blob: de485b9f95b76c90a965fd2beb85c04e57bba2b0 [file] [log] [blame]
Chad Rosier3d45a772012-08-17 21:27:25 +00001//===--- SemaStmtAsm.cpp - Semantic Analysis for Asm Statements -----------===//
Chad Rosier4b5e48d2012-08-17 21:19:40 +00002//
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 semantic analysis for inline asm statements.
11//
12//===----------------------------------------------------------------------===//
13
14#include "clang/Sema/SemaInternal.h"
15#include "clang/Sema/Scope.h"
16#include "clang/Sema/ScopeInfo.h"
17#include "clang/Sema/Initialization.h"
18#include "clang/Sema/Lookup.h"
19#include "clang/AST/TypeLoc.h"
20#include "clang/Lex/Preprocessor.h"
21#include "clang/Basic/TargetInfo.h"
22#include "llvm/ADT/ArrayRef.h"
23#include "llvm/ADT/BitVector.h"
24#include "llvm/ADT/SmallString.h"
25#include "llvm/MC/MCAsmInfo.h"
26#include "llvm/MC/MCContext.h"
Chad Rosier6e97be72012-08-22 23:42:09 +000027#include "llvm/MC/MCExpr.h"
Chad Rosier4b5e48d2012-08-17 21:19:40 +000028#include "llvm/MC/MCInst.h"
29#include "llvm/MC/MCInstPrinter.h"
30#include "llvm/MC/MCInstrInfo.h"
31#include "llvm/MC/MCObjectFileInfo.h"
32#include "llvm/MC/MCRegisterInfo.h"
33#include "llvm/MC/MCStreamer.h"
34#include "llvm/MC/MCSubtargetInfo.h"
Chad Rosier6e97be72012-08-22 23:42:09 +000035#include "llvm/MC/MCSymbol.h"
Chad Rosier4b5e48d2012-08-17 21:19:40 +000036#include "llvm/MC/MCTargetAsmParser.h"
37#include "llvm/MC/MCParser/MCAsmLexer.h"
38#include "llvm/MC/MCParser/MCAsmParser.h"
Chad Rosierfb700262012-09-11 23:13:15 +000039#include "llvm/MC/MCParser/MCParsedAsmOperand.h"
Chad Rosier4b5e48d2012-08-17 21:19:40 +000040#include "llvm/Support/SourceMgr.h"
41#include "llvm/Support/TargetRegistry.h"
42#include "llvm/Support/TargetSelect.h"
43using namespace clang;
44using namespace sema;
45
46/// CheckAsmLValue - GNU C has an extremely ugly extension whereby they silently
47/// ignore "noop" casts in places where an lvalue is required by an inline asm.
48/// We emulate this behavior when -fheinous-gnu-extensions is specified, but
49/// provide a strong guidance to not use it.
50///
51/// This method checks to see if the argument is an acceptable l-value and
52/// returns false if it is a case we can handle.
53static bool CheckAsmLValue(const Expr *E, Sema &S) {
54 // Type dependent expressions will be checked during instantiation.
55 if (E->isTypeDependent())
56 return false;
57
58 if (E->isLValue())
59 return false; // Cool, this is an lvalue.
60
61 // Okay, this is not an lvalue, but perhaps it is the result of a cast that we
62 // are supposed to allow.
63 const Expr *E2 = E->IgnoreParenNoopCasts(S.Context);
64 if (E != E2 && E2->isLValue()) {
65 if (!S.getLangOpts().HeinousExtensions)
66 S.Diag(E2->getLocStart(), diag::err_invalid_asm_cast_lvalue)
67 << E->getSourceRange();
68 else
69 S.Diag(E2->getLocStart(), diag::warn_invalid_asm_cast_lvalue)
70 << E->getSourceRange();
71 // Accept, even if we emitted an error diagnostic.
72 return false;
73 }
74
75 // None of the above, just randomly invalid non-lvalue.
76 return true;
77}
78
79/// isOperandMentioned - Return true if the specified operand # is mentioned
80/// anywhere in the decomposed asm string.
81static bool isOperandMentioned(unsigned OpNo,
Chad Rosierdf5faf52012-08-25 00:11:56 +000082 ArrayRef<GCCAsmStmt::AsmStringPiece> AsmStrPieces) {
Chad Rosier4b5e48d2012-08-17 21:19:40 +000083 for (unsigned p = 0, e = AsmStrPieces.size(); p != e; ++p) {
Chad Rosierdf5faf52012-08-25 00:11:56 +000084 const GCCAsmStmt::AsmStringPiece &Piece = AsmStrPieces[p];
Chad Rosier4b5e48d2012-08-17 21:19:40 +000085 if (!Piece.isOperand()) continue;
86
87 // If this is a reference to the input and if the input was the smaller
88 // one, then we have to reject this asm.
89 if (Piece.getOperandNo() == OpNo)
90 return true;
91 }
92 return false;
93}
94
Chad Rosierdf5faf52012-08-25 00:11:56 +000095StmtResult Sema::ActOnGCCAsmStmt(SourceLocation AsmLoc, bool IsSimple,
96 bool IsVolatile, unsigned NumOutputs,
97 unsigned NumInputs, IdentifierInfo **Names,
98 MultiExprArg constraints, MultiExprArg exprs,
99 Expr *asmString, MultiExprArg clobbers,
100 SourceLocation RParenLoc) {
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000101 unsigned NumClobbers = clobbers.size();
102 StringLiteral **Constraints =
Benjamin Kramer5354e772012-08-23 23:38:35 +0000103 reinterpret_cast<StringLiteral**>(constraints.data());
104 Expr **Exprs = exprs.data();
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000105 StringLiteral *AsmString = cast<StringLiteral>(asmString);
Benjamin Kramer5354e772012-08-23 23:38:35 +0000106 StringLiteral **Clobbers = reinterpret_cast<StringLiteral**>(clobbers.data());
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000107
108 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
109
110 // The parser verifies that there is a string literal here.
