1 //===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
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 // This tablegen backend emits an assembly printer for the current target.
10 // Note that this is currently fairly skeletal, but will grow over time.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "AsmWriterInst.h"
15 #include "CodeGenInstAlias.h"
16 #include "CodeGenInstruction.h"
17 #include "CodeGenRegisters.h"
18 #include "CodeGenTarget.h"
19 #include "SequenceToOffsetTable.h"
20 #include "Types.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/ADT/DenseMap.h"
23 #include "llvm/ADT/STLExtras.h"
24 #include "llvm/ADT/SmallString.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringRef.h"
28 #include "llvm/ADT/Twine.h"
29 #include "llvm/Support/Casting.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/Format.h"
32 #include "llvm/Support/FormatVariadic.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/TableGen/Error.h"
36 #include "llvm/TableGen/Record.h"
37 #include "llvm/TableGen/TableGenBackend.h"
38 #include <algorithm>
39 #include <cassert>
40 #include <cstddef>
41 #include <cstdint>
42 #include <deque>
43 #include <iterator>
44 #include <map>
45 #include <set>
46 #include <string>
47 #include <tuple>
48 #include <utility>
49 #include <vector>
50 
51 using namespace llvm;
52 
53 #define DEBUG_TYPE "asm-writer-emitter"
54 
55 namespace {
56 
57 class AsmWriterEmitter {
58   RecordKeeper &Records;
59   CodeGenTarget Target;
60   ArrayRef<const CodeGenInstruction *> NumberedInstructions;
61   std::vector<AsmWriterInst> Instructions;
62 
63 public:
64   AsmWriterEmitter(RecordKeeper &R);
65 
66   void run(raw_ostream &o);
67 private:
68   void EmitGetMnemonic(
69       raw_ostream &o,
70       std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
71       unsigned &BitsLeft, unsigned &AsmStrBits);
72   void EmitPrintInstruction(
73       raw_ostream &o,
74       std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
75       unsigned &BitsLeft, unsigned &AsmStrBits);
76   void EmitGetRegisterName(raw_ostream &o);
77   void EmitPrintAliasInstruction(raw_ostream &O);
78 
79   void FindUniqueOperandCommands(std::vector<std::string> &UOC,
80                                  std::vector<std::vector<unsigned>> &InstIdxs,
81                                  std::vector<unsigned> &InstOpsUsed,
82                                  bool PassSubtarget) const;
83 };
84 
85 } // end anonymous namespace
86 
PrintCases(std::vector<std::pair<std::string,AsmWriterOperand>> & OpsToPrint,raw_ostream & O,bool PassSubtarget)87 static void PrintCases(std::vector<std::pair<std::string,
88                        AsmWriterOperand>> &OpsToPrint, raw_ostream &O,
89                        bool PassSubtarget) {
90   O << "    case " << OpsToPrint.back().first << ":";
91   AsmWriterOperand TheOp = OpsToPrint.back().second;
92   OpsToPrint.pop_back();
93 
94   // Check to see if any other operands are identical in this list, and if so,
95   // emit a case label for them.
96   for (unsigned i = OpsToPrint.size(); i != 0; --i)
97     if (OpsToPrint[i-1].second == TheOp) {
98       O << "\n    case " << OpsToPrint[i-1].first << ":";
99       OpsToPrint.erase(OpsToPrint.begin()+i-1);
100     }
101 
102   // Finally, emit the code.
103   O << "\n      " << TheOp.getCode(PassSubtarget);
104   O << "\n      break;\n";
105 }
106 
107 /// EmitInstructions - Emit the last instruction in the vector and any other
108 /// instructions that are suitably similar to it.
EmitInstructions(std::vector<AsmWriterInst> & Insts,raw_ostream & O,bool PassSubtarget)109 static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
110                              raw_ostream &O, bool PassSubtarget) {
111   AsmWriterInst FirstInst = Insts.back();
112   Insts.pop_back();
113 
114   std::vector<AsmWriterInst> SimilarInsts;
115   unsigned DifferingOperand = ~0;
116   for (unsigned i = Insts.size(); i != 0; --i) {
117     unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
118     if (DiffOp != ~1U) {
119       if (DifferingOperand == ~0U)  // First match!
120         DifferingOperand = DiffOp;
121 
122       // If this differs in the same operand as the rest of the instructions in
123       // this class, move it to the SimilarInsts list.
124       if (DifferingOperand == DiffOp || DiffOp == ~0U) {
125         SimilarInsts.push_back(Insts[i-1]);
126         Insts.erase(Insts.begin()+i-1);
127       }
128     }
129   }
130 
131   O << "  case " << FirstInst.CGI->Namespace << "::"
132     << FirstInst.CGI->TheDef->getName() << ":\n";
133   for (const AsmWriterInst &AWI : SimilarInsts)
134     O << "  case " << AWI.CGI->Namespace << "::"
135       << AWI.CGI->TheDef->getName() << ":\n";
136   for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
137     if (i != DifferingOperand) {
138       // If the operand is the same for all instructions, just print it.
139       O << "    " << FirstInst.Operands[i].getCode(PassSubtarget);
140     } else {
141       // If this is the operand that varies between all of the instructions,
142       // emit a switch for just this operand now.
143       O << "    switch (MI->getOpcode()) {\n";
144       O << "    default: llvm_unreachable(\"Unexpected opcode.\");\n";
145       std::vector<std::pair<std::string, AsmWriterOperand>> OpsToPrint;
146       OpsToPrint.push_back(std::make_pair(FirstInst.CGI->Namespace.str() + "::" +
147                                           FirstInst.CGI->TheDef->getName().str(),
148                                           FirstInst.Operands[i]));
149 
150       for (const AsmWriterInst &AWI : SimilarInsts) {
151         OpsToPrint.push_back(std::make_pair(AWI.CGI->Namespace.str()+"::" +
152                                             AWI.CGI->TheDef->getName().str(),
153                                             AWI.Operands[i]));
154       }
155       std::reverse(OpsToPrint.begin(), OpsToPrint.end());
156       while (!OpsToPrint.empty())
157         PrintCases(OpsToPrint, O, PassSubtarget);
158       O << "    }";
159     }
160     O << "\n";
161   }
162   O << "    break;\n";
163 }
164 
165 void AsmWriterEmitter::
FindUniqueOperandCommands(std::vector<std::string> & UniqueOperandCommands,std::vector<std::vector<unsigned>> & InstIdxs,std::vector<unsigned> & InstOpsUsed,bool PassSubtarget) const166 FindUniqueOperandCommands(std::vector<std::string> &UniqueOperandCommands,
167                           std::vector<std::vector<unsigned>> &InstIdxs,
168                           std::vector<unsigned> &InstOpsUsed,
169                           bool PassSubtarget) const {
170   // This vector parallels UniqueOperandCommands, keeping track of which
171   // instructions each case are used for.  It is a comma separated string of
172   // enums.
173   std::vector<std::string> InstrsForCase;
174   InstrsForCase.resize(UniqueOperandCommands.size());
175   InstOpsUsed.assign(UniqueOperandCommands.size(), 0);
176 
177   for (size_t i = 0, e = Instructions.size(); i != e; ++i) {
178     const AsmWriterInst &Inst = Instructions[i];
179     if (Inst.Operands.empty())
180       continue;   // Instruction already done.
