AsmWriterEmitter.cpp revision 226633
1//===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This tablegen backend is emits an assembly printer for the current target.
11// Note that this is currently fairly skeletal, but will grow over time.
12//
13//===----------------------------------------------------------------------===//
14
15#include "AsmWriterEmitter.h"
16#include "AsmWriterInst.h"
17#include "CodeGenTarget.h"
18#include "StringToOffsetTable.h"
19#include "llvm/ADT/Twine.h"
20#include "llvm/Support/Debug.h"
21#include "llvm/Support/MathExtras.h"
22#include "llvm/TableGen/Error.h"
23#include "llvm/TableGen/Record.h"
24#include <algorithm>
25using namespace llvm;
26
27static void PrintCases(std::vector<std::pair<std::string,
28                       AsmWriterOperand> > &OpsToPrint, raw_ostream &O) {
29  O << "    case " << OpsToPrint.back().first << ": ";
30  AsmWriterOperand TheOp = OpsToPrint.back().second;
31  OpsToPrint.pop_back();
32
33  // Check to see if any other operands are identical in this list, and if so,
34  // emit a case label for them.
35  for (unsigned i = OpsToPrint.size(); i != 0; --i)
36    if (OpsToPrint[i-1].second == TheOp) {
37      O << "\n    case " << OpsToPrint[i-1].first << ": ";
38      OpsToPrint.erase(OpsToPrint.begin()+i-1);
39    }
40
41  // Finally, emit the code.
42  O << TheOp.getCode();
43  O << "break;\n";
44}
45
46
47/// EmitInstructions - Emit the last instruction in the vector and any other
48/// instructions that are suitably similar to it.
49static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
50                             raw_ostream &O) {
51  AsmWriterInst FirstInst = Insts.back();
52  Insts.pop_back();
53
54  std::vector<AsmWriterInst> SimilarInsts;
55  unsigned DifferingOperand = ~0;
56  for (unsigned i = Insts.size(); i != 0; --i) {
57    unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
58    if (DiffOp != ~1U) {
59      if (DifferingOperand == ~0U)  // First match!
60        DifferingOperand = DiffOp;
61
62      // If this differs in the same operand as the rest of the instructions in
63      // this class, move it to the SimilarInsts list.
64      if (DifferingOperand == DiffOp || DiffOp == ~0U) {
65        SimilarInsts.push_back(Insts[i-1]);
66        Insts.erase(Insts.begin()+i-1);
67      }
68    }
69  }
70
71  O << "  case " << FirstInst.CGI->Namespace << "::"
72    << FirstInst.CGI->TheDef->getName() << ":\n";
73  for (unsigned i = 0, e = SimilarInsts.size(); i != e; ++i)
74    O << "  case " << SimilarInsts[i].CGI->Namespace << "::"
75      << SimilarInsts[i].CGI->TheDef->getName() << ":\n";
76  for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
77    if (i != DifferingOperand) {
78      // If the operand is the same for all instructions, just print it.
79      O << "    " << FirstInst.Operands[i].getCode();
80    } else {
81      // If this is the operand that varies between all of the instructions,
82      // emit a switch for just this operand now.
83      O << "    switch (MI->getOpcode()) {\n";
84      std::vector<std::pair<std::string, AsmWriterOperand> > OpsToPrint;
85      OpsToPrint.push_back(std::make_pair(FirstInst.CGI->Namespace + "::" +
86                                          FirstInst.CGI->TheDef->getName(),
87                                          FirstInst.Operands[i]));
88
89      for (unsigned si = 0, e = SimilarInsts.size(); si != e; ++si) {
90        AsmWriterInst &AWI = SimilarInsts[si];
91        OpsToPrint.push_back(std::make_pair(AWI.CGI->Namespace+"::"+
92                                            AWI.CGI->TheDef->getName(),
93                                            AWI.Operands[i]));
94      }
95      std::reverse(OpsToPrint.begin(), OpsToPrint.end());
96      while (!OpsToPrint.empty())
97        PrintCases(OpsToPrint, O);
98      O << "    }";
99    }
100    O << "\n";
101  }
102  O << "    break;\n";
103}
104
105void AsmWriterEmitter::
106FindUniqueOperandCommands(std::vector<std::string> &UniqueOperandCommands,
107                          std::vector<unsigned> &InstIdxs,
108                          std::vector<unsigned> &InstOpsUsed) const {
109  InstIdxs.assign(NumberedInstructions.size(), ~0U);
110
111  // This vector parallels UniqueOperandCommands, keeping track of which
112  // instructions each case are used for.  It is a comma separated string of
113  // enums.
114  std::vector<std::string> InstrsForCase;
115  InstrsForCase.resize(UniqueOperandCommands.size());
116  InstOpsUsed.assign(UniqueOperandCommands.size(), 0);
117
118  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
119    const AsmWriterInst *Inst = getAsmWriterInstByID(i);
120    if (Inst == 0) continue;  // PHI, INLINEASM, PROLOG_LABEL, etc.
