#include "Record.h"
#include "llvm/ADT/OwningPtr.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
-#include <set>
#include <list>
+#include <map>
+#include <set>
using namespace llvm;
-namespace {
static cl::opt<std::string>
-MatchOneInstr("match-one-instr", cl::desc("Match only the named instruction"),
- cl::init(""));
-}
+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);
}
}
namespace {
+/// ClassInfo - Helper class for storing the information about a particular
+/// class of operands which can be matched.
+struct ClassInfo {
+ enum ClassInfoKind {
+ /// 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;
+
+ /// ClassName - The unadorned generic name for this class (e.g., Token).
+ std::string ClassName;
+
+ /// ValueName - The name of the value this class represents; for a token this
+ /// is the literal token string, for an operand it is the TableGen class (or
+ /// empty if this is a derived class).
+ std::string ValueName;
+
+ /// PredicateMethod - The name of the operand method to test whether the
+ /// 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 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 {
+ // 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 comparable by value.
+ //
+ // FIXME: Compare by enum value.
+ return ValueName < RHS.ValueName;
+
+ default:
+ // This class preceeds the RHS if it is a proper subset of the RHS.
+ return this != &RHS && isSubsetOf(RHS);
+ }
+ }
+};
+
+/// InstructionInfo - Helper class for storing the necessary information for an
+/// instruction which is capable of being matched.
struct InstructionInfo {
struct Operand {
- enum {
- Token,
- Class
- } Kind;
-
- struct ClassData {
- /// Operand - The tablegen operand this class corresponds to.
- const CodeGenInstruction::OperandInfo *Operand;
-
- /// ClassName - The name of this operand's class.
- std::string ClassName;
-
- /// PredicateMethod - The name of the operand method to test whether the
- /// operand matches this class.
- std::string PredicateMethod;
-
- /// RenderMethod - The name of the operand method to add this operand to
- /// an MCInst.
- std::string RenderMethod;
- } AsClass;
+ /// The unique class instance this operand should match.
+ ClassInfo *Class;
+
+ /// The original operand this corresponds to, if any.
+ const CodeGenInstruction::OperandInfo *OperandInfo;
};
/// InstrName - The target name for this instruction.
/// Operands - The operands that this instruction matches.
SmallVector<Operand, 4> Operands;
- /// ConversionFn - The name of the conversion function to convert parsed
- /// operands into an MCInst for this function.
- std::string ConversionFn;
+ /// ConversionFnKind - The enum value which is passed to the generated
+ /// ConvertToMCInst to convert parsed operands into an MCInst for this
+ /// function.
+ std::string ConversionFnKind;
+
+ /// operator< - Compare two instructions.
+ bool operator<(const InstructionInfo &RHS) const {
+ if (Operands.size() != RHS.Operands.size())
+ return Operands.size() < RHS.Operands.size();
+
+ // 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;
+ }
+
+ /// 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;
- /// OrderedClassOperands - The indices of the class operands, ordered by their
- /// MIOperandNo order (which is the order they should be passed to the
- /// conversion function).
- SmallVector<unsigned, 4> OrderedClassOperands;
+ // 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)
+ HasLT = true;
+ if (*RHS.Operands[i].Class < *Operands[i].Class)
+ HasGT = true;
+ }
+
+ return !(HasLT ^ HasGT);
+ }
public:
void dump();
};
+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 RegisterClass records to their class information.
+ std::map<Record*, ClassInfo*> RegisterClassClasses;
+
+ /// 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(StringRef Token);
+
+ /// getOperandClass - Lookup or create the class for the given operand.