181 
182     std::string Command = "    "+Inst.Operands[0].getCode(PassSubtarget)+"\n";
183 
184     // Check to see if we already have 'Command' in UniqueOperandCommands.
185     // If not, add it.
186     auto I = llvm::find(UniqueOperandCommands, Command);
187     if (I != UniqueOperandCommands.end()) {
188       size_t idx = I - UniqueOperandCommands.begin();
189       InstrsForCase[idx] += ", ";
190       InstrsForCase[idx] += Inst.CGI->TheDef->getName();
191       InstIdxs[idx].push_back(i);
192     } else {
193       UniqueOperandCommands.push_back(std::move(Command));
194       InstrsForCase.push_back(std::string(Inst.CGI->TheDef->getName()));
195       InstIdxs.emplace_back();
196       InstIdxs.back().push_back(i);
197 
198       // This command matches one operand so far.
199       InstOpsUsed.push_back(1);
200     }
201   }
202 
203   // For each entry of UniqueOperandCommands, there is a set of instructions
204   // that uses it.  If the next command of all instructions in the set are
205   // identical, fold it into the command.
206   for (size_t CommandIdx = 0, e = UniqueOperandCommands.size();
207        CommandIdx != e; ++CommandIdx) {
208 
209     const auto &Idxs = InstIdxs[CommandIdx];
210 
211     for (unsigned Op = 1; ; ++Op) {
212       // Find the first instruction in the set.
213       const AsmWriterInst &FirstInst = Instructions[Idxs.front()];
214       // If this instruction has no more operands, we isn't anything to merge
215       // into this command.
216       if (FirstInst.Operands.size() == Op)
217         break;
218 
219       // Otherwise, scan to see if all of the other instructions in this command
220       // set share the operand.
221       if (any_of(drop_begin(Idxs), [&](unsigned Idx) {
222             const AsmWriterInst &OtherInst = Instructions[Idx];
223             return OtherInst.Operands.size() == Op ||
224                    OtherInst.Operands[Op] != FirstInst.Operands[Op];
225           }))
226         break;
227 
228       // Okay, everything in this command set has the same next operand.  Add it
229       // to UniqueOperandCommands and remember that it was consumed.
230       std::string Command = "    " +
231         FirstInst.Operands[Op].getCode(PassSubtarget) + "\n";
232 
233       UniqueOperandCommands[CommandIdx] += Command;
234       InstOpsUsed[CommandIdx]++;
235     }
236   }
237 
238   // Prepend some of the instructions each case is used for onto the case val.
239   for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
240     std::string Instrs = InstrsForCase[i];
241     if (Instrs.size() > 70) {
242       Instrs.erase(Instrs.begin()+70, Instrs.end());
243       Instrs += "...";
244     }
245 
246     if (!Instrs.empty())
247       UniqueOperandCommands[i] = "    // " + Instrs + "\n" +
248         UniqueOperandCommands[i];
249   }
250 }
251 
UnescapeString(std::string & Str)252 static void UnescapeString(std::string &Str) {
253   for (unsigned i = 0; i != Str.size(); ++i) {
254     if (Str[i] == '\\' && i != Str.size()-1) {
255       switch (Str[i+1]) {
256       default: continue;  // Don't execute the code after the switch.
257       case 'a': Str[i] = '\a'; break;
258       case 'b': Str[i] = '\b'; break;
259       case 'e': Str[i] = 27; break;
260       case 'f': Str[i] = '\f'; break;
261       case 'n': Str[i] = '\n'; break;
262       case 'r': Str[i] = '\r'; break;
263       case 't': Str[i] = '\t'; break;
264       case 'v': Str[i] = '\v'; break;
265       case '"': Str[i] = '\"'; break;
266       case '\'': Str[i] = '\''; break;
267       case '\\': Str[i] = '\\'; break;
268       }
269       // Nuke the second character.
270       Str.erase(Str.begin()+i+1);
271     }
272   }
273 }
274 
275 /// UnescapeAliasString - Supports literal braces in InstAlias asm string which
276 /// are escaped with '\\' to avoid being interpreted as variants. Braces must
277 /// be unescaped before c++ code is generated as (e.g.):
278 ///
279 ///   AsmString = "foo \{$\x01\}";
280 ///
281 /// causes non-standard escape character warnings.
UnescapeAliasString(std::string & Str)282 static void UnescapeAliasString(std::string &Str) {
283   for (unsigned i = 0; i != Str.size(); ++i) {
284     if (Str[i] == '\\' && i != Str.size()-1) {
285       switch (Str[i+1]) {
286       default: continue;  // Don't execute the code after the switch.
287       case '{': Str[i] = '{'; break;
288       case '}': Str[i] = '}'; break;
289       }
290       // Nuke the second character.
291       Str.erase(Str.begin()+i+1);
292     }
293   }
294 }
295 
EmitGetMnemonic(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)296 void AsmWriterEmitter::EmitGetMnemonic(
297     raw_ostream &O,
298     std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
299     unsigned &BitsLeft, unsigned &AsmStrBits) {
300   Record *AsmWriter = Target.getAsmWriter();
301   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
302   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
303 
304   O << "/// getMnemonic - This method is automatically generated by "
305        "tablegen\n"
306        "/// from the instruction set description.\n"
307        "std::pair<const char *, uint64_t> "
308     << Target.getName() << ClassName << "::getMnemonic(const MCInst *MI) {\n";
309 
310   // Build an aggregate string, and build a table of offsets into it.
311   SequenceToOffsetTable<std::string> StringTable;
312 
313   /// OpcodeInfo - This encodes the index of the string to use for the first
314   /// chunk of the output as well as indices used for operand printing.
315   std::vector<uint64_t> OpcodeInfo(NumberedInstructions.size());
316   const unsigned OpcodeInfoBits = 64;
317 
318   // Add all strings to the string table upfront so it can generate an optimized
319   // representation.
320   for (AsmWriterInst &AWI : Instructions) {
321     if (AWI.Operands[0].OperandType ==
322                  AsmWriterOperand::isLiteralTextOperand &&
323         !AWI.Operands[0].Str.empty()) {
324       std::string Str = AWI.Operands[0].Str;
325       UnescapeString(Str);
326       StringTable.add(Str);
327     }
328   }
329 
330   StringTable.layout();
331 
332   unsigned MaxStringIdx = 0;
333   for (AsmWriterInst &AWI : Instructions) {
334     unsigned Idx;
335     if (AWI.Operands[0].OperandType != AsmWriterOperand::isLiteralTextOperand ||
336         AWI.Operands[0].Str.empty()) {
337       // Something handled by the asmwriter printer, but with no leading string.
338       Idx = StringTable.get("");
339     } else {
340       std::string Str = AWI.Operands[0].Str;
341       UnescapeString(Str);
342       Idx = StringTable.get(Str);
343       MaxStringIdx = std::max(MaxStringIdx, Idx);
344 
345       // Nuke the string from the operand list.  It is now handled!
346       AWI.Operands.erase(AWI.Operands.begin());
347     }
348 
349     // Bias offset by one since we want 0 as a sentinel.
