[llvm-readobj] - Implement LLVM-style dumping for .stack_sizes sections.
[llvm-complete.git] / utils / TableGen / AsmMatcherEmitter.cpp
blobd32366e2986d22a2e263090bb788f03a5a3b7dfb
1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This tablegen backend emits a target specifier matcher for converting parsed
10 // assembly operands in the MCInst structures. It also emits a matcher for
11 // custom operand parsing.
13 // Converting assembly operands into MCInst structures
14 // ---------------------------------------------------
16 // The input to the target specific matcher is a list of literal tokens and
17 // operands. The target specific parser should generally eliminate any syntax
18 // which is not relevant for matching; for example, comma tokens should have
19 // already been consumed and eliminated by the parser. Most instructions will
20 // end up with a single literal token (the instruction name) and some number of
21 // operands.
23 // Some example inputs, for X86:
24 // 'addl' (immediate ...) (register ...)
25 // 'add' (immediate ...) (memory ...)
26 // 'call' '*' %epc
28 // The assembly matcher is responsible for converting this input into a precise
29 // machine instruction (i.e., an instruction with a well defined encoding). This
30 // mapping has several properties which complicate matching:
32 // - It may be ambiguous; many architectures can legally encode particular
33 // variants of an instruction in different ways (for example, using a smaller
34 // encoding for small immediates). Such ambiguities should never be
35 // arbitrarily resolved by the assembler, the assembler is always responsible
36 // for choosing the "best" available instruction.
38 // - It may depend on the subtarget or the assembler context. Instructions
39 // which are invalid for the current mode, but otherwise unambiguous (e.g.,
40 // an SSE instruction in a file being assembled for i486) should be accepted
41 // and rejected by the assembler front end. However, if the proper encoding
42 // for an instruction is dependent on the assembler context then the matcher
43 // is responsible for selecting the correct machine instruction for the
44 // current mode.
46 // The core matching algorithm attempts to exploit the regularity in most
47 // instruction sets to quickly determine the set of possibly matching
48 // instructions, and the simplify the generated code. Additionally, this helps
49 // to ensure that the ambiguities are intentionally resolved by the user.
51 // The matching is divided into two distinct phases:
53 // 1. Classification: Each operand is mapped to the unique set which (a)
54 // contains it, and (b) is the largest such subset for which a single
55 // instruction could match all members.
57 // For register classes, we can generate these subgroups automatically. For
58 // arbitrary operands, we expect the user to define the classes and their
59 // relations to one another (for example, 8-bit signed immediates as a
60 // subset of 32-bit immediates).
62 // By partitioning the operands in this way, we guarantee that for any
63 // tuple of classes, any single instruction must match either all or none
64 // of the sets of operands which could classify to that tuple.
66 // In addition, the subset relation amongst classes induces a partial order
67 // on such tuples, which we use to resolve ambiguities.
69 // 2. The input can now be treated as a tuple of classes (static tokens are
70 // simple singleton sets). Each such tuple should generally map to a single
71 // instruction (we currently ignore cases where this isn't true, whee!!!),
72 // which we can emit a simple matcher for.
74 // Custom Operand Parsing
75 // ----------------------
77 // Some targets need a custom way to parse operands, some specific instructions
78 // can contain arguments that can represent processor flags and other kinds of
79 // identifiers that need to be mapped to specific values in the final encoded
80 // instructions. The target specific custom operand parsing works in the
81 // following way:
83 // 1. A operand match table is built, each entry contains a mnemonic, an
84 // operand class, a mask for all operand positions for that same
85 // class/mnemonic and target features to be checked while trying to match.
87 // 2. The operand matcher will try every possible entry with the same
88 // mnemonic and will check if the target feature for this mnemonic also
89 // matches. After that, if the operand to be matched has its index
90 // present in the mask, a successful match occurs. Otherwise, fallback
91 // to the regular operand parsing.
93 // 3. For a match success, each operand class that has a 'ParserMethod'
94 // becomes part of a switch from where the custom method is called.
96 //===----------------------------------------------------------------------===//
98 #include "CodeGenTarget.h"
99 #include "SubtargetFeatureInfo.h"
100 #include "Types.h"
101 #include "llvm/ADT/CachedHashString.h"
102 #include "llvm/ADT/PointerUnion.h"
103 #include "llvm/ADT/STLExtras.h"
104 #include "llvm/ADT/SmallPtrSet.h"
105 #include "llvm/ADT/SmallVector.h"
106 #include "llvm/ADT/StringExtras.h"
107 #include "llvm/Config/llvm-config.h"
108 #include "llvm/Support/CommandLine.h"
109 #include "llvm/Support/Debug.h"
110 #include "llvm/Support/ErrorHandling.h"
111 #include "llvm/TableGen/Error.h"
112 #include "llvm/TableGen/Record.h"
113 #include "llvm/TableGen/StringMatcher.h"
114 #include "llvm/TableGen/StringToOffsetTable.h"
115 #include "llvm/TableGen/TableGenBackend.h"
116 #include <cassert>
117 #include <cctype>
118 #include <forward_list>
119 #include <map>
120 #include <set>
122 using namespace llvm;
124 #define DEBUG_TYPE "asm-matcher-emitter"
126 cl::OptionCategory AsmMatcherEmitterCat("Options for -gen-asm-matcher");
128 static cl::opt<std::string>
129 MatchPrefix("match-prefix", cl::init(""),
130 cl::desc("Only match instructions with the given prefix"),
131 cl::cat(AsmMatcherEmitterCat));
133 namespace {
134 class AsmMatcherInfo;
136 // Register sets are used as keys in some second-order sets TableGen creates
137 // when generating its data structures. This means that the order of two
138 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and
139 // can even affect compiler output (at least seen in diagnostics produced when
140 // all matches fail). So we use a type that sorts them consistently.
141 typedef std::set<Record*, LessRecordByID> RegisterSet;
143 class AsmMatcherEmitter {
144 RecordKeeper &Records;
145 public:
146 AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
148 void run(raw_ostream &o);
151 /// ClassInfo - Helper class for storing the information about a particular
152 /// class of operands which can be matched.
153 struct ClassInfo {
154 enum ClassInfoKind {
155 /// Invalid kind, for use as a sentinel value.
156 Invalid = 0,
158 /// The class for a particular token.
159 Token,
161 /// The (first) register class, subsequent register classes are
162 /// RegisterClass0+1, and so on.
163 RegisterClass0,
165 /// The (first) user defined class, subsequent user defined classes are
166 /// UserClass0+1, and so on.
167 UserClass0 = 1<<16
170 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
171 /// N) for the Nth user defined class.
172 unsigned Kind;
174 /// SuperClasses - The super classes of this class. Note that for simplicities
175 /// sake user operands only record their immediate super class, while register
176 /// operands include all superclasses.
177 std::vector<ClassInfo*> SuperClasses;
179 /// Name - The full class name, suitable for use in an enum.
180 std::string Name;
182 /// ClassName - The unadorned generic name for this class (e.g., Token).
183 std::string ClassName;
185 /// ValueName - The name of the value this class represents; for a token this
186 /// is the literal token string, for an operand it is the TableGen class (or
187 /// empty if this is a derived class).
188 std::string ValueName;
190 /// PredicateMethod - The name of the operand method to test whether the
191 /// operand matches this class; this is not valid for Token or register kinds.
192 std::string PredicateMethod;
194 /// RenderMethod - The name of the operand method to add this operand to an
195 /// MCInst; this is not valid for Token or register kinds.
196 std::string RenderMethod;
198 /// ParserMethod - The name of the operand method to do a target specific
199 /// parsing on the operand.
200 std::string ParserMethod;
202 /// For register classes: the records for all the registers in this class.
203 RegisterSet Registers;
205 /// For custom match classes: the diagnostic kind for when the predicate fails.
206 std::string DiagnosticType;
208 /// For custom match classes: the diagnostic string for when the predicate fails.
209 std::string DiagnosticString;
211 /// Is this operand optional and not always required.
212 bool IsOptional;
214 /// DefaultMethod - The name of the method that returns the default operand
215 /// for optional operand
216 std::string DefaultMethod;
218 public:
219 /// isRegisterClass() - Check if this is a register class.
220 bool isRegisterClass() const {
221 return Kind >= RegisterClass0 && Kind < UserClass0;
224 /// isUserClass() - Check if this is a user defined class.
225 bool isUserClass() const {
226 return Kind >= UserClass0;
229 /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
230 /// are related if they are in the same class hierarchy.
231 bool isRelatedTo(const ClassInfo &RHS) const {
232 // Tokens are only related to tokens.
233 if (Kind == Token || RHS.Kind == Token)
234 return Kind == Token && RHS.Kind == Token;
236 // Registers classes are only related to registers classes, and only if
237 // their intersection is non-empty.
238 if (isRegisterClass() || RHS.isRegisterClass()) {
239 if (!isRegisterClass() || !RHS.isRegisterClass())
240 return false;
242 RegisterSet Tmp;
243 std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin());
244 std::set_intersection(Registers.begin(), Registers.end(),
245 RHS.Registers.begin(), RHS.Registers.end(),
246 II, LessRecordByID());
248 return !Tmp.empty();
251 // Otherwise we have two users operands; they are related if they are in the
252 // same class hierarchy.
254 // FIXME: This is an oversimplification, they should only be related if they
255 // intersect, however we don't have that information.
256 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
257 const ClassInfo *Root = this;
258 while (!Root->SuperClasses.empty())
259 Root = Root->SuperClasses.front();
261 const ClassInfo *RHSRoot = &RHS;
262 while (!RHSRoot->SuperClasses.empty())
263 RHSRoot = RHSRoot->SuperClasses.front();
265 return Root == RHSRoot;
268 /// isSubsetOf - Test whether this class is a subset of \p RHS.
269 bool isSubsetOf(const ClassInfo &RHS) const {
270 // This is a subset of RHS if it is the same class...
271 if (this == &RHS)
272 return true;
274 // ... or if any of its super classes are a subset of RHS.
275 SmallVector<const ClassInfo *, 16> Worklist(SuperClasses.begin(),
276 SuperClasses.end());
277 SmallPtrSet<const ClassInfo *, 16> Visited;
278 while (!Worklist.empty()) {
279 auto *CI = Worklist.pop_back_val();
280 if (CI == &RHS)
281 return true;
282 for (auto *Super : CI->SuperClasses)
283 if (Visited.insert(Super).second)
284 Worklist.push_back(Super);
287 return false;
290 int getTreeDepth() const {
291 int Depth = 0;
292 const ClassInfo *Root = this;
293 while (!Root->SuperClasses.empty()) {
294 Depth++;
295 Root = Root->SuperClasses.front();
297 return Depth;
300 const ClassInfo *findRoot() const {
301 const ClassInfo *Root = this;
302 while (!Root->SuperClasses.empty())
303 Root = Root->SuperClasses.front();
304 return Root;
307 /// Compare two classes. This does not produce a total ordering, but does
308 /// guarantee that subclasses are sorted before their parents, and that the
309 /// ordering is transitive.
310 bool operator<(const ClassInfo &RHS) const {
311 if (this == &RHS)
312 return false;
314 // First, enforce the ordering between the three different types of class.
315 // Tokens sort before registers, which sort before user classes.
316 if (Kind == Token) {
317 if (RHS.Kind != Token)
318 return true;
319 assert(RHS.Kind == Token);
320 } else if (isRegisterClass()) {
321 if (RHS.Kind == Token)
322 return false;
323 else if (RHS.isUserClass())
324 return true;
325 assert(RHS.isRegisterClass());
326 } else if (isUserClass()) {
327 if (!RHS.isUserClass())
328 return false;
329 assert(RHS.isUserClass());
330 } else {
331 llvm_unreachable("Unknown ClassInfoKind");
334 if (Kind == Token || isUserClass()) {
335 // Related tokens and user classes get sorted by depth in the inheritence
336 // tree (so that subclasses are before their parents).
337 if (isRelatedTo(RHS)) {
338 if (getTreeDepth() > RHS.getTreeDepth())
339 return true;
340 if (getTreeDepth() < RHS.getTreeDepth())
341 return false;
342 } else {
343 // Unrelated tokens and user classes are ordered by the name of their
344 // root nodes, so that there is a consistent ordering between
345 // unconnected trees.
346 return findRoot()->ValueName < RHS.findRoot()->ValueName;
348 } else if (isRegisterClass()) {
349 // For register sets, sort by number of registers. This guarantees that
350 // a set will always sort before all of it's strict supersets.
351 if (Registers.size() != RHS.Registers.size())
352 return Registers.size() < RHS.Registers.size();
353 } else {
354 llvm_unreachable("Unknown ClassInfoKind");
357 // FIXME: We should be able to just return false here, as we only need a
358 // partial order (we use stable sorts, so this is deterministic) and the
359 // name of a class shouldn't be significant. However, some of the backends
360 // accidentally rely on this behaviour, so it will have to stay like this
361 // until they are fixed.
362 return ValueName < RHS.ValueName;
366 class AsmVariantInfo {
367 public:
368 StringRef RegisterPrefix;
369 StringRef TokenizingCharacters;
370 StringRef SeparatorCharacters;
371 StringRef BreakCharacters;
372 StringRef Name;
373 int AsmVariantNo;
376 /// MatchableInfo - Helper class for storing the necessary information for an
377 /// instruction or alias which is capable of being matched.
378 struct MatchableInfo {
379 struct AsmOperand {
380 /// Token - This is the token that the operand came from.
381 StringRef Token;
383 /// The unique class instance this operand should match.
384 ClassInfo *Class;
386 /// The operand name this is, if anything.
387 StringRef SrcOpName;
389 /// The operand name this is, before renaming for tied operands.
390 StringRef OrigSrcOpName;
392 /// The suboperand index within SrcOpName, or -1 for the entire operand.
393 int SubOpIdx;
395 /// Whether the token is "isolated", i.e., it is preceded and followed
396 /// by separators.
397 bool IsIsolatedToken;
399 /// Register record if this token is singleton register.
400 Record *SingletonReg;
402 explicit AsmOperand(bool IsIsolatedToken, StringRef T)
403 : Token(T), Class(nullptr), SubOpIdx(-1),
404 IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {}
407 /// ResOperand - This represents a single operand in the result instruction
408 /// generated by the match. In cases (like addressing modes) where a single
409 /// assembler operand expands to multiple MCOperands, this represents the
410 /// single assembler operand, not the MCOperand.
411 struct ResOperand {
412 enum {
413 /// RenderAsmOperand - This represents an operand result that is
414 /// generated by calling the render method on the assembly operand. The
415 /// corresponding AsmOperand is specified by AsmOperandNum.
416 RenderAsmOperand,
418 /// TiedOperand - This represents a result operand that is a duplicate of
419 /// a previous result operand.
420 TiedOperand,
422 /// ImmOperand - This represents an immediate value that is dumped into
423 /// the operand.
424 ImmOperand,
426 /// RegOperand - This represents a fixed register that is dumped in.
427 RegOperand
428 } Kind;
430 /// Tuple containing the index of the (earlier) result operand that should
431 /// be copied from, as well as the indices of the corresponding (parsed)
432 /// operands in the asm string.
433 struct TiedOperandsTuple {
434 unsigned ResOpnd;
435 unsigned SrcOpnd1Idx;
436 unsigned SrcOpnd2Idx;
439 union {
440 /// This is the operand # in the AsmOperands list that this should be
441 /// copied from.
442 unsigned AsmOperandNum;
444 /// Description of tied operands.
445 TiedOperandsTuple TiedOperands;
447 /// ImmVal - This is the immediate value added to the instruction.
448 int64_t ImmVal;
450 /// Register - This is the register record.
451 Record *Register;
454 /// MINumOperands - The number of MCInst operands populated by this
455 /// operand.
456 unsigned MINumOperands;
458 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
459 ResOperand X;
460 X.Kind = RenderAsmOperand;
461 X.AsmOperandNum = AsmOpNum;
462 X.MINumOperands = NumOperands;
463 return X;
466 static ResOperand getTiedOp(unsigned TiedOperandNum, unsigned SrcOperand1,
467 unsigned SrcOperand2) {
468 ResOperand X;
469 X.Kind = TiedOperand;
470 X.TiedOperands = { TiedOperandNum, SrcOperand1, SrcOperand2 };
471 X.MINumOperands = 1;
472 return X;
475 static ResOperand getImmOp(int64_t Val) {
476 ResOperand X;
477 X.Kind = ImmOperand;
478 X.ImmVal = Val;
479 X.MINumOperands = 1;
480 return X;
483 static ResOperand getRegOp(Record *Reg) {
484 ResOperand X;
485 X.Kind = RegOperand;
486 X.Register = Reg;
487 X.MINumOperands = 1;
488 return X;
492 /// AsmVariantID - Target's assembly syntax variant no.
493 int AsmVariantID;
495 /// AsmString - The assembly string for this instruction (with variants
496 /// removed), e.g. "movsx $src, $dst".
497 std::string AsmString;
499 /// TheDef - This is the definition of the instruction or InstAlias that this
500 /// matchable came from.
501 Record *const TheDef;
503 /// DefRec - This is the definition that it came from.
504 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
506 const CodeGenInstruction *getResultInst() const {
507 if (DefRec.is<const CodeGenInstruction*>())
508 return DefRec.get<const CodeGenInstruction*>();
509 return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
512 /// ResOperands - This is the operand list that should be built for the result
513 /// MCInst.
514 SmallVector<ResOperand, 8> ResOperands;
516 /// Mnemonic - This is the first token of the matched instruction, its
517 /// mnemonic.
518 StringRef Mnemonic;
520 /// AsmOperands - The textual operands that this instruction matches,
521 /// annotated with a class and where in the OperandList they were defined.
522 /// This directly corresponds to the tokenized AsmString after the mnemonic is
523 /// removed.
524 SmallVector<AsmOperand, 8> AsmOperands;
526 /// Predicates - The required subtarget features to match this instruction.
527 SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
529 /// ConversionFnKind - The enum value which is passed to the generated
530 /// convertToMCInst to convert parsed operands into an MCInst for this
531 /// function.
532 std::string ConversionFnKind;
534 /// If this instruction is deprecated in some form.
