AsmWriterEmitter.cpp revision 234982
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 "SequenceToOffsetTable.h"
20#include "llvm/ADT/Twine.h"
21#include "llvm/Support/Debug.h"
22#include "llvm/Support/MathExtras.h"
23#include "llvm/TableGen/Error.h"
24#include "llvm/TableGen/Record.h"
25#include <algorithm>
26using namespace llvm;
27
28static void PrintCases(std::vector<std::pair<std::string,
29                       AsmWriterOperand> > &OpsToPrint, raw_ostream &O) {
30  O << "    case " << OpsToPrint.back().first << ": ";
31  AsmWriterOperand TheOp = OpsToPrint.back().second;
32  OpsToPrint.pop_back();
33
34  // Check to see if any other operands are identical in this list, and if so,
35  // emit a case label for them.
36  for (unsigned i = OpsToPrint.size(); i != 0; --i)
37    if (OpsToPrint[i-1].second == TheOp) {
38      O << "\n    case " << OpsToPrint[i-1].first << ": ";
39      OpsToPrint.erase(OpsToPrint.begin()+i-1);
40    }
41
42  // Finally, emit the code.
43  O << TheOp.getCode();
44  O << "break;\n";
45}
46
47
48/// EmitInstructions - Emit the last instruction in the vector and any other
49/// instructions that are suitably similar to it.
50static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
51                             raw_ostream &O) {
52  AsmWriterInst FirstInst = Insts.back();
53  Insts.pop_back();
54
55  std::vector<AsmWriterInst> SimilarInsts;
56  unsigned DifferingOperand = ~0;
57  for (unsigned i = Insts.size(); i != 0; --i) {
58    unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
59    if (DiffOp != ~1U) {
60      if (DifferingOperand == ~0U)  // First match!
61        DifferingOperand = DiffOp;
62
63      // If this differs in the same operand as the rest of the instructions in
64      // this class, move it to the SimilarInsts list.
65      if (DifferingOperand == DiffOp || DiffOp == ~0U) {
66        SimilarInsts.push_back(Insts[i-1]);
67        Insts.erase(Insts.begin()+i-1);
68      }
69    }
70  }
71
72  O << "  case " << FirstInst.CGI->Namespace << "::"
73    << FirstInst.CGI->TheDef->getName() << ":\n";
74  for (unsigned i = 0, e = SimilarInsts.size(); i != e; ++i)
75    O << "  case " << SimilarInsts[i].CGI->Namespace << "::"
76      << SimilarInsts[i].CGI->TheDef->getName() << ":\n";
77  for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
78    if (i != DifferingOperand) {
79      // If the operand is the same for all instructions, just print it.
80      O << "    " << FirstInst.Operands[i].getCode();
81    } else {
82      // If this is the operand that varies between all of the instructions,
83      // emit a switch for just this operand now.
84      O << "    switch (MI->getOpcode()) {\n";
85      std::vector<std::pair<std::string, AsmWriterOperand> > OpsToPrint;
86      OpsToPrint.push_back(std::make_pair(FirstInst.CGI->Namespace + "::" +
87                                          FirstInst.CGI->TheDef->getName(),
88                                          FirstInst.Operands[i]));
89
90      for (unsigned si = 0, e = SimilarInsts.size(); si != e; ++si) {
91        AsmWriterInst &AWI = SimilarInsts[si];
92        OpsToPrint.push_back(std::make_pair(AWI.CGI->Namespace+"::"+
93                                            AWI.CGI->TheDef->getName(),
94                                            AWI.Operands[i]));
95      }
96      std::reverse(OpsToPrint.begin(), OpsToPrint.end());
97      while (!OpsToPrint.empty())
98        PrintCases(OpsToPrint, O);
99      O << "    }";
100    }
101    O << "\n";
102  }
103  O << "    break;\n";
104}
105
106void AsmWriterEmitter::
107FindUniqueOperandCommands(std::vector<std::string> &UniqueOperandCommands,
108                          std::vector<unsigned> &InstIdxs,
109                          std::vector<unsigned> &InstOpsUsed) const {
110  InstIdxs.assign(NumberedInstructions.size(), ~0U);
111
112  // This vector parallels UniqueOperandCommands, keeping track of which
113  // instructions each case are used for.  It is a comma separated string of
114  // enums.
115  std::vector<std::string> InstrsForCase;
116  InstrsForCase.resize(UniqueOperandCommands.size());
117  InstOpsUsed.assign(UniqueOperandCommands.size(), 0);
118
119  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
120    const AsmWriterInst *Inst = getAsmWriterInstByID(i);
121    if (Inst == 0) continue;  // PHI, INLINEASM, PROLOG_LABEL, etc.
