1 //===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===//
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 contains code to emit Stmt nodes as LLVM code.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CodeGenFunction.h"
15 #include "CGDebugInfo.h"
16 #include "CodeGenModule.h"
17 #include "TargetInfo.h"
18 #include "clang/AST/StmtVisitor.h"
19 #include "clang/Basic/Builtins.h"
20 #include "clang/Basic/PrettyStackTrace.h"
21 #include "clang/Basic/TargetInfo.h"
22 #include "clang/Sema/LoopHint.h"
23 #include "clang/Sema/SemaDiagnostic.h"
24 #include "llvm/ADT/StringExtras.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/InlineAsm.h"
28 #include "llvm/IR/Intrinsics.h"
29 #include "llvm/IR/MDBuilder.h"
30
31 using namespace clang;
32 using namespace CodeGen;
33
34 //===----------------------------------------------------------------------===//
35 // Statement Emission
36 //===----------------------------------------------------------------------===//
37
EmitStopPoint(const Stmt * S)38 void CodeGenFunction::EmitStopPoint(const Stmt *S) {
39 if (CGDebugInfo *DI = getDebugInfo()) {
40 SourceLocation Loc;
41 Loc = S->getLocStart();
42 DI->EmitLocation(Builder, Loc);
43
44 LastStopPoint = Loc;
45 }
46 }
47
EmitStmt(const Stmt * S)48 void CodeGenFunction::EmitStmt(const Stmt *S) {
49 assert(S && "Null statement?");
50 PGO.setCurrentStmt(S);
51
52 // These statements have their own debug info handling.
53 if (EmitSimpleStmt(S))
54 return;
55
56 // Check if we are generating unreachable code.
57 if (!HaveInsertPoint()) {
58 // If so, and the statement doesn't contain a label, then we do not need to
59 // generate actual code. This is safe because (1) the current point is
60 // unreachable, so we don't need to execute the code, and (2) we've already
61 // handled the statements which update internal data structures (like the
62 // local variable map) which could be used by subsequent statements.
63 if (!ContainsLabel(S)) {
64 // Verify that any decl statements were handled as simple, they may be in
65 // scope of subsequent reachable statements.
66 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!");
67 return;
68 }
69
70 // Otherwise, make a new block to hold the code.
71 EnsureInsertPoint();
72 }
73
74 // Generate a stoppoint if we are emitting debug info.
75 EmitStopPoint(S);
76
77 switch (S->getStmtClass()) {
78 case Stmt::NoStmtClass:
79 case Stmt::CXXCatchStmtClass:
80 case Stmt::SEHExceptStmtClass:
81 case Stmt::SEHFinallyStmtClass:
82 case Stmt::MSDependentExistsStmtClass:
83 llvm_unreachable("invalid statement class to emit generically");
84 case Stmt::NullStmtClass:
85 case Stmt::CompoundStmtClass:
86 case Stmt::DeclStmtClass:
87 case Stmt::LabelStmtClass:
88 case Stmt::AttributedStmtClass:
89 case Stmt::GotoStmtClass:
90 case Stmt::BreakStmtClass:
91 case Stmt::ContinueStmtClass:
92 case Stmt::DefaultStmtClass:
93 case Stmt::CaseStmtClass:
94 case Stmt::SEHLeaveStmtClass:
95 llvm_unreachable("should have emitted these statements as simple");
96
97 #define STMT(Type, Base)
98 #define ABSTRACT_STMT(Op)
99 #define EXPR(Type, Base) \
100 case Stmt::Type##Class:
101 #include "clang/AST/StmtNodes.inc"
102 {
103 // Remember the block we came in on.
104 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
105 assert(incoming && "expression emission must have an insertion point");
106
107 EmitIgnoredExpr(cast<Expr>(S));
108
109 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
110 assert(outgoing && "expression emission cleared block!");
111
112 // The expression emitters assume (reasonably!) that the insertion
113 // point is always set. To maintain that, the call-emission code
114 // for noreturn functions has to enter a new block with no
115 // predecessors. We want to kill that block and mark the current
116 // insertion point unreachable in the common case of a call like
117 // "exit();". Since expression emission doesn't otherwise create
118 // blocks with no predecessors, we can just test for that.
119 // However, we must be careful not to do this to our incoming
120 // block, because *statement* emission does sometimes create
121 // reachable blocks which will have no predecessors until later in
122 // the function. This occurs with, e.g., labels that are not
123 // reachable by fallthrough.
124 if (incoming != outgoing && outgoing->use_empty()) {
125 outgoing->eraseFromParent();
126 Builder.ClearInsertionPoint();
127 }
128 break;
129 }
130
131 case Stmt::IndirectGotoStmtClass:
132 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break;
133
134 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break;
135 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S)); break;
136 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S)); break;
137 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S)); break;
138
139 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break;
140
141 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break;
142 case Stmt::GCCAsmStmtClass: // Intentional fall-through.
143 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break;
144 case Stmt::CoroutineBodyStmtClass:
145 case Stmt::CoreturnStmtClass:
146 CGM.ErrorUnsupported(S, "coroutine");
147 break;
148 case Stmt::CapturedStmtClass: {
149 const CapturedStmt *CS = cast<CapturedStmt>(S);
150 EmitCapturedStmt(*CS, CS->getCapturedRegionKind());
151 }
152 break;
153 case Stmt::ObjCAtTryStmtClass:
154 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S));
155 break;
156 case Stmt::ObjCAtCatchStmtClass:
157 llvm_unreachable(
158 "@catch statements should be handled by EmitObjCAtTryStmt");
159 case Stmt::ObjCAtFinallyStmtClass:
160 llvm_unreachable(
161 "@finally statements should be handled by EmitObjCAtTryStmt");
162 case Stmt::ObjCAtThrowStmtClass:
163 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S));
164 break;
165 case Stmt::ObjCAtSynchronizedStmtClass:
166 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S));
167 break;
168 case Stmt::ObjCForCollectionStmtClass:
169 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S));
170 break;
171 case Stmt::ObjCAutoreleasePoolStmtClass:
172 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S));
173 break;
174
175 case Stmt::CXXTryStmtClass:
176 EmitCXXTryStmt(cast<CXXTryStmt>(*S));
177 break;
178 case Stmt::CXXForRangeStmtClass:
179 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S));
180 break;
181 case Stmt::SEHTryStmtClass:
182 EmitSEHTryStmt(cast<SEHTryStmt>(*S));
183 break;
184 case Stmt::OMPParallelDirectiveClass:
185 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S));
186 break;
187 case Stmt::OMPSimdDirectiveClass:
188 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S));
189 break;
190 case Stmt::OMPForDirectiveClass:
191 EmitOMPForDirective(cast<OMPForDirective>(*S));
192 break;
193 case Stmt::OMPForSimdDirectiveClass:
194 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S));
195 break;
196 case Stmt::OMPSectionsDirectiveClass:
197 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S));
198 break;
199 case Stmt::OMPSectionDirectiveClass:
200 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S));
201 break;
202 case Stmt::OMPSingleDirectiveClass:
203 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S));
204 break;
205 case Stmt::OMPMasterDirectiveClass:
206 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S));
207 break;
208 case Stmt::OMPCriticalDirectiveClass:
209 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S));
210 break;
211 case Stmt::OMPParallelForDirectiveClass:
212 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S));
213 break;
214 case Stmt::OMPParallelForSimdDirectiveClass:
215 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S));
216 break;
217 case Stmt::OMPParallelSectionsDirectiveClass:
218 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S));
219 break;
220 case Stmt::OMPTaskDirectiveClass:
221 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S));
222 break;
223 case Stmt::OMPTaskyieldDirectiveClass:
224 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
225 break;
226 case Stmt::OMPBarrierDirectiveClass:
227 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
228 break;
229 case Stmt::OMPTaskwaitDirectiveClass:
230 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S));
231 break;
232 case Stmt::OMPTaskgroupDirectiveClass:
233 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S));
234 break;
235 case Stmt::OMPFlushDirectiveClass:
236 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S));
237 break;
238 case Stmt::OMPOrderedDirectiveClass:
239 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S));
240 break;
241 case Stmt::OMPAtomicDirectiveClass:
242 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S));
243 break;
244 case Stmt::OMPTargetDirectiveClass:
245 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S));
246 break;
247 case Stmt::OMPTeamsDirectiveClass:
248 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S));
249 break;
250 case Stmt::OMPCancellationPointDirectiveClass:
251 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S));
252 break;
253 case Stmt::OMPCancelDirectiveClass:
254 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S));
255 break;
256 case Stmt::OMPTargetDataDirectiveClass:
257 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S));
258 break;
259 case Stmt::OMPTargetEnterDataDirectiveClass:
260 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S));
261 break;
262 case Stmt::OMPTargetExitDataDirectiveClass:
263 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S));
264 break;
265 case Stmt::OMPTargetParallelDirectiveClass:
266 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S));
267 break;
268 case Stmt::OMPTargetParallelForDirectiveClass:
269 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S));
270 break;
271 case Stmt::OMPTaskLoopDirectiveClass:
272 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S));
273 break;
274 case Stmt::OMPTaskLoopSimdDirectiveClass:
275 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S));
276 break;
277 case Stmt::OMPDistributeDirectiveClass:
278 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S));
279 break;
280 case Stmt::OMPTargetUpdateDirectiveClass:
281 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S));
282 break;
283 case Stmt::OMPDistributeParallelForDirectiveClass:
284 EmitOMPDistributeParallelForDirective(
285 cast<OMPDistributeParallelForDirective>(*S));
286 break;
287 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
288 EmitOMPDistributeParallelForSimdDirective(
289 cast<OMPDistributeParallelForSimdDirective>(*S));
290 break;
291 case Stmt::OMPDistributeSimdDirectiveClass:
292 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S));
293 break;
294 case Stmt::OMPTargetParallelForSimdDirectiveClass:
295 EmitOMPTargetParallelForSimdDirective(
296 cast<OMPTargetParallelForSimdDirective>(*S));
297 break;
298 }
299 }
300
EmitSimpleStmt(const Stmt * S)301 bool CodeGenFunction::EmitSimpleStmt(const Stmt *S) {
302 switch (S->getStmtClass()) {
303 default: return false;
304 case Stmt::NullStmtClass: break;
305 case Stmt::CompoundStmtClass: EmitCompoundStmt(cast<CompoundStmt>(*S)); break;
306 case Stmt::DeclStmtClass: EmitDeclStmt(cast<DeclStmt>(*S)); break;
307 case Stmt::LabelStmtClass: EmitLabelStmt(cast<LabelStmt>(*S)); break;
308 case Stmt::AttributedStmtClass:
309 EmitAttributedStmt(cast<AttributedStmt>(*S)); break;
310 case Stmt::GotoStmtClass: EmitGotoStmt(cast<GotoStmt>(*S)); break;
311 case Stmt::BreakStmtClass: EmitBreakStmt(cast<BreakStmt>(*S)); break;
312 case Stmt::ContinueStmtClass: EmitContinueStmt(cast<ContinueStmt>(*S)); break;
313 case Stmt::DefaultStmtClass: EmitDefaultStmt(cast<DefaultStmt>(*S)); break;
314 case Stmt::CaseStmtClass: EmitCaseStmt(cast<CaseStmt>(*S)); break;
315 case Stmt::SEHLeaveStmtClass: EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S)); break;
316 }
317
318 return true;
319 }
320
321 /// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
322 /// this captures the expression result of the last sub-statement and returns it
323 /// (for use by the statement expression extension).
