1 //===-- SIMCCodeEmitter.cpp - SI Code Emitter -----------------------------===//
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 /// The SI code emitter produces machine code that can be executed
12 /// directly on the GPU device.
14 //===----------------------------------------------------------------------===//
17 #include "MCTargetDesc/AMDGPUFixupKinds.h"
18 #include "MCTargetDesc/AMDGPUMCCodeEmitter.h"
19 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
20 #include "Utils/AMDGPUBaseInfo.h"
21 #include "llvm/MC/MCCodeEmitter.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCExpr.h"
24 #include "llvm/MC/MCFixup.h"
25 #include "llvm/MC/MCInst.h"
26 #include "llvm/MC/MCInstrDesc.h"
27 #include "llvm/MC/MCInstrInfo.h"
28 #include "llvm/MC/MCRegisterInfo.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/MC/MCSymbol.h"
31 #include "llvm/Support/Casting.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
43 class SIMCCodeEmitter : public AMDGPUMCCodeEmitter {
44 const MCRegisterInfo &MRI;
46 /// Encode an fp or int literal
47 uint32_t getLitEncoding(const MCOperand &MO, const MCOperandInfo &OpInfo,
48 const MCSubtargetInfo &STI) const;
51 SIMCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
53 : AMDGPUMCCodeEmitter(mcii), MRI(mri) {}
54 SIMCCodeEmitter(const SIMCCodeEmitter &) = delete;
55 SIMCCodeEmitter &operator=(const SIMCCodeEmitter &) = delete;
57 /// Encode the instruction and write it to the OS.
58 void encodeInstruction(const MCInst &MI, raw_ostream &OS,
59 SmallVectorImpl<MCFixup> &Fixups,
60 const MCSubtargetInfo &STI) const override;
62 /// \returns the encoding for an MCOperand.
63 uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
64 SmallVectorImpl<MCFixup> &Fixups,
65 const MCSubtargetInfo &STI) const override;
67 /// Use a fixup to encode the simm16 field for SOPP branch
69 unsigned getSOPPBrEncoding(const MCInst &MI, unsigned OpNo,
70 SmallVectorImpl<MCFixup> &Fixups,
71 const MCSubtargetInfo &STI) const override;
73 unsigned getSDWASrcEncoding(const MCInst &MI, unsigned OpNo,
74 SmallVectorImpl<MCFixup> &Fixups,
75 const MCSubtargetInfo &STI) const override;
77 unsigned getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo,
78 SmallVectorImpl<MCFixup> &Fixups,
79 const MCSubtargetInfo &STI) const override;
82 } // end anonymous namespace
84 MCCodeEmitter *llvm::createSIMCCodeEmitter(const MCInstrInfo &MCII,
85 const MCRegisterInfo &MRI,
87 return new SIMCCodeEmitter(MCII, MRI, Ctx);
90 // Returns the encoding value to use if the given integer is an integer inline
91 // immediate value, or 0 if it is not.
92 template <typename IntTy>
93 static uint32_t getIntInlineImmEncoding(IntTy Imm) {
94 if (Imm >= 0 && Imm <= 64)
97 if (Imm >= -16 && Imm <= -1)
98 return 192 + std::abs(Imm);
103 static uint32_t getLit16Encoding(uint16_t Val, const MCSubtargetInfo &STI) {
104 uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val));
108 if (Val == 0x3800) // 0.5
111 if (Val == 0xB800) // -0.5
114 if (Val == 0x3C00) // 1.0
117 if (Val == 0xBC00) // -1.0
120 if (Val == 0x4000) // 2.0
123 if (Val == 0xC000) // -2.0
126 if (Val == 0x4400) // 4.0
129 if (Val == 0xC400) // -4.0
132 if (Val == 0x3118 && // 1.0 / (2.0 * pi)