122
123    std::string Command;
124    if (Inst->Operands.empty())
125      continue;   // Instruction already done.
126
127    Command = "    " + Inst->Operands[0].getCode() + "\n";
128
129    // Check to see if we already have 'Command' in UniqueOperandCommands.
130    // If not, add it.
131    bool FoundIt = false;
132    for (unsigned idx = 0, e = UniqueOperandCommands.size(); idx != e; ++idx)
133      if (UniqueOperandCommands[idx] == Command) {
134        InstIdxs[i] = idx;
135        InstrsForCase[idx] += ", ";
136        InstrsForCase[idx] += Inst->CGI->TheDef->getName();
137        FoundIt = true;
138        break;
139      }
140    if (!FoundIt) {
141      InstIdxs[i] = UniqueOperandCommands.size();
142      UniqueOperandCommands.push_back(Command);
143      InstrsForCase.push_back(Inst->CGI->TheDef->getName());
144
145      // This command matches one operand so far.
146      InstOpsUsed.push_back(1);
147    }
148  }
149
150  // For each entry of UniqueOperandCommands, there is a set of instructions
151  // that uses it.  If the next command of all instructions in the set are
152  // identical, fold it into the command.
153  for (unsigned CommandIdx = 0, e = UniqueOperandCommands.size();
154       CommandIdx != e; ++CommandIdx) {
155
156    for (unsigned Op = 1; ; ++Op) {
157      // Scan for the first instruction in the set.
158      std::vector<unsigned>::iterator NIT =
159        std::find(InstIdxs.begin(), InstIdxs.end(), CommandIdx);
160      if (NIT == InstIdxs.end()) break;  // No commonality.
161
162      // If this instruction has no more operands, we isn't anything to merge
163      // into this command.
164      const AsmWriterInst *FirstInst =
165        getAsmWriterInstByID(NIT-InstIdxs.begin());
166      if (!FirstInst || FirstInst->Operands.size() == Op)
167        break;
168
169      // Otherwise, scan to see if all of the other instructions in this command
170      // set share the operand.
171      bool AllSame = true;
172      // Keep track of the maximum, number of operands or any
173      // instruction we see in the group.
174      size_t MaxSize = FirstInst->Operands.size();
175
176      for (NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx);
177           NIT != InstIdxs.end();
178           NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx)) {
179        // Okay, found another instruction in this command set.  If the operand
180        // matches, we're ok, otherwise bail out.
181        const AsmWriterInst *OtherInst =
182          getAsmWriterInstByID(NIT-InstIdxs.begin());
183
184        if (OtherInst &&
185            OtherInst->Operands.size() > FirstInst->Operands.size())
186          MaxSize = std::max(MaxSize, OtherInst->Operands.size());
187
188        if (!OtherInst || OtherInst->Operands.size() == Op ||
189            OtherInst->Operands[Op] != FirstInst->Operands[Op]) {
190          AllSame = false;
191          break;
192        }
193      }
194      if (!AllSame) break;
195
196      // Okay, everything in this command set has the same next operand.  Add it
197      // to UniqueOperandCommands and remember that it was consumed.
198      std::string Command = "    " + FirstInst->Operands[Op].getCode() + "\n";
199
200      UniqueOperandCommands[CommandIdx] += Command;
201      InstOpsUsed[CommandIdx]++;
202    }
203  }
204
205  // Prepend some of the instructions each case is used for onto the case val.
206  for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
207    std::string Instrs = InstrsForCase[i];
208    if (Instrs.size() > 70) {
209      Instrs.erase(Instrs.begin()+70, Instrs.end());
210      Instrs += "...";
211    }
212
213    if (!Instrs.empty())
214      UniqueOperandCommands[i] = "    // " + Instrs + "\n" +
215        UniqueOperandCommands[i];
216  }
217}
218
219
220static void UnescapeString(std::string &Str) {
221  for (unsigned i = 0; i != Str.size(); ++i) {
222    if (Str[i] == '\\' && i != Str.size()-1) {
223      switch (Str[i+1]) {
224      default: continue;  // Don't execute the code after the switch.
225      case 'a': Str[i] = '\a'; break;
226      case 'b': Str[i] = '\b'; break;
227      case 'e': Str[i] = 27; break;
228      case 'f': Str[i] = '\f'; break;
229      case 'n': Str[i] = '\n'; break;
230      case 'r': Str[i] = '\r'; break;
231      case 't': Str[i] = '\t'; break;
232      case 'v': Str[i] = '\v'; break;
233      case '"': Str[i] = '\"'; break;
234      case '\'': Str[i] = '\''; break;
235      case '\\': Str[i] = '\\'; break;
236      }
237      // Nuke the second character.
