1 //===-- SIMCCodeEmitter.cpp - SI 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 SI 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/AMDGPUMCCodeEmitter.h"
17 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
18 #include "SIDefines.h"
19 #include "Utils/AMDGPUBaseInfo.h"
20 #include "llvm/ADT/APInt.h"
21 #include "llvm/MC/MCCodeEmitter.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCExpr.h"
24 #include "llvm/MC/MCInstrInfo.h"
25 #include "llvm/MC/MCRegisterInfo.h"
26 #include "llvm/MC/MCSubtargetInfo.h"
27 #include "llvm/MC/SubtargetFeature.h"
28 #include "llvm/Support/Casting.h"
29 #include <optional>
30 
31 using namespace llvm;
32 
33 namespace {
34 
35 class SIMCCodeEmitter : public  AMDGPUMCCodeEmitter {
36   const MCRegisterInfo &MRI;
37 
38   /// Encode an fp or int literal
39   std::optional<uint32_t> getLitEncoding(const MCOperand &MO,
40                                          const MCOperandInfo &OpInfo,
41                                          const MCSubtargetInfo &STI) const;
42 
43 public:
SIMCCodeEmitter(const MCInstrInfo & mcii,MCContext & ctx)44   SIMCCodeEmitter(const MCInstrInfo &mcii, MCContext &ctx)
45       : AMDGPUMCCodeEmitter(mcii), MRI(*ctx.getRegisterInfo()) {}
46   SIMCCodeEmitter(const SIMCCodeEmitter &) = delete;
47   SIMCCodeEmitter &operator=(const SIMCCodeEmitter &) = delete;
48 
49   /// Encode the instruction and write it to the OS.
50   void encodeInstruction(const MCInst &MI, raw_ostream &OS,
51                          SmallVectorImpl<MCFixup> &Fixups,
52                          const MCSubtargetInfo &STI) const override;
53 
54   void getMachineOpValue(const MCInst &MI, const MCOperand &MO, APInt &Op,
55                          SmallVectorImpl<MCFixup> &Fixups,
56                          const MCSubtargetInfo &STI) const override;
57 
58   /// Use a fixup to encode the simm16 field for SOPP branch
59   ///        instructions.
60   void getSOPPBrEncoding(const MCInst &MI, unsigned OpNo, APInt &Op,
61                          SmallVectorImpl<MCFixup> &Fixups,
62                          const MCSubtargetInfo &STI) const override;
63 
64   void getSMEMOffsetEncoding(const MCInst &MI, unsigned OpNo, APInt &Op,
65                              SmallVectorImpl<MCFixup> &Fixups,
66                              const MCSubtargetInfo &STI) const override;
67 
68   void getSDWASrcEncoding(const MCInst &MI, unsigned OpNo, APInt &Op,
69                           SmallVectorImpl<MCFixup> &Fixups,
70                           const MCSubtargetInfo &STI) const override;
71 
72   void getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo, APInt &Op,
73                               SmallVectorImpl<MCFixup> &Fixups,
74                               const MCSubtargetInfo &STI) const override;
75 
76   void getAVOperandEncoding(const MCInst &MI, unsigned OpNo, APInt &Op,
77                             SmallVectorImpl<MCFixup> &Fixups,
78                             const MCSubtargetInfo &STI) const override;
79 
80 private:
81   uint64_t getImplicitOpSelHiEncoding(int Opcode) const;
82   void getMachineOpValueCommon(const MCInst &MI, const MCOperand &MO,
83                                unsigned OpNo, APInt &Op,
84                                SmallVectorImpl<MCFixup> &Fixups,
85                                const MCSubtargetInfo &STI) const;
86 };
87 
88 } // end anonymous namespace
89 
createSIMCCodeEmitter(const MCInstrInfo & MCII,MCContext & Ctx)90 MCCodeEmitter *llvm::createSIMCCodeEmitter(const MCInstrInfo &MCII,
91                                            MCContext &Ctx) {
92   return new SIMCCodeEmitter(MCII, Ctx);
93 }
94 
95 // Returns the encoding value to use if the given integer is an integer inline
96 // immediate value, or 0 if it is not.
