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 addRegisterClass(MVT::v4f32, &AMDGPU::R600_Reg128RegClass);
31 addRegisterClass(MVT::f32, &AMDGPU::R600_Reg32RegClass);
32 addRegisterClass(MVT::v4i32, &AMDGPU::R600_Reg128RegClass);
33 addRegisterClass(MVT::i32, &AMDGPU::R600_Reg32RegClass);
34 computeRegisterProperties();
36 setOperationAction(ISD::FADD, MVT::v4f32, Expand);
37 setOperationAction(ISD::FMUL, MVT::v4f32, Expand);
38 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
39 setOperationAction(ISD::FSUB, MVT::v4f32, Expand);
41 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Expand);
42 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Expand);
43 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Expand);
44 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Expand);
45 setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
47 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
48 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
50 setOperationAction(ISD::FSUB, MVT::f32, Expand);
52 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
53 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
54 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
56 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
57 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
59 setOperationAction(ISD::SETCC, MVT::i32, Expand);
60 setOperationAction(ISD::SETCC, MVT::f32, Expand);
61 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
63 setOperationAction(ISD::SELECT, MVT::i32, Custom);
64 setOperationAction(ISD::SELECT, MVT::f32, Custom);
66 setOperationAction(ISD::VSELECT, MVT::v4i32, Expand);
67 setOperationAction(ISD::VSELECT, MVT::v2i32, Expand);
69 // Legalize loads and stores to the private address space.
70 setOperationAction(ISD::LOAD, MVT::i32, Custom);
71 setOperationAction(ISD::LOAD, MVT::v2i32, Expand);
72 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
73 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Custom);
74 setLoadExtAction(ISD::EXTLOAD, MVT::i8, Custom);
75 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
76 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Custom);
77 setOperationAction(ISD::STORE, MVT::i8, Custom);
78 setOperationAction(ISD::STORE, MVT::i32, Custom);
79 setOperationAction(ISD::STORE, MVT::v2i32, Expand);
80 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
82 setOperationAction(ISD::LOAD, MVT::i32, Custom);
83 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
84 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
86 setTargetDAGCombine(ISD::FP_ROUND);
87 setTargetDAGCombine(ISD::FP_TO_SINT);
88 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
89 setTargetDAGCombine(ISD::SELECT_CC);
91 setBooleanContents(ZeroOrNegativeOneBooleanContent);
92 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
93 setSchedulingPreference(Sched::VLIW);
96 MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
97 MachineInstr * MI, MachineBasicBlock * BB) const {
98 MachineFunction * MF = BB->getParent();
99 MachineRegisterInfo &MRI = MF->getRegInfo();
100 MachineBasicBlock::iterator I = *MI;
101 const R600InstrInfo *TII =
102 static_cast<const R600InstrInfo*>(MF->getTarget().getInstrInfo());
104 switch (MI->getOpcode()) {
105 default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
106 case AMDGPU::CLAMP_R600: {
107 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
109 MI->getOperand(0).getReg(),
110 MI->getOperand(1).getReg());
111 TII->addFlag(NewMI, 0, MO_FLAG_CLAMP);
115 case AMDGPU::FABS_R600: {
116 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
118 MI->getOperand(0).getReg(),
119 MI->getOperand(1).getReg());
120 TII->addFlag(NewMI, 0, MO_FLAG_ABS);
124 case AMDGPU::FNEG_R600: {
125 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
127 MI->getOperand(0).getReg(),
128 MI->getOperand(1).getReg());
129 TII->addFlag(NewMI, 0, MO_FLAG_NEG);
133 case AMDGPU::MASK_WRITE: {
134 unsigned maskedRegister = MI->getOperand(0).getReg();
135 assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
136 MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
137 TII->addFlag(defInstr, 0, MO_FLAG_MASK);
141 case AMDGPU::LDS_READ_RET: {
142 MachineInstrBuilder NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
143 TII->get(MI->getOpcode()),
145 for (unsigned i = 1, e = MI->getNumOperands(); i < e; ++i) {
146 NewMI.addOperand(MI->getOperand(i));
148 TII->buildDefaultInstruction(*BB, I, AMDGPU::MOV,
149 MI->getOperand(0).getReg(),
154 case AMDGPU::MOV_IMM_F32:
155 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
156 MI->getOperand(1).getFPImm()->getValueAPF()
157 .bitcastToAPInt().getZExtValue());
159 case AMDGPU::MOV_IMM_I32:
160 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
161 MI->getOperand(1).getImm());
163 case AMDGPU::CONST_COPY: {
164 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, MI, AMDGPU::MOV,
165 MI->getOperand(0).getReg(), AMDGPU::ALU_CONST);
166 TII->setImmOperand(NewMI, AMDGPU::OpName::src0_sel,
167 MI->getOperand(1).getImm());
171 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
172 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
173 unsigned EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN) ? 1 : 0;
175 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
176 .addOperand(MI->getOperand(0))
177 .addOperand(MI->getOperand(1))
178 .addImm(EOP); // Set End of program bit
183 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
184 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
185 MachineOperand &RID = MI->getOperand(4);
186 MachineOperand &SID = MI->getOperand(5);
187 unsigned TextureId = MI->getOperand(6).getImm();
188 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
189 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
201 case 8: // ShadowRect
212 case 11: // Shadow1DArray
216 case 12: // Shadow2DArray
220 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
221 .addOperand(MI->getOperand(3))
239 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
240 .addOperand(MI->getOperand(2))
258 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
259 .addOperand(MI->getOperand(0))
260 .addOperand(MI->getOperand(1))
278 .addReg(T0, RegState::Implicit)
279 .addReg(T1, RegState::Implicit);
283 case AMDGPU::TXD_SHADOW: {
284 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
285 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
286 MachineOperand &RID = MI->getOperand(4);
287 MachineOperand &SID = MI->getOperand(5);
288 unsigned TextureId = MI->getOperand(6).getImm();
289 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
290 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
302 case 8: // ShadowRect
313 case 11: // Shadow1DArray
317 case 12: // Shadow2DArray
322 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
323 .addOperand(MI->getOperand(3))
341 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
342 .addOperand(MI->getOperand(2))
360 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
361 .addOperand(MI->getOperand(0))
362 .addOperand(MI->getOperand(1))
380 .addReg(T0, RegState::Implicit)
381 .addReg(T1, RegState::Implicit);
386 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
387 .addOperand(MI->getOperand(0));
390 case AMDGPU::BRANCH_COND_f32: {
391 MachineInstr *NewMI =
392 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
393 AMDGPU::PREDICATE_BIT)
394 .addOperand(MI->getOperand(1))
395 .addImm(OPCODE_IS_NOT_ZERO)
397 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
398 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
399 .addOperand(MI->getOperand(0))
400 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
404 case AMDGPU::BRANCH_COND_i32: {
405 MachineInstr *NewMI =
406 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
407 AMDGPU::PREDICATE_BIT)
408 .addOperand(MI->getOperand(1))
409 .addImm(OPCODE_IS_NOT_ZERO_INT)
411 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
412 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
413 .addOperand(MI->getOperand(0))
414 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
418 case AMDGPU::EG_ExportSwz:
419 case AMDGPU::R600_ExportSwz: {
420 // Instruction is left unmodified if its not the last one of its type
421 bool isLastInstructionOfItsType = true;
422 unsigned InstExportType = MI->getOperand(1).getImm();
423 for (MachineBasicBlock::iterator NextExportInst = llvm::next(I),
424 EndBlock = BB->end(); NextExportInst != EndBlock;
425 NextExportInst = llvm::next(NextExportInst)) {
426 if (NextExportInst->getOpcode() == AMDGPU::EG_ExportSwz ||
427 NextExportInst->getOpcode() == AMDGPU::R600_ExportSwz) {
428 unsigned CurrentInstExportType = NextExportInst->getOperand(1)
430 if (CurrentInstExportType == InstExportType) {
431 isLastInstructionOfItsType = false;
436 bool EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN)? 1 : 0;
437 if (!EOP && !isLastInstructionOfItsType)
439 unsigned CfInst = (MI->getOpcode() == AMDGPU::EG_ExportSwz)? 84 : 40;
440 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
441 .addOperand(MI->getOperand(0))
442 .addOperand(MI->getOperand(1))
443 .addOperand(MI->getOperand(2))
444 .addOperand(MI->getOperand(3))
445 .addOperand(MI->getOperand(4))
446 .addOperand(MI->getOperand(5))
447 .addOperand(MI->getOperand(6))
452 case AMDGPU::RETURN: {
453 // RETURN instructions must have the live-out registers as implicit uses,
454 // otherwise they appear dead.
