1 | //===-- AMDGPUMCCodeEmitter.cpp - AMDGPU Code Emitter ---------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | /// \file |
10 | /// The AMDGPU code emitter produces machine code that can be executed |
11 | /// directly on the GPU device. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "MCTargetDesc/AMDGPUFixupKinds.h" |
16 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
17 | #include "SIDefines.h" |
18 | #include "Utils/AMDGPUBaseInfo.h" |
19 | #include "llvm/ADT/APInt.h" |
20 | #include "llvm/MC/MCCodeEmitter.h" |
21 | #include "llvm/MC/MCContext.h" |
22 | #include "llvm/MC/MCExpr.h" |
23 | #include "llvm/MC/MCInstrInfo.h" |
24 | #include "llvm/MC/MCRegisterInfo.h" |
25 | #include "llvm/MC/MCSubtargetInfo.h" |
26 | #include "llvm/Support/Casting.h" |
27 | #include "llvm/Support/EndianStream.h" |
28 | #include "llvm/TargetParser/SubtargetFeature.h" |
29 | #include <optional> |
30 | |
31 | using namespace llvm; |
32 | |
33 | namespace { |
34 | |
35 | class AMDGPUMCCodeEmitter : public MCCodeEmitter { |
36 | const MCRegisterInfo &MRI; |
37 | const MCInstrInfo &MCII; |
38 | |
39 | public: |
40 | AMDGPUMCCodeEmitter(const MCInstrInfo &MCII, const MCRegisterInfo &MRI) |
41 | : MRI(MRI), MCII(MCII) {} |
42 | |
43 | /// Encode the instruction and write it to the OS. |
44 | void encodeInstruction(const MCInst &MI, SmallVectorImpl<char> &CB, |
45 | SmallVectorImpl<MCFixup> &Fixups, |
46 | const MCSubtargetInfo &STI) const override; |
47 | |
48 | void getMachineOpValue(const MCInst &MI, const MCOperand &MO, APInt &Op, |
49 | SmallVectorImpl<MCFixup> &Fixups, |
50 | const MCSubtargetInfo &STI) const; |
51 | |
52 | void getMachineOpValueT16(const MCInst &MI, unsigned OpNo, APInt &Op, |
53 | SmallVectorImpl<MCFixup> &Fixups, |
54 | const MCSubtargetInfo &STI) const; |
55 | |
56 | void getMachineOpValueT16Lo128(const MCInst &MI, unsigned OpNo, APInt &Op, |
57 | SmallVectorImpl<MCFixup> &Fixups, |
58 | const MCSubtargetInfo &STI) const; |
59 | |
60 | /// Use a fixup to encode the simm16 field for SOPP branch |
61 | /// instructions. |
62 | void getSOPPBrEncoding(const MCInst &MI, unsigned OpNo, APInt &Op, |
63 | SmallVectorImpl<MCFixup> &Fixups, |
64 | const MCSubtargetInfo &STI) const; |
65 | |
66 | void getSMEMOffsetEncoding(const MCInst &MI, unsigned OpNo, APInt &Op, |
67 | SmallVectorImpl<MCFixup> &Fixups, |
68 | const MCSubtargetInfo &STI) const; |
69 | |
70 | void getSDWASrcEncoding(const MCInst &MI, unsigned OpNo, APInt &Op, |
71 | SmallVectorImpl<MCFixup> &Fixups, |
72 | const MCSubtargetInfo &STI) const; |
73 | |
74 | void getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo, APInt &Op, |
75 | SmallVectorImpl<MCFixup> &Fixups, |
76 | const MCSubtargetInfo &STI) const; |
77 | |
78 | void getAVOperandEncoding(const MCInst &MI, unsigned OpNo, APInt &Op, |
79 | SmallVectorImpl<MCFixup> &Fixups, |
80 | const MCSubtargetInfo &STI) const; |
81 | |
82 | private: |
83 | uint64_t getImplicitOpSelHiEncoding(int Opcode) const; |
84 | void getMachineOpValueCommon(const MCInst &MI, const MCOperand &MO, |
85 | unsigned OpNo, APInt &Op, |
86 | SmallVectorImpl<MCFixup> &Fixups, |
87 | const MCSubtargetInfo &STI) const; |
88 | |
89 | /// Encode an fp or int literal. |
90 | std::optional<uint32_t> getLitEncoding(const MCOperand &MO, |
91 | const MCOperandInfo &OpInfo, |
92 | const MCSubtargetInfo &STI) const; |
93 | |