97 template <typename IntTy>
getIntInlineImmEncoding(IntTy Imm)98 static uint32_t getIntInlineImmEncoding(IntTy Imm) {
99   if (Imm >= 0 && Imm <= 64)
100     return 128 + Imm;
101 
102   if (Imm >= -16 && Imm <= -1)
103     return 192 + std::abs(Imm);
104 
105   return 0;
106 }
107 
getLit16IntEncoding(uint16_t Val,const MCSubtargetInfo & STI)108 static uint32_t getLit16IntEncoding(uint16_t Val, const MCSubtargetInfo &STI) {
109   uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val));
110   return IntImm == 0 ? 255 : IntImm;
111 }
112 
getLit16Encoding(uint16_t Val,const MCSubtargetInfo & STI)113 static uint32_t getLit16Encoding(uint16_t Val, const MCSubtargetInfo &STI) {
114   uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val));
115   if (IntImm != 0)
116     return IntImm;
117 
118   if (Val == 0x3800) // 0.5
119     return 240;
120 
121   if (Val == 0xB800) // -0.5
122     return 241;
123 
124   if (Val == 0x3C00) // 1.0
125     return 242;
126 
127   if (Val == 0xBC00) // -1.0
128     return 243;
129 
130   if (Val == 0x4000) // 2.0
131     return 244;
132 
133   if (Val == 0xC000) // -2.0
134     return 245;
135 
136   if (Val == 0x4400) // 4.0
137     return 246;
138 
139   if (Val == 0xC400) // -4.0
140     return 247;
141 
142   if (Val == 0x3118 && // 1.0 / (2.0 * pi)
143       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
144     return 248;
145 
146   return 255;
147 }
148 
getLit32Encoding(uint32_t Val,const MCSubtargetInfo & STI)149 static uint32_t getLit32Encoding(uint32_t Val, const MCSubtargetInfo &STI) {
150   uint32_t IntImm = getIntInlineImmEncoding(static_cast<int32_t>(Val));
151   if (IntImm != 0)
152     return IntImm;
153 
154   if (Val == FloatToBits(0.5f))
155     return 240;
156 
157   if (Val == FloatToBits(-0.5f))
158     return 241;
159 
160   if (Val == FloatToBits(1.0f))
161     return 242;
162 
163   if (Val == FloatToBits(-1.0f))
164     return 243;
165 
166   if (Val == FloatToBits(2.0f))
167     return 244;
168 
169   if (Val == FloatToBits(-2.0f))
170     return 245;
171 
172   if (Val == FloatToBits(4.0f))
173     return 246;
174 
175   if (Val == FloatToBits(-4.0f))
176     return 247;
177 
178   if (Val == 0x3e22f983 && // 1.0 / (2.0 * pi)
179       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
180     return 248;
181 
182   return 255;
183 }
184 
getLit64Encoding(uint64_t Val,const MCSubtargetInfo & STI)185 static uint32_t getLit64Encoding(uint64_t Val, const MCSubtargetInfo &STI) {
186   uint32_t IntImm = getIntInlineImmEncoding(static_cast<int64_t>(Val));
187   if (IntImm != 0)
188     return IntImm;
189 
190   if (Val == DoubleToBits(0.5))
191     return 240;
192 
193   if (Val == DoubleToBits(-0.5))
194     return 241;
195 
196   if (Val == DoubleToBits(1.0))
197     return 242;
198 
199   if (Val == DoubleToBits(-1.0))
200     return 243;
201 
202   if (Val == DoubleToBits(2.0))
203     return 244;
204 
205   if (Val == DoubleToBits(-2.0))
206     return 245;
207 
208   if (Val == DoubleToBits(4.0))
209     return 246;
210 
211   if (Val == DoubleToBits(-4.0))
212     return 247;
213 
214   if (Val == 0x3fc45f306dc9c882 && // 1.0 / (2.0 * pi)
215       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
216     return 248;
217 
218   return 255;
219 }
220 
221 std::optional<uint32_t>
getLitEncoding(const MCOperand & MO,const MCOperandInfo & OpInfo,const MCSubtargetInfo & STI) const222 SIMCCodeEmitter::getLitEncoding(const MCOperand &MO,
223                                 const MCOperandInfo &OpInfo,
224                                 const MCSubtargetInfo &STI) const {
225   int64_t Imm;
226   if (MO.isExpr()) {
227     const auto *C = dyn_cast<MCConstantExpr>(MO.getExpr());
228     if (!C)
229       return 255;
230 
231     Imm = C->getValue();
232   } else {
233 
234     assert(!MO.isDFPImm());
235 
236     if (!MO.isImm())
237       return {};
238 
239     Imm = MO.getImm();
240   }
241 
242   switch (OpInfo.OperandType) {
243   case AMDGPU::OPERAND_REG_IMM_INT32:
244   case AMDGPU::OPERAND_REG_IMM_FP32:
245   case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
246   case AMDGPU::OPERAND_REG_INLINE_C_INT32:
247   case AMDGPU::OPERAND_REG_INLINE_C_FP32:
248   case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
249   case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
250   case AMDGPU::OPERAND_REG_IMM_V2INT32:
251   case AMDGPU::OPERAND_REG_IMM_V2FP32:
252   case AMDGPU::OPERAND_REG_INLINE_C_V2INT32:
253   case AMDGPU::OPERAND_REG_INLINE_C_V2FP32:
254     return getLit32Encoding(static_cast<uint32_t>(Imm), STI);
255 
256   case AMDGPU::OPERAND_REG_IMM_INT64:
257   case AMDGPU::OPERAND_REG_IMM_FP64:
258   case AMDGPU::OPERAND_REG_INLINE_C_INT64:
259   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
260   case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
261     return getLit64Encoding(static_cast<uint64_t>(Imm), STI);
262 
263   case AMDGPU::OPERAND_REG_IMM_INT16:
264   case AMDGPU::OPERAND_REG_INLINE_C_INT16:
265   case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
266     return getLit16IntEncoding(static_cast<uint16_t>(Imm), STI);
267   case AMDGPU::OPERAND_REG_IMM_FP16:
268   case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED:
269   case AMDGPU::OPERAND_REG_INLINE_C_FP16:
270   case AMDGPU::OPERAND_REG_INLINE_AC_FP16:
271     // FIXME Is this correct? What do inline immediates do on SI for f16 src
272     // which does not have f16 support?
273     return getLit16Encoding(static_cast<uint16_t>(Imm), STI);
274   case AMDGPU::OPERAND_REG_IMM_V2INT16:
275   case AMDGPU::OPERAND_REG_IMM_V2FP16: {
276     if (!isUInt<16>(Imm) && STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal])
277       return getLit32Encoding(static_cast<uint32_t>(Imm), STI);
278     if (OpInfo.OperandType == AMDGPU::OPERAND_REG_IMM_V2FP16)
279       return getLit16Encoding(static_cast<uint16_t>(Imm), STI);
280     [[fallthrough]];
281   }
282   case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
283   case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16:
284     return getLit16IntEncoding(static_cast<uint16_t>(Imm), STI);
285   case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
286   case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: {
287     uint16_t Lo16 = static_cast<uint16_t>(Imm);
288     uint32_t Encoding = getLit16Encoding(Lo16, STI);
289     return Encoding;
290   }
291   case AMDGPU::OPERAND_KIMM32:
292   case AMDGPU::OPERAND_KIMM16:
293     return MO.getImm();
294   default:
295     llvm_unreachable("invalid operand size");
296   }
297 }
298 
getImplicitOpSelHiEncoding(int Opcode) const299 uint64_t SIMCCodeEmitter::getImplicitOpSelHiEncoding(int Opcode) const {
300   using namespace AMDGPU::VOP3PEncoding;
301   using namespace AMDGPU::OpName;
302 
303   if (AMDGPU::hasNamedOperand(Opcode, op_sel_hi)) {
304     if (AMDGPU::hasNamedOperand(Opcode, src2))
305       return 0;
306     if (AMDGPU::hasNamedOperand(Opcode, src1))
307       return OP_SEL_HI_2;
308     if (AMDGPU::hasNamedOperand(Opcode, src0))
309       return OP_SEL_HI_1 | OP_SEL_HI_2;
310   }
311   return OP_SEL_HI_0 | OP_SEL_HI_1 | OP_SEL_HI_2;
312 }
313 
isVCMPX64(const MCInstrDesc & Desc)314 static bool isVCMPX64(const MCInstrDesc &Desc) {
315   return (Desc.TSFlags & SIInstrFlags::VOP3) &&
316          Desc.hasImplicitDefOfPhysReg(AMDGPU::EXEC);
317 }
318 
encodeInstruction(const MCInst & MI,raw_ostream & OS,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const319 void SIMCCodeEmitter::encodeInstruction(const MCInst &MI, raw_ostream &OS,
320                                         SmallVectorImpl<MCFixup> &Fixups,
321                                         const MCSubtargetInfo &STI) const {
322   int Opcode = MI.getOpcode();
323   APInt Encoding, Scratch;
324   getBinaryCodeForInstr(MI, Fixups, Encoding, Scratch,  STI);
325   const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
326   unsigned bytes = Desc.getSize();
327 
328   // Set unused op_sel_hi bits to 1 for VOP3P and MAI instructions.