350     OpcodeInfo[AWI.CGIIndex] = Idx+1;
351   }
352 
353   // Figure out how many bits we used for the string index.
354   AsmStrBits = Log2_32_Ceil(MaxStringIdx + 2);
355 
356   // To reduce code size, we compactify common instructions into a few bits
357   // in the opcode-indexed table.
358   BitsLeft = OpcodeInfoBits - AsmStrBits;
359 
360   while (true) {
361     std::vector<std::string> UniqueOperandCommands;
362     std::vector<std::vector<unsigned>> InstIdxs;
363     std::vector<unsigned> NumInstOpsHandled;
364     FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
365                               NumInstOpsHandled, PassSubtarget);
366 
367     // If we ran out of operands to print, we're done.
368     if (UniqueOperandCommands.empty()) break;
369 
370     // Compute the number of bits we need to represent these cases, this is
371     // ceil(log2(numentries)).
372     unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
373 
374     // If we don't have enough bits for this operand, don't include it.
375     if (NumBits > BitsLeft) {
376       LLVM_DEBUG(errs() << "Not enough bits to densely encode " << NumBits
377                         << " more bits\n");
378       break;
379     }
380 
381     // Otherwise, we can include this in the initial lookup table.  Add it in.
382     for (size_t i = 0, e = InstIdxs.size(); i != e; ++i) {
383       unsigned NumOps = NumInstOpsHandled[i];
384       for (unsigned Idx : InstIdxs[i]) {
385         OpcodeInfo[Instructions[Idx].CGIIndex] |=
386           (uint64_t)i << (OpcodeInfoBits-BitsLeft);
387         // Remove the info about this operand from the instruction.
388         AsmWriterInst &Inst = Instructions[Idx];
389         if (!Inst.Operands.empty()) {
390           assert(NumOps <= Inst.Operands.size() &&
391                  "Can't remove this many ops!");
392           Inst.Operands.erase(Inst.Operands.begin(),
393                               Inst.Operands.begin()+NumOps);
394         }
395       }
396     }
397     BitsLeft -= NumBits;
398 
399     // Remember the handlers for this set of operands.
400     TableDrivenOperandPrinters.push_back(std::move(UniqueOperandCommands));
401   }
402 
403   // Emit the string table itself.
404   StringTable.emitStringLiteralDef(O, "  static const char AsmStrs[]");
405 
406   // Emit the lookup tables in pieces to minimize wasted bytes.
407   unsigned BytesNeeded = ((OpcodeInfoBits - BitsLeft) + 7) / 8;
408   unsigned Table = 0, Shift = 0;
409   SmallString<128> BitsString;
410   raw_svector_ostream BitsOS(BitsString);
411   // If the total bits is more than 32-bits we need to use a 64-bit type.
412   BitsOS << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
413          << "_t Bits = 0;\n";
414   while (BytesNeeded != 0) {
415     // Figure out how big this table section needs to be, but no bigger than 4.
416     unsigned TableSize = std::min(llvm::bit_floor(BytesNeeded), 4u);
417     BytesNeeded -= TableSize;
418     TableSize *= 8; // Convert to bits;
419     uint64_t Mask = (1ULL << TableSize) - 1;
420     O << "  static const uint" << TableSize << "_t OpInfo" << Table
421       << "[] = {\n";
422     for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
423       O << "    " << ((OpcodeInfo[i] >> Shift) & Mask) << "U,\t// "
424         << NumberedInstructions[i]->TheDef->getName() << "\n";
425     }
426     O << "  };\n\n";
427     // Emit string to combine the individual table lookups.
428     BitsOS << "  Bits |= ";
429     // If the total bits is more than 32-bits we need to use a 64-bit type.
430     if (BitsLeft < (OpcodeInfoBits - 32))
431       BitsOS << "(uint64_t)";
432     BitsOS << "OpInfo" << Table << "[MI->getOpcode()] << " << Shift << ";\n";
433     // Prepare the shift for the next iteration and increment the table count.
434     Shift += TableSize;
435     ++Table;
436   }
437 
438   O << "  // Emit the opcode for the instruction.\n";
439   O << BitsString;
440 
441   // Make sure we don't return an invalid pointer if bits is 0
442   O << "  if (Bits == 0)\n"
443        "    return {nullptr, Bits};\n";
444 
445   // Return mnemonic string and bits.
446   O << "  return {AsmStrs+(Bits & " << (1 << AsmStrBits) - 1
447     << ")-1, Bits};\n\n";
448 
449   O << "}\n";
450 }
451 
452 /// EmitPrintInstruction - Generate the code for the "printInstruction" method
453 /// implementation. Destroys all instances of AsmWriterInst information, by
454 /// clearing the Instructions vector.
EmitPrintInstruction(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)455 void AsmWriterEmitter::EmitPrintInstruction(
456     raw_ostream &O,
457     std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
458     unsigned &BitsLeft, unsigned &AsmStrBits) {
459   const unsigned OpcodeInfoBits = 64;
460   Record *AsmWriter = Target.getAsmWriter();
461   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
462   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
463 
464   // This function has some huge switch statements that causing excessive
465   // compile time in LLVM profile instrumenation build. This print function
466   // usually is not frequently called in compilation. Here we disable the
467   // profile instrumenation for this function.
468   O << "/// printInstruction - This method is automatically generated by "
469        "tablegen\n"
470        "/// from the instruction set description.\n"
471        "LLVM_NO_PROFILE_INSTRUMENT_FUNCTION\n"
472        "void "
473     << Target.getName() << ClassName
474     << "::printInstruction(const MCInst *MI, uint64_t Address, "
475     << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
476     << "raw_ostream &O) {\n";
477 
478   // Emit the initial tab character.
479   O << "  O << \"\\t\";\n\n";
480 
481   // Emit the starting string.
482   O << "  auto MnemonicInfo = getMnemonic(MI);\n\n";
483   O << "  O << MnemonicInfo.first;\n\n";
484 
485   O << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
486     << "_t Bits = MnemonicInfo.second;\n"
487     << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n";
488 
489   // Output the table driven operand information.
490   BitsLeft = OpcodeInfoBits-AsmStrBits;
491   for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
492     std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
493 
494     // Compute the number of bits we need to represent these cases, this is
495     // ceil(log2(numentries)).
496     unsigned NumBits = Log2_32_Ceil(Commands.size());
497     assert(NumBits <= BitsLeft && "consistency error");
498 
499     // Emit code to extract this field from Bits.
500     O << "\n  // Fragment " << i << " encoded into " << NumBits
501       << " bits for " << Commands.size() << " unique commands.\n";
502 
503     if (Commands.size() == 2) {
504       // Emit two possibilitys with if/else.
505       O << "  if ((Bits >> "
506         << (OpcodeInfoBits-BitsLeft) << ") & "
507         << ((1 << NumBits)-1) << ") {\n"
508         << Commands[1]
509         << "  } else {\n"
510         << Commands[0]
511         << "  }\n\n";
512     } else if (Commands.size() == 1) {
513       // Emit a single possibility.
