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 SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
100 SDValue &SOffset, SDValue &ImmOffset) const;
101 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
102 SDValue &Offset, SDValue &GLC, SDValue &SLC,
104 SDNode *SelectAddrSpaceCast(SDNode *N);
105 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
106 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
107 SDValue &Clamp, SDValue &Omod) const;
109 SDNode *SelectADD_SUB_I64(SDNode *N);
110 SDNode *SelectDIV_SCALE(SDNode *N);
112 // Include the pieces autogenerated from the target description.
113 #include "AMDGPUGenDAGISel.inc"
115 } // end anonymous namespace
117 /// \brief This pass converts a legalized DAG into a AMDGPU-specific
118 // DAG, ready for instruction scheduling.
119 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) {
120 return new AMDGPUDAGToDAGISel(TM);
123 AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
124 : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) {
127 AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
130 bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const {
131 const SITargetLowering *TL
132 = static_cast<const SITargetLowering *>(getTargetLowering());
133 return TL->analyzeImmediate(N) == 0;
136 /// \brief Determine the register class for \p OpNo
137 /// \returns The register class of the virtual register that will be used for
138 /// the given operand number \OpNo or NULL if the register class cannot be
140 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
141 unsigned OpNo) const {
142 if (!N->isMachineOpcode())
145 switch (N->getMachineOpcode()) {
147 const MCInstrDesc &Desc =
148 TM.getSubtargetImpl()->getInstrInfo()->get(N->getMachineOpcode());
149 unsigned OpIdx = Desc.getNumDefs() + OpNo;
150 if (OpIdx >= Desc.getNumOperands())
152 int RegClass = Desc.OpInfo[OpIdx].RegClass;
156 return TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RegClass);
158 case AMDGPU::REG_SEQUENCE: {
159 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
160 const TargetRegisterClass *SuperRC =
161 TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RCID);
163 SDValue SubRegOp = N->getOperand(OpNo + 1);
164 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
165 return TM.getSubtargetImpl()->getRegisterInfo()->getSubClassWithSubReg(
171 SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) {
172 return CurDAG->getTargetConstant(Imm, MVT::i32);
175 bool AMDGPUDAGToDAGISel::SelectADDRParam(
176 SDValue Addr, SDValue& R1, SDValue& R2) {
178 if (Addr.getOpcode() == ISD::FrameIndex) {
179 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
180 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32);
181 R2 = CurDAG->getTargetConstant(0, MVT::i32);
184 R2 = CurDAG->getTargetConstant(0, MVT::i32);
186 } else if (Addr.getOpcode() == ISD::ADD) {
187 R1 = Addr.getOperand(0);
188 R2 = Addr.getOperand(1);
191 R2 = CurDAG->getTargetConstant(0, MVT::i32);
196 bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) {
197 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
198 Addr.getOpcode() == ISD::TargetGlobalAddress) {
201 return SelectADDRParam(Addr, R1, R2);
205 bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
206 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
207 Addr.getOpcode() == ISD::TargetGlobalAddress) {
211 if (Addr.getOpcode() == ISD::FrameIndex) {
212 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
213 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
214 R2 = CurDAG->getTargetConstant(0, MVT::i64);
217 R2 = CurDAG->getTargetConstant(0, MVT::i64);
219 } else if (Addr.getOpcode() == ISD::ADD) {
220 R1 = Addr.getOperand(0);
221 R2 = Addr.getOperand(1);
224 R2 = CurDAG->getTargetConstant(0, MVT::i64);
229 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
230 unsigned int Opc = N->getOpcode();
231 if (N->isMachineOpcode()) {
233 return nullptr; // Already selected.
236 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
239 // We are selecting i64 ADD here instead of custom lower it during
240 // DAG legalization, so we can fold some i64 ADDs used for address
241 // calculation into the LOAD and STORE instructions.