121
122    std::string Command;
123    if (Inst->Operands.empty())
124      continue;   // Instruction already done.
125
126    Command = "    " + Inst->Operands[0].getCode() + "\n";
127
128    // Check to see if we already have 'Command' in UniqueOperandCommands.
129    // If not, add it.
130    bool FoundIt = false;
131    for (unsigned idx = 0, e = UniqueOperandCommands.size(); idx != e; ++idx)
132      if (UniqueOperandCommands[idx] == Command) {
133        InstIdxs[i] = idx;
134        InstrsForCase[idx] += ", ";
135        InstrsForCase[idx] += Inst->CGI->TheDef->getName();
136        FoundIt = true;
137        break;
138      }
139    if (!FoundIt) {
140      InstIdxs[i] = UniqueOperandCommands.size();
141      UniqueOperandCommands.push_back(Command);
142      InstrsForCase.push_back(Inst->CGI->TheDef->getName());
143
144      // This command matches one operand so far.
145      InstOpsUsed.push_back(1);
146    }
147  }
148
149  // For each entry of UniqueOperandCommands, there is a set of instructions
150  // that uses it.  If the next command of all instructions in the set are
151  // identical, fold it into the command.
152  for (unsigned CommandIdx = 0, e = UniqueOperandCommands.size();
153       CommandIdx != e; ++CommandIdx) {
154
155    for (unsigned Op = 1; ; ++Op) {
156      // Scan for the first instruction in the set.
157      std::vector<unsigned>::iterator NIT =
158        std::find(InstIdxs.begin(), InstIdxs.end(), CommandIdx);
159      if (NIT == InstIdxs.end()) break;  // No commonality.
160
161      // If this instruction has no more operands, we isn't anything to merge
162      // into this command.
163      const AsmWriterInst *FirstInst =
164        getAsmWriterInstByID(NIT-InstIdxs.begin());
165      if (!FirstInst || FirstInst->Operands.size() == Op)
166        break;
167
168      // Otherwise, scan to see if all of the other instructions in this command
169      // set share the operand.
170      bool AllSame = true;
171      // Keep track of the maximum, number of operands or any
172      // instruction we see in the group.
173      size_t MaxSize = FirstInst->Operands.size();
174
175      for (NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx);
176           NIT != InstIdxs.end();
177           NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx)) {
178        // Okay, found another instruction in this command set.  If the operand
179        // matches, we're ok, otherwise bail out.
180        const AsmWriterInst *OtherInst =
181          getAsmWriterInstByID(NIT-InstIdxs.begin());
182
183        if (OtherInst &&
184            OtherInst->Operands.size() > FirstInst->Operands.size())
185          MaxSize = std::max(MaxSize, OtherInst->Operands.size());
186
187        if (!OtherInst || OtherInst->Operands.size() == Op ||
188            OtherInst->Operands[Op] != FirstInst->Operands[Op]) {
189          AllSame = false;
190          break;
191        }
192      }
193      if (!AllSame) break;
194
195      // Okay, everything in this command set has the same next operand.  Add it
196      // to UniqueOperandCommands and remember that it was consumed.
197      std::string Command = "    " + FirstInst->Operands[Op].getCode() + "\n";
198
199      UniqueOperandCommands[CommandIdx] += Command;
200      InstOpsUsed[CommandIdx]++;
201    }
202  }
203
204  // Prepend some of the instructions each case is used for onto the case val.
205  for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
206    std::string Instrs = InstrsForCase[i];
207    if (Instrs.size() > 70) {
208      Instrs.erase(Instrs.begin()+70, Instrs.end());
209      Instrs += "...";
210    }
211
212    if (!Instrs.empty())
213      UniqueOperandCommands[i] = "    // " + Instrs + "\n" +
214        UniqueOperandCommands[i];
215  }
216}
217
218
219static void UnescapeString(std::string &Str) {
220  for (unsigned i = 0; i != Str.size(); ++i) {
221    if (Str[i] == '\\' && i != Str.size()-1) {
222      switch (Str[i+1]) {
223      default: continue;  // Don't execute the code after the switch.
224      case 'a': Str[i] = '\a'; break;
225      case 'b': Str[i] = '\b'; break;
226      case 'e': Str[i] = 27; break;
227      case 'f': Str[i] = '\f'; break;
228      case 'n': Str[i] = '\n'; break;
229      case 'r': Str[i] = '\r'; break;
230      case 't': Str[i] = '\t'; break;
231      case 'v': Str[i] = '\v'; break;
232      case '"': Str[i] = '\"'; break;
233      case '\'': Str[i] = '\''; break;
234      case '\\': Str[i] = '\\'; break;
235      }
236      // Nuke the second character.
237      Str.erase(Str.begin()+i+1);
238    }
239  }
240}
241
242/// EmitPrintInstruction - Generate the code for the "printInstruction" method
243/// implementation.