455 R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
456 MachineInstrBuilder MIB(*MF, MI);
457 for (unsigned i = 0, e = MFI->LiveOuts.size(); i != e; ++i)
458 MIB.addReg(MFI->LiveOuts[i], RegState::Implicit);
463 MI->eraseFromParent();
467 //===----------------------------------------------------------------------===//
468 // Custom DAG Lowering Operations
469 //===----------------------------------------------------------------------===//
471 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
472 MachineFunction &MF = DAG.getMachineFunction();
473 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
474 switch (Op.getOpcode()) {
475 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
476 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
477 case ISD::SELECT: return LowerSELECT(Op, DAG);
478 case ISD::STORE: return LowerSTORE(Op, DAG);
479 case ISD::LOAD: return LowerLOAD(Op, DAG);
480 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
481 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
482 case ISD::INTRINSIC_VOID: {
483 SDValue Chain = Op.getOperand(0);
484 unsigned IntrinsicID =
485 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
486 switch (IntrinsicID) {
487 case AMDGPUIntrinsic::AMDGPU_store_output: {
488 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
489 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
490 MFI->LiveOuts.push_back(Reg);
491 return DAG.getCopyToReg(Chain, SDLoc(Op), Reg, Op.getOperand(2));
493 case AMDGPUIntrinsic::R600_store_swizzle: {
494 const SDValue Args[8] = {
496 Op.getOperand(2), // Export Value
497 Op.getOperand(3), // ArrayBase
498 Op.getOperand(4), // Type
499 DAG.getConstant(0, MVT::i32), // SWZ_X
500 DAG.getConstant(1, MVT::i32), // SWZ_Y
501 DAG.getConstant(2, MVT::i32), // SWZ_Z
502 DAG.getConstant(3, MVT::i32) // SWZ_W
504 return DAG.getNode(AMDGPUISD::EXPORT, SDLoc(Op), Op.getValueType(),
508 // default for switch(IntrinsicID)
511 // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
514 case ISD::INTRINSIC_WO_CHAIN: {
515 unsigned IntrinsicID =
516 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
517 EVT VT = Op.getValueType();
519 switch(IntrinsicID) {
520 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
521 case AMDGPUIntrinsic::R600_load_input: {
522 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
523 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
524 MachineFunction &MF = DAG.getMachineFunction();
525 MachineRegisterInfo &MRI = MF.getRegInfo();
527 return DAG.getCopyFromReg(DAG.getEntryNode(),
528 SDLoc(DAG.getEntryNode()), Reg, VT);
531 case AMDGPUIntrinsic::R600_interp_input: {
532 int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
533 int ijb = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
534 MachineSDNode *interp;
536 const MachineFunction &MF = DAG.getMachineFunction();
537 const R600InstrInfo *TII =
538 static_cast<const R600InstrInfo*>(MF.getTarget().getInstrInfo());
539 interp = DAG.getMachineNode(AMDGPU::INTERP_VEC_LOAD, DL,
540 MVT::v4f32, DAG.getTargetConstant(slot / 4 , MVT::i32));
541 return DAG.getTargetExtractSubreg(
542 TII->getRegisterInfo().getSubRegFromChannel(slot % 4),
543 DL, MVT::f32, SDValue(interp, 0));
546 MachineFunction &MF = DAG.getMachineFunction();
547 MachineRegisterInfo &MRI = MF.getRegInfo();
548 unsigned RegisterI = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb);
549 unsigned RegisterJ = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb + 1);
550 MRI.addLiveIn(RegisterI);
551 MRI.addLiveIn(RegisterJ);
552 SDValue RegisterINode = DAG.getCopyFromReg(DAG.getEntryNode(),
553 SDLoc(DAG.getEntryNode()), RegisterI, MVT::f32);
554 SDValue RegisterJNode = DAG.getCopyFromReg(DAG.getEntryNode(),
555 SDLoc(DAG.getEntryNode()), RegisterJ, MVT::f32);
558 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_XY, DL,
559 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
560 RegisterJNode, RegisterINode);
562 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_ZW, DL,
563 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
564 RegisterJNode, RegisterINode);
565 return SDValue(interp, slot % 2);
567 case AMDGPUIntrinsic::R600_tex:
568 case AMDGPUIntrinsic::R600_texc:
569 case AMDGPUIntrinsic::R600_txl:
570 case AMDGPUIntrinsic::R600_txlc:
571 case AMDGPUIntrinsic::R600_txb:
572 case AMDGPUIntrinsic::R600_txbc:
573 case AMDGPUIntrinsic::R600_txf:
574 case AMDGPUIntrinsic::R600_txq:
575 case AMDGPUIntrinsic::R600_ddx:
576 case AMDGPUIntrinsic::R600_ddy: {
578 switch (IntrinsicID) {
579 case AMDGPUIntrinsic::R600_tex:
582 case AMDGPUIntrinsic::R600_texc:
585 case AMDGPUIntrinsic::R600_txl:
588 case AMDGPUIntrinsic::R600_txlc:
591 case AMDGPUIntrinsic::R600_txb:
594 case AMDGPUIntrinsic::R600_txbc:
597 case AMDGPUIntrinsic::R600_txf:
600 case AMDGPUIntrinsic::R600_txq:
603 case AMDGPUIntrinsic::R600_ddx:
606 case AMDGPUIntrinsic::R600_ddy:
610 llvm_unreachable("Unknow Texture Operation");
613 SDValue TexArgs[19] = {
614 DAG.getConstant(TextureOp, MVT::i32),
616 DAG.getConstant(0, MVT::i32),
