xref: /freebsd/contrib/llvm-project/llvm/utils/TableGen/AsmWriterEmitter.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
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   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 
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.push_back(std::pair(FirstInst.CGI->Namespace.str() + "::" +
148                                          FirstInst.CGI->TheDef->getName().str(),
149                                      FirstInst.Operands[i]));
150 
151       for (const AsmWriterInst &AWI : SimilarInsts) {
152         OpsToPrint.push_back(std::pair(
153             AWI.CGI->Namespace.str() + "::" + 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(std::string(Inst.CGI->TheDef->getName()));
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   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> "
337     << Target.getName() << ClassName << "::getMnemonic(const MCInst *MI) {\n";
338 
339   // Build an aggregate string, and build a table of offsets into it.
340   SequenceToOffsetTable<std::string> StringTable;
341 
342   /// OpcodeInfo - This encodes the index of the string to use for the first
343   /// chunk of the output as well as indices used for operand printing.
344   std::vector<uint64_t> OpcodeInfo(NumberedInstructions.size());
345   const unsigned OpcodeInfoBits = 64;
346 
347   // Add all strings to the string table upfront so it can generate an optimized
348   // representation.
349   for (AsmWriterInst &AWI : Instructions) {
350     if (AWI.Operands[0].OperandType == AsmWriterOperand::isLiteralTextOperand &&
351         !AWI.Operands[0].Str.empty()) {
352       std::string Str = AWI.Operands[0].Str;
353       UnescapeString(Str);
354       StringTable.add(Str);
355     }
356   }
357 
358   StringTable.layout();
359 
360   unsigned MaxStringIdx = 0;
361   for (AsmWriterInst &AWI : Instructions) {
362     unsigned Idx;
363     if (AWI.Operands[0].OperandType != AsmWriterOperand::isLiteralTextOperand ||
364         AWI.Operands[0].Str.empty()) {
365       // Something handled by the asmwriter printer, but with no leading string.
366       Idx = StringTable.get("");
367     } else {
368       std::string Str = AWI.Operands[0].Str;
369       UnescapeString(Str);
370       Idx = StringTable.get(Str);
371       MaxStringIdx = std::max(MaxStringIdx, Idx);
372 
373       // Nuke the string from the operand list.  It is now handled!
374       AWI.Operands.erase(AWI.Operands.begin());
375     }
376 
377     // Bias offset by one since we want 0 as a sentinel.
378     OpcodeInfo[AWI.CGIIndex] = Idx + 1;
379   }
380 
381   // Figure out how many bits we used for the string index.
382   AsmStrBits = Log2_32_Ceil(MaxStringIdx + 2);
383 
384   // To reduce code size, we compactify common instructions into a few bits
385   // in the opcode-indexed table.
386   BitsLeft = OpcodeInfoBits - AsmStrBits;
387 
388   while (true) {
389     std::vector<std::string> UniqueOperandCommands;
390     std::vector<std::vector<unsigned>> InstIdxs;
391     std::vector<unsigned> NumInstOpsHandled;
392     FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
393                               NumInstOpsHandled, PassSubtarget);
394 
395     // If we ran out of operands to print, we're done.
396     if (UniqueOperandCommands.empty())
397       break;
398 
399     // Compute the number of bits we need to represent these cases, this is
400     // ceil(log2(numentries)).
401     unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
402 
403     // If we don't have enough bits for this operand, don't include it.
404     if (NumBits > BitsLeft) {
405       LLVM_DEBUG(errs() << "Not enough bits to densely encode " << NumBits
406                         << " more bits\n");
407       break;
408     }
409 
410     // Otherwise, we can include this in the initial lookup table.  Add it in.
411     for (size_t i = 0, e = InstIdxs.size(); i != e; ++i) {
412       unsigned NumOps = NumInstOpsHandled[i];
413       for (unsigned Idx : InstIdxs[i]) {
414         OpcodeInfo[Instructions[Idx].CGIIndex] |=
415             (uint64_t)i << (OpcodeInfoBits - BitsLeft);
416         // Remove the info about this operand from the instruction.
417         AsmWriterInst &Inst = Instructions[Idx];
418         if (!Inst.Operands.empty()) {
419           assert(NumOps <= Inst.Operands.size() &&
420                  "Can't remove this many ops!");
421           Inst.Operands.erase(Inst.Operands.begin(),
422                               Inst.Operands.begin() + NumOps);
423         }
424       }
425     }
426     BitsLeft -= NumBits;
427 
428     // Remember the handlers for this set of operands.
429     TableDrivenOperandPrinters.push_back(std::move(UniqueOperandCommands));
430   }
431 
432   // Emit the string table itself.
433   StringTable.emitStringLiteralDef(O, "  static const char AsmStrs[]");
434 
435   // Emit the lookup tables in pieces to minimize wasted bytes.
436   unsigned BytesNeeded = ((OpcodeInfoBits - BitsLeft) + 7) / 8;
437   unsigned Table = 0, Shift = 0;
438   SmallString<128> BitsString;
439   raw_svector_ostream BitsOS(BitsString);
440   // If the total bits is more than 32-bits we need to use a 64-bit type.
441   BitsOS << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
442          << "_t Bits = 0;\n";
443   while (BytesNeeded != 0) {
444     // Figure out how big this table section needs to be, but no bigger than 4.
445     unsigned TableSize = std::min(llvm::bit_floor(BytesNeeded), 4u);
446     BytesNeeded -= TableSize;
447     TableSize *= 8; // Convert to bits;
448     uint64_t Mask = (1ULL << TableSize) - 1;
449     O << "  static const uint" << TableSize << "_t OpInfo" << Table
450       << "[] = {\n";
451     for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
452       O << "    " << ((OpcodeInfo[i] >> Shift) & Mask) << "U,\t// "
453         << NumberedInstructions[i]->TheDef->getName() << "\n";
454     }
455     O << "  };\n\n";
456     // Emit string to combine the individual table lookups.
