1 //===---- ParseStmtAsm.cpp - Assembly Statement Parser --------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file implements parsing for GCC and Microsoft inline assembly.
11 //===----------------------------------------------------------------------===//
13 #include "clang/Parse/Parser.h"
14 #include "clang/AST/ASTContext.h"
15 #include "clang/Basic/Diagnostic.h"
16 #include "clang/Basic/TargetInfo.h"
17 #include "clang/Parse/RAIIObjectsForParser.h"
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/MC/MCAsmInfo.h"
21 #include "llvm/MC/MCContext.h"
22 #include "llvm/MC/MCInstPrinter.h"
23 #include "llvm/MC/MCInstrInfo.h"
24 #include "llvm/MC/MCObjectFileInfo.h"
25 #include "llvm/MC/MCParser/MCAsmParser.h"
26 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
27 #include "llvm/MC/MCRegisterInfo.h"
28 #include "llvm/MC/MCStreamer.h"
29 #include "llvm/MC/MCSubtargetInfo.h"
30 #include "llvm/MC/MCTargetOptions.h"
31 #include "llvm/Support/SourceMgr.h"
32 #include "llvm/Support/TargetRegistry.h"
33 #include "llvm/Support/TargetSelect.h"
34 using namespace clang;
37 class ClangAsmParserCallback : public llvm::MCAsmParserSemaCallback {
39 SourceLocation AsmLoc;
42 /// The tokens we streamed into AsmString and handed off to MC.
43 ArrayRef<Token> AsmToks;
45 /// The offset of each token in AsmToks within AsmString.
46 ArrayRef<unsigned> AsmTokOffsets;
49 ClangAsmParserCallback(Parser &P, SourceLocation Loc, StringRef AsmString,
50 ArrayRef<Token> Toks, ArrayRef<unsigned> Offsets)
51 : TheParser(P), AsmLoc(Loc), AsmString(AsmString), AsmToks(Toks),
52 AsmTokOffsets(Offsets) {
53 assert(AsmToks.size() == AsmTokOffsets.size());
56 void LookupInlineAsmIdentifier(StringRef &LineBuf,
57 llvm::InlineAsmIdentifierInfo &Info,
58 bool IsUnevaluatedContext) override;
60 StringRef LookupInlineAsmLabel(StringRef Identifier, llvm::SourceMgr &LSM,
62 bool Create) override;
64 bool LookupInlineAsmField(StringRef Base, StringRef Member,
65 unsigned &Offset) override {
66 return TheParser.getActions().LookupInlineAsmField(Base, Member, Offset,
70 static void DiagHandlerCallback(const llvm::SMDiagnostic &D, void *Context) {
71 ((ClangAsmParserCallback *)Context)->handleDiagnostic(D);
75 /// Collect the appropriate tokens for the given string.
76 void findTokensForString(StringRef Str, SmallVectorImpl<Token> &TempToks,
77 const Token *&FirstOrigToken) const;
79 SourceLocation translateLocation(const llvm::SourceMgr &LSM,
82 void handleDiagnostic(const llvm::SMDiagnostic &D);
86 void ClangAsmParserCallback::LookupInlineAsmIdentifier(
87 StringRef &LineBuf, llvm::InlineAsmIdentifierInfo &Info,
88 bool IsUnevaluatedContext) {
89 // Collect the desired tokens.
90 SmallVector<Token, 16> LineToks;
91 const Token *FirstOrigToken = nullptr;
92 findTokensForString(LineBuf, LineToks, FirstOrigToken);
94 unsigned NumConsumedToks;
95 ExprResult Result = TheParser.ParseMSAsmIdentifier(LineToks, NumConsumedToks,
96 IsUnevaluatedContext);
98 // If we consumed the entire line, tell MC that.
99 // Also do this if we consumed nothing as a way of reporting failure.
100 if (NumConsumedToks == 0 || NumConsumedToks == LineToks.size()) {
101 // By not modifying LineBuf, we're implicitly consuming it all.
103 // Otherwise, consume up to the original tokens.
105 assert(FirstOrigToken && "not using original tokens?");
107 // Since we're using original tokens, apply that offset.
108 assert(FirstOrigToken[NumConsumedToks].getLocation() ==
109 LineToks[NumConsumedToks].getLocation());
110 unsigned FirstIndex = FirstOrigToken - AsmToks.begin();
111 unsigned LastIndex = FirstIndex + NumConsumedToks - 1;
113 // The total length we've consumed is the relative offset
114 // of the last token we consumed plus its length.
