utils/TableGen/AsmWriterEmitter.cpp

//===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This tablegen backend is emits an assembly printer for the current target.
// Note that this is currently fairly skeletal, but will grow over time.
//
//===----------------------------------------------------------------------===//

#include "AsmWriterEmitter.h"
#include "AsmWriterInst.h"
#include "CodeGenTarget.h"
#include "Record.h"
#include "StringToOffsetTable.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include <algorithm>
using namespace llvm;

static void PrintCases(std::vector<std::pair<std::string,
                       AsmWriterOperand> > &OpsToPrint, raw_ostream &O) {
  O << "    case " << OpsToPrint.back().first << ": ";
  AsmWriterOperand TheOp = OpsToPrint.back().second;
  OpsToPrint.pop_back();

  // Check to see if any other operands are identical in this list, and if so,
  // emit a case label for them.
  for (unsigned i = OpsToPrint.size(); i != 0; --i)
    if (OpsToPrint[i-1].second == TheOp) {
      O << "\n    case " << OpsToPrint[i-1].first << ": ";
      OpsToPrint.erase(OpsToPrint.begin()+i-1);
    }

  // Finally, emit the code.
  O << TheOp.getCode();
  O << "break;\n";
}


/// EmitInstructions - Emit the last instruction in the vector and any other
/// instructions that are suitably similar to it.
static void EmitInstructions(std::vector<AsmWriterInst> &Insts,
                             raw_ostream &O) {
  AsmWriterInst FirstInst = Insts.back();
  Insts.pop_back();

  std::vector<AsmWriterInst> SimilarInsts;
  unsigned DifferingOperand = ~0;
  for (unsigned i = Insts.size(); i != 0; --i) {
    unsigned DiffOp = Insts[i-1].MatchesAllButOneOp(FirstInst);
    if (DiffOp != ~1U) {
      if (DifferingOperand == ~0U)  // First match!
        DifferingOperand = DiffOp;

      // If this differs in the same operand as the rest of the instructions in
      // this class, move it to the SimilarInsts list.
      if (DifferingOperand == DiffOp || DiffOp == ~0U) {
        SimilarInsts.push_back(Insts[i-1]);
        Insts.erase(Insts.begin()+i-1);
      }
    }
  }

  O << "  case " << FirstInst.CGI->Namespace << "::"
    << FirstInst.CGI->TheDef->getName() << ":\n";
  for (unsigned i = 0, e = SimilarInsts.size(); i != e; ++i)
    O << "  case " << SimilarInsts[i].CGI->Namespace << "::"
      << SimilarInsts[i].CGI->TheDef->getName() << ":\n";
  for (unsigned i = 0, e = FirstInst.Operands.size(); i != e; ++i) {
    if (i != DifferingOperand) {
      // If the operand is the same for all instructions, just print it.
      O << "    " << FirstInst.Operands[i].getCode();
    } else {
      // If this is the operand that varies between all of the instructions,
      // emit a switch for just this operand now.
      O << "    switch (MI->getOpcode()) {\n";
      std::vector<std::pair<std::string, AsmWriterOperand> > OpsToPrint;
      OpsToPrint.push_back(std::make_pair(FirstInst.CGI->Namespace + "::" +
                                          FirstInst.CGI->TheDef->getName(),
                                          FirstInst.Operands[i]));

      for (unsigned si = 0, e = SimilarInsts.size(); si != e; ++si) {
        AsmWriterInst &AWI = SimilarInsts[si];
        OpsToPrint.push_back(std::make_pair(AWI.CGI->Namespace+"::"+
                                            AWI.CGI->TheDef->getName(),
                                            AWI.Operands[i]));
      }
      std::reverse(OpsToPrint.begin(), OpsToPrint.end());
      while (!OpsToPrint.empty())
        PrintCases(OpsToPrint, O);
      O << "    }";
    }
    O << "\n";
  }
  O << "    break;\n";
}

void AsmWriterEmitter::
FindUniqueOperandCommands(std::vector<std::string> &UniqueOperandCommands, 
                          std::vector<unsigned> &InstIdxs,
                          std::vector<unsigned> &InstOpsUsed) const {
  InstIdxs.assign(NumberedInstructions.size(), ~0U);
  
  // This vector parallels UniqueOperandCommands, keeping track of which
  // instructions each case are used for.  It is a comma separated string of
  // enums.
  std::vector<std::string> InstrsForCase;
  InstrsForCase.resize(UniqueOperandCommands.size());
  InstOpsUsed.assign(UniqueOperandCommands.size(), 0);
  
  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
    const AsmWriterInst *Inst = getAsmWriterInstByID(i);
    if (Inst == 0) continue;  // PHI, INLINEASM, DBG_LABEL, etc.
    
    std::string Command;
    if (Inst->Operands.empty())
      continue;   // Instruction already done.

