1//===---- ParseStmtAsm.cpp - Assembly Statement Parser --------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements parsing for GCC and Microsoft inline assembly.
10//
11//===----------------------------------------------------------------------===//
12
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"
34using namespace clang;
35
36namespace {
37class 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
48public:
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
56  void LookupInlineAsmIdentifier(StringRef &LineBuf,
57                                 llvm::InlineAsmIdentifierInfo &Info,
58                                 bool IsUnevaluatedContext) override;
59
60  StringRef LookupInlineAsmLabel(StringRef Identifier, llvm::SourceMgr &LSM,
61                                 llvm::SMLoc Location,
62                                 bool Create) override;
63
64  bool LookupInlineAsmField(StringRef Base, StringRef Member,
65                            unsigned &Offset) override {
66    return TheParser.getActions().LookupInlineAsmField(Base, Member, Offset,
67                                                       AsmLoc);
68  }
69
70  static void DiagHandlerCallback(const llvm::SMDiagnostic &D, void *Context) {
71    ((ClangAsmParserCallback *)Context)->handleDiagnostic(D);
72  }
73
74private:
75  /// Collect the appropriate tokens for the given string.
76  void findTokensForString(StringRef Str, SmallVectorImpl<Token> &TempToks,
77                           const Token *&FirstOrigToken) const;
78
79  SourceLocation translateLocation(const llvm::SourceMgr &LSM,
80                                   llvm::SMLoc SMLoc);
81
82  void handleDiagnostic(const llvm::SMDiagnostic &D);
83};
84}
85
86void 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);
93
94  unsigned NumConsumedToks;
95  ExprResult Result = TheParser.ParseMSAsmIdentifier(LineToks, NumConsumedToks,
96                                                     IsUnevaluatedContext);
97
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.
102
103    // Otherwise, consume up to the original tokens.
104  } else {
105    assert(FirstOrigToken && "not using original tokens?");
106
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;
112
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);
119  }
120
121  // Initialize Info with the lookup result.
122  if (!Result.isUsable())
123    return;
124  TheParser.getActions().FillInlineAsmIdentifierInfo(Result.get(), Info);
125}
126
127StringRef ClangAsmParserCallback::LookupInlineAsmLabel(StringRef Identifier,
128                                                       llvm::SourceMgr &LSM,
129                                                       llvm::SMLoc Location,
130                                                       bool Create) {
131  SourceLocation Loc = translateLocation(LSM, Location);
132  LabelDecl *Label =
133      TheParser.getActions().GetOrCreateMSAsmLabel(Identifier, Loc, Create);
134  return Label->getMSAsmLabel();
135}
136
137void 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()));
144
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);
149
150  // For now, assert that the start of the string exactly
151  // corresponds to the start of a token.
152  assert(*FirstTokOffset == FirstCharOffset);
153
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)
161      break;
162    TempToks.push_back(AsmToks[i]);
163  }
164}
165
166SourceLocation
167ClangAsmParserCallback::translateLocation(const llvm::SourceMgr &LSM,
168                                          llvm::SMLoc SMLoc) {
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();
175
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;
180
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);
189  }
190  return Loc;
191}
192
193void 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();
197}
198
199/// Parse an identifier in an MS-style inline assembly block.
200ExprResult 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
205  // overrun it.
206  const tok::TokenKind EndOfStream = tok::semi;
207  Token EndOfStreamTok;
208  EndOfStreamTok.startToken();
209  EndOfStreamTok.setKind(EndOfStream);
210  LineToks.push_back(EndOfStreamTok);
211
212  // Also copy the current token over.
213  LineToks.push_back(Tok);
214
215  PP.EnterTokenStream(LineToks, /*DisableMacroExpansions*/ true,
216                      /*IsReinject*/ true);
217
218  // Clear the current token and advance to the first token in LineToks.
219  ConsumeAnyToken();
220
221  // Parse an optional scope-specifier if we're in C++.
222  CXXScopeSpec SS;
223  if (getLangOpts().CPlusPlus)
224    ParseOptionalCXXScopeSpecifier(SS, /*ObjectType=*/nullptr,
225                                   /*ObjectHadErrors=*/false,
226                                   /*EnteringContext=*/false);
227
228  // Require an identifier here.
