1 //===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
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 Custom DAG lowering for SI
13 //===----------------------------------------------------------------------===//
15 #include "SIISelLowering.h"
18 #include "AMDILIntrinsicInfo.h"
19 #include "SIInstrInfo.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "SIRegisterInfo.h"
22 #include "llvm/CodeGen/CallingConvLower.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/IR/Function.h"
28 const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL;
32 SITargetLowering::SITargetLowering(TargetMachine &TM) :
33 AMDGPUTargetLowering(TM),
34 TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())),
35 TRI(TM.getRegisterInfo()) {
37 addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
38 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
40 addRegisterClass(MVT::v16i8, &AMDGPU::SReg_128RegClass);
41 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
42 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
44 addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
45 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
47 addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
49 addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
50 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
52 addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
53 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
54 addRegisterClass(MVT::i128, &AMDGPU::SReg_128RegClass);
56 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
57 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
59 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
60 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
62 computeRegisterProperties();
64 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
65 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
66 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
67 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
69 setOperationAction(ISD::ADD, MVT::i64, Legal);
70 setOperationAction(ISD::ADD, MVT::i32, Legal);
72 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
73 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
75 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
77 setTargetDAGCombine(ISD::SELECT_CC);
79 setTargetDAGCombine(ISD::SETCC);
81 setSchedulingPreference(Sched::RegPressure);
84 SDValue SITargetLowering::LowerFormalArguments(
86 CallingConv::ID CallConv,
88 const SmallVectorImpl<ISD::InputArg> &Ins,
89 SDLoc DL, SelectionDAG &DAG,
90 SmallVectorImpl<SDValue> &InVals) const {
92 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
94 MachineFunction &MF = DAG.getMachineFunction();
95 FunctionType *FType = MF.getFunction()->getFunctionType();
96 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
98 assert(CallConv == CallingConv::C);
100 SmallVector<ISD::InputArg, 16> Splits;
101 uint32_t Skipped = 0;
103 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
104 const ISD::InputArg &Arg = Ins[i];
106 // First check if it's a PS input addr
107 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg()) {
109 assert((PSInputNum <= 15) && "Too many PS inputs!");
112 // We can savely skip PS inputs
118 Info->PSInputAddr |= 1 << PSInputNum++;
121 // Second split vertices into their elements
122 if (Arg.VT.isVector()) {
123 ISD::InputArg NewArg = Arg;
124 NewArg.Flags.setSplit();
125 NewArg.VT = Arg.VT.getVectorElementType();
127 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
128 // three or five element vertex only needs three or five registers,
129 // NOT four or eigth.
130 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
131 unsigned NumElements = ParamType->getVectorNumElements();
133 for (unsigned j = 0; j != NumElements; ++j) {
134 Splits.push_back(NewArg);
135 NewArg.PartOffset += NewArg.VT.getStoreSize();
139 Splits.push_back(Arg);
143 SmallVector<CCValAssign, 16> ArgLocs;
144 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
145 getTargetMachine(), ArgLocs, *DAG.getContext());
147 // At least one interpolation mode must be enabled or else the GPU will hang.
