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 const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
245 for (MVT VT : ScalarIntVTs) {
246 // GPU does not have divrem function for signed or unsigned.
247 setOperationAction(ISD::SDIVREM, VT, Expand);
249 // GPU does not have [S|U]MUL_LOHI functions as a single instruction.
250 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
251 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
253 setOperationAction(ISD::BSWAP, VT, Expand);
254 setOperationAction(ISD::CTTZ, VT, Expand);
255 setOperationAction(ISD::CTLZ, VT, Expand);
258 if (!Subtarget->hasBCNT(32))
259 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
261 if (!Subtarget->hasBCNT(64))
262 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
265 static const MVT::SimpleValueType VectorIntTypes[] = {
266 MVT::v2i32, MVT::v4i32
269 for (MVT VT : VectorIntTypes) {
270 // Expand the following operations for the current type by default.
271 setOperationAction(ISD::ADD, VT, Expand);
272 setOperationAction(ISD::AND, VT, Expand);
273 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
274 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
275 setOperationAction(ISD::MUL, VT, Expand);
276 setOperationAction(ISD::OR, VT, Expand);
277 setOperationAction(ISD::SHL, VT, Expand);
278 setOperationAction(ISD::SRA, VT, Expand);
279 setOperationAction(ISD::SRL, VT, Expand);
280 setOperationAction(ISD::ROTL, VT, Expand);
281 setOperationAction(ISD::ROTR, VT, Expand);
282 setOperationAction(ISD::SUB, VT, Expand);
283 setOperationAction(ISD::UDIV, VT, Expand);
284 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
285 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
286 // TODO: Implement custom UREM / SREM routines.
287 setOperationAction(ISD::SREM, VT, Expand);
288 setOperationAction(ISD::UREM, VT, Expand);
289 setOperationAction(ISD::SDIVREM, VT, Expand);
290 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
291 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
292 setOperationAction(ISD::SELECT, VT, Expand);
293 setOperationAction(ISD::VSELECT, VT, Expand);
294 setOperationAction(ISD::XOR, VT, Expand);
295 setOperationAction(ISD::BSWAP, VT, Expand);
296 setOperationAction(ISD::CTPOP, VT, Expand);
297 setOperationAction(ISD::CTTZ, VT, Expand);
298 setOperationAction(ISD::CTLZ, VT, Expand);
301 static const MVT::SimpleValueType FloatVectorTypes[] = {
302 MVT::v2f32, MVT::v4f32
305 for (MVT VT : FloatVectorTypes) {
306 setOperationAction(ISD::FABS, VT, Expand);
307 setOperationAction(ISD::FADD, VT, Expand);
308 setOperationAction(ISD::FCOS, VT, Expand);
309 setOperationAction(ISD::FDIV, VT, Expand);
310 setOperationAction(ISD::FPOW, VT, Expand);
311 setOperationAction(ISD::FFLOOR, VT, Expand);
312 setOperationAction(ISD::FTRUNC, VT, Expand);
313 setOperationAction(ISD::FMUL, VT, Expand);
314 setOperationAction(ISD::FRINT, VT, Expand);
315 setOperationAction(ISD::FSQRT, VT, Expand);
316 setOperationAction(ISD::FSIN, VT, Expand);
317 setOperationAction(ISD::FSUB, VT, Expand);
318 setOperationAction(ISD::FNEG, VT, Expand);
319 setOperationAction(ISD::SELECT, VT, Expand);
320 setOperationAction(ISD::VSELECT, VT, Expand);
321 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
324 setTargetDAGCombine(ISD::MUL);
325 setTargetDAGCombine(ISD::SELECT_CC);
327 setSchedulingPreference(Sched::RegPressure);
328 setJumpIsExpensive(true);
330 // FIXME: Need to really handle these.
331 MaxStoresPerMemcpy = 4096;
332 MaxStoresPerMemmove = 4096;
333 MaxStoresPerMemset = 4096;
336 //===----------------------------------------------------------------------===//
337 // Target Information
338 //===----------------------------------------------------------------------===//
340 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
344 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
346 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
349 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
350 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
352 return ((LScalarSize <= CastScalarSize) ||
353 (CastScalarSize >= 32) ||
357 //===---------------------------------------------------------------------===//
359 //===---------------------------------------------------------------------===//
361 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
362 assert(VT.isFloatingPoint());
363 return VT == MVT::f32;
366 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
367 assert(VT.isFloatingPoint());
368 return VT == MVT::f32;
371 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
372 // Truncate is just accessing a subregister.
373 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
376 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
377 // Truncate is just accessing a subregister.