514       O << Commands[0] << "\n\n";
515     } else {
516       O << "  switch ((Bits >> "
517         << (OpcodeInfoBits-BitsLeft) << ") & "
518         << ((1 << NumBits)-1) << ") {\n"
519         << "  default: llvm_unreachable(\"Invalid command number.\");\n";
520 
521       // Print out all the cases.
522       for (unsigned j = 0, e = Commands.size(); j != e; ++j) {
523         O << "  case " << j << ":\n";
524         O << Commands[j];
525         O << "    break;\n";
526       }
527       O << "  }\n\n";
528     }
529     BitsLeft -= NumBits;
530   }
531 
532   // Okay, delete instructions with no operand info left.
533   llvm::erase_if(Instructions,
534                  [](AsmWriterInst &Inst) { return Inst.Operands.empty(); });
535 
536   // Because this is a vector, we want to emit from the end.  Reverse all of the
537   // elements in the vector.
538   std::reverse(Instructions.begin(), Instructions.end());
539 
540 
541   // Now that we've emitted all of the operand info that fit into 64 bits, emit
542   // information for those instructions that are left.  This is a less dense
543   // encoding, but we expect the main 64-bit table to handle the majority of
544   // instructions.
545   if (!Instructions.empty()) {
546     // Find the opcode # of inline asm.
547     O << "  switch (MI->getOpcode()) {\n";
548     O << "  default: llvm_unreachable(\"Unexpected opcode.\");\n";
549     while (!Instructions.empty())
550       EmitInstructions(Instructions, O, PassSubtarget);
551 
552     O << "  }\n";
553   }
554 
555   O << "}\n";
556 }
557 
558 static void
emitRegisterNameString(raw_ostream & O,StringRef AltName,const std::deque<CodeGenRegister> & Registers)559 emitRegisterNameString(raw_ostream &O, StringRef AltName,
560                        const std::deque<CodeGenRegister> &Registers) {
561   SequenceToOffsetTable<std::string> StringTable;
562   SmallVector<std::string, 4> AsmNames(Registers.size());
563   unsigned i = 0;
564   for (const auto &Reg : Registers) {
565     std::string &AsmName = AsmNames[i++];
566 
567     // "NoRegAltName" is special. We don't need to do a lookup for that,
568     // as it's just a reference to the default register name.
569     if (AltName == "" || AltName == "NoRegAltName") {
570       AsmName = std::string(Reg.TheDef->getValueAsString("AsmName"));
571       if (AsmName.empty())
572         AsmName = std::string(Reg.getName());
573     } else {
574       // Make sure the register has an alternate name for this index.
575       std::vector<Record*> AltNameList =
576         Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
577       unsigned Idx = 0, e;
578       for (e = AltNameList.size();
579            Idx < e && (AltNameList[Idx]->getName() != AltName);
580            ++Idx)
581         ;
582       // If the register has an alternate name for this index, use it.
583       // Otherwise, leave it empty as an error flag.
584       if (Idx < e) {
585         std::vector<StringRef> AltNames =
586           Reg.TheDef->getValueAsListOfStrings("AltNames");
587         if (AltNames.size() <= Idx)
588           PrintFatalError(Reg.TheDef->getLoc(),
589                           "Register definition missing alt name for '" +
590                           AltName + "'.");
591         AsmName = std::string(AltNames[Idx]);
592       }
593     }
594     StringTable.add(AsmName);
595   }
596 
597   StringTable.layout();
598   StringTable.emitStringLiteralDef(O, Twine("  static const char AsmStrs") +
599                                           AltName + "[]");
600 
601   O << "  static const " << getMinimalTypeForRange(StringTable.size() - 1, 32)
602     << " RegAsmOffset" << AltName << "[] = {";
603   for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
604     if ((i % 14) == 0)
605       O << "\n    ";
606     O << StringTable.get(AsmNames[i]) << ", ";
607   }
608   O << "\n  };\n"
609     << "\n";
610 }
611 
EmitGetRegisterName(raw_ostream & O)612 void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
613   Record *AsmWriter = Target.getAsmWriter();
614   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
615   const auto &Registers = Target.getRegBank().getRegisters();
616   const std::vector<Record*> &AltNameIndices = Target.getRegAltNameIndices();
617   bool hasAltNames = AltNameIndices.size() > 1;
618   StringRef Namespace = Registers.front().TheDef->getValueAsString("Namespace");
619 
620   O <<
621   "\n\n/// getRegisterName - This method is automatically generated by tblgen\n"
622   "/// from the register set description.  This returns the assembler name\n"
623   "/// for the specified register.\n"
624   "const char *" << Target.getName() << ClassName << "::";
625   if (hasAltNames)
626     O << "\ngetRegisterName(MCRegister Reg, unsigned AltIdx) {\n";
627   else
628     O << "getRegisterName(MCRegister Reg) {\n";
629   O << "  unsigned RegNo = Reg.id();\n"
630     << "  assert(RegNo && RegNo < " << (Registers.size() + 1)
631     << " && \"Invalid register number!\");\n"
632     << "\n";
633 
634   if (hasAltNames) {
635     for (const Record *R : AltNameIndices)
636       emitRegisterNameString(O, R->getName(), Registers);
637   } else
638     emitRegisterNameString(O, "", Registers);
639 
640   if (hasAltNames) {
641     O << "  switch(AltIdx) {\n"
642       << "  default: llvm_unreachable(\"Invalid register alt name index!\");\n";
643     for (const Record *R : AltNameIndices) {
644       StringRef AltName = R->getName();
645       O << "  case ";
646       if (!Namespace.empty())
647         O << Namespace << "::";
648       O << AltName << ":\n";
649       if (R->isValueUnset("FallbackRegAltNameIndex"))
650         O << "    assert(*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
651           << "[RegNo-1]) &&\n"
652           << "           \"Invalid alt name index for register!\");\n";
653       else {
654         O << "    if (!*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
655           << "[RegNo-1]))\n"
656           << "      return getRegisterName(RegNo, ";
657         if (!Namespace.empty())
658           O << Namespace << "::";
659         O << R->getValueAsDef("FallbackRegAltNameIndex")->getName() << ");\n";
660       }
661       O << "    return AsmStrs" << AltName << "+RegAsmOffset" << AltName
662         << "[RegNo-1];\n";
663     }
664     O << "  }\n";
665   } else {
666     O << "  assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
667       << "          \"Invalid alt name index for register!\");\n"
668       << "  return AsmStrs+RegAsmOffset[RegNo-1];\n";
669   }
670   O << "}\n";
671 }
672 
673 namespace {
674 
675 // IAPrinter - Holds information about an InstAlias. Two InstAliases match if
676 // they both have the same conditionals. In which case, we cannot print out the
677 // alias for that pattern.
