#include <set>
using namespace llvm;
-namespace {
static cl::opt<std::string>
MatchPrefix("match-prefix", cl::init(""),
cl::desc("Only match instructions with the given prefix"));
-}
/// FlattenVariants - Flatten an .td file assembly string by selecting the
/// variant at index \arg N.
}
/// 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.
+
+ // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
//
- // FIXME: This is also a hack.
- if (Name == "PHI")
+ // FIXME: This is a total hack.
+ if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int"))
return false;
// Ignore instructions with no .s string.
//
// 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);
}
}
/// class of operands which can be matched.
struct ClassInfo {
enum ClassInfoKind {
- Invalid = 0, ///< Invalid kind, for use as a sentinel value.
- Token, ///< The class for a particular token.
- Register, ///< A register class.
- UserClass0 ///< The (first) user defined class, subsequent user defined
- /// classes are UserClass0+1, and so on.
+ /// Invalid kind, for use as a sentinel value.
+ Invalid = 0,
+
+ /// The class for a particular token.
+ Token,
+
+ /// The (first) register class, subsequent register classes are
+ /// RegisterClass0+1, and so on.
+ RegisterClass0,
+
+ /// The (first) user defined class, subsequent user defined classes are
+ /// UserClass0+1, and so on.
+ UserClass0 = 1<<16
};
/// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
/// N) for the Nth user defined class.
unsigned Kind;
+ /// SuperClasses - The super classes of this class. Note that for simplicities
+ /// sake user operands only record their immediate super class, while register
+ /// operands include all superclasses.
+ std::vector<ClassInfo*> SuperClasses;
+
/// Name - The full class name, suitable for use in an enum.
std::string Name;
std::string ValueName;
/// PredicateMethod - The name of the operand method to test whether the
- /// operand matches this class; this is not valid for Token kinds.
+ /// operand matches this class; this is not valid for Token or register kinds.
std::string PredicateMethod;
/// RenderMethod - The name of the operand method to add this operand to an
- /// MCInst; this is not valid for Token kinds.
+ /// MCInst; this is not valid for Token or register kinds.
std::string RenderMethod;
+ /// For register classes, the records for all the registers in this class.
+ std::set<Record*> Registers;
+
+public:
+ /// isRegisterClass() - Check if this is a register class.
+ bool isRegisterClass() const {
+ return Kind >= RegisterClass0 && Kind < UserClass0;
+ }
+
+ /// isUserClass() - Check if this is a user defined class.
+ bool isUserClass() const {
+ return Kind >= UserClass0;
+ }
+
+ /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
+ /// are related if they are in the same class hierarchy.
+ bool isRelatedTo(const ClassInfo &RHS) const {
+ // Tokens are only related to tokens.
+ if (Kind == Token || RHS.Kind == Token)
+ return Kind == Token && RHS.Kind == Token;
+
+ // Registers classes are only related to registers classes, and only if
+ // their intersection is non-empty.
+ if (isRegisterClass() || RHS.isRegisterClass()) {
+ if (!isRegisterClass() || !RHS.isRegisterClass())
+ return false;
+
+ std::set<Record*> Tmp;
+ std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
+ std::set_intersection(Registers.begin(), Registers.end(),
+ RHS.Registers.begin(), RHS.Registers.end(),
+ II);
+
+ return !Tmp.empty();
+ }
+
+ // Otherwise we have two users operands; they are related if they are in the
+ // same class hierarchy.
+ //
+ // FIXME: This is an oversimplification, they should only be related if they
+ // intersect, however we don't have that information.
+ assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
+ const ClassInfo *Root = this;
+ while (!Root->SuperClasses.empty())
+ Root = Root->SuperClasses.front();
+
+ const ClassInfo *RHSRoot = &RHS;
+ while (!RHSRoot->SuperClasses.empty())
+ RHSRoot = RHSRoot->SuperClasses.front();
+
+ return Root == RHSRoot;
+ }
+
+ /// isSubsetOf - Test whether this class is a subset of \arg RHS;
+ bool isSubsetOf(const ClassInfo &RHS) const {
+ // This is a subset of RHS if it is the same class...
