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/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/SelectionDAG.h"
23 #include "llvm/IR/Argument.h"
24 #include "llvm/IR/Function.h"
28 R600TargetLowering::R600TargetLowering(TargetMachine &TM) :
29 AMDGPUTargetLowering(TM),
30 TII(static_cast<const R600InstrInfo*>(TM.getInstrInfo())) {
31 addRegisterClass(MVT::v4f32, &AMDGPU::R600_Reg128RegClass);
32 addRegisterClass(MVT::f32, &AMDGPU::R600_Reg32RegClass);
33 addRegisterClass(MVT::v4i32, &AMDGPU::R600_Reg128RegClass);
34 addRegisterClass(MVT::i32, &AMDGPU::R600_Reg32RegClass);
35 computeRegisterProperties();
37 setOperationAction(ISD::FADD, MVT::v4f32, Expand);
38 setOperationAction(ISD::FMUL, MVT::v4f32, Expand);
39 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
40 setOperationAction(ISD::FSUB, MVT::v4f32, Expand);
42 setOperationAction(ISD::ADD, MVT::v4i32, Expand);
43 setOperationAction(ISD::AND, MVT::v4i32, Expand);
44 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Expand);
45 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Expand);
46 setOperationAction(ISD::MUL, MVT::v2i32, Expand);
47 setOperationAction(ISD::MUL, MVT::v4i32, Expand);
48 setOperationAction(ISD::OR, MVT::v4i32, Expand);
49 setOperationAction(ISD::OR, MVT::v2i32, Expand);
50 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Expand);
51 setOperationAction(ISD::SHL, MVT::v4i32, Expand);
52 setOperationAction(ISD::SHL, MVT::v2i32, Expand);
53 setOperationAction(ISD::SRL, MVT::v4i32, Expand);
54 setOperationAction(ISD::SRL, MVT::v2i32, Expand);
55 setOperationAction(ISD::SRA, MVT::v4i32, Expand);
56 setOperationAction(ISD::SRA, MVT::v2i32, Expand);
57 setOperationAction(ISD::SUB, MVT::v4i32, Expand);
58 setOperationAction(ISD::SUB, MVT::v2i32, Expand);
59 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Expand);
60 setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
61 setOperationAction(ISD::UREM, MVT::v4i32, Expand);
62 setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
63 setOperationAction(ISD::XOR, MVT::v4i32, Expand);
64 setOperationAction(ISD::XOR, 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::ROTL, MVT::i32, Custom);
77 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
78 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
80 setOperationAction(ISD::SETCC, MVT::i32, Expand);
81 setOperationAction(ISD::SETCC, MVT::f32, Expand);
82 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
84 setOperationAction(ISD::SELECT, MVT::i32, Custom);
85 setOperationAction(ISD::SELECT, MVT::f32, Custom);
87 setOperationAction(ISD::VSELECT, MVT::v4i32, Expand);
88 setOperationAction(ISD::VSELECT, MVT::v2i32, Expand);
90 // Legalize loads and stores to the private address space.
91 setOperationAction(ISD::LOAD, MVT::i32, Custom);
92 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
93 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
94 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Custom);
95 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
96 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
97 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Custom);
98 setOperationAction(ISD::STORE, MVT::i8, Custom);
99 setOperationAction(ISD::STORE, MVT::i32, Custom);
100 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
101 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
103 setOperationAction(ISD::LOAD, MVT::i32, Custom);
104 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
105 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
107 setTargetDAGCombine(ISD::FP_ROUND);
108 setTargetDAGCombine(ISD::FP_TO_SINT);
109 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
110 setTargetDAGCombine(ISD::SELECT_CC);
112 setBooleanContents(ZeroOrNegativeOneBooleanContent);
113 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
114 setSchedulingPreference(Sched::VLIW);
117 MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
118 MachineInstr * MI, MachineBasicBlock * BB) const {
119 MachineFunction * MF = BB->getParent();
120 MachineRegisterInfo &MRI = MF->getRegInfo();
121 MachineBasicBlock::iterator I = *MI;
123 switch (MI->getOpcode()) {
124 default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
125 case AMDGPU::CLAMP_R600: {
126 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
128 MI->getOperand(0).getReg(),
129 MI->getOperand(1).getReg());
130 TII->addFlag(NewMI, 0, MO_FLAG_CLAMP);
134 case AMDGPU::FABS_R600: {
135 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
137 MI->getOperand(0).getReg(),
138 MI->getOperand(1).getReg());
139 TII->addFlag(NewMI, 0, MO_FLAG_ABS);
143 case AMDGPU::FNEG_R600: {
144 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
146 MI->getOperand(0).getReg(),
147 MI->getOperand(1).getReg());
148 TII->addFlag(NewMI, 0, MO_FLAG_NEG);
152 case AMDGPU::MASK_WRITE: {
153 unsigned maskedRegister = MI->getOperand(0).getReg();
154 assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
