596 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
597 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), 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);
607 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
608 EVT PtrVT = InitPtr.getValueType();
610 unsigned NumElements;
611 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
612 NumElements = AT->getNumElements();
613 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
614 NumElements = VT->getNumElements();
616 llvm_unreachable("Unexpected type");
618 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
619 SmallVector<SDValue, 8> Chains;
620 for (unsigned i = 0; i < NumElements; ++i) {
621 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
622 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
624 Constant *Elt = Init->getAggregateElement(i);
625 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
628 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
632 llvm_unreachable("Unhandled constant initializer");
635 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
637 SelectionDAG &DAG) const {
639 const DataLayout *TD = getTargetMachine().getDataLayout();
640 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
641 const GlobalValue *GV = G->getGlobal();
643 switch (G->getAddressSpace()) {
644 default: llvm_unreachable("Global Address lowering not implemented for this "
646 case AMDGPUAS::LOCAL_ADDRESS: {
647 // XXX: What does the value of G->getOffset() mean?
648 assert(G->getOffset() == 0 &&
649 "Do not know what to do with an non-zero offset");
652 if (MFI->LocalMemoryObjects.count(GV) == 0) {
653 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
654 Offset = MFI->LDSSize;
655 MFI->LocalMemoryObjects[GV] = Offset;
656 // XXX: Account for alignment?
657 MFI->LDSSize += Size;
659 Offset = MFI->LocalMemoryObjects[GV];
662 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
664 case AMDGPUAS::CONSTANT_ADDRESS: {
665 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
666 Type *EltType = GV->getType()->getElementType();
667 unsigned Size = TD->getTypeAllocSize(EltType);
668 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
670 const GlobalVariable *Var = cast<GlobalVariable>(GV);
671 const Constant *Init = Var->getInitializer();
672 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
673 SDValue InitPtr = DAG.getFrameIndex(FI,
674 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
675 SmallVector<SDNode*, 8> WorkList;
677 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
678 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
679 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
681 WorkList.push_back(*I);
683 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
684 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
685 E = WorkList.end(); I != E; ++I) {
686 SmallVector<SDValue, 8> Ops;
687 Ops.push_back(Chain);
688 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
689 Ops.push_back((*I)->getOperand(i));
691 DAG.UpdateNodeOperands(*I, Ops);
693 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
694 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
699 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
700 SelectionDAG &DAG) const {
701 SmallVector<SDValue, 8> Args;
702 SDValue A = Op.getOperand(0);
703 SDValue B = Op.getOperand(1);
705 DAG.ExtractVectorElements(A, Args);
706 DAG.ExtractVectorElements(B, Args);
708 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
711 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
712 SelectionDAG &DAG) const {
714 SmallVector<SDValue, 8> Args;
715 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
716 EVT VT = Op.getValueType();
717 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
718 VT.getVectorNumElements());
720 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
723 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
724 SelectionDAG &DAG) const {
726 MachineFunction &MF = DAG.getMachineFunction();
727 const AMDGPUFrameLowering *TFL =
728 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
730 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
732 unsigned FrameIndex = FIN->getIndex();
733 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
734 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
738 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
739 SelectionDAG &DAG) const {
740 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
742 EVT VT = Op.getValueType();
744 switch (IntrinsicID) {
746 case AMDGPUIntrinsic::AMDIL_abs:
747 return LowerIntrinsicIABS(Op, DAG);
748 case AMDGPUIntrinsic::AMDIL_exp:
749 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
750 case AMDGPUIntrinsic::AMDGPU_lrp:
751 return LowerIntrinsicLRP(Op, DAG);
752 case AMDGPUIntrinsic::AMDIL_fraction:
753 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
754 case AMDGPUIntrinsic::AMDIL_max:
755 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