115 unsigned TotalOffset =
116 (AsmTokOffsets[LastIndex] + AsmToks[LastIndex].getLength() -
117 AsmTokOffsets[FirstIndex]);
118 LineBuf = LineBuf.substr(0, TotalOffset);
121 // Initialize Info with the lookup result.
122 if (!Result.isUsable())
124 TheParser.getActions().FillInlineAsmIdentifierInfo(Result.get(), Info);
127 StringRef ClangAsmParserCallback::LookupInlineAsmLabel(StringRef Identifier,
128 llvm::SourceMgr &LSM,
129 llvm::SMLoc Location,
131 SourceLocation Loc = translateLocation(LSM, Location);
133 TheParser.getActions().GetOrCreateMSAsmLabel(Identifier, Loc, Create);
134 return Label->getMSAsmLabel();
137 void ClangAsmParserCallback::findTokensForString(
138 StringRef Str, SmallVectorImpl<Token> &TempToks,
139 const Token *&FirstOrigToken) const {
140 // For now, assert that the string we're working with is a substring
141 // of what we gave to MC. This lets us use the original tokens.
142 assert(!std::less<const char *>()(Str.begin(), AsmString.begin()) &&
143 !std::less<const char *>()(AsmString.end(), Str.end()));
145 // Try to find a token whose offset matches the first token.
146 unsigned FirstCharOffset = Str.begin() - AsmString.begin();
147 const unsigned *FirstTokOffset =
148 llvm::lower_bound(AsmTokOffsets, FirstCharOffset);
150 // For now, assert that the start of the string exactly
151 // corresponds to the start of a token.
152 assert(*FirstTokOffset == FirstCharOffset);
154 // Use all the original tokens for this line. (We assume the
155 // end of the line corresponds cleanly to a token break.)
156 unsigned FirstTokIndex = FirstTokOffset - AsmTokOffsets.begin();
157 FirstOrigToken = &AsmToks[FirstTokIndex];
158 unsigned LastCharOffset = Str.end() - AsmString.begin();
159 for (unsigned i = FirstTokIndex, e = AsmTokOffsets.size(); i != e; ++i) {
160 if (AsmTokOffsets[i] >= LastCharOffset)
162 TempToks.push_back(AsmToks[i]);
167 ClangAsmParserCallback::translateLocation(const llvm::SourceMgr &LSM,
169 // Compute an offset into the inline asm buffer.
170 // FIXME: This isn't right if .macro is involved (but hopefully, no
171 // real-world code does that).
172 const llvm::MemoryBuffer *LBuf =
173 LSM.getMemoryBuffer(LSM.FindBufferContainingLoc(SMLoc));
174 unsigned Offset = SMLoc.getPointer() - LBuf->getBufferStart();
176 // Figure out which token that offset points into.
177 const unsigned *TokOffsetPtr = llvm::lower_bound(AsmTokOffsets, Offset);
178 unsigned TokIndex = TokOffsetPtr - AsmTokOffsets.begin();
179 unsigned TokOffset = *TokOffsetPtr;
181 // If we come up with an answer which seems sane, use it; otherwise,
182 // just point at the __asm keyword.
183 // FIXME: Assert the answer is sane once we handle .macro correctly.
184 SourceLocation Loc = AsmLoc;
185 if (TokIndex < AsmToks.size()) {
186 const Token &Tok = AsmToks[TokIndex];
187 Loc = Tok.getLocation();
188 Loc = Loc.getLocWithOffset(Offset - TokOffset);
193 void ClangAsmParserCallback::handleDiagnostic(const llvm::SMDiagnostic &D) {
194 const llvm::SourceMgr &LSM = *D.getSourceMgr();
195 SourceLocation Loc = translateLocation(LSM, D.getLoc());
196 TheParser.Diag(Loc, diag::err_inline_ms_asm_parsing) << D.getMessage();
199 /// Parse an identifier in an MS-style inline assembly block.