229  SourceLocation TemplateKWLoc;
230  UnqualifiedId Id;
231  bool Invalid = true;
232  ExprResult Result;
233  if (Tok.is(tok::kw_this)) {
234    Result = ParseCXXThis();
235    Invalid = false;
236  } else {
237    Invalid =
238        ParseUnqualifiedId(SS, /*ObjectType=*/nullptr,
239                           /*ObjectHadErrors=*/false,
240                           /*EnteringContext=*/false,
241                           /*AllowDestructorName=*/false,
242                           /*AllowConstructorName=*/false,
243                           /*AllowDeductionGuide=*/false, &TemplateKWLoc, Id);
244    // Perform the lookup.
245    Result = Actions.LookupInlineAsmIdentifier(SS, TemplateKWLoc, Id,
246                                               IsUnevaluatedContext);
247  }
248  // While the next two tokens are 'period' 'identifier', repeatedly parse it as
249  // a field access. We have to avoid consuming assembler directives that look
250  // like '.' 'else'.
251  while (Result.isUsable() && Tok.is(tok::period)) {
252    Token IdTok = PP.LookAhead(0);
253    if (IdTok.isNot(tok::identifier))
254      break;
255    ConsumeToken(); // Consume the period.
256    IdentifierInfo *Id = Tok.getIdentifierInfo();
257    ConsumeToken(); // Consume the identifier.
258    Result = Actions.LookupInlineAsmVarDeclField(Result.get(), Id->getName(),
259                                                 Tok.getLocation());
260  }
261
262  // Figure out how many tokens we are into LineToks.
263  unsigned LineIndex = 0;
264  if (Tok.is(EndOfStream)) {
265    LineIndex = LineToks.size() - 2;
266  } else {
267    while (LineToks[LineIndex].getLocation() != Tok.getLocation()) {
268      LineIndex++;
269      assert(LineIndex < LineToks.size() - 2); // we added two extra tokens
270    }
271  }
272
273  // If we've run into the poison token we inserted before, or there
274  // was a parsing error, then claim the entire line.
275  if (Invalid || Tok.is(EndOfStream)) {
276    NumLineToksConsumed = LineToks.size() - 2;
277  } else {
278    // Otherwise, claim up to the start of the next token.
279    NumLineToksConsumed = LineIndex;
280  }
281
282  // Finally, restore the old parsing state by consuming all the tokens we
283  // staged before, implicitly killing off the token-lexer we pushed.
284  for (unsigned i = 0, e = LineToks.size() - LineIndex - 2; i != e; ++i) {
285    ConsumeAnyToken();
286  }
287  assert(Tok.is(EndOfStream));
288  ConsumeToken();
289
290  // Leave LineToks in its original state.
291  LineToks.pop_back();
292  LineToks.pop_back();
293
294  return Result;
295}
296
297/// Turn a sequence of our tokens back into a string that we can hand
298/// to the MC asm parser.
299static bool buildMSAsmString(Preprocessor &PP, SourceLocation AsmLoc,
300                             ArrayRef<Token> AsmToks,
301                             SmallVectorImpl<unsigned> &TokOffsets,
302                             SmallString<512> &Asm) {
303  assert(!AsmToks.empty() && "Didn't expect an empty AsmToks!");
304
305  // Is this the start of a new assembly statement?
306  bool isNewStatement = true;
307
308  for (unsigned i = 0, e = AsmToks.size(); i < e; ++i) {
309    const Token &Tok = AsmToks[i];
310
311    // Start each new statement with a newline and a tab.
312    if (!isNewStatement && (Tok.is(tok::kw_asm) || Tok.isAtStartOfLine())) {
313      Asm += "\n\t";
314      isNewStatement = true;
315    }
316
317    // Preserve the existence of leading whitespace except at the
318    // start of a statement.
319    if (!isNewStatement && Tok.hasLeadingSpace())
320      Asm += ' ';
321
322    // Remember the offset of this token.
323    TokOffsets.push_back(Asm.size());
324
325    // Don't actually write '__asm' into the assembly stream.
326    if (Tok.is(tok::kw_asm)) {
327      // Complain about __asm at the end of the stream.
328      if (i + 1 == e) {
329        PP.Diag(AsmLoc, diag::err_asm_empty);
330        return true;
331      }
332
333      continue;
334    }
335
336    // Append the spelling of the token.
