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);
73 SDValue SITargetLowering::LowerFormalArguments(
75 CallingConv::ID CallConv,
77 const SmallVectorImpl<ISD::InputArg> &Ins,
78 DebugLoc DL, SelectionDAG &DAG,
79 SmallVectorImpl<SDValue> &InVals) const {
81 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
83 MachineFunction &MF = DAG.getMachineFunction();
84 FunctionType *FType = MF.getFunction()->getFunctionType();
85 SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
87 assert(CallConv == CallingConv::C);
89 SmallVector<ISD::InputArg, 16> Splits;
92 for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
93 const ISD::InputArg &Arg = Ins[i];
95 // First check if it's a PS input addr
96 if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg()) {
98 assert((PSInputNum <= 15) && "Too many PS inputs!");
101 // We can savely skip PS inputs
107 Info->PSInputAddr |= 1 << PSInputNum++;
110 // Second split vertices into their elements
111 if (Arg.VT.isVector()) {
112 ISD::InputArg NewArg = Arg;
113 NewArg.Flags.setSplit();
114 NewArg.VT = Arg.VT.getVectorElementType();
116 // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
117 // three or five element vertex only needs three or five registers,
118 // NOT four or eigth.
119 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
120 unsigned NumElements = ParamType->getVectorNumElements();
122 for (unsigned j = 0; j != NumElements; ++j) {
123 Splits.push_back(NewArg);
124 NewArg.PartOffset += NewArg.VT.getStoreSize();
128 Splits.push_back(Arg);
132 SmallVector<CCValAssign, 16> ArgLocs;
133 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
134 getTargetMachine(), ArgLocs, *DAG.getContext());
136 // At least one interpolation mode must be enabled or else the GPU will hang.
137 if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
138 Info->PSInputAddr |= 1;
139 CCInfo.AllocateReg(AMDGPU::VGPR0);
140 CCInfo.AllocateReg(AMDGPU::VGPR1);
143 AnalyzeFormalArguments(CCInfo, Splits);
145 for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
147 if (Skipped & (1 << i)) {
148 InVals.push_back(SDValue());
152 CCValAssign &VA = ArgLocs[ArgIdx++];
153 assert(VA.isRegLoc() && "Parameter must be in a register!");
155 unsigned Reg = VA.getLocReg();
156 MVT VT = VA.getLocVT();
158 if (VT == MVT::i64) {
159 // For now assume it is a pointer
160 Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
161 &AMDGPU::SReg_64RegClass);
162 Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
163 InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
167 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
169 Reg = MF.addLiveIn(Reg, RC);
170 SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
172 const ISD::InputArg &Arg = Ins[i];
173 if (Arg.VT.isVector()) {
175 // Build a vector from the registers
176 Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
177 unsigned NumElements = ParamType->getVectorNumElements();
179 SmallVector<SDValue, 4> Regs;
181 for (unsigned j = 1; j != NumElements; ++j) {
182 Reg = ArgLocs[ArgIdx++].getLocReg();
183 Reg = MF.addLiveIn(Reg, RC);
184 Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
187 // Fill up the missing vector elements
188 NumElements = Arg.VT.getVectorNumElements() - NumElements;
189 for (unsigned j = 0; j != NumElements; ++j)
190 Regs.push_back(DAG.getUNDEF(VT));
192 InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT,
193 Regs.data(), Regs.size()));
197 InVals.push_back(Val);
202 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
203 MachineInstr * MI, MachineBasicBlock * BB) const {
204 MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
205 MachineBasicBlock::iterator I = MI;
207 switch (MI->getOpcode()) {
209 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
210 case AMDGPU::BRANCH: return BB;
212 LowerSI_WQM(MI, *BB, I, MRI);
218 void SITargetLowering::LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
219 MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
220 BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WQM_B64), AMDGPU::EXEC)
221 .addReg(AMDGPU::EXEC);
223 MI->eraseFromParent();