757 case AMDGPUIntrinsic::AMDGPU_imax:
758 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
760 case AMDGPUIntrinsic::AMDGPU_umax:
761 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
763 case AMDGPUIntrinsic::AMDIL_min:
764 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
766 case AMDGPUIntrinsic::AMDGPU_imin:
767 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
769 case AMDGPUIntrinsic::AMDGPU_umin:
770 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
773 case AMDGPUIntrinsic::AMDGPU_umul24:
774 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
775 Op.getOperand(1), Op.getOperand(2));
777 case AMDGPUIntrinsic::AMDGPU_imul24:
778 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
779 Op.getOperand(1), Op.getOperand(2));
781 case AMDGPUIntrinsic::AMDGPU_umad24:
782 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
783 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
785 case AMDGPUIntrinsic::AMDGPU_imad24:
786 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
787 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
789 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
790 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
792 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
793 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
795 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
796 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
798 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
799 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
801 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
802 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
807 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
808 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
813 case AMDGPUIntrinsic::AMDGPU_bfi:
814 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
819 case AMDGPUIntrinsic::AMDGPU_bfm:
820 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
824 case AMDGPUIntrinsic::AMDIL_round_nearest:
825 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
829 ///IABS(a) = SMAX(sub(0, a), a)
830 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
831 SelectionDAG &DAG) const {
833 EVT VT = Op.getValueType();
834 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
837 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
840 /// Linear Interpolation
841 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
842 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
843 SelectionDAG &DAG) const {
845 EVT VT = Op.getValueType();
846 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
847 DAG.getConstantFP(1.0f, MVT::f32),
849 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
851 return DAG.getNode(ISD::FADD, DL, VT,
852 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
856 /// \brief Generate Min/Max node
857 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
858 SelectionDAG &DAG) const {
860 EVT VT = N->getValueType(0);
862 SDValue LHS = N->getOperand(0);
863 SDValue RHS = N->getOperand(1);
864 SDValue True = N->getOperand(2);
865 SDValue False = N->getOperand(3);
866 SDValue CC = N->getOperand(4);
868 if (VT != MVT::f32 ||
869 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
873 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
887 llvm_unreachable("Operation should already be optimised!");
894 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
895 return DAG.getNode(Opc, DL, VT, LHS, RHS);
903 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
904 return DAG.getNode(Opc, DL, VT, LHS, RHS);
906 case ISD::SETCC_INVALID:
907 llvm_unreachable("Invalid setcc condcode!");
912 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
913 SelectionDAG &DAG) const {
914 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
915 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
916 EVT EltVT = Op.getValueType().getVectorElementType();
917 EVT PtrVT = Load->getBasePtr().getValueType();
918 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
919 SmallVector<SDValue, 8> Loads;
922 for (unsigned i = 0, e = NumElts; i != e; ++i) {
923 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
924 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
925 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
926 Load->getChain(), Ptr,
927 MachinePointerInfo(Load->getMemOperand()->getValue()),
928 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
929 Load->getAlignment()));
931 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
934 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
935 SelectionDAG &DAG) const {
936 StoreSDNode *Store = cast<StoreSDNode>(Op);
937 EVT MemVT = Store->getMemoryVT();
938 unsigned MemBits = MemVT.getSizeInBits();
940 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
941 // truncating store into an i32 store.
942 // XXX: We could also handle optimize other vector bitwidths.
