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 setOperationAction(ISD::FADD, MVT::v4f32, Expand);
42 setOperationAction(ISD::FADD, MVT::v2f32, Expand);
43 setOperationAction(ISD::FMUL, MVT::v4f32, Expand);
44 setOperationAction(ISD::FMUL, MVT::v2f32, Expand);
45 setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
46 setOperationAction(ISD::FDIV, MVT::v2f32, Expand);
47 setOperationAction(ISD::FSUB, MVT::v4f32, Expand);
48 setOperationAction(ISD::FSUB, MVT::v2f32, Expand);
50 setOperationAction(ISD::FCOS, MVT::f32, Custom);
51 setOperationAction(ISD::FSIN, MVT::f32, Custom);
53 setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
54 setOperationAction(ISD::SETCC, MVT::v2i32, Expand);
56 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
57 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
59 setOperationAction(ISD::FSUB, MVT::f32, Expand);
61 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
62 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
63 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
65 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
66 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
68 setOperationAction(ISD::SETCC, MVT::i32, Expand);
69 setOperationAction(ISD::SETCC, MVT::f32, Expand);
70 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
72 setOperationAction(ISD::SELECT, MVT::i32, Custom);
73 setOperationAction(ISD::SELECT, MVT::f32, Custom);
75 // Legalize loads and stores to the private address space.
76 setOperationAction(ISD::LOAD, MVT::i32, Custom);
77 setOperationAction(ISD::LOAD, MVT::v2i32, Custom);
78 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
79 setLoadExtAction(ISD::SEXTLOAD, MVT::i8, Custom);
80 setLoadExtAction(ISD::SEXTLOAD, MVT::i16, Custom);
81 setLoadExtAction(ISD::ZEXTLOAD, MVT::i8, Custom);
82 setLoadExtAction(ISD::ZEXTLOAD, MVT::i16, Custom);
83 setOperationAction(ISD::STORE, MVT::i8, Custom);
84 setOperationAction(ISD::STORE, MVT::i32, Custom);
85 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
86 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
87 setTruncStoreAction(MVT::i32, MVT::i8, Custom);
88 setTruncStoreAction(MVT::i32, MVT::i16, Custom);
90 setOperationAction(ISD::LOAD, MVT::i32, Custom);
91 setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
92 setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
94 setTargetDAGCombine(ISD::FP_ROUND);
95 setTargetDAGCombine(ISD::FP_TO_SINT);
96 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
97 setTargetDAGCombine(ISD::SELECT_CC);
98 setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
100 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
102 setBooleanContents(ZeroOrNegativeOneBooleanContent);
103 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
104 setSchedulingPreference(Sched::Source);
107 MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
108 MachineInstr * MI, MachineBasicBlock * BB) const {
109 MachineFunction * MF = BB->getParent();
110 MachineRegisterInfo &MRI = MF->getRegInfo();
111 MachineBasicBlock::iterator I = *MI;
112 const R600InstrInfo *TII =
113 static_cast<const R600InstrInfo*>(MF->getTarget().getInstrInfo());
115 switch (MI->getOpcode()) {
116 default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
117 case AMDGPU::CLAMP_R600: {
118 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
120 MI->getOperand(0).getReg(),
121 MI->getOperand(1).getReg());
122 TII->addFlag(NewMI, 0, MO_FLAG_CLAMP);
126 case AMDGPU::FABS_R600: {
127 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
129 MI->getOperand(0).getReg(),
130 MI->getOperand(1).getReg());
131 TII->addFlag(NewMI, 0, MO_FLAG_ABS);
135 case AMDGPU::FNEG_R600: {
136 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
138 MI->getOperand(0).getReg(),
139 MI->getOperand(1).getReg());
140 TII->addFlag(NewMI, 0, MO_FLAG_NEG);
144 case AMDGPU::MASK_WRITE: {
145 unsigned maskedRegister = MI->getOperand(0).getReg();
146 assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
147 MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
148 TII->addFlag(defInstr, 0, MO_FLAG_MASK);
152 case AMDGPU::LDS_READ_RET: {
153 MachineInstrBuilder NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
154 TII->get(MI->getOpcode()),
156 for (unsigned i = 1, e = MI->getNumOperands(); i < e; ++i) {
157 NewMI.addOperand(MI->getOperand(i));
159 TII->buildDefaultInstruction(*BB, I, AMDGPU::MOV,
160 MI->getOperand(0).getReg(),
165 case AMDGPU::MOV_IMM_F32:
166 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
167 MI->getOperand(1).getFPImm()->getValueAPF()
168 .bitcastToAPInt().getZExtValue());
170 case AMDGPU::MOV_IMM_I32:
171 TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
172 MI->getOperand(1).getImm());
174 case AMDGPU::CONST_COPY: {
175 MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, MI, AMDGPU::MOV,
176 MI->getOperand(0).getReg(), AMDGPU::ALU_CONST);
177 TII->setImmOperand(NewMI, AMDGPU::OpName::src0_sel,
178 MI->getOperand(1).getImm());
182 case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
183 case AMDGPU::RAT_WRITE_CACHELESS_64_eg:
184 case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
185 unsigned EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN) ? 1 : 0;
187 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
188 .addOperand(MI->getOperand(0))
189 .addOperand(MI->getOperand(1))
190 .addImm(EOP); // Set End of program bit
195 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
196 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
197 MachineOperand &RID = MI->getOperand(4);
198 MachineOperand &SID = MI->getOperand(5);
199 unsigned TextureId = MI->getOperand(6).getImm();
200 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
201 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
213 case 8: // ShadowRect
224 case 11: // Shadow1DArray
228 case 12: // Shadow2DArray
232 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
233 .addOperand(MI->getOperand(3))
