1//===--- CGStmt.cpp - Emit LLVM Code from Statements ----------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This contains code to emit Stmt nodes as LLVM code.
10//
11//===----------------------------------------------------------------------===//
12
13#include "CGDebugInfo.h"
14#include "CGOpenMPRuntime.h"
15#include "CodeGenFunction.h"
16#include "CodeGenModule.h"
17#include "TargetInfo.h"
18#include "clang/AST/Attr.h"
19#include "clang/AST/Expr.h"
20#include "clang/AST/Stmt.h"
21#include "clang/AST/StmtVisitor.h"
22#include "clang/Basic/Builtins.h"
23#include "clang/Basic/DiagnosticSema.h"
24#include "clang/Basic/PrettyStackTrace.h"
25#include "clang/Basic/SourceManager.h"
26#include "clang/Basic/TargetInfo.h"
27#include "llvm/ADT/SmallSet.h"
28#include "llvm/ADT/StringExtras.h"
29#include "llvm/IR/DataLayout.h"
30#include "llvm/IR/InlineAsm.h"
31#include "llvm/IR/Intrinsics.h"
32#include "llvm/IR/MDBuilder.h"
33#include "llvm/Support/SaveAndRestore.h"
34
35using namespace clang;
36using namespace CodeGen;
37
38//===----------------------------------------------------------------------===//
39// Statement Emission
40//===----------------------------------------------------------------------===//
41
42void CodeGenFunction::EmitStopPoint(const Stmt *S) {
43 if (CGDebugInfo *DI = getDebugInfo()) {
44 SourceLocation Loc;
45 Loc = S->getBeginLoc();
46 DI->EmitLocation(Builder, Loc);
47
48 LastStopPoint = Loc;
49 }
50}
51
52void CodeGenFunction::EmitStmt(const Stmt *S, ArrayRef<const Attr *> Attrs) {
53 assert(S && "Null statement?");
54 PGO.setCurrentStmt(S);
55
56 // These statements have their own debug info handling.
57 if (EmitSimpleStmt(S, Attrs))
58 return;
59
60 // Check if we are generating unreachable code.
61 if (!HaveInsertPoint()) {
62 // If so, and the statement doesn't contain a label, then we do not need to
63 // generate actual code. This is safe because (1) the current point is
64 // unreachable, so we don't need to execute the code, and (2) we've already
65 // handled the statements which update internal data structures (like the
66 // local variable map) which could be used by subsequent statements.
67 if (!ContainsLabel(S)) {
68 // Verify that any decl statements were handled as simple, they may be in
69 // scope of subsequent reachable statements.
70 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!");
71 return;
72 }
73
74 // Otherwise, make a new block to hold the code.
75 EnsureInsertPoint();
76 }
77
78 // Generate a stoppoint if we are emitting debug info.
79 EmitStopPoint(S);
80
81 // Ignore all OpenMP directives except for simd if OpenMP with Simd is
82 // enabled.
83 if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) {
84 if (const auto *D = dyn_cast<OMPExecutableDirective>(S)) {
85 EmitSimpleOMPExecutableDirective(*D);
86 return;
87 }
88 }
89
90 switch (S->getStmtClass()) {
91 case Stmt::NoStmtClass:
92 case Stmt::CXXCatchStmtClass:
93 case Stmt::SEHExceptStmtClass:
94 case Stmt::SEHFinallyStmtClass:
95 case Stmt::MSDependentExistsStmtClass:
96 llvm_unreachable("invalid statement class to emit generically");
97 case Stmt::NullStmtClass:
98 case Stmt::CompoundStmtClass:
99 case Stmt::DeclStmtClass:
100 case Stmt::LabelStmtClass:
101 case Stmt::AttributedStmtClass:
102 case Stmt::GotoStmtClass:
103 case Stmt::BreakStmtClass:
104 case Stmt::ContinueStmtClass:
105 case Stmt::DefaultStmtClass:
106 case Stmt::CaseStmtClass:
107 case Stmt::SEHLeaveStmtClass:
108 llvm_unreachable("should have emitted these statements as simple");
109
110#define STMT(Type, Base)
111#define ABSTRACT_STMT(Op)
112#define EXPR(Type, Base) \
113 case Stmt::Type##Class:
114#include "clang/AST/StmtNodes.inc"
115 {
116 // Remember the block we came in on.
117 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
118 assert(incoming && "expression emission must have an insertion point");
119
120 EmitIgnoredExpr(cast<Expr>(S));
121
122 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
123 assert(outgoing && "expression emission cleared block!");
124
125 // The expression emitters assume (reasonably!) that the insertion
126 // point is always set. To maintain that, the call-emission code
127 // for noreturn functions has to enter a new block with no
128 // predecessors. We want to kill that block and mark the current
129 // insertion point unreachable in the common case of a call like
130 // "exit();". Since expression emission doesn't otherwise create
131 // blocks with no predecessors, we can just test for that.
132 // However, we must be careful not to do this to our incoming
133 // block, because *statement* emission does sometimes create
134 // reachable blocks which will have no predecessors until later in
135 // the function. This occurs with, e.g., labels that are not
136 // reachable by fallthrough.
137 if (incoming != outgoing && outgoing->use_empty()) {
138 outgoing->eraseFromParent();
139 Builder.ClearInsertionPoint();
140 }
141 break;
142 }
143
144 case Stmt::IndirectGotoStmtClass:
145 EmitIndirectGotoStmt(cast<IndirectGotoStmt>(*S)); break;
146
147 case Stmt::IfStmtClass: EmitIfStmt(cast<IfStmt>(*S)); break;
148 case Stmt::WhileStmtClass: EmitWhileStmt(cast<WhileStmt>(*S), Attrs); break;
149 case Stmt::DoStmtClass: EmitDoStmt(cast<DoStmt>(*S), Attrs); break;
150 case Stmt::ForStmtClass: EmitForStmt(cast<ForStmt>(*S), Attrs); break;
151
152 case Stmt::ReturnStmtClass: EmitReturnStmt(cast<ReturnStmt>(*S)); break;
153
154 case Stmt::SwitchStmtClass: EmitSwitchStmt(cast<SwitchStmt>(*S)); break;
155 case Stmt::GCCAsmStmtClass: // Intentional fall-through.
156 case Stmt::MSAsmStmtClass: EmitAsmStmt(cast<AsmStmt>(*S)); break;
157 case Stmt::CoroutineBodyStmtClass:
158 EmitCoroutineBody(cast<CoroutineBodyStmt>(*S));
159 break;
160 case Stmt::CoreturnStmtClass:
161 EmitCoreturnStmt(cast<CoreturnStmt>(*S));
162 break;
163 case Stmt::CapturedStmtClass: {
164 const CapturedStmt *CS = cast<CapturedStmt>(S);
165 EmitCapturedStmt(*CS, CS->getCapturedRegionKind());
166 }
167 break;
168 case Stmt::ObjCAtTryStmtClass:
169 EmitObjCAtTryStmt(cast<ObjCAtTryStmt>(*S));
170 break;
171 case Stmt::ObjCAtCatchStmtClass:
172 llvm_unreachable(
173 "@catch statements should be handled by EmitObjCAtTryStmt");
174 case Stmt::ObjCAtFinallyStmtClass:
175 llvm_unreachable(
176 "@finally statements should be handled by EmitObjCAtTryStmt");
177 case Stmt::ObjCAtThrowStmtClass:
178 EmitObjCAtThrowStmt(cast<ObjCAtThrowStmt>(*S));
179 break;
180 case Stmt::ObjCAtSynchronizedStmtClass:
181 EmitObjCAtSynchronizedStmt(cast<ObjCAtSynchronizedStmt>(*S));
182 break;
183 case Stmt::ObjCForCollectionStmtClass:
184 EmitObjCForCollectionStmt(cast<ObjCForCollectionStmt>(*S));
185 break;
186 case Stmt::ObjCAutoreleasePoolStmtClass:
187 EmitObjCAutoreleasePoolStmt(cast<ObjCAutoreleasePoolStmt>(*S));
188 break;
189
190 case Stmt::CXXTryStmtClass:
191 EmitCXXTryStmt(cast<CXXTryStmt>(*S));
192 break;
193 case Stmt::CXXForRangeStmtClass:
194 EmitCXXForRangeStmt(cast<CXXForRangeStmt>(*S), Attrs);
195 break;
196 case Stmt::SEHTryStmtClass:
197 EmitSEHTryStmt(cast<SEHTryStmt>(*S));
198 break;
199 case Stmt::OMPCanonicalLoopClass:
200 EmitOMPCanonicalLoop(cast<OMPCanonicalLoop>(S));
201 break;
202 case Stmt::OMPParallelDirectiveClass:
203 EmitOMPParallelDirective(cast<OMPParallelDirective>(*S));
204 break;
205 case Stmt::OMPSimdDirectiveClass:
206 EmitOMPSimdDirective(cast<OMPSimdDirective>(*S));
207 break;
208 case Stmt::OMPTileDirectiveClass:
209 EmitOMPTileDirective(cast<OMPTileDirective>(*S));
210 break;
211 case Stmt::OMPForDirectiveClass:
212 EmitOMPForDirective(cast<OMPForDirective>(*S));
213 break;
214 case Stmt::OMPForSimdDirectiveClass:
215 EmitOMPForSimdDirective(cast<OMPForSimdDirective>(*S));
216 break;
217 case Stmt::OMPSectionsDirectiveClass:
218 EmitOMPSectionsDirective(cast<OMPSectionsDirective>(*S));
219 break;
220 case Stmt::OMPSectionDirectiveClass:
221 EmitOMPSectionDirective(cast<OMPSectionDirective>(*S));
222 break;
223 case Stmt::OMPSingleDirectiveClass:
224 EmitOMPSingleDirective(cast<OMPSingleDirective>(*S));