238      Str.erase(Str.begin()+i+1);
239    }
240  }
241}
242
243/// EmitPrintInstruction - Generate the code for the "printInstruction" method
244/// implementation.
245void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) {
246  CodeGenTarget Target(Records);
247  Record *AsmWriter = Target.getAsmWriter();
248  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
249  bool isMC = AsmWriter->getValueAsBit("isMCAsmWriter");
250  const char *MachineInstrClassName = isMC ? "MCInst" : "MachineInstr";
251
252  O <<
253  "/// printInstruction - This method is automatically generated by tablegen\n"
254  "/// from the instruction set description.\n"
255    "void " << Target.getName() << ClassName
256            << "::printInstruction(const " << MachineInstrClassName
257            << " *MI, raw_ostream &O) {\n";
258
259  std::vector<AsmWriterInst> Instructions;
260
261  for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
262         E = Target.inst_end(); I != E; ++I)
263    if (!(*I)->AsmString.empty() &&
264        (*I)->TheDef->getName() != "PHI")
265      Instructions.push_back(
266        AsmWriterInst(**I,
267                      AsmWriter->getValueAsInt("Variant"),
268                      AsmWriter->getValueAsInt("FirstOperandColumn"),
269                      AsmWriter->getValueAsInt("OperandSpacing")));
270
271  // Get the instruction numbering.
272  NumberedInstructions = Target.getInstructionsByEnumValue();
273
274  // Compute the CodeGenInstruction -> AsmWriterInst mapping.  Note that not
275  // all machine instructions are necessarily being printed, so there may be
276  // target instructions not in this map.
277  for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
278    CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i]));
279
280  // Build an aggregate string, and build a table of offsets into it.
281  SequenceToOffsetTable<std::string> StringTable;
282
283  /// OpcodeInfo - This encodes the index of the string to use for the first
284  /// chunk of the output as well as indices used for operand printing.
285  std::vector<unsigned> OpcodeInfo;
286
287  // Add all strings to the string table upfront so it can generate an optimized
288  // representation.
289  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
290    AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
291    if (AWI != 0 &&
292        AWI->Operands[0].OperandType ==
293                 AsmWriterOperand::isLiteralTextOperand &&
294        !AWI->Operands[0].Str.empty()) {
295      std::string Str = AWI->Operands[0].Str;
296      UnescapeString(Str);
297      StringTable.add(Str);
298    }
299  }
300
301  StringTable.layout();
302
303  unsigned MaxStringIdx = 0;
304  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
305    AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
306    unsigned Idx;
307    if (AWI == 0) {
308      // Something not handled by the asmwriter printer.
309      Idx = ~0U;
310    } else if (AWI->Operands[0].OperandType !=
311                        AsmWriterOperand::isLiteralTextOperand ||
312               AWI->Operands[0].Str.empty()) {
313      // Something handled by the asmwriter printer, but with no leading string.
314      Idx = StringTable.get("");
315    } else {
316      std::string Str = AWI->Operands[0].Str;
317      UnescapeString(Str);
318      Idx = StringTable.get(Str);
319      MaxStringIdx = std::max(MaxStringIdx, Idx);
320
321      // Nuke the string from the operand list.  It is now handled!
322      AWI->Operands.erase(AWI->Operands.begin());
323    }
324
325    // Bias offset by one since we want 0 as a sentinel.
326    OpcodeInfo.push_back(Idx+1);
327  }
328
329  // Figure out how many bits we used for the string index.
330  unsigned AsmStrBits = Log2_32_Ceil(MaxStringIdx+2);
331
332  // To reduce code size, we compactify common instructions into a few bits
333  // in the opcode-indexed table.
334  unsigned BitsLeft = 32-AsmStrBits;
335
336  std::vector<std::vector<std::string> > TableDrivenOperandPrinters;
337
338  while (1) {
339    std::vector<std::string> UniqueOperandCommands;
340    std::vector<unsigned> InstIdxs;
341    std::vector<unsigned> NumInstOpsHandled;
342    FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
343                              NumInstOpsHandled);
344
345    // If we ran out of operands to print, we're done.
346    if (UniqueOperandCommands.empty()) break;
347
348    // Compute the number of bits we need to represent these cases, this is
349    // ceil(log2(numentries)).
350    unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
351
352    // If we don't have enough bits for this operand, don't include it.
353    if (NumBits > BitsLeft) {
354      DEBUG(errs() << "Not enough bits to densely encode " << NumBits
355                   << " more bits\n");
356      break;
357    }
358
359    // Otherwise, we can include this in the initial lookup table.  Add it in.
