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 /// \returns True if the current basic block being selected is at control
83 /// flow depth 0. Meaning that the current block dominates the
85 bool isCFDepth0() const;
87 const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
88 bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr);
89 bool SelectGlobalValueVariableOffset(SDValue Addr, SDValue &BaseReg,
91 bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
92 bool SelectADDRIndirect(SDValue Addr, SDValue &Base, SDValue &Offset);
93 bool isDSOffsetLegal(const SDValue &Base, unsigned Offset,
94 unsigned OffsetBits) const;
95 bool SelectDS1Addr1Offset(SDValue Ptr, SDValue &Base, SDValue &Offset) const;
96 bool SelectDS64Bit4ByteAligned(SDValue Ptr, SDValue &Base, SDValue &Offset0,
97 SDValue &Offset1) const;
98 void SelectMUBUF(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
99 SDValue &SOffset, SDValue &Offset, SDValue &Offen,
100 SDValue &Idxen, SDValue &Addr64, SDValue &GLC, SDValue &SLC,
102 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc, SDValue &VAddr,
103 SDValue &Offset) const;
104 bool SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
105 SDValue &VAddr, SDValue &Offset,
107 bool SelectMUBUFScratch(SDValue Addr, SDValue &RSrc, SDValue &VAddr,
108 SDValue &SOffset, SDValue &ImmOffset) const;
109 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &SOffset,
110 SDValue &Offset, SDValue &GLC, SDValue &SLC,
112 bool SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc, SDValue &Soffset,
113 SDValue &Offset, SDValue &GLC) const;
114 SDNode *SelectAddrSpaceCast(SDNode *N);
115 bool SelectVOP3Mods(SDValue In, SDValue &Src, SDValue &SrcMods) const;
116 bool SelectVOP3Mods0(SDValue In, SDValue &Src, SDValue &SrcMods,
117 SDValue &Clamp, SDValue &Omod) const;
119 SDNode *SelectADD_SUB_I64(SDNode *N);
120 SDNode *SelectDIV_SCALE(SDNode *N);
122 // Include the pieces autogenerated from the target description.
123 #include "AMDGPUGenDAGISel.inc"
125 } // end anonymous namespace
127 /// \brief This pass converts a legalized DAG into a AMDGPU-specific
128 // DAG, ready for instruction scheduling.
129 FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM) {
130 return new AMDGPUDAGToDAGISel(TM);
133 AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
134 : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) {
137 AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
140 bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const {
141 const SITargetLowering *TL
142 = static_cast<const SITargetLowering *>(getTargetLowering());
143 return TL->analyzeImmediate(N) == 0;
146 /// \brief Determine the register class for \p OpNo
147 /// \returns The register class of the virtual register that will be used for
148 /// the given operand number \OpNo or NULL if the register class cannot be
150 const TargetRegisterClass *AMDGPUDAGToDAGISel::getOperandRegClass(SDNode *N,
151 unsigned OpNo) const {
152 if (!N->isMachineOpcode())
155 switch (N->getMachineOpcode()) {
157 const MCInstrDesc &Desc =
158 TM.getSubtargetImpl()->getInstrInfo()->get(N->getMachineOpcode());
159 unsigned OpIdx = Desc.getNumDefs() + OpNo;
160 if (OpIdx >= Desc.getNumOperands())
162 int RegClass = Desc.OpInfo[OpIdx].RegClass;
166 return TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RegClass);
168 case AMDGPU::REG_SEQUENCE: {
169 unsigned RCID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
170 const TargetRegisterClass *SuperRC =
171 TM.getSubtargetImpl()->getRegisterInfo()->getRegClass(RCID);
173 SDValue SubRegOp = N->getOperand(OpNo + 1);
174 unsigned SubRegIdx = cast<ConstantSDNode>(SubRegOp)->getZExtValue();