617 DAG.getConstant(1, MVT::i32),
618 DAG.getConstant(2, MVT::i32),
619 DAG.getConstant(3, MVT::i32),
623 DAG.getConstant(0, MVT::i32),
624 DAG.getConstant(1, MVT::i32),
625 DAG.getConstant(2, MVT::i32),
626 DAG.getConstant(3, MVT::i32),
634 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs, 19);
636 case AMDGPUIntrinsic::AMDGPU_dp4: {
638 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
639 DAG.getConstant(0, MVT::i32)),
640 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
641 DAG.getConstant(0, MVT::i32)),
642 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
643 DAG.getConstant(1, MVT::i32)),
644 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
645 DAG.getConstant(1, MVT::i32)),
646 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
647 DAG.getConstant(2, MVT::i32)),
648 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
649 DAG.getConstant(2, MVT::i32)),
650 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
651 DAG.getConstant(3, MVT::i32)),
652 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
653 DAG.getConstant(3, MVT::i32))
655 return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args, 8);
658 case Intrinsic::r600_read_ngroups_x:
659 return LowerImplicitParameter(DAG, VT, DL, 0);
660 case Intrinsic::r600_read_ngroups_y:
661 return LowerImplicitParameter(DAG, VT, DL, 1);
662 case Intrinsic::r600_read_ngroups_z:
663 return LowerImplicitParameter(DAG, VT, DL, 2);
664 case Intrinsic::r600_read_global_size_x:
665 return LowerImplicitParameter(DAG, VT, DL, 3);
666 case Intrinsic::r600_read_global_size_y:
667 return LowerImplicitParameter(DAG, VT, DL, 4);
668 case Intrinsic::r600_read_global_size_z:
669 return LowerImplicitParameter(DAG, VT, DL, 5);
670 case Intrinsic::r600_read_local_size_x:
671 return LowerImplicitParameter(DAG, VT, DL, 6);
672 case Intrinsic::r600_read_local_size_y:
673 return LowerImplicitParameter(DAG, VT, DL, 7);
674 case Intrinsic::r600_read_local_size_z:
675 return LowerImplicitParameter(DAG, VT, DL, 8);
677 case Intrinsic::r600_read_tgid_x:
678 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
680 case Intrinsic::r600_read_tgid_y:
681 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
683 case Intrinsic::r600_read_tgid_z:
684 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
686 case Intrinsic::r600_read_tidig_x:
687 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
689 case Intrinsic::r600_read_tidig_y:
690 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
692 case Intrinsic::r600_read_tidig_z:
693 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
696 // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
699 } // end switch(Op.getOpcode())
703 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
704 SmallVectorImpl<SDValue> &Results,
705 SelectionDAG &DAG) const {
706 switch (N->getOpcode()) {
708 case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
711 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
712 Results.push_back(SDValue(Node, 0));
713 Results.push_back(SDValue(Node, 1));
714 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
716 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
720 SDNode *Node = LowerSTORE(SDValue(N, 0), DAG).getNode();
721 Results.push_back(SDValue(Node, 0));
726 SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
731 Op, DAG.getConstantFP(0.0f, MVT::f32),
732 DAG.getCondCode(ISD::SETNE)
736 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
738 unsigned DwordOffset) const {
739 unsigned ByteOffset = DwordOffset * 4;
740 PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
741 AMDGPUAS::PARAM_I_ADDRESS);
743 // We shouldn't be using an offset wider than 16-bits for implicit parameters.
744 assert(isInt<16>(ByteOffset));
746 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
747 DAG.getConstant(ByteOffset, MVT::i32), // PTR
748 MachinePointerInfo(ConstantPointerNull::get(PtrType)),
749 false, false, false, 0);
752 SDValue R600TargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
754 MachineFunction &MF = DAG.getMachineFunction();
755 const AMDGPUFrameLowering *TFL =
756 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
758 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
761 unsigned FrameIndex = FIN->getIndex();
762 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
763 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), MVT::i32);
766 bool R600TargetLowering::isZero(SDValue Op) const {
767 if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
768 return Cst->isNullValue();
769 } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
770 return CstFP->isZero();
776 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
778 EVT VT = Op.getValueType();
780 SDValue LHS = Op.getOperand(0);
781 SDValue RHS = Op.getOperand(1);
782 SDValue True = Op.getOperand(2);
783 SDValue False = Op.getOperand(3);
784 SDValue CC = Op.getOperand(4);
787 // LHS and RHS are guaranteed to be the same value type
788 EVT CompareVT = LHS.getValueType();
790 // Check if we can lower this to a native operation.
