}
/// TokenizeAsmString - Tokenize a simplified assembly string.
-static void TokenizeAsmString(const StringRef &AsmString,
+static void TokenizeAsmString(StringRef AsmString,
SmallVectorImpl<StringRef> &Tokens) {
unsigned Prev = 0;
bool InTok = true;
Tokens.push_back(AsmString.substr(Prev));
}
-static bool IsAssemblerInstruction(const StringRef &Name,
+static bool IsAssemblerInstruction(StringRef Name,
const CodeGenInstruction &CGI,
const SmallVectorImpl<StringRef> &Tokens) {
- // Ignore psuedo ops.
+ // Ignore "codegen only" instructions.
+ if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
+ return false;
+
+ // Ignore pseudo ops.
//
- // FIXME: This is a hack.
+ // FIXME: This is a hack; can we convert these instructions to set the
+ // "codegen only" bit instead?
if (const RecordVal *Form = CGI.TheDef->getValue("Form"))
if (Form->getValue()->getAsString() == "Pseudo")
return false;
-
- // Ignore "PHI" node.
- //
- // FIXME: This is also a hack.
- if (Name == "PHI")
- return false;
// Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
//
//
// FIXME: Is this true?
//
- // Also, we ignore instructions which reference the operand multiple times;
- // this implies a constraint we would not currently honor. These are
- // currently always fake instructions for simplifying codegen.
- //
- // FIXME: Encode this assumption in the .td, so we can error out here.
+ // Also, check for instructions which reference the operand multiple times;
+ // this implies a constraint we would not honor.
std::set<std::string> OperandNames;
for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
if (Tokens[i][0] == '$' &&
DEBUG({
errs() << "warning: '" << Name << "': "
<< "ignoring instruction; operand with attribute '"
- << Tokens[i] << "', \n";
+ << Tokens[i] << "'\n";
});
return false;
}
if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
- DEBUG({
- errs() << "warning: '" << Name << "': "
- << "ignoring instruction; tied operand '"
- << Tokens[i] << "'\n";
- });
- return false;
+ std::string Err = "'" + Name.str() + "': " +
+ "invalid assembler instruction; tied operand '" + Tokens[i].str() + "'";
+ throw TGError(CGI.TheDef->getLoc(), Err);
}
}
if (Operands.size() != RHS.Operands.size())
return false;
+ // Otherwise, make sure the ordering of the two instructions is unambiguous
+ // by checking that either (a) a token or operand kind discriminates them,
+ // or (b) the ordering among equivalent kinds is consistent.
+
// Tokens and operand kinds are unambiguous (assuming a correct target
// specific parser).
for (unsigned i = 0, e = Operands.size(); i != e; ++i)
private:
/// getTokenClass - Lookup or create the class for the given token.
- ClassInfo *getTokenClass(const StringRef &Token);
+ ClassInfo *getTokenClass(StringRef Token);
/// getOperandClass - Lookup or create the class for the given operand.
- ClassInfo *getOperandClass(const StringRef &Token,
+ ClassInfo *getOperandClass(StringRef Token,
const CodeGenInstruction::OperandInfo &OI);
/// BuildRegisterClasses - Build the ClassInfo* instances for register
/// classes.
- void BuildRegisterClasses(CodeGenTarget &Target);
+ void BuildRegisterClasses(CodeGenTarget &Target,
+ std::set<std::string> &SingletonRegisterNames);
/// BuildOperandClasses - Build the ClassInfo* instances for user defined
/// operand classes.
continue;
}
+ if (!Op.OperandInfo) {
+ errs() << "(singleton register)\n";
+ continue;
+ }
+
const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
errs() << OI.Name << " " << OI.Rec->getName()
<< " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
}
}
-static std::string getEnumNameForToken(const StringRef &Str) {
+static std::string getEnumNameForToken(StringRef Str) {
std::string Res;
for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
return Res;
}
-ClassInfo *AsmMatcherInfo::getTokenClass(const StringRef &Token) {
+/// getRegisterRecord - Get the register record for \arg name, or 0.