155 MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
156 TII->addFlag(defInstr, 0, MO_FLAG_MASK);
160 case AMDGPU::MOV_IMM_F32:
161 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
162 MI->getOperand(1).getFPImm()->getValueAPF()
163 .bitcastToAPInt().getZExtValue());
165 case AMDGPU::MOV_IMM_I32:
166 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
167 MI->getOperand(1).getImm());
169 case AMDGPU::CONST_COPY: {
170 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, MI, AMDGPU::MOV,
171 MI->getOperand(0).getReg(), AMDGPU::ALU_CONST);
172 TII->setImmOperand(NewMI, R600Operands::SRC0_SEL,
173 MI->getOperand(1).getImm());
177 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
178 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
179 unsigned EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN) ? 1 : 0;
181 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
182 .addOperand(MI->getOperand(0))
183 .addOperand(MI->getOperand(1))
184 .addImm(EOP); // Set End of program bit
189 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
190 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
191 MachineOperand &RID = MI->getOperand(4);
192 MachineOperand &SID = MI->getOperand(5);
193 unsigned TextureId = MI->getOperand(6).getImm();
194 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
195 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
207 case 8: // ShadowRect
218 case 11: // Shadow1DArray
222 case 12: // Shadow2DArray
226 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
227 .addOperand(MI->getOperand(3))
245 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
246 .addOperand(MI->getOperand(2))
264 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
265 .addOperand(MI->getOperand(0))
266 .addOperand(MI->getOperand(1))
284 .addReg(T0, RegState::Implicit)
285 .addReg(T1, RegState::Implicit);
289 case AMDGPU::TXD_SHADOW: {
290 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
291 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
292 MachineOperand &RID = MI->getOperand(4);
293 MachineOperand &SID = MI->getOperand(5);
294 unsigned TextureId = MI->getOperand(6).getImm();
295 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
296 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
308 case 8: // ShadowRect
319 case 11: // Shadow1DArray
323 case 12: // Shadow2DArray
328 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
329 .addOperand(MI->getOperand(3))
347 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
348 .addOperand(MI->getOperand(2))
366 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
367 .addOperand(MI->getOperand(0))
368 .addOperand(MI->getOperand(1))
386 .addReg(T0, RegState::Implicit)
387 .addReg(T1, RegState::Implicit);
392 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
393 .addOperand(MI->getOperand(0));
396 case AMDGPU::BRANCH_COND_f32: {
397 MachineInstr *NewMI =
398 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
399 AMDGPU::PREDICATE_BIT)
400 .addOperand(MI->getOperand(1))
401 .addImm(OPCODE_IS_NOT_ZERO)
403 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
404 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
405 .addOperand(MI->getOperand(0))
406 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
410 case AMDGPU::BRANCH_COND_i32: {
411 MachineInstr *NewMI =
412 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
413 AMDGPU::PREDICATE_BIT)
414 .addOperand(MI->getOperand(1))
415 .addImm(OPCODE_IS_NOT_ZERO_INT)
417 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
418 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
419 .addOperand(MI->getOperand(0))
420 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
424 case AMDGPU::EG_ExportSwz:
425 case AMDGPU::R600_ExportSwz: {
426 // Instruction is left unmodified if its not the last one of its type
427 bool isLastInstructionOfItsType = true;
428 unsigned InstExportType = MI->getOperand(1).getImm();
429 for (MachineBasicBlock::iterator NextExportInst = llvm::next(I),
430 EndBlock = BB->end(); NextExportInst != EndBlock;
431 NextExportInst = llvm::next(NextExportInst)) {
432 if (NextExportInst->getOpcode() == AMDGPU::EG_ExportSwz ||
433 NextExportInst->getOpcode() == AMDGPU::R600_ExportSwz) {
434 unsigned CurrentInstExportType = NextExportInst->getOperand(1)
436 if (CurrentInstExportType == InstExportType) {
437 isLastInstructionOfItsType = false;
442 bool EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN)? 1 : 0;
443 if (!EOP && !isLastInstructionOfItsType)
445 unsigned CfInst = (MI->getOpcode() == AMDGPU::EG_ExportSwz)? 84 : 40;
446 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
447 .addOperand(MI->getOperand(0))
448 .addOperand(MI->getOperand(1))
449 .addOperand(MI->getOperand(2))
450 .addOperand(MI->getOperand(3))
451 .addOperand(MI->getOperand(4))
452 .addOperand(MI->getOperand(5))
453 .addOperand(MI->getOperand(6))
458 case AMDGPU::RETURN: {
459 // RETURN instructions must have the live-out registers as implicit uses,
460 // otherwise they appear dead.