792 // Try to lower to a SET* instruction:
794 // SET* can match the following patterns:
796 // select_cc f32, f32, -1, 0, cc_any
797 // select_cc f32, f32, 1.0f, 0.0f, cc_any
798 // select_cc i32, i32, -1, 0, cc_any
801 // Move hardware True/False values to the correct operand.
802 if (isHWTrueValue(False) && isHWFalseValue(True)) {
803 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
804 std::swap(False, True);
805 CC = DAG.getCondCode(ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32));
808 if (isHWTrueValue(True) && isHWFalseValue(False) &&
809 (CompareVT == VT || VT == MVT::i32)) {
810 // This can be matched by a SET* instruction.
811 return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
814 // Try to lower to a CND* instruction:
816 // CND* can match the following patterns:
818 // select_cc f32, 0.0, f32, f32, cc_any
819 // select_cc f32, 0.0, i32, i32, cc_any
820 // select_cc i32, 0, f32, f32, cc_any
821 // select_cc i32, 0, i32, i32, cc_any
823 if (isZero(LHS) || isZero(RHS)) {
824 SDValue Cond = (isZero(LHS) ? RHS : LHS);
825 SDValue Zero = (isZero(LHS) ? LHS : RHS);
826 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
827 if (CompareVT != VT) {
828 // Bitcast True / False to the correct types. This will end up being
829 // a nop, but it allows us to define only a single pattern in the
830 // .TD files for each CND* instruction rather than having to have
831 // one pattern for integer True/False and one for fp True/False
832 True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
833 False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
836 CCOpcode = ISD::getSetCCSwappedOperands(CCOpcode);
849 CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
857 SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
860 DAG.getCondCode(CCOpcode));
861 return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
865 // Possible Min/Max pattern
866 SDValue MinMax = LowerMinMax(Op, DAG);
867 if (MinMax.getNode()) {
871 // If we make it this for it means we have no native instructions to handle
872 // this SELECT_CC, so we must lower it.
873 SDValue HWTrue, HWFalse;
875 if (CompareVT == MVT::f32) {
876 HWTrue = DAG.getConstantFP(1.0f, CompareVT);
877 HWFalse = DAG.getConstantFP(0.0f, CompareVT);
878 } else if (CompareVT == MVT::i32) {
879 HWTrue = DAG.getConstant(-1, CompareVT);
880 HWFalse = DAG.getConstant(0, CompareVT);
883 assert(!"Unhandled value type in LowerSELECT_CC");
886 // Lower this unsupported SELECT_CC into a combination of two supported
887 // SELECT_CC operations.
888 SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
890 return DAG.getNode(ISD::SELECT_CC, DL, VT,
893 DAG.getCondCode(ISD::SETNE));
896 SDValue R600TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
897 return DAG.getNode(ISD::SELECT_CC,
901 DAG.getConstant(0, MVT::i32),
904 DAG.getCondCode(ISD::SETNE));
907 /// LLVM generates byte-addresed pointers. For indirect addressing, we need to
908 /// convert these pointers to a register index. Each register holds
909 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
910 /// \p StackWidth, which tells us how many of the 4 sub-registrers will be used
911 /// for indirect addressing.
912 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
914 SelectionDAG &DAG) const {
926 default: llvm_unreachable("Invalid stack width");
929 return DAG.getNode(ISD::SRL, SDLoc(Ptr), Ptr.getValueType(), Ptr,
930 DAG.getConstant(SRLPad, MVT::i32));
933 void R600TargetLowering::getStackAddress(unsigned StackWidth,
936 unsigned &PtrIncr) const {
937 switch (StackWidth) {
948 Channel = ElemIdx % 2;
962 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
964 StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
965 SDValue Chain = Op.getOperand(0);
966 SDValue Value = Op.getOperand(1);
967 SDValue Ptr = Op.getOperand(2);
969 if (StoreNode->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
970 Ptr->getOpcode() != AMDGPUISD::DWORDADDR) {
971 // Convert pointer from byte address to dword address.