337    SmallString<32> SpellingBuffer;
338    bool SpellingInvalid = false;
339    Asm += PP.getSpelling(Tok, SpellingBuffer, &SpellingInvalid);
340    assert(!SpellingInvalid && "spelling was invalid after correct parse?");
341
342    // We are no longer at the start of a statement.
343    isNewStatement = false;
344  }
345
346  // Ensure that the buffer is null-terminated.
347  Asm.push_back('\0');
348  Asm.pop_back();
349
350  assert(TokOffsets.size() == AsmToks.size());
351  return false;
352}
353
354// Determine if this is a GCC-style asm statement.
355bool Parser::isGCCAsmStatement(const Token &TokAfterAsm) const {
356  return TokAfterAsm.is(tok::l_paren) || isGNUAsmQualifier(TokAfterAsm);
357}
358
359bool Parser::isGNUAsmQualifier(const Token &TokAfterAsm) const {
360  return getGNUAsmQualifier(TokAfterAsm) != GNUAsmQualifiers::AQ_unspecified;
361}
362
363/// ParseMicrosoftAsmStatement. When -fms-extensions/-fasm-blocks is enabled,
364/// this routine is called to collect the tokens for an MS asm statement.
365///
366/// [MS]  ms-asm-statement:
367///         ms-asm-block
368///         ms-asm-block ms-asm-statement
369///
370/// [MS]  ms-asm-block:
371///         '__asm' ms-asm-line '\n'
372///         '__asm' '{' ms-asm-instruction-block[opt] '}' ';'[opt]
373///
374/// [MS]  ms-asm-instruction-block
375///         ms-asm-line
376///         ms-asm-line '\n' ms-asm-instruction-block
377///
378StmtResult Parser::ParseMicrosoftAsmStatement(SourceLocation AsmLoc) {
379  SourceManager &SrcMgr = PP.getSourceManager();
380  SourceLocation EndLoc = AsmLoc;
381  SmallVector<Token, 4> AsmToks;
382
383  bool SingleLineMode = true;
384  unsigned BraceNesting = 0;
385  unsigned short savedBraceCount = BraceCount;
386  bool InAsmComment = false;
387  FileID FID;
388  unsigned LineNo = 0;
389  unsigned NumTokensRead = 0;
390  SmallVector<SourceLocation, 4> LBraceLocs;
391  bool SkippedStartOfLine = false;
392
393  if (Tok.is(tok::l_brace)) {
394    // Braced inline asm: consume the opening brace.
395    SingleLineMode = false;
396    BraceNesting = 1;
397    EndLoc = ConsumeBrace();
398    LBraceLocs.push_back(EndLoc);
399    ++NumTokensRead;
400  } else {
401    // Single-line inline asm; compute which line it is on.
402    std::pair<FileID, unsigned> ExpAsmLoc =
403        SrcMgr.getDecomposedExpansionLoc(EndLoc);
404    FID = ExpAsmLoc.first;
405    LineNo = SrcMgr.getLineNumber(FID, ExpAsmLoc.second);
406    LBraceLocs.push_back(SourceLocation());
407  }
408
409  SourceLocation TokLoc = Tok.getLocation();
410  do {
411    // If we hit EOF, we're done, period.
412    if (isEofOrEom())
413      break;
414
415    if (!InAsmComment && Tok.is(tok::l_brace)) {
416      // Consume the opening brace.
417      SkippedStartOfLine = Tok.isAtStartOfLine();
418      AsmToks.push_back(Tok);
419      EndLoc = ConsumeBrace();
420      BraceNesting++;
421      LBraceLocs.push_back(EndLoc);
422      TokLoc = Tok.getLocation();
423      ++NumTokensRead;
424      continue;
425    } else if (!InAsmComment && Tok.is(tok::semi)) {
426      // A semicolon in an asm is the start of a comment.
427      InAsmComment = true;
428      if (!SingleLineMode) {
429        // Compute which line the comment is on.
430        std::pair<FileID, unsigned> ExpSemiLoc =
431            SrcMgr.getDecomposedExpansionLoc(TokLoc);
432        FID = ExpSemiLoc.first;
433        LineNo = SrcMgr.getLineNumber(FID, ExpSemiLoc.second);
434      }
435    } else if (SingleLineMode || InAsmComment) {
436      // If end-of-line is significant, check whether this token is on a
437      // new line.
