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/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/MCRegisterInfo.h"
27 #include "llvm/MC/MCStreamer.h"
28 #include "llvm/MC/MCSubtargetInfo.h"
29 #include "llvm/MC/MCTargetAsmParser.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;
35 
36 namespace {
37 class ClangAsmParserCallback : public llvm::MCAsmParserSemaCallback {
38   Parser &TheParser;
39   SourceLocation AsmLoc;
40   StringRef AsmString;
41 
42   /// The tokens we streamed into AsmString and handed off to MC.
43   ArrayRef<Token> AsmToks;
44 
45   /// The offset of each token in AsmToks within AsmString.
46   ArrayRef<unsigned> AsmTokOffsets;
47 
48 public:
ClangAsmParserCallback(Parser & P,SourceLocation Loc,StringRef AsmString,ArrayRef<Token> Toks,ArrayRef<unsigned> Offsets)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());
54   }
55 
LookupInlineAsmIdentifier(StringRef & LineBuf,llvm::InlineAsmIdentifierInfo & Info,bool IsUnevaluatedContext)56   void *LookupInlineAsmIdentifier(StringRef &LineBuf,
57                                   llvm::InlineAsmIdentifierInfo &Info,
58                                   bool IsUnevaluatedContext) override {
59     // Collect the desired tokens.
60     SmallVector<Token, 16> LineToks;
61     const Token *FirstOrigToken = nullptr;
62     findTokensForString(LineBuf, LineToks, FirstOrigToken);
63 
64     unsigned NumConsumedToks;
65     ExprResult Result = TheParser.ParseMSAsmIdentifier(
66         LineToks, NumConsumedToks, &Info, IsUnevaluatedContext);
67 
68     // If we consumed the entire line, tell MC that.
69     // Also do this if we consumed nothing as a way of reporting failure.
70     if (NumConsumedToks == 0 || NumConsumedToks == LineToks.size()) {
71       // By not modifying LineBuf, we're implicitly consuming it all.
72 
73       // Otherwise, consume up to the original tokens.
74     } else {
75       assert(FirstOrigToken && "not using original tokens?");
76 
77       // Since we're using original tokens, apply that offset.
78       assert(FirstOrigToken[NumConsumedToks].getLocation() ==
79              LineToks[NumConsumedToks].getLocation());
80       unsigned FirstIndex = FirstOrigToken - AsmToks.begin();
81       unsigned LastIndex = FirstIndex + NumConsumedToks - 1;
82 
83       // The total length we've consumed is the relative offset
84       // of the last token we consumed plus its length.
85       unsigned TotalOffset =
86           (AsmTokOffsets[LastIndex] + AsmToks[LastIndex].getLength() -
87            AsmTokOffsets[FirstIndex]);
88       LineBuf = LineBuf.substr(0, TotalOffset);
89     }
90 
91     // Initialize the "decl" with the lookup result.
92     Info.OpDecl = static_cast<void *>(Result.get());
93     return Info.OpDecl;
94   }
95 
LookupInlineAsmLabel(StringRef Identifier,llvm::SourceMgr & LSM,llvm::SMLoc Location,bool Create)96   StringRef LookupInlineAsmLabel(StringRef Identifier, llvm::SourceMgr &LSM,
97                                  llvm::SMLoc Location,
98                                  bool Create) override {
99     SourceLocation Loc = translateLocation(LSM, Location);
100     LabelDecl *Label =
101       TheParser.getActions().GetOrCreateMSAsmLabel(Identifier, Loc, Create);
102     return Label->getMSAsmLabel();
103   }
104 
LookupInlineAsmField(StringRef Base,StringRef Member,unsigned & Offset)105   bool LookupInlineAsmField(StringRef Base, StringRef Member,
106                             unsigned &Offset) override {
107     return TheParser.getActions().LookupInlineAsmField(Base, Member, Offset,
108                                                        AsmLoc);
109   }
110 
DiagHandlerCallback(const llvm::SMDiagnostic & D,void * Context)111   static void DiagHandlerCallback(const llvm::SMDiagnostic &D, void *Context) {
112     ((ClangAsmParserCallback *)Context)->handleDiagnostic(D);
113   }
114 
115 private:
116   /// Collect the appropriate tokens for the given string.
findTokensForString(StringRef Str,SmallVectorImpl<Token> & TempToks,const Token * & FirstOrigToken) const117   void findTokensForString(StringRef Str, SmallVectorImpl<Token> &TempToks,
118                            const Token *&FirstOrigToken) const {
119     // For now, assert that the string we're working with is a substring
120     // of what we gave to MC.  This lets us use the original tokens.
