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 instructions with no .s string.
230 // FIXME: What are these?
231 if (CGI.AsmString.empty())
234 // FIXME: Hack; ignore any instructions with a newline in them.
235 if (std::find(CGI.AsmString.begin(),
236 CGI.AsmString.end(), '\n') != CGI.AsmString.end())
239 // Ignore instructions with attributes, these are always fake instructions for
240 // simplifying codegen.
242 // FIXME: Is this true?
244 // Also, we ignore instructions which reference the operand multiple times;
245 // this implies a constraint we would not currently honor. These are
246 // currently always fake instructions for simplifying codegen.
248 // FIXME: Encode this assumption in the .td, so we can error out here.
249 std::set<std::string> OperandNames;
250 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
251 if (Tokens[i][0] == '$' &&
252 std::find(Tokens[i].begin(),
253 Tokens[i].end(), ':') != Tokens[i].end()) {
255 errs() << "warning: '" << Name << "': "
256 << "ignoring instruction; operand with attribute '"
257 << Tokens[i] << "', \n";
262 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
264 errs() << "warning: '" << Name << "': "
265 << "ignoring instruction; tied operand '"
266 << Tokens[i] << "'\n";
277 /// ClassInfo - Helper class for storing the information about a particular
278 /// class of operands which can be matched.
281 Invalid = 0, ///< Invalid kind, for use as a sentinel value.
282 Token, ///< The class for a particular token.
283 Register, ///< A register class.
284 UserClass0 ///< The (first) user defined class, subsequent user defined
285 /// classes are UserClass0+1, and so on.
288 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
289 /// N) for the Nth user defined class.
292 /// SuperClassKind - The super class kind for user classes.
293 unsigned SuperClassKind;
295 /// SuperClass - The super class, or 0.
296 ClassInfo *SuperClass;
298 /// Name - The full class name, suitable for use in an enum.
301 /// ClassName - The unadorned generic name for this class (e.g., Token).
302 std::string ClassName;
304 /// ValueName - The name of the value this class represents; for a token this
305 /// is the literal token string, for an operand it is the TableGen class (or
306 /// empty if this is a derived class).
307 std::string ValueName;
309 /// PredicateMethod - The name of the operand method to test whether the
310 /// operand matches this class; this is not valid for Token kinds.
311 std::string PredicateMethod;
313 /// RenderMethod - The name of the operand method to add this operand to an
314 /// MCInst; this is not valid for Token kinds.
315 std::string RenderMethod;
317 /// isUserClass() - Check if this is a user defined class.
318 bool isUserClass() const {
319 return Kind >= UserClass0;
322 /// getRootClass - Return the root class of this one.
323 const ClassInfo *getRootClass() const {
324 const ClassInfo *CI = this;
325 while (CI->SuperClass)
330 /// operator< - Compare two classes.
331 bool operator<(const ClassInfo &RHS) const {
332 // Incompatible kinds are comparable for classes in disjoint hierarchies.
333 if (Kind != RHS.Kind && getRootClass() != RHS.getRootClass())
334 return Kind < RHS.Kind;
338 assert(0 && "Invalid kind!");
340 // Tokens are comparable by value.
342 // FIXME: Compare by enum value.
343 return ValueName < RHS.ValueName;
346 // This class preceeds the RHS if the RHS is a super class.
347 for (ClassInfo *Parent = SuperClass; Parent; Parent = Parent->SuperClass)
356 /// InstructionInfo - Helper class for storing the necessary information for an
357 /// instruction which is capable of being matched.
358 struct InstructionInfo {
360 /// The unique class instance this operand should match.
363 /// The original operand this corresponds to, if any.
364 const CodeGenInstruction::OperandInfo *OperandInfo;
367 /// InstrName - The target name for this instruction.
368 std::string InstrName;
370 /// Instr - The instruction this matches.