438      std::pair<FileID, unsigned> ExpLoc =
439          SrcMgr.getDecomposedExpansionLoc(TokLoc);
440      if (ExpLoc.first != FID ||
441          SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second) != LineNo) {
442        // If this is a single-line __asm, we're done, except if the next
443        // line is MS-style asm too, in which case we finish a comment
444        // if needed and then keep processing the next line as a single
445        // line __asm.
446        bool isAsm = Tok.is(tok::kw_asm);
447        if (SingleLineMode && (!isAsm || isGCCAsmStatement(NextToken())))
448          break;
449        // We're no longer in a comment.
450        InAsmComment = false;
451        if (isAsm) {
452          // If this is a new __asm {} block we want to process it separately
453          // from the single-line __asm statements
454          if (PP.LookAhead(0).is(tok::l_brace))
455            break;
456          LineNo = SrcMgr.getLineNumber(ExpLoc.first, ExpLoc.second);
457          SkippedStartOfLine = Tok.isAtStartOfLine();
458        } else if (Tok.is(tok::semi)) {
459          // A multi-line asm-statement, where next line is a comment
460          InAsmComment = true;
461          FID = ExpLoc.first;
462          LineNo = SrcMgr.getLineNumber(FID, ExpLoc.second);
463        }
464      } else if (!InAsmComment && Tok.is(tok::r_brace)) {
465        // In MSVC mode, braces only participate in brace matching and
466        // separating the asm statements.  This is an intentional
467        // departure from the Apple gcc behavior.
468        if (!BraceNesting)
469          break;
470      }
471    }
472    if (!InAsmComment && BraceNesting && Tok.is(tok::r_brace) &&
473        BraceCount == (savedBraceCount + BraceNesting)) {
474      // Consume the closing brace.
475      SkippedStartOfLine = Tok.isAtStartOfLine();
476      // Don't want to add the closing brace of the whole asm block
477      if (SingleLineMode || BraceNesting > 1) {
478        Tok.clearFlag(Token::LeadingSpace);
479        AsmToks.push_back(Tok);
480      }
481      EndLoc = ConsumeBrace();
482      BraceNesting--;
483      // Finish if all of the opened braces in the inline asm section were
484      // consumed.
485      if (BraceNesting == 0 && !SingleLineMode)
486        break;
487      else {
488        LBraceLocs.pop_back();
489        TokLoc = Tok.getLocation();
490        ++NumTokensRead;
491        continue;
492      }
493    }
494
495    // Consume the next token; make sure we don't modify the brace count etc.
496    // if we are in a comment.
497    EndLoc = TokLoc;
498    if (InAsmComment)
499      PP.Lex(Tok);
500    else {
501      // Set the token as the start of line if we skipped the original start
502      // of line token in case it was a nested brace.
503      if (SkippedStartOfLine)
504        Tok.setFlag(Token::StartOfLine);
505      AsmToks.push_back(Tok);
506      ConsumeAnyToken();
507    }
508    TokLoc = Tok.getLocation();
509    ++NumTokensRead;
510    SkippedStartOfLine = false;
511  } while (1);
512
513  if (BraceNesting && BraceCount != savedBraceCount) {
514    // __asm without closing brace (this can happen at EOF).
515    for (unsigned i = 0; i < BraceNesting; ++i) {
516      Diag(Tok, diag::err_expected) << tok::r_brace;
517      Diag(LBraceLocs.back(), diag::note_matching) << tok::l_brace;
518      LBraceLocs.pop_back();
519    }
520    return StmtError();
521  } else if (NumTokensRead == 0) {
522    // Empty __asm.
523    Diag(Tok, diag::err_expected) << tok::l_brace;
524    return StmtError();
525  }
526
527  // Okay, prepare to use MC to parse the assembly.
528  SmallVector<StringRef, 4> ConstraintRefs;
529  SmallVector<Expr *, 4> Exprs;
530  SmallVector<StringRef, 4> ClobberRefs;
531
532  // We need an actual supported target.
533  const llvm::Triple &TheTriple = Actions.Context.getTargetInfo().getTriple();
534  const std::string &TT = TheTriple.getTriple();
535  const llvm::Target *TheTarget = nullptr;
536  if (!TheTriple.isX86()) {
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  SmallString<512> AsmString;
548  auto EmptyStmt = [&] {
549    return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmString,
550                                  /*NumOutputs*/ 0, /*NumInputs*/ 0,
551                                  ConstraintRefs, ClobberRefs, Exprs, EndLoc);
552  };
553  // If we don't support assembly, or the assembly is empty, we don't
554  // need to instantiate the AsmParser, etc.
