xref: /freebsd/contrib/llvm-project/llvm/utils/TableGen/AsmWriterEmitter.cpp (revision 700637cbb5e582861067a11aaca4d053546871d2)
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 "Basic/SequenceToOffsetTable.h"
15 #include "Common/AsmWriterInst.h"
16 #include "Common/CodeGenInstAlias.h"
17 #include "Common/CodeGenInstruction.h"
18 #include "Common/CodeGenRegisters.h"
19 #include "Common/CodeGenTarget.h"
20 #include "Common/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   const RecordKeeper &Records;
59   CodeGenTarget Target;
60   ArrayRef<const CodeGenInstruction *> NumberedInstructions;
61   std::vector<AsmWriterInst> Instructions;
62 
63 public:
64   AsmWriterEmitter(const RecordKeeper &R);
65 
66   void run(raw_ostream &o);
67 
68 private:
69   void EmitGetMnemonic(
70       raw_ostream &o,
71       std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
72       unsigned &BitsLeft, unsigned &AsmStrBits);
73   void EmitPrintInstruction(
74       raw_ostream &o,
75       std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
76       unsigned &BitsLeft, unsigned &AsmStrBits);
77   void EmitGetRegisterName(raw_ostream &o);
78   void EmitPrintAliasInstruction(raw_ostream &O);
79 
80   void FindUniqueOperandCommands(std::vector<std::string> &UOC,
81                                  std::vector<std::vector<unsigned>> &InstIdxs,
82                                  std::vector<unsigned> &InstOpsUsed,
83                                  bool PassSubtarget) const;
84 };
85 
86 } // end anonymous namespace
87 
88 static void
PrintCases(std::vector<std::pair<std::string,AsmWriterOperand>> & OpsToPrint,raw_ostream & O,bool PassSubtarget)89 PrintCases(std::vector<std::pair<std::string, AsmWriterOperand>> &OpsToPrint,
90            raw_ostream &O, bool PassSubtarget) {
91   O << "    case " << OpsToPrint.back().first << ":";
92   AsmWriterOperand TheOp = OpsToPrint.back().second;
93   OpsToPrint.pop_back();
94 
95   // Check to see if any other operands are identical in this list, and if so,
96   // emit a case label for them.
97   for (unsigned i = OpsToPrint.size(); i != 0; --i)
98     if (OpsToPrint[i - 1].second == TheOp) {
99       O << "\n    case " << OpsToPrint[i - 1].first << ":";
100       OpsToPrint.erase(OpsToPrint.begin() + i - 1);
101     }
102 
103   // Finally, emit the code.
104   O << "\n      " << TheOp.getCode(PassSubtarget);
105   O << "\n      break;\n";
106 }
107 
108 /// EmitInstructions - Emit the last instruction in the vector and any other
109 /// instructions that are suitably similar to it.
EmitInstructions(std::vector<AsmWriterInst> & Insts,raw_ostream & O,bool PassSubtarget)110 static void EmitInstructions(std::vector<AsmWriterInst> &Insts, raw_ostream &O,
111                              bool PassSubtarget) {
112   AsmWriterInst FirstInst = Insts.back();
113   Insts.pop_back();
114 
115   std::vector<AsmWriterInst> SimilarInsts;
116   unsigned DifferingOperand = ~0;
117   for (unsigned i = Insts.size(); i != 0; --i) {
118     unsigned DiffOp = Insts[i - 1].MatchesAllButOneOp(FirstInst);
119     if (DiffOp != ~1U) {
120       if (DifferingOperand == ~0U) // First match!
121         DifferingOperand = DiffOp;
122 
123       // If this differs in the same operand as the rest of the instructions in
124       // this class, move it to the SimilarInsts list.
125       if (DifferingOperand == DiffOp || DiffOp == ~0U) {
126         SimilarInsts.push_back(Insts[i - 1]);
127         Insts.erase(Insts.begin() + i - 1);
128       }
129     }
130   }
131 
132   O << "  case " << FirstInst.CGI->Namespace
133     << "::" << FirstInst.CGI->TheDef->getName() << ":\n";
134   for (const AsmWriterInst &AWI : SimilarInsts)
135     O << "  case " << AWI.CGI->Namespace << "::" << AWI.CGI->TheDef->getName()
136       << ":\n";
137   for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
138     if (i != DifferingOperand) {
139       // If the operand is the same for all instructions, just print it.
140       O << "    " << FirstInst.Operands[i].getCode(PassSubtarget);
141     } else {
142       // If this is the operand that varies between all of the instructions,
143       // emit a switch for just this operand now.
144       O << "    switch (MI->getOpcode()) {\n";
145       O << "    default: llvm_unreachable(\"Unexpected opcode.\");\n";
146       std::vector<std::pair<std::string, AsmWriterOperand>> OpsToPrint;
147       OpsToPrint.emplace_back(FirstInst.CGI->Namespace.str() +
148                                   "::" + FirstInst.CGI->TheDef->getName().str(),
149                               FirstInst.Operands[i]);
150 
151       for (const AsmWriterInst &AWI : SimilarInsts) {
152         OpsToPrint.emplace_back(AWI.CGI->Namespace.str() +
153                                     "::" + AWI.CGI->TheDef->getName().str(),
154                                 AWI.Operands[i]);
155       }
156       std::reverse(OpsToPrint.begin(), OpsToPrint.end());
157       while (!OpsToPrint.empty())
158         PrintCases(OpsToPrint, O, PassSubtarget);
159       O << "    }";
160     }
161     O << "\n";
162   }
163   O << "    break;\n";
164 }
165 
FindUniqueOperandCommands(std::vector<std::string> & UniqueOperandCommands,std::vector<std::vector<unsigned>> & InstIdxs,std::vector<unsigned> & InstOpsUsed,bool PassSubtarget) const166 void AsmWriterEmitter::FindUniqueOperandCommands(
167     std::vector<std::string> &UniqueOperandCommands,
168     std::vector<std::vector<unsigned>> &InstIdxs,
169     std::vector<unsigned> &InstOpsUsed, 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 =
183         "    " + Inst.Operands[0].getCode(PassSubtarget) + "\n";
184 
185     // Check to see if we already have 'Command' in UniqueOperandCommands.
186     // If not, add it.
187     auto I = llvm::find(UniqueOperandCommands, Command);
188     if (I != UniqueOperandCommands.end()) {
189       size_t idx = I - UniqueOperandCommands.begin();
190       InstrsForCase[idx] += ", ";
191       InstrsForCase[idx] += Inst.CGI->TheDef->getName();
192       InstIdxs[idx].push_back(i);
193     } else {
194       UniqueOperandCommands.push_back(std::move(Command));
195       InstrsForCase.push_back(Inst.CGI->TheDef->getName().str());
196       InstIdxs.emplace_back();
197       InstIdxs.back().push_back(i);
198 
199       // This command matches one operand so far.