EmitCompoundStmt(const CompoundStmt & S,bool GetLast,AggValueSlot AggSlot)324 Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
325 AggValueSlot AggSlot) {
326 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
327 "LLVM IR generation of compound statement ('{}')");
328
329 // Keep track of the current cleanup stack depth, including debug scopes.
330 LexicalScope Scope(*this, S.getSourceRange());
331
332 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
333 }
334
335 Address
EmitCompoundStmtWithoutScope(const CompoundStmt & S,bool GetLast,AggValueSlot AggSlot)336 CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
337 bool GetLast,
338 AggValueSlot AggSlot) {
339
340 for (CompoundStmt::const_body_iterator I = S.body_begin(),
341 E = S.body_end()-GetLast; I != E; ++I)
342 EmitStmt(*I);
343
344 Address RetAlloca = Address::invalid();
345 if (GetLast) {
346 // We have to special case labels here. They are statements, but when put
347 // at the end of a statement expression, they yield the value of their
348 // subexpression. Handle this by walking through all labels we encounter,
349 // emitting them before we evaluate the subexpr.
350 const Stmt *LastStmt = S.body_back();
351 while (const LabelStmt *LS = dyn_cast<LabelStmt>(LastStmt)) {
352 EmitLabel(LS->getDecl());
353 LastStmt = LS->getSubStmt();
354 }
355
356 EnsureInsertPoint();
357
358 QualType ExprTy = cast<Expr>(LastStmt)->getType();
359 if (hasAggregateEvaluationKind(ExprTy)) {
360 EmitAggExpr(cast<Expr>(LastStmt), AggSlot);
361 } else {
362 // We can't return an RValue here because there might be cleanups at
363 // the end of the StmtExpr. Because of that, we have to emit the result
364 // here into a temporary alloca.
365 RetAlloca = CreateMemTemp(ExprTy);
366 EmitAnyExprToMem(cast<Expr>(LastStmt), RetAlloca, Qualifiers(),
367 /*IsInit*/false);
368 }
369
370 }
371
372 return RetAlloca;
373 }
374
SimplifyForwardingBlocks(llvm::BasicBlock * BB)375 void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
376 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator());
377
378 // If there is a cleanup stack, then we it isn't worth trying to
379 // simplify this block (we would need to remove it from the scope map
380 // and cleanup entry).
381 if (!EHStack.empty())
382 return;
383
384 // Can only simplify direct branches.
385 if (!BI || !BI->isUnconditional())
386 return;
387
388 // Can only simplify empty blocks.
389 if (BI->getIterator() != BB->begin())
390 return;
391
392 BB->replaceAllUsesWith(BI->getSuccessor(0));
393 BI->eraseFromParent();
394 BB->eraseFromParent();
395 }
396
EmitBlock(llvm::BasicBlock * BB,bool IsFinished)397 void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
398 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
399
400 // Fall out of the current block (if necessary).
401 EmitBranch(BB);
402
403 if (IsFinished && BB->use_empty()) {
404 delete BB;
405 return;
406 }
407
408 // Place the block after the current block, if possible, or else at
409 // the end of the function.
410 if (CurBB && CurBB->getParent())
411 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB);
412 else
413 CurFn->getBasicBlockList().push_back(BB);
414 Builder.SetInsertPoint(BB);
415 }
416
EmitBranch(llvm::BasicBlock * Target)417 void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
418 // Emit a branch from the current block to the target one if this
419 // was a real block. If this was just a fall-through block after a
420 // terminator, don't emit it.
421 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
422
423 if (!CurBB || CurBB->getTerminator()) {
424 // If there is no insert point or the previous block is already
425 // terminated, don't touch it.
426 } else {
427 // Otherwise, create a fall-through branch.
428 Builder.CreateBr(Target);
429 }
430
431 Builder.ClearInsertionPoint();
432 }
433
EmitBlockAfterUses(llvm::BasicBlock * block)434 void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
435 bool inserted = false;
436 for (llvm::User *u : block->users()) {
437 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) {
438 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(),
439 block);
440 inserted = true;
441 break;
442 }
443 }
444
445 if (!inserted)
446 CurFn->getBasicBlockList().push_back(block);
447
448 Builder.SetInsertPoint(block);
449 }
450
451 CodeGenFunction::JumpDest
getJumpDestForLabel(const LabelDecl * D)452 CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
453 JumpDest &Dest = LabelMap[D];
454 if (Dest.isValid()) return Dest;
455
456 // Create, but don't insert, the new block.
457 Dest = JumpDest(createBasicBlock(D->getName()),
458 EHScopeStack::stable_iterator::invalid(),
459 NextCleanupDestIndex++);
460 return Dest;
461 }
462
EmitLabel(const LabelDecl * D)463 void CodeGenFunction::EmitLabel(const LabelDecl *D) {
464 // Add this label to the current lexical scope if we're within any
465 // normal cleanups. Jumps "in" to this label --- when permitted by
466 // the language --- may need to be routed around such cleanups.
467 if (EHStack.hasNormalCleanups() && CurLexicalScope)
468 CurLexicalScope->addLabel(D);
469
470 JumpDest &Dest = LabelMap[D];
471
472 // If we didn't need a forward reference to this label, just go
473 // ahead and create a destination at the current scope.
474 if (!Dest.isValid()) {
475 Dest = getJumpDestInCurrentScope(D->getName());
476
477 // Otherwise, we need to give this label a target depth and remove
478 // it from the branch-fixups list.
479 } else {
480 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
481 Dest.setScopeDepth(EHStack.stable_begin());
482 ResolveBranchFixups(Dest.getBlock());
483 }
484
485 EmitBlock(Dest.getBlock());
486 incrementProfileCounter(D->getStmt());
487 }
488
489 /// Change the cleanup scope of the labels in this lexical scope to
490 /// match the scope of the enclosing context.
rescopeLabels()491 void CodeGenFunction::LexicalScope::rescopeLabels() {
492 assert(!Labels.empty());
493 EHScopeStack::stable_iterator innermostScope
494 = CGF.EHStack.getInnermostNormalCleanup();
495
496 // Change the scope depth of all the labels.
497 for (SmallVectorImpl<const LabelDecl*>::const_iterator
498 i = Labels.begin(), e = Labels.end(); i != e; ++i) {
499 assert(CGF.LabelMap.count(*i));
500 JumpDest &dest = CGF.LabelMap.find(*i)->second;
501 assert(dest.getScopeDepth().isValid());
502 assert(innermostScope.encloses(dest.getScopeDepth()));
503 dest.setScopeDepth(innermostScope);
504 }
505
506 // Reparent the labels if the new scope also has cleanups.
507 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
508 ParentScope->Labels.append(Labels.begin(), Labels.end());
509 }
510 }
511
512
EmitLabelStmt(const LabelStmt & S)513 void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
514 EmitLabel(S.getDecl());
515 EmitStmt(S.getSubStmt());
516 }
517
EmitAttributedStmt(const AttributedStmt & S)518 void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
519 const Stmt *SubStmt = S.getSubStmt();
520 switch (SubStmt->getStmtClass()) {
521 case Stmt::DoStmtClass:
522 EmitDoStmt(cast<DoStmt>(*SubStmt), S.getAttrs());
523 break;
524 case Stmt::ForStmtClass:
525 EmitForStmt(cast<ForStmt>(*SubStmt), S.getAttrs());
526 break;
527 case Stmt::WhileStmtClass:
528 EmitWhileStmt(cast<WhileStmt>(*SubStmt), S.getAttrs());
529 break;
530 case Stmt::CXXForRangeStmtClass:
531 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*SubStmt), S.getAttrs());
532 break;
533 default:
534 EmitStmt(SubStmt);
535 }
536 }
537
EmitGotoStmt(const GotoStmt & S)538 void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
539 // If this code is reachable then emit a stop point (if generating
540 // debug info). We have to do this ourselves because we are on the
541 // "simple" statement path.
542 if (HaveInsertPoint())
543 EmitStopPoint(&S);
544
545 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel()));
546 }
547
548
EmitIndirectGotoStmt(const IndirectGotoStmt & S)549 void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
550 if (const LabelDecl *Target = S.getConstantTarget()) {
551 EmitBranchThroughCleanup(getJumpDestForLabel(Target));
552 return;
553 }
554
555 // Ensure that we have an i8* for our PHI node.
556 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()),
557 Int8PtrTy, "addr");
558 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
559
560 // Get the basic block for the indirect goto.
561 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
562
563 // The first instruction in the block has to be the PHI for the switch dest,
564 // add an entry for this branch.
565 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB);
566
567 EmitBranch(IndGotoBB);
568 }
569
EmitIfStmt(const IfStmt & S)570 void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
571 // C99 6.8.4.1: The first substatement is executed if the expression compares
572 // unequal to 0. The condition must be a scalar type.
573 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
574
575 if (S.getInit())
576 EmitStmt(S.getInit());
577
578 if (S.getConditionVariable())
579 EmitAutoVarDecl(*S.getConditionVariable());
580
581 // If the condition constant folds and can be elided, try to avoid emitting
582 // the condition and the dead arm of the if/else.
583 bool CondConstant;
584 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant,
585 S.isConstexpr())) {
586 // Figure out which block (then or else) is executed.
587 const Stmt *Executed = S.getThen();
588 const Stmt *Skipped = S.getElse();
589 if (!CondConstant) // Condition false?
590 std::swap(Executed, Skipped);
591
592 // If the skipped block has no labels in it, just emit the executed block.
593 // This avoids emitting dead code and simplifies the CFG substantially.