175 return TM.getSubtargetImpl()->getRegisterInfo()->getSubClassWithSubReg(
181 SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) {
182 return CurDAG->getTargetConstant(Imm, MVT::i32);
185 bool AMDGPUDAGToDAGISel::SelectADDRParam(
186 SDValue Addr, SDValue& R1, SDValue& R2) {
188 if (Addr.getOpcode() == ISD::FrameIndex) {
189 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
190 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32);
191 R2 = CurDAG->getTargetConstant(0, MVT::i32);
194 R2 = CurDAG->getTargetConstant(0, MVT::i32);
196 } else if (Addr.getOpcode() == ISD::ADD) {
197 R1 = Addr.getOperand(0);
198 R2 = Addr.getOperand(1);
201 R2 = CurDAG->getTargetConstant(0, MVT::i32);
206 bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) {
207 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
208 Addr.getOpcode() == ISD::TargetGlobalAddress) {
211 return SelectADDRParam(Addr, R1, R2);
215 bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
216 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
217 Addr.getOpcode() == ISD::TargetGlobalAddress) {
221 if (Addr.getOpcode() == ISD::FrameIndex) {
222 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
223 R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
224 R2 = CurDAG->getTargetConstant(0, MVT::i64);
227 R2 = CurDAG->getTargetConstant(0, MVT::i64);
229 } else if (Addr.getOpcode() == ISD::ADD) {
230 R1 = Addr.getOperand(0);
231 R2 = Addr.getOperand(1);
234 R2 = CurDAG->getTargetConstant(0, MVT::i64);
239 SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
240 unsigned int Opc = N->getOpcode();
241 if (N->isMachineOpcode()) {
243 return nullptr; // Already selected.
246 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
249 // We are selecting i64 ADD here instead of custom lower it during
250 // DAG legalization, so we can fold some i64 ADDs used for address
251 // calculation into the LOAD and STORE instructions.
254 if (N->getValueType(0) != MVT::i64 ||
255 ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
258 return SelectADD_SUB_I64(N);
260 case ISD::SCALAR_TO_VECTOR:
261 case AMDGPUISD::BUILD_VERTICAL_VECTOR:
262 case ISD::BUILD_VECTOR: {
264 const AMDGPURegisterInfo *TRI = static_cast<const AMDGPURegisterInfo *>(
265 TM.getSubtargetImpl()->getRegisterInfo());
266 const SIRegisterInfo *SIRI = static_cast<const SIRegisterInfo *>(
267 TM.getSubtargetImpl()->getRegisterInfo());
268 EVT VT = N->getValueType(0);
269 unsigned NumVectorElts = VT.getVectorNumElements();
270 EVT EltVT = VT.getVectorElementType();
271 assert(EltVT.bitsEq(MVT::i32));
272 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
274 for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
276 if (!U->isMachineOpcode()) {
279 const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
283 if (SIRI->isSGPRClass(RC)) {
287 switch(NumVectorElts) {
288 case 1: RegClassID = UseVReg ? AMDGPU::VReg_32RegClassID :
289 AMDGPU::SReg_32RegClassID;
291 case 2: RegClassID = UseVReg ? AMDGPU::VReg_64RegClassID :
292 AMDGPU::SReg_64RegClassID;
294 case 4: RegClassID = UseVReg ? AMDGPU::VReg_128RegClassID :
295 AMDGPU::SReg_128RegClassID;
297 case 8: RegClassID = UseVReg ? AMDGPU::VReg_256RegClassID :
298 AMDGPU::SReg_256RegClassID;
300 case 16: RegClassID = UseVReg ? AMDGPU::VReg_512RegClassID :
301 AMDGPU::SReg_512RegClassID;
303 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
306 // BUILD_VECTOR was lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
307 // that adds a 128 bits reg copy when going through TwoAddressInstructions
308 // pass. We want to avoid 128 bits copies as much as possible because they
309 // can't be bundled by our scheduler.