94 | void getBinaryCodeForInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups, |
95 | APInt &Inst, APInt &Scratch, |
96 | const MCSubtargetInfo &STI) const; |
97 | }; |
98 | |
99 | } // end anonymous namespace |
100 | |
101 | MCCodeEmitter *llvm::createAMDGPUMCCodeEmitter(const MCInstrInfo &MCII, |
102 | MCContext &Ctx) { |
103 | return new AMDGPUMCCodeEmitter(MCII, *Ctx.getRegisterInfo()); |
104 | } |
105 | |
106 | // Returns the encoding value to use if the given integer is an integer inline |
107 | // immediate value, or 0 if it is not. |
108 | template <typename IntTy> |
109 | static uint32_t getIntInlineImmEncoding(IntTy Imm) { |
110 | if (Imm >= 0 && Imm <= 64) |
111 | return 128 + Imm; |
112 | |
113 | if (Imm >= -16 && Imm <= -1) |
114 | return 192 + std::abs(Imm); |
115 | |
116 | return 0; |
117 | } |
118 | |
119 | static uint32_t getLit16Encoding(uint16_t Val, const MCSubtargetInfo &STI) { |
120 | uint16_t IntImm = getIntInlineImmEncoding(Imm: static_cast<int16_t>(Val)); |
121 | if (IntImm != 0) |
122 | return IntImm; |
123 | |
124 | if (Val == 0x3800) // 0.5 |
125 | return 240; |
126 | |
127 | if (Val == 0xB800) // -0.5 |
128 | return 241; |
129 | |
130 | if (Val == 0x3C00) // 1.0 |
131 | return 242; |
132 | |
133 | if (Val == 0xBC00) // -1.0 |
134 | return 243; |
135 | |
136 | if (Val == 0x4000) // 2.0 |
137 | return 244; |
138 | |
139 | if (Val == 0xC000) // -2.0 |
140 | return 245; |
141 | |
142 | if (Val == 0x4400) // 4.0 |
143 | return 246; |
144 | |
145 | if (Val == 0xC400) // -4.0 |
146 | return 247; |
147 | |
148 | if (Val == 0x3118 && // 1.0 / (2.0 * pi) |
149 | STI.hasFeature(Feature: AMDGPU::FeatureInv2PiInlineImm)) |
150 | return 248; |
151 | |
152 | return 255; |
153 | } |
154 | |
155 | static uint32_t getLitBF16Encoding(uint16_t Val) { |
156 | uint16_t IntImm = getIntInlineImmEncoding(Imm: static_cast<int16_t>(Val)); |
157 | if (IntImm != 0) |
158 | return IntImm; |
159 | |
160 | // clang-format off |
161 | switch (Val) { |
162 | case 0x3F00: return 240; // 0.5 |
163 | case 0xBF00: return 241; // -0.5 |
164 | case 0x3F80: return 242; // 1.0 |
165 | case 0xBF80: return 243; // -1.0 |
166 | case 0x4000: return 244; // 2.0 |
167 | case 0xC000: return 245; // -2.0 |
168 | case 0x4080: return 246; // 4.0 |
169 | case 0xC080: return 247; // -4.0 |
170 | case 0x3E22: return 248; // 1.0 / (2.0 * pi) |
171 | default: return 255; |
172 | } |
173 | // clang-format on |
174 | } |
175 | |
176 | static uint32_t getLit32Encoding(uint32_t Val, const MCSubtargetInfo &STI) { |
177 | uint32_t IntImm = getIntInlineImmEncoding(Imm: static_cast<int32_t>(Val)); |
178 | if (IntImm != 0) |
179 | return IntImm; |
180 | |
181 | if (Val == llvm::bit_cast<uint32_t>(from: 0.5f)) |
182 | return 240; |
183 | |
184 | if (Val == llvm::bit_cast<uint32_t>(from: -0.5f)) |
185 | return 241; |
186 | |
187 | if (Val == llvm::bit_cast<uint32_t>(from: 1.0f)) |
188 | return 242; |
189 | |
190 | if (Val == llvm::bit_cast<uint32_t>(from: -1.0f)) |
191 | return 243; |
192 | |
193 | if (Val == llvm::bit_cast<uint32_t>(from: 2.0f)) |
194 | return 244; |
195 | |
196 | if (Val == llvm::bit_cast<uint32_t>(from: -2.0f)) |
197 | return 245; |
198 | |
199 | if (Val == llvm::bit_cast<uint32_t>(from: 4.0f)) |
200 | return 246; |
201 | |
202 | if (Val == llvm::bit_cast<uint32_t>(from: -4.0f)) |
203 | return 247; |
204 | |
205 | if (Val == 0x3e22f983 && // 1.0 / (2.0 * pi) |
206 | STI.hasFeature(Feature: AMDGPU::FeatureInv2PiInlineImm)) |
