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::printS16ImmDecOperand(const MCInst *MI, unsigned OpNo,
77 O << formatDec(static_cast<int16_t>(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 printS16ImmDecOperand(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";
304 // The low 8 bits of the encoding value is the register index, for both VGPRs
306 unsigned RegIdx = MRI.getEncodingValue(RegNo) & ((1 << 8) - 1);
309 if (MRI.getRegClass(AMDGPU::VGPR_32RegClassID).contains(RegNo)) {
312 } else if (MRI.getRegClass(AMDGPU::SGPR_32RegClassID).contains(RegNo)) {
315 } else if (MRI.getRegClass(AMDGPU::VReg_64RegClassID).contains(RegNo)) {
318 } else if (MRI.getRegClass(AMDGPU::SGPR_64RegClassID).contains(RegNo)) {
321 } else if (MRI.getRegClass(AMDGPU::VReg_128RegClassID).contains(RegNo)) {
324 } else if (MRI.getRegClass(AMDGPU::SGPR_128RegClassID).contains(RegNo)) {
327 } else if (MRI.getRegClass(AMDGPU::VReg_96RegClassID).contains(RegNo)) {
330 } else if (MRI.getRegClass(AMDGPU::VReg_256RegClassID).contains(RegNo)) {
333 } else if (MRI.getRegClass(AMDGPU::SReg_256RegClassID).contains(RegNo)) {
336 } else if (MRI.getRegClass(AMDGPU::VReg_512RegClassID).contains(RegNo)) {
339 } else if (MRI.getRegClass(AMDGPU::SReg_512RegClassID).contains(RegNo)) {
342 } else if (MRI.getRegClass(AMDGPU::TTMP_64RegClassID).contains(RegNo)) {
345 // Trap temps start at offset 112. TODO: Get this from tablegen.
347 } else if (MRI.getRegClass(AMDGPU::TTMP_128RegClassID).contains(RegNo)) {
350 // Trap temps start at offset 112. TODO: Get this from tablegen.
353 O << getRegisterName(RegNo);
362 O << '[' << RegIdx << ':' << (RegIdx + NumRegs - 1) << ']';
365 void AMDGPUInstPrinter::printVOPDst(const MCInst *MI, unsigned OpNo,
366 const MCSubtargetInfo &STI, raw_ostream &O) {
367 if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::VOP3)
369 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::DPP)
371 else if (MII.get(MI->getOpcode()).TSFlags & SIInstrFlags::SDWA)
376 printOperand(MI, OpNo, STI, O);
379 void AMDGPUInstPrinter::printImmediate16(uint32_t Imm,
380 const MCSubtargetInfo &STI,
382 int16_t SImm = static_cast<int16_t>(Imm);
383 if (SImm >= -16 && SImm <= 64) {
390 else if (Imm == 0xBC00)
392 else if (Imm == 0x3800)
394 else if (Imm == 0xB800)
396 else if (Imm == 0x4000)
398 else if (Imm == 0xC000)
400 else if (Imm == 0x4400)
402 else if (Imm == 0xC400)
404 else if (Imm == 0x3118) {
405 assert(STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm]);
408 O << formatHex(static_cast<uint64_t>(Imm));
411 void AMDGPUInstPrinter::printImmediateV216(uint32_t Imm,
412 const MCSubtargetInfo &STI,
414 uint16_t Lo16 = static_cast<uint16_t>(Imm);
415 printImmediate16(Lo16, STI, O);
418 void AMDGPUInstPrinter::printImmediate32(uint32_t Imm,
419 const MCSubtargetInfo &STI,
421 int32_t SImm = static_cast<int32_t>(Imm);
422 if (SImm >= -16 && SImm <= 64) {
427 if (Imm == FloatToBits(0.0f))
429 else if (Imm == FloatToBits(1.0f))
431 else if (Imm == FloatToBits(-1.0f))
433 else if (Imm == FloatToBits(0.5f))
435 else if (Imm == FloatToBits(-0.5f))