+ if (this == &RHS)
+ return true;
+
+ // ... or if any of its super classes are a subset of RHS.
+ for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
+ ie = SuperClasses.end(); it != ie; ++it)
+ if ((*it)->isSubsetOf(RHS))
+ return true;
+
+ return false;
+ }
+
/// operator< - Compare two classes.
bool operator<(const ClassInfo &RHS) const {
- // Incompatible kinds are comparable.
- if (Kind != RHS.Kind)
+ if (this == &RHS)
+ return false;
+
+ // Unrelated classes can be ordered by kind.
+ if (!isRelatedTo(RHS))
return Kind < RHS.Kind;
switch (Kind) {
case Invalid:
assert(0 && "Invalid kind!");
case Token:
- // Tokens are always comparable.
+ // Tokens are comparable by value.
//
// FIXME: Compare by enum value.
return ValueName < RHS.ValueName;
- case Register:
- // FIXME: Compare by subset relation.
- return false;
-
default:
- // FIXME: Allow user defined relation.
- return false;
+ // This class preceeds the RHS if it is a proper subset of the RHS.
+ if (isSubsetOf(RHS))
+ return true;
+ if (RHS.isSubsetOf(*this))
+ return false;
+
+ // Otherwise, order by name to ensure we have a total ordering.
+ return ValueName < RHS.ValueName;
}
}
};
/// operator< - Compare two instructions.
bool operator<(const InstructionInfo &RHS) const {
- // Order first by the number of operands (which is unambiguous).
if (Operands.size() != RHS.Operands.size())
return Operands.size() < RHS.Operands.size();
-
- // Otherwise, order by lexicographic comparison of tokens and operand kinds
- // (these can never be ambiguous).
- for (unsigned i = 0, e = Operands.size(); i != e; ++i)
- if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
- Operands[i].Class->Kind == ClassInfo::Token)
- if (*Operands[i].Class < *RHS.Operands[i].Class)
- return true;
-
- // Finally, order by the component wise comparison of operand classes. We
- // don't want to rely on the lexigraphic ordering of elements, so we define
- // only define the ordering when it is unambiguous. That is, when some pair
- // compares less than and no pair compares greater than.
- // Check that no pair compares greater than.
- for (unsigned i = 0, e = Operands.size(); i != e; ++i)
+ // Compare lexicographically by operand. The matcher validates that other
+ // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith().
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
+ if (*Operands[i].Class < *RHS.Operands[i].Class)
+ return true;
if (*RHS.Operands[i].Class < *Operands[i].Class)
return false;
+ }
+
+ return false;
+ }
- // Otherwise, return true if some pair compares less than.
+ /// CouldMatchAmiguouslyWith - Check whether this instruction could
+ /// ambiguously match the same set of operands as \arg RHS (without being a
+ /// strictly superior match).
+ bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
+ // The number of operands is unambiguous.
+ 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)
+ if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
+ Operands[i].Class->Kind == ClassInfo::Token)
+ if (*Operands[i].Class < *RHS.Operands[i].Class ||
+ *RHS.Operands[i].Class < *Operands[i].Class)
+ return false;
+
+ // Otherwise, this operand could commute if all operands are equivalent, or
+ // there is a pair of operands that compare less than and a pair that
+ // compare greater than.
+ bool HasLT = false, HasGT = false;
+ for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
if (*Operands[i].Class < *RHS.Operands[i].Class)
- return true;
+ HasLT = true;
+ if (*RHS.Operands[i].Class < *Operands[i].Class)
+ HasGT = true;
+ }
- return false;
+ return !(HasLT ^ HasGT);
}
public:
class AsmMatcherInfo {
public:
+ /// The tablegen AsmParser record.
+ Record *AsmParser;
+
+ /// The AsmParser "CommentDelimiter" value.
+ std::string CommentDelimiter;
+
+ /// The AsmParser "RegisterPrefix" value.
+ std::string RegisterPrefix;
+
/// The classes which are needed for matching.
std::vector<ClassInfo*> Classes;
/// The information on the instruction to match.
std::vector<InstructionInfo*> Instructions;
+ /// Map of Register records to their class information.