972 Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, Ptr.getValueType(),
973 DAG.getNode(ISD::SRL, DL, Ptr.getValueType(),
974 Ptr, DAG.getConstant(2, MVT::i32)));
976 if (StoreNode->isTruncatingStore() || StoreNode->isIndexed()) {
977 assert(!"Truncated and indexed stores not supported yet");
979 Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
984 EVT ValueVT = Value.getValueType();
986 if (StoreNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
990 // Lowering for indirect addressing
992 const MachineFunction &MF = DAG.getMachineFunction();
993 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
994 getTargetMachine().getFrameLowering());
995 unsigned StackWidth = TFL->getStackWidth(MF);
997 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
999 if (ValueVT.isVector()) {
1000 unsigned NumElemVT = ValueVT.getVectorNumElements();
1001 EVT ElemVT = ValueVT.getVectorElementType();
1004 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1005 "vector width in load");
1007 for (unsigned i = 0; i < NumElemVT; ++i) {
1008 unsigned Channel, PtrIncr;
1009 getStackAddress(StackWidth, i, Channel, PtrIncr);
1010 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1011 DAG.getConstant(PtrIncr, MVT::i32));
1012 SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT,
1013 Value, DAG.getConstant(i, MVT::i32));
1015 Stores[i] = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1017 DAG.getTargetConstant(Channel, MVT::i32));
1019 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores, NumElemVT);
1021 if (ValueVT == MVT::i8) {
1022 Value = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Value);
1024 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other, Chain, Value, Ptr,
1025 DAG.getTargetConstant(0, MVT::i32)); // Channel
1031 // return (512 + (kc_bank << 12)
1033 ConstantAddressBlock(unsigned AddressSpace) {
1034 switch (AddressSpace) {
1035 case AMDGPUAS::CONSTANT_BUFFER_0:
1037 case AMDGPUAS::CONSTANT_BUFFER_1:
1039 case AMDGPUAS::CONSTANT_BUFFER_2:
1040 return 512 + 4096 * 2;
1041 case AMDGPUAS::CONSTANT_BUFFER_3:
1042 return 512 + 4096 * 3;
1043 case AMDGPUAS::CONSTANT_BUFFER_4:
1044 return 512 + 4096 * 4;
1045 case AMDGPUAS::CONSTANT_BUFFER_5:
1046 return 512 + 4096 * 5;
1047 case AMDGPUAS::CONSTANT_BUFFER_6:
1048 return 512 + 4096 * 6;
1049 case AMDGPUAS::CONSTANT_BUFFER_7:
1050 return 512 + 4096 * 7;
1051 case AMDGPUAS::CONSTANT_BUFFER_8:
1052 return 512 + 4096 * 8;
1053 case AMDGPUAS::CONSTANT_BUFFER_9:
1054 return 512 + 4096 * 9;
1055 case AMDGPUAS::CONSTANT_BUFFER_10:
1056 return 512 + 4096 * 10;
1057 case AMDGPUAS::CONSTANT_BUFFER_11:
1058 return 512 + 4096 * 11;
1059 case AMDGPUAS::CONSTANT_BUFFER_12:
1060 return 512 + 4096 * 12;
1061 case AMDGPUAS::CONSTANT_BUFFER_13:
1062 return 512 + 4096 * 13;
1063 case AMDGPUAS::CONSTANT_BUFFER_14:
1064 return 512 + 4096 * 14;
1065 case AMDGPUAS::CONSTANT_BUFFER_15:
1066 return 512 + 4096 * 15;
1072 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
1074 EVT VT = Op.getValueType();
1076 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1077 SDValue Chain = Op.getOperand(0);
1078 SDValue Ptr = Op.getOperand(1);
1079 SDValue LoweredLoad;
1081 int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1082 if (ConstantBlock > -1) {
1084 if (dyn_cast<ConstantExpr>(LoadNode->getSrcValue()) ||
1085 dyn_cast<Constant>(LoadNode->getSrcValue()) ||
1086 dyn_cast<ConstantSDNode>(Ptr)) {
1088 for (unsigned i = 0; i < 4; i++) {
1089 // We want Const position encoded with the following formula :
1090 // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1091 // const_index is Ptr computed by llvm using an alignment of 16.