310 switch(NumVectorElts) {
311 case 2: RegClassID = AMDGPU::R600_Reg64RegClassID; break;
313 if (Opc == AMDGPUISD::BUILD_VERTICAL_VECTOR)
314 RegClassID = AMDGPU::R600_Reg128VerticalRegClassID;
316 RegClassID = AMDGPU::R600_Reg128RegClassID;
318 default: llvm_unreachable("Do not know how to lower this BUILD_VECTOR");
322 SDValue RegClass = CurDAG->getTargetConstant(RegClassID, MVT::i32);
324 if (NumVectorElts == 1) {
325 return CurDAG->SelectNodeTo(N, AMDGPU::COPY_TO_REGCLASS, EltVT,
326 N->getOperand(0), RegClass);
329 assert(NumVectorElts <= 16 && "Vectors with more than 16 elements not "
331 // 16 = Max Num Vector Elements
332 // 2 = 2 REG_SEQUENCE operands per element (value, subreg index)
333 // 1 = Vector Register Class
334 SmallVector<SDValue, 16 * 2 + 1> RegSeqArgs(NumVectorElts * 2 + 1);
336 RegSeqArgs[0] = CurDAG->getTargetConstant(RegClassID, MVT::i32);
337 bool IsRegSeq = true;
338 unsigned NOps = N->getNumOperands();
339 for (unsigned i = 0; i < NOps; i++) {
340 // XXX: Why is this here?
341 if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
345 RegSeqArgs[1 + (2 * i)] = N->getOperand(i);
346 RegSeqArgs[1 + (2 * i) + 1] =
347 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
350 if (NOps != NumVectorElts) {
351 // Fill in the missing undef elements if this was a scalar_to_vector.
352 assert(Opc == ISD::SCALAR_TO_VECTOR && NOps < NumVectorElts);
354 MachineSDNode *ImpDef = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
356 for (unsigned i = NOps; i < NumVectorElts; ++i) {
357 RegSeqArgs[1 + (2 * i)] = SDValue(ImpDef, 0);
358 RegSeqArgs[1 + (2 * i) + 1] =
359 CurDAG->getTargetConstant(TRI->getSubRegFromChannel(i), MVT::i32);
365 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
368 case ISD::BUILD_PAIR: {
369 SDValue RC, SubReg0, SubReg1;
370 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
373 if (N->getValueType(0) == MVT::i128) {
374 RC = CurDAG->getTargetConstant(AMDGPU::SReg_128RegClassID, MVT::i32);
375 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0_sub1, MVT::i32);
376 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub2_sub3, MVT::i32);
377 } else if (N->getValueType(0) == MVT::i64) {
378 RC = CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32);
379 SubReg0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
380 SubReg1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
382 llvm_unreachable("Unhandled value type for BUILD_PAIR");
384 const SDValue Ops[] = { RC, N->getOperand(0), SubReg0,
385 N->getOperand(1), SubReg1 };
386 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
387 SDLoc(N), N->getValueType(0), Ops);
391 case ISD::ConstantFP: {
392 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
393 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
394 N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
398 if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
399 Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
401 ConstantSDNode *C = cast<ConstantSDNode>(N);
402 Imm = C->getZExtValue();
405 SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
406 CurDAG->getConstant(Imm & 0xFFFFFFFF, MVT::i32));
407 SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
408 CurDAG->getConstant(Imm >> 32, MVT::i32));
409 const SDValue Ops[] = {
410 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
411 SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
412 SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
415 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, SDLoc(N),
416 N->getValueType(0), Ops);
419 case AMDGPUISD::REGISTER_LOAD: {
420 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
422 SDValue Addr, Offset;
424 SelectADDRIndirect(N->getOperand(1), Addr, Offset);
425 const SDValue Ops[] = {
428 CurDAG->getTargetConstant(0, MVT::i32),
431 return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, SDLoc(N),
432 CurDAG->getVTList(MVT::i32, MVT::i64, MVT::Other),
435 case AMDGPUISD::REGISTER_STORE: {
436 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
438 SDValue Addr, Offset;
439 SelectADDRIndirect(N->getOperand(2), Addr, Offset);
440 const SDValue Ops[] = {
444 CurDAG->getTargetConstant(0, MVT::i32),
447 return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, SDLoc(N),
448 CurDAG->getVTList(MVT::Other),
452 case AMDGPUISD::BFE_I32:
453 case AMDGPUISD::BFE_U32: {
454 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
457 // There is a scalar version available, but unlike the vector version which
458 // has a separate operand for the offset and width, the scalar version packs
459 // the width and offset into a single operand. Try to move to the scalar
460 // version if the offsets are constant, so that we can try to keep extended
461 // loads of kernel arguments in SGPRs.
463 // TODO: Technically we could try to pattern match scalar bitshifts of
464 // dynamic values, but it's probably not useful.
