1 //===-- SIMCCodeEmitter.cpp - SI Code Emitter -----------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief The SI code emitter produces machine code that can be executed
12 /// directly on the GPU device.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "AMDGPU.h"
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"
35 #include <cassert>
36 #include <cstdint>
37 #include <cstdlib>
38 
39 using namespace llvm;
40 
41 namespace {
42 
43 class SIMCCodeEmitter : public  AMDGPUMCCodeEmitter {
44   const MCRegisterInfo &MRI;
45 
46   /// \brief Encode an fp or int literal
47   uint32_t getLitEncoding(const MCOperand &MO, const MCOperandInfo &OpInfo,
48                           const MCSubtargetInfo &STI) const;
49 
50 public:
51   SIMCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
52                   MCContext &ctx)
53       : AMDGPUMCCodeEmitter(mcii), MRI(mri) {}
54   SIMCCodeEmitter(const SIMCCodeEmitter &) = delete;
55   SIMCCodeEmitter &operator=(const SIMCCodeEmitter &) = delete;
56 
57   /// \brief 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;
61 
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;
66 
67   /// \brief Use a fixup to encode the simm16 field for SOPP branch
68   ///        instructions.
69   unsigned getSOPPBrEncoding(const MCInst &MI, unsigned OpNo,
70                              SmallVectorImpl<MCFixup> &Fixups,
71                              const MCSubtargetInfo &STI) const override;
72 };
73 
74 } // end anonymous namespace
75 
76 MCCodeEmitter *llvm::createSIMCCodeEmitter(const MCInstrInfo &MCII,
77                                            const MCRegisterInfo &MRI,
78                                            MCContext &Ctx) {
79   return new SIMCCodeEmitter(MCII, MRI, Ctx);
80 }
81 
82 // Returns the encoding value to use if the given integer is an integer inline
83 // immediate value, or 0 if it is not.
84 template <typename IntTy>
85 static uint32_t getIntInlineImmEncoding(IntTy Imm) {
86   if (Imm >= 0 && Imm <= 64)
87     return 128 + Imm;
88 
89   if (Imm >= -16 && Imm <= -1)
90     return 192 + std::abs(Imm);
91 
92   return 0;
93 }
94 
95 static uint32_t getLit16Encoding(uint16_t Val, const MCSubtargetInfo &STI) {
96   uint16_t IntImm = getIntInlineImmEncoding(static_cast<int16_t>(Val));
97   if (IntImm != 0)
98     return IntImm;
99 
100   if (Val == 0x3800) // 0.5
101     return 240;
102 
103   if (Val == 0xB800) // -0.5
104     return 241;
105 
106   if (Val == 0x3C00) // 1.0
107     return 242;
108 
109   if (Val == 0xBC00) // -1.0
110     return 243;
111 
112   if (Val == 0x4000) // 2.0
113     return 244;
114 
115   if (Val == 0xC000) // -2.0
116     return 245;
117 
118   if (Val == 0x4400) // 4.0
119     return 246;
120 
121   if (Val == 0xC400) // -4.0
122     return 247;
123 
124   if (Val == 0x3118 && // 1.0 / (2.0 * pi)
125       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
126     return 248;
127 
128   return 255;
129 }
130 
131 static uint32_t getLit32Encoding(uint32_t Val, const MCSubtargetInfo &STI) {
132   uint32_t IntImm = getIntInlineImmEncoding(static_cast<int32_t>(Val));
133   if (IntImm != 0)
134     return IntImm;
135 
136   if (Val == FloatToBits(0.5f))
137     return 240;
138 
139   if (Val == FloatToBits(-0.5f))
140     return 241;
141 
142   if (Val == FloatToBits(1.0f))
143     return 242;
144 
145   if (Val == FloatToBits(-1.0f))
146     return 243;
147 
148   if (Val == FloatToBits(2.0f))
149     return 244;
150 
151   if (Val == FloatToBits(-2.0f))
152     return 245;
153 
154   if (Val == FloatToBits(4.0f))
155     return 246;
156 
157   if (Val == FloatToBits(-4.0f))
158     return 247;
159 
160   if (Val == 0x3e22f983 && // 1.0 / (2.0 * pi)
161       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
162     return 248;
163 
164   return 255;
165 }
