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/STLExtras.h"
82 #include "llvm/ADT/StringExtras.h"
83 #include "llvm/Support/CommandLine.h"
84 #include "llvm/Support/Debug.h"
91 static cl::opt<std::string>
92 MatchPrefix("match-prefix", cl::init(""),
93 cl::desc("Only match instructions with the given prefix"));
96 /// FlattenVariants - Flatten an .td file assembly string by selecting the
97 /// variant at index \arg N.
98 static std::string FlattenVariants(const std::string &AsmString,
100 StringRef Cur = AsmString;
101 std::string Res = "";
104 // Find the start of the next variant string.
105 size_t VariantsStart = 0;
106 for (size_t e = Cur.size(); VariantsStart != e; ++VariantsStart)
107 if (Cur[VariantsStart] == '{' &&
108 (VariantsStart == 0 || (Cur[VariantsStart-1] != '$' &&
109 Cur[VariantsStart-1] != '\\')))
112 // Add the prefix to the result.
113 Res += Cur.slice(0, VariantsStart);
114 if (VariantsStart == Cur.size())
117 ++VariantsStart; // Skip the '{'.
119 // Scan to the end of the variants string.
120 size_t VariantsEnd = VariantsStart;
121 unsigned NestedBraces = 1;
122 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) {
123 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') {
124 if (--NestedBraces == 0)
126 } else if (Cur[VariantsEnd] == '{')
130 // Select the Nth variant (or empty).
131 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd);
132 for (unsigned i = 0; i != N; ++i)
133 Selection = Selection.split('|').second;
134 Res += Selection.split('|').first;
136 assert(VariantsEnd != Cur.size() &&
137 "Unterminated variants in assembly string!");
138 Cur = Cur.substr(VariantsEnd + 1);
144 /// TokenizeAsmString - Tokenize a simplified assembly string.
145 static void TokenizeAsmString(const StringRef &AsmString,
146 SmallVectorImpl<StringRef> &Tokens) {
149 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) {
150 switch (AsmString[i]) {
159 Tokens.push_back(AsmString.slice(Prev, i));
162 if (!isspace(AsmString[i]) && AsmString[i] != ',')
163 Tokens.push_back(AsmString.substr(i, 1));
169 Tokens.push_back(AsmString.slice(Prev, i));
173 assert(i != AsmString.size() && "Invalid quoted character");
174 Tokens.push_back(AsmString.substr(i, 1));
179 // If this isn't "${", treat like a normal token.
180 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') {
182 Tokens.push_back(AsmString.slice(Prev, i));
190 Tokens.push_back(AsmString.slice(Prev, i));
194 StringRef::iterator End =
195 std::find(AsmString.begin() + i, AsmString.end(), '}');
196 assert(End != AsmString.end() && "Missing brace in operand reference!");
197 size_t EndPos = End - AsmString.begin();
198 Tokens.push_back(AsmString.slice(i, EndPos+1));
208 if (InTok && Prev != AsmString.size())
209 Tokens.push_back(AsmString.substr(Prev));
212 static bool IsAssemblerInstruction(const StringRef &Name,
213 const CodeGenInstruction &CGI,
214 const SmallVectorImpl<StringRef> &Tokens) {
215 // Ignore psuedo ops.
217 // FIXME: This is a hack.
218 if (const RecordVal *Form = CGI.TheDef->getValue("Form"))
219 if (Form->getValue()->getAsString() == "Pseudo")
222 // Ignore "PHI" node.
224 // FIXME: This is also a hack.
228 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
230 // FIXME: This is a total hack.
231 if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int"))
234 // Ignore instructions with no .s string.
236 // FIXME: What are these?
237 if (CGI.AsmString.empty())
240 // FIXME: Hack; ignore any instructions with a newline in them.
241 if (std::find(CGI.AsmString.begin(),
242 CGI.AsmString.end(), '\n') != CGI.AsmString.end())
245 // Ignore instructions with attributes, these are always fake instructions for
246 // simplifying codegen.