465 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
469 ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
473 bool Signed = Opc == AMDGPUISD::BFE_I32;
475 // Transformation function, pack the offset and width of a BFE into
476 // the format expected by the S_BFE_I32 / S_BFE_U32. In the second
477 // source, bits [5:0] contain the offset and bits [22:16] the width.
479 uint32_t OffsetVal = Offset->getZExtValue();
480 uint32_t WidthVal = Width->getZExtValue();
482 uint32_t PackedVal = OffsetVal | WidthVal << 16;
484 SDValue PackedOffsetWidth = CurDAG->getTargetConstant(PackedVal, MVT::i32);
485 return CurDAG->getMachineNode(Signed ? AMDGPU::S_BFE_I32 : AMDGPU::S_BFE_U32,
492 case AMDGPUISD::DIV_SCALE: {
493 return SelectDIV_SCALE(N);
495 case ISD::ADDRSPACECAST:
496 return SelectAddrSpaceCast(N);
498 return SelectCode(N);
502 bool AMDGPUDAGToDAGISel::checkType(const Value *Ptr, unsigned AS) {
503 assert(AS != 0 && "Use checkPrivateAddress instead.");
507 return Ptr->getType()->getPointerAddressSpace() == AS;
510 bool AMDGPUDAGToDAGISel::checkPrivateAddress(const MachineMemOperand *Op) {
511 if (Op->getPseudoValue())
514 if (PointerType *PT = dyn_cast<PointerType>(Op->getValue()->getType()))
515 return PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
520 bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
521 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
524 bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
525 const Value *MemVal = N->getMemOperand()->getValue();
526 return (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
527 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
528 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS));
531 bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
532 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
535 bool AMDGPUDAGToDAGISel::isFlatStore(const StoreSDNode *N) {
536 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
539 bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
540 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
543 bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int CbId) const {
544 const Value *MemVal = N->getMemOperand()->getValue();
546 return checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS);
548 return checkType(MemVal, AMDGPUAS::CONSTANT_BUFFER_0 + CbId);
551 bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) const {
552 if (N->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS) {
553 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
554 if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
555 N->getMemoryVT().bitsLT(MVT::i32)) {
559 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::GLOBAL_ADDRESS);
562 bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) const {
563 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::PARAM_I_ADDRESS);
566 bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) const {
567 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::LOCAL_ADDRESS);
570 bool AMDGPUDAGToDAGISel::isFlatLoad(const LoadSDNode *N) const {
571 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::FLAT_ADDRESS);
574 bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) const {
575 return checkType(N->getMemOperand()->getValue(), AMDGPUAS::REGION_ADDRESS);
578 bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) const {
579 MachineMemOperand *MMO = N->getMemOperand();
580 if (checkPrivateAddress(N->getMemOperand())) {
582 const PseudoSourceValue *PSV = MMO->getPseudoValue();
583 if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
591 bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
592 if (checkPrivateAddress(N->getMemOperand())) {
593 // Check to make sure we are not a constant pool load or a constant load
594 // that is marked as a private load
595 if (isCPLoad(N) || isConstantLoad(N, -1)) {
600 const Value *MemVal = N->getMemOperand()->getValue();
601 if (!checkType(MemVal, AMDGPUAS::LOCAL_ADDRESS) &&
602 !checkType(MemVal, AMDGPUAS::GLOBAL_ADDRESS) &&
603 !checkType(MemVal, AMDGPUAS::FLAT_ADDRESS) &&
604 !checkType(MemVal, AMDGPUAS::REGION_ADDRESS) &&
605 !checkType(MemVal, AMDGPUAS::CONSTANT_ADDRESS) &&
606 !checkType(MemVal, AMDGPUAS::PARAM_D_ADDRESS) &&
607 !checkType(MemVal, AMDGPUAS::PARAM_I_ADDRESS)) {
613 bool AMDGPUDAGToDAGISel::isCFDepth0() const {
614 // FIXME: Figure out a way to use DominatorTree analysis here.