    Command = "    " + Inst->Operands[0].getCode() + "\n";

    // Check to see if we already have 'Command' in UniqueOperandCommands.
    // If not, add it.
    bool FoundIt = false;
    for (unsigned idx = 0, e = UniqueOperandCommands.size(); idx != e; ++idx)
      if (UniqueOperandCommands[idx] == Command) {
        InstIdxs[i] = idx;
        InstrsForCase[idx] += ", ";
        InstrsForCase[idx] += Inst->CGI->TheDef->getName();
        FoundIt = true;
        break;
      }
    if (!FoundIt) {
      InstIdxs[i] = UniqueOperandCommands.size();
      UniqueOperandCommands.push_back(Command);
      InstrsForCase.push_back(Inst->CGI->TheDef->getName());

      // This command matches one operand so far.
      InstOpsUsed.push_back(1);
    }
  }
  
  // For each entry of UniqueOperandCommands, there is a set of instructions
  // that uses it.  If the next command of all instructions in the set are
  // identical, fold it into the command.
  for (unsigned CommandIdx = 0, e = UniqueOperandCommands.size();
       CommandIdx != e; ++CommandIdx) {
    
    for (unsigned Op = 1; ; ++Op) {
      // Scan for the first instruction in the set.
      std::vector<unsigned>::iterator NIT =
        std::find(InstIdxs.begin(), InstIdxs.end(), CommandIdx);
      if (NIT == InstIdxs.end()) break;  // No commonality.

      // If this instruction has no more operands, we isn't anything to merge
      // into this command.
      const AsmWriterInst *FirstInst = 
        getAsmWriterInstByID(NIT-InstIdxs.begin());
      if (!FirstInst || FirstInst->Operands.size() == Op)
        break;

      // Otherwise, scan to see if all of the other instructions in this command
      // set share the operand.
      bool AllSame = true;
      // Keep track of the maximum, number of operands or any
      // instruction we see in the group.
      size_t MaxSize = FirstInst->Operands.size();

      for (NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx);
           NIT != InstIdxs.end();
           NIT = std::find(NIT+1, InstIdxs.end(), CommandIdx)) {
        // Okay, found another instruction in this command set.  If the operand
        // matches, we're ok, otherwise bail out.
        const AsmWriterInst *OtherInst = 
          getAsmWriterInstByID(NIT-InstIdxs.begin());

        if (OtherInst &&
            OtherInst->Operands.size() > FirstInst->Operands.size())
          MaxSize = std::max(MaxSize, OtherInst->Operands.size());

        if (!OtherInst || OtherInst->Operands.size() == Op ||
            OtherInst->Operands[Op] != FirstInst->Operands[Op]) {
          AllSame = false;
          break;
        }
      }
      if (!AllSame) break;
      
      // Okay, everything in this command set has the same next operand.  Add it
      // to UniqueOperandCommands and remember that it was consumed.
      std::string Command = "    " + FirstInst->Operands[Op].getCode() + "\n";
      
      UniqueOperandCommands[CommandIdx] += Command;
      InstOpsUsed[CommandIdx]++;
    }
  }
  
  // Prepend some of the instructions each case is used for onto the case val.
  for (unsigned i = 0, e = InstrsForCase.size(); i != e; ++i) {
    std::string Instrs = InstrsForCase[i];
    if (Instrs.size() > 70) {
      Instrs.erase(Instrs.begin()+70, Instrs.end());
      Instrs += "...";
    }
    
    if (!Instrs.empty())
      UniqueOperandCommands[i] = "    // " + Instrs + "\n" + 
        UniqueOperandCommands[i];
  }
}


static void UnescapeString(std::string &Str) {
  for (unsigned i = 0; i != Str.size(); ++i) {
    if (Str[i] == '\\' && i != Str.size()-1) {
      switch (Str[i+1]) {
      default: continue;  // Don't execute the code after the switch.
      case 'a': Str[i] = '\a'; break;
      case 'b': Str[i] = '\b'; break;
      case 'e': Str[i] = 27; break;
      case 'f': Str[i] = '\f'; break;
      case 'n': Str[i] = '\n'; break;
      case 'r': Str[i] = '\r'; break;
      case 't': Str[i] = '\t'; break;
      case 'v': Str[i] = '\v'; break;
      case '"': Str[i] = '\"'; break;
      case '\'': Str[i] = '\''; break;
      case '\\': Str[i] = '\\'; break;
      }
      // Nuke the second character.
      Str.erase(Str.begin()+i+1);
    }
  }
}

