1 //===-- R600ISelLowering.cpp - R600 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 R600
13 //===----------------------------------------------------------------------===//
15 #include "R600ISelLowering.h"
16 #include "R600Defines.h"
17 #include "R600InstrInfo.h"
18 #include "R600MachineFunctionInfo.h"
19 #include "llvm/CodeGen/CallingConvLower.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/SelectionDAG.h"
24 #include "llvm/IR/Argument.h"
25 #include "llvm/IR/Function.h"
29 R600TargetLowering::R600TargetLowering(TargetMachine &TM) :
30 AMDGPUTargetLowering(TM),
31 Gen(TM.getSubtarget<AMDGPUSubtarget>().getGeneration()) {
32 addRegisterClass(MVT::v4f32, &AMDGPU::R600_Reg128RegClass);
33 addRegisterClass(MVT::f32, &AMDGPU::R600_Reg32RegClass);
34 addRegisterClass(MVT::v4i32, &AMDGPU::R600_Reg128RegClass);
35 addRegisterClass(MVT::i32, &AMDGPU::R600_Reg32RegClass);
36 addRegisterClass(MVT::v2f32, &AMDGPU::R600_Reg64RegClass);
37 addRegisterClass(MVT::v2i32, &AMDGPU::R600_Reg64RegClass);
39 computeRegisterProperties();
41 // Set condition code actions
42 setCondCodeAction(ISD::SETO, MVT::f32, Expand);
43 setCondCodeAction(ISD::SETUO, MVT::f32, Expand);
44 setCondCodeAction(ISD::SETLT, MVT::f32, Expand);
45 setCondCodeAction(ISD::SETLE, MVT::f32, Expand);
46 setCondCodeAction(ISD::SETOLT, MVT::f32, Expand);
47 setCondCodeAction(ISD::SETOLE, MVT::f32, Expand);
48 setCondCodeAction(ISD::SETONE, MVT::f32, Expand);
49 setCondCodeAction(ISD::SETUEQ, MVT::f32, Expand);
50 setCondCodeAction(ISD::SETUGE, MVT::f32, Expand);
51 setCondCodeAction(ISD::SETUGT, MVT::f32, Expand);
52 setCondCodeAction(ISD::SETULT, MVT::f32, Expand);
53 setCondCodeAction(ISD::SETULE, MVT::f32, Expand);
55 setCondCodeAction(ISD::SETLE, MVT::i32, Expand);
56 setCondCodeAction(ISD::SETLT, MVT::i32, Expand);
57 setCondCodeAction(ISD::SETULE, MVT::i32, Expand);
58 setCondCodeAction(ISD::SETULT, MVT::i32, Expand);
60 setOperationAction(ISD::FCOS, MVT::f32, Custom);
61 setOperationAction(ISD::FSIN, MVT::f32, Custom);
63 setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
64 setOperationAction(ISD::SETCC, MVT::v2i32, Expand);
66 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
67 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
69 setOperationAction(ISD::FSUB, MVT::f32, Expand);
71 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
72 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
73 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
75 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
76 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
78 setOperationAction(ISD::SETCC, MVT::i32, Expand);
79 setOperationAction(ISD::SETCC, MVT::f32, Expand);
80 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
82 setOperationAction(ISD::SELECT, MVT::i32, Expand);
83 setOperationAction(ISD::SELECT, MVT::f32, Expand);
84 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
85 setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
86 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
87 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
89 // Legalize loads and stores to the private address space.
90 setOperationAction(ISD::LOAD, MVT::i32, Custom);
91 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
92 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
93 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
94 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
95 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
96 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
97 setOperationAction(ISD::STORE, MVT::i8, Custom);
98 setOperationAction(ISD::STORE, MVT::i32, Custom);
99 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
100 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
101 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
102 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
104 setOperationAction(ISD::LOAD, MVT::i32, Custom);
105 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
106 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
108 setTargetDAGCombine(ISD::FP_ROUND);
109 setTargetDAGCombine(ISD::FP_TO_SINT);
110 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
111 setTargetDAGCombine(ISD::SELECT_CC);
112 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
114 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
116 setBooleanContents(ZeroOrNegativeOneBooleanContent);
117 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
118 setSchedulingPreference(Sched::Source);
121 MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
122 MachineInstr * MI, MachineBasicBlock * BB) const {
123 MachineFunction * MF = BB->getParent();
124 MachineRegisterInfo &MRI = MF->getRegInfo();
125 MachineBasicBlock::iterator I = *MI;
126 const R600InstrInfo *TII =
127 static_cast<const R600InstrInfo*>(MF->getTarget().getInstrInfo());
129 switch (MI->getOpcode()) {
131 if (TII->isLDSInstr(MI->getOpcode()) &&
132 TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst) != -1) {
133 int DstIdx = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
134 assert(DstIdx != -1);
135 MachineInstrBuilder NewMI;
136 if (!MRI.use_empty(MI->getOperand(DstIdx).getReg())) {
137 NewMI = BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()),
139 TII->buildDefaultInstruction(*BB, I, AMDGPU::MOV,
140 MI->getOperand(0).getReg(),
143 NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
144 TII->get(AMDGPU::getLDSNoRetOp(MI->getOpcode())));
146 for (unsigned i = 1, e = MI->getNumOperands(); i < e; ++i) {
147 NewMI.addOperand(MI->getOperand(i));
150 return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
153 case AMDGPU::CLAMP_R600: {
154 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
156 MI->getOperand(0).getReg(),
157 MI->getOperand(1).getReg());
158 TII->addFlag(NewMI, 0, MO_FLAG_CLAMP);
162 case AMDGPU::FABS_R600: {
163 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
165 MI->getOperand(0).getReg(),
166 MI->getOperand(1).getReg());
167 TII->addFlag(NewMI, 0, MO_FLAG_ABS);
171 case AMDGPU::FNEG_R600: {
172 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
174 MI->getOperand(0).getReg(),
175 MI->getOperand(1).getReg());
176 TII->addFlag(NewMI, 0, MO_FLAG_NEG);
180 case AMDGPU::MASK_WRITE: {
181 unsigned maskedRegister = MI->getOperand(0).getReg();
182 assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
183 MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
184 TII->addFlag(defInstr, 0, MO_FLAG_MASK);
188 case AMDGPU::MOV_IMM_F32:
189 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
190 MI->getOperand(1).getFPImm()->getValueAPF()
191 .bitcastToAPInt().getZExtValue());
193 case AMDGPU::MOV_IMM_I32:
194 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
195 MI->getOperand(1).getImm());
197 case AMDGPU::CONST_COPY: {
198 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, MI, AMDGPU::MOV,
199 MI->getOperand(0).getReg(), AMDGPU::ALU_CONST);
200 TII->setImmOperand(NewMI, AMDGPU::OpName::src0_sel,
201 MI->getOperand(1).getImm());
205 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
206 case AMDGPU::RAT_WRITE_CACHELESS_64_eg:
207 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