121     assert(!std::less<const char *>()(Str.begin(), AsmString.begin()) &&
122            !std::less<const char *>()(AsmString.end(), Str.end()));
123 
124     // Try to find a token whose offset matches the first token.
125     unsigned FirstCharOffset = Str.begin() - AsmString.begin();
126     const unsigned *FirstTokOffset = std::lower_bound(
127         AsmTokOffsets.begin(), AsmTokOffsets.end(), FirstCharOffset);
128 
129     // For now, assert that the start of the string exactly
130     // corresponds to the start of a token.
131     assert(*FirstTokOffset == FirstCharOffset);
132 
133     // Use all the original tokens for this line.  (We assume the
134     // end of the line corresponds cleanly to a token break.)
135     unsigned FirstTokIndex = FirstTokOffset - AsmTokOffsets.begin();
136     FirstOrigToken = &AsmToks[FirstTokIndex];
137     unsigned LastCharOffset = Str.end() - AsmString.begin();
138     for (unsigned i = FirstTokIndex, e = AsmTokOffsets.size(); i != e; ++i) {
139       if (AsmTokOffsets[i] >= LastCharOffset)
140         break;
141       TempToks.push_back(AsmToks[i]);
142     }
143   }
144 
translateLocation(const llvm::SourceMgr & LSM,llvm::SMLoc SMLoc)145   SourceLocation translateLocation(const llvm::SourceMgr &LSM, llvm::SMLoc SMLoc) {
146     // Compute an offset into the inline asm buffer.
147     // FIXME: This isn't right if .macro is involved (but hopefully, no
148     // real-world code does that).
149     const llvm::MemoryBuffer *LBuf =
150         LSM.getMemoryBuffer(LSM.FindBufferContainingLoc(SMLoc));
151     unsigned Offset = SMLoc.getPointer() - LBuf->getBufferStart();
152 
153     // Figure out which token that offset points into.
154     const unsigned *TokOffsetPtr =
155         std::lower_bound(AsmTokOffsets.begin(), AsmTokOffsets.end(), Offset);
156     unsigned TokIndex = TokOffsetPtr - AsmTokOffsets.begin();
157     unsigned TokOffset = *TokOffsetPtr;
158 
159     // If we come up with an answer which seems sane, use it; otherwise,
160     // just point at the __asm keyword.
161     // FIXME: Assert the answer is sane once we handle .macro correctly.
162     SourceLocation Loc = AsmLoc;
163     if (TokIndex < AsmToks.size()) {
164       const Token &Tok = AsmToks[TokIndex];
165       Loc = Tok.getLocation();
166       Loc = Loc.getLocWithOffset(Offset - TokOffset);
167     }
168     return Loc;
169   }
170 
handleDiagnostic(const llvm::SMDiagnostic & D)171   void handleDiagnostic(const llvm::SMDiagnostic &D) {
172     const llvm::SourceMgr &LSM = *D.getSourceMgr();
173     SourceLocation Loc = translateLocation(LSM, D.getLoc());
174     TheParser.Diag(Loc, diag::err_inline_ms_asm_parsing) << D.getMessage();
175   }
176 };
177 }
178 
179 /// Parse an identifier in an MS-style inline assembly block.
180 ///
181 /// \param CastInfo - a void* so that we don't have to teach Parser.h
182 ///   about the actual type.
ParseMSAsmIdentifier(llvm::SmallVectorImpl<Token> & LineToks,unsigned & NumLineToksConsumed,void * CastInfo,bool IsUnevaluatedContext)183 ExprResult Parser::ParseMSAsmIdentifier(llvm::SmallVectorImpl<Token> &LineToks,
184                                         unsigned &NumLineToksConsumed,
185                                         void *CastInfo,
186                                         bool IsUnevaluatedContext) {
187   llvm::InlineAsmIdentifierInfo &Info =
188       *(llvm::InlineAsmIdentifierInfo *)CastInfo;
189 
190   // Push a fake token on the end so that we don't overrun the token
191   // stream.  We use ';' because it expression-parsing should never
192   // overrun it.
193   const tok::TokenKind EndOfStream = tok::semi;
194   Token EndOfStreamTok;
195   EndOfStreamTok.startToken();
196   EndOfStreamTok.setKind(EndOfStream);
197   LineToks.push_back(EndOfStreamTok);
198 
199   // Also copy the current token over.