251 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
252 .addOperand(MI->getOperand(2))
270 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
271 .addOperand(MI->getOperand(0))
272 .addOperand(MI->getOperand(1))
290 .addReg(T0, RegState::Implicit)
291 .addReg(T1, RegState::Implicit);
295 case AMDGPU::TXD_SHADOW: {
296 unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
297 unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
298 MachineOperand &RID = MI->getOperand(4);
299 MachineOperand &SID = MI->getOperand(5);
300 unsigned TextureId = MI->getOperand(6).getImm();
301 unsigned SrcX = 0, SrcY = 1, SrcZ = 2, SrcW = 3;
302 unsigned CTX = 1, CTY = 1, CTZ = 1, CTW = 1;
314 case 8: // ShadowRect
325 case 11: // Shadow1DArray
329 case 12: // Shadow2DArray
334 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
335 .addOperand(MI->getOperand(3))
353 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
354 .addOperand(MI->getOperand(2))
372 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
373 .addOperand(MI->getOperand(0))
374 .addOperand(MI->getOperand(1))
392 .addReg(T0, RegState::Implicit)
393 .addReg(T1, RegState::Implicit);
398 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
399 .addOperand(MI->getOperand(0));
402 case AMDGPU::BRANCH_COND_f32: {
403 MachineInstr *NewMI =
404 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
405 AMDGPU::PREDICATE_BIT)
406 .addOperand(MI->getOperand(1))
407 .addImm(OPCODE_IS_NOT_ZERO)
409 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
410 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
411 .addOperand(MI->getOperand(0))
412 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
416 case AMDGPU::BRANCH_COND_i32: {
417 MachineInstr *NewMI =
418 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
419 AMDGPU::PREDICATE_BIT)
420 .addOperand(MI->getOperand(1))
421 .addImm(OPCODE_IS_NOT_ZERO_INT)
423 TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
424 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP_COND))
425 .addOperand(MI->getOperand(0))
426 .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
430 case AMDGPU::EG_ExportSwz:
431 case AMDGPU::R600_ExportSwz: {
432 // Instruction is left unmodified if its not the last one of its type
433 bool isLastInstructionOfItsType = true;
434 unsigned InstExportType = MI->getOperand(1).getImm();
435 for (MachineBasicBlock::iterator NextExportInst = llvm::next(I),
436 EndBlock = BB->end(); NextExportInst != EndBlock;
437 NextExportInst = llvm::next(NextExportInst)) {
438 if (NextExportInst->getOpcode() == AMDGPU::EG_ExportSwz ||
439 NextExportInst->getOpcode() == AMDGPU::R600_ExportSwz) {
440 unsigned CurrentInstExportType = NextExportInst->getOperand(1)
442 if (CurrentInstExportType == InstExportType) {
443 isLastInstructionOfItsType = false;
448 bool EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN)? 1 : 0;
449 if (!EOP && !isLastInstructionOfItsType)
451 unsigned CfInst = (MI->getOpcode() == AMDGPU::EG_ExportSwz)? 84 : 40;
452 BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
453 .addOperand(MI->getOperand(0))
454 .addOperand(MI->getOperand(1))
455 .addOperand(MI->getOperand(2))
456 .addOperand(MI->getOperand(3))
457 .addOperand(MI->getOperand(4))
458 .addOperand(MI->getOperand(5))
459 .addOperand(MI->getOperand(6))
464 case AMDGPU::RETURN: {
465 // RETURN instructions must have the live-out registers as implicit uses,
466 // otherwise they appear dead.
467 R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
468 MachineInstrBuilder MIB(*MF, MI);
469 for (unsigned i = 0, e = MFI->LiveOuts.size(); i != e; ++i)
470 MIB.addReg(MFI->LiveOuts[i], RegState::Implicit);
475 MI->eraseFromParent();
479 //===----------------------------------------------------------------------===//
480 // Custom DAG Lowering Operations
481 //===----------------------------------------------------------------------===//
483 SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
484 MachineFunction &MF = DAG.getMachineFunction();
485 R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
486 switch (Op.getOpcode()) {
487 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
489 case ISD::FSIN: return LowerTrig(Op, DAG);
490 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
491 case ISD::SELECT: return LowerSELECT(Op, DAG);
492 case ISD::STORE: return LowerSTORE(Op, DAG);
493 case ISD::LOAD: return LowerLOAD(Op, DAG);
494 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
495 case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
496 case ISD::INTRINSIC_VOID: {
497 SDValue Chain = Op.getOperand(0);
498 unsigned IntrinsicID =
499 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
500 switch (IntrinsicID) {
501 case AMDGPUIntrinsic::AMDGPU_store_output: {
502 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
503 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
504 MFI->LiveOuts.push_back(Reg);
505 return DAG.getCopyToReg(Chain, SDLoc(Op), Reg, Op.getOperand(2));
507 case AMDGPUIntrinsic::R600_store_swizzle: {
508 const SDValue Args[8] = {
510 Op.getOperand(2), // Export Value
511 Op.getOperand(3), // ArrayBase
512 Op.getOperand(4), // Type
513 DAG.getConstant(0, MVT::i32), // SWZ_X
514 DAG.getConstant(1, MVT::i32), // SWZ_Y
515 DAG.getConstant(2, MVT::i32), // SWZ_Z
516 DAG.getConstant(3, MVT::i32) // SWZ_W
518 return DAG.getNode(AMDGPUISD::EXPORT, SDLoc(Op), Op.getValueType(),
522 // default for switch(IntrinsicID)
525 // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
528 case ISD::INTRINSIC_WO_CHAIN: {
529 unsigned IntrinsicID =
530 cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
531 EVT VT = Op.getValueType();
533 switch(IntrinsicID) {
534 default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
535 case AMDGPUIntrinsic::R600_load_input: {
536 int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
537 unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
538 MachineFunction &MF = DAG.getMachineFunction();
539 MachineRegisterInfo &MRI = MF.getRegInfo();