437 else if (Imm == FloatToBits(2.0f))
439 else if (Imm == FloatToBits(-2.0f))
441 else if (Imm == FloatToBits(4.0f))
443 else if (Imm == FloatToBits(-4.0f))
445 else if (Imm == 0x3e22f983 &&
446 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
449 O << formatHex(static_cast<uint64_t>(Imm));
452 void AMDGPUInstPrinter::printImmediate64(uint64_t Imm,
453 const MCSubtargetInfo &STI,
455 int64_t SImm = static_cast<int64_t>(Imm);
456 if (SImm >= -16 && SImm <= 64) {
461 if (Imm == DoubleToBits(0.0))
463 else if (Imm == DoubleToBits(1.0))
465 else if (Imm == DoubleToBits(-1.0))
467 else if (Imm == DoubleToBits(0.5))
469 else if (Imm == DoubleToBits(-0.5))
471 else if (Imm == DoubleToBits(2.0))
473 else if (Imm == DoubleToBits(-2.0))
475 else if (Imm == DoubleToBits(4.0))
477 else if (Imm == DoubleToBits(-4.0))
479 else if (Imm == 0x3fc45f306dc9c882 &&
480 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
483 assert(isUInt<32>(Imm) || Imm == 0x3fc45f306dc9c882);
485 // In rare situations, we will have a 32-bit literal in a 64-bit
486 // operand. This is technically allowed for the encoding of s_mov_b64.
487 O << formatHex(static_cast<uint64_t>(Imm));
491 void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
492 const MCSubtargetInfo &STI,
494 if (OpNo >= MI->getNumOperands()) {
495 O << "/*Missing OP" << OpNo << "*/";
499 const MCOperand &Op = MI->getOperand(OpNo);
501 switch (Op.getReg()) {
502 // This is the default predicate state, so we don't need to print it.
503 case AMDGPU::PRED_SEL_OFF:
507 printRegOperand(Op.getReg(), O, MRI);
510 } else if (Op.isImm()) {
511 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
512 switch (Desc.OpInfo[OpNo].OperandType) {
513 case AMDGPU::OPERAND_REG_IMM_INT32:
514 case AMDGPU::OPERAND_REG_IMM_FP32:
515 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
516 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
517 case MCOI::OPERAND_IMMEDIATE:
518 printImmediate32(Op.getImm(), STI, O);
520 case AMDGPU::OPERAND_REG_IMM_INT64:
521 case AMDGPU::OPERAND_REG_IMM_FP64:
522 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
523 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
524 printImmediate64(Op.getImm(), STI, O);
526 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
527 case AMDGPU::OPERAND_REG_INLINE_C_FP16:
528 case AMDGPU::OPERAND_REG_IMM_INT16:
529 case AMDGPU::OPERAND_REG_IMM_FP16:
530 printImmediate16(Op.getImm(), STI, O);
532 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
533 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
534 printImmediateV216(Op.getImm(), STI, O);
536 case MCOI::OPERAND_UNKNOWN:
537 case MCOI::OPERAND_PCREL:
538 O << formatDec(Op.getImm());
540 case MCOI::OPERAND_REGISTER:
541 // FIXME: This should be removed and handled somewhere else. Seems to come
542 // from a disassembler bug.
543 O << "/*invalid immediate*/";
546 // We hit this for the immediate instruction bits that don't yet have a
548 llvm_unreachable("unexpected immediate operand type");
550 } else if (Op.isFPImm()) {
551 // We special case 0.0 because otherwise it will be printed as an integer.