200   LineToks.push_back(Tok);
201 
202   PP.EnterTokenStream(LineToks.begin(), LineToks.size(),
203                       /*disable macros*/ true,
204                       /*owns tokens*/ false);
205 
206   // Clear the current token and advance to the first token in LineToks.
207   ConsumeAnyToken();
208 
209   // Parse an optional scope-specifier if we're in C++.
210   CXXScopeSpec SS;
211   if (getLangOpts().CPlusPlus) {
212     ParseOptionalCXXScopeSpecifier(SS, ParsedType(), /*EnteringContext=*/false);
213   }
214 
215   // Require an identifier here.
216   SourceLocation TemplateKWLoc;
217   UnqualifiedId Id;
218   bool Invalid = true;
219   ExprResult Result;
220   if (Tok.is(tok::kw_this)) {
221     Result = ParseCXXThis();
222     Invalid = false;
223   } else {
224     Invalid =
225         ParseUnqualifiedId(SS,
226                            /*EnteringContext=*/false,
227                            /*AllowDestructorName=*/false,
228                            /*AllowConstructorName=*/false,
229                            /*ObjectType=*/ParsedType(), TemplateKWLoc, Id);
230     // Perform the lookup.
231     Result = Actions.LookupInlineAsmIdentifier(SS, TemplateKWLoc, Id, Info,
232                                                IsUnevaluatedContext);
233   }
234   // While the next two tokens are 'period' 'identifier', repeatedly parse it as
235   // a field access. We have to avoid consuming assembler directives that look
236   // like '.' 'else'.
237   while (Result.isUsable() && Tok.is(tok::period)) {
238     Token IdTok = PP.LookAhead(0);
239     if (IdTok.isNot(tok::identifier))
240       break;
241     ConsumeToken(); // Consume the period.
242     IdentifierInfo *Id = Tok.getIdentifierInfo();
243     ConsumeToken(); // Consume the identifier.
244     unsigned OffsetUnused;
245     Result = Actions.LookupInlineAsmVarDeclField(
246         Result.get(), Id->getName(), OffsetUnused, Info, Tok.getLocation());
247   }
248 
249   // Figure out how many tokens we are into LineToks.
250   unsigned LineIndex = 0;
251   if (Tok.is(EndOfStream)) {
252     LineIndex = LineToks.size() - 2;
253   } else {
254     while (LineToks[LineIndex].getLocation() != Tok.getLocation()) {
255       LineIndex++;
256       assert(LineIndex < LineToks.size() - 2); // we added two extra tokens
257     }
258   }
259 
260   // If we've run into the poison token we inserted before, or there
261   // was a parsing error, then claim the entire line.
262   if (Invalid || Tok.is(EndOfStream)) {
263     NumLineToksConsumed = LineToks.size() - 2;
264   } else {
265     // Otherwise, claim up to the start of the next token.
266     NumLineToksConsumed = LineIndex;
267   }
268 
269   // Finally, restore the old parsing state by consuming all the tokens we
270   // staged before, implicitly killing off the token-lexer we pushed.
271   for (unsigned i = 0, e = LineToks.size() - LineIndex - 2; i != e; ++i) {
272     ConsumeAnyToken();
273   }
274   assert(Tok.is(EndOfStream));
275   ConsumeToken();
276 
277   // Leave LineToks in its original state.
278   LineToks.pop_back();
279   LineToks.pop_back();
280 
281   return Result;
282 }
283 
284 /// Turn a sequence of our tokens back into a string that we can hand
285 /// to the MC asm parser.
buildMSAsmString(Preprocessor & PP,SourceLocation AsmLoc,ArrayRef<Token> AsmToks,SmallVectorImpl<unsigned> & TokOffsets,SmallString<512> & Asm)286 static bool buildMSAsmString(Preprocessor &PP, SourceLocation AsmLoc,
287                              ArrayRef<Token> AsmToks,
288                              SmallVectorImpl<unsigned> &TokOffsets,
289                              SmallString<512> &Asm) {
290   assert(!AsmToks.empty() && "Didn't expect an empty AsmToks!");
291 
292   // Is this the start of a new assembly statement?
293   bool isNewStatement = true;
294 
295   for (unsigned i = 0, e = AsmToks.size(); i < e; ++i) {
296     const Token &Tok = AsmToks[i];
297 
298     // Start each new statement with a newline and a tab.
299     if (!isNewStatement && (Tok.is(tok::kw_asm) || Tok.isAtStartOfLine())) {
300       Asm += "\n\t";
301       isNewStatement = true;
302     }
303 
304     // Preserve the existence of leading whitespace except at the
305     // start of a statement.
