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. It also emits a matcher for
12 // custom operand parsing.
14 // Converting assembly operands into MCInst structures
15 // ---------------------------------------------------
17 // The input to the target specific matcher is a list of literal tokens and
18 // operands. The target specific parser should generally eliminate any syntax
19 // which is not relevant for matching; for example, comma tokens should have
20 // already been consumed and eliminated by the parser. Most instructions will
21 // end up with a single literal token (the instruction name) and some number of
24 // Some example inputs, for X86:
25 // 'addl' (immediate ...) (register ...)
26 // 'add' (immediate ...) (memory ...)
29 // The assembly matcher is responsible for converting this input into a precise
30 // machine instruction (i.e., an instruction with a well defined encoding). This
31 // mapping has several properties which complicate matching:
33 // - It may be ambiguous; many architectures can legally encode particular
34 // variants of an instruction in different ways (for example, using a smaller
35 // encoding for small immediates). Such ambiguities should never be
36 // arbitrarily resolved by the assembler, the assembler is always responsible
37 // for choosing the "best" available instruction.
39 // - It may depend on the subtarget or the assembler context. Instructions
40 // which are invalid for the current mode, but otherwise unambiguous (e.g.,
41 // an SSE instruction in a file being assembled for i486) should be accepted
42 // and rejected by the assembler front end. However, if the proper encoding
43 // for an instruction is dependent on the assembler context then the matcher
44 // is responsible for selecting the correct machine instruction for the
47 // The core matching algorithm attempts to exploit the regularity in most
48 // instruction sets to quickly determine the set of possibly matching
49 // instructions, and the simplify the generated code. Additionally, this helps
50 // to ensure that the ambiguities are intentionally resolved by the user.
52 // The matching is divided into two distinct phases:
54 // 1. Classification: Each operand is mapped to the unique set which (a)
55 // contains it, and (b) is the largest such subset for which a single
56 // instruction could match all members.
58 // For register classes, we can generate these subgroups automatically. For
59 // arbitrary operands, we expect the user to define the classes and their
60 // relations to one another (for example, 8-bit signed immediates as a
61 // subset of 32-bit immediates).
63 // By partitioning the operands in this way, we guarantee that for any
64 // tuple of classes, any single instruction must match either all or none
65 // of the sets of operands which could classify to that tuple.
67 // In addition, the subset relation amongst classes induces a partial order
68 // on such tuples, which we use to resolve ambiguities.
70 // 2. The input can now be treated as a tuple of classes (static tokens are
71 // simple singleton sets). Each such tuple should generally map to a single
72 // instruction (we currently ignore cases where this isn't true, whee!!!),
73 // which we can emit a simple matcher for.
75 // Custom Operand Parsing
76 // ----------------------
78 // Some targets need a custom way to parse operands, some specific instructions
79 // can contain arguments that can represent processor flags and other kinds of
80 // identifiers that need to be mapped to specific values in the final encoded
81 // instructions. The target specific custom operand parsing works in the
84 // 1. A operand match table is built, each entry contains a mnemonic, an
85 // operand class, a mask for all operand positions for that same
86 // class/mnemonic and target features to be checked while trying to match.
88 // 2. The operand matcher will try every possible entry with the same
89 // mnemonic and will check if the target feature for this mnemonic also
90 // matches. After that, if the operand to be matched has its index
91 // present in the mask, a successful match occurs. Otherwise, fallback
92 // to the regular operand parsing.
94 // 3. For a match success, each operand class that has a 'ParserMethod'
95 // becomes part of a switch from where the custom method is called.
97 //===----------------------------------------------------------------------===//
99 #include "CodeGenTarget.h"
100 #include "llvm/ADT/PointerUnion.h"
101 #include "llvm/ADT/STLExtras.h"
102 #include "llvm/ADT/SmallPtrSet.h"
103 #include "llvm/ADT/SmallVector.h"
104 #include "llvm/ADT/StringExtras.h"
105 #include "llvm/Support/CommandLine.h"
106 #include "llvm/Support/Debug.h"
107 #include "llvm/Support/ErrorHandling.h"
108 #include "llvm/TableGen/Error.h"
109 #include "llvm/TableGen/Record.h"
110 #include "llvm/TableGen/StringMatcher.h"
111 #include "llvm/TableGen/StringToOffsetTable.h"
112 #include "llvm/TableGen/TableGenBackend.h"
118 #include <forward_list>
119 using namespace llvm;
121 #define DEBUG_TYPE "asm-matcher-emitter"
123 static cl::opt<std::string>
124 MatchPrefix("match-prefix", cl::init(""),
125 cl::desc("Only match instructions with the given prefix"));
128 class AsmMatcherInfo;
129 struct SubtargetFeatureInfo;
131 // Register sets are used as keys in some second-order sets TableGen creates
132 // when generating its data structures. This means that the order of two
133 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and
134 // can even affect compiler output (at least seen in diagnostics produced when
135 // all matches fail). So we use a type that sorts them consistently.
136 typedef std::set<Record*, LessRecordByID> RegisterSet;
138 class AsmMatcherEmitter {
139 RecordKeeper &Records;
141 AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
143 void run(raw_ostream &o);
146 /// ClassInfo - Helper class for storing the information about a particular
147 /// class of operands which can be matched.
150 /// Invalid kind, for use as a sentinel value.
153 /// The class for a particular token.
156 /// The (first) register class, subsequent register classes are
157 /// RegisterClass0+1, and so on.
160 /// The (first) user defined class, subsequent user defined classes are
161 /// UserClass0+1, and so on.
165 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
166 /// N) for the Nth user defined class.
169 /// SuperClasses - The super classes of this class. Note that for simplicities
170 /// sake user operands only record their immediate super class, while register
171 /// operands include all superclasses.
172 std::vector<ClassInfo*> SuperClasses;
174 /// Name - The full class name, suitable for use in an enum.
177 /// ClassName - The unadorned generic name for this class (e.g., Token).
178 std::string ClassName;
180 /// ValueName - The name of the value this class represents; for a token this
181 /// is the literal token string, for an operand it is the TableGen class (or
182 /// empty if this is a derived class).
183 std::string ValueName;
185 /// PredicateMethod - The name of the operand method to test whether the
186 /// operand matches this class; this is not valid for Token or register kinds.
187 std::string PredicateMethod;
189 /// RenderMethod - The name of the operand method to add this operand to an
190 /// MCInst; this is not valid for Token or register kinds.
191 std::string RenderMethod;
193 /// ParserMethod - The name of the operand method to do a target specific
194 /// parsing on the operand.
195 std::string ParserMethod;
197 /// For register classes: the records for all the registers in this class.
198 RegisterSet Registers;
200 /// For custom match classes: the diagnostic kind for when the predicate fails.
201 std::string DiagnosticType;
203 /// isRegisterClass() - Check if this is a register class.
204 bool isRegisterClass() const {
205 return Kind >= RegisterClass0 && Kind < UserClass0;
208 /// isUserClass() - Check if this is a user defined class.
209 bool isUserClass() const {
210 return Kind >= UserClass0;
213 /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
214 /// are related if they are in the same class hierarchy.
215 bool isRelatedTo(const ClassInfo &RHS) const {
216 // Tokens are only related to tokens.
217 if (Kind == Token || RHS.Kind == Token)
218 return Kind == Token && RHS.Kind == Token;
220 // Registers classes are only related to registers classes, and only if
221 // their intersection is non-empty.
222 if (isRegisterClass() || RHS.isRegisterClass()) {
223 if (!isRegisterClass() || !RHS.isRegisterClass())
227 std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin());
228 std::set_intersection(Registers.begin(), Registers.end(),
229 RHS.Registers.begin(), RHS.Registers.end(),
230 II, LessRecordByID());
235 // Otherwise we have two users operands; they are related if they are in the
236 // same class hierarchy.
238 // FIXME: This is an oversimplification, they should only be related if they
239 // intersect, however we don't have that information.
240 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
241 const ClassInfo *Root = this;
242 while (!Root->SuperClasses.empty())
243 Root = Root->SuperClasses.front();
245 const ClassInfo *RHSRoot = &RHS;
246 while (!RHSRoot->SuperClasses.empty())
247 RHSRoot = RHSRoot->SuperClasses.front();
249 return Root == RHSRoot;
252 /// isSubsetOf - Test whether this class is a subset of \p RHS.
253 bool isSubsetOf(const ClassInfo &RHS) const {
254 // This is a subset of RHS if it is the same class...
258 // ... or if any of its super classes are a subset of RHS.
259 for (const ClassInfo *CI : SuperClasses)
260 if (CI->isSubsetOf(RHS))
266 /// operator< - Compare two classes.
267 bool operator<(const ClassInfo &RHS) const {
271 // Unrelated classes can be ordered by kind.
272 if (!isRelatedTo(RHS))
273 return Kind < RHS.Kind;
277 llvm_unreachable("Invalid kind!");
280 // This class precedes the RHS if it is a proper subset of the RHS.
283 if (RHS.isSubsetOf(*this))
286 // Otherwise, order by name to ensure we have a total ordering.
287 return ValueName < RHS.ValueName;
292 /// MatchableInfo - Helper class for storing the necessary information for an
293 /// instruction or alias which is capable of being matched.
294 struct MatchableInfo {
296 /// Token - This is the token that the operand came from.
299 /// The unique class instance this operand should match.
302 /// The operand name this is, if anything.
305 /// The suboperand index within SrcOpName, or -1 for the entire operand.
308 /// Register record if this token is singleton register.
309 Record *SingletonReg;
311 explicit AsmOperand(StringRef T) : Token(T), Class(nullptr), SubOpIdx(-1),
312 SingletonReg(nullptr) {}
315 /// ResOperand - This represents a single operand in the result instruction
316 /// generated by the match. In cases (like addressing modes) where a single
317 /// assembler operand expands to multiple MCOperands, this represents the
318 /// single assembler operand, not the MCOperand.
321 /// RenderAsmOperand - This represents an operand result that is
322 /// generated by calling the render method on the assembly operand. The
323 /// corresponding AsmOperand is specified by AsmOperandNum.
326 /// TiedOperand - This represents a result operand that is a duplicate of
327 /// a previous result operand.
330 /// ImmOperand - This represents an immediate value that is dumped into
334 /// RegOperand - This represents a fixed register that is dumped in.
339 /// This is the operand # in the AsmOperands list that this should be
341 unsigned AsmOperandNum;
343 /// TiedOperandNum - This is the (earlier) result operand that should be
345 unsigned TiedOperandNum;
347 /// ImmVal - This is the immediate value added to the instruction.
350 /// Register - This is the register record.
354 /// MINumOperands - The number of MCInst operands populated by this
356 unsigned MINumOperands;
358 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
360 X.Kind = RenderAsmOperand;
361 X.AsmOperandNum = AsmOpNum;
362 X.MINumOperands = NumOperands;
366 static ResOperand getTiedOp(unsigned TiedOperandNum) {
368 X.Kind = TiedOperand;
369 X.TiedOperandNum = TiedOperandNum;
374 static ResOperand getImmOp(int64_t Val) {
382 static ResOperand getRegOp(Record *Reg) {
391 /// AsmVariantID - Target's assembly syntax variant no.
394 /// TheDef - This is the definition of the instruction or InstAlias that this
395 /// matchable came from.
396 Record *const TheDef;
398 /// AsmString - The assembly string for this instruction (with variants
399 /// removed), e.g. "movsx $src, $dst".