225 break;
226 case Stmt::OMPMasterDirectiveClass:
227 EmitOMPMasterDirective(cast<OMPMasterDirective>(*S));
228 break;
229 case Stmt::OMPCriticalDirectiveClass:
230 EmitOMPCriticalDirective(cast<OMPCriticalDirective>(*S));
231 break;
232 case Stmt::OMPParallelForDirectiveClass:
233 EmitOMPParallelForDirective(cast<OMPParallelForDirective>(*S));
234 break;
235 case Stmt::OMPParallelForSimdDirectiveClass:
236 EmitOMPParallelForSimdDirective(cast<OMPParallelForSimdDirective>(*S));
237 break;
238 case Stmt::OMPParallelMasterDirectiveClass:
239 EmitOMPParallelMasterDirective(cast<OMPParallelMasterDirective>(*S));
240 break;
241 case Stmt::OMPParallelSectionsDirectiveClass:
242 EmitOMPParallelSectionsDirective(cast<OMPParallelSectionsDirective>(*S));
243 break;
244 case Stmt::OMPTaskDirectiveClass:
245 EmitOMPTaskDirective(cast<OMPTaskDirective>(*S));
246 break;
247 case Stmt::OMPTaskyieldDirectiveClass:
248 EmitOMPTaskyieldDirective(cast<OMPTaskyieldDirective>(*S));
249 break;
250 case Stmt::OMPBarrierDirectiveClass:
251 EmitOMPBarrierDirective(cast<OMPBarrierDirective>(*S));
252 break;
253 case Stmt::OMPTaskwaitDirectiveClass:
254 EmitOMPTaskwaitDirective(cast<OMPTaskwaitDirective>(*S));
255 break;
256 case Stmt::OMPTaskgroupDirectiveClass:
257 EmitOMPTaskgroupDirective(cast<OMPTaskgroupDirective>(*S));
258 break;
259 case Stmt::OMPFlushDirectiveClass:
260 EmitOMPFlushDirective(cast<OMPFlushDirective>(*S));
261 break;
262 case Stmt::OMPDepobjDirectiveClass:
263 EmitOMPDepobjDirective(cast<OMPDepobjDirective>(*S));
264 break;
265 case Stmt::OMPScanDirectiveClass:
266 EmitOMPScanDirective(cast<OMPScanDirective>(*S));
267 break;
268 case Stmt::OMPOrderedDirectiveClass:
269 EmitOMPOrderedDirective(cast<OMPOrderedDirective>(*S));
270 break;
271 case Stmt::OMPAtomicDirectiveClass:
272 EmitOMPAtomicDirective(cast<OMPAtomicDirective>(*S));
273 break;
274 case Stmt::OMPTargetDirectiveClass:
275 EmitOMPTargetDirective(cast<OMPTargetDirective>(*S));
276 break;
277 case Stmt::OMPTeamsDirectiveClass:
278 EmitOMPTeamsDirective(cast<OMPTeamsDirective>(*S));
279 break;
280 case Stmt::OMPCancellationPointDirectiveClass:
281 EmitOMPCancellationPointDirective(cast<OMPCancellationPointDirective>(*S));
282 break;
283 case Stmt::OMPCancelDirectiveClass:
284 EmitOMPCancelDirective(cast<OMPCancelDirective>(*S));
285 break;
286 case Stmt::OMPTargetDataDirectiveClass:
287 EmitOMPTargetDataDirective(cast<OMPTargetDataDirective>(*S));
288 break;
289 case Stmt::OMPTargetEnterDataDirectiveClass:
290 EmitOMPTargetEnterDataDirective(cast<OMPTargetEnterDataDirective>(*S));
291 break;
292 case Stmt::OMPTargetExitDataDirectiveClass:
293 EmitOMPTargetExitDataDirective(cast<OMPTargetExitDataDirective>(*S));
294 break;
295 case Stmt::OMPTargetParallelDirectiveClass:
296 EmitOMPTargetParallelDirective(cast<OMPTargetParallelDirective>(*S));
297 break;
298 case Stmt::OMPTargetParallelForDirectiveClass:
299 EmitOMPTargetParallelForDirective(cast<OMPTargetParallelForDirective>(*S));
300 break;
301 case Stmt::OMPTaskLoopDirectiveClass:
302 EmitOMPTaskLoopDirective(cast<OMPTaskLoopDirective>(*S));
303 break;
304 case Stmt::OMPTaskLoopSimdDirectiveClass:
305 EmitOMPTaskLoopSimdDirective(cast<OMPTaskLoopSimdDirective>(*S));
306 break;
307 case Stmt::OMPMasterTaskLoopDirectiveClass:
308 EmitOMPMasterTaskLoopDirective(cast<OMPMasterTaskLoopDirective>(*S));
309 break;
310 case Stmt::OMPMasterTaskLoopSimdDirectiveClass:
311 EmitOMPMasterTaskLoopSimdDirective(
312 cast<OMPMasterTaskLoopSimdDirective>(*S));
313 break;
314 case Stmt::OMPParallelMasterTaskLoopDirectiveClass:
315 EmitOMPParallelMasterTaskLoopDirective(
316 cast<OMPParallelMasterTaskLoopDirective>(*S));
317 break;
318 case Stmt::OMPParallelMasterTaskLoopSimdDirectiveClass:
319 EmitOMPParallelMasterTaskLoopSimdDirective(
320 cast<OMPParallelMasterTaskLoopSimdDirective>(*S));
321 break;
322 case Stmt::OMPDistributeDirectiveClass:
323 EmitOMPDistributeDirective(cast<OMPDistributeDirective>(*S));
324 break;
325 case Stmt::OMPTargetUpdateDirectiveClass:
326 EmitOMPTargetUpdateDirective(cast<OMPTargetUpdateDirective>(*S));
327 break;
328 case Stmt::OMPDistributeParallelForDirectiveClass:
329 EmitOMPDistributeParallelForDirective(
330 cast<OMPDistributeParallelForDirective>(*S));
331 break;
332 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
333 EmitOMPDistributeParallelForSimdDirective(
334 cast<OMPDistributeParallelForSimdDirective>(*S));
335 break;
336 case Stmt::OMPDistributeSimdDirectiveClass:
337 EmitOMPDistributeSimdDirective(cast<OMPDistributeSimdDirective>(*S));
338 break;
339 case Stmt::OMPTargetParallelForSimdDirectiveClass:
340 EmitOMPTargetParallelForSimdDirective(
341 cast<OMPTargetParallelForSimdDirective>(*S));
342 break;
343 case Stmt::OMPTargetSimdDirectiveClass:
344 EmitOMPTargetSimdDirective(cast<OMPTargetSimdDirective>(*S));
345 break;
346 case Stmt::OMPTeamsDistributeDirectiveClass:
347 EmitOMPTeamsDistributeDirective(cast<OMPTeamsDistributeDirective>(*S));
348 break;
349 case Stmt::OMPTeamsDistributeSimdDirectiveClass:
350 EmitOMPTeamsDistributeSimdDirective(
351 cast<OMPTeamsDistributeSimdDirective>(*S));
352 break;
353 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass:
354 EmitOMPTeamsDistributeParallelForSimdDirective(
355 cast<OMPTeamsDistributeParallelForSimdDirective>(*S));
356 break;
357 case Stmt::OMPTeamsDistributeParallelForDirectiveClass:
358 EmitOMPTeamsDistributeParallelForDirective(
359 cast<OMPTeamsDistributeParallelForDirective>(*S));
360 break;
361 case Stmt::OMPTargetTeamsDirectiveClass:
362 EmitOMPTargetTeamsDirective(cast<OMPTargetTeamsDirective>(*S));
363 break;
364 case Stmt::OMPTargetTeamsDistributeDirectiveClass:
365 EmitOMPTargetTeamsDistributeDirective(
366 cast<OMPTargetTeamsDistributeDirective>(*S));
367 break;
368 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass:
369 EmitOMPTargetTeamsDistributeParallelForDirective(
370 cast<OMPTargetTeamsDistributeParallelForDirective>(*S));
371 break;
372 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass:
373 EmitOMPTargetTeamsDistributeParallelForSimdDirective(
374 cast<OMPTargetTeamsDistributeParallelForSimdDirective>(*S));
375 break;
376 case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass:
377 EmitOMPTargetTeamsDistributeSimdDirective(
378 cast<OMPTargetTeamsDistributeSimdDirective>(*S));
379 break;
380 case Stmt::OMPInteropDirectiveClass:
381 llvm_unreachable("Interop directive not supported yet.");
382 break;
383 case Stmt::OMPDispatchDirectiveClass:
384 llvm_unreachable("Dispatch directive not supported yet.");
385 break;
386 case Stmt::OMPMaskedDirectiveClass:
387 EmitOMPMaskedDirective(cast<OMPMaskedDirective>(*S));
388 break;
389 }
390}
391
392bool CodeGenFunction::EmitSimpleStmt(const Stmt *S,
393 ArrayRef<const Attr *> Attrs) {
394 switch (S->getStmtClass()) {
395 default:
396 return false;
397 case Stmt::NullStmtClass:
398 break;
399 case Stmt::CompoundStmtClass:
400 EmitCompoundStmt(cast<CompoundStmt>(*S));
401 break;
402 case Stmt::DeclStmtClass:
403 EmitDeclStmt(cast<DeclStmt>(*S));
404 break;
405 case Stmt::LabelStmtClass:
406 EmitLabelStmt(cast<LabelStmt>(*S));
407 break;
408 case Stmt::AttributedStmtClass:
409 EmitAttributedStmt(cast<AttributedStmt>(*S));
410 break;
411 case Stmt::GotoStmtClass:
412 EmitGotoStmt(cast<GotoStmt>(*S));
413 break;
414 case Stmt::BreakStmtClass:
415 EmitBreakStmt(cast<BreakStmt>(*S));
416 break;
417 case Stmt::ContinueStmtClass:
418 EmitContinueStmt(cast<ContinueStmt>(*S));
419 break;
420 case Stmt::DefaultStmtClass:
421 EmitDefaultStmt(cast<DefaultStmt>(*S), Attrs);
422 break;
423 case Stmt::CaseStmtClass:
424 EmitCaseStmt(cast<CaseStmt>(*S), Attrs);
425 break;
426 case Stmt::SEHLeaveStmtClass:
427 EmitSEHLeaveStmt(cast<SEHLeaveStmt>(*S));
428 break;
429 }
430 return true;
431}
432
433/// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
434/// this captures the expression result of the last sub-statement and returns it
435/// (for use by the statement expression extension).
436Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
437 AggValueSlot AggSlot) {
438 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
439 "LLVM IR generation of compound statement ('{}')");
440
441 // Keep track of the current cleanup stack depth, including debug scopes.
442 LexicalScope Scope(*this, S.getSourceRange());
443
444 return EmitCompoundStmtWithoutScope(S, GetLast, AggSlot);
445}
446
447Address
448CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
449 bool GetLast,
450 AggValueSlot AggSlot) {
451
452 const Stmt *ExprResult = S.getStmtExprResult();
453 assert((!GetLast || (GetLast && ExprResult)) &&
454 "If GetLast is true then the CompoundStmt must have a StmtExprResult");
455
456 Address RetAlloca = Address::invalid();
457
458 for (auto *CurStmt : S.body()) {
459 if (GetLast && ExprResult == CurStmt) {
460 // We have to special case labels here. They are statements, but when put
461 // at the end of a statement expression, they yield the value of their
462 // subexpression. Handle this by walking through all labels we encounter,
463 // emitting them before we evaluate the subexpr.
464 // Similar issues arise for attributed statements.
465 while (!isa<Expr>(ExprResult)) {
466 if (const auto *LS = dyn_cast<LabelStmt>(ExprResult)) {
467 EmitLabel(LS->getDecl());
468 ExprResult = LS->getSubStmt();
469 } else if (const auto *AS = dyn_cast<AttributedStmt>(ExprResult)) {
470 // FIXME: Update this if we ever have attributes that affect the
471 // semantics of an expression.
472 ExprResult = AS->getSubStmt();
473 } else {
474 llvm_unreachable("unknown value statement");
475 }
476 }
477
478 EnsureInsertPoint();
479
480 const Expr *E = cast<Expr>(ExprResult);
481 QualType ExprTy = E->getType();
482 if (hasAggregateEvaluationKind(ExprTy)) {
483 EmitAggExpr(E, AggSlot);
484 } else {
485 // We can't return an RValue here because there might be cleanups at
486 // the end of the StmtExpr. Because of that, we have to emit the result
487 // here into a temporary alloca.
488 RetAlloca = CreateMemTemp(ExprTy);
489 EmitAnyExprToMem(E, RetAlloca, Qualifiers(),
490 /*IsInit*/ false);
491 }
492 } else {
493 EmitStmt(CurStmt);
494 }
495 }
496
497 return RetAlloca;
498}
499
500void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
501 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(BB->getTerminator());
502
503 // If there is a cleanup stack, then we it isn't worth trying to
504 // simplify this block (we would need to remove it from the scope map
505 // and cleanup entry).
506 if (!EHStack.empty())
507 return;
508
509 // Can only simplify direct branches.
510 if (!BI || !BI->isUnconditional())
511 return;
512
513 // Can only simplify empty blocks.
514 if (BI->getIterator() != BB->begin())
515 return;
516
517 BB->replaceAllUsesWith(BI->getSuccessor(0));
518 BI->eraseFromParent();
519 BB->eraseFromParent();
520}
521
522void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
523 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
524
525 // Fall out of the current block (if necessary).
526 EmitBranch(BB);
527
528 if (IsFinished && BB->use_empty()) {
529 delete BB;
530 return;
531 }
532
533 // Place the block after the current block, if possible, or else at
534 // the end of the function.
535 if (CurBB && CurBB->getParent())
536 CurFn->getBasicBlockList().insertAfter(CurBB->getIterator(), BB);
537 else
538 CurFn->getBasicBlockList().push_back(BB);
539 Builder.SetInsertPoint(BB);
540}
541
542void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
543 // Emit a branch from the current block to the target one if this
544 // was a real block. If this was just a fall-through block after a
545 // terminator, don't emit it.
546 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
547
548 if (!CurBB || CurBB->getTerminator()) {
549 // If there is no insert point or the previous block is already
550 // terminated, don't touch it.
551 } else {
552 // Otherwise, create a fall-through branch.
553 Builder.CreateBr(Target);
554 }
555
556 Builder.ClearInsertionPoint();
557}
558
559void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
560 bool inserted = false;
561 for (llvm::User *u : block->users()) {
562 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(u)) {
563 CurFn->getBasicBlockList().insertAfter(insn->getParent()->getIterator(),
564 block);
565 inserted = true;
566 break;
567 }
568 }
569
570 if (!inserted)
571 CurFn->getBasicBlockList().push_back(block);
572
573 Builder.SetInsertPoint(block);
574}
575
576CodeGenFunction::JumpDest
577CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
578 JumpDest &Dest = LabelMap[D];
579 if (Dest.isValid()) return Dest;
580
581 // Create, but don't insert, the new block.
582 Dest = JumpDest(createBasicBlock(D->getName()),
583 EHScopeStack::stable_iterator::invalid(),
584 NextCleanupDestIndex++);
585 return Dest;
586}
587
588void CodeGenFunction::EmitLabel(const LabelDecl *D) {
589 // Add this label to the current lexical scope if we're within any
590 // normal cleanups. Jumps "in" to this label --- when permitted by
591 // the language --- may need to be routed around such cleanups.
592 if (EHStack.hasNormalCleanups() && CurLexicalScope)
593 CurLexicalScope->addLabel(D);
594
595 JumpDest &Dest = LabelMap[D];
596
597 // If we didn't need a forward reference to this label, just go
598 // ahead and create a destination at the current scope.
599 if (!Dest.isValid()) {
600 Dest = getJumpDestInCurrentScope(D->getName());
601
602 // Otherwise, we need to give this label a target depth and remove
603 // it from the branch-fixups list.
604 } else {
605 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
606 Dest.setScopeDepth(EHStack.stable_begin());
607 ResolveBranchFixups(Dest.getBlock());
608 }
609
610 EmitBlock(Dest.getBlock());
611
612 // Emit debug info for labels.
613 if (CGDebugInfo *DI = getDebugInfo()) {
614 if (CGM.getCodeGenOpts().hasReducedDebugInfo()) {
615 DI->setLocation(D->getLocation());
616 DI->EmitLabel(D, Builder);
617 }
618 }
619
620 incrementProfileCounter(D->getStmt());
621}
622
623/// Change the cleanup scope of the labels in this lexical scope to
624/// match the scope of the enclosing context.
625void CodeGenFunction::LexicalScope::rescopeLabels() {
626 assert(!Labels.empty());
627 EHScopeStack::stable_iterator innermostScope
628 = CGF.EHStack.getInnermostNormalCleanup();
629
630 // Change the scope depth of all the labels.
631 for (SmallVectorImpl<const LabelDecl*>::const_iterator
632 i = Labels.begin(), e = Labels.end(); i != e; ++i) {
633 assert(CGF.LabelMap.count(*i));
634 JumpDest &dest = CGF.LabelMap.find(*i)->second;
635 assert(dest.getScopeDepth().isValid());
636 assert(innermostScope.encloses(dest.getScopeDepth()));
637 dest.setScopeDepth(innermostScope);
638 }
639
640 // Reparent the labels if the new scope also has cleanups.