133 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
139 static uint32_t getLit32Encoding(uint32_t Val, const MCSubtargetInfo &STI) {
140 uint32_t IntImm = getIntInlineImmEncoding(static_cast<int32_t>(Val));
144 if (Val == FloatToBits(0.5f))
147 if (Val == FloatToBits(-0.5f))
150 if (Val == FloatToBits(1.0f))
153 if (Val == FloatToBits(-1.0f))
156 if (Val == FloatToBits(2.0f))
159 if (Val == FloatToBits(-2.0f))
162 if (Val == FloatToBits(4.0f))
165 if (Val == FloatToBits(-4.0f))
168 if (Val == 0x3e22f983 && // 1.0 / (2.0 * pi)
169 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
175 static uint32_t getLit64Encoding(uint64_t Val, const MCSubtargetInfo &STI) {
176 uint32_t IntImm = getIntInlineImmEncoding(static_cast<int64_t>(Val));
180 if (Val == DoubleToBits(0.5))
183 if (Val == DoubleToBits(-0.5))
186 if (Val == DoubleToBits(1.0))
189 if (Val == DoubleToBits(-1.0))
192 if (Val == DoubleToBits(2.0))
195 if (Val == DoubleToBits(-2.0))
198 if (Val == DoubleToBits(4.0))
201 if (Val == DoubleToBits(-4.0))
204 if (Val == 0x3fc45f306dc9c882 && // 1.0 / (2.0 * pi)
205 STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
211 uint32_t SIMCCodeEmitter::getLitEncoding(const MCOperand &MO,
212 const MCOperandInfo &OpInfo,
213 const MCSubtargetInfo &STI) const {
216 const auto *C = dyn_cast<MCConstantExpr>(MO.getExpr());
223 assert(!MO.isFPImm());
231 switch (OpInfo.OperandType) {
232 case AMDGPU::OPERAND_REG_IMM_INT32:
233 case AMDGPU::OPERAND_REG_IMM_FP32:
234 case AMDGPU::OPERAND_REG_INLINE_C_INT32:
235 case AMDGPU::OPERAND_REG_INLINE_C_FP32:
236 return getLit32Encoding(static_cast<uint32_t>(Imm), STI);
238 case AMDGPU::OPERAND_REG_IMM_INT64:
239 case AMDGPU::OPERAND_REG_IMM_FP64:
240 case AMDGPU::OPERAND_REG_INLINE_C_INT64:
241 case AMDGPU::OPERAND_REG_INLINE_C_FP64:
242 return getLit64Encoding(static_cast<uint64_t>(Imm), STI);
244 case AMDGPU::OPERAND_REG_IMM_INT16:
245 case AMDGPU::OPERAND_REG_IMM_FP16:
246 case AMDGPU::OPERAND_REG_INLINE_C_INT16:
247 case AMDGPU::OPERAND_REG_INLINE_C_FP16:
248 // FIXME Is this correct? What do inline immediates do on SI for f16 src
249 // which does not have f16 support?
250 return getLit16Encoding(static_cast<uint16_t>(Imm), STI);
252 case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
253 case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: {
254 uint16_t Lo16 = static_cast<uint16_t>(Imm);
255 uint32_t Encoding = getLit16Encoding(Lo16, STI);
259 llvm_unreachable("invalid operand size");
263 void SIMCCodeEmitter::encodeInstruction(const MCInst &MI, raw_ostream &OS,
264 SmallVectorImpl<MCFixup> &Fixups,
265 const MCSubtargetInfo &STI) const {
266 verifyInstructionPredicates(MI,
267 computeAvailableFeatures(STI.getFeatureBits()));
269 uint64_t Encoding = getBinaryCodeForInstr(MI, Fixups, STI);
270 const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
271 unsigned bytes = Desc.getSize();
273 for (unsigned i = 0; i < bytes; i++) {
274 OS.write((uint8_t) ((Encoding >> (8 * i)) & 0xff));
280 // Check for additional literals in SRC0/1/2 (Op 1/2/3)
281 for (unsigned i = 0, e = Desc.getNumOperands(); i < e; ++i) {
283 // Check if this operand should be encoded as [SV]Src
284 if (!AMDGPU::isSISrcOperand(Desc, i))
287 // Is this operand a literal immediate?