594 if (S.isConstexpr() || !ContainsLabel(Skipped)) {
595 if (CondConstant)
596 incrementProfileCounter(&S);
597 if (Executed) {
598 RunCleanupsScope ExecutedScope(*this);
599 EmitStmt(Executed);
600 }
601 return;
602 }
603 }
604
605 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
606 // the conditional branch.
607 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then");
608 llvm::BasicBlock *ContBlock = createBasicBlock("if.end");
609 llvm::BasicBlock *ElseBlock = ContBlock;
610 if (S.getElse())
611 ElseBlock = createBasicBlock("if.else");
612
613 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock,
614 getProfileCount(S.getThen()));
615
616 // Emit the 'then' code.
617 EmitBlock(ThenBlock);
618 incrementProfileCounter(&S);
619 {
620 RunCleanupsScope ThenScope(*this);
621 EmitStmt(S.getThen());
622 }
623 {
624 auto CurBlock = Builder.GetInsertBlock();
625 EmitBranch(ContBlock);
626 // Eliminate any empty blocks that may have been created by nested
627 // control flow statements in the 'then' clause.
628 if (CurBlock)
629 SimplifyForwardingBlocks(CurBlock);
630 }
631
632 // Emit the 'else' code if present.
633 if (const Stmt *Else = S.getElse()) {
634 {
635 // There is no need to emit line number for an unconditional branch.
636 auto NL = ApplyDebugLocation::CreateEmpty(*this);
637 EmitBlock(ElseBlock);
638 }
639 {
640 RunCleanupsScope ElseScope(*this);
641 EmitStmt(Else);
642 }
643 {
644 // There is no need to emit line number for an unconditional branch.
645 auto NL = ApplyDebugLocation::CreateEmpty(*this);
646 auto CurBlock = Builder.GetInsertBlock();
647 EmitBranch(ContBlock);
648 // Eliminate any empty blocks that may have been created by nested
649 // control flow statements emitted in the 'else' clause.
650 if (CurBlock)
651 SimplifyForwardingBlocks(CurBlock);
652 }
653 }
654
655 // Emit the continuation block for code after the if.
656 EmitBlock(ContBlock, true);
657 }
658
EmitWhileStmt(const WhileStmt & S,ArrayRef<const Attr * > WhileAttrs)659 void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
660 ArrayRef<const Attr *> WhileAttrs) {
661 // Emit the header for the loop, which will also become
662 // the continue target.
663 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond");
664 EmitBlock(LoopHeader.getBlock());
665
666 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), WhileAttrs,
667 Builder.getCurrentDebugLocation());
668
669 // Create an exit block for when the condition fails, which will
670 // also become the break target.
671 JumpDest LoopExit = getJumpDestInCurrentScope("while.end");
672
673 // Store the blocks to use for break and continue.
674 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader));
675
676 // C++ [stmt.while]p2:
677 // When the condition of a while statement is a declaration, the
678 // scope of the variable that is declared extends from its point
679 // of declaration (3.3.2) to the end of the while statement.
680 // [...]
681 // The object created in a condition is destroyed and created
682 // with each iteration of the loop.
683 RunCleanupsScope ConditionScope(*this);
684
685 if (S.getConditionVariable())
686 EmitAutoVarDecl(*S.getConditionVariable());
687
688 // Evaluate the conditional in the while header. C99 6.8.5.1: The
689 // evaluation of the controlling expression takes place before each
690 // execution of the loop body.
691 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
692
693 // while(1) is common, avoid extra exit blocks. Be sure
694 // to correctly handle break/continue though.
695 bool EmitBoolCondBranch = true;
696 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
697 if (C->isOne())
698 EmitBoolCondBranch = false;
699
700 // As long as the condition is true, go to the loop body.
701 llvm::BasicBlock *LoopBody = createBasicBlock("while.body");
702 if (EmitBoolCondBranch) {
703 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
704 if (ConditionScope.requiresCleanups())
705 ExitBlock = createBasicBlock("while.exit");
706 Builder.CreateCondBr(
707 BoolCondVal, LoopBody, ExitBlock,
708 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
709
710 if (ExitBlock != LoopExit.getBlock()) {
711 EmitBlock(ExitBlock);
712 EmitBranchThroughCleanup(LoopExit);
713 }
714 }
715
716 // Emit the loop body. We have to emit this in a cleanup scope
717 // because it might be a singleton DeclStmt.
718 {
719 RunCleanupsScope BodyScope(*this);
720 EmitBlock(LoopBody);
721 incrementProfileCounter(&S);
722 EmitStmt(S.getBody());
723 }
724
725 BreakContinueStack.pop_back();
726
727 // Immediately force cleanup.
728 ConditionScope.ForceCleanup();
729
730 EmitStopPoint(&S);
731 // Branch to the loop header again.
732 EmitBranch(LoopHeader.getBlock());
733
734 LoopStack.pop();
735
736 // Emit the exit block.
737 EmitBlock(LoopExit.getBlock(), true);
738
739 // The LoopHeader typically is just a branch if we skipped emitting
740 // a branch, try to erase it.
741 if (!EmitBoolCondBranch)
742 SimplifyForwardingBlocks(LoopHeader.getBlock());
743 }
744
EmitDoStmt(const DoStmt & S,ArrayRef<const Attr * > DoAttrs)745 void CodeGenFunction::EmitDoStmt(const DoStmt &S,
746 ArrayRef<const Attr *> DoAttrs) {
747 JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
748 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
749
750 uint64_t ParentCount = getCurrentProfileCount();
751
752 // Store the blocks to use for break and continue.
753 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond));
754
755 // Emit the body of the loop.
756 llvm::BasicBlock *LoopBody = createBasicBlock("do.body");
757
758 LoopStack.push(LoopBody, CGM.getContext(), DoAttrs,
759 Builder.getCurrentDebugLocation());
760
761 EmitBlockWithFallThrough(LoopBody, &S);
762 {
763 RunCleanupsScope BodyScope(*this);
764 EmitStmt(S.getBody());
765 }
766
767 EmitBlock(LoopCond.getBlock());
768
769 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
770 // after each execution of the loop body."
771
772 // Evaluate the conditional in the while header.
773 // C99 6.8.5p2/p4: The first substatement is executed if the expression
774 // compares unequal to 0. The condition must be a scalar type.
775 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
776
777 BreakContinueStack.pop_back();
778
779 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
780 // to correctly handle break/continue though.
781 bool EmitBoolCondBranch = true;
782 if (llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal))
783 if (C->isZero())
784 EmitBoolCondBranch = false;
785
786 // As long as the condition is true, iterate the loop.
787 if (EmitBoolCondBranch) {
788 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
789 Builder.CreateCondBr(
790 BoolCondVal, LoopBody, LoopExit.getBlock(),
791 createProfileWeightsForLoop(S.getCond(), BackedgeCount));
792 }
793
794 LoopStack.pop();
795
796 // Emit the exit block.
797 EmitBlock(LoopExit.getBlock());
798
799 // The DoCond block typically is just a branch if we skipped
800 // emitting a branch, try to erase it.
801 if (!EmitBoolCondBranch)
802 SimplifyForwardingBlocks(LoopCond.getBlock());
803 }
804
EmitForStmt(const ForStmt & S,ArrayRef<const Attr * > ForAttrs)805 void CodeGenFunction::EmitForStmt(const ForStmt &S,
806 ArrayRef<const Attr *> ForAttrs) {
807 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
808
809 LexicalScope ForScope(*this, S.getSourceRange());
810
811 llvm::DebugLoc DL = Builder.getCurrentDebugLocation();
812
813 // Evaluate the first part before the loop.
814 if (S.getInit())
815 EmitStmt(S.getInit());
816
817 // Start the loop with a block that tests the condition.
818 // If there's an increment, the continue scope will be overwritten
819 // later.
820 JumpDest Continue = getJumpDestInCurrentScope("for.cond");
821 llvm::BasicBlock *CondBlock = Continue.getBlock();
822 EmitBlock(CondBlock);
823
824 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, DL);
825
826 // If the for loop doesn't have an increment we can just use the
827 // condition as the continue block. Otherwise we'll need to create
828 // a block for it (in the current scope, i.e. in the scope of the
829 // condition), and that we will become our continue block.
830 if (S.getInc())
831 Continue = getJumpDestInCurrentScope("for.inc");
832
833 // Store the blocks to use for break and continue.
834 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
835
836 // Create a cleanup scope for the condition variable cleanups.
837 LexicalScope ConditionScope(*this, S.getSourceRange());
838
839 if (S.getCond()) {
840 // If the for statement has a condition scope, emit the local variable
841 // declaration.
842 if (S.getConditionVariable()) {
843 EmitAutoVarDecl(*S.getConditionVariable());
844 }
845
846 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
847 // If there are any cleanups between here and the loop-exit scope,
848 // create a block to stage a loop exit along.
849 if (ForScope.requiresCleanups())
850 ExitBlock = createBasicBlock("for.cond.cleanup");
851
852 // As long as the condition is true, iterate the loop.
853 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
854
855 // C99 6.8.5p2/p4: The first substatement is executed if the expression
856 // compares unequal to 0. The condition must be a scalar type.
857 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
858 Builder.CreateCondBr(
859 BoolCondVal, ForBody, ExitBlock,
860 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
861
862 if (ExitBlock != LoopExit.getBlock()) {
863 EmitBlock(ExitBlock);
864 EmitBranchThroughCleanup(LoopExit);
865 }
866
867 EmitBlock(ForBody);
868 } else {
869 // Treat it as a non-zero constant. Don't even create a new block for the
870 // body, just fall into it.
871 }
872 incrementProfileCounter(&S);
873
874 {
875 // Create a separate cleanup scope for the body, in case it is not
876 // a compound statement.
877 RunCleanupsScope BodyScope(*this);
878 EmitStmt(S.getBody());
879 }
880
881 // If there is an increment, emit it next.
882 if (S.getInc()) {
883 EmitBlock(Continue.getBlock());
884 EmitStmt(S.getInc());
885 }
886
887 BreakContinueStack.pop_back();
888
889 ConditionScope.ForceCleanup();
890
891 EmitStopPoint(&S);
892 EmitBranch(CondBlock);
893
894 ForScope.ForceCleanup();
895
896 LoopStack.pop();
897
898 // Emit the fall-through block.
899 EmitBlock(LoopExit.getBlock(), true);
900 }
901
902 void
EmitCXXForRangeStmt(const CXXForRangeStmt & S,ArrayRef<const Attr * > ForAttrs)903 CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
904 ArrayRef<const Attr *> ForAttrs) {
905 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
906
907 LexicalScope ForScope(*this, S.getSourceRange());
908
909 llvm::DebugLoc DL = Builder.getCurrentDebugLocation();
910
911 // Evaluate the first pieces before the loop.