641 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
642 ParentScope->Labels.append(Labels.begin(), Labels.end());
643 }
644}
645
646
647void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
648 EmitLabel(S.getDecl());
649 EmitStmt(S.getSubStmt());
650}
651
652void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
653 bool nomerge = false;
654 const CallExpr *musttail = nullptr;
655
656 for (const auto *A : S.getAttrs()) {
657 if (A->getKind() == attr::NoMerge) {
658 nomerge = true;
659 }
660 if (A->getKind() == attr::MustTail) {
661 const Stmt *Sub = S.getSubStmt();
662 const ReturnStmt *R = cast<ReturnStmt>(Sub);
663 musttail = cast<CallExpr>(R->getRetValue()->IgnoreParens());
664 }
665 }
666 SaveAndRestore<bool> save_nomerge(InNoMergeAttributedStmt, nomerge);
667 SaveAndRestore<const CallExpr *> save_musttail(MustTailCall, musttail);
668 EmitStmt(S.getSubStmt(), S.getAttrs());
669}
670
671void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
672 // If this code is reachable then emit a stop point (if generating
673 // debug info). We have to do this ourselves because we are on the
674 // "simple" statement path.
675 if (HaveInsertPoint())
676 EmitStopPoint(&S);
677
678 EmitBranchThroughCleanup(getJumpDestForLabel(S.getLabel()));
679}
680
681
682void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
683 if (const LabelDecl *Target = S.getConstantTarget()) {
684 EmitBranchThroughCleanup(getJumpDestForLabel(Target));
685 return;
686 }
687
688 // Ensure that we have an i8* for our PHI node.
689 llvm::Value *V = Builder.CreateBitCast(EmitScalarExpr(S.getTarget()),
690 Int8PtrTy, "addr");
691 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
692
693 // Get the basic block for the indirect goto.
694 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
695
696 // The first instruction in the block has to be the PHI for the switch dest,
697 // add an entry for this branch.
698 cast<llvm::PHINode>(IndGotoBB->begin())->addIncoming(V, CurBB);
699
700 EmitBranch(IndGotoBB);
701}
702
703void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
704 // C99 6.8.4.1: The first substatement is executed if the expression compares
705 // unequal to 0. The condition must be a scalar type.
706 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
707
708 if (S.getInit())
709 EmitStmt(S.getInit());
710
711 if (S.getConditionVariable())
712 EmitDecl(*S.getConditionVariable());
713
714 // If the condition constant folds and can be elided, try to avoid emitting
715 // the condition and the dead arm of the if/else.
716 bool CondConstant;
717 if (ConstantFoldsToSimpleInteger(S.getCond(), CondConstant,
718 S.isConstexpr())) {
719 // Figure out which block (then or else) is executed.
720 const Stmt *Executed = S.getThen();
721 const Stmt *Skipped = S.getElse();
722 if (!CondConstant) // Condition false?
723 std::swap(Executed, Skipped);
724
725 // If the skipped block has no labels in it, just emit the executed block.
726 // This avoids emitting dead code and simplifies the CFG substantially.
727 if (S.isConstexpr() || !ContainsLabel(Skipped)) {
728 if (CondConstant)
729 incrementProfileCounter(&S);
730 if (Executed) {
731 RunCleanupsScope ExecutedScope(*this);
732 EmitStmt(Executed);
733 }
734 return;
735 }
736 }
737
738 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
739 // the conditional branch.
740 llvm::BasicBlock *ThenBlock = createBasicBlock("if.then");
741 llvm::BasicBlock *ContBlock = createBasicBlock("if.end");
742 llvm::BasicBlock *ElseBlock = ContBlock;
743 if (S.getElse())
744 ElseBlock = createBasicBlock("if.else");
745
746 // Prefer the PGO based weights over the likelihood attribute.
747 // When the build isn't optimized the metadata isn't used, so don't generate
748 // it.
749 Stmt::Likelihood LH = Stmt::LH_None;
750 uint64_t Count = getProfileCount(S.getThen());
751 if (!Count && CGM.getCodeGenOpts().OptimizationLevel)
752 LH = Stmt::getLikelihood(S.getThen(), S.getElse());
753 EmitBranchOnBoolExpr(S.getCond(), ThenBlock, ElseBlock, Count, LH);
754
755 // Emit the 'then' code.
756 EmitBlock(ThenBlock);
757 incrementProfileCounter(&S);
758 {
759 RunCleanupsScope ThenScope(*this);
760 EmitStmt(S.getThen());
761 }
762 EmitBranch(ContBlock);
763
764 // Emit the 'else' code if present.
765 if (const Stmt *Else = S.getElse()) {
766 {
767 // There is no need to emit line number for an unconditional branch.
768 auto NL = ApplyDebugLocation::CreateEmpty(*this);
769 EmitBlock(ElseBlock);
770 }
771 {
772 RunCleanupsScope ElseScope(*this);
773 EmitStmt(Else);
774 }
775 {
776 // There is no need to emit line number for an unconditional branch.
777 auto NL = ApplyDebugLocation::CreateEmpty(*this);
778 EmitBranch(ContBlock);
779 }
780 }
781
782 // Emit the continuation block for code after the if.
783 EmitBlock(ContBlock, true);
784}
785
786void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
787 ArrayRef<const Attr *> WhileAttrs) {
788 // Emit the header for the loop, which will also become
789 // the continue target.
790 JumpDest LoopHeader = getJumpDestInCurrentScope("while.cond");
791 EmitBlock(LoopHeader.getBlock());
792
793 // Create an exit block for when the condition fails, which will
794 // also become the break target.
795 JumpDest LoopExit = getJumpDestInCurrentScope("while.end");
796
797 // Store the blocks to use for break and continue.
798 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopHeader));
799
800 // C++ [stmt.while]p2:
801 // When the condition of a while statement is a declaration, the
802 // scope of the variable that is declared extends from its point
803 // of declaration (3.3.2) to the end of the while statement.
804 // [...]
805 // The object created in a condition is destroyed and created
806 // with each iteration of the loop.
807 RunCleanupsScope ConditionScope(*this);
808
809 if (S.getConditionVariable())
810 EmitDecl(*S.getConditionVariable());
811
812 // Evaluate the conditional in the while header. C99 6.8.5.1: The
813 // evaluation of the controlling expression takes place before each
814 // execution of the loop body.
815 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
816
817 // while(1) is common, avoid extra exit blocks. Be sure
818 // to correctly handle break/continue though.
819 llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal);
820 bool CondIsConstInt = C != nullptr;
821 bool EmitBoolCondBranch = !CondIsConstInt || !C->isOne();
822 const SourceRange &R = S.getSourceRange();
823 LoopStack.push(LoopHeader.getBlock(), CGM.getContext(), CGM.getCodeGenOpts(),
824 WhileAttrs, SourceLocToDebugLoc(R.getBegin()),
825 SourceLocToDebugLoc(R.getEnd()),
826 checkIfLoopMustProgress(CondIsConstInt));
827
828 // As long as the condition is true, go to the loop body.
829 llvm::BasicBlock *LoopBody = createBasicBlock("while.body");
830 if (EmitBoolCondBranch) {
831 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
832 if (ConditionScope.requiresCleanups())
833 ExitBlock = createBasicBlock("while.exit");
834 llvm::MDNode *Weights =
835 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()));
836 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
837 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
838 BoolCondVal, Stmt::getLikelihood(S.getBody()));
839 Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock, Weights);
840
841 if (ExitBlock != LoopExit.getBlock()) {
842 EmitBlock(ExitBlock);
843 EmitBranchThroughCleanup(LoopExit);
844 }
845 } else if (const Attr *A = Stmt::getLikelihoodAttr(S.getBody())) {
846 CGM.getDiags().Report(A->getLocation(),
847 diag::warn_attribute_has_no_effect_on_infinite_loop)
848 << A << A->getRange();
849 CGM.getDiags().Report(
850 S.getWhileLoc(),
851 diag::note_attribute_has_no_effect_on_infinite_loop_here)
852 << SourceRange(S.getWhileLoc(), S.getRParenLoc());
853 }
854
855 // Emit the loop body. We have to emit this in a cleanup scope
856 // because it might be a singleton DeclStmt.
857 {
858 RunCleanupsScope BodyScope(*this);
859 EmitBlock(LoopBody);
860 incrementProfileCounter(&S);
861 EmitStmt(S.getBody());
862 }
863
864 BreakContinueStack.pop_back();
865
866 // Immediately force cleanup.
867 ConditionScope.ForceCleanup();
868
869 EmitStopPoint(&S);
870 // Branch to the loop header again.
871 EmitBranch(LoopHeader.getBlock());
872
873 LoopStack.pop();
874
875 // Emit the exit block.
876 EmitBlock(LoopExit.getBlock(), true);
877
878 // The LoopHeader typically is just a branch if we skipped emitting
879 // a branch, try to erase it.
880 if (!EmitBoolCondBranch)
881 SimplifyForwardingBlocks(LoopHeader.getBlock());
882}
883
884void CodeGenFunction::EmitDoStmt(const DoStmt &S,
885 ArrayRef<const Attr *> DoAttrs) {
886 JumpDest LoopExit = getJumpDestInCurrentScope("do.end");
887 JumpDest LoopCond = getJumpDestInCurrentScope("do.cond");
888
889 uint64_t ParentCount = getCurrentProfileCount();
890
891 // Store the blocks to use for break and continue.