208 unsigned EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN) ? 1 : 0;
210 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
211 .addOperand(MI->getOperand(0))
212 .addOperand(MI->getOperand(1))
213 .addImm(EOP); // Set End of program bit
218 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
219 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
220 MachineOperand &RID = MI->getOperand(4);
221 MachineOperand &SID = MI->getOperand(5);
222 unsigned TextureId = MI->getOperand(6).getImm();
223 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
224 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
236 case 8: // ShadowRect
247 case 11: // Shadow1DArray
251 case 12: // Shadow2DArray
255 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
256 .addOperand(MI->getOperand(3))
274 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
275 .addOperand(MI->getOperand(2))
293 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
294 .addOperand(MI->getOperand(0))
295 .addOperand(MI->getOperand(1))
313 .addReg(T0, RegState::Implicit)
314 .addReg(T1, RegState::Implicit);
318 case AMDGPU::TXD_SHADOW: {
319 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
320 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
321 MachineOperand &RID = MI->getOperand(4);
322 MachineOperand &SID = MI->getOperand(5);
323 unsigned TextureId = MI->getOperand(6).getImm();
324 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
325 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
337 case 8: // ShadowRect
348 case 11: // Shadow1DArray
352 case 12: // Shadow2DArray
357 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
358 .addOperand(MI->getOperand(3))
376 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
377 .addOperand(MI->getOperand(2))
395 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
396 .addOperand(MI->getOperand(0))
397 .addOperand(MI->getOperand(1))
415 .addReg(T0, RegState::Implicit)
416 .addReg(T1, RegState::Implicit);
421 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
422 .addOperand(MI->getOperand(0));
425 case AMDGPU::BRANCH_COND_f32: {
426 MachineInstr *NewMI =
427 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
428 AMDGPU::PREDICATE_BIT)
429 .addOperand(MI->getOperand(1))
430 .addImm(OPCODE_IS_NOT_ZERO)
432 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
433 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
434 .addOperand(MI->getOperand(0))
435 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
439 case AMDGPU::BRANCH_COND_i32: {
440 MachineInstr *NewMI =
441 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
442 AMDGPU::PREDICATE_BIT)
443 .addOperand(MI->getOperand(1))
444 .addImm(OPCODE_IS_NOT_ZERO_INT)
446 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
447 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
448 .addOperand(MI->getOperand(0))
449 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
453 case AMDGPU::EG_ExportSwz:
454 case AMDGPU::R600_ExportSwz: {
455 // Instruction is left unmodified if its not the last one of its type
456 bool isLastInstructionOfItsType = true;
457 unsigned InstExportType = MI->getOperand(1).getImm();
458 for (MachineBasicBlock::iterator NextExportInst = llvm::next(I),
459 EndBlock = BB->end(); NextExportInst != EndBlock;
460 NextExportInst = llvm::next(NextExportInst)) {
461 if (NextExportInst->getOpcode() == AMDGPU::EG_ExportSwz ||
462 NextExportInst->getOpcode() == AMDGPU::R600_ExportSwz) {
463 unsigned CurrentInstExportType = NextExportInst->getOperand(1)
465 if (CurrentInstExportType == InstExportType) {
466 isLastInstructionOfItsType = false;
471 bool EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN)? 1 : 0;
472 if (!EOP && !isLastInstructionOfItsType)
474 unsigned CfInst = (MI->getOpcode() == AMDGPU::EG_ExportSwz)? 84 : 40;
475 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
476 .addOperand(MI->getOperand(0))
477 .addOperand(MI->getOperand(1))
478 .addOperand(MI->getOperand(2))
479 .addOperand(MI->getOperand(3))
480 .addOperand(MI->getOperand(4))
481 .addOperand(MI->getOperand(5))
482 .addOperand(MI->getOperand(6))
487 case AMDGPU::RETURN: {
488 // RETURN instructions must have the live-out registers as implicit uses,
489 // otherwise they appear dead.
490 R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
491 MachineInstrBuilder MIB(*MF, MI);
492 for (unsigned i = 0, e = MFI->LiveOuts.size(); i != e; ++i)
493 MIB.addReg(MFI->LiveOuts[i], RegState::Implicit);
498 MI->eraseFromParent();
502 //===----------------------------------------------------------------------===//
503 // Custom DAG Lowering Operations
504 //===----------------------------------------------------------------------===//
506 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
507 MachineFunction &MF = DAG.getMachineFunction();
508 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
509 switch (Op.getOpcode()) {
510 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
512 case ISD::FSIN: return LowerTrig(Op, DAG);
513 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
514 case ISD::STORE: return LowerSTORE(Op, DAG);
515 case ISD::LOAD: return LowerLOAD(Op, DAG);
516 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
517 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
518 case ISD::INTRINSIC_VOID: {
519 SDValue Chain = Op.getOperand(0);
520 unsigned IntrinsicID =
521 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
522 switch (IntrinsicID) {
523 case AMDGPUIntrinsic::AMDGPU_store_output: {
524 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
525 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
526 MFI->LiveOuts.push_back(Reg);
527 return DAG.getCopyToReg(Chain, SDLoc(Op), Reg, Op.getOperand(2));
529 case AMDGPUIntrinsic::R600_store_swizzle: {
530 const SDValue Args[8] = {
532 Op.getOperand(2), // Export Value
533 Op.getOperand(3), // ArrayBase
534 Op.getOperand(4), // Type
535 DAG.getConstant(0, MVT::i32), // SWZ_X
536 DAG.getConstant(1, MVT::i32), // SWZ_Y
537 DAG.getConstant(2, MVT::i32), // SWZ_Z
538 DAG.getConstant(3, MVT::i32) // SWZ_W
540 return DAG.getNode(AMDGPUISD::EXPORT, SDLoc(Op), Op.getValueType(),
544 // default for switch(IntrinsicID)
547 // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
550 case ISD::INTRINSIC_WO_CHAIN: {
551 unsigned IntrinsicID =
552 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
553 EVT VT = Op.getValueType();
555 switch(IntrinsicID) {
556 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
557 case AMDGPUIntrinsic::R600_load_input: {
558 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
559 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
560 MachineFunction &MF = DAG.getMachineFunction();
561 MachineRegisterInfo &MRI = MF.getRegInfo();
563 return DAG.getCopyFromReg(DAG.getEntryNode(),
564 SDLoc(DAG.getEntryNode()), Reg, VT);
567 case AMDGPUIntrinsic::R600_interp_input: {
568 int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
569 int ijb = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
570 MachineSDNode *interp;
572 const MachineFunction &MF = DAG.getMachineFunction();
573 const R600InstrInfo *TII =
