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/IR/Function.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
30 SITargetLowering::SITargetLowering(TargetMachine &TM) :
31 AMDGPUTargetLowering(TM),
32 TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())),
33 TRI(TM.getRegisterInfo()) {
35 addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
36 addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
38 addRegisterClass(MVT::v16i8, &AMDGPU::SReg_128RegClass);
39 addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
40 addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
42 addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
43 addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
45 addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
47 addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
48 addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
50 addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
51 addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
53 addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
54 addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
56 addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
57 addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
59 computeRegisterProperties();
61 setOperationAction(ISD::ADD, MVT::i64, Legal);
62 setOperationAction(ISD::ADD, MVT::i32, Legal);
64 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
65 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
67 setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
68 setTargetDAGCombine(ISD::SELECT_CC);
70 setTargetDAGCombine(ISD::SETCC);
72 setSchedulingPreference(Sched::Source);
75 SDValue SITargetLowering::LowerFormalArguments(
77 CallingConv::ID CallConv,
79 const SmallVectorImpl<ISD::InputArg> &Ins,
80 DebugLoc DL, SelectionDAG &DAG,
81 SmallVectorImpl<SDValue> &InVals) const {
83 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
85 MachineFunction &MF = DAG.getMachineFunction();
86 FunctionType *FType = MF.getFunction()->getFunctionType();
87 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
89 assert(CallConv == CallingConv::C);
91 SmallVector<ISD::InputArg, 16> Splits;
94 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
95 const ISD::InputArg &Arg = Ins[i];
97 // First check if it's a PS input addr
98 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg()) {
100 assert((PSInputNum <= 15) && "Too many PS inputs!");
103 // We can savely skip PS inputs
109 Info->PSInputAddr |= 1 << PSInputNum++;
112 // Second split vertices into their elements
113 if (Arg.VT.isVector()) {
114 ISD::InputArg NewArg = Arg;
115 NewArg.Flags.setSplit();
116 NewArg.VT = Arg.VT.getVectorElementType();
118 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
119 // three or five element vertex only needs three or five registers,
120 // NOT four or eigth.
121 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
122 unsigned NumElements = ParamType->getVectorNumElements();
124 for (unsigned j = 0; j != NumElements; ++j) {
125 Splits.push_back(NewArg);
126 NewArg.PartOffset += NewArg.VT.getStoreSize();
130 Splits.push_back(Arg);
134 SmallVector<CCValAssign, 16> ArgLocs;
135 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
136 getTargetMachine(), ArgLocs, *DAG.getContext());
138 // At least one interpolation mode must be enabled or else the GPU will hang.
139 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
140 Info->PSInputAddr |= 1;
141 CCInfo.AllocateReg(AMDGPU::VGPR0);
142 CCInfo.AllocateReg(AMDGPU::VGPR1);
145 AnalyzeFormalArguments(CCInfo, Splits);
147 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
149 if (Skipped & (1 << i)) {
150 InVals.push_back(SDValue());
154 CCValAssign &VA = ArgLocs[ArgIdx++];
155 assert(VA.isRegLoc() && "Parameter must be in a register!");
157 unsigned Reg = VA.getLocReg();
158 MVT VT = VA.getLocVT();
160 if (VT == MVT::i64) {
161 // For now assume it is a pointer
162 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
163 &AMDGPU::SReg_64RegClass);
164 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
165 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
169 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
171 Reg = MF.addLiveIn(Reg, RC);
172 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
174 const ISD::InputArg &Arg = Ins[i];
175 if (Arg.VT.isVector()) {
177 // Build a vector from the registers
178 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
179 unsigned NumElements = ParamType->getVectorNumElements();
181 SmallVector<SDValue, 4> Regs;
183 for (unsigned j = 1; j != NumElements; ++j) {