244 if (N->getValueType(0) != MVT::i64 ||
245 ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
248 return SelectADD_SUB_I64(N);
250 case ISD::SCALAR_TO_VECTOR:
251 case AMDGPUISD::BUILD_VERTICAL_VECTOR:
252 case ISD::BUILD_VECTOR: {
254 const AMDGPURegisterInfo *TRI = static_cast<const AMDGPURegisterInfo *>(
255 TM.getSubtargetImpl()->getRegisterInfo());
256 const SIRegisterInfo *SIRI = static_cast<const SIRegisterInfo *>(
257 TM.getSubtargetImpl()->getRegisterInfo());
258 EVT VT = N->getValueType(0);
259 unsigned NumVectorElts = VT.getVectorNumElements();
260 EVT EltVT = VT.getVectorElementType();
261 assert(EltVT.bitsEq(MVT::i32));
262 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
264 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
266 if (!U->isMachineOpcode()) {
269 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
273 if (SIRI->isSGPRClass(RC)) {
277 switch(NumVectorElts) {
278 case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
279 AMDGPU::SReg_32RegClassID;
281 case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
282 AMDGPU::SReg_64RegClassID;
284 case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
285 AMDGPU::SReg_128RegClassID;
287 case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
288 AMDGPU::SReg_256RegClassID;
290 case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
291 AMDGPU::SReg_512RegClassID;
293 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
296 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
297 // that adds a 128 bits reg copy when going through TwoAddressInstructions
298 // pass. We want to avoid 128 bits copies as much as possible because they
299 // can't be bundled by our scheduler.
300 switch(NumVectorElts) {
301 case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
303 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
304 RegClassID = AMDGPU::R600_Reg128VerticalRegClassID;
306 RegClassID = AMDGPU::R600_Reg128RegClassID;
308 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
312 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
314 if (NumVectorElts == 1) {
315 return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT,
316 N->getOperand(0), RegClass);
319 assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
321 // 16 = Max Num Vector Elements
322 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
323 // 1 = Vector Register Class
324 SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
326 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
327 bool IsRegSeq = true;
328 unsigned NOps = N->getNumOperands();
329 for (unsigned i = 0; i < NOps; i++) {
330 // XXX: Why is this here?
331 if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
335 RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
336 RegSeqArgs[1 + (2 * i) + 1] =
337 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
340 if (NOps != NumVectorElts) {
341 // Fill in the missing undef elements if this was a scalar_to_vector.
342 assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
344 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
346 for (unsigned i = NOps; i < NumVectorElts; ++i) {
347 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
348 RegSeqArgs[1 + (2 * i) + 1] =
349 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
355 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
358 case ISD::BUILD_PAIR: {
359 SDValue RC, SubReg0, SubReg1;
360 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
363 if (N->getValueType(0) == MVT::i128) {
364 RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32);
365 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, MVT::i32);
366 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, MVT::i32);
367 } else if (N->getValueType(0) == MVT::i64) {
368 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32);
369 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
370 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
372 llvm_unreachable("Unhandled value type for BUILD_PAIR");
374 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
375 N->getOperand(1), SubReg1 };
376 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
377 SDLoc(N), N->getValueType(0), Ops);
381 case ISD::ConstantFP: {
382 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
383 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
384 N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
388 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
389 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
391 ConstantSDNode *C = cast<ConstantSDNode>(N);
392 Imm = C->getZExtValue();
395 SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
396 CurDAG->getConstant(Imm & 0xFFFFFFFF, MVT::i32));
397 SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
398 CurDAG->getConstant(Imm >> 32, MVT::i32));
399 const SDValue Ops[] = {
400 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
401 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
402 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
405 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, SDLoc(N),
406 N->getValueType(0), Ops);
409 case AMDGPUISD::REGISTER_LOAD: {
410 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
412 SDValue Addr, Offset;
414 SelectADDRIndirect(N->getOperand(1), Addr, Offset);
415 const SDValue Ops[] = {
418 CurDAG->getTargetConstant(0, MVT::i32),
421 return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, SDLoc(N),
422 CurDAG->getVTList(MVT::i32, MVT::i64, MVT::Other),
425 case AMDGPUISD::REGISTER_STORE: {
426 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
428 SDValue Addr, Offset;
429 SelectADDRIndirect(N->getOperand(2), Addr, Offset);
430 const SDValue Ops[] = {
434 CurDAG->getTargetConstant(0, MVT::i32),
437 return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, SDLoc(N),
438 CurDAG->getVTList(MVT::Other),
442 case AMDGPUISD::BFE_I32:
443 case AMDGPUISD::BFE_U32: {
444 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
447 // There is a scalar version available, but unlike the vector version which
448 // has a separate operand for the offset and width, the scalar version packs
449 // the width and offset into a single operand. Try to move to the scalar
450 // version if the offsets are constant, so that we can try to keep extended
451 // loads of kernel arguments in SGPRs.
