1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tablegen backend emits a target specifier matcher for converting parsed
11 // assembly operands in the MCInst structures.
13 // The input to the target specific matcher is a list of literal tokens and
14 // operands. The target specific parser should generally eliminate any syntax
15 // which is not relevant for matching; for example, comma tokens should have
16 // already been consumed and eliminated by the parser. Most instructions will
17 // end up with a single literal token (the instruction name) and some number of
20 // Some example inputs, for X86:
21 // 'addl' (immediate ...) (register ...)
22 // 'add' (immediate ...) (memory ...)
25 // The assembly matcher is responsible for converting this input into a precise
26 // machine instruction (i.e., an instruction with a well defined encoding). This
27 // mapping has several properties which complicate matching:
29 // - It may be ambiguous; many architectures can legally encode particular
30 // variants of an instruction in different ways (for example, using a smaller
31 // encoding for small immediates). Such ambiguities should never be
32 // arbitrarily resolved by the assembler, the assembler is always responsible
33 // for choosing the "best" available instruction.
35 // - It may depend on the subtarget or the assembler context. Instructions
36 // which are invalid for the current mode, but otherwise unambiguous (e.g.,
37 // an SSE instruction in a file being assembled for i486) should be accepted
38 // and rejected by the assembler front end. However, if the proper encoding
39 // for an instruction is dependent on the assembler context then the matcher
40 // is responsible for selecting the correct machine instruction for the
43 // The core matching algorithm attempts to exploit the regularity in most
44 // instruction sets to quickly determine the set of possibly matching
45 // instructions, and the simplify the generated code. Additionally, this helps
46 // to ensure that the ambiguities are intentionally resolved by the user.
48 // The matching is divided into two distinct phases:
50 // 1. Classification: Each operand is mapped to the unique set which (a)
51 // contains it, and (b) is the largest such subset for which a single
52 // instruction could match all members.
54 // For register classes, we can generate these subgroups automatically. For
55 // arbitrary operands, we expect the user to define the classes and their
56 // relations to one another (for example, 8-bit signed immediates as a
57 // subset of 32-bit immediates).
59 // By partitioning the operands in this way, we guarantee that for any
60 // tuple of classes, any single instruction must match either all or none
61 // of the sets of operands which could classify to that tuple.
63 // In addition, the subset relation amongst classes induces a partial order
64 // on such tuples, which we use to resolve ambiguities.
66 // FIXME: What do we do if a crazy case shows up where this is the wrong
69 // 2. The input can now be treated as a tuple of classes (static tokens are
70 // simple singleton sets). Each such tuple should generally map to a single
71 // instruction (we currently ignore cases where this isn't true, whee!!!),
72 // which we can emit a simple matcher for.
74 //===----------------------------------------------------------------------===//
76 #include "AsmMatcherEmitter.h"
77 #include "CodeGenTarget.h"
79 #include "llvm/ADT/OwningPtr.h"
80 #include "llvm/ADT/SmallVector.h"
81 #include "llvm/ADT/StringExtras.h"
82 #include "llvm/Support/CommandLine.h"
83 #include "llvm/Support/Debug.h"
90 static cl::opt<std::string>
91 MatchOneInstr("match-one-instr", cl::desc("Match only the named instruction"),
95 /// FlattenVariants - Flatten an .td file assembly string by selecting the
96 /// variant at index \arg N.
97 static std::string FlattenVariants(const std::string &AsmString,
99 StringRef Cur = AsmString;
100 std::string Res = "";
103 // Find the start of the next variant string.
104 size_t VariantsStart = 0;
105 for (size_t e = Cur.size(); VariantsStart != e; ++VariantsStart)
106 if (Cur[VariantsStart] == '{' &&
107 (VariantsStart == 0 || (Cur[VariantsStart-1] != '$' &&
108 Cur[VariantsStart-1] != '\\')))
111 // Add the prefix to the result.
112 Res += Cur.slice(0, VariantsStart);
113 if (VariantsStart == Cur.size())
116 ++VariantsStart; // Skip the '{'.
118 // Scan to the end of the variants string.
