• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===---- ParseStmtAsm.cpp - Assembly Statement Parser --------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements parsing for GCC and Microsoft inline assembly.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "clang/Parse/Parser.h"
15 #include "RAIIObjectsForParser.h"
16 #include "clang/AST/ASTContext.h"
17 #include "clang/Basic/Diagnostic.h"
18 #include "clang/Basic/TargetInfo.h"
19 #include "llvm/ADT/SmallString.h"
20 #include "llvm/ADT/StringExtras.h"
21 #include "llvm/MC/MCAsmInfo.h"
22 #include "llvm/MC/MCContext.h"
23 #include "llvm/MC/MCInstPrinter.h"
24 #include "llvm/MC/MCInstrInfo.h"
25 #include "llvm/MC/MCObjectFileInfo.h"
26 #include "llvm/MC/MCParser/MCAsmParser.h"
27 #include "llvm/MC/MCParser/MCTargetAsmParser.h"
28 #include "llvm/MC/MCRegisterInfo.h"
29 #include "llvm/MC/MCStreamer.h"
30 #include "llvm/MC/MCSubtargetInfo.h"
31 #include "llvm/MC/MCTargetOptions.h"
32 #include "llvm/Support/SourceMgr.h"
33 #include "llvm/Support/TargetRegistry.h"
34 #include "llvm/Support/TargetSelect.h"
35 using namespace clang;
36 
37 namespace {
38 class ClangAsmParserCallback : public llvm::MCAsmParserSemaCallback {
39   Parser &TheParser;
40   SourceLocation AsmLoc;
41   StringRef AsmString;
42 
43   /// The tokens we streamed into AsmString and handed off to MC.
44   ArrayRef<Token> AsmToks;
45 
46   /// The offset of each token in AsmToks within AsmString.
47   ArrayRef<unsigned> AsmTokOffsets;
48 
49 public:
ClangAsmParserCallback(Parser & P,SourceLocation Loc,StringRef AsmString,ArrayRef<Token> Toks,ArrayRef<unsigned> Offsets)50   ClangAsmParserCallback(Parser &P, SourceLocation Loc, StringRef AsmString,
51                          ArrayRef<Token> Toks, ArrayRef<unsigned> Offsets)
52       : TheParser(P), AsmLoc(Loc), AsmString(AsmString), AsmToks(Toks),
53         AsmTokOffsets(Offsets) {
54     assert(AsmToks.size() == AsmTokOffsets.size());
55   }
56 
LookupInlineAsmIdentifier(StringRef & LineBuf,llvm::InlineAsmIdentifierInfo & Info,bool IsUnevaluatedContext)57   void *LookupInlineAsmIdentifier(StringRef &LineBuf,
58                                   llvm::InlineAsmIdentifierInfo &Info,
59                                   bool IsUnevaluatedContext) override {
60     // Collect the desired tokens.
61     SmallVector<Token, 16> LineToks;
62     const Token *FirstOrigToken = nullptr;
63     findTokensForString(LineBuf, LineToks, FirstOrigToken);
64 
65     unsigned NumConsumedToks;
66     ExprResult Result = TheParser.ParseMSAsmIdentifier(
67         LineToks, NumConsumedToks, &Info, IsUnevaluatedContext);
68 
69     // If we consumed the entire line, tell MC that.
70     // Also do this if we consumed nothing as a way of reporting failure.
71     if (NumConsumedToks == 0 || NumConsumedToks == LineToks.size()) {
72       // By not modifying LineBuf, we're implicitly consuming it all.
73 
74       // Otherwise, consume up to the original tokens.
75     } else {
76       assert(FirstOrigToken && "not using original tokens?");
77 
78       // Since we're using original tokens, apply that offset.
79       assert(FirstOrigToken[NumConsumedToks].getLocation() ==
80              LineToks[NumConsumedToks].getLocation());
81       unsigned FirstIndex = FirstOrigToken - AsmToks.begin();
82       unsigned LastIndex = FirstIndex + NumConsumedToks - 1;
83 
84       // The total length we've consumed is the relative offset
85       // of the last token we consumed plus its length.
86       unsigned TotalOffset =
87           (AsmTokOffsets[LastIndex] + AsmToks[LastIndex].getLength() -
88            AsmTokOffsets[FirstIndex]);
89       LineBuf = LineBuf.substr(0, TotalOffset);
90     }
91 
92     // Initialize the "decl" with the lookup result.