457     BitsOS << "  Bits |= ";
458     // If the total bits is more than 32-bits we need to use a 64-bit type.
459     if (BitsLeft < (OpcodeInfoBits - 32))
460       BitsOS << "(uint64_t)";
461     BitsOS << "OpInfo" << Table << "[MI->getOpcode()] << " << Shift << ";\n";
462     // Prepare the shift for the next iteration and increment the table count.
463     Shift += TableSize;
464     ++Table;
465   }
466 
467   O << "  // Emit the opcode for the instruction.\n";
468   O << BitsString;
469 
470   // Make sure we don't return an invalid pointer if bits is 0
471   O << "  if (Bits == 0)\n"
472        "    return {nullptr, Bits};\n";
473 
474   // Return mnemonic string and bits.
475   O << "  return {AsmStrs+(Bits & " << (1 << AsmStrBits) - 1
476     << ")-1, Bits};\n\n";
477 
478   O << "}\n";
479 }
480 
481 /// EmitPrintInstruction - Generate the code for the "printInstruction" method
482 /// implementation. Destroys all instances of AsmWriterInst information, by
483 /// clearing the Instructions vector.
EmitPrintInstruction(raw_ostream & O,std::vector<std::vector<std::string>> & TableDrivenOperandPrinters,unsigned & BitsLeft,unsigned & AsmStrBits)484 void AsmWriterEmitter::EmitPrintInstruction(
485     raw_ostream &O,
486     std::vector<std::vector<std::string>> &TableDrivenOperandPrinters,
487     unsigned &BitsLeft, unsigned &AsmStrBits) {
488   const unsigned OpcodeInfoBits = 64;
489   Record *AsmWriter = Target.getAsmWriter();
490   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
491   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
492 
493   // This function has some huge switch statements that causing excessive
494   // compile time in LLVM profile instrumenation build. This print function
495   // usually is not frequently called in compilation. Here we disable the
496   // profile instrumenation for this function.
497   O << "/// printInstruction - This method is automatically generated by "
498        "tablegen\n"
499        "/// from the instruction set description.\n"
500        "LLVM_NO_PROFILE_INSTRUMENT_FUNCTION\n"
501        "void "
502     << Target.getName() << ClassName
503     << "::printInstruction(const MCInst *MI, uint64_t Address, "
504     << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
505     << "raw_ostream &O) {\n";
506 
507   // Emit the initial tab character.
508   O << "  O << \"\\t\";\n\n";
509 
510   // Emit the starting string.
511   O << "  auto MnemonicInfo = getMnemonic(MI);\n\n";
512   O << "  O << MnemonicInfo.first;\n\n";
513 
514   O << "  uint" << ((BitsLeft < (OpcodeInfoBits - 32)) ? 64 : 32)
515     << "_t Bits = MnemonicInfo.second;\n"
516     << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n";
517 
518   // Output the table driven operand information.
519   BitsLeft = OpcodeInfoBits - AsmStrBits;
520   for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
521     std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
522 
523     // Compute the number of bits we need to represent these cases, this is
524     // ceil(log2(numentries)).
525     unsigned NumBits = Log2_32_Ceil(Commands.size());
526     assert(NumBits <= BitsLeft && "consistency error");
527 
528     // Emit code to extract this field from Bits.
529     O << "\n  // Fragment " << i << " encoded into " << NumBits << " bits for "
530       << Commands.size() << " unique commands.\n";
531 
532     if (Commands.size() == 2) {
533       // Emit two possibilitys with if/else.
534       O << "  if ((Bits >> " << (OpcodeInfoBits - BitsLeft) << ") & "
535         << ((1 << NumBits) - 1) << ") {\n"
536         << Commands[1] << "  } else {\n"
537         << Commands[0] << "  }\n\n";
538     } else if (Commands.size() == 1) {
539       // Emit a single possibility.
540       O << Commands[0] << "\n\n";
541     } else {
542       O << "  switch ((Bits >> " << (OpcodeInfoBits - BitsLeft) << ") & "
543         << ((1 << NumBits) - 1) << ") {\n"
544         << "  default: llvm_unreachable(\"Invalid command number.\");\n";
545 
546       // Print out all the cases.
547       for (unsigned j = 0, e = Commands.size(); j != e; ++j) {
548         O << "  case " << j << ":\n";
549         O << Commands[j];
550         O << "    break;\n";
551       }
552       O << "  }\n\n";
553     }
554     BitsLeft -= NumBits;
555   }
556 
557   // Okay, delete instructions with no operand info left.
558   llvm::erase_if(Instructions,
559                  [](AsmWriterInst &Inst) { return Inst.Operands.empty(); });
560 
561   // Because this is a vector, we want to emit from the end.  Reverse all of the
562   // elements in the vector.
563   std::reverse(Instructions.begin(), Instructions.end());
564 
565   // Now that we've emitted all of the operand info that fit into 64 bits, emit
566   // information for those instructions that are left.  This is a less dense
567   // encoding, but we expect the main 64-bit table to handle the majority of
568   // instructions.
569   if (!Instructions.empty()) {
570     // Find the opcode # of inline asm.
571     O << "  switch (MI->getOpcode()) {\n";
572     O << "  default: llvm_unreachable(\"Unexpected opcode.\");\n";
573     while (!Instructions.empty())
574       EmitInstructions(Instructions, O, PassSubtarget);
575 
576     O << "  }\n";
577   }
578 
579   O << "}\n";
580 }
581 
582 static void
emitRegisterNameString(raw_ostream & O,StringRef AltName,const std::deque<CodeGenRegister> & Registers)583 emitRegisterNameString(raw_ostream &O, StringRef AltName,
584                        const std::deque<CodeGenRegister> &Registers) {
585   SequenceToOffsetTable<std::string> StringTable;
586   SmallVector<std::string, 4> AsmNames(Registers.size());
587   unsigned i = 0;
588   for (const auto &Reg : Registers) {
589     std::string &AsmName = AsmNames[i++];
590 
591     // "NoRegAltName" is special. We don't need to do a lookup for that,
592     // as it's just a reference to the default register name.