119 size_t VariantsEnd = VariantsStart;
120 unsigned NestedBraces = 1;
121 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) {
122 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') {
123 if (--NestedBraces == 0)
125 } else if (Cur[VariantsEnd] == '{')
129 // Select the Nth variant (or empty).
130 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd);
131 for (unsigned i = 0; i != N; ++i)
132 Selection = Selection.split('|').second;
133 Res += Selection.split('|').first;
135 assert(VariantsEnd != Cur.size() &&
136 "Unterminated variants in assembly string!");
137 Cur = Cur.substr(VariantsEnd + 1);
143 /// TokenizeAsmString - Tokenize a simplified assembly string.
144 static void TokenizeAsmString(const StringRef &AsmString,
145 SmallVectorImpl<StringRef> &Tokens) {
148 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) {
149 switch (AsmString[i]) {
158 Tokens.push_back(AsmString.slice(Prev, i));
161 if (!isspace(AsmString[i]) && AsmString[i] != ',')
162 Tokens.push_back(AsmString.substr(i, 1));
168 Tokens.push_back(AsmString.slice(Prev, i));
172 assert(i != AsmString.size() && "Invalid quoted character");
173 Tokens.push_back(AsmString.substr(i, 1));
178 // If this isn't "${", treat like a normal token.
179 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') {
181 Tokens.push_back(AsmString.slice(Prev, i));
189 Tokens.push_back(AsmString.slice(Prev, i));
193 StringRef::iterator End =
194 std::find(AsmString.begin() + i, AsmString.end(), '}');
195 assert(End != AsmString.end() && "Missing brace in operand reference!");
196 size_t EndPos = End - AsmString.begin();
197 Tokens.push_back(AsmString.slice(i, EndPos+1));
207 if (InTok && Prev != AsmString.size())
208 Tokens.push_back(AsmString.substr(Prev));
211 static bool IsAssemblerInstruction(const StringRef &Name,
212 const CodeGenInstruction &CGI,
213 const SmallVectorImpl<StringRef> &Tokens) {
214 // Ignore psuedo ops.
216 // FIXME: This is a hack.
217 if (const RecordVal *Form = CGI.TheDef->getValue("Form"))
218 if (Form->getValue()->getAsString() == "Pseudo")
221 // Ignore "PHI" node.
223 // FIXME: This is also a hack.
227 // Ignore instructions with no .s string.
229 // FIXME: What are these?
230 if (CGI.AsmString.empty())
233 // FIXME: Hack; ignore any instructions with a newline in them.
234 if (std::find(CGI.AsmString.begin(),
235 CGI.AsmString.end(), '\n') != CGI.AsmString.end())
238 // Ignore instructions with attributes, these are always fake instructions for
239 // simplifying codegen.
241 // FIXME: Is this true?
243 // Also, we ignore instructions which reference the operand multiple times;
244 // this implies a constraint we would not currently honor. These are
245 // currently always fake instructions for simplifying codegen.
247 // FIXME: Encode this assumption in the .td, so we can error out here.
248 std::set<std::string> OperandNames;
249 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
250 if (Tokens[i][0] == '$' &&
251 std::find(Tokens[i].begin(),
252 Tokens[i].end(), ':') != Tokens[i].end()) {
254 errs() << "warning: '" << Name << "': "
255 << "ignoring instruction; operand with attribute '"
256 << Tokens[i] << "', \n";
261 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
263 errs() << "warning: '" << Name << "': "
264 << "ignoring instruction; tied operand '"
265 << Tokens[i] << "', \n";
276 /// ClassInfo - Helper class for storing the information about a particular
277 /// class of operands which can be matched.
280 Token, ///< The class for a particular token.
281 Register, ///< A register class.
282 User ///< A user defined class.
285 /// Name - The class name, suitable for use as an enum.
288 /// ValueName - The name of the value this class represents; for a token this
289 /// is the literal token string, for an operand it is the TableGen class (or
290 /// empty if this is a derived class).
291 std::string ValueName;
293 /// PredicateMethod - The name of the operand method to test whether the
294 /// operand matches this class; this is not valid for Token kinds.
295 std::string PredicateMethod;
297 /// RenderMethod - The name of the operand method to add this operand to an
298 /// MCInst; this is not valid for Token kinds.