200       InstOpsUsed.push_back(1);
201     }
202   }
203 
204   // For each entry of UniqueOperandCommands, there is a set of instructions
205   // that uses it.  If the next command of all instructions in the set are
206   // identical, fold it into the command.
207   for (size_t CommandIdx = 0, e = UniqueOperandCommands.size(); CommandIdx != e;
208        ++CommandIdx) {
209 
210     const auto &Idxs = InstIdxs[CommandIdx];
211 
212     for (unsigned Op = 1;; ++Op) {
213       // Find the first instruction in the set.
214       const AsmWriterInst &FirstInst = Instructions[Idxs.front()];
215       // If this instruction has no more operands, we isn't anything to merge
216       // into this command.
217       if (FirstInst.Operands.size() == Op)
218         break;
219 
220       // Otherwise, scan to see if all of the other instructions in this command
221       // set share the operand.
222       if (any_of(drop_begin(Idxs), [&](unsigned Idx) {
223             const AsmWriterInst &OtherInst = Instructions[Idx];
224             return OtherInst.Operands.size() == Op ||
225                    OtherInst.Operands[Op] != FirstInst.Operands[Op];
226           }))
227         break;
228 
229       // Okay, everything in this command set has the same next operand.  Add it
230       // to UniqueOperandCommands and remember that it was consumed.
231       std::string Command =
232           "    " + FirstInst.Operands[Op].getCode(PassSubtarget) + "\n";
233 
234       UniqueOperandCommands[CommandIdx] += Command;
235       InstOpsUsed[CommandIdx]++;
236     }
237   }
238 
239   // Prepend some of the instructions each case is used for onto the case val.
240   for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
241     std::string Instrs = InstrsForCase[i];
242     if (Instrs.size() > 70) {
243       Instrs.erase(Instrs.begin() + 70, Instrs.end());
244       Instrs += "...";
245     }
246 
247     if (!Instrs.empty())
248       UniqueOperandCommands[i] =
249           "    // " + Instrs + "\n" + UniqueOperandCommands[i];
250   }
251 }
252 
UnescapeString(std::string & Str)253 static void UnescapeString(std::string &Str) {
254   for (unsigned i = 0; i != Str.size(); ++i) {
255     if (Str[i] == '\\' && i != Str.size() - 1) {
256       switch (Str[i + 1]) {
257       default:
258         continue; // Don't execute the code after the switch.
259       case 'a':
260         Str[i] = '\a';
261         break;
262       case 'b':
263         Str[i] = '\b';
264         break;
265       case 'e':
266         Str[i] = 27;
267         break;
268       case 'f':
269         Str[i] = '\f';
270         break;
271       case 'n':
272         Str[i] = '\n';
273         break;
274       case 'r':
275         Str[i] = '\r';
276         break;
277       case 't':
278         Str[i] = '\t';
279         break;
280       case 'v':
281         Str[i] = '\v';
282         break;
283       case '"':
284         Str[i] = '\"';
285         break;
286       case '\'':
287         Str[i] = '\'';
288         break;
289       case '\\':
290         Str[i] = '\\';
291         break;
292       }
293       // Nuke the second character.
294       Str.erase(Str.begin() + i + 1);
295     }
296   }
297 }
298 
299 /// UnescapeAliasString - Supports literal braces in InstAlias asm string which
300 /// are escaped with '\\' to avoid being interpreted as variants. Braces must
301 /// be unescaped before c++ code is generated as (e.g.):
302 ///
303 ///   AsmString = "foo \{$\x01\}";
304 ///
305 /// causes non-standard escape character warnings.
UnescapeAliasString(std::string & Str)306 static void UnescapeAliasString(std::string &Str) {
307   for (unsigned i = 0; i != Str.size(); ++i) {
308     if (Str[i] == '\\' && i != Str.size() - 1) {
309       switch (Str[i + 1]) {
310       default:
311         continue; // Don't execute the code after the switch.
312       case '{':
313         Str[i] = '{';
314         break;
315       case '}':
316         Str[i] = '}';
317         break;
318       }
319       // Nuke the second character.
320       Str.erase(Str.begin() + i + 1);
321     }
322   }
323 }
324 
EmitGetMnemonic(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)325 void AsmWriterEmitter::EmitGetMnemonic(
326     raw_ostream &O,
327     std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
328     unsigned &BitsLeft, unsigned &AsmStrBits) {
329   const Record *AsmWriter = Target.getAsmWriter();
330   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
331   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
332 
333   O << "/// getMnemonic - This method is automatically generated by "
334        "tablegen\n"
335        "/// from the instruction set description.\n"
336        "std::pair<const char *, uint64_t>\n"
337     << Target.getName() << ClassName
338     << "::getMnemonic(const MCInst &MI) const {\n";
339 
340   // Build an aggregate string, and build a table of offsets into it.
341   SequenceToOffsetTable<std::string> StringTable;
342 
343   /// OpcodeInfo - This encodes the index of the string to use for the first
344   /// chunk of the output as well as indices used for operand printing.
345   std::vector<uint64_t> OpcodeInfo(NumberedInstructions.size());
346   const unsigned OpcodeInfoBits = 64;
347 
348   // Add all strings to the string table upfront so it can generate an optimized
349   // representation.
350   for (AsmWriterInst &AWI : Instructions) {
351     if (AWI.Operands[0].OperandType == AsmWriterOperand::isLiteralTextOperand &&
352         !AWI.Operands[0].Str.empty()) {
353       std::string Str = AWI.Operands[0].Str;
354       UnescapeString(Str);
355       StringTable.add(Str);
356     }
357   }
358 
359   StringTable.layout();
360 
361   unsigned MaxStringIdx = 0;
362   for (AsmWriterInst &AWI : Instructions) {
363     unsigned Idx;
364     if (AWI.Operands[0].OperandType != AsmWriterOperand::isLiteralTextOperand ||
365         AWI.Operands[0].Str.empty()) {
366       // Something handled by the asmwriter printer, but with no leading string.
367       Idx = StringTable.get("");
368     } else {
369       std::string Str = AWI.Operands[0].Str;
370       UnescapeString(Str);
371       Idx = StringTable.get(Str);
372       MaxStringIdx = std::max(MaxStringIdx, Idx);
373 
374       // Nuke the string from the operand list.  It is now handled!
375       AWI.Operands.erase(AWI.Operands.begin());
376     }
377 
378     // Bias offset by one since we want 0 as a sentinel.
379     OpcodeInfo[AWI.CGIIndex] = Idx + 1;
380   }
381 
382   // Figure out how many bits we used for the string index.
383   AsmStrBits = Log2_32_Ceil(MaxStringIdx + 2);
384 
385   // To reduce code size, we compactify common instructions into a few bits
386   // in the opcode-indexed table.
387   BitsLeft = OpcodeInfoBits - AsmStrBits;
388 
389   while (true) {
390     std::vector<std::string> UniqueOperandCommands;
391     std::vector<std::vector<unsigned>> InstIdxs;
392     std::vector<unsigned> NumInstOpsHandled;
393     FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
394                               NumInstOpsHandled, PassSubtarget);
395 
396     // If we ran out of operands to print, we're done.