453 // TODO: Technically we could try to pattern match scalar bitshifts of
454 // dynamic values, but it's probably not useful.
455 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
459 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
463 bool Signed = Opc == AMDGPUISD::BFE_I32;
465 // Transformation function, pack the offset and width of a BFE into
466 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
467 // source, bits [5:0] contain the offset and bits [22:16] the width.
469 uint32_t OffsetVal = Offset->getZExtValue();
470 uint32_t WidthVal = Width->getZExtValue();
472 uint32_t PackedVal = OffsetVal | WidthVal << 16;
474 SDValue PackedOffsetWidth = CurDAG->getTargetConstant(PackedVal, MVT::i32);
475 return CurDAG->getMachineNode(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
482 case AMDGPUISD::DIV_SCALE: {
483 return SelectDIV_SCALE(N);
485 case ISD::ADDRSPACECAST:
486 return SelectAddrSpaceCast(N);
488 return SelectCode(N);
492 bool AMDGPUDAGToDAGISel::checkType(const Value *Ptr, unsigned AS) {
493 assert(AS != 0 && "Use checkPrivateAddress instead.");
497 return Ptr->getType()->getPointerAddressSpace() == AS;
500 bool AMDGPUDAGToDAGISel::checkPrivateAddress(const MachineMemOperand *Op) {
501 if (Op->getPseudoValue())
504 if (PointerType *PT = dyn_cast<PointerType>(Op->getValue()->getType()))
505 return PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
510 bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
511 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
514 bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
515 const Value *MemVal = N->getMemOperand()->getValue();
516 return (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
517 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
518 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS));
521 bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
522 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
525 bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
526 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
529 bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
530 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
533 bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
534 const Value *MemVal = N->getMemOperand()->getValue();
536 return checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS);
538 return checkType(MemVal, AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
541 bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
542 if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) {
543 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
544 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
545 N->getMemoryVT().bitsLT(MVT::i32)) {
549 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
552 bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) const {
553 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::PARAM_I_ADDRESS);
556 bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) const {
557 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
560 bool AMDGPUDAGToDAGISel::isFlatLoad(const LoadSDNode *N) const {
561 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
564 bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) const {
565 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
568 bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) const {
569 MachineMemOperand *MMO = N->getMemOperand();
570 if (checkPrivateAddress(N->getMemOperand())) {
572 const PseudoSourceValue *PSV = MMO->getPseudoValue();
573 if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
581 bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
582 if (checkPrivateAddress(N->getMemOperand())) {
583 // Check to make sure we are not a constant pool load or a constant load
584 // that is marked as a private load
585 if (isCPLoad(N) || isConstantLoad(N, -1)) {
590 const Value *MemVal = N->getMemOperand()->getValue();
591 if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
592 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
593 !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
594 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
595 !checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
596 !checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
597 !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
603 const char *AMDGPUDAGToDAGISel::getPassName() const {
604 return "AMDGPU DAG->DAG Pattern Instruction Selection";
612 //===----------------------------------------------------------------------===//
614 //===----------------------------------------------------------------------===//
616 bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
618 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
619 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, true);
625 bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
626 SDValue& BaseReg, SDValue &Offset) {
627 if (!isa<ConstantSDNode>(Addr)) {
629 Offset = CurDAG->getIntPtrConstant(0, true);
635 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
637 ConstantSDNode *IMMOffset;
639 if (Addr.getOpcode() == ISD::ADD
640 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
641 && isInt<16>(IMMOffset->getZExtValue())) {
643 Base = Addr.getOperand(0);
644 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
646 // If the pointer address is constant, we can move it to the offset field.