1092 // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1093 // then div by 4 at the ISel step
1094 SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1095 DAG.getConstant(4 * i + ConstantBlock * 16, MVT::i32));
1096 Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1098 Result = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Slots, 4);
1100 // non constant ptr cant be folded, keeps it as a v4f32 load
1101 Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1102 DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(4, MVT::i32)),
1103 DAG.getConstant(LoadNode->getAddressSpace() -
1104 AMDGPUAS::CONSTANT_BUFFER_0, MVT::i32)
1108 if (!VT.isVector()) {
1109 Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1110 DAG.getConstant(0, MVT::i32));
1113 SDValue MergedValues[2] = {
1117 return DAG.getMergeValues(MergedValues, 2, DL);
1120 if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1124 // Lowering for indirect addressing
1125 const MachineFunction &MF = DAG.getMachineFunction();
1126 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1127 getTargetMachine().getFrameLowering());
1128 unsigned StackWidth = TFL->getStackWidth(MF);
1130 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1132 if (VT.isVector()) {
1133 unsigned NumElemVT = VT.getVectorNumElements();
1134 EVT ElemVT = VT.getVectorElementType();
1137 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1138 "vector width in load");
1140 for (unsigned i = 0; i < NumElemVT; ++i) {
1141 unsigned Channel, PtrIncr;
1142 getStackAddress(StackWidth, i, Channel, PtrIncr);
1143 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1144 DAG.getConstant(PtrIncr, MVT::i32));
1145 Loads[i] = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, ElemVT,
1147 DAG.getTargetConstant(Channel, MVT::i32),
1150 for (unsigned i = NumElemVT; i < 4; ++i) {
1151 Loads[i] = DAG.getUNDEF(ElemVT);
1153 EVT TargetVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, 4);
1154 LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads, 4);
1156 LoweredLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, VT,
1158 DAG.getTargetConstant(0, MVT::i32), // Channel
1163 Ops[0] = LoweredLoad;
1166 return DAG.getMergeValues(Ops, 2, DL);
1169 /// XXX Only kernel functions are supported, so we can assume for now that
1170 /// every function is a kernel function, but in the future we should use
1171 /// separate calling conventions for kernel and non-kernel functions.
1172 SDValue R600TargetLowering::LowerFormalArguments(
1174 CallingConv::ID CallConv,
1176 const SmallVectorImpl<ISD::InputArg> &Ins,
1177 SDLoc DL, SelectionDAG &DAG,
1178 SmallVectorImpl<SDValue> &InVals) const {
1179 unsigned ParamOffsetBytes = 36;
1180 Function::const_arg_iterator FuncArg =
1181 DAG.getMachineFunction().getFunction()->arg_begin();
1182 for (unsigned i = 0, e = Ins.size(); i < e; ++i, ++FuncArg) {
1184 Type *ArgType = FuncArg->getType();
1185 unsigned ArgSizeInBits = ArgType->isPointerTy() ?
1186 32 : ArgType->getPrimitiveSizeInBits();
1187 unsigned ArgBytes = ArgSizeInBits >> 3;
1189 if (ArgSizeInBits < VT.getSizeInBits()) {
1190 assert(!ArgType->isFloatTy() &&
1191 "Extending floating point arguments not supported yet");
1192 ArgVT = MVT::getIntegerVT(ArgSizeInBits);
1196 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
1197 AMDGPUAS::PARAM_I_ADDRESS);
1198 SDValue Arg = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getRoot(),
1199 DAG.getConstant(ParamOffsetBytes, MVT::i32),
1200 MachinePointerInfo(UndefValue::get(PtrTy)),
1201 ArgVT, false, false, ArgBytes);
1202 InVals.push_back(Arg);
1203 ParamOffsetBytes += ArgBytes;
1208 EVT R600TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1209 if (!VT.isVector()) return MVT::i32;
1210 return VT.changeVectorElementTypeToInteger();
1214 CompactSwizzlableVector(SelectionDAG &DAG, SDValue VectorEntry,
1215 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1216 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1217 assert(RemapSwizzle.empty());
1218 SDValue NewBldVec[4] = {
1219 VectorEntry.getOperand(0),
1220 VectorEntry.getOperand(1),
1221 VectorEntry.getOperand(2),
1222 VectorEntry.getOperand(3)
1225 for (unsigned i = 0; i < 4; i++) {
1226 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(NewBldVec[i])) {
1228 RemapSwizzle[i] = 4; // SEL_0
1229 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1230 } else if (C->isExactlyValue(1.0)) {
1231 RemapSwizzle[i] = 5; // SEL_1
1232 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1236 if (NewBldVec[i].getOpcode() == ISD::UNDEF)
1238 for (unsigned j = 0; j < i; j++) {
1239 if (NewBldVec[i] == NewBldVec[j]) {
1240 NewBldVec[i] = DAG.getUNDEF(NewBldVec[i].getValueType());
1241 RemapSwizzle[i] = j;
1247 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1248 VectorEntry.getValueType(), NewBldVec, 4);
1251 static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
1252 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1253 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1254 assert(RemapSwizzle.empty());
1255 SDValue NewBldVec[4] = {
1256 VectorEntry.getOperand(0),
1257 VectorEntry.getOperand(1),
1258 VectorEntry.getOperand(2),
1259 VectorEntry.getOperand(3)
1261 bool isUnmovable[4] = { false, false, false, false };