397     if (UniqueOperandCommands.empty())
398       break;
399 
400     // Compute the number of bits we need to represent these cases, this is
401     // ceil(log2(numentries)).
402     unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
403 
404     // If we don't have enough bits for this operand, don't include it.
405     if (NumBits > BitsLeft) {
406       LLVM_DEBUG(errs() << "Not enough bits to densely encode " << NumBits
407                         << " more bits\n");
408       break;
409     }
410 
411     // Otherwise, we can include this in the initial lookup table.  Add it in.
412     for (size_t i = 0, e = InstIdxs.size(); i != e; ++i) {
413       unsigned NumOps = NumInstOpsHandled[i];
414       for (unsigned Idx : InstIdxs[i]) {
415         OpcodeInfo[Instructions[Idx].CGIIndex] |=
416             (uint64_t)i << (OpcodeInfoBits - BitsLeft);
417         // Remove the info about this operand from the instruction.
418         AsmWriterInst &Inst = Instructions[Idx];
419         if (!Inst.Operands.empty()) {
420           assert(NumOps <= Inst.Operands.size() &&
421                  "Can't remove this many ops!");
422           Inst.Operands.erase(Inst.Operands.begin(),
423                               Inst.Operands.begin() + NumOps);
424         }
425       }
426     }
427     BitsLeft -= NumBits;
428 
429     // Remember the handlers for this set of operands.
430     TableDrivenOperandPrinters.push_back(std::move(UniqueOperandCommands));
431   }
432 
433   // Emit the string table itself.
434   StringTable.emitStringLiteralDef(O, "  static const char AsmStrs[]");
435 
436   // Emit the lookup tables in pieces to minimize wasted bytes.
437   unsigned BytesNeeded = ((OpcodeInfoBits - BitsLeft) + 7) / 8;
438   unsigned Table = 0, Shift = 0;
439   SmallString<128> BitsString;
440   raw_svector_ostream BitsOS(BitsString);
441   // If the total bits is more than 32-bits we need to use a 64-bit type.
442   BitsOS << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
443          << "_t Bits = 0;\n";
444   while (BytesNeeded != 0) {
445     // Figure out how big this table section needs to be, but no bigger than 4.
446     unsigned TableSize = std::min(llvm::bit_floor(BytesNeeded), 4u);
447     BytesNeeded -= TableSize;
448     TableSize *= 8; // Convert to bits;
449     uint64_t Mask = (1ULL << TableSize) - 1;
450     O << "  static const uint" << TableSize << "_t OpInfo" << Table
451       << "[] = {\n";
452     for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
453       O << "    " << ((OpcodeInfo[i] >> Shift) & Mask) << "U,\t// "
454         << NumberedInstructions[i]->TheDef->getName() << "\n";
455     }
456     O << "  };\n\n";
457     // Emit string to combine the individual table lookups.
458     BitsOS << "  Bits |= ";
459     // If the total bits is more than 32-bits we need to use a 64-bit type.
460     if (BitsLeft < (OpcodeInfoBits - 32))
461       BitsOS << "(uint64_t)";
462     BitsOS << "OpInfo" << Table << "[MI.getOpcode()] << " << Shift << ";\n";
463     // Prepare the shift for the next iteration and increment the table count.
464     Shift += TableSize;
465     ++Table;
466   }
467 
468   O << "  // Emit the opcode for the instruction.\n";
469   O << BitsString;
470 
471   // Make sure we don't return an invalid pointer if bits is 0
472   O << "  if (Bits == 0)\n"
473        "    return {nullptr, Bits};\n";
474 
475   // Return mnemonic string and bits.
476   O << "  return {AsmStrs+(Bits & " << (1 << AsmStrBits) - 1
477     << ")-1, Bits};\n\n";
478 
479   O << "}\n";
480 }
481 
482 /// EmitPrintInstruction - Generate the code for the "printInstruction" method
483 /// implementation. Destroys all instances of AsmWriterInst information, by
484 /// clearing the Instructions vector.
EmitPrintInstruction(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)485 void AsmWriterEmitter::EmitPrintInstruction(
486     raw_ostream &O,
487     std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
488     unsigned &BitsLeft, unsigned &AsmStrBits) {
489   const unsigned OpcodeInfoBits = 64;
490   const Record *AsmWriter = Target.getAsmWriter();
491   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
492   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
493 
494   // This function has some huge switch statements that causing excessive
495   // compile time in LLVM profile instrumenation build. This print function
496   // usually is not frequently called in compilation. Here we disable the
497   // profile instrumenation for this function.
498   O << "/// printInstruction - This method is automatically generated by "
499        "tablegen\n"
500        "/// from the instruction set description.\n"
501        "LLVM_NO_PROFILE_INSTRUMENT_FUNCTION\n"
502        "void "
503     << Target.getName() << ClassName
504     << "::printInstruction(const MCInst *MI, uint64_t Address, "
505     << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
506     << "raw_ostream &O) {\n";
507 
508   // Emit the initial tab character.
509   O << "  O << \"\\t\";\n\n";
510 
511   // Emit the starting string.
512   O << "  auto MnemonicInfo = getMnemonic(*MI);\n\n";
513   O << "  O << MnemonicInfo.first;\n\n";
514 
515   O << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
516     << "_t Bits = MnemonicInfo.second;\n"
517     << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n";
518 
519   // Output the table driven operand information.
520   BitsLeft = OpcodeInfoBits - AsmStrBits;
521   for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
522     std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
523 
524     // Compute the number of bits we need to represent these cases, this is
525     // ceil(log2(numentries)).
526     unsigned NumBits = Log2_32_Ceil(Commands.size());
527     assert(NumBits <= BitsLeft && "consistency error");
528 
529     // Emit code to extract this field from Bits.
530     O << "\n  // Fragment " << i << " encoded into " << NumBits << " bits for "
531       << Commands.size() << " unique commands.\n";
532 
533     if (Commands.size() == 2) {
534       // Emit two possibilitys with if/else.
535       O << "  if ((Bits >> " << (OpcodeInfoBits - BitsLeft) << ") & "
536         << ((1 << NumBits) - 1) << ") {\n"
537         << Commands[1] << "  } else {\n"
538         << Commands[0] << "  }\n\n";
539     } else if (Commands.size() == 1) {
540       // Emit a single possibility.
541       O << Commands[0] << "\n\n";
542     } else {
543       O << "  switch ((Bits >> " << (OpcodeInfoBits - BitsLeft) << ") & "
544         << ((1 << NumBits) - 1) << ") {\n"
545         << "  default: llvm_unreachable(\"Invalid command number.\");\n";
546 
547       // Print out all the cases.
548       for (unsigned j = 0, e = Commands.size(); j != e; ++j) {
549         O << "  case " << j << ":\n";
550         O << Commands[j];
551         O << "    break;\n";
552       }
553       O << "  }\n\n";
554     }
555     BitsLeft -= NumBits;
556   }
557 
558   // Okay, delete instructions with no operand info left.