461 R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
462 MachineInstrBuilder MIB(*MF, MI);
463 for (unsigned i = 0, e = MFI->LiveOuts.size(); i != e; ++i)
464 MIB.addReg(MFI->LiveOuts[i], RegState::Implicit);
469 MI->eraseFromParent();
473 //===----------------------------------------------------------------------===//
474 // Custom DAG Lowering Operations
475 //===----------------------------------------------------------------------===//
477 using namespace llvm::Intrinsic;
478 using namespace llvm::AMDGPUIntrinsic;
480 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
481 switch (Op.getOpcode()) {
482 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
483 case ISD::ROTL: return LowerROTL(Op, DAG);
484 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
485 case ISD::SELECT: return LowerSELECT(Op, DAG);
486 case ISD::STORE: return LowerSTORE(Op, DAG);
487 case ISD::LOAD: return LowerLOAD(Op, DAG);
488 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
489 case ISD::INTRINSIC_VOID: {
490 SDValue Chain = Op.getOperand(0);
491 unsigned IntrinsicID =
492 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
493 switch (IntrinsicID) {
494 case AMDGPUIntrinsic::AMDGPU_store_output: {
495 MachineFunction &MF = DAG.getMachineFunction();
496 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
497 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
498 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
499 MFI->LiveOuts.push_back(Reg);
500 return DAG.getCopyToReg(Chain, Op.getDebugLoc(), Reg, Op.getOperand(2));
502 case AMDGPUIntrinsic::R600_store_swizzle: {
503 const SDValue Args[8] = {
505 Op.getOperand(2), // Export Value
506 Op.getOperand(3), // ArrayBase
507 Op.getOperand(4), // Type
508 DAG.getConstant(0, MVT::i32), // SWZ_X
509 DAG.getConstant(1, MVT::i32), // SWZ_Y
510 DAG.getConstant(2, MVT::i32), // SWZ_Z
511 DAG.getConstant(3, MVT::i32) // SWZ_W
513 return DAG.getNode(AMDGPUISD::EXPORT, Op.getDebugLoc(), Op.getValueType(),
517 // default for switch(IntrinsicID)
520 // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
523 case ISD::INTRINSIC_WO_CHAIN: {
524 unsigned IntrinsicID =
525 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
526 EVT VT = Op.getValueType();
527 DebugLoc DL = Op.getDebugLoc();
528 switch(IntrinsicID) {
529 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
530 case AMDGPUIntrinsic::R600_load_input: {
531 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
532 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
533 MachineFunction &MF = DAG.getMachineFunction();
534 MachineRegisterInfo &MRI = MF.getRegInfo();
536 return DAG.getCopyFromReg(DAG.getEntryNode(),
537 DAG.getEntryNode().getDebugLoc(), Reg, VT);
540 case AMDGPUIntrinsic::R600_interp_input: {
541 int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
542 int ijb = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
543 MachineSDNode *interp;
545 interp = DAG.getMachineNode(AMDGPU::INTERP_VEC_LOAD, DL,
546 MVT::v4f32, DAG.getTargetConstant(slot / 4 , MVT::i32));
547 return DAG.getTargetExtractSubreg(
548 TII->getRegisterInfo().getSubRegFromChannel(slot % 4),
549 DL, MVT::f32, SDValue(interp, 0));
553 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_XY, DL,
554 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
555 CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
556 AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb + 1), MVT::f32),
557 CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
558 AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb), MVT::f32));
560 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_ZW, DL,
561 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
562 CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
563 AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb + 1), MVT::f32),
564 CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
565 AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb), MVT::f32));
567 return SDValue(interp, slot % 2);
569 case AMDGPUIntrinsic::R600_tex:
570 case AMDGPUIntrinsic::R600_texc:
571 case AMDGPUIntrinsic::R600_txl:
572 case AMDGPUIntrinsic::R600_txlc:
573 case AMDGPUIntrinsic::R600_txb:
574 case AMDGPUIntrinsic::R600_txbc:
575 case AMDGPUIntrinsic::R600_txf:
576 case AMDGPUIntrinsic::R600_txq:
577 case AMDGPUIntrinsic::R600_ddx:
578 case AMDGPUIntrinsic::R600_ddy: {
580 switch (IntrinsicID) {
581 case AMDGPUIntrinsic::R600_tex:
584 case AMDGPUIntrinsic::R600_texc:
587 case AMDGPUIntrinsic::R600_txl:
590 case AMDGPUIntrinsic::R600_txlc:
593 case AMDGPUIntrinsic::R600_txb:
596 case AMDGPUIntrinsic::R600_txbc:
599 case AMDGPUIntrinsic::R600_txf:
602 case AMDGPUIntrinsic::R600_txq:
605 case AMDGPUIntrinsic::R600_ddx:
608 case AMDGPUIntrinsic::R600_ddy:
612 llvm_unreachable("Unknow Texture Operation");
615 SDValue TexArgs[19] = {
616 DAG.getConstant(TextureOp, MVT::i32),
618 DAG.getConstant(0, MVT::i32),
619 DAG.getConstant(1, MVT::i32),
620 DAG.getConstant(2, MVT::i32),
621 DAG.getConstant(3, MVT::i32),
625 DAG.getConstant(0, MVT::i32),
626 DAG.getConstant(1, MVT::i32),
627 DAG.getConstant(2, MVT::i32),
628 DAG.getConstant(3, MVT::i32),
636 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs, 19);
638 case AMDGPUIntrinsic::AMDGPU_dp4: {
640 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
641 DAG.getConstant(0, MVT::i32)),
642 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
643 DAG.getConstant(0, MVT::i32)),
644 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
645 DAG.getConstant(1, MVT::i32)),
646 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
647 DAG.getConstant(1, MVT::i32)),
648 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
649 DAG.getConstant(2, MVT::i32)),
650 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
651 DAG.getConstant(2, MVT::i32)),
652 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
653 DAG.getConstant(3, MVT::i32)),
654 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
655 DAG.getConstant(3, MVT::i32))
657 return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args, 8);
660 case r600_read_ngroups_x:
661 return LowerImplicitParameter(DAG, VT, DL, 0);
662 case r600_read_ngroups_y:
663 return LowerImplicitParameter(DAG, VT, DL, 1);
664 case r600_read_ngroups_z:
665 return LowerImplicitParameter(DAG, VT, DL, 2);
666 case r600_read_global_size_x:
667 return LowerImplicitParameter(DAG, VT, DL, 3);
668 case r600_read_global_size_y:
669 return LowerImplicitParameter(DAG, VT, DL, 4);
670 case r600_read_global_size_z:
671 return LowerImplicitParameter(DAG, VT, DL, 5);
672 case r600_read_local_size_x:
673 return LowerImplicitParameter(DAG, VT, DL, 6);
674 case r600_read_local_size_y:
675 return LowerImplicitParameter(DAG, VT, DL, 7);
676 case r600_read_local_size_z:
677 return LowerImplicitParameter(DAG, VT, DL, 8);
679 case r600_read_tgid_x:
680 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
682 case r600_read_tgid_y:
683 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
685 case r600_read_tgid_z:
686 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
688 case r600_read_tidig_x:
689 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
691 case r600_read_tidig_y:
692 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
694 case r600_read_tidig_z:
695 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
698 // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
701 } // end switch(Op.getOpcode())
705 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
706 SmallVectorImpl<SDValue> &Results,
707 SelectionDAG &DAG) const {
708 switch (N->getOpcode()) {
710 case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
713 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
714 Results.push_back(SDValue(Node, 0));
715 Results.push_back(SDValue(Node, 1));
716 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
718 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
722 SDNode *Node = LowerSTORE(SDValue(N, 0), DAG).getNode();
723 Results.push_back(SDValue(Node, 0));
728 SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
733 Op, DAG.getConstantFP(0.0f, MVT::f32),
734 DAG.getCondCode(ISD::SETNE)
738 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
740 unsigned DwordOffset) const {
741 unsigned ByteOffset = DwordOffset * 4;
742 PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
743 AMDGPUAS::PARAM_I_ADDRESS);
745 // We shouldn't be using an offset wider than 16-bits for implicit parameters.