678 class IAPrinter {
679   std::map<StringRef, std::pair<int, int>> OpMap;
680 
681   std::vector<std::string> Conds;
682 
683   std::string Result;
684   std::string AsmString;
685 
686   unsigned NumMIOps;
687 
688 public:
IAPrinter(std::string R,std::string AS,unsigned NumMIOps)689   IAPrinter(std::string R, std::string AS, unsigned NumMIOps)
690       : Result(std::move(R)), AsmString(std::move(AS)), NumMIOps(NumMIOps) {}
691 
addCond(std::string C)692   void addCond(std::string C) { Conds.push_back(std::move(C)); }
getConds() const693   ArrayRef<std::string> getConds() const { return Conds; }
getCondCount() const694   size_t getCondCount() const { return Conds.size(); }
695 
addOperand(StringRef Op,int OpIdx,int PrintMethodIdx=-1)696   void addOperand(StringRef Op, int OpIdx, int PrintMethodIdx = -1) {
697     assert(OpIdx >= 0 && OpIdx < 0xFE && "Idx out of range");
698     assert(PrintMethodIdx >= -1 && PrintMethodIdx < 0xFF &&
699            "Idx out of range");
700     OpMap[Op] = std::make_pair(OpIdx, PrintMethodIdx);
701   }
702 
getNumMIOps()703   unsigned getNumMIOps() { return NumMIOps; }
704 
getResult()705   StringRef getResult() { return Result; }
706 
isOpMapped(StringRef Op)707   bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
getOpIndex(StringRef Op)708   int getOpIndex(StringRef Op) { return OpMap[Op].first; }
getOpData(StringRef Op)709   std::pair<int, int> &getOpData(StringRef Op) { return OpMap[Op]; }
710 
parseName(StringRef::iterator Start,StringRef::iterator End)711   std::pair<StringRef, StringRef::iterator> parseName(StringRef::iterator Start,
712                                                       StringRef::iterator End) {
713     StringRef::iterator I = Start;
714     StringRef::iterator Next;
715     if (*I == '{') {
716       // ${some_name}
717       Start = ++I;
718       while (I != End && *I != '}')
719         ++I;
720       Next = I;
721       // eat the final '}'
722       if (Next != End)
723         ++Next;
724     } else {
725       // $name, just eat the usual suspects.
726       while (I != End && (isAlnum(*I) || *I == '_'))
727         ++I;
728       Next = I;
729     }
730 
731     return std::make_pair(StringRef(Start, I - Start), Next);
732   }
733 
formatAliasString(uint32_t & UnescapedSize)734   std::string formatAliasString(uint32_t &UnescapedSize) {
735     // Directly mangle mapped operands into the string. Each operand is
736     // identified by a '$' sign followed by a byte identifying the number of the
737     // operand. We add one to the index to avoid zero bytes.
738     StringRef ASM(AsmString);
739     std::string OutString;
740     raw_string_ostream OS(OutString);
741     for (StringRef::iterator I = ASM.begin(), E = ASM.end(); I != E;) {
742       OS << *I;
743       ++UnescapedSize;
744       if (*I == '$') {
745         StringRef Name;
746         std::tie(Name, I) = parseName(++I, E);
747         assert(isOpMapped(Name) && "Unmapped operand!");
748 
749         int OpIndex, PrintIndex;
750         std::tie(OpIndex, PrintIndex) = getOpData(Name);
751         if (PrintIndex == -1) {
752           // Can use the default printOperand route.
753           OS << format("\\x%02X", (unsigned char)OpIndex + 1);
754           ++UnescapedSize;
755         } else {
756           // 3 bytes if a PrintMethod is needed: 0xFF, the MCInst operand
757           // number, and which of our pre-detected Methods to call.
758           OS << format("\\xFF\\x%02X\\x%02X", OpIndex + 1, PrintIndex + 1);
759           UnescapedSize += 3;
760         }
761       } else {
762         ++I;
763       }
764     }
765     return OutString;
766   }
767 
operator ==(const IAPrinter & RHS) const768   bool operator==(const IAPrinter &RHS) const {
769     if (NumMIOps != RHS.NumMIOps)
770       return false;
771     if (Conds.size() != RHS.Conds.size())
772       return false;
773 
774     unsigned Idx = 0;
775     for (const auto &str : Conds)
776       if (str != RHS.Conds[Idx++])
777         return false;
778 
779     return true;
780   }
781 };
782 
783 } // end anonymous namespace
784 
CountNumOperands(StringRef AsmString,unsigned Variant)785 static unsigned CountNumOperands(StringRef AsmString, unsigned Variant) {
786   return AsmString.count(' ') + AsmString.count('\t');
787 }
788 
789 namespace {
790 
791 struct AliasPriorityComparator {
792   typedef std::pair<CodeGenInstAlias, int> ValueType;
operator ()__anond69e789b0511::AliasPriorityComparator793   bool operator()(const ValueType &LHS, const ValueType &RHS) const {
794     if (LHS.second ==  RHS.second) {
795       // We don't actually care about the order, but for consistency it
796       // shouldn't depend on pointer comparisons.
797       return LessRecordByID()(LHS.first.TheDef, RHS.first.TheDef);
798     }
799 
800     // Aliases with larger priorities should be considered first.
801     return LHS.second > RHS.second;
802   }
803 };
804 
805 } // end anonymous namespace
806 
EmitPrintAliasInstruction(raw_ostream & O)807 void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
808   Record *AsmWriter = Target.getAsmWriter();
809 
810   O << "\n#ifdef PRINT_ALIAS_INSTR\n";
811   O << "#undef PRINT_ALIAS_INSTR\n\n";
812 
813   //////////////////////////////
814   // Gather information about aliases we need to print
815   //////////////////////////////
816 
817   // Emit the method that prints the alias instruction.
818   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
819   unsigned Variant = AsmWriter->getValueAsInt("Variant");
820   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
821 
822   std::vector<Record*> AllInstAliases =
823     Records.getAllDerivedDefinitions("InstAlias");
824 
825   // Create a map from the qualified name to a list of potential matches.
826   typedef std::set<std::pair<CodeGenInstAlias, int>, AliasPriorityComparator>
827       AliasWithPriority;
828   std::map<std::string, AliasWithPriority> AliasMap;
829   for (Record *R : AllInstAliases) {
830     int Priority = R->getValueAsInt("EmitPriority");
831     if (Priority < 1)
832       continue; // Aliases with priority 0 are never emitted.
833 
834     const DagInit *DI = R->getValueAsDag("ResultInst");
835     AliasMap[getQualifiedName(DI->getOperatorAsDef(R->getLoc()))].insert(
836         std::make_pair(CodeGenInstAlias(R, Target), Priority));
837   }
838 
839   // A map of which conditions need to be met for each instruction operand
840   // before it can be matched to the mnemonic.
841   std::map<std::string, std::vector<IAPrinter>> IAPrinterMap;
842 
843   std::vector<std::pair<std::string, bool>> PrintMethods;
844 
845   // A list of MCOperandPredicates for all operands in use, and the reverse map
846   std::vector<const Record*> MCOpPredicates;
847   DenseMap<const Record*, unsigned> MCOpPredicateMap;
848 
849   for (auto &Aliases : AliasMap) {
850     // Collection of instruction alias rules. May contain ambiguous rules.