200 ExprResult Parser::ParseMSAsmIdentifier(llvm::SmallVectorImpl<Token> &LineToks,
201 unsigned &NumLineToksConsumed,
202 bool IsUnevaluatedContext) {
203 // Push a fake token on the end so that we don't overrun the token
204 // stream. We use ';' because it expression-parsing should never
206 const tok::TokenKind EndOfStream = tok::semi;
207 Token EndOfStreamTok;
208 EndOfStreamTok.startToken();
209 EndOfStreamTok.setKind(EndOfStream);
210 LineToks.push_back(EndOfStreamTok);
212 // Also copy the current token over.
213 LineToks.push_back(Tok);
215 PP.EnterTokenStream(LineToks, /*DisableMacroExpansions*/ true,
216 /*IsReinject*/ true);
218 // Clear the current token and advance to the first token in LineToks.
221 // Parse an optional scope-specifier if we're in C++.
223 if (getLangOpts().CPlusPlus) {
224 ParseOptionalCXXScopeSpecifier(SS, nullptr, /*EnteringContext=*/false);
227 // Require an identifier here.
228 SourceLocation TemplateKWLoc;
232 if (Tok.is(tok::kw_this)) {
233 Result = ParseCXXThis();
236 Invalid = ParseUnqualifiedId(SS,
237 /*EnteringContext=*/false,
238 /*AllowDestructorName=*/false,
239 /*AllowConstructorName=*/false,
240 /*AllowDeductionGuide=*/false,
241 /*ObjectType=*/nullptr, &TemplateKWLoc, Id);
242 // Perform the lookup.
243 Result = Actions.LookupInlineAsmIdentifier(SS, TemplateKWLoc, Id,
244 IsUnevaluatedContext);
246 // While the next two tokens are 'period' 'identifier', repeatedly parse it as
247 // a field access. We have to avoid consuming assembler directives that look
249 while (Result.isUsable() && Tok.is(tok::period)) {
250 Token IdTok = PP.LookAhead(0);
251 if (IdTok.isNot(tok::identifier))
253 ConsumeToken(); // Consume the period.
254 IdentifierInfo *Id = Tok.getIdentifierInfo();
255 ConsumeToken(); // Consume the identifier.
256 Result = Actions.LookupInlineAsmVarDeclField(Result.get(), Id->getName(),
260 // Figure out how many tokens we are into LineToks.
261 unsigned LineIndex = 0;
262 if (Tok.is(EndOfStream)) {
263 LineIndex = LineToks.size() - 2;
265 while (LineToks[LineIndex].getLocation() != Tok.getLocation()) {
267 assert(LineIndex < LineToks.size() - 2); // we added two extra tokens
271 // If we've run into the poison token we inserted before, or there
272 // was a parsing error, then claim the entire line.
273 if (Invalid || Tok.is(EndOfStream)) {
274 NumLineToksConsumed = LineToks.size() - 2;
276 // Otherwise, claim up to the start of the next token.
277 NumLineToksConsumed = LineIndex;
280 // Finally, restore the old parsing state by consuming all the tokens we
281 // staged before, implicitly killing off the token-lexer we pushed.
282 for (unsigned i = 0, e = LineToks.size() - LineIndex - 2; i != e; ++i) {
285 assert(Tok.is(EndOfStream));
288 // Leave LineToks in its original state.
295 /// Turn a sequence of our tokens back into a string that we can hand
296 /// to the MC asm parser.
297 static bool buildMSAsmString(Preprocessor &PP, SourceLocation AsmLoc,
298 ArrayRef<Token> AsmToks,
299 SmallVectorImpl<unsigned> &TokOffsets,
300 SmallString<512> &Asm) {
301 assert(!AsmToks.empty() && "Didn't expect an empty AsmToks!");
303 // Is this the start of a new assembly statement?
304 bool isNewStatement = true;
306 for (unsigned i = 0, e = AsmToks.size(); i < e; ++i) {
307 const Token &Tok = AsmToks[i];
309 // Start each new statement with a newline and a tab.
310 if (!isNewStatement && (Tok.is(tok::kw_asm) || Tok.isAtStartOfLine())) {
312 isNewStatement = true;
315 // Preserve the existence of leading whitespace except at the
316 // start of a statement.
317 if (!isNewStatement && Tok.hasLeadingSpace())
320 // Remember the offset of this token.
321 TokOffsets.push_back(Asm.size());
323 // Don't actually write '__asm' into the assembly stream.
324 if (Tok.is(tok::kw_asm)) {
325 // Complain about __asm at the end of the stream.