111 if (!AsmString->isAscii())
112 return StmtError(Diag(AsmString->getLocStart(),diag::err_asm_wide_character)
113 << AsmString->getSourceRange());
114
115 for (unsigned i = 0; i != NumOutputs; i++) {
116 StringLiteral *Literal = Constraints[i];
117 if (!Literal->isAscii())
118 return StmtError(Diag(Literal->getLocStart(),diag::err_asm_wide_character)
119 << Literal->getSourceRange());
120
121 StringRef OutputName;
122 if (Names[i])
123 OutputName = Names[i]->getName();
124
125 TargetInfo::ConstraintInfo Info(Literal->getString(), OutputName);
126 if (!Context.getTargetInfo().validateOutputConstraint(Info))
127 return StmtError(Diag(Literal->getLocStart(),
128 diag::err_asm_invalid_output_constraint)
129 << Info.getConstraintStr());
130
131 // Check that the output exprs are valid lvalues.
132 Expr *OutputExpr = Exprs[i];
133 if (CheckAsmLValue(OutputExpr, *this)) {
134 return StmtError(Diag(OutputExpr->getLocStart(),
135 diag::err_asm_invalid_lvalue_in_output)
136 << OutputExpr->getSourceRange());
137 }
138
139 OutputConstraintInfos.push_back(Info);
140 }
141
142 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
143
144 for (unsigned i = NumOutputs, e = NumOutputs + NumInputs; i != e; i++) {
145 StringLiteral *Literal = Constraints[i];
146 if (!Literal->isAscii())
147 return StmtError(Diag(Literal->getLocStart(),diag::err_asm_wide_character)
148 << Literal->getSourceRange());
149
150 StringRef InputName;
151 if (Names[i])
152 InputName = Names[i]->getName();
153
154 TargetInfo::ConstraintInfo Info(Literal->getString(), InputName);
155 if (!Context.getTargetInfo().validateInputConstraint(OutputConstraintInfos.data(),
156 NumOutputs, Info)) {
157 return StmtError(Diag(Literal->getLocStart(),
158 diag::err_asm_invalid_input_constraint)
159 << Info.getConstraintStr());
160 }
161
162 Expr *InputExpr = Exprs[i];
163
164 // Only allow void types for memory constraints.
165 if (Info.allowsMemory() && !Info.allowsRegister()) {
166 if (CheckAsmLValue(InputExpr, *this))
167 return StmtError(Diag(InputExpr->getLocStart(),
168 diag::err_asm_invalid_lvalue_in_input)
169 << Info.getConstraintStr()
170 << InputExpr->getSourceRange());
171 }
172
173 if (Info.allowsRegister()) {
174 if (InputExpr->getType()->isVoidType()) {
175 return StmtError(Diag(InputExpr->getLocStart(),
176 diag::err_asm_invalid_type_in_input)
177 << InputExpr->getType() << Info.getConstraintStr()
178 << InputExpr->getSourceRange());
179 }
180 }
181
182 ExprResult Result = DefaultFunctionArrayLvalueConversion(Exprs[i]);
183 if (Result.isInvalid())
184 return StmtError();
185
186 Exprs[i] = Result.take();
187 InputConstraintInfos.push_back(Info);
188 }
189
190 // Check that the clobbers are valid.
191 for (unsigned i = 0; i != NumClobbers; i++) {
192 StringLiteral *Literal = Clobbers[i];
193 if (!Literal->isAscii())
194 return StmtError(Diag(Literal->getLocStart(),diag::err_asm_wide_character)
195 << Literal->getSourceRange());
196
197 StringRef Clobber = Literal->getString();
198
199 if (!Context.getTargetInfo().isValidClobber(Clobber))
200 return StmtError(Diag(Literal->getLocStart(),
201 diag::err_asm_unknown_register_name) << Clobber);
202 }
203
Chad Rosierdf5faf52012-08-25 00:11:56 +0000204 GCCAsmStmt *NS =
205 new (Context) GCCAsmStmt(Context, AsmLoc, IsSimple, IsVolatile, NumOutputs,
206 NumInputs, Names, Constraints, Exprs, AsmString,
207 NumClobbers, Clobbers, RParenLoc);
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000208 // Validate the asm string, ensuring it makes sense given the operands we
209 // have.
Chad Rosierdf5faf52012-08-25 00:11:56 +0000210 SmallVector<GCCAsmStmt::AsmStringPiece, 8> Pieces;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000211 unsigned DiagOffs;
212 if (unsigned DiagID = NS->AnalyzeAsmString(Pieces, Context, DiagOffs)) {
213 Diag(getLocationOfStringLiteralByte(AsmString, DiagOffs), DiagID)
214 << AsmString->getSourceRange();
215 return StmtError();
216 }
217
218 // Validate tied input operands for type mismatches.
219 for (unsigned i = 0, e = InputConstraintInfos.size(); i != e; ++i) {
220 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
221
222 // If this is a tied constraint, verify that the output and input have
223 // either exactly the same type, or that they are int/ptr operands with the
224 // same size (int/long, int*/long, are ok etc).