184 Reg = ArgLocs[ArgIdx++].getLocReg();
185 Reg = MF.addLiveIn(Reg, RC);
186 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
189 // Fill up the missing vector elements
190 NumElements = Arg.VT.getVectorNumElements() - NumElements;
191 for (unsigned j = 0; j != NumElements; ++j)
192 Regs.push_back(DAG.getUNDEF(VT));
194 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT,
195 Regs.data(), Regs.size()));
199 InVals.push_back(Val);
204 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
205 MachineInstr * MI, MachineBasicBlock * BB) const {
206 MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
207 MachineBasicBlock::iterator I = MI;
209 switch (MI->getOpcode()) {
211 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
212 case AMDGPU::BRANCH: return BB;
214 LowerSI_WQM(MI, *BB, I, MRI);
220 void SITargetLowering::LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
221 MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
222 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WQM_B64), AMDGPU::EXEC)
223 .addReg(AMDGPU::EXEC);
225 MI->eraseFromParent();
228 EVT SITargetLowering::getSetCCResultType(EVT VT) const {
232 MVT SITargetLowering::getScalarShiftAmountTy(EVT VT) const {
236 //===----------------------------------------------------------------------===//
237 // Custom DAG Lowering Operations
238 //===----------------------------------------------------------------------===//
240 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
241 switch (Op.getOpcode()) {
242 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
243 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
244 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
249 /// \brief Helper function for LowerBRCOND
250 static SDNode *findUser(SDValue Value, unsigned Opcode) {
252 SDNode *Parent = Value.getNode();
253 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
256 if (I.getUse().get() != Value)
259 if (I->getOpcode() == Opcode)
265 /// This transforms the control flow intrinsics to get the branch destination as
266 /// last parameter, also switches branch target with BR if the need arise
267 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
268 SelectionDAG &DAG) const {
270 DebugLoc DL = BRCOND.getDebugLoc();
272 SDNode *Intr = BRCOND.getOperand(1).getNode();
273 SDValue Target = BRCOND.getOperand(2);
276 if (Intr->getOpcode() == ISD::SETCC) {
277 // As long as we negate the condition everything is fine
278 SDNode *SetCC = Intr;
279 assert(SetCC->getConstantOperandVal(1) == 1);
280 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
282 Intr = SetCC->getOperand(0).getNode();
285 // Get the target from BR if we don't negate the condition
286 BR = findUser(BRCOND, ISD::BR);
287 Target = BR->getOperand(1);
290 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
292 // Build the result and
293 SmallVector<EVT, 4> Res;
294 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
295 Res.push_back(Intr->getValueType(i));
297 // operands of the new intrinsic call
298 SmallVector<SDValue, 4> Ops;
299 Ops.push_back(BRCOND.getOperand(0));
300 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
301 Ops.push_back(Intr->getOperand(i));
302 Ops.push_back(Target);
304 // build the new intrinsic call
305 SDNode *Result = DAG.getNode(
306 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
307 DAG.getVTList(Res.data(), Res.size()), Ops.data(), Ops.size()).getNode();
310 // Give the branch instruction our target
315 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops, 2);
318 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
320 // Copy the intrinsic results to registers
321 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
322 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
326 Chain = DAG.getCopyToReg(
328 CopyToReg->getOperand(1),
329 SDValue(Result, i - 1),
332 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
335 // Remove the old intrinsic from the chain
336 DAG.ReplaceAllUsesOfValueWith(
337 SDValue(Intr, Intr->getNumValues() - 1),
338 Intr->getOperand(0));
343 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
344 SDValue LHS = Op.getOperand(0);
345 SDValue RHS = Op.getOperand(1);
346 SDValue True = Op.getOperand(2);
347 SDValue False = Op.getOperand(3);
348 SDValue CC = Op.getOperand(4);
349 EVT VT = Op.getValueType();
350 DebugLoc DL = Op.getDebugLoc();
352 // Possible Min/Max pattern
353 SDValue MinMax = LowerMinMax(Op, DAG);
354 if (MinMax.getNode()) {
358 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
359 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