912 EmitStmt(S.getRangeStmt());
913 EmitStmt(S.getBeginStmt());
914 EmitStmt(S.getEndStmt());
915
916 // Start the loop with a block that tests the condition.
917 // If there's an increment, the continue scope will be overwritten
918 // later.
919 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond");
920 EmitBlock(CondBlock);
921
922 LoopStack.push(CondBlock, CGM.getContext(), ForAttrs, DL);
923
924 // If there are any cleanups between here and the loop-exit scope,
925 // create a block to stage a loop exit along.
926 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
927 if (ForScope.requiresCleanups())
928 ExitBlock = createBasicBlock("for.cond.cleanup");
929
930 // The loop body, consisting of the specified body and the loop variable.
931 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
932
933 // The body is executed if the expression, contextually converted
934 // to bool, is true.
935 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
936 Builder.CreateCondBr(
937 BoolCondVal, ForBody, ExitBlock,
938 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody())));
939
940 if (ExitBlock != LoopExit.getBlock()) {
941 EmitBlock(ExitBlock);
942 EmitBranchThroughCleanup(LoopExit);
943 }
944
945 EmitBlock(ForBody);
946 incrementProfileCounter(&S);
947
948 // Create a block for the increment. In case of a 'continue', we jump there.
949 JumpDest Continue = getJumpDestInCurrentScope("for.inc");
950
951 // Store the blocks to use for break and continue.
952 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
953
954 {
955 // Create a separate cleanup scope for the loop variable and body.
956 LexicalScope BodyScope(*this, S.getSourceRange());
957 EmitStmt(S.getLoopVarStmt());
958 EmitStmt(S.getBody());
959 }
960
961 EmitStopPoint(&S);
962 // If there is an increment, emit it next.
963 EmitBlock(Continue.getBlock());
964 EmitStmt(S.getInc());
965
966 BreakContinueStack.pop_back();
967
968 EmitBranch(CondBlock);
969
970 ForScope.ForceCleanup();
971
972 LoopStack.pop();
973
974 // Emit the fall-through block.
975 EmitBlock(LoopExit.getBlock(), true);
976 }
977
EmitReturnOfRValue(RValue RV,QualType Ty)978 void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
979 if (RV.isScalar()) {
980 Builder.CreateStore(RV.getScalarVal(), ReturnValue);
981 } else if (RV.isAggregate()) {
982 EmitAggregateCopy(ReturnValue, RV.getAggregateAddress(), Ty);
983 } else {
984 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty),
985 /*init*/ true);
986 }
987 EmitBranchThroughCleanup(ReturnBlock);
988 }
989
990 /// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
991 /// if the function returns void, or may be missing one if the function returns
992 /// non-void. Fun stuff :).
EmitReturnStmt(const ReturnStmt & S)993 void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
994 // Returning from an outlined SEH helper is UB, and we already warn on it.
995 if (IsOutlinedSEHHelper) {
996 Builder.CreateUnreachable();
997 Builder.ClearInsertionPoint();
998 }
999
1000 // Emit the result value, even if unused, to evalute the side effects.
1001 const Expr *RV = S.getRetValue();
1002
1003 // Treat block literals in a return expression as if they appeared
1004 // in their own scope. This permits a small, easily-implemented
1005 // exception to our over-conservative rules about not jumping to
1006 // statements following block literals with non-trivial cleanups.
1007 RunCleanupsScope cleanupScope(*this);
1008 if (const ExprWithCleanups *cleanups =
1009 dyn_cast_or_null<ExprWithCleanups>(RV)) {
1010 enterFullExpression(cleanups);
1011 RV = cleanups->getSubExpr();
1012 }
1013
1014 // FIXME: Clean this up by using an LValue for ReturnTemp,
1015 // EmitStoreThroughLValue, and EmitAnyExpr.
1016 if (getLangOpts().ElideConstructors &&
1017 S.getNRVOCandidate() && S.getNRVOCandidate()->isNRVOVariable()) {
1018 // Apply the named return value optimization for this return statement,
1019 // which means doing nothing: the appropriate result has already been
1020 // constructed into the NRVO variable.
1021
1022 // If there is an NRVO flag for this variable, set it to 1 into indicate
1023 // that the cleanup code should not destroy the variable.
1024 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1025 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag);
1026 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1027 // Make sure not to return anything, but evaluate the expression
1028 // for side effects.
1029 if (RV)
1030 EmitAnyExpr(RV);
1031 } else if (!RV) {
1032 // Do nothing (return value is left uninitialized)
1033 } else if (FnRetTy->isReferenceType()) {
1034 // If this function returns a reference, take the address of the expression
1035 // rather than the value.
1036 RValue Result = EmitReferenceBindingToExpr(RV);
1037 Builder.CreateStore(Result.getScalarVal(), ReturnValue);
1038 } else {
1039 switch (getEvaluationKind(RV->getType())) {
1040 case TEK_Scalar:
1041 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue);
1042 break;
1043 case TEK_Complex:
1044 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()),
1045 /*isInit*/ true);
1046 break;
1047 case TEK_Aggregate:
1048 EmitAggExpr(RV, AggValueSlot::forAddr(ReturnValue,
1049 Qualifiers(),
1050 AggValueSlot::IsDestructed,
1051 AggValueSlot::DoesNotNeedGCBarriers,
1052 AggValueSlot::IsNotAliased));
1053 break;
1054 }
1055 }
1056
1057 ++NumReturnExprs;
1058 if (!RV || RV->isEvaluatable(getContext()))
1059 ++NumSimpleReturnExprs;
1060
1061 cleanupScope.ForceCleanup();
1062 EmitBranchThroughCleanup(ReturnBlock);
1063 }
1064
EmitDeclStmt(const DeclStmt & S)1065 void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1066 // As long as debug info is modeled with instructions, we have to ensure we
1067 // have a place to insert here and write the stop point here.
1068 if (HaveInsertPoint())
1069 EmitStopPoint(&S);
1070
1071 for (const auto *I : S.decls())
1072 EmitDecl(*I);
1073 }
1074
EmitBreakStmt(const BreakStmt & S)1075 void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1076 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1077
1078 // If this code is reachable then emit a stop point (if generating
1079 // debug info). We have to do this ourselves because we are on the
1080 // "simple" statement path.
1081 if (HaveInsertPoint())
1082 EmitStopPoint(&S);
1083
1084 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock);
1085 }
1086
EmitContinueStmt(const ContinueStmt & S)1087 void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1088 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1089
1090 // If this code is reachable then emit a stop point (if generating
1091 // debug info). We have to do this ourselves because we are on the
1092 // "simple" statement path.
1093 if (HaveInsertPoint())
1094 EmitStopPoint(&S);
1095
1096 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock);
1097 }
1098
1099 /// EmitCaseStmtRange - If case statement range is not too big then
1100 /// add multiple cases to switch instruction, one for each value within
1101 /// the range. If range is too big then emit "if" condition check.
EmitCaseStmtRange(const CaseStmt & S)1102 void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S) {
1103 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1104
1105 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext());
1106 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext());
1107
1108 // Emit the code for this case. We do this first to make sure it is
1109 // properly chained from our predecessor before generating the
1110 // switch machinery to enter this block.
1111 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1112 EmitBlockWithFallThrough(CaseDest, &S);
1113 EmitStmt(S.getSubStmt());
1114
1115 // If range is empty, do nothing.
1116 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS))
1117 return;
1118
1119 llvm::APInt Range = RHS - LHS;
1120 // FIXME: parameters such as this should not be hardcoded.
1121 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) {
1122 // Range is small enough to add multiple switch instruction cases.
1123 uint64_t Total = getProfileCount(&S);
1124 unsigned NCases = Range.getZExtValue() + 1;
1125 // We only have one region counter for the entire set of cases here, so we
1126 // need to divide the weights evenly between the generated cases, ensuring
1127 // that the total weight is preserved. E.g., a weight of 5 over three cases
1128 // will be distributed as weights of 2, 2, and 1.
1129 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1130 for (unsigned I = 0; I != NCases; ++I) {
1131 if (SwitchWeights)
1132 SwitchWeights->push_back(Weight + (Rem ? 1 : 0));
1133 if (Rem)
1134 Rem--;
1135 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest);
1136 LHS++;
1137 }
1138 return;
1139 }
1140
1141 // The range is too big. Emit "if" condition into a new block,
1142 // making sure to save and restore the current insertion point.
1143 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1144
1145 // Push this test onto the chain of range checks (which terminates
1146 // in the default basic block). The switch's default will be changed
1147 // to the top of this chain after switch emission is complete.
1148 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1149 CaseRangeBlock = createBasicBlock("sw.caserange");
1150
1151 CurFn->getBasicBlockList().push_back(CaseRangeBlock);
1152 Builder.SetInsertPoint(CaseRangeBlock);
1153
1154 // Emit range check.
1155 llvm::Value *Diff =
1156 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS));
1157 llvm::Value *Cond =
1158 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds");
1159
1160 llvm::MDNode *Weights = nullptr;
1161 if (SwitchWeights) {
1162 uint64_t ThisCount = getProfileCount(&S);
1163 uint64_t DefaultCount = (*SwitchWeights)[0];
1164 Weights = createProfileWeights(ThisCount, DefaultCount);
1165
1166 // Since we're chaining the switch default through each large case range, we
1167 // need to update the weight for the default, ie, the first case, to include
1168 // this case.
1169 (*SwitchWeights)[0] += ThisCount;
1170 }
1171 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights);
1172
1173 // Restore the appropriate insertion point.
1174 if (RestoreBB)
1175 Builder.SetInsertPoint(RestoreBB);
1176 else
1177 Builder.ClearInsertionPoint();
1178 }
1179
EmitCaseStmt(const CaseStmt & S)1180 void CodeGenFunction::EmitCaseStmt(const CaseStmt &S) {
1181 // If there is no enclosing switch instance that we're aware of, then this
1182 // case statement and its block can be elided. This situation only happens
1183 // when we've constant-folded the switch, are emitting the constant case,
1184 // and part of the constant case includes another case statement. For
1185 // instance: switch (4) { case 4: do { case 5: } while (1); }
1186 if (!SwitchInsn) {
1187 EmitStmt(S.getSubStmt());
1188 return;
1189 }
1190
1191 // Handle case ranges.
1192 if (S.getRHS()) {
1193 EmitCaseStmtRange(S);
1194 return;
1195 }
1196
1197 llvm::ConstantInt *CaseVal =
1198 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext()));
1199
1200 // If the body of the case is just a 'break', try to not emit an empty block.