225 if (!Info.hasTiedOperand()) continue;
226
227 unsigned TiedTo = Info.getTiedOperand();
228 unsigned InputOpNo = i+NumOutputs;
229 Expr *OutputExpr = Exprs[TiedTo];
230 Expr *InputExpr = Exprs[InputOpNo];
231
232 if (OutputExpr->isTypeDependent() || InputExpr->isTypeDependent())
233 continue;
234
235 QualType InTy = InputExpr->getType();
236 QualType OutTy = OutputExpr->getType();
237 if (Context.hasSameType(InTy, OutTy))
238 continue; // All types can be tied to themselves.
239
240 // Decide if the input and output are in the same domain (integer/ptr or
241 // floating point.
242 enum AsmDomain {
243 AD_Int, AD_FP, AD_Other
244 } InputDomain, OutputDomain;
245
246 if (InTy->isIntegerType() || InTy->isPointerType())
247 InputDomain = AD_Int;
248 else if (InTy->isRealFloatingType())
249 InputDomain = AD_FP;
250 else
251 InputDomain = AD_Other;
252
253 if (OutTy->isIntegerType() || OutTy->isPointerType())
254 OutputDomain = AD_Int;
255 else if (OutTy->isRealFloatingType())
256 OutputDomain = AD_FP;
257 else
258 OutputDomain = AD_Other;
259
260 // They are ok if they are the same size and in the same domain. This
261 // allows tying things like:
262 // void* to int*
263 // void* to int if they are the same size.
264 // double to long double if they are the same size.
265 //
266 uint64_t OutSize = Context.getTypeSize(OutTy);
267 uint64_t InSize = Context.getTypeSize(InTy);
268 if (OutSize == InSize && InputDomain == OutputDomain &&
269 InputDomain != AD_Other)
270 continue;
271
272 // If the smaller input/output operand is not mentioned in the asm string,
273 // then we can promote the smaller one to a larger input and the asm string
274 // won't notice.
275 bool SmallerValueMentioned = false;
276
277 // If this is a reference to the input and if the input was the smaller
278 // one, then we have to reject this asm.
279 if (isOperandMentioned(InputOpNo, Pieces)) {
280 // This is a use in the asm string of the smaller operand. Since we
281 // codegen this by promoting to a wider value, the asm will get printed
282 // "wrong".
283 SmallerValueMentioned |= InSize < OutSize;
284 }
285 if (isOperandMentioned(TiedTo, Pieces)) {
286 // If this is a reference to the output, and if the output is the larger
287 // value, then it's ok because we'll promote the input to the larger type.
288 SmallerValueMentioned |= OutSize < InSize;
289 }
290
291 // If the smaller value wasn't mentioned in the asm string, and if the
292 // output was a register, just extend the shorter one to the size of the
293 // larger one.
294 if (!SmallerValueMentioned && InputDomain != AD_Other &&
295 OutputConstraintInfos[TiedTo].allowsRegister())
296 continue;
297
298 // Either both of the operands were mentioned or the smaller one was
299 // mentioned. One more special case that we'll allow: if the tied input is
300 // integer, unmentioned, and is a constant, then we'll allow truncating it
301 // down to the size of the destination.
302 if (InputDomain == AD_Int && OutputDomain == AD_Int &&
303 !isOperandMentioned(InputOpNo, Pieces) &&
304 InputExpr->isEvaluatable(Context)) {
305 CastKind castKind =
306 (OutTy->isBooleanType() ? CK_IntegralToBoolean : CK_IntegralCast);
307 InputExpr = ImpCastExprToType(InputExpr, OutTy, castKind).take();
308 Exprs[InputOpNo] = InputExpr;
309 NS->setInputExpr(i, InputExpr);
310 continue;
311 }
312
313 Diag(InputExpr->getLocStart(),
314 diag::err_asm_tying_incompatible_types)
315 << InTy << OutTy << OutputExpr->getSourceRange()
316 << InputExpr->getSourceRange();
317 return StmtError();
318 }
319
320 return Owned(NS);
321}
322
323// isMSAsmKeyword - Return true if this is an MS-style inline asm keyword. These
324// require special handling.
325static bool isMSAsmKeyword(StringRef Name) {
326 bool Ret = llvm::StringSwitch<bool>(Name)
327 .Cases("EVEN", "ALIGN", true) // Alignment directives.
328 .Cases("LENGTH", "SIZE", "TYPE", true) // Type and variable sizes.
329 .Case("_emit", true) // _emit Pseudoinstruction.
330 .Default(false);
331 return Ret;
332}
333
Chad Rosier180d9d92012-10-02 18:51:05 +0000334// Check to see if the expression is a substring of the asm operand.
335static StringRef getMSInlineAsmExprName(StringRef Name) {
336 // Strip off the size directives.
337 // E.g., DWORD PTR [V] -> V
338 if (Name.startswith("BYTE") || Name.startswith("byte") ||
339 Name.startswith("WORD") || Name.startswith("word") ||
340 Name.startswith("DWORD") || Name.startswith("dword") ||
341 Name.startswith("QWORD") || Name.startswith("qword") ||
342 Name.startswith("XWORD") || Name.startswith("xword") ||
343 Name.startswith("XMMWORD") || Name.startswith("xmmword") ||
344 Name.startswith("YMMWORD") || Name.startswith("ymmword")) {
345 std::pair< StringRef, StringRef > SplitName = Name.split(' ');
346 assert((SplitName.second.startswith("PTR") ||
347 SplitName.second.startswith("ptr")) &&
348 "Expected PTR/ptr!");
349 SplitName = SplitName.second.split('[');
350 SplitName = SplitName.second.split(']');
351 return SplitName.first;
352 }
353 return Name;
354}
355
Chad Rosier6e97be72012-08-22 23:42:09 +0000356// getIdentifierInfo - Given a Name and a range of tokens, find the associated
357// IdentifierInfo*.