244void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) {
245  CodeGenTarget Target(Records);
246  Record *AsmWriter = Target.getAsmWriter();
247  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
248  bool isMC = AsmWriter->getValueAsBit("isMCAsmWriter");
249  const char *MachineInstrClassName = isMC ? "MCInst" : "MachineInstr";
250
251  O <<
252  "/// printInstruction - This method is automatically generated by tablegen\n"
253  "/// from the instruction set description.\n"
254    "void " << Target.getName() << ClassName
255            << "::printInstruction(const " << MachineInstrClassName
256            << " *MI, raw_ostream &O) {\n";
257
258  std::vector<AsmWriterInst> Instructions;
259
260  for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
261         E = Target.inst_end(); I != E; ++I)
262    if (!(*I)->AsmString.empty() &&
263        (*I)->TheDef->getName() != "PHI")
264      Instructions.push_back(
265        AsmWriterInst(**I,
266                      AsmWriter->getValueAsInt("Variant"),
267                      AsmWriter->getValueAsInt("FirstOperandColumn"),
268                      AsmWriter->getValueAsInt("OperandSpacing")));
269
270  // Get the instruction numbering.
271  NumberedInstructions = Target.getInstructionsByEnumValue();
272
273  // Compute the CodeGenInstruction -> AsmWriterInst mapping.  Note that not
274  // all machine instructions are necessarily being printed, so there may be
275  // target instructions not in this map.
276  for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
277    CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i]));
278
279  // Build an aggregate string, and build a table of offsets into it.
280  StringToOffsetTable StringTable;
281
282  /// OpcodeInfo - This encodes the index of the string to use for the first
283  /// chunk of the output as well as indices used for operand printing.
284  std::vector<unsigned> OpcodeInfo;
285
286  unsigned MaxStringIdx = 0;
287  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
288    AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
289    unsigned Idx;
290    if (AWI == 0) {
291      // Something not handled by the asmwriter printer.
292      Idx = ~0U;
293    } else if (AWI->Operands[0].OperandType !=
294                        AsmWriterOperand::isLiteralTextOperand ||
295               AWI->Operands[0].Str.empty()) {
296      // Something handled by the asmwriter printer, but with no leading string.
297      Idx = StringTable.GetOrAddStringOffset("");
298    } else {
299      std::string Str = AWI->Operands[0].Str;
300      UnescapeString(Str);
301      Idx = StringTable.GetOrAddStringOffset(Str);
302      MaxStringIdx = std::max(MaxStringIdx, Idx);
303
304      // Nuke the string from the operand list.  It is now handled!
305      AWI->Operands.erase(AWI->Operands.begin());
306    }
307
308    // Bias offset by one since we want 0 as a sentinel.
309    OpcodeInfo.push_back(Idx+1);
310  }
311
312  // Figure out how many bits we used for the string index.
313  unsigned AsmStrBits = Log2_32_Ceil(MaxStringIdx+2);
314
315  // To reduce code size, we compactify common instructions into a few bits
316  // in the opcode-indexed table.
317  unsigned BitsLeft = 32-AsmStrBits;
318
319  std::vector<std::vector<std::string> > TableDrivenOperandPrinters;
320
321  while (1) {
322    std::vector<std::string> UniqueOperandCommands;
323    std::vector<unsigned> InstIdxs;
324    std::vector<unsigned> NumInstOpsHandled;
325    FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
326                              NumInstOpsHandled);
327
328    // If we ran out of operands to print, we're done.
329    if (UniqueOperandCommands.empty()) break;
330
331    // Compute the number of bits we need to represent these cases, this is
332    // ceil(log2(numentries)).
333    unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
334
335    // If we don't have enough bits for this operand, don't include it.
336    if (NumBits > BitsLeft) {
337      DEBUG(errs() << "Not enough bits to densely encode " << NumBits
338                   << " more bits\n");
339      break;
340    }
341
342    // Otherwise, we can include this in the initial lookup table.  Add it in.
343    BitsLeft -= NumBits;
344    for (unsigned i = 0, e = InstIdxs.size(); i != e; ++i)
345      if (InstIdxs[i] != ~0U)
346        OpcodeInfo[i] |= InstIdxs[i] << (BitsLeft+AsmStrBits);
347
348    // Remove the info about this operand.
349    for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
350      if (AsmWriterInst *Inst = getAsmWriterInstByID(i))
351        if (!Inst->Operands.empty()) {
352          unsigned NumOps = NumInstOpsHandled[InstIdxs[i]];
353          assert(NumOps <= Inst->Operands.size() &&
354                 "Can't remove this many ops!");
355          Inst->Operands.erase(Inst->Operands.begin(),
356                               Inst->Operands.begin()+NumOps);
357        }
358    }
359
360    // Remember the handlers for this set of operands.
361    TableDrivenOperandPrinters.push_back(UniqueOperandCommands);
362  }
363
364
365
366  O<<"  static const unsigned OpInfo[] = {\n";
367  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
368    O << "    " << OpcodeInfo[i] << "U,\t// "
369      << NumberedInstructions[i]->TheDef->getName() << "\n";
370  }
371  // Add a dummy entry so the array init doesn't end with a comma.