+ std::map<Record*, ClassInfo*> RegisterClasses;
+
private:
/// Map of token to class information which has already been constructed.
std::map<std::string, ClassInfo*> TokenClasses;
- /// Map of operand name to class information which has already been
- /// constructed.
- std::map<std::string, ClassInfo*> OperandClasses;
+ /// Map of RegisterClass records to their class information.
+ std::map<Record*, ClassInfo*> RegisterClassClasses;
- /// Map of user class names to kind value.
- std::map<std::string, unsigned> UserClasses;
+ /// Map of AsmOperandClass records to their class information.
+ std::map<Record*, ClassInfo*> AsmOperandClasses;
private:
/// getTokenClass - Lookup or create the class for the given token.
- ClassInfo *getTokenClass(const StringRef &Token);
-
- /// getUserClassKind - Lookup or create the kind value for the given class
- /// name.
- unsigned getUserClassKind(const StringRef &Name);
+ 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,
+ std::set<std::string> &SingletonRegisterNames);
+
+ /// BuildOperandClasses - Build the ClassInfo* instances for user defined
+ /// operand classes.
+ void BuildOperandClasses(CodeGenTarget &Target);
+
public:
+ AsmMatcherInfo(Record *_AsmParser);
+
/// BuildInfo - Construct the various tables used during matching.
void BuildInfo(CodeGenTarget &Target);
};
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) {
return Entry;
}
-unsigned AsmMatcherInfo::getUserClassKind(const StringRef &Name) {
- unsigned &Entry = UserClasses[Name];
-
- if (!Entry)
- Entry = ClassInfo::UserClass0 + UserClasses.size() - 1;
-
- return Entry;
-}
-
ClassInfo *
-AsmMatcherInfo::getOperandClass(const StringRef &Token,
+AsmMatcherInfo::getOperandClass(StringRef Token,
const CodeGenInstruction::OperandInfo &OI) {
- std::string ClassName;
if (OI.Rec->isSubClassOf("RegisterClass")) {
- ClassName = "Reg";
- } else {
- try {
- ClassName = OI.Rec->getValueAsString("ParserMatchClass");
- assert(ClassName != "Reg" && "'Reg' class name is reserved!");
- } catch(...) {
- PrintError(OI.Rec->getLoc(), "operand has no match class!");
- ClassName = "Invalid";
+ ClassInfo *CI = RegisterClassClasses[OI.Rec];
+
+ if (!CI) {
+ PrintError(OI.Rec->getLoc(), "register class has no class info!");
+ throw std::string("ERROR: Missing register class!");
}
+
+ return CI;
}
- ClassInfo *&Entry = OperandClasses[ClassName];
-
- if (!Entry) {
- Entry = new ClassInfo();
- // FIXME: Hack.
- if (ClassName == "Reg") {
- Entry->Kind = ClassInfo::Register;
+ assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!");
+ Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass");
+ ClassInfo *CI = AsmOperandClasses[MatchClass];
+
+ if (!CI) {
+ PrintError(OI.Rec->getLoc(), "operand has no match class!");
+ throw std::string("ERROR: Missing match class!");
+ }
+
+ return CI;
+}
+
+void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target,
+ std::set<std::string>
+ &SingletonRegisterNames) {
+ std::vector<CodeGenRegisterClass> RegisterClasses;
+ std::vector<CodeGenRegister> Registers;
+
+ RegisterClasses = Target.getRegisterClasses();
+ Registers = Target.getRegisters();
+
+ // The register sets used for matching.
+ std::set< std::set<Record*> > RegisterSets;
+
+ // Gather the defined sets.
+ for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
+ 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.
+ std::map<Record*, std::set<Record*> > RegisterMap;
+ for (std::vector<CodeGenRegister>::iterator it = Registers.begin(),
+ ie = Registers.end(); it != ie; ++it) {
+ CodeGenRegister &CGR = *it;
+ // Compute the intersection of all sets containing this register.