400 std::string AsmString;
402 /// DefRec - This is the definition that it came from.
403 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
405 const CodeGenInstruction *getResultInst() const {
406 if (DefRec.is<const CodeGenInstruction*>())
407 return DefRec.get<const CodeGenInstruction*>();
408 return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
411 /// ResOperands - This is the operand list that should be built for the result
413 SmallVector<ResOperand, 8> ResOperands;
415 /// Mnemonic - This is the first token of the matched instruction, its
419 /// AsmOperands - The textual operands that this instruction matches,
420 /// annotated with a class and where in the OperandList they were defined.
421 /// This directly corresponds to the tokenized AsmString after the mnemonic is
423 SmallVector<AsmOperand, 8> AsmOperands;
425 /// Predicates - The required subtarget features to match this instruction.
426 SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
428 /// ConversionFnKind - The enum value which is passed to the generated
429 /// convertToMCInst to convert parsed operands into an MCInst for this
431 std::string ConversionFnKind;
433 /// If this instruction is deprecated in some form.
436 MatchableInfo(const CodeGenInstruction &CGI)
437 : AsmVariantID(0), TheDef(CGI.TheDef), AsmString(CGI.AsmString),
440 MatchableInfo(std::unique_ptr<CodeGenInstAlias> Alias)
441 : AsmVariantID(0), TheDef(Alias->TheDef), AsmString(Alias->AsmString),
442 DefRec(Alias.release()) {}
444 ~MatchableInfo() { delete DefRec.dyn_cast<const CodeGenInstAlias *>(); }
446 // Two-operand aliases clone from the main matchable, but mark the second
447 // operand as a tied operand of the first for purposes of the assembler.
448 void formTwoOperandAlias(StringRef Constraint);
450 void initialize(const AsmMatcherInfo &Info,
451 SmallPtrSetImpl<Record*> &SingletonRegisters,
452 int AsmVariantNo, std::string &RegisterPrefix);
454 /// validate - Return true if this matchable is a valid thing to match against
455 /// and perform a bunch of validity checking.
456 bool validate(StringRef CommentDelimiter, bool Hack) const;
458 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
459 /// if present, from specified token.
461 extractSingletonRegisterForAsmOperand(unsigned i, const AsmMatcherInfo &Info,
462 std::string &RegisterPrefix);
464 /// findAsmOperand - Find the AsmOperand with the specified name and
465 /// suboperand index.
466 int findAsmOperand(StringRef N, int SubOpIdx) const {
467 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
468 if (N == AsmOperands[i].SrcOpName &&
469 SubOpIdx == AsmOperands[i].SubOpIdx)
474 /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
475 /// This does not check the suboperand index.
476 int findAsmOperandNamed(StringRef N) const {
477 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
478 if (N == AsmOperands[i].SrcOpName)
483 void buildInstructionResultOperands();
484 void buildAliasResultOperands();
486 /// operator< - Compare two matchables.
487 bool operator<(const MatchableInfo &RHS) const {
488 // The primary comparator is the instruction mnemonic.
489 if (Mnemonic != RHS.Mnemonic)
490 return Mnemonic < RHS.Mnemonic;
492 if (AsmOperands.size() != RHS.AsmOperands.size())
493 return AsmOperands.size() < RHS.AsmOperands.size();
495 // Compare lexicographically by operand. The matcher validates that other
496 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
497 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
498 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
500 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
504 // Give matches that require more features higher precedence. This is useful
505 // because we cannot define AssemblerPredicates with the negation of
506 // processor features. For example, ARM v6 "nop" may be either a HINT or
507 // MOV. With v6, we want to match HINT. The assembler has no way to
508 // predicate MOV under "NoV6", but HINT will always match first because it
509 // requires V6 while MOV does not.
510 if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
511 return RequiredFeatures.size() > RHS.RequiredFeatures.size();
516 /// couldMatchAmbiguouslyWith - Check whether this matchable could
517 /// ambiguously match the same set of operands as \p RHS (without being a
518 /// strictly superior match).
519 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const {
520 // The primary comparator is the instruction mnemonic.
521 if (Mnemonic != RHS.Mnemonic)
524 // The number of operands is unambiguous.
525 if (AsmOperands.size() != RHS.AsmOperands.size())
528 // Otherwise, make sure the ordering of the two instructions is unambiguous
529 // by checking that either (a) a token or operand kind discriminates them,
530 // or (b) the ordering among equivalent kinds is consistent.
532 // Tokens and operand kinds are unambiguous (assuming a correct target
534 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
535 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
536 AsmOperands[i].Class->Kind == ClassInfo::Token)
537 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
538 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
541 // Otherwise, this operand could commute if all operands are equivalent, or
542 // there is a pair of operands that compare less than and a pair that
543 // compare greater than.
544 bool HasLT = false, HasGT = false;
545 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
546 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
548 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
552 return !(HasLT ^ HasGT);
558 void tokenizeAsmString(const AsmMatcherInfo &Info);
561 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
562 /// feature which participates in instruction matching.
563 struct SubtargetFeatureInfo {
564 /// \brief The predicate record for this feature.
567 /// \brief An unique index assigned to represent this feature.
570 SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {}
572 /// \brief The name of the enumerated constant identifying this feature.
573 std::string getEnumName() const {
574 return "Feature_" + TheDef->getName();
578 errs() << getEnumName() << " " << Index << "\n";
583 struct OperandMatchEntry {
584 unsigned OperandMask;
585 const MatchableInfo* MI;
588 static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
591 X.OperandMask = opMask;
599 class AsmMatcherInfo {
602 RecordKeeper &Records;
604 /// The tablegen AsmParser record.
607 /// Target - The target information.
608 CodeGenTarget &Target;
610 /// The classes which are needed for matching.
611 std::forward_list<ClassInfo> Classes;
613 /// The information on the matchables to match.
614 std::forward_list<MatchableInfo> Matchables;
616 /// Info for custom matching operands by user defined methods.
617 std::vector<OperandMatchEntry> OperandMatchInfo;
619 /// Map of Register records to their class information.
620 typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy;
621 RegisterClassesTy RegisterClasses;
623 /// Map of Predicate records to their subtarget information.
624 std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
626 /// Map of AsmOperandClass records to their class information.
627 std::map<Record*, ClassInfo*> AsmOperandClasses;
630 /// Map of token to class information which has already been constructed.
631 std::map<std::string, ClassInfo*> TokenClasses;
633 /// Map of RegisterClass records to their class information.
634 std::map<Record*, ClassInfo*> RegisterClassClasses;
637 /// getTokenClass - Lookup or create the class for the given token.
638 ClassInfo *getTokenClass(StringRef Token);
640 /// getOperandClass - Lookup or create the class for the given operand.
641 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
643 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
645 /// buildRegisterClasses - Build the ClassInfo* instances for register
647 void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters);
649 /// buildOperandClasses - Build the ClassInfo* instances for user defined
651 void buildOperandClasses();
653 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
655 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
656 MatchableInfo::AsmOperand &Op);
659 AsmMatcherInfo(Record *AsmParser,
660 CodeGenTarget &Target,
661 RecordKeeper &Records);
663 /// buildInfo - Construct the various tables used during matching.
666 /// buildOperandMatchInfo - Build the necessary information to handle user
667 /// defined operand parsing methods.
668 void buildOperandMatchInfo();
670 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
672 const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
673 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
674 const auto &I = SubtargetFeatures.find(Def);
675 return I == SubtargetFeatures.end() ? nullptr : &I->second;
678 RecordKeeper &getRecords() const {
683 } // End anonymous namespace
685 void MatchableInfo::dump() const {
686 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
688 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
689 const AsmOperand &Op = AsmOperands[i];
690 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
691 errs() << '\"' << Op.Token << "\"\n";
695 static std::pair<StringRef, StringRef>
696 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
697 // Split via the '='.
698 std::pair<StringRef, StringRef> Ops = S.split('=');
699 if (Ops.second == "")
700 PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
701 // Trim whitespace and the leading '$' on the operand names.
702 size_t start = Ops.first.find_first_of('$');
703 if (start == std::string::npos)
704 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
705 Ops.first = Ops.first.slice(start + 1, std::string::npos);
706 size_t end = Ops.first.find_last_of(" \t");
707 Ops.first = Ops.first.slice(0, end);
708 // Now the second operand.
709 start = Ops.second.find_first_of('$');
710 if (start == std::string::npos)
711 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
712 Ops.second = Ops.second.slice(start + 1, std::string::npos);
713 end = Ops.second.find_last_of(" \t");
714 Ops.first = Ops.first.slice(0, end);
718 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
719 // Figure out which operands are aliased and mark them as tied.
720 std::pair<StringRef, StringRef> Ops =
721 parseTwoOperandConstraint(Constraint, TheDef->getLoc());
723 // Find the AsmOperands that refer to the operands we're aliasing.
724 int SrcAsmOperand = findAsmOperandNamed(Ops.first);
725 int DstAsmOperand = findAsmOperandNamed(Ops.second);
726 if (SrcAsmOperand == -1)
727 PrintFatalError(TheDef->getLoc(),
728 "unknown source two-operand alias operand '" + Ops.first +
730 if (DstAsmOperand == -1)
731 PrintFatalError(TheDef->getLoc(),
732 "unknown destination two-operand alias operand '" +
735 // Find the ResOperand that refers to the operand we're aliasing away
736 // and update it to refer to the combined operand instead.
737 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
738 ResOperand &Op = ResOperands[i];
739 if (Op.Kind == ResOperand::RenderAsmOperand &&
740 Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
741 Op.AsmOperandNum = DstAsmOperand;
745 // Remove the AsmOperand for the alias operand.
746 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
747 // Adjust the ResOperand references to any AsmOperands that followed
748 // the one we just deleted.
749 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
750 ResOperand &Op = ResOperands[i];
753 // Nothing to do for operands that don't reference AsmOperands.
755 case ResOperand::RenderAsmOperand:
756 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
759 case ResOperand::TiedOperand:
760 if (Op.TiedOperandNum > (unsigned)SrcAsmOperand)
767 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
768 SmallPtrSetImpl<Record*> &SingletonRegisters,
769 int AsmVariantNo, std::string &RegisterPrefix) {
770 AsmVariantID = AsmVariantNo;
772 CodeGenInstruction::FlattenAsmStringVariants(AsmString, AsmVariantNo);
774 tokenizeAsmString(Info);
776 // Compute the require features.
777 std::vector<Record*> Predicates =TheDef->getValueAsListOfDefs("Predicates");
778 for (unsigned i = 0, e = Predicates.size(); i != e; ++i)
779 if (const SubtargetFeatureInfo *Feature =
780 Info.getSubtargetFeature(Predicates[i]))
781 RequiredFeatures.push_back(Feature);
783 // Collect singleton registers, if used.
784 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
785 extractSingletonRegisterForAsmOperand(i, Info, RegisterPrefix);
786 if (Record *Reg = AsmOperands[i].SingletonReg)
787 SingletonRegisters.insert(Reg);
790 const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
792 DepMask = TheDef->getValue("ComplexDeprecationPredicate");
795 DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
798 /// tokenizeAsmString - Tokenize a simplified assembly string.
799 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info) {
800 StringRef String = AsmString;
803 for (unsigned i = 0, e = String.size(); i != e; ++i) {
813 AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
816 if (!isspace(String[i]) && String[i] != ',')
817 AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
823 AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
827 assert(i != String.size() && "Invalid quoted character");
828 AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
834 AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
838 // If this isn't "${", treat like a normal token.
