1 //===-- AMDILISelDAGToDAG.cpp - A dag to dag inst selector for AMDIL ------===//
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 Defines an instruction selector for the AMDGPU target.
13 //===----------------------------------------------------------------------===//
14 #include "AMDGPUInstrInfo.h"
15 #include "AMDGPUISelLowering.h" // For AMDGPUISD
16 #include "AMDGPURegisterInfo.h"
17 #include "AMDGPUSubtarget.h"
18 #include "R600InstrInfo.h"
19 #include "SIDefines.h"
20 #include "SIISelLowering.h"
21 #include "SIMachineFunctionInfo.h"
22 #include "llvm/CodeGen/FunctionLoweringInfo.h"
23 #include "llvm/CodeGen/PseudoSourceValue.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/CodeGen/SelectionDAGISel.h"
28 #include "llvm/IR/Function.h"
32 //===----------------------------------------------------------------------===//
33 // Instruction Selector Implementation
34 //===----------------------------------------------------------------------===//
37 /// AMDGPU specific code to select AMDGPU machine instructions for
38 /// SelectionDAG operations.
39 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
40 // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
41 // make the right decision when generating code for different targets.
42 const AMDGPUSubtarget &Subtarget;
44 AMDGPUDAGToDAGISel(TargetMachine &TM);
45 virtual ~AMDGPUDAGToDAGISel();
47 SDNode *Select(SDNode *N) override;
48 const char *getPassName() const override;
49 void PostprocessISelDAG() override;
52 bool isInlineImmediate(SDNode *N) const;
53 inline SDValue getSmallIPtrImm(unsigned Imm);
54 bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs,
55 const R600InstrInfo *TII);
56 bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
57 bool FoldDotOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &);
59 // Complex pattern selectors
60 bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2);
61 bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2);
62 bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2);
64 static bool checkType(const Value *ptr, unsigned int addrspace);
65 static bool checkPrivateAddress(const MachineMemOperand *Op);
67 static bool isGlobalStore(const StoreSDNode *N);
68 static bool isFlatStore(const StoreSDNode *N);
69 static bool isPrivateStore(const StoreSDNode *N);
70 static bool isLocalStore(const StoreSDNode *N);
71 static bool isRegionStore(const StoreSDNode *N);
73 bool isCPLoad(const LoadSDNode *N) const;
74 bool isConstantLoad(const LoadSDNode *N, int cbID) const;
75 bool isGlobalLoad(const LoadSDNode *N) const;
76 bool isFlatLoad(const LoadSDNode *N) const;
77 bool isParamLoad(const LoadSDNode *N) const;
78 bool isPrivateLoad(const LoadSDNode *N) const;
79 bool isLocalLoad(const LoadSDNode *N) const;
80 bool isRegionLoad(const LoadSDNode *N) const;
82 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
83 bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
84 bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
86 bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
87 bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
88 bool isDSOffsetLegal(const SDValue &Base, unsigned Offset,
89 unsigned OffsetBits) const;
90 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
91 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
92 SDValue &Offset1) const;
93 void SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
94 SDValue &SOffset, SDValue &Offset, SDValue &Offen,
95 SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
97 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
98 SDValue &Offset) const;
99 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
100 SDValue &VAddr, SDValue &Offset,
102 bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
103 SDValue &SOffset, SDValue &ImmOffset) const;
104 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
105 SDValue &Offset, SDValue &GLC, SDValue &SLC,
107 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
108 SDValue &Offset, SDValue &GLC) const;
109 SDNode *SelectAddrSpaceCast(SDNode *N);
110 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
111 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
112 SDValue &Clamp, SDValue &Omod) const;
114 SDNode *SelectADD_SUB_I64(SDNode *N);
115 SDNode *SelectDIV_SCALE(SDNode *N);
117 // Include the pieces autogenerated from the target description.
118 #include "AMDGPUGenDAGISel.inc"
120 } // end anonymous namespace
122 /// \brief This pass converts a legalized DAG into a AMDGPU-specific
123 // DAG, ready for instruction scheduling.