372  O << "    0U\n";
373  O << "  };\n\n";
374
375  // Emit the string itself.
376  O << "  const char *AsmStrs = \n";
377  StringTable.EmitString(O);
378  O << ";\n\n";
379
380  O << "  O << \"\\t\";\n\n";
381
382  O << "  // Emit the opcode for the instruction.\n"
383    << "  unsigned Bits = OpInfo[MI->getOpcode()];\n"
384    << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n"
385    << "  O << AsmStrs+(Bits & " << (1 << AsmStrBits)-1 << ")-1;\n\n";
386
387  // Output the table driven operand information.
388  BitsLeft = 32-AsmStrBits;
389  for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
390    std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
391
392    // Compute the number of bits we need to represent these cases, this is
393    // ceil(log2(numentries)).
394    unsigned NumBits = Log2_32_Ceil(Commands.size());
395    assert(NumBits <= BitsLeft && "consistency error");
396
397    // Emit code to extract this field from Bits.
398    BitsLeft -= NumBits;
399
400    O << "\n  // Fragment " << i << " encoded into " << NumBits
401      << " bits for " << Commands.size() << " unique commands.\n";
402
403    if (Commands.size() == 2) {
404      // Emit two possibilitys with if/else.
405      O << "  if ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
406        << ((1 << NumBits)-1) << ") {\n"
407        << Commands[1]
408        << "  } else {\n"
409        << Commands[0]
410        << "  }\n\n";
411    } else if (Commands.size() == 1) {
412      // Emit a single possibility.
413      O << Commands[0] << "\n\n";
414    } else {
415      O << "  switch ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
416        << ((1 << NumBits)-1) << ") {\n"
417        << "  default:   // unreachable.\n";
418
419      // Print out all the cases.
420      for (unsigned i = 0, e = Commands.size(); i != e; ++i) {
421        O << "  case " << i << ":\n";
422        O << Commands[i];
423        O << "    break;\n";
424      }
425      O << "  }\n\n";
426    }
427  }
428
429  // Okay, delete instructions with no operand info left.
430  for (unsigned i = 0, e = Instructions.size(); i != e; ++i) {
431    // Entire instruction has been emitted?
432    AsmWriterInst &Inst = Instructions[i];
433    if (Inst.Operands.empty()) {
434      Instructions.erase(Instructions.begin()+i);
435      --i; --e;
436    }
437  }
438
439
440  // Because this is a vector, we want to emit from the end.  Reverse all of the
441  // elements in the vector.
442  std::reverse(Instructions.begin(), Instructions.end());
443
444
445  // Now that we've emitted all of the operand info that fit into 32 bits, emit
446  // information for those instructions that are left.  This is a less dense
447  // encoding, but we expect the main 32-bit table to handle the majority of
448  // instructions.
449  if (!Instructions.empty()) {
450    // Find the opcode # of inline asm.
451    O << "  switch (MI->getOpcode()) {\n";
452    while (!Instructions.empty())
453      EmitInstructions(Instructions, O);
454
455    O << "  }\n";
456    O << "  return;\n";
457  }
458
459  O << "}\n";
460}
461
462static void
463emitRegisterNameString(raw_ostream &O, StringRef AltName,
464  const std::vector<CodeGenRegister*> &Registers) {
465  StringToOffsetTable StringTable;
466  O << "  static const unsigned RegAsmOffset" << AltName << "[] = {\n    ";
467  for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
468    const CodeGenRegister &Reg = *Registers[i];
469
470    std::string AsmName;
471    // "NoRegAltName" is special. We don't need to do a lookup for that,
472    // as it's just a reference to the default register name.
473    if (AltName == "" || AltName == "NoRegAltName") {
474      AsmName = Reg.TheDef->getValueAsString("AsmName");
475      if (AsmName.empty())
476        AsmName = Reg.getName();
477    } else {
478      // Make sure the register has an alternate name for this index.
479      std::vector<Record*> AltNameList =
480        Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
481      unsigned Idx = 0, e;
482      for (e = AltNameList.size();
483           Idx < e && (AltNameList[Idx]->getName() != AltName);
484           ++Idx)
485        ;
486      // If the register has an alternate name for this index, use it.
487      // Otherwise, leave it empty as an error flag.
