1 //===-- AMDGPUInstPrinter.cpp - AMDGPU MC Inst -> ASM ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
9 //===----------------------------------------------------------------------===//
11 #include "AMDGPUInstPrinter.h"
12 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
13 #include "SIDefines.h"
14 #include "Utils/AMDGPUAsmUtils.h"
15 #include "Utils/AMDGPUBaseInfo.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCInst.h"
18 #include "llvm/MC/MCInstrDesc.h"
19 #include "llvm/MC/MCInstrInfo.h"
20 #include "llvm/MC/MCRegisterInfo.h"
21 #include "llvm/MC/MCSubtargetInfo.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/MathExtras.h"
24 #include "llvm/Support/raw_ostream.h"
28 using namespace llvm::AMDGPU;
30 void AMDGPUInstPrinter::printInst(const MCInst *MI, raw_ostream &OS,
31 StringRef Annot, const MCSubtargetInfo &STI) {
33 printInstruction(MI, STI, OS);
34 printAnnotation(OS, Annot);
37 void AMDGPUInstPrinter::printU4ImmOperand(const MCInst *MI, unsigned OpNo,
38 const MCSubtargetInfo &STI,
40 O << formatHex(MI->getOperand(OpNo).getImm() & 0xf);
43 void AMDGPUInstPrinter::printU8ImmOperand(const MCInst *MI, unsigned OpNo,
45 O << formatHex(MI->getOperand(OpNo).getImm() & 0xff);
48 void AMDGPUInstPrinter::printU16ImmOperand(const MCInst *MI, unsigned OpNo,
49 const MCSubtargetInfo &STI,
51 // It's possible to end up with a 32-bit literal used with a 16-bit operand
52 // with ignored high bits. Print as 32-bit anyway in that case.
53 int64_t Imm = MI->getOperand(OpNo).getImm();
54 if (isInt<16>(Imm) || isUInt<16>(Imm))
55 O << formatHex(static_cast<uint64_t>(Imm & 0xffff));
57 printU32ImmOperand(MI, OpNo, STI, O);
60 void AMDGPUInstPrinter::printU4ImmDecOperand(const MCInst *MI, unsigned OpNo,
62 O << formatDec(MI->getOperand(OpNo).getImm() & 0xf);
65 void AMDGPUInstPrinter::printU8ImmDecOperand(const MCInst *MI, unsigned OpNo,
67 O << formatDec(MI->getOperand(OpNo).getImm() & 0xff);
70 void AMDGPUInstPrinter::printU16ImmDecOperand(const MCInst *MI, unsigned OpNo,
72 O << formatDec(MI->getOperand(OpNo).getImm() & 0xffff);
75 void AMDGPUInstPrinter::printS13ImmDecOperand(const MCInst *MI, unsigned OpNo,
77 O << formatDec(SignExtend32<13>(MI->getOperand(OpNo).getImm()));
80 void AMDGPUInstPrinter::printU32ImmOperand(const MCInst *MI, unsigned OpNo,
81 const MCSubtargetInfo &STI,
83 O << formatHex(MI->getOperand(OpNo).getImm() & 0xffffffff);
86 void AMDGPUInstPrinter::printNamedBit(const MCInst *MI, unsigned OpNo,
87 raw_ostream &O, StringRef BitName) {
88 if (MI->getOperand(OpNo).getImm()) {
93 void AMDGPUInstPrinter::printOffen(const MCInst *MI, unsigned OpNo,
95 printNamedBit(MI, OpNo, O, "offen");
98 void AMDGPUInstPrinter::printIdxen(const MCInst *MI, unsigned OpNo,
100 printNamedBit(MI, OpNo, O, "idxen");
103 void AMDGPUInstPrinter::printAddr64(const MCInst *MI, unsigned OpNo,
105 printNamedBit(MI, OpNo, O, "addr64");
108 void AMDGPUInstPrinter::printMBUFOffset(const MCInst *MI, unsigned OpNo,
110 if (MI->getOperand(OpNo).getImm()) {
112 printU16ImmDecOperand(MI, OpNo, O);
116 void AMDGPUInstPrinter::printOffset(const MCInst *MI, unsigned OpNo,
117 const MCSubtargetInfo &STI,
119 uint16_t Imm = MI->getOperand(OpNo).getImm();
121 O << ((OpNo == 0)? "offset:" : " offset:");
122 printU16ImmDecOperand(MI, OpNo, O);
126 void AMDGPUInstPrinter::printOffsetS13(const MCInst *MI, unsigned OpNo,
127 const MCSubtargetInfo &STI,
129 uint16_t Imm = MI->getOperand(OpNo).getImm();
131 O << ((OpNo == 0)? "offset:" : " offset:");
132 printS13ImmDecOperand(MI, OpNo, O);
136 void AMDGPUInstPrinter::printOffset0(const MCInst *MI, unsigned OpNo,
137 const MCSubtargetInfo &STI,
139 if (MI->getOperand(OpNo).getImm()) {
141 printU8ImmDecOperand(MI, OpNo, O);
145 void AMDGPUInstPrinter::printOffset1(const MCInst *MI, unsigned OpNo,
146 const MCSubtargetInfo &STI,
148 if (MI->getOperand(OpNo).getImm()) {
150 printU8ImmDecOperand(MI, OpNo, O);
154 void AMDGPUInstPrinter::printSMRDOffset8(const MCInst *MI, unsigned OpNo,
155 const MCSubtargetInfo &STI,
157 printU32ImmOperand(MI, OpNo, STI, O);