207 | return 248; |
208 | |
209 | return 255; |
210 | } |
211 | |
212 | static uint32_t getLit16IntEncoding(uint32_t Val, const MCSubtargetInfo &STI) { |
213 | return getLit32Encoding(Val, STI); |
214 | } |
215 | |
216 | static uint32_t getLit64Encoding(uint64_t Val, const MCSubtargetInfo &STI) { |
217 | uint32_t IntImm = getIntInlineImmEncoding(Imm: static_cast<int64_t>(Val)); |
218 | if (IntImm != 0) |
219 | return IntImm; |
220 | |
221 | if (Val == llvm::bit_cast<uint64_t>(from: 0.5)) |
222 | return 240; |
223 | |
224 | if (Val == llvm::bit_cast<uint64_t>(from: -0.5)) |
225 | return 241; |
226 | |
227 | if (Val == llvm::bit_cast<uint64_t>(from: 1.0)) |
228 | return 242; |
229 | |
230 | if (Val == llvm::bit_cast<uint64_t>(from: -1.0)) |
231 | return 243; |
232 | |
233 | if (Val == llvm::bit_cast<uint64_t>(from: 2.0)) |
234 | return 244; |
235 | |
236 | if (Val == llvm::bit_cast<uint64_t>(from: -2.0)) |
237 | return 245; |
238 | |
239 | if (Val == llvm::bit_cast<uint64_t>(from: 4.0)) |
240 | return 246; |
241 | |
242 | if (Val == llvm::bit_cast<uint64_t>(from: -4.0)) |
243 | return 247; |
244 | |
245 | if (Val == 0x3fc45f306dc9c882 && // 1.0 / (2.0 * pi) |
246 | STI.hasFeature(Feature: AMDGPU::FeatureInv2PiInlineImm)) |
247 | return 248; |
248 | |
249 | return 255; |
250 | } |
251 | |
252 | std::optional<uint32_t> |
253 | AMDGPUMCCodeEmitter::getLitEncoding(const MCOperand &MO, |
254 | const MCOperandInfo &OpInfo, |
255 | const MCSubtargetInfo &STI) const { |
256 | int64_t Imm; |
257 | if (MO.isExpr()) { |
258 | const auto *C = dyn_cast<MCConstantExpr>(Val: MO.getExpr()); |
259 | if (!C) |
260 | return 255; |
261 | |
262 | Imm = C->getValue(); |
263 | } else { |
264 | |
265 | assert(!MO.isDFPImm()); |
266 | |
267 | if (!MO.isImm()) |
268 | return {}; |
269 | |
270 | Imm = MO.getImm(); |
271 | } |
272 | |
273 | switch (OpInfo.OperandType) { |
274 | case AMDGPU::OPERAND_REG_IMM_INT32: |
275 | case AMDGPU::OPERAND_REG_IMM_FP32: |
276 | case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED: |
277 | case AMDGPU::OPERAND_REG_INLINE_C_INT32: |
278 | case AMDGPU::OPERAND_REG_INLINE_C_FP32: |
279 | case AMDGPU::OPERAND_REG_INLINE_AC_INT32: |
280 | case AMDGPU::OPERAND_REG_INLINE_AC_FP32: |
281 | case AMDGPU::OPERAND_REG_IMM_V2INT32: |
282 | case AMDGPU::OPERAND_REG_IMM_V2FP32: |
283 | case AMDGPU::OPERAND_REG_INLINE_C_V2INT32: |
284 | case AMDGPU::OPERAND_REG_INLINE_C_V2FP32: |
285 | case AMDGPU::OPERAND_INLINE_SPLIT_BARRIER_INT32: |
286 | return getLit32Encoding(Val: static_cast<uint32_t>(Imm), STI); |
287 | |
288 | case AMDGPU::OPERAND_REG_IMM_INT64: |
289 | case AMDGPU::OPERAND_REG_IMM_FP64: |
290 | case AMDGPU::OPERAND_REG_INLINE_C_INT64: |
291 | case AMDGPU::OPERAND_REG_INLINE_C_FP64: |
292 | case AMDGPU::OPERAND_REG_INLINE_AC_FP64: |
293 | return getLit64Encoding(Val: static_cast<uint64_t>(Imm), STI); |
294 | |
295 | case AMDGPU::OPERAND_REG_IMM_INT16: |
296 | case AMDGPU::OPERAND_REG_INLINE_C_INT16: |
297 | case AMDGPU::OPERAND_REG_INLINE_AC_INT16: |
298 | return getLit16IntEncoding(Val: static_cast<uint32_t>(Imm), STI); |
299 | |
300 | case AMDGPU::OPERAND_REG_IMM_FP16: |
301 | case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED: |
302 | case AMDGPU::OPERAND_REG_INLINE_C_FP16: |
303 | case AMDGPU::OPERAND_REG_INLINE_AC_FP16: |
304 | // FIXME Is this correct? What do inline immediates do on SI for f16 src |