306     if (!isNewStatement && Tok.hasLeadingSpace())
307       Asm += ' ';
308 
309     // Remember the offset of this token.
310     TokOffsets.push_back(Asm.size());
311 
312     // Don't actually write '__asm' into the assembly stream.
313     if (Tok.is(tok::kw_asm)) {
314       // Complain about __asm at the end of the stream.
315       if (i + 1 == e) {
316         PP.Diag(AsmLoc, diag::err_asm_empty);
317         return true;
318       }
319 
320       continue;
321     }
322 
323     // Append the spelling of the token.
324     SmallString<32> SpellingBuffer;
325     bool SpellingInvalid = false;
326     Asm += PP.getSpelling(Tok, SpellingBuffer, &SpellingInvalid);
327     assert(!SpellingInvalid && "spelling was invalid after correct parse?");
328 
329     // We are no longer at the start of a statement.
330     isNewStatement = false;
331   }
332 
333   // Ensure that the buffer is null-terminated.
334   Asm.push_back('\0');
335   Asm.pop_back();
336 
337   assert(TokOffsets.size() == AsmToks.size());
338   return false;
339 }
340 
341 /// ParseMicrosoftAsmStatement. When -fms-extensions/-fasm-blocks is enabled,
342 /// this routine is called to collect the tokens for an MS asm statement.
343 ///
344 /// [MS]  ms-asm-statement:
345 ///         ms-asm-block
346 ///         ms-asm-block ms-asm-statement
347 ///
348 /// [MS]  ms-asm-block:
349 ///         '__asm' ms-asm-line '\n'
350 ///         '__asm' '{' ms-asm-instruction-block[opt] '}' ';'[opt]
351 ///
352 /// [MS]  ms-asm-instruction-block
353 ///         ms-asm-line
354 ///         ms-asm-line '\n' ms-asm-instruction-block
355 ///
ParseMicrosoftAsmStatement(SourceLocation AsmLoc)356 StmtResult Parser::ParseMicrosoftAsmStatement(SourceLocation AsmLoc) {
357   SourceManager &SrcMgr = PP.getSourceManager();
358   SourceLocation EndLoc = AsmLoc;
359   SmallVector<Token, 4> AsmToks;
360 
361   bool SingleLineMode = true;
362   unsigned BraceNesting = 0;
363   unsigned short savedBraceCount = BraceCount;
364   bool InAsmComment = false;
365   FileID FID;
366   unsigned LineNo = 0;
367   unsigned NumTokensRead = 0;
368   SmallVector<SourceLocation, 4> LBraceLocs;
369   bool SkippedStartOfLine = false;
370 
371   if (Tok.is(tok::l_brace)) {
372     // Braced inline asm: consume the opening brace.
373     SingleLineMode = false;
374     BraceNesting = 1;
375     EndLoc = ConsumeBrace();
376     LBraceLocs.push_back(EndLoc);
377     ++NumTokensRead;
378   } else {
379     // Single-line inline asm; compute which line it is on.
380     std::pair<FileID, unsigned> ExpAsmLoc =
381         SrcMgr.getDecomposedExpansionLoc(EndLoc);
382     FID = ExpAsmLoc.first;
383     LineNo = SrcMgr.getLineNumber(FID, ExpAsmLoc.second);
384     LBraceLocs.push_back(SourceLocation());
385   }
386 
387   SourceLocation TokLoc = Tok.getLocation();
388   do {
389     // If we hit EOF, we're done, period.
390     if (isEofOrEom())
391       break;
392 
393     if (!InAsmComment && Tok.is(tok::l_brace)) {
394       // Consume the opening brace.
395       SkippedStartOfLine = Tok.isAtStartOfLine();
396       EndLoc = ConsumeBrace();
397       BraceNesting++;
398       LBraceLocs.push_back(EndLoc);
399       TokLoc = Tok.getLocation();
400       ++NumTokensRead;
401       continue;
402     } else if (!InAsmComment && Tok.is(tok::semi)) {
403       // A semicolon in an asm is the start of a comment.
404       InAsmComment = true;
405       if (!SingleLineMode) {
406         // Compute which line the comment is on.
407         std::pair<FileID, unsigned> ExpSemiLoc =
408             SrcMgr.getDecomposedExpansionLoc(TokLoc);
409         FID = ExpSemiLoc.first;
410         LineNo = SrcMgr.getLineNumber(FID, ExpSemiLoc.second);
411       }
412     } else if (SingleLineMode || InAsmComment) {
413       // If end-of-line is significant, check whether this token is on a
414       // new line.