160 void AMDGPUInstPrinter::printSMRDOffset20(const MCInst *MI, unsigned OpNo,
161 const MCSubtargetInfo &STI,
163 printU32ImmOperand(MI, OpNo, STI, O);
166 void AMDGPUInstPrinter::printSMRDLiteralOffset(const MCInst *MI, unsigned OpNo,
167 const MCSubtargetInfo &STI,
169 printU32ImmOperand(MI, OpNo, STI, O);
172 void AMDGPUInstPrinter::printGDS(const MCInst *MI, unsigned OpNo,
173 const MCSubtargetInfo &STI, raw_ostream &O) {
174 printNamedBit(MI, OpNo, O, "gds");
177 void AMDGPUInstPrinter::printGLC(const MCInst *MI, unsigned OpNo,
178 const MCSubtargetInfo &STI, raw_ostream &O) {
179 printNamedBit(MI, OpNo, O, "glc");
182 void AMDGPUInstPrinter::printSLC(const MCInst *MI, unsigned OpNo,
183 const MCSubtargetInfo &STI, raw_ostream &O) {
184 printNamedBit(MI, OpNo, O, "slc");
187 void AMDGPUInstPrinter::printTFE(const MCInst *MI, unsigned OpNo,
188 const MCSubtargetInfo &STI, raw_ostream &O) {
189 printNamedBit(MI, OpNo, O, "tfe");
192 void AMDGPUInstPrinter::printDMask(const MCInst *MI, unsigned OpNo,
193 const MCSubtargetInfo &STI, raw_ostream &O) {
194 if (MI->getOperand(OpNo).getImm()) {
196 printU16ImmOperand(MI, OpNo, STI, O);
200 void AMDGPUInstPrinter::printUNorm(const MCInst *MI, unsigned OpNo,
201 const MCSubtargetInfo &STI, raw_ostream &O) {
202 printNamedBit(MI, OpNo, O, "unorm");
205 void AMDGPUInstPrinter::printDA(const MCInst *MI, unsigned OpNo,
206 const MCSubtargetInfo &STI, raw_ostream &O) {
207 printNamedBit(MI, OpNo, O, "da");
210 void AMDGPUInstPrinter::printR128(const MCInst *MI, unsigned OpNo,
211 const MCSubtargetInfo &STI, raw_ostream &O) {
212 printNamedBit(MI, OpNo, O, "r128");
215 void AMDGPUInstPrinter::printLWE(const MCInst *MI, unsigned OpNo,
216 const MCSubtargetInfo &STI, raw_ostream &O) {
217 printNamedBit(MI, OpNo, O, "lwe");
220 void AMDGPUInstPrinter::printExpCompr(const MCInst *MI, unsigned OpNo,
221 const MCSubtargetInfo &STI,
223 if (MI->getOperand(OpNo).getImm())
227 void AMDGPUInstPrinter::printExpVM(const MCInst *MI, unsigned OpNo,
228 const MCSubtargetInfo &STI,
230 if (MI->getOperand(OpNo).getImm())
234 void AMDGPUInstPrinter::printDFMT(const MCInst *MI, unsigned OpNo,
235 const MCSubtargetInfo &STI,
237 if (MI->getOperand(OpNo).getImm()) {
239 printU8ImmDecOperand(MI, OpNo, O);
243 void AMDGPUInstPrinter::printNFMT(const MCInst *MI, unsigned OpNo,
244 const MCSubtargetInfo &STI,
246 if (MI->getOperand(OpNo).getImm()) {
248 printU8ImmDecOperand(MI, OpNo, O);
252 void AMDGPUInstPrinter::printRegOperand(unsigned RegNo, raw_ostream &O,
253 const MCRegisterInfo &MRI) {
267 case AMDGPU::FLAT_SCR:
288 case AMDGPU::EXEC_LO:
291 case AMDGPU::EXEC_HI:
294 case AMDGPU::FLAT_SCR_LO:
295 O << "flat_scratch_lo";
297 case AMDGPU::FLAT_SCR_HI:
298 O << "flat_scratch_hi";
302 case AMDGPU::SCRATCH_WAVE_OFFSET_REG:
303 case AMDGPU::PRIVATE_RSRC_REG:
304 llvm_unreachable("pseudo-register should not ever be emitted");
309 // The low 8 bits of the encoding value is the register index, for both VGPRs
311 unsigned RegIdx = MRI.getEncodingValue(RegNo) & ((1 << 8) - 1);
314 if (MRI.getRegClass(AMDGPU::VGPR_32RegClassID).contains(RegNo)) {
317 } else if (MRI.getRegClass(AMDGPU::SGPR_32RegClassID).contains(RegNo)) {
320 } else if (MRI.getRegClass(AMDGPU::VReg_64RegClassID).contains(RegNo)) {
323 } else if (MRI.getRegClass(AMDGPU::SGPR_64RegClassID).contains(RegNo)) {
326 } else if (MRI.getRegClass(AMDGPU::VReg_128RegClassID).contains(RegNo)) {
329 } else if (MRI.getRegClass(AMDGPU::SGPR_128RegClassID).contains(RegNo)) {
332 } else if (MRI.getRegClass(AMDGPU::VReg_96RegClassID).contains(RegNo)) {
335 } else if (MRI.getRegClass(AMDGPU::VReg_256RegClassID).contains(RegNo)) {
338 } else if (MRI.getRegClass(AMDGPU::SGPR_256RegClassID).contains(RegNo)) {
341 } else if (MRI.getRegClass(AMDGPU::VReg_512RegClassID).contains(RegNo)) {
344 } else if (MRI.getRegClass(AMDGPU::SGPR_512RegClassID).contains(RegNo)) {
348 O << getRegisterName(RegNo);
357 O << '[' << RegIdx << ':' << (RegIdx + NumRegs - 1) << ']';
360 void AMDGPUInstPrinter::printVOPDst(const MCInst *MI, unsigned OpNo,
361 const MCSubtargetInfo &STI, raw_ostream &O) {