839 if (i + 1 == String.size() || String[i + 1] != '{') {
844 StringRef::iterator End = std::find(String.begin() + i, String.end(),'}');
845 assert(End != String.end() && "Missing brace in operand reference!");
846 size_t EndPos = End - String.begin();
847 AsmOperands.push_back(AsmOperand(String.slice(i, EndPos+1)));
854 if (!Info.AsmParser->getValueAsBit("MnemonicContainsDot")) {
856 AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
866 if (InTok && Prev != String.size())
867 AsmOperands.push_back(AsmOperand(String.substr(Prev)));
869 // The first token of the instruction is the mnemonic, which must be a
870 // simple string, not a $foo variable or a singleton register.
871 if (AsmOperands.empty())
872 PrintFatalError(TheDef->getLoc(),
873 "Instruction '" + TheDef->getName() + "' has no tokens");
874 Mnemonic = AsmOperands[0].Token;
875 if (Mnemonic.empty())
876 PrintFatalError(TheDef->getLoc(),
877 "Missing instruction mnemonic");
878 // FIXME : Check and raise an error if it is a register.
879 if (Mnemonic[0] == '$')
880 PrintFatalError(TheDef->getLoc(),
881 "Invalid instruction mnemonic '" + Mnemonic + "'!");
883 // Remove the first operand, it is tracked in the mnemonic field.
884 AsmOperands.erase(AsmOperands.begin());
887 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const {
888 // Reject matchables with no .s string.
889 if (AsmString.empty())
890 PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
892 // Reject any matchables with a newline in them, they should be marked
893 // isCodeGenOnly if they are pseudo instructions.
894 if (AsmString.find('\n') != std::string::npos)
895 PrintFatalError(TheDef->getLoc(),
896 "multiline instruction is not valid for the asmparser, "
897 "mark it isCodeGenOnly");
899 // Remove comments from the asm string. We know that the asmstring only
901 if (!CommentDelimiter.empty() &&
902 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
903 PrintFatalError(TheDef->getLoc(),
904 "asmstring for instruction has comment character in it, "
905 "mark it isCodeGenOnly");
907 // Reject matchables with operand modifiers, these aren't something we can
908 // handle, the target should be refactored to use operands instead of
911 // Also, check for instructions which reference the operand multiple times;
912 // this implies a constraint we would not honor.
913 std::set<std::string> OperandNames;
914 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
915 StringRef Tok = AsmOperands[i].Token;
916 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
917 PrintFatalError(TheDef->getLoc(),
918 "matchable with operand modifier '" + Tok +
919 "' not supported by asm matcher. Mark isCodeGenOnly!");
921 // Verify that any operand is only mentioned once.
922 // We reject aliases and ignore instructions for now.
923 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
925 PrintFatalError(TheDef->getLoc(),
926 "ERROR: matchable with tied operand '" + Tok +
927 "' can never be matched!");
928 // FIXME: Should reject these. The ARM backend hits this with $lane in a
929 // bunch of instructions. It is unclear what the right answer is.
931 errs() << "warning: '" << TheDef->getName() << "': "
932 << "ignoring instruction with tied operand '"
942 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
943 /// if present, from specified token.
945 extractSingletonRegisterForAsmOperand(unsigned OperandNo,
946 const AsmMatcherInfo &Info,
947 std::string &RegisterPrefix) {
948 StringRef Tok = AsmOperands[OperandNo].Token;
949 if (RegisterPrefix.empty()) {
950 std::string LoweredTok = Tok.lower();
951 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
952 AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
956 if (!Tok.startswith(RegisterPrefix))
959 StringRef RegName = Tok.substr(RegisterPrefix.size());
960 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
961 AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
963 // If there is no register prefix (i.e. "%" in "%eax"), then this may
964 // be some random non-register token, just ignore it.
968 static std::string getEnumNameForToken(StringRef Str) {
971 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
973 case '*': Res += "_STAR_"; break;
974 case '%': Res += "_PCT_"; break;
975 case ':': Res += "_COLON_"; break;
976 case '!': Res += "_EXCLAIM_"; break;
977 case '.': Res += "_DOT_"; break;
978 case '<': Res += "_LT_"; break;
979 case '>': Res += "_GT_"; break;
981 if ((*it >= 'A' && *it <= 'Z') ||
982 (*it >= 'a' && *it <= 'z') ||
983 (*it >= '0' && *it <= '9'))
986 Res += "_" + utostr((unsigned) *it) + "_";
993 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
994 ClassInfo *&Entry = TokenClasses[Token];
997 Classes.emplace_front();
998 Entry = &Classes.front();
999 Entry->Kind = ClassInfo::Token;
1000 Entry->ClassName = "Token";
1001 Entry->Name = "MCK_" + getEnumNameForToken(Token);
1002 Entry->ValueName = Token;
1003 Entry->PredicateMethod = "<invalid>";
1004 Entry->RenderMethod = "<invalid>";
1005 Entry->ParserMethod = "";
1006 Entry->DiagnosticType = "";
1013 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1015 Record *Rec = OI.Rec;
1017 Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1018 return getOperandClass(Rec, SubOpIdx);
1022 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1023 if (Rec->isSubClassOf("RegisterOperand")) {
1024 // RegisterOperand may have an associated ParserMatchClass. If it does,
1025 // use it, else just fall back to the underlying register class.
1026 const RecordVal *R = Rec->getValue("ParserMatchClass");
1027 if (!R || !R->getValue())
1028 PrintFatalError("Record `" + Rec->getName() +
1029 "' does not have a ParserMatchClass!\n");
1031 if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) {
1032 Record *MatchClass = DI->getDef();
1033 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1037 // No custom match class. Just use the register class.
1038 Record *ClassRec = Rec->getValueAsDef("RegClass");
1040 PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1041 "' has no associated register class!\n");
1042 if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1044 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1048 if (Rec->isSubClassOf("RegisterClass")) {
1049 if (ClassInfo *CI = RegisterClassClasses[Rec])
1051 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1054 if (!Rec->isSubClassOf("Operand"))
1055 PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() +
1056 "' does not derive from class Operand!\n");
1057 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1058 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1061 PrintFatalError(Rec->getLoc(), "operand has no match class!");
1064 struct LessRegisterSet {
1065 bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const {
1066 // std::set<T> defines its own compariso "operator<", but it
1067 // performs a lexicographical comparison by T's innate comparison
1068 // for some reason. We don't want non-deterministic pointer
1069 // comparisons so use this instead.
1070 return std::lexicographical_compare(LHS.begin(), LHS.end(),
1071 RHS.begin(), RHS.end(),
1076 void AsmMatcherInfo::
1077 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) {
1078 const std::vector<CodeGenRegister*> &Registers =
1079 Target.getRegBank().getRegisters();
1080 ArrayRef<CodeGenRegisterClass*> RegClassList =
1081 Target.getRegBank().getRegClasses();
1083 typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1085 // The register sets used for matching.
1086 RegisterSetSet RegisterSets;
1088 // Gather the defined sets.
1089 for (const CodeGenRegisterClass *RC : RegClassList)
1090 RegisterSets.insert(RegisterSet(RC->getOrder().begin(),
1091 RC->getOrder().end()));
1093 // Add any required singleton sets.
1094 for (Record *Rec : SingletonRegisters) {
1095 RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1098 // Introduce derived sets where necessary (when a register does not determine
1099 // a unique register set class), and build the mapping of registers to the set
1100 // they should classify to.
1101 std::map<Record*, RegisterSet> RegisterMap;
1102 for (const CodeGenRegister *CGR : Registers) {
1103 // Compute the intersection of all sets containing this register.
1104 RegisterSet ContainingSet;
1106 for (const RegisterSet &RS : RegisterSets) {
1107 if (!RS.count(CGR->TheDef))
1110 if (ContainingSet.empty()) {
1116 std::swap(Tmp, ContainingSet);
1117 std::insert_iterator<RegisterSet> II(ContainingSet,
1118 ContainingSet.begin());
1119 std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II,
1123 if (!ContainingSet.empty()) {
1124 RegisterSets.insert(ContainingSet);
1125 RegisterMap.insert(std::make_pair(CGR->TheDef, ContainingSet));
1129 // Construct the register classes.
1130 std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses;
1132 for (const RegisterSet &RS : RegisterSets) {
1133 Classes.emplace_front();
1134 ClassInfo *CI = &Classes.front();
1135 CI->Kind = ClassInfo::RegisterClass0 + Index;
1136 CI->ClassName = "Reg" + utostr(Index);
1137 CI->Name = "MCK_Reg" + utostr(Index);
1139 CI->PredicateMethod = ""; // unused
1140 CI->RenderMethod = "addRegOperands";
1142 // FIXME: diagnostic type.
1143 CI->DiagnosticType = "";
1144 RegisterSetClasses.insert(std::make_pair(RS, CI));
1148 // Find the superclasses; we could compute only the subgroup lattice edges,
1149 // but there isn't really a point.
1150 for (const RegisterSet &RS : RegisterSets) {
1151 ClassInfo *CI = RegisterSetClasses[RS];
1152 for (const RegisterSet &RS2 : RegisterSets)
1154 std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(),
1156 CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1159 // Name the register classes which correspond to a user defined RegisterClass.
1160 for (const CodeGenRegisterClass *RC : RegClassList) {
1161 // Def will be NULL for non-user defined register classes.
1162 Record *Def = RC->getDef();
1165 ClassInfo *CI = RegisterSetClasses[RegisterSet(RC->getOrder().begin(),
1166 RC->getOrder().end())];
1167 if (CI->ValueName.empty()) {
1168 CI->ClassName = RC->getName();
1169 CI->Name = "MCK_" + RC->getName();
1170 CI->ValueName = RC->getName();
1172 CI->ValueName = CI->ValueName + "," + RC->getName();
1174 RegisterClassClasses.insert(std::make_pair(Def, CI));
1177 // Populate the map for individual registers.
1178 for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(),
1179 ie = RegisterMap.end(); it != ie; ++it)
1180 RegisterClasses[it->first] = RegisterSetClasses[it->second];
1182 // Name the register classes which correspond to singleton registers.
1183 for (Record *Rec : SingletonRegisters) {
1184 ClassInfo *CI = RegisterClasses[Rec];
1185 assert(CI && "Missing singleton register class info!");
1187 if (CI->ValueName.empty()) {
1188 CI->ClassName = Rec->getName();
1189 CI->Name = "MCK_" + Rec->getName();
1190 CI->ValueName = Rec->getName();
1192 CI->ValueName = CI->ValueName + "," + Rec->getName();
1196 void AsmMatcherInfo::buildOperandClasses() {
1197 std::vector<Record*> AsmOperands =
1198 Records.getAllDerivedDefinitions("AsmOperandClass");
1200 // Pre-populate AsmOperandClasses map.
1201 for (Record *Rec : AsmOperands) {
1202 Classes.emplace_front();
1203 AsmOperandClasses[Rec] = &Classes.front();
1207 for (Record *Rec : AsmOperands) {
1208 ClassInfo *CI = AsmOperandClasses[Rec];
1209 CI->Kind = ClassInfo::UserClass0 + Index;
1211 ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1212 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
1213 DefInit *DI = dyn_cast<DefInit>(Supers->getElement(i));
1215 PrintError(Rec->getLoc(), "Invalid super class reference!");
1219 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1221 PrintError(Rec->getLoc(), "Invalid super class reference!");
1223 CI->SuperClasses.push_back(SC);
1225 CI->ClassName = Rec->getValueAsString("Name");
1226 CI->Name = "MCK_" + CI->ClassName;
1227 CI->ValueName = Rec->getName();
1229 // Get or construct the predicate method name.