93     Info.OpDecl = static_cast<void *>(Result.get());
94     return Info.OpDecl;
95   }
96 
LookupInlineAsmLabel(StringRef Identifier,llvm::SourceMgr & LSM,llvm::SMLoc Location,bool Create)97   StringRef LookupInlineAsmLabel(StringRef Identifier, llvm::SourceMgr &LSM,
98                                  llvm::SMLoc Location,
99                                  bool Create) override {
100     SourceLocation Loc = translateLocation(LSM, Location);
101     LabelDecl *Label =
102       TheParser.getActions().GetOrCreateMSAsmLabel(Identifier, Loc, Create);
103     return Label->getMSAsmLabel();
104   }
105 
LookupInlineAsmField(StringRef Base,StringRef Member,unsigned & Offset)106   bool LookupInlineAsmField(StringRef Base, StringRef Member,
107                             unsigned &Offset) override {
108     return TheParser.getActions().LookupInlineAsmField(Base, Member, Offset,
109                                                        AsmLoc);
110   }
111 
DiagHandlerCallback(const llvm::SMDiagnostic & D,void * Context)112   static void DiagHandlerCallback(const llvm::SMDiagnostic &D, void *Context) {
113     ((ClangAsmParserCallback *)Context)->handleDiagnostic(D);
114   }
115 
116 private:
117   /// Collect the appropriate tokens for the given string.
findTokensForString(StringRef Str,SmallVectorImpl<Token> & TempToks,const Token * & FirstOrigToken) const118   void findTokensForString(StringRef Str, SmallVectorImpl<Token> &TempToks,
119                            const Token *&FirstOrigToken) const {
120     // For now, assert that the string we're working with is a substring
121     // of what we gave to MC.  This lets us use the original tokens.
122     assert(!std::less<const char *>()(Str.begin(), AsmString.begin()) &&
123            !std::less<const char *>()(AsmString.end(), Str.end()));
124 
125     // Try to find a token whose offset matches the first token.
126     unsigned FirstCharOffset = Str.begin() - AsmString.begin();
127     const unsigned *FirstTokOffset = std::lower_bound(
128         AsmTokOffsets.begin(), AsmTokOffsets.end(), FirstCharOffset);
129 
130     // For now, assert that the start of the string exactly
131     // corresponds to the start of a token.
132     assert(*FirstTokOffset == FirstCharOffset);
133 
134     // Use all the original tokens for this line.  (We assume the
135     // end of the line corresponds cleanly to a token break.)
136     unsigned FirstTokIndex = FirstTokOffset - AsmTokOffsets.begin();
137     FirstOrigToken = &AsmToks[FirstTokIndex];
138     unsigned LastCharOffset = Str.end() - AsmString.begin();
139     for (unsigned i = FirstTokIndex, e = AsmTokOffsets.size(); i != e; ++i) {
140       if (AsmTokOffsets[i] >= LastCharOffset)
141         break;
142       TempToks.push_back(AsmToks[i]);
143     }
144   }
145 
translateLocation(const llvm::SourceMgr & LSM,llvm::SMLoc SMLoc)146   SourceLocation translateLocation(const llvm::SourceMgr &LSM, llvm::SMLoc SMLoc) {
147     // Compute an offset into the inline asm buffer.
148     // FIXME: This isn't right if .macro is involved (but hopefully, no
149     // real-world code does that).
150     const llvm::MemoryBuffer *LBuf =
151         LSM.getMemoryBuffer(LSM.FindBufferContainingLoc(SMLoc));
152     unsigned Offset = SMLoc.getPointer() - LBuf->getBufferStart();
153 
154     // Figure out which token that offset points into.
155     const unsigned *TokOffsetPtr =
156         std::lower_bound(AsmTokOffsets.begin(), AsmTokOffsets.end(), Offset);
157     unsigned TokIndex = TokOffsetPtr - AsmTokOffsets.begin();
158     unsigned TokOffset = *TokOffsetPtr;
159 
160     // If we come up with an answer which seems sane, use it; otherwise,
161     // just point at the __asm keyword.
162     // FIXME: Assert the answer is sane once we handle .macro correctly.
163     SourceLocation Loc = AsmLoc;
164     if (TokIndex < AsmToks.size()) {
165       const Token &Tok = AsmToks[TokIndex];
166       Loc = Tok.getLocation();
167       Loc = Loc.getLocWithOffset(Offset - TokOffset);
168     }
169     return Loc;
170   }
171 
handleDiagnostic(const llvm::SMDiagnostic & D)172   void handleDiagnostic(const llvm::SMDiagnostic &D) {
173     const llvm::SourceMgr &LSM = *D.getSourceMgr();
174     SourceLocation Loc = translateLocation(LSM, D.getLoc());
175     TheParser.Diag(Loc, diag::err_inline_ms_asm_parsing) << D.getMessage();
176   }
177 };
178 }
179 
180 /// Parse an identifier in an MS-style inline assembly block.
181 ///
182 /// \param CastInfo - a void* so that we don't have to teach Parser.h
183 ///   about the actual type.
ParseMSAsmIdentifier(llvm::SmallVectorImpl<Token> & LineToks,unsigned & NumLineToksConsumed,void * CastInfo,bool IsUnevaluatedContext)184 ExprResult Parser::ParseMSAsmIdentifier(llvm::SmallVectorImpl<Token> &LineToks,
185                                         unsigned &NumLineToksConsumed,
186                                         void *CastInfo,
187                                         bool IsUnevaluatedContext) {
188   llvm::InlineAsmIdentifierInfo &Info =
189       *(llvm::InlineAsmIdentifierInfo *)CastInfo;
190 
191   // Push a fake token on the end so that we don't overrun the token
192   // stream.  We use ';' because it expression-parsing should never
193   // overrun it.