943 if (!MemVT.isVector() || MemBits > 32) {
948 SDValue Value = Store->getValue();
949 EVT VT = Value.getValueType();
950 EVT ElemVT = VT.getVectorElementType();
951 SDValue Ptr = Store->getBasePtr();
952 EVT MemEltVT = MemVT.getVectorElementType();
953 unsigned MemEltBits = MemEltVT.getSizeInBits();
954 unsigned MemNumElements = MemVT.getVectorNumElements();
955 unsigned PackedSize = MemVT.getStoreSizeInBits();
956 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
958 assert(Value.getValueType().getScalarSizeInBits() >= 32);
961 for (unsigned i = 0; i < MemNumElements; ++i) {
962 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
963 DAG.getConstant(i, MVT::i32));
964 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
965 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
967 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
968 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
973 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
977 if (PackedSize < 32) {
978 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
979 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
980 Store->getMemOperand()->getPointerInfo(),
982 Store->isNonTemporal(), Store->isVolatile(),
983 Store->getAlignment());
986 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
987 Store->getMemOperand()->getPointerInfo(),
988 Store->isVolatile(), Store->isNonTemporal(),
989 Store->getAlignment());
992 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
993 SelectionDAG &DAG) const {
994 StoreSDNode *Store = cast<StoreSDNode>(Op);
995 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
996 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
997 EVT PtrVT = Store->getBasePtr().getValueType();
998 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1001 SmallVector<SDValue, 8> Chains;
1003 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1004 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1005 Store->getValue(), DAG.getConstant(i, MVT::i32));
1006 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1007 Store->getBasePtr(),
1008 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1010 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1011 MachinePointerInfo(Store->getMemOperand()->getValue()),
1012 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1013 Store->getAlignment()));
1015 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1018 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1020 LoadSDNode *Load = cast<LoadSDNode>(Op);
1021 ISD::LoadExtType ExtType = Load->getExtensionType();
1022 EVT VT = Op.getValueType();
1023 EVT MemVT = Load->getMemoryVT();
1025 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1026 // We can do the extload to 32-bits, and then need to separately extend to
1029 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1033 Load->getMemOperand());
1034 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1037 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1038 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1039 // FIXME: Copied from PPC
1040 // First, load into 32 bits, then truncate to 1 bit.
1042 SDValue Chain = Load->getChain();
1043 SDValue BasePtr = Load->getBasePtr();
1044 MachineMemOperand *MMO = Load->getMemOperand();
1046 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1047 BasePtr, MVT::i8, MMO);
1048 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1051 // Lower loads constant address space global variable loads
1052 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1053 isa<GlobalVariable>(
1054 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1056 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1057 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1058 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1059 DAG.getConstant(2, MVT::i32));
1060 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1061 Load->getChain(), Ptr,
1062 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1065 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1066 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1070 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1071 DAG.getConstant(2, MVT::i32));
1072 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1073 Load->getChain(), Ptr,
1074 DAG.getTargetConstant(0, MVT::i32),
1076 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1078 DAG.getConstant(0x3, MVT::i32));
1079 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1080 DAG.getConstant(3, MVT::i32));
1082 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1084 EVT MemEltVT = MemVT.getScalarType();
1085 if (ExtType == ISD::SEXTLOAD) {
1086 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1087 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1090 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1093 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1095 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1096 if (Result.getNode()) {
1100 StoreSDNode *Store = cast<StoreSDNode>(Op);
1101 SDValue Chain = Store->getChain();
1102 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1103 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1104 Store->getValue().getValueType().isVector()) {
1105 return SplitVectorStore(Op, DAG);
1108 EVT MemVT = Store->getMemoryVT();
1109 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1110 MemVT.bitsLT(MVT::i32)) {
1112 if (Store->getMemoryVT() == MVT::i8) {
1114 } else if (Store->getMemoryVT() == MVT::i16) {
1117 SDValue BasePtr = Store->getBasePtr();
1118 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1119 DAG.getConstant(2, MVT::i32));
1120 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1121 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1123 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1124 DAG.getConstant(0x3, MVT::i32));
1126 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1127 DAG.getConstant(3, MVT::i32));
1129 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1132 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1134 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1135 MaskedValue, ShiftAmt);
1137 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1139 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1140 DAG.getConstant(0xffffffff, MVT::i32));
1141 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1143 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1144 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1145 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1150 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1151 SelectionDAG &DAG) const {
1153 EVT VT = Op.getValueType();
1155 SDValue Num = Op.getOperand(0);
1156 SDValue Den = Op.getOperand(1);
1158 // RCP = URECIP(Den) = 2^32 / Den + e
1159 // e is rounding error.
