//===- AsmWriterEmitter.cpp - Generate an assembly writer -----------------===//
-//
+//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and 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.
#include "AsmWriterEmitter.h"
#include "CodeGenTarget.h"
+#include "Record.h"
+#include <algorithm>
#include <ostream>
using namespace llvm;
C == '_';
}
+namespace {
+ struct AsmWriterOperand {
+ enum { isLiteralTextOperand, isMachineInstrOperand } OperandType;
+
+ /// Str - For isLiteralTextOperand, this IS the literal text. For
+ /// isMachineInstrOperand, this is the PrinterMethodName for the operand.
+ std::string Str;
+
+ /// MiOpNo - For isMachineInstrOperand, this is the operand number of the
+ /// machine instruction.
+ unsigned MIOpNo;
+
+ /// MiModifier - For isMachineInstrOperand, this is the modifier string for
+ /// an operand, specified with syntax like ${opname:modifier}.
+ std::string MiModifier;
+
+ AsmWriterOperand(const std::string &LitStr)
+ : OperandType(isLiteralTextOperand), Str(LitStr) {}
+
+ AsmWriterOperand(const std::string &Printer, unsigned OpNo,
+ const std::string &Modifier)
+ : OperandType(isMachineInstrOperand), Str(Printer), MIOpNo(OpNo),
+ MiModifier(Modifier) {}
+
+ bool operator!=(const AsmWriterOperand &Other) const {
+ if (OperandType != Other.OperandType || Str != Other.Str) return true;
+ if (OperandType == isMachineInstrOperand)
+ return MIOpNo != Other.MIOpNo || MiModifier != Other.MiModifier;
+ return false;
+ }
+ bool operator==(const AsmWriterOperand &Other) const {
+ return !operator!=(Other);
+ }
+ void EmitCode(std::ostream &OS) const;
+ };
+
+ struct AsmWriterInst {
+ std::vector<AsmWriterOperand> Operands;
+ const CodeGenInstruction *CGI;
+
+ AsmWriterInst(const CodeGenInstruction &CGI, unsigned Variant);
+
+ /// MatchesAllButOneOp - If this instruction is exactly identical to the
+ /// specified instruction except for one differing operand, return the
+ /// differing operand number. Otherwise return ~0.
+ unsigned MatchesAllButOneOp(const AsmWriterInst &Other) const;
+
+ private:
+ void AddLiteralString(const std::string &Str) {
+ // If the last operand was already a literal text string, append this to
+ // it, otherwise add a new operand.
+ if (!Operands.empty() &&
+ Operands.back().OperandType == AsmWriterOperand::isLiteralTextOperand)
+ Operands.back().Str.append(Str);
+ else
+ Operands.push_back(AsmWriterOperand(Str));
+ }
+ };
+}
+
+
+void AsmWriterOperand::EmitCode(std::ostream &OS) const {
+ if (OperandType == isLiteralTextOperand)
+ OS << "O << \"" << Str << "\"; ";
+ else {
+ OS << Str << "(MI, " << MIOpNo;
+ if (!MiModifier.empty())
+ OS << ", \"" << MiModifier << '"';
+ OS << "); ";
+ }
+}
+
+
+/// ParseAsmString - Parse the specified Instruction's AsmString into this
+/// AsmWriterInst.
+///
+AsmWriterInst::AsmWriterInst(const CodeGenInstruction &CGI, unsigned Variant) {
+ this->CGI = &CGI;
+ unsigned CurVariant = ~0U; // ~0 if we are outside a {.|.|.} region, other #.
+
+ // NOTE: Any extensions to this code need to be mirrored in the
+ // AsmPrinter::printInlineAsm code that executes as compile time (assuming
+ // that inline asm strings should also get the new feature)!
+ const std::string &AsmString = CGI.AsmString;
+ std::string::size_type LastEmitted = 0;
+ while (LastEmitted != AsmString.size()) {
+ std::string::size_type DollarPos =
+ AsmString.find_first_of("${|}", LastEmitted);
+ if (DollarPos == std::string::npos) DollarPos = AsmString.size();
+
+ // Emit a constant string fragment.