892 BreakContinueStack.push_back(BreakContinue(LoopExit, LoopCond));
893
894 // Emit the body of the loop.
895 llvm::BasicBlock *LoopBody = createBasicBlock("do.body");
896
897 EmitBlockWithFallThrough(LoopBody, &S);
898 {
899 RunCleanupsScope BodyScope(*this);
900 EmitStmt(S.getBody());
901 }
902
903 EmitBlock(LoopCond.getBlock());
904
905 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
906 // after each execution of the loop body."
907
908 // Evaluate the conditional in the while header.
909 // C99 6.8.5p2/p4: The first substatement is executed if the expression
910 // compares unequal to 0. The condition must be a scalar type.
911 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
912
913 BreakContinueStack.pop_back();
914
915 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
916 // to correctly handle break/continue though.
917 llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(BoolCondVal);
918 bool CondIsConstInt = C;
919 bool EmitBoolCondBranch = !C || !C->isZero();
920
921 const SourceRange &R = S.getSourceRange();
922 LoopStack.push(LoopBody, CGM.getContext(), CGM.getCodeGenOpts(), DoAttrs,
923 SourceLocToDebugLoc(R.getBegin()),
924 SourceLocToDebugLoc(R.getEnd()),
925 checkIfLoopMustProgress(CondIsConstInt));
926
927 // As long as the condition is true, iterate the loop.
928 if (EmitBoolCondBranch) {
929 uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
930 Builder.CreateCondBr(
931 BoolCondVal, LoopBody, LoopExit.getBlock(),
932 createProfileWeightsForLoop(S.getCond(), BackedgeCount));
933 }
934
935 LoopStack.pop();
936
937 // Emit the exit block.
938 EmitBlock(LoopExit.getBlock());
939
940 // The DoCond block typically is just a branch if we skipped
941 // emitting a branch, try to erase it.
942 if (!EmitBoolCondBranch)
943 SimplifyForwardingBlocks(LoopCond.getBlock());
944}
945
946void CodeGenFunction::EmitForStmt(const ForStmt &S,
947 ArrayRef<const Attr *> ForAttrs) {
948 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
949
950 LexicalScope ForScope(*this, S.getSourceRange());
951
952 // Evaluate the first part before the loop.
953 if (S.getInit())
954 EmitStmt(S.getInit());
955
956 // Start the loop with a block that tests the condition.
957 // If there's an increment, the continue scope will be overwritten
958 // later.
959 JumpDest CondDest = getJumpDestInCurrentScope("for.cond");
960 llvm::BasicBlock *CondBlock = CondDest.getBlock();
961 EmitBlock(CondBlock);
962
963 Expr::EvalResult Result;
964 bool CondIsConstInt =
965 !S.getCond() || S.getCond()->EvaluateAsInt(Result, getContext());
966
967 const SourceRange &R = S.getSourceRange();
968 LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs,
969 SourceLocToDebugLoc(R.getBegin()),
970 SourceLocToDebugLoc(R.getEnd()),
971 checkIfLoopMustProgress(CondIsConstInt));
972
973 // Create a cleanup scope for the condition variable cleanups.
974 LexicalScope ConditionScope(*this, S.getSourceRange());
975
976 // If the for loop doesn't have an increment we can just use the condition as
977 // the continue block. Otherwise, if there is no condition variable, we can
978 // form the continue block now. If there is a condition variable, we can't
979 // form the continue block until after we've emitted the condition, because
980 // the condition is in scope in the increment, but Sema's jump diagnostics
981 // ensure that there are no continues from the condition variable that jump
982 // to the loop increment.
983 JumpDest Continue;
984 if (!S.getInc())
985 Continue = CondDest;
986 else if (!S.getConditionVariable())
987 Continue = getJumpDestInCurrentScope("for.inc");
988 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
989
990 if (S.getCond()) {
991 // If the for statement has a condition scope, emit the local variable
992 // declaration.
993 if (S.getConditionVariable()) {
994 EmitDecl(*S.getConditionVariable());
995
996 // We have entered the condition variable's scope, so we're now able to
997 // jump to the continue block.
998 Continue = S.getInc() ? getJumpDestInCurrentScope("for.inc") : CondDest;
999 BreakContinueStack.back().ContinueBlock = Continue;
1000 }
1001
1002 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1003 // If there are any cleanups between here and the loop-exit scope,
1004 // create a block to stage a loop exit along.
1005 if (ForScope.requiresCleanups())
1006 ExitBlock = createBasicBlock("for.cond.cleanup");
1007
1008 // As long as the condition is true, iterate the loop.
1009 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
1010
1011 // C99 6.8.5p2/p4: The first substatement is executed if the expression
1012 // compares unequal to 0. The condition must be a scalar type.
1013 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
1014 llvm::MDNode *Weights =
1015 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()));
1016 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
1017 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
1018 BoolCondVal, Stmt::getLikelihood(S.getBody()));
1019
1020 Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights);
1021
1022 if (ExitBlock != LoopExit.getBlock()) {
1023 EmitBlock(ExitBlock);
1024 EmitBranchThroughCleanup(LoopExit);
1025 }
1026
1027 EmitBlock(ForBody);
1028 } else {
1029 // Treat it as a non-zero constant. Don't even create a new block for the
1030 // body, just fall into it.
1031 }
1032 incrementProfileCounter(&S);
1033
1034 {
1035 // Create a separate cleanup scope for the body, in case it is not
1036 // a compound statement.
1037 RunCleanupsScope BodyScope(*this);
1038 EmitStmt(S.getBody());
1039 }
1040
1041 // If there is an increment, emit it next.
1042 if (S.getInc()) {
1043 EmitBlock(Continue.getBlock());
1044 EmitStmt(S.getInc());
1045 }
1046
1047 BreakContinueStack.pop_back();
1048
1049 ConditionScope.ForceCleanup();
1050
1051 EmitStopPoint(&S);
1052 EmitBranch(CondBlock);
1053
1054 ForScope.ForceCleanup();
1055
1056 LoopStack.pop();
1057
1058 // Emit the fall-through block.
1059 EmitBlock(LoopExit.getBlock(), true);
1060}
1061
1062void
1063CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
1064 ArrayRef<const Attr *> ForAttrs) {
1065 JumpDest LoopExit = getJumpDestInCurrentScope("for.end");
1066
1067 LexicalScope ForScope(*this, S.getSourceRange());
1068
1069 // Evaluate the first pieces before the loop.
1070 if (S.getInit())
1071 EmitStmt(S.getInit());
1072 EmitStmt(S.getRangeStmt());
1073 EmitStmt(S.getBeginStmt());
1074 EmitStmt(S.getEndStmt());
1075
1076 // Start the loop with a block that tests the condition.
1077 // If there's an increment, the continue scope will be overwritten
1078 // later.
1079 llvm::BasicBlock *CondBlock = createBasicBlock("for.cond");
1080 EmitBlock(CondBlock);
1081
1082 const SourceRange &R = S.getSourceRange();
1083 LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), ForAttrs,
1084 SourceLocToDebugLoc(R.getBegin()),
1085 SourceLocToDebugLoc(R.getEnd()));
1086
1087 // If there are any cleanups between here and the loop-exit scope,
1088 // create a block to stage a loop exit along.
1089 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1090 if (ForScope.requiresCleanups())
1091 ExitBlock = createBasicBlock("for.cond.cleanup");
1092
1093 // The loop body, consisting of the specified body and the loop variable.
1094 llvm::BasicBlock *ForBody = createBasicBlock("for.body");
1095
1096 // The body is executed if the expression, contextually converted
1097 // to bool, is true.
1098 llvm::Value *BoolCondVal = EvaluateExprAsBool(S.getCond());
1099 llvm::MDNode *Weights =
1100 createProfileWeightsForLoop(S.getCond(), getProfileCount(S.getBody()));
1101 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
1102 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
1103 BoolCondVal, Stmt::getLikelihood(S.getBody()));
1104 Builder.CreateCondBr(BoolCondVal, ForBody, ExitBlock, Weights);
1105
1106 if (ExitBlock != LoopExit.getBlock()) {
1107 EmitBlock(ExitBlock);
1108 EmitBranchThroughCleanup(LoopExit);
1109 }
1110
1111 EmitBlock(ForBody);
1112 incrementProfileCounter(&S);
1113
1114 // Create a block for the increment. In case of a 'continue', we jump there.
1115 JumpDest Continue = getJumpDestInCurrentScope("for.inc");
1116
1117 // Store the blocks to use for break and continue.
1118 BreakContinueStack.push_back(BreakContinue(LoopExit, Continue));
1119
1120 {
1121 // Create a separate cleanup scope for the loop variable and body.
1122 LexicalScope BodyScope(*this, S.getSourceRange());
1123 EmitStmt(S.getLoopVarStmt());
1124 EmitStmt(S.getBody());
1125 }
1126
1127 EmitStopPoint(&S);
1128 // If there is an increment, emit it next.