488      if (Idx < e) {
489        std::vector<std::string> AltNames =
490          Reg.TheDef->getValueAsListOfStrings("AltNames");
491        if (AltNames.size() <= Idx)
492          throw TGError(Reg.TheDef->getLoc(),
493                        (Twine("Register definition missing alt name for '") +
494                        AltName + "'.").str());
495        AsmName = AltNames[Idx];
496      }
497    }
498
499    O << StringTable.GetOrAddStringOffset(AsmName);
500    if (((i + 1) % 14) == 0)
501      O << ",\n    ";
502    else
503      O << ", ";
504
505  }
506  O << "0\n"
507    << "  };\n"
508    << "\n";
509
510  O << "  const char *AsmStrs" << AltName << " =\n";
511  StringTable.EmitString(O);
512  O << ";\n";
513}
514
515void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
516  CodeGenTarget Target(Records);
517  Record *AsmWriter = Target.getAsmWriter();
518  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
519  const std::vector<CodeGenRegister*> &Registers =
520    Target.getRegBank().getRegisters();
521  std::vector<Record*> AltNameIndices = Target.getRegAltNameIndices();
522  bool hasAltNames = AltNameIndices.size() > 1;
523
524  O <<
525  "\n\n/// getRegisterName - This method is automatically generated by tblgen\n"
526  "/// from the register set description.  This returns the assembler name\n"
527  "/// for the specified register.\n"
528  "const char *" << Target.getName() << ClassName << "::";
529  if (hasAltNames)
530    O << "\ngetRegisterName(unsigned RegNo, unsigned AltIdx) {\n";
531  else
532    O << "getRegisterName(unsigned RegNo) {\n";
533  O << "  assert(RegNo && RegNo < " << (Registers.size()+1)
534    << " && \"Invalid register number!\");\n"
535    << "\n";
536
537  if (hasAltNames) {
538    for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i)
539      emitRegisterNameString(O, AltNameIndices[i]->getName(), Registers);
540  } else
541    emitRegisterNameString(O, "", Registers);
542
543  if (hasAltNames) {
544    O << "  const unsigned *RegAsmOffset;\n"
545      << "  const char *AsmStrs;\n"
546      << "  switch(AltIdx) {\n"
547      << "  default: assert(0 && \"Invalid register alt name index!\");\n";
548    for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i) {
549      StringRef Namespace = AltNameIndices[1]->getValueAsString("Namespace");
550      StringRef AltName(AltNameIndices[i]->getName());
551      O << "  case " << Namespace << "::" << AltName
552        << ":\n"
553        << "    AsmStrs = AsmStrs" << AltName  << ";\n"
554        << "    RegAsmOffset = RegAsmOffset" << AltName << ";\n"
555        << "    break;\n";
556    }
557    O << "}\n";
558  }
559
560  O << "  assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
561    << "          \"Invalid alt name index for register!\");\n"
562    << "  return AsmStrs+RegAsmOffset[RegNo-1];\n"
563    << "}\n";
564}
565
566void AsmWriterEmitter::EmitGetInstructionName(raw_ostream &O) {
567  CodeGenTarget Target(Records);
568  Record *AsmWriter = Target.getAsmWriter();
569  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
570
571  const std::vector<const CodeGenInstruction*> &NumberedInstructions =
572    Target.getInstructionsByEnumValue();
573
574  StringToOffsetTable StringTable;
575  O <<
576"\n\n#ifdef GET_INSTRUCTION_NAME\n"
577"#undef GET_INSTRUCTION_NAME\n\n"
578"/// getInstructionName: This method is automatically generated by tblgen\n"
579"/// from the instruction set description.  This returns the enum name of the\n"
580"/// specified instruction.\n"
581  "const char *" << Target.getName() << ClassName
582  << "::getInstructionName(unsigned Opcode) {\n"
583  << "  assert(Opcode < " << NumberedInstructions.size()
584  << " && \"Invalid instruction number!\");\n"
585  << "\n"
586  << "  static const unsigned InstAsmOffset[] = {";
587  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
588    const CodeGenInstruction &Inst = *NumberedInstructions[i];
589
590    std::string AsmName = Inst.TheDef->getName();
591    if ((i % 14) == 0)
592      O << "\n    ";
593
594    O << StringTable.GetOrAddStringOffset(AsmName) << ", ";
595  }
596  O << "0\n"
597  << "  };\n"
598  << "\n";
599
600  O << "  const char *Strs =\n";
601  StringTable.EmitString(O);
602  O << ";\n";
603
604  O << "  return Strs+InstAsmOffset[Opcode];\n"
605  << "}\n\n#endif\n";
606}
607
608namespace {
609// IAPrinter - Holds information about an InstAlias. Two InstAliases match if
610// they both have the same conditionals. In which case, we cannot print out the
611// alias for that pattern.