124 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) {
125 return new AMDGPUDAGToDAGISel(TM);
128 AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
129 : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) {
132 AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
135 bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const {
136 const SITargetLowering *TL
137 = static_cast<const SITargetLowering *>(getTargetLowering());
138 return TL->analyzeImmediate(N) == 0;
141 /// \brief Determine the register class for \p OpNo
142 /// \returns The register class of the virtual register that will be used for
143 /// the given operand number \OpNo or NULL if the register class cannot be
145 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
146 unsigned OpNo) const {
147 if (!N->isMachineOpcode())
150 switch (N->getMachineOpcode()) {
152 const MCInstrDesc &Desc =
153 TM.getSubtargetImpl()->getInstrInfo()->get(N->getMachineOpcode());
154 unsigned OpIdx = Desc.getNumDefs() + OpNo;
155 if (OpIdx >= Desc.getNumOperands())
157 int RegClass = Desc.OpInfo[OpIdx].RegClass;
161 return TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RegClass);
163 case AMDGPU::REG_SEQUENCE: {
164 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
165 const TargetRegisterClass *SuperRC =
166 TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RCID);
168 SDValue SubRegOp = N->getOperand(OpNo + 1);
169 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
170 return TM.getSubtargetImpl()->getRegisterInfo()->getSubClassWithSubReg(
176 SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) {
177 return CurDAG->getTargetConstant(Imm, MVT::i32);
180 bool AMDGPUDAGToDAGISel::SelectADDRParam(
181 SDValue Addr, SDValue& R1, SDValue& R2) {
183 if (Addr.getOpcode() == ISD::FrameIndex) {
184 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
185 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32);
186 R2 = CurDAG->getTargetConstant(0, MVT::i32);
189 R2 = CurDAG->getTargetConstant(0, MVT::i32);
191 } else if (Addr.getOpcode() == ISD::ADD) {
192 R1 = Addr.getOperand(0);
193 R2 = Addr.getOperand(1);
196 R2 = CurDAG->getTargetConstant(0, MVT::i32);
201 bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) {
202 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
203 Addr.getOpcode() == ISD::TargetGlobalAddress) {
206 return SelectADDRParam(Addr, R1, R2);
210 bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
211 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
212 Addr.getOpcode() == ISD::TargetGlobalAddress) {
216 if (Addr.getOpcode() == ISD::FrameIndex) {
217 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
218 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
219 R2 = CurDAG->getTargetConstant(0, MVT::i64);
222 R2 = CurDAG->getTargetConstant(0, MVT::i64);
224 } else if (Addr.getOpcode() == ISD::ADD) {
225 R1 = Addr.getOperand(0);
226 R2 = Addr.getOperand(1);
229 R2 = CurDAG->getTargetConstant(0, MVT::i64);
234 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
235 unsigned int Opc = N->getOpcode();
236 if (N->isMachineOpcode()) {
238 return nullptr; // Already selected.
241 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
244 // We are selecting i64 ADD here instead of custom lower it during
245 // DAG legalization, so we can fold some i64 ADDs used for address
246 // calculation into the LOAD and STORE instructions.
249 if (N->getValueType(0) != MVT::i64 ||
250 ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
253 return SelectADD_SUB_I64(N);
255 case ISD::SCALAR_TO_VECTOR:
256 case AMDGPUISD::BUILD_VERTICAL_VECTOR:
257 case ISD::BUILD_VECTOR: {
259 const AMDGPURegisterInfo *TRI = static_cast<const AMDGPURegisterInfo *>(
260 TM.getSubtargetImpl()->getRegisterInfo());
261 const SIRegisterInfo *SIRI = static_cast<const SIRegisterInfo *>(
262 TM.getSubtargetImpl()->getRegisterInfo());
263 EVT VT = N->getValueType(0);
264 unsigned NumVectorElts = VT.getVectorNumElements();
265 EVT EltVT = VT.getVectorElementType();
266 assert(EltVT.bitsEq(MVT::i32));
267 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
269 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
271 if (!U->isMachineOpcode()) {
274 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
278 if (SIRI->isSGPRClass(RC)) {
282 switch(NumVectorElts) {
283 case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
284 AMDGPU::SReg_32RegClassID;
286 case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
287 AMDGPU::SReg_64RegClassID;
289 case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
290 AMDGPU::SReg_128RegClassID;
292 case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
293 AMDGPU::SReg_256RegClassID;
295 case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
296 AMDGPU::SReg_512RegClassID;
298 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
301 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
302 // that adds a 128 bits reg copy when going through TwoAddressInstructions
303 // pass. We want to avoid 128 bits copies as much as possible because they
304 // can't be bundled by our scheduler.