535 bool HasDeprecation;
537 /// If this is an alias, this is use to determine whether or not to using
538 /// the conversion function defined by the instruction's AsmMatchConverter
539 /// or to use the function generated by the alias.
540 bool UseInstAsmMatchConverter;
542 MatchableInfo(const CodeGenInstruction &CGI)
543 : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI),
544 UseInstAsmMatchConverter(true) {
547 MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias)
548 : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef),
549 DefRec(Alias.release()),
550 UseInstAsmMatchConverter(
551 TheDef->getValueAsBit("UseInstAsmMatchConverter")) {
554 // Could remove this and the dtor if PointerUnion supported unique_ptr
555 // elements with a dynamic failure/assertion (like the one below) in the case
556 // where it was copied while being in an owning state.
557 MatchableInfo(const MatchableInfo &RHS)
558 : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString),
559 TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands),
560 Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands),
561 RequiredFeatures(RHS.RequiredFeatures),
562 ConversionFnKind(RHS.ConversionFnKind),
563 HasDeprecation(RHS.HasDeprecation),
564 UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) {
565 assert(!DefRec.is<const CodeGenInstAlias *>());
568 ~MatchableInfo() {
569 delete DefRec.dyn_cast<const CodeGenInstAlias*>();
572 // Two-operand aliases clone from the main matchable, but mark the second
573 // operand as a tied operand of the first for purposes of the assembler.
574 void formTwoOperandAlias(StringRef Constraint);
576 void initialize(const AsmMatcherInfo &Info,
577 SmallPtrSetImpl<Record*> &SingletonRegisters,
578 AsmVariantInfo const &Variant,
579 bool HasMnemonicFirst);
581 /// validate - Return true if this matchable is a valid thing to match against
582 /// and perform a bunch of validity checking.
583 bool validate(StringRef CommentDelimiter, bool IsAlias) const;
585 /// findAsmOperand - Find the AsmOperand with the specified name and
586 /// suboperand index.
587 int findAsmOperand(StringRef N, int SubOpIdx) const {
588 auto I = find_if(AsmOperands, [&](const AsmOperand &Op) {
589 return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx;
591 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
594 /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
595 /// This does not check the suboperand index.
596 int findAsmOperandNamed(StringRef N, int LastIdx = -1) const {
597 auto I = std::find_if(AsmOperands.begin() + LastIdx + 1, AsmOperands.end(),
598 [&](const AsmOperand &Op) { return Op.SrcOpName == N; });
599 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
602 int findAsmOperandOriginallyNamed(StringRef N) const {
603 auto I =
604 find_if(AsmOperands,
605 [&](const AsmOperand &Op) { return Op.OrigSrcOpName == N; });
606 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
609 void buildInstructionResultOperands();
610 void buildAliasResultOperands(bool AliasConstraintsAreChecked);
612 /// operator< - Compare two matchables.
613 bool operator<(const MatchableInfo &RHS) const {
614 // The primary comparator is the instruction mnemonic.
615 if (int Cmp = Mnemonic.compare(RHS.Mnemonic))
616 return Cmp == -1;
618 if (AsmOperands.size() != RHS.AsmOperands.size())
619 return AsmOperands.size() < RHS.AsmOperands.size();
621 // Compare lexicographically by operand. The matcher validates that other
622 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
623 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
624 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
625 return true;
626 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
627 return false;
630 // Give matches that require more features higher precedence. This is useful
631 // because we cannot define AssemblerPredicates with the negation of
632 // processor features. For example, ARM v6 "nop" may be either a HINT or
633 // MOV. With v6, we want to match HINT. The assembler has no way to
634 // predicate MOV under "NoV6", but HINT will always match first because it
635 // requires V6 while MOV does not.
636 if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
637 return RequiredFeatures.size() > RHS.RequiredFeatures.size();
639 return false;
642 /// couldMatchAmbiguouslyWith - Check whether this matchable could
643 /// ambiguously match the same set of operands as \p RHS (without being a
644 /// strictly superior match).
645 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const {
646 // The primary comparator is the instruction mnemonic.
647 if (Mnemonic != RHS.Mnemonic)
648 return false;
650 // Different variants can't conflict.
651 if (AsmVariantID != RHS.AsmVariantID)
652 return false;
654 // The number of operands is unambiguous.
655 if (AsmOperands.size() != RHS.AsmOperands.size())
656 return false;
658 // Otherwise, make sure the ordering of the two instructions is unambiguous
659 // by checking that either (a) a token or operand kind discriminates them,
660 // or (b) the ordering among equivalent kinds is consistent.
662 // Tokens and operand kinds are unambiguous (assuming a correct target
663 // specific parser).
664 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
665 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
666 AsmOperands[i].Class->Kind == ClassInfo::Token)
667 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
668 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
669 return false;
671 // Otherwise, this operand could commute if all operands are equivalent, or
672 // there is a pair of operands that compare less than and a pair that
673 // compare greater than.
674 bool HasLT = false, HasGT = false;
675 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
676 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
677 HasLT = true;
678 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
679 HasGT = true;
682 return HasLT == HasGT;
685 void dump() const;
687 private:
688 void tokenizeAsmString(AsmMatcherInfo const &Info,
689 AsmVariantInfo const &Variant);
690 void addAsmOperand(StringRef Token, bool IsIsolatedToken = false);
693 struct OperandMatchEntry {
694 unsigned OperandMask;
695 const MatchableInfo* MI;
696 ClassInfo *CI;
698 static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
699 unsigned opMask) {
700 OperandMatchEntry X;
701 X.OperandMask = opMask;
702 X.CI = ci;
703 X.MI = mi;
704 return X;
708 class AsmMatcherInfo {
709 public:
710 /// Tracked Records
711 RecordKeeper &Records;
713 /// The tablegen AsmParser record.
714 Record *AsmParser;
716 /// Target - The target information.
717 CodeGenTarget &Target;
719 /// The classes which are needed for matching.
720 std::forward_list<ClassInfo> Classes;
722 /// The information on the matchables to match.
723 std::vector<std::unique_ptr<MatchableInfo>> Matchables;
725 /// Info for custom matching operands by user defined methods.
726 std::vector<OperandMatchEntry> OperandMatchInfo;
728 /// Map of Register records to their class information.
729 typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy;
730 RegisterClassesTy RegisterClasses;
732 /// Map of Predicate records to their subtarget information.
733 std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
735 /// Map of AsmOperandClass records to their class information.
736 std::map<Record*, ClassInfo*> AsmOperandClasses;
738 /// Map of RegisterClass records to their class information.
739 std::map<Record*, ClassInfo*> RegisterClassClasses;
741 private:
742 /// Map of token to class information which has already been constructed.
743 std::map<std::string, ClassInfo*> TokenClasses;
745 private:
746 /// getTokenClass - Lookup or create the class for the given token.
747 ClassInfo *getTokenClass(StringRef Token);
749 /// getOperandClass - Lookup or create the class for the given operand.
750 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
751 int SubOpIdx);
752 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
754 /// buildRegisterClasses - Build the ClassInfo* instances for register
755 /// classes.
756 void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters);
758 /// buildOperandClasses - Build the ClassInfo* instances for user defined
759 /// operand classes.
760 void buildOperandClasses();
762 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
763 unsigned AsmOpIdx);
764 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
765 MatchableInfo::AsmOperand &Op);
767 public:
768 AsmMatcherInfo(Record *AsmParser,
769 CodeGenTarget &Target,
770 RecordKeeper &Records);
772 /// Construct the various tables used during matching.
773 void buildInfo();
775 /// buildOperandMatchInfo - Build the necessary information to handle user
776 /// defined operand parsing methods.
777 void buildOperandMatchInfo();
779 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
780 /// given operand.
781 const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
782 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
783 const auto &I = SubtargetFeatures.find(Def);
784 return I == SubtargetFeatures.end() ? nullptr : &I->second;
787 RecordKeeper &getRecords() const {
788 return Records;
791 bool hasOptionalOperands() const {
792 return find_if(Classes, [](const ClassInfo &Class) {
793 return Class.IsOptional;
794 }) != Classes.end();
798 } // end anonymous namespace
800 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
801 LLVM_DUMP_METHOD void MatchableInfo::dump() const {
802 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
804 errs() << " variant: " << AsmVariantID << "\n";
806 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
807 const AsmOperand &Op = AsmOperands[i];
808 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
809 errs() << '\"' << Op.Token << "\"\n";
812 #endif
814 static std::pair<StringRef, StringRef>
815 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
816 // Split via the '='.
817 std::pair<StringRef, StringRef> Ops = S.split('=');
818 if (Ops.second == "")
819 PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
820 // Trim whitespace and the leading '$' on the operand names.
821 size_t start = Ops.first.find_first_of('$');
822 if (start == std::string::npos)
823 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
824 Ops.first = Ops.first.slice(start + 1, std::string::npos);
825 size_t end = Ops.first.find_last_of(" \t");
826 Ops.first = Ops.first.slice(0, end);
827 // Now the second operand.
828 start = Ops.second.find_first_of('$');
829 if (start == std::string::npos)
830 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
831 Ops.second = Ops.second.slice(start + 1, std::string::npos);
832 end = Ops.second.find_last_of(" \t");
833 Ops.first = Ops.first.slice(0, end);
834 return Ops;
837 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
838 // Figure out which operands are aliased and mark them as tied.
839 std::pair<StringRef, StringRef> Ops =
840 parseTwoOperandConstraint(Constraint, TheDef->getLoc());
842 // Find the AsmOperands that refer to the operands we're aliasing.
843 int SrcAsmOperand = findAsmOperandNamed(Ops.first);
844 int DstAsmOperand = findAsmOperandNamed(Ops.second);
845 if (SrcAsmOperand == -1)
846 PrintFatalError(TheDef->getLoc(),
847 "unknown source two-operand alias operand '" + Ops.first +
848 "'.");
849 if (DstAsmOperand == -1)
850 PrintFatalError(TheDef->getLoc(),
851 "unknown destination two-operand alias operand '" +
852 Ops.second + "'.");
854 // Find the ResOperand that refers to the operand we're aliasing away
855 // and update it to refer to the combined operand instead.
856 for (ResOperand &Op : ResOperands) {
857 if (Op.Kind == ResOperand::RenderAsmOperand &&
858 Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
859 Op.AsmOperandNum = DstAsmOperand;
860 break;
863 // Remove the AsmOperand for the alias operand.
864 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
865 // Adjust the ResOperand references to any AsmOperands that followed
866 // the one we just deleted.
867 for (ResOperand &Op : ResOperands) {
868 switch(Op.Kind) {
869 default:
870 // Nothing to do for operands that don't reference AsmOperands.
871 break;
872 case ResOperand::RenderAsmOperand:
873 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
874 --Op.AsmOperandNum;
875 break;
880 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
881 /// if present, from specified token.
882 static void
883 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op,
884 const AsmMatcherInfo &Info,
885 StringRef RegisterPrefix) {
886 StringRef Tok = Op.Token;
888 // If this token is not an isolated token, i.e., it isn't separated from
889 // other tokens (e.g. with whitespace), don't interpret it as a register name.
890 if (!Op.IsIsolatedToken)
891 return;
893 if (RegisterPrefix.empty()) {
894 std::string LoweredTok = Tok.lower();
895 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
896 Op.SingletonReg = Reg->TheDef;
897 return;
900 if (!Tok.startswith(RegisterPrefix))
901 return;
903 StringRef RegName = Tok.substr(RegisterPrefix.size());
904 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
905 Op.SingletonReg = Reg->TheDef;
907 // If there is no register prefix (i.e. "%" in "%eax"), then this may
908 // be some random non-register token, just ignore it.
911 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
912 SmallPtrSetImpl<Record*> &SingletonRegisters,
913 AsmVariantInfo const &Variant,
914 bool HasMnemonicFirst) {
915 AsmVariantID = Variant.AsmVariantNo;
916 AsmString =
917 CodeGenInstruction::FlattenAsmStringVariants(AsmString,
918 Variant.AsmVariantNo);
920 tokenizeAsmString(Info, Variant);
922 // The first token of the instruction is the mnemonic, which must be a
923 // simple string, not a $foo variable or a singleton register.
924 if (AsmOperands.empty())
925 PrintFatalError(TheDef->getLoc(),
926 "Instruction '" + TheDef->getName() + "' has no tokens");
928 assert(!AsmOperands[0].Token.empty());
929 if (HasMnemonicFirst) {
930 Mnemonic = AsmOperands[0].Token;
931 if (Mnemonic[0] == '$')
932 PrintFatalError(TheDef->getLoc(),
933 "Invalid instruction mnemonic '" + Mnemonic + "'!");
935 // Remove the first operand, it is tracked in the mnemonic field.
936 AsmOperands.erase(AsmOperands.begin());
937 } else if (AsmOperands[0].Token[0] != '$')
938 Mnemonic = AsmOperands[0].Token;
940 // Compute the require features.
941 for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates"))
942 if (const SubtargetFeatureInfo *Feature =
943 Info.getSubtargetFeature(Predicate))
944 RequiredFeatures.push_back(Feature);
946 // Collect singleton registers, if used.
947 for (MatchableInfo::AsmOperand &Op : AsmOperands) {
948 extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix);
949 if (Record *Reg = Op.SingletonReg)
950 SingletonRegisters.insert(Reg);
953 const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
954 if (!DepMask)
955 DepMask = TheDef->getValue("ComplexDeprecationPredicate");
957 HasDeprecation =
958 DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
961 /// Append an AsmOperand for the given substring of AsmString.
962 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) {
963 AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token));
966 /// tokenizeAsmString - Tokenize a simplified assembly string.
967 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info,
968 AsmVariantInfo const &Variant) {
969 StringRef String = AsmString;
970 size_t Prev = 0;
971 bool InTok = false;
972 bool IsIsolatedToken = true;
973 for (size_t i = 0, e = String.size(); i != e; ++i) {
974 char Char = String[i];
975 if (Variant.BreakCharacters.find(Char) != std::string::npos) {
976 if (InTok) {
977 addAsmOperand(String.slice(Prev, i), false);
978 Prev = i;
979 IsIsolatedToken = false;
981 InTok = true;
982 continue;
984 if (Variant.TokenizingCharacters.find(Char) != std::string::npos) {
985 if (InTok) {
986 addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
987 InTok = false;
988 IsIsolatedToken = false;
990 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
991 Prev = i + 1;
992 IsIsolatedToken = true;
993 continue;
995 if (Variant.SeparatorCharacters.find(Char) != std::string::npos) {
996 if (InTok) {
997 addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
998 InTok = false;
1000 Prev = i + 1;
1001 IsIsolatedToken = true;
1002 continue;
1005 switch (Char) {
1006 case '\\':
1007 if (InTok) {
1008 addAsmOperand(String.slice(Prev, i), false);
1009 InTok = false;
1010 IsIsolatedToken = false;
1012 ++i;
1013 assert(i != String.size() && "Invalid quoted character");
1014 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
1015 Prev = i + 1;
1016 IsIsolatedToken = false;
1017 break;
1019 case '$': {
1020 if (InTok) {
1021 addAsmOperand(String.slice(Prev, i), false);
1022 InTok = false;
1023 IsIsolatedToken = false;
1026 // If this isn't "${", start new identifier looking like "$xxx"
1027 if (i + 1 == String.size() || String[i + 1] != '{') {
1028 Prev = i;
1029 break;
1032 size_t EndPos = String.find('}', i);
1033 assert(EndPos != StringRef::npos &&
1034 "Missing brace in operand reference!");
1035 addAsmOperand(String.slice(i, EndPos+1), IsIsolatedToken);
1036 Prev = EndPos + 1;
1037 i = EndPos;
1038 IsIsolatedToken = false;
1039 break;
1042 default:
1043 InTok = true;
1044 break;
1047 if (InTok && Prev != String.size())
1048 addAsmOperand(String.substr(Prev), IsIsolatedToken);
1051 bool MatchableInfo::validate(StringRef CommentDelimiter, bool IsAlias) const {
1052 // Reject matchables with no .s string.
1053 if (AsmString.empty())
1054 PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
1056 // Reject any matchables with a newline in them, they should be marked
1057 // isCodeGenOnly if they are pseudo instructions.
1058 if (AsmString.find('\n') != std::string::npos)
1059 PrintFatalError(TheDef->getLoc(),
1060 "multiline instruction is not valid for the asmparser, "
1061 "mark it isCodeGenOnly");
1063 // Remove comments from the asm string. We know that the asmstring only
1064 // has one line.
1065 if (!CommentDelimiter.empty() &&
1066 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
1067 PrintFatalError(TheDef->getLoc(),
1068 "asmstring for instruction has comment character in it, "
1069 "mark it isCodeGenOnly");
1071 // Reject matchables with operand modifiers, these aren't something we can
1072 // handle, the target should be refactored to use operands instead of
1073 // modifiers.
1075 // Also, check for instructions which reference the operand multiple times,
1076 // if they don't define a custom AsmMatcher: this implies a constraint that
1077 // the built-in matching code would not honor.
1078 std::set<std::string> OperandNames;
1079 for (const AsmOperand &Op : AsmOperands) {
1080 StringRef Tok = Op.Token;
1081 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
1082 PrintFatalError(TheDef->getLoc(),
1083 "matchable with operand modifier '" + Tok +
1084 "' not supported by asm matcher. Mark isCodeGenOnly!");
1085 // Verify that any operand is only mentioned once.
1086 // We reject aliases and ignore instructions for now.