299 std::string RenderMethod;
302 /// InstructionInfo - Helper class for storing the necessary information for an
303 /// instruction which is capable of being matched.
304 struct InstructionInfo {
306 /// The unique class instance this operand should match.
309 /// The original operand this corresponds to, if any.
310 const CodeGenInstruction::OperandInfo *OperandInfo;
313 /// InstrName - The target name for this instruction.
314 std::string InstrName;
316 /// Instr - The instruction this matches.
317 const CodeGenInstruction *Instr;
319 /// AsmString - The assembly string for this instruction (with variants
321 std::string AsmString;
323 /// Tokens - The tokenized assembly pattern that this instruction matches.
324 SmallVector<StringRef, 4> Tokens;
326 /// Operands - The operands that this instruction matches.
327 SmallVector<Operand, 4> Operands;
329 /// ConversionFnKind - The enum value which is passed to the generated
330 /// ConvertToMCInst to convert parsed operands into an MCInst for this
332 std::string ConversionFnKind;
338 class AsmMatcherInfo {
340 /// The classes which are needed for matching.
341 std::vector<ClassInfo*> Classes;
343 /// The information on the instruction to match.
344 std::vector<InstructionInfo*> Instructions;
347 /// Map of token to class information which has already been constructed.
348 std::map<std::string, ClassInfo*> TokenClasses;
350 /// Map of operand name to class information which has already been
352 std::map<std::string, ClassInfo*> OperandClasses;
355 /// getTokenClass - Lookup or create the class for the given token.
356 ClassInfo *getTokenClass(const StringRef &Token);
358 /// getOperandClass - Lookup or create the class for the given operand.
359 ClassInfo *getOperandClass(const StringRef &Token,
360 const CodeGenInstruction::OperandInfo &OI);
363 /// BuildInfo - Construct the various tables used during matching.
364 void BuildInfo(CodeGenTarget &Target);
369 void InstructionInfo::dump() {
370 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
372 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
379 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
380 Operand &Op = Operands[i];
381 errs() << " op[" << i << "] = ";
382 if (Op.Class->Kind == ClassInfo::Token) {
383 errs() << '\"' << Tokens[i] << "\"\n";
387 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
388 errs() << OI.Name << " " << OI.Rec->getName()
389 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
393 static std::string getEnumNameForToken(const StringRef &Str) {
396 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
398 case '*': Res += "_STAR_"; break;
399 case '%': Res += "_PCT_"; break;
400 case ':': Res += "_COLON_"; break;
406 Res += "_" + utostr((unsigned) *it) + "_";
414 ClassInfo *AsmMatcherInfo::getTokenClass(const StringRef &Token) {
415 ClassInfo *&Entry = TokenClasses[Token];
418 Entry = new ClassInfo();
419 Entry->Kind = ClassInfo::Token;
420 Entry->Name = "MCK_" + getEnumNameForToken(Token);
421 Entry->ValueName = Token;
422 Entry->PredicateMethod = "<invalid>";
423 Entry->RenderMethod = "<invalid>";
424 Classes.push_back(Entry);
431 AsmMatcherInfo::getOperandClass(const StringRef &Token,
432 const CodeGenInstruction::OperandInfo &OI) {
433 std::string ClassName;
434 if (OI.Rec->isSubClassOf("RegisterClass")) {
436 } else if (OI.Rec->isSubClassOf("Operand")) {
437 // FIXME: This should not be hard coded.
438 const RecordVal *RV = OI.Rec->getValue("Type");
440 // FIXME: Yet another total hack.