647 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
648 && isInt<16>(IMMOffset->getZExtValue())) {
649 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
650 SDLoc(CurDAG->getEntryNode()),
651 AMDGPU::ZERO, MVT::i32);
652 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
656 // Default case, no offset
658 Offset = CurDAG->getTargetConstant(0, MVT::i32);
662 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
666 if ((C = dyn_cast<ConstantSDNode>(Addr))) {
667 Base = CurDAG->getRegister(AMDGPU::INDIRECT_BASE_ADDR, MVT::i32);
668 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
669 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
670 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
671 Base = Addr.getOperand(0);
672 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
675 Offset = CurDAG->getTargetConstant(0, MVT::i32);
681 SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
683 SDValue LHS = N->getOperand(0);
684 SDValue RHS = N->getOperand(1);
686 bool IsAdd = (N->getOpcode() == ISD::ADD);
688 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
689 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
691 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
692 DL, MVT::i32, LHS, Sub0);
693 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
694 DL, MVT::i32, LHS, Sub1);
696 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
697 DL, MVT::i32, RHS, Sub0);
698 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
699 DL, MVT::i32, RHS, Sub1);
701 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
702 SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
705 unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
706 unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
708 SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs);
709 SDValue Carry(AddLo, 1);
711 = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32,
712 SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
715 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
721 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
724 SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
726 EVT VT = N->getValueType(0);
728 assert(VT == MVT::f32 || VT == MVT::f64);
731 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
733 const SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
745 return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
748 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
749 unsigned OffsetBits) const {
750 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
751 if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
752 (OffsetBits == 8 && !isUInt<8>(Offset)))
755 if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
758 // On Southern Islands instruction with a negative base value and an offset
759 // don't seem to work.
760 return CurDAG->SignBitIsZero(Base);
763 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
764 SDValue &Offset) const {
765 if (CurDAG->isBaseWithConstantOffset(Addr)) {
766 SDValue N0 = Addr.getOperand(0);
767 SDValue N1 = Addr.getOperand(1);
768 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
769 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
779 Offset = CurDAG->getTargetConstant(0, MVT::i16);
783 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
785 SDValue &Offset1) const {
786 if (CurDAG->isBaseWithConstantOffset(Addr)) {
787 SDValue N0 = Addr.getOperand(0);
788 SDValue N1 = Addr.getOperand(1);
789 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
790 unsigned DWordOffset0 = C1->getZExtValue() / 4;
791 unsigned DWordOffset1 = DWordOffset0 + 1;
793 if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
795 Offset0 = CurDAG->getTargetConstant(DWordOffset0, MVT::i8);
796 Offset1 = CurDAG->getTargetConstant(DWordOffset1, MVT::i8);
803 Offset0 = CurDAG->getTargetConstant(0, MVT::i8);
804 Offset1 = CurDAG->getTargetConstant(1, MVT::i8);
808 static SDValue wrapAddr64Rsrc(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
809 return SDValue(DAG->getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::v4i32,
813 static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
814 return isUInt<12>(Imm->getZExtValue());
817 void AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
818 SDValue &VAddr, SDValue &SOffset,
819 SDValue &Offset, SDValue &Offen,
820 SDValue &Idxen, SDValue &Addr64,
821 SDValue &GLC, SDValue &SLC,
822 SDValue &TFE) const {
825 GLC = CurDAG->getTargetConstant(0, MVT::i1);
826 SLC = CurDAG->getTargetConstant(0, MVT::i1);
827 TFE = CurDAG->getTargetConstant(0, MVT::i1);
829 Idxen = CurDAG->getTargetConstant(0, MVT::i1);
830 Offen = CurDAG->getTargetConstant(0, MVT::i1);
831 Addr64 = CurDAG->getTargetConstant(0, MVT::i1);
832 SOffset = CurDAG->getTargetConstant(0, MVT::i32);
834 if (CurDAG->isBaseWithConstantOffset(Addr)) {
835 SDValue N0 = Addr.getOperand(0);