360    BitsLeft -= NumBits;
361    for (unsigned i = 0, e = InstIdxs.size(); i != e; ++i)
362      if (InstIdxs[i] != ~0U)
363        OpcodeInfo[i] |= InstIdxs[i] << (BitsLeft+AsmStrBits);
364
365    // Remove the info about this operand.
366    for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
367      if (AsmWriterInst *Inst = getAsmWriterInstByID(i))
368        if (!Inst->Operands.empty()) {
369          unsigned NumOps = NumInstOpsHandled[InstIdxs[i]];
370          assert(NumOps <= Inst->Operands.size() &&
371                 "Can't remove this many ops!");
372          Inst->Operands.erase(Inst->Operands.begin(),
373                               Inst->Operands.begin()+NumOps);
374        }
375    }
376
377    // Remember the handlers for this set of operands.
378    TableDrivenOperandPrinters.push_back(UniqueOperandCommands);
379  }
380
381
382
383  O<<"  static const unsigned OpInfo[] = {\n";
384  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
385    O << "    " << OpcodeInfo[i] << "U,\t// "
386      << NumberedInstructions[i]->TheDef->getName() << "\n";
387  }
388  // Add a dummy entry so the array init doesn't end with a comma.
389  O << "    0U\n";
390  O << "  };\n\n";
391
392  // Emit the string itself.
393  O << "  const char AsmStrs[] = {\n";
394  StringTable.emit(O, printChar);
395  O << "  };\n\n";
396
397  O << "  O << \"\\t\";\n\n";
398
399  O << "  // Emit the opcode for the instruction.\n"
400    << "  unsigned Bits = OpInfo[MI->getOpcode()];\n"
401    << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n"
402    << "  O << AsmStrs+(Bits & " << (1 << AsmStrBits)-1 << ")-1;\n\n";
403
404  // Output the table driven operand information.
405  BitsLeft = 32-AsmStrBits;
406  for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
407    std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];
408
409    // Compute the number of bits we need to represent these cases, this is
410    // ceil(log2(numentries)).
411    unsigned NumBits = Log2_32_Ceil(Commands.size());
412    assert(NumBits <= BitsLeft && "consistency error");
413
414    // Emit code to extract this field from Bits.
415    BitsLeft -= NumBits;
416
417    O << "\n  // Fragment " << i << " encoded into " << NumBits
418      << " bits for " << Commands.size() << " unique commands.\n";
419
420    if (Commands.size() == 2) {
421      // Emit two possibilitys with if/else.
422      O << "  if ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
423        << ((1 << NumBits)-1) << ") {\n"
424        << Commands[1]
425        << "  } else {\n"
426        << Commands[0]
427        << "  }\n\n";
428    } else if (Commands.size() == 1) {
429      // Emit a single possibility.
430      O << Commands[0] << "\n\n";
431    } else {
432      O << "  switch ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
433        << ((1 << NumBits)-1) << ") {\n"
434        << "  default:   // unreachable.\n";
435
436      // Print out all the cases.
437      for (unsigned i = 0, e = Commands.size(); i != e; ++i) {
438        O << "  case " << i << ":\n";
439        O << Commands[i];
440        O << "    break;\n";
441      }
442      O << "  }\n\n";
443    }
444  }
445
446  // Okay, delete instructions with no operand info left.
447  for (unsigned i = 0, e = Instructions.size(); i != e; ++i) {
448    // Entire instruction has been emitted?
449    AsmWriterInst &Inst = Instructions[i];
450    if (Inst.Operands.empty()) {
451      Instructions.erase(Instructions.begin()+i);
452      --i; --e;
453    }
454  }
455
456
457  // Because this is a vector, we want to emit from the end.  Reverse all of the
458  // elements in the vector.
459  std::reverse(Instructions.begin(), Instructions.end());
460
461
462  // Now that we've emitted all of the operand info that fit into 32 bits, emit
463  // information for those instructions that are left.  This is a less dense
464  // encoding, but we expect the main 32-bit table to handle the majority of
465  // instructions.
466  if (!Instructions.empty()) {
467    // Find the opcode # of inline asm.
468    O << "  switch (MI->getOpcode()) {\n";
469    while (!Instructions.empty())
470      EmitInstructions(Instructions, O);
471
472    O << "  }\n";
473    O << "  return;\n";
474  }
475
476  O << "}\n";
477}
478
479static void
480emitRegisterNameString(raw_ostream &O, StringRef AltName,
481                       const std::vector<CodeGenRegister*> &Registers) {
482  SequenceToOffsetTable<std::string> StringTable;
483  SmallVector<std::string, 4> AsmNames(Registers.size());
484  for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
485    const CodeGenRegister &Reg = *Registers[i];
486    std::string &AsmName = AsmNames[i];
487
488    // "NoRegAltName" is special. We don't need to do a lookup for that,
489    // as it's just a reference to the default register name.