574 static_cast<const R600InstrInfo*>(MF.getTarget().getInstrInfo());
575 interp = DAG.getMachineNode(AMDGPU::INTERP_VEC_LOAD, DL,
576 MVT::v4f32, DAG.getTargetConstant(slot / 4 , MVT::i32));
577 return DAG.getTargetExtractSubreg(
578 TII->getRegisterInfo().getSubRegFromChannel(slot % 4),
579 DL, MVT::f32, SDValue(interp, 0));
581 MachineFunction &MF = DAG.getMachineFunction();
582 MachineRegisterInfo &MRI = MF.getRegInfo();
583 unsigned RegisterI = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb);
584 unsigned RegisterJ = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb + 1);
585 MRI.addLiveIn(RegisterI);
586 MRI.addLiveIn(RegisterJ);
587 SDValue RegisterINode = DAG.getCopyFromReg(DAG.getEntryNode(),
588 SDLoc(DAG.getEntryNode()), RegisterI, MVT::f32);
589 SDValue RegisterJNode = DAG.getCopyFromReg(DAG.getEntryNode(),
590 SDLoc(DAG.getEntryNode()), RegisterJ, MVT::f32);
593 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_XY, DL,
594 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
595 RegisterJNode, RegisterINode);
597 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_ZW, DL,
598 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
599 RegisterJNode, RegisterINode);
600 return SDValue(interp, slot % 2);
602 case AMDGPUIntrinsic::R600_interp_xy:
603 case AMDGPUIntrinsic::R600_interp_zw: {
604 int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
605 MachineSDNode *interp;
606 SDValue RegisterINode = Op.getOperand(2);
607 SDValue RegisterJNode = Op.getOperand(3);
609 if (IntrinsicID == AMDGPUIntrinsic::R600_interp_xy)
610 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_XY, DL,
611 MVT::f32, MVT::f32, DAG.getTargetConstant(slot, MVT::i32),
612 RegisterJNode, RegisterINode);
614 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_ZW, DL,
615 MVT::f32, MVT::f32, DAG.getTargetConstant(slot, MVT::i32),
616 RegisterJNode, RegisterINode);
617 return DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v2f32,
618 SDValue(interp, 0), SDValue(interp, 1));
620 case AMDGPUIntrinsic::R600_tex:
621 case AMDGPUIntrinsic::R600_texc:
622 case AMDGPUIntrinsic::R600_txl:
623 case AMDGPUIntrinsic::R600_txlc:
624 case AMDGPUIntrinsic::R600_txb:
625 case AMDGPUIntrinsic::R600_txbc:
626 case AMDGPUIntrinsic::R600_txf:
627 case AMDGPUIntrinsic::R600_txq:
628 case AMDGPUIntrinsic::R600_ddx:
629 case AMDGPUIntrinsic::R600_ddy:
630 case AMDGPUIntrinsic::R600_ldptr: {
632 switch (IntrinsicID) {
633 case AMDGPUIntrinsic::R600_tex:
636 case AMDGPUIntrinsic::R600_texc:
639 case AMDGPUIntrinsic::R600_txl:
642 case AMDGPUIntrinsic::R600_txlc:
645 case AMDGPUIntrinsic::R600_txb:
648 case AMDGPUIntrinsic::R600_txbc:
651 case AMDGPUIntrinsic::R600_txf:
654 case AMDGPUIntrinsic::R600_txq:
657 case AMDGPUIntrinsic::R600_ddx:
660 case AMDGPUIntrinsic::R600_ddy:
663 case AMDGPUIntrinsic::R600_ldptr:
667 llvm_unreachable("Unknow Texture Operation");
670 SDValue TexArgs[19] = {
671 DAG.getConstant(TextureOp, MVT::i32),
673 DAG.getConstant(0, MVT::i32),
674 DAG.getConstant(1, MVT::i32),
675 DAG.getConstant(2, MVT::i32),
676 DAG.getConstant(3, MVT::i32),
680 DAG.getConstant(0, MVT::i32),
681 DAG.getConstant(1, MVT::i32),
682 DAG.getConstant(2, MVT::i32),
683 DAG.getConstant(3, MVT::i32),
691 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs, 19);
693 case AMDGPUIntrinsic::AMDGPU_dp4: {
695 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
696 DAG.getConstant(0, MVT::i32)),
697 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
698 DAG.getConstant(0, MVT::i32)),
699 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
700 DAG.getConstant(1, MVT::i32)),
701 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
702 DAG.getConstant(1, MVT::i32)),
703 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
704 DAG.getConstant(2, MVT::i32)),
705 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
706 DAG.getConstant(2, MVT::i32)),
707 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
708 DAG.getConstant(3, MVT::i32)),
709 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
710 DAG.getConstant(3, MVT::i32))
712 return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args, 8);
715 case Intrinsic::r600_read_ngroups_x:
716 return LowerImplicitParameter(DAG, VT, DL, 0);
717 case Intrinsic::r600_read_ngroups_y:
718 return LowerImplicitParameter(DAG, VT, DL, 1);
719 case Intrinsic::r600_read_ngroups_z:
720 return LowerImplicitParameter(DAG, VT, DL, 2);
721 case Intrinsic::r600_read_global_size_x:
722 return LowerImplicitParameter(DAG, VT, DL, 3);
723 case Intrinsic::r600_read_global_size_y:
724 return LowerImplicitParameter(DAG, VT, DL, 4);
725 case Intrinsic::r600_read_global_size_z:
726 return LowerImplicitParameter(DAG, VT, DL, 5);
727 case Intrinsic::r600_read_local_size_x:
728 return LowerImplicitParameter(DAG, VT, DL, 6);
729 case Intrinsic::r600_read_local_size_y:
730 return LowerImplicitParameter(DAG, VT, DL, 7);
731 case Intrinsic::r600_read_local_size_z:
732 return LowerImplicitParameter(DAG, VT, DL, 8);
734 case Intrinsic::r600_read_tgid_x:
735 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
737 case Intrinsic::r600_read_tgid_y:
738 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
740 case Intrinsic::r600_read_tgid_z:
741 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
743 case Intrinsic::r600_read_tidig_x:
744 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
746 case Intrinsic::r600_read_tidig_y:
747 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
749 case Intrinsic::r600_read_tidig_z:
750 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
753 // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
756 } // end switch(Op.getOpcode())
760 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
761 SmallVectorImpl<SDValue> &Results,
762 SelectionDAG &DAG) const {
763 switch (N->getOpcode()) {
765 case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
768 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
769 Results.push_back(SDValue(Node, 0));
770 Results.push_back(SDValue(Node, 1));
771 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
773 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
777 SDNode *Node = LowerSTORE(SDValue(N, 0), DAG).getNode();
778 Results.push_back(SDValue(Node, 0));
783 SDValue R600TargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
784 // On hw >= R700, COS/SIN input must be between -1. and 1.
785 // Thus we lower them to TRIG ( FRACT ( x / 2Pi + 0.5) - 0.5)
786 EVT VT = Op.getValueType();
787 SDValue Arg = Op.getOperand(0);
788 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, SDLoc(Op), VT,
789 DAG.getNode(ISD::FADD, SDLoc(Op), VT,
790 DAG.getNode(ISD::FMUL, SDLoc(Op), VT, Arg,
791 DAG.getConstantFP(0.15915494309, MVT::f32)),
792 DAG.getConstantFP(0.5, MVT::f32)));
794 switch (Op.getOpcode()) {
796 TrigNode = AMDGPUISD::COS_HW;
799 TrigNode = AMDGPUISD::SIN_HW;
802 llvm_unreachable("Wrong trig opcode");
804 SDValue TrigVal = DAG.getNode(TrigNode, SDLoc(Op), VT,
805 DAG.getNode(ISD::FADD, SDLoc(Op), VT, FractPart,
806 DAG.getConstantFP(-0.5, MVT::f32)));
807 if (Gen >= AMDGPUSubtarget::R700)
809 // On R600 hw, COS/SIN input must be between -Pi and Pi.