746 assert(isInt<16>(ByteOffset));
748 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
749 DAG.getConstant(ByteOffset, MVT::i32), // PTR
750 MachinePointerInfo(ConstantPointerNull::get(PtrType)),
751 false, false, false, 0);
754 SDValue R600TargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
756 MachineFunction &MF = DAG.getMachineFunction();
757 const AMDGPUFrameLowering *TFL =
758 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
760 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
763 unsigned FrameIndex = FIN->getIndex();
764 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
765 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), MVT::i32);
768 SDValue R600TargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const {
769 DebugLoc DL = Op.getDebugLoc();
770 EVT VT = Op.getValueType();
772 return DAG.getNode(AMDGPUISD::BITALIGN, DL, VT,
775 DAG.getNode(ISD::SUB, DL, VT,
776 DAG.getConstant(32, MVT::i32),
780 bool R600TargetLowering::isZero(SDValue Op) const {
781 if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
782 return Cst->isNullValue();
783 } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
784 return CstFP->isZero();
790 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
791 DebugLoc DL = Op.getDebugLoc();
792 EVT VT = Op.getValueType();
794 SDValue LHS = Op.getOperand(0);
795 SDValue RHS = Op.getOperand(1);
796 SDValue True = Op.getOperand(2);
797 SDValue False = Op.getOperand(3);
798 SDValue CC = Op.getOperand(4);
801 // LHS and RHS are guaranteed to be the same value type
802 EVT CompareVT = LHS.getValueType();
804 // Check if we can lower this to a native operation.
806 // Try to lower to a SET* instruction:
808 // SET* can match the following patterns:
810 // select_cc f32, f32, -1, 0, cc_any
811 // select_cc f32, f32, 1.0f, 0.0f, cc_any
812 // select_cc i32, i32, -1, 0, cc_any
815 // Move hardware True/False values to the correct operand.
816 if (isHWTrueValue(False) && isHWFalseValue(True)) {
817 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
818 std::swap(False, True);
819 CC = DAG.getCondCode(ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32));
822 if (isHWTrueValue(True) && isHWFalseValue(False) &&
823 (CompareVT == VT || VT == MVT::i32)) {
824 // This can be matched by a SET* instruction.
825 return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
828 // Try to lower to a CND* instruction:
830 // CND* can match the following patterns:
832 // select_cc f32, 0.0, f32, f32, cc_any
833 // select_cc f32, 0.0, i32, i32, cc_any
834 // select_cc i32, 0, f32, f32, cc_any
835 // select_cc i32, 0, i32, i32, cc_any
837 if (isZero(LHS) || isZero(RHS)) {
838 SDValue Cond = (isZero(LHS) ? RHS : LHS);
839 SDValue Zero = (isZero(LHS) ? LHS : RHS);
840 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
841 if (CompareVT != VT) {
842 // Bitcast True / False to the correct types. This will end up being
843 // a nop, but it allows us to define only a single pattern in the
844 // .TD files for each CND* instruction rather than having to have
845 // one pattern for integer True/False and one for fp True/False
846 True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
847 False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
850 CCOpcode = ISD::getSetCCSwappedOperands(CCOpcode);
863 CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
871 SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
874 DAG.getCondCode(CCOpcode));
875 return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
879 // Possible Min/Max pattern
880 SDValue MinMax = LowerMinMax(Op, DAG);
881 if (MinMax.getNode()) {
885 // If we make it this for it means we have no native instructions to handle
886 // this SELECT_CC, so we must lower it.
887 SDValue HWTrue, HWFalse;
889 if (CompareVT == MVT::f32) {
890 HWTrue = DAG.getConstantFP(1.0f, CompareVT);
891 HWFalse = DAG.getConstantFP(0.0f, CompareVT);
892 } else if (CompareVT == MVT::i32) {
893 HWTrue = DAG.getConstant(-1, CompareVT);
894 HWFalse = DAG.getConstant(0, CompareVT);
897 assert(!"Unhandled value type in LowerSELECT_CC");
900 // Lower this unsupported SELECT_CC into a combination of two supported
901 // SELECT_CC operations.
902 SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
904 return DAG.getNode(ISD::SELECT_CC, DL, VT,
907 DAG.getCondCode(ISD::SETNE));
910 SDValue R600TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
911 return DAG.getNode(ISD::SELECT_CC,
915 DAG.getConstant(0, MVT::i32),
918 DAG.getCondCode(ISD::SETNE));
921 /// LLVM generates byte-addresed pointers. For indirect addressing, we need to
922 /// convert these pointers to a register index. Each register holds
923 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
924 /// \p StackWidth, which tells us how many of the 4 sub-registrers will be used
925 /// for indirect addressing.