1087 if (!IsAlias && TheDef->getValueAsString("AsmMatchConverter").empty() &&
1088 Tok[0] == '$' && !OperandNames.insert(Tok).second) {
1089 LLVM_DEBUG({
1090 errs() << "warning: '" << TheDef->getName() << "': "
1091 << "ignoring instruction with tied operand '"
1092 << Tok << "'\n";
1094 return false;
1098 return true;
1101 static std::string getEnumNameForToken(StringRef Str) {
1102 std::string Res;
1104 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
1105 switch (*it) {
1106 case '*': Res += "_STAR_"; break;
1107 case '%': Res += "_PCT_"; break;
1108 case ':': Res += "_COLON_"; break;
1109 case '!': Res += "_EXCLAIM_"; break;
1110 case '.': Res += "_DOT_"; break;
1111 case '<': Res += "_LT_"; break;
1112 case '>': Res += "_GT_"; break;
1113 case '-': Res += "_MINUS_"; break;
1114 default:
1115 if ((*it >= 'A' && *it <= 'Z') ||
1116 (*it >= 'a' && *it <= 'z') ||
1117 (*it >= '0' && *it <= '9'))
1118 Res += *it;
1119 else
1120 Res += "_" + utostr((unsigned) *it) + "_";
1124 return Res;
1127 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
1128 ClassInfo *&Entry = TokenClasses[Token];
1130 if (!Entry) {
1131 Classes.emplace_front();
1132 Entry = &Classes.front();
1133 Entry->Kind = ClassInfo::Token;
1134 Entry->ClassName = "Token";
1135 Entry->Name = "MCK_" + getEnumNameForToken(Token);
1136 Entry->ValueName = Token;
1137 Entry->PredicateMethod = "<invalid>";
1138 Entry->RenderMethod = "<invalid>";
1139 Entry->ParserMethod = "";
1140 Entry->DiagnosticType = "";
1141 Entry->IsOptional = false;
1142 Entry->DefaultMethod = "<invalid>";
1145 return Entry;
1148 ClassInfo *
1149 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1150 int SubOpIdx) {
1151 Record *Rec = OI.Rec;
1152 if (SubOpIdx != -1)
1153 Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1154 return getOperandClass(Rec, SubOpIdx);
1157 ClassInfo *
1158 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1159 if (Rec->isSubClassOf("RegisterOperand")) {
1160 // RegisterOperand may have an associated ParserMatchClass. If it does,
1161 // use it, else just fall back to the underlying register class.
1162 const RecordVal *R = Rec->getValue("ParserMatchClass");
1163 if (!R || !R->getValue())
1164 PrintFatalError(Rec->getLoc(),
1165 "Record `" + Rec->getName() +
1166 "' does not have a ParserMatchClass!\n");
1168 if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) {
1169 Record *MatchClass = DI->getDef();
1170 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1171 return CI;
1174 // No custom match class. Just use the register class.
1175 Record *ClassRec = Rec->getValueAsDef("RegClass");
1176 if (!ClassRec)
1177 PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1178 "' has no associated register class!\n");
1179 if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1180 return CI;
1181 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1184 if (Rec->isSubClassOf("RegisterClass")) {
1185 if (ClassInfo *CI = RegisterClassClasses[Rec])
1186 return CI;
1187 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1190 if (!Rec->isSubClassOf("Operand"))
1191 PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() +
1192 "' does not derive from class Operand!\n");
1193 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1194 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1195 return CI;
1197 PrintFatalError(Rec->getLoc(), "operand has no match class!");
1200 struct LessRegisterSet {
1201 bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const {
1202 // std::set<T> defines its own compariso "operator<", but it
1203 // performs a lexicographical comparison by T's innate comparison
1204 // for some reason. We don't want non-deterministic pointer
1205 // comparisons so use this instead.
1206 return std::lexicographical_compare(LHS.begin(), LHS.end(),
1207 RHS.begin(), RHS.end(),
1208 LessRecordByID());
1212 void AsmMatcherInfo::
1213 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) {
1214 const auto &Registers = Target.getRegBank().getRegisters();
1215 auto &RegClassList = Target.getRegBank().getRegClasses();
1217 typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1219 // The register sets used for matching.
1220 RegisterSetSet RegisterSets;
1222 // Gather the defined sets.
1223 for (const CodeGenRegisterClass &RC : RegClassList)
1224 RegisterSets.insert(
1225 RegisterSet(RC.getOrder().begin(), RC.getOrder().end()));
1227 // Add any required singleton sets.
1228 for (Record *Rec : SingletonRegisters) {
1229 RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1232 // Introduce derived sets where necessary (when a register does not determine
1233 // a unique register set class), and build the mapping of registers to the set
1234 // they should classify to.
1235 std::map<Record*, RegisterSet> RegisterMap;
1236 for (const CodeGenRegister &CGR : Registers) {
1237 // Compute the intersection of all sets containing this register.
1238 RegisterSet ContainingSet;
1240 for (const RegisterSet &RS : RegisterSets) {
1241 if (!RS.count(CGR.TheDef))
1242 continue;
1244 if (ContainingSet.empty()) {
1245 ContainingSet = RS;
1246 continue;
1249 RegisterSet Tmp;
1250 std::swap(Tmp, ContainingSet);
1251 std::insert_iterator<RegisterSet> II(ContainingSet,
1252 ContainingSet.begin());
1253 std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II,
1254 LessRecordByID());
1257 if (!ContainingSet.empty()) {
1258 RegisterSets.insert(ContainingSet);
1259 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
1263 // Construct the register classes.
1264 std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses;
1265 unsigned Index = 0;
1266 for (const RegisterSet &RS : RegisterSets) {
1267 Classes.emplace_front();
1268 ClassInfo *CI = &Classes.front();
1269 CI->Kind = ClassInfo::RegisterClass0 + Index;
1270 CI->ClassName = "Reg" + utostr(Index);
1271 CI->Name = "MCK_Reg" + utostr(Index);
1272 CI->ValueName = "";
1273 CI->PredicateMethod = ""; // unused
1274 CI->RenderMethod = "addRegOperands";
1275 CI->Registers = RS;
1276 // FIXME: diagnostic type.
1277 CI->DiagnosticType = "";
1278 CI->IsOptional = false;
1279 CI->DefaultMethod = ""; // unused
1280 RegisterSetClasses.insert(std::make_pair(RS, CI));
1281 ++Index;
1284 // Find the superclasses; we could compute only the subgroup lattice edges,
1285 // but there isn't really a point.
1286 for (const RegisterSet &RS : RegisterSets) {
1287 ClassInfo *CI = RegisterSetClasses[RS];
1288 for (const RegisterSet &RS2 : RegisterSets)
1289 if (RS != RS2 &&
1290 std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(),
1291 LessRecordByID()))
1292 CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1295 // Name the register classes which correspond to a user defined RegisterClass.
1296 for (const CodeGenRegisterClass &RC : RegClassList) {
1297 // Def will be NULL for non-user defined register classes.
1298 Record *Def = RC.getDef();
1299 if (!Def)
1300 continue;
1301 ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(),
1302 RC.getOrder().end())];
1303 if (CI->ValueName.empty()) {
1304 CI->ClassName = RC.getName();
1305 CI->Name = "MCK_" + RC.getName();
1306 CI->ValueName = RC.getName();
1307 } else
1308 CI->ValueName = CI->ValueName + "," + RC.getName();
1310 Init *DiagnosticType = Def->getValueInit("DiagnosticType");
1311 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1312 CI->DiagnosticType = SI->getValue();
1314 Init *DiagnosticString = Def->getValueInit("DiagnosticString");
1315 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticString))
1316 CI->DiagnosticString = SI->getValue();
1318 // If we have a diagnostic string but the diagnostic type is not specified
1319 // explicitly, create an anonymous diagnostic type.
1320 if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty())
1321 CI->DiagnosticType = RC.getName();
1323 RegisterClassClasses.insert(std::make_pair(Def, CI));
1326 // Populate the map for individual registers.
1327 for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(),
1328 ie = RegisterMap.end(); it != ie; ++it)
1329 RegisterClasses[it->first] = RegisterSetClasses[it->second];
1331 // Name the register classes which correspond to singleton registers.
1332 for (Record *Rec : SingletonRegisters) {
1333 ClassInfo *CI = RegisterClasses[Rec];
1334 assert(CI && "Missing singleton register class info!");
1336 if (CI->ValueName.empty()) {
1337 CI->ClassName = Rec->getName();
1338 CI->Name = "MCK_" + Rec->getName().str();
1339 CI->ValueName = Rec->getName();
1340 } else
1341 CI->ValueName = CI->ValueName + "," + Rec->getName().str();
1345 void AsmMatcherInfo::buildOperandClasses() {
1346 std::vector<Record*> AsmOperands =
1347 Records.getAllDerivedDefinitions("AsmOperandClass");
1349 // Pre-populate AsmOperandClasses map.
1350 for (Record *Rec : AsmOperands) {
1351 Classes.emplace_front();
1352 AsmOperandClasses[Rec] = &Classes.front();
1355 unsigned Index = 0;
1356 for (Record *Rec : AsmOperands) {
1357 ClassInfo *CI = AsmOperandClasses[Rec];
1358 CI->Kind = ClassInfo::UserClass0 + Index;
1360 ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1361 for (Init *I : Supers->getValues()) {
1362 DefInit *DI = dyn_cast<DefInit>(I);
1363 if (!DI) {
1364 PrintError(Rec->getLoc(), "Invalid super class reference!");
1365 continue;
1368 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1369 if (!SC)
1370 PrintError(Rec->getLoc(), "Invalid super class reference!");
1371 else
1372 CI->SuperClasses.push_back(SC);
1374 CI->ClassName = Rec->getValueAsString("Name");
1375 CI->Name = "MCK_" + CI->ClassName;
1376 CI->ValueName = Rec->getName();
1378 // Get or construct the predicate method name.
1379 Init *PMName = Rec->getValueInit("PredicateMethod");
1380 if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1381 CI->PredicateMethod = SI->getValue();
1382 } else {
1383 assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1384 CI->PredicateMethod = "is" + CI->ClassName;
1387 // Get or construct the render method name.
1388 Init *RMName = Rec->getValueInit("RenderMethod");
1389 if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1390 CI->RenderMethod = SI->getValue();
1391 } else {
1392 assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1393 CI->RenderMethod = "add" + CI->ClassName + "Operands";
1396 // Get the parse method name or leave it as empty.
1397 Init *PRMName = Rec->getValueInit("ParserMethod");
1398 if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1399 CI->ParserMethod = SI->getValue();
1401 // Get the diagnostic type and string or leave them as empty.
1402 Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1403 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1404 CI->DiagnosticType = SI->getValue();
1405 Init *DiagnosticString = Rec->getValueInit("DiagnosticString");
1406 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticString))
1407 CI->DiagnosticString = SI->getValue();
1408 // If we have a DiagnosticString, we need a DiagnosticType for use within
1409 // the matcher.
1410 if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty())
1411 CI->DiagnosticType = CI->ClassName;
1413 Init *IsOptional = Rec->getValueInit("IsOptional");
1414 if (BitInit *BI = dyn_cast<BitInit>(IsOptional))
1415 CI->IsOptional = BI->getValue();
1417 // Get or construct the default method name.
1418 Init *DMName = Rec->getValueInit("DefaultMethod");
1419 if (StringInit *SI = dyn_cast<StringInit>(DMName)) {
1420 CI->DefaultMethod = SI->getValue();
1421 } else {
1422 assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!");
1423 CI->DefaultMethod = "default" + CI->ClassName + "Operands";
1426 ++Index;
1430 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1431 CodeGenTarget &target,
1432 RecordKeeper &records)
1433 : Records(records), AsmParser(asmParser), Target(target) {
1436 /// buildOperandMatchInfo - Build the necessary information to handle user
1437 /// defined operand parsing methods.
1438 void AsmMatcherInfo::buildOperandMatchInfo() {
1440 /// Map containing a mask with all operands indices that can be found for
1441 /// that class inside a instruction.
1442 typedef std::map<ClassInfo *, unsigned, deref<std::less<>>> OpClassMaskTy;
1443 OpClassMaskTy OpClassMask;
1445 for (const auto &MI : Matchables) {
1446 OpClassMask.clear();
1448 // Keep track of all operands of this instructions which belong to the
1449 // same class.
1450 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
1451 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
1452 if (Op.Class->ParserMethod.empty())
1453 continue;
1454 unsigned &OperandMask = OpClassMask[Op.Class];
1455 OperandMask |= (1 << i);
1458 // Generate operand match info for each mnemonic/operand class pair.
1459 for (const auto &OCM : OpClassMask) {
1460 unsigned OpMask = OCM.second;
1461 ClassInfo *CI = OCM.first;
1462 OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI,
1463 OpMask));
1468 void AsmMatcherInfo::buildInfo() {
1469 // Build information about all of the AssemblerPredicates.
1470 const std::vector<std::pair<Record *, SubtargetFeatureInfo>>
1471 &SubtargetFeaturePairs = SubtargetFeatureInfo::getAll(Records);
1472 SubtargetFeatures.insert(SubtargetFeaturePairs.begin(),
1473 SubtargetFeaturePairs.end());
1474 #ifndef NDEBUG
1475 for (const auto &Pair : SubtargetFeatures)
1476 LLVM_DEBUG(Pair.second.dump());
1477 #endif // NDEBUG
1479 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
1480 bool ReportMultipleNearMisses =
1481 AsmParser->getValueAsBit("ReportMultipleNearMisses");
1483 // Parse the instructions; we need to do this first so that we can gather the
1484 // singleton register classes.
1485 SmallPtrSet<Record*, 16> SingletonRegisters;
1486 unsigned VariantCount = Target.getAsmParserVariantCount();
1487 for (unsigned VC = 0; VC != VariantCount; ++VC) {
1488 Record *AsmVariant = Target.getAsmParserVariant(VC);
1489 StringRef CommentDelimiter =
1490 AsmVariant->getValueAsString("CommentDelimiter");
1491 AsmVariantInfo Variant;
1492 Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1493 Variant.TokenizingCharacters =
1494 AsmVariant->getValueAsString("TokenizingCharacters");
1495 Variant.SeparatorCharacters =
1496 AsmVariant->getValueAsString("SeparatorCharacters");
1497 Variant.BreakCharacters =
1498 AsmVariant->getValueAsString("BreakCharacters");
1499 Variant.Name = AsmVariant->getValueAsString("Name");
1500 Variant.AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1502 for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) {
1504 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1505 // filter the set of instructions we consider.
1506 if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix))
1507 continue;
1509 // Ignore "codegen only" instructions.
1510 if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1511 continue;
1513 // Ignore instructions for different instructions
1514 StringRef V = CGI->TheDef->getValueAsString("AsmVariantName");
1515 if (!V.empty() && V != Variant.Name)
1516 continue;
1518 auto II = std::make_unique<MatchableInfo>(*CGI);
1520 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1522 // Ignore instructions which shouldn't be matched and diagnose invalid
1523 // instruction definitions with an error.
1524 if (!II->validate(CommentDelimiter, false))
1525 continue;
1527 Matchables.push_back(std::move(II));
1530 // Parse all of the InstAlias definitions and stick them in the list of
1531 // matchables.
1532 std::vector<Record*> AllInstAliases =
1533 Records.getAllDerivedDefinitions("InstAlias");
1534 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1535 auto Alias = std::make_unique<CodeGenInstAlias>(AllInstAliases[i],
1536 Target);
1538 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1539 // filter the set of instruction aliases we consider, based on the target
1540 // instruction.
1541 if (!StringRef(Alias->ResultInst->TheDef->getName())
1542 .startswith( MatchPrefix))
1543 continue;
1545 StringRef V = Alias->TheDef->getValueAsString("AsmVariantName");
1546 if (!V.empty() && V != Variant.Name)
1547 continue;
1549 auto II = std::make_unique<MatchableInfo>(std::move(Alias));
1551 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1553 // Validate the alias definitions.
1554 II->validate(CommentDelimiter, true);
1556 Matchables.push_back(std::move(II));
1560 // Build info for the register classes.
1561 buildRegisterClasses(SingletonRegisters);
1563 // Build info for the user defined assembly operand classes.
1564 buildOperandClasses();
1566 // Build the information about matchables, now that we have fully formed
1567 // classes.
1568 std::vector<std::unique_ptr<MatchableInfo>> NewMatchables;
1569 for (auto &II : Matchables) {
1570 // Parse the tokens after the mnemonic.
1571 // Note: buildInstructionOperandReference may insert new AsmOperands, so
1572 // don't precompute the loop bound.
1573 for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1574 MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1575 StringRef Token = Op.Token;
1577 // Check for singleton registers.
1578 if (Record *RegRecord = Op.SingletonReg) {
1579 Op.Class = RegisterClasses[RegRecord];
1580 assert(Op.Class && Op.Class->Registers.size() == 1 &&
1581 "Unexpected class for singleton register");
1582 continue;
1585 // Check for simple tokens.
1586 if (Token[0] != '$') {
1587 Op.Class = getTokenClass(Token);
1588 continue;
1591 if (Token.size() > 1 && isdigit(Token[1])) {
1592 Op.Class = getTokenClass(Token);
1593 continue;
1596 // Otherwise this is an operand reference.
1597 StringRef OperandName;
1598 if (Token[1] == '{')
1599 OperandName = Token.substr(2, Token.size() - 3);
1600 else
1601 OperandName = Token.substr(1);
1603 if (II->DefRec.is<const CodeGenInstruction*>())
1604 buildInstructionOperandReference(II.get(), OperandName, i);
1605 else
1606 buildAliasOperandReference(II.get(), OperandName, Op);
1609 if (II->DefRec.is<const CodeGenInstruction*>()) {
1610 II->buildInstructionResultOperands();
1611 // If the instruction has a two-operand alias, build up the
1612 // matchable here. We'll add them in bulk at the end to avoid
1613 // confusing this loop.
1614 StringRef Constraint =
1615 II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1616 if (Constraint != "") {
1617 // Start by making a copy of the original matchable.
1618 auto AliasII = std::make_unique<MatchableInfo>(*II);
1620 // Adjust it to be a two-operand alias.
1621 AliasII->formTwoOperandAlias(Constraint);
1623 // Add the alias to the matchables list.
1624 NewMatchables.push_back(std::move(AliasII));
1626 } else
1627 // FIXME: The tied operands checking is not yet integrated with the
1628 // framework for reporting multiple near misses. To prevent invalid
1629 // formats from being matched with an alias if a tied-operands check
1630 // would otherwise have disallowed it, we just disallow such constructs
1631 // in TableGen completely.
1632 II->buildAliasResultOperands(!ReportMultipleNearMisses);
1634 if (!NewMatchables.empty())
1635 Matchables.insert(Matchables.end(),
1636 std::make_move_iterator(NewMatchables.begin()),
1637 std::make_move_iterator(NewMatchables.end()));
1639 // Process token alias definitions and set up the associated superclass
1640 // information.