541 return DAG.getCopyFromReg(DAG.getEntryNode(),
542 SDLoc(DAG.getEntryNode()), Reg, VT);
545 case AMDGPUIntrinsic::R600_interp_input: {
546 int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
547 int ijb = cast<ConstantSDNode>(Op.getOperand(2))->getSExtValue();
548 MachineSDNode *interp;
550 const MachineFunction &MF = DAG.getMachineFunction();
551 const R600InstrInfo *TII =
552 static_cast<const R600InstrInfo*>(MF.getTarget().getInstrInfo());
553 interp = DAG.getMachineNode(AMDGPU::INTERP_VEC_LOAD, DL,
554 MVT::v4f32, DAG.getTargetConstant(slot / 4 , MVT::i32));
555 return DAG.getTargetExtractSubreg(
556 TII->getRegisterInfo().getSubRegFromChannel(slot % 4),
557 DL, MVT::f32, SDValue(interp, 0));
560 MachineFunction &MF = DAG.getMachineFunction();
561 MachineRegisterInfo &MRI = MF.getRegInfo();
562 unsigned RegisterI = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb);
563 unsigned RegisterJ = AMDGPU::R600_TReg32RegClass.getRegister(2 * ijb + 1);
564 MRI.addLiveIn(RegisterI);
565 MRI.addLiveIn(RegisterJ);
566 SDValue RegisterINode = DAG.getCopyFromReg(DAG.getEntryNode(),
567 SDLoc(DAG.getEntryNode()), RegisterI, MVT::f32);
568 SDValue RegisterJNode = DAG.getCopyFromReg(DAG.getEntryNode(),
569 SDLoc(DAG.getEntryNode()), RegisterJ, MVT::f32);
572 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_XY, DL,
573 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
574 RegisterJNode, RegisterINode);
576 interp = DAG.getMachineNode(AMDGPU::INTERP_PAIR_ZW, DL,
577 MVT::f32, MVT::f32, DAG.getTargetConstant(slot / 4 , MVT::i32),
578 RegisterJNode, RegisterINode);
579 return SDValue(interp, slot % 2);
581 case AMDGPUIntrinsic::R600_tex:
582 case AMDGPUIntrinsic::R600_texc:
583 case AMDGPUIntrinsic::R600_txl:
584 case AMDGPUIntrinsic::R600_txlc:
585 case AMDGPUIntrinsic::R600_txb:
586 case AMDGPUIntrinsic::R600_txbc:
587 case AMDGPUIntrinsic::R600_txf:
588 case AMDGPUIntrinsic::R600_txq:
589 case AMDGPUIntrinsic::R600_ddx:
590 case AMDGPUIntrinsic::R600_ddy: {
592 switch (IntrinsicID) {
593 case AMDGPUIntrinsic::R600_tex:
596 case AMDGPUIntrinsic::R600_texc:
599 case AMDGPUIntrinsic::R600_txl:
602 case AMDGPUIntrinsic::R600_txlc:
605 case AMDGPUIntrinsic::R600_txb:
608 case AMDGPUIntrinsic::R600_txbc:
611 case AMDGPUIntrinsic::R600_txf:
614 case AMDGPUIntrinsic::R600_txq:
617 case AMDGPUIntrinsic::R600_ddx:
620 case AMDGPUIntrinsic::R600_ddy:
624 llvm_unreachable("Unknow Texture Operation");
627 SDValue TexArgs[19] = {
628 DAG.getConstant(TextureOp, MVT::i32),
630 DAG.getConstant(0, MVT::i32),
631 DAG.getConstant(1, MVT::i32),
632 DAG.getConstant(2, MVT::i32),
633 DAG.getConstant(3, MVT::i32),
637 DAG.getConstant(0, MVT::i32),
638 DAG.getConstant(1, MVT::i32),
639 DAG.getConstant(2, MVT::i32),
640 DAG.getConstant(3, MVT::i32),
648 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs, 19);
650 case AMDGPUIntrinsic::AMDGPU_dp4: {
652 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
653 DAG.getConstant(0, MVT::i32)),
654 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
655 DAG.getConstant(0, MVT::i32)),
656 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
657 DAG.getConstant(1, MVT::i32)),
658 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
659 DAG.getConstant(1, MVT::i32)),
660 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
661 DAG.getConstant(2, MVT::i32)),
662 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
663 DAG.getConstant(2, MVT::i32)),
664 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(1),
665 DAG.getConstant(3, MVT::i32)),
666 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
667 DAG.getConstant(3, MVT::i32))
669 return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args, 8);
672 case Intrinsic::r600_read_ngroups_x:
673 return LowerImplicitParameter(DAG, VT, DL, 0);
674 case Intrinsic::r600_read_ngroups_y:
675 return LowerImplicitParameter(DAG, VT, DL, 1);
676 case Intrinsic::r600_read_ngroups_z:
677 return LowerImplicitParameter(DAG, VT, DL, 2);
678 case Intrinsic::r600_read_global_size_x:
679 return LowerImplicitParameter(DAG, VT, DL, 3);
680 case Intrinsic::r600_read_global_size_y:
681 return LowerImplicitParameter(DAG, VT, DL, 4);
682 case Intrinsic::r600_read_global_size_z:
683 return LowerImplicitParameter(DAG, VT, DL, 5);
684 case Intrinsic::r600_read_local_size_x:
685 return LowerImplicitParameter(DAG, VT, DL, 6);
686 case Intrinsic::r600_read_local_size_y:
687 return LowerImplicitParameter(DAG, VT, DL, 7);
688 case Intrinsic::r600_read_local_size_z:
689 return LowerImplicitParameter(DAG, VT, DL, 8);
691 case Intrinsic::r600_read_tgid_x:
692 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
694 case Intrinsic::r600_read_tgid_y:
695 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
697 case Intrinsic::r600_read_tgid_z:
698 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
700 case Intrinsic::r600_read_tidig_x:
701 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
703 case Intrinsic::r600_read_tidig_y:
704 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
706 case Intrinsic::r600_read_tidig_z:
707 return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
710 // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
713 } // end switch(Op.getOpcode())
717 void R600TargetLowering::ReplaceNodeResults(SDNode *N,
718 SmallVectorImpl<SDValue> &Results,
719 SelectionDAG &DAG) const {
720 switch (N->getOpcode()) {
722 case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
725 SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
726 Results.push_back(SDValue(Node, 0));
727 Results.push_back(SDValue(Node, 1));
728 // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
730 DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
734 SDNode *Node = LowerSTORE(SDValue(N, 0), DAG).getNode();
735 Results.push_back(SDValue(Node, 0));
740 SDValue R600TargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
741 // On hw >= R700, COS/SIN input must be between -1. and 1.