+static Record *getRegisterRecord(CodeGenTarget &Target, StringRef Name) {
+ for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
+ const CodeGenRegister &Reg = Target.getRegisters()[i];
+ if (Name == Reg.TheDef->getValueAsString("AsmName"))
+ return Reg.TheDef;
+ }
+
+ return 0;
+}
+
+ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
ClassInfo *&Entry = TokenClasses[Token];
if (!Entry) {
}
ClassInfo *
-AsmMatcherInfo::getOperandClass(const StringRef &Token,
+AsmMatcherInfo::getOperandClass(StringRef Token,
const CodeGenInstruction::OperandInfo &OI) {
if (OI.Rec->isSubClassOf("RegisterClass")) {
ClassInfo *CI = RegisterClassClasses[OI.Rec];
return CI;
}
-void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target) {
+void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target,
+ std::set<std::string>
+ &SingletonRegisterNames) {
std::vector<CodeGenRegisterClass> RegisterClasses;
std::vector<CodeGenRegister> Registers;
ie = RegisterClasses.end(); it != ie; ++it)
RegisterSets.insert(std::set<Record*>(it->Elements.begin(),
it->Elements.end()));
-
+
+ // Add any required singleton sets.
+ for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
+ ie = SingletonRegisterNames.end(); it != ie; ++it)
+ if (Record *Rec = getRegisterRecord(Target, *it))
+ RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
+
// Introduce derived sets where necessary (when a register does not determine
// a unique register set class), and build the mapping of registers to the set
// they should classify to.
for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
ie = RegisterMap.end(); it != ie; ++it)
this->RegisterClasses[it->first] = RegisterSetClasses[it->second];
+
+ // Name the register classes which correspond to singleton registers.
+ for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
+ ie = SingletonRegisterNames.end(); it != ie; ++it) {
+ if (Record *Rec = getRegisterRecord(Target, *it)) {
+ ClassInfo *CI = this->RegisterClasses[Rec];
+ assert(CI && "Missing singleton register class info!");
+
+ if (CI->ValueName.empty()) {
+ CI->ClassName = Rec->getName();
+ CI->Name = "MCK_" + Rec->getName();
+ CI->ValueName = Rec->getName();
+ } else
+ CI->ValueName = CI->ValueName + "," + Rec->getName();
+ }
+ }
}
void AsmMatcherInfo::BuildOperandClasses(CodeGenTarget &Target) {
std::vector<Record*> AsmOperands;
AsmOperands = Records.getAllDerivedDefinitions("AsmOperandClass");
+
+ // Pre-populate AsmOperandClasses map.
+ for (std::vector<Record*>::iterator it = AsmOperands.begin(),
+ ie = AsmOperands.end(); it != ie; ++it)
+ AsmOperandClasses[*it] = new ClassInfo();
+
unsigned Index = 0;
for (std::vector<Record*>::iterator it = AsmOperands.begin(),
ie = AsmOperands.end(); it != ie; ++it, ++Index) {
- ClassInfo *CI = new ClassInfo();
+ ClassInfo *CI = AsmOperandClasses[*it];
CI->Kind = ClassInfo::UserClass0 + Index;
Init *Super = (*it)->getValueInit("SuperClass");
}
void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
- // Build info for the register classes.
- BuildRegisterClasses(Target);
-
- // Build info for the user defined assembly operand classes.
- BuildOperandClasses(Target);
-
- // Build the instruction information.
- for (std::map<std::string, CodeGenInstruction>::const_iterator
- it = Target.getInstructions().begin(),
- ie = Target.getInstructions().end();
- it != ie; ++it) {
- const CodeGenInstruction &CGI = it->second;
+ // Parse the instructions; we need to do this first so that we can gather the
+ // singleton register classes.