148 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
149 Info->PSInputAddr |= 1;
150 CCInfo.AllocateReg(AMDGPU::VGPR0);
151 CCInfo.AllocateReg(AMDGPU::VGPR1);
154 AnalyzeFormalArguments(CCInfo, Splits);
156 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
158 const ISD::InputArg &Arg = Ins[i];
159 if (Skipped & (1 << i)) {
160 InVals.push_back(DAG.getUNDEF(Arg.VT));
164 CCValAssign &VA = ArgLocs[ArgIdx++];
165 assert(VA.isRegLoc() && "Parameter must be in a register!");
167 unsigned Reg = VA.getLocReg();
168 MVT VT = VA.getLocVT();
170 if (VT == MVT::i64) {
171 // For now assume it is a pointer
172 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
173 &AMDGPU::SReg_64RegClass);
174 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
175 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
179 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
181 Reg = MF.addLiveIn(Reg, RC);
182 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
184 if (Arg.VT.isVector()) {
186 // Build a vector from the registers
187 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
188 unsigned NumElements = ParamType->getVectorNumElements();
190 SmallVector<SDValue, 4> Regs;
192 for (unsigned j = 1; j != NumElements; ++j) {
193 Reg = ArgLocs[ArgIdx++].getLocReg();
194 Reg = MF.addLiveIn(Reg, RC);
195 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
198 // Fill up the missing vector elements
199 NumElements = Arg.VT.getVectorNumElements() - NumElements;
200 for (unsigned j = 0; j != NumElements; ++j)
201 Regs.push_back(DAG.getUNDEF(VT));
203 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT,
204 Regs.data(), Regs.size()));
208 InVals.push_back(Val);
213 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
214 MachineInstr * MI, MachineBasicBlock * BB) const {
216 MachineBasicBlock::iterator I = *MI;
218 switch (MI->getOpcode()) {
220 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
221 case AMDGPU::BRANCH: return BB;
222 case AMDGPU::SI_ADDR64_RSRC: {
223 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
224 unsigned SuperReg = MI->getOperand(0).getReg();
225 unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
226 unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
227 unsigned SubRegHiHi = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
228 unsigned SubRegHiLo = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
229 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B64), SubRegLo)
230 .addOperand(MI->getOperand(1));
231 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiLo)
233 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), SubRegHiHi)
234 .addImm(RSRC_DATA_FORMAT >> 32);
235 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SubRegHi)
237 .addImm(AMDGPU::sub0)
239 .addImm(AMDGPU::sub1);
240 BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE), SuperReg)
242 .addImm(AMDGPU::sub0_sub1)
244 .addImm(AMDGPU::sub2_sub3);
245 MI->eraseFromParent();
252 EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
256 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
260 //===----------------------------------------------------------------------===//
261 // Custom DAG Lowering Operations
262 //===----------------------------------------------------------------------===//
264 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
265 switch (Op.getOpcode()) {
266 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
267 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
268 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
273 /// \brief Helper function for LowerBRCOND
274 static SDNode *findUser(SDValue Value, unsigned Opcode) {
276 SDNode *Parent = Value.getNode();
277 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
280 if (I.getUse().get() != Value)
283 if (I->getOpcode() == Opcode)
289 /// This transforms the control flow intrinsics to get the branch destination as
290 /// last parameter, also switches branch target with BR if the need arise
291 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
292 SelectionDAG &DAG) const {
296 SDNode *Intr = BRCOND.getOperand(1).getNode();
297 SDValue Target = BRCOND.getOperand(2);
300 if (Intr->getOpcode() == ISD::SETCC) {
301 // As long as we negate the condition everything is fine
302 SDNode *SetCC = Intr;
303 assert(SetCC->getConstantOperandVal(1) == 1);
304 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
306 Intr = SetCC->getOperand(0).getNode();
309 // Get the target from BR if we don't negate the condition
310 BR = findUser(BRCOND, ISD::BR);
311 Target = BR->getOperand(1);
314 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
316 // Build the result and
317 SmallVector<EVT, 4> Res;
318 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
319 Res.push_back(Intr->getValueType(i));