378 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
379 (Dest->getPrimitiveSizeInBits() % 32 == 0);
382 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
383 const DataLayout *DL = getDataLayout();
384 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
385 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
387 return SrcSize == 32 && DestSize == 64;
390 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
391 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
392 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
393 // this will enable reducing 64-bit operations the 32-bit, which is always
395 return Src == MVT::i32 && Dest == MVT::i64;
398 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
399 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
400 // limited number of native 64-bit operations. Shrinking an operation to fit
401 // in a single 32-bit register should always be helpful. As currently used,
402 // this is much less general than the name suggests, and is only used in
403 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
404 // not profitable, and may actually be harmful.
405 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
408 //===---------------------------------------------------------------------===//
409 // TargetLowering Callbacks
410 //===---------------------------------------------------------------------===//
412 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
413 const SmallVectorImpl<ISD::InputArg> &Ins) const {
415 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
418 SDValue AMDGPUTargetLowering::LowerReturn(
420 CallingConv::ID CallConv,
422 const SmallVectorImpl<ISD::OutputArg> &Outs,
423 const SmallVectorImpl<SDValue> &OutVals,
424 SDLoc DL, SelectionDAG &DAG) const {
425 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
428 //===---------------------------------------------------------------------===//
429 // Target specific lowering
430 //===---------------------------------------------------------------------===//
432 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
433 SmallVectorImpl<SDValue> &InVals) const {
434 SDValue Callee = CLI.Callee;
435 SelectionDAG &DAG = CLI.DAG;
437 const Function &Fn = *DAG.getMachineFunction().getFunction();
439 StringRef FuncName("<unknown>");
441 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
442 FuncName = G->getSymbol();
443 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
444 FuncName = G->getGlobal()->getName();
446 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
447 DAG.getContext()->diagnose(NoCalls);
451 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
453 switch (Op.getOpcode()) {
455 Op.getNode()->dump();
456 llvm_unreachable("Custom lowering code for this"
457 "instruction is not implemented yet!");
459 // AMDIL DAG lowering
460 case ISD::SDIV: return LowerSDIV(Op, DAG);
461 case ISD::SREM: return LowerSREM(Op, DAG);
462 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
463 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
464 // AMDGPU DAG lowering
465 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
466 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
467 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
468 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
469 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
470 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
475 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
476 SmallVectorImpl<SDValue> &Results,
477 SelectionDAG &DAG) const {
478 switch (N->getOpcode()) {
479 case ISD::SIGN_EXTEND_INREG:
480 // Different parts of legalization seem to interpret which type of
481 // sign_extend_inreg is the one to check for custom lowering. The extended
482 // from type is what really matters, but some places check for custom
483 // lowering of the result type. This results in trying to use
484 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
485 // nothing here and let the illegal result integer be handled normally.
488 SDValue Op = SDValue(N, 0);
490 EVT VT = Op.getValueType();
491 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
492 N->getOperand(0), N->getOperand(1));
493 Results.push_back(UDIVREM);
497 SDValue Op = SDValue(N, 0);
499 EVT VT = Op.getValueType();
500 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
501 N->getOperand(0), N->getOperand(1));
502 Results.push_back(UDIVREM.getValue(1));
506 SDValue Op = SDValue(N, 0);
508 EVT VT = Op.getValueType();
509 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
511 SDValue one = DAG.getConstant(1, HalfVT);
512 SDValue zero = DAG.getConstant(0, HalfVT);
515 SDValue LHS = N->getOperand(0);
516 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
517 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
519 SDValue RHS = N->getOperand(1);
520 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
521 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
523 // Get Speculative values
524 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
525 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
527 SDValue REM_Hi = zero;
528 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
530 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
531 SDValue DIV_Lo = zero;
533 const unsigned halfBitWidth = HalfVT.getSizeInBits();
535 for (unsigned i = 0; i < halfBitWidth; ++i) {
536 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
537 // Get Value of high bit
539 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
540 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
542 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
543 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
546 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
547 DAG.getConstant(halfBitWidth - 1, HalfVT));
548 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
549 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
551 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
552 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
555 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
557 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
558 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
560 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
564 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
566 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
567 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
568 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
571 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
572 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
573 Results.push_back(DIV);
574 Results.push_back(REM);
582 // FIXME: This implements accesses to initialized globals in the constant
583 // address space by copying them to private and accessing that. It does not
584 // properly handle illegal types or vectors. The private vector loads are not
585 // scalarized, and the illegal scalars hit an assertion. This technique will not
586 // work well with large initializers, and this should eventually be
587 // removed. Initialized globals should be placed into a data section that the
588 // runtime will load into a buffer before the kernel is executed. Uses of the
589 // global need to be replaced with a pointer loaded from an implicit kernel
590 // argument into this buffer holding the copy of the data, which will remove the
591 // need for any of this.