612class IAPrinter {
613  std::vector<std::string> Conds;
614  std::map<StringRef, unsigned> OpMap;
615  std::string Result;
616  std::string AsmString;
617  std::vector<Record*> ReqFeatures;
618public:
619  IAPrinter(std::string R, std::string AS)
620    : Result(R), AsmString(AS) {}
621
622  void addCond(const std::string &C) { Conds.push_back(C); }
623
624  void addOperand(StringRef Op, unsigned Idx) { OpMap[Op] = Idx; }
625  unsigned getOpIndex(StringRef Op) { return OpMap[Op]; }
626  bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
627
628  void print(raw_ostream &O) {
629    if (Conds.empty() && ReqFeatures.empty()) {
630      O.indent(6) << "return true;\n";
631      return;
632    }
633
634    O << "if (";
635
636    for (std::vector<std::string>::iterator
637           I = Conds.begin(), E = Conds.end(); I != E; ++I) {
638      if (I != Conds.begin()) {
639        O << " &&\n";
640        O.indent(8);
641      }
642
643      O << *I;
644    }
645
646    O << ") {\n";
647    O.indent(6) << "// " << Result << "\n";
648    O.indent(6) << "AsmString = \"" << AsmString << "\";\n";
649
650    for (std::map<StringRef, unsigned>::iterator
651           I = OpMap.begin(), E = OpMap.end(); I != E; ++I)
652      O.indent(6) << "OpMap.push_back(std::make_pair(\"" << I->first << "\", "
653                  << I->second << "));\n";
654
655    O.indent(6) << "break;\n";
656    O.indent(4) << '}';
657  }
658
659  bool operator==(const IAPrinter &RHS) {
660    if (Conds.size() != RHS.Conds.size())
661      return false;
662
663    unsigned Idx = 0;
664    for (std::vector<std::string>::iterator
665           I = Conds.begin(), E = Conds.end(); I != E; ++I)
666      if (*I != RHS.Conds[Idx++])
667        return false;
668
669    return true;
670  }
671
672  bool operator()(const IAPrinter &RHS) {
673    if (Conds.size() < RHS.Conds.size())
674      return true;
675
676    unsigned Idx = 0;
677    for (std::vector<std::string>::iterator
678           I = Conds.begin(), E = Conds.end(); I != E; ++I)
679      if (*I != RHS.Conds[Idx++])
680        return *I < RHS.Conds[Idx++];
681
682    return false;
683  }
684};
685
686} // end anonymous namespace
687
688static void EmitGetMapOperandNumber(raw_ostream &O) {
689  O << "static unsigned getMapOperandNumber("
690    << "const SmallVectorImpl<std::pair<StringRef, unsigned> > &OpMap,\n";
691  O << "                                    StringRef Name) {\n";
692  O << "  for (SmallVectorImpl<std::pair<StringRef, unsigned> >::"
693    << "const_iterator\n";
694  O << "         I = OpMap.begin(), E = OpMap.end(); I != E; ++I)\n";
695  O << "    if (I->first == Name)\n";
696  O << "      return I->second;\n";
697  O << "  assert(false && \"Operand not in map!\");\n";
698  O << "  return 0;\n";
699  O << "}\n\n";
700}
701
702void AsmWriterEmitter::EmitRegIsInRegClass(raw_ostream &O) {
703  CodeGenTarget Target(Records);
704
705  // Enumerate the register classes.
706  ArrayRef<CodeGenRegisterClass*> RegisterClasses =
707    Target.getRegBank().getRegClasses();
708
709  O << "namespace { // Register classes\n";
710  O << "  enum RegClass {\n";
711
712  // Emit the register enum value for each RegisterClass.
713  for (unsigned I = 0, E = RegisterClasses.size(); I != E; ++I) {
714    if (I != 0) O << ",\n";
715    O << "    RC_" << RegisterClasses[I]->getName();
716  }
717
718  O << "\n  };\n";
719  O << "} // end anonymous namespace\n\n";
720
721  // Emit a function that returns 'true' if a regsiter is part of a particular
722  // register class. I.e., RAX is part of GR64 on X86.
723  O << "static bool regIsInRegisterClass"
724    << "(unsigned RegClass, unsigned Reg) {\n";
725
726  // Emit the switch that checks if a register belongs to a particular register
727  // class.
728  O << "  switch (RegClass) {\n";
729  O << "  default: break;\n";
730
731  for (unsigned I = 0, E = RegisterClasses.size(); I != E; ++I) {
732    const CodeGenRegisterClass &RC = *RegisterClasses[I];
733
734    // Give the register class a legal C name if it's anonymous.
735    std::string Name = RC.getName();
736    O << "  case RC_" << Name << ":\n";
737
738    // Emit the register list now.
739    unsigned IE = RC.getOrder().size();
740    if (IE == 1) {
741      O << "    if (Reg == " << getQualifiedName(RC.getOrder()[0]) << ")\n";
742      O << "      return true;\n";
743    } else {
744      O << "    switch (Reg) {\n";
745      O << "    default: break;\n";
746
747      for (unsigned II = 0; II != IE; ++II) {
748        Record *Reg = RC.getOrder()[II];
749        O << "    case " << getQualifiedName(Reg) << ":\n";
750      }
751
752      O << "      return true;\n";
753      O << "    }\n";
754    }
755
756    O << "    break;\n";
757  }
758
759  O << "  }\n\n";
760  O << "  return false;\n";
761  O << "}\n\n";
762}
763
764static unsigned CountNumOperands(StringRef AsmString) {
765  unsigned NumOps = 0;
766  std::pair<StringRef, StringRef> ASM = AsmString.split(' ');
767
768  while (!ASM.second.empty()) {
769    ++NumOps;
770    ASM = ASM.second.split(' ');
771  }
772
773  return NumOps;
774}
775
776static unsigned CountResultNumOperands(StringRef AsmString) {
777  unsigned NumOps = 0;
778  std::pair<StringRef, StringRef> ASM = AsmString.split('\t');
779
780  if (!ASM.second.empty()) {
781    size_t I = ASM.second.find('{');
782    StringRef Str = ASM.second;
783    if (I != StringRef::npos)
784      Str = ASM.second.substr(I, ASM.second.find('|', I));
785
786    ASM = Str.split(' ');
787
788    do {
789      ++NumOps;
790      ASM = ASM.second.split(' ');
791    } while (!ASM.second.empty());
792  }
793
794  return NumOps;
795}
796
797void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
798  CodeGenTarget Target(Records);
799  Record *AsmWriter = Target.getAsmWriter();
800
801  if (!AsmWriter->getValueAsBit("isMCAsmWriter"))
802    return;
803
804  O << "\n#ifdef PRINT_ALIAS_INSTR\n";
805  O << "#undef PRINT_ALIAS_INSTR\n\n";
806
807  EmitRegIsInRegClass(O);
808
809  // Emit the method that prints the alias instruction.