559   llvm::erase_if(Instructions,
560                  [](AsmWriterInst &Inst) { return Inst.Operands.empty(); });
561 
562   // Because this is a vector, we want to emit from the end.  Reverse all of the
563   // elements in the vector.
564   std::reverse(Instructions.begin(), Instructions.end());
565 
566   // Now that we've emitted all of the operand info that fit into 64 bits, emit
567   // information for those instructions that are left.  This is a less dense
568   // encoding, but we expect the main 64-bit table to handle the majority of
569   // instructions.
570   if (!Instructions.empty()) {
571     // Find the opcode # of inline asm.
572     O << "  switch (MI->getOpcode()) {\n";
573     O << "  default: llvm_unreachable(\"Unexpected opcode.\");\n";
574     while (!Instructions.empty())
575       EmitInstructions(Instructions, O, PassSubtarget);
576 
577     O << "  }\n";
578   }
579 
580   O << "}\n";
581 }
582 
583 static void
emitRegisterNameString(raw_ostream & O,StringRef AltName,const std::deque<CodeGenRegister> & Registers)584 emitRegisterNameString(raw_ostream &O, StringRef AltName,
585                        const std::deque<CodeGenRegister> &Registers) {
586   SequenceToOffsetTable<std::string> StringTable;
587   SmallVector<std::string, 4> AsmNames(Registers.size());
588   unsigned i = 0;
589   for (const auto &Reg : Registers) {
590     std::string &AsmName = AsmNames[i++];
591 
592     // "NoRegAltName" is special. We don't need to do a lookup for that,
593     // as it's just a reference to the default register name.
594     if (AltName == "" || AltName == "NoRegAltName") {
595       AsmName = Reg.TheDef->getValueAsString("AsmName").str();
596       if (AsmName.empty())
597         AsmName = Reg.getName().str();
598     } else {
599       // Make sure the register has an alternate name for this index.
600       std::vector<const Record *> AltNameList =
601           Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
602       unsigned Idx = 0, e;
603       for (e = AltNameList.size();
604            Idx < e && (AltNameList[Idx]->getName() != AltName); ++Idx)
605         ;
606       // If the register has an alternate name for this index, use it.
607       // Otherwise, leave it empty as an error flag.
608       if (Idx < e) {
609         std::vector<StringRef> AltNames =
610             Reg.TheDef->getValueAsListOfStrings("AltNames");
611         if (AltNames.size() <= Idx)
612           PrintFatalError(Reg.TheDef->getLoc(),
613                           "Register definition missing alt name for '" +
614                               AltName + "'.");
615         AsmName = AltNames[Idx].str();
616       }
617     }
618     StringTable.add(AsmName);
619   }
620 
621   StringTable.layout();
622   StringTable.emitStringLiteralDef(O, Twine("  static const char AsmStrs") +
623                                           AltName + "[]");
624 
625   O << "  static const " << getMinimalTypeForRange(StringTable.size() - 1, 32)
626     << " RegAsmOffset" << AltName << "[] = {";
627   for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
628     if ((i % 14) == 0)
629       O << "\n    ";
630     O << StringTable.get(AsmNames[i]) << ", ";
631   }
632   O << "\n  };\n"
633     << "\n";
634 }
635 
EmitGetRegisterName(raw_ostream & O)636 void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
637   const Record *AsmWriter = Target.getAsmWriter();
638   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
639   const auto &Registers = Target.getRegBank().getRegisters();
640   ArrayRef<const Record *> AltNameIndices = Target.getRegAltNameIndices();
641   bool hasAltNames = AltNameIndices.size() > 1;
642   StringRef Namespace = Registers.front().TheDef->getValueAsString("Namespace");
643 
644   O << "\n\n/// getRegisterName - This method is automatically generated by "
645        "tblgen\n"
646        "/// from the register set description.  This returns the assembler "
647        "name\n"
648        "/// for the specified register.\n"
649        "const char *"
650     << Target.getName() << ClassName << "::";
651   if (hasAltNames)
652     O << "\ngetRegisterName(MCRegister Reg, unsigned AltIdx) {\n";
653   else
654     O << "getRegisterName(MCRegister Reg) {\n";
655   O << "  unsigned RegNo = Reg.id();\n"
656     << "  assert(RegNo && RegNo < " << (Registers.size() + 1)
657     << " && \"Invalid register number!\");\n"
658     << "\n";
659 
660   if (hasAltNames) {
661     for (const Record *R : AltNameIndices)
662       emitRegisterNameString(O, R->getName(), Registers);
663   } else {
664     emitRegisterNameString(O, "", Registers);
665   }
666 
667   if (hasAltNames) {
668     O << "  switch(AltIdx) {\n"
669       << "  default: llvm_unreachable(\"Invalid register alt name index!\");\n";
670     for (const Record *R : AltNameIndices) {
671       StringRef AltName = R->getName();
672       O << "  case ";
673       if (!Namespace.empty())
674         O << Namespace << "::";
675       O << AltName << ":\n";
676       if (R->isValueUnset("FallbackRegAltNameIndex"))
677         O << "    assert(*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
678           << "[RegNo-1]) &&\n"
679           << "           \"Invalid alt name index for register!\");\n";
680       else {
681         O << "    if (!*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
682           << "[RegNo-1]))\n"
683           << "      return getRegisterName(RegNo, ";
684         if (!Namespace.empty())
685           O << Namespace << "::";
686         O << R->getValueAsDef("FallbackRegAltNameIndex")->getName() << ");\n";
687       }
688       O << "    return AsmStrs" << AltName << "+RegAsmOffset" << AltName
689         << "[RegNo-1];\n";
690     }
691     O << "  }\n";
692   } else {
693     O << "  assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
694       << "          \"Invalid alt name index for register!\");\n"
695       << "  return AsmStrs+RegAsmOffset[RegNo-1];\n";
696   }
697   O << "}\n";
698 }
699 
700 namespace {
701 
702 // IAPrinter - Holds information about an InstAlias. Two InstAliases match if
703 // they both have the same conditionals. In which case, we cannot print out the
704 // alias for that pattern.