305 | // which does not have f16 support? |
306 | return getLit16Encoding(Val: static_cast<uint16_t>(Imm), STI); |
307 | |
308 | case AMDGPU::OPERAND_REG_IMM_BF16: |
309 | case AMDGPU::OPERAND_REG_IMM_BF16_DEFERRED: |
310 | case AMDGPU::OPERAND_REG_INLINE_C_BF16: |
311 | case AMDGPU::OPERAND_REG_INLINE_AC_BF16: |
312 | // We don't actually need to check Inv2Pi here because BF16 instructions can |
313 | // only be emitted for targets that already support the feature. |
314 | return getLitBF16Encoding(Val: static_cast<uint16_t>(Imm)); |
315 | |
316 | case AMDGPU::OPERAND_REG_IMM_V2INT16: |
317 | case AMDGPU::OPERAND_REG_INLINE_C_V2INT16: |
318 | case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16: |
319 | return AMDGPU::getInlineEncodingV2I16(Literal: static_cast<uint32_t>(Imm)) |
320 | .value_or(u: 255); |
321 | |
322 | case AMDGPU::OPERAND_REG_IMM_V2FP16: |
323 | case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: |
324 | case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: |
325 | return AMDGPU::getInlineEncodingV2F16(Literal: static_cast<uint32_t>(Imm)) |
326 | .value_or(u: 255); |
327 | |
328 | case AMDGPU::OPERAND_REG_IMM_V2BF16: |
329 | case AMDGPU::OPERAND_REG_INLINE_C_V2BF16: |
330 | case AMDGPU::OPERAND_REG_INLINE_AC_V2BF16: |
331 | return AMDGPU::getInlineEncodingV2BF16(Literal: static_cast<uint32_t>(Imm)) |
332 | .value_or(u: 255); |
333 | |
334 | case AMDGPU::OPERAND_KIMM32: |
335 | case AMDGPU::OPERAND_KIMM16: |
336 | return MO.getImm(); |
337 | default: |
338 | llvm_unreachable("invalid operand size" ); |
339 | } |
340 | } |
341 | |
342 | uint64_t AMDGPUMCCodeEmitter::getImplicitOpSelHiEncoding(int Opcode) const { |
343 | using namespace AMDGPU::VOP3PEncoding; |
344 | using namespace AMDGPU::OpName; |
345 | |
346 | if (AMDGPU::hasNamedOperand(Opcode, NamedIdx: op_sel_hi)) { |
347 | if (AMDGPU::hasNamedOperand(Opcode, NamedIdx: src2)) |
348 | return 0; |
349 | if (AMDGPU::hasNamedOperand(Opcode, NamedIdx: src1)) |
350 | return OP_SEL_HI_2; |
351 | if (AMDGPU::hasNamedOperand(Opcode, NamedIdx: src0)) |
352 | return OP_SEL_HI_1 | OP_SEL_HI_2; |
353 | } |
354 | return OP_SEL_HI_0 | OP_SEL_HI_1 | OP_SEL_HI_2; |
355 | } |
356 | |
357 | static bool isVCMPX64(const MCInstrDesc &Desc) { |
358 | return (Desc.TSFlags & SIInstrFlags::VOP3) && |
359 | Desc.hasImplicitDefOfPhysReg(Reg: AMDGPU::EXEC); |
360 | } |
361 | |
362 | void AMDGPUMCCodeEmitter::encodeInstruction(const MCInst &MI, |
363 | SmallVectorImpl<char> &CB, |
364 | SmallVectorImpl<MCFixup> &Fixups, |
365 | const MCSubtargetInfo &STI) const { |
366 | int Opcode = MI.getOpcode(); |
367 | APInt Encoding, Scratch; |
368 | getBinaryCodeForInstr(MI, Fixups, Inst&: Encoding, Scratch, STI); |
369 | const MCInstrDesc &Desc = MCII.get(Opcode: MI.getOpcode()); |
370 | unsigned bytes = Desc.getSize(); |
371 | |
372 | // Set unused op_sel_hi bits to 1 for VOP3P and MAI instructions. |
373 | // Note that accvgpr_read/write are MAI, have src0, but do not use op_sel. |
374 | if ((Desc.TSFlags & SIInstrFlags::VOP3P) || |
375 | Opcode == AMDGPU::V_ACCVGPR_READ_B32_vi || |
376 | Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_vi) { |
377 | Encoding |= getImplicitOpSelHiEncoding(Opcode); |
378 | } |
379 | |
380 | // GFX10+ v_cmpx opcodes promoted to VOP3 have implied dst=EXEC. |
381 | // Documentation requires dst to be encoded as EXEC (0x7E), |
382 | // but it looks like the actual value encoded for dst operand |