194   const tok::TokenKind EndOfStream = tok::semi;
195   Token EndOfStreamTok;
196   EndOfStreamTok.startToken();
197   EndOfStreamTok.setKind(EndOfStream);
198   LineToks.push_back(EndOfStreamTok);
199 
200   // Also copy the current token over.
201   LineToks.push_back(Tok);
202 
203   PP.EnterTokenStream(LineToks, /*DisableMacroExpansions*/ true);
204 
205   // Clear the current token and advance to the first token in LineToks.
206   ConsumeAnyToken();
207 
208   // Parse an optional scope-specifier if we're in C++.
209   CXXScopeSpec SS;
210   if (getLangOpts().CPlusPlus) {
211     ParseOptionalCXXScopeSpecifier(SS, nullptr, /*EnteringContext=*/false);
212   }
213 
214   // Require an identifier here.
215   SourceLocation TemplateKWLoc;
216   UnqualifiedId Id;
217   bool Invalid = true;
218   ExprResult Result;
219   if (Tok.is(tok::kw_this)) {
220     Result = ParseCXXThis();
221     Invalid = false;
222   } else {
223     Invalid = ParseUnqualifiedId(SS,
224                                  /*EnteringContext=*/false,
225                                  /*AllowDestructorName=*/false,
226                                  /*AllowConstructorName=*/false,
227                                  /*ObjectType=*/nullptr, TemplateKWLoc, Id);
228     // Perform the lookup.
229     Result = Actions.LookupInlineAsmIdentifier(SS, TemplateKWLoc, Id, Info,
230                                                IsUnevaluatedContext);
231   }
232   // While the next two tokens are 'period' 'identifier', repeatedly parse it as
233   // a field access. We have to avoid consuming assembler directives that look
234   // like '.' 'else'.
235   while (Result.isUsable() && Tok.is(tok::period)) {
236     Token IdTok = PP.LookAhead(0);
237     if (IdTok.isNot(tok::identifier))
238       break;
239     ConsumeToken(); // Consume the period.
240     IdentifierInfo *Id = Tok.getIdentifierInfo();
241     ConsumeToken(); // Consume the identifier.
242     Result = Actions.LookupInlineAsmVarDeclField(Result.get(), Id->getName(),
243                                                  Info, Tok.getLocation());
244   }
245 
246   // Figure out how many tokens we are into LineToks.
247   unsigned LineIndex = 0;
248   if (Tok.is(EndOfStream)) {
249     LineIndex = LineToks.size() - 2;
250   } else {
251     while (LineToks[LineIndex].getLocation() != Tok.getLocation()) {
252       LineIndex++;
253       assert(LineIndex < LineToks.size() - 2); // we added two extra tokens
254     }
255   }
256 
257   // If we've run into the poison token we inserted before, or there
258   // was a parsing error, then claim the entire line.
259   if (Invalid || Tok.is(EndOfStream)) {
260     NumLineToksConsumed = LineToks.size() - 2;
261   } else {
262     // Otherwise, claim up to the start of the next token.
263     NumLineToksConsumed = LineIndex;
264   }
265 
266   // Finally, restore the old parsing state by consuming all the tokens we
267   // staged before, implicitly killing off the token-lexer we pushed.
268   for (unsigned i = 0, e = LineToks.size() - LineIndex - 2; i != e; ++i) {
269     ConsumeAnyToken();
270   }
271   assert(Tok.is(EndOfStream));
272   ConsumeToken();
273 
274   // Leave LineToks in its original state.
275   LineToks.pop_back();
276   LineToks.pop_back();
277 
278   return Result;
279 }
280 
281 /// Turn a sequence of our tokens back into a string that we can hand
282 /// to the MC asm parser.
buildMSAsmString(Preprocessor & PP,SourceLocation AsmLoc,ArrayRef<Token> AsmToks,SmallVectorImpl<unsigned> & TokOffsets,SmallString<512> & Asm)283 static bool buildMSAsmString(Preprocessor &PP, SourceLocation AsmLoc,
284                              ArrayRef<Token> AsmToks,
285                              SmallVectorImpl<unsigned> &TokOffsets,
286                              SmallString<512> &Asm) {
287   assert(!AsmToks.empty() && "Didn't expect an empty AsmToks!");
288 
289   // Is this the start of a new assembly statement?
290   bool isNewStatement = true;
291 
292   for (unsigned i = 0, e = AsmToks.size(); i < e; ++i) {
293     const Token &Tok = AsmToks[i];
294 
295     // Start each new statement with a newline and a tab.