226 EVT SITargetLowering::getSetCCResultType(EVT VT) const {
230 //===----------------------------------------------------------------------===//
231 // Custom DAG Lowering Operations
232 //===----------------------------------------------------------------------===//
234 SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
235 switch (Op.getOpcode()) {
236 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
237 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
238 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
243 /// \brief Helper function for LowerBRCOND
244 static SDNode *findUser(SDValue Value, unsigned Opcode) {
246 SDNode *Parent = Value.getNode();
247 for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
250 if (I.getUse().get() != Value)
253 if (I->getOpcode() == Opcode)
259 /// This transforms the control flow intrinsics to get the branch destination as
260 /// last parameter, also switches branch target with BR if the need arise
261 SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
262 SelectionDAG &DAG) const {
264 DebugLoc DL = BRCOND.getDebugLoc();
266 SDNode *Intr = BRCOND.getOperand(1).getNode();
267 SDValue Target = BRCOND.getOperand(2);
270 if (Intr->getOpcode() == ISD::SETCC) {
271 // As long as we negate the condition everything is fine
272 SDNode *SetCC = Intr;
273 assert(SetCC->getConstantOperandVal(1) == 1);
274 assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
276 Intr = SetCC->getOperand(0).getNode();
279 // Get the target from BR if we don't negate the condition
280 BR = findUser(BRCOND, ISD::BR);
281 Target = BR->getOperand(1);
284 assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
286 // Build the result and
287 SmallVector<EVT, 4> Res;
288 for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
289 Res.push_back(Intr->getValueType(i));
291 // operands of the new intrinsic call
292 SmallVector<SDValue, 4> Ops;
293 Ops.push_back(BRCOND.getOperand(0));
294 for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
295 Ops.push_back(Intr->getOperand(i));
296 Ops.push_back(Target);
298 // build the new intrinsic call
299 SDNode *Result = DAG.getNode(
300 Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
301 DAG.getVTList(Res.data(), Res.size()), Ops.data(), Ops.size()).getNode();
304 // Give the branch instruction our target
309 DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops, 2);
312 SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
314 // Copy the intrinsic results to registers
315 for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
316 SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
320 Chain = DAG.getCopyToReg(
322 CopyToReg->getOperand(1),
323 SDValue(Result, i - 1),
326 DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
329 // Remove the old intrinsic from the chain
330 DAG.ReplaceAllUsesOfValueWith(
331 SDValue(Intr, Intr->getNumValues() - 1),
332 Intr->getOperand(0));
337 SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
338 SDValue LHS = Op.getOperand(0);
339 SDValue RHS = Op.getOperand(1);
340 SDValue True = Op.getOperand(2);
341 SDValue False = Op.getOperand(3);
342 SDValue CC = Op.getOperand(4);
343 EVT VT = Op.getValueType();
344 DebugLoc DL = Op.getDebugLoc();
346 // Possible Min/Max pattern
347 SDValue MinMax = LowerMinMax(Op, DAG);
348 if (MinMax.getNode()) {
352 SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
353 return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
356 //===----------------------------------------------------------------------===//
357 // Custom DAG optimizations
358 //===----------------------------------------------------------------------===//
360 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
361 DAGCombinerInfo &DCI) const {
362 SelectionDAG &DAG = DCI.DAG;
363 DebugLoc DL = N->getDebugLoc();
364 EVT VT = N->getValueType(0);
366 switch (N->getOpcode()) {
368 case ISD::SELECT_CC: {
370 ConstantSDNode *True, *False;
371 // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
372 if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