166 
167 static uint32_t getLit64Encoding(uint64_t Val, const MCSubtargetInfo &STI) {
168   uint32_t IntImm = getIntInlineImmEncoding(static_cast<int64_t>(Val));
169   if (IntImm != 0)
170     return IntImm;
171 
172   if (Val == DoubleToBits(0.5))
173     return 240;
174 
175   if (Val == DoubleToBits(-0.5))
176     return 241;
177 
178   if (Val == DoubleToBits(1.0))
179     return 242;
180 
181   if (Val == DoubleToBits(-1.0))
182     return 243;
183 
184   if (Val == DoubleToBits(2.0))
185     return 244;
186 
187   if (Val == DoubleToBits(-2.0))
188     return 245;
189 
190   if (Val == DoubleToBits(4.0))
191     return 246;
192 
193   if (Val == DoubleToBits(-4.0))
194     return 247;
195 
196   if (Val == 0x3fc45f306dc9c882 && // 1.0 / (2.0 * pi)
197       STI.getFeatureBits()[AMDGPU::FeatureInv2PiInlineImm])
198     return 248;
199 
200   return 255;
201 }
202 
203 uint32_t SIMCCodeEmitter::getLitEncoding(const MCOperand &MO,
204                                          const MCOperandInfo &OpInfo,
205                                          const MCSubtargetInfo &STI) const {
206   int64_t Imm;
207   if (MO.isExpr()) {
208     const auto *C = dyn_cast<MCConstantExpr>(MO.getExpr());
209     if (!C)
210       return 255;
211 
212     Imm = C->getValue();
213   } else {
214 
215     assert(!MO.isFPImm());
216 
217     if (!MO.isImm())
218       return ~0;
219 
220     Imm = MO.getImm();
221   }
222 
223   switch (OpInfo.OperandType) {
224   case AMDGPU::OPERAND_REG_IMM_INT32:
225   case AMDGPU::OPERAND_REG_IMM_FP32:
226   case AMDGPU::OPERAND_REG_INLINE_C_INT32:
227   case AMDGPU::OPERAND_REG_INLINE_C_FP32:
228     return getLit32Encoding(static_cast<uint32_t>(Imm), STI);
229 
230   case AMDGPU::OPERAND_REG_IMM_INT64:
231   case AMDGPU::OPERAND_REG_IMM_FP64:
232   case AMDGPU::OPERAND_REG_INLINE_C_INT64:
233   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
234     return getLit64Encoding(static_cast<uint64_t>(Imm), STI);
235 
236   case AMDGPU::OPERAND_REG_IMM_INT16:
237   case AMDGPU::OPERAND_REG_IMM_FP16:
238   case AMDGPU::OPERAND_REG_INLINE_C_INT16:
239   case AMDGPU::OPERAND_REG_INLINE_C_FP16:
240     // FIXME Is this correct? What do inline immediates do on SI for f16 src
241     // which does not have f16 support?
242     return getLit16Encoding(static_cast<uint16_t>(Imm), STI);
243 
244   case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
245   case AMDGPU::OPERAND_REG_INLINE_C_V2FP16: {
246     uint16_t Lo16 = static_cast<uint16_t>(Imm);
247     assert(Lo16 == static_cast<uint16_t>(Imm >> 16));
248     uint32_t Encoding = getLit16Encoding(Lo16, STI);
249     assert(Encoding != 255 && "packed constants can only be inline immediates");
250     return Encoding;
251   }
252   default:
253     llvm_unreachable("invalid operand size");
254   }
255 }
256 
257 void SIMCCodeEmitter::encodeInstruction(const MCInst &MI, raw_ostream &OS,
258                                        SmallVectorImpl<MCFixup> &Fixups,
259                                        const MCSubtargetInfo &STI) const {
260   verifyInstructionPredicates(MI,
261                               computeAvailableFeatures(STI.getFeatureBits()));
262 
263   uint64_t Encoding = getBinaryCodeForInstr(MI, Fixups, STI);
264   const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
265   unsigned bytes = Desc.getSize();
266 
267   for (unsigned i = 0; i < bytes; i++) {
268     OS.write((uint8_t) ((Encoding >> (8 * i)) & 0xff));
269   }
270 
271   if (bytes > 4)
272     return;
273 
274   // Check for additional literals in SRC0/1/2 (Op 1/2/3)
275   for (unsigned i = 0, e = MI.getNumOperands(); i < e; ++i) {
276 
277     // Check if this operand should be encoded as [SV]Src
278     if (!AMDGPU::isSISrcOperand(Desc, i))
279       continue;
280 
281     // Is this operand a literal immediate?