305 switch(NumVectorElts) {
306 case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
308 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
309 RegClassID = AMDGPU::R600_Reg128VerticalRegClassID;
311 RegClassID = AMDGPU::R600_Reg128RegClassID;
313 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
317 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
319 if (NumVectorElts == 1) {
320 return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT,
321 N->getOperand(0), RegClass);
324 assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
326 // 16 = Max Num Vector Elements
327 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
328 // 1 = Vector Register Class
329 SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
331 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
332 bool IsRegSeq = true;
333 unsigned NOps = N->getNumOperands();
334 for (unsigned i = 0; i < NOps; i++) {
335 // XXX: Why is this here?
336 if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
340 RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
341 RegSeqArgs[1 + (2 * i) + 1] =
342 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
345 if (NOps != NumVectorElts) {
346 // Fill in the missing undef elements if this was a scalar_to_vector.
347 assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
349 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
351 for (unsigned i = NOps; i < NumVectorElts; ++i) {
352 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
353 RegSeqArgs[1 + (2 * i) + 1] =
354 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
360 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
363 case ISD::BUILD_PAIR: {
364 SDValue RC, SubReg0, SubReg1;
365 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
368 if (N->getValueType(0) == MVT::i128) {
369 RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32);
370 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, MVT::i32);
371 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, MVT::i32);
372 } else if (N->getValueType(0) == MVT::i64) {
373 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32);
374 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
375 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
377 llvm_unreachable("Unhandled value type for BUILD_PAIR");
379 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
380 N->getOperand(1), SubReg1 };
381 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
382 SDLoc(N), N->getValueType(0), Ops);
386 case ISD::ConstantFP: {
387 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
388 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
389 N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
393 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
394 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
396 ConstantSDNode *C = cast<ConstantSDNode>(N);
397 Imm = C->getZExtValue();
400 SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
401 CurDAG->getConstant(Imm & 0xFFFFFFFF, MVT::i32));
402 SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
403 CurDAG->getConstant(Imm >> 32, MVT::i32));
404 const SDValue Ops[] = {
405 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
406 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
407 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
410 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, SDLoc(N),
411 N->getValueType(0), Ops);
414 case AMDGPUISD::REGISTER_LOAD: {
415 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
417 SDValue Addr, Offset;
419 SelectADDRIndirect(N->getOperand(1), Addr, Offset);
420 const SDValue Ops[] = {
423 CurDAG->getTargetConstant(0, MVT::i32),
426 return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, SDLoc(N),
427 CurDAG->getVTList(MVT::i32, MVT::i64, MVT::Other),
430 case AMDGPUISD::REGISTER_STORE: {
431 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
433 SDValue Addr, Offset;
434 SelectADDRIndirect(N->getOperand(2), Addr, Offset);
435 const SDValue Ops[] = {
439 CurDAG->getTargetConstant(0, MVT::i32),
442 return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, SDLoc(N),
443 CurDAG->getVTList(MVT::Other),
447 case AMDGPUISD::BFE_I32:
448 case AMDGPUISD::BFE_U32: {
449 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
452 // There is a scalar version available, but unlike the vector version which
453 // has a separate operand for the offset and width, the scalar version packs
454 // the width and offset into a single operand. Try to move to the scalar
455 // version if the offsets are constant, so that we can try to keep extended
456 // loads of kernel arguments in SGPRs.