371 const CodeGenInstruction *Instr;
373 /// AsmString - The assembly string for this instruction (with variants
375 std::string AsmString;
377 /// Tokens - The tokenized assembly pattern that this instruction matches.
378 SmallVector<StringRef, 4> Tokens;
380 /// Operands - The operands that this instruction matches.
381 SmallVector<Operand, 4> Operands;
383 /// ConversionFnKind - The enum value which is passed to the generated
384 /// ConvertToMCInst to convert parsed operands into an MCInst for this
386 std::string ConversionFnKind;
388 /// operator< - Compare two instructions.
389 bool operator<(const InstructionInfo &RHS) const {
390 if (Operands.size() != RHS.Operands.size())
391 return Operands.size() < RHS.Operands.size();
393 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
394 if (*Operands[i].Class < *RHS.Operands[i].Class)
400 /// CouldMatchAmiguouslyWith - Check whether this instruction could
401 /// ambiguously match the same set of operands as \arg RHS (without being a
402 /// strictly superior match).
403 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
404 // The number of operands is unambiguous.
405 if (Operands.size() != RHS.Operands.size())
408 // Tokens and operand kinds are unambiguous (assuming a correct target
410 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
411 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
412 Operands[i].Class->Kind == ClassInfo::Token)
413 if (*Operands[i].Class < *RHS.Operands[i].Class ||
414 *RHS.Operands[i].Class < *Operands[i].Class)
417 // Otherwise, this operand could commute if all operands are equivalent, or
418 // there is a pair of operands that compare less than and a pair that
419 // compare greater than.
420 bool HasLT = false, HasGT = false;
421 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
422 if (*Operands[i].Class < *RHS.Operands[i].Class)
424 if (*RHS.Operands[i].Class < *Operands[i].Class)
428 return !(HasLT ^ HasGT);
435 class AsmMatcherInfo {
437 /// The classes which are needed for matching.
438 std::vector<ClassInfo*> Classes;
440 /// The information on the instruction to match.
441 std::vector<InstructionInfo*> Instructions;
444 /// Map of token to class information which has already been constructed.
445 std::map<std::string, ClassInfo*> TokenClasses;
447 /// Map of operand name to class information which has already been
449 std::map<std::string, ClassInfo*> OperandClasses;
451 /// Map of user class names to kind value.
452 std::map<std::string, unsigned> UserClasses;
455 /// getTokenClass - Lookup or create the class for the given token.
456 ClassInfo *getTokenClass(const StringRef &Token);
458 /// getUserClassKind - Lookup or create the kind value for the given class
460 unsigned getUserClassKind(const StringRef &Name);
462 /// getOperandClass - Lookup or create the class for the given operand.
463 ClassInfo *getOperandClass(const StringRef &Token,
464 const CodeGenInstruction::OperandInfo &OI);
467 /// BuildInfo - Construct the various tables used during matching.
468 void BuildInfo(CodeGenTarget &Target);
473 void InstructionInfo::dump() {
474 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
476 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
483 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
484 Operand &Op = Operands[i];
485 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
486 if (Op.Class->Kind == ClassInfo::Token) {
487 errs() << '\"' << Tokens[i] << "\"\n";
491 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
492 errs() << OI.Name << " " << OI.Rec->getName()
493 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
497 static std::string getEnumNameForToken(const StringRef &Str) {
500 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
502 case '*': Res += "_STAR_"; break;
503 case '%': Res += "_PCT_"; break;
504 case ':': Res += "_COLON_"; break;
510 Res += "_" + utostr((unsigned) *it) + "_";
518 ClassInfo *AsmMatcherInfo::getTokenClass(const StringRef &Token) {
519 ClassInfo *&Entry = TokenClasses[Token];
522 Entry = new ClassInfo();
523 Entry->Kind = ClassInfo::Token;
524 Entry->ClassName = "Token";
525 Entry->Name = "MCK_" + getEnumNameForToken(Token);
526 Entry->ValueName = Token;
527 Entry->PredicateMethod = "<invalid>";
528 Entry->RenderMethod = "<invalid>";
529 Classes.push_back(Entry);
535 unsigned AsmMatcherInfo::getUserClassKind(const StringRef &Name) {
536 unsigned &Entry = UserClasses[Name];
539 Entry = ClassInfo::UserClass0 + UserClasses.size() - 1;
545 AsmMatcherInfo::getOperandClass(const StringRef &Token,
546 const CodeGenInstruction::OperandInfo &OI) {
547 unsigned SuperClass = ClassInfo::Invalid;
548 std::string ClassName;
549 if (OI.Rec->isSubClassOf("RegisterClass")) {
553 ClassName = OI.Rec->getValueAsString("ParserMatchClass");
554 assert(ClassName != "Reg" && "'Reg' class name is reserved!");
556 PrintError(OI.Rec->getLoc(), "operand has no match class!");
557 ClassName = "Invalid";
560 // Determine the super class.