810  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
811
812  std::vector<Record*> AllInstAliases =
813    Records.getAllDerivedDefinitions("InstAlias");
814
815  // Create a map from the qualified name to a list of potential matches.
816  std::map<std::string, std::vector<CodeGenInstAlias*> > AliasMap;
817  for (std::vector<Record*>::iterator
818         I = AllInstAliases.begin(), E = AllInstAliases.end(); I != E; ++I) {
819    CodeGenInstAlias *Alias = new CodeGenInstAlias(*I, Target);
820    const Record *R = *I;
821    if (!R->getValueAsBit("EmitAlias"))
822      continue; // We were told not to emit the alias, but to emit the aliasee.
823    const DagInit *DI = R->getValueAsDag("ResultInst");
824    const DefInit *Op = dynamic_cast<const DefInit*>(DI->getOperator());
825    AliasMap[getQualifiedName(Op->getDef())].push_back(Alias);
826  }
827
828  // A map of which conditions need to be met for each instruction operand
829  // before it can be matched to the mnemonic.
830  std::map<std::string, std::vector<IAPrinter*> > IAPrinterMap;
831
832  for (std::map<std::string, std::vector<CodeGenInstAlias*> >::iterator
833         I = AliasMap.begin(), E = AliasMap.end(); I != E; ++I) {
834    std::vector<CodeGenInstAlias*> &Aliases = I->second;
835
836    for (std::vector<CodeGenInstAlias*>::iterator
837           II = Aliases.begin(), IE = Aliases.end(); II != IE; ++II) {
838      const CodeGenInstAlias *CGA = *II;
839      unsigned LastOpNo = CGA->ResultInstOperandIndex.size();
840      unsigned NumResultOps =
841        CountResultNumOperands(CGA->ResultInst->AsmString);
842
843      // Don't emit the alias if it has more operands than what it's aliasing.
844      if (NumResultOps < CountNumOperands(CGA->AsmString))
845        continue;
846
847      IAPrinter *IAP = new IAPrinter(CGA->Result->getAsString(),
848                                     CGA->AsmString);
849
850      std::string Cond;
851      Cond = std::string("MI->getNumOperands() == ") + llvm::utostr(LastOpNo);
852      IAP->addCond(Cond);
853
854      std::map<StringRef, unsigned> OpMap;
855      bool CantHandle = false;
856
857      for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
858        const CodeGenInstAlias::ResultOperand &RO = CGA->ResultOperands[i];
859
860        switch (RO.Kind) {
861        default: assert(0 && "unexpected InstAlias operand kind");
862        case CodeGenInstAlias::ResultOperand::K_Record: {
863          const Record *Rec = RO.getRecord();
864          StringRef ROName = RO.getName();
865
866
867          if (Rec->isSubClassOf("RegisterOperand"))
868            Rec = Rec->getValueAsDef("RegClass");
869          if (Rec->isSubClassOf("RegisterClass")) {
870            Cond = std::string("MI->getOperand(")+llvm::utostr(i)+").isReg()";
871            IAP->addCond(Cond);
872
873            if (!IAP->isOpMapped(ROName)) {
874              IAP->addOperand(ROName, i);
875              Cond = std::string("regIsInRegisterClass(RC_") +
876                CGA->ResultOperands[i].getRecord()->getName() +
877                ", MI->getOperand(" + llvm::utostr(i) + ").getReg())";
878              IAP->addCond(Cond);
879            } else {
880              Cond = std::string("MI->getOperand(") +
881                llvm::utostr(i) + ").getReg() == MI->getOperand(" +
882                llvm::utostr(IAP->getOpIndex(ROName)) + ").getReg()";
883              IAP->addCond(Cond);
884            }
885          } else {
886            assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
887            // FIXME: We may need to handle these situations.