1129 EmitBlock(Continue.getBlock());
1130 EmitStmt(S.getInc());
1131
1132 BreakContinueStack.pop_back();
1133
1134 EmitBranch(CondBlock);
1135
1136 ForScope.ForceCleanup();
1137
1138 LoopStack.pop();
1139
1140 // Emit the fall-through block.
1141 EmitBlock(LoopExit.getBlock(), true);
1142}
1143
1144void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
1145 if (RV.isScalar()) {
1146 Builder.CreateStore(RV.getScalarVal(), ReturnValue);
1147 } else if (RV.isAggregate()) {
1148 LValue Dest = MakeAddrLValue(ReturnValue, Ty);
1149 LValue Src = MakeAddrLValue(RV.getAggregateAddress(), Ty);
1150 EmitAggregateCopy(Dest, Src, Ty, getOverlapForReturnValue());
1151 } else {
1152 EmitStoreOfComplex(RV.getComplexVal(), MakeAddrLValue(ReturnValue, Ty),
1153 /*init*/ true);
1154 }
1155 EmitBranchThroughCleanup(ReturnBlock);
1156}
1157
1158namespace {
1159// RAII struct used to save and restore a return statment's result expression.
1160struct SaveRetExprRAII {
1161 SaveRetExprRAII(const Expr *RetExpr, CodeGenFunction &CGF)
1162 : OldRetExpr(CGF.RetExpr), CGF(CGF) {
1163 CGF.RetExpr = RetExpr;
1164 }
1165 ~SaveRetExprRAII() { CGF.RetExpr = OldRetExpr; }
1166 const Expr *OldRetExpr;
1167 CodeGenFunction &CGF;
1168};
1169} // namespace
1170
1171/// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
1172/// if the function returns void, or may be missing one if the function returns
1173/// non-void. Fun stuff :).
1174void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
1175 if (requiresReturnValueCheck()) {
1176 llvm::Constant *SLoc = EmitCheckSourceLocation(S.getBeginLoc());
1177 auto *SLocPtr =
1178 new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false,
1179 llvm::GlobalVariable::PrivateLinkage, SLoc);
1180 SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
1181 CGM.getSanitizerMetadata()->disableSanitizerForGlobal(SLocPtr);
1182 assert(ReturnLocation.isValid() && "No valid return location");
1183 Builder.CreateStore(Builder.CreateBitCast(SLocPtr, Int8PtrTy),
1184 ReturnLocation);
1185 }
1186
1187 // Returning from an outlined SEH helper is UB, and we already warn on it.
1188 if (IsOutlinedSEHHelper) {
1189 Builder.CreateUnreachable();
1190 Builder.ClearInsertionPoint();
1191 }
1192
1193 // Emit the result value, even if unused, to evaluate the side effects.
1194 const Expr *RV = S.getRetValue();
1195
1196 // Record the result expression of the return statement. The recorded
1197 // expression is used to determine whether a block capture's lifetime should
1198 // end at the end of the full expression as opposed to the end of the scope
1199 // enclosing the block expression.
1200 //
1201 // This permits a small, easily-implemented exception to our over-conservative
1202 // rules about not jumping to statements following block literals with
1203 // non-trivial cleanups.
1204 SaveRetExprRAII SaveRetExpr(RV, *this);
1205
1206 RunCleanupsScope cleanupScope(*this);
1207 if (const auto *EWC = dyn_cast_or_null<ExprWithCleanups>(RV))
1208 RV = EWC->getSubExpr();
1209 // FIXME: Clean this up by using an LValue for ReturnTemp,
1210 // EmitStoreThroughLValue, and EmitAnyExpr.
1211 // Check if the NRVO candidate was not globalized in OpenMP mode.
1212 if (getLangOpts().ElideConstructors && S.getNRVOCandidate() &&
1213 S.getNRVOCandidate()->isNRVOVariable() &&
1214 (!getLangOpts().OpenMP ||
1215 !CGM.getOpenMPRuntime()
1216 .getAddressOfLocalVariable(*this, S.getNRVOCandidate())
1217 .isValid())) {
1218 // Apply the named return value optimization for this return statement,
1219 // which means doing nothing: the appropriate result has already been
1220 // constructed into the NRVO variable.
1221
1222 // If there is an NRVO flag for this variable, set it to 1 into indicate
1223 // that the cleanup code should not destroy the variable.
1224 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1225 Builder.CreateFlagStore(Builder.getTrue(), NRVOFlag);
1226 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1227 // Make sure not to return anything, but evaluate the expression
1228 // for side effects.
1229 if (RV)
1230 EmitAnyExpr(RV);
1231 } else if (!RV) {
1232 // Do nothing (return value is left uninitialized)
1233 } else if (FnRetTy->isReferenceType()) {
1234 // If this function returns a reference, take the address of the expression
1235 // rather than the value.
1236 RValue Result = EmitReferenceBindingToExpr(RV);
1237 Builder.CreateStore(Result.getScalarVal(), ReturnValue);
1238 } else {
1239 switch (getEvaluationKind(RV->getType())) {
1240 case TEK_Scalar:
1241 Builder.CreateStore(EmitScalarExpr(RV), ReturnValue);
1242 break;
1243 case TEK_Complex:
1244 EmitComplexExprIntoLValue(RV, MakeAddrLValue(ReturnValue, RV->getType()),
1245 /*isInit*/ true);
1246 break;
1247 case TEK_Aggregate:
1248 EmitAggExpr(RV, AggValueSlot::forAddr(
1249 ReturnValue, Qualifiers(),
1250 AggValueSlot::IsDestructed,
1251 AggValueSlot::DoesNotNeedGCBarriers,
1252 AggValueSlot::IsNotAliased,
1253 getOverlapForReturnValue()));
1254 break;
1255 }
1256 }
1257
1258 ++NumReturnExprs;
1259 if (!RV || RV->isEvaluatable(getContext()))
1260 ++NumSimpleReturnExprs;
1261
1262 cleanupScope.ForceCleanup();
1263 EmitBranchThroughCleanup(ReturnBlock);
1264}
1265
1266void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1267 // As long as debug info is modeled with instructions, we have to ensure we
1268 // have a place to insert here and write the stop point here.
1269 if (HaveInsertPoint())
1270 EmitStopPoint(&S);
1271
1272 for (const auto *I : S.decls())
1273 EmitDecl(*I);
1274}
1275
1276void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1277 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1278
1279 // If this code is reachable then emit a stop point (if generating
1280 // debug info). We have to do this ourselves because we are on the
1281 // "simple" statement path.
1282 if (HaveInsertPoint())
1283 EmitStopPoint(&S);
1284
1285 EmitBranchThroughCleanup(BreakContinueStack.back().BreakBlock);
1286}
1287
1288void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1289 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1290
1291 // If this code is reachable then emit a stop point (if generating
1292 // debug info). We have to do this ourselves because we are on the
1293 // "simple" statement path.
1294 if (HaveInsertPoint())
1295 EmitStopPoint(&S);
1296
1297 EmitBranchThroughCleanup(BreakContinueStack.back().ContinueBlock);
1298}
1299
1300/// EmitCaseStmtRange - If case statement range is not too big then
1301/// add multiple cases to switch instruction, one for each value within
1302/// the range. If range is too big then emit "if" condition check.
1303void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S,
1304 ArrayRef<const Attr *> Attrs) {
1305 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1306
1307 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(getContext());
1308 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(getContext());
1309
1310 // Emit the code for this case. We do this first to make sure it is
1311 // properly chained from our predecessor before generating the
1312 // switch machinery to enter this block.
1313 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1314 EmitBlockWithFallThrough(CaseDest, &S);
1315 EmitStmt(S.getSubStmt());
1316
1317 // If range is empty, do nothing.
1318 if (LHS.isSigned() ? RHS.slt(LHS) : RHS.ult(LHS))
1319 return;
1320
1321 Stmt::Likelihood LH = Stmt::getLikelihood(Attrs);
1322 llvm::APInt Range = RHS - LHS;
1323 // FIXME: parameters such as this should not be hardcoded.
1324 if (Range.ult(llvm::APInt(Range.getBitWidth(), 64))) {
1325 // Range is small enough to add multiple switch instruction cases.
1326 uint64_t Total = getProfileCount(&S);
1327 unsigned NCases = Range.getZExtValue() + 1;
1328 // We only have one region counter for the entire set of cases here, so we
1329 // need to divide the weights evenly between the generated cases, ensuring
1330 // that the total weight is preserved. E.g., a weight of 5 over three cases
1331 // will be distributed as weights of 2, 2, and 1.
1332 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1333 for (unsigned I = 0; I != NCases; ++I) {
1334 if (SwitchWeights)
1335 SwitchWeights->push_back(Weight + (Rem ? 1 : 0));
1336 else if (SwitchLikelihood)
1337 SwitchLikelihood->push_back(LH);
1338
1339 if (Rem)
1340 Rem--;
1341 SwitchInsn->addCase(Builder.getInt(LHS), CaseDest);
1342 ++LHS;
1343 }
1344 return;
1345 }
1346
1347 // The range is too big. Emit "if" condition into a new block,
1348 // making sure to save and restore the current insertion point.