810 return DAG.getNode(ISD::FMUL, SDLoc(Op), VT, TrigVal,
811 DAG.getConstantFP(3.14159265359, MVT::f32));
814 SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
819 Op, DAG.getConstantFP(0.0f, MVT::f32),
820 DAG.getCondCode(ISD::SETNE)
824 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
826 unsigned DwordOffset) const {
827 unsigned ByteOffset = DwordOffset * 4;
828 PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
829 AMDGPUAS::CONSTANT_BUFFER_0);
831 // We shouldn't be using an offset wider than 16-bits for implicit parameters.
832 assert(isInt<16>(ByteOffset));
834 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
835 DAG.getConstant(ByteOffset, MVT::i32), // PTR
836 MachinePointerInfo(ConstantPointerNull::get(PtrType)),
837 false, false, false, 0);
840 SDValue R600TargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
842 MachineFunction &MF = DAG.getMachineFunction();
843 const AMDGPUFrameLowering *TFL =
844 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
846 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
849 unsigned FrameIndex = FIN->getIndex();
850 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
851 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), MVT::i32);
854 bool R600TargetLowering::isZero(SDValue Op) const {
855 if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
856 return Cst->isNullValue();
857 } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
858 return CstFP->isZero();
864 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
866 EVT VT = Op.getValueType();
868 SDValue LHS = Op.getOperand(0);
869 SDValue RHS = Op.getOperand(1);
870 SDValue True = Op.getOperand(2);
871 SDValue False = Op.getOperand(3);
872 SDValue CC = Op.getOperand(4);
875 // LHS and RHS are guaranteed to be the same value type
876 EVT CompareVT = LHS.getValueType();
878 // Check if we can lower this to a native operation.
880 // Try to lower to a SET* instruction:
882 // SET* can match the following patterns:
884 // select_cc f32, f32, -1, 0, cc_supported
885 // select_cc f32, f32, 1.0f, 0.0f, cc_supported
886 // select_cc i32, i32, -1, 0, cc_supported
889 // Move hardware True/False values to the correct operand.
890 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
891 ISD::CondCode InverseCC =
892 ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
893 if (isHWTrueValue(False) && isHWFalseValue(True)) {
894 if (isCondCodeLegal(InverseCC, CompareVT.getSimpleVT())) {
895 std::swap(False, True);
896 CC = DAG.getCondCode(InverseCC);
898 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InverseCC);
899 if (isCondCodeLegal(SwapInvCC, CompareVT.getSimpleVT())) {
900 std::swap(False, True);
902 CC = DAG.getCondCode(SwapInvCC);
907 if (isHWTrueValue(True) && isHWFalseValue(False) &&
908 (CompareVT == VT || VT == MVT::i32)) {
909 // This can be matched by a SET* instruction.
910 return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
913 // Try to lower to a CND* instruction:
915 // CND* can match the following patterns:
917 // select_cc f32, 0.0, f32, f32, cc_supported
918 // select_cc f32, 0.0, i32, i32, cc_supported
919 // select_cc i32, 0, f32, f32, cc_supported
920 // select_cc i32, 0, i32, i32, cc_supported
923 // Try to move the zero value to the RHS
925 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
926 // Try swapping the operands
927 ISD::CondCode CCSwapped = ISD::getSetCCSwappedOperands(CCOpcode);
928 if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
930 CC = DAG.getCondCode(CCSwapped);
932 // Try inverting the conditon and then swapping the operands
933 ISD::CondCode CCInv = ISD::getSetCCInverse(CCOpcode, CompareVT.isInteger());
934 CCSwapped = ISD::getSetCCSwappedOperands(CCInv);
935 if (isCondCodeLegal(CCSwapped, CompareVT.getSimpleVT())) {
936 std::swap(True, False);
938 CC = DAG.getCondCode(CCSwapped);
945 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
946 if (CompareVT != VT) {
947 // Bitcast True / False to the correct types. This will end up being
948 // a nop, but it allows us to define only a single pattern in the
949 // .TD files for each CND* instruction rather than having to have
950 // one pattern for integer True/False and one for fp True/False
951 True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
952 False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
959 CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
967 SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
970 DAG.getCondCode(CCOpcode));
971 return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
975 // Possible Min/Max pattern
976 SDValue MinMax = LowerMinMax(Op, DAG);
977 if (MinMax.getNode()) {
981 // If we make it this for it means we have no native instructions to handle
982 // this SELECT_CC, so we must lower it.
983 SDValue HWTrue, HWFalse;
985 if (CompareVT == MVT::f32) {
986 HWTrue = DAG.getConstantFP(1.0f, CompareVT);
987 HWFalse = DAG.getConstantFP(0.0f, CompareVT);
988 } else if (CompareVT == MVT::i32) {
989 HWTrue = DAG.getConstant(-1, CompareVT);
990 HWFalse = DAG.getConstant(0, CompareVT);
993 assert(!"Unhandled value type in LowerSELECT_CC");
996 // Lower this unsupported SELECT_CC into a combination of two supported
997 // SELECT_CC operations.
998 SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
1000 return DAG.getNode(ISD::SELECT_CC, DL, VT,
1003 DAG.getCondCode(ISD::SETNE));
1006 /// LLVM generates byte-addresed pointers. For indirect addressing, we need to
1007 /// convert these pointers to a register index. Each register holds
1008 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
1009 /// \p StackWidth, which tells us how many of the 4 sub-registrers will be used
1010 /// for indirect addressing.
1011 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
1012 unsigned StackWidth,
1013 SelectionDAG &DAG) const {
1015 switch(StackWidth) {
1025 default: llvm_unreachable("Invalid stack width");
1028 return DAG.getNode(ISD::SRL, SDLoc(Ptr), Ptr.getValueType(), Ptr,
1029 DAG.getConstant(SRLPad, MVT::i32));
1032 void R600TargetLowering::getStackAddress(unsigned StackWidth,
1035 unsigned &PtrIncr) const {
1036 switch (StackWidth) {
1047 Channel = ElemIdx % 2;
1061 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1063 StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
1064 SDValue Chain = Op.getOperand(0);
1065 SDValue Value = Op.getOperand(1);
1066 SDValue Ptr = Op.getOperand(2);
1068 SDValue Result = AMDGPUTargetLowering::LowerSTORE(Op, DAG);
1069 if (Result.getNode()) {
1073 if (StoreNode->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS) {
1074 if (StoreNode->isTruncatingStore()) {
1075 EVT VT = Value.getValueType();
1076 assert(VT.bitsLE(MVT::i32));
1077 EVT MemVT = StoreNode->getMemoryVT();
1078 SDValue MaskConstant;
1079 if (MemVT == MVT::i8) {
1080 MaskConstant = DAG.getConstant(0xFF, MVT::i32);
1082 assert(MemVT == MVT::i16);
1083 MaskConstant = DAG.getConstant(0xFFFF, MVT::i32);
1085 SDValue DWordAddr = DAG.getNode(ISD::SRL, DL, VT, Ptr,
1086 DAG.getConstant(2, MVT::i32));
1087 SDValue ByteIndex = DAG.getNode(ISD::AND, DL, Ptr.getValueType(), Ptr,
1088 DAG.getConstant(0x00000003, VT));
1089 SDValue TruncValue = DAG.getNode(ISD::AND, DL, VT, Value, MaskConstant);
1090 SDValue Shift = DAG.getNode(ISD::SHL, DL, VT, ByteIndex,
1091 DAG.getConstant(3, VT));
1092 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, VT, TruncValue, Shift);
1093 SDValue Mask = DAG.getNode(ISD::SHL, DL, VT, MaskConstant, Shift);
1094 // XXX: If we add a 64-bit ZW register class, then we could use a 2 x i32
1098 DAG.getConstant(0, MVT::i32),
1099 DAG.getConstant(0, MVT::i32),
1102 SDValue Input = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Src, 4);
1103 SDValue Args[3] = { Chain, Input, DWordAddr };
1104 return DAG.getMemIntrinsicNode(AMDGPUISD::STORE_MSKOR, DL,
1105 Op->getVTList(), Args, 3, MemVT,
1106 StoreNode->getMemOperand());
1107 } else if (Ptr->getOpcode() != AMDGPUISD::DWORDADDR &&
1108 Value.getValueType().bitsGE(MVT::i32)) {
1109 // Convert pointer from byte address to dword address.