248 // FIXME: Is this true?
250 // Also, we ignore instructions which reference the operand multiple times;
251 // this implies a constraint we would not currently honor. These are
252 // currently always fake instructions for simplifying codegen.
254 // FIXME: Encode this assumption in the .td, so we can error out here.
255 std::set<std::string> OperandNames;
256 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
257 if (Tokens[i][0] == '$' &&
258 std::find(Tokens[i].begin(),
259 Tokens[i].end(), ':') != Tokens[i].end()) {
261 errs() << "warning: '" << Name << "': "
262 << "ignoring instruction; operand with attribute '"
263 << Tokens[i] << "', \n";
268 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
270 errs() << "warning: '" << Name << "': "
271 << "ignoring instruction; tied operand '"
272 << Tokens[i] << "'\n";
283 /// ClassInfo - Helper class for storing the information about a particular
284 /// class of operands which can be matched.
287 Invalid = 0, ///< Invalid kind, for use as a sentinel value.
288 Token, ///< The class for a particular token.
289 Register, ///< A register class.
290 UserClass0 ///< The (first) user defined class, subsequent user defined
291 /// classes are UserClass0+1, and so on.
294 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
295 /// N) for the Nth user defined class.
298 /// SuperClassKind - The super class kind for user classes.
299 unsigned SuperClassKind;
301 /// SuperClass - The super class, or 0.
302 ClassInfo *SuperClass;
304 /// Name - The full class name, suitable for use in an enum.
307 /// ClassName - The unadorned generic name for this class (e.g., Token).
308 std::string ClassName;
310 /// ValueName - The name of the value this class represents; for a token this
311 /// is the literal token string, for an operand it is the TableGen class (or
312 /// empty if this is a derived class).
313 std::string ValueName;
315 /// PredicateMethod - The name of the operand method to test whether the
316 /// operand matches this class; this is not valid for Token kinds.
317 std::string PredicateMethod;
319 /// RenderMethod - The name of the operand method to add this operand to an
320 /// MCInst; this is not valid for Token kinds.
321 std::string RenderMethod;
323 /// isUserClass() - Check if this is a user defined class.
324 bool isUserClass() const {
325 return Kind >= UserClass0;
328 /// getRootClass - Return the root class of this one.
329 const ClassInfo *getRootClass() const {
330 const ClassInfo *CI = this;
331 while (CI->SuperClass)
336 /// operator< - Compare two classes.
337 bool operator<(const ClassInfo &RHS) const {
338 // Incompatible kinds are comparable for classes in disjoint hierarchies.
339 if (Kind != RHS.Kind && getRootClass() != RHS.getRootClass())
340 return Kind < RHS.Kind;
344 assert(0 && "Invalid kind!");
346 // Tokens are comparable by value.
348 // FIXME: Compare by enum value.
349 return ValueName < RHS.ValueName;
352 // This class preceeds the RHS if the RHS is a super class.
353 for (ClassInfo *Parent = SuperClass; Parent; Parent = Parent->SuperClass)
362 /// InstructionInfo - Helper class for storing the necessary information for an
363 /// instruction which is capable of being matched.
364 struct InstructionInfo {
366 /// The unique class instance this operand should match.
369 /// The original operand this corresponds to, if any.
370 const CodeGenInstruction::OperandInfo *OperandInfo;
373 /// InstrName - The target name for this instruction.
374 std::string InstrName;
376 /// Instr - The instruction this matches.
377 const CodeGenInstruction *Instr;
379 /// AsmString - The assembly string for this instruction (with variants
381 std::string AsmString;
383 /// Tokens - The tokenized assembly pattern that this instruction matches.
384 SmallVector<StringRef, 4> Tokens;
386 /// Operands - The operands that this instruction matches.