415       std::pair<FileID, unsigned> ExpLoc =
416           SrcMgr.getDecomposedExpansionLoc(TokLoc);
417       if (ExpLoc.first != FID ||
418           SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second) != LineNo) {
419         // If this is a single-line __asm, we're done, except if the next
420         // line begins with an __asm too, in which case we finish a comment
421         // if needed and then keep processing the next line as a single
422         // line __asm.
423         bool isAsm = Tok.is(tok::kw_asm);
424         if (SingleLineMode && !isAsm)
425           break;
426         // We're no longer in a comment.
427         InAsmComment = false;
428         if (isAsm) {
429           LineNo = SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second);
430           SkippedStartOfLine = Tok.isAtStartOfLine();
431         }
432       } else if (!InAsmComment && Tok.is(tok::r_brace)) {
433         // In MSVC mode, braces only participate in brace matching and
434         // separating the asm statements.  This is an intentional
435         // departure from the Apple gcc behavior.
436         if (!BraceNesting)
437           break;
438       }
439     }
440     if (!InAsmComment && BraceNesting && Tok.is(tok::r_brace) &&
441         BraceCount == (savedBraceCount + BraceNesting)) {
442       // Consume the closing brace.
443       SkippedStartOfLine = Tok.isAtStartOfLine();
444       EndLoc = ConsumeBrace();
445       BraceNesting--;
446       // Finish if all of the opened braces in the inline asm section were
447       // consumed.
448       if (BraceNesting == 0 && !SingleLineMode)
449         break;
450       else {
451         LBraceLocs.pop_back();
452         TokLoc = Tok.getLocation();
453         ++NumTokensRead;
454         continue;
455       }
456     }
457 
458     // Consume the next token; make sure we don't modify the brace count etc.
459     // if we are in a comment.
460     EndLoc = TokLoc;
461     if (InAsmComment)
462       PP.Lex(Tok);
463     else {
464       // Set the token as the start of line if we skipped the original start
465       // of line token in case it was a nested brace.
466       if (SkippedStartOfLine)
467         Tok.setFlag(Token::StartOfLine);
468       AsmToks.push_back(Tok);
469       ConsumeAnyToken();
470     }
471     TokLoc = Tok.getLocation();
472     ++NumTokensRead;
473     SkippedStartOfLine = false;
474   } while (1);
475 
476   if (BraceNesting && BraceCount != savedBraceCount) {
477     // __asm without closing brace (this can happen at EOF).
478     for (unsigned i = 0; i < BraceNesting; ++i) {
479       Diag(Tok, diag::err_expected) << tok::r_brace;
480       Diag(LBraceLocs.back(), diag::note_matching) << tok::l_brace;
481       LBraceLocs.pop_back();
482     }
483     return StmtError();
484   } else if (NumTokensRead == 0) {
485     // Empty __asm.
486     Diag(Tok, diag::err_expected) << tok::l_brace;
487     return StmtError();
488   }
489 
490   // Okay, prepare to use MC to parse the assembly.
491   SmallVector<StringRef, 4> ConstraintRefs;
492   SmallVector<Expr *, 4> Exprs;
493   SmallVector<StringRef, 4> ClobberRefs;
494 
495   // We need an actual supported target.
496   const llvm::Triple &TheTriple = Actions.Context.getTargetInfo().getTriple();
497   llvm::Triple::ArchType ArchTy = TheTriple.getArch();
498   const std::string &TT = TheTriple.getTriple();
499   const llvm::Target *TheTarget = nullptr;
500   bool UnsupportedArch =
501       (ArchTy != llvm::Triple::x86 && ArchTy != llvm::Triple::x86_64);
502   if (UnsupportedArch) {
503     Diag(AsmLoc, diag::err_msasm_unsupported_arch) << TheTriple.getArchName();
504   } else {
505     std::string Error;
506     TheTarget = llvm::TargetRegistry::lookupTarget(TT, Error);
507     if (!TheTarget)
508       Diag(AsmLoc, diag::err_msasm_unable_to_create_target) << Error;
509   }
510 
511   assert(!LBraceLocs.empty() && "Should have at least one location here");
512 
513   // If we don't support assembly, or the assembly is empty, we don't
514   // need to instantiate the AsmParser, etc.