552 if (Op.getFPImm() == 0.0)
555 const MCInstrDesc &Desc = MII.get(MI->getOpcode());
556 int RCID = Desc.OpInfo[OpNo].RegClass;
557 unsigned RCBits = AMDGPU::getRegBitWidth(MRI.getRegClass(RCID));
559 printImmediate32(FloatToBits(Op.getFPImm()), STI, O);
560 else if (RCBits == 64)
561 printImmediate64(DoubleToBits(Op.getFPImm()), STI, O);
563 llvm_unreachable("Invalid register class size");
565 } else if (Op.isExpr()) {
566 const MCExpr *Exp = Op.getExpr();
573 void AMDGPUInstPrinter::printOperandAndFPInputMods(const MCInst *MI,
575 const MCSubtargetInfo &STI,
577 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
579 // Use 'neg(...)' instead of '-' to avoid ambiguity.
580 // This is important for integer literals because
581 // -1 is not the same value as neg(1).
582 bool NegMnemo = false;
584 if (InputModifiers & SISrcMods::NEG) {
585 if (OpNo + 1 < MI->getNumOperands() &&
586 (InputModifiers & SISrcMods::ABS) == 0) {
587 const MCOperand &Op = MI->getOperand(OpNo + 1);
588 NegMnemo = Op.isImm() || Op.isFPImm();
597 if (InputModifiers & SISrcMods::ABS)
599 printOperand(MI, OpNo + 1, STI, O);
600 if (InputModifiers & SISrcMods::ABS)
608 void AMDGPUInstPrinter::printOperandAndIntInputMods(const MCInst *MI,
610 const MCSubtargetInfo &STI,
612 unsigned InputModifiers = MI->getOperand(OpNo).getImm();
613 if (InputModifiers & SISrcMods::SEXT)
615 printOperand(MI, OpNo + 1, STI, O);
616 if (InputModifiers & SISrcMods::SEXT)
620 void AMDGPUInstPrinter::printDPPCtrl(const MCInst *MI, unsigned OpNo,
621 const MCSubtargetInfo &STI,
623 unsigned Imm = MI->getOperand(OpNo).getImm();
626 O << formatDec(Imm & 0x3) << ',';
627 O << formatDec((Imm & 0xc) >> 2) << ',';
628 O << formatDec((Imm & 0x30) >> 4) << ',';
629 O << formatDec((Imm & 0xc0) >> 6) << ']';
630 } else if ((Imm >= 0x101) && (Imm <= 0x10f)) {
632 printU4ImmDecOperand(MI, OpNo, O);
633 } else if ((Imm >= 0x111) && (Imm <= 0x11f)) {
635 printU4ImmDecOperand(MI, OpNo, O);
636 } else if ((Imm >= 0x121) && (Imm <= 0x12f)) {
638 printU4ImmDecOperand(MI, OpNo, O);
639 } else if (Imm == 0x130) {
641 } else if (Imm == 0x134) {
643 } else if (Imm == 0x138) {
645 } else if (Imm == 0x13c) {
647 } else if (Imm == 0x140) {
649 } else if (Imm == 0x141) {
650 O << " row_half_mirror";
651 } else if (Imm == 0x142) {
652 O << " row_bcast:15";
653 } else if (Imm == 0x143) {
654 O << " row_bcast:31";
656 llvm_unreachable("Invalid dpp_ctrl value");
660 void AMDGPUInstPrinter::printRowMask(const MCInst *MI, unsigned OpNo,
661 const MCSubtargetInfo &STI,
664 printU4ImmOperand(MI, OpNo, STI, O);
667 void AMDGPUInstPrinter::printBankMask(const MCInst *MI, unsigned OpNo,
668 const MCSubtargetInfo &STI,
671 printU4ImmOperand(MI, OpNo, STI, O);
674 void AMDGPUInstPrinter::printBoundCtrl(const MCInst *MI, unsigned OpNo,
675 const MCSubtargetInfo &STI,