362 if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3)
364 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::DPP)
366 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::SDWA)
371 printOperand(MI, OpNo, STI, O);
374 void AMDGPUInstPrinter::printImmediate16(uint32_t Imm,
375 const MCSubtargetInfo &STI,
377 int16_t SImm = static_cast<int16_t>(Imm);
378 if (SImm >= -16 && SImm <= 64) {
385 else if (Imm == 0xBC00)
387 else if (Imm == 0x3800)
389 else if (Imm == 0xB800)
391 else if (Imm == 0x4000)
393 else if (Imm == 0xC000)
395 else if (Imm == 0x4400)
397 else if (Imm == 0xC400)
399 else if (Imm == 0x3118) {
400 assert(STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]);
403 O << formatHex(static_cast<uint64_t>(Imm));
406 void AMDGPUInstPrinter::printImmediateV216(uint32_t Imm,
407 const MCSubtargetInfo &STI,
409 uint16_t Lo16 = static_cast<uint16_t>(Imm);
410 printImmediate16(Lo16, STI, O);
413 void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
414 const MCSubtargetInfo &STI,
416 int32_t SImm = static_cast<int32_t>(Imm);
417 if (SImm >= -16 && SImm <= 64) {
422 if (Imm == FloatToBits(0.0f))
424 else if (Imm == FloatToBits(1.0f))
426 else if (Imm == FloatToBits(-1.0f))
428 else if (Imm == FloatToBits(0.5f))
430 else if (Imm == FloatToBits(-0.5f))
432 else if (Imm == FloatToBits(2.0f))
434 else if (Imm == FloatToBits(-2.0f))
436 else if (Imm == FloatToBits(4.0f))
438 else if (Imm == FloatToBits(-4.0f))
440 else if (Imm == 0x3e22f983 &&
441 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
444 O << formatHex(static_cast<uint64_t>(Imm));
447 void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
448 const MCSubtargetInfo &STI,
450 int64_t SImm = static_cast<int64_t>(Imm);
451 if (SImm >= -16 && SImm <= 64) {
456 if (Imm == DoubleToBits(0.0))
458 else if (Imm == DoubleToBits(1.0))
460 else if (Imm == DoubleToBits(-1.0))
462 else if (Imm == DoubleToBits(0.5))
464 else if (Imm == DoubleToBits(-0.5))
466 else if (Imm == DoubleToBits(2.0))
468 else if (Imm == DoubleToBits(-2.0))
470 else if (Imm == DoubleToBits(4.0))
472 else if (Imm == DoubleToBits(-4.0))
474 else if (Imm == 0x3fc45f306dc9c882 &&
475 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
478 assert(isUInt<32>(Imm) || Imm == 0x3fc45f306dc9c882);
480 // In rare situations, we will have a 32-bit literal in a 64-bit
481 // operand. This is technically allowed for the encoding of s_mov_b64.
482 O << formatHex(static_cast<uint64_t>(Imm));
486 void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
487 const MCSubtargetInfo &STI,
489 if (!STI.getFeatureBits()[AMDGPU::FeatureGCN]) {
490 static_cast<R600InstPrinter*>(this)->printOperand(MI, OpNo, O);
494 if (OpNo >= MI->getNumOperands()) {
495 O << "/*Missing OP" << OpNo << "*/";
499 const MCOperand &Op = MI->getOperand(OpNo);
501 printRegOperand(Op.getReg(), O, MRI);
502 } else if (Op.isImm()) {
503 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
504 switch (Desc.OpInfo[OpNo].OperandType) {
505 case AMDGPU::OPERAND_REG_IMM_INT32:
506 case AMDGPU::OPERAND_REG_IMM_FP32:
507 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
508 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
509 case MCOI::OPERAND_IMMEDIATE:
510 printImmediate32(Op.getImm(), STI, O);
512 case AMDGPU::OPERAND_REG_IMM_INT64:
513 case AMDGPU::OPERAND_REG_IMM_FP64:
514 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
515 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
516 printImmediate64(Op.getImm(), STI, O);
518 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
519 case AMDGPU::OPERAND_REG_INLINE_C_FP16:
520 case AMDGPU::OPERAND_REG_IMM_INT16:
521 case AMDGPU::OPERAND_REG_IMM_FP16:
522 printImmediate16(Op.getImm(), STI, O);
524 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
525 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
526 printImmediateV216(Op.getImm(), STI, O);
528 case MCOI::OPERAND_UNKNOWN:
529 case MCOI::OPERAND_PCREL:
530 O << formatDec(Op.getImm());
532 case MCOI::OPERAND_REGISTER:
533 // FIXME: This should be removed and handled somewhere else. Seems to come
534 // from a disassembler bug.