490    if (AltName == "" || AltName == "NoRegAltName") {
491      AsmName = Reg.TheDef->getValueAsString("AsmName");
492      if (AsmName.empty())
493        AsmName = Reg.getName();
494    } else {
495      // Make sure the register has an alternate name for this index.
496      std::vector<Record*> AltNameList =
497        Reg.TheDef->getValueAsListOfDefs("RegAltNameIndices");
498      unsigned Idx = 0, e;
499      for (e = AltNameList.size();
500           Idx < e && (AltNameList[Idx]->getName() != AltName);
501           ++Idx)
502        ;
503      // If the register has an alternate name for this index, use it.
504      // Otherwise, leave it empty as an error flag.
505      if (Idx < e) {
506        std::vector<std::string> AltNames =
507          Reg.TheDef->getValueAsListOfStrings("AltNames");
508        if (AltNames.size() <= Idx)
509          throw TGError(Reg.TheDef->getLoc(),
510                        (Twine("Register definition missing alt name for '") +
511                        AltName + "'.").str());
512        AsmName = AltNames[Idx];
513      }
514    }
515    StringTable.add(AsmName);
516  }
517
518  StringTable.layout();
519  O << "  static const char AsmStrs" << AltName << "[] = {\n";
520  StringTable.emit(O, printChar);
521  O << "  };\n\n";
522
523  O << "  static const unsigned RegAsmOffset" << AltName << "[] = {";
524  for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
525    if ((i % 14) == 0)
526      O << "\n    ";
527    O << StringTable.get(AsmNames[i]) << ", ";
528  }
529  O << "\n  };\n"
530    << "\n";
531}
532
533void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
534  CodeGenTarget Target(Records);
535  Record *AsmWriter = Target.getAsmWriter();
536  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
537  const std::vector<CodeGenRegister*> &Registers =
538    Target.getRegBank().getRegisters();
539  std::vector<Record*> AltNameIndices = Target.getRegAltNameIndices();
540  bool hasAltNames = AltNameIndices.size() > 1;
541
542  O <<
543  "\n\n/// getRegisterName - This method is automatically generated by tblgen\n"
544  "/// from the register set description.  This returns the assembler name\n"
545  "/// for the specified register.\n"
546  "const char *" << Target.getName() << ClassName << "::";
547  if (hasAltNames)
548    O << "\ngetRegisterName(unsigned RegNo, unsigned AltIdx) {\n";
549  else
550    O << "getRegisterName(unsigned RegNo) {\n";
551  O << "  assert(RegNo && RegNo < " << (Registers.size()+1)
552    << " && \"Invalid register number!\");\n"
553    << "\n";
554
555  if (hasAltNames) {
556    for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i)
557      emitRegisterNameString(O, AltNameIndices[i]->getName(), Registers);
558  } else
559    emitRegisterNameString(O, "", Registers);
560
561  if (hasAltNames) {
562    O << "  const unsigned *RegAsmOffset;\n"
563      << "  const char *AsmStrs;\n"
564      << "  switch(AltIdx) {\n"
565      << "  default: llvm_unreachable(\"Invalid register alt name index!\");\n";
566    for (unsigned i = 0, e = AltNameIndices.size(); i < e; ++i) {
567      StringRef Namespace = AltNameIndices[1]->getValueAsString("Namespace");
568      StringRef AltName(AltNameIndices[i]->getName());
569      O << "  case " << Namespace << "::" << AltName
570        << ":\n"
571        << "    AsmStrs = AsmStrs" << AltName  << ";\n"
572        << "    RegAsmOffset = RegAsmOffset" << AltName << ";\n"
573        << "    break;\n";
574    }
575    O << "}\n";
576  }
577
578  O << "  assert (*(AsmStrs+RegAsmOffset[RegNo-1]) &&\n"
579    << "          \"Invalid alt name index for register!\");\n"
580    << "  return AsmStrs+RegAsmOffset[RegNo-1];\n"
581    << "}\n";
582}
583
584namespace {
585// IAPrinter - Holds information about an InstAlias. Two InstAliases match if
586// they both have the same conditionals. In which case, we cannot print out the
587// alias for that pattern.