+ std::set<Record*> ContainingSet;
+
+ for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
+ ie = RegisterSets.end(); it != ie; ++it) {
+ if (!it->count(CGR.TheDef))
+ continue;
+
+ if (ContainingSet.empty()) {
+ ContainingSet = *it;
+ } else {
+ std::set<Record*> Tmp;
+ std::swap(Tmp, ContainingSet);
+ std::insert_iterator< std::set<Record*> > II(ContainingSet,
+ ContainingSet.begin());
+ std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(),
+ II);
+ }
+ }
+
+ if (!ContainingSet.empty()) {
+ RegisterSets.insert(ContainingSet);
+ RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
+ }
+ }
+
+ // Construct the register classes.
+ std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
+ unsigned Index = 0;
+ for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
+ ie = RegisterSets.end(); it != ie; ++it, ++Index) {
+ ClassInfo *CI = new ClassInfo();
+ CI->Kind = ClassInfo::RegisterClass0 + Index;
+ CI->ClassName = "Reg" + utostr(Index);
+ CI->Name = "MCK_Reg" + utostr(Index);
+ CI->ValueName = "";
+ CI->PredicateMethod = ""; // unused
+ CI->RenderMethod = "addRegOperands";
+ CI->Registers = *it;
+ Classes.push_back(CI);
+ RegisterSetClasses.insert(std::make_pair(*it, CI));
+ }
+
+ // Find the superclasses; we could compute only the subgroup lattice edges,
+ // but there isn't really a point.
+ for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
+ ie = RegisterSets.end(); it != ie; ++it) {
+ ClassInfo *CI = RegisterSetClasses[*it];
+ for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
+ ie2 = RegisterSets.end(); it2 != ie2; ++it2)
+ if (*it != *it2 &&
+ std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
+ CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
+ }
+
+ // Name the register classes which correspond to a user defined RegisterClass.
+ for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
+ ie = RegisterClasses.end(); it != ie; ++it) {
+ ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(),
+ it->Elements.end())];
+ if (CI->ValueName.empty()) {
+ CI->ClassName = it->getName();
+ CI->Name = "MCK_" + it->getName();
+ CI->ValueName = it->getName();
+ } else
+ CI->ValueName = CI->ValueName + "," + it->getName();
+
+ RegisterClassClasses.insert(std::make_pair(it->TheDef, CI));
+ }
+
+ // Populate the map for individual registers.
+ 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 = AsmOperandClasses[*it];
+ CI->Kind = ClassInfo::UserClass0 + Index;
+
+ ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
+ for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
+ DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
+ if (!DI) {
+ PrintError((*it)->getLoc(), "Invalid super class reference!");
+ continue;
+ }
+
+ ClassInfo *SC = AsmOperandClasses[DI->getDef()];
+ if (!SC)
+ PrintError((*it)->getLoc(), "Invalid super class reference!");
+ else
+ CI->SuperClasses.push_back(SC);
+ }
+ CI->ClassName = (*it)->getValueAsString("Name");
+ CI->Name = "MCK_" + CI->ClassName;
+ CI->ValueName = (*it)->getName();
+
+ // Get or construct the predicate method name.
+ Init *PMName = (*it)->getValueInit("PredicateMethod");
+ if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
+ CI->PredicateMethod = SI->getValue();
} else {
- Entry->Kind = getUserClassKind(ClassName);
+ assert(dynamic_cast<UnsetInit*>(PMName) &&
+ "Unexpected PredicateMethod field!");
+ CI->PredicateMethod = "is" + CI->ClassName;
}
- Entry->ClassName = ClassName;
- Entry->Name = "MCK_" + ClassName;
- Entry->ValueName = OI.Rec->getName();
- Entry->PredicateMethod = "is" + ClassName;
- Entry->RenderMethod = "add" + ClassName + "Operands";
- Classes.push_back(Entry);
+
+ // Get or construct the render method name.