535 O << "/*invalid immediate*/";
538 // We hit this for the immediate instruction bits that don't yet have a
540 llvm_unreachable("unexpected immediate operand type");
542 } else if (Op.isFPImm()) {
543 // We special case 0.0 because otherwise it will be printed as an integer.
544 if (Op.getFPImm() == 0.0)
547 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
548 int RCID = Desc.OpInfo[OpNo].RegClass;
549 unsigned RCBits = AMDGPU::getRegBitWidth(MRI.getRegClass(RCID));
551 printImmediate32(FloatToBits(Op.getFPImm()), STI, O);
552 else if (RCBits == 64)
553 printImmediate64(DoubleToBits(Op.getFPImm()), STI, O);
555 llvm_unreachable("Invalid register class size");
557 } else if (Op.isExpr()) {
558 const MCExpr *Exp = Op.getExpr();
565 void AMDGPUInstPrinter::printOperandAndFPInputMods(const MCInst *MI,
567 const MCSubtargetInfo &STI,
569 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
571 // Use 'neg(...)' instead of '-' to avoid ambiguity.
572 // This is important for integer literals because
573 // -1 is not the same value as neg(1).
574 bool NegMnemo = false;
576 if (InputModifiers & SISrcMods::NEG) {
577 if (OpNo + 1 < MI->getNumOperands() &&
578 (InputModifiers & SISrcMods::ABS) == 0) {
579 const MCOperand &Op = MI->getOperand(OpNo + 1);
580 NegMnemo = Op.isImm() || Op.isFPImm();
589 if (InputModifiers & SISrcMods::ABS)
591 printOperand(MI, OpNo + 1, STI, O);
592 if (InputModifiers & SISrcMods::ABS)
600 void AMDGPUInstPrinter::printOperandAndIntInputMods(const MCInst *MI,
602 const MCSubtargetInfo &STI,
604 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
605 if (InputModifiers & SISrcMods::SEXT)
607 printOperand(MI, OpNo + 1, STI, O);
608 if (InputModifiers & SISrcMods::SEXT)
612 void AMDGPUInstPrinter::printDPPCtrl(const MCInst *MI, unsigned OpNo,
613 const MCSubtargetInfo &STI,
615 unsigned Imm = MI->getOperand(OpNo).getImm();
618 O << formatDec(Imm & 0x3) << ',';
619 O << formatDec((Imm & 0xc) >> 2) << ',';
620 O << formatDec((Imm & 0x30) >> 4) << ',';
621 O << formatDec((Imm & 0xc0) >> 6) << ']';
622 } else if ((Imm >= 0x101) && (Imm <= 0x10f)) {
624 printU4ImmDecOperand(MI, OpNo, O);
625 } else if ((Imm >= 0x111) && (Imm <= 0x11f)) {
627 printU4ImmDecOperand(MI, OpNo, O);
628 } else if ((Imm >= 0x121) && (Imm <= 0x12f)) {
630 printU4ImmDecOperand(MI, OpNo, O);
631 } else if (Imm == 0x130) {
633 } else if (Imm == 0x134) {
635 } else if (Imm == 0x138) {
637 } else if (Imm == 0x13c) {
639 } else if (Imm == 0x140) {
641 } else if (Imm == 0x141) {
642 O << " row_half_mirror";
643 } else if (Imm == 0x142) {
644 O << " row_bcast:15";
645 } else if (Imm == 0x143) {
646 O << " row_bcast:31";
648 llvm_unreachable("Invalid dpp_ctrl value");
652 void AMDGPUInstPrinter::printRowMask(const MCInst *MI, unsigned OpNo,
653 const MCSubtargetInfo &STI,
656 printU4ImmOperand(MI, OpNo, STI, O);
659 void AMDGPUInstPrinter::printBankMask(const MCInst *MI, unsigned OpNo,
660 const MCSubtargetInfo &STI,
663 printU4ImmOperand(MI, OpNo, STI, O);
666 void AMDGPUInstPrinter::printBoundCtrl(const MCInst *MI, unsigned OpNo,
667 const MCSubtargetInfo &STI,
669 unsigned Imm = MI->getOperand(OpNo).getImm();
671 O << " bound_ctrl:0"; // XXX - this syntax is used in sp3
675 void AMDGPUInstPrinter::printSDWASel(const MCInst *MI, unsigned OpNo,
677 using namespace llvm::AMDGPU::SDWA;
679 unsigned Imm = MI->getOperand(OpNo).getImm();
681 case SdwaSel::BYTE_0: O << "BYTE_0"; break;
682 case SdwaSel::BYTE_1: O << "BYTE_1"; break;
683 case SdwaSel::BYTE_2: O << "BYTE_2"; break;
684 case SdwaSel::BYTE_3: O << "BYTE_3"; break;
685 case SdwaSel::WORD_0: O << "WORD_0"; break;
686 case SdwaSel::WORD_1: O << "WORD_1"; break;
687 case SdwaSel::DWORD: O << "DWORD"; break;
688 default: llvm_unreachable("Invalid SDWA data select operand");
692 void AMDGPUInstPrinter::printSDWADstSel(const MCInst *MI, unsigned OpNo,
693 const MCSubtargetInfo &STI,
696 printSDWASel(MI, OpNo, O);
699 void AMDGPUInstPrinter::printSDWASrc0Sel(const MCInst *MI, unsigned OpNo,
700 const MCSubtargetInfo &STI,
703 printSDWASel(MI, OpNo, O);
706 void AMDGPUInstPrinter::printSDWASrc1Sel(const MCInst *MI, unsigned OpNo,
707 const MCSubtargetInfo &STI,
710 printSDWASel(MI, OpNo, O);
713 void AMDGPUInstPrinter::printSDWADstUnused(const MCInst *MI, unsigned OpNo,
714 const MCSubtargetInfo &STI,