562 std::string SuperClassName =
563 OI.Rec->getValueAsString("ParserMatchSuperClass");
564 SuperClass = getUserClassKind(SuperClassName);
568 ClassInfo *&Entry = OperandClasses[ClassName];
571 Entry = new ClassInfo();
572 if (ClassName == "Reg") {
573 Entry->Kind = ClassInfo::Register;
575 Entry->Kind = getUserClassKind(ClassName);
576 Entry->SuperClassKind = SuperClass;
578 Entry->ClassName = ClassName;
579 Entry->Name = "MCK_" + ClassName;
580 Entry->ValueName = OI.Rec->getName();
581 Entry->PredicateMethod = "is" + ClassName;
582 Entry->RenderMethod = "add" + ClassName + "Operands";
583 Classes.push_back(Entry);
589 void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
590 for (std::map<std::string, CodeGenInstruction>::const_iterator
591 it = Target.getInstructions().begin(),
592 ie = Target.getInstructions().end();
594 const CodeGenInstruction &CGI = it->second;
596 if (!StringRef(it->first).startswith(MatchPrefix))
599 OwningPtr<InstructionInfo> II(new InstructionInfo);
601 II->InstrName = it->first;
602 II->Instr = &it->second;
603 II->AsmString = FlattenVariants(CGI.AsmString, 0);
605 TokenizeAsmString(II->AsmString, II->Tokens);
607 // Ignore instructions which shouldn't be matched.
608 if (!IsAssemblerInstruction(it->first, CGI, II->Tokens))
611 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
612 StringRef Token = II->Tokens[i];
614 // Check for simple tokens.
615 if (Token[0] != '$') {
616 InstructionInfo::Operand Op;
617 Op.Class = getTokenClass(Token);
619 II->Operands.push_back(Op);
623 // Otherwise this is an operand reference.
624 StringRef OperandName;
626 OperandName = Token.substr(2, Token.size() - 3);
628 OperandName = Token.substr(1);
630 // Map this token to an operand. FIXME: Move elsewhere.
633 Idx = CGI.getOperandNamed(OperandName);
635 errs() << "error: unable to find operand: '" << OperandName << "'!\n";
639 const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx];
640 InstructionInfo::Operand Op;
641 Op.Class = getOperandClass(Token, OI);
642 Op.OperandInfo = &OI;
643 II->Operands.push_back(Op);
646 // If we broke out, ignore the instruction.
647 if (II->Operands.size() != II->Tokens.size())
650 Instructions.push_back(II.take());
653 // Bind user super classes.
654 std::map<unsigned, ClassInfo*> UserClasses;
655 for (unsigned i = 0, e = Classes.size(); i != e; ++i) {
656 ClassInfo &CI = *Classes[i];
657 if (CI.isUserClass())
658 UserClasses[CI.Kind] = &CI;
661 for (unsigned i = 0, e = Classes.size(); i != e; ++i) {
662 ClassInfo &CI = *Classes[i];
663 if (CI.isUserClass() && CI.SuperClassKind != ClassInfo::Invalid) {
664 CI.SuperClass = UserClasses[CI.SuperClassKind];
665 assert(CI.SuperClass && "Missing super class definition!");
671 // Reorder classes so that classes preceed super classes.