836 SDValue N1 = Addr.getOperand(1);
837 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
839 if (isLegalMUBUFImmOffset(C1)) {
841 if (N0.getOpcode() == ISD::ADD) {
842 // (add (add N2, N3), C1) -> addr64
843 SDValue N2 = N0.getOperand(0);
844 SDValue N3 = N0.getOperand(1);
845 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
848 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
852 // (add N0, C1) -> offset
853 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
855 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
859 if (Addr.getOpcode() == ISD::ADD) {
860 // (add N0, N1) -> addr64
861 SDValue N0 = Addr.getOperand(0);
862 SDValue N1 = Addr.getOperand(1);
863 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
866 Offset = CurDAG->getTargetConstant(0, MVT::i16);
870 // default case -> offset
871 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
873 Offset = CurDAG->getTargetConstant(0, MVT::i16);
877 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
879 SDValue &Offset) const {
880 SDValue Ptr, SOffset, Offen, Idxen, Addr64, GLC, SLC, TFE;
882 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
885 ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
886 if (C->getSExtValue()) {
888 SRsrc = wrapAddr64Rsrc(CurDAG, DL, Ptr);
894 static SDValue buildRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr,
895 uint32_t RsrcDword1, uint64_t RsrcDword2And3) {
897 SDValue PtrLo = DAG->getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
898 SDValue PtrHi = DAG->getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
900 PtrHi = SDValue(DAG->getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
901 DAG->getConstant(RsrcDword1, MVT::i32)), 0);
903 SDValue DataLo = DAG->getTargetConstant(
904 RsrcDword2And3 & APInt::getAllOnesValue(32).getZExtValue(), MVT::i32);
905 SDValue DataHi = DAG->getTargetConstant(RsrcDword2And3 >> 32, MVT::i32);
907 const SDValue Ops[] = { PtrLo, PtrHi, DataLo, DataHi };
908 return SDValue(DAG->getMachineNode(AMDGPU::SI_BUFFER_RSRC, DL,
909 MVT::v4i32, Ops), 0);
912 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
913 /// The TID (Thread ID) is multipled by the stride value (bits [61:48]
914 /// of the resource descriptor) to create an offset, which is added to the
916 static SDValue buildScratchRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
918 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT | AMDGPU::RSRC_TID_ENABLE |
921 return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
924 bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
925 SDValue &VAddr, SDValue &SOffset,
926 SDValue &ImmOffset) const {
929 MachineFunction &MF = CurDAG->getMachineFunction();
930 const SIRegisterInfo *TRI =
931 static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
932 MachineRegisterInfo &MRI = MF.getRegInfo();
933 const SITargetLowering& Lowering =
934 *static_cast<const SITargetLowering*>(getTargetLowering());
936 unsigned ScratchPtrReg =
937 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
938 unsigned ScratchOffsetReg =
939 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
940 Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
941 ScratchOffsetReg, MVT::i32);
943 Rsrc = buildScratchRSRC(CurDAG, DL,
944 CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
945 MRI.getLiveInVirtReg(ScratchPtrReg), MVT::i64));
946 SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
947 MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
950 if (CurDAG->isBaseWithConstantOffset(Addr)) {
951 SDValue N1 = Addr.getOperand(1);
952 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
954 if (isLegalMUBUFImmOffset(C1)) {
955 VAddr = Addr.getOperand(0);
956 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
962 if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
963 isa<FrameIndexSDNode>(Addr.getOperand(0))) {
964 VAddr = Addr.getOperand(1);
965 ImmOffset = Addr.getOperand(0);
970 if (isa<FrameIndexSDNode>(Addr)) {
971 VAddr = SDValue(CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, DL, MVT::i32,
972 CurDAG->getConstant(0, MVT::i32)), 0);
979 ImmOffset = CurDAG->getTargetConstant(0, MVT::i16);
983 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
984 SDValue &SOffset, SDValue &Offset,
985 SDValue &GLC, SDValue &SLC,
986 SDValue &TFE) const {
987 SDValue Ptr, VAddr, Offen, Idxen, Addr64;
989 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
992 if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
993 !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
994 !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
995 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT |
996 APInt::getAllOnesValue(32).getZExtValue(); // Size
998 SRsrc = buildRSRC(CurDAG, DL, Ptr, 0, Rsrc);
1004 // FIXME: This is incorrect and only enough to be able to compile.