1230 Init *PMName = Rec->getValueInit("PredicateMethod");
1231 if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1232 CI->PredicateMethod = SI->getValue();
1234 assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1235 CI->PredicateMethod = "is" + CI->ClassName;
1238 // Get or construct the render method name.
1239 Init *RMName = Rec->getValueInit("RenderMethod");
1240 if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1241 CI->RenderMethod = SI->getValue();
1243 assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1244 CI->RenderMethod = "add" + CI->ClassName + "Operands";
1247 // Get the parse method name or leave it as empty.
1248 Init *PRMName = Rec->getValueInit("ParserMethod");
1249 if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1250 CI->ParserMethod = SI->getValue();
1252 // Get the diagnostic type or leave it as empty.
1253 // Get the parse method name or leave it as empty.
1254 Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1255 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1256 CI->DiagnosticType = SI->getValue();
1262 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1263 CodeGenTarget &target,
1264 RecordKeeper &records)
1265 : Records(records), AsmParser(asmParser), Target(target) {
1268 /// buildOperandMatchInfo - Build the necessary information to handle user
1269 /// defined operand parsing methods.
1270 void AsmMatcherInfo::buildOperandMatchInfo() {
1272 /// Map containing a mask with all operands indices that can be found for
1273 /// that class inside a instruction.
1274 typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy;
1275 OpClassMaskTy OpClassMask;
1277 for (const auto &MI : Matchables) {
1278 OpClassMask.clear();
1280 // Keep track of all operands of this instructions which belong to the
1282 for (unsigned i = 0, e = MI.AsmOperands.size(); i != e; ++i) {
1283 const MatchableInfo::AsmOperand &Op = MI.AsmOperands[i];
1284 if (Op.Class->ParserMethod.empty())
1286 unsigned &OperandMask = OpClassMask[Op.Class];
1287 OperandMask |= (1 << i);
1290 // Generate operand match info for each mnemonic/operand class pair.
1291 for (const auto &OCM : OpClassMask) {
1292 unsigned OpMask = OCM.second;
1293 ClassInfo *CI = OCM.first;
1294 OperandMatchInfo.push_back(OperandMatchEntry::create(&MI, CI, OpMask));
1299 void AsmMatcherInfo::buildInfo() {
1300 // Build information about all of the AssemblerPredicates.
1301 std::vector<Record*> AllPredicates =
1302 Records.getAllDerivedDefinitions("Predicate");
1303 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
1304 Record *Pred = AllPredicates[i];
1305 // Ignore predicates that are not intended for the assembler.
1306 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1309 if (Pred->getName().empty())
1310 PrintFatalError(Pred->getLoc(), "Predicate has no name!");
1312 SubtargetFeatures.emplace(
1313 std::piecewise_construct, std::forward_as_tuple(Pred),
1314 std::forward_as_tuple(Pred, SubtargetFeatures.size()));
1315 DEBUG(SubtargetFeatures.find(Pred)->second.dump());
1316 assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!");
1319 // Parse the instructions; we need to do this first so that we can gather the
1320 // singleton register classes.
1321 SmallPtrSet<Record*, 16> SingletonRegisters;
1322 unsigned VariantCount = Target.getAsmParserVariantCount();
1323 for (unsigned VC = 0; VC != VariantCount; ++VC) {
1324 Record *AsmVariant = Target.getAsmParserVariant(VC);
1325 std::string CommentDelimiter =
1326 AsmVariant->getValueAsString("CommentDelimiter");
1327 std::string RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1328 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1330 for (const CodeGenInstruction *CGI : Target.instructions()) {
1332 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1333 // filter the set of instructions we consider.
1334 if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix))
1337 // Ignore "codegen only" instructions.
1338 if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1341 Matchables.emplace_front(*CGI);
1342 MatchableInfo *II = &Matchables.front();
1344 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1346 // Ignore instructions which shouldn't be matched and diagnose invalid
1347 // instruction definitions with an error.
1348 if (!II->validate(CommentDelimiter, true))
1349 Matchables.pop_front();
1352 // Parse all of the InstAlias definitions and stick them in the list of
1354 std::vector<Record*> AllInstAliases =
1355 Records.getAllDerivedDefinitions("InstAlias");
1356 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1357 auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i],
1358 AsmVariantNo, Target);
1360 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1361 // filter the set of instruction aliases we consider, based on the target
1363 if (!StringRef(Alias->ResultInst->TheDef->getName())
1364 .startswith( MatchPrefix))
1367 Matchables.emplace_front(std::move(Alias));
1368 MatchableInfo *II = &Matchables.front();
1370 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1372 // Validate the alias definitions.
1373 II->validate(CommentDelimiter, false);
1377 // Build info for the register classes.
1378 buildRegisterClasses(SingletonRegisters);
1380 // Build info for the user defined assembly operand classes.
1381 buildOperandClasses();
1383 // Build the information about matchables, now that we have fully formed
1385 std::forward_list<MatchableInfo> NewMatchables;
1386 for (auto &II : Matchables) {
1387 // Parse the tokens after the mnemonic.
1388 // Note: buildInstructionOperandReference may insert new AsmOperands, so
1389 // don't precompute the loop bound.
1390 for (unsigned i = 0; i != II.AsmOperands.size(); ++i) {
1391 MatchableInfo::AsmOperand &Op = II.AsmOperands[i];
1392 StringRef Token = Op.Token;
1394 // Check for singleton registers.
1395 if (Record *RegRecord = II.AsmOperands[i].SingletonReg) {
1396 Op.Class = RegisterClasses[RegRecord];
1397 assert(Op.Class && Op.Class->Registers.size() == 1 &&
1398 "Unexpected class for singleton register");
1402 // Check for simple tokens.
1403 if (Token[0] != '$') {
1404 Op.Class = getTokenClass(Token);
1408 if (Token.size() > 1 && isdigit(Token[1])) {
1409 Op.Class = getTokenClass(Token);
1413 // Otherwise this is an operand reference.
1414 StringRef OperandName;
1415 if (Token[1] == '{')
1416 OperandName = Token.substr(2, Token.size() - 3);
1418 OperandName = Token.substr(1);
1420 if (II.DefRec.is<const CodeGenInstruction *>())
1421 buildInstructionOperandReference(&II, OperandName, i);
1423 buildAliasOperandReference(&II, OperandName, Op);
1426 if (II.DefRec.is<const CodeGenInstruction *>()) {
1427 II.buildInstructionResultOperands();
1428 // If the instruction has a two-operand alias, build up the
1429 // matchable here. We'll add them in bulk at the end to avoid
1430 // confusing this loop.
1431 std::string Constraint =
1432 II.TheDef->getValueAsString("TwoOperandAliasConstraint");
1433 if (Constraint != "") {
1434 // Start by making a copy of the original matchable.
1435 NewMatchables.emplace_front(II);
1437 // Adjust it to be a two-operand alias.
1438 NewMatchables.front().formTwoOperandAlias(Constraint);
1441 II.buildAliasResultOperands();
1443 Matchables.splice_after(Matchables.before_begin(), NewMatchables);
1445 // Process token alias definitions and set up the associated superclass
1447 std::vector<Record*> AllTokenAliases =
1448 Records.getAllDerivedDefinitions("TokenAlias");
1449 for (unsigned i = 0, e = AllTokenAliases.size(); i != e; ++i) {
1450 Record *Rec = AllTokenAliases[i];
1451 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1452 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1453 if (FromClass == ToClass)
1454 PrintFatalError(Rec->getLoc(),
1455 "error: Destination value identical to source value.");
1456 FromClass->SuperClasses.push_back(ToClass);
1459 // Reorder classes so that classes precede super classes.
1463 /// buildInstructionOperandReference - The specified operand is a reference to a
1464 /// named operand such as $src. Resolve the Class and OperandInfo pointers.
1465 void AsmMatcherInfo::
1466 buildInstructionOperandReference(MatchableInfo *II,
1467 StringRef OperandName,
1468 unsigned AsmOpIdx) {
1469 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1470 const CGIOperandList &Operands = CGI.Operands;
1471 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1473 // Map this token to an operand.
1475 if (!Operands.hasOperandNamed(OperandName, Idx))
1476 PrintFatalError(II->TheDef->getLoc(),
1477 "error: unable to find operand: '" + OperandName + "'");
1479 // If the instruction operand has multiple suboperands, but the parser
1480 // match class for the asm operand is still the default "ImmAsmOperand",
1481 // then handle each suboperand separately.
1482 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1483 Record *Rec = Operands[Idx].Rec;
1484 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1485 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1486 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1487 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1488 StringRef Token = Op->Token; // save this in case Op gets moved
1489 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1490 MatchableInfo::AsmOperand NewAsmOp(Token);
1491 NewAsmOp.SubOpIdx = SI;
1492 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1494 // Replace Op with first suboperand.
1495 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1500 // Set up the operand class.
1501 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1503 // If the named operand is tied, canonicalize it to the untied operand.
1504 // For example, something like:
1505 // (outs GPR:$dst), (ins GPR:$src)
1506 // with an asmstring of
1508 // we want to canonicalize to:
1510 // so that we know how to provide the $dst operand when filling in the result.
1512 if (Operands[Idx].MINumOperands == 1)
1513 OITied = Operands[Idx].getTiedRegister();
1515 // The tied operand index is an MIOperand index, find the operand that
1517 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1518 OperandName = Operands[Idx.first].Name;
1519 Op->SubOpIdx = Idx.second;
1522 Op->SrcOpName = OperandName;
1525 /// buildAliasOperandReference - When parsing an operand reference out of the
1526 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1527 /// operand reference is by looking it up in the result pattern definition.
1528 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1529 StringRef OperandName,
1530 MatchableInfo::AsmOperand &Op) {
1531 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1533 // Set up the operand class.
1534 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1535 if (CGA.ResultOperands[i].isRecord() &&
1536 CGA.ResultOperands[i].getName() == OperandName) {
1537 // It's safe to go with the first one we find, because CodeGenInstAlias
1538 // validates that all operands with the same name have the same record.
1539 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1540 // Use the match class from the Alias definition, not the
1541 // destination instruction, as we may have an immediate that's
1542 // being munged by the match class.
1543 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1545 Op.SrcOpName = OperandName;
1549 PrintFatalError(II->TheDef->getLoc(),
1550 "error: unable to find operand: '" + OperandName + "'");
1553 void MatchableInfo::buildInstructionResultOperands() {
1554 const CodeGenInstruction *ResultInst = getResultInst();
1556 // Loop over all operands of the result instruction, determining how to
1558 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1559 const CGIOperandList::OperandInfo &OpInfo = ResultInst->Operands[i];
1561 // If this is a tied operand, just copy from the previously handled operand.
1563 if (OpInfo.MINumOperands == 1)
1564 TiedOp = OpInfo.getTiedRegister();
1566 ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1570 // Find out what operand from the asmparser this MCInst operand comes from.
1571 int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1572 if (OpInfo.Name.empty() || SrcOperand == -1) {
1573 // This may happen for operands that are tied to a suboperand of a
1574 // complex operand. Simply use a dummy value here; nobody should
1575 // use this operand slot.
1576 // FIXME: The long term goal is for the MCOperand list to not contain
1577 // tied operands at all.