672 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
675 static void EmitConvertToMCInst(CodeGenTarget &Target,
676 std::vector<InstructionInfo*> &Infos,
678 // Write the convert function to a separate stream, so we can drop it after
680 std::string ConvertFnBody;
681 raw_string_ostream CvtOS(ConvertFnBody);
683 // Function we have already generated.
684 std::set<std::string> GeneratedFns;
686 // Start the unified conversion function.
688 CvtOS << "static bool ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
689 << "unsigned Opcode,\n"
690 << " SmallVectorImpl<"
691 << Target.getName() << "Operand> &Operands) {\n";
692 CvtOS << " Inst.setOpcode(Opcode);\n";
693 CvtOS << " switch (Kind) {\n";
694 CvtOS << " default:\n";
696 // Start the enum, which we will generate inline.
698 OS << "// Unified function for converting operants to MCInst instances.\n\n";
699 OS << "enum ConversionKind {\n";
701 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
702 ie = Infos.end(); it != ie; ++it) {
703 InstructionInfo &II = **it;
705 // Order the (class) operands by the order to convert them into an MCInst.
706 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
707 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
708 InstructionInfo::Operand &Op = II.Operands[i];
710 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
712 std::sort(MIOperandList.begin(), MIOperandList.end());
714 // Compute the total number of operands.
715 unsigned NumMIOperands = 0;
716 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
717 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i];
718 NumMIOperands = std::max(NumMIOperands,
719 OI.MIOperandNo + OI.MINumOperands);
722 // Build the conversion function signature.
723 std::string Signature = "Convert";
724 unsigned CurIndex = 0;
725 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
726 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
727 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
728 "Duplicate match for instruction operand!");
732 // Skip operands which weren't matched by anything, this occurs when the
733 // .td file encodes "implicit" operands as explicit ones.
735 // FIXME: This should be removed from the MCInst structure.
736 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
739 Signature += Op.Class->ClassName;
740 Signature += utostr(Op.OperandInfo->MINumOperands);
741 Signature += "_" + utostr(MIOperandList[i].second);
743 CurIndex += Op.OperandInfo->MINumOperands;
746 // Add any trailing implicit operands.
747 for (; CurIndex != NumMIOperands; ++CurIndex)
750 II.ConversionFnKind = Signature;
752 // Check if we have already generated this signature.
753 if (!GeneratedFns.insert(Signature).second)
756 // If not, emit it now.
758 // Add to the enum list.
759 OS << " " << Signature << ",\n";
761 // And to the convert function.
762 CvtOS << " case " << Signature << ":\n";
764 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
765 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
767 // Add the implicit operands.
768 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex)
769 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
771 CvtOS << " Operands[" << MIOperandList[i].second
772 << "]." << Op.Class->RenderMethod
773 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
774 CurIndex += Op.OperandInfo->MINumOperands;
777 // And add trailing implicit operands.
778 for (; CurIndex != NumMIOperands; ++CurIndex)
779 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
780 CvtOS << " break;\n";
783 // Finish the convert function.
786 CvtOS << " return false;\n";
789 // Finish the enum, and drop the convert function after it.
791 OS << " NumConversionVariants\n";
797 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
798 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
799 std::vector<ClassInfo*> &Infos,
801 OS << "namespace {\n\n";
803 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
804 << "/// instruction matching.\n";
805 OS << "enum MatchClassKind {\n";
806 OS << " InvalidMatchClass = 0,\n";
807 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
808 ie = Infos.end(); it != ie; ++it) {
809 ClassInfo &CI = **it;
810 OS << " " << CI.Name << ", // ";
811 if (CI.Kind == ClassInfo::Token) {
812 OS << "'" << CI.ValueName << "'\n";
813 } else if (CI.Kind == ClassInfo::Register) {
814 if (!CI.ValueName.empty())
815 OS << "register class '" << CI.ValueName << "'\n";
817 OS << "derived register class\n";
819 OS << "user defined class '" << CI.ValueName << "'\n";
822 OS << " NumMatchClassKinds\n";
828 /// EmitClassifyOperand - Emit the function to classify an operand.