362 //===----------------------------------------------------------------------===//
363 // Custom DAG optimizations
364 //===----------------------------------------------------------------------===//
366 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
367 DAGCombinerInfo &DCI) const {
368 SelectionDAG &DAG = DCI.DAG;
369 DebugLoc DL = N->getDebugLoc();
370 EVT VT = N->getValueType(0);
372 switch (N->getOpcode()) {
374 case ISD::SELECT_CC: {
376 ConstantSDNode *True, *False;
377 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
378 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
379 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
380 && True->isAllOnesValue()
381 && False->isNullValue()
383 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
384 N->getOperand(1), N->getOperand(4));
390 SDValue Arg0 = N->getOperand(0);
391 SDValue Arg1 = N->getOperand(1);
392 SDValue CC = N->getOperand(2);
393 ConstantSDNode * C = NULL;
394 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
396 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
398 && Arg0.getOpcode() == ISD::SIGN_EXTEND
399 && Arg0.getOperand(0).getValueType() == MVT::i1
400 && (C = dyn_cast<ConstantSDNode>(Arg1))
402 && CCOp == ISD::SETNE) {
403 return SimplifySetCC(VT, Arg0.getOperand(0),
404 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
412 /// \brief Test if RegClass is one of the VSrc classes
413 static bool isVSrc(unsigned RegClass) {
414 return AMDGPU::VSrc_32RegClassID == RegClass ||
415 AMDGPU::VSrc_64RegClassID == RegClass;
418 /// \brief Test if RegClass is one of the SSrc classes
419 static bool isSSrc(unsigned RegClass) {
420 return AMDGPU::SSrc_32RegClassID == RegClass ||
421 AMDGPU::SSrc_64RegClassID == RegClass;
424 /// \brief Analyze the possible immediate value Op
426 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
427 /// and the immediate value if it's a literal immediate
428 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
435 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N))
436 Imm.I = Node->getSExtValue();
437 else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N))
438 Imm.F = Node->getValueAPF().convertToFloat();
440 return -1; // It isn't an immediate
442 if ((Imm.I >= -16 && Imm.I <= 64) ||
443 Imm.F == 0.5f || Imm.F == -0.5f ||
444 Imm.F == 1.0f || Imm.F == -1.0f ||
445 Imm.F == 2.0f || Imm.F == -2.0f ||
446 Imm.F == 4.0f || Imm.F == -4.0f)
447 return 0; // It's an inline immediate
449 return Imm.I; // It's a literal immediate
452 /// \brief Try to fold an immediate directly into an instruction
453 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
454 bool &ScalarSlotUsed) const {
456 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
457 if (Mov == 0 || !TII->isMov(Mov->getMachineOpcode()))
460 const SDValue &Op = Mov->getOperand(0);
461 int32_t Value = analyzeImmediate(Op.getNode());
463 // Not an immediate at all
466 } else if (Value == 0) {
467 // Inline immediates can always be fold
471 } else if (Value == Immediate) {
472 // Already fold literal immediate
476 } else if (!ScalarSlotUsed && !Immediate) {
477 // Fold this literal immediate
478 ScalarSlotUsed = true;
488 /// \brief Does "Op" fit into register class "RegClass" ?
489 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, SDValue &Op,
490 unsigned RegClass) const {
492 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
493 SDNode *Node = Op.getNode();
496 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
497 const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
498 OpClass = Desc.OpInfo[Op.getResNo()].RegClass;
500 } else if (Node->getOpcode() == ISD::CopyFromReg) {
501 RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
502 OpClass = MRI.getRegClass(Reg->getReg())->getID();
510 return TRI->getRegClass(RegClass)->hasSubClassEq(TRI->getRegClass(OpClass));
513 /// \brief Make sure that we don't exeed the number of allowed scalars
514 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
516 bool &ScalarSlotUsed) const {
518 // First map the operands register class to a destination class
519 if (RegClass == AMDGPU::VSrc_32RegClassID)
520 RegClass = AMDGPU::VReg_32RegClassID;
521 else if (RegClass == AMDGPU::VSrc_64RegClassID)
522 RegClass = AMDGPU::VReg_64RegClassID;
526 // Nothing todo if they fit naturaly
527 if (fitsRegClass(DAG, Operand, RegClass))
530 // If the scalar slot isn't used yet use it now
531 if (!ScalarSlotUsed) {
532 ScalarSlotUsed = true;
536 // This is a conservative aproach, it is possible that we can't determine
537 // the correct register class and copy too often, but better save than sorry.