441 if (RV->getValue()->getAsString() == "iPTR" ||
442 OI.Rec->getName() == "i8mem_NOREX" ||
443 OI.Rec->getName() == "lea32mem" ||
444 OI.Rec->getName() == "lea64mem" ||
445 OI.Rec->getName() == "i128mem" ||
446 OI.Rec->getName() == "sdmem" ||
447 OI.Rec->getName() == "ssmem" ||
448 OI.Rec->getName() == "lea64_32mem") {
455 ClassInfo *&Entry = OperandClasses[ClassName];
458 Entry = new ClassInfo();
460 if (ClassName == "Reg") {
461 Entry->Kind = ClassInfo::Register;
463 Entry->Kind = ClassInfo::User;
465 Entry->Name = "MCK_" + ClassName;
466 Entry->ValueName = OI.Rec->getName();
467 Entry->PredicateMethod = "is" + ClassName;
468 Entry->RenderMethod = "add" + ClassName + "Operands";
469 Classes.push_back(Entry);
475 void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
476 for (std::map<std::string, CodeGenInstruction>::const_iterator
477 it = Target.getInstructions().begin(),
478 ie = Target.getInstructions().end();
480 const CodeGenInstruction &CGI = it->second;
482 if (!MatchOneInstr.empty() && it->first != MatchOneInstr)
485 OwningPtr<InstructionInfo> II(new InstructionInfo);
487 II->InstrName = it->first;
488 II->Instr = &it->second;
489 II->AsmString = FlattenVariants(CGI.AsmString, 0);
491 TokenizeAsmString(II->AsmString, II->Tokens);
493 // Ignore instructions which shouldn't be matched.
494 if (!IsAssemblerInstruction(it->first, CGI, II->Tokens))
497 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
498 StringRef Token = II->Tokens[i];
500 // Check for simple tokens.
501 if (Token[0] != '$') {
502 InstructionInfo::Operand Op;
503 Op.Class = getTokenClass(Token);
505 II->Operands.push_back(Op);
509 // Otherwise this is an operand reference.
510 StringRef OperandName;
512 OperandName = Token.substr(2, Token.size() - 3);
514 OperandName = Token.substr(1);
516 // Map this token to an operand. FIXME: Move elsewhere.
519 Idx = CGI.getOperandNamed(OperandName);
521 errs() << "error: unable to find operand: '" << OperandName << "'!\n";
525 const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx];
526 InstructionInfo::Operand Op;
527 Op.Class = getOperandClass(Token, OI);
528 Op.OperandInfo = &OI;
529 II->Operands.push_back(Op);
532 // If we broke out, ignore the instruction.
533 if (II->Operands.size() != II->Tokens.size())
536 Instructions.push_back(II.take());
540 static void ConstructConversionFunctions(CodeGenTarget &Target,
541 std::vector<InstructionInfo*> &Infos,
543 // Write the convert function to a separate stream, so we can drop it after
545 std::string ConvertFnBody;
546 raw_string_ostream CvtOS(ConvertFnBody);
548 // Function we have already generated.
549 std::set<std::string> GeneratedFns;
551 // Start the unified conversion function.
553 CvtOS << "static bool ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
554 << "unsigned Opcode,\n"
555 << " SmallVectorImpl<"
556 << Target.getName() << "Operand> &Operands) {\n";
557 CvtOS << " Inst.setOpcode(Opcode);\n";
558 CvtOS << " switch (Kind) {\n";
559 CvtOS << " default:\n";
561 // Start the enum, which we will generate inline.
563 OS << "// Unified function for converting operants to MCInst instances.\n\n";
564 OS << "enum ConversionKind {\n";
566 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
567 ie = Infos.end(); it != ie; ++it) {
568 InstructionInfo &II = **it;
570 // Order the (class) operands by the order to convert them into an MCInst.
571 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
572 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
573 InstructionInfo::Operand &Op = II.Operands[i];
575 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
577 std::sort(MIOperandList.begin(), MIOperandList.end());
579 // Compute the total number of operands.
580 unsigned NumMIOperands = 0;
581 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
582 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i];
583 NumMIOperands = std::max(NumMIOperands,
584 OI.MIOperandNo + OI.MINumOperands);
587 // Build the conversion function signature.
588 std::string Signature = "Convert";
589 unsigned CurIndex = 0;
590 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
591 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
592 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
593 "Duplicate match for instruction operand!");
597 // Skip operands which weren't matched by anything, this occurs when the
598 // .td file encodes "implicit" operands as explicit ones.
600 // FIXME: This should be removed from the MCInst structure.