592 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
593 const GlobalValue *GV,
594 const SDValue &InitPtr,
596 SelectionDAG &DAG) const {
597 const DataLayout *TD = getTargetMachine().getDataLayout();
599 Type *InitTy = Init->getType();
601 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
602 EVT VT = EVT::getEVT(InitTy);
603 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
604 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
605 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
606 TD->getPrefTypeAlignment(InitTy));
609 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
610 EVT VT = EVT::getEVT(CFP->getType());
611 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
612 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
613 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
614 TD->getPrefTypeAlignment(CFP->getType()));
617 if (StructType *ST = dyn_cast<StructType>(InitTy)) {
618 const StructLayout *SL = TD->getStructLayout(ST);
620 EVT PtrVT = InitPtr.getValueType();
621 SmallVector<SDValue, 8> Chains;
623 for (unsigned I = 0, N = ST->getNumElements(); I != N; ++I) {
624 SDValue Offset = DAG.getConstant(SL->getElementOffset(I), PtrVT);
625 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
627 Constant *Elt = Init->getAggregateElement(I);
628 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
631 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
634 if (SequentialType *SeqTy = dyn_cast<SequentialType>(InitTy)) {
635 EVT PtrVT = InitPtr.getValueType();
637 unsigned NumElements;
638 if (ArrayType *AT = dyn_cast<ArrayType>(SeqTy))
639 NumElements = AT->getNumElements();
640 else if (VectorType *VT = dyn_cast<VectorType>(SeqTy))
641 NumElements = VT->getNumElements();
643 llvm_unreachable("Unexpected type");
645 unsigned EltSize = TD->getTypeAllocSize(SeqTy->getElementType());
646 SmallVector<SDValue, 8> Chains;
647 for (unsigned i = 0; i < NumElements; ++i) {
648 SDValue Offset = DAG.getConstant(i * EltSize, PtrVT);
649 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
651 Constant *Elt = Init->getAggregateElement(i);
652 Chains.push_back(LowerConstantInitializer(Elt, GV, Ptr, Chain, DAG));
655 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
658 if (isa<UndefValue>(Init)) {
659 EVT VT = EVT::getEVT(InitTy);
660 PointerType *PtrTy = PointerType::get(InitTy, AMDGPUAS::PRIVATE_ADDRESS);
661 return DAG.getStore(Chain, DL, DAG.getUNDEF(VT), InitPtr,
662 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
663 TD->getPrefTypeAlignment(InitTy));
667 llvm_unreachable("Unhandled constant initializer");
670 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
672 SelectionDAG &DAG) const {
674 const DataLayout *TD = getTargetMachine().getDataLayout();
675 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
676 const GlobalValue *GV = G->getGlobal();
678 switch (G->getAddressSpace()) {
679 default: llvm_unreachable("Global Address lowering not implemented for this "
681 case AMDGPUAS::LOCAL_ADDRESS: {
682 // XXX: What does the value of G->getOffset() mean?
683 assert(G->getOffset() == 0 &&
684 "Do not know what to do with an non-zero offset");
687 if (MFI->LocalMemoryObjects.count(GV) == 0) {
688 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
689 Offset = MFI->LDSSize;
690 MFI->LocalMemoryObjects[GV] = Offset;
691 // XXX: Account for alignment?
692 MFI->LDSSize += Size;
694 Offset = MFI->LocalMemoryObjects[GV];
697 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
699 case AMDGPUAS::CONSTANT_ADDRESS: {
700 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
701 Type *EltType = GV->getType()->getElementType();
702 unsigned Size = TD->getTypeAllocSize(EltType);
703 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
705 MVT PrivPtrVT = getPointerTy(AMDGPUAS::PRIVATE_ADDRESS);
706 MVT ConstPtrVT = getPointerTy(AMDGPUAS::CONSTANT_ADDRESS);
708 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
709 SDValue InitPtr = DAG.getFrameIndex(FI, PrivPtrVT);
711 const GlobalVariable *Var = cast<GlobalVariable>(GV);
712 if (!Var->hasInitializer()) {
713 // This has no use, but bugpoint will hit it.