615 const BasicBlock *CurBlock = FuncInfo->MBB->getBasicBlock();
616 const Function *Fn = FuncInfo->Fn;
617 return &Fn->front() == CurBlock || &Fn->back() == CurBlock;
621 const char *AMDGPUDAGToDAGISel::getPassName() const {
622 return "AMDGPU DAG->DAG Pattern Instruction Selection";
630 //===----------------------------------------------------------------------===//
632 //===----------------------------------------------------------------------===//
634 bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr,
636 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) {
637 IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, true);
643 bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr,
644 SDValue& BaseReg, SDValue &Offset) {
645 if (!isa<ConstantSDNode>(Addr)) {
647 Offset = CurDAG->getIntPtrConstant(0, true);
653 bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
655 ConstantSDNode *IMMOffset;
657 if (Addr.getOpcode() == ISD::ADD
658 && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
659 && isInt<16>(IMMOffset->getZExtValue())) {
661 Base = Addr.getOperand(0);
662 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
664 // If the pointer address is constant, we can move it to the offset field.
665 } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
666 && isInt<16>(IMMOffset->getZExtValue())) {
667 Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
668 SDLoc(CurDAG->getEntryNode()),
669 AMDGPU::ZERO, MVT::i32);
670 Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
674 // Default case, no offset
676 Offset = CurDAG->getTargetConstant(0, MVT::i32);
680 bool AMDGPUDAGToDAGISel::SelectADDRIndirect(SDValue Addr, SDValue &Base,
684 if ((C = dyn_cast<ConstantSDNode>(Addr))) {
685 Base = CurDAG->getRegister(AMDGPU::INDIRECT_BASE_ADDR, MVT::i32);
686 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
687 } else if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
688 (C = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))) {
689 Base = Addr.getOperand(0);
690 Offset = CurDAG->getTargetConstant(C->getZExtValue(), MVT::i32);
693 Offset = CurDAG->getTargetConstant(0, MVT::i32);
699 SDNode *AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
701 SDValue LHS = N->getOperand(0);
702 SDValue RHS = N->getOperand(1);
704 bool IsAdd = (N->getOpcode() == ISD::ADD);
706 SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32);
707 SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32);
709 SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
710 DL, MVT::i32, LHS, Sub0);
711 SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
712 DL, MVT::i32, LHS, Sub1);
714 SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
715 DL, MVT::i32, RHS, Sub0);
716 SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
717 DL, MVT::i32, RHS, Sub1);
719 SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
720 SDValue AddLoArgs[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
723 unsigned Opc = IsAdd ? AMDGPU::S_ADD_U32 : AMDGPU::S_SUB_U32;
724 unsigned CarryOpc = IsAdd ? AMDGPU::S_ADDC_U32 : AMDGPU::S_SUBB_U32;
727 Opc = IsAdd ? AMDGPU::V_ADD_I32_e32 : AMDGPU::V_SUB_I32_e32;
728 CarryOpc = IsAdd ? AMDGPU::V_ADDC_U32_e32 : AMDGPU::V_SUBB_U32_e32;
731 SDNode *AddLo = CurDAG->getMachineNode( Opc, DL, VTList, AddLoArgs);
732 SDValue Carry(AddLo, 1);
734 = CurDAG->getMachineNode(CarryOpc, DL, MVT::i32,
735 SDValue(Hi0, 0), SDValue(Hi1, 0), Carry);
738 CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
744 return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, MVT::i64, Args);
747 SDNode *AMDGPUDAGToDAGISel::SelectDIV_SCALE(SDNode *N) {
749 EVT VT = N->getValueType(0);
751 assert(VT == MVT::f32 || VT == MVT::f64);
754 = (VT == MVT::f64) ? AMDGPU::V_DIV_SCALE_F64 : AMDGPU::V_DIV_SCALE_F32;
756 const SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
768 return CurDAG->SelectNodeTo(N, Opc, VT, MVT::i1, Ops);
771 bool AMDGPUDAGToDAGISel::isDSOffsetLegal(const SDValue &Base, unsigned Offset,
772 unsigned OffsetBits) const {
773 const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
774 if ((OffsetBits == 16 && !isUInt<16>(Offset)) ||
775 (OffsetBits == 8 && !isUInt<8>(Offset)))
778 if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS)
781 // On Southern Islands instruction with a negative base value and an offset
782 // don't seem to work.