593     if (AltName == "" || AltName == "NoRegAltName") {
594       AsmName = std::string(Reg.TheDef->getValueAsString("AsmName"));
595       if (AsmName.empty())
596         AsmName = std::string(Reg.getName());
597     } else {
598       // Make sure the register has an alternate name for this index.
599       std::vector<Record *> AltNameList =
600           Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
601       unsigned Idx = 0, e;
602       for (e = AltNameList.size();
603            Idx < e && (AltNameList[Idx]->getName() != AltName); ++Idx)
604         ;
605       // If the register has an alternate name for this index, use it.
606       // Otherwise, leave it empty as an error flag.
607       if (Idx < e) {
608         std::vector<StringRef> AltNames =
609             Reg.TheDef->getValueAsListOfStrings("AltNames");
610         if (AltNames.size() <= Idx)
611           PrintFatalError(Reg.TheDef->getLoc(),
612                           "Register definition missing alt name for '" +
613                               AltName + "'.");
614         AsmName = std::string(AltNames[Idx]);
615       }
616     }
617     StringTable.add(AsmName);
618   }
619 
620   StringTable.layout();
621   StringTable.emitStringLiteralDef(O, Twine("  static const char AsmStrs") +
622                                           AltName + "[]");
623 
624   O << "  static const " << getMinimalTypeForRange(StringTable.size() - 1, 32)
625     << " RegAsmOffset" << AltName << "[] = {";
626   for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
627     if ((i % 14) == 0)
628       O << "\n    ";
629     O << StringTable.get(AsmNames[i]) << ", ";
630   }
631   O << "\n  };\n"
632     << "\n";
633 }
634 
EmitGetRegisterName(raw_ostream & O)635 void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
636   Record *AsmWriter = Target.getAsmWriter();
637   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
638   const auto &Registers = Target.getRegBank().getRegisters();
639   const std::vector<Record *> &AltNameIndices = Target.getRegAltNameIndices();
640   bool hasAltNames = AltNameIndices.size() > 1;
641   StringRef Namespace = Registers.front().TheDef->getValueAsString("Namespace");
642 
643   O << "\n\n/// getRegisterName - This method is automatically generated by "
644        "tblgen\n"
645        "/// from the register set description.  This returns the assembler "
646        "name\n"
647        "/// for the specified register.\n"
648        "const char *"
649     << Target.getName() << ClassName << "::";
650   if (hasAltNames)
651     O << "\ngetRegisterName(MCRegister Reg, unsigned AltIdx) {\n";
652   else
653     O << "getRegisterName(MCRegister Reg) {\n";
654   O << "  unsigned RegNo = Reg.id();\n"
655     << "  assert(RegNo && RegNo < " << (Registers.size() + 1)
656     << " && \"Invalid register number!\");\n"
657     << "\n";
658 
659   if (hasAltNames) {
660     for (const Record *R : AltNameIndices)
661       emitRegisterNameString(O, R->getName(), Registers);
662   } else
663     emitRegisterNameString(O, "", Registers);
664 
665   if (hasAltNames) {
666     O << "  switch(AltIdx) {\n"
667       << "  default: llvm_unreachable(\"Invalid register alt name index!\");\n";
668     for (const Record *R : AltNameIndices) {
669       StringRef AltName = R->getName();
670       O << "  case ";
671       if (!Namespace.empty())
672         O << Namespace << "::";
673       O << AltName << ":\n";
674       if (R->isValueUnset("FallbackRegAltNameIndex"))
675         O << "    assert(*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
676           << "[RegNo-1]) &&\n"
677           << "           \"Invalid alt name index for register!\");\n";
678       else {
679         O << "    if (!*(AsmStrs" << AltName << "+RegAsmOffset" << AltName
680           << "[RegNo-1]))\n"
681           << "      return getRegisterName(RegNo, ";
682         if (!Namespace.empty())
683           O << Namespace << "::";
684         O << R->getValueAsDef("FallbackRegAltNameIndex")->getName() << ");\n";
685       }
686       O << "    return AsmStrs" << AltName << "+RegAsmOffset" << AltName
687         << "[RegNo-1];\n";
688     }
689     O << "  }\n";
690   } else {
691     O << "  assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
692       << "          \"Invalid alt name index for register!\");\n"
693       << "  return AsmStrs+RegAsmOffset[RegNo-1];\n";
694   }
695   O << "}\n";
696 }
697 
698 namespace {
699 
700 // IAPrinter - Holds information about an InstAlias. Two InstAliases match if
701 // they both have the same conditionals. In which case, we cannot print out the
702 // alias for that pattern.