851     std::vector<IAPrinter> IAPs;
852 
853     for (auto &Alias : Aliases.second) {
854       const CodeGenInstAlias &CGA = Alias.first;
855       unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
856       std::string FlatInstAsmString =
857          CodeGenInstruction::FlattenAsmStringVariants(CGA.ResultInst->AsmString,
858                                                       Variant);
859       unsigned NumResultOps = CountNumOperands(FlatInstAsmString, Variant);
860 
861       std::string FlatAliasAsmString =
862           CodeGenInstruction::FlattenAsmStringVariants(CGA.AsmString, Variant);
863       UnescapeAliasString(FlatAliasAsmString);
864 
865       // Don't emit the alias if it has more operands than what it's aliasing.
866       if (NumResultOps < CountNumOperands(FlatAliasAsmString, Variant))
867         continue;
868 
869       StringRef Namespace = Target.getName();
870       unsigned NumMIOps = 0;
871       for (auto &ResultInstOpnd : CGA.ResultInst->Operands)
872         NumMIOps += ResultInstOpnd.MINumOperands;
873 
874       IAPrinter IAP(CGA.Result->getAsString(), FlatAliasAsmString, NumMIOps);
875 
876       unsigned MIOpNum = 0;
877       for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
878         // Skip over tied operands as they're not part of an alias declaration.
879         auto &Operands = CGA.ResultInst->Operands;
880         while (true) {
881           unsigned OpNum = Operands.getSubOperandNumber(MIOpNum).first;
882           if (Operands[OpNum].MINumOperands == 1 &&
883               Operands[OpNum].getTiedRegister() != -1) {
884             // Tied operands of different RegisterClass should be explicit within
885             // an instruction's syntax and so cannot be skipped.
886             int TiedOpNum = Operands[OpNum].getTiedRegister();
887             if (Operands[OpNum].Rec->getName() ==
888                 Operands[TiedOpNum].Rec->getName()) {
889               ++MIOpNum;
890               continue;
891             }
892           }
893           break;
894         }
895 
896         // Ignore unchecked result operands.
897         while (IAP.getCondCount() < MIOpNum)
898           IAP.addCond("AliasPatternCond::K_Ignore, 0");
899 
900         const CodeGenInstAlias::ResultOperand &RO = CGA.ResultOperands[i];
901 
902         switch (RO.Kind) {
903         case CodeGenInstAlias::ResultOperand::K_Record: {
904           const Record *Rec = RO.getRecord();
905           StringRef ROName = RO.getName();
906           int PrintMethodIdx = -1;
907 
908           // These two may have a PrintMethod, which we want to record (if it's
909           // the first time we've seen it) and provide an index for the aliasing
910           // code to use.
911           if (Rec->isSubClassOf("RegisterOperand") ||
912               Rec->isSubClassOf("Operand")) {
913             StringRef PrintMethod = Rec->getValueAsString("PrintMethod");
914             bool IsPCRel =
915                 Rec->getValueAsString("OperandType") == "OPERAND_PCREL";
916             if (PrintMethod != "" && PrintMethod != "printOperand") {
917               PrintMethodIdx = llvm::find_if(PrintMethods,
918                                              [&](auto &X) {
919                                                return X.first == PrintMethod;
920                                              }) -
921                                PrintMethods.begin();
922               if (static_cast<unsigned>(PrintMethodIdx) == PrintMethods.size())
923                 PrintMethods.emplace_back(std::string(PrintMethod), IsPCRel);
924             }
925           }
926 
927           if (Rec->isSubClassOf("RegisterOperand"))
928             Rec = Rec->getValueAsDef("RegClass");
929           if (Rec->isSubClassOf("RegisterClass")) {
930             if (!IAP.isOpMapped(ROName)) {
931               IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
932               Record *R = CGA.ResultOperands[i].getRecord();
933               if (R->isSubClassOf("RegisterOperand"))
934                 R = R->getValueAsDef("RegClass");
935               IAP.addCond(std::string(
936                   formatv("AliasPatternCond::K_RegClass, {0}::{1}RegClassID",
937                           Namespace, R->getName())));
938             } else {
939               IAP.addCond(std::string(formatv(
940                   "AliasPatternCond::K_TiedReg, {0}", IAP.getOpIndex(ROName))));
941             }
942           } else {
943             // Assume all printable operands are desired for now. This can be
944             // overridden in the InstAlias instantiation if necessary.
945             IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
946 
947             // There might be an additional predicate on the MCOperand
948             unsigned Entry = MCOpPredicateMap[Rec];
949             if (!Entry) {
950               if (!Rec->isValueUnset("MCOperandPredicate")) {
951                 MCOpPredicates.push_back(Rec);
952                 Entry = MCOpPredicates.size();
953                 MCOpPredicateMap[Rec] = Entry;
954               } else
955                 break; // No conditions on this operand at all
956             }
957             IAP.addCond(
958                 std::string(formatv("AliasPatternCond::K_Custom, {0}", Entry)));
959           }
960           break;
961         }
962         case CodeGenInstAlias::ResultOperand::K_Imm: {
963           // Just because the alias has an immediate result, doesn't mean the
964           // MCInst will. An MCExpr could be present, for example.
965           auto Imm = CGA.ResultOperands[i].getImm();
966           int32_t Imm32 = int32_t(Imm);
967           if (Imm != Imm32)
968             PrintFatalError("Matching an alias with an immediate out of the "
969                             "range of int32_t is not supported");
970           IAP.addCond(std::string(
971               formatv("AliasPatternCond::K_Imm, uint32_t({0})", Imm32)));
972           break;
973         }
974         case CodeGenInstAlias::ResultOperand::K_Reg:
975           if (!CGA.ResultOperands[i].getRegister()) {
976             IAP.addCond(std::string(formatv(
977                 "AliasPatternCond::K_Reg, {0}::NoRegister", Namespace)));
978             break;
979           }
980 
981           StringRef Reg = CGA.ResultOperands[i].getRegister()->getName();
982           IAP.addCond(std::string(
983               formatv("AliasPatternCond::K_Reg, {0}::{1}", Namespace, Reg)));
984           break;
985         }
986 
987         MIOpNum += RO.getMINumOperands();
988       }
989 
990       std::vector<Record *> ReqFeatures;
991       if (PassSubtarget) {
992         // We only consider ReqFeatures predicates if PassSubtarget
993         std::vector<Record *> RF =
994             CGA.TheDef->getValueAsListOfDefs("Predicates");
995         copy_if(RF, std::back_inserter(ReqFeatures), [](Record *R) {
996           return R->getValueAsBit("AssemblerMatcherPredicate");
997         });
998       }
999 
1000       for (Record *const R : ReqFeatures) {
1001         const DagInit *D = R->getValueAsDag("AssemblerCondDag");
1002         auto *Op = dyn_cast<DefInit>(D->getOperator());
1003         if (!Op)
1004           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1005         StringRef CombineType = Op->getDef()->getName();
1006         if (CombineType != "any_of" && CombineType != "all_of")
1007           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1008         if (D->getNumArgs() == 0)
1009           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1010         bool IsOr = CombineType == "any_of";
1011         // Change (any_of FeatureAll, (any_of ...)) to (any_of FeatureAll, ...).