321 // operands of the new intrinsic call
322 SmallVector<SDValue, 4> Ops;
323 Ops.push_back(BRCOND.getOperand(0));
324 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
325 Ops.push_back(Intr->getOperand(i));
326 Ops.push_back(Target);
328 // build the new intrinsic call
329 SDNode *Result = DAG.getNode(
330 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
331 DAG.getVTList(Res.data(), Res.size()), Ops.data(), Ops.size()).getNode();
334 // Give the branch instruction our target
339 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops, 2);
342 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
344 // Copy the intrinsic results to registers
345 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
346 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
350 Chain = DAG.getCopyToReg(
352 CopyToReg->getOperand(1),
353 SDValue(Result, i - 1),
356 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
359 // Remove the old intrinsic from the chain
360 DAG.ReplaceAllUsesOfValueWith(
361 SDValue(Intr, Intr->getNumValues() - 1),
362 Intr->getOperand(0));
367 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
368 SDValue LHS = Op.getOperand(0);
369 SDValue RHS = Op.getOperand(1);
370 SDValue True = Op.getOperand(2);
371 SDValue False = Op.getOperand(3);
372 SDValue CC = Op.getOperand(4);
373 EVT VT = Op.getValueType();
376 // Possible Min/Max pattern
377 SDValue MinMax = LowerMinMax(Op, DAG);
378 if (MinMax.getNode()) {
382 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
383 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
386 //===----------------------------------------------------------------------===//
387 // Custom DAG optimizations
388 //===----------------------------------------------------------------------===//
390 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
391 DAGCombinerInfo &DCI) const {
392 SelectionDAG &DAG = DCI.DAG;
394 EVT VT = N->getValueType(0);
396 switch (N->getOpcode()) {
398 case ISD::SELECT_CC: {
400 ConstantSDNode *True, *False;
401 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
402 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
403 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
404 && True->isAllOnesValue()
405 && False->isNullValue()
407 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
408 N->getOperand(1), N->getOperand(4));
414 SDValue Arg0 = N->getOperand(0);
415 SDValue Arg1 = N->getOperand(1);
416 SDValue CC = N->getOperand(2);
417 ConstantSDNode * C = NULL;
418 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
420 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
422 && Arg0.getOpcode() == ISD::SIGN_EXTEND
423 && Arg0.getOperand(0).getValueType() == MVT::i1
424 && (C = dyn_cast<ConstantSDNode>(Arg1))
426 && CCOp == ISD::SETNE) {
427 return SimplifySetCC(VT, Arg0.getOperand(0),
428 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
436 /// \brief Test if RegClass is one of the VSrc classes
437 static bool isVSrc(unsigned RegClass) {
438 return AMDGPU::VSrc_32RegClassID == RegClass ||
439 AMDGPU::VSrc_64RegClassID == RegClass;
442 /// \brief Test if RegClass is one of the SSrc classes
443 static bool isSSrc(unsigned RegClass) {
444 return AMDGPU::SSrc_32RegClassID == RegClass ||
445 AMDGPU::SSrc_64RegClassID == RegClass;
448 /// \brief Analyze the possible immediate value Op
450 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
451 /// and the immediate value if it's a literal immediate
452 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
459 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
460 if (Node->getZExtValue() >> 32) {
463 Imm.I = Node->getSExtValue();
464 } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N))
465 Imm.F = Node->getValueAPF().convertToFloat();
467 return -1; // It isn't an immediate
469 if ((Imm.I >= -16 && Imm.I <= 64) ||
470 Imm.F == 0.5f || Imm.F == -0.5f ||
471 Imm.F == 1.0f || Imm.F == -1.0f ||
472 Imm.F == 2.0f || Imm.F == -2.0f ||
473 Imm.F == 4.0f || Imm.F == -4.0f)
474 return 0; // It's an inline immediate
476 return Imm.I; // It's a literal immediate
479 /// \brief Try to fold an immediate directly into an instruction
480 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
481 bool &ScalarSlotUsed) const {
483 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
484 if (Mov == 0 || !TII->isMov(Mov->getMachineOpcode()))
487 const SDValue &Op = Mov->getOperand(0);
488 int32_t Value = analyzeImmediate(Op.getNode());
490 // Not an immediate at all
493 } else if (Value == 0) {
494 // Inline immediates can always be fold
498 } else if (Value == Immediate) {
499 // Already fold literal immediate
503 } else if (!ScalarSlotUsed && !Immediate) {
504 // Fold this literal immediate
505 ScalarSlotUsed = true;
515 /// \brief Does "Op" fit into register class "RegClass" ?