677 unsigned Imm = MI->getOperand(OpNo).getImm();
679 O << " bound_ctrl:0"; // XXX - this syntax is used in sp3
683 void AMDGPUInstPrinter::printSDWASel(const MCInst *MI, unsigned OpNo,
685 using namespace llvm::AMDGPU::SDWA;
687 unsigned Imm = MI->getOperand(OpNo).getImm();
689 case SdwaSel::BYTE_0: O << "BYTE_0"; break;
690 case SdwaSel::BYTE_1: O << "BYTE_1"; break;
691 case SdwaSel::BYTE_2: O << "BYTE_2"; break;
692 case SdwaSel::BYTE_3: O << "BYTE_3"; break;
693 case SdwaSel::WORD_0: O << "WORD_0"; break;
694 case SdwaSel::WORD_1: O << "WORD_1"; break;
695 case SdwaSel::DWORD: O << "DWORD"; break;
696 default: llvm_unreachable("Invalid SDWA data select operand");
700 void AMDGPUInstPrinter::printSDWADstSel(const MCInst *MI, unsigned OpNo,
701 const MCSubtargetInfo &STI,
704 printSDWASel(MI, OpNo, O);
707 void AMDGPUInstPrinter::printSDWASrc0Sel(const MCInst *MI, unsigned OpNo,
708 const MCSubtargetInfo &STI,
711 printSDWASel(MI, OpNo, O);
714 void AMDGPUInstPrinter::printSDWASrc1Sel(const MCInst *MI, unsigned OpNo,
715 const MCSubtargetInfo &STI,
718 printSDWASel(MI, OpNo, O);
721 void AMDGPUInstPrinter::printSDWADstUnused(const MCInst *MI, unsigned OpNo,
722 const MCSubtargetInfo &STI,
724 using namespace llvm::AMDGPU::SDWA;
727 unsigned Imm = MI->getOperand(OpNo).getImm();
729 case DstUnused::UNUSED_PAD: O << "UNUSED_PAD"; break;
730 case DstUnused::UNUSED_SEXT: O << "UNUSED_SEXT"; break;
731 case DstUnused::UNUSED_PRESERVE: O << "UNUSED_PRESERVE"; break;
732 default: llvm_unreachable("Invalid SDWA dest_unused operand");
736 template <unsigned N>
737 void AMDGPUInstPrinter::printExpSrcN(const MCInst *MI, unsigned OpNo,
738 const MCSubtargetInfo &STI,
740 unsigned Opc = MI->getOpcode();
741 int EnIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::en);
742 unsigned En = MI->getOperand(EnIdx).getImm();
744 int ComprIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::compr);
746 // If compr is set, print as src0, src0, src1, src1
747 if (MI->getOperand(ComprIdx).getImm()) {
748 if (N == 1 || N == 2)
755 printRegOperand(MI->getOperand(OpNo).getReg(), O, MRI);
760 void AMDGPUInstPrinter::printExpSrc0(const MCInst *MI, unsigned OpNo,
761 const MCSubtargetInfo &STI,
763 printExpSrcN<0>(MI, OpNo, STI, O);
766 void AMDGPUInstPrinter::printExpSrc1(const MCInst *MI, unsigned OpNo,
767 const MCSubtargetInfo &STI,
769 printExpSrcN<1>(MI, OpNo, STI, O);
772 void AMDGPUInstPrinter::printExpSrc2(const MCInst *MI, unsigned OpNo,
773 const MCSubtargetInfo &STI,
775 printExpSrcN<2>(MI, OpNo, STI, O);
778 void AMDGPUInstPrinter::printExpSrc3(const MCInst *MI, unsigned OpNo,
779 const MCSubtargetInfo &STI,
781 printExpSrcN<3>(MI, OpNo, STI, O);
784 void AMDGPUInstPrinter::printExpTgt(const MCInst *MI, unsigned OpNo,
785 const MCSubtargetInfo &STI,
787 // This is really a 6 bit field.