703 class IAPrinter {
704   std::map<StringRef, std::pair<int, int>> OpMap;
705 
706   std::vector<std::string> Conds;
707 
708   std::string Result;
709   std::string AsmString;
710 
711   unsigned NumMIOps;
712 
713 public:
IAPrinter(std::string R,std::string AS,unsigned NumMIOps)714   IAPrinter(std::string R, std::string AS, unsigned NumMIOps)
715       : Result(std::move(R)), AsmString(std::move(AS)), NumMIOps(NumMIOps) {}
716 
addCond(std::string C)717   void addCond(std::string C) { Conds.push_back(std::move(C)); }
getConds() const718   ArrayRef<std::string> getConds() const { return Conds; }
getCondCount() const719   size_t getCondCount() const { return Conds.size(); }
720 
addOperand(StringRef Op,int OpIdx,int PrintMethodIdx=-1)721   void addOperand(StringRef Op, int OpIdx, int PrintMethodIdx = -1) {
722     assert(OpIdx >= 0 && OpIdx < 0xFE && "Idx out of range");
723     assert(PrintMethodIdx >= -1 && PrintMethodIdx < 0xFF && "Idx out of range");
724     OpMap[Op] = std::pair(OpIdx, PrintMethodIdx);
725   }
726 
getNumMIOps()727   unsigned getNumMIOps() { return NumMIOps; }
728 
getResult()729   StringRef getResult() { return Result; }
730 
isOpMapped(StringRef Op)731   bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
getOpIndex(StringRef Op)732   int getOpIndex(StringRef Op) { return OpMap[Op].first; }
getOpData(StringRef Op)733   std::pair<int, int> &getOpData(StringRef Op) { return OpMap[Op]; }
734 
parseName(StringRef::iterator Start,StringRef::iterator End)735   std::pair<StringRef, StringRef::iterator> parseName(StringRef::iterator Start,
736                                                       StringRef::iterator End) {
737     StringRef::iterator I = Start;
738     StringRef::iterator Next;
739     if (*I == '{') {
740       // ${some_name}
741       Start = ++I;
742       while (I != End && *I != '}')
743         ++I;
744       Next = I;
745       // eat the final '}'
746       if (Next != End)
747         ++Next;
748     } else {
749       // $name, just eat the usual suspects.
750       while (I != End && (isAlnum(*I) || *I == '_'))
751         ++I;
752       Next = I;
753     }
754 
755     return std::pair(StringRef(Start, I - Start), Next);
756   }
757 
formatAliasString(uint32_t & UnescapedSize)758   std::string formatAliasString(uint32_t &UnescapedSize) {
759     // Directly mangle mapped operands into the string. Each operand is
760     // identified by a '$' sign followed by a byte identifying the number of the
761     // operand. We add one to the index to avoid zero bytes.
762     StringRef ASM(AsmString);
763     std::string OutString;
764     raw_string_ostream OS(OutString);
765     for (StringRef::iterator I = ASM.begin(), E = ASM.end(); I != E;) {
766       OS << *I;
767       ++UnescapedSize;
768       if (*I == '$') {
769         StringRef Name;
770         std::tie(Name, I) = parseName(++I, E);
771         assert(isOpMapped(Name) && "Unmapped operand!");
772 
773         int OpIndex, PrintIndex;
774         std::tie(OpIndex, PrintIndex) = getOpData(Name);
775         if (PrintIndex == -1) {
776           // Can use the default printOperand route.
777           OS << format("\\x%02X", (unsigned char)OpIndex + 1);
778           ++UnescapedSize;
779         } else {
780           // 3 bytes if a PrintMethod is needed: 0xFF, the MCInst operand
781           // number, and which of our pre-detected Methods to call.
782           OS << format("\\xFF\\x%02X\\x%02X", OpIndex + 1, PrintIndex + 1);
783           UnescapedSize += 3;
784         }
785       } else {
786         ++I;
787       }
788     }
789     return OutString;
790   }
791 
operator ==(const IAPrinter & RHS) const792   bool operator==(const IAPrinter &RHS) const {
793     if (NumMIOps != RHS.NumMIOps)
794       return false;
795     if (Conds.size() != RHS.Conds.size())
796       return false;
797 
798     unsigned Idx = 0;
799     for (const auto &str : Conds)
800       if (str != RHS.Conds[Idx++])
801         return false;
802 
803     return true;
804   }
805 };
806 
807 } // end anonymous namespace
808 
CountNumOperands(StringRef AsmString,unsigned Variant)809 static unsigned CountNumOperands(StringRef AsmString, unsigned Variant) {
810   return AsmString.count(' ') + AsmString.count('\t');
811 }
812 
813 namespace {
814 
815 struct AliasPriorityComparator {
816   typedef std::pair<CodeGenInstAlias, int> ValueType;
operator ()__anon992240d30511::AliasPriorityComparator817   bool operator()(const ValueType &LHS, const ValueType &RHS) const {
818     if (LHS.second == RHS.second) {
819       // We don't actually care about the order, but for consistency it
820       // shouldn't depend on pointer comparisons.
821       return LessRecordByID()(LHS.first.TheDef, RHS.first.TheDef);
822     }
823 
824     // Aliases with larger priorities should be considered first.
825     return LHS.second > RHS.second;
826   }
827 };
828 
829 } // end anonymous namespace
830 
EmitPrintAliasInstruction(raw_ostream & O)831 void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
832   Record *AsmWriter = Target.getAsmWriter();
833 
834   O << "\n#ifdef PRINT_ALIAS_INSTR\n";
835   O << "#undef PRINT_ALIAS_INSTR\n\n";
836 
837   //////////////////////////////
838   // Gather information about aliases we need to print
839   //////////////////////////////
840 
841   // Emit the method that prints the alias instruction.
842   StringRef ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
843   unsigned Variant = AsmWriter->getValueAsInt("Variant");
844   bool PassSubtarget = AsmWriter->getValueAsInt("PassSubtarget");
845 
846   std::vector<Record *> AllInstAliases =
847       Records.getAllDerivedDefinitions("InstAlias");
848 
849   // Create a map from the qualified name to a list of potential matches.
850   typedef std::set<std::pair<CodeGenInstAlias, int>, AliasPriorityComparator>
851       AliasWithPriority;
852   std::map<std::string, AliasWithPriority> AliasMap;
853   for (Record *R : AllInstAliases) {
854     int Priority = R->getValueAsInt("EmitPriority");
855     if (Priority < 1)
856       continue; // Aliases with priority 0 are never emitted.