555  if (!TheTarget || AsmToks.empty()) {
556    return EmptyStmt();
557  }
558
559  // Expand the tokens into a string buffer.
560  SmallVector<unsigned, 8> TokOffsets;
561  if (buildMSAsmString(PP, AsmLoc, AsmToks, TokOffsets, AsmString))
562    return StmtError();
563
564  const TargetOptions &TO = Actions.Context.getTargetInfo().getTargetOpts();
565  std::string FeaturesStr =
566      llvm::join(TO.Features.begin(), TO.Features.end(), ",");
567
568  std::unique_ptr<llvm::MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TT));
569  if (!MRI) {
570    Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
571        << "target MC unavailable";
572    return EmptyStmt();
573  }
574  // FIXME: init MCOptions from sanitizer flags here.
575  llvm::MCTargetOptions MCOptions;
576  std::unique_ptr<llvm::MCAsmInfo> MAI(
577      TheTarget->createMCAsmInfo(*MRI, TT, MCOptions));
578  // Get the instruction descriptor.
579  std::unique_ptr<llvm::MCInstrInfo> MII(TheTarget->createMCInstrInfo());
580  std::unique_ptr<llvm::MCObjectFileInfo> MOFI(new llvm::MCObjectFileInfo());
581  std::unique_ptr<llvm::MCSubtargetInfo> STI(
582      TheTarget->createMCSubtargetInfo(TT, TO.CPU, FeaturesStr));
583  // Target MCTargetDesc may not be linked in clang-based tools.
584  if (!MAI || !MII | !MOFI || !STI) {
585    Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
586        << "target MC unavailable";
587    return EmptyStmt();
588  }
589
590  llvm::SourceMgr TempSrcMgr;
591  llvm::MCContext Ctx(MAI.get(), MRI.get(), MOFI.get(), &TempSrcMgr);
592  MOFI->InitMCObjectFileInfo(TheTriple, /*PIC*/ false, Ctx);
593  std::unique_ptr<llvm::MemoryBuffer> Buffer =
594      llvm::MemoryBuffer::getMemBuffer(AsmString, "<MS inline asm>");
595
596  // Tell SrcMgr about this buffer, which is what the parser will pick up.
597  TempSrcMgr.AddNewSourceBuffer(std::move(Buffer), llvm::SMLoc());
598
599  std::unique_ptr<llvm::MCStreamer> Str(createNullStreamer(Ctx));
600  std::unique_ptr<llvm::MCAsmParser> Parser(
601      createMCAsmParser(TempSrcMgr, Ctx, *Str.get(), *MAI));
602
603  std::unique_ptr<llvm::MCTargetAsmParser> TargetParser(
604      TheTarget->createMCAsmParser(*STI, *Parser, *MII, MCOptions));
605  // Target AsmParser may not be linked in clang-based tools.
606  if (!TargetParser) {
607    Diag(AsmLoc, diag::err_msasm_unable_to_create_target)
608        << "target ASM parser unavailable";
609    return EmptyStmt();
610  }
611
612  std::unique_ptr<llvm::MCInstPrinter> IP(
613      TheTarget->createMCInstPrinter(llvm::Triple(TT), 1, *MAI, *MII, *MRI));
614
615  // Change to the Intel dialect.
616  Parser->setAssemblerDialect(1);
617  Parser->setTargetParser(*TargetParser.get());
618  Parser->setParsingMSInlineAsm(true);
619  TargetParser->setParsingMSInlineAsm(true);
620
621  ClangAsmParserCallback Callback(*this, AsmLoc, AsmString, AsmToks,
622                                  TokOffsets);
623  TargetParser->setSemaCallback(&Callback);
624  TempSrcMgr.setDiagHandler(ClangAsmParserCallback::DiagHandlerCallback,
625                            &Callback);
626
627  unsigned NumOutputs;
628  unsigned NumInputs;
629  std::string AsmStringIR;
630  SmallVector<std::pair<void *, bool>, 4> OpExprs;
631  SmallVector<std::string, 4> Constraints;
632  SmallVector<std::string, 4> Clobbers;
633  if (Parser->parseMSInlineAsm(AsmLoc.getPtrEncoding(), AsmStringIR, NumOutputs,
634                               NumInputs, OpExprs, Constraints, Clobbers,
635                               MII.get(), IP.get(), Callback))
636    return StmtError();
637
638  // Filter out "fpsw" and "mxcsr". They aren't valid GCC asm clobber
639  // constraints. Clang always adds fpsr to the clobber list anyway.