387 SmallVector<Operand, 4> Operands;
389 /// ConversionFnKind - The enum value which is passed to the generated
390 /// ConvertToMCInst to convert parsed operands into an MCInst for this
392 std::string ConversionFnKind;
394 /// operator< - Compare two instructions.
395 bool operator<(const InstructionInfo &RHS) const {
396 if (Operands.size() != RHS.Operands.size())
397 return Operands.size() < RHS.Operands.size();
399 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
400 if (*Operands[i].Class < *RHS.Operands[i].Class)
406 /// CouldMatchAmiguouslyWith - Check whether this instruction could
407 /// ambiguously match the same set of operands as \arg RHS (without being a
408 /// strictly superior match).
409 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
410 // The number of operands is unambiguous.
411 if (Operands.size() != RHS.Operands.size())
414 // Tokens and operand kinds are unambiguous (assuming a correct target
416 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
417 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
418 Operands[i].Class->Kind == ClassInfo::Token)
419 if (*Operands[i].Class < *RHS.Operands[i].Class ||
420 *RHS.Operands[i].Class < *Operands[i].Class)
423 // Otherwise, this operand could commute if all operands are equivalent, or
424 // there is a pair of operands that compare less than and a pair that
425 // compare greater than.
426 bool HasLT = false, HasGT = false;
427 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
428 if (*Operands[i].Class < *RHS.Operands[i].Class)
430 if (*RHS.Operands[i].Class < *Operands[i].Class)
434 return !(HasLT ^ HasGT);
441 class AsmMatcherInfo {
443 /// The classes which are needed for matching.
444 std::vector<ClassInfo*> Classes;
446 /// The information on the instruction to match.
447 std::vector<InstructionInfo*> Instructions;
450 /// Map of token to class information which has already been constructed.
451 std::map<std::string, ClassInfo*> TokenClasses;
453 /// Map of operand name to class information which has already been
455 std::map<std::string, ClassInfo*> OperandClasses;
457 /// Map of user class names to kind value.
458 std::map<std::string, unsigned> UserClasses;
461 /// getTokenClass - Lookup or create the class for the given token.
462 ClassInfo *getTokenClass(const StringRef &Token);
464 /// getUserClassKind - Lookup or create the kind value for the given class
466 unsigned getUserClassKind(const StringRef &Name);
468 /// getOperandClass - Lookup or create the class for the given operand.
469 ClassInfo *getOperandClass(const StringRef &Token,
470 const CodeGenInstruction::OperandInfo &OI);
473 /// BuildInfo - Construct the various tables used during matching.
474 void BuildInfo(CodeGenTarget &Target);
479 void InstructionInfo::dump() {
480 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
482 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
489 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
490 Operand &Op = Operands[i];
491 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
492 if (Op.Class->Kind == ClassInfo::Token) {
493 errs() << '\"' << Tokens[i] << "\"\n";
497 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
498 errs() << OI.Name << " " << OI.Rec->getName()
499 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
503 static std::string getEnumNameForToken(const StringRef &Str) {
506 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
508 case '*': Res += "_STAR_"; break;
509 case '%': Res += "_PCT_"; break;
510 case ':': Res += "_COLON_"; break;
516 Res += "_" + utostr((unsigned) *it) + "_";
524 ClassInfo *AsmMatcherInfo::getTokenClass(const StringRef &Token) {
525 ClassInfo *&Entry = TokenClasses[Token];
528 Entry = new ClassInfo();
529 Entry->Kind = ClassInfo::Token;
530 Entry->ClassName = "Token";
531 Entry->Name = "MCK_" + getEnumNameForToken(Token);
532 Entry->ValueName = Token;
533 Entry->PredicateMethod = "<invalid>";
534 Entry->RenderMethod = "<invalid>";
535 Classes.push_back(Entry);
541 unsigned AsmMatcherInfo::getUserClassKind(const StringRef &Name) {
542 unsigned &Entry = UserClasses[Name];
545 Entry = ClassInfo::UserClass0 + UserClasses.size() - 1;
551 AsmMatcherInfo::getOperandClass(const StringRef &Token,
552 const CodeGenInstruction::OperandInfo &OI) {
553 unsigned SuperClass = ClassInfo::Invalid;
554 std::string ClassName;
555 if (OI.Rec->isSubClassOf("RegisterClass")) {
559 ClassName = OI.Rec->getValueAsString("ParserMatchClass");
560 assert(ClassName != "Reg" && "'Reg' class name is reserved!");
562 PrintError(OI.Rec->getLoc(), "operand has no match class!");
563 ClassName = "Invalid";
566 // Determine the super class.