742 // Thus we lower them to TRIG ( FRACT ( x / 2Pi + 0.5) - 0.5)
743 EVT VT = Op.getValueType();
744 SDValue Arg = Op.getOperand(0);
745 SDValue FractPart = DAG.getNode(AMDGPUISD::FRACT, SDLoc(Op), VT,
746 DAG.getNode(ISD::FADD, SDLoc(Op), VT,
747 DAG.getNode(ISD::FMUL, SDLoc(Op), VT, Arg,
748 DAG.getConstantFP(0.15915494309, MVT::f32)),
749 DAG.getConstantFP(0.5, MVT::f32)));
751 switch (Op.getOpcode()) {
753 TrigNode = AMDGPUISD::COS_HW;
756 TrigNode = AMDGPUISD::SIN_HW;
759 llvm_unreachable("Wrong trig opcode");
761 SDValue TrigVal = DAG.getNode(TrigNode, SDLoc(Op), VT,
762 DAG.getNode(ISD::FADD, SDLoc(Op), VT, FractPart,
763 DAG.getConstantFP(-0.5, MVT::f32)));
764 if (Gen >= AMDGPUSubtarget::R700)
766 // On R600 hw, COS/SIN input must be between -Pi and Pi.
767 return DAG.getNode(ISD::FMUL, SDLoc(Op), VT, TrigVal,
768 DAG.getConstantFP(3.14159265359, MVT::f32));
771 SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
776 Op, DAG.getConstantFP(0.0f, MVT::f32),
777 DAG.getCondCode(ISD::SETNE)
781 SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
783 unsigned DwordOffset) const {
784 unsigned ByteOffset = DwordOffset * 4;
785 PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
786 AMDGPUAS::CONSTANT_BUFFER_0);
788 // We shouldn't be using an offset wider than 16-bits for implicit parameters.
789 assert(isInt<16>(ByteOffset));
791 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
792 DAG.getConstant(ByteOffset, MVT::i32), // PTR
793 MachinePointerInfo(ConstantPointerNull::get(PtrType)),
794 false, false, false, 0);
797 SDValue R600TargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
799 MachineFunction &MF = DAG.getMachineFunction();
800 const AMDGPUFrameLowering *TFL =
801 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
803 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
806 unsigned FrameIndex = FIN->getIndex();
807 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
808 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF), MVT::i32);
811 bool R600TargetLowering::isZero(SDValue Op) const {
812 if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
813 return Cst->isNullValue();
814 } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
815 return CstFP->isZero();
821 SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
823 EVT VT = Op.getValueType();
825 SDValue LHS = Op.getOperand(0);
826 SDValue RHS = Op.getOperand(1);
827 SDValue True = Op.getOperand(2);
828 SDValue False = Op.getOperand(3);
829 SDValue CC = Op.getOperand(4);
832 // LHS and RHS are guaranteed to be the same value type
833 EVT CompareVT = LHS.getValueType();
835 // Check if we can lower this to a native operation.
837 // Try to lower to a SET* instruction:
839 // SET* can match the following patterns:
841 // select_cc f32, f32, -1, 0, cc_any
842 // select_cc f32, f32, 1.0f, 0.0f, cc_any
843 // select_cc i32, i32, -1, 0, cc_any
846 // Move hardware True/False values to the correct operand.
847 if (isHWTrueValue(False) && isHWFalseValue(True)) {
848 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
849 std::swap(False, True);
850 CC = DAG.getCondCode(ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32));
853 if (isHWTrueValue(True) && isHWFalseValue(False) &&
854 (CompareVT == VT || VT == MVT::i32)) {
855 // This can be matched by a SET* instruction.
856 return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
859 // Try to lower to a CND* instruction:
861 // CND* can match the following patterns:
863 // select_cc f32, 0.0, f32, f32, cc_any
864 // select_cc f32, 0.0, i32, i32, cc_any
865 // select_cc i32, 0, f32, f32, cc_any
866 // select_cc i32, 0, i32, i32, cc_any
868 if (isZero(LHS) || isZero(RHS)) {
869 SDValue Cond = (isZero(LHS) ? RHS : LHS);
870 SDValue Zero = (isZero(LHS) ? LHS : RHS);
871 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
872 if (CompareVT != VT) {
873 // Bitcast True / False to the correct types. This will end up being
874 // a nop, but it allows us to define only a single pattern in the
875 // .TD files for each CND* instruction rather than having to have
876 // one pattern for integer True/False and one for fp True/False
877 True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
878 False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
881 CCOpcode = ISD::getSetCCSwappedOperands(CCOpcode);
894 CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
902 SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
905 DAG.getCondCode(CCOpcode));
906 return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
910 // Possible Min/Max pattern
911 SDValue MinMax = LowerMinMax(Op, DAG);
912 if (MinMax.getNode()) {
916 // If we make it this for it means we have no native instructions to handle
917 // this SELECT_CC, so we must lower it.