+ Init *RMName = (*it)->getValueInit("RenderMethod");
+ if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
+ CI->RenderMethod = SI->getValue();
+ } else {
+ assert(dynamic_cast<UnsetInit*>(RMName) &&
+ "Unexpected RenderMethod field!");
+ CI->RenderMethod = "add" + CI->ClassName + "Operands";
+ }
+
+ AsmOperandClasses[*it] = CI;
+ Classes.push_back(CI);
}
-
- return Entry;
+}
+
+AsmMatcherInfo::AsmMatcherInfo(Record *_AsmParser)
+ : AsmParser(_AsmParser),
+ CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")),
+ RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix"))
+{
}
void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
- 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.
+ if (!CommentDelimiter.empty()) {
+ size_t Idx = StringRef(II->AsmString).find(CommentDelimiter);
+ if (Idx != StringRef::npos)
+ II->AsmString = II->AsmString.substr(0, Idx);
+ }
+
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;
+ // Reorder classes so that classes preceed super classes.
+ std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
+}
- Instructions.push_back(II.take());
- }
+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,
// 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 += Op.Class->ClassName;
+ Signature += "__";
+
+ // Registers are always converted the same, don't duplicate the conversion
+ // function based on them.
+ //
+ // FIXME: We could generalize this based on the render method, if it
+ // mattered.
+ if (Op.Class->isRegisterClass())
+ Signature += "Reg";
+ else
+ Signature += Op.Class->ClassName;
Signature += utostr(Op.OperandInfo->MINumOperands);
Signature += "_" + utostr(MIOperandList[i].second);
}
// 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.
OS << " " << CI.Name << ", // ";
if (CI.Kind == ClassInfo::Token) {
OS << "'" << CI.ValueName << "'\n";
- } else if (CI.Kind == ClassInfo::Register) {
+ } else if (CI.isRegisterClass()) {
if (!CI.ValueName.empty())
OS << "register class '" << CI.ValueName << "'\n";
else
/// EmitClassifyOperand - Emit the function to classify an operand.
static void EmitClassifyOperand(CodeGenTarget &Target,
- std::vector<ClassInfo*> &Infos,
+ 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";
OS << " return MatchTokenString(Operand.getToken());\n\n";
- for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
- ie = Infos.end(); it != ie; ++it) {
+
+ // Classify registers.
+ //
+ // FIXME: Don't hardcode isReg, getReg.
+ OS << " if (Operand.isReg()) {\n";
+ OS << " switch (Operand.getReg()) {\n";
+ OS << " default: return InvalidMatchClass;\n";
+ for (std::map<Record*, ClassInfo*>::iterator
+ it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
+ it != ie; ++it)
+ OS << " case " << Target.getName() << "::"
+ << it->first->getName() << ": return " << it->second->Name << ";\n";
+ OS << " }\n";
+ OS << " }\n\n";
+
+ // Classify user defined operands.
+ for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
+ ie = Info.Classes.end(); it != ie; ++it) {
ClassInfo &CI = **it;
- if (CI.Kind != ClassInfo::Token) {
- OS << " if (Operand." << CI.PredicateMethod << "())\n";
- OS << " return " << CI.Name << ";\n\n";
+ if (!CI.isUserClass())
+ continue;
+
+ OS << " // '" << CI.ClassName << "' class";
+ if (!CI.SuperClasses.empty()) {
+ OS << ", subclass of ";
+ for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) {
+ if (i) OS << ", ";
+ OS << "'" << CI.SuperClasses[i]->ClassName << "'";
+ assert(CI < *CI.SuperClasses[i] && "Invalid class relation!");
+ }
}
+ OS << "\n";
+
+ OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
+
+ // Validate subclass relationships.
+ if (!CI.SuperClasses.empty()) {
+ for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i)
+ OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod
+ << "() && \"Invalid class relationship!\");\n";
+ }
+
+ OS << " return " << CI.Name << ";\n";
+ OS << " }\n\n";
}
OS << " return InvalidMatchClass;\n";
OS << "}\n\n";
}
+/// EmitIsSubclass - Emit the subclass predicate function.