+ std::set<std::string> SingletonRegisterNames;
+
+ const std::vector<const CodeGenInstruction*> &InstrList =
+ Target.getInstructionsByEnumValue();
+
+ for (unsigned i = 0, e = InstrList.size(); i != e; ++i) {
+ const CodeGenInstruction &CGI = *InstrList[i];
- if (!StringRef(it->first).startswith(MatchPrefix))
+ if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
continue;
- OwningPtr<InstructionInfo> II(new InstructionInfo);
+ OwningPtr<InstructionInfo> II(new InstructionInfo());
- II->InstrName = it->first;
- II->Instr = &it->second;
+ II->InstrName = CGI.TheDef->getName();
+ II->Instr = &CGI;
II->AsmString = FlattenVariants(CGI.AsmString, 0);
// Remove comments from the asm string.
TokenizeAsmString(II->AsmString, II->Tokens);
// Ignore instructions which shouldn't be matched.
- if (!IsAssemblerInstruction(it->first, CGI, II->Tokens))
+ if (!IsAssemblerInstruction(CGI.TheDef->getName(), CGI, II->Tokens))
continue;
+ // Collect singleton registers, if used.
+ if (!RegisterPrefix.empty()) {
+ for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
+ if (II->Tokens[i].startswith(RegisterPrefix)) {
+ StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
+ Record *Rec = getRegisterRecord(Target, RegName);
+
+ if (!Rec) {
+ std::string Err = "unable to find register for '" + RegName.str() +
+ "' (which matches register prefix)";
+ throw TGError(CGI.TheDef->getLoc(), Err);
+ }
+
+ SingletonRegisterNames.insert(RegName);
+ }
+ }
+ }
+
+ Instructions.push_back(II.take());
+ }
+
+ // Build info for the register classes.
+ BuildRegisterClasses(Target, SingletonRegisterNames);
+
+ // Build info for the user defined assembly operand classes.
+ BuildOperandClasses(Target);
+
+ // Build the instruction information.
+ for (std::vector<InstructionInfo*>::iterator it = Instructions.begin(),
+ ie = Instructions.end(); it != ie; ++it) {
+ InstructionInfo *II = *it;
+
for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
StringRef Token = II->Tokens[i];
+ // Check for singleton registers.
+ if (!RegisterPrefix.empty() && Token.startswith(RegisterPrefix)) {
+ StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
+ InstructionInfo::Operand Op;
+ Op.Class = RegisterClasses[getRegisterRecord(Target, RegName)];
+ Op.OperandInfo = 0;
+ assert(Op.Class && Op.Class->Registers.size() == 1 &&
+ "Unexpected class for singleton register");
+ II->Operands.push_back(Op);
+ continue;
+ }
+
// Check for simple tokens.
if (Token[0] != '$') {
InstructionInfo::Operand Op;
// Map this token to an operand. FIXME: Move elsewhere.
unsigned Idx;
try {
- Idx = CGI.getOperandNamed(OperandName);
+ Idx = II->Instr->getOperandNamed(OperandName);
} catch(...) {
- errs() << "error: unable to find operand: '" << OperandName << "'!\n";
- break;
+ throw std::string("error: unable to find operand: '" +
+ OperandName.str() + "'");
+ }
+
+ // FIXME: This is annoying, the named operand may be tied (e.g.,
+ // XCHG8rm). What we want is the untied operand, which we now have to
+ // grovel for. Only worry about this for single entry operands, we have to
+ // clean this up anyway.
+ const CodeGenInstruction::OperandInfo *OI = &II->Instr->OperandList[Idx];
+ if (OI->Constraints[0].isTied()) {
+ unsigned TiedOp = OI->Constraints[0].getTiedOperand();
+
+ // The tied operand index is an MIOperand index, find the operand that
+ // contains it.
+ for (unsigned i = 0, e = II->Instr->OperandList.size(); i != e; ++i) {
+ if (II->Instr->OperandList[i].MIOperandNo == TiedOp) {
+ OI = &II->Instr->OperandList[i];
+ break;
+ }
+ }
+
+ assert(OI && "Unable to find tied operand target!");
}
- const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx];
InstructionInfo::Operand Op;
- Op.Class = getOperandClass(Token, OI);
- Op.OperandInfo = &OI;
+ Op.Class = getOperandClass(Token, *OI);
+ Op.OperandInfo = OI;
II->Operands.push_back(Op);
}
-
- // If we broke out, ignore the instruction.