288 const MCOperand &Op = MI.getOperand(i);
289 if (getLitEncoding(Op, Desc.OpInfo[i], STI) != 255)
297 else if (Op.isExpr()) {
298 if (const auto *C = dyn_cast<MCConstantExpr>(Op.getExpr()))
301 } else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
302 llvm_unreachable("Must be immediate or expr");
304 for (unsigned j = 0; j < 4; j++) {
305 OS.write((uint8_t) ((Imm >> (8 * j)) & 0xff));
308 // Only one literal value allowed
313 unsigned SIMCCodeEmitter::getSOPPBrEncoding(const MCInst &MI, unsigned OpNo,
314 SmallVectorImpl<MCFixup> &Fixups,
315 const MCSubtargetInfo &STI) const {
316 const MCOperand &MO = MI.getOperand(OpNo);
319 const MCExpr *Expr = MO.getExpr();
320 MCFixupKind Kind = (MCFixupKind)AMDGPU::fixup_si_sopp_br;
321 Fixups.push_back(MCFixup::create(0, Expr, Kind, MI.getLoc()));
325 return getMachineOpValue(MI, MO, Fixups, STI);
329 SIMCCodeEmitter::getSDWASrcEncoding(const MCInst &MI, unsigned OpNo,
330 SmallVectorImpl<MCFixup> &Fixups,
331 const MCSubtargetInfo &STI) const {
332 using namespace AMDGPU::SDWA;
336 const MCOperand &MO = MI.getOperand(OpNo);
339 unsigned Reg = MO.getReg();
340 RegEnc |= MRI.getEncodingValue(Reg);
341 RegEnc &= SDWA9EncValues::SRC_VGPR_MASK;
342 if (AMDGPU::isSGPR(AMDGPU::mc2PseudoReg(Reg), &MRI)) {
343 RegEnc |= SDWA9EncValues::SRC_SGPR_MASK;
347 const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
348 uint32_t Enc = getLitEncoding(MO, Desc.OpInfo[OpNo], STI);
349 if (Enc != ~0U && Enc != 255) {
350 return Enc | SDWA9EncValues::SRC_SGPR_MASK;
354 llvm_unreachable("Unsupported operand kind");
359 SIMCCodeEmitter::getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo,
360 SmallVectorImpl<MCFixup> &Fixups,
361 const MCSubtargetInfo &STI) const {
362 using namespace AMDGPU::SDWA;
366 const MCOperand &MO = MI.getOperand(OpNo);
368 unsigned Reg = MO.getReg();
369 if (Reg != AMDGPU::VCC) {
370 RegEnc |= MRI.getEncodingValue(Reg);
371 RegEnc &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
372 RegEnc |= SDWA9EncValues::VOPC_DST_VCC_MASK;
377 static bool needsPCRel(const MCExpr *Expr) {
378 switch (Expr->getKind()) {
379 case MCExpr::SymbolRef:
381 case MCExpr::Binary: {
382 auto *BE = cast<MCBinaryExpr>(Expr);
383 if (BE->getOpcode() == MCBinaryExpr::Sub)
385 return needsPCRel(BE->getLHS()) || needsPCRel(BE->getRHS());
388 return needsPCRel(cast<MCUnaryExpr>(Expr)->getSubExpr());
390 case MCExpr::Constant:
393 llvm_unreachable("invalid kind");
396 uint64_t SIMCCodeEmitter::getMachineOpValue(const MCInst &MI,
398 SmallVectorImpl<MCFixup> &Fixups,
399 const MCSubtargetInfo &STI) const {
401 return MRI.getEncodingValue(MO.getReg());
403 if (MO.isExpr() && MO.getExpr()->getKind() != MCExpr::Constant) {
404 // FIXME: If this is expression is PCRel or not should not depend on what
405 // the expression looks like. Given that this is just a general expression,
406 // it should probably be FK_Data_4 and whatever is producing
408 // s_add_u32 s2, s2, (extern_const_addrspace+16
410 // And expecting a PCRel should instead produce
413 // s_add_u32 s2, s2, (extern_const_addrspace+16)-.Ltmp1
415 if (needsPCRel(MO.getExpr()))
419 Fixups.push_back(MCFixup::create(4, MO.getExpr(), Kind, MI.getLoc()));
422 // Figure out the operand number, needed for isSrcOperand check
424 for (unsigned e = MI.getNumOperands(); OpNo < e; ++OpNo) {
425 if (&MO == &MI.getOperand(OpNo))
429 const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
430 if (AMDGPU::isSISrcOperand(Desc, OpNo)) {
431 uint32_t Enc = getLitEncoding(MO, Desc.OpInfo[OpNo], STI);
432 if (Enc != ~0U && (Enc != 255 || Desc.getSize() == 4))
435 } else if (MO.isImm())
438 llvm_unreachable("Encoding of this operand type is not supported yet.");
442 #define ENABLE_INSTR_PREDICATE_VERIFIER
443 #include "AMDGPUGenMCCodeEmitter.inc"