716 using namespace llvm::AMDGPU::SDWA;
719 unsigned Imm = MI->getOperand(OpNo).getImm();
721 case DstUnused::UNUSED_PAD: O << "UNUSED_PAD"; break;
722 case DstUnused::UNUSED_SEXT: O << "UNUSED_SEXT"; break;
723 case DstUnused::UNUSED_PRESERVE: O << "UNUSED_PRESERVE"; break;
724 default: llvm_unreachable("Invalid SDWA dest_unused operand");
728 template <unsigned N>
729 void AMDGPUInstPrinter::printExpSrcN(const MCInst *MI, unsigned OpNo,
730 const MCSubtargetInfo &STI,
732 unsigned Opc = MI->getOpcode();
733 int EnIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::en);
734 unsigned En = MI->getOperand(EnIdx).getImm();
736 int ComprIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::compr);
738 // If compr is set, print as src0, src0, src1, src1
739 if (MI->getOperand(ComprIdx).getImm()) {
740 if (N == 1 || N == 2)
747 printRegOperand(MI->getOperand(OpNo).getReg(), O, MRI);
752 void AMDGPUInstPrinter::printExpSrc0(const MCInst *MI, unsigned OpNo,
753 const MCSubtargetInfo &STI,
755 printExpSrcN<0>(MI, OpNo, STI, O);
758 void AMDGPUInstPrinter::printExpSrc1(const MCInst *MI, unsigned OpNo,
759 const MCSubtargetInfo &STI,
761 printExpSrcN<1>(MI, OpNo, STI, O);
764 void AMDGPUInstPrinter::printExpSrc2(const MCInst *MI, unsigned OpNo,
765 const MCSubtargetInfo &STI,
767 printExpSrcN<2>(MI, OpNo, STI, O);
770 void AMDGPUInstPrinter::printExpSrc3(const MCInst *MI, unsigned OpNo,
771 const MCSubtargetInfo &STI,
773 printExpSrcN<3>(MI, OpNo, STI, O);
776 void AMDGPUInstPrinter::printExpTgt(const MCInst *MI, unsigned OpNo,
777 const MCSubtargetInfo &STI,
779 // This is really a 6 bit field.
780 uint32_t Tgt = MI->getOperand(OpNo).getImm() & ((1 << 6) - 1);
788 else if (Tgt >= 12 && Tgt <= 15)
789 O << " pos" << Tgt - 12;
790 else if (Tgt >= 32 && Tgt <= 63)
791 O << " param" << Tgt - 32;
793 // Reserved values 10, 11
794 O << " invalid_target_" << Tgt;
798 static bool allOpsDefaultValue(const int* Ops, int NumOps, int Mod,
799 bool IsPacked, bool HasDstSel) {
800 int DefaultValue = IsPacked && (Mod == SISrcMods::OP_SEL_1);
802 for (int I = 0; I < NumOps; ++I) {
803 if (!!(Ops[I] & Mod) != DefaultValue)
807 if (HasDstSel && (Ops[0] & SISrcMods::DST_OP_SEL) != 0)
813 void AMDGPUInstPrinter::printPackedModifier(const MCInst *MI,
817 unsigned Opc = MI->getOpcode();
821 for (int OpName : { AMDGPU::OpName::src0_modifiers,
822 AMDGPU::OpName::src1_modifiers,
823 AMDGPU::OpName::src2_modifiers }) {
824 int Idx = AMDGPU::getNamedOperandIdx(Opc, OpName);
828 Ops[NumOps++] = MI->getOperand(Idx).getImm();
831 const bool HasDstSel =
833 Mod == SISrcMods::OP_SEL_0 &&
834 MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3_OPSEL;
836 const bool IsPacked =
837 MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::IsPacked;
839 if (allOpsDefaultValue(Ops, NumOps, Mod, IsPacked, HasDstSel))
843 for (int I = 0; I < NumOps; ++I) {
847 O << !!(Ops[I] & Mod);
851 O << ',' << !!(Ops[0] & SISrcMods::DST_OP_SEL);
857 void AMDGPUInstPrinter::printOpSel(const MCInst *MI, unsigned,
858 const MCSubtargetInfo &STI,
860 printPackedModifier(MI, " op_sel:[", SISrcMods::OP_SEL_0, O);
863 void AMDGPUInstPrinter::printOpSelHi(const MCInst *MI, unsigned OpNo,
864 const MCSubtargetInfo &STI,
866 printPackedModifier(MI, " op_sel_hi:[", SISrcMods::OP_SEL_1, O);
869 void AMDGPUInstPrinter::printNegLo(const MCInst *MI, unsigned OpNo,
870 const MCSubtargetInfo &STI,
872 printPackedModifier(MI, " neg_lo:[", SISrcMods::NEG, O);
875 void AMDGPUInstPrinter::printNegHi(const MCInst *MI, unsigned OpNo,
876 const MCSubtargetInfo &STI,
878 printPackedModifier(MI, " neg_hi:[", SISrcMods::NEG_HI, O);
881 void AMDGPUInstPrinter::printInterpSlot(const MCInst *MI, unsigned OpNum,
882 const MCSubtargetInfo &STI,
884 unsigned Imm = MI->getOperand(OpNum).getImm();
896 O << "invalid_param_" << Imm;
900 void AMDGPUInstPrinter::printInterpAttr(const MCInst *MI, unsigned OpNum,
901 const MCSubtargetInfo &STI,
903 unsigned Attr = MI->getOperand(OpNum).getImm();
907 void AMDGPUInstPrinter::printInterpAttrChan(const MCInst *MI, unsigned OpNum,
908 const MCSubtargetInfo &STI,
910 unsigned Chan = MI->getOperand(OpNum).getImm();
911 O << '.' << "xyzw"[Chan & 0x3];
914 void AMDGPUInstPrinter::printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
915 const MCSubtargetInfo &STI,
917 unsigned Val = MI->getOperand(OpNo).getImm();