296     if (!isNewStatement && (Tok.is(tok::kw_asm) || Tok.isAtStartOfLine())) {
297       Asm += "\n\t";
298       isNewStatement = true;
299     }
300 
301     // Preserve the existence of leading whitespace except at the
302     // start of a statement.
303     if (!isNewStatement && Tok.hasLeadingSpace())
304       Asm += ' ';
305 
306     // Remember the offset of this token.
307     TokOffsets.push_back(Asm.size());
308 
309     // Don't actually write '__asm' into the assembly stream.
310     if (Tok.is(tok::kw_asm)) {
311       // Complain about __asm at the end of the stream.
312       if (i + 1 == e) {
313         PP.Diag(AsmLoc, diag::err_asm_empty);
314         return true;
315       }
316 
317       continue;
318     }
319 
320     // Append the spelling of the token.
321     SmallString<32> SpellingBuffer;
322     bool SpellingInvalid = false;
323     Asm += PP.getSpelling(Tok, SpellingBuffer, &SpellingInvalid);
324     assert(!SpellingInvalid && "spelling was invalid after correct parse?");
325 
326     // We are no longer at the start of a statement.
327     isNewStatement = false;
328   }
329 
330   // Ensure that the buffer is null-terminated.
331   Asm.push_back('\0');
332   Asm.pop_back();
333 
334   assert(TokOffsets.size() == AsmToks.size());
335   return false;
336 }
337 
338 /// isTypeQualifier - Return true if the current token could be the
339 /// start of a type-qualifier-list.
isTypeQualifier(const Token & Tok)340 static bool isTypeQualifier(const Token &Tok) {
341   switch (Tok.getKind()) {
342   default: return false;
343   // type-qualifier
344   case tok::kw_const:
345   case tok::kw_volatile:
346   case tok::kw_restrict:
347   case tok::kw___private:
348   case tok::kw___local:
349   case tok::kw___global:
350   case tok::kw___constant:
351   case tok::kw___generic:
352   case tok::kw___read_only:
353   case tok::kw___read_write:
354   case tok::kw___write_only:
355     return true;
356   }
357 }
358 
359 // Determine if this is a GCC-style asm statement.
isGCCAsmStatement(const Token & TokAfterAsm)360 static bool isGCCAsmStatement(const Token &TokAfterAsm) {
361   return TokAfterAsm.is(tok::l_paren) || TokAfterAsm.is(tok::kw_goto) ||
362          isTypeQualifier(TokAfterAsm);
363 }
364 
365 /// ParseMicrosoftAsmStatement. When -fms-extensions/-fasm-blocks is enabled,
366 /// this routine is called to collect the tokens for an MS asm statement.
367 ///
368 /// [MS]  ms-asm-statement:
369 ///         ms-asm-block
370 ///         ms-asm-block ms-asm-statement
371 ///
372 /// [MS]  ms-asm-block:
373 ///         '__asm' ms-asm-line '\n'
374 ///         '__asm' '{' ms-asm-instruction-block[opt] '}' ';'[opt]
375 ///
376 /// [MS]  ms-asm-instruction-block
377 ///         ms-asm-line
378 ///         ms-asm-line '\n' ms-asm-instruction-block
379 ///
ParseMicrosoftAsmStatement(SourceLocation AsmLoc)380 StmtResult Parser::ParseMicrosoftAsmStatement(SourceLocation AsmLoc) {
381   SourceManager &SrcMgr = PP.getSourceManager();
382   SourceLocation EndLoc = AsmLoc;
383   SmallVector<Token, 4> AsmToks;
384 
385   bool SingleLineMode = true;
386   unsigned BraceNesting = 0;
387   unsigned short savedBraceCount = BraceCount;
388   bool InAsmComment = false;
389   FileID FID;
390   unsigned LineNo = 0;
391   unsigned NumTokensRead = 0;
392   SmallVector<SourceLocation, 4> LBraceLocs;
393   bool SkippedStartOfLine = false;
394 
395   if (Tok.is(tok::l_brace)) {
396     // Braced inline asm: consume the opening brace.
397     SingleLineMode = false;
398     BraceNesting = 1;
399     EndLoc = ConsumeBrace();
400     LBraceLocs.push_back(EndLoc);
401     ++NumTokensRead;
402   } else {
403     // Single-line inline asm; compute which line it is on.
404     std::pair<FileID, unsigned> ExpAsmLoc =
405         SrcMgr.getDecomposedExpansionLoc(EndLoc);
406     FID = ExpAsmLoc.first;
407     LineNo = SrcMgr.getLineNumber(FID, ExpAsmLoc.second);
408     LBraceLocs.push_back(SourceLocation());
409   }
410 
411   SourceLocation TokLoc = Tok.getLocation();
412   do {
413     // If we hit EOF, we're done, period.
414     if (isEofOrEom())
415       break;
416 
417     if (!InAsmComment && Tok.is(tok::l_brace)) {
418       // Consume the opening brace.