1201 // If we're profiling or we're not optimizing, leave the block in for better
1202 // debug and coverage analysis.
1203 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1204 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1205 isa<BreakStmt>(S.getSubStmt())) {
1206 JumpDest Block = BreakContinueStack.back().BreakBlock;
1207
1208 // Only do this optimization if there are no cleanups that need emitting.
1209 if (isObviouslyBranchWithoutCleanups(Block)) {
1210 if (SwitchWeights)
1211 SwitchWeights->push_back(getProfileCount(&S));
1212 SwitchInsn->addCase(CaseVal, Block.getBlock());
1213
1214 // If there was a fallthrough into this case, make sure to redirect it to
1215 // the end of the switch as well.
1216 if (Builder.GetInsertBlock()) {
1217 Builder.CreateBr(Block.getBlock());
1218 Builder.ClearInsertionPoint();
1219 }
1220 return;
1221 }
1222 }
1223
1224 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1225 EmitBlockWithFallThrough(CaseDest, &S);
1226 if (SwitchWeights)
1227 SwitchWeights->push_back(getProfileCount(&S));
1228 SwitchInsn->addCase(CaseVal, CaseDest);
1229
1230 // Recursively emitting the statement is acceptable, but is not wonderful for
1231 // code where we have many case statements nested together, i.e.:
1232 // case 1:
1233 // case 2:
1234 // case 3: etc.
1235 // Handling this recursively will create a new block for each case statement
1236 // that falls through to the next case which is IR intensive. It also causes
1237 // deep recursion which can run into stack depth limitations. Handle
1238 // sequential non-range case statements specially.
1239 const CaseStmt *CurCase = &S;
1240 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt());
1241
1242 // Otherwise, iteratively add consecutive cases to this switch stmt.
1243 while (NextCase && NextCase->getRHS() == nullptr) {
1244 CurCase = NextCase;
1245 llvm::ConstantInt *CaseVal =
1246 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext()));
1247
1248 if (SwitchWeights)
1249 SwitchWeights->push_back(getProfileCount(NextCase));
1250 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1251 CaseDest = createBasicBlock("sw.bb");
1252 EmitBlockWithFallThrough(CaseDest, &S);
1253 }
1254
1255 SwitchInsn->addCase(CaseVal, CaseDest);
1256 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt());
1257 }
1258
1259 // Normal default recursion for non-cases.
1260 EmitStmt(CurCase->getSubStmt());
1261 }
1262
EmitDefaultStmt(const DefaultStmt & S)1263 void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S) {
1264 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1265 assert(DefaultBlock->empty() &&
1266 "EmitDefaultStmt: Default block already defined?");
1267
1268 EmitBlockWithFallThrough(DefaultBlock, &S);
1269
1270 EmitStmt(S.getSubStmt());
1271 }
1272
1273 /// CollectStatementsForCase - Given the body of a 'switch' statement and a
1274 /// constant value that is being switched on, see if we can dead code eliminate
1275 /// the body of the switch to a simple series of statements to emit. Basically,
1276 /// on a switch (5) we want to find these statements:
1277 /// case 5:
1278 /// printf(...); <--
1279 /// ++i; <--
1280 /// break;
1281 ///
1282 /// and add them to the ResultStmts vector. If it is unsafe to do this
1283 /// transformation (for example, one of the elided statements contains a label
1284 /// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
1285 /// should include statements after it (e.g. the printf() line is a substmt of
1286 /// the case) then return CSFC_FallThrough. If we handled it and found a break
1287 /// statement, then return CSFC_Success.
1288 ///
1289 /// If Case is non-null, then we are looking for the specified case, checking
1290 /// that nothing we jump over contains labels. If Case is null, then we found
1291 /// the case and are looking for the break.
1292 ///
1293 /// If the recursive walk actually finds our Case, then we set FoundCase to
1294 /// true.
1295 ///
1296 enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
CollectStatementsForCase(const Stmt * S,const SwitchCase * Case,bool & FoundCase,SmallVectorImpl<const Stmt * > & ResultStmts)1297 static CSFC_Result CollectStatementsForCase(const Stmt *S,
1298 const SwitchCase *Case,
1299 bool &FoundCase,
1300 SmallVectorImpl<const Stmt*> &ResultStmts) {
1301 // If this is a null statement, just succeed.
1302 if (!S)
1303 return Case ? CSFC_Success : CSFC_FallThrough;
1304
1305 // If this is the switchcase (case 4: or default) that we're looking for, then
1306 // we're in business. Just add the substatement.
1307 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) {
1308 if (S == Case) {
1309 FoundCase = true;
1310 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase,
1311 ResultStmts);
1312 }
1313
1314 // Otherwise, this is some other case or default statement, just ignore it.
1315 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase,
1316 ResultStmts);
1317 }
1318
1319 // If we are in the live part of the code and we found our break statement,
1320 // return a success!
1321 if (!Case && isa<BreakStmt>(S))
1322 return CSFC_Success;
1323
1324 // If this is a switch statement, then it might contain the SwitchCase, the
1325 // break, or neither.
1326 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) {
1327 // Handle this as two cases: we might be looking for the SwitchCase (if so
1328 // the skipped statements must be skippable) or we might already have it.
1329 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
1330 if (Case) {
1331 // Keep track of whether we see a skipped declaration. The code could be
1332 // using the declaration even if it is skipped, so we can't optimize out
1333 // the decl if the kept statements might refer to it.
1334 bool HadSkippedDecl = false;
1335
1336 // If we're looking for the case, just see if we can skip each of the
1337 // substatements.
1338 for (; Case && I != E; ++I) {
1339 HadSkippedDecl |= isa<DeclStmt>(*I);
1340
1341 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) {
1342 case CSFC_Failure: return CSFC_Failure;
1343 case CSFC_Success:
1344 // A successful result means that either 1) that the statement doesn't
1345 // have the case and is skippable, or 2) does contain the case value
1346 // and also contains the break to exit the switch. In the later case,
1347 // we just verify the rest of the statements are elidable.
1348 if (FoundCase) {
1349 // If we found the case and skipped declarations, we can't do the
1350 // optimization.
1351 if (HadSkippedDecl)
1352 return CSFC_Failure;
1353
1354 for (++I; I != E; ++I)
1355 if (CodeGenFunction::ContainsLabel(*I, true))
1356 return CSFC_Failure;
1357 return CSFC_Success;
1358 }
1359 break;
1360 case CSFC_FallThrough:
1361 // If we have a fallthrough condition, then we must have found the
1362 // case started to include statements. Consider the rest of the
1363 // statements in the compound statement as candidates for inclusion.
1364 assert(FoundCase && "Didn't find case but returned fallthrough?");
1365 // We recursively found Case, so we're not looking for it anymore.
1366 Case = nullptr;
1367
1368 // If we found the case and skipped declarations, we can't do the
1369 // optimization.
1370 if (HadSkippedDecl)
1371 return CSFC_Failure;
1372 break;
1373 }
1374 }
1375 }
1376
1377 // If we have statements in our range, then we know that the statements are
1378 // live and need to be added to the set of statements we're tracking.
1379 for (; I != E; ++I) {
1380 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) {
1381 case CSFC_Failure: return CSFC_Failure;
1382 case CSFC_FallThrough:
1383 // A fallthrough result means that the statement was simple and just
1384 // included in ResultStmt, keep adding them afterwards.
1385 break;
1386 case CSFC_Success:
1387 // A successful result means that we found the break statement and
1388 // stopped statement inclusion. We just ensure that any leftover stmts
1389 // are skippable and return success ourselves.
1390 for (++I; I != E; ++I)
1391 if (CodeGenFunction::ContainsLabel(*I, true))
1392 return CSFC_Failure;
1393 return CSFC_Success;
1394 }
1395 }
1396
1397 return Case ? CSFC_Success : CSFC_FallThrough;
1398 }
1399
1400 // Okay, this is some other statement that we don't handle explicitly, like a
1401 // for statement or increment etc. If we are skipping over this statement,
1402 // just verify it doesn't have labels, which would make it invalid to elide.
1403 if (Case) {
1404 if (CodeGenFunction::ContainsLabel(S, true))
1405 return CSFC_Failure;
1406 return CSFC_Success;
1407 }
1408
1409 // Otherwise, we want to include this statement. Everything is cool with that
1410 // so long as it doesn't contain a break out of the switch we're in.
1411 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
1412
1413 // Otherwise, everything is great. Include the statement and tell the caller
1414 // that we fall through and include the next statement as well.
1415 ResultStmts.push_back(S);
1416 return CSFC_FallThrough;
1417 }
1418
1419 /// FindCaseStatementsForValue - Find the case statement being jumped to and
1420 /// then invoke CollectStatementsForCase to find the list of statements to emit
1421 /// for a switch on constant. See the comment above CollectStatementsForCase
1422 /// for more details.
FindCaseStatementsForValue(const SwitchStmt & S,const llvm::APSInt & ConstantCondValue,SmallVectorImpl<const Stmt * > & ResultStmts,ASTContext & C,const SwitchCase * & ResultCase)1423 static bool FindCaseStatementsForValue(const SwitchStmt &S,
1424 const llvm::APSInt &ConstantCondValue,
1425 SmallVectorImpl<const Stmt*> &ResultStmts,
1426 ASTContext &C,
1427 const SwitchCase *&ResultCase) {
1428 // First step, find the switch case that is being branched to. We can do this
1429 // efficiently by scanning the SwitchCase list.
1430 const SwitchCase *Case = S.getSwitchCaseList();
1431 const DefaultStmt *DefaultCase = nullptr;
1432
1433 for (; Case; Case = Case->getNextSwitchCase()) {
1434 // It's either a default or case. Just remember the default statement in
1435 // case we're not jumping to any numbered cases.
1436 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) {
1437 DefaultCase = DS;
1438 continue;
1439 }
1440
1441 // Check to see if this case is the one we're looking for.
1442 const CaseStmt *CS = cast<CaseStmt>(Case);
1443 // Don't handle case ranges yet.
1444 if (CS->getRHS()) return false;
1445
1446 // If we found our case, remember it as 'case'.
1447 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue)
1448 break;
1449 }
1450
1451 // If we didn't find a matching case, we use a default if it exists, or we
1452 // elide the whole switch body!
1453 if (!Case) {
1454 // It is safe to elide the body of the switch if it doesn't contain labels
1455 // etc. If it is safe, return successfully with an empty ResultStmts list.