1349 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1350
1351 // Push this test onto the chain of range checks (which terminates
1352 // in the default basic block). The switch's default will be changed
1353 // to the top of this chain after switch emission is complete.
1354 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1355 CaseRangeBlock = createBasicBlock("sw.caserange");
1356
1357 CurFn->getBasicBlockList().push_back(CaseRangeBlock);
1358 Builder.SetInsertPoint(CaseRangeBlock);
1359
1360 // Emit range check.
1361 llvm::Value *Diff =
1362 Builder.CreateSub(SwitchInsn->getCondition(), Builder.getInt(LHS));
1363 llvm::Value *Cond =
1364 Builder.CreateICmpULE(Diff, Builder.getInt(Range), "inbounds");
1365
1366 llvm::MDNode *Weights = nullptr;
1367 if (SwitchWeights) {
1368 uint64_t ThisCount = getProfileCount(&S);
1369 uint64_t DefaultCount = (*SwitchWeights)[0];
1370 Weights = createProfileWeights(ThisCount, DefaultCount);
1371
1372 // Since we're chaining the switch default through each large case range, we
1373 // need to update the weight for the default, ie, the first case, to include
1374 // this case.
1375 (*SwitchWeights)[0] += ThisCount;
1376 } else if (SwitchLikelihood)
1377 Cond = emitCondLikelihoodViaExpectIntrinsic(Cond, LH);
1378
1379 Builder.CreateCondBr(Cond, CaseDest, FalseDest, Weights);
1380
1381 // Restore the appropriate insertion point.
1382 if (RestoreBB)
1383 Builder.SetInsertPoint(RestoreBB);
1384 else
1385 Builder.ClearInsertionPoint();
1386}
1387
1388void CodeGenFunction::EmitCaseStmt(const CaseStmt &S,
1389 ArrayRef<const Attr *> Attrs) {
1390 // If there is no enclosing switch instance that we're aware of, then this
1391 // case statement and its block can be elided. This situation only happens
1392 // when we've constant-folded the switch, are emitting the constant case,
1393 // and part of the constant case includes another case statement. For
1394 // instance: switch (4) { case 4: do { case 5: } while (1); }
1395 if (!SwitchInsn) {
1396 EmitStmt(S.getSubStmt());
1397 return;
1398 }
1399
1400 // Handle case ranges.
1401 if (S.getRHS()) {
1402 EmitCaseStmtRange(S, Attrs);
1403 return;
1404 }
1405
1406 llvm::ConstantInt *CaseVal =
1407 Builder.getInt(S.getLHS()->EvaluateKnownConstInt(getContext()));
1408 if (SwitchLikelihood)
1409 SwitchLikelihood->push_back(Stmt::getLikelihood(Attrs));
1410
1411 // If the body of the case is just a 'break', try to not emit an empty block.
1412 // If we're profiling or we're not optimizing, leave the block in for better
1413 // debug and coverage analysis.
1414 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1415 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1416 isa<BreakStmt>(S.getSubStmt())) {
1417 JumpDest Block = BreakContinueStack.back().BreakBlock;
1418
1419 // Only do this optimization if there are no cleanups that need emitting.
1420 if (isObviouslyBranchWithoutCleanups(Block)) {
1421 if (SwitchWeights)
1422 SwitchWeights->push_back(getProfileCount(&S));
1423 SwitchInsn->addCase(CaseVal, Block.getBlock());
1424
1425 // If there was a fallthrough into this case, make sure to redirect it to
1426 // the end of the switch as well.
1427 if (Builder.GetInsertBlock()) {
1428 Builder.CreateBr(Block.getBlock());
1429 Builder.ClearInsertionPoint();
1430 }
1431 return;
1432 }
1433 }
1434
1435 llvm::BasicBlock *CaseDest = createBasicBlock("sw.bb");
1436 EmitBlockWithFallThrough(CaseDest, &S);
1437 if (SwitchWeights)
1438 SwitchWeights->push_back(getProfileCount(&S));
1439 SwitchInsn->addCase(CaseVal, CaseDest);
1440
1441 // Recursively emitting the statement is acceptable, but is not wonderful for
1442 // code where we have many case statements nested together, i.e.:
1443 // case 1:
1444 // case 2:
1445 // case 3: etc.
1446 // Handling this recursively will create a new block for each case statement
1447 // that falls through to the next case which is IR intensive. It also causes
1448 // deep recursion which can run into stack depth limitations. Handle
1449 // sequential non-range case statements specially.
1450 //
1451 // TODO When the next case has a likelihood attribute the code returns to the
1452 // recursive algorithm. Maybe improve this case if it becomes common practice
1453 // to use a lot of attributes.
1454 const CaseStmt *CurCase = &S;
1455 const CaseStmt *NextCase = dyn_cast<CaseStmt>(S.getSubStmt());
1456
1457 // Otherwise, iteratively add consecutive cases to this switch stmt.
1458 while (NextCase && NextCase->getRHS() == nullptr) {
1459 CurCase = NextCase;
1460 llvm::ConstantInt *CaseVal =
1461 Builder.getInt(CurCase->getLHS()->EvaluateKnownConstInt(getContext()));
1462
1463 if (SwitchWeights)
1464 SwitchWeights->push_back(getProfileCount(NextCase));
1465 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1466 CaseDest = createBasicBlock("sw.bb");
1467 EmitBlockWithFallThrough(CaseDest, CurCase);
1468 }
1469 // Since this loop is only executed when the CaseStmt has no attributes
1470 // use a hard-coded value.
1471 if (SwitchLikelihood)
1472 SwitchLikelihood->push_back(Stmt::LH_None);
1473
1474 SwitchInsn->addCase(CaseVal, CaseDest);
1475 NextCase = dyn_cast<CaseStmt>(CurCase->getSubStmt());
1476 }
1477
1478 // Normal default recursion for non-cases.
1479 EmitStmt(CurCase->getSubStmt());
1480}
1481
1482void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S,
1483 ArrayRef<const Attr *> Attrs) {
1484 // If there is no enclosing switch instance that we're aware of, then this
1485 // default statement can be elided. This situation only happens when we've
1486 // constant-folded the switch.
1487 if (!SwitchInsn) {
1488 EmitStmt(S.getSubStmt());
1489 return;
1490 }
1491
1492 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1493 assert(DefaultBlock->empty() &&
1494 "EmitDefaultStmt: Default block already defined?");
1495
1496 if (SwitchLikelihood)
1497 SwitchLikelihood->front() = Stmt::getLikelihood(Attrs);
1498
1499 EmitBlockWithFallThrough(DefaultBlock, &S);
1500
1501 EmitStmt(S.getSubStmt());
1502}
1503
1504/// CollectStatementsForCase - Given the body of a 'switch' statement and a
1505/// constant value that is being switched on, see if we can dead code eliminate
1506/// the body of the switch to a simple series of statements to emit. Basically,
1507/// on a switch (5) we want to find these statements:
1508/// case 5:
1509/// printf(...); <--
1510/// ++i; <--
1511/// break;
1512///
1513/// and add them to the ResultStmts vector. If it is unsafe to do this
1514/// transformation (for example, one of the elided statements contains a label
1515/// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
1516/// should include statements after it (e.g. the printf() line is a substmt of
1517/// the case) then return CSFC_FallThrough. If we handled it and found a break
1518/// statement, then return CSFC_Success.
1519///
1520/// If Case is non-null, then we are looking for the specified case, checking
1521/// that nothing we jump over contains labels. If Case is null, then we found
1522/// the case and are looking for the break.
1523///
1524/// If the recursive walk actually finds our Case, then we set FoundCase to
1525/// true.
1526///
1527enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
1528static CSFC_Result CollectStatementsForCase(const Stmt *S,
1529 const SwitchCase *Case,
1530 bool &FoundCase,
1531 SmallVectorImpl<const Stmt*> &ResultStmts) {
1532 // If this is a null statement, just succeed.
1533 if (!S)
1534 return Case ? CSFC_Success : CSFC_FallThrough;
1535
1536 // If this is the switchcase (case 4: or default) that we're looking for, then
1537 // we're in business. Just add the substatement.
1538 if (const SwitchCase *SC = dyn_cast<SwitchCase>(S)) {
1539 if (S == Case) {
1540 FoundCase = true;
1541 return CollectStatementsForCase(SC->getSubStmt(), nullptr, FoundCase,
1542 ResultStmts);
1543 }
1544
1545 // Otherwise, this is some other case or default statement, just ignore it.
1546 return CollectStatementsForCase(SC->getSubStmt(), Case, FoundCase,
1547 ResultStmts);
1548 }
1549
1550 // If we are in the live part of the code and we found our break statement,
1551 // return a success!
1552 if (!Case && isa<BreakStmt>(S))
1553 return CSFC_Success;
1554
1555 // If this is a switch statement, then it might contain the SwitchCase, the
1556 // break, or neither.
1557 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(S)) {
1558 // Handle this as two cases: we might be looking for the SwitchCase (if so
1559 // the skipped statements must be skippable) or we might already have it.
1560 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
1561 bool StartedInLiveCode = FoundCase;
1562 unsigned StartSize = ResultStmts.size();
1563
1564 // If we've not found the case yet, scan through looking for it.