327 PP.Diag(AsmLoc, diag::err_asm_empty);
334 // Append the spelling of the token.
335 SmallString<32> SpellingBuffer;
336 bool SpellingInvalid = false;
337 Asm += PP.getSpelling(Tok, SpellingBuffer, &SpellingInvalid);
338 assert(!SpellingInvalid && "spelling was invalid after correct parse?");
340 // We are no longer at the start of a statement.
341 isNewStatement = false;
344 // Ensure that the buffer is null-terminated.
348 assert(TokOffsets.size() == AsmToks.size());
352 /// isTypeQualifier - Return true if the current token could be the
353 /// start of a type-qualifier-list.
354 static bool isTypeQualifier(const Token &Tok) {
355 switch (Tok.getKind()) {
356 default: return false;
359 case tok::kw_volatile:
360 case tok::kw_restrict:
361 case tok::kw___private:
362 case tok::kw___local:
363 case tok::kw___global:
364 case tok::kw___constant:
365 case tok::kw___generic:
366 case tok::kw___read_only:
367 case tok::kw___read_write:
368 case tok::kw___write_only:
373 // Determine if this is a GCC-style asm statement.
374 static bool isGCCAsmStatement(const Token &TokAfterAsm) {
375 return TokAfterAsm.is(tok::l_paren) || TokAfterAsm.is(tok::kw_goto) ||
376 isTypeQualifier(TokAfterAsm);
379 /// ParseMicrosoftAsmStatement. When -fms-extensions/-fasm-blocks is enabled,
380 /// this routine is called to collect the tokens for an MS asm statement.
382 /// [MS] ms-asm-statement:
384 /// ms-asm-block ms-asm-statement
386 /// [MS] ms-asm-block:
387 /// '__asm' ms-asm-line '\n'
388 /// '__asm' '{' ms-asm-instruction-block[opt] '}' ';'[opt]
390 /// [MS] ms-asm-instruction-block
392 /// ms-asm-line '\n' ms-asm-instruction-block
394 StmtResult Parser::ParseMicrosoftAsmStatement(SourceLocation AsmLoc) {
395 SourceManager &SrcMgr = PP.getSourceManager();
396 SourceLocation EndLoc = AsmLoc;
397 SmallVector<Token, 4> AsmToks;
399 bool SingleLineMode = true;
400 unsigned BraceNesting = 0;
401 unsigned short savedBraceCount = BraceCount;
402 bool InAsmComment = false;
405 unsigned NumTokensRead = 0;
406 SmallVector<SourceLocation, 4> LBraceLocs;
407 bool SkippedStartOfLine = false;
409 if (Tok.is(tok::l_brace)) {
410 // Braced inline asm: consume the opening brace.
411 SingleLineMode = false;
413 EndLoc = ConsumeBrace();
414 LBraceLocs.push_back(EndLoc);
417 // Single-line inline asm; compute which line it is on.
418 std::pair<FileID, unsigned> ExpAsmLoc =
419 SrcMgr.getDecomposedExpansionLoc(EndLoc);
420 FID = ExpAsmLoc.first;
421 LineNo = SrcMgr.getLineNumber(FID, ExpAsmLoc.second);
422 LBraceLocs.push_back(SourceLocation());
425 SourceLocation TokLoc = Tok.getLocation();
427 // If we hit EOF, we're done, period.
431 if (!InAsmComment && Tok.is(tok::l_brace)) {
432 // Consume the opening brace.
433 SkippedStartOfLine = Tok.isAtStartOfLine();
434 AsmToks.push_back(Tok);
435 EndLoc = ConsumeBrace();
437 LBraceLocs.push_back(EndLoc);
438 TokLoc = Tok.getLocation();
441 } else if (!InAsmComment && Tok.is(tok::semi)) {
442 // A semicolon in an asm is the start of a comment.
444 if (!SingleLineMode) {
445 // Compute which line the comment is on.
446 std::pair<FileID, unsigned> ExpSemiLoc =
447 SrcMgr.getDecomposedExpansionLoc(TokLoc);
448 FID = ExpSemiLoc.first;
449 LineNo = SrcMgr.getLineNumber(FID, ExpSemiLoc.second);
451 } else if (SingleLineMode || InAsmComment) {
452 // If end-of-line is significant, check whether this token is on a
454 std::pair<FileID, unsigned> ExpLoc =
455 SrcMgr.getDecomposedExpansionLoc(TokLoc);
456 if (ExpLoc.first != FID ||
457 SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second) != LineNo) {
458 // If this is a single-line __asm, we're done, except if the next
459 // line is MS-style asm too, in which case we finish a comment
460 // if needed and then keep processing the next line as a single
462 bool isAsm = Tok.is(tok::kw_asm);
463 if (SingleLineMode && (!isAsm || isGCCAsmStatement(NextToken())))
465 // We're no longer in a comment.