1012         if (IsOr && D->getNumArgs() == 2 && isa<DagInit>(D->getArg(1))) {
1013           DagInit *RHS = cast<DagInit>(D->getArg(1));
1014           SmallVector<Init *> Args{D->getArg(0)};
1015           SmallVector<StringInit *> ArgNames{D->getArgName(0)};
1016           for (unsigned i = 0, e = RHS->getNumArgs(); i != e; ++i) {
1017             Args.push_back(RHS->getArg(i));
1018             ArgNames.push_back(RHS->getArgName(i));
1019           }
1020           D = DagInit::get(D->getOperator(), nullptr, Args, ArgNames);
1021         }
1022 
1023         for (auto *Arg : D->getArgs()) {
1024           bool IsNeg = false;
1025           if (auto *NotArg = dyn_cast<DagInit>(Arg)) {
1026             if (NotArg->getOperator()->getAsString() != "not" ||
1027                 NotArg->getNumArgs() != 1)
1028               PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1029             Arg = NotArg->getArg(0);
1030             IsNeg = true;
1031           }
1032           if (!isa<DefInit>(Arg) ||
1033               !cast<DefInit>(Arg)->getDef()->isSubClassOf("SubtargetFeature"))
1034             PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1035 
1036           IAP.addCond(std::string(formatv(
1037               "AliasPatternCond::K_{0}{1}Feature, {2}::{3}", IsOr ? "Or" : "",
1038               IsNeg ? "Neg" : "", Namespace, Arg->getAsString())));
1039         }
1040         // If an AssemblerPredicate with ors is used, note end of list should
1041         // these be combined.
1042         if (IsOr)
1043           IAP.addCond("AliasPatternCond::K_EndOrFeatures, 0");
1044       }
1045 
1046       IAPrinterMap[Aliases.first].push_back(std::move(IAP));
1047     }
1048   }
1049 
1050   //////////////////////////////
1051   // Write out the printAliasInstr function
1052   //////////////////////////////
1053 
1054   std::string Header;
1055   raw_string_ostream HeaderO(Header);
1056 
1057   HeaderO << "bool " << Target.getName() << ClassName
1058           << "::printAliasInstr(const MCInst"
1059           << " *MI, uint64_t Address, "
1060           << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
1061           << "raw_ostream &OS) {\n";
1062 
1063   std::string PatternsForOpcode;
1064   raw_string_ostream OpcodeO(PatternsForOpcode);
1065 
1066   unsigned PatternCount = 0;
1067   std::string Patterns;
1068   raw_string_ostream PatternO(Patterns);
1069 
1070   unsigned CondCount = 0;
1071   std::string Conds;
1072   raw_string_ostream CondO(Conds);
1073 
1074   // All flattened alias strings.
1075   std::map<std::string, uint32_t> AsmStringOffsets;
1076   std::vector<std::pair<uint32_t, std::string>> AsmStrings;
1077   size_t AsmStringsSize = 0;
1078 
1079   // Iterate over the opcodes in enum order so they are sorted by opcode for
1080   // binary search.
1081   for (const CodeGenInstruction *Inst : NumberedInstructions) {
1082     auto It = IAPrinterMap.find(getQualifiedName(Inst->TheDef));
1083     if (It == IAPrinterMap.end())
1084       continue;
1085     std::vector<IAPrinter> &IAPs = It->second;
1086     std::vector<IAPrinter*> UniqueIAPs;
1087 
1088     // Remove any ambiguous alias rules.
1089     for (auto &LHS : IAPs) {
1090       bool IsDup = false;
1091       for (const auto &RHS : IAPs) {
1092         if (&LHS != &RHS && LHS == RHS) {
1093           IsDup = true;
1094           break;
1095         }
1096       }
1097 
1098       if (!IsDup)
1099         UniqueIAPs.push_back(&LHS);
1100     }
1101 
1102     if (UniqueIAPs.empty()) continue;
1103 
1104     unsigned PatternStart = PatternCount;
1105 
1106     // Insert the pattern start and opcode in the pattern list for debugging.
1107     PatternO << formatv("    // {0} - {1}\n", It->first, PatternStart);
1108 
1109     for (IAPrinter *IAP : UniqueIAPs) {
1110       // Start each condition list with a comment of the resulting pattern that
1111       // we're trying to match.
1112       unsigned CondStart = CondCount;
1113       CondO << formatv("    // {0} - {1}\n", IAP->getResult(), CondStart);
1114       for (const auto &Cond : IAP->getConds())
1115         CondO << "    {" << Cond << "},\n";
1116       CondCount += IAP->getCondCount();
1117 
1118       // After operands have been examined, re-encode the alias string with
1119       // escapes indicating how operands should be printed.
1120       uint32_t UnescapedSize = 0;
1121       std::string EncodedAsmString = IAP->formatAliasString(UnescapedSize);
1122       auto Insertion =
1123           AsmStringOffsets.insert({EncodedAsmString, AsmStringsSize});
1124       if (Insertion.second) {
1125         // If the string is new, add it to the vector.
1126         AsmStrings.push_back({AsmStringsSize, EncodedAsmString});
1127         AsmStringsSize += UnescapedSize + 1;
1128       }
1129       unsigned AsmStrOffset = Insertion.first->second;
1130 
1131       PatternO << formatv("    {{{0}, {1}, {2}, {3} },\n", AsmStrOffset,
1132                           CondStart, IAP->getNumMIOps(), IAP->getCondCount());
1133       ++PatternCount;
1134     }
1135 
1136     OpcodeO << formatv("    {{{0}, {1}, {2} },\n", It->first, PatternStart,
1137                        PatternCount - PatternStart);
1138   }
1139 
1140   if (OpcodeO.str().empty()) {
1141     O << HeaderO.str();
1142     O << "  return false;\n";
1143     O << "}\n\n";
1144     O << "#endif // PRINT_ALIAS_INSTR\n";
1145     return;
1146   }
1147 
1148   // Forward declare the validation method if needed.
1149   if (!MCOpPredicates.empty())
1150     O << "static bool " << Target.getName() << ClassName
1151       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1152       << "                  const MCSubtargetInfo &STI,\n"
1153       << "                  unsigned PredicateIndex);\n";
1154 
1155   O << HeaderO.str();
1156   O.indent(2) << "static const PatternsForOpcode OpToPatterns[] = {\n";
1157   O << OpcodeO.str();
1158   O.indent(2) << "};\n\n";
1159   O.indent(2) << "static const AliasPattern Patterns[] = {\n";
1160   O << PatternO.str();
1161   O.indent(2) << "};\n\n";
1162   O.indent(2) << "static const AliasPatternCond Conds[] = {\n";
1163   O << CondO.str();
1164   O.indent(2) << "};\n\n";
1165   O.indent(2) << "static const char AsmStrings[] =\n";
1166   for (const auto &P : AsmStrings) {
1167     O.indent(4) << "/* " << P.first << " */ \"" << P.second << "\\0\"\n";
1168   }
1169 
1170   O.indent(2) << ";\n\n";
1171 
1172   // Assert that the opcode table is sorted. Use a static local constructor to
1173   // ensure that the check only happens once on first run.