1263 for (unsigned i = 0; i < 4; i++) {
1264 if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1265 unsigned Idx = dyn_cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1267 if (!isUnmovable[Idx]) {
1269 std::swap(NewBldVec[Idx], NewBldVec[i]);
1270 RemapSwizzle[Idx] = i;
1271 RemapSwizzle[i] = Idx;
1273 isUnmovable[Idx] = true;
1277 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1278 VectorEntry.getValueType(), NewBldVec, 4);
1282 SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector,
1283 SDValue Swz[4], SelectionDAG &DAG) const {
1284 assert(BuildVector.getOpcode() == ISD::BUILD_VECTOR);
1285 // Old -> New swizzle values
1286 DenseMap<unsigned, unsigned> SwizzleRemap;
1288 BuildVector = CompactSwizzlableVector(DAG, BuildVector, SwizzleRemap);
1289 for (unsigned i = 0; i < 4; i++) {
1290 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1291 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1292 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1295 SwizzleRemap.clear();
1296 BuildVector = ReorganizeVector(DAG, BuildVector, SwizzleRemap);
1297 for (unsigned i = 0; i < 4; i++) {
1298 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1299 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1300 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1307 //===----------------------------------------------------------------------===//
1308 // Custom DAG Optimizations
1309 //===----------------------------------------------------------------------===//
1311 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1312 DAGCombinerInfo &DCI) const {
1313 SelectionDAG &DAG = DCI.DAG;
1315 switch (N->getOpcode()) {
1316 // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1317 case ISD::FP_ROUND: {
1318 SDValue Arg = N->getOperand(0);
1319 if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1320 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), N->getValueType(0),
1326 // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1327 // (i32 select_cc f32, f32, -1, 0 cc)
1329 // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1330 // this to one of the SET*_DX10 instructions.
1331 case ISD::FP_TO_SINT: {
1332 SDValue FNeg = N->getOperand(0);
1333 if (FNeg.getOpcode() != ISD::FNEG) {
1336 SDValue SelectCC = FNeg.getOperand(0);
1337 if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1338 SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1339 SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1340 !isHWTrueValue(SelectCC.getOperand(2)) ||
1341 !isHWFalseValue(SelectCC.getOperand(3))) {
1345 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N->getValueType(0),
1346 SelectCC.getOperand(0), // LHS
1347 SelectCC.getOperand(1), // RHS
1348 DAG.getConstant(-1, MVT::i32), // True
1349 DAG.getConstant(0, MVT::i32), // Flase
1350 SelectCC.getOperand(4)); // CC
1354 // Extract_vec (Build_vector) generated by custom lowering
1355 // also needs to be customly combined
1356 case ISD::EXTRACT_VECTOR_ELT: {
1357 SDValue Arg = N->getOperand(0);
1358 if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1359 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1360 unsigned Element = Const->getZExtValue();
1361 return Arg->getOperand(Element);
1364 if (Arg.getOpcode() == ISD::BITCAST &&
1365 Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
1366 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1367 unsigned Element = Const->getZExtValue();
1368 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getVTList(),
1369 Arg->getOperand(0).getOperand(Element));
1374 case ISD::SELECT_CC: {
1375 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1376 // selectcc x, y, a, b, inv(cc)
1378 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1379 // selectcc x, y, a, b, cc
1380 SDValue LHS = N->getOperand(0);
1381 if (LHS.getOpcode() != ISD::SELECT_CC) {
1385 SDValue RHS = N->getOperand(1);
1386 SDValue True = N->getOperand(2);
1387 SDValue False = N->getOperand(3);
1388 ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1390 if (LHS.getOperand(2).getNode() != True.getNode() ||
1391 LHS.getOperand(3).getNode() != False.getNode() ||
1392 RHS.getNode() != False.getNode()) {
1397 default: return SDValue();
1398 case ISD::SETNE: return LHS;
1400 ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1401 LHSCC = ISD::getSetCCInverse(LHSCC,
1402 LHS.getOperand(0).getValueType().isInteger());
1403 return DAG.getSelectCC(SDLoc(N),
1412 case AMDGPUISD::EXPORT: {
1413 SDValue Arg = N->getOperand(1);
1414 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1417 SDValue NewArgs[8] = {
1418 N->getOperand(0), // Chain
1420 N->getOperand(2), // ArrayBase
1421 N->getOperand(3), // Type
1422 N->getOperand(4), // SWZ_X
1423 N->getOperand(5), // SWZ_Y
1424 N->getOperand(6), // SWZ_Z
1425 N->getOperand(7) // SWZ_W
1428 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG);
1429 return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs, 8);
1431 case AMDGPUISD::TEXTURE_FETCH: {
1432 SDValue Arg = N->getOperand(1);
1433 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1436 SDValue NewArgs[19] = {
1457 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG);
1458 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, SDLoc(N), N->getVTList(),