857 
858     const DagInit *DI = R->getValueAsDag("ResultInst");
859     AliasMap[getQualifiedName(DI->getOperatorAsDef(R->getLoc()))].insert(
860         std::pair(CodeGenInstAlias(R, Target), Priority));
861   }
862 
863   // A map of which conditions need to be met for each instruction operand
864   // before it can be matched to the mnemonic.
865   std::map<std::string, std::vector<IAPrinter>> IAPrinterMap;
866 
867   std::vector<std::pair<std::string, bool>> PrintMethods;
868 
869   // A list of MCOperandPredicates for all operands in use, and the reverse map
870   std::vector<const Record *> MCOpPredicates;
871   DenseMap<const Record *, unsigned> MCOpPredicateMap;
872 
873   for (auto &Aliases : AliasMap) {
874     // Collection of instruction alias rules. May contain ambiguous rules.
875     std::vector<IAPrinter> IAPs;
876 
877     for (auto &Alias : Aliases.second) {
878       const CodeGenInstAlias &CGA = Alias.first;
879       unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
880       std::string FlatInstAsmString =
881           CodeGenInstruction::FlattenAsmStringVariants(
882               CGA.ResultInst->AsmString, Variant);
883       unsigned NumResultOps = CountNumOperands(FlatInstAsmString, Variant);
884 
885       std::string FlatAliasAsmString =
886           CodeGenInstruction::FlattenAsmStringVariants(CGA.AsmString, Variant);
887       UnescapeAliasString(FlatAliasAsmString);
888 
889       // Don't emit the alias if it has more operands than what it's aliasing.
890       if (NumResultOps < CountNumOperands(FlatAliasAsmString, Variant))
891         continue;
892 
893       StringRef Namespace = Target.getName();
894       unsigned NumMIOps = 0;
895       for (auto &ResultInstOpnd : CGA.ResultInst->Operands)
896         NumMIOps += ResultInstOpnd.MINumOperands;
897 
898       IAPrinter IAP(CGA.Result->getAsString(), FlatAliasAsmString, NumMIOps);
899 
900       unsigned MIOpNum = 0;
901       for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
902         // Skip over tied operands as they're not part of an alias declaration.
903         auto &Operands = CGA.ResultInst->Operands;
904         while (true) {
905           unsigned OpNum = Operands.getSubOperandNumber(MIOpNum).first;
906           if (Operands[OpNum].MINumOperands == 1 &&
907               Operands[OpNum].getTiedRegister() != -1) {
908             // Tied operands of different RegisterClass should be explicit
909             // within an instruction's syntax and so cannot be skipped.
910             int TiedOpNum = Operands[OpNum].getTiedRegister();
911             if (Operands[OpNum].Rec->getName() ==
912                 Operands[TiedOpNum].Rec->getName()) {
913               ++MIOpNum;
914               continue;
915             }
916           }
917           break;
918         }
919 
920         // Ignore unchecked result operands.
921         while (IAP.getCondCount() < MIOpNum)
922           IAP.addCond("AliasPatternCond::K_Ignore, 0");
923 
924         const CodeGenInstAlias::ResultOperand &RO = CGA.ResultOperands[i];
925 
926         switch (RO.Kind) {
927         case CodeGenInstAlias::ResultOperand::K_Record: {
928           const Record *Rec = RO.getRecord();
929           StringRef ROName = RO.getName();
930           int PrintMethodIdx = -1;
931 
932           // These two may have a PrintMethod, which we want to record (if it's
933           // the first time we've seen it) and provide an index for the aliasing
934           // code to use.
935           if (Rec->isSubClassOf("RegisterOperand") ||
936               Rec->isSubClassOf("Operand")) {
937             StringRef PrintMethod = Rec->getValueAsString("PrintMethod");
938             bool IsPCRel =
939                 Rec->getValueAsString("OperandType") == "OPERAND_PCREL";
940             if (PrintMethod != "" && PrintMethod != "printOperand") {
941               PrintMethodIdx = llvm::find_if(PrintMethods,
942                                              [&](auto &X) {
943                                                return X.first == PrintMethod;
944                                              }) -
945                                PrintMethods.begin();
946               if (static_cast<unsigned>(PrintMethodIdx) == PrintMethods.size())
947                 PrintMethods.emplace_back(std::string(PrintMethod), IsPCRel);
948             }
949           }
950 
951           if (Rec->isSubClassOf("RegisterOperand"))
952             Rec = Rec->getValueAsDef("RegClass");
953           if (Rec->isSubClassOf("RegisterClass")) {
954             if (!IAP.isOpMapped(ROName)) {
955               IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
956               Record *R = CGA.ResultOperands[i].getRecord();
957               if (R->isSubClassOf("RegisterOperand"))
958                 R = R->getValueAsDef("RegClass");
959               IAP.addCond(std::string(
960                   formatv("AliasPatternCond::K_RegClass, {0}::{1}RegClassID",
961                           Namespace, R->getName())));
962             } else {
963               IAP.addCond(std::string(formatv(
964                   "AliasPatternCond::K_TiedReg, {0}", IAP.getOpIndex(ROName))));
965             }
966           } else {
967             // Assume all printable operands are desired for now. This can be
968             // overridden in the InstAlias instantiation if necessary.
969             IAP.addOperand(ROName, MIOpNum, PrintMethodIdx);
970 
971             // There might be an additional predicate on the MCOperand
972             unsigned Entry = MCOpPredicateMap[Rec];
973             if (!Entry) {
974               if (!Rec->isValueUnset("MCOperandPredicate")) {
975                 MCOpPredicates.push_back(Rec);
976                 Entry = MCOpPredicates.size();
977                 MCOpPredicateMap[Rec] = Entry;
978               } else
979                 break; // No conditions on this operand at all
980             }
981             IAP.addCond(
982                 std::string(formatv("AliasPatternCond::K_Custom, {0}", Entry)));
983           }
984           break;
985         }
986         case CodeGenInstAlias::ResultOperand::K_Imm: {
987           // Just because the alias has an immediate result, doesn't mean the
988           // MCInst will. An MCExpr could be present, for example.