1174   O << "#ifndef NDEBUG\n";
1175   O.indent(2) << "static struct SortCheck {\n";
1176   O.indent(2) << "  SortCheck(ArrayRef<PatternsForOpcode> OpToPatterns) {\n";
1177   O.indent(2) << "    assert(std::is_sorted(\n";
1178   O.indent(2) << "               OpToPatterns.begin(), OpToPatterns.end(),\n";
1179   O.indent(2) << "               [](const PatternsForOpcode &L, const "
1180                  "PatternsForOpcode &R) {\n";
1181   O.indent(2) << "                 return L.Opcode < R.Opcode;\n";
1182   O.indent(2) << "               }) &&\n";
1183   O.indent(2) << "           \"tablegen failed to sort opcode patterns\");\n";
1184   O.indent(2) << "  }\n";
1185   O.indent(2) << "} sortCheckVar(OpToPatterns);\n";
1186   O << "#endif\n\n";
1187 
1188   O.indent(2) << "AliasMatchingData M {\n";
1189   O.indent(2) << "  ArrayRef(OpToPatterns),\n";
1190   O.indent(2) << "  ArrayRef(Patterns),\n";
1191   O.indent(2) << "  ArrayRef(Conds),\n";
1192   O.indent(2) << "  StringRef(AsmStrings, std::size(AsmStrings)),\n";
1193   if (MCOpPredicates.empty())
1194     O.indent(2) << "  nullptr,\n";
1195   else
1196     O.indent(2) << "  &" << Target.getName() << ClassName << "ValidateMCOperand,\n";
1197   O.indent(2) << "};\n";
1198 
1199   O.indent(2) << "const char *AsmString = matchAliasPatterns(MI, "
1200               << (PassSubtarget ? "&STI" : "nullptr") << ", M);\n";
1201   O.indent(2) << "if (!AsmString) return false;\n\n";
1202 
1203   // Code that prints the alias, replacing the operands with the ones from the
1204   // MCInst.
1205   O << "  unsigned I = 0;\n";
1206   O << "  while (AsmString[I] != ' ' && AsmString[I] != '\\t' &&\n";
1207   O << "         AsmString[I] != '$' && AsmString[I] != '\\0')\n";
1208   O << "    ++I;\n";
1209   O << "  OS << '\\t' << StringRef(AsmString, I);\n";
1210 
1211   O << "  if (AsmString[I] != '\\0') {\n";
1212   O << "    if (AsmString[I] == ' ' || AsmString[I] == '\\t') {\n";
1213   O << "      OS << '\\t';\n";
1214   O << "      ++I;\n";
1215   O << "    }\n";
1216   O << "    do {\n";
1217   O << "      if (AsmString[I] == '$') {\n";
1218   O << "        ++I;\n";
1219   O << "        if (AsmString[I] == (char)0xff) {\n";
1220   O << "          ++I;\n";
1221   O << "          int OpIdx = AsmString[I++] - 1;\n";
1222   O << "          int PrintMethodIdx = AsmString[I++] - 1;\n";
1223   O << "          printCustomAliasOperand(MI, Address, OpIdx, PrintMethodIdx, ";
1224   O << (PassSubtarget ? "STI, " : "");
1225   O << "OS);\n";
1226   O << "        } else\n";
1227   O << "          printOperand(MI, unsigned(AsmString[I++]) - 1, ";
1228   O << (PassSubtarget ? "STI, " : "");
1229   O << "OS);\n";
1230   O << "      } else {\n";
1231   O << "        OS << AsmString[I++];\n";
1232   O << "      }\n";
1233   O << "    } while (AsmString[I] != '\\0');\n";
1234   O << "  }\n\n";
1235 
1236   O << "  return true;\n";
1237   O << "}\n\n";
1238 
1239   //////////////////////////////
1240   // Write out the printCustomAliasOperand function
1241   //////////////////////////////
1242 
1243   O << "void " << Target.getName() << ClassName << "::"
1244     << "printCustomAliasOperand(\n"
1245     << "         const MCInst *MI, uint64_t Address, unsigned OpIdx,\n"
1246     << "         unsigned PrintMethodIdx,\n"
1247     << (PassSubtarget ? "         const MCSubtargetInfo &STI,\n" : "")
1248     << "         raw_ostream &OS) {\n";
1249   if (PrintMethods.empty())
1250     O << "  llvm_unreachable(\"Unknown PrintMethod kind\");\n";
1251   else {
1252     O << "  switch (PrintMethodIdx) {\n"
1253       << "  default:\n"
1254       << "    llvm_unreachable(\"Unknown PrintMethod kind\");\n"
1255       << "    break;\n";
1256 
1257     for (unsigned i = 0; i < PrintMethods.size(); ++i) {
1258       O << "  case " << i << ":\n"
1259         << "    " << PrintMethods[i].first << "(MI, "
1260         << (PrintMethods[i].second ? "Address, " : "") << "OpIdx, "
1261         << (PassSubtarget ? "STI, " : "") << "OS);\n"
1262         << "    break;\n";
1263     }
1264     O << "  }\n";
1265   }
1266   O << "}\n\n";
1267 
1268   if (!MCOpPredicates.empty()) {
1269     O << "static bool " << Target.getName() << ClassName
1270       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1271       << "                  const MCSubtargetInfo &STI,\n"
1272       << "                  unsigned PredicateIndex) {\n"
1273       << "  switch (PredicateIndex) {\n"
1274       << "  default:\n"
1275       << "    llvm_unreachable(\"Unknown MCOperandPredicate kind\");\n"
1276       << "    break;\n";
1277 
1278     for (unsigned i = 0; i < MCOpPredicates.size(); ++i) {
1279       StringRef MCOpPred = MCOpPredicates[i]->getValueAsString("MCOperandPredicate");
1280       O << "  case " << i + 1 << ": {\n"
1281         << MCOpPred.data() << "\n"
1282         << "    }\n";
1283     }
1284     O << "  }\n"
1285       << "}\n\n";
1286   }
1287 
1288   O << "#endif // PRINT_ALIAS_INSTR\n";
1289 }
1290 
AsmWriterEmitter(RecordKeeper & R)1291 AsmWriterEmitter::AsmWriterEmitter(RecordKeeper &R) : Records(R), Target(R) {
1292   Record *AsmWriter = Target.getAsmWriter();
1293   unsigned Variant = AsmWriter->getValueAsInt("Variant");
1294 
1295   // Get the instruction numbering.
1296   NumberedInstructions = Target.getInstructionsByEnumValue();
1297 
1298   for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
1299     const CodeGenInstruction *I = NumberedInstructions[i];
1300     if (!I->AsmString.empty() && I->TheDef->getName() != "PHI")
1301       Instructions.emplace_back(*I, i, Variant);
1302   }
1303 }
1304 
run(raw_ostream & O)1305 void AsmWriterEmitter::run(raw_ostream &O) {
1306   std::vector<std::vector<std::string>> TableDrivenOperandPrinters;
1307   unsigned BitsLeft = 0;
1308   unsigned AsmStrBits = 0;
1309   emitSourceFileHeader("Assembly Writer Source Fragment", O, Records);
1310   EmitGetMnemonic(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1311   EmitPrintInstruction(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1312   EmitGetRegisterName(O);
1313   EmitPrintAliasInstruction(O);
1314 }
1315 
1316 static TableGen::Emitter::OptClass<AsmWriterEmitter>
1317     X("gen-asm-writer", "Generate assembly writer");
1318