588class IAPrinter {
589  std::vector<std::string> Conds;
590  std::map<StringRef, unsigned> OpMap;
591  std::string Result;
592  std::string AsmString;
593  SmallVector<Record*, 4> ReqFeatures;
594public:
595  IAPrinter(std::string R, std::string AS)
596    : Result(R), AsmString(AS) {}
597
598  void addCond(const std::string &C) { Conds.push_back(C); }
599
600  void addOperand(StringRef Op, unsigned Idx) { OpMap[Op] = Idx; }
601  unsigned getOpIndex(StringRef Op) { return OpMap[Op]; }
602  bool isOpMapped(StringRef Op) { return OpMap.find(Op) != OpMap.end(); }
603
604  void print(raw_ostream &O) {
605    if (Conds.empty() && ReqFeatures.empty()) {
606      O.indent(6) << "return true;\n";
607      return;
608    }
609
610    O << "if (";
611
612    for (std::vector<std::string>::iterator
613           I = Conds.begin(), E = Conds.end(); I != E; ++I) {
614      if (I != Conds.begin()) {
615        O << " &&\n";
616        O.indent(8);
617      }
618
619      O << *I;
620    }
621
622    O << ") {\n";
623    O.indent(6) << "// " << Result << "\n";
624    O.indent(6) << "AsmString = \"" << AsmString << "\";\n";
625
626    for (std::map<StringRef, unsigned>::iterator
627           I = OpMap.begin(), E = OpMap.end(); I != E; ++I)
628      O.indent(6) << "OpMap.push_back(std::make_pair(\"" << I->first << "\", "
629                  << I->second << "));\n";
630
631    O.indent(6) << "break;\n";
632    O.indent(4) << '}';
633  }
634
635  bool operator==(const IAPrinter &RHS) {
636    if (Conds.size() != RHS.Conds.size())
637      return false;
638
639    unsigned Idx = 0;
640    for (std::vector<std::string>::iterator
641           I = Conds.begin(), E = Conds.end(); I != E; ++I)
642      if (*I != RHS.Conds[Idx++])
643        return false;
644
645    return true;
646  }
647
648  bool operator()(const IAPrinter &RHS) {
649    if (Conds.size() < RHS.Conds.size())
650      return true;
651
652    unsigned Idx = 0;
653    for (std::vector<std::string>::iterator
654           I = Conds.begin(), E = Conds.end(); I != E; ++I)
655      if (*I != RHS.Conds[Idx++])
656        return *I < RHS.Conds[Idx++];
657
658    return false;
659  }
660};
661
662} // end anonymous namespace
663
664static void EmitGetMapOperandNumber(raw_ostream &O) {
665  O << "static unsigned getMapOperandNumber("
666    << "const SmallVectorImpl<std::pair<StringRef, unsigned> > &OpMap,\n";
667  O << "                                    StringRef Name) {\n";
668  O << "  for (SmallVectorImpl<std::pair<StringRef, unsigned> >::"
669    << "const_iterator\n";
670  O << "         I = OpMap.begin(), E = OpMap.end(); I != E; ++I)\n";
671  O << "    if (I->first == Name)\n";
672  O << "      return I->second;\n";
673  O << "  llvm_unreachable(\"Operand not in map!\");\n";
674  O << "}\n\n";
675}
676
677static unsigned CountNumOperands(StringRef AsmString) {
678  unsigned NumOps = 0;
679  std::pair<StringRef, StringRef> ASM = AsmString.split(' ');
680
681  while (!ASM.second.empty()) {
682    ++NumOps;
683    ASM = ASM.second.split(' ');
684  }
685
686  return NumOps;
687}
688
689static unsigned CountResultNumOperands(StringRef AsmString) {
690  unsigned NumOps = 0;
691  std::pair<StringRef, StringRef> ASM = AsmString.split('\t');
692
693  if (!ASM.second.empty()) {
694    size_t I = ASM.second.find('{');
695    StringRef Str = ASM.second;
696    if (I != StringRef::npos)
697      Str = ASM.second.substr(I, ASM.second.find('|', I));
698
699    ASM = Str.split(' ');
700
701    do {
702      ++NumOps;
703      ASM = ASM.second.split(' ');
704    } while (!ASM.second.empty());
705  }
706
707  return NumOps;
708}
709
710void AsmWriterEmitter::EmitPrintAliasInstruction(raw_ostream &O) {
711  CodeGenTarget Target(Records);
712  Record *AsmWriter = Target.getAsmWriter();
713
714  if (!AsmWriter->getValueAsBit("isMCAsmWriter"))
715    return;
716
717  O << "\n#ifdef PRINT_ALIAS_INSTR\n";
718  O << "#undef PRINT_ALIAS_INSTR\n\n";
719
720  // Emit the method that prints the alias instruction.