1641 std::vector<Record*> AllTokenAliases =
1642 Records.getAllDerivedDefinitions("TokenAlias");
1643 for (Record *Rec : AllTokenAliases) {
1644 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1645 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1646 if (FromClass == ToClass)
1647 PrintFatalError(Rec->getLoc(),
1648 "error: Destination value identical to source value.");
1649 FromClass->SuperClasses.push_back(ToClass);
1652 // Reorder classes so that classes precede super classes.
1653 Classes.sort();
1655 #ifdef EXPENSIVE_CHECKS
1656 // Verify that the table is sorted and operator < works transitively.
1657 for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) {
1658 for (auto J = I; J != E; ++J) {
1659 assert(!(*J < *I));
1660 assert(I == J || !J->isSubsetOf(*I));
1663 #endif
1666 /// buildInstructionOperandReference - The specified operand is a reference to a
1667 /// named operand such as $src. Resolve the Class and OperandInfo pointers.
1668 void AsmMatcherInfo::
1669 buildInstructionOperandReference(MatchableInfo *II,
1670 StringRef OperandName,
1671 unsigned AsmOpIdx) {
1672 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1673 const CGIOperandList &Operands = CGI.Operands;
1674 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1676 // Map this token to an operand.
1677 unsigned Idx;
1678 if (!Operands.hasOperandNamed(OperandName, Idx))
1679 PrintFatalError(II->TheDef->getLoc(),
1680 "error: unable to find operand: '" + OperandName + "'");
1682 // If the instruction operand has multiple suboperands, but the parser
1683 // match class for the asm operand is still the default "ImmAsmOperand",
1684 // then handle each suboperand separately.
1685 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1686 Record *Rec = Operands[Idx].Rec;
1687 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1688 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1689 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1690 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1691 StringRef Token = Op->Token; // save this in case Op gets moved
1692 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1693 MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token);
1694 NewAsmOp.SubOpIdx = SI;
1695 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1697 // Replace Op with first suboperand.
1698 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1699 Op->SubOpIdx = 0;
1703 // Set up the operand class.
1704 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1705 Op->OrigSrcOpName = OperandName;
1707 // If the named operand is tied, canonicalize it to the untied operand.
1708 // For example, something like:
1709 // (outs GPR:$dst), (ins GPR:$src)
1710 // with an asmstring of
1711 // "inc $src"
1712 // we want to canonicalize to:
1713 // "inc $dst"
1714 // so that we know how to provide the $dst operand when filling in the result.
1715 int OITied = -1;
1716 if (Operands[Idx].MINumOperands == 1)
1717 OITied = Operands[Idx].getTiedRegister();
1718 if (OITied != -1) {
1719 // The tied operand index is an MIOperand index, find the operand that
1720 // contains it.
1721 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1722 OperandName = Operands[Idx.first].Name;
1723 Op->SubOpIdx = Idx.second;
1726 Op->SrcOpName = OperandName;
1729 /// buildAliasOperandReference - When parsing an operand reference out of the
1730 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1731 /// operand reference is by looking it up in the result pattern definition.
1732 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1733 StringRef OperandName,
1734 MatchableInfo::AsmOperand &Op) {
1735 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1737 // Set up the operand class.
1738 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1739 if (CGA.ResultOperands[i].isRecord() &&
1740 CGA.ResultOperands[i].getName() == OperandName) {
1741 // It's safe to go with the first one we find, because CodeGenInstAlias
1742 // validates that all operands with the same name have the same record.
1743 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1744 // Use the match class from the Alias definition, not the
1745 // destination instruction, as we may have an immediate that's
1746 // being munged by the match class.
1747 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1748 Op.SubOpIdx);
1749 Op.SrcOpName = OperandName;
1750 Op.OrigSrcOpName = OperandName;
1751 return;
1754 PrintFatalError(II->TheDef->getLoc(),
1755 "error: unable to find operand: '" + OperandName + "'");
1758 void MatchableInfo::buildInstructionResultOperands() {
1759 const CodeGenInstruction *ResultInst = getResultInst();
1761 // Loop over all operands of the result instruction, determining how to
1762 // populate them.
1763 for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) {
1764 // If this is a tied operand, just copy from the previously handled operand.
1765 int TiedOp = -1;
1766 if (OpInfo.MINumOperands == 1)
1767 TiedOp = OpInfo.getTiedRegister();
1768 if (TiedOp != -1) {
1769 int TiedSrcOperand = findAsmOperandOriginallyNamed(OpInfo.Name);
1770 if (TiedSrcOperand != -1 &&
1771 ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand)
1772 ResOperands.push_back(ResOperand::getTiedOp(
1773 TiedOp, ResOperands[TiedOp].AsmOperandNum, TiedSrcOperand));
1774 else
1775 ResOperands.push_back(ResOperand::getTiedOp(TiedOp, 0, 0));
1776 continue;
1779 int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1780 if (OpInfo.Name.empty() || SrcOperand == -1) {
1781 // This may happen for operands that are tied to a suboperand of a
1782 // complex operand. Simply use a dummy value here; nobody should
1783 // use this operand slot.
1784 // FIXME: The long term goal is for the MCOperand list to not contain
1785 // tied operands at all.
1786 ResOperands.push_back(ResOperand::getImmOp(0));
1787 continue;
1790 // Check if the one AsmOperand populates the entire operand.
1791 unsigned NumOperands = OpInfo.MINumOperands;
1792 if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1793 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1794 continue;
1797 // Add a separate ResOperand for each suboperand.
1798 for (unsigned AI = 0; AI < NumOperands; ++AI) {
1799 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1800 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1801 "unexpected AsmOperands for suboperands");
1802 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1807 void MatchableInfo::buildAliasResultOperands(bool AliasConstraintsAreChecked) {
1808 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1809 const CodeGenInstruction *ResultInst = getResultInst();
1811 // Map of: $reg -> #lastref
1812 // where $reg is the name of the operand in the asm string
1813 // where #lastref is the last processed index where $reg was referenced in
1814 // the asm string.
1815 SmallDenseMap<StringRef, int> OperandRefs;
1817 // Loop over all operands of the result instruction, determining how to
1818 // populate them.
1819 unsigned AliasOpNo = 0;
1820 unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1821 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1822 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1824 // If this is a tied operand, just copy from the previously handled operand.
1825 int TiedOp = -1;
1826 if (OpInfo->MINumOperands == 1)
1827 TiedOp = OpInfo->getTiedRegister();
1828 if (TiedOp != -1) {
1829 unsigned SrcOp1 = 0;
1830 unsigned SrcOp2 = 0;
1832 // If an operand has been specified twice in the asm string,
1833 // add the two source operand's indices to the TiedOp so that
1834 // at runtime the 'tied' constraint is checked.
1835 if (ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand) {
1836 SrcOp1 = ResOperands[TiedOp].AsmOperandNum;
1838 // Find the next operand (similarly named operand) in the string.
1839 StringRef Name = AsmOperands[SrcOp1].SrcOpName;
1840 auto Insert = OperandRefs.try_emplace(Name, SrcOp1);
1841 SrcOp2 = findAsmOperandNamed(Name, Insert.first->second);
1843 // Not updating the record in OperandRefs will cause TableGen
1844 // to fail with an error at the end of this function.
1845 if (AliasConstraintsAreChecked)
1846 Insert.first->second = SrcOp2;
1848 // In case it only has one reference in the asm string,
1849 // it doesn't need to be checked for tied constraints.
1850 SrcOp2 = (SrcOp2 == (unsigned)-1) ? SrcOp1 : SrcOp2;
1853 // If the alias operand is of a different operand class, we only want
1854 // to benefit from the tied-operands check and just match the operand
1855 // as a normal, but not copy the original (TiedOp) to the result
1856 // instruction. We do this by passing -1 as the tied operand to copy.
1857 if (ResultInst->Operands[i].Rec->getName() !=
1858 ResultInst->Operands[TiedOp].Rec->getName()) {
1859 SrcOp1 = ResOperands[TiedOp].AsmOperandNum;
1860 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1861 StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1862 SrcOp2 = findAsmOperand(Name, SubIdx);
1863 ResOperands.push_back(
1864 ResOperand::getTiedOp((unsigned)-1, SrcOp1, SrcOp2));
1865 } else {
1866 ResOperands.push_back(ResOperand::getTiedOp(TiedOp, SrcOp1, SrcOp2));
1867 continue;
1871 // Handle all the suboperands for this operand.
1872 const std::string &OpName = OpInfo->Name;
1873 for ( ; AliasOpNo < LastOpNo &&
1874 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1875 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1877 // Find out what operand from the asmparser that this MCInst operand
1878 // comes from.
1879 switch (CGA.ResultOperands[AliasOpNo].Kind) {
1880 case CodeGenInstAlias::ResultOperand::K_Record: {
1881 StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1882 int SrcOperand = findAsmOperand(Name, SubIdx);
1883 if (SrcOperand == -1)
1884 PrintFatalError(TheDef->getLoc(), "Instruction '" +
1885 TheDef->getName() + "' has operand '" + OpName +
1886 "' that doesn't appear in asm string!");
1888 // Add it to the operand references. If it is added a second time, the
1889 // record won't be updated and it will fail later on.
1890 OperandRefs.try_emplace(Name, SrcOperand);
1892 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1893 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1894 NumOperands));
1895 break;
1897 case CodeGenInstAlias::ResultOperand::K_Imm: {
1898 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1899 ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1900 break;
1902 case CodeGenInstAlias::ResultOperand::K_Reg: {
1903 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1904 ResOperands.push_back(ResOperand::getRegOp(Reg));
1905 break;
1911 // Check that operands are not repeated more times than is supported.
1912 for (auto &T : OperandRefs) {
1913 if (T.second != -1 && findAsmOperandNamed(T.first, T.second) != -1)
1914 PrintFatalError(TheDef->getLoc(),
1915 "Operand '" + T.first + "' can never be matched");
1919 static unsigned
1920 getConverterOperandID(const std::string &Name,
1921 SmallSetVector<CachedHashString, 16> &Table,
1922 bool &IsNew) {
1923 IsNew = Table.insert(CachedHashString(Name));
1925 unsigned ID = IsNew ? Table.size() - 1 : find(Table, Name) - Table.begin();
1927 assert(ID < Table.size());
1929 return ID;
1932 static unsigned
1933 emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1934 std::vector<std::unique_ptr<MatchableInfo>> &Infos,
1935 bool HasMnemonicFirst, bool HasOptionalOperands,
1936 raw_ostream &OS) {
1937 SmallSetVector<CachedHashString, 16> OperandConversionKinds;
1938 SmallSetVector<CachedHashString, 16> InstructionConversionKinds;
1939 std::vector<std::vector<uint8_t> > ConversionTable;
1940 size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1942 // TargetOperandClass - This is the target's operand class, like X86Operand.
1943 std::string TargetOperandClass = Target.getName().str() + "Operand";
1945 // Write the convert function to a separate stream, so we can drop it after
1946 // the enum. We'll build up the conversion handlers for the individual
1947 // operand types opportunistically as we encounter them.
1948 std::string ConvertFnBody;
1949 raw_string_ostream CvtOS(ConvertFnBody);
1950 // Start the unified conversion function.
1951 if (HasOptionalOperands) {
1952 CvtOS << "void " << Target.getName() << ClassName << "::\n"
1953 << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1954 << "unsigned Opcode,\n"
1955 << " const OperandVector &Operands,\n"
1956 << " const SmallBitVector &OptionalOperandsMask) {\n";
1957 } else {
1958 CvtOS << "void " << Target.getName() << ClassName << "::\n"
1959 << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1960 << "unsigned Opcode,\n"
1961 << " const OperandVector &Operands) {\n";
1963 CvtOS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n";
1964 CvtOS << " const uint8_t *Converter = ConversionTable[Kind];\n";
1965 if (HasOptionalOperands) {
1966 size_t MaxNumOperands = 0;
1967 for (const auto &MI : Infos) {
1968 MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
1970 CvtOS << " unsigned DefaultsOffset[" << (MaxNumOperands + 1)
1971 << "] = { 0 };\n";
1972 CvtOS << " assert(OptionalOperandsMask.size() == " << (MaxNumOperands)
1973 << ");\n";
1974 CvtOS << " for (unsigned i = 0, NumDefaults = 0; i < " << (MaxNumOperands)
1975 << "; ++i) {\n";
1976 CvtOS << " DefaultsOffset[i + 1] = NumDefaults;\n";
1977 CvtOS << " NumDefaults += (OptionalOperandsMask[i] ? 1 : 0);\n";
1978 CvtOS << " }\n";
1980 CvtOS << " unsigned OpIdx;\n";
1981 CvtOS << " Inst.setOpcode(Opcode);\n";
1982 CvtOS << " for (const uint8_t *p = Converter; *p; p+= 2) {\n";
1983 if (HasOptionalOperands) {
1984 CvtOS << " OpIdx = *(p + 1) - DefaultsOffset[*(p + 1)];\n";
1985 } else {
1986 CvtOS << " OpIdx = *(p + 1);\n";
1988 CvtOS << " switch (*p) {\n";
1989 CvtOS << " default: llvm_unreachable(\"invalid conversion entry!\");\n";
1990 CvtOS << " case CVT_Reg:\n";
1991 CvtOS << " static_cast<" << TargetOperandClass
1992 << "&>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n";
1993 CvtOS << " break;\n";
1994 CvtOS << " case CVT_Tied: {\n";
1995 CvtOS << " assert(OpIdx < (size_t)(std::end(TiedAsmOperandTable) -\n";
1996 CvtOS << " std::begin(TiedAsmOperandTable)) &&\n";
1997 CvtOS << " \"Tied operand not found\");\n";
1998 CvtOS << " unsigned TiedResOpnd = TiedAsmOperandTable[OpIdx][0];\n";
1999 CvtOS << " if (TiedResOpnd != (uint8_t) -1)\n";
2000 CvtOS << " Inst.addOperand(Inst.getOperand(TiedResOpnd));\n";
2001 CvtOS << " break;\n";
2002 CvtOS << " }\n";
2004 std::string OperandFnBody;
2005 raw_string_ostream OpOS(OperandFnBody);
2006 // Start the operand number lookup function.
2007 OpOS << "void " << Target.getName() << ClassName << "::\n"
2008 << "convertToMapAndConstraints(unsigned Kind,\n";
2009 OpOS.indent(27);
2010 OpOS << "const OperandVector &Operands) {\n"
2011 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
2012 << " unsigned NumMCOperands = 0;\n"
2013 << " const uint8_t *Converter = ConversionTable[Kind];\n"
2014 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"
2015 << " switch (*p) {\n"
2016 << " default: llvm_unreachable(\"invalid conversion entry!\");\n"
2017 << " case CVT_Reg:\n"
2018 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2019 << " Operands[*(p + 1)]->setConstraint(\"r\");\n"
2020 << " ++NumMCOperands;\n"
2021 << " break;\n"
2022 << " case CVT_Tied:\n"
2023 << " ++NumMCOperands;\n"
2024 << " break;\n";
2026 // Pre-populate the operand conversion kinds with the standard always
2027 // available entries.
2028 OperandConversionKinds.insert(CachedHashString("CVT_Done"));
2029 OperandConversionKinds.insert(CachedHashString("CVT_Reg"));
2030 OperandConversionKinds.insert(CachedHashString("CVT_Tied"));
2031 enum { CVT_Done, CVT_Reg, CVT_Tied };
2033 // Map of e.g. <0, 2, 3> -> "Tie_0_2_3" enum label.
2034 std::map<std::tuple<uint8_t, uint8_t, uint8_t>, std::string>
2035 TiedOperandsEnumMap;
2037 for (auto &II : Infos) {
2038 // Check if we have a custom match function.
2039 StringRef AsmMatchConverter =
2040 II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
2041 if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) {
2042 std::string Signature = ("ConvertCustom_" + AsmMatchConverter).str();
2043 II->ConversionFnKind = Signature;
2045 // Check if we have already generated this signature.
2046 if (!InstructionConversionKinds.insert(CachedHashString(Signature)))
2047 continue;
2049 // Remember this converter for the kind enum.
2050 unsigned KindID = OperandConversionKinds.size();
2051 OperandConversionKinds.insert(
2052 CachedHashString("CVT_" + getEnumNameForToken(AsmMatchConverter)));
2054 // Add the converter row for this instruction.
2055 ConversionTable.emplace_back();
2056 ConversionTable.back().push_back(KindID);
2057 ConversionTable.back().push_back(CVT_Done);
2059 // Add the handler to the conversion driver function.
2060 CvtOS << " case CVT_"
2061 << getEnumNameForToken(AsmMatchConverter) << ":\n"
2062 << " " << AsmMatchConverter << "(Inst, Operands);\n"
2063 << " break;\n";
2065 // FIXME: Handle the operand number lookup for custom match functions.
2066 continue;
2069 // Build the conversion function signature.
2070 std::string Signature = "Convert";
2072 std::vector<uint8_t> ConversionRow;
2074 // Compute the convert enum and the case body.
2075 MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 );
2077 for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) {
2078 const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i];
2080 // Generate code to populate each result operand.
2081 switch (OpInfo.Kind) {
2082 case MatchableInfo::ResOperand::RenderAsmOperand: {
2083 // This comes from something we parsed.
2084 const MatchableInfo::AsmOperand &Op =
2085 II->AsmOperands[OpInfo.AsmOperandNum];
2087 // Registers are always converted the same, don't duplicate the
2088 // conversion function based on them.
2089 Signature += "__";
2090 std::string Class;
2091 Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
2092 Signature += Class;
2093 Signature += utostr(OpInfo.MINumOperands);
2094 Signature += "_" + itostr(OpInfo.AsmOperandNum);
2096 // Add the conversion kind, if necessary, and get the associated ID
2097 // the index of its entry in the vector).
2098 std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" :
2099 Op.Class->RenderMethod);
2100 if (Op.Class->IsOptional) {
2101 // For optional operands we must also care about DefaultMethod
2102 assert(HasOptionalOperands);
2103 Name += "_" + Op.Class->DefaultMethod;
2105 Name = getEnumNameForToken(Name);
2107 bool IsNewConverter = false;
2108 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2109 IsNewConverter);
2111 // Add the operand entry to the instruction kind conversion row.