515   if (!TheTarget || AsmToks.empty()) {
516     return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, StringRef(),
517                                   /*NumOutputs*/ 0, /*NumInputs*/ 0,
518                                   ConstraintRefs, ClobberRefs, Exprs, EndLoc);
519   }
520 
521   // Expand the tokens into a string buffer.
522   SmallString<512> AsmString;
523   SmallVector<unsigned, 8> TokOffsets;
524   if (buildMSAsmString(PP, AsmLoc, AsmToks, TokOffsets, AsmString))
525     return StmtError();
526 
527   std::unique_ptr<llvm::MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TT));
528   std::unique_ptr<llvm::MCAsmInfo> MAI(TheTarget->createMCAsmInfo(*MRI, TT));
529   // Get the instruction descriptor.
530   std::unique_ptr<llvm::MCInstrInfo> MII(TheTarget->createMCInstrInfo());
531   std::unique_ptr<llvm::MCObjectFileInfo> MOFI(new llvm::MCObjectFileInfo());
532   std::unique_ptr<llvm::MCSubtargetInfo> STI(
533       TheTarget->createMCSubtargetInfo(TT, "", ""));
534 
535   llvm::SourceMgr TempSrcMgr;
536   llvm::MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &TempSrcMgr);
537   MOFI->InitMCObjectFileInfo(TheTriple, llvm::Reloc::Default,
538                              llvm::CodeModel::Default, Ctx);
539   std::unique_ptr<llvm::MemoryBuffer> Buffer =
540       llvm::MemoryBuffer::getMemBuffer(AsmString, "<MS inline asm>");
541 
542   // Tell SrcMgr about this buffer, which is what the parser will pick up.
543   TempSrcMgr.AddNewSourceBuffer(std::move(Buffer), llvm::SMLoc());
544 
545   std::unique_ptr<llvm::MCStreamer> Str(createNullStreamer(Ctx));
546   std::unique_ptr<llvm::MCAsmParser> Parser(
547       createMCAsmParser(TempSrcMgr, Ctx, *Str.get(), *MAI));
548 
549   // FIXME: init MCOptions from sanitizer flags here.
550   llvm::MCTargetOptions MCOptions;
551   std::unique_ptr<llvm::MCTargetAsmParser> TargetParser(
552       TheTarget->createMCAsmParser(*STI, *Parser, *MII, MCOptions));
553 
554   std::unique_ptr<llvm::MCInstPrinter> IP(
555       TheTarget->createMCInstPrinter(llvm::Triple(TT), 1, *MAI, *MII, *MRI));
556 
557   // Change to the Intel dialect.
558   Parser->setAssemblerDialect(1);
559   Parser->setTargetParser(*TargetParser.get());
560   Parser->setParsingInlineAsm(true);
561   TargetParser->setParsingInlineAsm(true);
562 
563   ClangAsmParserCallback Callback(*this, AsmLoc, AsmString, AsmToks,
564                                   TokOffsets);
565   TargetParser->setSemaCallback(&Callback);
566   TempSrcMgr.setDiagHandler(ClangAsmParserCallback::DiagHandlerCallback,
567                             &Callback);
568 
569   unsigned NumOutputs;
570   unsigned NumInputs;
571   std::string AsmStringIR;
572   SmallVector<std::pair<void *, bool>, 4> OpExprs;
573   SmallVector<std::string, 4> Constraints;
574   SmallVector<std::string, 4> Clobbers;
575   if (Parser->parseMSInlineAsm(AsmLoc.getPtrEncoding(), AsmStringIR, NumOutputs,
576                                NumInputs, OpExprs, Constraints, Clobbers,
577                                MII.get(), IP.get(), Callback))
578     return StmtError();
579 
580   // Filter out "fpsw".  Clang doesn't accept it, and it always lists flags and
581   // fpsr as clobbers.
582   auto End = std::remove(Clobbers.begin(), Clobbers.end(), "fpsw");
583   Clobbers.erase(End, Clobbers.end());
584 
585   // Build the vector of clobber StringRefs.
586   ClobberRefs.insert(ClobberRefs.end(), Clobbers.begin(), Clobbers.end());
587 
588   // Recast the void pointers and build the vector of constraint StringRefs.
589   unsigned NumExprs = NumOutputs + NumInputs;
590   ConstraintRefs.resize(NumExprs);
591   Exprs.resize(NumExprs);
592   for (unsigned i = 0, e = NumExprs; i != e; ++i) {
593     Expr *OpExpr = static_cast<Expr *>(OpExprs[i].first);
594     if (!OpExpr)
595       return StmtError();
596 
597     // Need address of variable.