788 uint32_t Tgt = MI->getOperand(OpNo).getImm() & ((1 << 6) - 1);
796 else if (Tgt >= 12 && Tgt <= 15)
797 O << " pos" << Tgt - 12;
798 else if (Tgt >= 32 && Tgt <= 63)
799 O << " param" << Tgt - 32;
801 // Reserved values 10, 11
802 O << " invalid_target_" << Tgt;
806 static bool allOpsDefaultValue(const int* Ops, int NumOps, int Mod) {
807 int DefaultValue = (Mod == SISrcMods::OP_SEL_1);
809 for (int I = 0; I < NumOps; ++I) {
810 if (!!(Ops[I] & Mod) != DefaultValue)
817 static void printPackedModifier(const MCInst *MI, StringRef Name, unsigned Mod,
819 unsigned Opc = MI->getOpcode();
823 for (int OpName : { AMDGPU::OpName::src0_modifiers,
824 AMDGPU::OpName::src1_modifiers,
825 AMDGPU::OpName::src2_modifiers }) {
826 int Idx = AMDGPU::getNamedOperandIdx(Opc, OpName);
830 Ops[NumOps++] = MI->getOperand(Idx).getImm();
833 if (allOpsDefaultValue(Ops, NumOps, Mod))
837 for (int I = 0; I < NumOps; ++I) {
841 O << !!(Ops[I] & Mod);
847 void AMDGPUInstPrinter::printOpSel(const MCInst *MI, unsigned,
848 const MCSubtargetInfo &STI,
850 printPackedModifier(MI, " op_sel:[", SISrcMods::OP_SEL_0, O);
853 void AMDGPUInstPrinter::printOpSelHi(const MCInst *MI, unsigned OpNo,
854 const MCSubtargetInfo &STI,
856 printPackedModifier(MI, " op_sel_hi:[", SISrcMods::OP_SEL_1, O);
859 void AMDGPUInstPrinter::printNegLo(const MCInst *MI, unsigned OpNo,
860 const MCSubtargetInfo &STI,
862 printPackedModifier(MI, " neg_lo:[", SISrcMods::NEG, O);
865 void AMDGPUInstPrinter::printNegHi(const MCInst *MI, unsigned OpNo,
866 const MCSubtargetInfo &STI,
868 printPackedModifier(MI, " neg_hi:[", SISrcMods::NEG_HI, O);
871 void AMDGPUInstPrinter::printInterpSlot(const MCInst *MI, unsigned OpNum,
872 const MCSubtargetInfo &STI,
874 unsigned Imm = MI->getOperand(OpNum).getImm();
886 O << "invalid_param_" << Imm;
890 void AMDGPUInstPrinter::printInterpAttr(const MCInst *MI, unsigned OpNum,
891 const MCSubtargetInfo &STI,
893 unsigned Attr = MI->getOperand(OpNum).getImm();
897 void AMDGPUInstPrinter::printInterpAttrChan(const MCInst *MI, unsigned OpNum,
898 const MCSubtargetInfo &STI,
900 unsigned Chan = MI->getOperand(OpNum).getImm();
901 O << '.' << "xyzw"[Chan & 0x3];
904 void AMDGPUInstPrinter::printVGPRIndexMode(const MCInst *MI, unsigned OpNo,
905 const MCSubtargetInfo &STI,
907 unsigned Val = MI->getOperand(OpNo).getImm();
913 if (Val & VGPRIndexMode::DST_ENABLE)
916 if (Val & VGPRIndexMode::SRC0_ENABLE)
919 if (Val & VGPRIndexMode::SRC1_ENABLE)
922 if (Val & VGPRIndexMode::SRC2_ENABLE)
926 void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
927 const MCSubtargetInfo &STI,