419       SkippedStartOfLine = Tok.isAtStartOfLine();
420       AsmToks.push_back(Tok);
421       EndLoc = ConsumeBrace();
422       BraceNesting++;
423       LBraceLocs.push_back(EndLoc);
424       TokLoc = Tok.getLocation();
425       ++NumTokensRead;
426       continue;
427     } else if (!InAsmComment && Tok.is(tok::semi)) {
428       // A semicolon in an asm is the start of a comment.
429       InAsmComment = true;
430       if (!SingleLineMode) {
431         // Compute which line the comment is on.
432         std::pair<FileID, unsigned> ExpSemiLoc =
433             SrcMgr.getDecomposedExpansionLoc(TokLoc);
434         FID = ExpSemiLoc.first;
435         LineNo = SrcMgr.getLineNumber(FID, ExpSemiLoc.second);
436       }
437     } else if (SingleLineMode || InAsmComment) {
438       // If end-of-line is significant, check whether this token is on a
439       // new line.
440       std::pair<FileID, unsigned> ExpLoc =
441           SrcMgr.getDecomposedExpansionLoc(TokLoc);
442       if (ExpLoc.first != FID ||
443           SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second) != LineNo) {
444         // If this is a single-line __asm, we're done, except if the next
445         // line is MS-style asm too, in which case we finish a comment
446         // if needed and then keep processing the next line as a single
447         // line __asm.
448         bool isAsm = Tok.is(tok::kw_asm);
449         if (SingleLineMode && (!isAsm || isGCCAsmStatement(NextToken())))
450           break;
451         // We're no longer in a comment.
452         InAsmComment = false;
453         if (isAsm) {
454           // If this is a new __asm {} block we want to process it seperately
455           // from the single-line __asm statements
456           if (PP.LookAhead(0).is(tok::l_brace))
457             break;
458           LineNo = SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second);
459           SkippedStartOfLine = Tok.isAtStartOfLine();
460         }
461       } else if (!InAsmComment && Tok.is(tok::r_brace)) {
462         // In MSVC mode, braces only participate in brace matching and
463         // separating the asm statements.  This is an intentional
464         // departure from the Apple gcc behavior.
465         if (!BraceNesting)
466           break;
467       }
468     }
469     if (!InAsmComment && BraceNesting && Tok.is(tok::r_brace) &&
470         BraceCount == (savedBraceCount + BraceNesting)) {
471       // Consume the closing brace.
472       SkippedStartOfLine = Tok.isAtStartOfLine();
473       // Don't want to add the closing brace of the whole asm block
474       if (SingleLineMode || BraceNesting > 1) {
475         Tok.clearFlag(Token::LeadingSpace);
476         AsmToks.push_back(Tok);
477       }
478       EndLoc = ConsumeBrace();
479       BraceNesting--;
480       // Finish if all of the opened braces in the inline asm section were
481       // consumed.
482       if (BraceNesting == 0 && !SingleLineMode)
483         break;
484       else {
485         LBraceLocs.pop_back();
486         TokLoc = Tok.getLocation();
487         ++NumTokensRead;
488         continue;
489       }
490     }
491 
492     // Consume the next token; make sure we don't modify the brace count etc.
493     // if we are in a comment.
494     EndLoc = TokLoc;
495     if (InAsmComment)
496       PP.Lex(Tok);
497     else {
498       // Set the token as the start of line if we skipped the original start
499       // of line token in case it was a nested brace.
500       if (SkippedStartOfLine)
501         Tok.setFlag(Token::StartOfLine);
502       AsmToks.push_back(Tok);
503       ConsumeAnyToken();
504     }
505     TokLoc = Tok.getLocation();
506     ++NumTokensRead;
507     SkippedStartOfLine = false;
508   } while (1);
509 
510   if (BraceNesting && BraceCount != savedBraceCount) {
511     // __asm without closing brace (this can happen at EOF).
512     for (unsigned i = 0; i < BraceNesting; ++i) {
513       Diag(Tok, diag::err_expected) << tok::r_brace;
514       Diag(LBraceLocs.back(), diag::note_matching) << tok::l_brace;
515       LBraceLocs.pop_back();
516     }
517     return StmtError();
518   } else if (NumTokensRead == 0) {
519     // Empty __asm.
520     Diag(Tok, diag::err_expected) << tok::l_brace;
521     return StmtError();
522   }
523 
524   // Okay, prepare to use MC to parse the assembly.
525   SmallVector<StringRef, 4> ConstraintRefs;
526   SmallVector<Expr *, 4> Exprs;
527   SmallVector<StringRef, 4> ClobberRefs;
528 
529   // We need an actual supported target.