516 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, const SDValue &Op,
517 unsigned RegClass) const {
519 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
520 SDNode *Node = Op.getNode();
522 const TargetRegisterClass *OpClass;
523 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
524 const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
525 int OpClassID = Desc.OpInfo[Op.getResNo()].RegClass;
527 OpClass = getRegClassFor(Op.getSimpleValueType());
529 OpClass = TRI->getRegClass(OpClassID);
531 } else if (Node->getOpcode() == ISD::CopyFromReg) {
532 RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
533 OpClass = MRI.getRegClass(Reg->getReg());
538 return TRI->getRegClass(RegClass)->hasSubClassEq(OpClass);
541 /// \brief Make sure that we don't exeed the number of allowed scalars
542 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
544 bool &ScalarSlotUsed) const {
546 // First map the operands register class to a destination class
547 if (RegClass == AMDGPU::VSrc_32RegClassID)
548 RegClass = AMDGPU::VReg_32RegClassID;
549 else if (RegClass == AMDGPU::VSrc_64RegClassID)
550 RegClass = AMDGPU::VReg_64RegClassID;
554 // Nothing todo if they fit naturaly
555 if (fitsRegClass(DAG, Operand, RegClass))
558 // If the scalar slot isn't used yet use it now
559 if (!ScalarSlotUsed) {
560 ScalarSlotUsed = true;
564 // This is a conservative aproach, it is possible that we can't determine
565 // the correct register class and copy too often, but better save than sorry.
566 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
567 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
568 Operand.getValueType(), Operand, RC);
569 Operand = SDValue(Node, 0);
572 /// \brief Try to fold the Nodes operands into the Node
573 SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
574 SelectionDAG &DAG) const {
576 // Original encoding (either e32 or e64)
577 int Opcode = Node->getMachineOpcode();
578 const MCInstrDesc *Desc = &TII->get(Opcode);
580 unsigned NumDefs = Desc->getNumDefs();
581 unsigned NumOps = Desc->getNumOperands();
583 // Commuted opcode if available
584 int OpcodeRev = Desc->isCommutable() ? TII->commuteOpcode(Opcode) : -1;
585 const MCInstrDesc *DescRev = OpcodeRev == -1 ? 0 : &TII->get(OpcodeRev);
587 assert(!DescRev || DescRev->getNumDefs() == NumDefs);
588 assert(!DescRev || DescRev->getNumOperands() == NumOps);
590 // e64 version if available, -1 otherwise
591 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
592 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? 0 : &TII->get(OpcodeE64);
594 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
595 assert(!DescE64 || DescE64->getNumOperands() == (NumOps + 4));
597 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
598 bool HaveVSrc = false, HaveSSrc = false;
600 // First figure out what we alread have in this instruction
601 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
602 i != e && Op < NumOps; ++i, ++Op) {
604 unsigned RegClass = Desc->OpInfo[Op].RegClass;
605 if (isVSrc(RegClass))
607 else if (isSSrc(RegClass))
612 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
613 if (Imm != -1 && Imm != 0) {
619 // If we neither have VSrc nor SSrc it makes no sense to continue
620 if (!HaveVSrc && !HaveSSrc)
623 // No scalar allowed when we have both VSrc and SSrc
624 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
626 // Second go over the operands and try to fold them
627 std::vector<SDValue> Ops;
628 bool Promote2e64 = false;
629 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
630 i != e && Op < NumOps; ++i, ++Op) {
632 const SDValue &Operand = Node->getOperand(i);
633 Ops.push_back(Operand);
635 // Already folded immediate ?
636 if (isa<ConstantSDNode>(Operand.getNode()) ||
637 isa<ConstantFPSDNode>(Operand.getNode()))
640 // Is this a VSrc or SSrc operand ?