989           auto Imm = CGA.ResultOperands[i].getImm();
990           int32_t Imm32 = int32_t(Imm);
991           if (Imm != Imm32)
992             PrintFatalError("Matching an alias with an immediate out of the "
993                             "range of int32_t is not supported");
994           IAP.addCond(std::string(
995               formatv("AliasPatternCond::K_Imm, uint32_t({0})", Imm32)));
996           break;
997         }
998         case CodeGenInstAlias::ResultOperand::K_Reg:
999           if (!CGA.ResultOperands[i].getRegister()) {
1000             IAP.addCond(std::string(formatv(
1001                 "AliasPatternCond::K_Reg, {0}::NoRegister", Namespace)));
1002             break;
1003           }
1004 
1005           StringRef Reg = CGA.ResultOperands[i].getRegister()->getName();
1006           IAP.addCond(std::string(
1007               formatv("AliasPatternCond::K_Reg, {0}::{1}", Namespace, Reg)));
1008           break;
1009         }
1010 
1011         MIOpNum += RO.getMINumOperands();
1012       }
1013 
1014       std::vector<Record *> ReqFeatures;
1015       if (PassSubtarget) {
1016         // We only consider ReqFeatures predicates if PassSubtarget
1017         std::vector<Record *> RF =
1018             CGA.TheDef->getValueAsListOfDefs("Predicates");
1019         copy_if(RF, std::back_inserter(ReqFeatures), [](Record *R) {
1020           return R->getValueAsBit("AssemblerMatcherPredicate");
1021         });
1022       }
1023 
1024       for (Record *const R : ReqFeatures) {
1025         const DagInit *D = R->getValueAsDag("AssemblerCondDag");
1026         auto *Op = dyn_cast<DefInit>(D->getOperator());
1027         if (!Op)
1028           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1029         StringRef CombineType = Op->getDef()->getName();
1030         if (CombineType != "any_of" && CombineType != "all_of")
1031           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1032         if (D->getNumArgs() == 0)
1033           PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1034         bool IsOr = CombineType == "any_of";
1035         // Change (any_of FeatureAll, (any_of ...)) to (any_of FeatureAll, ...).
1036         if (IsOr && D->getNumArgs() == 2 && isa<DagInit>(D->getArg(1))) {
1037           DagInit *RHS = cast<DagInit>(D->getArg(1));
1038           SmallVector<Init *> Args{D->getArg(0)};
1039           SmallVector<StringInit *> ArgNames{D->getArgName(0)};
1040           for (unsigned i = 0, e = RHS->getNumArgs(); i != e; ++i) {
1041             Args.push_back(RHS->getArg(i));
1042             ArgNames.push_back(RHS->getArgName(i));
1043           }
1044           D = DagInit::get(D->getOperator(), nullptr, Args, ArgNames);
1045         }
1046 
1047         for (auto *Arg : D->getArgs()) {
1048           bool IsNeg = false;
1049           if (auto *NotArg = dyn_cast<DagInit>(Arg)) {
1050             if (NotArg->getOperator()->getAsString() != "not" ||
1051                 NotArg->getNumArgs() != 1)
1052               PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1053             Arg = NotArg->getArg(0);
1054             IsNeg = true;
1055           }
1056           if (!isa<DefInit>(Arg) ||
1057               !cast<DefInit>(Arg)->getDef()->isSubClassOf("SubtargetFeature"))
1058             PrintFatalError(R->getLoc(), "Invalid AssemblerCondDag!");
1059 
1060           IAP.addCond(std::string(formatv(
1061               "AliasPatternCond::K_{0}{1}Feature, {2}::{3}", IsOr ? "Or" : "",
1062               IsNeg ? "Neg" : "", Namespace, Arg->getAsString())));
1063         }
1064         // If an AssemblerPredicate with ors is used, note end of list should
1065         // these be combined.
1066         if (IsOr)
1067           IAP.addCond("AliasPatternCond::K_EndOrFeatures, 0");
1068       }
1069 
1070       IAPrinterMap[Aliases.first].push_back(std::move(IAP));
1071     }
1072   }
1073 
1074   //////////////////////////////
1075   // Write out the printAliasInstr function
1076   //////////////////////////////
1077 
1078   std::string Header;
1079   raw_string_ostream HeaderO(Header);
1080 
1081   HeaderO << "bool " << Target.getName() << ClassName
1082           << "::printAliasInstr(const MCInst"
1083           << " *MI, uint64_t Address, "
1084           << (PassSubtarget ? "const MCSubtargetInfo &STI, " : "")
1085           << "raw_ostream &OS) {\n";
1086 
1087   std::string PatternsForOpcode;
1088   raw_string_ostream OpcodeO(PatternsForOpcode);
1089 
1090   unsigned PatternCount = 0;
1091   std::string Patterns;
1092   raw_string_ostream PatternO(Patterns);
1093 
1094   unsigned CondCount = 0;
1095   std::string Conds;
1096   raw_string_ostream CondO(Conds);
1097 
1098   // All flattened alias strings.
1099   std::map<std::string, uint32_t> AsmStringOffsets;
1100   std::vector<std::pair<uint32_t, std::string>> AsmStrings;
1101   size_t AsmStringsSize = 0;
1102 
1103   // Iterate over the opcodes in enum order so they are sorted by opcode for
1104   // binary search.