705 class IAPrinter {
706   std::map<StringRef, std::pair<int, int>> OpMap;
707 
708   std::vector<std::string> Conds;
709 
710   std::string Result;
711   std::string AsmString;
712 
713   unsigned NumMIOps;
714 
715 public:
IAPrinter(std::string R,std::string AS,unsigned NumMIOps)716   IAPrinter(std::string R, std::string AS, unsigned NumMIOps)
717       : Result(std::move(R)), AsmString(std::move(AS)), NumMIOps(NumMIOps) {}
718 
addCond(std::string C)719   void addCond(std::string C) { Conds.push_back(std::move(C)); }
getConds() const720   ArrayRef<std::string> getConds() const { return Conds; }
getCondCount() const721   size_t getCondCount() const { return Conds.size(); }
722 
addOperand(StringRef Op,int OpIdx,int PrintMethodIdx=-1)723   void addOperand(StringRef Op, int OpIdx, int PrintMethodIdx = -1) {
724     assert(OpIdx >= 0 && OpIdx < 0xFE && "Idx out of range");
725     assert(PrintMethodIdx >= -1 && PrintMethodIdx < 0xFF && "Idx out of range");
726     OpMap[Op] = {OpIdx, PrintMethodIdx};
727   }
728 
getNumMIOps()729   unsigned getNumMIOps() { return NumMIOps; }
730 
getResult()731   StringRef getResult() { return Result; }
732 
isOpMapped(StringRef Op)733   bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
getOpIndex(StringRef Op)734   int getOpIndex(StringRef Op) { return OpMap[Op].first; }
getOpData(StringRef Op)735   std::pair<int, int> &getOpData(StringRef Op) { return OpMap[Op]; }
736 
parseName(StringRef::iterator Start,StringRef::iterator End)737   std::pair<StringRef, StringRef::iterator> parseName(StringRef::iterator Start,
738                                                       StringRef::iterator End) {
739     StringRef::iterator I = Start;
740     StringRef::iterator Next;
741     if (*I == '{') {
742       // ${some_name}
743       Start = ++I;
744       while (I != End && *I != '}')
745         ++I;
746       Next = I;
747       // eat the final '}'
748       if (Next != End)
749         ++Next;
750     } else {
751       // $name, just eat the usual suspects.
752       while (I != End && (isAlnum(*I) || *I == '_'))
753         ++I;
754       Next = I;
755     }
756 
757     return {StringRef(Start, I - Start), Next};
758   }
759 
formatAliasString(uint32_t & UnescapedSize)760   std::string formatAliasString(uint32_t &UnescapedSize) {
761     // Directly mangle mapped operands into the string. Each operand is
762     // identified by a '$' sign followed by a byte identifying the number of the
763     // operand. We add one to the index to avoid zero bytes.
764     StringRef ASM(AsmString);
765     std::string OutString;
766     raw_string_ostream OS(OutString);
767     for (StringRef::iterator I = ASM.begin(), E = ASM.end(); I != E;) {
768       OS << *I;
769       ++UnescapedSize;
770       if (*I == '$') {
771         StringRef Name;
772         std::tie(Name, I) = parseName(++I, E);
773         assert(isOpMapped(Name) && "Unmapped operand!");
774 
775         int OpIndex, PrintIndex;
776         std::tie(OpIndex, PrintIndex) = getOpData(Name);
777         if (PrintIndex == -1) {
778           // Can use the default printOperand route.
779           OS << format("\\x%02X", (unsigned char)OpIndex + 1);
780           ++UnescapedSize;
781         } else {
782           // 3 bytes if a PrintMethod is needed: 0xFF, the MCInst operand
783           // number, and which of our pre-detected Methods to call.
784           OS << format("\\xFF\\x%02X\\x%02X", OpIndex + 1, PrintIndex + 1);
785           UnescapedSize += 3;
786         }
787       } else {
788         ++I;
789       }
790     }
791     return OutString;
792   }
793 
operator ==(const IAPrinter & RHS) const794   bool operator==(const IAPrinter &RHS) const {
795     if (NumMIOps != RHS.NumMIOps)
796       return false;
797     if (Conds.size() != RHS.Conds.size())
798       return false;
799 
800     unsigned Idx = 0;
801     for (const auto &str : Conds)
802       if (str != RHS.Conds[Idx++])
803         return false;
804 
805     return true;
806   }
807 };
808 
809 } // end anonymous namespace
810 
CountNumOperands(StringRef AsmString,unsigned Variant)811 static unsigned CountNumOperands(StringRef AsmString, unsigned Variant) {
812   return AsmString.count(' ') + AsmString.count('\t');
813 }
814 
815 namespace {
816 
817 struct AliasPriorityComparator {
818   typedef std::pair<CodeGenInstAlias, int> ValueType;
operator ()__anon992240d30511::AliasPriorityComparator819   bool operator()(const ValueType &LHS, const ValueType &RHS) const {
820     if (LHS.second == RHS.second) {
821       // We don't actually care about the order, but for consistency it
822       // shouldn't depend on pointer comparisons.
823       return LessRecordByID()(LHS.first.TheDef, RHS.first.TheDef);
824     }
825 
826     // Aliases with larger priorities should be considered first.
827     return LHS.second > RHS.second;
828   }
829 };
830 
831 } // end anonymous namespace
832 
EmitPrintAliasInstruction(raw_ostream & O)833 void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
834   const Record *AsmWriter = Target.getAsmWriter();
835 
836   O << "\n#ifdef PRINT_ALIAS_INSTR\n";
837   O << "#undef PRINT_ALIAS_INSTR\n\n";
838 
839   //////////////////////////////
840   // Gather information about aliases we need to print
841   //////////////////////////////
842 
843   // Emit the method that prints the alias instruction.
844   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
845   unsigned Variant = AsmWriter->getValueAsInt("Variant");
846   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
847 
848   // Create a map from the qualified name to a list of potential matches.
849   typedef std::set<std::pair<CodeGenInstAlias, int>, AliasPriorityComparator>
850       AliasWithPriority;
851   std::map<std::string, AliasWithPriority> AliasMap;
852   for (const Record *R : Records.getAllDerivedDefinitions("InstAlias")) {
853     int Priority = R->getValueAsInt("EmitPriority");
854     if (Priority < 1)
855       continue; // Aliases with priority 0 are never emitted.
856 
857     const DagInit *DI = R->getValueAsDag("ResultInst");
858     AliasMap[getQualifiedName(DI->getOperatorAsDef(R->getLoc()))].emplace(
859         CodeGenInstAlias(R, Target), Priority);
860   }
861 
862   // A map of which conditions need to be met for each instruction operand
863   // before it can be matched to the mnemonic.
864   std::map<std::string, std::vector<IAPrinter>> IAPrinterMap;
865 
866   std::vector<std::pair<std::string, bool>> PrintMethods;
867 
868   // A list of MCOperandPredicates for all operands in use, and the reverse map
869   std::vector<const Record *> MCOpPredicates;
870   DenseMap<const Record *, unsigned> MCOpPredicateMap;
871 
872   for (auto &Aliases : AliasMap) {
873     for (auto &Alias : Aliases.second) {
874       const CodeGenInstAlias &CGA = Alias.first;
875       unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
876       std::string FlatInstAsmString =
877           CodeGenInstruction::FlattenAsmStringVariants(
878               CGA.ResultInst->AsmString, Variant);
879       unsigned NumResultOps = CountNumOperands(FlatInstAsmString, Variant);
880 
881       std::string FlatAliasAsmString =
882           CodeGenInstruction::FlattenAsmStringVariants(CGA.AsmString, Variant);
883       UnescapeAliasString(FlatAliasAsmString);
884 
885       // Don't emit the alias if it has more operands than what it's aliasing.