923 if (Val & VGPRIndexMode::DST_ENABLE)
926 if (Val & VGPRIndexMode::SRC0_ENABLE)
929 if (Val & VGPRIndexMode::SRC1_ENABLE)
932 if (Val & VGPRIndexMode::SRC2_ENABLE)
936 void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
937 const MCSubtargetInfo &STI,
939 if (!STI.getFeatureBits()[AMDGPU::FeatureGCN]) {
940 static_cast<R600InstPrinter*>(this)->printMemOperand(MI, OpNo, O);
944 printOperand(MI, OpNo, STI, O);
946 printOperand(MI, OpNo + 1, STI, O);
949 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
950 raw_ostream &O, StringRef Asm,
952 const MCOperand &Op = MI->getOperand(OpNo);
954 if (Op.getImm() == 1) {
961 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
962 raw_ostream &O, char Asm) {
963 const MCOperand &Op = MI->getOperand(OpNo);
965 if (Op.getImm() == 1)
969 void AMDGPUInstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
970 const MCSubtargetInfo &STI, raw_ostream &O) {
971 static_cast<R600InstPrinter*>(this)->printAbs(MI, OpNo, O);
974 void AMDGPUInstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
975 const MCSubtargetInfo &STI, raw_ostream &O) {
976 static_cast<R600InstPrinter*>(this)->printClamp(MI, OpNo, O);
979 void AMDGPUInstPrinter::printHigh(const MCInst *MI, unsigned OpNo,
980 const MCSubtargetInfo &STI,
982 if (MI->getOperand(OpNo).getImm())
986 void AMDGPUInstPrinter::printClampSI(const MCInst *MI, unsigned OpNo,
987 const MCSubtargetInfo &STI,
989 if (MI->getOperand(OpNo).getImm())
993 void AMDGPUInstPrinter::printOModSI(const MCInst *MI, unsigned OpNo,
994 const MCSubtargetInfo &STI,
996 int Imm = MI->getOperand(OpNo).getImm();
997 if (Imm == SIOutMods::MUL2)
999 else if (Imm == SIOutMods::MUL4)
1001 else if (Imm == SIOutMods::DIV2)
1005 void AMDGPUInstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
1006 const MCSubtargetInfo &STI,
1008 static_cast<R600InstPrinter*>(this)->printLiteral(MI, OpNo, O);
1011 void AMDGPUInstPrinter::printLast(const MCInst *MI, unsigned OpNo,
1012 const MCSubtargetInfo &STI, raw_ostream &O) {
1013 static_cast<R600InstPrinter*>(this)->printLast(MI, OpNo, O);
1016 void AMDGPUInstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
1017 const MCSubtargetInfo &STI, raw_ostream &O) {
1018 static_cast<R600InstPrinter*>(this)->printNeg(MI, OpNo, O);
1021 void AMDGPUInstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
1022 const MCSubtargetInfo &STI, raw_ostream &O) {
1023 static_cast<R600InstPrinter*>(this)->printOMOD(MI, OpNo, O);
1026 void AMDGPUInstPrinter::printRel(const MCInst *MI, unsigned OpNo,
1027 const MCSubtargetInfo &STI, raw_ostream &O) {
1028 static_cast<R600InstPrinter*>(this)->printRel(MI, OpNo, O);
1031 void AMDGPUInstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
1032 const MCSubtargetInfo &STI,
1034 static_cast<R600InstPrinter*>(this)->printUpdateExecMask(MI, OpNo, O);
1037 void AMDGPUInstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
1038 const MCSubtargetInfo &STI,
1040 static_cast<R600InstPrinter*>(this)->printUpdatePred(MI, OpNo, O);
1043 void AMDGPUInstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
1044 const MCSubtargetInfo &STI, raw_ostream &O) {
1045 static_cast<R600InstPrinter*>(this)->printWrite(MI, OpNo, O);
1048 void AMDGPUInstPrinter::printBankSwizzle(const MCInst *MI, unsigned OpNo,
1049 const MCSubtargetInfo &STI,
1051 static_cast<R600InstPrinter*>(this)->printBankSwizzle(MI, OpNo, O);
1054 void AMDGPUInstPrinter::printRSel(const MCInst *MI, unsigned OpNo,
1055 const MCSubtargetInfo &STI, raw_ostream &O) {
1056 static_cast<R600InstPrinter*>(this)->printRSel(MI, OpNo, O);
1059 void AMDGPUInstPrinter::printCT(const MCInst *MI, unsigned OpNo,
1060 const MCSubtargetInfo &STI, raw_ostream &O) {
1061 static_cast<R600InstPrinter*>(this)->printCT(MI, OpNo, O);
1064 void AMDGPUInstPrinter::printKCache(const MCInst *MI, unsigned OpNo,
1065 const MCSubtargetInfo &STI, raw_ostream &O) {
1066 static_cast<R600InstPrinter*>(this)->printKCache(MI, OpNo, O);
1069 void AMDGPUInstPrinter::printSendMsg(const MCInst *MI, unsigned OpNo,
1070 const MCSubtargetInfo &STI,
1072 using namespace llvm::AMDGPU::SendMsg;
1074 const unsigned SImm16 = MI->getOperand(OpNo).getImm();
1075 const unsigned Id = SImm16 & ID_MASK_;
1077 if (Id == ID_INTERRUPT) {
1078 if ((SImm16 & ~ID_MASK_) != 0) // Unused/unknown bits must be 0.