466 InAsmComment = false;
468 // If this is a new __asm {} block we want to process it separately
469 // from the single-line __asm statements
470 if (PP.LookAhead(0).is(tok::l_brace))
472 LineNo = SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second);
473 SkippedStartOfLine = Tok.isAtStartOfLine();
474 } else if (Tok.is(tok::semi)) {
475 // A multi-line asm-statement, where next line is a comment
478 LineNo = SrcMgr.getLineNumber(FID, ExpLoc.second);
480 } else if (!InAsmComment && Tok.is(tok::r_brace)) {
481 // In MSVC mode, braces only participate in brace matching and
482 // separating the asm statements. This is an intentional
483 // departure from the Apple gcc behavior.
488 if (!InAsmComment && BraceNesting && Tok.is(tok::r_brace) &&
489 BraceCount == (savedBraceCount + BraceNesting)) {
490 // Consume the closing brace.
491 SkippedStartOfLine = Tok.isAtStartOfLine();
492 // Don't want to add the closing brace of the whole asm block
493 if (SingleLineMode || BraceNesting > 1) {
494 Tok.clearFlag(Token::LeadingSpace);
495 AsmToks.push_back(Tok);
497 EndLoc = ConsumeBrace();
499 // Finish if all of the opened braces in the inline asm section were
501 if (BraceNesting == 0 && !SingleLineMode)
504 LBraceLocs.pop_back();
505 TokLoc = Tok.getLocation();
511 // Consume the next token; make sure we don't modify the brace count etc.
512 // if we are in a comment.
517 // Set the token as the start of line if we skipped the original start
518 // of line token in case it was a nested brace.
519 if (SkippedStartOfLine)
520 Tok.setFlag(Token::StartOfLine);
521 AsmToks.push_back(Tok);
524 TokLoc = Tok.getLocation();
526 SkippedStartOfLine = false;
529 if (BraceNesting && BraceCount != savedBraceCount) {
530 // __asm without closing brace (this can happen at EOF).
531 for (unsigned i = 0; i < BraceNesting; ++i) {
532 Diag(Tok, diag::err_expected) << tok::r_brace;
533 Diag(LBraceLocs.back(), diag::note_matching) << tok::l_brace;
534 LBraceLocs.pop_back();
537 } else if (NumTokensRead == 0) {
539 Diag(Tok, diag::err_expected) << tok::l_brace;
543 // Okay, prepare to use MC to parse the assembly.
544 SmallVector<StringRef, 4> ConstraintRefs;
545 SmallVector<Expr *, 4> Exprs;
546 SmallVector<StringRef, 4> ClobberRefs;
548 // We need an actual supported target.
549 const llvm::Triple &TheTriple = Actions.Context.getTargetInfo().getTriple();
550 const std::string &TT = TheTriple.getTriple();
551 const llvm::Target *TheTarget = nullptr;
552 if (!TheTriple.isX86()) {
553 Diag(AsmLoc, diag::err_msasm_unsupported_arch) << TheTriple.getArchName();
556 TheTarget = llvm::TargetRegistry::lookupTarget(TT, Error);
558 Diag(AsmLoc, diag::err_msasm_unable_to_create_target) << Error;
561 assert(!LBraceLocs.empty() && "Should have at least one location here");
563 SmallString<512> AsmString;
564 auto EmptyStmt = [&] {
565 return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmString,
566 /*NumOutputs*/ 0, /*NumInputs*/ 0,
567 ConstraintRefs, ClobberRefs, Exprs, EndLoc);
569 // If we don't support assembly, or the assembly is empty, we don't
570 // need to instantiate the AsmParser, etc.
571 if (!TheTarget || AsmToks.empty()) {
575 // Expand the tokens into a string buffer.