329   // Note that accvgpr_read/write are MAI, have src0, but do not use op_sel.
330   if ((Desc.TSFlags & SIInstrFlags::VOP3P) ||
331       Opcode == AMDGPU::V_ACCVGPR_READ_B32_vi ||
332       Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_vi) {
333     Encoding |= getImplicitOpSelHiEncoding(Opcode);
334   }
335 
336   // GFX10+ v_cmpx opcodes promoted to VOP3 have implied dst=EXEC.
337   // Documentation requires dst to be encoded as EXEC (0x7E),
338   // but it looks like the actual value encoded for dst operand
339   // is ignored by HW. It was decided to define dst as "do not care"
340   // in td files to allow disassembler accept any dst value.
341   // However, dst is encoded as EXEC for compatibility with SP3.
342   if (AMDGPU::isGFX10Plus(STI) && isVCMPX64(Desc)) {
343     assert((Encoding & 0xFF) == 0);
344     Encoding |= MRI.getEncodingValue(AMDGPU::EXEC_LO);
345   }
346 
347   for (unsigned i = 0; i < bytes; i++) {
348     OS.write((uint8_t)Encoding.extractBitsAsZExtValue(8, 8 * i));
349   }
350 
351   // NSA encoding.
352   if (AMDGPU::isGFX10Plus(STI) && Desc.TSFlags & SIInstrFlags::MIMG) {
353     int vaddr0 = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
354                                             AMDGPU::OpName::vaddr0);
355     int srsrc = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
356                                            AMDGPU::OpName::srsrc);
357     assert(vaddr0 >= 0 && srsrc > vaddr0);
358     unsigned NumExtraAddrs = srsrc - vaddr0 - 1;
359     unsigned NumPadding = (-NumExtraAddrs) & 3;
360 
361     for (unsigned i = 0; i < NumExtraAddrs; ++i) {
362       getMachineOpValue(MI, MI.getOperand(vaddr0 + 1 + i), Encoding, Fixups,
363                         STI);
364       OS.write((uint8_t)Encoding.getLimitedValue());
365     }
366     for (unsigned i = 0; i < NumPadding; ++i)
367       OS.write(0);
368   }
369 
370   if ((bytes > 8 && STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal]) ||
371       (bytes > 4 && !STI.getFeatureBits()[AMDGPU::FeatureVOP3Literal]))
372     return;
373 
374   // Do not print literals from SISrc Operands for insts with mandatory literals
375   if (AMDGPU::hasNamedOperand(MI.getOpcode(), AMDGPU::OpName::imm))
376     return;
377 
378   // Check for additional literals
379   for (unsigned i = 0, e = Desc.getNumOperands(); i < e; ++i) {
380 
381     // Check if this operand should be encoded as [SV]Src
382     if (!AMDGPU::isSISrcOperand(Desc, i))
383       continue;
384 
385     // Is this operand a literal immediate?
386     const MCOperand &Op = MI.getOperand(i);
387     auto Enc = getLitEncoding(Op, Desc.operands()[i], STI);
388     if (!Enc || *Enc != 255)
389       continue;
390 
391     // Yes! Encode it
392     int64_t Imm = 0;
393 
394     if (Op.isImm())
395       Imm = Op.getImm();
396     else if (Op.isExpr()) {
397       if (const auto *C = dyn_cast<MCConstantExpr>(Op.getExpr()))
398         Imm = C->getValue();
399 
400     } else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
401       llvm_unreachable("Must be immediate or expr");
402 
403     for (unsigned j = 0; j < 4; j++) {
404       OS.write((uint8_t) ((Imm >> (8 * j)) & 0xff));
405     }
406 
407     // Only one literal value allowed
408     break;
409   }
410 }
411 
getSOPPBrEncoding(const MCInst & MI,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const412 void SIMCCodeEmitter::getSOPPBrEncoding(const MCInst &MI, unsigned OpNo,
413                                         APInt &Op,
414                                         SmallVectorImpl<MCFixup> &Fixups,
415                                         const MCSubtargetInfo &STI) const {
416   const MCOperand &MO = MI.getOperand(OpNo);
417 
418   if (MO.isExpr()) {
419     const MCExpr *Expr = MO.getExpr();
420     MCFixupKind Kind = (MCFixupKind)AMDGPU::fixup_si_sopp_br;
421     Fixups.push_back(MCFixup::create(0, Expr, Kind, MI.getLoc()));
422     Op = APInt::getNullValue(96);
423   } else {
424     getMachineOpValue(MI, MO, Op, Fixups, STI);
425   }
426 }
427 
getSMEMOffsetEncoding(const MCInst & MI,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const428 void SIMCCodeEmitter::getSMEMOffsetEncoding(const MCInst &MI, unsigned OpNo,
429                                             APInt &Op,
430                                             SmallVectorImpl<MCFixup> &Fixups,
431                                             const MCSubtargetInfo &STI) const {
432   auto Offset = MI.getOperand(OpNo).getImm();
433   // VI only supports 20-bit unsigned offsets.