358static IdentifierInfo *getIdentifierInfo(StringRef Name,
359 ArrayRef<Token> AsmToks,
360 unsigned Begin, unsigned End) {
361 for (unsigned i = Begin; i <= End; ++i) {
362 IdentifierInfo *II = AsmToks[i].getIdentifierInfo();
363 if (II && II->getName() == Name)
364 return II;
365 }
366 return 0;
367}
368
Chad Rosier358ab762012-08-22 21:08:06 +0000369// getSpelling - Get the spelling of the AsmTok token.
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000370static StringRef getSpelling(Sema &SemaRef, Token AsmTok) {
371 StringRef Asm;
372 SmallString<512> TokenBuf;
373 TokenBuf.resize(512);
374 bool StringInvalid = false;
375 Asm = SemaRef.PP.getSpelling(AsmTok, TokenBuf, &StringInvalid);
376 assert (!StringInvalid && "Expected valid string!");
377 return Asm;
378}
379
Chad Rosier358ab762012-08-22 21:08:06 +0000380// Determine if we should bail on this MSAsm instruction.
Chad Rosier2735df22012-08-22 19:18:30 +0000381static bool bailOnMSAsm(std::vector<StringRef> Piece) {
382 for (unsigned i = 0, e = Piece.size(); i != e; ++i)
383 if (isMSAsmKeyword(Piece[i]))
384 return true;
385 return false;
386}
387
Chad Rosier358ab762012-08-22 21:08:06 +0000388// Determine if we should bail on this MSAsm block.
Chad Rosier2735df22012-08-22 19:18:30 +0000389static bool bailOnMSAsm(std::vector<std::vector<StringRef> > Pieces) {
390 for (unsigned i = 0, e = Pieces.size(); i != e; ++i)
391 if (bailOnMSAsm(Pieces[i]))
392 return true;
393 return false;
394}
395
Chad Rosier358ab762012-08-22 21:08:06 +0000396// Determine if this is a simple MSAsm instruction.
Chad Rosier98ac6082012-08-21 23:09:21 +0000397static bool isSimpleMSAsm(std::vector<StringRef> &Pieces,
398 const TargetInfo &TI) {
399 if (isMSAsmKeyword(Pieces[0]))
400 return false;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000401
Chad Rosier180d9d92012-10-02 18:51:05 +0000402 for (unsigned i = 1, e = Pieces.size(); i != e; ++i) {
403 StringRef Op = getMSInlineAsmExprName(Pieces[i]);
404 if (!TI.isValidGCCRegisterName(Op))
Chad Rosier98ac6082012-08-21 23:09:21 +0000405 return false;
Chad Rosier180d9d92012-10-02 18:51:05 +0000406 }
Chad Rosier153f8ec2012-08-22 19:50:28 +0000407 return true;
408}
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000409
Chad Rosier358ab762012-08-22 21:08:06 +0000410// Determine if this is a simple MSAsm block.
Chad Rosier153f8ec2012-08-22 19:50:28 +0000411static bool isSimpleMSAsm(std::vector<std::vector<StringRef> > Pieces,
412 const TargetInfo &TI) {
413 for (unsigned i = 0, e = Pieces.size(); i != e; ++i)
414 if (!isSimpleMSAsm(Pieces[i], TI))
415 return false;
Chad Rosier98ac6082012-08-21 23:09:21 +0000416 return true;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000417}
418
Bob Wilson40d39e32012-09-24 19:57:55 +0000419// Break the AsmString into pieces (i.e., mnemonic and operands).
Chad Rosier38c71d32012-08-21 21:56:39 +0000420static void buildMSAsmPieces(StringRef Asm, std::vector<StringRef> &Pieces) {
421 std::pair<StringRef,StringRef> Split = Asm.split(' ');
422
423 // Mnemonic
424 Pieces.push_back(Split.first);
425 Asm = Split.second;
426
427 // Operands
428 while (!Asm.empty()) {
429 Split = Asm.split(", ");
430 Pieces.push_back(Split.first);
431 Asm = Split.second;
432 }
433}
434
Chad Rosierf0fbd772012-08-22 21:04:07 +0000435static void buildMSAsmPieces(std::vector<std::string> &AsmStrings,
436 std::vector<std::vector<StringRef> > &Pieces) {
437 for (unsigned i = 0, e = AsmStrings.size(); i != e; ++i)
438 buildMSAsmPieces(AsmStrings[i], Pieces[i]);
439}
440
Chad Rosierb55e6022012-09-13 00:06:55 +0000441// Build the individual assembly instruction(s) and place them in the AsmStrings
Bob Wilsonb0f6b9c2012-09-24 19:57:59 +0000442// vector. These strings are fed to the AsmParser. Returns true on error.