458 // TODO: Technically we could try to pattern match scalar bitshifts of
459 // dynamic values, but it's probably not useful.
460 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
464 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
468 bool Signed = Opc == AMDGPUISD::BFE_I32;
470 // Transformation function, pack the offset and width of a BFE into
471 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
472 // source, bits [5:0] contain the offset and bits [22:16] the width.
474 uint32_t OffsetVal = Offset->getZExtValue();
475 uint32_t WidthVal = Width->getZExtValue();
477 uint32_t PackedVal = OffsetVal | WidthVal << 16;
479 SDValue PackedOffsetWidth = CurDAG->getTargetConstant(PackedVal, MVT::i32);
480 return CurDAG->getMachineNode(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
487 case AMDGPUISD::DIV_SCALE: {
488 return SelectDIV_SCALE(N);
490 case ISD::CopyToReg: {
491 const SITargetLowering& Lowering =
492 *static_cast<const SITargetLowering*>(getTargetLowering());
493 Lowering.legalizeTargetIndependentNode(N, *CurDAG);
496 case ISD::ADDRSPACECAST:
497 return SelectAddrSpaceCast(N);
500 return SelectCode(N);
504 bool AMDGPUDAGToDAGISel::checkType(const Value *Ptr, unsigned AS) {
505 assert(AS != 0 && "Use checkPrivateAddress instead.");
509 return Ptr->getType()->getPointerAddressSpace() == AS;
512 bool AMDGPUDAGToDAGISel::checkPrivateAddress(const MachineMemOperand *Op) {
513 if (Op->getPseudoValue())
516 if (PointerType *PT = dyn_cast<PointerType>(Op->getValue()->getType()))
517 return PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
522 bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
523 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
526 bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
527 const Value *MemVal = N->getMemOperand()->getValue();
528 return (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
529 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
530 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS));
533 bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
534 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
537 bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
538 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
541 bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
542 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
545 bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
546 const Value *MemVal = N->getMemOperand()->getValue();
548 return checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS);
550 return checkType(MemVal, AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
553 bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
554 if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) {
555 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
556 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
557 N->getMemoryVT().bitsLT(MVT::i32)) {
561 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
564 bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) const {
565 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::PARAM_I_ADDRESS);
568 bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) const {
569 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
572 bool AMDGPUDAGToDAGISel::isFlatLoad(const LoadSDNode *N) const {
573 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
576 bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) const {
577 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
580 bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) const {
581 MachineMemOperand *MMO = N->getMemOperand();
582 if (checkPrivateAddress(N->getMemOperand())) {
584 const PseudoSourceValue *PSV = MMO->getPseudoValue();
585 if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
593 bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
594 if (checkPrivateAddress(N->getMemOperand())) {
595 // Check to make sure we are not a constant pool load or a constant load
596 // that is marked as a private load
597 if (isCPLoad(N) || isConstantLoad(N, -1)) {
602 const Value *MemVal = N->getMemOperand()->getValue();
603 if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
604 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
605 !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
606 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
607 !checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
608 !checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
609 !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
615 const char *AMDGPUDAGToDAGISel::getPassName() const {
616 return "AMDGPU DAG->DAG Pattern Instruction Selection";
624 //===----------------------------------------------------------------------===//
626 //===----------------------------------------------------------------------===//
628 bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
630 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
631 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, true);
637 bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
638 SDValue& BaseReg, SDValue &Offset) {
639 if (!isa<ConstantSDNode>(Addr)) {
641 Offset = CurDAG->getIntPtrConstant(0, true);
647 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
649 ConstantSDNode *IMMOffset;
651 if (Addr.getOpcode() == ISD::ADD
652 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
653 && isInt<16>(IMMOffset->getZExtValue())) {
655 Base = Addr.getOperand(0);
656 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
658 // If the pointer address is constant, we can move it to the offset field.