1080 O << "sendmsg(" << IdSymbolic[Id] << ')';
1083 if (Id == ID_GS || Id == ID_GS_DONE) {
1084 if ((SImm16 & ~(ID_MASK_|OP_GS_MASK_|STREAM_ID_MASK_)) != 0) // Unused/unknown bits must be 0.
1086 const unsigned OpGs = (SImm16 & OP_GS_MASK_) >> OP_SHIFT_;
1087 const unsigned StreamId = (SImm16 & STREAM_ID_MASK_) >> STREAM_ID_SHIFT_;
1088 if (OpGs == OP_GS_NOP && Id != ID_GS_DONE) // NOP to be used for GS_DONE only.
1090 if (OpGs == OP_GS_NOP && StreamId != 0) // NOP does not use/define stream id bits.
1092 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpGsSymbolic[OpGs];
1093 if (OpGs != OP_GS_NOP) { O << ", " << StreamId; }
1097 if (Id == ID_SYSMSG) {
1098 if ((SImm16 & ~(ID_MASK_|OP_SYS_MASK_)) != 0) // Unused/unknown bits must be 0.
1100 const unsigned OpSys = (SImm16 & OP_SYS_MASK_) >> OP_SHIFT_;
1101 if (! (OP_SYS_FIRST_ <= OpSys && OpSys < OP_SYS_LAST_)) // Unused/unknown.
1103 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpSysSymbolic[OpSys] << ')';
1107 O << SImm16; // Unknown simm16 code.
1110 static void printSwizzleBitmask(const uint16_t AndMask,
1111 const uint16_t OrMask,
1112 const uint16_t XorMask,
1114 using namespace llvm::AMDGPU::Swizzle;
1116 uint16_t Probe0 = ((0 & AndMask) | OrMask) ^ XorMask;
1117 uint16_t Probe1 = ((BITMASK_MASK & AndMask) | OrMask) ^ XorMask;
1121 for (unsigned Mask = 1 << (BITMASK_WIDTH - 1); Mask > 0; Mask >>= 1) {
1122 uint16_t p0 = Probe0 & Mask;
1123 uint16_t p1 = Probe1 & Mask;
1143 void AMDGPUInstPrinter::printSwizzle(const MCInst *MI, unsigned OpNo,
1144 const MCSubtargetInfo &STI,
1146 using namespace llvm::AMDGPU::Swizzle;
1148 uint16_t Imm = MI->getOperand(OpNo).getImm();
1155 if ((Imm & QUAD_PERM_ENC_MASK) == QUAD_PERM_ENC) {
1157 O << "swizzle(" << IdSymbolic[ID_QUAD_PERM];
1158 for (auto i = 0; i < LANE_NUM; ++i) {
1160 O << formatDec(Imm & LANE_MASK);
1165 } else if ((Imm & BITMASK_PERM_ENC_MASK) == BITMASK_PERM_ENC) {
1167 uint16_t AndMask = (Imm >> BITMASK_AND_SHIFT) & BITMASK_MASK;
1168 uint16_t OrMask = (Imm >> BITMASK_OR_SHIFT) & BITMASK_MASK;
1169 uint16_t XorMask = (Imm >> BITMASK_XOR_SHIFT) & BITMASK_MASK;
1171 if (AndMask == BITMASK_MAX &&
1173 countPopulation(XorMask) == 1) {
1175 O << "swizzle(" << IdSymbolic[ID_SWAP];
1177 O << formatDec(XorMask);
1180 } else if (AndMask == BITMASK_MAX &&
1181 OrMask == 0 && XorMask > 0 &&
1182 isPowerOf2_64(XorMask + 1)) {
1184 O << "swizzle(" << IdSymbolic[ID_REVERSE];
1186 O << formatDec(XorMask + 1);
1191 uint16_t GroupSize = BITMASK_MAX - AndMask + 1;
1192 if (GroupSize > 1 &&
1193 isPowerOf2_64(GroupSize) &&
1194 OrMask < GroupSize &&
1197 O << "swizzle(" << IdSymbolic[ID_BROADCAST];
1199 O << formatDec(GroupSize);
1201 O << formatDec(OrMask);
1205 O << "swizzle(" << IdSymbolic[ID_BITMASK_PERM];
1207 printSwizzleBitmask(AndMask, OrMask, XorMask, O);
1212 printU16ImmDecOperand(MI, OpNo, O);
1216 void AMDGPUInstPrinter::printWaitFlag(const MCInst *MI, unsigned OpNo,
1217 const MCSubtargetInfo &STI,
1219 AMDGPU::IsaInfo::IsaVersion ISA =
1220 AMDGPU::IsaInfo::getIsaVersion(STI.getFeatureBits());
1222 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1223 unsigned Vmcnt, Expcnt, Lgkmcnt;
1224 decodeWaitcnt(ISA, SImm16, Vmcnt, Expcnt, Lgkmcnt);
1226 bool NeedSpace = false;
1228 if (Vmcnt != getVmcntBitMask(ISA)) {
1229 O << "vmcnt(" << Vmcnt << ')';
1233 if (Expcnt != getExpcntBitMask(ISA)) {
1236 O << "expcnt(" << Expcnt << ')';
1240 if (Lgkmcnt != getLgkmcntBitMask(ISA)) {
1243 O << "lgkmcnt(" << Lgkmcnt << ')';
1247 void AMDGPUInstPrinter::printHwreg(const MCInst *MI, unsigned OpNo,
1248 const MCSubtargetInfo &STI, raw_ostream &O) {
1249 using namespace llvm::AMDGPU::Hwreg;