1456 if (!DefaultCase)
1457 return !CodeGenFunction::ContainsLabel(&S);
1458 Case = DefaultCase;
1459 }
1460
1461 // Ok, we know which case is being jumped to, try to collect all the
1462 // statements that follow it. This can fail for a variety of reasons. Also,
1463 // check to see that the recursive walk actually found our case statement.
1464 // Insane cases like this can fail to find it in the recursive walk since we
1465 // don't handle every stmt kind:
1466 // switch (4) {
1467 // while (1) {
1468 // case 4: ...
1469 bool FoundCase = false;
1470 ResultCase = Case;
1471 return CollectStatementsForCase(S.getBody(), Case, FoundCase,
1472 ResultStmts) != CSFC_Failure &&
1473 FoundCase;
1474 }
1475
EmitSwitchStmt(const SwitchStmt & S)1476 void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
1477 // Handle nested switch statements.
1478 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
1479 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
1480 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
1481
1482 // See if we can constant fold the condition of the switch and therefore only
1483 // emit the live case statement (if any) of the switch.
1484 llvm::APSInt ConstantCondValue;
1485 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) {
1486 SmallVector<const Stmt*, 4> CaseStmts;
1487 const SwitchCase *Case = nullptr;
1488 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts,
1489 getContext(), Case)) {
1490 if (Case)
1491 incrementProfileCounter(Case);
1492 RunCleanupsScope ExecutedScope(*this);
1493
1494 if (S.getInit())
1495 EmitStmt(S.getInit());
1496
1497 // Emit the condition variable if needed inside the entire cleanup scope
1498 // used by this special case for constant folded switches.
1499 if (S.getConditionVariable())
1500 EmitAutoVarDecl(*S.getConditionVariable());
1501
1502 // At this point, we are no longer "within" a switch instance, so
1503 // we can temporarily enforce this to ensure that any embedded case
1504 // statements are not emitted.
1505 SwitchInsn = nullptr;
1506
1507 // Okay, we can dead code eliminate everything except this case. Emit the
1508 // specified series of statements and we're good.
1509 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i)
1510 EmitStmt(CaseStmts[i]);
1511 incrementProfileCounter(&S);
1512
1513 // Now we want to restore the saved switch instance so that nested
1514 // switches continue to function properly
1515 SwitchInsn = SavedSwitchInsn;
1516
1517 return;
1518 }
1519 }
1520
1521 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog");
1522
1523 RunCleanupsScope ConditionScope(*this);
1524
1525 if (S.getInit())
1526 EmitStmt(S.getInit());
1527
1528 if (S.getConditionVariable())
1529 EmitAutoVarDecl(*S.getConditionVariable());
1530 llvm::Value *CondV = EmitScalarExpr(S.getCond());
1531
1532 // Create basic block to hold stuff that comes after switch
1533 // statement. We also need to create a default block now so that
1534 // explicit case ranges tests can have a place to jump to on
1535 // failure.
1536 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default");
1537 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock);
1538 if (PGO.haveRegionCounts()) {
1539 // Walk the SwitchCase list to find how many there are.
1540 uint64_t DefaultCount = 0;
1541 unsigned NumCases = 0;
1542 for (const SwitchCase *Case = S.getSwitchCaseList();
1543 Case;
1544 Case = Case->getNextSwitchCase()) {
1545 if (isa<DefaultStmt>(Case))
1546 DefaultCount = getProfileCount(Case);
1547 NumCases += 1;
1548 }
1549 SwitchWeights = new SmallVector<uint64_t, 16>();
1550 SwitchWeights->reserve(NumCases);
1551 // The default needs to be first. We store the edge count, so we already
1552 // know the right weight.
1553 SwitchWeights->push_back(DefaultCount);
1554 }
1555 CaseRangeBlock = DefaultBlock;
1556
1557 // Clear the insertion point to indicate we are in unreachable code.
1558 Builder.ClearInsertionPoint();
1559
1560 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
1561 // then reuse last ContinueBlock.
1562 JumpDest OuterContinue;
1563 if (!BreakContinueStack.empty())
1564 OuterContinue = BreakContinueStack.back().ContinueBlock;
1565
1566 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue));
1567
1568 // Emit switch body.
1569 EmitStmt(S.getBody());
1570
1571 BreakContinueStack.pop_back();
1572
1573 // Update the default block in case explicit case range tests have
1574 // been chained on top.
1575 SwitchInsn->setDefaultDest(CaseRangeBlock);
1576
1577 // If a default was never emitted:
1578 if (!DefaultBlock->getParent()) {
1579 // If we have cleanups, emit the default block so that there's a
1580 // place to jump through the cleanups from.
1581 if (ConditionScope.requiresCleanups()) {
1582 EmitBlock(DefaultBlock);
1583
1584 // Otherwise, just forward the default block to the switch end.
1585 } else {
1586 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock());
1587 delete DefaultBlock;
1588 }
1589 }
1590
1591 ConditionScope.ForceCleanup();
1592
1593 // Emit continuation.
1594 EmitBlock(SwitchExit.getBlock(), true);
1595 incrementProfileCounter(&S);
1596
1597 // If the switch has a condition wrapped by __builtin_unpredictable,
1598 // create metadata that specifies that the switch is unpredictable.
1599 // Don't bother if not optimizing because that metadata would not be used.
1600 auto *Call = dyn_cast<CallExpr>(S.getCond());
1601 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1602 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1603 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1604 llvm::MDBuilder MDHelper(getLLVMContext());
1605 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable,
1606 MDHelper.createUnpredictable());
1607 }
1608 }
1609
1610 if (SwitchWeights) {
1611 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
1612 "switch weights do not match switch cases");
1613 // If there's only one jump destination there's no sense weighting it.
1614 if (SwitchWeights->size() > 1)
1615 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1616 createProfileWeights(*SwitchWeights));
1617 delete SwitchWeights;
1618 }
1619 SwitchInsn = SavedSwitchInsn;
1620 SwitchWeights = SavedSwitchWeights;
1621 CaseRangeBlock = SavedCRBlock;
1622 }
1623
1624 static std::string
SimplifyConstraint(const char * Constraint,const TargetInfo & Target,SmallVectorImpl<TargetInfo::ConstraintInfo> * OutCons=nullptr)1625 SimplifyConstraint(const char *Constraint, const TargetInfo &Target,
1626 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) {
1627 std::string Result;
1628
1629 while (*Constraint) {
1630 switch (*Constraint) {
1631 default:
1632 Result += Target.convertConstraint(Constraint);
1633 break;
1634 // Ignore these
1635 case '*':
1636 case '?':
1637 case '!':
1638 case '=': // Will see this and the following in mult-alt constraints.
1639 case '+':
1640 break;
1641 case '#': // Ignore the rest of the constraint alternative.
1642 while (Constraint[1] && Constraint[1] != ',')
1643 Constraint++;
1644 break;
1645 case '&':
1646 case '%':
1647 Result += *Constraint;
1648 while (Constraint[1] && Constraint[1] == *Constraint)
1649 Constraint++;
1650 break;
1651 case ',':
1652 Result += "|";
1653 break;
1654 case 'g':
1655 Result += "imr";
1656 break;
1657 case '[': {
1658 assert(OutCons &&
1659 "Must pass output names to constraints with a symbolic name");
1660 unsigned Index;
1661 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index);
1662 assert(result && "Could not resolve symbolic name"); (void)result;
1663 Result += llvm::utostr(Index);
1664 break;
1665 }
1666 }
1667
1668 Constraint++;
1669 }
1670
1671 return Result;
1672 }
1673
1674 /// AddVariableConstraints - Look at AsmExpr and if it is a variable declared
1675 /// as using a particular register add that as a constraint that will be used
1676 /// in this asm stmt.
1677 static std::string
AddVariableConstraints(const std::string & Constraint,const Expr & AsmExpr,const TargetInfo & Target,CodeGenModule & CGM,const AsmStmt & Stmt,const bool EarlyClobber)1678 AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr,
1679 const TargetInfo &Target, CodeGenModule &CGM,
1680 const AsmStmt &Stmt, const bool EarlyClobber) {
1681 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr);
1682 if (!AsmDeclRef)
1683 return Constraint;
1684 const ValueDecl &Value = *AsmDeclRef->getDecl();
1685 const VarDecl *Variable = dyn_cast<VarDecl>(&Value);
1686 if (!Variable)
1687 return Constraint;
1688 if (Variable->getStorageClass() != SC_Register)
1689 return Constraint;
1690 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>();
1691 if (!Attr)
1692 return Constraint;
1693 StringRef Register = Attr->getLabel();
1694 assert(Target.isValidGCCRegisterName(Register));
1695 // We're using validateOutputConstraint here because we only care if
1696 // this is a register constraint.
1697 TargetInfo::ConstraintInfo Info(Constraint, "");
1698 if (Target.validateOutputConstraint(Info) &&
1699 !Info.allowsRegister()) {
1700 CGM.ErrorUnsupported(&Stmt, "__asm__");
1701 return Constraint;
1702 }
1703 // Canonicalize the register here before returning it.
1704 Register = Target.getNormalizedGCCRegisterName(Register);
1705 return (EarlyClobber ? "&{" : "{") + Register.str() + "}";
1706 }
1707
1708 llvm::Value*
EmitAsmInputLValue(const TargetInfo::ConstraintInfo & Info,LValue InputValue,QualType InputType,std::string & ConstraintStr,SourceLocation Loc)1709 CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info,
1710 LValue InputValue, QualType InputType,
1711 std::string &ConstraintStr,
1712 SourceLocation Loc) {
1713 llvm::Value *Arg;
1714 if (Info.allowsRegister() || !Info.allowsMemory()) {
1715 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) {
1716 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal();
1717 } else {
1718 llvm::Type *Ty = ConvertType(InputType);
1719 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
1720 if (Size <= 64 && llvm::isPowerOf2_64(Size)) {
1721 Ty = llvm::IntegerType::get(getLLVMContext(), Size);
1722 Ty = llvm::PointerType::getUnqual(Ty);
1723
1724 Arg = Builder.CreateLoad(Builder.CreateBitCast(InputValue.getAddress(),
1725 Ty));
1726 } else {
1727 Arg = InputValue.getPointer();
1728 ConstraintStr += '*';
1729 }
1730 }
1731 } else {
1732 Arg = InputValue.getPointer();
1733 ConstraintStr += '*';
1734 }
1735
1736 return Arg;
1737 }
1738
EmitAsmInput(const TargetInfo::ConstraintInfo & Info,const Expr * InputExpr,std::string & ConstraintStr)1739 llvm::Value* CodeGenFunction::EmitAsmInput(
1740 const TargetInfo::ConstraintInfo &Info,
1741 const Expr *InputExpr,
1742 std::string &ConstraintStr) {
1743 // If this can't be a register or memory, i.e., has to be a constant
1744 // (immediate or symbolic), try to emit it as such.