1578 ResOperands.push_back(ResOperand::getImmOp(0));
1582 // Check if the one AsmOperand populates the entire operand.
1583 unsigned NumOperands = OpInfo.MINumOperands;
1584 if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1585 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1589 // Add a separate ResOperand for each suboperand.
1590 for (unsigned AI = 0; AI < NumOperands; ++AI) {
1591 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1592 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1593 "unexpected AsmOperands for suboperands");
1594 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1599 void MatchableInfo::buildAliasResultOperands() {
1600 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1601 const CodeGenInstruction *ResultInst = getResultInst();
1603 // Loop over all operands of the result instruction, determining how to
1605 unsigned AliasOpNo = 0;
1606 unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1607 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1608 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1610 // If this is a tied operand, just copy from the previously handled operand.
1612 if (OpInfo->MINumOperands == 1)
1613 TiedOp = OpInfo->getTiedRegister();
1615 ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1619 // Handle all the suboperands for this operand.
1620 const std::string &OpName = OpInfo->Name;
1621 for ( ; AliasOpNo < LastOpNo &&
1622 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1623 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1625 // Find out what operand from the asmparser that this MCInst operand
1627 switch (CGA.ResultOperands[AliasOpNo].Kind) {
1628 case CodeGenInstAlias::ResultOperand::K_Record: {
1629 StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1630 int SrcOperand = findAsmOperand(Name, SubIdx);
1631 if (SrcOperand == -1)
1632 PrintFatalError(TheDef->getLoc(), "Instruction '" +
1633 TheDef->getName() + "' has operand '" + OpName +
1634 "' that doesn't appear in asm string!");
1635 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1636 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1640 case CodeGenInstAlias::ResultOperand::K_Imm: {
1641 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1642 ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1645 case CodeGenInstAlias::ResultOperand::K_Reg: {
1646 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1647 ResOperands.push_back(ResOperand::getRegOp(Reg));
1655 static unsigned getConverterOperandID(const std::string &Name,
1656 SetVector<std::string> &Table,
1658 IsNew = Table.insert(Name);
1660 unsigned ID = IsNew ? Table.size() - 1 :
1661 std::find(Table.begin(), Table.end(), Name) - Table.begin();
1663 assert(ID < Table.size());
1668 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1669 std::forward_list<MatchableInfo> &Infos,
1671 SetVector<std::string> OperandConversionKinds;
1672 SetVector<std::string> InstructionConversionKinds;
1673 std::vector<std::vector<uint8_t> > ConversionTable;
1674 size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1676 // TargetOperandClass - This is the target's operand class, like X86Operand.
1677 std::string TargetOperandClass = Target.getName() + "Operand";
1679 // Write the convert function to a separate stream, so we can drop it after
1680 // the enum. We'll build up the conversion handlers for the individual
1681 // operand types opportunistically as we encounter them.
1682 std::string ConvertFnBody;
1683 raw_string_ostream CvtOS(ConvertFnBody);
1684 // Start the unified conversion function.
1685 CvtOS << "void " << Target.getName() << ClassName << "::\n"
1686 << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1687 << "unsigned Opcode,\n"
1688 << " const OperandVector"
1689 << " &Operands) {\n"
1690 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1691 << " const uint8_t *Converter = ConversionTable[Kind];\n"
1692 << " Inst.setOpcode(Opcode);\n"
1693 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1694 << " switch (*p) {\n"
1695 << " default: llvm_unreachable(\"invalid conversion entry!\");\n"
1696 << " case CVT_Reg:\n"
1697 << " static_cast<" << TargetOperandClass
1698 << "&>(*Operands[*(p + 1)]).addRegOperands(Inst, 1);\n"
1700 << " case CVT_Tied:\n"
1701 << " Inst.addOperand(Inst.getOperand(*(p + 1)));\n"
1704 std::string OperandFnBody;
1705 raw_string_ostream OpOS(OperandFnBody);
1706 // Start the operand number lookup function.
1707 OpOS << "void " << Target.getName() << ClassName << "::\n"
1708 << "convertToMapAndConstraints(unsigned Kind,\n";
1710 OpOS << "const OperandVector &Operands) {\n"
1711 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1712 << " unsigned NumMCOperands = 0;\n"
1713 << " const uint8_t *Converter = ConversionTable[Kind];\n"
1714 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1715 << " switch (*p) {\n"
1716 << " default: llvm_unreachable(\"invalid conversion entry!\");\n"
1717 << " case CVT_Reg:\n"
1718 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1719 << " Operands[*(p + 1)]->setConstraint(\"r\");\n"
1720 << " ++NumMCOperands;\n"
1722 << " case CVT_Tied:\n"
1723 << " ++NumMCOperands;\n"
1726 // Pre-populate the operand conversion kinds with the standard always
1727 // available entries.
1728 OperandConversionKinds.insert("CVT_Done");
1729 OperandConversionKinds.insert("CVT_Reg");
1730 OperandConversionKinds.insert("CVT_Tied");
1731 enum { CVT_Done, CVT_Reg, CVT_Tied };
1733 for (auto &II : Infos) {
1734 // Check if we have a custom match function.
1735 std::string AsmMatchConverter =
1736 II.getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1737 if (!AsmMatchConverter.empty()) {
1738 std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1739 II.ConversionFnKind = Signature;
1741 // Check if we have already generated this signature.
1742 if (!InstructionConversionKinds.insert(Signature))
1745 // Remember this converter for the kind enum.
1746 unsigned KindID = OperandConversionKinds.size();
1747 OperandConversionKinds.insert("CVT_" +
1748 getEnumNameForToken(AsmMatchConverter));
1750 // Add the converter row for this instruction.
1751 ConversionTable.push_back(std::vector<uint8_t>());
1752 ConversionTable.back().push_back(KindID);
1753 ConversionTable.back().push_back(CVT_Done);
1755 // Add the handler to the conversion driver function.
1756 CvtOS << " case CVT_"
1757 << getEnumNameForToken(AsmMatchConverter) << ":\n"
1758 << " " << AsmMatchConverter << "(Inst, Operands);\n"
1761 // FIXME: Handle the operand number lookup for custom match functions.
1765 // Build the conversion function signature.
1766 std::string Signature = "Convert";
1768 std::vector<uint8_t> ConversionRow;
1770 // Compute the convert enum and the case body.
1771 MaxRowLength = std::max(MaxRowLength, II.ResOperands.size() * 2 + 1);
1773 for (unsigned i = 0, e = II.ResOperands.size(); i != e; ++i) {
1774 const MatchableInfo::ResOperand &OpInfo = II.ResOperands[i];
1776 // Generate code to populate each result operand.
1777 switch (OpInfo.Kind) {
1778 case MatchableInfo::ResOperand::RenderAsmOperand: {
1779 // This comes from something we parsed.
1780 const MatchableInfo::AsmOperand &Op =
1781 II.AsmOperands[OpInfo.AsmOperandNum];
1783 // Registers are always converted the same, don't duplicate the
1784 // conversion function based on them.
1787 Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
1789 Signature += utostr(OpInfo.MINumOperands);
1790 Signature += "_" + itostr(OpInfo.AsmOperandNum);
1792 // Add the conversion kind, if necessary, and get the associated ID
1793 // the index of its entry in the vector).
1794 std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" :
1795 Op.Class->RenderMethod);
1796 Name = getEnumNameForToken(Name);
1798 bool IsNewConverter = false;
1799 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1802 // Add the operand entry to the instruction kind conversion row.
1803 ConversionRow.push_back(ID);
1804 ConversionRow.push_back(OpInfo.AsmOperandNum + 1);
1806 if (!IsNewConverter)
1809 // This is a new operand kind. Add a handler for it to the
1810 // converter driver.
1811 CvtOS << " case " << Name << ":\n"
1812 << " static_cast<" << TargetOperandClass
1813 << "&>(*Operands[*(p + 1)])." << Op.Class->RenderMethod
1814 << "(Inst, " << OpInfo.MINumOperands << ");\n"
1817 // Add a handler for the operand number lookup.
1818 OpOS << " case " << Name << ":\n"
1819 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
1821 if (Op.Class->isRegisterClass())
1822 OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n";
1824 OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n";
1825 OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n"
1829 case MatchableInfo::ResOperand::TiedOperand: {
1830 // If this operand is tied to a previous one, just copy the MCInst
1831 // operand from the earlier one.We can only tie single MCOperand values.
1832 assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
1833 unsigned TiedOp = OpInfo.TiedOperandNum;
1834 assert(i > TiedOp && "Tied operand precedes its target!");
1835 Signature += "__Tie" + utostr(TiedOp);
1836 ConversionRow.push_back(CVT_Tied);
1837 ConversionRow.push_back(TiedOp);
1840 case MatchableInfo::ResOperand::ImmOperand: {
1841 int64_t Val = OpInfo.ImmVal;
1842 std::string Ty = "imm_" + itostr(Val);
1843 Signature += "__" + Ty;
1845 std::string Name = "CVT_" + Ty;
1846 bool IsNewConverter = false;
1847 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1849 // Add the operand entry to the instruction kind conversion row.
1850 ConversionRow.push_back(ID);
1851 ConversionRow.push_back(0);
1853 if (!IsNewConverter)
1856 CvtOS << " case " << Name << ":\n"
1857 << " Inst.addOperand(MCOperand::CreateImm(" << Val << "));\n"
1860 OpOS << " case " << Name << ":\n"
1861 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1862 << " Operands[*(p + 1)]->setConstraint(\"\");\n"
1863 << " ++NumMCOperands;\n"
1867 case MatchableInfo::ResOperand::RegOperand: {
1868 std::string Reg, Name;
1869 if (!OpInfo.Register) {
1873 Reg = getQualifiedName(OpInfo.Register);
1874 Name = "reg" + OpInfo.Register->getName();
1876 Signature += "__" + Name;
1877 Name = "CVT_" + Name;
1878 bool IsNewConverter = false;
1879 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1881 // Add the operand entry to the instruction kind conversion row.
1882 ConversionRow.push_back(ID);
1883 ConversionRow.push_back(0);
1885 if (!IsNewConverter)
1887 CvtOS << " case " << Name << ":\n"
1888 << " Inst.addOperand(MCOperand::CreateReg(" << Reg << "));\n"
1891 OpOS << " case " << Name << ":\n"
1892 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1893 << " Operands[*(p + 1)]->setConstraint(\"m\");\n"
1894 << " ++NumMCOperands;\n"
1900 // If there were no operands, add to the signature to that effect
1901 if (Signature == "Convert")
1902 Signature += "_NoOperands";
1904 II.ConversionFnKind = Signature;
1906 // Save the signature. If we already have it, don't add a new row
1908 if (!InstructionConversionKinds.insert(Signature))
1911 // Add the row to the table.
1912 ConversionTable.push_back(ConversionRow);
1915 // Finish up the converter driver function.
1916 CvtOS << " }\n }\n}\n\n";
1918 // Finish up the operand number lookup function.
1919 OpOS << " }\n }\n}\n\n";
1921 OS << "namespace {\n";
1923 // Output the operand conversion kind enum.
1924 OS << "enum OperatorConversionKind {\n";
1925 for (unsigned i = 0, e = OperandConversionKinds.size(); i != e; ++i)
1926 OS << " " << OperandConversionKinds[i] << ",\n";
1927 OS << " CVT_NUM_CONVERTERS\n";
1930 // Output the instruction conversion kind enum.