886       if (NumResultOps < CountNumOperands(FlatAliasAsmString, Variant))
887         continue;
888 
889       StringRef Namespace = Target.getName();
890       unsigned NumMIOps = 0;
891       for (auto &ResultInstOpnd : CGA.ResultInst->Operands)
892         NumMIOps += ResultInstOpnd.MINumOperands;
893 
894       IAPrinter IAP(CGA.Result->getAsString(), FlatAliasAsmString, NumMIOps);
895 
896       unsigned MIOpNum = 0;
897       for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
898         // Skip over tied operands as they're not part of an alias declaration.
899         auto &Operands = CGA.ResultInst->Operands;
900         while (true) {
901           unsigned OpNum = Operands.getSubOperandNumber(MIOpNum).first;
902           if (Operands[OpNum].MINumOperands == 1 &&
903               Operands[OpNum].getTiedRegister() != -1) {
904             // Tied operands of different RegisterClass should be explicit
905             // within an instruction's syntax and so cannot be skipped.
906             int TiedOpNum = Operands[OpNum].getTiedRegister();
907             if (Operands[OpNum].Rec->getName() ==
908                 Operands[TiedOpNum].Rec->getName()) {
909               ++MIOpNum;
910               continue;
911             }
912           }
913           break;
914         }
915 
916         // Ignore unchecked result operands.
917         while (IAP.getCondCount() < MIOpNum)
918           IAP.addCond("AliasPatternCond::K_Ignore, 0");
919 
920         const CodeGenInstAlias::ResultOperand &RO = CGA.ResultOperands[i];
921 
922         switch (RO.Kind) {
923         case CodeGenInstAlias::ResultOperand::K_Record: {
924           const Record *Rec = RO.getRecord();
925           StringRef ROName = RO.getName();
926           int PrintMethodIdx = -1;
927 
928           // These two may have a PrintMethod, which we want to record (if it's
929           // the first time we've seen it) and provide an index for the aliasing
930           // code to use.
931           if (Rec->isSubClassOf("RegisterOperand") ||
932               Rec->isSubClassOf("Operand")) {
933             StringRef PrintMethod = Rec->getValueAsString("PrintMethod");
934             bool IsPCRel =
935                 Rec->getValueAsString("OperandType") == "OPERAND_PCREL";
936             if (PrintMethod != "" && PrintMethod != "printOperand") {
937               PrintMethodIdx = llvm::find_if(PrintMethods,
938                                              [&](auto &X) {
939                                                return X.first == PrintMethod;
940                                              }) -
941                                PrintMethods.begin();
942               if (static_cast<unsigned>(PrintMethodIdx) == PrintMethods.size())
943                 PrintMethods.emplace_back(PrintMethod.str(), IsPCRel);
944             }
945           }
946 
947           if (Rec->isSubClassOf("RegisterOperand"))
948             Rec = Rec->getValueAsDef("RegClass");
949           if (Rec->isSubClassOf("RegisterClass")) {
950             if (!IAP.isOpMapped(ROName)) {
951               IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
952               const Record *R = CGA.ResultOperands[i].getRecord();
953               if (R->isSubClassOf("RegisterOperand"))
954                 R = R->getValueAsDef("RegClass");
955               IAP.addCond(std::string(
956                   formatv("AliasPatternCond::K_RegClass, {}::{}RegClassID",
957                           Namespace, R->getName())));
958             } else {
959               IAP.addCond(std::string(formatv("AliasPatternCond::K_TiedReg, {}",
960                                               IAP.getOpIndex(ROName))));
961             }
962           } else {
963             // Assume all printable operands are desired for now. This can be
964             // overridden in the InstAlias instantiation if necessary.
965             IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
966 
967             // There might be an additional predicate on the MCOperand
968             unsigned &Entry = MCOpPredicateMap[Rec];
969             if (!Entry) {
970               if (!Rec->isValueUnset("MCOperandPredicate")) {
971                 MCOpPredicates.push_back(Rec);
972                 Entry = MCOpPredicates.size();
973               } else {
974                 break; // No conditions on this operand at all
975               }
976             }
977             IAP.addCond(
978                 std::string(formatv("AliasPatternCond::K_Custom, {}", Entry)));
979           }
980           break;
981         }
982         case CodeGenInstAlias::ResultOperand::K_Imm: {
983           // Just because the alias has an immediate result, doesn't mean the
984           // MCInst will. An MCExpr could be present, for example.
985           auto Imm = CGA.ResultOperands[i].getImm();
986           int32_t Imm32 = int32_t(Imm);
987           if (Imm != Imm32)
988             PrintFatalError("Matching an alias with an immediate out of the "
989                             "range of int32_t is not supported");
990           IAP.addCond(std::string(
991               formatv("AliasPatternCond::K_Imm, uint32_t({})", Imm32)));
992           break;
993         }
994         case CodeGenInstAlias::ResultOperand::K_Reg:
995           if (!CGA.ResultOperands[i].getRegister()) {
996             IAP.addCond(std::string(
997                 formatv("AliasPatternCond::K_Reg, {}::NoRegister", Namespace)));
998             break;
999           }
1000 
1001           StringRef Reg = CGA.ResultOperands[i].getRegister()->getName();
1002           IAP.addCond(std::string(
1003               formatv("AliasPatternCond::K_Reg, {}::{}", Namespace, Reg)));
1004           break;
1005         }
1006 
1007         MIOpNum += RO.getMINumOperands();
1008       }
1009 
1010       std::vector<const Record *> ReqFeatures;
1011       if (PassSubtarget) {
1012         // We only consider ReqFeatures predicates if PassSubtarget
1013         std::vector<const Record *> RF =
1014             CGA.TheDef->getValueAsListOfDefs("Predicates");
1015         copy_if(RF, std::back_inserter(ReqFeatures), [](const Record *R) {
1016           return R->getValueAsBit("AssemblerMatcherPredicate");
1017         });
1018       }
1019 
1020       for (const Record *R : ReqFeatures) {
1021         const DagInit *D = R->getValueAsDag("AssemblerCondDag");
1022         auto *Op = dyn_cast<DefInit>(D->getOperator());
1023         if (!Op)
1024           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1025         StringRef CombineType = Op->getDef()->getName();
1026         if (CombineType != "any_of" && CombineType != "all_of")
1027           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1028         if (D->getNumArgs() == 0)
1029           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1030         bool IsOr = CombineType == "any_of";
1031         // Change (any_of FeatureAll, (any_of ...)) to (any_of FeatureAll, ...).