+ if (DollarPos != LastEmitted) {
+ // TODO: this should eventually handle escaping.
+ if (CurVariant == Variant || CurVariant == ~0U)
+ AddLiteralString(std::string(AsmString.begin()+LastEmitted,
+ AsmString.begin()+DollarPos));
+ LastEmitted = DollarPos;
+ } else if (AsmString[DollarPos] == '{') {
+ if (CurVariant != ~0U)
+ throw "Nested variants found for instruction '" +
+ CGI.TheDef->getName() + "'!";
+ LastEmitted = DollarPos+1;
+ CurVariant = 0; // We are now inside of the variant!
+ } else if (AsmString[DollarPos] == '|') {
+ if (CurVariant == ~0U)
+ throw "'|' character found outside of a variant in instruction '"
+ + CGI.TheDef->getName() + "'!";
+ ++CurVariant;
+ ++LastEmitted;
+ } else if (AsmString[DollarPos] == '}') {
+ if (CurVariant == ~0U)
+ throw "'}' character found outside of a variant in instruction '"
+ + CGI.TheDef->getName() + "'!";
+ ++LastEmitted;
+ CurVariant = ~0U;
+ } else if (DollarPos+1 != AsmString.size() &&
+ AsmString[DollarPos+1] == '$') {
+ if (CurVariant == Variant || CurVariant == ~0U)
+ AddLiteralString("$"); // "$$" -> $
+ LastEmitted = DollarPos+2;
+ } else {
+ // Get the name of the variable.
+ std::string::size_type VarEnd = DollarPos+1;
+
+ // handle ${foo}bar as $foo by detecting whether the character following
+ // the dollar sign is a curly brace. If so, advance VarEnd and DollarPos
+ // so the variable name does not contain the leading curly brace.
+ bool hasCurlyBraces = false;
+ if (VarEnd < AsmString.size() && '{' == AsmString[VarEnd]) {
+ hasCurlyBraces = true;
+ ++DollarPos;
+ ++VarEnd;
+ }
+
+ while (VarEnd < AsmString.size() && isIdentChar(AsmString[VarEnd]))
+ ++VarEnd;
+ std::string VarName(AsmString.begin()+DollarPos+1,
+ AsmString.begin()+VarEnd);
+
+ // Modifier - Support ${foo:modifier} syntax, where "modifier" is passed
+ // into printOperand.
+ std::string Modifier;
+
+ // In order to avoid starting the next string at the terminating curly
+ // brace, advance the end position past it if we found an opening curly
+ // brace.
+ if (hasCurlyBraces) {
+ if (VarEnd >= AsmString.size())
+ throw "Reached end of string before terminating curly brace in '"
+ + CGI.TheDef->getName() + "'";
+
+ // Look for a modifier string.
+ if (AsmString[VarEnd] == ':') {
+ ++VarEnd;
+ if (VarEnd >= AsmString.size())
+ throw "Reached end of string before terminating curly brace in '"
+ + CGI.TheDef->getName() + "'";
+
+ unsigned ModifierStart = VarEnd;
+ while (VarEnd < AsmString.size() && isIdentChar(AsmString[VarEnd]))
+ ++VarEnd;
+ Modifier = std::string(AsmString.begin()+ModifierStart,
+ AsmString.begin()+VarEnd);
+ if (Modifier.empty())
+ throw "Bad operand modifier name in '"+ CGI.TheDef->getName() + "'";
+ }
+
+ if (AsmString[VarEnd] != '}')
+ throw "Variable name beginning with '{' did not end with '}' in '"
+ + CGI.TheDef->getName() + "'";
+ ++VarEnd;
+ }
+ if (VarName.empty())
+ throw "Stray '$' in '" + CGI.TheDef->getName() +
+ "' asm string, maybe you want $$?";
+
+ unsigned OpNo = CGI.getOperandNamed(VarName);
+ CodeGenInstruction::OperandInfo OpInfo = CGI.OperandList[OpNo];
+
+ // If this is a two-address instruction and we are not accessing the
+ // 0th operand, remove an operand.