721  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
722
723  std::vector<Record*> AllInstAliases =
724    Records.getAllDerivedDefinitions("InstAlias");
725
726  // Create a map from the qualified name to a list of potential matches.
727  std::map<std::string, std::vector<CodeGenInstAlias*> > AliasMap;
728  for (std::vector<Record*>::iterator
729         I = AllInstAliases.begin(), E = AllInstAliases.end(); I != E; ++I) {
730    CodeGenInstAlias *Alias = new CodeGenInstAlias(*I, Target);
731    const Record *R = *I;
732    if (!R->getValueAsBit("EmitAlias"))
733      continue; // We were told not to emit the alias, but to emit the aliasee.
734    const DagInit *DI = R->getValueAsDag("ResultInst");
735    const DefInit *Op = dynamic_cast<const DefInit*>(DI->getOperator());
736    AliasMap[getQualifiedName(Op->getDef())].push_back(Alias);
737  }
738
739  // A map of which conditions need to be met for each instruction operand
740  // before it can be matched to the mnemonic.
741  std::map<std::string, std::vector<IAPrinter*> > IAPrinterMap;
742
743  for (std::map<std::string, std::vector<CodeGenInstAlias*> >::iterator
744         I = AliasMap.begin(), E = AliasMap.end(); I != E; ++I) {
745    std::vector<CodeGenInstAlias*> &Aliases = I->second;
746
747    for (std::vector<CodeGenInstAlias*>::iterator
748           II = Aliases.begin(), IE = Aliases.end(); II != IE; ++II) {
749      const CodeGenInstAlias *CGA = *II;
750      unsigned LastOpNo = CGA->ResultInstOperandIndex.size();
751      unsigned NumResultOps =
752        CountResultNumOperands(CGA->ResultInst->AsmString);
753
754      // Don't emit the alias if it has more operands than what it's aliasing.
755      if (NumResultOps < CountNumOperands(CGA->AsmString))
756        continue;
757
758      IAPrinter *IAP = new IAPrinter(CGA->Result->getAsString(),
759                                     CGA->AsmString);
760
761      std::string Cond;
762      Cond = std::string("MI->getNumOperands() == ") + llvm::utostr(LastOpNo);
763      IAP->addCond(Cond);
764
765      std::map<StringRef, unsigned> OpMap;
766      bool CantHandle = false;
767
768      for (unsigned i = 0, e = LastOpNo; i != e; ++i) {
769        const CodeGenInstAlias::ResultOperand &RO = CGA->ResultOperands[i];
770
771        switch (RO.Kind) {
772        case CodeGenInstAlias::ResultOperand::K_Record: {
773          const Record *Rec = RO.getRecord();
774          StringRef ROName = RO.getName();
775
776
777          if (Rec->isSubClassOf("RegisterOperand"))
778            Rec = Rec->getValueAsDef("RegClass");
779          if (Rec->isSubClassOf("RegisterClass")) {
780            Cond = std::string("MI->getOperand(")+llvm::utostr(i)+").isReg()";
781            IAP->addCond(Cond);
782
783            if (!IAP->isOpMapped(ROName)) {
784              IAP->addOperand(ROName, i);
785              Cond = std::string("MRI.getRegClass(") + Target.getName() + "::" +
786                CGA->ResultOperands[i].getRecord()->getName() + "RegClassID)"
787                ".contains(MI->getOperand(" + llvm::utostr(i) + ").getReg())";
788              IAP->addCond(Cond);
789            } else {
790              Cond = std::string("MI->getOperand(") +
791                llvm::utostr(i) + ").getReg() == MI->getOperand(" +
792                llvm::utostr(IAP->getOpIndex(ROName)) + ").getReg()";
793              IAP->addCond(Cond);
794            }
795          } else {
796            assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
797            // FIXME: We may need to handle these situations.
798            delete IAP;
799            IAP = 0;
800            CantHandle = true;
801            break;
802          }
803
804          break;
805        }
806        case CodeGenInstAlias::ResultOperand::K_Imm:
807          Cond = std::string("MI->getOperand(") +
808            llvm::utostr(i) + ").getImm() == " +
809            llvm::utostr(CGA->ResultOperands[i].getImm());
810          IAP->addCond(Cond);
811          break;
812        case CodeGenInstAlias::ResultOperand::K_Reg:
813          // If this is zero_reg, something's playing tricks we're not
814          // equipped to handle.