1745 if (!Info.allowsRegister() && !Info.allowsMemory()) {
1746 llvm::APSInt Result;
1747 if (InputExpr->EvaluateAsInt(Result, getContext()))
1748 return llvm::ConstantInt::get(getLLVMContext(), Result);
1749 assert(!Info.requiresImmediateConstant() &&
1750 "Required-immediate inlineasm arg isn't constant?");
1751 }
1752
1753 if (Info.allowsRegister() || !Info.allowsMemory())
1754 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType()))
1755 return EmitScalarExpr(InputExpr);
1756 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
1757 return EmitScalarExpr(InputExpr);
1758 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext());
1759 LValue Dest = EmitLValue(InputExpr);
1760 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr,
1761 InputExpr->getExprLoc());
1762 }
1763
1764 /// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
1765 /// asm call instruction. The !srcloc MDNode contains a list of constant
1766 /// integers which are the source locations of the start of each line in the
1767 /// asm.
getAsmSrcLocInfo(const StringLiteral * Str,CodeGenFunction & CGF)1768 static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
1769 CodeGenFunction &CGF) {
1770 SmallVector<llvm::Metadata *, 8> Locs;
1771 // Add the location of the first line to the MDNode.
1772 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
1773 CGF.Int32Ty, Str->getLocStart().getRawEncoding())));
1774 StringRef StrVal = Str->getString();
1775 if (!StrVal.empty()) {
1776 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
1777 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
1778 unsigned StartToken = 0;
1779 unsigned ByteOffset = 0;
1780
1781 // Add the location of the start of each subsequent line of the asm to the
1782 // MDNode.
1783 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
1784 if (StrVal[i] != '\n') continue;
1785 SourceLocation LineLoc = Str->getLocationOfByte(
1786 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset);
1787 Locs.push_back(llvm::ConstantAsMetadata::get(
1788 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding())));
1789 }
1790 }
1791
1792 return llvm::MDNode::get(CGF.getLLVMContext(), Locs);
1793 }
1794
EmitAsmStmt(const AsmStmt & S)1795 void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
1796 // Assemble the final asm string.
1797 std::string AsmString = S.generateAsmString(getContext());
1798
1799 // Get all the output and input constraints together.
1800 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
1801 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
1802
1803 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1804 StringRef Name;
1805 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1806 Name = GAS->getOutputName(i);
1807 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
1808 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
1809 assert(IsValid && "Failed to parse output constraint");
1810 OutputConstraintInfos.push_back(Info);
1811 }
1812
1813 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1814 StringRef Name;
1815 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
1816 Name = GAS->getInputName(i);
1817 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
1818 bool IsValid =
1819 getTarget().validateInputConstraint(OutputConstraintInfos, Info);
1820 assert(IsValid && "Failed to parse input constraint"); (void)IsValid;
1821 InputConstraintInfos.push_back(Info);
1822 }
1823
1824 std::string Constraints;
1825
1826 std::vector<LValue> ResultRegDests;
1827 std::vector<QualType> ResultRegQualTys;
1828 std::vector<llvm::Type *> ResultRegTypes;
1829 std::vector<llvm::Type *> ResultTruncRegTypes;
1830 std::vector<llvm::Type *> ArgTypes;
1831 std::vector<llvm::Value*> Args;
1832
1833 // Keep track of inout constraints.
1834 std::string InOutConstraints;
1835 std::vector<llvm::Value*> InOutArgs;
1836 std::vector<llvm::Type*> InOutArgTypes;
1837
1838 // An inline asm can be marked readonly if it meets the following conditions:
1839 // - it doesn't have any sideeffects
1840 // - it doesn't clobber memory
1841 // - it doesn't return a value by-reference
1842 // It can be marked readnone if it doesn't have any input memory constraints
1843 // in addition to meeting the conditions listed above.
1844 bool ReadOnly = true, ReadNone = true;
1845
1846 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
1847 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
1848
1849 // Simplify the output constraint.
1850 std::string OutputConstraint(S.getOutputConstraint(i));
1851 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1,
1852 getTarget());
1853
1854 const Expr *OutExpr = S.getOutputExpr(i);
1855 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext());
1856
1857 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr,
1858 getTarget(), CGM, S,
1859 Info.earlyClobber());
1860
1861 LValue Dest = EmitLValue(OutExpr);
1862 if (!Constraints.empty())
1863 Constraints += ',';
1864
1865 // If this is a register output, then make the inline asm return it
1866 // by-value. If this is a memory result, return the value by-reference.
1867 if (!Info.allowsMemory() && hasScalarEvaluationKind(OutExpr->getType())) {
1868 Constraints += "=" + OutputConstraint;
1869 ResultRegQualTys.push_back(OutExpr->getType());
1870 ResultRegDests.push_back(Dest);
1871 ResultRegTypes.push_back(ConvertTypeForMem(OutExpr->getType()));
1872 ResultTruncRegTypes.push_back(ResultRegTypes.back());
1873
1874 // If this output is tied to an input, and if the input is larger, then
1875 // we need to set the actual result type of the inline asm node to be the
1876 // same as the input type.
1877 if (Info.hasMatchingInput()) {
1878 unsigned InputNo;
1879 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
1880 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
1881 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
1882 break;
1883 }
1884 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
1885
1886 QualType InputTy = S.getInputExpr(InputNo)->getType();
1887 QualType OutputType = OutExpr->getType();
1888
1889 uint64_t InputSize = getContext().getTypeSize(InputTy);
1890 if (getContext().getTypeSize(OutputType) < InputSize) {
1891 // Form the asm to return the value as a larger integer or fp type.
1892 ResultRegTypes.back() = ConvertType(InputTy);
1893 }
1894 }
1895 if (llvm::Type* AdjTy =
1896 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1897 ResultRegTypes.back()))
1898 ResultRegTypes.back() = AdjTy;
1899 else {
1900 CGM.getDiags().Report(S.getAsmLoc(),
1901 diag::err_asm_invalid_type_in_input)
1902 << OutExpr->getType() << OutputConstraint;
1903 }
1904 } else {
1905 ArgTypes.push_back(Dest.getAddress().getType());
1906 Args.push_back(Dest.getPointer());
1907 Constraints += "=*";
1908 Constraints += OutputConstraint;
1909 ReadOnly = ReadNone = false;
1910 }
1911
1912 if (Info.isReadWrite()) {
1913 InOutConstraints += ',';
1914
1915 const Expr *InputExpr = S.getOutputExpr(i);
1916 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
1917 InOutConstraints,
1918 InputExpr->getExprLoc());
1919
1920 if (llvm::Type* AdjTy =
1921 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
1922 Arg->getType()))
1923 Arg = Builder.CreateBitCast(Arg, AdjTy);
1924
1925 if (Info.allowsRegister())
1926 InOutConstraints += llvm::utostr(i);
1927 else
1928 InOutConstraints += OutputConstraint;
1929
1930 InOutArgTypes.push_back(Arg->getType());
1931 InOutArgs.push_back(Arg);
1932 }
1933 }
1934
1935 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
1936 // to the return value slot. Only do this when returning in registers.
1937 if (isa<MSAsmStmt>(&S)) {
1938 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
1939 if (RetAI.isDirect() || RetAI.isExtend()) {
1940 // Make a fake lvalue for the return value slot.
1941 LValue ReturnSlot = MakeAddrLValue(ReturnValue, FnRetTy);
1942 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs(
1943 *this, ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes,
1944 ResultRegDests, AsmString, S.getNumOutputs());
1945 SawAsmBlock = true;
1946 }
1947 }
1948
1949 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
1950 const Expr *InputExpr = S.getInputExpr(i);
1951
1952 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
1953
1954 if (Info.allowsMemory())
1955 ReadNone = false;
1956
1957 if (!Constraints.empty())
1958 Constraints += ',';
1959
1960 // Simplify the input constraint.
1961 std::string InputConstraint(S.getInputConstraint(i));
1962 InputConstraint = SimplifyConstraint(InputConstraint.c_str(), getTarget(),
1963 &OutputConstraintInfos);
1964
1965 InputConstraint = AddVariableConstraints(
1966 InputConstraint, *InputExpr->IgnoreParenNoopCasts(getContext()),
1967 getTarget(), CGM, S, false /* No EarlyClobber */);
1968
1969 llvm::Value *Arg = EmitAsmInput(Info, InputExpr, Constraints);
1970
1971 // If this input argument is tied to a larger output result, extend the
1972 // input to be the same size as the output. The LLVM backend wants to see
1973 // the input and output of a matching constraint be the same size. Note
1974 // that GCC does not define what the top bits are here. We use zext because
1975 // that is usually cheaper, but LLVM IR should really get an anyext someday.
1976 if (Info.hasTiedOperand()) {
1977 unsigned Output = Info.getTiedOperand();
1978 QualType OutputType = S.getOutputExpr(Output)->getType();
1979 QualType InputTy = InputExpr->getType();
1980
1981 if (getContext().getTypeSize(OutputType) >
1982 getContext().getTypeSize(InputTy)) {
1983 // Use ptrtoint as appropriate so that we can do our extension.
1984 if (isa<llvm::PointerType>(Arg->getType()))
1985 Arg = Builder.CreatePtrToInt(Arg, IntPtrTy);
1986 llvm::Type *OutputTy = ConvertType(OutputType);
1987 if (isa<llvm::IntegerType>(OutputTy))
1988 Arg = Builder.CreateZExt(Arg, OutputTy);
1989 else if (isa<llvm::PointerType>(OutputTy))
1990 Arg = Builder.CreateZExt(Arg, IntPtrTy);
1991 else {
1992 assert(OutputTy->isFloatingPointTy() && "Unexpected output type");
1993 Arg = Builder.CreateFPExt(Arg, OutputTy);
1994 }
1995 }
1996 }
1997 if (llvm::Type* AdjTy =
1998 getTargetHooks().adjustInlineAsmType(*this, InputConstraint,
1999 Arg->getType()))
2000 Arg = Builder.CreateBitCast(Arg, AdjTy);
2001 else
2002 CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input)
2003 << InputExpr->getType() << InputConstraint;
2004
2005 ArgTypes.push_back(Arg->getType());
2006 Args.push_back(Arg);
2007 Constraints += InputConstraint;
2008 }
2009
2010 // Append the "input" part of inout constraints last.