- if (II->Operands.size() != II->Tokens.size())
- continue;
-
- Instructions.push_back(II.take());
}
// Reorder classes so that classes preceed super classes.
std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
}
+static std::pair<unsigned, unsigned> *
+GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List,
+ unsigned Index) {
+ for (unsigned i = 0, e = List.size(); i != e; ++i)
+ if (Index == List[i].first)
+ return &List[i];
+
+ return 0;
+}
+
static void EmitConvertToMCInst(CodeGenTarget &Target,
std::vector<InstructionInfo*> &Infos,
raw_ostream &OS) {
// Start the unified conversion function.
- CvtOS << "static bool ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
+ CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
<< "unsigned Opcode,\n"
- << " SmallVectorImpl<"
- << Target.getName() << "Operand> &Operands) {\n";
+ << " const SmallVectorImpl<MCParsedAsmOperand*"
+ << "> &Operands) {\n";
CvtOS << " Inst.setOpcode(Opcode);\n";
CvtOS << " switch (Kind) {\n";
CvtOS << " default:\n";
OS << "// Unified function for converting operants to MCInst instances.\n\n";
OS << "enum ConversionKind {\n";
+ // TargetOperandClass - This is the target's operand class, like X86Operand.
+ std::string TargetOperandClass = Target.getName() + "Operand";
+
for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
ie = Infos.end(); it != ie; ++it) {
InstructionInfo &II = **it;
if (Op.OperandInfo)
MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
}
+
+ // Find any tied operands.
+ SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands;
+ for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
+ const CodeGenInstruction::OperandInfo &OpInfo = II.Instr->OperandList[i];
+ for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) {
+ const CodeGenInstruction::ConstraintInfo &CI = OpInfo.Constraints[j];
+ if (CI.isTied())
+ TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo + j,
+ CI.getTiedOperand()));
+ }
+ }
+
std::sort(MIOperandList.begin(), MIOperandList.end());
// Compute the total number of operands.
assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
"Duplicate match for instruction operand!");
- Signature += "_";
-
// Skip operands which weren't matched by anything, this occurs when the
// .td file encodes "implicit" operands as explicit ones.
//
// FIXME: This should be removed from the MCInst structure.
- for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
- Signature += "Imp";
+ for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
+ std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
+ CurIndex);
+ if (!Tie)
+ Signature += "__Imp";
+ else
+ Signature += "__Tie" + utostr(Tie->second);
+ }
+
+ Signature += "__";
// Registers are always converted the same, don't duplicate the conversion
// function based on them.
}
// Add any trailing implicit operands.
- for (; CurIndex != NumMIOperands; ++CurIndex)
- Signature += "Imp";
+ for (; CurIndex != NumMIOperands; ++CurIndex) {
+ std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
+ CurIndex);
+ if (!Tie)
+ Signature += "__Imp";
+ else
+ Signature += "__Tie" + utostr(Tie->second);
+ }
II.ConversionFnKind = Signature;
InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
// Add the implicit operands.
- for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
- CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
+ for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
+ // See if this is a tied operand.
+ std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
+ CurIndex);
+
+ if (!Tie) {
+ // If not, this is some implicit operand. Just assume it is a register
+ // for now.
+ CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
+ } else {
+ // Copy the tied operand.
+ assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
+ CvtOS << " Inst.addOperand(Inst.getOperand("
+ << Tie->second << "));\n";
+ }
+ }
- CvtOS << " Operands[" << MIOperandList[i].second
- << "]." << Op.Class->RenderMethod
+ CvtOS << " ((" << TargetOperandClass << "*)Operands["
+ << MIOperandList[i].second
+ << "])->" << Op.Class->RenderMethod
<< "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
CurIndex += Op.OperandInfo->MINumOperands;
}
// And add trailing implicit operands.