1032         if (IsOr && D->getNumArgs() == 2 && isa<DagInit>(D->getArg(1))) {
1033           const DagInit *RHS = cast<DagInit>(D->getArg(1));
1034           SmallVector<std::pair<const Init *, const StringInit *>> Args{
1035               *D->getArgAndNames().begin()};
1036           llvm::append_range(Args, RHS->getArgAndNames());
1037           D = DagInit::get(D->getOperator(), Args);
1038         }
1039 
1040         for (auto *Arg : D->getArgs()) {
1041           bool IsNeg = false;
1042           if (auto *NotArg = dyn_cast<DagInit>(Arg)) {
1043             if (NotArg->getOperator()->getAsString() != "not" ||
1044                 NotArg->getNumArgs() != 1)
1045               PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1046             Arg = NotArg->getArg(0);
1047             IsNeg = true;
1048           }
1049           if (!isa<DefInit>(Arg) ||
1050               !cast<DefInit>(Arg)->getDef()->isSubClassOf("SubtargetFeature"))
1051             PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1052 
1053           IAP.addCond(std::string(formatv(
1054               "AliasPatternCond::K_{}{}Feature, {}::{}", IsOr ? "Or" : "",
1055               IsNeg ? "Neg" : "", Namespace, Arg->getAsString())));
1056         }
1057         // If an AssemblerPredicate with ors is used, note end of list should
1058         // these be combined.
1059         if (IsOr)
1060           IAP.addCond("AliasPatternCond::K_EndOrFeatures, 0");
1061       }
1062 
1063       IAPrinterMap[Aliases.first].push_back(std::move(IAP));
1064     }
1065   }
1066 
1067   //////////////////////////////
1068   // Write out the printAliasInstr function
1069   //////////////////////////////
1070 
1071   std::string Header;
1072   raw_string_ostream HeaderO(Header);
1073 
1074   HeaderO << "bool " << Target.getName() << ClassName
1075           << "::printAliasInstr(const MCInst"
1076           << " *MI, uint64_t Address, "
1077           << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
1078           << "raw_ostream &OS) {\n";
1079 
1080   std::string PatternsForOpcode;
1081   raw_string_ostream OpcodeO(PatternsForOpcode);
1082 
1083   unsigned PatternCount = 0;
1084   std::string Patterns;
1085   raw_string_ostream PatternO(Patterns);
1086 
1087   unsigned CondCount = 0;
1088   std::string Conds;
1089   raw_string_ostream CondO(Conds);
1090 
1091   // All flattened alias strings.
1092   std::map<std::string, uint32_t> AsmStringOffsets;
1093   std::vector<std::pair<uint32_t, std::string>> AsmStrings;
1094   size_t AsmStringsSize = 0;
1095 
1096   // Iterate over the opcodes in enum order so they are sorted by opcode for
1097   // binary search.
1098   for (const CodeGenInstruction *Inst : NumberedInstructions) {
1099     auto It = IAPrinterMap.find(getQualifiedName(Inst->TheDef));
1100     if (It == IAPrinterMap.end())
1101       continue;
1102     std::vector<IAPrinter> &IAPs = It->second;
1103     std::vector<IAPrinter *> UniqueIAPs;
1104 
1105     // Remove any ambiguous alias rules.
1106     for (auto &LHS : IAPs) {
1107       bool IsDup = false;
1108       for (const auto &RHS : IAPs) {
1109         if (&LHS != &RHS && LHS == RHS) {
1110           IsDup = true;
1111           break;
1112         }
1113       }
1114 
1115       if (!IsDup)
1116         UniqueIAPs.push_back(&LHS);
1117     }
1118 
1119     if (UniqueIAPs.empty())
1120       continue;
1121 
1122     unsigned PatternStart = PatternCount;
1123 
1124     // Insert the pattern start and opcode in the pattern list for debugging.
1125     PatternO << formatv("    // {} - {}\n", It->first, PatternStart);
1126 
1127     for (IAPrinter *IAP : UniqueIAPs) {
1128       // Start each condition list with a comment of the resulting pattern that
1129       // we're trying to match.
1130       unsigned CondStart = CondCount;
1131       CondO << formatv("    // {} - {}\n", IAP->getResult(), CondStart);
1132       for (const auto &Cond : IAP->getConds())
1133         CondO << "    {" << Cond << "},\n";
1134       CondCount += IAP->getCondCount();
1135 
1136       // After operands have been examined, re-encode the alias string with
1137       // escapes indicating how operands should be printed.
1138       uint32_t UnescapedSize = 0;
1139       std::string EncodedAsmString = IAP->formatAliasString(UnescapedSize);
1140       auto Insertion =
1141           AsmStringOffsets.try_emplace(EncodedAsmString, AsmStringsSize);
1142       if (Insertion.second) {
1143         // If the string is new, add it to the vector.
1144         AsmStrings.emplace_back(AsmStringsSize, EncodedAsmString);
1145         AsmStringsSize += UnescapedSize + 1;
1146       }
1147       unsigned AsmStrOffset = Insertion.first->second;
1148 
1149       PatternO << formatv("    {{{}, {}, {}, {} },\n", AsmStrOffset, CondStart,
1150                           IAP->getNumMIOps(), IAP->getCondCount());
1151       ++PatternCount;
1152     }
1153 
1154     OpcodeO << formatv("    {{{}, {}, {} },\n", It->first, PatternStart,
1155                        PatternCount - PatternStart);
1156   }
1157 
1158   if (PatternsForOpcode.empty()) {
1159     O << Header;
1160     O << "  return false;\n";
1161     O << "}\n\n";
1162     O << "#endif // PRINT_ALIAS_INSTR\n";
1163     return;
1164   }
1165 
1166   // Forward declare the validation method if needed.
1167   if (!MCOpPredicates.empty())
1168     O << "static bool " << Target.getName() << ClassName
1169       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1170       << "                  const MCSubtargetInfo &STI,\n"
1171       << "                  unsigned PredicateIndex);\n";
1172 
1173   O << Header;
1174   O.indent(2) << "static const PatternsForOpcode OpToPatterns[] = {\n";
1175   O << PatternsForOpcode;
1176   O.indent(2) << "};\n\n";
1177   O.indent(2) << "static const AliasPattern Patterns[] = {\n";
1178   O << Patterns;
1179   O.indent(2) << "};\n\n";
1180   O.indent(2) << "static const AliasPatternCond Conds[] = {\n";
1181   O << Conds;
1182   O.indent(2) << "};\n\n";
1183   O.indent(2) << "static const char AsmStrings[] =\n";
1184   for (const auto &P : AsmStrings) {
1185     O.indent(4) << "/* " << P.first << " */ \"" << P.second << "\\0\"\n";
1186   }
1187 
1188   O.indent(2) << ";\n\n";
1189 
1190   // Assert that the opcode table is sorted. Use a static local constructor to
1191   // ensure that the check only happens once on first run.