783 return CurDAG->SignBitIsZero(Base);
786 bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
787 SDValue &Offset) const {
788 if (CurDAG->isBaseWithConstantOffset(Addr)) {
789 SDValue N0 = Addr.getOperand(0);
790 SDValue N1 = Addr.getOperand(1);
791 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
792 if (isDSOffsetLegal(N0, C1->getSExtValue(), 16)) {
802 Offset = CurDAG->getTargetConstant(0, MVT::i16);
806 bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
808 SDValue &Offset1) const {
809 if (CurDAG->isBaseWithConstantOffset(Addr)) {
810 SDValue N0 = Addr.getOperand(0);
811 SDValue N1 = Addr.getOperand(1);
812 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
813 unsigned DWordOffset0 = C1->getZExtValue() / 4;
814 unsigned DWordOffset1 = DWordOffset0 + 1;
816 if (isDSOffsetLegal(N0, DWordOffset1, 8)) {
818 Offset0 = CurDAG->getTargetConstant(DWordOffset0, MVT::i8);
819 Offset1 = CurDAG->getTargetConstant(DWordOffset1, MVT::i8);
826 Offset0 = CurDAG->getTargetConstant(0, MVT::i8);
827 Offset1 = CurDAG->getTargetConstant(1, MVT::i8);
831 static SDValue wrapAddr64Rsrc(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
832 return SDValue(DAG->getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::v4i32,
836 static bool isLegalMUBUFImmOffset(const ConstantSDNode *Imm) {
837 return isUInt<12>(Imm->getZExtValue());
840 void AMDGPUDAGToDAGISel::SelectMUBUF(SDValue Addr, SDValue &Ptr,
841 SDValue &VAddr, SDValue &SOffset,
842 SDValue &Offset, SDValue &Offen,
843 SDValue &Idxen, SDValue &Addr64,
844 SDValue &GLC, SDValue &SLC,
845 SDValue &TFE) const {
848 GLC = CurDAG->getTargetConstant(0, MVT::i1);
849 SLC = CurDAG->getTargetConstant(0, MVT::i1);
850 TFE = CurDAG->getTargetConstant(0, MVT::i1);
852 Idxen = CurDAG->getTargetConstant(0, MVT::i1);
853 Offen = CurDAG->getTargetConstant(0, MVT::i1);
854 Addr64 = CurDAG->getTargetConstant(0, MVT::i1);
855 SOffset = CurDAG->getTargetConstant(0, MVT::i32);
857 if (CurDAG->isBaseWithConstantOffset(Addr)) {
858 SDValue N0 = Addr.getOperand(0);
859 SDValue N1 = Addr.getOperand(1);
860 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
862 if (isLegalMUBUFImmOffset(C1)) {
864 if (N0.getOpcode() == ISD::ADD) {
865 // (add (add N2, N3), C1) -> addr64
866 SDValue N2 = N0.getOperand(0);
867 SDValue N3 = N0.getOperand(1);
868 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
871 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
875 // (add N0, C1) -> offset
876 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
878 Offset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
882 if (Addr.getOpcode() == ISD::ADD) {
883 // (add N0, N1) -> addr64
884 SDValue N0 = Addr.getOperand(0);
885 SDValue N1 = Addr.getOperand(1);
886 Addr64 = CurDAG->getTargetConstant(1, MVT::i1);
889 Offset = CurDAG->getTargetConstant(0, MVT::i16);
893 // default case -> offset
894 VAddr = CurDAG->getTargetConstant(0, MVT::i32);
896 Offset = CurDAG->getTargetConstant(0, MVT::i16);
900 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
902 SDValue &Offset) const {
903 SDValue Ptr, SOffset, Offen, Idxen, Addr64, GLC, SLC, TFE;
905 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
908 ConstantSDNode *C = cast<ConstantSDNode>(Addr64);
909 if (C->getSExtValue()) {
911 SRsrc = wrapAddr64Rsrc(CurDAG, DL, Ptr);
917 bool AMDGPUDAGToDAGISel::SelectMUBUFAddr64(SDValue Addr, SDValue &SRsrc,
918 SDValue &VAddr, SDValue &Offset,
919 SDValue &SLC) const {
920 SLC = CurDAG->getTargetConstant(0, MVT::i1);
922 return SelectMUBUFAddr64(Addr, SRsrc, VAddr, Offset);
925 static SDValue buildRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr,
926 uint32_t RsrcDword1, uint64_t RsrcDword2And3) {