598     if (OpExprs[i].second)
599       OpExpr =
600           Actions.BuildUnaryOp(getCurScope(), AsmLoc, UO_AddrOf, OpExpr).get();
601 
602     ConstraintRefs[i] = StringRef(Constraints[i]);
603     Exprs[i] = OpExpr;
604   }
605 
606   // FIXME: We should be passing source locations for better diagnostics.
607   return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmStringIR,
608                                 NumOutputs, NumInputs, ConstraintRefs,
609                                 ClobberRefs, Exprs, EndLoc);
610 }
611 
612 /// ParseAsmStatement - Parse a GNU extended asm statement.
613 ///       asm-statement:
614 ///         gnu-asm-statement
615 ///         ms-asm-statement
616 ///
617 /// [GNU] gnu-asm-statement:
618 ///         'asm' type-qualifier[opt] '(' asm-argument ')' ';'
619 ///
620 /// [GNU] asm-argument:
621 ///         asm-string-literal
622 ///         asm-string-literal ':' asm-operands[opt]
623 ///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
624 ///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
625 ///                 ':' asm-clobbers
626 ///
627 /// [GNU] asm-clobbers:
628 ///         asm-string-literal
629 ///         asm-clobbers ',' asm-string-literal
630 ///
ParseAsmStatement(bool & msAsm)631 StmtResult Parser::ParseAsmStatement(bool &msAsm) {
632   assert(Tok.is(tok::kw_asm) && "Not an asm stmt");
633   SourceLocation AsmLoc = ConsumeToken();
634 
635   if (getLangOpts().AsmBlocks && Tok.isNot(tok::l_paren) &&
636       !isTypeQualifier()) {
637     msAsm = true;
638     return ParseMicrosoftAsmStatement(AsmLoc);
639   }
640 
641   DeclSpec DS(AttrFactory);
642   SourceLocation Loc = Tok.getLocation();
643   ParseTypeQualifierListOpt(DS, AR_VendorAttributesParsed);
644 
645   // GNU asms accept, but warn, about type-qualifiers other than volatile.
646   if (DS.getTypeQualifiers() & DeclSpec::TQ_const)
647     Diag(Loc, diag::w_asm_qualifier_ignored) << "const";
648   if (DS.getTypeQualifiers() & DeclSpec::TQ_restrict)
649     Diag(Loc, diag::w_asm_qualifier_ignored) << "restrict";
650   // FIXME: Once GCC supports _Atomic, check whether it permits it here.
651   if (DS.getTypeQualifiers() & DeclSpec::TQ_atomic)
652     Diag(Loc, diag::w_asm_qualifier_ignored) << "_Atomic";
653 
654   // Remember if this was a volatile asm.
655   bool isVolatile = DS.getTypeQualifiers() & DeclSpec::TQ_volatile;
656   if (Tok.isNot(tok::l_paren)) {
657     Diag(Tok, diag::err_expected_lparen_after) << "asm";
658     SkipUntil(tok::r_paren, StopAtSemi);
659     return StmtError();
660   }
661   BalancedDelimiterTracker T(*this, tok::l_paren);
662   T.consumeOpen();
663 
664   ExprResult AsmString(ParseAsmStringLiteral());
665 
666   // Check if GNU-style InlineAsm is disabled.
667   // Error on anything other than empty string.
668   if (!(getLangOpts().GNUAsm || AsmString.isInvalid())) {
669     const auto *SL = cast<StringLiteral>(AsmString.get());
670     if (!SL->getString().trim().empty())
671       Diag(Loc, diag::err_gnu_inline_asm_disabled);
672   }
673 
674   if (AsmString.isInvalid()) {
675     // Consume up to and including the closing paren.
676     T.skipToEnd();
677     return StmtError();
678   }
679 
680   SmallVector<IdentifierInfo *, 4> Names;
681   ExprVector Constraints;
682   ExprVector Exprs;
683   ExprVector Clobbers;
684 
685   if (Tok.is(tok::r_paren)) {
686     // We have a simple asm expression like 'asm("foo")'.
687     T.consumeClose();
688     return Actions.ActOnGCCAsmStmt(AsmLoc, /*isSimple*/ true, isVolatile,
689                                    /*NumOutputs*/ 0, /*NumInputs*/ 0, nullptr,
690                                    Constraints, Exprs, AsmString.get(),
691                                    Clobbers, T.getCloseLocation());
692   }
693 
694   // Parse Outputs, if present.