829 static void EmitClassifyOperand(CodeGenTarget &Target,
830 std::vector<ClassInfo*> &Infos,
832 OS << "static MatchClassKind ClassifyOperand("
833 << Target.getName() << "Operand &Operand) {\n";
834 OS << " if (Operand.isToken())\n";
835 OS << " return MatchTokenString(Operand.getToken());\n\n";
836 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
837 ie = Infos.end(); it != ie; ++it) {
838 ClassInfo &CI = **it;
840 if (CI.Kind != ClassInfo::Token) {
841 OS << " // '" << CI.ClassName << "' class";
843 OS << ", subclass of '" << CI.SuperClass->ClassName << "'";
844 assert(CI < *CI.SuperClass && "Invalid class relation!");
848 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
850 // Validate subclass relationships.
852 OS << " assert(Operand." << CI.SuperClass->PredicateMethod
853 << "() && \"Invalid class relationship!\");\n";
855 OS << " return " << CI.Name << ";\n\n";
859 OS << " return InvalidMatchClass;\n";
863 typedef std::pair<std::string, std::string> StringPair;
865 /// FindFirstNonCommonLetter - Find the first character in the keys of the
866 /// string pairs that is not shared across the whole set of strings. All
867 /// strings are assumed to have the same length.
869 FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) {
870 assert(!Matches.empty());
871 for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) {
872 // Check to see if letter i is the same across the set.
873 char Letter = Matches[0]->first[i];
875 for (unsigned str = 0, e = Matches.size(); str != e; ++str)
876 if (Matches[str]->first[i] != Letter)
880 return Matches[0]->first.size();
883 /// EmitStringMatcherForChar - Given a set of strings that are known to be the
884 /// same length and whose characters leading up to CharNo are the same, emit
885 /// code to verify that CharNo and later are the same.
887 /// \return - True if control can leave the emitted code fragment.
888 static bool EmitStringMatcherForChar(const std::string &StrVariableName,
889 const std::vector<const StringPair*> &Matches,
890 unsigned CharNo, unsigned IndentCount,
892 assert(!Matches.empty() && "Must have at least one string to match!");
893 std::string Indent(IndentCount*2+4, ' ');
895 // If we have verified that the entire string matches, we're done: output the
897 if (CharNo == Matches[0]->first.size()) {
898 assert(Matches.size() == 1 && "Had duplicate keys to match on");
900 // FIXME: If Matches[0].first has embeded \n, this will be bad.
901 OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first
906 // Bucket the matches by the character we are comparing.
907 std::map<char, std::vector<const StringPair*> > MatchesByLetter;
909 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
910 MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]);
913 // If we have exactly one bucket to match, see how many characters are common
914 // across the whole set and match all of them at once.
915 if (MatchesByLetter.size() == 1) {
916 unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches);
917 unsigned NumChars = FirstNonCommonLetter-CharNo;
919 // Emit code to break out if the prefix doesn't match.
921 // Do the comparison with if (Str[1] != 'f')
922 // FIXME: Need to escape general characters.
923 OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '"
924 << Matches[0]->first[CharNo] << "')\n";
925 OS << Indent << " break;\n";
927 // Do the comparison with if (Str.substr(1,3) != "foo").
928 // FIXME: Need to escape general strings.