1110 Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, Ptr.getValueType(),
1111 DAG.getNode(ISD::SRL, DL, Ptr.getValueType(),
1112 Ptr, DAG.getConstant(2, MVT::i32)));
1114 if (StoreNode->isTruncatingStore() || StoreNode->isIndexed()) {
1115 assert(!"Truncated and indexed stores not supported yet");
1117 Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1123 EVT ValueVT = Value.getValueType();
1125 if (StoreNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1129 // Lowering for indirect addressing
1131 const MachineFunction &MF = DAG.getMachineFunction();
1132 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1133 getTargetMachine().getFrameLowering());
1134 unsigned StackWidth = TFL->getStackWidth(MF);
1136 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1138 if (ValueVT.isVector()) {
1139 unsigned NumElemVT = ValueVT.getVectorNumElements();
1140 EVT ElemVT = ValueVT.getVectorElementType();
1143 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1144 "vector width in load");
1146 for (unsigned i = 0; i < NumElemVT; ++i) {
1147 unsigned Channel, PtrIncr;
1148 getStackAddress(StackWidth, i, Channel, PtrIncr);
1149 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1150 DAG.getConstant(PtrIncr, MVT::i32));
1151 SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT,
1152 Value, DAG.getConstant(i, MVT::i32));
1154 Stores[i] = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1156 DAG.getTargetConstant(Channel, MVT::i32));
1158 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores, NumElemVT);
1160 if (ValueVT == MVT::i8) {
1161 Value = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Value);
1163 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other, Chain, Value, Ptr,
1164 DAG.getTargetConstant(0, MVT::i32)); // Channel
1170 // return (512 + (kc_bank << 12)
1172 ConstantAddressBlock(unsigned AddressSpace) {
1173 switch (AddressSpace) {
1174 case AMDGPUAS::CONSTANT_BUFFER_0:
1176 case AMDGPUAS::CONSTANT_BUFFER_1:
1178 case AMDGPUAS::CONSTANT_BUFFER_2:
1179 return 512 + 4096 * 2;
1180 case AMDGPUAS::CONSTANT_BUFFER_3:
1181 return 512 + 4096 * 3;
1182 case AMDGPUAS::CONSTANT_BUFFER_4:
1183 return 512 + 4096 * 4;
1184 case AMDGPUAS::CONSTANT_BUFFER_5:
1185 return 512 + 4096 * 5;
1186 case AMDGPUAS::CONSTANT_BUFFER_6:
1187 return 512 + 4096 * 6;
1188 case AMDGPUAS::CONSTANT_BUFFER_7:
1189 return 512 + 4096 * 7;
1190 case AMDGPUAS::CONSTANT_BUFFER_8:
1191 return 512 + 4096 * 8;
1192 case AMDGPUAS::CONSTANT_BUFFER_9:
1193 return 512 + 4096 * 9;
1194 case AMDGPUAS::CONSTANT_BUFFER_10:
1195 return 512 + 4096 * 10;
1196 case AMDGPUAS::CONSTANT_BUFFER_11:
1197 return 512 + 4096 * 11;
1198 case AMDGPUAS::CONSTANT_BUFFER_12:
1199 return 512 + 4096 * 12;
1200 case AMDGPUAS::CONSTANT_BUFFER_13:
1201 return 512 + 4096 * 13;
1202 case AMDGPUAS::CONSTANT_BUFFER_14:
1203 return 512 + 4096 * 14;
1204 case AMDGPUAS::CONSTANT_BUFFER_15:
1205 return 512 + 4096 * 15;
1211 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
1213 EVT VT = Op.getValueType();
1215 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1216 SDValue Chain = Op.getOperand(0);
1217 SDValue Ptr = Op.getOperand(1);
1218 SDValue LoweredLoad;
1220 if (LoadNode->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS && VT.isVector()) {
1221 SDValue MergedValues[2] = {
1222 SplitVectorLoad(Op, DAG),
1225 return DAG.getMergeValues(MergedValues, 2, DL);
1228 int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1229 if (ConstantBlock > -1 && LoadNode->getExtensionType() != ISD::SEXTLOAD) {
1231 if (isa<ConstantExpr>(LoadNode->getSrcValue()) ||
1232 isa<Constant>(LoadNode->getSrcValue()) ||
1233 isa<ConstantSDNode>(Ptr)) {
1235 for (unsigned i = 0; i < 4; i++) {
1236 // We want Const position encoded with the following formula :
1237 // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1238 // const_index is Ptr computed by llvm using an alignment of 16.
1239 // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1240 // then div by 4 at the ISel step
1241 SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1242 DAG.getConstant(4 * i + ConstantBlock * 16, MVT::i32));
1243 Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1245 EVT NewVT = MVT::v4i32;
1246 unsigned NumElements = 4;
1247 if (VT.isVector()) {
1249 NumElements = VT.getVectorNumElements();
1251 Result = DAG.getNode(ISD::BUILD_VECTOR, DL, NewVT, Slots, NumElements);
1253 // non constant ptr cant be folded, keeps it as a v4f32 load
1254 Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1255 DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(4, MVT::i32)),
1256 DAG.getConstant(LoadNode->getAddressSpace() -
1257 AMDGPUAS::CONSTANT_BUFFER_0, MVT::i32)
1261 if (!VT.isVector()) {
1262 Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1263 DAG.getConstant(0, MVT::i32));
1266 SDValue MergedValues[2] = {
1270 return DAG.getMergeValues(MergedValues, 2, DL);
1273 // For most operations returning SDValue() will result in the node being
1274 // expanded by the DAG Legalizer. This is not the case for ISD::LOAD, so we
1275 // need to manually expand loads that may be legal in some address spaces and
1276 // illegal in others. SEXT loads from CONSTANT_BUFFER_0 are supported for
1277 // compute shaders, since the data is sign extended when it is uploaded to the
1278 // buffer. However SEXT loads from other address spaces are not supported, so
1279 // we need to expand them here.