926 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
928 SelectionDAG &DAG) const {
940 default: llvm_unreachable("Invalid stack width");
943 return DAG.getNode(ISD::SRL, Ptr.getDebugLoc(), Ptr.getValueType(), Ptr,
944 DAG.getConstant(SRLPad, MVT::i32));
947 void R600TargetLowering::getStackAddress(unsigned StackWidth,
950 unsigned &PtrIncr) const {
951 switch (StackWidth) {
962 Channel = ElemIdx % 2;
976 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
977 DebugLoc DL = Op.getDebugLoc();
978 StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
979 SDValue Chain = Op.getOperand(0);
980 SDValue Value = Op.getOperand(1);
981 SDValue Ptr = Op.getOperand(2);
983 if (StoreNode->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
984 Ptr->getOpcode() != AMDGPUISD::DWORDADDR) {
985 // Convert pointer from byte address to dword address.
986 Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, Ptr.getValueType(),
987 DAG.getNode(ISD::SRL, DL, Ptr.getValueType(),
988 Ptr, DAG.getConstant(2, MVT::i32)));
990 if (StoreNode->isTruncatingStore() || StoreNode->isIndexed()) {
991 assert(!"Truncated and indexed stores not supported yet");
993 Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
998 EVT ValueVT = Value.getValueType();
1000 if (StoreNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1004 // Lowering for indirect addressing
1006 const MachineFunction &MF = DAG.getMachineFunction();
1007 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1008 getTargetMachine().getFrameLowering());
1009 unsigned StackWidth = TFL->getStackWidth(MF);
1011 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1013 if (ValueVT.isVector()) {
1014 unsigned NumElemVT = ValueVT.getVectorNumElements();
1015 EVT ElemVT = ValueVT.getVectorElementType();
1018 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1019 "vector width in load");
1021 for (unsigned i = 0; i < NumElemVT; ++i) {
1022 unsigned Channel, PtrIncr;
1023 getStackAddress(StackWidth, i, Channel, PtrIncr);
1024 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1025 DAG.getConstant(PtrIncr, MVT::i32));
1026 SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT,
1027 Value, DAG.getConstant(i, MVT::i32));
1029 Stores[i] = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1031 DAG.getTargetConstant(Channel, MVT::i32));
1033 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores, NumElemVT);
1035 if (ValueVT == MVT::i8) {
1036 Value = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Value);
1038 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other, Chain, Value, Ptr,
1039 DAG.getTargetConstant(0, MVT::i32)); // Channel
1045 // return (512 + (kc_bank << 12)
1047 ConstantAddressBlock(unsigned AddressSpace) {
1048 switch (AddressSpace) {
1049 case AMDGPUAS::CONSTANT_BUFFER_0:
1051 case AMDGPUAS::CONSTANT_BUFFER_1:
1053 case AMDGPUAS::CONSTANT_BUFFER_2:
1054 return 512 + 4096 * 2;
1055 case AMDGPUAS::CONSTANT_BUFFER_3:
1056 return 512 + 4096 * 3;
1057 case AMDGPUAS::CONSTANT_BUFFER_4:
1058 return 512 + 4096 * 4;
1059 case AMDGPUAS::CONSTANT_BUFFER_5:
1060 return 512 + 4096 * 5;
1061 case AMDGPUAS::CONSTANT_BUFFER_6:
1062 return 512 + 4096 * 6;
1063 case AMDGPUAS::CONSTANT_BUFFER_7:
1064 return 512 + 4096 * 7;
1065 case AMDGPUAS::CONSTANT_BUFFER_8:
1066 return 512 + 4096 * 8;
1067 case AMDGPUAS::CONSTANT_BUFFER_9:
1068 return 512 + 4096 * 9;
1069 case AMDGPUAS::CONSTANT_BUFFER_10:
1070 return 512 + 4096 * 10;
1071 case AMDGPUAS::CONSTANT_BUFFER_11:
1072 return 512 + 4096 * 11;
1073 case AMDGPUAS::CONSTANT_BUFFER_12:
1074 return 512 + 4096 * 12;
1075 case AMDGPUAS::CONSTANT_BUFFER_13:
1076 return 512 + 4096 * 13;
1077 case AMDGPUAS::CONSTANT_BUFFER_14:
1078 return 512 + 4096 * 14;
1079 case AMDGPUAS::CONSTANT_BUFFER_15:
1080 return 512 + 4096 * 15;
1086 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
1088 EVT VT = Op.getValueType();
1089 DebugLoc DL = Op.getDebugLoc();
1090 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1091 SDValue Chain = Op.getOperand(0);
1092 SDValue Ptr = Op.getOperand(1);
1093 SDValue LoweredLoad;
1095 int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1096 if (ConstantBlock > -1) {
1098 if (dyn_cast<ConstantExpr>(LoadNode->getSrcValue()) ||
1099 dyn_cast<Constant>(LoadNode->getSrcValue()) ||
1100 dyn_cast<ConstantSDNode>(Ptr)) {
1102 for (unsigned i = 0; i < 4; i++) {
1103 // We want Const position encoded with the following formula :
1104 // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1105 // const_index is Ptr computed by llvm using an alignment of 16.