383 | // is ignored by HW. It was decided to define dst as "do not care" |
384 | // in td files to allow disassembler accept any dst value. |
385 | // However, dst is encoded as EXEC for compatibility with SP3. |
386 | if (AMDGPU::isGFX10Plus(STI) && isVCMPX64(Desc)) { |
387 | assert((Encoding & 0xFF) == 0); |
388 | Encoding |= MRI.getEncodingValue(RegNo: AMDGPU::EXEC_LO) & |
389 | AMDGPU::HWEncoding::REG_IDX_MASK; |
390 | } |
391 | |
392 | for (unsigned i = 0; i < bytes; i++) { |
393 | CB.push_back(Elt: (uint8_t)Encoding.extractBitsAsZExtValue(numBits: 8, bitPosition: 8 * i)); |
394 | } |
395 | |
396 | // NSA encoding. |
397 | if (AMDGPU::isGFX10Plus(STI) && Desc.TSFlags & SIInstrFlags::MIMG) { |
398 | int vaddr0 = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), |
399 | NamedIdx: AMDGPU::OpName::vaddr0); |
400 | int srsrc = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), |
401 | NamedIdx: AMDGPU::OpName::srsrc); |
402 | assert(vaddr0 >= 0 && srsrc > vaddr0); |
403 | unsigned = srsrc - vaddr0 - 1; |
404 | unsigned NumPadding = (-NumExtraAddrs) & 3; |
405 | |
406 | for (unsigned i = 0; i < NumExtraAddrs; ++i) { |
407 | getMachineOpValue(MI, MO: MI.getOperand(i: vaddr0 + 1 + i), Op&: Encoding, Fixups, |
408 | STI); |
409 | CB.push_back(Elt: (uint8_t)Encoding.getLimitedValue()); |
410 | } |
411 | CB.append(NumInputs: NumPadding, Elt: 0); |
412 | } |
413 | |
414 | if ((bytes > 8 && STI.hasFeature(Feature: AMDGPU::FeatureVOP3Literal)) || |
415 | (bytes > 4 && !STI.hasFeature(Feature: AMDGPU::FeatureVOP3Literal))) |
416 | return; |
417 | |
418 | // Do not print literals from SISrc Operands for insts with mandatory literals |
419 | if (AMDGPU::hasNamedOperand(Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::imm)) |
420 | return; |
421 | |
422 | // Check for additional literals |
423 | for (unsigned i = 0, e = Desc.getNumOperands(); i < e; ++i) { |
424 | |
425 | // Check if this operand should be encoded as [SV]Src |
426 | if (!AMDGPU::isSISrcOperand(Desc, OpNo: i)) |
427 | continue; |
428 | |
429 | // Is this operand a literal immediate? |
430 | const MCOperand &Op = MI.getOperand(i); |
431 | auto Enc = getLitEncoding(MO: Op, OpInfo: Desc.operands()[i], STI); |
432 | if (!Enc || *Enc != 255) |
433 | continue; |
434 | |
435 | // Yes! Encode it |
436 | int64_t Imm = 0; |
437 | |
438 | if (Op.isImm()) |
439 | Imm = Op.getImm(); |
440 | else if (Op.isExpr()) { |
441 | if (const auto *C = dyn_cast<MCConstantExpr>(Val: Op.getExpr())) |
442 | Imm = C->getValue(); |
443 | } else // Exprs will be replaced with a fixup value. |
444 | llvm_unreachable("Must be immediate or expr" ); |
445 | |
446 | if (Desc.operands()[i].OperandType == AMDGPU::OPERAND_REG_IMM_FP64) |
447 | Imm = Hi_32(Value: Imm); |
448 | |
449 | support::endian::write<uint32_t>(Out&: CB, V: Imm, E: llvm::endianness::little); |
450 | |
451 | // Only one literal value allowed |
452 | break; |
453 | } |
454 | } |
455 | |
456 | void AMDGPUMCCodeEmitter::getSOPPBrEncoding(const MCInst &MI, unsigned OpNo, |
457 | APInt &Op, |
458 | SmallVectorImpl<MCFixup> &Fixups, |
459 | const MCSubtargetInfo &STI) const { |
460 | const MCOperand &MO = MI.getOperand(i: OpNo); |
461 | |
462 | if (MO.isExpr()) { |
463 | const MCExpr *Expr = MO.getExpr(); |
464 | MCFixupKind Kind = (MCFixupKind)AMDGPU::fixup_si_sopp_br; |
465 | Fixups.push_back(Elt: MCFixup::create(Offset: 0, Value: Expr, Kind, Loc: MI.getLoc())); |