- for (; CurIndex != NumMIOperands; ++CurIndex)
- CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
- CvtOS << " break;\n";
+ for (; CurIndex != NumMIOperands; ++CurIndex) {
+ std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
+ CurIndex);
+
+ if (!Tie) {
+ // If not, this is some implicit operand. Just assume it is a register
+ // for now.
+ CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
+ } else {
+ // Copy the tied operand.
+ assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
+ CvtOS << " Inst.addOperand(Inst.getOperand("
+ << Tie->second << "));\n";
+ }
+ }
+
+ CvtOS << " return;\n";
}
// Finish the convert function.
CvtOS << " }\n";
- CvtOS << " return false;\n";
CvtOS << "}\n\n";
// Finish the enum, and drop the convert function after it.
static void EmitClassifyOperand(CodeGenTarget &Target,
AsmMatcherInfo &Info,
raw_ostream &OS) {
- OS << "static MatchClassKind ClassifyOperand("
- << Target.getName() << "Operand &Operand) {\n";
+ OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n"
+ << " " << Target.getName() << "Operand &Operand = *("
+ << Target.getName() << "Operand*)GOp;\n";
// Classify tokens.
OS << " if (Operand.isToken())\n";
Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";"));
}
- OS << "static MatchClassKind MatchTokenString(const StringRef &Name) {\n";
+ OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
EmitStringMatcher("Name", Matches, OS);
"return " + utostr(i + 1) + ";"));
}
- OS << "unsigned " << Target.getName()
- << AsmParser->getValueAsString("AsmParserClassName")
- << "::MatchRegisterName(const StringRef &Name) {\n";
+ OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
EmitStringMatcher("Name", Matches, OS);
Record *AsmParser = Target.getAsmParser();
std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
- EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
-
- // Emit the function to match a register name to number.
- EmitMatchRegisterName(Target, AsmParser, OS);
-
// Compute the information on the instructions to match.
AsmMatcherInfo Info(AsmParser);
Info.BuildInfo(Target);
- // Sort the instruction table using the partial order on classes.
- std::sort(Info.Instructions.begin(), Info.Instructions.end(),
- less_ptr<InstructionInfo>());
+ // Sort the instruction table using the partial order on classes. We use
+ // stable_sort to ensure that ambiguous instructions are still
+ // deterministically ordered.
+ std::stable_sort(Info.Instructions.begin(), Info.Instructions.end(),
+ less_ptr<InstructionInfo>());
DEBUG_WITH_TYPE("instruction_info", {
for (std::vector<InstructionInfo*>::iterator
<< " ambiguous instructions!\n";
});
+ // Write the output.
+
+ EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
+
+ // Emit the function to match a register name to number.
+ EmitMatchRegisterName(Target, AsmParser, OS);
+
+ OS << "#ifndef REGISTERS_ONLY\n\n";
+
// Generate the unified function to convert operands into an MCInst.
EmitConvertToMCInst(Target, Info.Instructions, OS);
MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
OS << "bool " << Target.getName() << ClassName
- << "::MatchInstruction("
- << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
- << "MCInst &Inst) {\n";
+ << "::\nMatchInstruction(const SmallVectorImpl<MCParsedAsmOperand*> "
+ "&Operands,\n MCInst &Inst) {\n";
// Emit the static match table; unused classes get initalized to 0 which is
// guaranteed to be InvalidMatchClass.
OS << " continue;\n";
}
OS << "\n";
- OS << " return ConvertToMCInst(it->ConvertFn, Inst, "
- << "it->Opcode, Operands);\n";
+ OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
+
+ // Call the post-processing function, if used.
+ std::string InsnCleanupFn =
+ AsmParser->getValueAsString("AsmParserInstCleanup");
+ if (!InsnCleanupFn.empty())
+ OS << " " << InsnCleanupFn << "(Inst);\n";
+
+ OS << " return false;\n";
OS << " }\n\n";
OS << " return true;\n";
OS << "}\n\n";
+
+ OS << "#endif // REGISTERS_ONLY\n";
}