1931 OS << "enum InstructionConversionKind {\n";
1932 for (SetVector<std::string>::const_iterator
1933 i = InstructionConversionKinds.begin(),
1934 e = InstructionConversionKinds.end(); i != e; ++i)
1935 OS << " " << *i << ",\n";
1936 OS << " CVT_NUM_SIGNATURES\n";
1940 OS << "} // end anonymous namespace\n\n";
1942 // Output the conversion table.
1943 OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
1944 << MaxRowLength << "] = {\n";
1946 for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
1947 assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
1948 OS << " // " << InstructionConversionKinds[Row] << "\n";
1950 for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2)
1951 OS << OperandConversionKinds[ConversionTable[Row][i]] << ", "
1952 << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
1953 OS << "CVT_Done },\n";
1958 // Spit out the conversion driver function.
1961 // Spit out the operand number lookup function.
1965 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
1966 static void emitMatchClassEnumeration(CodeGenTarget &Target,
1967 std::forward_list<ClassInfo> &Infos,
1969 OS << "namespace {\n\n";
1971 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1972 << "/// instruction matching.\n";
1973 OS << "enum MatchClassKind {\n";
1974 OS << " InvalidMatchClass = 0,\n";
1975 for (const auto &CI : Infos) {
1976 OS << " " << CI.Name << ", // ";
1977 if (CI.Kind == ClassInfo::Token) {
1978 OS << "'" << CI.ValueName << "'\n";
1979 } else if (CI.isRegisterClass()) {
1980 if (!CI.ValueName.empty())
1981 OS << "register class '" << CI.ValueName << "'\n";
1983 OS << "derived register class\n";
1985 OS << "user defined class '" << CI.ValueName << "'\n";
1988 OS << " NumMatchClassKinds\n";
1994 /// emitValidateOperandClass - Emit the function to validate an operand class.
1995 static void emitValidateOperandClass(AsmMatcherInfo &Info,
1997 OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
1998 << "MatchClassKind Kind) {\n";
1999 OS << " " << Info.Target.getName() << "Operand &Operand = ("
2000 << Info.Target.getName() << "Operand&)GOp;\n";
2002 // The InvalidMatchClass is not to match any operand.
2003 OS << " if (Kind == InvalidMatchClass)\n";
2004 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2006 // Check for Token operands first.
2007 // FIXME: Use a more specific diagnostic type.
2008 OS << " if (Operand.isToken())\n";
2009 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2010 << " MCTargetAsmParser::Match_Success :\n"
2011 << " MCTargetAsmParser::Match_InvalidOperand;\n\n";
2013 // Check the user classes. We don't care what order since we're only
2014 // actually matching against one of them.
2015 for (const auto &CI : Info.Classes) {
2016 if (!CI.isUserClass())
2019 OS << " // '" << CI.ClassName << "' class\n";
2020 OS << " if (Kind == " << CI.Name << ") {\n";
2021 OS << " if (Operand." << CI.PredicateMethod << "())\n";
2022 OS << " return MCTargetAsmParser::Match_Success;\n";
2023 if (!CI.DiagnosticType.empty())
2024 OS << " return " << Info.Target.getName() << "AsmParser::Match_"
2025 << CI.DiagnosticType << ";\n";
2029 // Check for register operands, including sub-classes.
2030 OS << " if (Operand.isReg()) {\n";
2031 OS << " MatchClassKind OpKind;\n";
2032 OS << " switch (Operand.getReg()) {\n";
2033 OS << " default: OpKind = InvalidMatchClass; break;\n";
2034 for (const auto &RC : Info.RegisterClasses)
2035 OS << " case " << Info.Target.getName() << "::"
2036 << RC.first->getName() << ": OpKind = " << RC.second->Name
2039 OS << " return isSubclass(OpKind, Kind) ? "
2040 << "MCTargetAsmParser::Match_Success :\n "
2041 << " MCTargetAsmParser::Match_InvalidOperand;\n }\n\n";
2043 // Generic fallthrough match failure case for operands that don't have
2044 // specialized diagnostic types.
2045 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n";
2049 /// emitIsSubclass - Emit the subclass predicate function.
2050 static void emitIsSubclass(CodeGenTarget &Target,
2051 std::forward_list<ClassInfo> &Infos,
2053 OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2054 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2055 OS << " if (A == B)\n";
2056 OS << " return true;\n\n";
2059 raw_string_ostream SS(OStr);
2061 SS << " switch (A) {\n";
2062 SS << " default:\n";
2063 SS << " return false;\n";
2064 for (const auto &A : Infos) {
2065 std::vector<StringRef> SuperClasses;
2066 for (const auto &B : Infos) {
2067 if (&A != &B && A.isSubsetOf(B))
2068 SuperClasses.push_back(B.Name);
2071 if (SuperClasses.empty())
2075 SS << "\n case " << A.Name << ":\n";
2077 if (SuperClasses.size() == 1) {
2078 SS << " return B == " << SuperClasses.back().str() << ";\n";
2082 if (!SuperClasses.empty()) {
2083 SS << " switch (B) {\n";
2084 SS << " default: return false;\n";
2085 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
2086 SS << " case " << SuperClasses[i].str() << ": return true;\n";
2089 // No case statement to emit
2090 SS << " return false;\n";
2095 // If there were case statements emitted into the string stream, write them
2096 // to the output stream, otherwise write the default.
2100 OS << " return false;\n";
2105 /// emitMatchTokenString - Emit the function to match a token string to the
2106 /// appropriate match class value.
2107 static void emitMatchTokenString(CodeGenTarget &Target,
2108 std::forward_list<ClassInfo> &Infos,
2110 // Construct the match list.
2111 std::vector<StringMatcher::StringPair> Matches;
2112 for (const auto &CI : Infos) {
2113 if (CI.Kind == ClassInfo::Token)
2115 StringMatcher::StringPair(CI.ValueName, "return " + CI.Name + ";"));
2118 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2120 StringMatcher("Name", Matches, OS).Emit();
2122 OS << " return InvalidMatchClass;\n";
2126 /// emitMatchRegisterName - Emit the function to match a string to the target
2127 /// specific register enum.
2128 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2130 // Construct the match list.
2131 std::vector<StringMatcher::StringPair> Matches;
2132 const std::vector<CodeGenRegister*> &Regs =
2133 Target.getRegBank().getRegisters();
2134 for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
2135 const CodeGenRegister *Reg = Regs[i];
2136 if (Reg->TheDef->getValueAsString("AsmName").empty())
2139 Matches.push_back(StringMatcher::StringPair(
2140 Reg->TheDef->getValueAsString("AsmName"),
2141 "return " + utostr(Reg->EnumValue) + ";"));
2144 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
2146 StringMatcher("Name", Matches, OS).Emit();
2148 OS << " return 0;\n";
2152 static const char *getMinimalTypeForRange(uint64_t Range) {
2153 assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large");
2154 if (Range > 0xFFFFFFFFULL)
2163 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) {
2164 uint64_t MaxIndex = Info.SubtargetFeatures.size();
2167 return getMinimalTypeForRange(1ULL << MaxIndex);
2170 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
2172 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
2174 OS << "// Flags for subtarget features that participate in "
2175 << "instruction matching.\n";
2176 OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info)
2178 for (const auto &SF : Info.SubtargetFeatures) {
2179 const SubtargetFeatureInfo &SFI = SF.second;
2180 OS << " " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n";
2182 OS << " Feature_None = 0\n";
2186 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
2187 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2188 // Get the set of diagnostic types from all of the operand classes.
2189 std::set<StringRef> Types;
2190 for (std::map<Record*, ClassInfo*>::const_iterator
2191 I = Info.AsmOperandClasses.begin(),
2192 E = Info.AsmOperandClasses.end(); I != E; ++I) {
2193 if (!I->second->DiagnosticType.empty())
2194 Types.insert(I->second->DiagnosticType);
2197 if (Types.empty()) return;
2199 // Now emit the enum entries.
2200 for (std::set<StringRef>::const_iterator I = Types.begin(), E = Types.end();
2202 OS << " Match_" << *I << ",\n";
2203 OS << " END_OPERAND_DIAGNOSTIC_TYPES\n";
2206 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2207 /// user-level name for a subtarget feature.
2208 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2209 OS << "// User-level names for subtarget features that participate in\n"
2210 << "// instruction matching.\n"
2211 << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2212 if (!Info.SubtargetFeatures.empty()) {
2213 OS << " switch(Val) {\n";
2214 for (const auto &SF : Info.SubtargetFeatures) {
2215 const SubtargetFeatureInfo &SFI = SF.second;
2216 // FIXME: Totally just a placeholder name to get the algorithm working.
2217 OS << " case " << SFI.getEnumName() << ": return \""
2218 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2220 OS << " default: return \"(unknown)\";\n";
2223 // Nothing to emit, so skip the switch
2224 OS << " return \"(unknown)\";\n";
2229 /// emitComputeAvailableFeatures - Emit the function to compute the list of
2230 /// available features given a subtarget.
2231 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info,
2233 std::string ClassName =
2234 Info.AsmParser->getValueAsString("AsmParserClassName");
2236 OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n"
2237 << "ComputeAvailableFeatures(uint64_t FB) const {\n";
2238 OS << " uint64_t Features = 0;\n";
2239 for (const auto &SF : Info.SubtargetFeatures) {
2240 const SubtargetFeatureInfo &SFI = SF.second;
2243 std::string CondStorage =
2244 SFI.TheDef->getValueAsString("AssemblerCondString");
2245 StringRef Conds = CondStorage;
2246 std::pair<StringRef,StringRef> Comma = Conds.split(',');
2253 StringRef Cond = Comma.first;
2254 if (Cond[0] == '!') {
2256 Cond = Cond.substr(1);
2259 OS << "((FB & " << Info.Target.getName() << "::" << Cond << ")";
2266 if (Comma.second.empty())
2270 Comma = Comma.second.split(',');
2274 OS << " Features |= " << SFI.getEnumName() << ";\n";
2276 OS << " return Features;\n";
2280 static std::string GetAliasRequiredFeatures(Record *R,
2281 const AsmMatcherInfo &Info) {
2282 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2284 unsigned NumFeatures = 0;
2285 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2286 const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2289 PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2290 "' is not marked as an AssemblerPredicate!");
2295 Result += F->getEnumName();
2299 if (NumFeatures > 1)
2300 Result = '(' + Result + ')';
2304 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info,
2305 std::vector<Record*> &Aliases,
2306 unsigned Indent = 0,
2307 StringRef AsmParserVariantName = StringRef()){
2308 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
2309 // iteration order of the map is stable.
2310 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2312 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
2313 Record *R = Aliases[i];
2314 // FIXME: Allow AssemblerVariantName to be a comma separated list.
2315 std::string AsmVariantName = R->getValueAsString("AsmVariantName");
2316 if (AsmVariantName != AsmParserVariantName)
2318 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2320 if (AliasesFromMnemonic.empty())
2323 // Process each alias a "from" mnemonic at a time, building the code executed
2324 // by the string remapper.
2325 std::vector<StringMatcher::StringPair> Cases;
2326 for (std::map<std::string, std::vector<Record*> >::iterator
2327 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
2329 const std::vector<Record*> &ToVec = I->second;
2331 // Loop through each alias and emit code that handles each case. If there
2332 // are two instructions without predicates, emit an error. If there is one,
2334 std::string MatchCode;
2335 int AliasWithNoPredicate = -1;
2337 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2338 Record *R = ToVec[i];
2339 std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2341 // If this unconditionally matches, remember it for later and diagnose
2343 if (FeatureMask.empty()) {
2344 if (AliasWithNoPredicate != -1) {
2345 // We can't have two aliases from the same mnemonic with no predicate.