443static bool buildMSAsmStrings(Sema &SemaRef,
444 SourceLocation AsmLoc,
445 ArrayRef<Token> AsmToks,
Chad Rosierb55e6022012-09-13 00:06:55 +0000446 std::vector<std::string> &AsmStrings,
Chad Rosier9072a022012-08-22 20:30:58 +0000447 std::vector<std::pair<unsigned,unsigned> > &AsmTokRanges) {
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000448 assert (!AsmToks.empty() && "Didn't expect an empty AsmToks!");
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000449
450 SmallString<512> Asm;
Chad Rosier9072a022012-08-22 20:30:58 +0000451 unsigned startTok = 0;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000452 for (unsigned i = 0, e = AsmToks.size(); i < e; ++i) {
Bob Wilson40d39e32012-09-24 19:57:55 +0000453 bool isNewAsm = ((i == 0) ||
454 AsmToks[i].isAtStartOfLine() ||
455 AsmToks[i].is(tok::kw_asm));
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000456
457 if (isNewAsm) {
Chad Rosier38c71d32012-08-21 21:56:39 +0000458 if (i) {
Benjamin Kramer32f3acc2012-08-24 20:43:21 +0000459 AsmStrings.push_back(Asm.str());
Chad Rosier9072a022012-08-22 20:30:58 +0000460 AsmTokRanges.push_back(std::make_pair(startTok, i-1));
461 startTok = i;
Chad Rosier38c71d32012-08-21 21:56:39 +0000462 Asm.clear();
Chad Rosier38c71d32012-08-21 21:56:39 +0000463 }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000464 if (AsmToks[i].is(tok::kw_asm)) {
465 i++; // Skip __asm
Bob Wilsonb0f6b9c2012-09-24 19:57:59 +0000466 if (i == e) {
467 SemaRef.Diag(AsmLoc, diag::err_asm_empty);
468 return true;
469 }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000470 }
471 }
472
Chad Rosierb55e6022012-09-13 00:06:55 +0000473 if (i && AsmToks[i].hasLeadingSpace() && !isNewAsm)
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000474 Asm += ' ';
Chad Rosier4de97162012-09-11 00:51:28 +0000475
476 StringRef Spelling = getSpelling(SemaRef, AsmToks[i]);
477 Asm += Spelling;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000478 }
Benjamin Kramer32f3acc2012-08-24 20:43:21 +0000479 AsmStrings.push_back(Asm.str());
Chad Rosier9072a022012-08-22 20:30:58 +0000480 AsmTokRanges.push_back(std::make_pair(startTok, AsmToks.size()-1));
Bob Wilsonb0f6b9c2012-09-24 19:57:59 +0000481
482 return false;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000483}
484
Chad Rosierc6916492012-09-06 19:56:25 +0000485#define DEF_SIMPLE_MSASM(STR) \
Chad Rosier89fb6d72012-08-28 20:28:20 +0000486 MSAsmStmt *NS = \
487 new (Context) MSAsmStmt(Context, AsmLoc, LBraceLoc, /*IsSimple*/ true, \
488 /*IsVolatile*/ true, AsmToks, Inputs, Outputs, \
Chad Rosierc6916492012-09-06 19:56:25 +0000489 InputExprs, OutputExprs, STR, Constraints, \
Chad Rosier89fb6d72012-08-28 20:28:20 +0000490 Clobbers, EndLoc);
Chad Rosier2735df22012-08-22 19:18:30 +0000491
Chad Rosierb55e6022012-09-13 00:06:55 +0000492StmtResult Sema::ActOnMSAsmStmt(SourceLocation AsmLoc, SourceLocation LBraceLoc,
493 ArrayRef<Token> AsmToks,SourceLocation EndLoc) {
Chad Rosier89fb6d72012-08-28 20:28:20 +0000494 SmallVector<StringRef, 4> Constraints;
495 std::vector<std::string> InputConstraints;
496 std::vector<std::string> OutputConstraints;
Chad Rosier4112a4c2012-08-28 20:35:06 +0000497 SmallVector<StringRef, 4> Clobbers;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000498 std::set<std::string> ClobberRegs;
Chad Rosierb55e6022012-09-13 00:06:55 +0000499
500 // FIXME: Use a struct to hold the various expression information.
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000501 SmallVector<IdentifierInfo*, 4> Inputs;
502 SmallVector<IdentifierInfo*, 4> Outputs;
Chad Rosier633abb02012-08-24 00:07:09 +0000503 SmallVector<Expr*, 4> InputExprs;
504 SmallVector<Expr*, 4> OutputExprs;
Chad Rosieracc22b62012-09-06 19:35:00 +0000505 SmallVector<std::string, 4> InputExprNames;
506 SmallVector<std::string, 4> OutputExprNames;
Chad Rosierb55e6022012-09-13 00:06:55 +0000507 SmallVector<unsigned, 4> InputExprStrIdx;
508 SmallVector<unsigned, 4> OutputExprStrIdx;
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000509
510 // Empty asm statements don't need to instantiate the AsmParser, etc.
Chad Rosierc6916492012-09-06 19:56:25 +0000511 StringRef EmptyAsmStr;
512 if (AsmToks.empty()) { DEF_SIMPLE_MSASM(EmptyAsmStr); return Owned(NS); }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000513
Chad Rosier38c71d32012-08-21 21:56:39 +0000514 std::vector<std::string> AsmStrings;
Chad Rosier9072a022012-08-22 20:30:58 +0000515 std::vector<std::pair<unsigned,unsigned> > AsmTokRanges;
Bob Wilsonb0f6b9c2012-09-24 19:57:59 +0000516 if (buildMSAsmStrings(*this, AsmLoc, AsmToks, AsmStrings, AsmTokRanges))
517 return StmtError();
Chad Rosier38c71d32012-08-21 21:56:39 +0000518
Chad Rosiere78460f2012-08-22 20:57:07 +0000519 std::vector<std::vector<StringRef> > Pieces(AsmStrings.size());
Chad Rosierf0fbd772012-08-22 21:04:07 +0000520 buildMSAsmPieces(AsmStrings, Pieces);
Chad Rosier153f8ec2012-08-22 19:50:28 +0000521
522 bool IsSimple = isSimpleMSAsm(Pieces, Context.getTargetInfo());
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000523
Chad Rosier2735df22012-08-22 19:18:30 +0000524 // AsmParser doesn't fully support these asm statements.