530   const llvm::Triple &TheTriple = Actions.Context.getTargetInfo().getTriple();
531   llvm::Triple::ArchType ArchTy = TheTriple.getArch();
532   const std::string &TT = TheTriple.getTriple();
533   const llvm::Target *TheTarget = nullptr;
534   bool UnsupportedArch =
535       (ArchTy != llvm::Triple::x86 && ArchTy != llvm::Triple::x86_64);
536   if (UnsupportedArch) {
537     Diag(AsmLoc, diag::err_msasm_unsupported_arch) << TheTriple.getArchName();
538   } else {
539     std::string Error;
540     TheTarget = llvm::TargetRegistry::lookupTarget(TT, Error);
541     if (!TheTarget)
542       Diag(AsmLoc, diag::err_msasm_unable_to_create_target) << Error;
543   }
544 
545   assert(!LBraceLocs.empty() && "Should have at least one location here");
546 
547   // If we don't support assembly, or the assembly is empty, we don't
548   // need to instantiate the AsmParser, etc.
549   if (!TheTarget || AsmToks.empty()) {
550     return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, StringRef(),
551                                   /*NumOutputs*/ 0, /*NumInputs*/ 0,
552                                   ConstraintRefs, ClobberRefs, Exprs, EndLoc);
553   }
554 
555   // Expand the tokens into a string buffer.
556   SmallString<512> AsmString;
557   SmallVector<unsigned, 8> TokOffsets;
558   if (buildMSAsmString(PP, AsmLoc, AsmToks, TokOffsets, AsmString))
559     return StmtError();
560 
561   TargetOptions TO = Actions.Context.getTargetInfo().getTargetOpts();
562   std::string FeaturesStr =
563       llvm::join(TO.Features.begin(), TO.Features.end(), ",");
564 
565   std::unique_ptr<llvm::MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TT));
566   std::unique_ptr<llvm::MCAsmInfo> MAI(TheTarget->createMCAsmInfo(*MRI, TT));
567   // Get the instruction descriptor.
568   std::unique_ptr<llvm::MCInstrInfo> MII(TheTarget->createMCInstrInfo());
569   std::unique_ptr<llvm::MCObjectFileInfo> MOFI(new llvm::MCObjectFileInfo());
570   std::unique_ptr<llvm::MCSubtargetInfo> STI(
571       TheTarget->createMCSubtargetInfo(TT, TO.CPU, FeaturesStr));
572 
573   llvm::SourceMgr TempSrcMgr;
574   llvm::MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &TempSrcMgr);
575   MOFI->InitMCObjectFileInfo(TheTriple, /*PIC*/ false, llvm::CodeModel::Default,
576                              Ctx);
577   std::unique_ptr<llvm::MemoryBuffer> Buffer =
578       llvm::MemoryBuffer::getMemBuffer(AsmString, "<MS inline asm>");
579 
580   // Tell SrcMgr about this buffer, which is what the parser will pick up.
581   TempSrcMgr.AddNewSourceBuffer(std::move(Buffer), llvm::SMLoc());
582 
583   std::unique_ptr<llvm::MCStreamer> Str(createNullStreamer(Ctx));
584   std::unique_ptr<llvm::MCAsmParser> Parser(
585       createMCAsmParser(TempSrcMgr, Ctx, *Str.get(), *MAI));
586 
587   // FIXME: init MCOptions from sanitizer flags here.
588   llvm::MCTargetOptions MCOptions;
589   std::unique_ptr<llvm::MCTargetAsmParser> TargetParser(
590       TheTarget->createMCAsmParser(*STI, *Parser, *MII, MCOptions));
591 
592   std::unique_ptr<llvm::MCInstPrinter> IP(
593       TheTarget->createMCInstPrinter(llvm::Triple(TT), 1, *MAI, *MII, *MRI));
594 
595   // Change to the Intel dialect.
596   Parser->setAssemblerDialect(1);
597   Parser->setTargetParser(*TargetParser.get());
598   Parser->setParsingInlineAsm(true);
599   TargetParser->setParsingInlineAsm(true);
600 
601   ClangAsmParserCallback Callback(*this, AsmLoc, AsmString, AsmToks,
602                                   TokOffsets);
603   TargetParser->setSemaCallback(&Callback);
604   TempSrcMgr.setDiagHandler(ClangAsmParserCallback::DiagHandlerCallback,
605                             &Callback);
606 
607   unsigned NumOutputs;
608   unsigned NumInputs;
609   std::string AsmStringIR;
610   SmallVector<std::pair<void *, bool>, 4> OpExprs;
611   SmallVector<std::string, 4> Constraints;
612   SmallVector<std::string, 4> Clobbers;
613   if (Parser->parseMSInlineAsm(AsmLoc.getPtrEncoding(), AsmStringIR, NumOutputs,
614                                NumInputs, OpExprs, Constraints, Clobbers,
615                                MII.get(), IP.get(), Callback))
616     return StmtError();
617 
618   // Filter out "fpsw".  Clang doesn't accept it, and it always lists flags and
619   // fpsr as clobbers.
620   auto End = std::remove(Clobbers.begin(), Clobbers.end(), "fpsw");
621   Clobbers.erase(End, Clobbers.end());
622 
623   // Build the vector of clobber StringRefs.