2112 ConversionRow.push_back(ID);
2113 ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst);
2115 if (!IsNewConverter)
2116 break;
2118 // This is a new operand kind. Add a handler for it to the
2119 // converter driver.
2120 CvtOS << " case " << Name << ":\n";
2121 if (Op.Class->IsOptional) {
2122 // If optional operand is not present in actual instruction then we
2123 // should call its DefaultMethod before RenderMethod
2124 assert(HasOptionalOperands);
2125 CvtOS << " if (OptionalOperandsMask[*(p + 1) - 1]) {\n"
2126 << " " << Op.Class->DefaultMethod << "()"
2127 << "->" << Op.Class->RenderMethod << "(Inst, "
2128 << OpInfo.MINumOperands << ");\n"
2129 << " } else {\n"
2130 << " static_cast<" << TargetOperandClass
2131 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod
2132 << "(Inst, " << OpInfo.MINumOperands << ");\n"
2133 << " }\n";
2134 } else {
2135 CvtOS << " static_cast<" << TargetOperandClass
2136 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod
2137 << "(Inst, " << OpInfo.MINumOperands << ");\n";
2139 CvtOS << " break;\n";
2141 // Add a handler for the operand number lookup.
2142 OpOS << " case " << Name << ":\n"
2143 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
2145 if (Op.Class->isRegisterClass())
2146 OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n";
2147 else
2148 OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n";
2149 OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n"
2150 << " break;\n";
2151 break;
2153 case MatchableInfo::ResOperand::TiedOperand: {
2154 // If this operand is tied to a previous one, just copy the MCInst
2155 // operand from the earlier one.We can only tie single MCOperand values.
2156 assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
2157 uint8_t TiedOp = OpInfo.TiedOperands.ResOpnd;
2158 uint8_t SrcOp1 =
2159 OpInfo.TiedOperands.SrcOpnd1Idx + HasMnemonicFirst;
2160 uint8_t SrcOp2 =
2161 OpInfo.TiedOperands.SrcOpnd2Idx + HasMnemonicFirst;
2162 assert((i > TiedOp || TiedOp == (uint8_t)-1) &&
2163 "Tied operand precedes its target!");
2164 auto TiedTupleName = std::string("Tie") + utostr(TiedOp) + '_' +
2165 utostr(SrcOp1) + '_' + utostr(SrcOp2);
2166 Signature += "__" + TiedTupleName;
2167 ConversionRow.push_back(CVT_Tied);
2168 ConversionRow.push_back(TiedOp);
2169 ConversionRow.push_back(SrcOp1);
2170 ConversionRow.push_back(SrcOp2);
2172 // Also create an 'enum' for this combination of tied operands.
2173 auto Key = std::make_tuple(TiedOp, SrcOp1, SrcOp2);
2174 TiedOperandsEnumMap.emplace(Key, TiedTupleName);
2175 break;
2177 case MatchableInfo::ResOperand::ImmOperand: {
2178 int64_t Val = OpInfo.ImmVal;
2179 std::string Ty = "imm_" + itostr(Val);
2180 Ty = getEnumNameForToken(Ty);
2181 Signature += "__" + Ty;
2183 std::string Name = "CVT_" + Ty;
2184 bool IsNewConverter = false;
2185 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2186 IsNewConverter);
2187 // Add the operand entry to the instruction kind conversion row.
2188 ConversionRow.push_back(ID);
2189 ConversionRow.push_back(0);
2191 if (!IsNewConverter)
2192 break;
2194 CvtOS << " case " << Name << ":\n"
2195 << " Inst.addOperand(MCOperand::createImm(" << Val << "));\n"
2196 << " break;\n";
2198 OpOS << " case " << Name << ":\n"
2199 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2200 << " Operands[*(p + 1)]->setConstraint(\"\");\n"
2201 << " ++NumMCOperands;\n"
2202 << " break;\n";
2203 break;
2205 case MatchableInfo::ResOperand::RegOperand: {
2206 std::string Reg, Name;
2207 if (!OpInfo.Register) {
2208 Name = "reg0";
2209 Reg = "0";
2210 } else {
2211 Reg = getQualifiedName(OpInfo.Register);
2212 Name = "reg" + OpInfo.Register->getName().str();
2214 Signature += "__" + Name;
2215 Name = "CVT_" + Name;
2216 bool IsNewConverter = false;
2217 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2218 IsNewConverter);
2219 // Add the operand entry to the instruction kind conversion row.
2220 ConversionRow.push_back(ID);
2221 ConversionRow.push_back(0);
2223 if (!IsNewConverter)
2224 break;
2225 CvtOS << " case " << Name << ":\n"
2226 << " Inst.addOperand(MCOperand::createReg(" << Reg << "));\n"
2227 << " break;\n";
2229 OpOS << " case " << Name << ":\n"
2230 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2231 << " Operands[*(p + 1)]->setConstraint(\"m\");\n"
2232 << " ++NumMCOperands;\n"
2233 << " break;\n";
2238 // If there were no operands, add to the signature to that effect
2239 if (Signature == "Convert")
2240 Signature += "_NoOperands";
2242 II->ConversionFnKind = Signature;
2244 // Save the signature. If we already have it, don't add a new row
2245 // to the table.
2246 if (!InstructionConversionKinds.insert(CachedHashString(Signature)))
2247 continue;
2249 // Add the row to the table.
2250 ConversionTable.push_back(std::move(ConversionRow));
2253 // Finish up the converter driver function.
2254 CvtOS << " }\n }\n}\n\n";
2256 // Finish up the operand number lookup function.
2257 OpOS << " }\n }\n}\n\n";
2259 // Output a static table for tied operands.
2260 if (TiedOperandsEnumMap.size()) {
2261 // The number of tied operand combinations will be small in practice,
2262 // but just add the assert to be sure.
2263 assert(TiedOperandsEnumMap.size() <= 254 &&
2264 "Too many tied-operand combinations to reference with "
2265 "an 8bit offset from the conversion table, where index "
2266 "'255' is reserved as operand not to be copied.");
2268 OS << "enum {\n";
2269 for (auto &KV : TiedOperandsEnumMap) {
2270 OS << " " << KV.second << ",\n";
2272 OS << "};\n\n";
2274 OS << "static const uint8_t TiedAsmOperandTable[][3] = {\n";
2275 for (auto &KV : TiedOperandsEnumMap) {
2276 OS << " /* " << KV.second << " */ { "
2277 << utostr(std::get<0>(KV.first)) << ", "
2278 << utostr(std::get<1>(KV.first)) << ", "
2279 << utostr(std::get<2>(KV.first)) << " },\n";
2281 OS << "};\n\n";
2282 } else
2283 OS << "static const uint8_t TiedAsmOperandTable[][3] = "
2284 "{ /* empty */ {0, 0, 0} };\n\n";
2286 OS << "namespace {\n";
2288 // Output the operand conversion kind enum.
2289 OS << "enum OperatorConversionKind {\n";
2290 for (const auto &Converter : OperandConversionKinds)
2291 OS << " " << Converter << ",\n";
2292 OS << " CVT_NUM_CONVERTERS\n";
2293 OS << "};\n\n";
2295 // Output the instruction conversion kind enum.
2296 OS << "enum InstructionConversionKind {\n";
2297 for (const auto &Signature : InstructionConversionKinds)
2298 OS << " " << Signature << ",\n";
2299 OS << " CVT_NUM_SIGNATURES\n";
2300 OS << "};\n\n";
2302 OS << "} // end anonymous namespace\n\n";
2304 // Output the conversion table.
2305 OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
2306 << MaxRowLength << "] = {\n";
2308 for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
2309 assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
2310 OS << " // " << InstructionConversionKinds[Row] << "\n";
2311 OS << " { ";
2312 for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2) {
2313 OS << OperandConversionKinds[ConversionTable[Row][i]] << ", ";
2314 if (OperandConversionKinds[ConversionTable[Row][i]] !=
2315 CachedHashString("CVT_Tied")) {
2316 OS << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
2317 continue;
2320 // For a tied operand, emit a reference to the TiedAsmOperandTable
2321 // that contains the operand to copy, and the parsed operands to
2322 // check for their tied constraints.
2323 auto Key = std::make_tuple((uint8_t)ConversionTable[Row][i + 1],
2324 (uint8_t)ConversionTable[Row][i + 2],
2325 (uint8_t)ConversionTable[Row][i + 3]);
2326 auto TiedOpndEnum = TiedOperandsEnumMap.find(Key);
2327 assert(TiedOpndEnum != TiedOperandsEnumMap.end() &&
2328 "No record for tied operand pair");
2329 OS << TiedOpndEnum->second << ", ";
2330 i += 2;
2332 OS << "CVT_Done },\n";
2335 OS << "};\n\n";
2337 // Spit out the conversion driver function.
2338 OS << CvtOS.str();
2340 // Spit out the operand number lookup function.
2341 OS << OpOS.str();
2343 return ConversionTable.size();
2346 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
2347 static void emitMatchClassEnumeration(CodeGenTarget &Target,
2348 std::forward_list<ClassInfo> &Infos,
2349 raw_ostream &OS) {
2350 OS << "namespace {\n\n";
2352 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
2353 << "/// instruction matching.\n";
2354 OS << "enum MatchClassKind {\n";
2355 OS << " InvalidMatchClass = 0,\n";
2356 OS << " OptionalMatchClass = 1,\n";
2357 ClassInfo::ClassInfoKind LastKind = ClassInfo::Token;
2358 StringRef LastName = "OptionalMatchClass";
2359 for (const auto &CI : Infos) {
2360 if (LastKind == ClassInfo::Token && CI.Kind != ClassInfo::Token) {
2361 OS << " MCK_LAST_TOKEN = " << LastName << ",\n";
2362 } else if (LastKind < ClassInfo::UserClass0 &&
2363 CI.Kind >= ClassInfo::UserClass0) {
2364 OS << " MCK_LAST_REGISTER = " << LastName << ",\n";
2366 LastKind = (ClassInfo::ClassInfoKind)CI.Kind;
2367 LastName = CI.Name;
2369 OS << " " << CI.Name << ", // ";
2370 if (CI.Kind == ClassInfo::Token) {
2371 OS << "'" << CI.ValueName << "'\n";
2372 } else if (CI.isRegisterClass()) {
2373 if (!CI.ValueName.empty())
2374 OS << "register class '" << CI.ValueName << "'\n";
2375 else
2376 OS << "derived register class\n";
2377 } else {
2378 OS << "user defined class '" << CI.ValueName << "'\n";
2381 OS << " NumMatchClassKinds\n";
2382 OS << "};\n\n";
2384 OS << "} // end anonymous namespace\n\n";
2387 /// emitMatchClassDiagStrings - Emit a function to get the diagnostic text to be
2388 /// used when an assembly operand does not match the expected operand class.
2389 static void emitOperandMatchErrorDiagStrings(AsmMatcherInfo &Info, raw_ostream &OS) {
2390 // If the target does not use DiagnosticString for any operands, don't emit
2391 // an unused function.
2392 if (std::all_of(
2393 Info.Classes.begin(), Info.Classes.end(),
2394 [](const ClassInfo &CI) { return CI.DiagnosticString.empty(); }))
2395 return;
2397 OS << "static const char *getMatchKindDiag(" << Info.Target.getName()
2398 << "AsmParser::" << Info.Target.getName()
2399 << "MatchResultTy MatchResult) {\n";
2400 OS << " switch (MatchResult) {\n";
2402 for (const auto &CI: Info.Classes) {
2403 if (!CI.DiagnosticString.empty()) {
2404 assert(!CI.DiagnosticType.empty() &&
2405 "DiagnosticString set without DiagnosticType");
2406 OS << " case " << Info.Target.getName()
2407 << "AsmParser::Match_" << CI.DiagnosticType << ":\n";
2408 OS << " return \"" << CI.DiagnosticString << "\";\n";
2412 OS << " default:\n";
2413 OS << " return nullptr;\n";
2415 OS << " }\n";
2416 OS << "}\n\n";
2419 static void emitRegisterMatchErrorFunc(AsmMatcherInfo &Info, raw_ostream &OS) {
2420 OS << "static unsigned getDiagKindFromRegisterClass(MatchClassKind "
2421 "RegisterClass) {\n";
2422 if (none_of(Info.Classes, [](const ClassInfo &CI) {
2423 return CI.isRegisterClass() && !CI.DiagnosticType.empty();
2424 })) {
2425 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n";
2426 } else {
2427 OS << " switch (RegisterClass) {\n";
2428 for (const auto &CI: Info.Classes) {
2429 if (CI.isRegisterClass() && !CI.DiagnosticType.empty()) {
2430 OS << " case " << CI.Name << ":\n";
2431 OS << " return " << Info.Target.getName() << "AsmParser::Match_"
2432 << CI.DiagnosticType << ";\n";
2436 OS << " default:\n";
2437 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n";
2439 OS << " }\n";
2441 OS << "}\n\n";
2444 /// emitValidateOperandClass - Emit the function to validate an operand class.
2445 static void emitValidateOperandClass(AsmMatcherInfo &Info,
2446 raw_ostream &OS) {
2447 OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
2448 << "MatchClassKind Kind) {\n";
2449 OS << " " << Info.Target.getName() << "Operand &Operand = ("
2450 << Info.Target.getName() << "Operand&)GOp;\n";
2452 // The InvalidMatchClass is not to match any operand.
2453 OS << " if (Kind == InvalidMatchClass)\n";
2454 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2456 // Check for Token operands first.
2457 // FIXME: Use a more specific diagnostic type.
2458 OS << " if (Operand.isToken() && Kind <= MCK_LAST_TOKEN)\n";
2459 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2460 << " MCTargetAsmParser::Match_Success :\n"
2461 << " MCTargetAsmParser::Match_InvalidOperand;\n\n";
2463 // Check the user classes. We don't care what order since we're only
2464 // actually matching against one of them.
2465 OS << " switch (Kind) {\n"
2466 " default: break;\n";
2467 for (const auto &CI : Info.Classes) {
2468 if (!CI.isUserClass())
2469 continue;
2471 OS << " // '" << CI.ClassName << "' class\n";
2472 OS << " case " << CI.Name << ": {\n";
2473 OS << " DiagnosticPredicate DP(Operand." << CI.PredicateMethod
2474 << "());\n";
2475 OS << " if (DP.isMatch())\n";
2476 OS << " return MCTargetAsmParser::Match_Success;\n";
2477 if (!CI.DiagnosticType.empty()) {
2478 OS << " if (DP.isNearMatch())\n";
2479 OS << " return " << Info.Target.getName() << "AsmParser::Match_"
2480 << CI.DiagnosticType << ";\n";
2481 OS << " break;\n";
2483 else
2484 OS << " break;\n";
2485 OS << " }\n";
2487 OS << " } // end switch (Kind)\n\n";
2489 // Check for register operands, including sub-classes.
2490 OS << " if (Operand.isReg()) {\n";
2491 OS << " MatchClassKind OpKind;\n";
2492 OS << " switch (Operand.getReg()) {\n";
2493 OS << " default: OpKind = InvalidMatchClass; break;\n";
2494 for (const auto &RC : Info.RegisterClasses)
2495 OS << " case " << RC.first->getValueAsString("Namespace") << "::"
2496 << RC.first->getName() << ": OpKind = " << RC.second->Name
2497 << "; break;\n";
2498 OS << " }\n";
2499 OS << " return isSubclass(OpKind, Kind) ? "
2500 << "(unsigned)MCTargetAsmParser::Match_Success :\n "
2501 << " getDiagKindFromRegisterClass(Kind);\n }\n\n";
2503 // Expected operand is a register, but actual is not.
2504 OS << " if (Kind > MCK_LAST_TOKEN && Kind <= MCK_LAST_REGISTER)\n";
2505 OS << " return getDiagKindFromRegisterClass(Kind);\n\n";
2507 // Generic fallthrough match failure case for operands that don't have
2508 // specialized diagnostic types.
2509 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n";
2510 OS << "}\n\n";
2513 /// emitIsSubclass - Emit the subclass predicate function.
2514 static void emitIsSubclass(CodeGenTarget &Target,
2515 std::forward_list<ClassInfo> &Infos,
2516 raw_ostream &OS) {
2517 OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2518 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2519 OS << " if (A == B)\n";
2520 OS << " return true;\n\n";
2522 bool EmittedSwitch = false;
2523 for (const auto &A : Infos) {
2524 std::vector<StringRef> SuperClasses;
2525 if (A.IsOptional)
2526 SuperClasses.push_back("OptionalMatchClass");
2527 for (const auto &B : Infos) {
2528 if (&A != &B && A.isSubsetOf(B))
2529 SuperClasses.push_back(B.Name);
2532 if (SuperClasses.empty())
2533 continue;
2535 // If this is the first SuperClass, emit the switch header.
2536 if (!EmittedSwitch) {
2537 OS << " switch (A) {\n";
2538 OS << " default:\n";
2539 OS << " return false;\n";
2540 EmittedSwitch = true;
2543 OS << "\n case " << A.Name << ":\n";
2545 if (SuperClasses.size() == 1) {
2546 OS << " return B == " << SuperClasses.back() << ";\n";
2547 continue;
2550 if (!SuperClasses.empty()) {
2551 OS << " switch (B) {\n";
2552 OS << " default: return false;\n";
2553 for (StringRef SC : SuperClasses)
2554 OS << " case " << SC << ": return true;\n";
2555 OS << " }\n";
2556 } else {
2557 // No case statement to emit
2558 OS << " return false;\n";
2562 // If there were case statements emitted into the string stream write the
2563 // default.
2564 if (EmittedSwitch)
2565 OS << " }\n";
2566 else
2567 OS << " return false;\n";
2569 OS << "}\n\n";
2572 /// emitMatchTokenString - Emit the function to match a token string to the
2573 /// appropriate match class value.
2574 static void emitMatchTokenString(CodeGenTarget &Target,
2575 std::forward_list<ClassInfo> &Infos,
2576 raw_ostream &OS) {
2577 // Construct the match list.
2578 std::vector<StringMatcher::StringPair> Matches;
2579 for (const auto &CI : Infos) {
2580 if (CI.Kind == ClassInfo::Token)
2581 Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";");
2584 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2586 StringMatcher("Name", Matches, OS).Emit();
2588 OS << " return InvalidMatchClass;\n";
2589 OS << "}\n\n";
2592 /// emitMatchRegisterName - Emit the function to match a string to the target
2593 /// specific register enum.