568 std::string SuperClassName =
569 OI.Rec->getValueAsString("ParserMatchSuperClass");
570 SuperClass = getUserClassKind(SuperClassName);
574 ClassInfo *&Entry = OperandClasses[ClassName];
577 Entry = new ClassInfo();
578 if (ClassName == "Reg") {
579 Entry->Kind = ClassInfo::Register;
580 Entry->SuperClassKind = SuperClass;
582 Entry->Kind = getUserClassKind(ClassName);
583 Entry->SuperClassKind = SuperClass;
585 Entry->ClassName = ClassName;
586 Entry->Name = "MCK_" + ClassName;
587 Entry->ValueName = OI.Rec->getName();
588 Entry->PredicateMethod = "is" + ClassName;
589 Entry->RenderMethod = "add" + ClassName + "Operands";
590 Classes.push_back(Entry);
592 // Verify the super class matches.
593 assert(SuperClass == Entry->SuperClassKind &&
594 "Cannot redefine super class kind!");
600 void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
601 for (std::map<std::string, CodeGenInstruction>::const_iterator
602 it = Target.getInstructions().begin(),
603 ie = Target.getInstructions().end();
605 const CodeGenInstruction &CGI = it->second;
607 if (!StringRef(it->first).startswith(MatchPrefix))
610 OwningPtr<InstructionInfo> II(new InstructionInfo);
612 II->InstrName = it->first;
613 II->Instr = &it->second;
614 II->AsmString = FlattenVariants(CGI.AsmString, 0);
616 TokenizeAsmString(II->AsmString, II->Tokens);
618 // Ignore instructions which shouldn't be matched.
619 if (!IsAssemblerInstruction(it->first, CGI, II->Tokens))
622 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
623 StringRef Token = II->Tokens[i];
625 // Check for simple tokens.
626 if (Token[0] != '$') {
627 InstructionInfo::Operand Op;
628 Op.Class = getTokenClass(Token);
630 II->Operands.push_back(Op);
634 // Otherwise this is an operand reference.
635 StringRef OperandName;
637 OperandName = Token.substr(2, Token.size() - 3);
639 OperandName = Token.substr(1);
641 // Map this token to an operand. FIXME: Move elsewhere.
644 Idx = CGI.getOperandNamed(OperandName);
646 errs() << "error: unable to find operand: '" << OperandName << "'!\n";
650 const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx];
651 InstructionInfo::Operand Op;
652 Op.Class = getOperandClass(Token, OI);
653 Op.OperandInfo = &OI;
654 II->Operands.push_back(Op);
657 // If we broke out, ignore the instruction.
658 if (II->Operands.size() != II->Tokens.size())
661 Instructions.push_back(II.take());
664 // Bind user super classes.
665 std::map<unsigned, ClassInfo*> UserClasses;
666 for (unsigned i = 0, e = Classes.size(); i != e; ++i) {
667 ClassInfo &CI = *Classes[i];
668 if (CI.isUserClass())
669 UserClasses[CI.Kind] = &CI;
672 for (unsigned i = 0, e = Classes.size(); i != e; ++i) {
673 ClassInfo &CI = *Classes[i];
674 if (CI.isUserClass() && CI.SuperClassKind != ClassInfo::Invalid) {
675 CI.SuperClass = UserClasses[CI.SuperClassKind];
676 assert(CI.SuperClass && "Missing super class definition!");
682 // Reorder classes so that classes preceed super classes.