+ 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);
+};
+
}
void InstructionInfo::dump() {
for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
Operand &Op = Operands[i];
- errs() << " op[" << i << "] = ";
- if (Op.Kind == Operand::Token) {
+ errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
+ if (Op.Class->Kind == ClassInfo::Token) {
errs() << '\"' << Tokens[i] << "\"\n";
continue;
}
- assert(Op.Kind == Operand::Class && "Invalid kind!");
- const CodeGenInstruction::OperandInfo &OI = *Op.AsClass.Operand;
+ 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 void BuildInstructionInfos(CodeGenTarget &Target,
- std::vector<InstructionInfo*> &Infos) {
- const std::map<std::string, CodeGenInstruction> &Instructions =
- Target.getInstructions();
+static std::string getEnumNameForToken(StringRef Str) {
+ std::string Res;
+
+ for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
+ switch (*it) {
+ case '*': Res += "_STAR_"; break;
+ case '%': Res += "_PCT_"; break;
+ case ':': Res += "_COLON_"; break;
+
+ default:
+ if (isalnum(*it)) {
+ Res += *it;
+ } else {
+ Res += "_" + utostr((unsigned) *it) + "_";
+ }
+ }
+ }
+
+ return Res;
+}
+
+/// 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) {
+ Entry = new ClassInfo();
+ Entry->Kind = ClassInfo::Token;
+ Entry->ClassName = "Token";
+ Entry->Name = "MCK_" + getEnumNameForToken(Token);
+ Entry->ValueName = Token;
+ Entry->PredicateMethod = "<invalid>";
+ Entry->RenderMethod = "<invalid>";
+ Classes.push_back(Entry);
+ }
+
+ return Entry;
+}
+
+ClassInfo *
+AsmMatcherInfo::getOperandClass(StringRef Token,
+ const CodeGenInstruction::OperandInfo &OI) {
+ if (OI.Rec->isSubClassOf("RegisterClass")) {
+ 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;
+ }
+
+ 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);
+ }
+ }
- for (std::map<std::string, CodeGenInstruction>::const_iterator
- it = Instructions.begin(), ie = Instructions.end(); it != ie; ++it) {
- const CodeGenInstruction &CGI = it->second;
+ if (!ContainingSet.empty()) {
+ RegisterSets.insert(ContainingSet);
+ RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
+ }
+ }
- if (!MatchOneInstr.empty() && it->first != MatchOneInstr)
+ // 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;
+
+ Init *Super = (*it)->getValueInit("SuperClass");
+ if (DefInit *DI = dynamic_cast<DefInit*>(Super)) {
+ ClassInfo *SC = AsmOperandClasses[DI->getDef()];
+ if (!SC)
+ PrintError((*it)->getLoc(), "Invalid super class reference!");
+ else
+ CI->SuperClasses.push_back(SC);
+ } else {
+ assert(dynamic_cast<UnsetInit*>(Super) && "Unexpected SuperClass field!");
+ }
+ 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 {
+ assert(dynamic_cast<UnsetInit*>(PMName) &&
+ "Unexpected PredicateMethod field!");
+ CI->PredicateMethod = "is" + CI->ClassName;
+ }
+
+ // 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);
+ }
+}
+
+AsmMatcherInfo::AsmMatcherInfo(Record *_AsmParser)
+ : AsmParser(_AsmParser),
+ CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")),
+ RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix"))
+{
+}
+
+void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
+ // 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(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;
- Op.Kind = InstructionInfo::Operand::Token;
+ Op.Class = getTokenClass(Token);
+ Op.OperandInfo = 0;
II->Operands.push_back(Op);
continue;
}
// Otherwise this is an operand reference.
- InstructionInfo::Operand Op;
- Op.Kind = InstructionInfo::Operand::Class;
-
StringRef OperandName;
if (Token[1] == '{')
OperandName = Token.substr(2, Token.size() - 3);
// 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() + "'");
}
- const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx];
- Op.AsClass.Operand = &OI;
-
- if (OI.Rec->isSubClassOf("RegisterClass")) {
- Op.AsClass.ClassName = "Reg";
- Op.AsClass.PredicateMethod = "isReg";
- Op.AsClass.RenderMethod = "addRegOperands";
- } else if (OI.Rec->isSubClassOf("Operand")) {
- // FIXME: This should not be hard coded.
- const RecordVal *RV = OI.Rec->getValue("Type");
-
- // FIXME: Yet another total hack.
- if (RV->getValue()->getAsString() == "iPTR" ||
- OI.Rec->getName() == "lea32mem" ||
- OI.Rec->getName() == "lea64_32mem") {
- Op.AsClass.ClassName = "Mem";
- Op.AsClass.PredicateMethod = "isMem";
- Op.AsClass.RenderMethod = "addMemOperands";
- } else {
- Op.AsClass.ClassName = "Imm";
- Op.AsClass.PredicateMethod = "isImm";
- Op.AsClass.RenderMethod = "addImmOperands";
+ // 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;
+ }
}
- } else {
- OI.Rec->dump();
- assert(0 && "Unexpected instruction operand record!");
+
+ assert(OI && "Unable to find tied operand target!");
}
+ InstructionInfo::Operand Op;
+ 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>());
+}
- Infos.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 ConstructConversionFunctions(CodeGenTarget &Target,
- std::vector<InstructionInfo*> &Infos,
- raw_ostream &OS) {
+static void EmitConvertToMCInst(CodeGenTarget &Target,
+ std::vector<InstructionInfo*> &Infos,
+ raw_ostream &OS) {
+ // Write the convert function to a separate stream, so we can drop it after
+ // the enum.