714 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
717 const Constant *Init = Var->getInitializer();
718 SmallVector<SDNode*, 8> WorkList;
720 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
721 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
722 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
724 WorkList.push_back(*I);
726 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
727 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
728 E = WorkList.end(); I != E; ++I) {
729 SmallVector<SDValue, 8> Ops;
730 Ops.push_back(Chain);
731 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
732 Ops.push_back((*I)->getOperand(i));
734 DAG.UpdateNodeOperands(*I, Ops);
736 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op), ConstPtrVT);
741 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
742 SelectionDAG &DAG) const {
743 SmallVector<SDValue, 8> Args;
744 SDValue A = Op.getOperand(0);
745 SDValue B = Op.getOperand(1);
747 DAG.ExtractVectorElements(A, Args);
748 DAG.ExtractVectorElements(B, Args);
750 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
753 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
754 SelectionDAG &DAG) const {
756 SmallVector<SDValue, 8> Args;
757 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
758 EVT VT = Op.getValueType();
759 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
760 VT.getVectorNumElements());
762 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
765 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
766 SelectionDAG &DAG) const {
768 MachineFunction &MF = DAG.getMachineFunction();
769 const AMDGPUFrameLowering *TFL =
770 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
772 FrameIndexSDNode *FIN = cast<FrameIndexSDNode>(Op);
774 unsigned FrameIndex = FIN->getIndex();
775 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
776 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
780 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
781 SelectionDAG &DAG) const {
782 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
784 EVT VT = Op.getValueType();
786 switch (IntrinsicID) {
788 case AMDGPUIntrinsic::AMDGPU_abs:
789 case AMDGPUIntrinsic::AMDIL_abs: // Legacy name.
790 return LowerIntrinsicIABS(Op, DAG);
791 case AMDGPUIntrinsic::AMDGPU_lrp:
792 return LowerIntrinsicLRP(Op, DAG);
793 case AMDGPUIntrinsic::AMDGPU_fract:
794 case AMDGPUIntrinsic::AMDIL_fraction: // Legacy name.
795 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
797 case AMDGPUIntrinsic::AMDGPU_clamp:
798 case AMDGPUIntrinsic::AMDIL_clamp: // Legacy name.
799 return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
800 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
802 case AMDGPUIntrinsic::AMDGPU_imax:
803 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
805 case AMDGPUIntrinsic::AMDGPU_umax:
806 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
808 case AMDGPUIntrinsic::AMDGPU_imin:
809 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
811 case AMDGPUIntrinsic::AMDGPU_umin:
812 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
815 case AMDGPUIntrinsic::AMDGPU_umul24:
816 return DAG.getNode(AMDGPUISD::MUL_U24, DL, VT,
817 Op.getOperand(1), Op.getOperand(2));
819 case AMDGPUIntrinsic::AMDGPU_imul24:
820 return DAG.getNode(AMDGPUISD::MUL_I24, DL, VT,
821 Op.getOperand(1), Op.getOperand(2));
823 case AMDGPUIntrinsic::AMDGPU_umad24:
824 return DAG.getNode(AMDGPUISD::MAD_U24, DL, VT,
825 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
827 case AMDGPUIntrinsic::AMDGPU_imad24:
828 return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT,
829 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
831 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0:
832 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1));
834 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1:
835 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1));
837 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2:
838 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1));
840 case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3:
841 return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1));
843 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
844 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
849 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
850 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
855 case AMDGPUIntrinsic::AMDGPU_bfi:
856 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
861 case AMDGPUIntrinsic::AMDGPU_bfm:
862 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
866 case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
867 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
869 case AMDGPUIntrinsic::AMDIL_round_nearest: // Legacy name.
870 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
874 ///IABS(a) = SMAX(sub(0, a), a)
875 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
876 SelectionDAG &DAG) const {
878 EVT VT = Op.getValueType();
879 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
882 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
885 /// Linear Interpolation
886 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
887 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
888 SelectionDAG &DAG) const {
890 EVT VT = Op.getValueType();
891 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
892 DAG.getConstantFP(1.0f, MVT::f32),
894 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
896 return DAG.getNode(ISD::FADD, DL, VT,
897 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
901 /// \brief Generate Min/Max node
902 SDValue AMDGPUTargetLowering::CombineMinMax(SDNode *N,
903 SelectionDAG &DAG) const {
905 EVT VT = N->getValueType(0);
907 SDValue LHS = N->getOperand(0);
908 SDValue RHS = N->getOperand(1);
909 SDValue True = N->getOperand(2);
910 SDValue False = N->getOperand(3);
911 SDValue CC = N->getOperand(4);
913 if (VT != MVT::f32 ||
914 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
918 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
932 llvm_unreachable("Operation should already be optimised!");
939 unsigned Opc = (LHS == True) ? AMDGPUISD::FMIN : AMDGPUISD::FMAX;
940 return DAG.getNode(Opc, DL, VT, LHS, RHS);
948 unsigned Opc = (LHS == True) ? AMDGPUISD::FMAX : AMDGPUISD::FMIN;
949 return DAG.getNode(Opc, DL, VT, LHS, RHS);
951 case ISD::SETCC_INVALID:
952 llvm_unreachable("Invalid setcc condcode!");
957 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
958 SelectionDAG &DAG) const {
959 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
960 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
961 EVT EltVT = Op.getValueType().getVectorElementType();
962 EVT PtrVT = Load->getBasePtr().getValueType();
963 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
964 SmallVector<SDValue, 8> Loads;
967 for (unsigned i = 0, e = NumElts; i != e; ++i) {
968 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
969 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
970 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
971 Load->getChain(), Ptr,
972 MachinePointerInfo(Load->getMemOperand()->getValue()),
973 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
974 Load->getAlignment()));
976 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
979 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
980 SelectionDAG &DAG) const {
981 StoreSDNode *Store = cast<StoreSDNode>(Op);
982 EVT MemVT = Store->getMemoryVT();
983 unsigned MemBits = MemVT.getSizeInBits();
985 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
986 // truncating store into an i32 store.