1280 if (LoadNode->getExtensionType() == ISD::SEXTLOAD) {
1281 EVT MemVT = LoadNode->getMemoryVT();
1282 assert(!MemVT.isVector() && (MemVT == MVT::i16 || MemVT == MVT::i8));
1283 SDValue ShiftAmount =
1284 DAG.getConstant(VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
1285 SDValue NewLoad = DAG.getExtLoad(ISD::EXTLOAD, DL, VT, Chain, Ptr,
1286 LoadNode->getPointerInfo(), MemVT,
1287 LoadNode->isVolatile(),
1288 LoadNode->isNonTemporal(),
1289 LoadNode->getAlignment());
1290 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, NewLoad, ShiftAmount);
1291 SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Shl, ShiftAmount);
1293 SDValue MergedValues[2] = { Sra, Chain };
1294 return DAG.getMergeValues(MergedValues, 2, DL);
1297 if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1301 // Lowering for indirect addressing
1302 const MachineFunction &MF = DAG.getMachineFunction();
1303 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1304 getTargetMachine().getFrameLowering());
1305 unsigned StackWidth = TFL->getStackWidth(MF);
1307 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1309 if (VT.isVector()) {
1310 unsigned NumElemVT = VT.getVectorNumElements();
1311 EVT ElemVT = VT.getVectorElementType();
1314 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1315 "vector width in load");
1317 for (unsigned i = 0; i < NumElemVT; ++i) {
1318 unsigned Channel, PtrIncr;
1319 getStackAddress(StackWidth, i, Channel, PtrIncr);
1320 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1321 DAG.getConstant(PtrIncr, MVT::i32));
1322 Loads[i] = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, ElemVT,
1324 DAG.getTargetConstant(Channel, MVT::i32),
1327 for (unsigned i = NumElemVT; i < 4; ++i) {
1328 Loads[i] = DAG.getUNDEF(ElemVT);
1330 EVT TargetVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, 4);
1331 LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads, 4);
1333 LoweredLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, VT,
1335 DAG.getTargetConstant(0, MVT::i32), // Channel
1340 Ops[0] = LoweredLoad;
1343 return DAG.getMergeValues(Ops, 2, DL);
1346 /// XXX Only kernel functions are supported, so we can assume for now that
1347 /// every function is a kernel function, but in the future we should use
1348 /// separate calling conventions for kernel and non-kernel functions.
1349 SDValue R600TargetLowering::LowerFormalArguments(
1351 CallingConv::ID CallConv,
1353 const SmallVectorImpl<ISD::InputArg> &Ins,
1354 SDLoc DL, SelectionDAG &DAG,
1355 SmallVectorImpl<SDValue> &InVals) const {
1356 SmallVector<CCValAssign, 16> ArgLocs;
1357 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1358 getTargetMachine(), ArgLocs, *DAG.getContext());
1359 MachineFunction &MF = DAG.getMachineFunction();
1360 unsigned ShaderType = MF.getInfo<R600MachineFunctionInfo>()->ShaderType;
1362 SmallVector<ISD::InputArg, 8> LocalIns;
1364 getOriginalFunctionArgs(DAG, DAG.getMachineFunction().getFunction(), Ins,
1367 AnalyzeFormalArguments(CCInfo, LocalIns);
1369 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1370 CCValAssign &VA = ArgLocs[i];
1372 EVT MemVT = LocalIns[i].VT;
1374 if (ShaderType != ShaderType::COMPUTE) {
1375 unsigned Reg = MF.addLiveIn(VA.getLocReg(), &AMDGPU::R600_Reg128RegClass);
1376 SDValue Register = DAG.getCopyFromReg(Chain, DL, Reg, VT);
1377 InVals.push_back(Register);
1381 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
1382 AMDGPUAS::CONSTANT_BUFFER_0);
1384 // The first 36 bytes of the input buffer contains information about
1385 // thread group and global sizes.
1386 SDValue Arg = DAG.getExtLoad(ISD::SEXTLOAD, DL, VT, Chain,
1387 DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
1388 MachinePointerInfo(UndefValue::get(PtrTy)),
1389 MemVT, false, false, 4);
1390 // 4 is the prefered alignment for
1391 // the CONSTANT memory space.
1392 InVals.push_back(Arg);
1397 EVT R600TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1398 if (!VT.isVector()) return MVT::i32;
1399 return VT.changeVectorElementTypeToInteger();
1403 CompactSwizzlableVector(SelectionDAG &DAG, SDValue VectorEntry,
1404 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1405 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1406 assert(RemapSwizzle.empty());
1407 SDValue NewBldVec[4] = {
1408 VectorEntry.getOperand(0),
1409 VectorEntry.getOperand(1),
1410 VectorEntry.getOperand(2),
1411 VectorEntry.getOperand(3)
1414 for (unsigned i = 0; i < 4; i++) {
1415 if (NewBldVec[i].getOpcode() == ISD::UNDEF)
1416 // We mask write here to teach later passes that the ith element of this
1417 // vector is undef. Thus we can use it to reduce 128 bits reg usage,
1418 // break false dependencies and additionnaly make assembly easier to read.
1419 RemapSwizzle[i] = 7; // SEL_MASK_WRITE
1420 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(NewBldVec[i])) {
1422 RemapSwizzle[i] = 4; // SEL_0
1423 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1424 } else if (C->isExactlyValue(1.0)) {
1425 RemapSwizzle[i] = 5; // SEL_1
1426 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1430 if (NewBldVec[i].getOpcode() == ISD::UNDEF)
1432 for (unsigned j = 0; j < i; j++) {
1433 if (NewBldVec[i] == NewBldVec[j]) {
1434 NewBldVec[i] = DAG.getUNDEF(NewBldVec[i].getValueType());
1435 RemapSwizzle[i] = j;
1441 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1442 VectorEntry.getValueType(), NewBldVec, 4);
1445 static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
1446 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1447 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1448 assert(RemapSwizzle.empty());
1449 SDValue NewBldVec[4] = {
1450 VectorEntry.getOperand(0),
1451 VectorEntry.getOperand(1),
1452 VectorEntry.getOperand(2),
1453 VectorEntry.getOperand(3)
1455 bool isUnmovable[4] = { false, false, false, false };
1456 for (unsigned i = 0; i < 4; i++)
1457 RemapSwizzle[i] = i;
1459 for (unsigned i = 0; i < 4; i++) {
1460 if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1461 unsigned Idx = dyn_cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1464 isUnmovable[Idx] = true;
1467 if (isUnmovable[Idx])
1470 std::swap(NewBldVec[Idx], NewBldVec[i]);
1471 std::swap(RemapSwizzle[i], RemapSwizzle[Idx]);
1476 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1477 VectorEntry.getValueType(), NewBldVec, 4);
1481 SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector,
1482 SDValue Swz[4], SelectionDAG &DAG) const {
1483 assert(BuildVector.getOpcode() == ISD::BUILD_VECTOR);
1484 // Old -> New swizzle values
1485 DenseMap<unsigned, unsigned> SwizzleRemap;
1487 BuildVector = CompactSwizzlableVector(DAG, BuildVector, SwizzleRemap);
1488 for (unsigned i = 0; i < 4; i++) {
1489 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1490 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1491 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1494 SwizzleRemap.clear();
1495 BuildVector = ReorganizeVector(DAG, BuildVector, SwizzleRemap);
1496 for (unsigned i = 0; i < 4; i++) {
1497 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1498 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1499 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1506 //===----------------------------------------------------------------------===//
1507 // Custom DAG Optimizations
1508 //===----------------------------------------------------------------------===//
1510 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1511 DAGCombinerInfo &DCI) const {
1512 SelectionDAG &DAG = DCI.DAG;
1514 switch (N->getOpcode()) {
1515 // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1516 case ISD::FP_ROUND: {
1517 SDValue Arg = N->getOperand(0);
1518 if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1519 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), N->getValueType(0),
1525 // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1526 // (i32 select_cc f32, f32, -1, 0 cc)
1528 // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1529 // this to one of the SET*_DX10 instructions.