659 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
660 && isInt<16>(IMMOffset->getZExtValue())) {
661 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
662 SDLoc(CurDAG->getEntryNode()),
663 AMDGPU::ZERO, MVT::i32);
664 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
668 // Default case, no offset
670 Offset = CurDAG->getTargetConstant(0, MVT::i32);
674 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
678 if ((C = dyn_cast<ConstantSDNode>(Addr))) {
679 Base = CurDAG->getRegister(AMDGPU::INDIRECT_BASE_ADDR, MVT::i32);
680 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
681 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
682 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
683 Base = Addr.getOperand(0);
684 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
687 Offset = CurDAG->getTargetConstant(0, MVT::i32);
693 SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
695 SDValue LHS = N->getOperand(0);
696 SDValue RHS = N->getOperand(1);
698 bool IsAdd = (N->getOpcode() == ISD::ADD);
700 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
701 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
703 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
704 DL, MVT::i32, LHS, Sub0);
705 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
706 DL, MVT::i32, LHS, Sub1);
708 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
709 DL, MVT::i32, RHS, Sub0);
710 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
711 DL, MVT::i32, RHS, Sub1);
713 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
714 SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
717 unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
718 unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
720 SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs);
721 SDValue Carry(AddLo, 1);
723 = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32,
724 SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
727 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
733 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
736 SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
738 EVT VT = N->getValueType(0);
740 assert(VT == MVT::f32 || VT == MVT::f64);
743 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
745 const SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
746 const SDValue False = CurDAG->getTargetConstant(0, MVT::i1);
748 Zero, // src0_modifiers
749 N->getOperand(0), // src0
750 Zero, // src1_modifiers
751 N->getOperand(1), // src1
752 Zero, // src2_modifiers
753 N->getOperand(2), // src2
758 return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
761 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
762 unsigned OffsetBits) const {
763 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
764 if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
765 (OffsetBits == 8 && !isUInt<8>(Offset)))
768 if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
771 // On Southern Islands instruction with a negative base value and an offset
772 // don't seem to work.
773 return CurDAG->SignBitIsZero(Base);
776 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
777 SDValue &Offset) const {
778 if (CurDAG->isBaseWithConstantOffset(Addr)) {
779 SDValue N0 = Addr.getOperand(0);
780 SDValue N1 = Addr.getOperand(1);
781 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
782 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
790 // If we have a constant address, prefer to put the constant into the
791 // offset. This can save moves to load the constant address since multiple
792 // operations can share the zero base address register, and enables merging
793 // into read2 / write2 instructions.
794 if (const ConstantSDNode *CAddr = dyn_cast<ConstantSDNode>(Addr)) {
795 if (isUInt<16>(CAddr->getZExtValue())) {
796 Base = CurDAG->getConstant(0, MVT::i32);
804 Offset = CurDAG->getTargetConstant(0, MVT::i16);
808 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
810 SDValue &Offset1) const {
811 if (CurDAG->isBaseWithConstantOffset(Addr)) {
812 SDValue N0 = Addr.getOperand(0);
813 SDValue N1 = Addr.getOperand(1);
814 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
815 unsigned DWordOffset0 = C1->getZExtValue() / 4;
816 unsigned DWordOffset1 = DWordOffset0 + 1;
818 if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
820 Offset0 = CurDAG->getTargetConstant(DWordOffset0, MVT::i8);
821 Offset1 = CurDAG->getTargetConstant(DWordOffset1, MVT::i8);
828 Offset0 = CurDAG->getTargetConstant(0, MVT::i8);
829 Offset1 = CurDAG->getTargetConstant(1, MVT::i8);
833 static SDValue wrapAddr64Rsrc(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