1005 SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
1006 AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
1009 assert(Subtarget.hasFlatAddressSpace() &&
1010 "addrspacecast only supported with flat address space!");
1012 assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
1013 ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
1014 "Cannot cast address space to / from constant address!");
1016 assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
1017 ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
1018 "Can only cast to / from flat address space!");
1020 // The flat instructions read the address as the index of the VGPR holding the
1021 // address, so casting should just be reinterpreting the base VGPR, so just
1022 // insert trunc / bitcast / zext.
1024 SDValue Src = ASC->getOperand(0);
1025 EVT DestVT = ASC->getValueType(0);
1026 EVT SrcVT = Src.getValueType();
1028 unsigned SrcSize = SrcVT.getSizeInBits();
1029 unsigned DestSize = DestVT.getSizeInBits();
1031 if (SrcSize > DestSize) {
1032 assert(SrcSize == 64 && DestSize == 32);
1033 return CurDAG->getMachineNode(
1034 TargetOpcode::EXTRACT_SUBREG,
1038 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32));
1042 if (DestSize > SrcSize) {
1043 assert(SrcSize == 32 && DestSize == 64);
1045 SDValue RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
1047 const SDValue Ops[] = {
1050 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
1051 SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
1052 CurDAG->getConstant(0, MVT::i32)), 0),
1053 CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
1056 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1057 SDLoc(N), N->getValueType(0), Ops);
1060 assert(SrcSize == 64 && DestSize == 64);
1061 return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
1064 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
1065 SDValue &SrcMods) const {
1071 if (Src.getOpcode() == ISD::FNEG) {
1072 Mods |= SISrcMods::NEG;
1073 Src = Src.getOperand(0);
1076 if (Src.getOpcode() == ISD::FABS) {
1077 Mods |= SISrcMods::ABS;
1078 Src = Src.getOperand(0);
1081 SrcMods = CurDAG->getTargetConstant(Mods, MVT::i32);
1086 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
1087 SDValue &SrcMods, SDValue &Clamp,
1088 SDValue &Omod) const {
1089 // FIXME: Handle Clamp and Omod
1090 Clamp = CurDAG->getTargetConstant(0, MVT::i32);
1091 Omod = CurDAG->getTargetConstant(0, MVT::i32);
1093 return SelectVOP3Mods(In, Src, SrcMods);
1096 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
1097 const AMDGPUTargetLowering& Lowering =
1098 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
1099 bool IsModified = false;
1102 // Go over all selected nodes and try to fold them a bit more
1103 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1104 E = CurDAG->allnodes_end(); I != E; ++I) {
1108 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1112 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
1113 if (ResNode != Node) {
1114 ReplaceUses(Node, ResNode);
1118 CurDAG->RemoveDeadNodes();
1119 } while (IsModified);