987 // XXX: We could also handle optimize other vector bitwidths.
988 if (!MemVT.isVector() || MemBits > 32) {
993 SDValue Value = Store->getValue();
994 EVT VT = Value.getValueType();
995 EVT ElemVT = VT.getVectorElementType();
996 SDValue Ptr = Store->getBasePtr();
997 EVT MemEltVT = MemVT.getVectorElementType();
998 unsigned MemEltBits = MemEltVT.getSizeInBits();
999 unsigned MemNumElements = MemVT.getVectorNumElements();
1000 unsigned PackedSize = MemVT.getStoreSizeInBits();
1001 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
1003 assert(Value.getValueType().getScalarSizeInBits() >= 32);
1005 SDValue PackedValue;
1006 for (unsigned i = 0; i < MemNumElements; ++i) {
1007 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
1008 DAG.getConstant(i, MVT::i32));
1009 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
1010 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
1012 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
1013 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
1018 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
1022 if (PackedSize < 32) {
1023 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
1024 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
1025 Store->getMemOperand()->getPointerInfo(),
1027 Store->isNonTemporal(), Store->isVolatile(),
1028 Store->getAlignment());
1031 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
1032 Store->getMemOperand()->getPointerInfo(),
1033 Store->isVolatile(), Store->isNonTemporal(),
1034 Store->getAlignment());
1037 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
1038 SelectionDAG &DAG) const {
1039 StoreSDNode *Store = cast<StoreSDNode>(Op);
1040 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
1041 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
1042 EVT PtrVT = Store->getBasePtr().getValueType();
1043 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
1046 SmallVector<SDValue, 8> Chains;
1048 for (unsigned i = 0, e = NumElts; i != e; ++i) {
1049 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
1050 Store->getValue(), DAG.getConstant(i, MVT::i32));
1051 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
1052 Store->getBasePtr(),
1053 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
1055 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
1056 MachinePointerInfo(Store->getMemOperand()->getValue()),
1057 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
1058 Store->getAlignment()));
1060 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
1063 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1065 LoadSDNode *Load = cast<LoadSDNode>(Op);
1066 ISD::LoadExtType ExtType = Load->getExtensionType();
1067 EVT VT = Op.getValueType();
1068 EVT MemVT = Load->getMemoryVT();
1070 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
1071 // We can do the extload to 32-bits, and then need to separately extend to
1074 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
1078 Load->getMemOperand());
1079 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
1082 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
1083 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
1084 // FIXME: Copied from PPC
1085 // First, load into 32 bits, then truncate to 1 bit.
1087 SDValue Chain = Load->getChain();
1088 SDValue BasePtr = Load->getBasePtr();
1089 MachineMemOperand *MMO = Load->getMemOperand();
1091 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
1092 BasePtr, MVT::i8, MMO);
1093 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
1096 // Lower loads constant address space global variable loads
1097 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
1098 isa<GlobalVariable>(
1099 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
1101 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
1102 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
1103 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
1104 DAG.getConstant(2, MVT::i32));
1105 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1106 Load->getChain(), Ptr,
1107 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
1110 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
1111 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
1115 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
1116 DAG.getConstant(2, MVT::i32));
1117 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
1118 Load->getChain(), Ptr,
1119 DAG.getTargetConstant(0, MVT::i32),
1121 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
1123 DAG.getConstant(0x3, MVT::i32));
1124 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1125 DAG.getConstant(3, MVT::i32));
1127 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
1129 EVT MemEltVT = MemVT.getScalarType();
1130 if (ExtType == ISD::SEXTLOAD) {
1131 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
1132 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
1135 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
1138 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1140 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
1141 if (Result.getNode()) {
1145 StoreSDNode *Store = cast<StoreSDNode>(Op);
1146 SDValue Chain = Store->getChain();
1147 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
1148 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
1149 Store->getValue().getValueType().isVector()) {
1150 return SplitVectorStore(Op, DAG);
1153 EVT MemVT = Store->getMemoryVT();
1154 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1155 MemVT.bitsLT(MVT::i32)) {
1157 if (Store->getMemoryVT() == MVT::i8) {
1159 } else if (Store->getMemoryVT() == MVT::i16) {
1162 SDValue BasePtr = Store->getBasePtr();
1163 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1164 DAG.getConstant(2, MVT::i32));
1165 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1166 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1168 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1169 DAG.getConstant(0x3, MVT::i32));
1171 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1172 DAG.getConstant(3, MVT::i32));
1174 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1177 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1179 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1180 MaskedValue, ShiftAmt);
1182 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1184 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1185 DAG.getConstant(0xffffffff, MVT::i32));
1186 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1188 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1189 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1190 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1195 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1196 SelectionDAG &DAG) const {
1198 EVT VT = Op.getValueType();
1200 SDValue Num = Op.getOperand(0);
1201 SDValue Den = Op.getOperand(1);
1203 // RCP = URECIP(Den) = 2^32 / Den + e
1204 // e is rounding error.