2594 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2595 raw_ostream &OS) {
2596 // Construct the match list.
2597 std::vector<StringMatcher::StringPair> Matches;
2598 const auto &Regs = Target.getRegBank().getRegisters();
2599 for (const CodeGenRegister &Reg : Regs) {
2600 if (Reg.TheDef->getValueAsString("AsmName").empty())
2601 continue;
2603 Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"),
2604 "return " + utostr(Reg.EnumValue) + ";");
2607 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
2609 bool IgnoreDuplicates =
2610 AsmParser->getValueAsBit("AllowDuplicateRegisterNames");
2611 StringMatcher("Name", Matches, OS).Emit(0, IgnoreDuplicates);
2613 OS << " return 0;\n";
2614 OS << "}\n\n";
2617 /// Emit the function to match a string to the target
2618 /// specific register enum.
2619 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser,
2620 raw_ostream &OS) {
2621 // Construct the match list.
2622 std::vector<StringMatcher::StringPair> Matches;
2623 const auto &Regs = Target.getRegBank().getRegisters();
2624 for (const CodeGenRegister &Reg : Regs) {
2626 auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames");
2628 for (auto AltName : AltNames) {
2629 AltName = StringRef(AltName).trim();
2631 // don't handle empty alternative names
2632 if (AltName.empty())
2633 continue;
2635 Matches.emplace_back(AltName,
2636 "return " + utostr(Reg.EnumValue) + ";");
2640 OS << "static unsigned MatchRegisterAltName(StringRef Name) {\n";
2642 bool IgnoreDuplicates =
2643 AsmParser->getValueAsBit("AllowDuplicateRegisterNames");
2644 StringMatcher("Name", Matches, OS).Emit(0, IgnoreDuplicates);
2646 OS << " return 0;\n";
2647 OS << "}\n\n";
2650 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
2651 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2652 // Get the set of diagnostic types from all of the operand classes.
2653 std::set<StringRef> Types;
2654 for (const auto &OpClassEntry : Info.AsmOperandClasses) {
2655 if (!OpClassEntry.second->DiagnosticType.empty())
2656 Types.insert(OpClassEntry.second->DiagnosticType);
2658 for (const auto &OpClassEntry : Info.RegisterClassClasses) {
2659 if (!OpClassEntry.second->DiagnosticType.empty())
2660 Types.insert(OpClassEntry.second->DiagnosticType);
2663 if (Types.empty()) return;
2665 // Now emit the enum entries.
2666 for (StringRef Type : Types)
2667 OS << " Match_" << Type << ",\n";
2668 OS << " END_OPERAND_DIAGNOSTIC_TYPES\n";
2671 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2672 /// user-level name for a subtarget feature.
2673 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2674 OS << "// User-level names for subtarget features that participate in\n"
2675 << "// instruction matching.\n"
2676 << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2677 if (!Info.SubtargetFeatures.empty()) {
2678 OS << " switch(Val) {\n";
2679 for (const auto &SF : Info.SubtargetFeatures) {
2680 const SubtargetFeatureInfo &SFI = SF.second;
2681 // FIXME: Totally just a placeholder name to get the algorithm working.
2682 OS << " case " << SFI.getEnumBitName() << ": return \""
2683 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2685 OS << " default: return \"(unknown)\";\n";
2686 OS << " }\n";
2687 } else {
2688 // Nothing to emit, so skip the switch
2689 OS << " return \"(unknown)\";\n";
2691 OS << "}\n\n";
2694 static std::string GetAliasRequiredFeatures(Record *R,
2695 const AsmMatcherInfo &Info) {
2696 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2697 std::string Result;
2699 if (ReqFeatures.empty())
2700 return Result;
2702 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2703 const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2705 if (!F)
2706 PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2707 "' is not marked as an AssemblerPredicate!");
2709 if (i)
2710 Result += " && ";
2712 Result += "Features.test(" + F->getEnumBitName() + ')';
2715 return Result;
2718 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info,
2719 std::vector<Record*> &Aliases,
2720 unsigned Indent = 0,
2721 StringRef AsmParserVariantName = StringRef()){
2722 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
2723 // iteration order of the map is stable.
2724 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2726 for (Record *R : Aliases) {
2727 // FIXME: Allow AssemblerVariantName to be a comma separated list.
2728 StringRef AsmVariantName = R->getValueAsString("AsmVariantName");
2729 if (AsmVariantName != AsmParserVariantName)
2730 continue;
2731 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2733 if (AliasesFromMnemonic.empty())
2734 return;
2736 // Process each alias a "from" mnemonic at a time, building the code executed
2737 // by the string remapper.
2738 std::vector<StringMatcher::StringPair> Cases;
2739 for (const auto &AliasEntry : AliasesFromMnemonic) {
2740 const std::vector<Record*> &ToVec = AliasEntry.second;
2742 // Loop through each alias and emit code that handles each case. If there
2743 // are two instructions without predicates, emit an error. If there is one,
2744 // emit it last.
2745 std::string MatchCode;
2746 int AliasWithNoPredicate = -1;
2748 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2749 Record *R = ToVec[i];
2750 std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2752 // If this unconditionally matches, remember it for later and diagnose
2753 // duplicates.
2754 if (FeatureMask.empty()) {
2755 if (AliasWithNoPredicate != -1) {
2756 // We can't have two aliases from the same mnemonic with no predicate.
2757 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2758 "two MnemonicAliases with the same 'from' mnemonic!");
2759 PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2762 AliasWithNoPredicate = i;
2763 continue;
2765 if (R->getValueAsString("ToMnemonic") == AliasEntry.first)
2766 PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2768 if (!MatchCode.empty())
2769 MatchCode += "else ";
2770 MatchCode += "if (" + FeatureMask + ")\n";
2771 MatchCode += " Mnemonic = \"";
2772 MatchCode += R->getValueAsString("ToMnemonic");
2773 MatchCode += "\";\n";
2776 if (AliasWithNoPredicate != -1) {
2777 Record *R = ToVec[AliasWithNoPredicate];
2778 if (!MatchCode.empty())
2779 MatchCode += "else\n ";
2780 MatchCode += "Mnemonic = \"";
2781 MatchCode += R->getValueAsString("ToMnemonic");
2782 MatchCode += "\";\n";
2785 MatchCode += "return;";
2787 Cases.push_back(std::make_pair(AliasEntry.first, MatchCode));
2789 StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2792 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2793 /// emit a function for them and return true, otherwise return false.
2794 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2795 CodeGenTarget &Target) {
2796 // Ignore aliases when match-prefix is set.
2797 if (!MatchPrefix.empty())
2798 return false;
2800 std::vector<Record*> Aliases =
2801 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2802 if (Aliases.empty()) return false;
2804 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2805 "const FeatureBitset &Features, unsigned VariantID) {\n";
2806 OS << " switch (VariantID) {\n";
2807 unsigned VariantCount = Target.getAsmParserVariantCount();
2808 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2809 Record *AsmVariant = Target.getAsmParserVariant(VC);
2810 int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2811 StringRef AsmParserVariantName = AsmVariant->getValueAsString("Name");
2812 OS << " case " << AsmParserVariantNo << ":\n";
2813 emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2814 AsmParserVariantName);
2815 OS << " break;\n";
2817 OS << " }\n";
2819 // Emit aliases that apply to all variants.
2820 emitMnemonicAliasVariant(OS, Info, Aliases);
2822 OS << "}\n\n";
2824 return true;
2827 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2828 const AsmMatcherInfo &Info, StringRef ClassName,
2829 StringToOffsetTable &StringTable,
2830 unsigned MaxMnemonicIndex,
2831 unsigned MaxFeaturesIndex,
2832 bool HasMnemonicFirst) {
2833 unsigned MaxMask = 0;
2834 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2835 MaxMask |= OMI.OperandMask;
2838 // Emit the static custom operand parsing table;
2839 OS << "namespace {\n";
2840 OS << " struct OperandMatchEntry {\n";
2841 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
2842 << " Mnemonic;\n";
2843 OS << " " << getMinimalTypeForRange(MaxMask)
2844 << " OperandMask;\n";
2845 OS << " " << getMinimalTypeForRange(std::distance(
2846 Info.Classes.begin(), Info.Classes.end())) << " Class;\n";
2847 OS << " " << getMinimalTypeForRange(MaxFeaturesIndex)
2848 << " RequiredFeaturesIdx;\n\n";
2849 OS << " StringRef getMnemonic() const {\n";
2850 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
2851 OS << " MnemonicTable[Mnemonic]);\n";
2852 OS << " }\n";
2853 OS << " };\n\n";
2855 OS << " // Predicate for searching for an opcode.\n";
2856 OS << " struct LessOpcodeOperand {\n";
2857 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2858 OS << " return LHS.getMnemonic() < RHS;\n";
2859 OS << " }\n";
2860 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2861 OS << " return LHS < RHS.getMnemonic();\n";
2862 OS << " }\n";
2863 OS << " bool operator()(const OperandMatchEntry &LHS,";
2864 OS << " const OperandMatchEntry &RHS) {\n";
2865 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
2866 OS << " }\n";
2867 OS << " };\n";
2869 OS << "} // end anonymous namespace\n\n";
2871 OS << "static const OperandMatchEntry OperandMatchTable["
2872 << Info.OperandMatchInfo.size() << "] = {\n";
2874 OS << " /* Operand List Mnemonic, Mask, Operand Class, Features */\n";
2875 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2876 const MatchableInfo &II = *OMI.MI;
2878 OS << " { ";
2880 // Store a pascal-style length byte in the mnemonic.
2881 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2882 OS << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2883 << " /* " << II.Mnemonic << " */, ";
2885 OS << OMI.OperandMask;
2886 OS << " /* ";
2887 bool printComma = false;
2888 for (int i = 0, e = 31; i !=e; ++i)
2889 if (OMI.OperandMask & (1 << i)) {
2890 if (printComma)
2891 OS << ", ";
2892 OS << i;
2893 printComma = true;
2895 OS << " */, ";
2897 OS << OMI.CI->Name;
2899 // Write the required features mask.
2900 OS << ", AMFBS";
2901 if (II.RequiredFeatures.empty())
2902 OS << "_None";
2903 else
2904 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i)
2905 OS << '_' << II.RequiredFeatures[i]->TheDef->getName();
2907 OS << " },\n";
2909 OS << "};\n\n";
2911 // Emit the operand class switch to call the correct custom parser for
2912 // the found operand class.
2913 OS << "OperandMatchResultTy " << Target.getName() << ClassName << "::\n"
2914 << "tryCustomParseOperand(OperandVector"
2915 << " &Operands,\n unsigned MCK) {\n\n"
2916 << " switch(MCK) {\n";
2918 for (const auto &CI : Info.Classes) {
2919 if (CI.ParserMethod.empty())
2920 continue;
2921 OS << " case " << CI.Name << ":\n"
2922 << " return " << CI.ParserMethod << "(Operands);\n";
2925 OS << " default:\n";
2926 OS << " return MatchOperand_NoMatch;\n";
2927 OS << " }\n";
2928 OS << " return MatchOperand_NoMatch;\n";
2929 OS << "}\n\n";
2931 // Emit the static custom operand parser. This code is very similar with
2932 // the other matcher. Also use MatchResultTy here just in case we go for
2933 // a better error handling.
2934 OS << "OperandMatchResultTy " << Target.getName() << ClassName << "::\n"
2935 << "MatchOperandParserImpl(OperandVector"
2936 << " &Operands,\n StringRef Mnemonic,\n"
2937 << " bool ParseForAllFeatures) {\n";
2939 // Emit code to get the available features.
2940 OS << " // Get the current feature set.\n";
2941 OS << " const FeatureBitset &AvailableFeatures = getAvailableFeatures();\n\n";
2943 OS << " // Get the next operand index.\n";
2944 OS << " unsigned NextOpNum = Operands.size()"
2945 << (HasMnemonicFirst ? " - 1" : "") << ";\n";
2947 // Emit code to search the table.
2948 OS << " // Search the table.\n";
2949 if (HasMnemonicFirst) {
2950 OS << " auto MnemonicRange =\n";
2951 OS << " std::equal_range(std::begin(OperandMatchTable), "
2952 "std::end(OperandMatchTable),\n";
2953 OS << " Mnemonic, LessOpcodeOperand());\n\n";
2954 } else {
2955 OS << " auto MnemonicRange = std::make_pair(std::begin(OperandMatchTable),"
2956 " std::end(OperandMatchTable));\n";
2957 OS << " if (!Mnemonic.empty())\n";
2958 OS << " MnemonicRange =\n";
2959 OS << " std::equal_range(std::begin(OperandMatchTable), "
2960 "std::end(OperandMatchTable),\n";
2961 OS << " Mnemonic, LessOpcodeOperand());\n\n";
2964 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
2965 OS << " return MatchOperand_NoMatch;\n\n";
2967 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2968 << " *ie = MnemonicRange.second; it != ie; ++it) {\n";
2970 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
2971 OS << " assert(Mnemonic == it->getMnemonic());\n\n";
2973 // Emit check that the required features are available.
2974 OS << " // check if the available features match\n";
2975 OS << " const FeatureBitset &RequiredFeatures = "
2976 "FeatureBitsets[it->RequiredFeaturesIdx];\n";
2977 OS << " if (!ParseForAllFeatures && (AvailableFeatures & "
2978 "RequiredFeatures) != RequiredFeatures)\n";
2979 OS << " continue;\n\n";
2981 // Emit check to ensure the operand number matches.
2982 OS << " // check if the operand in question has a custom parser.\n";
2983 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n";
2984 OS << " continue;\n\n";
2986 // Emit call to the custom parser method
2987 OS << " // call custom parse method to handle the operand\n";
2988 OS << " OperandMatchResultTy Result = ";
2989 OS << "tryCustomParseOperand(Operands, it->Class);\n";
2990 OS << " if (Result != MatchOperand_NoMatch)\n";
2991 OS << " return Result;\n";
2992 OS << " }\n\n";
2994 OS << " // Okay, we had no match.\n";
2995 OS << " return MatchOperand_NoMatch;\n";
2996 OS << "}\n\n";
2999 static void emitAsmTiedOperandConstraints(CodeGenTarget &Target,
3000 AsmMatcherInfo &Info,
3001 raw_ostream &OS) {
3002 std::string AsmParserName =
3003 Info.AsmParser->getValueAsString("AsmParserClassName");
3004 OS << "static bool ";
3005 OS << "checkAsmTiedOperandConstraints(const " << Target.getName()
3006 << AsmParserName << "&AsmParser,\n";
3007 OS << " unsigned Kind,\n";
3008 OS << " const OperandVector &Operands,\n";
3009 OS << " uint64_t &ErrorInfo) {\n";
3010 OS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n";
3011 OS << " const uint8_t *Converter = ConversionTable[Kind];\n";
3012 OS << " for (const uint8_t *p = Converter; *p; p+= 2) {\n";
3013 OS << " switch (*p) {\n";
3014 OS << " case CVT_Tied: {\n";
3015 OS << " unsigned OpIdx = *(p+1);\n";
3016 OS << " assert(OpIdx < (size_t)(std::end(TiedAsmOperandTable) -\n";
3017 OS << " std::begin(TiedAsmOperandTable)) &&\n";
3018 OS << " \"Tied operand not found\");\n";
3019 OS << " unsigned OpndNum1 = TiedAsmOperandTable[OpIdx][1];\n";
3020 OS << " unsigned OpndNum2 = TiedAsmOperandTable[OpIdx][2];\n";
3021 OS << " if (OpndNum1 != OpndNum2) {\n";
3022 OS << " auto &SrcOp1 = Operands[OpndNum1];\n";
3023 OS << " auto &SrcOp2 = Operands[OpndNum2];\n";
3024 OS << " if (SrcOp1->isReg() && SrcOp2->isReg()) {\n";
3025 OS << " if (!AsmParser.regsEqual(*SrcOp1, *SrcOp2)) {\n";
3026 OS << " ErrorInfo = OpndNum2;\n";
3027 OS << " return false;\n";
3028 OS << " }\n";
3029 OS << " }\n";
3030 OS << " }\n";
3031 OS << " break;\n";
3032 OS << " }\n";
3033 OS << " default:\n";
3034 OS << " break;\n";
3035 OS << " }\n";
3036 OS << " }\n";
3037 OS << " return true;\n";
3038 OS << "}\n\n";
3041 static void emitMnemonicSpellChecker(raw_ostream &OS, CodeGenTarget &Target,
3042 unsigned VariantCount) {
3043 OS << "static std::string " << Target.getName()
3044 << "MnemonicSpellCheck(StringRef S, const FeatureBitset &FBS,"
3045 << " unsigned VariantID) {\n";
3046 if (!VariantCount)
3047 OS << " return \"\";";
3048 else {
3049 OS << " const unsigned MaxEditDist = 2;\n";
3050 OS << " std::vector<StringRef> Candidates;\n";
3051 OS << " StringRef Prev = \"\";\n\n";
3053 OS << " // Find the appropriate table for this asm variant.\n";
3054 OS << " const MatchEntry *Start, *End;\n";
3055 OS << " switch (VariantID) {\n";
3056 OS << " default: llvm_unreachable(\"invalid variant!\");\n";
3057 for (unsigned VC = 0; VC != VariantCount; ++VC) {
3058 Record *AsmVariant = Target.getAsmParserVariant(VC);
3059 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3060 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3061 << "); End = std::end(MatchTable" << VC << "); break;\n";
3063 OS << " }\n\n";
3064 OS << " for (auto I = Start; I < End; I++) {\n";
3065 OS << " // Ignore unsupported instructions.\n";
3066 OS << " const FeatureBitset &RequiredFeatures = "
3067 "FeatureBitsets[I->RequiredFeaturesIdx];\n";
3068 OS << " if ((FBS & RequiredFeatures) != RequiredFeatures)\n";
3069 OS << " continue;\n";
3070 OS << "\n";
3071 OS << " StringRef T = I->getMnemonic();\n";
3072 OS << " // Avoid recomputing the edit distance for the same string.\n";
3073 OS << " if (T.equals(Prev))\n";
3074 OS << " continue;\n";
3075 OS << "\n";
3076 OS << " Prev = T;\n";
3077 OS << " unsigned Dist = S.edit_distance(T, false, MaxEditDist);\n";
3078 OS << " if (Dist <= MaxEditDist)\n";
3079 OS << " Candidates.push_back(T);\n";
3080 OS << " }\n";
3081 OS << "\n";
3082 OS << " if (Candidates.empty())\n";
3083 OS << " return \"\";\n";
3084 OS << "\n";
3085 OS << " std::string Res = \", did you mean: \";\n";
3086 OS << " unsigned i = 0;\n";
3087 OS << " for( ; i < Candidates.size() - 1; i++)\n";
3088 OS << " Res += Candidates[i].str() + \", \";\n";
3089 OS << " return Res + Candidates[i].str() + \"?\";\n";
3091 OS << "}\n";
3092 OS << "\n";
3096 // Emit a function mapping match classes to strings, for debugging.