1530 case ISD::FP_TO_SINT: {
1531 SDValue FNeg = N->getOperand(0);
1532 if (FNeg.getOpcode() != ISD::FNEG) {
1535 SDValue SelectCC = FNeg.getOperand(0);
1536 if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1537 SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1538 SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1539 !isHWTrueValue(SelectCC.getOperand(2)) ||
1540 !isHWFalseValue(SelectCC.getOperand(3))) {
1544 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N->getValueType(0),
1545 SelectCC.getOperand(0), // LHS
1546 SelectCC.getOperand(1), // RHS
1547 DAG.getConstant(-1, MVT::i32), // True
1548 DAG.getConstant(0, MVT::i32), // Flase
1549 SelectCC.getOperand(4)); // CC
1554 // insert_vector_elt (build_vector elt0, ... , eltN), NewEltIdx, idx
1555 // => build_vector elt0, ... , NewEltIdx, ... , eltN
1556 case ISD::INSERT_VECTOR_ELT: {
1557 SDValue InVec = N->getOperand(0);
1558 SDValue InVal = N->getOperand(1);
1559 SDValue EltNo = N->getOperand(2);
1562 // If the inserted element is an UNDEF, just use the input vector.
1563 if (InVal.getOpcode() == ISD::UNDEF)
1566 EVT VT = InVec.getValueType();
1568 // If we can't generate a legal BUILD_VECTOR, exit
1569 if (!isOperationLegal(ISD::BUILD_VECTOR, VT))
1572 // Check that we know which element is being inserted
1573 if (!isa<ConstantSDNode>(EltNo))
1575 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
1577 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
1578 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
1580 SmallVector<SDValue, 8> Ops;
1581 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
1582 Ops.append(InVec.getNode()->op_begin(),
1583 InVec.getNode()->op_end());
1584 } else if (InVec.getOpcode() == ISD::UNDEF) {
1585 unsigned NElts = VT.getVectorNumElements();
1586 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
1591 // Insert the element
1592 if (Elt < Ops.size()) {
1593 // All the operands of BUILD_VECTOR must have the same type;
1594 // we enforce that here.
1595 EVT OpVT = Ops[0].getValueType();
1596 if (InVal.getValueType() != OpVT)
1597 InVal = OpVT.bitsGT(InVal.getValueType()) ?
1598 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
1599 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
1603 // Return the new vector
1604 return DAG.getNode(ISD::BUILD_VECTOR, dl,
1605 VT, &Ops[0], Ops.size());
1608 // Extract_vec (Build_vector) generated by custom lowering
1609 // also needs to be customly combined
1610 case ISD::EXTRACT_VECTOR_ELT: {
1611 SDValue Arg = N->getOperand(0);
1612 if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1613 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1614 unsigned Element = Const->getZExtValue();
1615 return Arg->getOperand(Element);
1618 if (Arg.getOpcode() == ISD::BITCAST &&
1619 Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
1620 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1621 unsigned Element = Const->getZExtValue();
1622 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getVTList(),
1623 Arg->getOperand(0).getOperand(Element));
1628 case ISD::SELECT_CC: {
1629 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1630 // selectcc x, y, a, b, inv(cc)
1632 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1633 // selectcc x, y, a, b, cc
1634 SDValue LHS = N->getOperand(0);
1635 if (LHS.getOpcode() != ISD::SELECT_CC) {
1639 SDValue RHS = N->getOperand(1);
1640 SDValue True = N->getOperand(2);
1641 SDValue False = N->getOperand(3);
1642 ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1644 if (LHS.getOperand(2).getNode() != True.getNode() ||
1645 LHS.getOperand(3).getNode() != False.getNode() ||
1646 RHS.getNode() != False.getNode()) {
1651 default: return SDValue();
1652 case ISD::SETNE: return LHS;
1654 ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1655 LHSCC = ISD::getSetCCInverse(LHSCC,
1656 LHS.getOperand(0).getValueType().isInteger());
1657 if (DCI.isBeforeLegalizeOps() ||
1658 isCondCodeLegal(LHSCC, LHS.getOperand(0).getSimpleValueType()))
1659 return DAG.getSelectCC(SDLoc(N),
1671 case AMDGPUISD::EXPORT: {
1672 SDValue Arg = N->getOperand(1);
1673 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1676 SDValue NewArgs[8] = {
1677 N->getOperand(0), // Chain
1679 N->getOperand(2), // ArrayBase
1680 N->getOperand(3), // Type
1681 N->getOperand(4), // SWZ_X
1682 N->getOperand(5), // SWZ_Y
1683 N->getOperand(6), // SWZ_Z
1684 N->getOperand(7) // SWZ_W
1687 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG);
1688 return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs, 8);
1690 case AMDGPUISD::TEXTURE_FETCH: {
1691 SDValue Arg = N->getOperand(1);
1692 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1695 SDValue NewArgs[19] = {
1716 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG);
1717 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, SDLoc(N), N->getVTList(),
1725 FoldOperand(SDNode *ParentNode, unsigned SrcIdx, SDValue &Src, SDValue &Neg,
1726 SDValue &Abs, SDValue &Sel, SDValue &Imm, SelectionDAG &DAG) {
1727 const R600InstrInfo *TII =
1728 static_cast<const R600InstrInfo *>(DAG.getTarget().getInstrInfo());
1729 if (!Src.isMachineOpcode())
1731 switch (Src.getMachineOpcode()) {
1732 case AMDGPU::FNEG_R600:
1735 Src = Src.getOperand(0);
1736 Neg = DAG.getTargetConstant(1, MVT::i32);
1738 case AMDGPU::FABS_R600:
1741 Src = Src.getOperand(0);
1742 Abs = DAG.getTargetConstant(1, MVT::i32);
1744 case AMDGPU::CONST_COPY: {
1745 unsigned Opcode = ParentNode->getMachineOpcode();
1746 bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
1751 SDValue CstOffset = Src.getOperand(0);
1752 if (ParentNode->getValueType(0).isVector())
1755 // Gather constants values
1756 int SrcIndices[] = {
1757 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0),
1758 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1),
1759 TII->getOperandIdx(Opcode, AMDGPU::OpName::src2),
1760 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_X),
1761 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Y),
1762 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Z),
1763 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_W),
1764 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_X),
1765 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Y),
1766 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Z),
1767 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
1769 std::vector<unsigned> Consts;
1770 for (unsigned i = 0; i < sizeof(SrcIndices) / sizeof(int); i++) {
1771 int OtherSrcIdx = SrcIndices[i];