+ std::string ConvertFnBody;
+ raw_string_ostream CvtOS(ConvertFnBody);
+
// Function we have already generated.
std::set<std::string> GeneratedFns;
+ // Start the unified conversion function.
+
+ CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
+ << "unsigned Opcode,\n"
+ << " const SmallVectorImpl<MCParsedAsmOperand*"
+ << "> &Operands) {\n";
+ CvtOS << " Inst.setOpcode(Opcode);\n";
+ CvtOS << " switch (Kind) {\n";
+ CvtOS << " default:\n";
+
+ // Start the enum, which we will generate inline.
+
+ 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;
SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
InstructionInfo::Operand &Op = II.Operands[i];
- if (Op.Kind == InstructionInfo::Operand::Class)
- MIOperandList.push_back(std::make_pair(Op.AsClass.Operand->MIOperandNo,
- i));
+ 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.
unsigned CurIndex = 0;
for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
- assert(CurIndex <= Op.AsClass.Operand->MIOperandNo &&
+ assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
"Duplicate match for instruction operand!");
-
- // Save the conversion index, for use by the matcher.
- II.OrderedClassOperands.push_back(MIOperandList[i].second);
// 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.AsClass.Operand->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.
+ //
+ // 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);
- Signature += Op.AsClass.ClassName;
- Signature += utostr(Op.AsClass.Operand->MINumOperands);
- CurIndex += Op.AsClass.Operand->MINumOperands;
+ CurIndex += Op.OperandInfo->MINumOperands;
}
// 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);
+ }
- // Save the conversion function, for use by the matcher.
- II.ConversionFn = Signature;
+ II.ConversionFnKind = Signature;
- // Check if we have already generated this function.
+ // Check if we have already generated this signature.
if (!GeneratedFns.insert(Signature).second)
continue;
// If not, emit it now.
- //
- // FIXME: There should be no need to pass the number of operands to fill;
- // this should always be implicit in the class.
- OS << "static bool " << Signature << "(MCInst &Inst, unsigned Opcode";
- for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i)
- OS << ", " << Target.getName() << "Operand Op" << i;
- OS << ") {\n";
- OS << " Inst.setOpcode(Opcode);\n";
+
+ // Add to the enum list.
+ OS << " " << Signature << ",\n";
+
+ // And to the convert function.
+ CvtOS << " case " << Signature << ":\n";
CurIndex = 0;
for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
// Add the implicit operands.
- for (; CurIndex != Op.AsClass.Operand->MIOperandNo; ++CurIndex)
- OS << " 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";
+ }
+ }
- OS << " Op" << i << "." << Op.AsClass.RenderMethod
- << "(Inst, " << Op.AsClass.Operand->MINumOperands << ");\n";
- CurIndex += Op.AsClass.Operand->MINumOperands;
+ 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)
- OS << " Inst.addOperand(MCOperand::CreateReg(0));\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 << "}\n\n";
+
+ // Finish the enum, and drop the convert function after it.
+
+ OS << " NumConversionVariants\n";
+ OS << "};\n\n";
+
+ OS << CvtOS.str();
+}
+
+/// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
+static void EmitMatchClassEnumeration(CodeGenTarget &Target,
+ std::vector<ClassInfo*> &Infos,
+ raw_ostream &OS) {
+ OS << "namespace {\n\n";
+
+ OS << "/// MatchClassKind - The kinds of classes which participate in\n"
+ << "/// instruction matching.\n";
+ OS << "enum MatchClassKind {\n";
+ OS << " InvalidMatchClass = 0,\n";
+ for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
+ ie = Infos.end(); it != ie; ++it) {
+ ClassInfo &CI = **it;
+ OS << " " << CI.Name << ", // ";
+ if (CI.Kind == ClassInfo::Token) {
+ OS << "'" << CI.ValueName << "'\n";
+ } else if (CI.isRegisterClass()) {
+ if (!CI.ValueName.empty())
+ OS << "register class '" << CI.ValueName << "'\n";
+ else
+ OS << "derived register class\n";
+ } else {
+ OS << "user defined class '" << CI.ValueName << "'\n";
+ }
+ }
+ OS << " NumMatchClassKinds\n";
+ OS << "};\n\n";
+
+ OS << "}\n\n";
+}
+
+/// EmitClassifyOperand - Emit the function to classify an operand.