576 SmallVector<unsigned, 8> TokOffsets;
577 if (buildMSAsmString(PP, AsmLoc, AsmToks, TokOffsets, AsmString))
580 const TargetOptions &TO = Actions.Context.getTargetInfo().getTargetOpts();
581 std::string FeaturesStr =
582 llvm::join(TO.Features.begin(), TO.Features.end(), ",");
584 std::unique_ptr<llvm::MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TT));
586 Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
587 << "target MC unavailable";
590 // FIXME: init MCOptions from sanitizer flags here.
591 llvm::MCTargetOptions MCOptions;
592 std::unique_ptr<llvm::MCAsmInfo> MAI(
593 TheTarget->createMCAsmInfo(*MRI, TT, MCOptions));
594 // Get the instruction descriptor.
595 std::unique_ptr<llvm::MCInstrInfo> MII(TheTarget->createMCInstrInfo());
596 std::unique_ptr<llvm::MCObjectFileInfo> MOFI(new llvm::MCObjectFileInfo());
597 std::unique_ptr<llvm::MCSubtargetInfo> STI(
598 TheTarget->createMCSubtargetInfo(TT, TO.CPU, FeaturesStr));
599 // Target MCTargetDesc may not be linked in clang-based tools.
600 if (!MAI || !MII | !MOFI || !STI) {
601 Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
602 << "target MC unavailable";
606 llvm::SourceMgr TempSrcMgr;
607 llvm::MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &TempSrcMgr);
608 MOFI->InitMCObjectFileInfo(TheTriple, /*PIC*/ false, Ctx);
609 std::unique_ptr<llvm::MemoryBuffer> Buffer =
610 llvm::MemoryBuffer::getMemBuffer(AsmString, "<MS inline asm>");
612 // Tell SrcMgr about this buffer, which is what the parser will pick up.
613 TempSrcMgr.AddNewSourceBuffer(std::move(Buffer), llvm::SMLoc());
615 std::unique_ptr<llvm::MCStreamer> Str(createNullStreamer(Ctx));
616 std::unique_ptr<llvm::MCAsmParser> Parser(
617 createMCAsmParser(TempSrcMgr, Ctx, *Str.get(), *MAI));
619 std::unique_ptr<llvm::MCTargetAsmParser> TargetParser(
620 TheTarget->createMCAsmParser(*STI, *Parser, *MII, MCOptions));
621 // Target AsmParser may not be linked in clang-based tools.
623 Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
624 << "target ASM parser unavailable";
628 std::unique_ptr<llvm::MCInstPrinter> IP(
629 TheTarget->createMCInstPrinter(llvm::Triple(TT), 1, *MAI, *MII, *MRI));
631 // Change to the Intel dialect.
632 Parser->setAssemblerDialect(1);
633 Parser->setTargetParser(*TargetParser.get());
634 Parser->setParsingInlineAsm(true);
635 TargetParser->setParsingInlineAsm(true);
637 ClangAsmParserCallback Callback(*this, AsmLoc, AsmString, AsmToks,
639 TargetParser->setSemaCallback(&Callback);
640 TempSrcMgr.setDiagHandler(ClangAsmParserCallback::DiagHandlerCallback,
645 std::string AsmStringIR;
646 SmallVector<std::pair<void *, bool>, 4> OpExprs;
647 SmallVector<std::string, 4> Constraints;
648 SmallVector<std::string, 4> Clobbers;
649 if (Parser->parseMSInlineAsm(AsmLoc.getPtrEncoding(), AsmStringIR, NumOutputs,
650 NumInputs, OpExprs, Constraints, Clobbers,
651 MII.get(), IP.get(), Callback))
654 // Filter out "fpsw" and "mxcsr". They aren't valid GCC asm clobber
655 // constraints. Clang always adds fpsr to the clobber list anyway.
656 llvm::erase_if(Clobbers, [](const std::string &C) {
657 return C == "fpsr" || C == "mxcsr";
660 // Build the vector of clobber StringRefs.
661 ClobberRefs.insert(ClobberRefs.end(), Clobbers.begin(), Clobbers.end());
663 // Recast the void pointers and build the vector of constraint StringRefs.
664 unsigned NumExprs = NumOutputs + NumInputs;
665 ConstraintRefs.resize(NumExprs);
666 Exprs.resize(NumExprs);
667 for (unsigned i = 0, e = NumExprs; i != e; ++i) {
668 Expr *OpExpr = static_cast<Expr *>(OpExprs[i].first);
672 // Need address of variable.