640  llvm::erase_if(Clobbers, [](const std::string &C) {
641    return C == "fpsr" || C == "mxcsr";
642  });
643
644  // Build the vector of clobber StringRefs.
645  ClobberRefs.insert(ClobberRefs.end(), Clobbers.begin(), Clobbers.end());
646
647  // Recast the void pointers and build the vector of constraint StringRefs.
648  unsigned NumExprs = NumOutputs + NumInputs;
649  ConstraintRefs.resize(NumExprs);
650  Exprs.resize(NumExprs);
651  for (unsigned i = 0, e = NumExprs; i != e; ++i) {
652    Expr *OpExpr = static_cast<Expr *>(OpExprs[i].first);
653    if (!OpExpr)
654      return StmtError();
655
656    // Need address of variable.
657    if (OpExprs[i].second)
658      OpExpr =
659          Actions.BuildUnaryOp(getCurScope(), AsmLoc, UO_AddrOf, OpExpr).get();
660
661    ConstraintRefs[i] = StringRef(Constraints[i]);
662    Exprs[i] = OpExpr;
663  }
664
665  // FIXME: We should be passing source locations for better diagnostics.
666  return Actions.ActOnMSAsmStmt(AsmLoc, LBraceLocs[0], AsmToks, AsmStringIR,
667                                NumOutputs, NumInputs, ConstraintRefs,
668                                ClobberRefs, Exprs, EndLoc);
669}
670
671/// parseGNUAsmQualifierListOpt - Parse a GNU extended asm qualifier list.
672///       asm-qualifier:
673///         volatile
674///         inline
675///         goto
676///
677///       asm-qualifier-list:
678///         asm-qualifier
679///         asm-qualifier-list asm-qualifier
680bool Parser::parseGNUAsmQualifierListOpt(GNUAsmQualifiers &AQ) {
681  while (1) {
682    const GNUAsmQualifiers::AQ A = getGNUAsmQualifier(Tok);
683    if (A == GNUAsmQualifiers::AQ_unspecified) {
684      if (Tok.isNot(tok::l_paren)) {
685        Diag(Tok.getLocation(), diag::err_asm_qualifier_ignored);
686        SkipUntil(tok::r_paren, StopAtSemi);
687        return true;
688      }
689      return false;
690    }
691    if (AQ.setAsmQualifier(A))
692      Diag(Tok.getLocation(), diag::err_asm_duplicate_qual)
693          << GNUAsmQualifiers::getQualifierName(A);
694    ConsumeToken();
695  }
696  return false;
697}
698
699/// ParseAsmStatement - Parse a GNU extended asm statement.
700///       asm-statement:
701///         gnu-asm-statement
702///         ms-asm-statement
703///
704/// [GNU] gnu-asm-statement:
705///         'asm' asm-qualifier-list[opt] '(' asm-argument ')' ';'
706///
707/// [GNU] asm-argument:
708///         asm-string-literal
709///         asm-string-literal ':' asm-operands[opt]
710///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
711///         asm-string-literal ':' asm-operands[opt] ':' asm-operands[opt]
712///                 ':' asm-clobbers
713///
714/// [GNU] asm-clobbers:
715///         asm-string-literal
716///         asm-clobbers ',' asm-string-literal
717///
718StmtResult Parser::ParseAsmStatement(bool &msAsm) {
719  assert(Tok.is(tok::kw_asm) && "Not an asm stmt");
720  SourceLocation AsmLoc = ConsumeToken();
721
722  if (getLangOpts().AsmBlocks && !isGCCAsmStatement(Tok)) {
723    msAsm = true;
724    return ParseMicrosoftAsmStatement(AsmLoc);
725  }
726
727  SourceLocation Loc = Tok.getLocation();
728  GNUAsmQualifiers GAQ;
729  if (parseGNUAsmQualifierListOpt(GAQ))
730    return StmtError();
731
732  if (GAQ.isGoto() && getLangOpts().SpeculativeLoadHardening)
733    Diag(Loc, diag::warn_slh_does_not_support_asm_goto);
734
735  BalancedDelimiterTracker T(*this, tok::l_paren);
736  T.consumeOpen();
737
738  ExprResult AsmString(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
739
740  // Check if GNU-style InlineAsm is disabled.