929 printOperand(MI, OpNo, STI, O);
931 printOperand(MI, OpNo + 1, STI, O);
934 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
935 raw_ostream &O, StringRef Asm,
937 const MCOperand &Op = MI->getOperand(OpNo);
939 if (Op.getImm() == 1) {
946 void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
947 raw_ostream &O, char Asm) {
948 const MCOperand &Op = MI->getOperand(OpNo);
950 if (Op.getImm() == 1)
954 void AMDGPUInstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
955 const MCSubtargetInfo &STI, raw_ostream &O) {
956 printIfSet(MI, OpNo, O, '|');
959 void AMDGPUInstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
960 const MCSubtargetInfo &STI, raw_ostream &O) {
961 printIfSet(MI, OpNo, O, "_SAT");
964 void AMDGPUInstPrinter::printClampSI(const MCInst *MI, unsigned OpNo,
965 const MCSubtargetInfo &STI,
967 if (MI->getOperand(OpNo).getImm())
971 void AMDGPUInstPrinter::printOModSI(const MCInst *MI, unsigned OpNo,
972 const MCSubtargetInfo &STI,
974 int Imm = MI->getOperand(OpNo).getImm();
975 if (Imm == SIOutMods::MUL2)
977 else if (Imm == SIOutMods::MUL4)
979 else if (Imm == SIOutMods::DIV2)
983 void AMDGPUInstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
984 const MCSubtargetInfo &STI,
986 const MCOperand &Op = MI->getOperand(OpNo);
987 assert(Op.isImm() || Op.isExpr());
989 int64_t Imm = Op.getImm();
990 O << Imm << '(' << BitsToFloat(Imm) << ')';
993 Op.getExpr()->print(O << '@', &MAI);
997 void AMDGPUInstPrinter::printLast(const MCInst *MI, unsigned OpNo,
998 const MCSubtargetInfo &STI, raw_ostream &O) {
999 printIfSet(MI, OpNo, O, "*", " ");
1002 void AMDGPUInstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
1003 const MCSubtargetInfo &STI, raw_ostream &O) {
1004 printIfSet(MI, OpNo, O, '-');
1007 void AMDGPUInstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
1008 const MCSubtargetInfo &STI, raw_ostream &O) {
1009 switch (MI->getOperand(OpNo).getImm()) {
1023 void AMDGPUInstPrinter::printRel(const MCInst *MI, unsigned OpNo,
1024 const MCSubtargetInfo &STI, raw_ostream &O) {
1025 printIfSet(MI, OpNo, O, '+');
1028 void AMDGPUInstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
1029 const MCSubtargetInfo &STI,
1031 printIfSet(MI, OpNo, O, "ExecMask,");
1034 void AMDGPUInstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
1035 const MCSubtargetInfo &STI,
1037 printIfSet(MI, OpNo, O, "Pred,");
1040 void AMDGPUInstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
1041 const MCSubtargetInfo &STI, raw_ostream &O) {
1042 const MCOperand &Op = MI->getOperand(OpNo);
1043 if (Op.getImm() == 0) {