2346 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2347 "two MnemonicAliases with the same 'from' mnemonic!");
2348 PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2351 AliasWithNoPredicate = i;
2354 if (R->getValueAsString("ToMnemonic") == I->first)
2355 PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2357 if (!MatchCode.empty())
2358 MatchCode += "else ";
2359 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
2360 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
2363 if (AliasWithNoPredicate != -1) {
2364 Record *R = ToVec[AliasWithNoPredicate];
2365 if (!MatchCode.empty())
2366 MatchCode += "else\n ";
2367 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2370 MatchCode += "return;";
2372 Cases.push_back(std::make_pair(I->first, MatchCode));
2374 StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2377 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2378 /// emit a function for them and return true, otherwise return false.
2379 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2380 CodeGenTarget &Target) {
2381 // Ignore aliases when match-prefix is set.
2382 if (!MatchPrefix.empty())
2385 std::vector<Record*> Aliases =
2386 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2387 if (Aliases.empty()) return false;
2389 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2390 "uint64_t Features, unsigned VariantID) {\n";
2391 OS << " switch (VariantID) {\n";
2392 unsigned VariantCount = Target.getAsmParserVariantCount();
2393 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2394 Record *AsmVariant = Target.getAsmParserVariant(VC);
2395 int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2396 std::string AsmParserVariantName = AsmVariant->getValueAsString("Name");
2397 OS << " case " << AsmParserVariantNo << ":\n";
2398 emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2399 AsmParserVariantName);
2404 // Emit aliases that apply to all variants.
2405 emitMnemonicAliasVariant(OS, Info, Aliases);
2412 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2413 const AsmMatcherInfo &Info, StringRef ClassName,
2414 StringToOffsetTable &StringTable,
2415 unsigned MaxMnemonicIndex) {
2416 unsigned MaxMask = 0;
2417 for (std::vector<OperandMatchEntry>::const_iterator it =
2418 Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2420 MaxMask |= it->OperandMask;
2423 // Emit the static custom operand parsing table;
2424 OS << "namespace {\n";
2425 OS << " struct OperandMatchEntry {\n";
2426 OS << " " << getMinimalRequiredFeaturesType(Info)
2427 << " RequiredFeatures;\n";
2428 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
2430 OS << " " << getMinimalTypeForRange(std::distance(
2431 Info.Classes.begin(), Info.Classes.end())) << " Class;\n";
2432 OS << " " << getMinimalTypeForRange(MaxMask)
2433 << " OperandMask;\n\n";
2434 OS << " StringRef getMnemonic() const {\n";
2435 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
2436 OS << " MnemonicTable[Mnemonic]);\n";
2440 OS << " // Predicate for searching for an opcode.\n";
2441 OS << " struct LessOpcodeOperand {\n";
2442 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2443 OS << " return LHS.getMnemonic() < RHS;\n";
2445 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2446 OS << " return LHS < RHS.getMnemonic();\n";
2448 OS << " bool operator()(const OperandMatchEntry &LHS,";
2449 OS << " const OperandMatchEntry &RHS) {\n";
2450 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
2454 OS << "} // end anonymous namespace.\n\n";
2456 OS << "static const OperandMatchEntry OperandMatchTable["
2457 << Info.OperandMatchInfo.size() << "] = {\n";
2459 OS << " /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2460 for (std::vector<OperandMatchEntry>::const_iterator it =
2461 Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2463 const OperandMatchEntry &OMI = *it;
2464 const MatchableInfo &II = *OMI.MI;
2468 // Write the required features mask.
2469 if (!II.RequiredFeatures.empty()) {
2470 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2472 OS << II.RequiredFeatures[i]->getEnumName();
2477 // Store a pascal-style length byte in the mnemonic.
2478 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2479 OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2480 << " /* " << II.Mnemonic << " */, ";
2484 OS << ", " << OMI.OperandMask;
2486 bool printComma = false;
2487 for (int i = 0, e = 31; i !=e; ++i)
2488 if (OMI.OperandMask & (1 << i)) {
2500 // Emit the operand class switch to call the correct custom parser for
2501 // the found operand class.
2502 OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2503 << Target.getName() << ClassName << "::\n"
2504 << "tryCustomParseOperand(OperandVector"
2505 << " &Operands,\n unsigned MCK) {\n\n"
2506 << " switch(MCK) {\n";
2508 for (const auto &CI : Info.Classes) {
2509 if (CI.ParserMethod.empty())
2511 OS << " case " << CI.Name << ":\n"
2512 << " return " << CI.ParserMethod << "(Operands);\n";
2515 OS << " default:\n";
2516 OS << " return MatchOperand_NoMatch;\n";
2518 OS << " return MatchOperand_NoMatch;\n";
2521 // Emit the static custom operand parser. This code is very similar with
2522 // the other matcher. Also use MatchResultTy here just in case we go for
2523 // a better error handling.
2524 OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2525 << Target.getName() << ClassName << "::\n"
2526 << "MatchOperandParserImpl(OperandVector"
2527 << " &Operands,\n StringRef Mnemonic) {\n";
2529 // Emit code to get the available features.
2530 OS << " // Get the current feature set.\n";
2531 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2533 OS << " // Get the next operand index.\n";
2534 OS << " unsigned NextOpNum = Operands.size()-1;\n";
2536 // Emit code to search the table.
2537 OS << " // Search the table.\n";
2538 OS << " std::pair<const OperandMatchEntry*, const OperandMatchEntry*>";
2539 OS << " MnemonicRange =\n";
2540 OS << " std::equal_range(OperandMatchTable, OperandMatchTable+"
2541 << Info.OperandMatchInfo.size() << ", Mnemonic,\n"
2542 << " LessOpcodeOperand());\n\n";
2544 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
2545 OS << " return MatchOperand_NoMatch;\n\n";
2547 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2548 << " *ie = MnemonicRange.second; it != ie; ++it) {\n";
2550 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
2551 OS << " assert(Mnemonic == it->getMnemonic());\n\n";
2553 // Emit check that the required features are available.
2554 OS << " // check if the available features match\n";
2555 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
2556 << "!= it->RequiredFeatures) {\n";
2557 OS << " continue;\n";
2560 // Emit check to ensure the operand number matches.
2561 OS << " // check if the operand in question has a custom parser.\n";
2562 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n";
2563 OS << " continue;\n\n";
2565 // Emit call to the custom parser method
2566 OS << " // call custom parse method to handle the operand\n";
2567 OS << " OperandMatchResultTy Result = ";
2568 OS << "tryCustomParseOperand(Operands, it->Class);\n";
2569 OS << " if (Result != MatchOperand_NoMatch)\n";
2570 OS << " return Result;\n";
2573 OS << " // Okay, we had no match.\n";
2574 OS << " return MatchOperand_NoMatch;\n";
2578 void AsmMatcherEmitter::run(raw_ostream &OS) {
2579 CodeGenTarget Target(Records);
2580 Record *AsmParser = Target.getAsmParser();
2581 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2583 // Compute the information on the instructions to match.
2584 AsmMatcherInfo Info(AsmParser, Target, Records);
2587 // Sort the instruction table using the partial order on classes. We use
2588 // stable_sort to ensure that ambiguous instructions are still
2589 // deterministically ordered.
2590 Info.Matchables.sort();
2592 DEBUG_WITH_TYPE("instruction_info", {
2593 for (const auto &MI : Info.Matchables)
2597 // Check for ambiguous matchables.
2598 DEBUG_WITH_TYPE("ambiguous_instrs", {
2599 unsigned NumAmbiguous = 0;
2600 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2602 for (auto J = std::next(I); J != E; ++J) {
2603 const MatchableInfo &A = *I;
2604 const MatchableInfo &B = *J;
2606 if (A.couldMatchAmbiguouslyWith(B)) {
2607 errs() << "warning: ambiguous matchables:\n";
2609 errs() << "\nis incomparable with:\n";
2617 errs() << "warning: " << NumAmbiguous
2618 << " ambiguous matchables!\n";
2621 // Compute the information on the custom operand parsing.
2622 Info.buildOperandMatchInfo();
2624 // Write the output.
2626 // Information for the class declaration.
2627 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2628 OS << "#undef GET_ASSEMBLER_HEADER\n";
2629 OS << " // This should be included into the middle of the declaration of\n";
2630 OS << " // your subclasses implementation of MCTargetAsmParser.\n";
2631 OS << " uint64_t ComputeAvailableFeatures(uint64_t FeatureBits) const;\n";
2632 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, "
2633 << "unsigned Opcode,\n"
2634 << " const OperandVector "
2636 OS << " void convertToMapAndConstraints(unsigned Kind,\n ";
2637 OS << " const OperandVector &Operands) override;\n";
2638 OS << " bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) override;\n";
2639 OS << " unsigned MatchInstructionImpl(\n";
2641 OS << "const OperandVector &Operands,\n"
2642 << " MCInst &Inst,\n"
2643 << " uint64_t &ErrorInfo,"
2644 << " bool matchingInlineAsm,\n"
2645 << " unsigned VariantID = 0);\n";
2647 if (Info.OperandMatchInfo.size()) {
2648 OS << "\n enum OperandMatchResultTy {\n";
2649 OS << " MatchOperand_Success, // operand matched successfully\n";
2650 OS << " MatchOperand_NoMatch, // operand did not match\n";
2651 OS << " MatchOperand_ParseFail // operand matched but had errors\n";
2653 OS << " OperandMatchResultTy MatchOperandParserImpl(\n";
2654 OS << " OperandVector &Operands,\n";
2655 OS << " StringRef Mnemonic);\n";
2657 OS << " OperandMatchResultTy tryCustomParseOperand(\n";
2658 OS << " OperandVector &Operands,\n";
2659 OS << " unsigned MCK);\n\n";
2662 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2664 // Emit the operand match diagnostic enum names.
2665 OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
2666 OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2667 emitOperandDiagnosticTypes(Info, OS);
2668 OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2671 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2672 OS << "#undef GET_REGISTER_MATCHER\n\n";
2674 // Emit the subtarget feature enumeration.
2675 emitSubtargetFeatureFlagEnumeration(Info, OS);
2677 // Emit the function to match a register name to number.
2678 // This should be omitted for Mips target
2679 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
2680 emitMatchRegisterName(Target, AsmParser, OS);
2682 OS << "#endif // GET_REGISTER_MATCHER\n\n";
2684 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
2685 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
2687 // Generate the helper function to get the names for subtarget features.
2688 emitGetSubtargetFeatureName(Info, OS);
2690 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
2692 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2693 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2695 // Generate the function that remaps for mnemonic aliases.
2696 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
2698 // Generate the convertToMCInst function to convert operands into an MCInst.
2699 // Also, generate the convertToMapAndConstraints function for MS-style inline
2700 // assembly. The latter doesn't actually generate a MCInst.
2701 emitConvertFuncs(Target, ClassName, Info.Matchables, OS);
2703 // Emit the enumeration for classes which participate in matching.
2704 emitMatchClassEnumeration(Target, Info.Classes, OS);
2706 // Emit the routine to match token strings to their match class.
2707 emitMatchTokenString(Target, Info.Classes, OS);
2709 // Emit the subclass predicate routine.