1106 // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1107 // then div by 4 at the ISel step
1108 SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1109 DAG.getConstant(4 * i + ConstantBlock * 16, MVT::i32));
1110 Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1112 Result = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Slots, 4);
1114 // non constant ptr cant be folded, keeps it as a v4f32 load
1115 Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1116 DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(4, MVT::i32)),
1117 DAG.getConstant(LoadNode->getAddressSpace() -
1118 AMDGPUAS::CONSTANT_BUFFER_0, MVT::i32)
1122 if (!VT.isVector()) {
1123 Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1124 DAG.getConstant(0, MVT::i32));
1127 SDValue MergedValues[2] = {
1131 return DAG.getMergeValues(MergedValues, 2, DL);
1134 if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1138 // Lowering for indirect addressing
1139 const MachineFunction &MF = DAG.getMachineFunction();
1140 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1141 getTargetMachine().getFrameLowering());
1142 unsigned StackWidth = TFL->getStackWidth(MF);
1144 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1146 if (VT.isVector()) {
1147 unsigned NumElemVT = VT.getVectorNumElements();
1148 EVT ElemVT = VT.getVectorElementType();
1151 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1152 "vector width in load");
1154 for (unsigned i = 0; i < NumElemVT; ++i) {
1155 unsigned Channel, PtrIncr;
1156 getStackAddress(StackWidth, i, Channel, PtrIncr);
1157 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1158 DAG.getConstant(PtrIncr, MVT::i32));
1159 Loads[i] = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, ElemVT,
1161 DAG.getTargetConstant(Channel, MVT::i32),
1164 for (unsigned i = NumElemVT; i < 4; ++i) {
1165 Loads[i] = DAG.getUNDEF(ElemVT);
1167 EVT TargetVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, 4);
1168 LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads, 4);
1170 LoweredLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, VT,
1172 DAG.getTargetConstant(0, MVT::i32), // Channel
1177 Ops[0] = LoweredLoad;
1180 return DAG.getMergeValues(Ops, 2, DL);
1183 /// XXX Only kernel functions are supported, so we can assume for now that
1184 /// every function is a kernel function, but in the future we should use
1185 /// separate calling conventions for kernel and non-kernel functions.
1186 SDValue R600TargetLowering::LowerFormalArguments(
1188 CallingConv::ID CallConv,
1190 const SmallVectorImpl<ISD::InputArg> &Ins,
1191 DebugLoc DL, SelectionDAG &DAG,
1192 SmallVectorImpl<SDValue> &InVals) const {
1193 unsigned ParamOffsetBytes = 36;
1194 Function::const_arg_iterator FuncArg =
1195 DAG.getMachineFunction().getFunction()->arg_begin();
1196 for (unsigned i = 0, e = Ins.size(); i < e; ++i, ++FuncArg) {
1198 Type *ArgType = FuncArg->getType();
1199 unsigned ArgSizeInBits = ArgType->isPointerTy() ?
1200 32 : ArgType->getPrimitiveSizeInBits();
1201 unsigned ArgBytes = ArgSizeInBits >> 3;
1203 if (ArgSizeInBits < VT.getSizeInBits()) {
1204 assert(!ArgType->isFloatTy() &&
1205 "Extending floating point arguments not supported yet");
1206 ArgVT = MVT::getIntegerVT(ArgSizeInBits);
1210 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
1211 AMDGPUAS::PARAM_I_ADDRESS);
1212 SDValue Arg = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getRoot(),
1213 DAG.getConstant(ParamOffsetBytes, MVT::i32),
1214 MachinePointerInfo(UndefValue::get(PtrTy)),
1215 ArgVT, false, false, ArgBytes);
1216 InVals.push_back(Arg);
1217 ParamOffsetBytes += ArgBytes;
1222 EVT R600TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1223 if (!VT.isVector()) return MVT::i32;
1224 return VT.changeVectorElementTypeToInteger();
1227 //===----------------------------------------------------------------------===//
1228 // Custom DAG Optimizations
1229 //===----------------------------------------------------------------------===//
1231 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1232 DAGCombinerInfo &DCI) const {
1233 SelectionDAG &DAG = DCI.DAG;
1235 switch (N->getOpcode()) {
1236 // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1237 case ISD::FP_ROUND: {
1238 SDValue Arg = N->getOperand(0);
1239 if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1240 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), N->getValueType(0),
1246 // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1247 // (i32 select_cc f32, f32, -1, 0 cc)
1249 // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1250 // this to one of the SET*_DX10 instructions.