741  // Error on anything other than empty string.
742  if (!(getLangOpts().GNUAsm || AsmString.isInvalid())) {
743    const auto *SL = cast<StringLiteral>(AsmString.get());
744    if (!SL->getString().trim().empty())
745      Diag(Loc, diag::err_gnu_inline_asm_disabled);
746  }
747
748  if (AsmString.isInvalid()) {
749    // Consume up to and including the closing paren.
750    T.skipToEnd();
751    return StmtError();
752  }
753
754  SmallVector<IdentifierInfo *, 4> Names;
755  ExprVector Constraints;
756  ExprVector Exprs;
757  ExprVector Clobbers;
758
759  if (Tok.is(tok::r_paren)) {
760    // We have a simple asm expression like 'asm("foo")'.
761    T.consumeClose();
762    return Actions.ActOnGCCAsmStmt(
763        AsmLoc, /*isSimple*/ true, GAQ.isVolatile(),
764        /*NumOutputs*/ 0, /*NumInputs*/ 0, nullptr, Constraints, Exprs,
765        AsmString.get(), Clobbers, /*NumLabels*/ 0, T.getCloseLocation());
766  }
767
768  // Parse Outputs, if present.
769  bool AteExtraColon = false;
770  if (Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
771    // In C++ mode, parse "::" like ": :".
772    AteExtraColon = Tok.is(tok::coloncolon);
773    ConsumeToken();
774
775    if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
776      return StmtError();
777  }
778
779  unsigned NumOutputs = Names.size();
780
781  // Parse Inputs, if present.
782  if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
783    // In C++ mode, parse "::" like ": :".
784    if (AteExtraColon)
785      AteExtraColon = false;
786    else {
787      AteExtraColon = Tok.is(tok::coloncolon);
788      ConsumeToken();
789    }
790
791    if (!AteExtraColon && ParseAsmOperandsOpt(Names, Constraints, Exprs))
792      return StmtError();
793  }
794
795  assert(Names.size() == Constraints.size() &&
796         Constraints.size() == Exprs.size() && "Input operand size mismatch!");
797
798  unsigned NumInputs = Names.size() - NumOutputs;
799
800  // Parse the clobbers, if present.
801  if (AteExtraColon || Tok.is(tok::colon) || Tok.is(tok::coloncolon)) {
802    if (AteExtraColon)
803      AteExtraColon = false;
804    else {
805      AteExtraColon = Tok.is(tok::coloncolon);
806      ConsumeToken();
807    }
808    // Parse the asm-string list for clobbers if present.
809    if (!AteExtraColon && isTokenStringLiteral()) {
810      while (1) {
811        ExprResult Clobber(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
812
813        if (Clobber.isInvalid())
814          break;
815
816        Clobbers.push_back(Clobber.get());
817
818        if (!TryConsumeToken(tok::comma))
819          break;
820      }
821    }
822  }
823  if (!GAQ.isGoto() && (Tok.isNot(tok::r_paren) || AteExtraColon)) {
824    Diag(Tok, diag::err_expected) << tok::r_paren;
825    SkipUntil(tok::r_paren, StopAtSemi);
826    return StmtError();
827  }
828
829  // Parse the goto label, if present.
830  unsigned NumLabels = 0;
831  if (AteExtraColon || Tok.is(tok::colon)) {
832    if (!AteExtraColon)
833      ConsumeToken();
834
835    while (true) {
836      if (Tok.isNot(tok::identifier)) {
837        Diag(Tok, diag::err_expected) << tok::identifier;
838        SkipUntil(tok::r_paren, StopAtSemi);
839        return StmtError();
840      }
841      LabelDecl *LD = Actions.LookupOrCreateLabel(Tok.getIdentifierInfo(),
842                                                  Tok.getLocation());
843      Names.push_back(Tok.getIdentifierInfo());
844      if (!LD) {
845        SkipUntil(tok::r_paren, StopAtSemi);
846        return StmtError();
847      }
848      ExprResult Res =
849          Actions.ActOnAddrLabel(Tok.getLocation(), Tok.getLocation(), LD);
850      Exprs.push_back(Res.get());
851      NumLabels++;
852      ConsumeToken();
853      if (!TryConsumeToken(tok::comma))
854        break;
855    }
856  } else if (GAQ.isGoto()) {
857    Diag(Tok, diag::err_expected) << tok::colon;
858    SkipUntil(tok::r_paren, StopAtSemi);
859    return StmtError();
860  }
861  T.consumeClose();
862  return Actions.ActOnGCCAsmStmt(AsmLoc, false, GAQ.isVolatile(), NumOutputs,
863                                 NumInputs, Names.data(), Constraints, Exprs,
864                                 AsmString.get(), Clobbers, NumLabels,
865                                 T.getCloseLocation());
866}
867
868/// ParseAsmOperands - Parse the asm-operands production as used by
869/// asm-statement, assuming the leading ':' token was eaten.