1048 void AMDGPUInstPrinter::printSel(const MCInst *MI, unsigned OpNo,
1050 const char * chans = "XYZW";
1051 int sel = MI->getOperand(OpNo).getImm();
1060 O << cb << '[' << sel << ']';
1061 } else if (sel >= 448) {
1064 } else if (sel >= 0){
1069 O << '.' << chans[chan];
1072 void AMDGPUInstPrinter::printBankSwizzle(const MCInst *MI, unsigned OpNo,
1073 const MCSubtargetInfo &STI,
1075 int BankSwizzle = MI->getOperand(OpNo).getImm();
1076 switch (BankSwizzle) {
1078 O << "BS:VEC_021/SCL_122";
1081 O << "BS:VEC_120/SCL_212";
1084 O << "BS:VEC_102/SCL_221";
1097 void AMDGPUInstPrinter::printRSel(const MCInst *MI, unsigned OpNo,
1098 const MCSubtargetInfo &STI, raw_ostream &O) {
1099 unsigned Sel = MI->getOperand(OpNo).getImm();
1127 void AMDGPUInstPrinter::printCT(const MCInst *MI, unsigned OpNo,
1128 const MCSubtargetInfo &STI, raw_ostream &O) {
1129 unsigned CT = MI->getOperand(OpNo).getImm();
1142 void AMDGPUInstPrinter::printKCache(const MCInst *MI, unsigned OpNo,
1143 const MCSubtargetInfo &STI, raw_ostream &O) {
1144 int KCacheMode = MI->getOperand(OpNo).getImm();
1145 if (KCacheMode > 0) {
1146 int KCacheBank = MI->getOperand(OpNo - 2).getImm();
1147 O << "CB" << KCacheBank << ':';
1148 int KCacheAddr = MI->getOperand(OpNo + 2).getImm();
1149 int LineSize = (KCacheMode == 1) ? 16 : 32;
1150 O << KCacheAddr * 16 << '-' << KCacheAddr * 16 + LineSize;
1154 void AMDGPUInstPrinter::printSendMsg(const MCInst *MI, unsigned OpNo,
1155 const MCSubtargetInfo &STI,
1157 using namespace llvm::AMDGPU::SendMsg;
1159 const unsigned SImm16 = MI->getOperand(OpNo).getImm();
1160 const unsigned Id = SImm16 & ID_MASK_;
1162 if (Id == ID_INTERRUPT) {
1163 if ((SImm16 & ~ID_MASK_) != 0) // Unused/unknown bits must be 0.
1165 O << "sendmsg(" << IdSymbolic[Id] << ')';
1168 if (Id == ID_GS || Id == ID_GS_DONE) {
1169 if ((SImm16 & ~(ID_MASK_|OP_GS_MASK_|STREAM_ID_MASK_)) != 0) // Unused/unknown bits must be 0.
1171 const unsigned OpGs = (SImm16 & OP_GS_MASK_) >> OP_SHIFT_;
1172 const unsigned StreamId = (SImm16 & STREAM_ID_MASK_) >> STREAM_ID_SHIFT_;
1173 if (OpGs == OP_GS_NOP && Id != ID_GS_DONE) // NOP to be used for GS_DONE only.
1175 if (OpGs == OP_GS_NOP && StreamId != 0) // NOP does not use/define stream id bits.
1177 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpGsSymbolic[OpGs];
1178 if (OpGs != OP_GS_NOP) { O << ", " << StreamId; }
1182 if (Id == ID_SYSMSG) {
1183 if ((SImm16 & ~(ID_MASK_|OP_SYS_MASK_)) != 0) // Unused/unknown bits must be 0.
1185 const unsigned OpSys = (SImm16 & OP_SYS_MASK_) >> OP_SHIFT_;
1186 if (! (OP_SYS_FIRST_ <= OpSys && OpSys < OP_SYS_LAST_)) // Unused/unknown.