624   ClobberRefs.insert(ClobberRefs.end(), Clobbers.begin(), Clobbers.end());
625 
626   // Recast the void pointers and build the vector of constraint StringRefs.
627   unsigned NumExprs = NumOutputs + NumInputs;
628   ConstraintRefs.resize(NumExprs);
629   Exprs.resize(NumExprs);
630   for (unsigned i = 0, e = NumExprs; i != e; ++i) {
631     Expr *OpExpr = static_cast<Expr *>(OpExprs[i].first);
632     if (!OpExpr)
633       return StmtError();
634 
635     // Need address of variable.
636     if (OpExprs[i].second)
637       OpExpr =
638           Actions.BuildUnaryOp(getCurScope(), AsmLoc, UO_AddrOf, OpExpr).get();
639 
640     ConstraintRefs[i] = StringRef(Constraints[i]);
641     Exprs[i] = OpExpr;
642   }
643 
644   // FIXME: We should be passing source locations for better diagnostics.
645   return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmStringIR,
646                                 NumOutputs, NumInputs, ConstraintRefs,
647                                 ClobberRefs, Exprs, EndLoc);
648 }
649 
650 /// ParseAsmStatement - Parse a GNU extended asm statement.
651 ///       asm-statement:
652 ///         gnu-asm-statement
653 ///         ms-asm-statement
654 ///
655 /// [GNU] gnu-asm-statement:
656 ///         'asm' type-qualifier[opt] '(' asm-argument ')' ';'
657 ///
658 /// [GNU] asm-argument:
659 ///         asm-string-literal
660 ///         asm-string-literal ':' asm-operands[opt]
661 ///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
662 ///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
663 ///                 ':' asm-clobbers
664 ///
665 /// [GNU] asm-clobbers:
666 ///         asm-string-literal
667 ///         asm-clobbers ',' asm-string-literal
668 ///
ParseAsmStatement(bool & msAsm)669 StmtResult Parser::ParseAsmStatement(bool &msAsm) {
670   assert(Tok.is(tok::kw_asm) && "Not an asm stmt");
671   SourceLocation AsmLoc = ConsumeToken();
672 
673   if (getLangOpts().AsmBlocks && !isGCCAsmStatement(Tok)) {
674     msAsm = true;
675     return ParseMicrosoftAsmStatement(AsmLoc);
676   }
677 
678   DeclSpec DS(AttrFactory);
679   SourceLocation Loc = Tok.getLocation();
680   ParseTypeQualifierListOpt(DS, AR_VendorAttributesParsed);
681 
682   // GNU asms accept, but warn, about type-qualifiers other than volatile.
683   if (DS.getTypeQualifiers() & DeclSpec::TQ_const)
684     Diag(Loc, diag::w_asm_qualifier_ignored) << "const";
685   if (DS.getTypeQualifiers() & DeclSpec::TQ_restrict)
686     Diag(Loc, diag::w_asm_qualifier_ignored) << "restrict";
687   // FIXME: Once GCC supports _Atomic, check whether it permits it here.
688   if (DS.getTypeQualifiers() & DeclSpec::TQ_atomic)
689     Diag(Loc, diag::w_asm_qualifier_ignored) << "_Atomic";
690 
691   // Remember if this was a volatile asm.
692   bool isVolatile = DS.getTypeQualifiers() & DeclSpec::TQ_volatile;
693 
694   // TODO: support "asm goto" constructs (PR#9295).
695   if (Tok.is(tok::kw_goto)) {
696     Diag(Tok, diag::err_asm_goto_not_supported_yet);
697     SkipUntil(tok::r_paren, StopAtSemi);
698     return StmtError();
699   }
700 
701   if (Tok.isNot(tok::l_paren)) {
702     Diag(Tok, diag::err_expected_lparen_after) << "asm";
703     SkipUntil(tok::r_paren, StopAtSemi);
704     return StmtError();
705   }
706   BalancedDelimiterTracker T(*this, tok::l_paren);
707   T.consumeOpen();
708 
709   ExprResult AsmString(ParseAsmStringLiteral());
710 
711   // Check if GNU-style InlineAsm is disabled.
712   // Error on anything other than empty string.
713   if (!(getLangOpts().GNUAsm || AsmString.isInvalid())) {
714     const auto *SL = cast<StringLiteral>(AsmString.get());
715     if (!SL->getString().trim().empty())
716       Diag(Loc, diag::err_gnu_inline_asm_disabled);
717   }
718 
719   if (AsmString.isInvalid()) {
720     // Consume up to and including the closing paren.
721     T.skipToEnd();
722     return StmtError();
723   }
724 
725   SmallVector<IdentifierInfo *, 4> Names;
726   ExprVector Constraints;
727   ExprVector Exprs;
728   ExprVector Clobbers;
729 
730   if (Tok.is(tok::r_paren)) {
731     // We have a simple asm expression like 'asm("foo")'.