466 | Op = APInt::getZero(numBits: 96); |
467 | } else { |
468 | getMachineOpValue(MI, MO, Op, Fixups, STI); |
469 | } |
470 | } |
471 | |
472 | void AMDGPUMCCodeEmitter::getSMEMOffsetEncoding( |
473 | const MCInst &MI, unsigned OpNo, APInt &Op, |
474 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
475 | auto Offset = MI.getOperand(i: OpNo).getImm(); |
476 | // VI only supports 20-bit unsigned offsets. |
477 | assert(!AMDGPU::isVI(STI) || isUInt<20>(Offset)); |
478 | Op = Offset; |
479 | } |
480 | |
481 | void AMDGPUMCCodeEmitter::getSDWASrcEncoding(const MCInst &MI, unsigned OpNo, |
482 | APInt &Op, |
483 | SmallVectorImpl<MCFixup> &Fixups, |
484 | const MCSubtargetInfo &STI) const { |
485 | using namespace AMDGPU::SDWA; |
486 | |
487 | uint64_t RegEnc = 0; |
488 | |
489 | const MCOperand &MO = MI.getOperand(i: OpNo); |
490 | |
491 | if (MO.isReg()) { |
492 | unsigned Reg = MO.getReg(); |
493 | RegEnc |= MRI.getEncodingValue(RegNo: Reg); |
494 | RegEnc &= SDWA9EncValues::SRC_VGPR_MASK; |
495 | if (AMDGPU::isSGPR(Reg: AMDGPU::mc2PseudoReg(Reg), TRI: &MRI)) { |
496 | RegEnc |= SDWA9EncValues::SRC_SGPR_MASK; |
497 | } |
498 | Op = RegEnc; |
499 | return; |
500 | } else { |
501 | const MCInstrDesc &Desc = MCII.get(Opcode: MI.getOpcode()); |
502 | auto Enc = getLitEncoding(MO, OpInfo: Desc.operands()[OpNo], STI); |
503 | if (Enc && *Enc != 255) { |
504 | Op = *Enc | SDWA9EncValues::SRC_SGPR_MASK; |
505 | return; |
506 | } |
507 | } |
508 | |
509 | llvm_unreachable("Unsupported operand kind" ); |
510 | } |
511 | |
512 | void AMDGPUMCCodeEmitter::getSDWAVopcDstEncoding( |
513 | const MCInst &MI, unsigned OpNo, APInt &Op, |
514 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
515 | using namespace AMDGPU::SDWA; |
516 | |
517 | uint64_t RegEnc = 0; |
518 | |
519 | const MCOperand &MO = MI.getOperand(i: OpNo); |
520 | |
521 | unsigned Reg = MO.getReg(); |
522 | if (Reg != AMDGPU::VCC && Reg != AMDGPU::VCC_LO) { |
523 | RegEnc |= MRI.getEncodingValue(RegNo: Reg); |
524 | RegEnc &= SDWA9EncValues::VOPC_DST_SGPR_MASK; |
525 | RegEnc |= SDWA9EncValues::VOPC_DST_VCC_MASK; |
526 | } |
527 | Op = RegEnc; |
528 | } |
529 | |
530 | void AMDGPUMCCodeEmitter::getAVOperandEncoding( |
531 | const MCInst &MI, unsigned OpNo, APInt &Op, |
532 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
533 | unsigned Reg = MI.getOperand(i: OpNo).getReg(); |
534 | unsigned Enc = MRI.getEncodingValue(RegNo: Reg); |
535 | unsigned Idx = Enc & AMDGPU::HWEncoding::REG_IDX_MASK; |
536 | bool IsVGPROrAGPR = Enc & AMDGPU::HWEncoding::IS_VGPR_OR_AGPR; |
537 | |
538 | // VGPR and AGPR have the same encoding, but SrcA and SrcB operands of mfma |
539 | // instructions use acc[0:1] modifier bits to distinguish. These bits are |
540 | // encoded as a virtual 9th bit of the register for these operands. |
541 | bool IsAGPR = false; |
542 | if (MRI.getRegClass(i: AMDGPU::AGPR_32RegClassID).contains(Reg) || |
543 | MRI.getRegClass(i: AMDGPU::AReg_64RegClassID).contains(Reg) || |
544 | MRI.getRegClass(i: AMDGPU::AReg_96RegClassID).contains(Reg) || |
545 | MRI.getRegClass(i: AMDGPU::AReg_128RegClassID).contains(Reg) || |
546 | MRI.getRegClass(i: AMDGPU::AReg_160RegClassID).contains(Reg) || |
547 | MRI.getRegClass(i: AMDGPU::AReg_192RegClassID).contains(Reg) || |
548 | MRI.getRegClass(i: AMDGPU::AReg_224RegClassID).contains(Reg) || |
549 | MRI.getRegClass(i: AMDGPU::AReg_256RegClassID).contains(Reg) || |