1188 O << "sendmsg(" << IdSymbolic[Id] << ", " << OpSysSymbolic[OpSys] << ')';
1192 O << SImm16; // Unknown simm16 code.
1195 static void printSwizzleBitmask(const uint16_t AndMask,
1196 const uint16_t OrMask,
1197 const uint16_t XorMask,
1199 using namespace llvm::AMDGPU::Swizzle;
1201 uint16_t Probe0 = ((0 & AndMask) | OrMask) ^ XorMask;
1202 uint16_t Probe1 = ((BITMASK_MASK & AndMask) | OrMask) ^ XorMask;
1206 for (unsigned Mask = 1 << (BITMASK_WIDTH - 1); Mask > 0; Mask >>= 1) {
1207 uint16_t p0 = Probe0 & Mask;
1208 uint16_t p1 = Probe1 & Mask;
1228 void AMDGPUInstPrinter::printSwizzle(const MCInst *MI, unsigned OpNo,
1229 const MCSubtargetInfo &STI,
1231 using namespace llvm::AMDGPU::Swizzle;
1233 uint16_t Imm = MI->getOperand(OpNo).getImm();
1240 if ((Imm & QUAD_PERM_ENC_MASK) == QUAD_PERM_ENC) {
1242 O << "swizzle(" << IdSymbolic[ID_QUAD_PERM];
1243 for (auto i = 0; i < LANE_NUM; ++i) {
1245 O << formatDec(Imm & LANE_MASK);
1250 } else if ((Imm & BITMASK_PERM_ENC_MASK) == BITMASK_PERM_ENC) {
1252 uint16_t AndMask = (Imm >> BITMASK_AND_SHIFT) & BITMASK_MASK;
1253 uint16_t OrMask = (Imm >> BITMASK_OR_SHIFT) & BITMASK_MASK;
1254 uint16_t XorMask = (Imm >> BITMASK_XOR_SHIFT) & BITMASK_MASK;
1256 if (AndMask == BITMASK_MAX &&
1258 countPopulation(XorMask) == 1) {
1260 O << "swizzle(" << IdSymbolic[ID_SWAP];
1262 O << formatDec(XorMask);
1265 } else if (AndMask == BITMASK_MAX &&
1266 OrMask == 0 && XorMask > 0 &&
1267 isPowerOf2_64(XorMask + 1)) {
1269 O << "swizzle(" << IdSymbolic[ID_REVERSE];
1271 O << formatDec(XorMask + 1);
1276 uint16_t GroupSize = BITMASK_MAX - AndMask + 1;
1277 if (GroupSize > 1 &&
1278 isPowerOf2_64(GroupSize) &&
1279 OrMask < GroupSize &&
1282 O << "swizzle(" << IdSymbolic[ID_BROADCAST];
1284 O << formatDec(GroupSize);
1286 O << formatDec(OrMask);
1290 O << "swizzle(" << IdSymbolic[ID_BITMASK_PERM];
1292 printSwizzleBitmask(AndMask, OrMask, XorMask, O);
1297 printU16ImmDecOperand(MI, OpNo, O);
1301 void AMDGPUInstPrinter::printWaitFlag(const MCInst *MI, unsigned OpNo,
1302 const MCSubtargetInfo &STI,
1304 AMDGPU::IsaInfo::IsaVersion ISA =
1305 AMDGPU::IsaInfo::getIsaVersion(STI.getFeatureBits());
1307 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1308 unsigned Vmcnt, Expcnt, Lgkmcnt;
1309 decodeWaitcnt(ISA, SImm16, Vmcnt, Expcnt, Lgkmcnt);
1311 bool NeedSpace = false;
1313 if (Vmcnt != getVmcntBitMask(ISA)) {
1314 O << "vmcnt(" << Vmcnt << ')';
1318 if (Expcnt != getExpcntBitMask(ISA)) {
1321 O << "expcnt(" << Expcnt << ')';
1325 if (Lgkmcnt != getLgkmcntBitMask(ISA)) {
1328 O << "lgkmcnt(" << Lgkmcnt << ')';
1332 void AMDGPUInstPrinter::printHwreg(const MCInst *MI, unsigned OpNo,
1333 const MCSubtargetInfo &STI, raw_ostream &O) {
1334 using namespace llvm::AMDGPU::Hwreg;
1336 unsigned SImm16 = MI->getOperand(OpNo).getImm();
1337 const unsigned Id = (SImm16 & ID_MASK_) >> ID_SHIFT_;
1338 const unsigned Offset = (SImm16 & OFFSET_MASK_) >> OFFSET_SHIFT_;
1339 const unsigned Width = ((SImm16 & WIDTH_M1_MASK_) >> WIDTH_M1_SHIFT_) + 1;
1342 if (ID_SYMBOLIC_FIRST_ <= Id && Id < ID_SYMBOLIC_LAST_) {
1343 O << IdSymbolic[Id];
1347 if (Width != WIDTH_M1_DEFAULT_ + 1 || Offset != OFFSET_DEFAULT_) {
1348 O << ", " << Offset << ", " << Width;
1353 #include "AMDGPUGenAsmWriter.inc"