1772 int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
1773 if (OtherSrcIdx < 0 || OtherSelIdx < 0)
1779 if (RegisterSDNode *Reg =
1780 dyn_cast<RegisterSDNode>(ParentNode->getOperand(OtherSrcIdx))) {
1781 if (Reg->getReg() == AMDGPU::ALU_CONST) {
1782 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(
1783 ParentNode->getOperand(OtherSelIdx));
1784 Consts.push_back(Cst->getZExtValue());
1789 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
1790 Consts.push_back(Cst->getZExtValue());
1791 if (!TII->fitsConstReadLimitations(Consts)) {
1796 Src = DAG.getRegister(AMDGPU::ALU_CONST, MVT::f32);
1799 case AMDGPU::MOV_IMM_I32:
1800 case AMDGPU::MOV_IMM_F32: {
1801 unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
1802 uint64_t ImmValue = 0;
1805 if (Src.getMachineOpcode() == AMDGPU::MOV_IMM_F32) {
1806 ConstantFPSDNode *FPC = dyn_cast<ConstantFPSDNode>(Src.getOperand(0));
1807 float FloatValue = FPC->getValueAPF().convertToFloat();
1808 if (FloatValue == 0.0) {
1809 ImmReg = AMDGPU::ZERO;
1810 } else if (FloatValue == 0.5) {
1811 ImmReg = AMDGPU::HALF;
1812 } else if (FloatValue == 1.0) {
1813 ImmReg = AMDGPU::ONE;
1815 ImmValue = FPC->getValueAPF().bitcastToAPInt().getZExtValue();
1818 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Src.getOperand(0));
1819 uint64_t Value = C->getZExtValue();
1821 ImmReg = AMDGPU::ZERO;
1822 } else if (Value == 1) {
1823 ImmReg = AMDGPU::ONE_INT;
1829 // Check that we aren't already using an immediate.
1830 // XXX: It's possible for an instruction to have more than one
1831 // immediate operand, but this is not supported yet.
1832 if (ImmReg == AMDGPU::ALU_LITERAL_X) {
1835 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Imm);
1837 if (C->getZExtValue())
1839 Imm = DAG.getTargetConstant(ImmValue, MVT::i32);
1841 Src = DAG.getRegister(ImmReg, MVT::i32);
1850 /// \brief Fold the instructions after selecting them
1851 SDNode *R600TargetLowering::PostISelFolding(MachineSDNode *Node,
1852 SelectionDAG &DAG) const {
1853 const R600InstrInfo *TII =
1854 static_cast<const R600InstrInfo *>(DAG.getTarget().getInstrInfo());
1855 if (!Node->isMachineOpcode())
1857 unsigned Opcode = Node->getMachineOpcode();
1860 std::vector<SDValue> Ops;
1861 for(SDNode::op_iterator I = Node->op_begin(), E = Node->op_end();
1865 if (Opcode == AMDGPU::DOT_4) {
1866 int OperandIdx[] = {
1867 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_X),
1868 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Y),
1869 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_Z),
1870 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_W),
1871 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_X),
1872 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Y),
1873 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_Z),
1874 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
1877 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_neg_X),
1878 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_neg_Y),
1879 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_neg_Z),
1880 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_neg_W),
1881 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_neg_X),
1882 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_neg_Y),
1883 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_neg_Z),
1884 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_neg_W)
1887 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_abs_X),
1888 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_abs_Y),
1889 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_abs_Z),
1890 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_abs_W),
1891 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs_X),
1892 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs_Y),
1893 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs_Z),
1894 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs_W)
1896 for (unsigned i = 0; i < 8; i++) {
1897 if (OperandIdx[i] < 0)
1899 SDValue &Src = Ops[OperandIdx[i] - 1];
1900 SDValue &Neg = Ops[NegIdx[i] - 1];
1901 SDValue &Abs = Ops[AbsIdx[i] - 1];
1902 bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
1903 int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
1906 SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
1907 if (FoldOperand(Node, i, Src, Neg, Abs, Sel, FakeOp, DAG))
1908 return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
1910 } else if (Opcode == AMDGPU::REG_SEQUENCE) {
1911 for (unsigned i = 1, e = Node->getNumOperands(); i < e; i += 2) {
1912 SDValue &Src = Ops[i];
1913 if (FoldOperand(Node, i, Src, FakeOp, FakeOp, FakeOp, FakeOp, DAG))
1914 return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);
1916 } else if (Opcode == AMDGPU::CLAMP_R600) {
1917 SDValue Src = Node->getOperand(0);
1918 if (!Src.isMachineOpcode() ||
1919 !TII->hasInstrModifiers(Src.getMachineOpcode()))
1921 int ClampIdx = TII->getOperandIdx(Src.getMachineOpcode(),
1922 AMDGPU::OpName::clamp);
1925 std::vector<SDValue> Ops;
1926 unsigned NumOp = Src.getNumOperands();
1927 for(unsigned i = 0; i < NumOp; ++i)
1928 Ops.push_back(Src.getOperand(i));
1929 Ops[ClampIdx - 1] = DAG.getTargetConstant(1, MVT::i32);
1930 return DAG.getMachineNode(Src.getMachineOpcode(), SDLoc(Node),
1931 Node->getVTList(), Ops);
1933 if (!TII->hasInstrModifiers(Opcode))
1935 int OperandIdx[] = {
1936 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0),
1937 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1),
1938 TII->getOperandIdx(Opcode, AMDGPU::OpName::src2)
1941 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_neg),
1942 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_neg),
1943 TII->getOperandIdx(Opcode, AMDGPU::OpName::src2_neg)
1946 TII->getOperandIdx(Opcode, AMDGPU::OpName::src0_abs),
1947 TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_abs),
1950 for (unsigned i = 0; i < 3; i++) {
1951 if (OperandIdx[i] < 0)
1953 SDValue &Src = Ops[OperandIdx[i] - 1];
1954 SDValue &Neg = Ops[NegIdx[i] - 1];
1956 SDValue &Abs = (AbsIdx[i] > -1) ? Ops[AbsIdx[i] - 1] : FakeAbs;
1957 bool HasDst = TII->getOperandIdx(Opcode, AMDGPU::OpName::dst) > -1;
1958 int SelIdx = TII->getSelIdx(Opcode, OperandIdx[i]);
1959 int ImmIdx = TII->getOperandIdx(Opcode, AMDGPU::OpName::literal);
1964 SDValue &Sel = (SelIdx > -1) ? Ops[SelIdx] : FakeOp;
1965 SDValue &Imm = Ops[ImmIdx];
1966 if (FoldOperand(Node, i, Src, Neg, Abs, Sel, Imm, DAG))
1967 return DAG.getMachineNode(Opcode, SDLoc(Node), Node->getVTList(), Ops);