928 SDValue PtrLo = DAG->getTargetExtractSubreg(AMDGPU::sub0, DL, MVT::i32, Ptr);
929 SDValue PtrHi = DAG->getTargetExtractSubreg(AMDGPU::sub1, DL, MVT::i32, Ptr);
931 PtrHi = SDValue(DAG->getMachineNode(AMDGPU::S_OR_B32, DL, MVT::i32, PtrHi,
932 DAG->getConstant(RsrcDword1, MVT::i32)), 0);
934 SDValue DataLo = DAG->getTargetConstant(
935 RsrcDword2And3 & APInt::getAllOnesValue(32).getZExtValue(), MVT::i32);
936 SDValue DataHi = DAG->getTargetConstant(RsrcDword2And3 >> 32, MVT::i32);
938 const SDValue Ops[] = { PtrLo, PtrHi, DataLo, DataHi };
939 return SDValue(DAG->getMachineNode(AMDGPU::SI_BUFFER_RSRC, DL,
940 MVT::v4i32, Ops), 0);
943 /// \brief Return a resource descriptor with the 'Add TID' bit enabled
944 /// The TID (Thread ID) is multipled by the stride value (bits [61:48]
945 /// of the resource descriptor) to create an offset, which is added to the
947 static SDValue buildScratchRSRC(SelectionDAG *DAG, SDLoc DL, SDValue Ptr) {
949 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT | AMDGPU::RSRC_TID_ENABLE |
952 return buildRSRC(DAG, DL, Ptr, 0, Rsrc);
955 bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
956 SDValue &VAddr, SDValue &SOffset,
957 SDValue &ImmOffset) const {
960 MachineFunction &MF = CurDAG->getMachineFunction();
961 const SIRegisterInfo *TRI =
962 static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
963 MachineRegisterInfo &MRI = MF.getRegInfo();
964 const SITargetLowering& Lowering =
965 *static_cast<const SITargetLowering*>(getTargetLowering());
967 unsigned ScratchPtrReg =
968 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
969 unsigned ScratchOffsetReg =
970 TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
971 Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
972 ScratchOffsetReg, MVT::i32);
974 Rsrc = buildScratchRSRC(CurDAG, DL,
975 CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
976 MRI.getLiveInVirtReg(ScratchPtrReg), MVT::i64));
977 SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
978 MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
981 if (CurDAG->isBaseWithConstantOffset(Addr)) {
982 SDValue N1 = Addr.getOperand(1);
983 ConstantSDNode *C1 = cast<ConstantSDNode>(N1);
985 if (isLegalMUBUFImmOffset(C1)) {
986 VAddr = Addr.getOperand(0);
987 ImmOffset = CurDAG->getTargetConstant(C1->getZExtValue(), MVT::i16);
993 if ((Addr.getOpcode() == ISD::ADD || Addr.getOpcode() == ISD::OR) &&
994 isa<FrameIndexSDNode>(Addr.getOperand(0))) {
995 VAddr = Addr.getOperand(1);
996 ImmOffset = Addr.getOperand(0);
1001 if (isa<FrameIndexSDNode>(Addr)) {
1002 VAddr = SDValue(CurDAG->getMachineNode(AMDGPU::V_MOV_B32_e32, DL, MVT::i32,
1003 CurDAG->getConstant(0, MVT::i32)), 0);
1010 ImmOffset = CurDAG->getTargetConstant(0, MVT::i16);
1014 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1015 SDValue &SOffset, SDValue &Offset,
1016 SDValue &GLC, SDValue &SLC,
1017 SDValue &TFE) const {
1018 SDValue Ptr, VAddr, Offen, Idxen, Addr64;
1020 SelectMUBUF(Addr, Ptr, VAddr, SOffset, Offset, Offen, Idxen, Addr64,
1023 if (!cast<ConstantSDNode>(Offen)->getSExtValue() &&
1024 !cast<ConstantSDNode>(Idxen)->getSExtValue() &&
1025 !cast<ConstantSDNode>(Addr64)->getSExtValue()) {
1026 uint64_t Rsrc = AMDGPU::RSRC_DATA_FORMAT |
1027 APInt::getAllOnesValue(32).getZExtValue(); // Size
1029 SRsrc = buildRSRC(CurDAG, DL, Ptr, 0, Rsrc);
1035 bool AMDGPUDAGToDAGISel::SelectMUBUFOffset(SDValue Addr, SDValue &SRsrc,
1036 SDValue &Soffset, SDValue &Offset,
1037 SDValue &GLC) const {
1040 return SelectMUBUFOffset(Addr, SRsrc, Soffset, Offset, GLC, SLC, TFE);
1043 // FIXME: This is incorrect and only enough to be able to compile.