1205 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1207 // RCP_LO = umulo(RCP, Den) */
1208 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1210 // RCP_HI = mulhu (RCP, Den) */
1211 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1213 // NEG_RCP_LO = -RCP_LO
1214 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1217 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1218 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1221 // Calculate the rounding error from the URECIP instruction
1222 // E = mulhu(ABS_RCP_LO, RCP)
1223 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1225 // RCP_A_E = RCP + E
1226 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1228 // RCP_S_E = RCP - E
1229 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1231 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1232 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1235 // Quotient = mulhu(Tmp0, Num)
1236 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1238 // Num_S_Remainder = Quotient * Den
1239 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1241 // Remainder = Num - Num_S_Remainder
1242 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1244 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1245 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1246 DAG.getConstant(-1, VT),
1247 DAG.getConstant(0, VT),
1249 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1250 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1252 DAG.getConstant(-1, VT),
1253 DAG.getConstant(0, VT),
1255 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1256 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1259 // Calculate Division result:
1261 // Quotient_A_One = Quotient + 1
1262 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1263 DAG.getConstant(1, VT));
1265 // Quotient_S_One = Quotient - 1
1266 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1267 DAG.getConstant(1, VT));
1269 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1270 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1271 Quotient, Quotient_A_One, ISD::SETEQ);
1273 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1274 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1275 Quotient_S_One, Div, ISD::SETEQ);
1277 // Calculate Rem result:
1279 // Remainder_S_Den = Remainder - Den
1280 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1282 // Remainder_A_Den = Remainder + Den
1283 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1285 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1286 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1287 Remainder, Remainder_S_Den, ISD::SETEQ);
1289 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1290 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1291 Remainder_A_Den, Rem, ISD::SETEQ);
1296 return DAG.getMergeValues(Ops, DL);
1299 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1300 SelectionDAG &DAG) const {
1301 SDValue S0 = Op.getOperand(0);
1303 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1306 // f32 uint_to_fp i64
1307 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1308 DAG.getConstant(0, MVT::i32));
1309 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1310 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1311 DAG.getConstant(1, MVT::i32));
1312 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1313 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1314 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1315 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1318 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1320 SelectionDAG &DAG) const {
1321 MVT VT = Op.getSimpleValueType();
1323 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1324 // Shift left by 'Shift' bits.
1325 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1326 // Signed shift Right by 'Shift' bits.
1327 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1330 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1331 SelectionDAG &DAG) const {
1332 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1333 MVT VT = Op.getSimpleValueType();
1334 MVT ScalarVT = VT.getScalarType();
1339 SDValue Src = Op.getOperand(0);
1342 // TODO: Don't scalarize on Evergreen?
1343 unsigned NElts = VT.getVectorNumElements();
1344 SmallVector<SDValue, 8> Args;
1345 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1347 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1348 for (unsigned I = 0; I < NElts; ++I)
1349 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1351 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1354 //===----------------------------------------------------------------------===//
1355 // Custom DAG optimizations
1356 //===----------------------------------------------------------------------===//
1358 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1359 APInt KnownZero, KnownOne;
1360 EVT VT = Op.getValueType();
1361 DAG.computeKnownBits(Op, KnownZero, KnownOne);
1363 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1366 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1367 EVT VT = Op.getValueType();
1369 // In order for this to be a signed 24-bit value, bit 23, must
1371 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1372 // as unsigned 24-bit values.
1373 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1376 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1378 SelectionDAG &DAG = DCI.DAG;
1379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1380 EVT VT = Op.getValueType();
1382 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1383 APInt KnownZero, KnownOne;
1384 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1385 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1386 DCI.CommitTargetLoweringOpt(TLO);
1389 template <typename IntTy>
1390 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
1391 uint32_t Offset, uint32_t Width) {
1392 if (Width + Offset < 32) {
1393 IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
1394 return DAG.getConstant(Result, MVT::i32);
1397 return DAG.getConstant(Src0 >> Offset, MVT::i32);
1400 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1401 DAGCombinerInfo &DCI) const {
1402 SelectionDAG &DAG = DCI.DAG;
1405 switch(N->getOpcode()) {
1408 EVT VT = N->getValueType(0);
1409 SDValue N0 = N->getOperand(0);
1410 SDValue N1 = N->getOperand(1);
1413 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1414 if (VT.isVector() || VT.getSizeInBits() > 32)
1417 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1418 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1419 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1420 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1421 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1422 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1423 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1424 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1429 // We need to use sext even for MUL_U24, because MUL_U24 is used
1430 // for signed multiply of 8 and 16-bit types.