373 && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
374 && True->isAllOnesValue()
375 && False->isNullValue()
377 return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
378 N->getOperand(1), N->getOperand(4));
384 SDValue Arg0 = N->getOperand(0);
385 SDValue Arg1 = N->getOperand(1);
386 SDValue CC = N->getOperand(2);
387 ConstantSDNode * C = NULL;
388 ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
390 // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
392 && Arg0.getOpcode() == ISD::SIGN_EXTEND
393 && Arg0.getOperand(0).getValueType() == MVT::i1
394 && (C = dyn_cast<ConstantSDNode>(Arg1))
396 && CCOp == ISD::SETNE) {
397 return SimplifySetCC(VT, Arg0.getOperand(0),
398 DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
406 /// \brief Test if RegClass is one of the VSrc classes
407 static bool isVSrc(unsigned RegClass) {
408 return AMDGPU::VSrc_32RegClassID == RegClass ||
409 AMDGPU::VSrc_64RegClassID == RegClass;
412 /// \brief Test if RegClass is one of the SSrc classes
413 static bool isSSrc(unsigned RegClass) {
414 return AMDGPU::SSrc_32RegClassID == RegClass ||
415 AMDGPU::SSrc_64RegClassID == RegClass;
418 /// \brief Analyze the possible immediate value Op
420 /// Returns -1 if it isn't an immediate, 0 if it's and inline immediate
421 /// and the immediate value if it's a literal immediate
422 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
429 if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N))
430 Imm.I = Node->getSExtValue();
431 else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N))
432 Imm.F = Node->getValueAPF().convertToFloat();
434 return -1; // It isn't an immediate
436 if ((Imm.I >= -16 && Imm.I <= 64) ||
437 Imm.F == 0.5f || Imm.F == -0.5f ||
438 Imm.F == 1.0f || Imm.F == -1.0f ||
439 Imm.F == 2.0f || Imm.F == -2.0f ||
440 Imm.F == 4.0f || Imm.F == -4.0f)
441 return 0; // It's an inline immediate
443 return Imm.I; // It's a literal immediate
446 /// \brief Try to fold an immediate directly into an instruction
447 bool SITargetLowering::foldImm(SDValue &Operand, int32_t &Immediate,
448 bool &ScalarSlotUsed) const {
450 MachineSDNode *Mov = dyn_cast<MachineSDNode>(Operand);
451 if (Mov == 0 || !TII->isMov(Mov->getMachineOpcode()))
454 const SDValue &Op = Mov->getOperand(0);
455 int32_t Value = analyzeImmediate(Op.getNode());
457 // Not an immediate at all
460 } else if (Value == 0) {
461 // Inline immediates can always be fold
465 } else if (Value == Immediate) {
466 // Already fold literal immediate
470 } else if (!ScalarSlotUsed && !Immediate) {
471 // Fold this literal immediate
472 ScalarSlotUsed = true;
482 /// \brief Does "Op" fit into register class "RegClass" ?
483 bool SITargetLowering::fitsRegClass(SelectionDAG &DAG, SDValue &Op,
484 unsigned RegClass) const {
486 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
487 SDNode *Node = Op.getNode();
490 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(Node)) {
491 const MCInstrDesc &Desc = TII->get(MN->getMachineOpcode());
492 OpClass = Desc.OpInfo[Op.getResNo()].RegClass;
494 } else if (Node->getOpcode() == ISD::CopyFromReg) {
495 RegisterSDNode *Reg = cast<RegisterSDNode>(Node->getOperand(1).getNode());
496 OpClass = MRI.getRegClass(Reg->getReg())->getID();
504 return TRI->getRegClass(RegClass)->hasSubClassEq(TRI->getRegClass(OpClass));
507 /// \brief Make sure that we don't exeed the number of allowed scalars
508 void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
510 bool &ScalarSlotUsed) const {
512 // First map the operands register class to a destination class
513 if (RegClass == AMDGPU::VSrc_32RegClassID)
514 RegClass = AMDGPU::VReg_32RegClassID;
515 else if (RegClass == AMDGPU::VSrc_64RegClassID)
516 RegClass = AMDGPU::VReg_64RegClassID;
520 // Nothing todo if they fit naturaly
521 if (fitsRegClass(DAG, Operand, RegClass))
524 // If the scalar slot isn't used yet use it now
525 if (!ScalarSlotUsed) {
526 ScalarSlotUsed = true;
530 // This is a conservative aproach, it is possible that we can't determine
531 // the correct register class and copy too often, but better save than sorry.