1251 case ISD::FP_TO_SINT: {
1252 SDValue FNeg = N->getOperand(0);
1253 if (FNeg.getOpcode() != ISD::FNEG) {
1256 SDValue SelectCC = FNeg.getOperand(0);
1257 if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1258 SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1259 SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1260 !isHWTrueValue(SelectCC.getOperand(2)) ||
1261 !isHWFalseValue(SelectCC.getOperand(3))) {
1265 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N->getValueType(0),
1266 SelectCC.getOperand(0), // LHS
1267 SelectCC.getOperand(1), // RHS
1268 DAG.getConstant(-1, MVT::i32), // True
1269 DAG.getConstant(0, MVT::i32), // Flase
1270 SelectCC.getOperand(4)); // CC
1274 // Extract_vec (Build_vector) generated by custom lowering
1275 // also needs to be customly combined
1276 case ISD::EXTRACT_VECTOR_ELT: {
1277 SDValue Arg = N->getOperand(0);
1278 if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1279 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1280 unsigned Element = Const->getZExtValue();
1281 return Arg->getOperand(Element);
1284 if (Arg.getOpcode() == ISD::BITCAST &&
1285 Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
1286 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1287 unsigned Element = Const->getZExtValue();
1288 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), N->getVTList(),
1289 Arg->getOperand(0).getOperand(Element));
1294 case ISD::SELECT_CC: {
1295 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1296 // selectcc x, y, a, b, inv(cc)
1298 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1299 // selectcc x, y, a, b, cc
1300 SDValue LHS = N->getOperand(0);
1301 if (LHS.getOpcode() != ISD::SELECT_CC) {
1305 SDValue RHS = N->getOperand(1);
1306 SDValue True = N->getOperand(2);
1307 SDValue False = N->getOperand(3);
1308 ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1310 if (LHS.getOperand(2).getNode() != True.getNode() ||
1311 LHS.getOperand(3).getNode() != False.getNode() ||
1312 RHS.getNode() != False.getNode()) {
1317 default: return SDValue();
1318 case ISD::SETNE: return LHS;
1320 ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1321 LHSCC = ISD::getSetCCInverse(LHSCC,
1322 LHS.getOperand(0).getValueType().isInteger());
1323 return DAG.getSelectCC(N->getDebugLoc(),
1332 case AMDGPUISD::EXPORT: {
1333 SDValue Arg = N->getOperand(1);
1334 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1336 SDValue NewBldVec[4] = {
1337 DAG.getUNDEF(MVT::f32),
1338 DAG.getUNDEF(MVT::f32),
1339 DAG.getUNDEF(MVT::f32),
1340 DAG.getUNDEF(MVT::f32)
1342 SDValue NewArgs[8] = {
1343 N->getOperand(0), // Chain
1345 N->getOperand(2), // ArrayBase
1346 N->getOperand(3), // Type
1347 N->getOperand(4), // SWZ_X
1348 N->getOperand(5), // SWZ_Y
1349 N->getOperand(6), // SWZ_Z
1350 N->getOperand(7) // SWZ_W
1352 for (unsigned i = 0; i < Arg.getNumOperands(); i++) {
1353 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Arg.getOperand(i))) {
1355 NewArgs[4 + i] = DAG.getConstant(4, MVT::i32); // SEL_0
1356 } else if (C->isExactlyValue(1.0)) {
1357 NewArgs[4 + i] = DAG.getConstant(5, MVT::i32); // SEL_0
1359 NewBldVec[i] = Arg.getOperand(i);
1362 NewBldVec[i] = Arg.getOperand(i);
1365 DebugLoc DL = N->getDebugLoc();
1366 NewArgs[1] = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4f32, NewBldVec, 4);
1367 return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs, 8);