1431 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1435 case AMDGPUISD::MUL_I24:
1436 case AMDGPUISD::MUL_U24: {
1437 SDValue N0 = N->getOperand(0);
1438 SDValue N1 = N->getOperand(1);
1439 simplifyI24(N0, DCI);
1440 simplifyI24(N1, DCI);
1443 case ISD::SELECT_CC: {
1444 return CombineMinMax(N, DAG);
1446 case AMDGPUISD::BFE_I32:
1447 case AMDGPUISD::BFE_U32: {
1448 assert(!N->getValueType(0).isVector() &&
1449 "Vector handling of BFE not implemented");
1450 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
1454 uint32_t WidthVal = Width->getZExtValue() & 0x1f;
1456 return DAG.getConstant(0, MVT::i32);
1458 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
1462 SDValue BitsFrom = N->getOperand(0);
1463 uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
1465 bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
1467 if (OffsetVal == 0) {
1468 // This is already sign / zero extended, so try to fold away extra BFEs.
1469 unsigned SignBits = Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
1471 unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
1472 if (OpSignBits >= SignBits)
1475 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), WidthVal);
1477 // This is a sign_extend_inreg. Replace it to take advantage of existing
1478 // DAG Combines. If not eliminated, we will match back to BFE during
1481 // TODO: The sext_inreg of extended types ends, although we can could
1482 // handle them in a single BFE.
1483 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, BitsFrom,
1484 DAG.getValueType(SmallVT));
1487 return DAG.getZeroExtendInReg(BitsFrom, DL, SmallVT);
1490 if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
1492 return constantFoldBFE<int32_t>(DAG,
1493 Val->getSExtValue(),
1498 return constantFoldBFE<uint32_t>(DAG,
1499 Val->getZExtValue(),
1504 APInt Demanded = APInt::getBitsSet(32,
1506 OffsetVal + WidthVal);
1508 if ((OffsetVal + WidthVal) >= 32) {
1509 SDValue ShiftVal = DAG.getConstant(OffsetVal, MVT::i32);
1510 return DAG.getNode(Signed ? ISD::SRA : ISD::SRL, DL, MVT::i32,
1511 BitsFrom, ShiftVal);
1514 APInt KnownZero, KnownOne;
1515 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
1516 !DCI.isBeforeLegalizeOps());
1517 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1518 if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
1519 TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
1520 DCI.CommitTargetLoweringOpt(TLO);
1529 //===----------------------------------------------------------------------===//
1531 //===----------------------------------------------------------------------===//
1533 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1536 const SmallVectorImpl<ISD::InputArg> &Ins,
1537 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1539 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1540 if (Ins[i].ArgVT == Ins[i].VT) {
1541 OrigIns.push_back(Ins[i]);
1546 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1547 // Vector has been split into scalars.
1548 VT = Ins[i].ArgVT.getVectorElementType();
1549 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1550 Ins[i].ArgVT.getVectorElementType() !=
1551 Ins[i].VT.getVectorElementType()) {
1552 // Vector elements have been promoted
1555 // Vector has been spilt into smaller vectors.