1160 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1162 // RCP_LO = umulo(RCP, Den) */
1163 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1165 // RCP_HI = mulhu (RCP, Den) */
1166 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1168 // NEG_RCP_LO = -RCP_LO
1169 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1172 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1173 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1176 // Calculate the rounding error from the URECIP instruction
1177 // E = mulhu(ABS_RCP_LO, RCP)
1178 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1180 // RCP_A_E = RCP + E
1181 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1183 // RCP_S_E = RCP - E
1184 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1186 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1187 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1190 // Quotient = mulhu(Tmp0, Num)
1191 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1193 // Num_S_Remainder = Quotient * Den
1194 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1196 // Remainder = Num - Num_S_Remainder
1197 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1199 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1200 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1201 DAG.getConstant(-1, VT),
1202 DAG.getConstant(0, VT),
1204 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1205 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1207 DAG.getConstant(-1, VT),
1208 DAG.getConstant(0, VT),
1210 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1211 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1214 // Calculate Division result:
1216 // Quotient_A_One = Quotient + 1
1217 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1218 DAG.getConstant(1, VT));
1220 // Quotient_S_One = Quotient - 1
1221 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1222 DAG.getConstant(1, VT));
1224 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1225 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1226 Quotient, Quotient_A_One, ISD::SETEQ);
1228 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1229 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1230 Quotient_S_One, Div, ISD::SETEQ);
1232 // Calculate Rem result:
1234 // Remainder_S_Den = Remainder - Den
1235 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1237 // Remainder_A_Den = Remainder + Den
1238 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1240 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1241 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1242 Remainder, Remainder_S_Den, ISD::SETEQ);
1244 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1245 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1246 Remainder_A_Den, Rem, ISD::SETEQ);
1251 return DAG.getMergeValues(Ops, DL);
1254 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1255 SelectionDAG &DAG) const {
1256 SDValue S0 = Op.getOperand(0);
1258 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1261 // f32 uint_to_fp i64
1262 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1263 DAG.getConstant(0, MVT::i32));
1264 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1265 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1266 DAG.getConstant(1, MVT::i32));
1267 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1268 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1269 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1270 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1273 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1275 SelectionDAG &DAG) const {
1276 MVT VT = Op.getSimpleValueType();
1278 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1279 // Shift left by 'Shift' bits.
1280 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1281 // Signed shift Right by 'Shift' bits.
1282 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1285 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1286 SelectionDAG &DAG) const {
1287 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1288 MVT VT = Op.getSimpleValueType();
1289 MVT ScalarVT = VT.getScalarType();
1294 SDValue Src = Op.getOperand(0);
1297 // TODO: Don't scalarize on Evergreen?
1298 unsigned NElts = VT.getVectorNumElements();
1299 SmallVector<SDValue, 8> Args;
1300 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1302 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1303 for (unsigned I = 0; I < NElts; ++I)
1304 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1306 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1309 //===----------------------------------------------------------------------===//
1310 // Custom DAG optimizations
1311 //===----------------------------------------------------------------------===//
1313 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1314 APInt KnownZero, KnownOne;
1315 EVT VT = Op.getValueType();
1316 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1318 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1321 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1322 EVT VT = Op.getValueType();
1324 // In order for this to be a signed 24-bit value, bit 23, must
1326 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1327 // as unsigned 24-bit values.
1328 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1331 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1333 SelectionDAG &DAG = DCI.DAG;
1334 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1335 EVT VT = Op.getValueType();
1337 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1338 APInt KnownZero, KnownOne;
1339 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1340 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1341 DCI.CommitTargetLoweringOpt(TLO);
1344 template <typename IntTy>
1345 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1346 uint32_t Offset, uint32_t Width) {
1347 if (Width + Offset < 32) {
1348 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1349 return DAG.getConstant(Result, MVT::i32);
1352 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1355 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1356 DAGCombinerInfo &DCI) const {
1357 SelectionDAG &DAG = DCI.DAG;
1360 switch(N->getOpcode()) {
1363 EVT VT = N->getValueType(0);
1364 SDValue N0 = N->getOperand(0);
1365 SDValue N1 = N->getOperand(1);
1368 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1369 if (VT.isVector() || VT.getSizeInBits() > 32)
1372 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1373 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1374 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1375 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1376 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1377 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1378 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1379 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1384 // We need to use sext even for MUL_U24, because MUL_U24 is used
1385 // for signed multiply of 8 and 16-bit types.