695   bool AteExtraColon = false;
696   if (Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
697     // In C++ mode, parse "::" like ": :".
698     AteExtraColon = Tok.is(tok::coloncolon);
699     ConsumeToken();
700 
701     if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
702       return StmtError();
703   }
704 
705   unsigned NumOutputs = Names.size();
706 
707   // Parse Inputs, if present.
708   if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
709     // In C++ mode, parse "::" like ": :".
710     if (AteExtraColon)
711       AteExtraColon = false;
712     else {
713       AteExtraColon = Tok.is(tok::coloncolon);
714       ConsumeToken();
715     }
716 
717     if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
718       return StmtError();
719   }
720 
721   assert(Names.size() == Constraints.size() &&
722          Constraints.size() == Exprs.size() && "Input operand size mismatch!");
723 
724   unsigned NumInputs = Names.size() - NumOutputs;
725 
726   // Parse the clobbers, if present.
727   if (AteExtraColon || Tok.is(tok::colon)) {
728     if (!AteExtraColon)
729       ConsumeToken();
730 
731     // Parse the asm-string list for clobbers if present.
732     if (Tok.isNot(tok::r_paren)) {
733       while (1) {
734         ExprResult Clobber(ParseAsmStringLiteral());
735 
736         if (Clobber.isInvalid())
737           break;
738 
739         Clobbers.push_back(Clobber.get());
740 
741         if (!TryConsumeToken(tok::comma))
742           break;
743       }
744     }
745   }
746 
747   T.consumeClose();
748   return Actions.ActOnGCCAsmStmt(
749       AsmLoc, false, isVolatile, NumOutputs, NumInputs, Names.data(),
750       Constraints, Exprs, AsmString.get(), Clobbers, T.getCloseLocation());
751 }
752 
753 /// ParseAsmOperands - Parse the asm-operands production as used by
754 /// asm-statement, assuming the leading ':' token was eaten.
755 ///
756 /// [GNU] asm-operands:
757 ///         asm-operand
758 ///         asm-operands ',' asm-operand
759 ///
760 /// [GNU] asm-operand:
761 ///         asm-string-literal '(' expression ')'
762 ///         '[' identifier ']' asm-string-literal '(' expression ')'
763 ///
764 //
765 // FIXME: Avoid unnecessary std::string trashing.
ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo * > & Names,SmallVectorImpl<Expr * > & Constraints,SmallVectorImpl<Expr * > & Exprs)766 bool Parser::ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo *> &Names,
767                                  SmallVectorImpl<Expr *> &Constraints,
768                                  SmallVectorImpl<Expr *> &Exprs) {
769   // 'asm-operands' isn't present?
770   if (!isTokenStringLiteral() && Tok.isNot(tok::l_square))
771     return false;
772 
773   while (1) {
774     // Read the [id] if present.
775     if (Tok.is(tok::l_square)) {
776       BalancedDelimiterTracker T(*this, tok::l_square);
777       T.consumeOpen();
778 
779       if (Tok.isNot(tok::identifier)) {
780         Diag(Tok, diag::err_expected) << tok::identifier;
781         SkipUntil(tok::r_paren, StopAtSemi);
782         return true;
783       }
784 
785       IdentifierInfo *II = Tok.getIdentifierInfo();
786       ConsumeToken();
787 
788       Names.push_back(II);
789       T.consumeClose();
790     } else
791       Names.push_back(nullptr);
792 
793     ExprResult Constraint(ParseAsmStringLiteral());
794     if (Constraint.isInvalid()) {
795       SkipUntil(tok::r_paren, StopAtSemi);
796       return true;
797     }
798     Constraints.push_back(Constraint.get());
799 
800     if (Tok.isNot(tok::l_paren)) {
801       Diag(Tok, diag::err_expected_lparen_after) << "asm operand";
802       SkipUntil(tok::r_paren, StopAtSemi);
803       return true;
804     }
805 
806     // Read the parenthesized expression.
807     BalancedDelimiterTracker T(*this, tok::l_paren);
808     T.consumeOpen();
809     ExprResult Res = Actions.CorrectDelayedTyposInExpr(ParseExpression());
810     T.consumeClose();
811     if (Res.isInvalid()) {
812       SkipUntil(tok::r_paren, StopAtSemi);
813       return true;
814     }
815     Exprs.push_back(Res.get());
816     // Eat the comma and continue parsing if it exists.
817     if (!TryConsumeToken(tok::comma))
818       return false;
819   }
820 }
821