918 SDValue HWTrue, HWFalse;
920 if (CompareVT == MVT::f32) {
921 HWTrue = DAG.getConstantFP(1.0f, CompareVT);
922 HWFalse = DAG.getConstantFP(0.0f, CompareVT);
923 } else if (CompareVT == MVT::i32) {
924 HWTrue = DAG.getConstant(-1, CompareVT);
925 HWFalse = DAG.getConstant(0, CompareVT);
928 assert(!"Unhandled value type in LowerSELECT_CC");
931 // Lower this unsupported SELECT_CC into a combination of two supported
932 // SELECT_CC operations.
933 SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
935 return DAG.getNode(ISD::SELECT_CC, DL, VT,
938 DAG.getCondCode(ISD::SETNE));
941 SDValue R600TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
942 return DAG.getNode(ISD::SELECT_CC,
946 DAG.getConstant(0, MVT::i32),
949 DAG.getCondCode(ISD::SETNE));
952 /// LLVM generates byte-addresed pointers. For indirect addressing, we need to
953 /// convert these pointers to a register index. Each register holds
954 /// 16 bytes, (4 x 32bit sub-register), but we need to take into account the
955 /// \p StackWidth, which tells us how many of the 4 sub-registrers will be used
956 /// for indirect addressing.
957 SDValue R600TargetLowering::stackPtrToRegIndex(SDValue Ptr,
959 SelectionDAG &DAG) const {
971 default: llvm_unreachable("Invalid stack width");
974 return DAG.getNode(ISD::SRL, SDLoc(Ptr), Ptr.getValueType(), Ptr,
975 DAG.getConstant(SRLPad, MVT::i32));
978 void R600TargetLowering::getStackAddress(unsigned StackWidth,
981 unsigned &PtrIncr) const {
982 switch (StackWidth) {
993 Channel = ElemIdx % 2;
1007 SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1009 StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
1010 SDValue Chain = Op.getOperand(0);
1011 SDValue Value = Op.getOperand(1);
1012 SDValue Ptr = Op.getOperand(2);
1014 if (StoreNode->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS) {
1015 if (StoreNode->isTruncatingStore()) {
1016 EVT VT = Value.getValueType();
1017 assert(VT == MVT::i32);
1018 EVT MemVT = StoreNode->getMemoryVT();
1019 SDValue MaskConstant;
1020 if (MemVT == MVT::i8) {
1021 MaskConstant = DAG.getConstant(0xFF, MVT::i32);
1023 assert(MemVT == MVT::i16);
1024 MaskConstant = DAG.getConstant(0xFFFF, MVT::i32);
1026 SDValue DWordAddr = DAG.getNode(ISD::SRL, DL, VT, Ptr,
1027 DAG.getConstant(2, MVT::i32));
1028 SDValue ByteIndex = DAG.getNode(ISD::AND, DL, Ptr.getValueType(), Ptr,
1029 DAG.getConstant(0x00000003, VT));
1030 SDValue TruncValue = DAG.getNode(ISD::AND, DL, VT, Value, MaskConstant);
1031 SDValue Shift = DAG.getNode(ISD::SHL, DL, VT, ByteIndex,
1032 DAG.getConstant(3, VT));
1033 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, VT, TruncValue, Shift);
1034 SDValue Mask = DAG.getNode(ISD::SHL, DL, VT, MaskConstant, Shift);
1035 // XXX: If we add a 64-bit ZW register class, then we could use a 2 x i32
1039 DAG.getConstant(0, MVT::i32),
1040 DAG.getConstant(0, MVT::i32),
1043 SDValue Input = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Src, 4);
1044 SDValue Args[3] = { Chain, Input, DWordAddr };
1045 return DAG.getMemIntrinsicNode(AMDGPUISD::STORE_MSKOR, DL,
1046 Op->getVTList(), Args, 3, MemVT,
1047 StoreNode->getMemOperand());
1048 } else if (Ptr->getOpcode() != AMDGPUISD::DWORDADDR &&
1049 Value.getValueType().bitsGE(MVT::i32)) {
1050 // Convert pointer from byte address to dword address.
1051 Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, Ptr.getValueType(),
1052 DAG.getNode(ISD::SRL, DL, Ptr.getValueType(),
1053 Ptr, DAG.getConstant(2, MVT::i32)));
1055 if (StoreNode->isTruncatingStore() || StoreNode->isIndexed()) {
1056 assert(!"Truncated and indexed stores not supported yet");
1058 Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
1064 EVT ValueVT = Value.getValueType();
1066 if (StoreNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1070 // Lowering for indirect addressing
1072 const MachineFunction &MF = DAG.getMachineFunction();
1073 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1074 getTargetMachine().getFrameLowering());
1075 unsigned StackWidth = TFL->getStackWidth(MF);
1077 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1079 if (ValueVT.isVector()) {
1080 unsigned NumElemVT = ValueVT.getVectorNumElements();
1081 EVT ElemVT = ValueVT.getVectorElementType();
1084 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1085 "vector width in load");
1087 for (unsigned i = 0; i < NumElemVT; ++i) {
1088 unsigned Channel, PtrIncr;
1089 getStackAddress(StackWidth, i, Channel, PtrIncr);
1090 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1091 DAG.getConstant(PtrIncr, MVT::i32));
1092 SDValue Elem = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT,
1093 Value, DAG.getConstant(i, MVT::i32));
1095 Stores[i] = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1097 DAG.getTargetConstant(Channel, MVT::i32));
1099 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores, NumElemVT);
1101 if (ValueVT == MVT::i8) {
1102 Value = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Value);
1104 Chain = DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other, Chain, Value, Ptr,
1105 DAG.getTargetConstant(0, MVT::i32)); // Channel
1111 // return (512 + (kc_bank << 12)
1113 ConstantAddressBlock(unsigned AddressSpace) {
1114 switch (AddressSpace) {
1115 case AMDGPUAS::CONSTANT_BUFFER_0:
1117 case AMDGPUAS::CONSTANT_BUFFER_1:
1119 case AMDGPUAS::CONSTANT_BUFFER_2:
1120 return 512 + 4096 * 2;
1121 case AMDGPUAS::CONSTANT_BUFFER_3:
1122 return 512 + 4096 * 3;
1123 case AMDGPUAS::CONSTANT_BUFFER_4:
1124 return 512 + 4096 * 4;
1125 case AMDGPUAS::CONSTANT_BUFFER_5:
1126 return 512 + 4096 * 5;
1127 case AMDGPUAS::CONSTANT_BUFFER_6:
1128 return 512 + 4096 * 6;
1129 case AMDGPUAS::CONSTANT_BUFFER_7:
1130 return 512 + 4096 * 7;
1131 case AMDGPUAS::CONSTANT_BUFFER_8:
1132 return 512 + 4096 * 8;
1133 case AMDGPUAS::CONSTANT_BUFFER_9:
1134 return 512 + 4096 * 9;
1135 case AMDGPUAS::CONSTANT_BUFFER_10:
1136 return 512 + 4096 * 10;
1137 case AMDGPUAS::CONSTANT_BUFFER_11:
1138 return 512 + 4096 * 11;
1139 case AMDGPUAS::CONSTANT_BUFFER_12:
1140 return 512 + 4096 * 12;
1141 case AMDGPUAS::CONSTANT_BUFFER_13:
1142 return 512 + 4096 * 13;
1143 case AMDGPUAS::CONSTANT_BUFFER_14:
1144 return 512 + 4096 * 14;
1145 case AMDGPUAS::CONSTANT_BUFFER_15:
1146 return 512 + 4096 * 15;
1152 SDValue R600TargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
1154 EVT VT = Op.getValueType();
1156 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1157 SDValue Chain = Op.getOperand(0);
1158 SDValue Ptr = Op.getOperand(1);
1159 SDValue LoweredLoad;
1161 int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
1162 if (ConstantBlock > -1) {
1164 if (dyn_cast<ConstantExpr>(LoadNode->getSrcValue()) ||
1165 dyn_cast<Constant>(LoadNode->getSrcValue()) ||
1166 dyn_cast<ConstantSDNode>(Ptr)) {
1168 for (unsigned i = 0; i < 4; i++) {
1169 // We want Const position encoded with the following formula :
1170 // (((512 + (kc_bank << 12) + const_index) << 2) + chan)
1171 // const_index is Ptr computed by llvm using an alignment of 16.
1172 // Thus we add (((512 + (kc_bank << 12)) + chan ) * 4 here and
1173 // then div by 4 at the ISel step
1174 SDValue NewPtr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
1175 DAG.getConstant(4 * i + ConstantBlock * 16, MVT::i32));
1176 Slots[i] = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::i32, NewPtr);
1178 EVT NewVT = MVT::v4i32;
1179 unsigned NumElements = 4;
1180 if (VT.isVector()) {
1182 NumElements = VT.getVectorNumElements();
1184 Result = DAG.getNode(ISD::BUILD_VECTOR, DL, NewVT, Slots, NumElements);
1186 // non constant ptr cant be folded, keeps it as a v4f32 load
1187 Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
1188 DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr, DAG.getConstant(4, MVT::i32)),
1189 DAG.getConstant(LoadNode->getAddressSpace() -
1190 AMDGPUAS::CONSTANT_BUFFER_0, MVT::i32)
1194 if (!VT.isVector()) {
1195 Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32, Result,
1196 DAG.getConstant(0, MVT::i32));
1199 SDValue MergedValues[2] = {
1203 return DAG.getMergeValues(MergedValues, 2, DL);
1206 // For most operations returning SDValue() will result int he node being
1207 // expanded by the DAG Legalizer. This is not the case for ISD::LOAD, so
1208 // we need to manually expand loads that may be legal in some address spaces
1209 // and illegal in others. SEXT loads from CONSTANT_BUFFER_0 are supported
1210 // for compute shaders, since the data is sign extended when it is uploaded
1211 // to the buffer. Howerver SEXT loads from other addresspaces are not
1212 // supported, so we need to expand them here.
1213 if (LoadNode->getExtensionType() == ISD::SEXTLOAD) {
1214 EVT MemVT = LoadNode->getMemoryVT();
1215 assert(!MemVT.isVector() && (MemVT == MVT::i16 || MemVT == MVT::i8));
1216 SDValue ShiftAmount =
1217 DAG.getConstant(VT.getSizeInBits() - MemVT.getSizeInBits(), MVT::i32);
1218 SDValue NewLoad = DAG.getExtLoad(ISD::EXTLOAD, DL, VT, Chain, Ptr,
1219 LoadNode->getPointerInfo(), MemVT,
1220 LoadNode->isVolatile(),
1221 LoadNode->isNonTemporal(),
1222 LoadNode->getAlignment());
1223 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, NewLoad, ShiftAmount);
1224 SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Shl, ShiftAmount);
1226 SDValue MergedValues[2] = { Sra, Chain };
1227 return DAG.getMergeValues(MergedValues, 2, DL);
1230 if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
1234 // Lowering for indirect addressing
1235 const MachineFunction &MF = DAG.getMachineFunction();
1236 const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
1237 getTargetMachine().getFrameLowering());
1238 unsigned StackWidth = TFL->getStackWidth(MF);
1240 Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
1242 if (VT.isVector()) {
1243 unsigned NumElemVT = VT.getVectorNumElements();
1244 EVT ElemVT = VT.getVectorElementType();
1247 assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
1248 "vector width in load");
1250 for (unsigned i = 0; i < NumElemVT; ++i) {
1251 unsigned Channel, PtrIncr;
1252 getStackAddress(StackWidth, i, Channel, PtrIncr);
1253 Ptr = DAG.getNode(ISD::ADD, DL, MVT::i32, Ptr,
1254 DAG.getConstant(PtrIncr, MVT::i32));
1255 Loads[i] = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, ElemVT,
1257 DAG.getTargetConstant(Channel, MVT::i32),
1260 for (unsigned i = NumElemVT; i < 4; ++i) {
1261 Loads[i] = DAG.getUNDEF(ElemVT);
1263 EVT TargetVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, 4);
1264 LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads, 4);
1266 LoweredLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, VT,
1268 DAG.getTargetConstant(0, MVT::i32), // Channel
1273 Ops[0] = LoweredLoad;
1276 return DAG.getMergeValues(Ops, 2, DL);
1279 /// XXX Only kernel functions are supported, so we can assume for now that
1280 /// every function is a kernel function, but in the future we should use
1281 /// separate calling conventions for kernel and non-kernel functions.