1105   for (const CodeGenInstruction *Inst : NumberedInstructions) {
1106     auto It = IAPrinterMap.find(getQualifiedName(Inst->TheDef));
1107     if (It == IAPrinterMap.end())
1108       continue;
1109     std::vector<IAPrinter> &IAPs = It->second;
1110     std::vector<IAPrinter *> UniqueIAPs;
1111 
1112     // Remove any ambiguous alias rules.
1113     for (auto &LHS : IAPs) {
1114       bool IsDup = false;
1115       for (const auto &RHS : IAPs) {
1116         if (&LHS != &RHS && LHS == RHS) {
1117           IsDup = true;
1118           break;
1119         }
1120       }
1121 
1122       if (!IsDup)
1123         UniqueIAPs.push_back(&LHS);
1124     }
1125 
1126     if (UniqueIAPs.empty())
1127       continue;
1128 
1129     unsigned PatternStart = PatternCount;
1130 
1131     // Insert the pattern start and opcode in the pattern list for debugging.
1132     PatternO << formatv("    // {0} - {1}\n", It->first, PatternStart);
1133 
1134     for (IAPrinter *IAP : UniqueIAPs) {
1135       // Start each condition list with a comment of the resulting pattern that
1136       // we're trying to match.
1137       unsigned CondStart = CondCount;
1138       CondO << formatv("    // {0} - {1}\n", IAP->getResult(), CondStart);
1139       for (const auto &Cond : IAP->getConds())
1140         CondO << "    {" << Cond << "},\n";
1141       CondCount += IAP->getCondCount();
1142 
1143       // After operands have been examined, re-encode the alias string with
1144       // escapes indicating how operands should be printed.
1145       uint32_t UnescapedSize = 0;
1146       std::string EncodedAsmString = IAP->formatAliasString(UnescapedSize);
1147       auto Insertion =
1148           AsmStringOffsets.insert({EncodedAsmString, AsmStringsSize});
1149       if (Insertion.second) {
1150         // If the string is new, add it to the vector.
1151         AsmStrings.push_back({AsmStringsSize, EncodedAsmString});
1152         AsmStringsSize += UnescapedSize + 1;
1153       }
1154       unsigned AsmStrOffset = Insertion.first->second;
1155 
1156       PatternO << formatv("    {{{0}, {1}, {2}, {3} },\n", AsmStrOffset,
1157                           CondStart, IAP->getNumMIOps(), IAP->getCondCount());
1158       ++PatternCount;
1159     }
1160 
1161     OpcodeO << formatv("    {{{0}, {1}, {2} },\n", It->first, PatternStart,
1162                        PatternCount - PatternStart);
1163   }
1164 
1165   if (PatternsForOpcode.empty()) {
1166     O << Header;
1167     O << "  return false;\n";
1168     O << "}\n\n";
1169     O << "#endif // PRINT_ALIAS_INSTR\n";
1170     return;
1171   }
1172 
1173   // Forward declare the validation method if needed.
1174   if (!MCOpPredicates.empty())
1175     O << "static bool " << Target.getName() << ClassName
1176       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1177       << "                  const MCSubtargetInfo &STI,\n"
1178       << "                  unsigned PredicateIndex);\n";
1179 
1180   O << Header;
1181   O.indent(2) << "static const PatternsForOpcode OpToPatterns[] = {\n";
1182   O << PatternsForOpcode;
1183   O.indent(2) << "};\n\n";
1184   O.indent(2) << "static const AliasPattern Patterns[] = {\n";
1185   O << Patterns;
1186   O.indent(2) << "};\n\n";
1187   O.indent(2) << "static const AliasPatternCond Conds[] = {\n";
1188   O << Conds;
1189   O.indent(2) << "};\n\n";
1190   O.indent(2) << "static const char AsmStrings[] =\n";
1191   for (const auto &P : AsmStrings) {
1192     O.indent(4) << "/* " << P.first << " */ \"" << P.second << "\\0\"\n";
1193   }
1194 
1195   O.indent(2) << ";\n\n";
1196 
1197   // Assert that the opcode table is sorted. Use a static local constructor to
1198   // ensure that the check only happens once on first run.
1199   O << "#ifndef NDEBUG\n";
1200   O.indent(2) << "static struct SortCheck {\n";
1201   O.indent(2) << "  SortCheck(ArrayRef<PatternsForOpcode> OpToPatterns) {\n";
1202   O.indent(2) << "    assert(std::is_sorted(\n";
1203   O.indent(2) << "               OpToPatterns.begin(), OpToPatterns.end(),\n";
1204   O.indent(2) << "               [](const PatternsForOpcode &L, const "
1205                  "PatternsForOpcode &R) {\n";
1206   O.indent(2) << "                 return L.Opcode < R.Opcode;\n";
1207   O.indent(2) << "               }) &&\n";
1208   O.indent(2) << "           \"tablegen failed to sort opcode patterns\");\n";
1209   O.indent(2) << "  }\n";
1210   O.indent(2) << "} sortCheckVar(OpToPatterns);\n";
1211   O << "#endif\n\n";
1212 
1213   O.indent(2) << "AliasMatchingData M {\n";
1214   O.indent(2) << "  ArrayRef(OpToPatterns),\n";
1215   O.indent(2) << "  ArrayRef(Patterns),\n";
1216   O.indent(2) << "  ArrayRef(Conds),\n";
1217   O.indent(2) << "  StringRef(AsmStrings, std::size(AsmStrings)),\n";
1218   if (MCOpPredicates.empty())
1219     O.indent(2) << "  nullptr,\n";
1220   else
1221     O.indent(2) << "  &" << Target.getName() << ClassName
1222                 << "ValidateMCOperand,\n";
1223   O.indent(2) << "};\n";
1224 
1225   O.indent(2) << "const char *AsmString = matchAliasPatterns(MI, "
1226               << (PassSubtarget ? "&STI" : "nullptr") << ", M);\n";
1227   O.indent(2) << "if (!AsmString) return false;\n\n";
1228 
1229   // Code that prints the alias, replacing the operands with the ones from the
1230   // MCInst.