834 return SDValue(DAG->getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::v4i32,
838 static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
839 return isUInt<12>(Imm->getZExtValue());
842 void AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
843 SDValue &VAddr, SDValue &SOffset,
844 SDValue &Offset, SDValue &Offen,
845 SDValue &Idxen, SDValue &Addr64,
846 SDValue &GLC, SDValue &SLC,
847 SDValue &TFE) const {
850 GLC = CurDAG->getTargetConstant(0, MVT::i1);
851 SLC = CurDAG->getTargetConstant(0, MVT::i1);
852 TFE = CurDAG->getTargetConstant(0, MVT::i1);
854 Idxen = CurDAG->getTargetConstant(0, MVT::i1);
855 Offen = CurDAG->getTargetConstant(0, MVT::i1);
856 Addr64 = CurDAG->getTargetConstant(0, MVT::i1);
857 SOffset = CurDAG->getTargetConstant(0, MVT::i32);
859 if (CurDAG->isBaseWithConstantOffset(Addr)) {
860 SDValue N0 = Addr.getOperand(0);
861 SDValue N1 = Addr.getOperand(1);
862 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
864 if (isLegalMUBUFImmOffset(C1)) {
866 if (N0.getOpcode() == ISD::ADD) {
867 // (add (add N2, N3), C1) -> addr64
868 SDValue N2 = N0.getOperand(0);
869 SDValue N3 = N0.getOperand(1);
870 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
873 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
877 // (add N0, C1) -> offset
878 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
880 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
884 if (Addr.getOpcode() == ISD::ADD) {
885 // (add N0, N1) -> addr64
886 SDValue N0 = Addr.getOperand(0);
887 SDValue N1 = Addr.getOperand(1);
888 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
891 Offset = CurDAG->getTargetConstant(0, MVT::i16);
895 // default case -> offset
896 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
898 Offset = CurDAG->getTargetConstant(0, MVT::i16);
902 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
904 SDValue &Offset) const {
905 SDValue Ptr, SOffset, Offen, Idxen, Addr64, GLC, SLC, TFE;
907 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
910 ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
911 if (C->getSExtValue()) {
913 SRsrc = wrapAddr64Rsrc(CurDAG, DL, Ptr);
919 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
920 SDValue &VAddr, SDValue &Offset,
921 SDValue &SLC) const {
922 SLC = CurDAG->getTargetConstant(0, MVT::i1);
924 return SelectMUBUFAddr64(Addr, SRsrc, VAddr, Offset);
927 static SDValue buildRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr,
928 uint32_t RsrcDword1, uint64_t RsrcDword2And3) {
930 SDValue PtrLo = DAG->getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
931 SDValue PtrHi = DAG->getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
933 PtrHi = SDValue(DAG->getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
934 DAG->getConstant(RsrcDword1, MVT::i32)), 0);
936 SDValue DataLo = DAG->getTargetConstant(
937 RsrcDword2And3 & APInt::getAllOnesValue(32).getZExtValue(), MVT::i32);
938 SDValue DataHi = DAG->getTargetConstant(RsrcDword2And3 >> 32, MVT::i32);
940 const SDValue Ops[] = { PtrLo, PtrHi, DataLo, DataHi };
941 return SDValue(DAG->getMachineNode(AMDGPU::SI_BUFFER_RSRC, DL,
942 MVT::v4i32, Ops), 0);
945 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
946 /// The TID (Thread ID) is multipled by the stride value (bits [61:48]
947 /// of the resource descriptor) to create an offset, which is added to the
949 static SDValue buildScratchRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
951 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT | AMDGPU::RSRC_TID_ENABLE |
954 return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
957 bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
958 SDValue &VAddr, SDValue &SOffset,
959 SDValue &ImmOffset) const {
962 MachineFunction &MF = CurDAG->getMachineFunction();
963 const SIRegisterInfo *TRI =
964 static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
965 MachineRegisterInfo &MRI = MF.getRegInfo();
966 const SITargetLowering& Lowering =
967 *static_cast<const SITargetLowering*>(getTargetLowering());
969 unsigned ScratchPtrReg =
970 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
971 unsigned ScratchOffsetReg =
972 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
973 Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
974 ScratchOffsetReg, MVT::i32);
976 Rsrc = buildScratchRSRC(CurDAG, DL,
977 CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
978 MRI.getLiveInVirtReg(ScratchPtrReg), MVT::i64));
979 SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
980 MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