3097 static void emitMatchClassKindNames(std::forward_list<ClassInfo> &Infos,
3098 raw_ostream &OS) {
3099 OS << "#ifndef NDEBUG\n";
3100 OS << "const char *getMatchClassName(MatchClassKind Kind) {\n";
3101 OS << " switch (Kind) {\n";
3103 OS << " case InvalidMatchClass: return \"InvalidMatchClass\";\n";
3104 OS << " case OptionalMatchClass: return \"OptionalMatchClass\";\n";
3105 for (const auto &CI : Infos) {
3106 OS << " case " << CI.Name << ": return \"" << CI.Name << "\";\n";
3108 OS << " case NumMatchClassKinds: return \"NumMatchClassKinds\";\n";
3110 OS << " }\n";
3111 OS << " llvm_unreachable(\"unhandled MatchClassKind!\");\n";
3112 OS << "}\n\n";
3113 OS << "#endif // NDEBUG\n";
3116 static std::string
3117 getNameForFeatureBitset(const std::vector<Record *> &FeatureBitset) {
3118 std::string Name = "AMFBS";
3119 for (const auto &Feature : FeatureBitset)
3120 Name += ("_" + Feature->getName()).str();
3121 return Name;
3124 void AsmMatcherEmitter::run(raw_ostream &OS) {
3125 CodeGenTarget Target(Records);
3126 Record *AsmParser = Target.getAsmParser();
3127 StringRef ClassName = AsmParser->getValueAsString("AsmParserClassName");
3129 // Compute the information on the instructions to match.
3130 AsmMatcherInfo Info(AsmParser, Target, Records);
3131 Info.buildInfo();
3133 // Sort the instruction table using the partial order on classes. We use
3134 // stable_sort to ensure that ambiguous instructions are still
3135 // deterministically ordered.
3136 llvm::stable_sort(
3137 Info.Matchables,
3138 [](const std::unique_ptr<MatchableInfo> &a,
3139 const std::unique_ptr<MatchableInfo> &b) { return *a < *b; });
3141 #ifdef EXPENSIVE_CHECKS
3142 // Verify that the table is sorted and operator < works transitively.
3143 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
3144 ++I) {
3145 for (auto J = I; J != E; ++J) {
3146 assert(!(**J < **I));
3149 #endif
3151 DEBUG_WITH_TYPE("instruction_info", {
3152 for (const auto &MI : Info.Matchables)
3153 MI->dump();
3156 // Check for ambiguous matchables.
3157 DEBUG_WITH_TYPE("ambiguous_instrs", {
3158 unsigned NumAmbiguous = 0;
3159 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
3160 ++I) {
3161 for (auto J = std::next(I); J != E; ++J) {
3162 const MatchableInfo &A = **I;
3163 const MatchableInfo &B = **J;
3165 if (A.couldMatchAmbiguouslyWith(B)) {
3166 errs() << "warning: ambiguous matchables:\n";
3167 A.dump();
3168 errs() << "\nis incomparable with:\n";
3169 B.dump();
3170 errs() << "\n\n";
3171 ++NumAmbiguous;
3175 if (NumAmbiguous)
3176 errs() << "warning: " << NumAmbiguous
3177 << " ambiguous matchables!\n";
3180 // Compute the information on the custom operand parsing.
3181 Info.buildOperandMatchInfo();
3183 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
3184 bool HasOptionalOperands = Info.hasOptionalOperands();
3185 bool ReportMultipleNearMisses =
3186 AsmParser->getValueAsBit("ReportMultipleNearMisses");
3188 // Write the output.
3190 // Information for the class declaration.
3191 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
3192 OS << "#undef GET_ASSEMBLER_HEADER\n";
3193 OS << " // This should be included into the middle of the declaration of\n";
3194 OS << " // your subclasses implementation of MCTargetAsmParser.\n";
3195 OS << " FeatureBitset ComputeAvailableFeatures(const FeatureBitset& FB) const;\n";
3196 if (HasOptionalOperands) {
3197 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, "
3198 << "unsigned Opcode,\n"
3199 << " const OperandVector &Operands,\n"
3200 << " const SmallBitVector &OptionalOperandsMask);\n";
3201 } else {
3202 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, "
3203 << "unsigned Opcode,\n"
3204 << " const OperandVector &Operands);\n";
3206 OS << " void convertToMapAndConstraints(unsigned Kind,\n ";
3207 OS << " const OperandVector &Operands) override;\n";
3208 OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
3209 << " MCInst &Inst,\n";
3210 if (ReportMultipleNearMisses)
3211 OS << " SmallVectorImpl<NearMissInfo> *NearMisses,\n";
3212 else
3213 OS << " uint64_t &ErrorInfo,\n"
3214 << " FeatureBitset &MissingFeatures,\n";
3215 OS << " bool matchingInlineAsm,\n"
3216 << " unsigned VariantID = 0);\n";
3217 if (!ReportMultipleNearMisses)
3218 OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
3219 << " MCInst &Inst,\n"
3220 << " uint64_t &ErrorInfo,\n"
3221 << " bool matchingInlineAsm,\n"
3222 << " unsigned VariantID = 0) {\n"
3223 << " FeatureBitset MissingFeatures;\n"
3224 << " return MatchInstructionImpl(Operands, Inst, ErrorInfo, MissingFeatures,\n"
3225 << " matchingInlineAsm, VariantID);\n"
3226 << " }\n\n";
3229 if (!Info.OperandMatchInfo.empty()) {
3230 OS << " OperandMatchResultTy MatchOperandParserImpl(\n";
3231 OS << " OperandVector &Operands,\n";
3232 OS << " StringRef Mnemonic,\n";
3233 OS << " bool ParseForAllFeatures = false);\n";
3235 OS << " OperandMatchResultTy tryCustomParseOperand(\n";
3236 OS << " OperandVector &Operands,\n";
3237 OS << " unsigned MCK);\n\n";
3240 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
3242 // Emit the operand match diagnostic enum names.
3243 OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
3244 OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
3245 emitOperandDiagnosticTypes(Info, OS);
3246 OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
3248 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
3249 OS << "#undef GET_REGISTER_MATCHER\n\n";
3251 // Emit the subtarget feature enumeration.
3252 SubtargetFeatureInfo::emitSubtargetFeatureBitEnumeration(
3253 Info.SubtargetFeatures, OS);
3255 // Emit the function to match a register name to number.
3256 // This should be omitted for Mips target
3257 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
3258 emitMatchRegisterName(Target, AsmParser, OS);
3260 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName"))
3261 emitMatchRegisterAltName(Target, AsmParser, OS);
3263 OS << "#endif // GET_REGISTER_MATCHER\n\n";
3265 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
3266 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
3268 // Generate the helper function to get the names for subtarget features.
3269 emitGetSubtargetFeatureName(Info, OS);
3271 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
3273 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
3274 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
3276 // Generate the function that remaps for mnemonic aliases.
3277 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
3279 // Generate the convertToMCInst function to convert operands into an MCInst.
3280 // Also, generate the convertToMapAndConstraints function for MS-style inline
3281 // assembly. The latter doesn't actually generate a MCInst.
3282 unsigned NumConverters = emitConvertFuncs(Target, ClassName, Info.Matchables,
3283 HasMnemonicFirst,
3284 HasOptionalOperands, OS);
3286 // Emit the enumeration for classes which participate in matching.
3287 emitMatchClassEnumeration(Target, Info.Classes, OS);
3289 // Emit a function to get the user-visible string to describe an operand
3290 // match failure in diagnostics.
3291 emitOperandMatchErrorDiagStrings(Info, OS);
3293 // Emit a function to map register classes to operand match failure codes.
3294 emitRegisterMatchErrorFunc(Info, OS);
3296 // Emit the routine to match token strings to their match class.
3297 emitMatchTokenString(Target, Info.Classes, OS);
3299 // Emit the subclass predicate routine.
3300 emitIsSubclass(Target, Info.Classes, OS);
3302 // Emit the routine to validate an operand against a match class.
3303 emitValidateOperandClass(Info, OS);
3305 emitMatchClassKindNames(Info.Classes, OS);
3307 // Emit the available features compute function.
3308 SubtargetFeatureInfo::emitComputeAssemblerAvailableFeatures(
3309 Info.Target.getName(), ClassName, "ComputeAvailableFeatures",
3310 Info.SubtargetFeatures, OS);
3312 if (!ReportMultipleNearMisses)
3313 emitAsmTiedOperandConstraints(Target, Info, OS);
3315 StringToOffsetTable StringTable;
3317 size_t MaxNumOperands = 0;
3318 unsigned MaxMnemonicIndex = 0;
3319 bool HasDeprecation = false;
3320 for (const auto &MI : Info.Matchables) {
3321 MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
3322 HasDeprecation |= MI->HasDeprecation;
3324 // Store a pascal-style length byte in the mnemonic.
3325 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
3326 MaxMnemonicIndex = std::max(MaxMnemonicIndex,
3327 StringTable.GetOrAddStringOffset(LenMnemonic, false));
3330 OS << "static const char *const MnemonicTable =\n";
3331 StringTable.EmitString(OS);
3332 OS << ";\n\n";
3334 std::vector<std::vector<Record *>> FeatureBitsets;
3335 for (const auto &MI : Info.Matchables) {
3336 if (MI->RequiredFeatures.empty())
3337 continue;
3338 FeatureBitsets.emplace_back();
3339 for (unsigned I = 0, E = MI->RequiredFeatures.size(); I != E; ++I)
3340 FeatureBitsets.back().push_back(MI->RequiredFeatures[I]->TheDef);
3343 llvm::sort(FeatureBitsets, [&](const std::vector<Record *> &A,
3344 const std::vector<Record *> &B) {
3345 if (A.size() < B.size())
3346 return true;
3347 if (A.size() > B.size())
3348 return false;
3349 for (const auto &Pair : zip(A, B)) {
3350 if (std::get<0>(Pair)->getName() < std::get<1>(Pair)->getName())
3351 return true;
3352 if (std::get<0>(Pair)->getName() > std::get<1>(Pair)->getName())
3353 return false;
3355 return false;
3357 FeatureBitsets.erase(
3358 std::unique(FeatureBitsets.begin(), FeatureBitsets.end()),
3359 FeatureBitsets.end());
3360 OS << "// Feature bitsets.\n"
3361 << "enum : " << getMinimalTypeForRange(FeatureBitsets.size()) << " {\n"
3362 << " AMFBS_None,\n";
3363 for (const auto &FeatureBitset : FeatureBitsets) {
3364 if (FeatureBitset.empty())
3365 continue;
3366 OS << " " << getNameForFeatureBitset(FeatureBitset) << ",\n";
3368 OS << "};\n\n"
3369 << "static constexpr FeatureBitset FeatureBitsets[] = {\n"
3370 << " {}, // AMFBS_None\n";
3371 for (const auto &FeatureBitset : FeatureBitsets) {
3372 if (FeatureBitset.empty())
3373 continue;
3374 OS << " {";
3375 for (const auto &Feature : FeatureBitset) {
3376 const auto &I = Info.SubtargetFeatures.find(Feature);
3377 assert(I != Info.SubtargetFeatures.end() && "Didn't import predicate?");
3378 OS << I->second.getEnumBitName() << ", ";
3380 OS << "},\n";
3382 OS << "};\n\n";
3384 // Emit the static match table; unused classes get initialized to 0 which is
3385 // guaranteed to be InvalidMatchClass.
3387 // FIXME: We can reduce the size of this table very easily. First, we change
3388 // it so that store the kinds in separate bit-fields for each index, which
3389 // only needs to be the max width used for classes at that index (we also need
3390 // to reject based on this during classification). If we then make sure to
3391 // order the match kinds appropriately (putting mnemonics last), then we
3392 // should only end up using a few bits for each class, especially the ones
3393 // following the mnemonic.
3394 OS << "namespace {\n";
3395 OS << " struct MatchEntry {\n";
3396 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
3397 << " Mnemonic;\n";
3398 OS << " uint16_t Opcode;\n";
3399 OS << " " << getMinimalTypeForRange(NumConverters)
3400 << " ConvertFn;\n";
3401 OS << " " << getMinimalTypeForRange(FeatureBitsets.size())
3402 << " RequiredFeaturesIdx;\n";
3403 OS << " " << getMinimalTypeForRange(
3404 std::distance(Info.Classes.begin(), Info.Classes.end()))
3405 << " Classes[" << MaxNumOperands << "];\n";
3406 OS << " StringRef getMnemonic() const {\n";
3407 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
3408 OS << " MnemonicTable[Mnemonic]);\n";
3409 OS << " }\n";
3410 OS << " };\n\n";
3412 OS << " // Predicate for searching for an opcode.\n";
3413 OS << " struct LessOpcode {\n";
3414 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
3415 OS << " return LHS.getMnemonic() < RHS;\n";
3416 OS << " }\n";
3417 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
3418 OS << " return LHS < RHS.getMnemonic();\n";
3419 OS << " }\n";
3420 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
3421 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
3422 OS << " }\n";
3423 OS << " };\n";
3425 OS << "} // end anonymous namespace\n\n";
3427 unsigned VariantCount = Target.getAsmParserVariantCount();
3428 for (unsigned VC = 0; VC != VariantCount; ++VC) {
3429 Record *AsmVariant = Target.getAsmParserVariant(VC);
3430 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3432 OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
3434 for (const auto &MI : Info.Matchables) {
3435 if (MI->AsmVariantID != AsmVariantNo)
3436 continue;
3438 // Store a pascal-style length byte in the mnemonic.
3439 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
3440 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
3441 << " /* " << MI->Mnemonic << " */, "
3442 << Target.getInstNamespace() << "::"
3443 << MI->getResultInst()->TheDef->getName() << ", "
3444 << MI->ConversionFnKind << ", ";
3446 // Write the required features mask.
3447 OS << "AMFBS";
3448 if (MI->RequiredFeatures.empty())
3449 OS << "_None";
3450 else
3451 for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i)
3452 OS << '_' << MI->RequiredFeatures[i]->TheDef->getName();
3454 OS << ", { ";
3455 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
3456 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
3458 if (i) OS << ", ";
3459 OS << Op.Class->Name;
3461 OS << " }, },\n";
3464 OS << "};\n\n";
3467 OS << "#include \"llvm/Support/Debug.h\"\n";
3468 OS << "#include \"llvm/Support/Format.h\"\n\n";
3470 // Finally, build the match function.
3471 OS << "unsigned " << Target.getName() << ClassName << "::\n"
3472 << "MatchInstructionImpl(const OperandVector &Operands,\n";
3473 OS << " MCInst &Inst,\n";
3474 if (ReportMultipleNearMisses)
3475 OS << " SmallVectorImpl<NearMissInfo> *NearMisses,\n";
3476 else
3477 OS << " uint64_t &ErrorInfo,\n"
3478 << " FeatureBitset &MissingFeatures,\n";
3479 OS << " bool matchingInlineAsm, unsigned VariantID) {\n";
3481 if (!ReportMultipleNearMisses) {
3482 OS << " // Eliminate obvious mismatches.\n";
3483 OS << " if (Operands.size() > "
3484 << (MaxNumOperands + HasMnemonicFirst) << ") {\n";
3485 OS << " ErrorInfo = "
3486 << (MaxNumOperands + HasMnemonicFirst) << ";\n";
3487 OS << " return Match_InvalidOperand;\n";
3488 OS << " }\n\n";
3491 // Emit code to get the available features.
3492 OS << " // Get the current feature set.\n";
3493 OS << " const FeatureBitset &AvailableFeatures = getAvailableFeatures();\n\n";
3495 OS << " // Get the instruction mnemonic, which is the first token.\n";
3496 if (HasMnemonicFirst) {
3497 OS << " StringRef Mnemonic = ((" << Target.getName()
3498 << "Operand&)*Operands[0]).getToken();\n\n";
3499 } else {
3500 OS << " StringRef Mnemonic;\n";
3501 OS << " if (Operands[0]->isToken())\n";
3502 OS << " Mnemonic = ((" << Target.getName()
3503 << "Operand&)*Operands[0]).getToken();\n\n";
3506 if (HasMnemonicAliases) {
3507 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
3508 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
3511 // Emit code to compute the class list for this operand vector.
3512 if (!ReportMultipleNearMisses) {
3513 OS << " // Some state to try to produce better error messages.\n";
3514 OS << " bool HadMatchOtherThanFeatures = false;\n";
3515 OS << " bool HadMatchOtherThanPredicate = false;\n";
3516 OS << " unsigned RetCode = Match_InvalidOperand;\n";
3517 OS << " MissingFeatures.set();\n";
3518 OS << " // Set ErrorInfo to the operand that mismatches if it is\n";
3519 OS << " // wrong for all instances of the instruction.\n";
3520 OS << " ErrorInfo = ~0ULL;\n";
3523 if (HasOptionalOperands) {
3524 OS << " SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n";
3527 // Emit code to search the table.