1231   O << "  unsigned I = 0;\n";
1232   O << "  while (AsmString[I] != ' ' && AsmString[I] != '\\t' &&\n";
1233   O << "         AsmString[I] != '$' && AsmString[I] != '\\0')\n";
1234   O << "    ++I;\n";
1235   O << "  OS << '\\t' << StringRef(AsmString, I);\n";
1236 
1237   O << "  if (AsmString[I] != '\\0') {\n";
1238   O << "    if (AsmString[I] == ' ' || AsmString[I] == '\\t') {\n";
1239   O << "      OS << '\\t';\n";
1240   O << "      ++I;\n";
1241   O << "    }\n";
1242   O << "    do {\n";
1243   O << "      if (AsmString[I] == '$') {\n";
1244   O << "        ++I;\n";
1245   O << "        if (AsmString[I] == (char)0xff) {\n";
1246   O << "          ++I;\n";
1247   O << "          int OpIdx = AsmString[I++] - 1;\n";
1248   O << "          int PrintMethodIdx = AsmString[I++] - 1;\n";
1249   O << "          printCustomAliasOperand(MI, Address, OpIdx, PrintMethodIdx, ";
1250   O << (PassSubtarget ? "STI, " : "");
1251   O << "OS);\n";
1252   O << "        } else\n";
1253   O << "          printOperand(MI, unsigned(AsmString[I++]) - 1, ";
1254   O << (PassSubtarget ? "STI, " : "");
1255   O << "OS);\n";
1256   O << "      } else {\n";
1257   O << "        OS << AsmString[I++];\n";
1258   O << "      }\n";
1259   O << "    } while (AsmString[I] != '\\0');\n";
1260   O << "  }\n\n";
1261 
1262   O << "  return true;\n";
1263   O << "}\n\n";
1264 
1265   //////////////////////////////
1266   // Write out the printCustomAliasOperand function
1267   //////////////////////////////
1268 
1269   O << "void " << Target.getName() << ClassName << "::"
1270     << "printCustomAliasOperand(\n"
1271     << "         const MCInst *MI, uint64_t Address, unsigned OpIdx,\n"
1272     << "         unsigned PrintMethodIdx,\n"
1273     << (PassSubtarget ? "         const MCSubtargetInfo &STI,\n" : "")
1274     << "         raw_ostream &OS) {\n";
1275   if (PrintMethods.empty())
1276     O << "  llvm_unreachable(\"Unknown PrintMethod kind\");\n";
1277   else {
1278     O << "  switch (PrintMethodIdx) {\n"
1279       << "  default:\n"
1280       << "    llvm_unreachable(\"Unknown PrintMethod kind\");\n"
1281       << "    break;\n";
1282 
1283     for (unsigned i = 0; i < PrintMethods.size(); ++i) {
1284       O << "  case " << i << ":\n"
1285         << "    " << PrintMethods[i].first << "(MI, "
1286         << (PrintMethods[i].second ? "Address, " : "") << "OpIdx, "
1287         << (PassSubtarget ? "STI, " : "") << "OS);\n"
1288         << "    break;\n";
1289     }
1290     O << "  }\n";
1291   }
1292   O << "}\n\n";
1293 
1294   if (!MCOpPredicates.empty()) {
1295     O << "static bool " << Target.getName() << ClassName
1296       << "ValidateMCOperand(const MCOperand &MCOp,\n"
1297       << "                  const MCSubtargetInfo &STI,\n"
1298       << "                  unsigned PredicateIndex) {\n"
1299       << "  switch (PredicateIndex) {\n"
1300       << "  default:\n"
1301       << "    llvm_unreachable(\"Unknown MCOperandPredicate kind\");\n"
1302       << "    break;\n";
1303 
1304     for (unsigned i = 0; i < MCOpPredicates.size(); ++i) {
1305       StringRef MCOpPred =
1306           MCOpPredicates[i]->getValueAsString("MCOperandPredicate");
1307       O << "  case " << i + 1 << ": {\n"
1308         << MCOpPred.data() << "\n"
1309         << "    }\n";
1310     }
1311     O << "  }\n"
1312       << "}\n\n";
1313   }
1314 
1315   O << "#endif // PRINT_ALIAS_INSTR\n";
1316 }
1317 
AsmWriterEmitter(RecordKeeper & R)1318 AsmWriterEmitter::AsmWriterEmitter(RecordKeeper &R) : Records(R), Target(R) {
1319   Record *AsmWriter = Target.getAsmWriter();
1320   unsigned Variant = AsmWriter->getValueAsInt("Variant");
1321 
1322   // Get the instruction numbering.
1323   NumberedInstructions = Target.getInstructionsByEnumValue();
1324 
1325   for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
1326     const CodeGenInstruction *I = NumberedInstructions[i];
1327     if (!I->AsmString.empty() && I->TheDef->getName() != "PHI")
1328       Instructions.emplace_back(*I, i, Variant);
1329   }
1330 }
1331 
run(raw_ostream & O)1332 void AsmWriterEmitter::run(raw_ostream &O) {
1333   std::vector<std::vector<std::string>> TableDrivenOperandPrinters;
1334   unsigned BitsLeft = 0;
1335   unsigned AsmStrBits = 0;
1336   emitSourceFileHeader("Assembly Writer Source Fragment", O, Records);
1337   EmitGetMnemonic(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1338   EmitPrintInstruction(O, TableDrivenOperandPrinters, BitsLeft, AsmStrBits);
1339   EmitGetRegisterName(O);
1340   EmitPrintAliasInstruction(O);
1341 }
1342 
1343 static TableGen::Emitter::OptClass<AsmWriterEmitter>
1344     X("gen-asm-writer", "Generate assembly writer");
1345