1282 SDValue R600TargetLowering::LowerFormalArguments(
1284 CallingConv::ID CallConv,
1286 const SmallVectorImpl<ISD::InputArg> &Ins,
1287 SDLoc DL, SelectionDAG &DAG,
1288 SmallVectorImpl<SDValue> &InVals) const {
1289 SmallVector<CCValAssign, 16> ArgLocs;
1290 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1291 getTargetMachine(), ArgLocs, *DAG.getContext());
1293 AnalyzeFormalArguments(CCInfo, Ins);
1295 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1296 CCValAssign &VA = ArgLocs[i];
1297 EVT VT = VA.getLocVT();
1299 PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
1300 AMDGPUAS::CONSTANT_BUFFER_0);
1302 // The first 36 bytes of the input buffer contains information about
1303 // thread group and global sizes.
1304 SDValue Arg = DAG.getLoad(VT, DL, Chain,
1305 DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
1306 MachinePointerInfo(UndefValue::get(PtrTy)), false,
1307 false, false, 4); // 4 is the prefered alignment for
1308 // the CONSTANT memory space.
1309 InVals.push_back(Arg);
1314 EVT R600TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1315 if (!VT.isVector()) return MVT::i32;
1316 return VT.changeVectorElementTypeToInteger();
1320 CompactSwizzlableVector(SelectionDAG &DAG, SDValue VectorEntry,
1321 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1322 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1323 assert(RemapSwizzle.empty());
1324 SDValue NewBldVec[4] = {
1325 VectorEntry.getOperand(0),
1326 VectorEntry.getOperand(1),
1327 VectorEntry.getOperand(2),
1328 VectorEntry.getOperand(3)
1331 for (unsigned i = 0; i < 4; i++) {
1332 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(NewBldVec[i])) {
1334 RemapSwizzle[i] = 4; // SEL_0
1335 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1336 } else if (C->isExactlyValue(1.0)) {
1337 RemapSwizzle[i] = 5; // SEL_1
1338 NewBldVec[i] = DAG.getUNDEF(MVT::f32);
1342 if (NewBldVec[i].getOpcode() == ISD::UNDEF)
1344 for (unsigned j = 0; j < i; j++) {
1345 if (NewBldVec[i] == NewBldVec[j]) {
1346 NewBldVec[i] = DAG.getUNDEF(NewBldVec[i].getValueType());
1347 RemapSwizzle[i] = j;
1353 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1354 VectorEntry.getValueType(), NewBldVec, 4);
1357 static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
1358 DenseMap<unsigned, unsigned> &RemapSwizzle) {
1359 assert(VectorEntry.getOpcode() == ISD::BUILD_VECTOR);
1360 assert(RemapSwizzle.empty());
1361 SDValue NewBldVec[4] = {
1362 VectorEntry.getOperand(0),
1363 VectorEntry.getOperand(1),
1364 VectorEntry.getOperand(2),
1365 VectorEntry.getOperand(3)
1367 bool isUnmovable[4] = { false, false, false, false };
1368 for (unsigned i = 0; i < 4; i++)
1369 RemapSwizzle[i] = i;
1371 for (unsigned i = 0; i < 4; i++) {
1372 if (NewBldVec[i].getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
1373 unsigned Idx = dyn_cast<ConstantSDNode>(NewBldVec[i].getOperand(1))
1375 if (!isUnmovable[Idx]) {
1377 std::swap(NewBldVec[Idx], NewBldVec[i]);
1378 std::swap(RemapSwizzle[RemapSwizzle[Idx]], RemapSwizzle[RemapSwizzle[i]]);
1380 isUnmovable[Idx] = true;
1384 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
1385 VectorEntry.getValueType(), NewBldVec, 4);
1389 SDValue R600TargetLowering::OptimizeSwizzle(SDValue BuildVector,
1390 SDValue Swz[4], SelectionDAG &DAG) const {
1391 assert(BuildVector.getOpcode() == ISD::BUILD_VECTOR);
1392 // Old -> New swizzle values
1393 DenseMap<unsigned, unsigned> SwizzleRemap;
1395 BuildVector = CompactSwizzlableVector(DAG, BuildVector, SwizzleRemap);
1396 for (unsigned i = 0; i < 4; i++) {
1397 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1398 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1399 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1402 SwizzleRemap.clear();
1403 BuildVector = ReorganizeVector(DAG, BuildVector, SwizzleRemap);
1404 for (unsigned i = 0; i < 4; i++) {
1405 unsigned Idx = dyn_cast<ConstantSDNode>(Swz[i])->getZExtValue();
1406 if (SwizzleRemap.find(Idx) != SwizzleRemap.end())
1407 Swz[i] = DAG.getConstant(SwizzleRemap[Idx], MVT::i32);
1414 //===----------------------------------------------------------------------===//
1415 // Custom DAG Optimizations
1416 //===----------------------------------------------------------------------===//
1418 SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
1419 DAGCombinerInfo &DCI) const {
1420 SelectionDAG &DAG = DCI.DAG;
1422 switch (N->getOpcode()) {
1423 // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
1424 case ISD::FP_ROUND: {
1425 SDValue Arg = N->getOperand(0);
1426 if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
1427 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), N->getValueType(0),
1433 // (i32 fp_to_sint (fneg (select_cc f32, f32, 1.0, 0.0 cc))) ->
1434 // (i32 select_cc f32, f32, -1, 0 cc)
1436 // Mesa's GLSL frontend generates the above pattern a lot and we can lower
1437 // this to one of the SET*_DX10 instructions.