601 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
604 Signature += Op.Class->Name;
605 Signature += utostr(Op.OperandInfo->MINumOperands);
606 Signature += "_" + utostr(MIOperandList[i].second);
608 CurIndex += Op.OperandInfo->MINumOperands;
611 // Add any trailing implicit operands.
612 for (; CurIndex != NumMIOperands; ++CurIndex)
615 II.ConversionFnKind = Signature;
617 // Check if we have already generated this signature.
618 if (!GeneratedFns.insert(Signature).second)
621 // If not, emit it now.
623 // Add to the enum list.
624 OS << " " << Signature << ",\n";
626 // And to the convert function.
627 CvtOS << " case " << Signature << ":\n";
629 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
630 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
632 // Add the implicit operands.
633 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
634 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
636 CvtOS << " Operands[" << MIOperandList[i].second
637 << "]." << Op.Class->RenderMethod
638 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
639 CurIndex += Op.OperandInfo->MINumOperands;
642 // And add trailing implicit operands.
643 for (; CurIndex != NumMIOperands; ++CurIndex)
644 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
645 CvtOS << " break;\n";
648 // Finish the convert function.
651 CvtOS << " return false;\n";
654 // Finish the enum, and drop the convert function after it.
656 OS << " NumConversionVariants\n";
662 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
663 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
664 std::vector<ClassInfo*> &Infos,
666 OS << "namespace {\n\n";
668 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
669 << "/// instruction matching.\n";
670 OS << "enum MatchClassKind {\n";
671 OS << " InvalidMatchClass = 0,\n";
672 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
673 ie = Infos.end(); it != ie; ++it) {
674 ClassInfo &CI = **it;
675 OS << " " << CI.Name << ", // ";
676 if (CI.Kind == ClassInfo::Token) {
677 OS << "'" << CI.ValueName << "'\n";
678 } else if (CI.Kind == ClassInfo::Register) {
679 if (!CI.ValueName.empty())
680 OS << "register class '" << CI.ValueName << "'\n";
682 OS << "derived register class\n";
684 OS << "user defined class '" << CI.ValueName << "'\n";
687 OS << " NumMatchClassKinds\n";
693 /// EmitMatchRegisterName - Emit the function to match a string to appropriate
694 /// match class value.
695 static void EmitMatchTokenString(CodeGenTarget &Target,
696 std::vector<ClassInfo*> &Infos,
698 // FIXME: TableGen should have a fast string matcher generator.
699 OS << "static MatchClassKind MatchTokenString(const StringRef &Name) {\n";
700 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
701 ie = Infos.end(); it != ie; ++it) {
702 ClassInfo &CI = **it;
704 if (CI.Kind == ClassInfo::Token)
705 OS << " if (Name == \"" << CI.ValueName << "\")\n"
706 << " return " << CI.Name << ";\n\n";
708 OS << " return InvalidMatchClass;\n";
712 /// EmitClassifyOperand - Emit the function to classify an operand.
713 static void EmitClassifyOperand(CodeGenTarget &Target,
714 std::vector<ClassInfo*> &Infos,
716 OS << "static MatchClassKind ClassifyOperand("
717 << Target.getName() << "Operand &Operand) {\n";
718 OS << " if (Operand.isToken())\n";
719 OS << " return MatchTokenString(Operand.getToken());\n\n";
720 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
721 ie = Infos.end(); it != ie; ++it) {
722 ClassInfo &CI = **it;
724 if (CI.Kind != ClassInfo::Token) {
725 OS << " if (Operand." << CI.PredicateMethod << "())\n";
726 OS << " return " << CI.Name << ";\n\n";
729 OS << " return InvalidMatchClass;\n";
733 /// EmitMatchRegisterName - Emit the function to match a string to the target
734 /// specific register enum.
735 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
737 const std::vector<CodeGenRegister> &Registers = Target.getRegisters();
739 OS << "bool " << Target.getName()
740 << AsmParser->getValueAsString("AsmParserClassName")
741 << "::MatchRegisterName(const StringRef &Name, unsigned &RegNo) {\n";
743 // FIXME: TableGen should have a fast string matcher generator.