+ unsigned MIOp = OpInfo.MIOperandNo;
+ if (CGI.isTwoAddress && MIOp != 0) {
+ if (MIOp == 1)
+ throw "Should refer to operand #0 instead of #1 for two-address"
+ " instruction '" + CGI.TheDef->getName() + "'!";
+ --MIOp;
+ }
+
+ if (CurVariant == Variant || CurVariant == ~0U)
+ Operands.push_back(AsmWriterOperand(OpInfo.PrinterMethodName, MIOp,
+ Modifier));
+ LastEmitted = VarEnd;
+ }
+ }
+
+ AddLiteralString("\\n");
+}
+
+/// MatchesAllButOneOp - If this instruction is exactly identical to the
+/// specified instruction except for one differing operand, return the differing
+/// operand number. If more than one operand mismatches, return ~1, otherwise
+/// if the instructions are identical return ~0.
+unsigned AsmWriterInst::MatchesAllButOneOp(const AsmWriterInst &Other)const{
+ if (Operands.size() != Other.Operands.size()) return ~1;
+
+ unsigned MismatchOperand = ~0U;
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
+ if (Operands[i] != Other.Operands[i])
+ if (MismatchOperand != ~0U) // Already have one mismatch?
+ return ~1U;
+ else
+ MismatchOperand = i;
+ }
+ return MismatchOperand;
+}
+
+static void PrintCases(std::vector<std::pair<std::string,
+ AsmWriterOperand> > &OpsToPrint, std::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.
+ TheOp.EmitCode(O);
+ 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,
+ std::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);
+ }
+ }
+ }
+
+ std::string Namespace = FirstInst.CGI->Namespace;
+
+ O << " case " << Namespace << "::"
+ << FirstInst.CGI->TheDef->getName() << ":\n";
+ for (unsigned i = 0, e = SimilarInsts.size(); i != e; ++i)
+ O << " case " << 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].EmitCode(O);
+ } 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(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(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::run(std::ostream &O) {
EmitSourceFileHeader("Assembly Writer Source Fragment", O);
CodeGenTarget Target;
+ Record *AsmWriter = Target.getAsmWriter();
+ std::string ClassName = AsmWriter->getValueAsString("AsmWriterClassName");
+ unsigned Variant = AsmWriter->getValueAsInt("Variant");
+
O <<
"/// printInstruction - This method is automatically generated by tablegen\n"
"/// from the instruction set description. This method returns true if the\n"
"/// machine instruction was sufficiently described to print it, otherwise\n"
"/// it returns false.\n"
- "bool " << Target.getName()
- << "AsmPrinter::printInstruction(const MachineInstr *MI) {\n";
- O << " switch (MI->getOpcode()) {\n"
- " default: return false;\n";
+ "bool " << Target.getName() << ClassName
+ << "::printInstruction(const MachineInstr *MI) {\n";
std::string Namespace = Target.inst_begin()->second.Namespace;
+ std::vector<AsmWriterInst> Instructions;
+
for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
E = Target.inst_end(); I != E; ++I)
- if (!I->second.AsmString.empty()) {
- const std::string &AsmString = I->second.AsmString;
- O << " case " << Namespace << "::" << I->first << ": O ";
-
- std::string::size_type LastEmitted = 0;
- while (LastEmitted != AsmString.size()) {
- std::string::size_type DollarPos = AsmString.find('$', LastEmitted);
- if (DollarPos == std::string::npos) DollarPos = AsmString.size();
-
- // Emit a constant string fragment.
- if (DollarPos != LastEmitted) {
- // TODO: this should eventually handle escaping.