1559 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1560 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1561 OrigIns.push_back(Arg);
1565 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1566 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1567 return CFP->isExactlyValue(1.0);
1569 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1570 return C->isAllOnesValue();
1575 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1576 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1577 return CFP->getValueAPF().isZero();
1579 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1580 return C->isNullValue();
1585 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1586 const TargetRegisterClass *RC,
1587 unsigned Reg, EVT VT) const {
1588 MachineFunction &MF = DAG.getMachineFunction();
1589 MachineRegisterInfo &MRI = MF.getRegInfo();
1590 unsigned VirtualRegister;
1591 if (!MRI.isLiveIn(Reg)) {
1592 VirtualRegister = MRI.createVirtualRegister(RC);
1593 MRI.addLiveIn(Reg, VirtualRegister);
1595 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1597 return DAG.getRegister(VirtualRegister, VT);
1600 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1602 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1604 default: return nullptr;
1606 NODE_NAME_CASE(CALL);
1607 NODE_NAME_CASE(UMUL);
1608 NODE_NAME_CASE(DIV_INF);
1609 NODE_NAME_CASE(RET_FLAG);
1610 NODE_NAME_CASE(BRANCH_COND);
1613 NODE_NAME_CASE(DWORDADDR)
1614 NODE_NAME_CASE(FRACT)
1615 NODE_NAME_CASE(CLAMP)
1616 NODE_NAME_CASE(FMAX)
1617 NODE_NAME_CASE(SMAX)
1618 NODE_NAME_CASE(UMAX)
1619 NODE_NAME_CASE(FMIN)
1620 NODE_NAME_CASE(SMIN)
1621 NODE_NAME_CASE(UMIN)
1622 NODE_NAME_CASE(BFE_U32)
1623 NODE_NAME_CASE(BFE_I32)
1626 NODE_NAME_CASE(MUL_U24)
1627 NODE_NAME_CASE(MUL_I24)
1628 NODE_NAME_CASE(MAD_U24)
1629 NODE_NAME_CASE(MAD_I24)
1630 NODE_NAME_CASE(URECIP)
1631 NODE_NAME_CASE(DOT4)
1632 NODE_NAME_CASE(EXPORT)
1633 NODE_NAME_CASE(CONST_ADDRESS)
1634 NODE_NAME_CASE(REGISTER_LOAD)
1635 NODE_NAME_CASE(REGISTER_STORE)
1636 NODE_NAME_CASE(LOAD_CONSTANT)
1637 NODE_NAME_CASE(LOAD_INPUT)
1638 NODE_NAME_CASE(SAMPLE)
1639 NODE_NAME_CASE(SAMPLEB)
1640 NODE_NAME_CASE(SAMPLED)
1641 NODE_NAME_CASE(SAMPLEL)
1642 NODE_NAME_CASE(CVT_F32_UBYTE0)
1643 NODE_NAME_CASE(CVT_F32_UBYTE1)
1644 NODE_NAME_CASE(CVT_F32_UBYTE2)
1645 NODE_NAME_CASE(CVT_F32_UBYTE3)
1646 NODE_NAME_CASE(STORE_MSKOR)
1647 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1651 static void computeKnownBitsForMinMax(const SDValue Op0,
1655 const SelectionDAG &DAG,
1657 APInt Op0Zero, Op0One;
1658 APInt Op1Zero, Op1One;
1659 DAG.computeKnownBits(Op0, Op0Zero, Op0One, Depth);
1660 DAG.computeKnownBits(Op1, Op1Zero, Op1One, Depth);
1662 KnownZero = Op0Zero & Op1Zero;
1663 KnownOne = Op0One & Op1One;
1666 void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
1670 const SelectionDAG &DAG,
1671 unsigned Depth) const {
1673 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1677 unsigned Opc = Op.getOpcode();
1682 case ISD::INTRINSIC_WO_CHAIN: {
1683 // FIXME: The intrinsic should just use the node.
1684 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1685 case AMDGPUIntrinsic::AMDGPU_imax:
1686 case AMDGPUIntrinsic::AMDGPU_umax:
1687 case AMDGPUIntrinsic::AMDGPU_imin:
1688 case AMDGPUIntrinsic::AMDGPU_umin:
1689 computeKnownBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1690 KnownZero, KnownOne, DAG, Depth);
1698 case AMDGPUISD::SMAX:
1699 case AMDGPUISD::UMAX:
1700 case AMDGPUISD::SMIN:
1701 case AMDGPUISD::UMIN:
1702 computeKnownBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1703 KnownZero, KnownOne, DAG, Depth);
1706 case AMDGPUISD::BFE_I32:
1707 case AMDGPUISD::BFE_U32: {
1708 ConstantSDNode *CWidth = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1712 unsigned BitWidth = 32;
1713 uint32_t Width = CWidth->getZExtValue() & 0x1f;
1715 KnownZero = APInt::getAllOnesValue(BitWidth);
1716 KnownOne = APInt::getNullValue(BitWidth);
1720 // FIXME: This could do a lot more. If offset is 0, should be the same as
1721 // sign_extend_inreg implementation, but that involves duplicating it.
1722 if (Opc == AMDGPUISD::BFE_I32)
1723 KnownOne = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1725 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - Width);
1732 unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
1734 const SelectionDAG &DAG,
1735 unsigned Depth) const {
1736 switch (Op.getOpcode()) {
1737 case AMDGPUISD::BFE_I32: {
1738 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1742 unsigned SignBits = 32 - Width->getZExtValue() + 1;
1743 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(Op.getOperand(1));
1744 if (!Offset || !Offset->isNullValue())
1747 // TODO: Could probably figure something out with non-0 offsets.
1748 unsigned Op0SignBits = DAG.ComputeNumSignBits(Op.getOperand(0), Depth + 1);
1749 return std::max(SignBits, Op0SignBits);
1752 case AMDGPUISD::BFE_U32: {
1753 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
1754 return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;