983 if (CurDAG->isBaseWithConstantOffset(Addr)) {
984 SDValue N1 = Addr.getOperand(1);
985 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
987 if (isLegalMUBUFImmOffset(C1)) {
988 VAddr = Addr.getOperand(0);
989 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
995 if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
996 isa<FrameIndexSDNode>(Addr.getOperand(0))) {
997 VAddr = Addr.getOperand(1);
998 ImmOffset = Addr.getOperand(0);
1003 if (isa<FrameIndexSDNode>(Addr)) {
1004 VAddr = SDValue(CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, DL, MVT::i32,
1005 CurDAG->getConstant(0, MVT::i32)), 0);
1012 ImmOffset = CurDAG->getTargetConstant(0, MVT::i16);
1016 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1017 SDValue &SOffset, SDValue &Offset,
1018 SDValue &GLC, SDValue &SLC,
1019 SDValue &TFE) const {
1020 SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1022 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1025 if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1026 !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1027 !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1028 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT |
1029 APInt::getAllOnesValue(32).getZExtValue(); // Size
1031 SRsrc = buildRSRC(CurDAG, DL, Ptr, 0, Rsrc);
1037 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1038 SDValue &Soffset, SDValue &Offset,
1039 SDValue &GLC) const {
1042 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
1045 // FIXME: This is incorrect and only enough to be able to compile.
1046 SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
1047 AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
1050 assert(Subtarget.hasFlatAddressSpace() &&
1051 "addrspacecast only supported with flat address space!");
1053 assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
1054 ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
1055 "Cannot cast address space to / from constant address!");
1057 assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
1058 ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
1059 "Can only cast to / from flat address space!");
1061 // The flat instructions read the address as the index of the VGPR holding the
1062 // address, so casting should just be reinterpreting the base VGPR, so just
1063 // insert trunc / bitcast / zext.
1065 SDValue Src = ASC->getOperand(0);
1066 EVT DestVT = ASC->getValueType(0);
1067 EVT SrcVT = Src.getValueType();
1069 unsigned SrcSize = SrcVT.getSizeInBits();
1070 unsigned DestSize = DestVT.getSizeInBits();
1072 if (SrcSize > DestSize) {
1073 assert(SrcSize == 64 && DestSize == 32);
1074 return CurDAG->getMachineNode(
1075 TargetOpcode::EXTRACT_SUBREG,
1079 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32));
1083 if (DestSize > SrcSize) {
1084 assert(SrcSize == 32 && DestSize == 64);
1086 SDValue RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
1088 const SDValue Ops[] = {
1091 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
1092 SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
1093 CurDAG->getConstant(0, MVT::i32)), 0),
1094 CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
1097 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1098 SDLoc(N), N->getValueType(0), Ops);
1101 assert(SrcSize == 64 && DestSize == 64);
1102 return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
1105 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
1106 SDValue &SrcMods) const {
1112 if (Src.getOpcode() == ISD::FNEG) {
1113 Mods |= SISrcMods::NEG;
1114 Src = Src.getOperand(0);
1117 if (Src.getOpcode() == ISD::FABS) {
1118 Mods |= SISrcMods::ABS;
1119 Src = Src.getOperand(0);
1122 SrcMods = CurDAG->getTargetConstant(Mods, MVT::i32);
1127 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
1128 SDValue &SrcMods, SDValue &Clamp,
1129 SDValue &Omod) const {
1130 // FIXME: Handle Clamp and Omod
1131 Clamp = CurDAG->getTargetConstant(0, MVT::i32);
1132 Omod = CurDAG->getTargetConstant(0, MVT::i32);
1134 return SelectVOP3Mods(In, Src, SrcMods);
1137 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
1138 const AMDGPUTargetLowering& Lowering =
1139 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
1140 bool IsModified = false;
1143 // Go over all selected nodes and try to fold them a bit more
1144 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1145 E = CurDAG->allnodes_end(); I != E; ++I) {
1149 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1153 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
1154 if (ResNode != Node) {
1155 ReplaceUses(Node, ResNode);
1159 CurDAG->RemoveDeadNodes();
1160 } while (IsModified);