+static void EmitIsSubclass(CodeGenTarget &Target,
+ std::vector<ClassInfo*> &Infos,
+ raw_ostream &OS) {
+ OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
+ OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n";
+ OS << " if (A == B)\n";
+ OS << " return true;\n\n";
+
+ OS << " switch (A) {\n";
+ OS << " default:\n";
+ OS << " return false;\n";
+ for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
+ ie = Infos.end(); it != ie; ++it) {
+ ClassInfo &A = **it;
+
+ if (A.Kind != ClassInfo::Token) {
+ std::vector<StringRef> SuperClasses;
+ for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
+ ie = Infos.end(); it != ie; ++it) {
+ ClassInfo &B = **it;
+
+ if (&A != &B && A.isSubsetOf(B))
+ SuperClasses.push_back(B.Name);
+ }
+
+ if (SuperClasses.empty())
+ continue;
+
+ OS << "\n case " << A.Name << ":\n";
+
+ if (SuperClasses.size() == 1) {
+ OS << " return B == " << SuperClasses.back() << ";\n";
+ continue;
+ }
+
+ OS << " switch (B) {\n";
+ OS << " default: return false;\n";
+ for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
+ OS << " case " << SuperClasses[i] << ": return true;\n";
+ OS << " }\n";
+ }
+ }
+ OS << " }\n";
+ OS << "}\n\n";
+}
+
typedef std::pair<std::string, std::string> StringPair;
/// FindFirstNonCommonLetter - Find the first character in the keys of the
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;
+ 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
// Check for ambiguous instructions.
unsigned NumAmbiguous = 0;
- for (std::vector<InstructionInfo*>::const_iterator it =
- Info.Instructions.begin(), ie = Info.Instructions.end() - 1;
- it != ie;) {
- InstructionInfo &II = **it;
- ++it;
-
- InstructionInfo &Next = **it;
+ for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
+ for (unsigned j = i + 1; j != e; ++j) {
+ InstructionInfo &A = *Info.Instructions[i];
+ InstructionInfo &B = *Info.Instructions[j];
- if (!(II < Next)){
- DEBUG_WITH_TYPE("ambiguous_instrs", {
- errs() << "warning: ambiguous instruction match:\n";
- II.dump();
- errs() << "\nis incomparable with:\n";
- Next.dump();
- errs() << "\n\n";
- });
- ++NumAmbiguous;
+ if (A.CouldMatchAmiguouslyWith(B)) {
+ DEBUG_WITH_TYPE("ambiguous_instrs", {
+ errs() << "warning: ambiguous instruction match:\n";
+ A.dump();
+ errs() << "\nis incomparable with:\n";
+ B.dump();
+ errs() << "\n\n";
+ });
+ ++NumAmbiguous;
+ }
}
}
if (NumAmbiguous)
<< " 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);
EmitMatchTokenString(Target, Info.Classes, OS);
// Emit the routine to classify an operand.
- EmitClassifyOperand(Target, Info.Classes, OS);
+ EmitClassifyOperand(Target, Info, OS);
+
+ // Emit the subclass predicate routine.
+ EmitIsSubclass(Target, Info.Classes, OS);
// Finally, build the match function.
Info.Instructions.begin(), ie = Info.Instructions.end();
it != ie; ++it)
MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
-
- OS << "bool " << Target.getName() << ClassName
- << "::MatchInstruction("
- << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
- << "MCInst &Inst) {\n";
+
+ const std::string &MatchName =
+ AsmParser->getValueAsString("MatchInstructionName");
+ OS << "bool " << Target.getName() << ClassName << "::\n"
+ << MatchName
+ << "(const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
+ OS.indent(MatchName.size() + 1);
+ OS << "MCInst &Inst) {\n";
// Emit the static match table; unused classes get initalized to 0 which is
// guaranteed to be InvalidMatchClass.
<< "*ie = MatchTable + " << Info.Instructions.size()
<< "; it != ie; ++it) {\n";
for (unsigned i = 0; i != MaxNumOperands; ++i) {
- OS << " if (Classes[" << i << "] != it->Classes[" << i << "])\n";
+ OS << " if (!IsSubclass(Classes["
+ << i << "], it->Classes[" << i << "]))\n";
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";
}
projects / oota-llvm.git / blobdiff