+static void EmitClassifyOperand(CodeGenTarget &Target,
+ AsmMatcherInfo &Info,
+ raw_ostream &OS) {
+ 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";
+
+ // 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.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 << " return false;\n";
- OS << "}\n\n";
+ 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
+/// string pairs that is not shared across the whole set of strings. All
+/// strings are assumed to have the same length.
+static unsigned
+FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) {
+ assert(!Matches.empty());
+ for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) {
+ // Check to see if letter i is the same across the set.
+ char Letter = Matches[0]->first[i];
+
+ for (unsigned str = 0, e = Matches.size(); str != e; ++str)
+ if (Matches[str]->first[i] != Letter)
+ return i;
+ }
+
+ return Matches[0]->first.size();
+}
+
+/// EmitStringMatcherForChar - Given a set of strings that are known to be the
+/// same length and whose characters leading up to CharNo are the same, emit
+/// code to verify that CharNo and later are the same.
+///
+/// \return - True if control can leave the emitted code fragment.
+static bool EmitStringMatcherForChar(const std::string &StrVariableName,
+ const std::vector<const StringPair*> &Matches,
+ unsigned CharNo, unsigned IndentCount,
+ raw_ostream &OS) {
+ assert(!Matches.empty() && "Must have at least one string to match!");
+ std::string Indent(IndentCount*2+4, ' ');
+
+ // If we have verified that the entire string matches, we're done: output the
+ // matching code.
+ if (CharNo == Matches[0]->first.size()) {
+ assert(Matches.size() == 1 && "Had duplicate keys to match on");
+
+ // FIXME: If Matches[0].first has embeded \n, this will be bad.
+ OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first
+ << "\"\n";
+ return false;
+ }
+
+ // Bucket the matches by the character we are comparing.
+ std::map<char, std::vector<const StringPair*> > MatchesByLetter;
+
+ for (unsigned i = 0, e = Matches.size(); i != e; ++i)
+ MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]);
+
+
+ // If we have exactly one bucket to match, see how many characters are common
+ // across the whole set and match all of them at once.
+ if (MatchesByLetter.size() == 1) {
+ unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches);
+ unsigned NumChars = FirstNonCommonLetter-CharNo;
+
+ // Emit code to break out if the prefix doesn't match.
+ if (NumChars == 1) {
+ // Do the comparison with if (Str[1] != 'f')
+ // FIXME: Need to escape general characters.
+ OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '"
+ << Matches[0]->first[CharNo] << "')\n";
+ OS << Indent << " break;\n";
+ } else {
+ // Do the comparison with if (Str.substr(1,3) != "foo").
+ // FIXME: Need to escape general strings.
+ OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << ","
+ << NumChars << ") != \"";
+ OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n";
+ OS << Indent << " break;\n";
+ }
+
+ return EmitStringMatcherForChar(StrVariableName, Matches,
+ FirstNonCommonLetter, IndentCount, OS);
+ }
+
+ // Otherwise, we have multiple possible things, emit a switch on the
+ // character.
+ OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n";
+ OS << Indent << "default: break;\n";
+
+ for (std::map<char, std::vector<const StringPair*> >::iterator LI =
+ MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) {
+ // TODO: escape hard stuff (like \n) if we ever care about it.
+ OS << Indent << "case '" << LI->first << "':\t // "
+ << LI->second.size() << " strings to match.\n";
+ if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1,
+ IndentCount+1, OS))
+ OS << Indent << " break;\n";
+ }
+
+ OS << Indent << "}\n";
+ return true;
+}
+
+
+/// EmitStringMatcher - Given a list of strings and code to execute when they
+/// match, output a simple switch tree to classify the input string.
+///
+/// If a match is found, the code in Vals[i].second is executed; control must
+/// not exit this code fragment. If nothing matches, execution falls through.