/// EmitPrintInstruction - Generate the code for the "printInstruction" method
/// implementation.
void AsmWriterEmitter::EmitPrintInstruction(raw_ostream &O) {
  CodeGenTarget Target;
  Record *AsmWriter = Target.getAsmWriter();
  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
  
  O <<
  "/// printInstruction - This method is automatically generated by tablegen\n"
  "/// from the instruction set description.\n"
    "void " << Target.getName() << ClassName
            << "::printInstruction(const MachineInstr *MI) {\n";

  std::vector<AsmWriterInst> Instructions;

  for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
         E = Target.inst_end(); I != E; ++I)
    if (!I->second.AsmString.empty() &&
        I->second.TheDef->getName() != "PHI")
      Instructions.push_back(AsmWriterInst(I->second, AsmWriter));

  // Get the instruction numbering.
  Target.getInstructionsByEnumValue(NumberedInstructions);
  
  // Compute the CodeGenInstruction -> AsmWriterInst mapping.  Note that not
  // all machine instructions are necessarily being printed, so there may be
  // target instructions not in this map.
  for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
    CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i]));

  // Build an aggregate string, and build a table of offsets into it.
  StringToOffsetTable StringTable;
  
  /// OpcodeInfo - This encodes the index of the string to use for the first
  /// chunk of the output as well as indices used for operand printing.
  std::vector<unsigned> OpcodeInfo;
  
  unsigned MaxStringIdx = 0;
  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
    AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
    unsigned Idx;
    if (AWI == 0) {
      // Something not handled by the asmwriter printer.
      Idx = ~0U;
    } else if (AWI->Operands[0].OperandType != 
                        AsmWriterOperand::isLiteralTextOperand ||
               AWI->Operands[0].Str.empty()) {
      // Something handled by the asmwriter printer, but with no leading string.
      Idx = StringTable.GetOrAddStringOffset("");
    } else {
      std::string Str = AWI->Operands[0].Str;
      UnescapeString(Str);
      Idx = StringTable.GetOrAddStringOffset(Str);
      MaxStringIdx = std::max(MaxStringIdx, Idx);
      
      // Nuke the string from the operand list.  It is now handled!
      AWI->Operands.erase(AWI->Operands.begin());
    }
    
    // Bias offset by one since we want 0 as a sentinel.
    OpcodeInfo.push_back(Idx+1);
  }
  
  // Figure out how many bits we used for the string index.
  unsigned AsmStrBits = Log2_32_Ceil(MaxStringIdx+2);
  
  // To reduce code size, we compactify common instructions into a few bits
  // in the opcode-indexed table.
  unsigned BitsLeft = 32-AsmStrBits;

  std::vector<std::vector<std::string> > TableDrivenOperandPrinters;
  
  while (1) {
    std::vector<std::string> UniqueOperandCommands;
    std::vector<unsigned> InstIdxs;
    std::vector<unsigned> NumInstOpsHandled;
    FindUniqueOperandCommands(UniqueOperandCommands, InstIdxs,
                              NumInstOpsHandled);
    
    // If we ran out of operands to print, we're done.
    if (UniqueOperandCommands.empty()) break;
    
    // Compute the number of bits we need to represent these cases, this is
    // ceil(log2(numentries)).
    unsigned NumBits = Log2_32_Ceil(UniqueOperandCommands.size());
    
    // If we don't have enough bits for this operand, don't include it.
    if (NumBits > BitsLeft) {
      DEBUG(errs() << "Not enough bits to densely encode " << NumBits
                   << " more bits\n");
      break;
    }
    
    // Otherwise, we can include this in the initial lookup table.  Add it in.
    BitsLeft -= NumBits;
    for (unsigned i = 0, e = InstIdxs.size(); i != e; ++i)
      if (InstIdxs[i] != ~0U)
        OpcodeInfo[i] |= InstIdxs[i] << (BitsLeft+AsmStrBits);
    
    // Remove the info about this operand.
    for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
      if (AsmWriterInst *Inst = getAsmWriterInstByID(i))
        if (!Inst->Operands.empty()) {
          unsigned NumOps = NumInstOpsHandled[InstIdxs[i]];
          assert(NumOps <= Inst->Operands.size() &&
                 "Can't remove this many ops!");
          Inst->Operands.erase(Inst->Operands.begin(),
                               Inst->Operands.begin()+NumOps);
        }
    }
    
    // Remember the handlers for this set of operands.
    TableDrivenOperandPrinters.push_back(UniqueOperandCommands);
  }
  
  
  
  O<<"  static const unsigned OpInfo[] = {\n";
  for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
    O << "    " << OpcodeInfo[i] << "U,\t// "
      << NumberedInstructions[i]->TheDef->getName() << "\n";
  }
  // Add a dummy entry so the array init doesn't end with a comma.
  O << "    0U\n";
  O << "  };\n\n";
  