550 | MRI.getRegClass(i: AMDGPU::AReg_288RegClassID).contains(Reg) || |
551 | MRI.getRegClass(i: AMDGPU::AReg_320RegClassID).contains(Reg) || |
552 | MRI.getRegClass(i: AMDGPU::AReg_352RegClassID).contains(Reg) || |
553 | MRI.getRegClass(i: AMDGPU::AReg_384RegClassID).contains(Reg) || |
554 | MRI.getRegClass(i: AMDGPU::AReg_512RegClassID).contains(Reg) || |
555 | MRI.getRegClass(i: AMDGPU::AGPR_LO16RegClassID).contains(Reg)) |
556 | IsAGPR = true; |
557 | |
558 | Op = Idx | (IsVGPROrAGPR << 8) | (IsAGPR << 9); |
559 | } |
560 | |
561 | static bool needsPCRel(const MCExpr *Expr) { |
562 | switch (Expr->getKind()) { |
563 | case MCExpr::SymbolRef: { |
564 | auto *SE = cast<MCSymbolRefExpr>(Val: Expr); |
565 | MCSymbolRefExpr::VariantKind Kind = SE->getKind(); |
566 | return Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_LO && |
567 | Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_HI; |
568 | } |
569 | case MCExpr::Binary: { |
570 | auto *BE = cast<MCBinaryExpr>(Val: Expr); |
571 | if (BE->getOpcode() == MCBinaryExpr::Sub) |
572 | return false; |
573 | return needsPCRel(Expr: BE->getLHS()) || needsPCRel(Expr: BE->getRHS()); |
574 | } |
575 | case MCExpr::Unary: |
576 | return needsPCRel(Expr: cast<MCUnaryExpr>(Val: Expr)->getSubExpr()); |
577 | case MCExpr::Target: |
578 | case MCExpr::Constant: |
579 | return false; |
580 | } |
581 | llvm_unreachable("invalid kind" ); |
582 | } |
583 | |
584 | void AMDGPUMCCodeEmitter::getMachineOpValue(const MCInst &MI, |
585 | const MCOperand &MO, APInt &Op, |
586 | SmallVectorImpl<MCFixup> &Fixups, |
587 | const MCSubtargetInfo &STI) const { |
588 | if (MO.isReg()){ |
589 | unsigned Enc = MRI.getEncodingValue(RegNo: MO.getReg()); |
590 | unsigned Idx = Enc & AMDGPU::HWEncoding::REG_IDX_MASK; |
591 | bool IsVGPR = Enc & AMDGPU::HWEncoding::IS_VGPR_OR_AGPR; |
592 | Op = Idx | (IsVGPR << 8); |
593 | return; |
594 | } |
595 | unsigned OpNo = &MO - MI.begin(); |
596 | getMachineOpValueCommon(MI, MO, OpNo, Op, Fixups, STI); |
597 | } |
598 | |
599 | void AMDGPUMCCodeEmitter::getMachineOpValueT16( |
600 | const MCInst &MI, unsigned OpNo, APInt &Op, |
601 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
602 | const MCOperand &MO = MI.getOperand(i: OpNo); |
603 | if (MO.isReg()) { |
604 | unsigned Enc = MRI.getEncodingValue(RegNo: MO.getReg()); |
605 | unsigned Idx = Enc & AMDGPU::HWEncoding::REG_IDX_MASK; |
606 | bool IsVGPR = Enc & AMDGPU::HWEncoding::IS_VGPR_OR_AGPR; |
607 | Op = Idx | (IsVGPR << 8); |
608 | return; |
609 | } |
610 | getMachineOpValueCommon(MI, MO, OpNo, Op, Fixups, STI); |
611 | // VGPRs include the suffix/op_sel bit in the register encoding, but |
612 | // immediates and SGPRs include it in src_modifiers. Therefore, copy the |
613 | // op_sel bit from the src operands into src_modifier operands if Op is |
614 | // src_modifiers and the corresponding src is a VGPR |
615 | int SrcMOIdx = -1; |
616 | assert(OpNo < INT_MAX); |
617 | if ((int)OpNo == AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), |
618 | NamedIdx: AMDGPU::OpName::src0_modifiers)) { |
619 | SrcMOIdx = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::src0); |
620 | int VDstMOIdx = |
621 | AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::vdst); |
622 | if (VDstMOIdx != -1) { |
623 | auto DstReg = MI.getOperand(i: VDstMOIdx).getReg(); |
624 | if (AMDGPU::isHi(Reg: DstReg, MRI)) |