+///
+/// \param StrVariableName - The name of the variable to test.
+static void EmitStringMatcher(const std::string &StrVariableName,
+ const std::vector<StringPair> &Matches,
+ raw_ostream &OS) {
+ // First level categorization: group strings by length.
+ std::map<unsigned, std::vector<const StringPair*> > MatchesByLength;
+
+ for (unsigned i = 0, e = Matches.size(); i != e; ++i)
+ MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]);
+
+ // Output a switch statement on length and categorize the elements within each
+ // bin.
+ OS << " switch (" << StrVariableName << ".size()) {\n";
+ OS << " default: break;\n";
+
+ for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI =
+ MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) {
+ OS << " case " << LI->first << ":\t // " << LI->second.size()
+ << " strings to match.\n";
+ if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS))
+ OS << " break;\n";
}
+
+ OS << " }\n";
+}
+
+
+/// EmitMatchTokenString - Emit the function to match a token string to the
+/// appropriate match class value.
+static void EmitMatchTokenString(CodeGenTarget &Target,
+ std::vector<ClassInfo*> &Infos,
+ raw_ostream &OS) {
+ // Construct the match list.
+ std::vector<StringPair> Matches;
+ for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
+ ie = Infos.end(); it != ie; ++it) {
+ ClassInfo &CI = **it;
+
+ if (CI.Kind == ClassInfo::Token)
+ Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";"));
+ }
+
+ OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
+
+ EmitStringMatcher("Name", Matches, OS);
+
+ OS << " return InvalidMatchClass;\n";
+ OS << "}\n\n";
}
/// EmitMatchRegisterName - Emit the function to match a string to the target
/// specific register enum.
static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
raw_ostream &OS) {
- const std::vector<CodeGenRegister> &Registers = Target.getRegisters();
-
- OS << "bool " << Target.getName()
- << AsmParser->getValueAsString("AsmParserClassName")
- << "::MatchRegisterName(const StringRef &Name, unsigned &RegNo) {\n";
-
- // FIXME: TableGen should have a fast string matcher generator.
- for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
- const CodeGenRegister &Reg = Registers[i];
+ // Construct the match list.
+ std::vector<StringPair> Matches;
+ for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
+ const CodeGenRegister &Reg = Target.getRegisters()[i];
if (Reg.TheDef->getValueAsString("AsmName").empty())
continue;
- OS << " if (Name == \""
- << Reg.TheDef->getValueAsString("AsmName") << "\")\n"
- << " return RegNo=" << i + 1 << ", false;\n";
+ Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"),
+ "return " + utostr(i + 1) + ";"));
}
- OS << " return true;\n";
+
+ OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
+
+ EmitStringMatcher("Name", Matches, OS);
+
+ OS << " return 0;\n";
OS << "}\n\n";
}
Record *AsmParser = Target.getAsmParser();
std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
+ // Compute the information on the instructions to match.
+ AsmMatcherInfo Info(AsmParser);
+ Info.BuildInfo(Target);
+
+ // 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
+ it = Info.Instructions.begin(), ie = Info.Instructions.end();
+ it != ie; ++it)
+ (*it)->dump();
+ });
+
+ // Check for ambiguous instructions.
+ unsigned NumAmbiguous = 0;
+ 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 (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)
+ DEBUG_WITH_TYPE("ambiguous_instrs", {
+ errs() << "warning: " << 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";
- // Compute the information on the list of instructions to match.
- std::vector<InstructionInfo*> Infos;
- BuildInstructionInfos(Target, Infos);
+ // Generate the unified function to convert operands into an MCInst.
+ EmitConvertToMCInst(Target, Info.Instructions, OS);
-#undef DEBUG_TYPE
-#define DEBUG_TYPE "instruction_info"
- DEBUG({
- for (std::vector<InstructionInfo*>::iterator it = Infos.begin(),
- ie = Infos.end(); it != ie; ++it)
- (*it)->dump();
- });
-#undef DEBUG_TYPE
-#define DEBUG_TYPE ""
+ // Emit the enumeration for classes which participate in matching.
+ EmitMatchClassEnumeration(Target, Info.Classes, OS);
- // FIXME: At this point we should be able to totally order Infos, if not then
- // we have an ambiguity which the .td file should be forced to resolve.