683 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
686 static void EmitConvertToMCInst(CodeGenTarget &Target,
687 std::vector<InstructionInfo*> &Infos,
689 // Write the convert function to a separate stream, so we can drop it after
691 std::string ConvertFnBody;
692 raw_string_ostream CvtOS(ConvertFnBody);
694 // Function we have already generated.
695 std::set<std::string> GeneratedFns;
697 // Start the unified conversion function.
699 CvtOS << "static bool ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
700 << "unsigned Opcode,\n"
701 << " SmallVectorImpl<"
702 << Target.getName() << "Operand> &Operands) {\n";
703 CvtOS << " Inst.setOpcode(Opcode);\n";
704 CvtOS << " switch (Kind) {\n";
705 CvtOS << " default:\n";
707 // Start the enum, which we will generate inline.
709 OS << "// Unified function for converting operants to MCInst instances.\n\n";
710 OS << "enum ConversionKind {\n";
712 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
713 ie = Infos.end(); it != ie; ++it) {
714 InstructionInfo &II = **it;
716 // Order the (class) operands by the order to convert them into an MCInst.
717 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
718 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
719 InstructionInfo::Operand &Op = II.Operands[i];
721 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
723 std::sort(MIOperandList.begin(), MIOperandList.end());
725 // Compute the total number of operands.
726 unsigned NumMIOperands = 0;
727 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
728 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i];
729 NumMIOperands = std::max(NumMIOperands,
730 OI.MIOperandNo + OI.MINumOperands);
733 // Build the conversion function signature.
734 std::string Signature = "Convert";
735 unsigned CurIndex = 0;
736 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
737 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
738 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
739 "Duplicate match for instruction operand!");
743 // Skip operands which weren't matched by anything, this occurs when the
744 // .td file encodes "implicit" operands as explicit ones.
746 // FIXME: This should be removed from the MCInst structure.
747 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
750 Signature += Op.Class->ClassName;
751 Signature += utostr(Op.OperandInfo->MINumOperands);
752 Signature += "_" + utostr(MIOperandList[i].second);
754 CurIndex += Op.OperandInfo->MINumOperands;
757 // Add any trailing implicit operands.
758 for (; CurIndex != NumMIOperands; ++CurIndex)
761 II.ConversionFnKind = Signature;
763 // Check if we have already generated this signature.
764 if (!GeneratedFns.insert(Signature).second)
767 // If not, emit it now.
769 // Add to the enum list.
770 OS << " " << Signature << ",\n";
772 // And to the convert function.
773 CvtOS << " case " << Signature << ":\n";
775 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
776 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
778 // Add the implicit operands.
779 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
780 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
782 CvtOS << " Operands[" << MIOperandList[i].second
783 << "]." << Op.Class->RenderMethod
784 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
785 CurIndex += Op.OperandInfo->MINumOperands;
788 // And add trailing implicit operands.
789 for (; CurIndex != NumMIOperands; ++CurIndex)
790 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
791 CvtOS << " break;\n";
794 // Finish the convert function.
797 CvtOS << " return false;\n";
800 // Finish the enum, and drop the convert function after it.
802 OS << " NumConversionVariants\n";
808 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
809 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
810 std::vector<ClassInfo*> &Infos,
812 OS << "namespace {\n\n";
814 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
815 << "/// instruction matching.\n";
816 OS << "enum MatchClassKind {\n";
817 OS << " InvalidMatchClass = 0,\n";
818 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
819 ie = Infos.end(); it != ie; ++it) {
820 ClassInfo &CI = **it;
821 OS << " " << CI.Name << ", // ";
822 if (CI.Kind == ClassInfo::Token) {
823 OS << "'" << CI.ValueName << "'\n";
824 } else if (CI.Kind == ClassInfo::Register) {
825 if (!CI.ValueName.empty())
826 OS << "register class '" << CI.ValueName << "'\n";
828 OS << "derived register class\n";
830 OS << "user defined class '" << CI.ValueName << "'\n";
833 OS << " NumMatchClassKinds\n";
839 /// EmitClassifyOperand - Emit the function to classify an operand.