434   assert(!AMDGPU::isVI(STI) || isUInt<20>(Offset));
435   Op = Offset;
436 }
437 
getSDWASrcEncoding(const MCInst & MI,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const438 void SIMCCodeEmitter::getSDWASrcEncoding(const MCInst &MI, unsigned OpNo,
439                                          APInt &Op,
440                                          SmallVectorImpl<MCFixup> &Fixups,
441                                          const MCSubtargetInfo &STI) const {
442   using namespace AMDGPU::SDWA;
443 
444   uint64_t RegEnc = 0;
445 
446   const MCOperand &MO = MI.getOperand(OpNo);
447 
448   if (MO.isReg()) {
449     unsigned Reg = MO.getReg();
450     RegEnc |= MRI.getEncodingValue(Reg);
451     RegEnc &= SDWA9EncValues::SRC_VGPR_MASK;
452     if (AMDGPU::isSGPR(AMDGPU::mc2PseudoReg(Reg), &MRI)) {
453       RegEnc |= SDWA9EncValues::SRC_SGPR_MASK;
454     }
455     Op = RegEnc;
456     return;
457   } else {
458     const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
459     auto Enc = getLitEncoding(MO, Desc.operands()[OpNo], STI);
460     if (Enc && *Enc != 255) {
461       Op = *Enc | SDWA9EncValues::SRC_SGPR_MASK;
462       return;
463     }
464   }
465 
466   llvm_unreachable("Unsupported operand kind");
467 }
468 
getSDWAVopcDstEncoding(const MCInst & MI,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const469 void SIMCCodeEmitter::getSDWAVopcDstEncoding(const MCInst &MI, unsigned OpNo,
470                                              APInt &Op,
471                                              SmallVectorImpl<MCFixup> &Fixups,
472                                              const MCSubtargetInfo &STI) const {
473   using namespace AMDGPU::SDWA;
474 
475   uint64_t RegEnc = 0;
476 
477   const MCOperand &MO = MI.getOperand(OpNo);
478 
479   unsigned Reg = MO.getReg();
480   if (Reg != AMDGPU::VCC && Reg != AMDGPU::VCC_LO) {
481     RegEnc |= MRI.getEncodingValue(Reg);
482     RegEnc &= SDWA9EncValues::VOPC_DST_SGPR_MASK;
483     RegEnc |= SDWA9EncValues::VOPC_DST_VCC_MASK;
484   }
485   Op = RegEnc;
486 }
487 
getAVOperandEncoding(const MCInst & MI,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const488 void SIMCCodeEmitter::getAVOperandEncoding(const MCInst &MI, unsigned OpNo,
489                                            APInt &Op,
490                                            SmallVectorImpl<MCFixup> &Fixups,
491                                            const MCSubtargetInfo &STI) const {
492   unsigned Reg = MI.getOperand(OpNo).getReg();
493   uint64_t Enc = MRI.getEncodingValue(Reg);
494 
495   // VGPR and AGPR have the same encoding, but SrcA and SrcB operands of mfma
496   // instructions use acc[0:1] modifier bits to distinguish. These bits are
497   // encoded as a virtual 9th bit of the register for these operands.