532 SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
533 SDNode *Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, DebugLoc(),
534 Operand.getValueType(), Operand, RC);
535 Operand = SDValue(Node, 0);
538 SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
539 SelectionDAG &DAG) const {
541 // Original encoding (either e32 or e64)
542 int Opcode = Node->getMachineOpcode();
543 const MCInstrDesc *Desc = &TII->get(Opcode);
545 unsigned NumDefs = Desc->getNumDefs();
546 unsigned NumOps = Desc->getNumOperands();
548 // e64 version if available, -1 otherwise
549 int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
550 const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? 0 : &TII->get(OpcodeE64);
552 assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
553 assert(!DescE64 || DescE64->getNumOperands() == (NumOps + 4));
555 int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
556 bool HaveVSrc = false, HaveSSrc = false;
558 // First figure out what we alread have in this instruction
559 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
560 i != e && Op < NumOps; ++i, ++Op) {
562 unsigned RegClass = Desc->OpInfo[Op].RegClass;
563 if (isVSrc(RegClass))
565 else if (isSSrc(RegClass))
570 int32_t Imm = analyzeImmediate(Node->getOperand(i).getNode());
571 if (Imm != -1 && Imm != 0) {
577 // If we neither have VSrc nor SSrc it makes no sense to continue
578 if (!HaveVSrc && !HaveSSrc)
581 // No scalar allowed when we have both VSrc and SSrc
582 bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
584 // Second go over the operands and try to fold them
585 std::vector<SDValue> Ops;
586 bool Promote2e64 = false;
587 for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
588 i != e && Op < NumOps; ++i, ++Op) {
590 const SDValue &Operand = Node->getOperand(i);
591 Ops.push_back(Operand);
593 // Already folded immediate ?
594 if (isa<ConstantSDNode>(Operand.getNode()) ||
595 isa<ConstantFPSDNode>(Operand.getNode()))
598 // Is this a VSrc or SSrc operand ?
599 unsigned RegClass = Desc->OpInfo[Op].RegClass;
600 if (!isVSrc(RegClass) && !isSSrc(RegClass)) {
602 if (i == 1 && Desc->isCommutable() &&
603 fitsRegClass(DAG, Ops[0], RegClass) &&
604 foldImm(Ops[1], Immediate, ScalarSlotUsed)) {
606 assert(isVSrc(Desc->OpInfo[NumDefs].RegClass) ||
607 isSSrc(Desc->OpInfo[NumDefs].RegClass));
609 // Swap commutable operands
610 SDValue Tmp = Ops[1];
614 } else if (DescE64 && !Immediate) {
615 // Test if it makes sense to switch to e64 encoding
617 RegClass = DescE64->OpInfo[Op].RegClass;
619 if ((isVSrc(RegClass) || isSSrc(RegClass)) &&
620 foldImm(Ops[i], TmpImm, ScalarSlotUsed)) {
631 // Try to fold the immediates
632 if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
633 // Folding didn't worked, make sure we don't hit the SReg limit
634 ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
639 // Add the modifier flags while promoting
640 for (unsigned i = 0; i < 4; ++i)
641 Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
644 // Add optional chain and glue
645 for (unsigned i = NumOps - NumDefs, e = Node->getNumOperands(); i < e; ++i)
646 Ops.push_back(Node->getOperand(i));
648 // Either create a complete new or update the current instruction
650 return DAG.getMachineNode(OpcodeE64, Node->getDebugLoc(),
651 Node->getVTList(), Ops.data(), Ops.size());
653 return DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());