732     T.consumeClose();
733     return Actions.ActOnGCCAsmStmt(AsmLoc, /*isSimple*/ true, isVolatile,
734                                    /*NumOutputs*/ 0, /*NumInputs*/ 0, nullptr,
735                                    Constraints, Exprs, AsmString.get(),
736                                    Clobbers, T.getCloseLocation());
737   }
738 
739   // Parse Outputs, if present.
740   bool AteExtraColon = false;
741   if (Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
742     // In C++ mode, parse "::" like ": :".
743     AteExtraColon = Tok.is(tok::coloncolon);
744     ConsumeToken();
745 
746     if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
747       return StmtError();
748   }
749 
750   unsigned NumOutputs = Names.size();
751 
752   // Parse Inputs, if present.
753   if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
754     // In C++ mode, parse "::" like ": :".
755     if (AteExtraColon)
756       AteExtraColon = false;
757     else {
758       AteExtraColon = Tok.is(tok::coloncolon);
759       ConsumeToken();
760     }
761 
762     if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
763       return StmtError();
764   }
765 
766   assert(Names.size() == Constraints.size() &&
767          Constraints.size() == Exprs.size() && "Input operand size mismatch!");
768 
769   unsigned NumInputs = Names.size() - NumOutputs;
770 
771   // Parse the clobbers, if present.
772   if (AteExtraColon || Tok.is(tok::colon)) {
773     if (!AteExtraColon)
774       ConsumeToken();
775 
776     // Parse the asm-string list for clobbers if present.
777     if (Tok.isNot(tok::r_paren)) {
778       while (1) {
779         ExprResult Clobber(ParseAsmStringLiteral());
780 
781         if (Clobber.isInvalid())
782           break;
783 
784         Clobbers.push_back(Clobber.get());
785 
786         if (!TryConsumeToken(tok::comma))
787           break;
788       }
789     }
790   }
791 
792   T.consumeClose();
793   return Actions.ActOnGCCAsmStmt(
794       AsmLoc, false, isVolatile, NumOutputs, NumInputs, Names.data(),
795       Constraints, Exprs, AsmString.get(), Clobbers, T.getCloseLocation());
796 }
797 
798 /// ParseAsmOperands - Parse the asm-operands production as used by
799 /// asm-statement, assuming the leading ':' token was eaten.
800 ///
801 /// [GNU] asm-operands:
802 ///         asm-operand
803 ///         asm-operands ',' asm-operand
804 ///
805 /// [GNU] asm-operand:
806 ///         asm-string-literal '(' expression ')'
807 ///         '[' identifier ']' asm-string-literal '(' expression ')'
808 ///
809 //
810 // FIXME: Avoid unnecessary std::string trashing.
ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo * > & Names,SmallVectorImpl<Expr * > & Constraints,SmallVectorImpl<Expr * > & Exprs)811 bool Parser::ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo *> &Names,
812                                  SmallVectorImpl<Expr *> &Constraints,
813                                  SmallVectorImpl<Expr *> &Exprs) {
814   // 'asm-operands' isn't present?
815   if (!isTokenStringLiteral() && Tok.isNot(tok::l_square))
816     return false;
817 
818   while (1) {
819     // Read the [id] if present.
820     if (Tok.is(tok::l_square)) {
821       BalancedDelimiterTracker T(*this, tok::l_square);
822       T.consumeOpen();
823 
824       if (Tok.isNot(tok::identifier)) {
825         Diag(Tok, diag::err_expected) << tok::identifier;
826         SkipUntil(tok::r_paren, StopAtSemi);
827         return true;
828       }
829 
830       IdentifierInfo *II = Tok.getIdentifierInfo();
831       ConsumeToken();
832 
833       Names.push_back(II);
834       T.consumeClose();
835     } else
836       Names.push_back(nullptr);
837 
838     ExprResult Constraint(ParseAsmStringLiteral());
839     if (Constraint.isInvalid()) {
840       SkipUntil(tok::r_paren, StopAtSemi);
841       return true;
842     }
843     Constraints.push_back(Constraint.get());
844 
845     if (Tok.isNot(tok::l_paren)) {
846       Diag(Tok, diag::err_expected_lparen_after) << "asm operand";
847       SkipUntil(tok::r_paren, StopAtSemi);
848       return true;
849     }
850 
851     // Read the parenthesized expression.
852     BalancedDelimiterTracker T(*this, tok::l_paren);
853     T.consumeOpen();
854     ExprResult Res = Actions.CorrectDelayedTyposInExpr(ParseExpression());
855     T.consumeClose();
856     if (Res.isInvalid()) {
857       SkipUntil(tok::r_paren, StopAtSemi);
858       return true;
859     }
860     Exprs.push_back(Res.get());
861     // Eat the comma and continue parsing if it exists.
862     if (!TryConsumeToken(tok::comma))
863       return false;
864   }
865 }
866