1386 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1390 case AMDGPUISD::MUL_I24:
1391 case AMDGPUISD::MUL_U24: {
1392 SDValue N0 = N->getOperand(0);
1393 SDValue N1 = N->getOperand(1);
1394 simplifyI24(N0, DCI);
1395 simplifyI24(N1, DCI);
1398 case ISD::SELECT_CC: {
1399 return CombineMinMax(N, DAG);
1401 case AMDGPUISD::BFE_I32:
1402 case AMDGPUISD::BFE_U32: {
1403 assert(!N->getValueType(0).isVector() &&
1404 "Vector handling of BFE not implemented");
1405 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1409 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1411 return DAG.getConstant(0, MVT::i32);
1413 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1417 SDValue BitsFrom = N->getOperand(0);
1418 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1420 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1422 if (OffsetVal == 0) {
1423 // This is already sign / zero extended, so try to fold away extra BFEs.
1424 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1426 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1427 if (OpSignBits >= SignBits)
1430 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1432 // This is a sign_extend_inreg. Replace it to take advantage of existing
1433 // DAG Combines. If not eliminated, we will match back to BFE during
1436 // TODO: The sext_inreg of extended types ends, although we can could
1437 // handle them in a single BFE.
1438 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1439 DAG.getValueType(SmallVT));
1442 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1445 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1447 return constantFoldBFE<int32_t>(DAG,
1448 Val->getSExtValue(),
1453 return constantFoldBFE<uint32_t>(DAG,
1454 Val->getZExtValue(),
1459 APInt Demanded = APInt::getBitsSet(32,
1461 OffsetVal + WidthVal);
1463 if ((OffsetVal + WidthVal) >= 32) {
1464 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1465 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1466 BitsFrom, ShiftVal);
1469 APInt KnownZero, KnownOne;
1470 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1471 !DCI.isBeforeLegalizeOps());
1472 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1473 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1474 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1475 DCI.CommitTargetLoweringOpt(TLO);
1484 //===----------------------------------------------------------------------===//
1486 //===----------------------------------------------------------------------===//
1488 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1491 const SmallVectorImpl<ISD::InputArg> &Ins,
1492 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1494 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1495 if (Ins[i].ArgVT == Ins[i].VT) {
1496 OrigIns.push_back(Ins[i]);
1501 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1502 // Vector has been split into scalars.
1503 VT = Ins[i].ArgVT.getVectorElementType();
1504 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1505 Ins[i].ArgVT.getVectorElementType() !=
1506 Ins[i].VT.getVectorElementType()) {
1507 // Vector elements have been promoted
1510 // Vector has been spilt into smaller vectors.
1514 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1515 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1516 OrigIns.push_back(Arg);
1520 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1521 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1522 return CFP->isExactlyValue(1.0);
1524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1525 return C->isAllOnesValue();
1530 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1531 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1532 return CFP->getValueAPF().isZero();
1534 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1535 return C->isNullValue();
1540 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1541 const TargetRegisterClass *RC,
1542 unsigned Reg, EVT VT) const {
1543 MachineFunction &MF = DAG.getMachineFunction();
1544 MachineRegisterInfo &MRI = MF.getRegInfo();