+ // Emit the routine to match token strings to their match class.
+ EmitMatchTokenString(Target, Info.Classes, OS);
- // Generate the terminal actions to convert operands into an MCInst. We still
- // pass the operands in to these functions individually (as opposed to the
- // array) so that we do not need to worry about the operand order.
- ConstructConversionFunctions(Target, Infos, OS);
+ // Emit the routine to classify an operand.
+ EmitClassifyOperand(Target, Info, OS);
- // Build a very stupid version of the match function which just checks each
- // instruction in order.
+ // Emit the subclass predicate routine.
+ EmitIsSubclass(Target, Info.Classes, OS);
- OS << "bool " << Target.getName() << ClassName
- << "::MatchInstruction("
- << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
- << "MCInst &Inst) {\n";
+ // Finally, build the match function.
- for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
- ie = Infos.end(); it != ie; ++it) {
+ size_t MaxNumOperands = 0;
+ for (std::vector<InstructionInfo*>::const_iterator it =
+ Info.Instructions.begin(), ie = Info.Instructions.end();
+ it != ie; ++it)
+ MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
+
+ 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.
+ //
+ // FIXME: We can reduce the size of this table very easily. First, we change
+ // it so that store the kinds in separate bit-fields for each index, which
+ // only needs to be the max width used for classes at that index (we also need
+ // to reject based on this during classification). If we then make sure to
+ // order the match kinds appropriately (putting mnemonics last), then we
+ // should only end up using a few bits for each class, especially the ones
+ // following the mnemonic.
+ OS << " static const struct MatchEntry {\n";
+ OS << " unsigned Opcode;\n";
+ OS << " ConversionKind ConvertFn;\n";
+ OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
+ OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n";
+
+ for (std::vector<InstructionInfo*>::const_iterator it =
+ Info.Instructions.begin(), ie = Info.Instructions.end();
+ it != ie; ++it) {
InstructionInfo &II = **it;
- // The parser is expected to arrange things so that each "token" matches
- // exactly one target specific operand.
- OS << " if (Operands.size() == " << II.Operands.size();
+ OS << " { " << Target.getName() << "::" << II.InstrName
+ << ", " << II.ConversionFnKind << ", { ";
for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
InstructionInfo::Operand &Op = II.Operands[i];
- OS << " &&\n";
- OS << " ";
-
- if (Op.Kind == InstructionInfo::Operand::Token)
- OS << "Operands[" << i << "].isToken(\"" << II.Tokens[i] << "\")";
- else
- OS << "Operands[" << i << "]."
- << Op.AsClass.PredicateMethod << "()";
+ if (i) OS << ", ";
+ OS << Op.Class->Name;
}
- OS << ")\n";
- OS << " return " << II.ConversionFn << "(Inst, "
- << Target.getName() << "::" << II.InstrName;
- for (unsigned i = 0, e = II.OrderedClassOperands.size(); i != e; ++i)
- OS << ", Operands[" << II.OrderedClassOperands[i] << "]";
- OS << ");\n\n";
+ OS << " } },\n";
}
+ OS << " };\n\n";
+
+ // Emit code to compute the class list for this operand vector.
+ OS << " // Eliminate obvious mismatches.\n";
+ OS << " if (Operands.size() > " << MaxNumOperands << ")\n";
+ OS << " return true;\n\n";
+
+ OS << " // Compute the class list for this operand vector.\n";
+ OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
+ OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n";
+ OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n";
+
+ OS << " // Check for invalid operands before matching.\n";
+ OS << " if (Classes[i] == InvalidMatchClass)\n";
+ OS << " return true;\n";
+ OS << " }\n\n";
+
+ OS << " // Mark unused classes.\n";
+ OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; "
+ << "i != e; ++i)\n";
+ OS << " Classes[i] = InvalidMatchClass;\n\n";
+
+ // Emit code to search the table.
+ OS << " // Search the table.\n";
+ OS << " for (const MatchEntry *it = MatchTable, "
+ << "*ie = MatchTable + " << Info.Instructions.size()
+ << "; it != ie; ++it) {\n";
+ for (unsigned i = 0; i != MaxNumOperands; ++i) {
+ OS << " if (!IsSubclass(Classes["
+ << i << "], it->Classes[" << i << "]))\n";
+ OS << " continue;\n";
+ }
+ OS << "\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