2710 emitIsSubclass(Target, Info.Classes, OS);
2712 // Emit the routine to validate an operand against a match class.
2713 emitValidateOperandClass(Info, OS);
2715 // Emit the available features compute function.
2716 emitComputeAvailableFeatures(Info, OS);
2719 StringToOffsetTable StringTable;
2721 size_t MaxNumOperands = 0;
2722 unsigned MaxMnemonicIndex = 0;
2723 bool HasDeprecation = false;
2724 for (const auto &MI : Info.Matchables) {
2725 MaxNumOperands = std::max(MaxNumOperands, MI.AsmOperands.size());
2726 HasDeprecation |= MI.HasDeprecation;
2728 // Store a pascal-style length byte in the mnemonic.
2729 std::string LenMnemonic = char(MI.Mnemonic.size()) + MI.Mnemonic.str();
2730 MaxMnemonicIndex = std::max(MaxMnemonicIndex,
2731 StringTable.GetOrAddStringOffset(LenMnemonic, false));
2734 OS << "static const char *const MnemonicTable =\n";
2735 StringTable.EmitString(OS);
2738 // Emit the static match table; unused classes get initalized to 0 which is
2739 // guaranteed to be InvalidMatchClass.
2741 // FIXME: We can reduce the size of this table very easily. First, we change
2742 // it so that store the kinds in separate bit-fields for each index, which
2743 // only needs to be the max width used for classes at that index (we also need
2744 // to reject based on this during classification). If we then make sure to
2745 // order the match kinds appropriately (putting mnemonics last), then we
2746 // should only end up using a few bits for each class, especially the ones
2747 // following the mnemonic.
2748 OS << "namespace {\n";
2749 OS << " struct MatchEntry {\n";
2750 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
2752 OS << " uint16_t Opcode;\n";
2753 OS << " " << getMinimalTypeForRange(std::distance(Info.Matchables.begin(),
2754 Info.Matchables.end()))
2756 OS << " " << getMinimalRequiredFeaturesType(Info)
2757 << " RequiredFeatures;\n";
2758 OS << " " << getMinimalTypeForRange(
2759 std::distance(Info.Classes.begin(), Info.Classes.end()))
2760 << " Classes[" << MaxNumOperands << "];\n";
2761 OS << " StringRef getMnemonic() const {\n";
2762 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
2763 OS << " MnemonicTable[Mnemonic]);\n";
2767 OS << " // Predicate for searching for an opcode.\n";
2768 OS << " struct LessOpcode {\n";
2769 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
2770 OS << " return LHS.getMnemonic() < RHS;\n";
2772 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
2773 OS << " return LHS < RHS.getMnemonic();\n";
2775 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
2776 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
2780 OS << "} // end anonymous namespace.\n\n";
2782 unsigned VariantCount = Target.getAsmParserVariantCount();
2783 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2784 Record *AsmVariant = Target.getAsmParserVariant(VC);
2785 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2787 OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
2789 for (const auto &MI : Info.Matchables) {
2790 if (MI.AsmVariantID != AsmVariantNo)
2793 // Store a pascal-style length byte in the mnemonic.
2794 std::string LenMnemonic = char(MI.Mnemonic.size()) + MI.Mnemonic.str();
2795 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2796 << " /* " << MI.Mnemonic << " */, " << Target.getName()
2797 << "::" << MI.getResultInst()->TheDef->getName() << ", "
2798 << MI.ConversionFnKind << ", ";
2800 // Write the required features mask.
2801 if (!MI.RequiredFeatures.empty()) {
2802 for (unsigned i = 0, e = MI.RequiredFeatures.size(); i != e; ++i) {
2804 OS << MI.RequiredFeatures[i]->getEnumName();
2810 for (unsigned i = 0, e = MI.AsmOperands.size(); i != e; ++i) {
2811 const MatchableInfo::AsmOperand &Op = MI.AsmOperands[i];
2814 OS << Op.Class->Name;
2822 // A method to determine if a mnemonic is in the list.
2823 OS << "bool " << Target.getName() << ClassName << "::\n"
2824 << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n";
2825 OS << " // Find the appropriate table for this asm variant.\n";
2826 OS << " const MatchEntry *Start, *End;\n";
2827 OS << " switch (VariantID) {\n";
2828 OS << " default: // unreachable\n";
2829 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2830 Record *AsmVariant = Target.getAsmParserVariant(VC);
2831 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2832 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
2833 << "); End = std::end(MatchTable" << VC << "); break;\n";
2836 OS << " // Search the table.\n";
2837 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2838 OS << " std::equal_range(Start, End, Mnemonic, LessOpcode());\n";
2839 OS << " return MnemonicRange.first != MnemonicRange.second;\n";
2842 // Finally, build the match function.
2843 OS << "unsigned " << Target.getName() << ClassName << "::\n"
2844 << "MatchInstructionImpl(const OperandVector"
2846 OS << " MCInst &Inst,\n"
2847 << "uint64_t &ErrorInfo, bool matchingInlineAsm, unsigned VariantID) {\n";
2849 OS << " // Eliminate obvious mismatches.\n";
2850 OS << " if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
2851 OS << " ErrorInfo = " << (MaxNumOperands+1) << ";\n";
2852 OS << " return Match_InvalidOperand;\n";
2855 // Emit code to get the available features.
2856 OS << " // Get the current feature set.\n";
2857 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2859 OS << " // Get the instruction mnemonic, which is the first token.\n";
2860 OS << " StringRef Mnemonic = ((" << Target.getName()
2861 << "Operand&)*Operands[0]).getToken();\n\n";
2863 if (HasMnemonicAliases) {
2864 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
2865 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
2868 // Emit code to compute the class list for this operand vector.
2869 OS << " // Some state to try to produce better error messages.\n";
2870 OS << " bool HadMatchOtherThanFeatures = false;\n";
2871 OS << " bool HadMatchOtherThanPredicate = false;\n";
2872 OS << " unsigned RetCode = Match_InvalidOperand;\n";
2873 OS << " uint64_t MissingFeatures = ~0ULL;\n";
2874 OS << " // Set ErrorInfo to the operand that mismatches if it is\n";
2875 OS << " // wrong for all instances of the instruction.\n";
2876 OS << " ErrorInfo = ~0U;\n";
2878 // Emit code to search the table.
2879 OS << " // Find the appropriate table for this asm variant.\n";
2880 OS << " const MatchEntry *Start, *End;\n";
2881 OS << " switch (VariantID) {\n";
2882 OS << " default: // unreachable\n";
2883 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2884 Record *AsmVariant = Target.getAsmParserVariant(VC);
2885 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2886 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
2887 << "); End = std::end(MatchTable" << VC << "); break;\n";
2890 OS << " // Search the table.\n";
2891 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2892 OS << " std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
2894 OS << " // Return a more specific error code if no mnemonics match.\n";
2895 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
2896 OS << " return Match_MnemonicFail;\n\n";
2898 OS << " for (const MatchEntry *it = MnemonicRange.first, "
2899 << "*ie = MnemonicRange.second;\n";
2900 OS << " it != ie; ++it) {\n";
2902 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
2903 OS << " assert(Mnemonic == it->getMnemonic());\n";
2905 // Emit check that the subclasses match.
2906 OS << " bool OperandsValid = true;\n";
2907 OS << " for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
2908 OS << " if (i + 1 >= Operands.size()) {\n";
2909 OS << " OperandsValid = (it->Classes[i] == " <<"InvalidMatchClass);\n";
2910 OS << " if (!OperandsValid) ErrorInfo = i + 1;\n";
2913 OS << " unsigned Diag = validateOperandClass(*Operands[i+1],\n";
2915 OS << "(MatchClassKind)it->Classes[i]);\n";
2916 OS << " if (Diag == Match_Success)\n";
2917 OS << " continue;\n";
2918 OS << " // If the generic handler indicates an invalid operand\n";
2919 OS << " // failure, check for a special case.\n";
2920 OS << " if (Diag == Match_InvalidOperand) {\n";
2921 OS << " Diag = validateTargetOperandClass(*Operands[i+1],\n";
2923 OS << "(MatchClassKind)it->Classes[i]);\n";
2924 OS << " if (Diag == Match_Success)\n";
2925 OS << " continue;\n";
2927 OS << " // If this operand is broken for all of the instances of this\n";
2928 OS << " // mnemonic, keep track of it so we can report loc info.\n";
2929 OS << " // If we already had a match that only failed due to a\n";
2930 OS << " // target predicate, that diagnostic is preferred.\n";
2931 OS << " if (!HadMatchOtherThanPredicate &&\n";
2932 OS << " (it == MnemonicRange.first || ErrorInfo <= i+1)) {\n";
2933 OS << " ErrorInfo = i+1;\n";
2934 OS << " // InvalidOperand is the default. Prefer specificity.\n";
2935 OS << " if (Diag != Match_InvalidOperand)\n";
2936 OS << " RetCode = Diag;\n";
2938 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
2939 OS << " OperandsValid = false;\n";
2943 OS << " if (!OperandsValid) continue;\n";
2945 // Emit check that the required features are available.
2946 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
2947 << "!= it->RequiredFeatures) {\n";
2948 OS << " HadMatchOtherThanFeatures = true;\n";
2949 OS << " uint64_t NewMissingFeatures = it->RequiredFeatures & "
2950 "~AvailableFeatures;\n";
2951 OS << " if (CountPopulation_64(NewMissingFeatures) <=\n"
2952 " CountPopulation_64(MissingFeatures))\n";
2953 OS << " MissingFeatures = NewMissingFeatures;\n";
2954 OS << " continue;\n";
2957 OS << " if (matchingInlineAsm) {\n";
2958 OS << " Inst.setOpcode(it->Opcode);\n";
2959 OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n";
2960 OS << " return Match_Success;\n";
2962 OS << " // We have selected a definite instruction, convert the parsed\n"
2963 << " // operands into the appropriate MCInst.\n";
2964 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
2967 // Verify the instruction with the target-specific match predicate function.
2968 OS << " // We have a potential match. Check the target predicate to\n"
2969 << " // handle any context sensitive constraints.\n"
2970 << " unsigned MatchResult;\n"
2971 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
2972 << " Match_Success) {\n"
2973 << " Inst.clear();\n"
2974 << " RetCode = MatchResult;\n"
2975 << " HadMatchOtherThanPredicate = true;\n"
2979 // Call the post-processing function, if used.
2980 std::string InsnCleanupFn =
2981 AsmParser->getValueAsString("AsmParserInstCleanup");
2982 if (!InsnCleanupFn.empty())
2983 OS << " " << InsnCleanupFn << "(Inst);\n";
2985 if (HasDeprecation) {
2986 OS << " std::string Info;\n";
2987 OS << " if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, STI, Info)) {\n";
2988 OS << " SMLoc Loc = ((" << Target.getName()
2989 << "Operand&)*Operands[0]).getStartLoc();\n";
2990 OS << " getParser().Warning(Loc, Info, None);\n";
2994 OS << " return Match_Success;\n";
2997 OS << " // Okay, we had no match. Try to return a useful error code.\n";
2998 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
2999 OS << " return RetCode;\n\n";
3000 OS << " // Missing feature matches return which features were missing\n";
3001 OS << " ErrorInfo = MissingFeatures;\n";
3002 OS << " return Match_MissingFeature;\n";
3005 if (Info.OperandMatchInfo.size())
3006 emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
3009 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
3014 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
3015 emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
3016 AsmMatcherEmitter(RK).run(OS);
3019 } // End llvm namespace