Chad Rosierc6916492012-09-06 19:56:25 +0000525 if (bailOnMSAsm(Pieces)) { DEF_SIMPLE_MSASM(EmptyAsmStr); return Owned(NS); }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000526
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000527 // Get the target specific parser.
528 std::string Error;
529 const std::string &TT = Context.getTargetInfo().getTriple().getTriple();
530 const llvm::Target *TheTarget(llvm::TargetRegistry::lookupTarget(TT, Error));
531
532 OwningPtr<llvm::MCAsmInfo> MAI(TheTarget->createMCAsmInfo(TT));
533 OwningPtr<llvm::MCRegisterInfo> MRI(TheTarget->createMCRegInfo(TT));
534 OwningPtr<llvm::MCObjectFileInfo> MOFI(new llvm::MCObjectFileInfo());
535 OwningPtr<llvm::MCSubtargetInfo>
536 STI(TheTarget->createMCSubtargetInfo(TT, "", ""));
537
Chad Rosier25bd2982012-08-23 15:44:35 +0000538 for (unsigned StrIdx = 0, e = AsmStrings.size(); StrIdx != e; ++StrIdx) {
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000539 llvm::SourceMgr SrcMgr;
540 llvm::MCContext Ctx(*MAI, *MRI, MOFI.get(), &SrcMgr);
541 llvm::MemoryBuffer *Buffer =
Chad Rosier25bd2982012-08-23 15:44:35 +0000542 llvm::MemoryBuffer::getMemBuffer(AsmStrings[StrIdx], "<inline asm>");
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000543
544 // Tell SrcMgr about this buffer, which is what the parser will pick up.
545 SrcMgr.AddNewSourceBuffer(Buffer, llvm::SMLoc());
546
547 OwningPtr<llvm::MCStreamer> Str(createNullStreamer(Ctx));
548 OwningPtr<llvm::MCAsmParser>
549 Parser(createMCAsmParser(SrcMgr, Ctx, *Str.get(), *MAI));
550 OwningPtr<llvm::MCTargetAsmParser>
551 TargetParser(TheTarget->createMCAsmParser(*STI, *Parser));
552 // Change to the Intel dialect.
553 Parser->setAssemblerDialect(1);
554 Parser->setTargetParser(*TargetParser.get());
555
556 // Prime the lexer.
557 Parser->Lex();
558
559 // Parse the opcode.
560 StringRef IDVal;
561 Parser->ParseIdentifier(IDVal);
562
563 // Canonicalize the opcode to lower case.
Chad Rosierbe5c3fb2012-09-03 02:30:13 +0000564 SmallString<128> OpcodeStr;
Chad Rosierb706d902012-08-28 22:08:58 +0000565 for (unsigned i = 0, e = IDVal.size(); i != e; ++i)
Chad Rosierbe5c3fb2012-09-03 02:30:13 +0000566 OpcodeStr.push_back(tolower(IDVal[i]));
Chad Rosier3d442192012-09-21 22:22:39 +0000567 // FIXME: Convert to a StmtError.
568 assert(TargetParser->mnemonicIsValid(OpcodeStr) && "Invalid mnemonic!");
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000569
570 // Parse the operands.
571 llvm::SMLoc IDLoc;
572 SmallVector<llvm::MCParsedAsmOperand*, 8> Operands;
Chad Rosierbe5c3fb2012-09-03 02:30:13 +0000573 bool HadError = TargetParser->ParseInstruction(OpcodeStr.str(), IDLoc,
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000574 Operands);
Chad Rosier2735df22012-08-22 19:18:30 +0000575 // If we had an error parsing the operands, fail gracefully.
Chad Rosierc6916492012-09-06 19:56:25 +0000576 if (HadError) { DEF_SIMPLE_MSASM(EmptyAsmStr); return Owned(NS); }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000577
Chad Rosier3329ad12012-10-02 21:49:14 +0000578 // Rewrite the symbol references as wildcard MCParsedAsmOperands.
579 for (unsigned i = 1, e = Operands.size(); i != e; ++i)
580 if (Operands[i]->isMem()) {
581 StringRef Name = getMSInlineAsmExprName(Pieces[StrIdx][i]);
582
583 // The expr may be a register. E.g., DWORD PTR [eax]
584 if (Context.getTargetInfo().isValidGCCRegisterName(Name))
585 continue;
586
587 IdentifierInfo *II = getIdentifierInfo(Name, AsmToks,
588 AsmTokRanges[StrIdx].first,
589 AsmTokRanges[StrIdx].second);
590 // Lookup the identifier.
591 // TODO: Someone with more experience with clang should verify this the
592 // proper way of doing a symbol lookup.
593 DeclarationName DeclName(II);
594 Scope *CurScope = getCurScope();
595 LookupResult R(*this, DeclName, AsmLoc, Sema::LookupOrdinaryName);
596 if (!this->LookupName(R, CurScope, false/*AllowBuiltinCreation*/))
597 assert(0 && "Sema::LookupName failed!");
598 assert (R.isSingleResult() && "Expected a single result?!");
599 NamedDecl *Decl = R.getFoundDecl();
600 switch (Decl->getKind()) {
601 default:
602 assert(0 && "Unknown decl kind.");
603 break;
604 case Decl::Var: {
605 case Decl::ParmVar:
606 VarDecl *Var = cast<VarDecl>(Decl);
607 QualType Ty = Var->getType();
608 // Set the expected operand size.
609 Operands[i]->setMSAsmWildcard(Context.getTypeInfo(Ty).first);
610 break;
611 }
612 }
613 }
614
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000615 // Match the MCInstr.