282     const MCOperand &Op = MI.getOperand(i);
283     if (getLitEncoding(Op, Desc.OpInfo[i], STI) != 255)
284       continue;
285 
286     // Yes! Encode it
287     int64_t Imm = 0;
288 
289     if (Op.isImm())
290       Imm = Op.getImm();
291     else if (Op.isExpr()) {
292       if (const auto *C = dyn_cast<MCConstantExpr>(Op.getExpr()))
293         Imm = C->getValue();
294 
295     } else if (!Op.isExpr()) // Exprs will be replaced with a fixup value.
296       llvm_unreachable("Must be immediate or expr");
297 
298     for (unsigned j = 0; j < 4; j++) {
299       OS.write((uint8_t) ((Imm >> (8 * j)) & 0xff));
300     }
301 
302     // Only one literal value allowed
303     break;
304   }
305 }
306 
307 unsigned SIMCCodeEmitter::getSOPPBrEncoding(const MCInst &MI, unsigned OpNo,
308                                             SmallVectorImpl<MCFixup> &Fixups,
309                                             const MCSubtargetInfo &STI) const {
310   const MCOperand &MO = MI.getOperand(OpNo);
311 
312   if (MO.isExpr()) {
313     const MCExpr *Expr = MO.getExpr();
314     MCFixupKind Kind = (MCFixupKind)AMDGPU::fixup_si_sopp_br;
315     Fixups.push_back(MCFixup::create(0, Expr, Kind, MI.getLoc()));
316     return 0;
317   }
318 
319   return getMachineOpValue(MI, MO, Fixups, STI);
320 }
321 
322 uint64_t SIMCCodeEmitter::getMachineOpValue(const MCInst &MI,
323                                             const MCOperand &MO,
324                                        SmallVectorImpl<MCFixup> &Fixups,
325                                        const MCSubtargetInfo &STI) const {
326   if (MO.isReg())
327     return MRI.getEncodingValue(MO.getReg());
328 
329   if (MO.isExpr() && MO.getExpr()->getKind() != MCExpr::Constant) {
330     const auto *Expr = dyn_cast<MCSymbolRefExpr>(MO.getExpr());
331     MCFixupKind Kind;
332     if (Expr && Expr->getSymbol().isExternal())
333       Kind = FK_Data_4;
334     else
335       Kind = FK_PCRel_4;
336     Fixups.push_back(MCFixup::create(4, MO.getExpr(), Kind, MI.getLoc()));
337   }
338 
339   // Figure out the operand number, needed for isSrcOperand check
340   unsigned OpNo = 0;
341   for (unsigned e = MI.getNumOperands(); OpNo < e; ++OpNo) {
342     if (&MO == &MI.getOperand(OpNo))
343       break;
344   }
345 
346   const MCInstrDesc &Desc = MCII.get(MI.getOpcode());
347   if (AMDGPU::isSISrcOperand(Desc, OpNo)) {
348     uint32_t Enc = getLitEncoding(MO, Desc.OpInfo[OpNo], STI);
349     if (Enc != ~0U && (Enc != 255 || Desc.getSize() == 4))
350       return Enc;
351 
352   } else if (MO.isImm())
353     return MO.getImm();
354 
355   llvm_unreachable("Encoding of this operand type is not supported yet.");
356   return 0;
357 }
358