840 static void EmitClassifyOperand(CodeGenTarget &Target,
841 std::vector<ClassInfo*> &Infos,
843 OS << "static MatchClassKind ClassifyOperand("
844 << Target.getName() << "Operand &Operand) {\n";
845 OS << " if (Operand.isToken())\n";
846 OS << " return MatchTokenString(Operand.getToken());\n\n";
847 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
848 ie = Infos.end(); it != ie; ++it) {
849 ClassInfo &CI = **it;
851 if (CI.Kind != ClassInfo::Token) {
852 OS << " // '" << CI.ClassName << "' class";
854 OS << ", subclass of '" << CI.SuperClass->ClassName << "'";
855 assert(CI < *CI.SuperClass && "Invalid class relation!");
859 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
861 // Validate subclass relationships.
863 OS << " assert(Operand." << CI.SuperClass->PredicateMethod
864 << "() && \"Invalid class relationship!\");\n";
866 OS << " return " << CI.Name << ";\n\n";
870 OS << " return InvalidMatchClass;\n";
874 typedef std::pair<std::string, std::string> StringPair;
876 /// FindFirstNonCommonLetter - Find the first character in the keys of the
877 /// string pairs that is not shared across the whole set of strings. All
878 /// strings are assumed to have the same length.
880 FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) {
881 assert(!Matches.empty());
882 for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) {
883 // Check to see if letter i is the same across the set.
884 char Letter = Matches[0]->first[i];
886 for (unsigned str = 0, e = Matches.size(); str != e; ++str)
887 if (Matches[str]->first[i] != Letter)
891 return Matches[0]->first.size();
894 /// EmitStringMatcherForChar - Given a set of strings that are known to be the
895 /// same length and whose characters leading up to CharNo are the same, emit
896 /// code to verify that CharNo and later are the same.
898 /// \return - True if control can leave the emitted code fragment.
899 static bool EmitStringMatcherForChar(const std::string &StrVariableName,
900 const std::vector<const StringPair*> &Matches,
901 unsigned CharNo, unsigned IndentCount,
903 assert(!Matches.empty() && "Must have at least one string to match!");
904 std::string Indent(IndentCount*2+4, ' ');
906 // If we have verified that the entire string matches, we're done: output the
908 if (CharNo == Matches[0]->first.size()) {
909 assert(Matches.size() == 1 && "Had duplicate keys to match on");
911 // FIXME: If Matches[0].first has embeded \n, this will be bad.
912 OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first
917 // Bucket the matches by the character we are comparing.
918 std::map<char, std::vector<const StringPair*> > MatchesByLetter;
920 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
921 MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]);
924 // If we have exactly one bucket to match, see how many characters are common
925 // across the whole set and match all of them at once.
926 if (MatchesByLetter.size() == 1) {
927 unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches);
928 unsigned NumChars = FirstNonCommonLetter-CharNo;
930 // Emit code to break out if the prefix doesn't match.
932 // Do the comparison with if (Str[1] != 'f')
933 // FIXME: Need to escape general characters.
934 OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '"
935 << Matches[0]->first[CharNo] << "')\n";
936 OS << Indent << " break;\n";
938 // Do the comparison with if (Str.substr(1,3) != "foo").
939 // FIXME: Need to escape general strings.