1545 unsigned VirtualRegister;
1546 if (!MRI.isLiveIn(Reg)) {
1547 VirtualRegister = MRI.createVirtualRegister(RC);
1548 MRI.addLiveIn(Reg, VirtualRegister);
1550 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1552 return DAG.getRegister(VirtualRegister, VT);
1555 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1557 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1559 default: return nullptr;
1561 NODE_NAME_CASE(CALL);
1562 NODE_NAME_CASE(UMUL);
1563 NODE_NAME_CASE(DIV_INF);
1564 NODE_NAME_CASE(RET_FLAG);
1565 NODE_NAME_CASE(BRANCH_COND);
1568 NODE_NAME_CASE(DWORDADDR)
1569 NODE_NAME_CASE(FRACT)
1570 NODE_NAME_CASE(FMAX)
1571 NODE_NAME_CASE(SMAX)
1572 NODE_NAME_CASE(UMAX)
1573 NODE_NAME_CASE(FMIN)
1574 NODE_NAME_CASE(SMIN)
1575 NODE_NAME_CASE(UMIN)
1576 NODE_NAME_CASE(BFE_U32)
1577 NODE_NAME_CASE(BFE_I32)
1580 NODE_NAME_CASE(MUL_U24)
1581 NODE_NAME_CASE(MUL_I24)
1582 NODE_NAME_CASE(MAD_U24)
1583 NODE_NAME_CASE(MAD_I24)
1584 NODE_NAME_CASE(URECIP)
1585 NODE_NAME_CASE(DOT4)
1586 NODE_NAME_CASE(EXPORT)
1587 NODE_NAME_CASE(CONST_ADDRESS)
1588 NODE_NAME_CASE(REGISTER_LOAD)
1589 NODE_NAME_CASE(REGISTER_STORE)
1590 NODE_NAME_CASE(LOAD_CONSTANT)
1591 NODE_NAME_CASE(LOAD_INPUT)
1592 NODE_NAME_CASE(SAMPLE)
1593 NODE_NAME_CASE(SAMPLEB)
1594 NODE_NAME_CASE(SAMPLED)
1595 NODE_NAME_CASE(SAMPLEL)
1596 NODE_NAME_CASE(CVT_F32_UBYTE0)
1597 NODE_NAME_CASE(CVT_F32_UBYTE1)
1598 NODE_NAME_CASE(CVT_F32_UBYTE2)
1599 NODE_NAME_CASE(CVT_F32_UBYTE3)
1600 NODE_NAME_CASE(STORE_MSKOR)
1601 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1605 static void computeKnownBitsForMinMax(const SDValue Op0,
1609 const SelectionDAG &DAG,
1611 APInt Op0Zero, Op0One;
1612 APInt Op1Zero, Op1One;
1613 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1614 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1616 KnownZero = Op0Zero & Op1Zero;
1617 KnownOne = Op0One & Op1One;
1620 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1624 const SelectionDAG &DAG,
1625 unsigned Depth) const {
1627 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1631 unsigned Opc = Op.getOpcode();
1636 case ISD::INTRINSIC_WO_CHAIN: {
1637 // FIXME: The intrinsic should just use the node.
1638 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1639 case AMDGPUIntrinsic::AMDGPU_imax:
1640 case AMDGPUIntrinsic::AMDGPU_umax:
1641 case AMDGPUIntrinsic::AMDGPU_imin:
1642 case AMDGPUIntrinsic::AMDGPU_umin:
1643 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1644 KnownZero, KnownOne, DAG, Depth);
1652 case AMDGPUISD::SMAX:
1653 case AMDGPUISD::UMAX:
1654 case AMDGPUISD::SMIN:
1655 case AMDGPUISD::UMIN:
1656 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1657 KnownZero, KnownOne, DAG, Depth);
1660 case AMDGPUISD::BFE_I32:
1661 case AMDGPUISD::BFE_U32: {
1662 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1666 unsigned BitWidth = 32;
1667 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1669 KnownZero = APInt::getAllOnesValue(BitWidth);
1670 KnownOne = APInt::getNullValue(BitWidth);
1674 // FIXME: This could do a lot more. If offset is 0, should be the same as
1675 // sign_extend_inreg implementation, but that involves duplicating it.
1676 if (Opc == AMDGPUISD::BFE_I32)
1677 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1679 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1686 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1688 const SelectionDAG &DAG,
1689 unsigned Depth) const {
1690 switch (Op.getOpcode()) {
1691 case AMDGPUISD::BFE_I32: {
1692 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1696 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1697 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1698 if (!Offset || !Offset->isNullValue())
1701 // TODO: Could probably figure something out with non-0 offsets.
1702 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1703 return std::max(SignBits, Op0SignBits);
1706 case AMDGPUISD::BFE_U32: {
1707 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1708 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;