1438 case ISD::FP_TO_SINT: {
1439 SDValue FNeg = N->getOperand(0);
1440 if (FNeg.getOpcode() != ISD::FNEG) {
1443 SDValue SelectCC = FNeg.getOperand(0);
1444 if (SelectCC.getOpcode() != ISD::SELECT_CC ||
1445 SelectCC.getOperand(0).getValueType() != MVT::f32 || // LHS
1446 SelectCC.getOperand(2).getValueType() != MVT::f32 || // True
1447 !isHWTrueValue(SelectCC.getOperand(2)) ||
1448 !isHWFalseValue(SelectCC.getOperand(3))) {
1452 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N->getValueType(0),
1453 SelectCC.getOperand(0), // LHS
1454 SelectCC.getOperand(1), // RHS
1455 DAG.getConstant(-1, MVT::i32), // True
1456 DAG.getConstant(0, MVT::i32), // Flase
1457 SelectCC.getOperand(4)); // CC
1462 // insert_vector_elt (build_vector elt0, …, eltN), NewEltIdx, idx
1463 // => build_vector elt0, …, NewEltIdx, …, eltN
1464 case ISD::INSERT_VECTOR_ELT: {
1465 SDValue InVec = N->getOperand(0);
1466 SDValue InVal = N->getOperand(1);
1467 SDValue EltNo = N->getOperand(2);
1470 // If the inserted element is an UNDEF, just use the input vector.
1471 if (InVal.getOpcode() == ISD::UNDEF)
1474 EVT VT = InVec.getValueType();
1476 // If we can't generate a legal BUILD_VECTOR, exit
1477 if (!isOperationLegal(ISD::BUILD_VECTOR, VT))
1480 // Check that we know which element is being inserted
1481 if (!isa<ConstantSDNode>(EltNo))
1483 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
1485 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
1486 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
1488 SmallVector<SDValue, 8> Ops;
1489 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
1490 Ops.append(InVec.getNode()->op_begin(),
1491 InVec.getNode()->op_end());
1492 } else if (InVec.getOpcode() == ISD::UNDEF) {
1493 unsigned NElts = VT.getVectorNumElements();
1494 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
1499 // Insert the element
1500 if (Elt < Ops.size()) {
1501 // All the operands of BUILD_VECTOR must have the same type;
1502 // we enforce that here.
1503 EVT OpVT = Ops[0].getValueType();
1504 if (InVal.getValueType() != OpVT)
1505 InVal = OpVT.bitsGT(InVal.getValueType()) ?
1506 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
1507 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
1511 // Return the new vector
1512 return DAG.getNode(ISD::BUILD_VECTOR, dl,
1513 VT, &Ops[0], Ops.size());
1516 // Extract_vec (Build_vector) generated by custom lowering
1517 // also needs to be customly combined
1518 case ISD::EXTRACT_VECTOR_ELT: {
1519 SDValue Arg = N->getOperand(0);
1520 if (Arg.getOpcode() == ISD::BUILD_VECTOR) {
1521 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1522 unsigned Element = Const->getZExtValue();
1523 return Arg->getOperand(Element);
1526 if (Arg.getOpcode() == ISD::BITCAST &&
1527 Arg.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
1528 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
1529 unsigned Element = Const->getZExtValue();
1530 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getVTList(),
1531 Arg->getOperand(0).getOperand(Element));
1536 case ISD::SELECT_CC: {
1537 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
1538 // selectcc x, y, a, b, inv(cc)
1540 // fold selectcc (selectcc x, y, a, b, cc), b, a, b, setne ->
1541 // selectcc x, y, a, b, cc
1542 SDValue LHS = N->getOperand(0);
1543 if (LHS.getOpcode() != ISD::SELECT_CC) {
1547 SDValue RHS = N->getOperand(1);
1548 SDValue True = N->getOperand(2);
1549 SDValue False = N->getOperand(3);
1550 ISD::CondCode NCC = cast<CondCodeSDNode>(N->getOperand(4))->get();
1552 if (LHS.getOperand(2).getNode() != True.getNode() ||
1553 LHS.getOperand(3).getNode() != False.getNode() ||
1554 RHS.getNode() != False.getNode()) {
1559 default: return SDValue();
1560 case ISD::SETNE: return LHS;
1562 ISD::CondCode LHSCC = cast<CondCodeSDNode>(LHS.getOperand(4))->get();
1563 LHSCC = ISD::getSetCCInverse(LHSCC,
1564 LHS.getOperand(0).getValueType().isInteger());
1565 return DAG.getSelectCC(SDLoc(N),
1574 case AMDGPUISD::EXPORT: {
1575 SDValue Arg = N->getOperand(1);
1576 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1579 SDValue NewArgs[8] = {
1580 N->getOperand(0), // Chain
1582 N->getOperand(2), // ArrayBase
1583 N->getOperand(3), // Type
1584 N->getOperand(4), // SWZ_X
1585 N->getOperand(5), // SWZ_Y
1586 N->getOperand(6), // SWZ_Z
1587 N->getOperand(7) // SWZ_W
1590 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG);
1591 return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs, 8);
1593 case AMDGPUISD::TEXTURE_FETCH: {
1594 SDValue Arg = N->getOperand(1);
1595 if (Arg.getOpcode() != ISD::BUILD_VECTOR)
1598 SDValue NewArgs[19] = {
1619 NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG);
1620 return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, SDLoc(N), N->getVTList(),