940 OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << ","
941 << NumChars << ") != \"";
942 OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n";
943 OS << Indent << " break;\n";
946 return EmitStringMatcherForChar(StrVariableName, Matches,
947 FirstNonCommonLetter, IndentCount, OS);
950 // Otherwise, we have multiple possible things, emit a switch on the
952 OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n";
953 OS << Indent << "default: break;\n";
955 for (std::map<char, std::vector<const StringPair*> >::iterator LI =
956 MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) {
957 // TODO: escape hard stuff (like \n) if we ever care about it.
958 OS << Indent << "case '" << LI->first << "':\t // "
959 << LI->second.size() << " strings to match.\n";
960 if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1,
962 OS << Indent << " break;\n";
965 OS << Indent << "}\n";
970 /// EmitStringMatcher - Given a list of strings and code to execute when they
971 /// match, output a simple switch tree to classify the input string.
973 /// If a match is found, the code in Vals[i].second is executed; control must
974 /// not exit this code fragment. If nothing matches, execution falls through.
976 /// \param StrVariableName - The name of the variable to test.
977 static void EmitStringMatcher(const std::string &StrVariableName,
978 const std::vector<StringPair> &Matches,
980 // First level categorization: group strings by length.
981 std::map<unsigned, std::vector<const StringPair*> > MatchesByLength;
983 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
984 MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]);
986 // Output a switch statement on length and categorize the elements within each
988 OS << " switch (" << StrVariableName << ".size()) {\n";
989 OS << " default: break;\n";
991 for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI =
992 MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) {
993 OS << " case " << LI->first << ":\t // " << LI->second.size()
994 << " strings to match.\n";
995 if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS))
1003 /// EmitMatchTokenString - Emit the function to match a token string to the
1004 /// appropriate match class value.
1005 static void EmitMatchTokenString(CodeGenTarget &Target,
1006 std::vector<ClassInfo*> &Infos,
1008 // Construct the match list.
1009 std::vector<StringPair> Matches;
1010 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1011 ie = Infos.end(); it != ie; ++it) {
1012 ClassInfo &CI = **it;
1014 if (CI.Kind == ClassInfo::Token)
1015 Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";"));
1018 OS << "static MatchClassKind MatchTokenString(const StringRef &Name) {\n";
1020 EmitStringMatcher("Name", Matches, OS);
1022 OS << " return InvalidMatchClass;\n";
1026 /// EmitMatchRegisterName - Emit the function to match a string to the target
1027 /// specific register enum.
1028 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1030 // Construct the match list.
1031 std::vector<StringPair> Matches;
1032 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1033 const CodeGenRegister &Reg = Target.getRegisters()[i];
1034 if (Reg.TheDef->getValueAsString("AsmName").empty())
1037 Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"),
1038 "return " + utostr(i + 1) + ";"));
1041 OS << "unsigned " << Target.getName()
1042 << AsmParser->getValueAsString("AsmParserClassName")
1043 << "::MatchRegisterName(const StringRef &Name) {\n";
1045 EmitStringMatcher("Name", Matches, OS);
1047 OS << " return 0;\n";
1051 void AsmMatcherEmitter::run(raw_ostream &OS) {
1052 CodeGenTarget Target;
1053 Record *AsmParser = Target.getAsmParser();
1054 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1056 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1058 // Emit the function to match a register name to number.
1059 EmitMatchRegisterName(Target, AsmParser, OS);
1061 // Compute the information on the instructions to match.
1062 AsmMatcherInfo Info;
1063 Info.BuildInfo(Target);
1065 // Sort the instruction table using the partial order on classes.
1066 std::sort(Info.Instructions.begin(), Info.Instructions.end(),
1067 less_ptr<InstructionInfo>());
1069 DEBUG_WITH_TYPE("instruction_info", {
1070 for (std::vector<InstructionInfo*>::iterator
1071 it = Info.Instructions.begin(), ie = Info.Instructions.end();
1076 // Check for ambiguous instructions.