815          if (!CGA->ResultOperands[i].getRegister()) {
816            CantHandle = true;
817            break;
818          }
819
820          Cond = std::string("MI->getOperand(") +
821            llvm::utostr(i) + ").getReg() == " + Target.getName() +
822            "::" + CGA->ResultOperands[i].getRegister()->getName();
823          IAP->addCond(Cond);
824          break;
825        }
826
827        if (!IAP) break;
828      }
829
830      if (CantHandle) continue;
831      IAPrinterMap[I->first].push_back(IAP);
832    }
833  }
834
835  std::string Header;
836  raw_string_ostream HeaderO(Header);
837
838  HeaderO << "bool " << Target.getName() << ClassName
839          << "::printAliasInstr(const MCInst"
840          << " *MI, raw_ostream &OS) {\n";
841
842  std::string Cases;
843  raw_string_ostream CasesO(Cases);
844
845  for (std::map<std::string, std::vector<IAPrinter*> >::iterator
846         I = IAPrinterMap.begin(), E = IAPrinterMap.end(); I != E; ++I) {
847    std::vector<IAPrinter*> &IAPs = I->second;
848    std::vector<IAPrinter*> UniqueIAPs;
849
850    for (std::vector<IAPrinter*>::iterator
851           II = IAPs.begin(), IE = IAPs.end(); II != IE; ++II) {
852      IAPrinter *LHS = *II;
853      bool IsDup = false;
854      for (std::vector<IAPrinter*>::iterator
855             III = IAPs.begin(), IIE = IAPs.end(); III != IIE; ++III) {
856        IAPrinter *RHS = *III;
857        if (LHS != RHS && *LHS == *RHS) {
858          IsDup = true;
859          break;
860        }
861      }
862
863      if (!IsDup) UniqueIAPs.push_back(LHS);
864    }
865
866    if (UniqueIAPs.empty()) continue;
867
868    CasesO.indent(2) << "case " << I->first << ":\n";
869
870    for (std::vector<IAPrinter*>::iterator
871           II = UniqueIAPs.begin(), IE = UniqueIAPs.end(); II != IE; ++II) {
872      IAPrinter *IAP = *II;
873      CasesO.indent(4);
874      IAP->print(CasesO);
875      CasesO << '\n';
876    }
877
878    CasesO.indent(4) << "return false;\n";
879  }
880
881  if (CasesO.str().empty()) {
882    O << HeaderO.str();
883    O << "  return false;\n";
884    O << "}\n\n";
885    O << "#endif // PRINT_ALIAS_INSTR\n";
886    return;
887  }
888
889  EmitGetMapOperandNumber(O);
890
891  O << HeaderO.str();
892  O.indent(2) << "StringRef AsmString;\n";
893  O.indent(2) << "SmallVector<std::pair<StringRef, unsigned>, 4> OpMap;\n";
894  O.indent(2) << "switch (MI->getOpcode()) {\n";
895  O.indent(2) << "default: return false;\n";
896  O << CasesO.str();
897  O.indent(2) << "}\n\n";
898
899  // Code that prints the alias, replacing the operands with the ones from the
900  // MCInst.
901  O << "  std::pair<StringRef, StringRef> ASM = AsmString.split(' ');\n";
902  O << "  OS << '\\t' << ASM.first;\n";
903
904  O << "  if (!ASM.second.empty()) {\n";
905  O << "    OS << '\\t';\n";
906  O << "    for (StringRef::iterator\n";
907  O << "         I = ASM.second.begin(), E = ASM.second.end(); I != E; ) {\n";
908  O << "      if (*I == '$') {\n";
909  O << "        StringRef::iterator Start = ++I;\n";
910  O << "        while (I != E &&\n";
911  O << "               ((*I >= 'a' && *I <= 'z') ||\n";
912  O << "                (*I >= 'A' && *I <= 'Z') ||\n";
913  O << "                (*I >= '0' && *I <= '9') ||\n";
914  O << "                *I == '_'))\n";
915  O << "          ++I;\n";
916  O << "        StringRef Name(Start, I - Start);\n";
917  O << "        printOperand(MI, getMapOperandNumber(OpMap, Name), OS);\n";
918  O << "      } else {\n";
919  O << "        OS << *I++;\n";
920  O << "      }\n";
921  O << "    }\n";
922  O << "  }\n\n";
923
924  O << "  return true;\n";
925  O << "}\n\n";
926
927  O << "#endif // PRINT_ALIAS_INSTR\n";
928}
929
930void AsmWriterEmitter::run(raw_ostream &O) {
931  EmitSourceFileHeader("Assembly Writer Source Fragment", O);
932
933  EmitPrintInstruction(O);
934  EmitGetRegisterName(O);
935  EmitPrintAliasInstruction(O);
936}
937
938