- O << " << \"" << std::string(AsmString.begin()+LastEmitted,
- AsmString.begin()+DollarPos) << "\"";
- LastEmitted = DollarPos;
- } else if (DollarPos+1 != AsmString.size() &&
- AsmString[DollarPos+1] == '$') {
- O << " << '$'"; // "$$" -> $
- } else {
- // Get the name of the variable.
- // TODO: should eventually handle ${foo}bar as $foo
- std::string::size_type VarEnd = DollarPos+1;
- while (VarEnd < AsmString.size() && isIdentChar(AsmString[VarEnd]))
- ++VarEnd;
- std::string VarName(AsmString.begin()+DollarPos+1,
- AsmString.begin()+VarEnd);
- if (VarName.empty())
- throw "Stray '$' in '" + I->first +
- "' asm string, maybe you want $$?";
- unsigned OpNo = I->second.getOperandNamed(VarName);
-
- // If this is a two-address instruction and we are not accessing the
- // 0th operand, remove an operand.
- if (I->second.isTwoAddress && OpNo != 0) {
- if (OpNo == 1)
- throw "Should refer to operand #0 instead of #1 for two-address"
- " instruction '" + I->first + "'!";
- --OpNo;
- }
-
- O << "; " << I->second.OperandList[OpNo].PrinterMethodName
- << "(MI, " << I->second.OperandList[OpNo].MIOperandNo << ", MVT::"
- << getName(I->second.OperandList[OpNo].Ty) << "); O ";
- LastEmitted = VarEnd;
- }
- }
+ if (!I->second.AsmString.empty())
+ Instructions.push_back(AsmWriterInst(I->second, Variant));
- O << " << '\\n'; break;\n";
+ // If all of the instructions start with a constant string (a very very common
+ // occurance), emit all of the constant strings as a big table lookup instead
+ // of requiring a switch for them.
+ bool AllStartWithString = true;
+
+ for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
+ if (Instructions[i].Operands.empty() ||
+ Instructions[i].Operands[0].OperandType !=
+ AsmWriterOperand::isLiteralTextOperand) {
+ AllStartWithString = false;
+ break;
+ }
+
+ std::vector<const CodeGenInstruction*> NumberedInstructions;
+ Target.getInstructionsByEnumValue(NumberedInstructions);
+
+ if (AllStartWithString) {
+ // 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.
+ std::map<const CodeGenInstruction*, AsmWriterInst*> CGIAWIMap;
+ for (unsigned i = 0, e = Instructions.size(); i != e; ++i)
+ CGIAWIMap.insert(std::make_pair(Instructions[i].CGI, &Instructions[i]));
+
+ // Emit a table of constant strings.
+ O << " static const char * const OpStrs[] = {\n";
+ for (unsigned i = 0, e = NumberedInstructions.size(); i != e; ++i) {
+ AsmWriterInst *AWI = CGIAWIMap[NumberedInstructions[i]];
+ if (AWI == 0) {
+ // Something not handled by the asmwriter printer.
+ O << " 0,\t// ";
+ } else {
+ O << " \"" << AWI->Operands[0].Str << "\",\t// ";
+ // Nuke the string from the operand list. It is now handled!
+ AWI->Operands.erase(AWI->Operands.begin());
+ }
+ O << NumberedInstructions[i]->TheDef->getName() << "\n";
}
+ O << " };\n\n"
+ << " // Emit the opcode for the instruction.\n"
+ << " if (const char *AsmStr = OpStrs[MI->getOpcode()])\n"
+ << " O << AsmStr;\n\n";
+ }
+
+ // 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());
+
+ // Find the opcode # of inline asm
+ O << " switch (MI->getOpcode()) {\n"
+ " default: return false;\n"
+ " case " << NumberedInstructions.back()->Namespace
+ << "::INLINEASM: printInlineAsm(MI); break;\n";
+
+ while (!Instructions.empty())
+ EmitInstructions(Instructions, O);
O << " }\n"
" return true;\n"
"}\n";
- EmitSourceFileTail(O);
}
projects / oota-llvm.git / blobdiff