3528 OS << " // Find the appropriate table for this asm variant.\n";
3529 OS << " const MatchEntry *Start, *End;\n";
3530 OS << " switch (VariantID) {\n";
3531 OS << " default: llvm_unreachable(\"invalid variant!\");\n";
3532 for (unsigned VC = 0; VC != VariantCount; ++VC) {
3533 Record *AsmVariant = Target.getAsmParserVariant(VC);
3534 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3535 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3536 << "); End = std::end(MatchTable" << VC << "); break;\n";
3538 OS << " }\n";
3540 OS << " // Search the table.\n";
3541 if (HasMnemonicFirst) {
3542 OS << " auto MnemonicRange = "
3543 "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
3544 } else {
3545 OS << " auto MnemonicRange = std::make_pair(Start, End);\n";
3546 OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
3547 OS << " if (!Mnemonic.empty())\n";
3548 OS << " MnemonicRange = "
3549 "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
3552 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"AsmMatcher: found \" <<\n"
3553 << " std::distance(MnemonicRange.first, MnemonicRange.second) << \n"
3554 << " \" encodings with mnemonic '\" << Mnemonic << \"'\\n\");\n\n";
3556 OS << " // Return a more specific error code if no mnemonics match.\n";
3557 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
3558 OS << " return Match_MnemonicFail;\n\n";
3560 OS << " for (const MatchEntry *it = MnemonicRange.first, "
3561 << "*ie = MnemonicRange.second;\n";
3562 OS << " it != ie; ++it) {\n";
3563 OS << " const FeatureBitset &RequiredFeatures = "
3564 "FeatureBitsets[it->RequiredFeaturesIdx];\n";
3565 OS << " bool HasRequiredFeatures =\n";
3566 OS << " (AvailableFeatures & RequiredFeatures) == RequiredFeatures;\n";
3567 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Trying to match opcode \"\n";
3568 OS << " << MII.getName(it->Opcode) << \"\\n\");\n";
3570 if (ReportMultipleNearMisses) {
3571 OS << " // Some state to record ways in which this instruction did not match.\n";
3572 OS << " NearMissInfo OperandNearMiss = NearMissInfo::getSuccess();\n";
3573 OS << " NearMissInfo FeaturesNearMiss = NearMissInfo::getSuccess();\n";
3574 OS << " NearMissInfo EarlyPredicateNearMiss = NearMissInfo::getSuccess();\n";
3575 OS << " NearMissInfo LatePredicateNearMiss = NearMissInfo::getSuccess();\n";
3576 OS << " bool MultipleInvalidOperands = false;\n";
3579 if (HasMnemonicFirst) {
3580 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
3581 OS << " assert(Mnemonic == it->getMnemonic());\n";
3584 // Emit check that the subclasses match.
3585 if (!ReportMultipleNearMisses)
3586 OS << " bool OperandsValid = true;\n";
3587 if (HasOptionalOperands) {
3588 OS << " OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n";
3590 OS << " for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex")
3591 << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex")
3592 << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n";
3593 OS << " auto Formal = "
3594 << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n";
3595 OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3596 OS << " dbgs() << \" Matching formal operand class \" << getMatchClassName(Formal)\n";
3597 OS << " << \" against actual operand at index \" << ActualIdx);\n";
3598 OS << " if (ActualIdx < Operands.size())\n";
3599 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \" (\";\n";
3600 OS << " Operands[ActualIdx]->print(dbgs()); dbgs() << \"): \");\n";
3601 OS << " else\n";
3602 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \": \");\n";
3603 OS << " if (ActualIdx >= Operands.size()) {\n";
3604 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"actual operand index out of range \");\n";
3605 if (ReportMultipleNearMisses) {
3606 OS << " bool ThisOperandValid = (Formal == " <<"InvalidMatchClass) || "
3607 "isSubclass(Formal, OptionalMatchClass);\n";
3608 OS << " if (!ThisOperandValid) {\n";
3609 OS << " if (!OperandNearMiss) {\n";
3610 OS << " // Record info about match failure for later use.\n";
3611 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"recording too-few-operands near miss\\n\");\n";
3612 OS << " OperandNearMiss =\n";
3613 OS << " NearMissInfo::getTooFewOperands(Formal, it->Opcode);\n";
3614 OS << " } else if (OperandNearMiss.getKind() != NearMissInfo::NearMissTooFewOperands) {\n";
3615 OS << " // If more than one operand is invalid, give up on this match entry.\n";
3616 OS << " DEBUG_WITH_TYPE(\n";
3617 OS << " \"asm-matcher\",\n";
3618 OS << " dbgs() << \"second invalid operand, giving up on this opcode\\n\");\n";
3619 OS << " MultipleInvalidOperands = true;\n";
3620 OS << " break;\n";
3621 OS << " }\n";
3622 OS << " } else {\n";
3623 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"but formal operand not required\\n\");\n";
3624 OS << " break;\n";
3625 OS << " }\n";
3626 OS << " continue;\n";
3627 } else {
3628 OS << " OperandsValid = (Formal == InvalidMatchClass) || isSubclass(Formal, OptionalMatchClass);\n";
3629 OS << " if (!OperandsValid) ErrorInfo = ActualIdx;\n";
3630 if (HasOptionalOperands) {
3631 OS << " OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands
3632 << ");\n";
3634 OS << " break;\n";
3636 OS << " }\n";
3637 OS << " MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n";
3638 OS << " unsigned Diag = validateOperandClass(Actual, Formal);\n";
3639 OS << " if (Diag == Match_Success) {\n";
3640 OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3641 OS << " dbgs() << \"match success using generic matcher\\n\");\n";
3642 OS << " ++ActualIdx;\n";
3643 OS << " continue;\n";
3644 OS << " }\n";
3645 OS << " // If the generic handler indicates an invalid operand\n";
3646 OS << " // failure, check for a special case.\n";
3647 OS << " if (Diag != Match_Success) {\n";
3648 OS << " unsigned TargetDiag = validateTargetOperandClass(Actual, Formal);\n";
3649 OS << " if (TargetDiag == Match_Success) {\n";
3650 OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3651 OS << " dbgs() << \"match success using target matcher\\n\");\n";
3652 OS << " ++ActualIdx;\n";
3653 OS << " continue;\n";
3654 OS << " }\n";
3655 OS << " // If the target matcher returned a specific error code use\n";
3656 OS << " // that, else use the one from the generic matcher.\n";
3657 OS << " if (TargetDiag != Match_InvalidOperand && "
3658 "HasRequiredFeatures)\n";
3659 OS << " Diag = TargetDiag;\n";
3660 OS << " }\n";
3661 OS << " // If current formal operand wasn't matched and it is optional\n"
3662 << " // then try to match next formal operand\n";
3663 OS << " if (Diag == Match_InvalidOperand "
3664 << "&& isSubclass(Formal, OptionalMatchClass)) {\n";
3665 if (HasOptionalOperands) {
3666 OS << " OptionalOperandsMask.set(FormalIdx);\n";
3668 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"ignoring optional operand\\n\");\n";
3669 OS << " continue;\n";
3670 OS << " }\n";
3672 if (ReportMultipleNearMisses) {
3673 OS << " if (!OperandNearMiss) {\n";
3674 OS << " // If this is the first invalid operand we have seen, record some\n";
3675 OS << " // information about it.\n";
3676 OS << " DEBUG_WITH_TYPE(\n";
3677 OS << " \"asm-matcher\",\n";
3678 OS << " dbgs()\n";
3679 OS << " << \"operand match failed, recording near-miss with diag code \"\n";
3680 OS << " << Diag << \"\\n\");\n";
3681 OS << " OperandNearMiss =\n";
3682 OS << " NearMissInfo::getMissedOperand(Diag, Formal, it->Opcode, ActualIdx);\n";
3683 OS << " ++ActualIdx;\n";
3684 OS << " } else {\n";
3685 OS << " // If more than one operand is invalid, give up on this match entry.\n";
3686 OS << " DEBUG_WITH_TYPE(\n";
3687 OS << " \"asm-matcher\",\n";
3688 OS << " dbgs() << \"second operand mismatch, skipping this opcode\\n\");\n";
3689 OS << " MultipleInvalidOperands = true;\n";
3690 OS << " break;\n";
3691 OS << " }\n";
3692 OS << " }\n\n";
3693 } else {
3694 OS << " // If this operand is broken for all of the instances of this\n";
3695 OS << " // mnemonic, keep track of it so we can report loc info.\n";
3696 OS << " // If we already had a match that only failed due to a\n";
3697 OS << " // target predicate, that diagnostic is preferred.\n";
3698 OS << " if (!HadMatchOtherThanPredicate &&\n";
3699 OS << " (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) {\n";
3700 OS << " if (HasRequiredFeatures && (ErrorInfo != ActualIdx || Diag "
3701 "!= Match_InvalidOperand))\n";
3702 OS << " RetCode = Diag;\n";
3703 OS << " ErrorInfo = ActualIdx;\n";
3704 OS << " }\n";
3705 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
3706 OS << " OperandsValid = false;\n";
3707 OS << " break;\n";
3708 OS << " }\n\n";
3711 if (ReportMultipleNearMisses)
3712 OS << " if (MultipleInvalidOperands) {\n";
3713 else
3714 OS << " if (!OperandsValid) {\n";
3715 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: multiple \"\n";
3716 OS << " \"operand mismatches, ignoring \"\n";
3717 OS << " \"this opcode\\n\");\n";
3718 OS << " continue;\n";
3719 OS << " }\n";
3721 // Emit check that the required features are available.
3722 OS << " if (!HasRequiredFeatures) {\n";
3723 if (!ReportMultipleNearMisses)
3724 OS << " HadMatchOtherThanFeatures = true;\n";
3725 OS << " FeatureBitset NewMissingFeatures = RequiredFeatures & "
3726 "~AvailableFeatures;\n";
3727 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Missing target features:\";\n";
3728 OS << " for (unsigned I = 0, E = NewMissingFeatures.size(); I != E; ++I)\n";
3729 OS << " if (NewMissingFeatures[I])\n";
3730 OS << " dbgs() << ' ' << I;\n";
3731 OS << " dbgs() << \"\\n\");\n";
3732 if (ReportMultipleNearMisses) {
3733 OS << " FeaturesNearMiss = NearMissInfo::getMissedFeature(NewMissingFeatures);\n";
3734 } else {
3735 OS << " if (NewMissingFeatures.count() <=\n"
3736 " MissingFeatures.count())\n";
3737 OS << " MissingFeatures = NewMissingFeatures;\n";
3738 OS << " continue;\n";
3740 OS << " }\n";
3741 OS << "\n";
3742 OS << " Inst.clear();\n\n";
3743 OS << " Inst.setOpcode(it->Opcode);\n";
3744 // Verify the instruction with the target-specific match predicate function.
3745 OS << " // We have a potential match but have not rendered the operands.\n"
3746 << " // Check the target predicate to handle any context sensitive\n"
3747 " // constraints.\n"
3748 << " // For example, Ties that are referenced multiple times must be\n"
3749 " // checked here to ensure the input is the same for each match\n"
3750 " // constraints. If we leave it any later the ties will have been\n"
3751 " // canonicalized\n"
3752 << " unsigned MatchResult;\n"
3753 << " if ((MatchResult = checkEarlyTargetMatchPredicate(Inst, "
3754 "Operands)) != Match_Success) {\n"
3755 << " Inst.clear();\n";
3756 OS << " DEBUG_WITH_TYPE(\n";
3757 OS << " \"asm-matcher\",\n";
3758 OS << " dbgs() << \"Early target match predicate failed with diag code \"\n";
3759 OS << " << MatchResult << \"\\n\");\n";
3760 if (ReportMultipleNearMisses) {
3761 OS << " EarlyPredicateNearMiss = NearMissInfo::getMissedPredicate(MatchResult);\n";
3762 } else {
3763 OS << " RetCode = MatchResult;\n"
3764 << " HadMatchOtherThanPredicate = true;\n"
3765 << " continue;\n";
3767 OS << " }\n\n";
3769 if (ReportMultipleNearMisses) {
3770 OS << " // If we did not successfully match the operands, then we can't convert to\n";
3771 OS << " // an MCInst, so bail out on this instruction variant now.\n";
3772 OS << " if (OperandNearMiss) {\n";
3773 OS << " // If the operand mismatch was the only problem, reprrt it as a near-miss.\n";
3774 OS << " if (NearMisses && !FeaturesNearMiss && !EarlyPredicateNearMiss) {\n";
3775 OS << " DEBUG_WITH_TYPE(\n";
3776 OS << " \"asm-matcher\",\n";
3777 OS << " dbgs()\n";
3778 OS << " << \"Opcode result: one mismatched operand, adding near-miss\\n\");\n";
3779 OS << " NearMisses->push_back(OperandNearMiss);\n";
3780 OS << " } else {\n";
3781 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: multiple \"\n";
3782 OS << " \"types of mismatch, so not \"\n";
3783 OS << " \"reporting near-miss\\n\");\n";
3784 OS << " }\n";
3785 OS << " continue;\n";
3786 OS << " }\n\n";
3789 OS << " if (matchingInlineAsm) {\n";
3790 OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n";
3791 if (!ReportMultipleNearMisses) {
3792 OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
3793 "Operands, ErrorInfo))\n";
3794 OS << " return Match_InvalidTiedOperand;\n";
3795 OS << "\n";
3797 OS << " return Match_Success;\n";
3798 OS << " }\n\n";
3799 OS << " // We have selected a definite instruction, convert the parsed\n"
3800 << " // operands into the appropriate MCInst.\n";
3801 if (HasOptionalOperands) {
3802 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n"
3803 << " OptionalOperandsMask);\n";
3804 } else {
3805 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
3807 OS << "\n";
3809 // Verify the instruction with the target-specific match predicate function.
3810 OS << " // We have a potential match. Check the target predicate to\n"
3811 << " // handle any context sensitive constraints.\n"
3812 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
3813 << " Match_Success) {\n"
3814 << " DEBUG_WITH_TYPE(\"asm-matcher\",\n"
3815 << " dbgs() << \"Target match predicate failed with diag code \"\n"
3816 << " << MatchResult << \"\\n\");\n"
3817 << " Inst.clear();\n";
3818 if (ReportMultipleNearMisses) {
3819 OS << " LatePredicateNearMiss = NearMissInfo::getMissedPredicate(MatchResult);\n";
3820 } else {
3821 OS << " RetCode = MatchResult;\n"
3822 << " HadMatchOtherThanPredicate = true;\n"
3823 << " continue;\n";
3825 OS << " }\n\n";
3827 if (ReportMultipleNearMisses) {
3828 OS << " int NumNearMisses = ((int)(bool)OperandNearMiss +\n";
3829 OS << " (int)(bool)FeaturesNearMiss +\n";
3830 OS << " (int)(bool)EarlyPredicateNearMiss +\n";
3831 OS << " (int)(bool)LatePredicateNearMiss);\n";
3832 OS << " if (NumNearMisses == 1) {\n";
3833 OS << " // We had exactly one type of near-miss, so add that to the list.\n";
3834 OS << " assert(!OperandNearMiss && \"OperandNearMiss was handled earlier\");\n";
3835 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: found one type of \"\n";
3836 OS << " \"mismatch, so reporting a \"\n";
3837 OS << " \"near-miss\\n\");\n";
3838 OS << " if (NearMisses && FeaturesNearMiss)\n";
3839 OS << " NearMisses->push_back(FeaturesNearMiss);\n";
3840 OS << " else if (NearMisses && EarlyPredicateNearMiss)\n";
3841 OS << " NearMisses->push_back(EarlyPredicateNearMiss);\n";
3842 OS << " else if (NearMisses && LatePredicateNearMiss)\n";
3843 OS << " NearMisses->push_back(LatePredicateNearMiss);\n";
3844 OS << "\n";
3845 OS << " continue;\n";
3846 OS << " } else if (NumNearMisses > 1) {\n";
3847 OS << " // This instruction missed in more than one way, so ignore it.\n";
3848 OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: multiple \"\n";
3849 OS << " \"types of mismatch, so not \"\n";
3850 OS << " \"reporting near-miss\\n\");\n";
3851 OS << " continue;\n";
3852 OS << " }\n";
3855 // Call the post-processing function, if used.
3856 StringRef InsnCleanupFn = AsmParser->getValueAsString("AsmParserInstCleanup");
3857 if (!InsnCleanupFn.empty())
3858 OS << " " << InsnCleanupFn << "(Inst);\n";
3860 if (HasDeprecation) {
3861 OS << " std::string Info;\n";
3862 OS << " if (!getParser().getTargetParser().\n";
3863 OS << " getTargetOptions().MCNoDeprecatedWarn &&\n";
3864 OS << " MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n";
3865 OS << " SMLoc Loc = ((" << Target.getName()
3866 << "Operand&)*Operands[0]).getStartLoc();\n";
3867 OS << " getParser().Warning(Loc, Info, None);\n";
3868 OS << " }\n";
3871 if (!ReportMultipleNearMisses) {
3872 OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
3873 "Operands, ErrorInfo))\n";
3874 OS << " return Match_InvalidTiedOperand;\n";
3875 OS << "\n";
3878 OS << " DEBUG_WITH_TYPE(\n";
3879 OS << " \"asm-matcher\",\n";
3880 OS << " dbgs() << \"Opcode result: complete match, selecting this opcode\\n\");\n";
3881 OS << " return Match_Success;\n";
3882 OS << " }\n\n";
3884 if (ReportMultipleNearMisses) {
3885 OS << " // No instruction variants matched exactly.\n";
3886 OS << " return Match_NearMisses;\n";
3887 } else {
3888 OS << " // Okay, we had no match. Try to return a useful error code.\n";
3889 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
3890 OS << " return RetCode;\n\n";
3891 OS << " ErrorInfo = 0;\n";
3892 OS << " return Match_MissingFeature;\n";
3894 OS << "}\n\n";
3896 if (!Info.OperandMatchInfo.empty())
3897 emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
3898 MaxMnemonicIndex, FeatureBitsets.size(),
3899 HasMnemonicFirst);
3901 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
3903 OS << "\n#ifdef GET_MNEMONIC_SPELL_CHECKER\n";
3904 OS << "#undef GET_MNEMONIC_SPELL_CHECKER\n\n";
3906 emitMnemonicSpellChecker(OS, Target, VariantCount);
3908 OS << "#endif // GET_MNEMONIC_SPELL_CHECKER\n\n";
3911 namespace llvm {
3913 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
3914 emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
3915 AsmMatcherEmitter(RK).run(OS);
3918 } // end namespace llvm