673 if (OpExprs[i].second)
675 Actions.BuildUnaryOp(getCurScope(), AsmLoc, UO_AddrOf, OpExpr).get();
677 ConstraintRefs[i] = StringRef(Constraints[i]);
681 // FIXME: We should be passing source locations for better diagnostics.
682 return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmStringIR,
683 NumOutputs, NumInputs, ConstraintRefs,
684 ClobberRefs, Exprs, EndLoc);
687 /// ParseAsmStatement - Parse a GNU extended asm statement.
689 /// gnu-asm-statement
692 /// [GNU] gnu-asm-statement:
693 /// 'asm' type-qualifier[opt] '(' asm-argument ')' ';'
695 /// [GNU] asm-argument:
696 /// asm-string-literal
697 /// asm-string-literal ':' asm-operands[opt]
698 /// asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
699 /// asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
702 /// [GNU] asm-clobbers:
703 /// asm-string-literal
704 /// asm-clobbers ',' asm-string-literal
706 StmtResult Parser::ParseAsmStatement(bool &msAsm) {
707 assert(Tok.is(tok::kw_asm) && "Not an asm stmt");
708 SourceLocation AsmLoc = ConsumeToken();
710 if (getLangOpts().AsmBlocks && !isGCCAsmStatement(Tok)) {
712 return ParseMicrosoftAsmStatement(AsmLoc);
715 DeclSpec DS(AttrFactory);
716 SourceLocation Loc = Tok.getLocation();
717 ParseTypeQualifierListOpt(DS, AR_VendorAttributesParsed);
719 // GNU asms accept, but warn, about type-qualifiers other than volatile.
720 if (DS.getTypeQualifiers() & DeclSpec::TQ_const)
721 Diag(Loc, diag::warn_asm_qualifier_ignored) << "const";
722 if (DS.getTypeQualifiers() & DeclSpec::TQ_restrict)
723 Diag(Loc, diag::warn_asm_qualifier_ignored) << "restrict";
724 // FIXME: Once GCC supports _Atomic, check whether it permits it here.
725 if (DS.getTypeQualifiers() & DeclSpec::TQ_atomic)
726 Diag(Loc, diag::warn_asm_qualifier_ignored) << "_Atomic";
728 // Remember if this was a volatile asm.
729 bool isVolatile = DS.getTypeQualifiers() & DeclSpec::TQ_volatile;
730 // Remember if this was a goto asm.
731 bool isGotoAsm = false;
733 if (Tok.is(tok::kw_goto)) {
738 if (Tok.isNot(tok::l_paren)) {
739 Diag(Tok, diag::err_expected_lparen_after) << "asm";
740 SkipUntil(tok::r_paren, StopAtSemi);
743 BalancedDelimiterTracker T(*this, tok::l_paren);
746 ExprResult AsmString(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
748 // Check if GNU-style InlineAsm is disabled.
749 // Error on anything other than empty string.
750 if (!(getLangOpts().GNUAsm || AsmString.isInvalid())) {
751 const auto *SL = cast<StringLiteral>(AsmString.get());
752 if (!SL->getString().trim().empty())
753 Diag(Loc, diag::err_gnu_inline_asm_disabled);
756 if (AsmString.isInvalid()) {
757 // Consume up to and including the closing paren.
762 SmallVector<IdentifierInfo *, 4> Names;
763 ExprVector Constraints;
767 if (Tok.is(tok::r_paren)) {
768 // We have a simple asm expression like 'asm("foo")'.
770 return Actions.ActOnGCCAsmStmt(AsmLoc, /*isSimple*/ true, isVolatile,
771 /*NumOutputs*/ 0, /*NumInputs*/ 0, nullptr,
772 Constraints, Exprs, AsmString.get(),
773 Clobbers, /*NumLabels*/ 0,
774 T.getCloseLocation());
777 // Parse Outputs, if present.
778 bool AteExtraColon = false;
779 if (Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
780 // In C++ mode, parse "::" like ": :".
781 AteExtraColon = Tok.is(tok::coloncolon);
784 if (!AteExtraColon && isGotoAsm && Tok.isNot(tok::colon)) {
785 Diag(Tok, diag::err_asm_goto_cannot_have_output);
786 SkipUntil(tok::r_paren, StopAtSemi);
790 if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
794 unsigned NumOutputs = Names.size();
796 // Parse Inputs, if present.