2011 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) {
2012 ArgTypes.push_back(InOutArgTypes[i]);
2013 Args.push_back(InOutArgs[i]);
2014 }
2015 Constraints += InOutConstraints;
2016
2017 // Clobbers
2018 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) {
2019 StringRef Clobber = S.getClobber(i);
2020
2021 if (Clobber == "memory")
2022 ReadOnly = ReadNone = false;
2023 else if (Clobber != "cc")
2024 Clobber = getTarget().getNormalizedGCCRegisterName(Clobber);
2025
2026 if (!Constraints.empty())
2027 Constraints += ',';
2028
2029 Constraints += "~{";
2030 Constraints += Clobber;
2031 Constraints += '}';
2032 }
2033
2034 // Add machine specific clobbers
2035 std::string MachineClobbers = getTarget().getClobbers();
2036 if (!MachineClobbers.empty()) {
2037 if (!Constraints.empty())
2038 Constraints += ',';
2039 Constraints += MachineClobbers;
2040 }
2041
2042 llvm::Type *ResultType;
2043 if (ResultRegTypes.empty())
2044 ResultType = VoidTy;
2045 else if (ResultRegTypes.size() == 1)
2046 ResultType = ResultRegTypes[0];
2047 else
2048 ResultType = llvm::StructType::get(getLLVMContext(), ResultRegTypes);
2049
2050 llvm::FunctionType *FTy =
2051 llvm::FunctionType::get(ResultType, ArgTypes, false);
2052
2053 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0;
2054 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(&S) ?
2055 llvm::InlineAsm::AD_Intel : llvm::InlineAsm::AD_ATT;
2056 llvm::InlineAsm *IA =
2057 llvm::InlineAsm::get(FTy, AsmString, Constraints, HasSideEffect,
2058 /* IsAlignStack */ false, AsmDialect);
2059 llvm::CallInst *Result = Builder.CreateCall(IA, Args);
2060 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2061 llvm::Attribute::NoUnwind);
2062
2063 if (isa<MSAsmStmt>(&S)) {
2064 // If the assembly contains any labels, mark the call noduplicate to prevent
2065 // defining the same ASM label twice (PR23715). This is pretty hacky, but it
2066 // works.
2067 if (AsmString.find("__MSASMLABEL_") != std::string::npos)
2068 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2069 llvm::Attribute::NoDuplicate);
2070 }
2071
2072 // Attach readnone and readonly attributes.
2073 if (!HasSideEffect) {
2074 if (ReadNone)
2075 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2076 llvm::Attribute::ReadNone);
2077 else if (ReadOnly)
2078 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2079 llvm::Attribute::ReadOnly);
2080 }
2081
2082 // Slap the source location of the inline asm into a !srcloc metadata on the
2083 // call.
2084 if (const GCCAsmStmt *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S)) {
2085 Result->setMetadata("srcloc", getAsmSrcLocInfo(gccAsmStmt->getAsmString(),
2086 *this));
2087 } else {
2088 // At least put the line number on MS inline asm blobs.
2089 auto Loc = llvm::ConstantInt::get(Int32Ty, S.getAsmLoc().getRawEncoding());
2090 Result->setMetadata("srcloc",
2091 llvm::MDNode::get(getLLVMContext(),
2092 llvm::ConstantAsMetadata::get(Loc)));
2093 }
2094
2095 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
2096 // Conservatively, mark all inline asm blocks in CUDA as convergent
2097 // (meaning, they may call an intrinsically convergent op, such as bar.sync,
2098 // and so can't have certain optimizations applied around them).
2099 Result->addAttribute(llvm::AttributeSet::FunctionIndex,
2100 llvm::Attribute::Convergent);
2101 }
2102
2103 // Extract all of the register value results from the asm.
2104 std::vector<llvm::Value*> RegResults;
2105 if (ResultRegTypes.size() == 1) {
2106 RegResults.push_back(Result);
2107 } else {
2108 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2109 llvm::Value *Tmp = Builder.CreateExtractValue(Result, i, "asmresult");
2110 RegResults.push_back(Tmp);
2111 }
2112 }
2113
2114 assert(RegResults.size() == ResultRegTypes.size());
2115 assert(RegResults.size() == ResultTruncRegTypes.size());
2116 assert(RegResults.size() == ResultRegDests.size());
2117 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) {
2118 llvm::Value *Tmp = RegResults[i];
2119
2120 // If the result type of the LLVM IR asm doesn't match the result type of
2121 // the expression, do the conversion.
2122 if (ResultRegTypes[i] != ResultTruncRegTypes[i]) {
2123 llvm::Type *TruncTy = ResultTruncRegTypes[i];
2124
2125 // Truncate the integer result to the right size, note that TruncTy can be
2126 // a pointer.
2127 if (TruncTy->isFloatingPointTy())
2128 Tmp = Builder.CreateFPTrunc(Tmp, TruncTy);
2129 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) {
2130 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(TruncTy);
2131 Tmp = Builder.CreateTrunc(Tmp,
2132 llvm::IntegerType::get(getLLVMContext(), (unsigned)ResSize));
2133 Tmp = Builder.CreateIntToPtr(Tmp, TruncTy);
2134 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) {
2135 uint64_t TmpSize =CGM.getDataLayout().getTypeSizeInBits(Tmp->getType());
2136 Tmp = Builder.CreatePtrToInt(Tmp,
2137 llvm::IntegerType::get(getLLVMContext(), (unsigned)TmpSize));
2138 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2139 } else if (TruncTy->isIntegerTy()) {
2140 Tmp = Builder.CreateTrunc(Tmp, TruncTy);
2141 } else if (TruncTy->isVectorTy()) {
2142 Tmp = Builder.CreateBitCast(Tmp, TruncTy);
2143 }
2144 }
2145
2146 EmitStoreThroughLValue(RValue::get(Tmp), ResultRegDests[i]);
2147 }
2148 }
2149
InitCapturedStruct(const CapturedStmt & S)2150 LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) {
2151 const RecordDecl *RD = S.getCapturedRecordDecl();
2152 QualType RecordTy = getContext().getRecordType(RD);
2153
2154 // Initialize the captured struct.
2155 LValue SlotLV =
2156 MakeAddrLValue(CreateMemTemp(RecordTy, "agg.captured"), RecordTy);
2157
2158 RecordDecl::field_iterator CurField = RD->field_begin();
2159 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
2160 E = S.capture_init_end();
2161 I != E; ++I, ++CurField) {
2162 LValue LV = EmitLValueForFieldInitialization(SlotLV, *CurField);
2163 if (CurField->hasCapturedVLAType()) {
2164 auto VAT = CurField->getCapturedVLAType();
2165 EmitStoreThroughLValue(RValue::get(VLASizeMap[VAT->getSizeExpr()]), LV);
2166 } else {
2167 EmitInitializerForField(*CurField, LV, *I, None);
2168 }
2169 }
2170
2171 return SlotLV;
2172 }
2173
2174 /// Generate an outlined function for the body of a CapturedStmt, store any
2175 /// captured variables into the captured struct, and call the outlined function.
2176 llvm::Function *
EmitCapturedStmt(const CapturedStmt & S,CapturedRegionKind K)2177 CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) {
2178 LValue CapStruct = InitCapturedStruct(S);
2179
2180 // Emit the CapturedDecl
2181 CodeGenFunction CGF(CGM, true);
2182 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K));
2183 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S);
2184 delete CGF.CapturedStmtInfo;
2185
2186 // Emit call to the helper function.
2187 EmitCallOrInvoke(F, CapStruct.getPointer());
2188
2189 return F;
2190 }
2191
GenerateCapturedStmtArgument(const CapturedStmt & S)2192 Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) {
2193 LValue CapStruct = InitCapturedStruct(S);
2194 return CapStruct.getAddress();
2195 }
2196
2197 /// Creates the outlined function for a CapturedStmt.
2198 llvm::Function *
GenerateCapturedStmtFunction(const CapturedStmt & S)2199 CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) {
2200 assert(CapturedStmtInfo &&
2201 "CapturedStmtInfo should be set when generating the captured function");
2202 const CapturedDecl *CD = S.getCapturedDecl();
2203 const RecordDecl *RD = S.getCapturedRecordDecl();
2204 SourceLocation Loc = S.getLocStart();
2205 assert(CD->hasBody() && "missing CapturedDecl body");
2206
2207 // Build the argument list.
2208 ASTContext &Ctx = CGM.getContext();
2209 FunctionArgList Args;
2210 Args.append(CD->param_begin(), CD->param_end());
2211
2212 // Create the function declaration.
2213 FunctionType::ExtInfo ExtInfo;
2214 const CGFunctionInfo &FuncInfo =
2215 CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, Args);
2216 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo);
2217
2218 llvm::Function *F =
2219 llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage,
2220 CapturedStmtInfo->getHelperName(), &CGM.getModule());
2221 CGM.SetInternalFunctionAttributes(CD, F, FuncInfo);
2222 if (CD->isNothrow())
2223 F->addFnAttr(llvm::Attribute::NoUnwind);
2224
2225 // Generate the function.
2226 StartFunction(CD, Ctx.VoidTy, F, FuncInfo, Args,
2227 CD->getLocation(),
2228 CD->getBody()->getLocStart());
2229 // Set the context parameter in CapturedStmtInfo.
2230 Address DeclPtr = GetAddrOfLocalVar(CD->getContextParam());
2231 CapturedStmtInfo->setContextValue(Builder.CreateLoad(DeclPtr));
2232
2233 // Initialize variable-length arrays.
2234 LValue Base = MakeNaturalAlignAddrLValue(CapturedStmtInfo->getContextValue(),
2235 Ctx.getTagDeclType(RD));
2236 for (auto *FD : RD->fields()) {
2237 if (FD->hasCapturedVLAType()) {
2238 auto *ExprArg = EmitLoadOfLValue(EmitLValueForField(Base, FD),
2239 S.getLocStart()).getScalarVal();
2240 auto VAT = FD->getCapturedVLAType();
2241 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
2242 }
2243 }
2244
2245 // If 'this' is captured, load it into CXXThisValue.
2246 if (CapturedStmtInfo->isCXXThisExprCaptured()) {
2247 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl();
2248 LValue ThisLValue = EmitLValueForField(Base, FD);
2249 CXXThisValue = EmitLoadOfLValue(ThisLValue, Loc).getScalarVal();
2250 }
2251
2252 PGO.assignRegionCounters(GlobalDecl(CD), F);
2253 CapturedStmtInfo->EmitBody(*this, CD->getBody());
2254 FinishFunction(CD->getBodyRBrace());
2255
2256 return F;
2257 }
2258