641 unsigned RegClass = Desc->OpInfo[Op].RegClass;
642 if (isVSrc(RegClass) || isSSrc(RegClass)) {
643 // Try to fold the immediates
644 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
645 // Folding didn't worked, make sure we don't hit the SReg limit
646 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
651 if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
653 unsigned OtherRegClass = Desc->OpInfo[NumDefs].RegClass;
654 assert(isVSrc(OtherRegClass) || isSSrc(OtherRegClass));
656 // Test if it makes sense to swap operands
657 if (foldImm(Ops[1], Immediate, ScalarSlotUsed) ||
658 (!fitsRegClass(DAG, Ops[1], RegClass) &&
659 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
661 // Swap commutable operands
662 SDValue Tmp = Ops[1];
672 if (DescE64 && !Immediate) {
674 // Test if it makes sense to switch to e64 encoding
675 unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
676 if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
680 if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
681 (!fitsRegClass(DAG, Ops[i], RegClass) &&
682 fitsRegClass(DAG, Ops[1], OtherRegClass))) {
684 // Switch to e64 encoding
694 // Add the modifier flags while promoting
695 for (unsigned i = 0; i < 4; ++i)
696 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
699 // Add optional chain and glue
700 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
701 Ops.push_back(Node->getOperand(i));
703 // Nodes that have a glue result are not CSE'd by getMachineNode(), so in
704 // this case a brand new node is always be created, even if the operands
705 // are the same as before. So, manually check if anything has been changed.
706 if (Desc->Opcode == Opcode) {
707 bool Changed = false;
708 for (unsigned i = 0, e = Node->getNumOperands(); i < e; ++i) {
709 if (Ops[i].getNode() != Node->getOperand(i).getNode()) {
719 // Create a complete new instruction
720 return DAG.getMachineNode(Desc->Opcode, SDLoc(Node), Node->getVTList(), Ops);
723 /// \brief Helper function for adjustWritemask
724 static unsigned SubIdx2Lane(unsigned Idx) {
727 case AMDGPU::sub0: return 0;
728 case AMDGPU::sub1: return 1;
729 case AMDGPU::sub2: return 2;
730 case AMDGPU::sub3: return 3;
734 /// \brief Adjust the writemask of MIMG instructions
735 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
736 SelectionDAG &DAG) const {
737 SDNode *Users[4] = { };
738 unsigned Writemask = 0, Lane = 0;
740 // Try to figure out the used register components
741 for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
744 // Abort if we can't understand the usage
745 if (!I->isMachineOpcode() ||
746 I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
749 Lane = SubIdx2Lane(I->getConstantOperandVal(1));
751 // Abort if we have more than one user per component
756 Writemask |= 1 << Lane;
759 // Abort if all components are used
760 if (Writemask == 0xf)
763 // Adjust the writemask in the node
764 std::vector<SDValue> Ops;
765 Ops.push_back(DAG.getTargetConstant(Writemask, MVT::i32));
766 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
767 Ops.push_back(Node->getOperand(i));
768 Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());
770 // If we only got one lane, replace it with a copy
771 if (Writemask == (1U << Lane)) {
772 SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
773 SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
774 SDLoc(), Users[Lane]->getValueType(0),
775 SDValue(Node, 0), RC);
776 DAG.ReplaceAllUsesWith(Users[Lane], Copy);
780 // Update the users of the node with the new indices
781 for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
783 SDNode *User = Users[i];
787 SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
788 DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
792 case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
793 case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
794 case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
799 /// \brief Fold the instructions after slecting them
800 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
801 SelectionDAG &DAG) const {
803 if (AMDGPU::isMIMG(Node->getMachineOpcode()) != -1)
804 adjustWritemask(Node, DAG);
806 return foldOperands(Node, DAG);
809 /// \brief Assign the register class depending on the number of
810 /// bits set in the writemask
811 void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
812 SDNode *Node) const {
813 if (AMDGPU::isMIMG(MI->getOpcode()) == -1)
816 unsigned VReg = MI->getOperand(0).getReg();
817 unsigned Writemask = MI->getOperand(1).getImm();
818 unsigned BitsSet = 0;
819 for (unsigned i = 0; i < 4; ++i)
820 BitsSet += Writemask & (1 << i) ? 1 : 0;
822 const TargetRegisterClass *RC;
825 case 1: RC = &AMDGPU::VReg_32RegClass; break;
826 case 2: RC = &AMDGPU::VReg_64RegClass; break;
827 case 3: RC = &AMDGPU::VReg_96RegClass; break;
830 MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
831 MRI.setRegClass(VReg, RC);