888            delete IAP;
889            IAP = 0;
890            CantHandle = true;
891            break;
892          }
893
894          break;
895        }
896        case CodeGenInstAlias::ResultOperand::K_Imm:
897          Cond = std::string("MI->getOperand(") +
898            llvm::utostr(i) + ").getImm() == " +
899            llvm::utostr(CGA->ResultOperands[i].getImm());
900          IAP->addCond(Cond);
901          break;
902        case CodeGenInstAlias::ResultOperand::K_Reg:
903          Cond = std::string("MI->getOperand(") +
904            llvm::utostr(i) + ").getReg() == " + Target.getName() +
905            "::" + CGA->ResultOperands[i].getRegister()->getName();
906          IAP->addCond(Cond);
907          break;
908        }
909
910        if (!IAP) break;
911      }
912
913      if (CantHandle) continue;
914      IAPrinterMap[I->first].push_back(IAP);
915    }
916  }
917
918  std::string Header;
919  raw_string_ostream HeaderO(Header);
920
921  HeaderO << "bool " << Target.getName() << ClassName
922          << "::printAliasInstr(const MCInst"
923          << " *MI, raw_ostream &OS) {\n";
924
925  std::string Cases;
926  raw_string_ostream CasesO(Cases);
927
928  for (std::map<std::string, std::vector<IAPrinter*> >::iterator
929         I = IAPrinterMap.begin(), E = IAPrinterMap.end(); I != E; ++I) {
930    std::vector<IAPrinter*> &IAPs = I->second;
931    std::vector<IAPrinter*> UniqueIAPs;
932
933    for (std::vector<IAPrinter*>::iterator
934           II = IAPs.begin(), IE = IAPs.end(); II != IE; ++II) {
935      IAPrinter *LHS = *II;
936      bool IsDup = false;
937      for (std::vector<IAPrinter*>::iterator
938             III = IAPs.begin(), IIE = IAPs.end(); III != IIE; ++III) {
939        IAPrinter *RHS = *III;
940        if (LHS != RHS && *LHS == *RHS) {
941          IsDup = true;
942          break;
943        }
944      }
945
946      if (!IsDup) UniqueIAPs.push_back(LHS);
947    }
948
949    if (UniqueIAPs.empty()) continue;
950
951    CasesO.indent(2) << "case " << I->first << ":\n";
952
953    for (std::vector<IAPrinter*>::iterator
954           II = UniqueIAPs.begin(), IE = UniqueIAPs.end(); II != IE; ++II) {
955      IAPrinter *IAP = *II;
956      CasesO.indent(4);
957      IAP->print(CasesO);
958      CasesO << '\n';
959    }
960
961    CasesO.indent(4) << "return false;\n";
962  }
963
964  if (CasesO.str().empty()) {
965    O << HeaderO.str();
966    O << "  return false;\n";
967    O << "}\n\n";
968    O << "#endif // PRINT_ALIAS_INSTR\n";
969    return;
970  }
971
972  EmitGetMapOperandNumber(O);
973
974  O << HeaderO.str();
975  O.indent(2) << "StringRef AsmString;\n";
976  O.indent(2) << "SmallVector<std::pair<StringRef, unsigned>, 4> OpMap;\n";
977  O.indent(2) << "switch (MI->getOpcode()) {\n";
978  O.indent(2) << "default: return false;\n";
979  O << CasesO.str();
980  O.indent(2) << "}\n\n";
981
982  // Code that prints the alias, replacing the operands with the ones from the
983  // MCInst.
984  O << "  std::pair<StringRef, StringRef> ASM = AsmString.split(' ');\n";
985  O << "  OS << '\\t' << ASM.first;\n";
986
987  O << "  if (!ASM.second.empty()) {\n";
988  O << "    OS << '\\t';\n";
989  O << "    for (StringRef::iterator\n";
990  O << "         I = ASM.second.begin(), E = ASM.second.end(); I != E; ) {\n";
991  O << "      if (*I == '$') {\n";
992  O << "        StringRef::iterator Start = ++I;\n";
993  O << "        while (I != E &&\n";
994  O << "               ((*I >= 'a' && *I <= 'z') ||\n";
995  O << "                (*I >= 'A' && *I <= 'Z') ||\n";
996  O << "                (*I >= '0' && *I <= '9') ||\n";
997  O << "                *I == '_'))\n";
998  O << "          ++I;\n";
999  O << "        StringRef Name(Start, I - Start);\n";
1000  O << "        printOperand(MI, getMapOperandNumber(OpMap, Name), OS);\n";
1001  O << "      } else {\n";
1002  O << "        OS << *I++;\n";
1003  O << "      }\n";
1004  O << "    }\n";
1005  O << "  }\n\n";
1006
1007  O << "  return true;\n";
1008  O << "}\n\n";
1009
1010  O << "#endif // PRINT_ALIAS_INSTR\n";
1011}
1012
1013void AsmWriterEmitter::run(raw_ostream &O) {
1014  EmitSourceFileHeader("Assembly Writer Source Fragment", O);
1015
1016  EmitPrintInstruction(O);
1017  EmitGetRegisterName(O);
1018  EmitGetInstructionName(O);
1019  EmitPrintAliasInstruction(O);
1020}
1021
1022