1192   O << "#ifndef NDEBUG\n";
1193   O.indent(2) << "static struct SortCheck {\n";
1194   O.indent(2) << "  SortCheck(ArrayRef<PatternsForOpcode> OpToPatterns) {\n";
1195   O.indent(2) << "    assert(std::is_sorted(\n";
1196   O.indent(2) << "               OpToPatterns.begin(), OpToPatterns.end(),\n";
1197   O.indent(2) << "               [](const PatternsForOpcode &L, const "
1198                  "PatternsForOpcode &R) {\n";
1199   O.indent(2) << "                 return L.Opcode < R.Opcode;\n";
1200   O.indent(2) << "               }) &&\n";
1201   O.indent(2) << "           \"tablegen failed to sort opcode patterns\");\n";
1202   O.indent(2) << "  }\n";
1203   O.indent(2) << "} sortCheckVar(OpToPatterns);\n";
1204   O << "#endif\n\n";
1205 
1206   O.indent(2) << "AliasMatchingData M {\n";
1207   O.indent(2) << "  ArrayRef(OpToPatterns),\n";
1208   O.indent(2) << "  ArrayRef(Patterns),\n";
1209   O.indent(2) << "  ArrayRef(Conds),\n";
1210   O.indent(2) << "  StringRef(AsmStrings, std::size(AsmStrings)),\n";
1211   if (MCOpPredicates.empty())
1212     O.indent(2) << "  nullptr,\n";
1213   else
1214     O.indent(2) << "  &" << Target.getName() << ClassName
1215                 << "ValidateMCOperand,\n";
1216   O.indent(2) << "};\n";
1217 
1218   O.indent(2) << "const char *AsmString = matchAliasPatterns(MI, "
1219               << (PassSubtarget ? "&STI" : "nullptr") << ", M);\n";
1220   O.indent(2) << "if (!AsmString) return false;\n\n";
1221 
1222   // Code that prints the alias, replacing the operands with the ones from the
1223   // MCInst.
1224   O << "  unsigned I = 0;\n";
1225   O << "  while (AsmString[I] != ' ' && AsmString[I] != '\\t' &&\n";
1226   O << "         AsmString[I] != '$' && AsmString[I] != '\\0')\n";
1227   O << "    ++I;\n";
1228   O << "  OS << '\\t' << StringRef(AsmString, I);\n";
1229 
1230   O << "  if (AsmString[I] != '\\0') {\n";
1231   O << "    if (AsmString[I] == ' ' || AsmString[I] == '\\t') {\n";
1232   O << "      OS << '\\t';\n";
1233   O << "      ++I;\n";
1234   O << "    }\n";
1235   O << "    do {\n";
1236   O << "      if (AsmString[I] == '$') {\n";
1237   O << "        ++I;\n";
1238   O << "        if (AsmString[I] == (char)0xff) {\n";
1239   O << "          ++I;\n";
1240   O << "          int OpIdx = AsmString[I++] - 1;\n";
1241   O << "          int PrintMethodIdx = AsmString[I++] - 1;\n";
1242   O << "          printCustomAliasOperand(MI, Address, OpIdx, PrintMethodIdx, ";
1243   O << (PassSubtarget ? "STI, " : "");
1244   O << "OS);\n";
1245   O << "        } else\n";
1246   O << "          printOperand(MI, unsigned(AsmString[I++]) - 1, ";
1247   O << (PassSubtarget ? "STI, " : "");
1248   O << "OS);\n";
1249   O << "      } else {\n";
1250   O << "        OS << AsmString[I++];\n";
1251   O << "      }\n";
1252   O << "    } while (AsmString[I] != '\\0');\n";
1253   O << "  }\n\n";
1254 
1255   O << "  return true;\n";
1256   O << "}\n\n";
1257 
1258   //////////////////////////////
1259   // Write out the printCustomAliasOperand function
1260   //////////////////////////////
1261 
1262   O << "void " << Target.getName() << ClassName << "::"
1263     << "printCustomAliasOperand(\n"
1264     << "         const MCInst *MI, uint64_t Address, unsigned OpIdx,\n"
1265     << "         unsigned PrintMethodIdx,\n"
1266     << (PassSubtarget ? "         const MCSubtargetInfo &STI,\n" : "")
1267     << "         raw_ostream &OS) {\n";
1268   if (PrintMethods.empty())
1269     O << "  llvm_unreachable(\"Unknown PrintMethod kind\");\n";
1270   else {
1271     O << "  switch (PrintMethodIdx) {\n"
1272       << "  default:\n"
1273       << "    llvm_unreachable(\"Unknown PrintMethod kind\");\n"
1274       << "    break;\n";
1275 
1276     for (unsigned i = 0; i < PrintMethods.size(); ++i) {
1277       O << "  case " << i << ":\n"
1278         << "    " << PrintMethods[i].first << "(MI, "
1279         << (PrintMethods[i].second ? "Address, " : "") << "OpIdx, "
1280         << (PassSubtarget ? "STI, " : "") << "OS);\n"
1281         << "    break;\n";
1282     }
1283     O << "  }\n";
1284   }
1285   O << "}\n\n";
1286 
1287   if (!MCOpPredicates.empty()) {
1288     O << "static bool " << Target.getName() << ClassName
1289       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1290       << "                  const MCSubtargetInfo &STI,\n"
1291       << "                  unsigned PredicateIndex) {\n"
1292       << "  switch (PredicateIndex) {\n"
1293       << "  default:\n"
1294       << "    llvm_unreachable(\"Unknown MCOperandPredicate kind\");\n"
1295       << "    break;\n";
1296 
1297     for (unsigned i = 0; i < MCOpPredicates.size(); ++i) {
1298       StringRef MCOpPred =
1299           MCOpPredicates[i]->getValueAsString("MCOperandPredicate");
1300       O << "  case " << i + 1 << ": {\n"
1301         << MCOpPred.data() << "\n"
1302         << "    }\n";
1303     }
1304     O << "  }\n"
1305       << "}\n\n";
1306   }
1307 
1308   O << "#endif // PRINT_ALIAS_INSTR\n";
1309 }
1310 
AsmWriterEmitter(const RecordKeeper & R)1311 AsmWriterEmitter::AsmWriterEmitter(const RecordKeeper &R)
1312     : Records(R), Target(R) {
1313   const Record *AsmWriter = Target.getAsmWriter();
1314   unsigned Variant = AsmWriter->getValueAsInt("Variant");
1315 
1316   // Get the instruction numbering.
1317   NumberedInstructions = Target.getInstructions();
1318 
1319   for (const auto &[Idx, I] : enumerate(NumberedInstructions)) {
1320     if (!I->AsmString.empty() && I->TheDef->getName() != "PHI")
1321       Instructions.emplace_back(*I, Idx, Variant);
1322   }
1323 }
1324 
run(raw_ostream & O)1325 void AsmWriterEmitter::run(raw_ostream &O) {
1326   std::vector<std::vector<std::string>> TableDrivenOperandPrinters;
1327   unsigned BitsLeft = 0;
1328   unsigned AsmStrBits = 0;
1329   emitSourceFileHeader("Assembly Writer Source Fragment", O, Records);
1330   EmitGetMnemonic(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1331   EmitPrintInstruction(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1332   EmitGetRegisterName(O);
1333   EmitPrintAliasInstruction(O);
1334 }
1335 
1336 static TableGen::Emitter::OptClass<AsmWriterEmitter>
1337     X("gen-asm-writer", "Generate assembly writer");
1338