1077 unsigned NumAmbiguous = 0;
1078 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
1079 for (unsigned j = i + 1; j != e; ++j) {
1080 InstructionInfo &A = *Info.Instructions[i];
1081 InstructionInfo &B = *Info.Instructions[j];
1083 if (A.CouldMatchAmiguouslyWith(B)) {
1084 DEBUG_WITH_TYPE("ambiguous_instrs", {
1085 errs() << "warning: ambiguous instruction match:\n";
1087 errs() << "\nis incomparable with:\n";
1096 DEBUG_WITH_TYPE("ambiguous_instrs", {
1097 errs() << "warning: " << NumAmbiguous
1098 << " ambiguous instructions!\n";
1101 // Generate the unified function to convert operands into an MCInst.
1102 EmitConvertToMCInst(Target, Info.Instructions, OS);
1104 // Emit the enumeration for classes which participate in matching.
1105 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1107 // Emit the routine to match token strings to their match class.
1108 EmitMatchTokenString(Target, Info.Classes, OS);
1110 // Emit the routine to classify an operand.
1111 EmitClassifyOperand(Target, Info.Classes, OS);
1113 // Finally, build the match function.
1115 size_t MaxNumOperands = 0;
1116 for (std::vector<InstructionInfo*>::const_iterator it =
1117 Info.Instructions.begin(), ie = Info.Instructions.end();
1119 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
1121 OS << "bool " << Target.getName() << ClassName
1122 << "::MatchInstruction("
1123 << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
1124 << "MCInst &Inst) {\n";
1126 // Emit the static match table; unused classes get initalized to 0 which is
1127 // guaranteed to be InvalidMatchClass.
1129 // FIXME: We can reduce the size of this table very easily. First, we change
1130 // it so that store the kinds in separate bit-fields for each index, which
1131 // only needs to be the max width used for classes at that index (we also need
1132 // to reject based on this during classification). If we then make sure to
1133 // order the match kinds appropriately (putting mnemonics last), then we
1134 // should only end up using a few bits for each class, especially the ones
1135 // following the mnemonic.
1136 OS << " static const struct MatchEntry {\n";
1137 OS << " unsigned Opcode;\n";
1138 OS << " ConversionKind ConvertFn;\n";
1139 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1140 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n";
1142 for (std::vector<InstructionInfo*>::const_iterator it =
1143 Info.Instructions.begin(), ie = Info.Instructions.end();
1145 InstructionInfo &II = **it;
1147 OS << " { " << Target.getName() << "::" << II.InstrName
1148 << ", " << II.ConversionFnKind << ", { ";
1149 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1150 InstructionInfo::Operand &Op = II.Operands[i];
1153 OS << Op.Class->Name;
1160 // Emit code to compute the class list for this operand vector.
1161 OS << " // Eliminate obvious mismatches.\n";
1162 OS << " if (Operands.size() > " << MaxNumOperands << ")\n";
1163 OS << " return true;\n\n";
1165 OS << " // Compute the class list for this operand vector.\n";
1166 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1167 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n";
1168 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n";
1170 OS << " // Check for invalid operands before matching.\n";
1171 OS << " if (Classes[i] == InvalidMatchClass)\n";
1172 OS << " return true;\n";
1175 OS << " // Mark unused classes.\n";
1176 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; "
1177 << "i != e; ++i)\n";
1178 OS << " Classes[i] = InvalidMatchClass;\n\n";
1180 // Emit code to search the table.
1181 OS << " // Search the table.\n";
1182 OS << " for (const MatchEntry *it = MatchTable, "
1183 << "*ie = MatchTable + " << Info.Instructions.size()
1184 << "; it != ie; ++it) {\n";
1185 for (unsigned i = 0; i != MaxNumOperands; ++i) {
1186 OS << " if (Classes[" << i << "] != it->Classes[" << i << "])\n";
1187 OS << " continue;\n";
1190 OS << " return ConvertToMCInst(it->ConvertFn, Inst, "
1191 << "it->Opcode, Operands);\n";
1194 OS << " return true;\n";