625 | Op |= SISrcMods::DST_OP_SEL; |
626 | } |
627 | } else if ((int)OpNo == AMDGPU::getNamedOperandIdx( |
628 | Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::src1_modifiers)) |
629 | SrcMOIdx = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::src1); |
630 | else if ((int)OpNo == AMDGPU::getNamedOperandIdx( |
631 | Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::src2_modifiers)) |
632 | SrcMOIdx = AMDGPU::getNamedOperandIdx(Opcode: MI.getOpcode(), NamedIdx: AMDGPU::OpName::src2); |
633 | if (SrcMOIdx == -1) |
634 | return; |
635 | |
636 | const MCOperand &SrcMO = MI.getOperand(i: SrcMOIdx); |
637 | if (!SrcMO.isReg()) |
638 | return; |
639 | auto SrcReg = SrcMO.getReg(); |
640 | if (AMDGPU::isSGPR(Reg: SrcReg, TRI: &MRI)) |
641 | return; |
642 | if (AMDGPU::isHi(Reg: SrcReg, MRI)) |
643 | Op |= SISrcMods::OP_SEL_0; |
644 | } |
645 | |
646 | void AMDGPUMCCodeEmitter::getMachineOpValueT16Lo128( |
647 | const MCInst &MI, unsigned OpNo, APInt &Op, |
648 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
649 | const MCOperand &MO = MI.getOperand(i: OpNo); |
650 | if (MO.isReg()) { |
651 | uint16_t Encoding = MRI.getEncodingValue(RegNo: MO.getReg()); |
652 | unsigned RegIdx = Encoding & AMDGPU::HWEncoding::REG_IDX_MASK; |
653 | bool IsHi = Encoding & AMDGPU::HWEncoding::IS_HI; |
654 | bool IsVGPR = Encoding & AMDGPU::HWEncoding::IS_VGPR_OR_AGPR; |
655 | assert((!IsVGPR || isUInt<7>(RegIdx)) && "VGPR0-VGPR127 expected!" ); |
656 | Op = (IsVGPR ? 0x100 : 0) | (IsHi ? 0x80 : 0) | RegIdx; |
657 | return; |
658 | } |
659 | getMachineOpValueCommon(MI, MO, OpNo, Op, Fixups, STI); |
660 | } |
661 | |
662 | void AMDGPUMCCodeEmitter::getMachineOpValueCommon( |
663 | const MCInst &MI, const MCOperand &MO, unsigned OpNo, APInt &Op, |
664 | SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const { |
665 | int64_t Val; |
666 | if (MO.isExpr() && MO.getExpr()->evaluateAsAbsolute(Res&: Val)) { |
667 | Op = Val; |
668 | return; |
669 | } |
670 | |
671 | if (MO.isExpr() && MO.getExpr()->getKind() != MCExpr::Constant) { |
672 | // FIXME: If this is expression is PCRel or not should not depend on what |
673 | // the expression looks like. Given that this is just a general expression, |
674 | // it should probably be FK_Data_4 and whatever is producing |
675 | // |
676 | // s_add_u32 s2, s2, (extern_const_addrspace+16 |
677 | // |
678 | // And expecting a PCRel should instead produce |
679 | // |
680 | // .Ltmp1: |
681 | // s_add_u32 s2, s2, (extern_const_addrspace+16)-.Ltmp1 |
682 | MCFixupKind Kind; |
683 | if (needsPCRel(Expr: MO.getExpr())) |
684 | Kind = FK_PCRel_4; |
685 | else |
686 | Kind = FK_Data_4; |
687 | |
688 | const MCInstrDesc &Desc = MCII.get(Opcode: MI.getOpcode()); |
689 | uint32_t Offset = Desc.getSize(); |
690 | assert(Offset == 4 || Offset == 8); |
691 | |
692 | Fixups.push_back(Elt: MCFixup::create(Offset, Value: MO.getExpr(), Kind, Loc: MI.getLoc())); |
693 | } |
694 | |
695 | const MCInstrDesc &Desc = MCII.get(Opcode: MI.getOpcode()); |
696 | if (AMDGPU::isSISrcOperand(Desc, OpNo)) { |
697 | if (auto Enc = getLitEncoding(MO, OpInfo: Desc.operands()[OpNo], STI)) { |
698 | Op = *Enc; |
699 | return; |
700 | } |
701 | } else if (MO.isImm()) { |
702 | Op = MO.getImm(); |
703 | return; |
704 | } |
705 | |
706 | llvm_unreachable("Encoding of this operand type is not supported yet." ); |
707 | } |
708 | |
709 | #include "AMDGPUGenMCCodeEmitter.inc" |
710 | |