929 OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << ","
930 << NumChars << ") != \"";
931 OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n";
932 OS << Indent << " break;\n";
935 return EmitStringMatcherForChar(StrVariableName, Matches,
936 FirstNonCommonLetter, IndentCount, OS);
939 // Otherwise, we have multiple possible things, emit a switch on the
941 OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n";
942 OS << Indent << "default: break;\n";
944 for (std::map<char, std::vector<const StringPair*> >::iterator LI =
945 MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) {
946 // TODO: escape hard stuff (like \n) if we ever care about it.
947 OS << Indent << "case '" << LI->first << "':\t // "
948 << LI->second.size() << " strings to match.\n";
949 if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1,
951 OS << Indent << " break;\n";
954 OS << Indent << "}\n";
959 /// EmitStringMatcher - Given a list of strings and code to execute when they
960 /// match, output a simple switch tree to classify the input string.
962 /// If a match is found, the code in Vals[i].second is executed; control must
963 /// not exit this code fragment. If nothing matches, execution falls through.
965 /// \param StrVariableName - The name of the variable to test.
966 static void EmitStringMatcher(const std::string &StrVariableName,
967 const std::vector<StringPair> &Matches,
969 // First level categorization: group strings by length.
970 std::map<unsigned, std::vector<const StringPair*> > MatchesByLength;
972 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
973 MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]);
975 // Output a switch statement on length and categorize the elements within each
977 OS << " switch (" << StrVariableName << ".size()) {\n";
978 OS << " default: break;\n";
980 for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI =
981 MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) {
982 OS << " case " << LI->first << ":\t // " << LI->second.size()
983 << " strings to match.\n";
984 if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS))
992 /// EmitMatchTokenString - Emit the function to match a token string to the
993 /// appropriate match class value.
994 static void EmitMatchTokenString(CodeGenTarget &Target,
995 std::vector<ClassInfo*> &Infos,
997 // Construct the match list.
998 std::vector<StringPair> Matches;
999 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1000 ie = Infos.end(); it != ie; ++it) {
1001 ClassInfo &CI = **it;
1003 if (CI.Kind == ClassInfo::Token)
1004 Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";"));
1007 OS << "static MatchClassKind MatchTokenString(const StringRef &Name) {\n";
1009 EmitStringMatcher("Name", Matches, OS);
1011 OS << " return InvalidMatchClass;\n";
1015 /// EmitMatchRegisterName - Emit the function to match a string to the target
1016 /// specific register enum.
1017 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1019 // Construct the match list.
1020 std::vector<StringPair> Matches;
1021 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1022 const CodeGenRegister &Reg = Target.getRegisters()[i];
1023 if (Reg.TheDef->getValueAsString("AsmName").empty())
1026 Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"),
1027 "return " + utostr(i + 1) + ";"));
1030 OS << "unsigned " << Target.getName()
1031 << AsmParser->getValueAsString("AsmParserClassName")
1032 << "::MatchRegisterName(const StringRef &Name) {\n";
1034 EmitStringMatcher("Name", Matches, OS);
1036 OS << " return 0;\n";
1040 void AsmMatcherEmitter::run(raw_ostream &OS) {
1041 CodeGenTarget Target;
1042 Record *AsmParser = Target.getAsmParser();
1043 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1045 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1047 // Emit the function to match a register name to number.
1048 EmitMatchRegisterName(Target, AsmParser, OS);
1050 // Compute the information on the instructions to match.
1051 AsmMatcherInfo Info;
1052 Info.BuildInfo(Target);
1054 // Sort the instruction table using the partial order on classes.
1055 std::sort(Info.Instructions.begin(), Info.Instructions.end(),
1056 less_ptr<InstructionInfo>());
1058 DEBUG_WITH_TYPE("instruction_info", {
1059 for (std::vector<InstructionInfo*>::iterator
1060 it = Info.Instructions.begin(), ie = Info.Instructions.end();
1065 // Check for ambiguous instructions.