498   if (MRI.getRegClass(AMDGPU::AGPR_32RegClassID).contains(Reg) ||
499       MRI.getRegClass(AMDGPU::AReg_64RegClassID).contains(Reg) ||
500       MRI.getRegClass(AMDGPU::AReg_96RegClassID).contains(Reg) ||
501       MRI.getRegClass(AMDGPU::AReg_128RegClassID).contains(Reg) ||
502       MRI.getRegClass(AMDGPU::AReg_160RegClassID).contains(Reg) ||
503       MRI.getRegClass(AMDGPU::AReg_192RegClassID).contains(Reg) ||
504       MRI.getRegClass(AMDGPU::AReg_224RegClassID).contains(Reg) ||
505       MRI.getRegClass(AMDGPU::AReg_256RegClassID).contains(Reg) ||
506       MRI.getRegClass(AMDGPU::AReg_288RegClassID).contains(Reg) ||
507       MRI.getRegClass(AMDGPU::AReg_320RegClassID).contains(Reg) ||
508       MRI.getRegClass(AMDGPU::AReg_352RegClassID).contains(Reg) ||
509       MRI.getRegClass(AMDGPU::AReg_384RegClassID).contains(Reg) ||
510       MRI.getRegClass(AMDGPU::AReg_512RegClassID).contains(Reg) ||
511       MRI.getRegClass(AMDGPU::AGPR_LO16RegClassID).contains(Reg))
512     Enc |= 512;
513 
514   Op = Enc;
515 }
516 
needsPCRel(const MCExpr * Expr)517 static bool needsPCRel(const MCExpr *Expr) {
518   switch (Expr->getKind()) {
519   case MCExpr::SymbolRef: {
520     auto *SE = cast<MCSymbolRefExpr>(Expr);
521     MCSymbolRefExpr::VariantKind Kind = SE->getKind();
522     return Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_LO &&
523            Kind != MCSymbolRefExpr::VK_AMDGPU_ABS32_HI;
524   }
525   case MCExpr::Binary: {
526     auto *BE = cast<MCBinaryExpr>(Expr);
527     if (BE->getOpcode() == MCBinaryExpr::Sub)
528       return false;
529     return needsPCRel(BE->getLHS()) || needsPCRel(BE->getRHS());
530   }
531   case MCExpr::Unary:
532     return needsPCRel(cast<MCUnaryExpr>(Expr)->getSubExpr());
533   case MCExpr::Target:
534   case MCExpr::Constant:
535     return false;
536   }
537   llvm_unreachable("invalid kind");
538 }
539 
getMachineOpValue(const MCInst & MI,const MCOperand & MO,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const540 void SIMCCodeEmitter::getMachineOpValue(const MCInst &MI,
541                                         const MCOperand &MO, APInt &Op,
542                                         SmallVectorImpl<MCFixup> &Fixups,
543                                         const MCSubtargetInfo &STI) const {
544   if (MO.isReg()){
545     Op = MRI.getEncodingValue(MO.getReg());
546     return;
547   }
548   unsigned OpNo = &MO - MI.begin();
549   getMachineOpValueCommon(MI, MO, OpNo, Op, Fixups, STI);
550 }
551 
getMachineOpValueCommon(const MCInst & MI,const MCOperand & MO,unsigned OpNo,APInt & Op,SmallVectorImpl<MCFixup> & Fixups,const MCSubtargetInfo & STI) const552 void SIMCCodeEmitter::getMachineOpValueCommon(
553     const MCInst &MI, const MCOperand &MO, unsigned OpNo, APInt &Op,
554     SmallVectorImpl<MCFixup> &Fixups, const MCSubtargetInfo &STI) const {
555 
556   if (MO.isExpr() && MO.getExpr()->getKind() != MCExpr::Constant) {
557     // FIXME: If this is expression is PCRel or not should not depend on what
558     // the expression looks like. Given that this is just a general expression,
559     // it should probably be FK_Data_4 and whatever is producing
560     //
561     //    s_add_u32 s2, s2, (extern_const_addrspace+16
562     //
563     // And expecting a PCRel should instead produce
564     //
565     // .Ltmp1:
566     //   s_add_u32 s2, s2, (extern_const_addrspace+16)-.Ltmp1
567     MCFixupKind Kind;
568     if (needsPCRel(MO.getExpr()))
569       Kind = FK_PCRel_4;
570     else
571       Kind = FK_Data_4;
572 
573     const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
574     uint32_t Offset = Desc.getSize();
575     assert(Offset == 4 || Offset == 8);
576 
577     Fixups.push_back(MCFixup::create(Offset, MO.getExpr(), Kind, MI.getLoc()));
578   }
579 
580   const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
581   if (AMDGPU::isSISrcOperand(Desc, OpNo)) {
582     if (auto Enc = getLitEncoding(MO, Desc.operands()[OpNo], STI)) {
583       Op = *Enc;
584       return;
585     }
586   } else if (MO.isImm()) {
587     Op = MO.getImm();
588     return;
589   }
590 
591   llvm_unreachable("Encoding of this operand type is not supported yet.");
592 }
593 
594 #include "AMDGPUGenMCCodeEmitter.inc"
595