1251 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1252 const unsigned Id = (SImm16 & ID_MASK_) >> ID_SHIFT_;
1253 const unsigned Offset = (SImm16 & OFFSET_MASK_) >> OFFSET_SHIFT_;
1254 const unsigned Width = ((SImm16 & WIDTH_M1_MASK_) >> WIDTH_M1_SHIFT_) + 1;
1257 if (ID_SYMBOLIC_FIRST_ <= Id && Id < ID_SYMBOLIC_LAST_) {
1258 O << IdSymbolic[Id];
1262 if (Width != WIDTH_M1_DEFAULT_ + 1 || Offset != OFFSET_DEFAULT_) {
1263 O << ", " << Offset << ", " << Width;
1268 #include "AMDGPUGenAsmWriter.inc"
1270 void R600InstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
1272 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '|');
1275 void R600InstPrinter::printBankSwizzle(const MCInst *MI, unsigned OpNo,
1277 int BankSwizzle = MI->getOperand(OpNo).getImm();
1278 switch (BankSwizzle) {
1280 O << "BS:VEC_021/SCL_122";
1283 O << "BS:VEC_120/SCL_212";
1286 O << "BS:VEC_102/SCL_221";
1299 void R600InstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
1301 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "_SAT");
1304 void R600InstPrinter::printCT(const MCInst *MI, unsigned OpNo,
1306 unsigned CT = MI->getOperand(OpNo).getImm();
1319 void R600InstPrinter::printKCache(const MCInst *MI, unsigned OpNo,
1321 int KCacheMode = MI->getOperand(OpNo).getImm();
1322 if (KCacheMode > 0) {
1323 int KCacheBank = MI->getOperand(OpNo - 2).getImm();
1324 O << "CB" << KCacheBank << ':';
1325 int KCacheAddr = MI->getOperand(OpNo + 2).getImm();
1326 int LineSize = (KCacheMode == 1) ? 16 : 32;
1327 O << KCacheAddr * 16 << '-' << KCacheAddr * 16 + LineSize;
1331 void R600InstPrinter::printLast(const MCInst *MI, unsigned OpNo,
1333 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "*", " ");
1336 void R600InstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
1338 const MCOperand &Op = MI->getOperand(OpNo);
1339 assert(Op.isImm() || Op.isExpr());
1341 int64_t Imm = Op.getImm();
1342 O << Imm << '(' << BitsToFloat(Imm) << ')';
1345 Op.getExpr()->print(O << '@', &MAI);
1349 void R600InstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
1351 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '-');
1354 void R600InstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
1356 switch (MI->getOperand(OpNo).getImm()) {
1370 void R600InstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
1372 printOperand(MI, OpNo, O);
1374 printOperand(MI, OpNo + 1, O);
1377 void R600InstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
1379 if (OpNo >= MI->getNumOperands()) {
1380 O << "/*Missing OP" << OpNo << "*/";
1384 const MCOperand &Op = MI->getOperand(OpNo);
1386 switch (Op.getReg()) {
1387 // This is the default predicate state, so we don't need to print it.
1388 case AMDGPU::PRED_SEL_OFF:
1392 O << getRegisterName(Op.getReg());
1395 } else if (Op.isImm()) {
1397 } else if (Op.isFPImm()) {
1398 // We special case 0.0 because otherwise it will be printed as an integer.
1399 if (Op.getFPImm() == 0.0)
1404 } else if (Op.isExpr()) {
1405 const MCExpr *Exp = Op.getExpr();
1406 Exp->print(O, &MAI);
1412 void R600InstPrinter::printRel(const MCInst *MI, unsigned OpNo,
1414 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, '+');
1417 void R600InstPrinter::printRSel(const MCInst *MI, unsigned OpNo,
1419 unsigned Sel = MI->getOperand(OpNo).getImm();
1447 void R600InstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
1449 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "ExecMask,");
1452 void R600InstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
1454 AMDGPUInstPrinter::printIfSet(MI, OpNo, O, "Pred,");
1457 void R600InstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
1459 const MCOperand &Op = MI->getOperand(OpNo);
1460 if (Op.getImm() == 0) {