Chad Rosierbe5c3fb2012-09-03 02:30:13 +0000616 unsigned Kind;
Chad Rosier3034d882012-10-01 23:45:59 +0000617 unsigned Opcode;
Chad Rosier83591b62012-08-21 18:15:08 +0000618 unsigned ErrorInfo;
Chad Rosier3034d882012-10-01 23:45:59 +0000619 SmallVector<std::pair< unsigned, std::string >, 4> MapAndConstraints;
620 HadError = TargetParser->MatchInstruction(IDLoc, Operands, *Str.get(), Kind,
621 Opcode, MapAndConstraints,
Chad Rosier7065c522012-09-03 03:16:15 +0000622 ErrorInfo,
Chad Rosier51a6b3f2012-08-21 19:37:55 +0000623 /*matchingInlineAsm*/ true);
Chad Rosier2735df22012-08-22 19:18:30 +0000624 // If we had an error parsing the operands, fail gracefully.
Chad Rosierc6916492012-09-06 19:56:25 +0000625 if (HadError) { DEF_SIMPLE_MSASM(EmptyAsmStr); return Owned(NS); }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000626
627 // Get the instruction descriptor.
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000628 const llvm::MCInstrInfo *MII = TheTarget->createMCInstrInfo();
Chad Rosier3034d882012-10-01 23:45:59 +0000629 const llvm::MCInstrDesc &Desc = MII->get(Opcode);
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000630 llvm::MCInstPrinter *IP =
631 TheTarget->createMCInstPrinter(1, *MAI, *MII, *MRI, *STI);
632
Chad Rosier6e97be72012-08-22 23:42:09 +0000633 // Build the list of clobbers, outputs and inputs.
Chad Rosierfd5e56e2012-08-22 22:10:51 +0000634 unsigned NumDefs = Desc.getNumDefs();
Chad Rosier1b497f22012-09-03 20:40:52 +0000635 for (unsigned i = 1, e = Operands.size(); i != e; ++i) {
Chad Rosier3034d882012-10-01 23:45:59 +0000636 // Skip immediates.
637 if (Operands[i]->isImm())
Chad Rosierfb700262012-09-11 23:13:15 +0000638 continue;
639
Chad Rosier3034d882012-10-01 23:45:59 +0000640 // Register.
641 if (Operands[i]->isReg()) {
642 // Clobber.
643 if (NumDefs && (MapAndConstraints[i-1].first < NumDefs)) {
Chad Rosier7f9678b2012-09-12 18:14:25 +0000644 std::string Reg;
645 llvm::raw_string_ostream OS(Reg);
Chad Rosier3034d882012-10-01 23:45:59 +0000646 IP->printRegName(OS, Operands[i]->getReg());
Chad Rosier7f9678b2012-09-12 18:14:25 +0000647 StringRef Clobber(OS.str());
648 if (!Context.getTargetInfo().isValidClobber(Clobber))
649 return StmtError(
650 Diag(AsmLoc, diag::err_asm_unknown_register_name) << Clobber);
651 ClobberRegs.insert(Reg);
Chad Rosier1b497f22012-09-03 20:40:52 +0000652 }
Chad Rosier3034d882012-10-01 23:45:59 +0000653 continue;
654 }
655
656 // Expr/Input or Output.
657 StringRef Name = getMSInlineAsmExprName(Pieces[StrIdx][i]);
658
Chad Rosier3329ad12012-10-02 21:49:14 +0000659 // The expr may be a register. E.g., DWORD PTR [eax]
Chad Rosier3034d882012-10-01 23:45:59 +0000660 if (Context.getTargetInfo().isValidGCCRegisterName(Name))
661 continue;
662
663 IdentifierInfo *II = getIdentifierInfo(Name, AsmToks,
664 AsmTokRanges[StrIdx].first,
665 AsmTokRanges[StrIdx].second);
666 if (II) {
667 CXXScopeSpec SS;
668 UnqualifiedId Id;
669 SourceLocation Loc;
670 Id.setIdentifier(II, AsmLoc);
671 ExprResult Result = ActOnIdExpression(getCurScope(), SS, Loc, Id,
672 false, false);
673 if (!Result.isInvalid()) {
674 // FIXME: Determine the proper constraints.
675 bool isMemDef = (i == 1) && Desc.mayStore();
676 if (isMemDef) {
677 Outputs.push_back(II);
678 OutputExprs.push_back(Result.take());
679 OutputExprNames.push_back(Name.str());
680 OutputExprStrIdx.push_back(StrIdx);
681
682 std::string Constraint = "=" + MapAndConstraints[i-1].second;
683 OutputConstraints.push_back(Constraint);
684 } else {
685 Inputs.push_back(II);
686 InputExprs.push_back(Result.take());
687 InputExprNames.push_back(Name.str());
688 InputExprStrIdx.push_back(StrIdx);
689 InputConstraints.push_back(MapAndConstraints[i-1].second);
Chad Rosier633abb02012-08-24 00:07:09 +0000690 }
Chad Rosier6e97be72012-08-22 23:42:09 +0000691 }
Chad Rosierfd5e56e2012-08-22 22:10:51 +0000692 }
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000693 }
694 }
695 for (std::set<std::string>::iterator I = ClobberRegs.begin(),
696 E = ClobberRegs.end(); I != E; ++I)
697 Clobbers.push_back(*I);
698
Chad Rosier89fb6d72012-08-28 20:28:20 +0000699 // Merge the output and input constraints. Output constraints are expected
700 // first.
701 for (std::vector<std::string>::iterator I = OutputConstraints.begin(),
702 E = OutputConstraints.end(); I != E; ++I)
703 Constraints.push_back(*I);
704
705 for (std::vector<std::string>::iterator I = InputConstraints.begin(),
706 E = InputConstraints.end(); I != E; ++I)
707 Constraints.push_back(*I);
708
Chad Rosieracc22b62012-09-06 19:35:00 +0000709 // Enumerate the AsmString expressions.