797 if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
798 // In C++ mode, parse "::" like ": :".
800 AteExtraColon = false;
802 AteExtraColon = Tok.is(tok::coloncolon);
806 if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
810 assert(Names.size() == Constraints.size() &&
811 Constraints.size() == Exprs.size() && "Input operand size mismatch!");
813 unsigned NumInputs = Names.size() - NumOutputs;
815 // Parse the clobbers, if present.
816 if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
818 AteExtraColon = false;
820 AteExtraColon = Tok.is(tok::coloncolon);
823 // Parse the asm-string list for clobbers if present.
824 if (!AteExtraColon && isTokenStringLiteral()) {
826 ExprResult Clobber(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
828 if (Clobber.isInvalid())
831 Clobbers.push_back(Clobber.get());
833 if (!TryConsumeToken(tok::comma))
838 if (!isGotoAsm && (Tok.isNot(tok::r_paren) || AteExtraColon)) {
839 Diag(Tok, diag::err_expected) << tok::r_paren;
840 SkipUntil(tok::r_paren, StopAtSemi);
844 // Parse the goto label, if present.
845 unsigned NumLabels = 0;
846 if (AteExtraColon || Tok.is(tok::colon)) {
851 if (Tok.isNot(tok::identifier)) {
852 Diag(Tok, diag::err_expected) << tok::identifier;
853 SkipUntil(tok::r_paren, StopAtSemi);
856 LabelDecl *LD = Actions.LookupOrCreateLabel(Tok.getIdentifierInfo(),
858 Names.push_back(Tok.getIdentifierInfo());
860 SkipUntil(tok::r_paren, StopAtSemi);
864 Actions.ActOnAddrLabel(Tok.getLocation(), Tok.getLocation(), LD);
865 Exprs.push_back(Res.get());
868 if (!TryConsumeToken(tok::comma))
871 } else if (isGotoAsm) {
872 Diag(Tok, diag::err_expected) << tok::colon;
873 SkipUntil(tok::r_paren, StopAtSemi);
877 return Actions.ActOnGCCAsmStmt(
878 AsmLoc, false, isVolatile, NumOutputs, NumInputs, Names.data(),
879 Constraints, Exprs, AsmString.get(), Clobbers, NumLabels,
880 T.getCloseLocation());
883 /// ParseAsmOperands - Parse the asm-operands production as used by
884 /// asm-statement, assuming the leading ':' token was eaten.
886 /// [GNU] asm-operands:
888 /// asm-operands ',' asm-operand
890 /// [GNU] asm-operand:
891 /// asm-string-literal '(' expression ')'
892 /// '[' identifier ']' asm-string-literal '(' expression ')'
895 // FIXME: Avoid unnecessary std::string trashing.
896 bool Parser::ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo *> &Names,
897 SmallVectorImpl<Expr *> &Constraints,
898 SmallVectorImpl<Expr *> &Exprs) {
899 // 'asm-operands' isn't present?
900 if (!isTokenStringLiteral() && Tok.isNot(tok::l_square))
904 // Read the [id] if present.
905 if (Tok.is(tok::l_square)) {
906 BalancedDelimiterTracker T(*this, tok::l_square);
909 if (Tok.isNot(tok::identifier)) {
910 Diag(Tok, diag::err_expected) << tok::identifier;
911 SkipUntil(tok::r_paren, StopAtSemi);
915 IdentifierInfo *II = Tok.getIdentifierInfo();
921 Names.push_back(nullptr);
923 ExprResult Constraint(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
924 if (Constraint.isInvalid()) {
925 SkipUntil(tok::r_paren, StopAtSemi);
928 Constraints.push_back(Constraint.get());
930 if (Tok.isNot(tok::l_paren)) {
931 Diag(Tok, diag::err_expected_lparen_after) << "asm operand";
932 SkipUntil(tok::r_paren, StopAtSemi);
936 // Read the parenthesized expression.
937 BalancedDelimiterTracker T(*this, tok::l_paren);
939 ExprResult Res = Actions.CorrectDelayedTyposInExpr(ParseExpression());
941 if (Res.isInvalid()) {
942 SkipUntil(tok::r_paren, StopAtSemi);
945 Exprs.push_back(Res.get());
946 // Eat the comma and continue parsing if it exists.
947 if (!TryConsumeToken(tok::comma))