  // Emit the string itself.
  O << "  const char *AsmStrs = \n";
  StringTable.EmitString(O);
  O << ";\n\n";

  O << "  O << \"\\t\";\n\n";

  O << "  // Emit the opcode for the instruction.\n"
    << "  unsigned Bits = OpInfo[MI->getOpcode()];\n"
    << "  assert(Bits != 0 && \"Cannot print this instruction.\");\n"
    << "  O << AsmStrs+(Bits & " << (1 << AsmStrBits)-1 << ")-1;\n\n";

  // Output the table driven operand information.
  BitsLeft = 32-AsmStrBits;
  for (unsigned i = 0, e = TableDrivenOperandPrinters.size(); i != e; ++i) {
    std::vector<std::string> &Commands = TableDrivenOperandPrinters[i];

    // Compute the number of bits we need to represent these cases, this is
    // ceil(log2(numentries)).
    unsigned NumBits = Log2_32_Ceil(Commands.size());
    assert(NumBits <= BitsLeft && "consistency error");
    
    // Emit code to extract this field from Bits.
    BitsLeft -= NumBits;
    
    O << "\n  // Fragment " << i << " encoded into " << NumBits
      << " bits for " << Commands.size() << " unique commands.\n";
    
    if (Commands.size() == 2) {
      // Emit two possibilitys with if/else.
      O << "  if ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
        << ((1 << NumBits)-1) << ") {\n"
        << Commands[1]
        << "  } else {\n"
        << Commands[0]
        << "  }\n\n";
    } else {
      O << "  switch ((Bits >> " << (BitsLeft+AsmStrBits) << ") & "
        << ((1 << NumBits)-1) << ") {\n"
        << "  default:   // unreachable.\n";
      
      // Print out all the cases.
      for (unsigned i = 0, e = Commands.size(); i != e; ++i) {
        O << "  case " << i << ":\n";
        O << Commands[i];
        O << "    break;\n";
      }
      O << "  }\n\n";
    }
  }
  
  // Okay, delete instructions with no operand info left.
  for (unsigned i = 0, e = Instructions.size(); i != e; ++i) {
    // Entire instruction has been emitted?
    AsmWriterInst &Inst = Instructions[i];
    if (Inst.Operands.empty()) {
      Instructions.erase(Instructions.begin()+i);
      --i; --e;
    }
  }

    
  // Because this is a vector, we want to emit from the end.  Reverse all of the
  // elements in the vector.
  std::reverse(Instructions.begin(), Instructions.end());
  
  
  // Now that we've emitted all of the operand info that fit into 32 bits, emit
  // information for those instructions that are left.  This is a less dense
  // encoding, but we expect the main 32-bit table to handle the majority of
  // instructions.
  if (!Instructions.empty()) {
    // Find the opcode # of inline asm.
    O << "  switch (MI->getOpcode()) {\n";
    while (!Instructions.empty())
      EmitInstructions(Instructions, O);

    O << "  }\n";
    O << "  return;\n";
  }

  O << "}\n";
}


void AsmWriterEmitter::EmitGetRegisterName(raw_ostream &O) {
  CodeGenTarget Target;
  Record *AsmWriter = Target.getAsmWriter();
  std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
  const std::vector<CodeGenRegister> &Registers = Target.getRegisters();
  
  StringToOffsetTable StringTable;
  O <<
  "\n\n/// getRegisterName - This method is automatically generated by tblgen\n"
  "/// from the register set description.  This returns the assembler name\n"
  "/// for the specified register.\n"
  "const char *" << Target.getName() << ClassName
  << "::getRegisterName(unsigned RegNo) {\n"
  << "  assert(RegNo && RegNo < " << (Registers.size()+1)
  << " && \"Invalid register number!\");\n"
  << "\n"
  << "  static const unsigned RegAsmOffset[] = {";
  for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
    const CodeGenRegister &Reg = Registers[i];

    std::string AsmName = Reg.TheDef->getValueAsString("AsmName");
    if (AsmName.empty())
      AsmName = Reg.getName();
    
    
    if ((i % 14) == 0)
      O << "\n    ";
    
    O << StringTable.GetOrAddStringOffset(AsmName) << ", ";
  }
  O << "0\n"
    << "  };\n"
    << "\n";
  
  O << "  const char *AsmStrs =\n";
  StringTable.EmitString(O);
  O << ";\n";
  
  O << "  return AsmStrs+RegAsmOffset[RegNo-1];\n"
    << "}\n";
}


void AsmWriterEmitter::run(raw_ostream &O) {
  EmitSourceFileHeader("Assembly Writer Source Fragment", O);
  
  EmitPrintInstruction(O);
  EmitGetRegisterName(O);
}