870///
871/// [GNU] asm-operands:
872///         asm-operand
873///         asm-operands ',' asm-operand
874///
875/// [GNU] asm-operand:
876///         asm-string-literal '(' expression ')'
877///         '[' identifier ']' asm-string-literal '(' expression ')'
878///
879//
880// FIXME: Avoid unnecessary std::string trashing.
881bool Parser::ParseAsmOperandsOpt(SmallVectorImpl<IdentifierInfo *> &Names,
882                                 SmallVectorImpl<Expr *> &Constraints,
883                                 SmallVectorImpl<Expr *> &Exprs) {
884  // 'asm-operands' isn't present?
885  if (!isTokenStringLiteral() && Tok.isNot(tok::l_square))
886    return false;
887
888  while (1) {
889    // Read the [id] if present.
890    if (Tok.is(tok::l_square)) {
891      BalancedDelimiterTracker T(*this, tok::l_square);
892      T.consumeOpen();
893
894      if (Tok.isNot(tok::identifier)) {
895        Diag(Tok, diag::err_expected) << tok::identifier;
896        SkipUntil(tok::r_paren, StopAtSemi);
897        return true;
898      }
899
900      IdentifierInfo *II = Tok.getIdentifierInfo();
901      ConsumeToken();
902
903      Names.push_back(II);
904      T.consumeClose();
905    } else
906      Names.push_back(nullptr);
907
908    ExprResult Constraint(ParseAsmStringLiteral(/*ForAsmLabel*/ false));
909    if (Constraint.isInvalid()) {
910      SkipUntil(tok::r_paren, StopAtSemi);
911      return true;
912    }
913    Constraints.push_back(Constraint.get());
914
915    if (Tok.isNot(tok::l_paren)) {
916      Diag(Tok, diag::err_expected_lparen_after) << "asm operand";
917      SkipUntil(tok::r_paren, StopAtSemi);
918      return true;
919    }
920
921    // Read the parenthesized expression.
922    BalancedDelimiterTracker T(*this, tok::l_paren);
923    T.consumeOpen();
924    ExprResult Res = Actions.CorrectDelayedTyposInExpr(ParseExpression());
925    T.consumeClose();
926    if (Res.isInvalid()) {
927      SkipUntil(tok::r_paren, StopAtSemi);
928      return true;
929    }
930    Exprs.push_back(Res.get());
931    // Eat the comma and continue parsing if it exists.
932    if (!TryConsumeToken(tok::comma))
933      return false;
934  }
935}
936
937const char *Parser::GNUAsmQualifiers::getQualifierName(AQ Qualifier) {
938  switch (Qualifier) {
939    case AQ_volatile: return "volatile";
940    case AQ_inline: return "inline";
941    case AQ_goto: return "goto";
942    case AQ_unspecified: return "unspecified";
943  }
944  llvm_unreachable("Unknown GNUAsmQualifier");
945}
946
947Parser::GNUAsmQualifiers::AQ
948Parser::getGNUAsmQualifier(const Token &Tok) const {
949  switch (Tok.getKind()) {
950    case tok::kw_volatile: return GNUAsmQualifiers::AQ_volatile;
951    case tok::kw_inline: return GNUAsmQualifiers::AQ_inline;
952    case tok::kw_goto: return GNUAsmQualifiers::AQ_goto;
953    default: return GNUAsmQualifiers::AQ_unspecified;
954  }
955}
956bool Parser::GNUAsmQualifiers::setAsmQualifier(AQ Qualifier) {
957  bool IsDuplicate = Qualifiers & Qualifier;
958  Qualifiers |= Qualifier;
959  return IsDuplicate;
960}
961