744 for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
745 const CodeGenRegister &Reg = Registers[i];
746 if (Reg.TheDef->getValueAsString("AsmName").empty())
749 OS << " if (Name == \""
750 << Reg.TheDef->getValueAsString("AsmName") << "\")\n"
751 << " return RegNo=" << i + 1 << ", false;\n";
753 OS << " return true;\n";
757 void AsmMatcherEmitter::run(raw_ostream &OS) {
758 CodeGenTarget Target;
759 Record *AsmParser = Target.getAsmParser();
760 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
762 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
764 // Emit the function to match a register name to number.
765 EmitMatchRegisterName(Target, AsmParser, OS);
767 // Compute the information on the instructions to match.
769 Info.BuildInfo(Target);
771 DEBUG_WITH_TYPE("instruction_info", {
772 for (std::vector<InstructionInfo*>::iterator
773 it = Info.Instructions.begin(), ie = Info.Instructions.end();
778 // FIXME: At this point we should be able to totally order Infos, if not then
779 // we have an ambiguity which the .td file should be forced to resolve.
781 // Generate the terminal actions to convert operands into an MCInst.
782 ConstructConversionFunctions(Target, Info.Instructions, OS);
784 // Emit the enumeration for classes which participate in matching.
785 EmitMatchClassEnumeration(Target, Info.Classes, OS);
787 // Emit the routine to match token strings to their match class.
788 EmitMatchTokenString(Target, Info.Classes, OS);
790 // Emit the routine to classify an operand.
791 EmitClassifyOperand(Target, Info.Classes, OS);
793 // Finally, build the match function.
795 size_t MaxNumOperands = 0;
796 for (std::vector<InstructionInfo*>::const_iterator it =
797 Info.Instructions.begin(), ie = Info.Instructions.end();
799 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
801 OS << "bool " << Target.getName() << ClassName
802 << "::MatchInstruction("
803 << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
804 << "MCInst &Inst) {\n";
806 // Emit the static match table; unused classes get initalized to 0 which is
807 // guaranteed to be InvalidMatchClass.
809 // FIXME: We can reduce the size of this table very easily. First, we change
810 // it so that store the kinds in separate bit-fields for each index, which
811 // only needs to be the max width used for classes at that index (we also need
812 // to reject based on this during classification). If we then make sure to
813 // order the match kinds appropriately (putting mnemonics last), then we
814 // should only end up using a few bits for each class, especially the ones
815 // following the mnemonic.
816 OS << " static const struct MatchEntry {\n";
817 OS << " unsigned Opcode;\n";
818 OS << " ConversionKind ConvertFn;\n";
819 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
820 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n";
822 for (std::vector<InstructionInfo*>::const_iterator it =
823 Info.Instructions.begin(), ie = Info.Instructions.end();
825 InstructionInfo &II = **it;
827 OS << " { " << Target.getName() << "::" << II.InstrName
828 << ", " << II.ConversionFnKind << ", { ";
829 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
830 InstructionInfo::Operand &Op = II.Operands[i];
833 OS << Op.Class->Name;
840 // Emit code to compute the class list for this operand vector.
841 OS << " // Eliminate obvious mismatches.\n";
842 OS << " if (Operands.size() > " << MaxNumOperands << ")\n";
843 OS << " return true;\n\n";
845 OS << " // Compute the class list for this operand vector.\n";
846 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
847 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n";
848 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n";
850 OS << " // Check for invalid operands before matching.\n";
851 OS << " if (Classes[i] == InvalidMatchClass)\n";
852 OS << " return true;\n";
855 OS << " // Mark unused classes.\n";
856 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; "
858 OS << " Classes[i] = InvalidMatchClass;\n\n";
860 // Emit code to search the table.
861 OS << " // Search the table.\n";
862 OS << " for (MatchEntry *it = MatchTable, "
863 << "*ie = MatchTable + " << Info.Instructions.size()
864 << "; it != ie; ++it) {\n";
865 for (unsigned i = 0; i != MaxNumOperands; ++i) {
866 OS << " if (Classes[" << i << "] != it->Classes[" << i << "])\n";
867 OS << " continue;\n";
870 OS << " return ConvertToMCInst(it->ConvertFn, Inst, "
871 << "it->Opcode, Operands);\n";
874 OS << " return true;\n";