538 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
539 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DebugLoc(),
540 Operand.getValueType(), Operand, RC);
541 Operand = SDValue(Node, 0);
544 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
545 SelectionDAG &DAG) const {
547 // Original encoding (either e32 or e64)
548 int Opcode = Node->getMachineOpcode();
549 const MCInstrDesc *Desc = &TII->get(Opcode);
551 unsigned NumDefs = Desc->getNumDefs();
552 unsigned NumOps = Desc->getNumOperands();
554 // e64 version if available, -1 otherwise
555 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
556 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? 0 : &TII->get(OpcodeE64);
558 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
559 assert(!DescE64 || DescE64->getNumOperands() == (NumOps + 4));
561 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
562 bool HaveVSrc = false, HaveSSrc = false;
564 // First figure out what we alread have in this instruction
565 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
566 i != e && Op < NumOps; ++i, ++Op) {
568 unsigned RegClass = Desc->OpInfo[Op].RegClass;
569 if (isVSrc(RegClass))
571 else if (isSSrc(RegClass))
576 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
577 if (Imm != -1 && Imm != 0) {
583 // If we neither have VSrc nor SSrc it makes no sense to continue
584 if (!HaveVSrc && !HaveSSrc)
587 // No scalar allowed when we have both VSrc and SSrc
588 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
590 // Second go over the operands and try to fold them
591 std::vector<SDValue> Ops;
592 bool Promote2e64 = false;
593 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
594 i != e && Op < NumOps; ++i, ++Op) {
596 const SDValue &Operand = Node->getOperand(i);
597 Ops.push_back(Operand);
599 // Already folded immediate ?
600 if (isa<ConstantSDNode>(Operand.getNode()) ||
601 isa<ConstantFPSDNode>(Operand.getNode()))
604 // Is this a VSrc or SSrc operand ?
605 unsigned RegClass = Desc->OpInfo[Op].RegClass;
606 if (!isVSrc(RegClass) && !isSSrc(RegClass)) {
608 if (i == 1 && Desc->isCommutable() &&
609 fitsRegClass(DAG, Ops[0], RegClass) &&
610 foldImm(Ops[1], Immediate, ScalarSlotUsed)) {
612 assert(isVSrc(Desc->OpInfo[NumDefs].RegClass) ||
613 isSSrc(Desc->OpInfo[NumDefs].RegClass));
615 // Swap commutable operands
616 SDValue Tmp = Ops[1];
620 } else if (DescE64 && !Immediate) {
621 // Test if it makes sense to switch to e64 encoding
623 RegClass = DescE64->OpInfo[Op].RegClass;
625 if ((isVSrc(RegClass) || isSSrc(RegClass)) &&
626 foldImm(Ops[i], TmpImm, ScalarSlotUsed)) {
637 // Try to fold the immediates
638 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
639 // Folding didn't worked, make sure we don't hit the SReg limit
640 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
645 // Add the modifier flags while promoting
646 for (unsigned i = 0; i < 4; ++i)
647 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
650 // Add optional chain and glue
651 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
652 Ops.push_back(Node->getOperand(i));
654 // Either create a complete new or update the current instruction
656 return DAG.getMachineNode(OpcodeE64, Node->getDebugLoc(),
657 Node->getVTList(), Ops.data(), Ops.size());
659 return DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());