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