1044 SDNode *AMDGPUDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
1045 AddrSpaceCastSDNode *ASC = cast<AddrSpaceCastSDNode>(N);
1048 assert(Subtarget.hasFlatAddressSpace() &&
1049 "addrspacecast only supported with flat address space!");
1051 assert((ASC->getSrcAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS &&
1052 ASC->getDestAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS) &&
1053 "Cannot cast address space to / from constant address!");
1055 assert((ASC->getSrcAddressSpace() == AMDGPUAS::FLAT_ADDRESS ||
1056 ASC->getDestAddressSpace() == AMDGPUAS::FLAT_ADDRESS) &&
1057 "Can only cast to / from flat address space!");
1059 // The flat instructions read the address as the index of the VGPR holding the
1060 // address, so casting should just be reinterpreting the base VGPR, so just
1061 // insert trunc / bitcast / zext.
1063 SDValue Src = ASC->getOperand(0);
1064 EVT DestVT = ASC->getValueType(0);
1065 EVT SrcVT = Src.getValueType();
1067 unsigned SrcSize = SrcVT.getSizeInBits();
1068 unsigned DestSize = DestVT.getSizeInBits();
1070 if (SrcSize > DestSize) {
1071 assert(SrcSize == 64 && DestSize == 32);
1072 return CurDAG->getMachineNode(
1073 TargetOpcode::EXTRACT_SUBREG,
1077 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32));
1081 if (DestSize > SrcSize) {
1082 assert(SrcSize == 32 && DestSize == 64);
1084 SDValue RC = CurDAG->getTargetConstant(AMDGPU::VSrc_64RegClassID, MVT::i32);
1086 const SDValue Ops[] = {
1089 CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
1090 SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
1091 CurDAG->getConstant(0, MVT::i32)), 0),
1092 CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
1095 return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1096 SDLoc(N), N->getValueType(0), Ops);
1099 assert(SrcSize == 64 && DestSize == 64);
1100 return CurDAG->getNode(ISD::BITCAST, DL, DestVT, Src).getNode();
1103 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
1104 SDValue &SrcMods) const {
1110 if (Src.getOpcode() == ISD::FNEG) {
1111 Mods |= SISrcMods::NEG;
1112 Src = Src.getOperand(0);
1115 if (Src.getOpcode() == ISD::FABS) {
1116 Mods |= SISrcMods::ABS;
1117 Src = Src.getOperand(0);
1120 SrcMods = CurDAG->getTargetConstant(Mods, MVT::i32);
1125 bool AMDGPUDAGToDAGISel::SelectVOP3Mods0(SDValue In, SDValue &Src,
1126 SDValue &SrcMods, SDValue &Clamp,
1127 SDValue &Omod) const {
1128 // FIXME: Handle Clamp and Omod
1129 Clamp = CurDAG->getTargetConstant(0, MVT::i32);
1130 Omod = CurDAG->getTargetConstant(0, MVT::i32);
1132 return SelectVOP3Mods(In, Src, SrcMods);
1135 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
1136 const AMDGPUTargetLowering& Lowering =
1137 *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());
1138 bool IsModified = false;
1141 // Go over all selected nodes and try to fold them a bit more
1142 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
1143 E = CurDAG->allnodes_end(); I != E; ++I) {
1147 MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
1151 SDNode *ResNode = Lowering.PostISelFolding(MachineNode, *CurDAG);
1152 if (ResNode != Node) {
1153 ReplaceUses(Node, ResNode);
1157 CurDAG->RemoveDeadNodes();
1158 } while (IsModified);