Chad Rosieracc22b62012-09-06 19:35:00 +0000710 unsigned OpNum = 0;
711 for (unsigned i = 0, e = OutputExprNames.size(); i != e; ++i, ++OpNum) {
Chad Rosierb55e6022012-09-13 00:06:55 +0000712 unsigned StrIdx = OutputExprStrIdx[i];
713 // Iterate over the assembly instruction pieces, skipping the mnemonic.
714 for (unsigned j = 1, f = Pieces[StrIdx].size(); j != f; ++j) {
715 // If the operand and the expression name match, then rewrite the operand.
716 if (OutputExprNames[i] == Pieces[StrIdx][j]) {
717 SmallString<32> Res;
718 llvm::raw_svector_ostream OS(Res);
719 OS << '$' << OpNum;
720 OutputExprNames[i] = OS.str();
721 Pieces[StrIdx][j] = OutputExprNames[i];
722 break;
723 }
724 // Check to see if the expression is a substring of the asm piece.
725 std::pair< StringRef, StringRef > Split = Pieces[StrIdx][j].split(' ');
726 bool isKeyword = llvm::StringSwitch<bool>(Split.first)
727 .Cases("BYTE", "byte", "WORD", "word", "DWORD", true)
728 .Cases("dword", "QWORD", "qword", "XWORD", "xword", true)
729 .Cases("XMMWORD", "xmmword", "YMMWORD", "ymmword", true)
730 .Default(false);
731 if (isKeyword &&
732 Split.second.find_first_of(OutputExprNames[i]) != StringRef::npos) {
733 // Is is a substring, do the replacement.
734 SmallString<32> Res;
735 llvm::raw_svector_ostream OS(Res);
736 OS << '$' << OpNum;
737 std::string piece = Pieces[StrIdx][j].str();
738 size_t found = piece.find(InputExprNames[i]);
739 piece.replace(found, InputExprNames[i].size(), OS.str());
740 OutputExprNames[i] = piece;
741 Pieces[StrIdx][j] = OutputExprNames[i];
742 break;
743 }
744 }
Chad Rosieracc22b62012-09-06 19:35:00 +0000745 }
746 for (unsigned i = 0, e = InputExprNames.size(); i != e; ++i, ++OpNum) {
Chad Rosierb55e6022012-09-13 00:06:55 +0000747 unsigned StrIdx = InputExprStrIdx[i];
748 // Iterate over the assembly instruction pieces, skipping the mnemonic.
749 for (unsigned j = 1, f = Pieces[StrIdx].size(); j != f; ++j) {
750 // If the operand and the expression name match, then rewrite the operand.
751 if (InputExprNames[i] == Pieces[StrIdx][j]) {
752 SmallString<32> Res;
753 llvm::raw_svector_ostream OS(Res);
754 OS << '$' << OpNum;
755 InputExprNames[i] = OS.str();
756 Pieces[StrIdx][j] = InputExprNames[i];
757 break;
758 }
759 // Check to see if the expression is a substring of the asm piece.
760 std::pair< StringRef, StringRef > Split = Pieces[StrIdx][j].split(' ');
761 bool isKeyword = llvm::StringSwitch<bool>(Split.first)
762 .Cases("BYTE", "byte", "WORD", "word", "DWORD", true)
763 .Cases("dword", "QWORD", "qword", "XWORD", "xword", true)
764 .Cases("XMMWORD", "xmmword", "YMMWORD", "ymmword", true)
765 .Default(false);
766 if (isKeyword &&
767 Split.second.find_first_of(InputExprNames[i]) != StringRef::npos) {
768 // It is a substring, do the replacement.
769 SmallString<32> Res;
770 llvm::raw_svector_ostream OS(Res);
771 OS << '$' << OpNum;
772 std::string piece = Pieces[StrIdx][j].str();
773 size_t found = piece.find(InputExprNames[i]);
774 piece.replace(found, InputExprNames[i].size(), OS.str());
775 InputExprNames[i] = piece;
776 Pieces[StrIdx][j] = InputExprNames[i];
777 break;
778 }
779 }
780 }
781
782 // Emit the IR assembly string.
783 std::string AsmString;
784 for (unsigned i = 0, e = Pieces.size(); i != e; ++i) {
785 // Skip empty asm stmts.
786 if (Pieces[i].empty()) continue;
787
788 if (i > 0)
789 AsmString += "\n\t";
790
791 // Emit the mnemonic.
792 AsmString += Pieces[i][0];
793 if (Pieces[i].size() > 1)
794 AsmString += ' ';
795
796 // Emit the operands adding $$ to constants.
797 for (unsigned j = 1, f = Pieces[i].size(); j != f; ++j) {
798 if (j > 1) AsmString += ", ";
799 unsigned Val;
800 if (!Pieces[i][j].getAsInteger(0, Val))
801 AsmString += "$$";
802
803 AsmString += Pieces[i][j];
804 }
Chad Rosieracc22b62012-09-06 19:35:00 +0000805 }
806
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000807 MSAsmStmt *NS =
808 new (Context) MSAsmStmt(Context, AsmLoc, LBraceLoc, IsSimple,
809 /*IsVolatile*/ true, AsmToks, Inputs, Outputs,
Chad Rosier89fb6d72012-08-28 20:28:20 +0000810 InputExprs, OutputExprs, AsmString, Constraints,
811 Clobbers, EndLoc);
Chad Rosier4b5e48d2012-08-17 21:19:40 +0000812 return Owned(NS);
813}