1066 unsigned NumAmbiguous = 0;
1067 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
1068 for (unsigned j = i + 1; j != e; ++j) {
1069 InstructionInfo &A = *Info.Instructions[i];
1070 InstructionInfo &B = *Info.Instructions[j];
1072 if (A.CouldMatchAmiguouslyWith(B)) {
1073 DEBUG_WITH_TYPE("ambiguous_instrs", {
1074 errs() << "warning: ambiguous instruction match:\n";
1076 errs() << "\nis incomparable with:\n";
1085 DEBUG_WITH_TYPE("ambiguous_instrs", {
1086 errs() << "warning: " << NumAmbiguous
1087 << " ambiguous instructions!\n";
1090 // Generate the unified function to convert operands into an MCInst.
1091 EmitConvertToMCInst(Target, Info.Instructions, OS);
1093 // Emit the enumeration for classes which participate in matching.
1094 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1096 // Emit the routine to match token strings to their match class.
1097 EmitMatchTokenString(Target, Info.Classes, OS);
1099 // Emit the routine to classify an operand.
1100 EmitClassifyOperand(Target, Info.Classes, OS);
1102 // Finally, build the match function.
1104 size_t MaxNumOperands = 0;
1105 for (std::vector<InstructionInfo*>::const_iterator it =
1106 Info.Instructions.begin(), ie = Info.Instructions.end();
1108 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
1110 OS << "bool " << Target.getName() << ClassName
1111 << "::MatchInstruction("
1112 << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, "
1113 << "MCInst &Inst) {\n";
1115 // Emit the static match table; unused classes get initalized to 0 which is
1116 // guaranteed to be InvalidMatchClass.
1118 // FIXME: We can reduce the size of this table very easily. First, we change
1119 // it so that store the kinds in separate bit-fields for each index, which
1120 // only needs to be the max width used for classes at that index (we also need
1121 // to reject based on this during classification). If we then make sure to
1122 // order the match kinds appropriately (putting mnemonics last), then we
1123 // should only end up using a few bits for each class, especially the ones
1124 // following the mnemonic.
1125 OS << " static const struct MatchEntry {\n";
1126 OS << " unsigned Opcode;\n";
1127 OS << " ConversionKind ConvertFn;\n";
1128 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1129 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n";
1131 for (std::vector<InstructionInfo*>::const_iterator it =
1132 Info.Instructions.begin(), ie = Info.Instructions.end();
1134 InstructionInfo &II = **it;
1136 OS << " { " << Target.getName() << "::" << II.InstrName
1137 << ", " << II.ConversionFnKind << ", { ";
1138 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1139 InstructionInfo::Operand &Op = II.Operands[i];
1142 OS << Op.Class->Name;
1149 // Emit code to compute the class list for this operand vector.
1150 OS << " // Eliminate obvious mismatches.\n";
1151 OS << " if (Operands.size() > " << MaxNumOperands << ")\n";
1152 OS << " return true;\n\n";
1154 OS << " // Compute the class list for this operand vector.\n";
1155 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1156 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n";
1157 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n";
1159 OS << " // Check for invalid operands before matching.\n";
1160 OS << " if (Classes[i] == InvalidMatchClass)\n";
1161 OS << " return true;\n";
1164 OS << " // Mark unused classes.\n";
1165 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; "
1166 << "i != e; ++i)\n";
1167 OS << " Classes[i] = InvalidMatchClass;\n\n";
1169 // Emit code to search the table.
1170 OS << " // Search the table.\n";
1171 OS << " for (const MatchEntry *it = MatchTable, "
1172 << "*ie = MatchTable + " << Info.Instructions.size()
1173 << "; it != ie; ++it) {\n";
1174 for (unsigned i = 0; i != MaxNumOperands; ++i) {
1175 OS << " if (Classes[" << i << "] != it->Classes[" << i << "])\n";
1176 OS << " continue;\n";
1179 OS << " return ConvertToMCInst(it->ConvertFn, Inst, "
1180 << "it->Opcode, Operands);\n";
1183 OS << " return true;\n";