1565 if (Case) {
1566 // Keep track of whether we see a skipped declaration. The code could be
1567 // using the declaration even if it is skipped, so we can't optimize out
1568 // the decl if the kept statements might refer to it.
1569 bool HadSkippedDecl = false;
1570
1571 // If we're looking for the case, just see if we can skip each of the
1572 // substatements.
1573 for (; Case && I != E; ++I) {
1574 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(*I);
1575
1576 switch (CollectStatementsForCase(*I, Case, FoundCase, ResultStmts)) {
1577 case CSFC_Failure: return CSFC_Failure;
1578 case CSFC_Success:
1579 // A successful result means that either 1) that the statement doesn't
1580 // have the case and is skippable, or 2) does contain the case value
1581 // and also contains the break to exit the switch. In the later case,
1582 // we just verify the rest of the statements are elidable.
1583 if (FoundCase) {
1584 // If we found the case and skipped declarations, we can't do the
1585 // optimization.
1586 if (HadSkippedDecl)
1587 return CSFC_Failure;
1588
1589 for (++I; I != E; ++I)
1590 if (CodeGenFunction::ContainsLabel(*I, true))
1591 return CSFC_Failure;
1592 return CSFC_Success;
1593 }
1594 break;
1595 case CSFC_FallThrough:
1596 // If we have a fallthrough condition, then we must have found the
1597 // case started to include statements. Consider the rest of the
1598 // statements in the compound statement as candidates for inclusion.
1599 assert(FoundCase && "Didn't find case but returned fallthrough?");
1600 // We recursively found Case, so we're not looking for it anymore.
1601 Case = nullptr;
1602
1603 // If we found the case and skipped declarations, we can't do the
1604 // optimization.
1605 if (HadSkippedDecl)
1606 return CSFC_Failure;
1607 break;
1608 }
1609 }
1610
1611 if (!FoundCase)
1612 return CSFC_Success;
1613
1614 assert(!HadSkippedDecl && "fallthrough after skipping decl");
1615 }
1616
1617 // If we have statements in our range, then we know that the statements are
1618 // live and need to be added to the set of statements we're tracking.
1619 bool AnyDecls = false;
1620 for (; I != E; ++I) {
1621 AnyDecls |= CodeGenFunction::mightAddDeclToScope(*I);
1622
1623 switch (CollectStatementsForCase(*I, nullptr, FoundCase, ResultStmts)) {
1624 case CSFC_Failure: return CSFC_Failure;
1625 case CSFC_FallThrough:
1626 // A fallthrough result means that the statement was simple and just
1627 // included in ResultStmt, keep adding them afterwards.
1628 break;
1629 case CSFC_Success:
1630 // A successful result means that we found the break statement and
1631 // stopped statement inclusion. We just ensure that any leftover stmts
1632 // are skippable and return success ourselves.
1633 for (++I; I != E; ++I)
1634 if (CodeGenFunction::ContainsLabel(*I, true))
1635 return CSFC_Failure;
1636 return CSFC_Success;
1637 }
1638 }
1639
1640 // If we're about to fall out of a scope without hitting a 'break;', we
1641 // can't perform the optimization if there were any decls in that scope
1642 // (we'd lose their end-of-lifetime).
1643 if (AnyDecls) {
1644 // If the entire compound statement was live, there's one more thing we
1645 // can try before giving up: emit the whole thing as a single statement.
1646 // We can do that unless the statement contains a 'break;'.
1647 // FIXME: Such a break must be at the end of a construct within this one.
1648 // We could emit this by just ignoring the BreakStmts entirely.
1649 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) {
1650 ResultStmts.resize(StartSize);
1651 ResultStmts.push_back(S);
1652 } else {
1653 return CSFC_Failure;
1654 }
1655 }
1656
1657 return CSFC_FallThrough;
1658 }
1659
1660 // Okay, this is some other statement that we don't handle explicitly, like a
1661 // for statement or increment etc. If we are skipping over this statement,
1662 // just verify it doesn't have labels, which would make it invalid to elide.
1663 if (Case) {
1664 if (CodeGenFunction::ContainsLabel(S, true))
1665 return CSFC_Failure;
1666 return CSFC_Success;
1667 }
1668
1669 // Otherwise, we want to include this statement. Everything is cool with that
1670 // so long as it doesn't contain a break out of the switch we're in.
1671 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
1672
1673 // Otherwise, everything is great. Include the statement and tell the caller
1674 // that we fall through and include the next statement as well.
1675 ResultStmts.push_back(S);
1676 return CSFC_FallThrough;
1677}
1678
1679/// FindCaseStatementsForValue - Find the case statement being jumped to and
1680/// then invoke CollectStatementsForCase to find the list of statements to emit
1681/// for a switch on constant. See the comment above CollectStatementsForCase
1682/// for more details.
1683static bool FindCaseStatementsForValue(const SwitchStmt &S,
1684 const llvm::APSInt &ConstantCondValue,
1685 SmallVectorImpl<const Stmt*> &ResultStmts,
1686 ASTContext &C,
1687 const SwitchCase *&ResultCase) {
1688 // First step, find the switch case that is being branched to. We can do this
1689 // efficiently by scanning the SwitchCase list.
1690 const SwitchCase *Case = S.getSwitchCaseList();
1691 const DefaultStmt *DefaultCase = nullptr;
1692
1693 for (; Case; Case = Case->getNextSwitchCase()) {
1694 // It's either a default or case. Just remember the default statement in
1695 // case we're not jumping to any numbered cases.
1696 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Case)) {
1697 DefaultCase = DS;
1698 continue;
1699 }
1700
1701 // Check to see if this case is the one we're looking for.
1702 const CaseStmt *CS = cast<CaseStmt>(Case);
1703 // Don't handle case ranges yet.
1704 if (CS->getRHS()) return false;
1705
1706 // If we found our case, remember it as 'case'.
1707 if (CS->getLHS()->EvaluateKnownConstInt(C) == ConstantCondValue)
1708 break;
1709 }
1710
1711 // If we didn't find a matching case, we use a default if it exists, or we
1712 // elide the whole switch body!
1713 if (!Case) {
1714 // It is safe to elide the body of the switch if it doesn't contain labels
1715 // etc. If it is safe, return successfully with an empty ResultStmts list.
1716 if (!DefaultCase)
1717 return !CodeGenFunction::ContainsLabel(&S);
1718 Case = DefaultCase;
1719 }
1720
1721 // Ok, we know which case is being jumped to, try to collect all the
1722 // statements that follow it. This can fail for a variety of reasons. Also,
1723 // check to see that the recursive walk actually found our case statement.
1724 // Insane cases like this can fail to find it in the recursive walk since we
1725 // don't handle every stmt kind:
1726 // switch (4) {
1727 // while (1) {
1728 // case 4: ...
1729 bool FoundCase = false;
1730 ResultCase = Case;
1731 return CollectStatementsForCase(S.getBody(), Case, FoundCase,
1732 ResultStmts) != CSFC_Failure &&
1733 FoundCase;
1734}
1735
1736static Optional<SmallVector<uint64_t, 16>>
1737getLikelihoodWeights(ArrayRef<Stmt::Likelihood> Likelihoods) {
1738 // Are there enough branches to weight them?
1739 if (Likelihoods.size() <= 1)
1740 return None;
1741
1742 uint64_t NumUnlikely = 0;
1743 uint64_t NumNone = 0;
1744 uint64_t NumLikely = 0;
1745 for (const auto LH : Likelihoods) {
1746 switch (LH) {
1747 case Stmt::LH_Unlikely:
1748 ++NumUnlikely;
1749 break;
1750 case Stmt::LH_None:
1751 ++NumNone;
1752 break;
1753 case Stmt::LH_Likely:
1754 ++NumLikely;
1755 break;
1756 }
1757 }
1758
1759 // Is there a likelihood attribute used?
1760 if (NumUnlikely == 0 && NumLikely == 0)
1761 return None;
1762
1763 // When multiple cases share the same code they can be combined during
1764 // optimization. In that case the weights of the branch will be the sum of
1765 // the individual weights. Make sure the combined sum of all neutral cases
1766 // doesn't exceed the value of a single likely attribute.
1767 // The additions both avoid divisions by 0 and make sure the weights of None
1768 // don't exceed the weight of Likely.
1769 const uint64_t Likely = INT32_MAX / (NumLikely + 2);
1770 const uint64_t None = Likely / (NumNone + 1);
1771 const uint64_t Unlikely = 0;
1772
1773 SmallVector<uint64_t, 16> Result;
1774 Result.reserve(Likelihoods.size());
1775 for (const auto LH : Likelihoods) {
1776 switch (LH) {
1777 case Stmt::LH_Unlikely:
1778 Result.push_back(Unlikely);
1779 break;
1780 case Stmt::LH_None:
1781 Result.push_back(None);
1782 break;
1783 case Stmt::LH_Likely:
1784 Result.push_back(Likely);
1785 break;
1786 }
1787 }
1788
1789 return Result;
1790}
1791
1792void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
1793 // Handle nested switch statements.
1794 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
1795 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
1796 SmallVector<Stmt::Likelihood, 16> *SavedSwitchLikelihood = SwitchLikelihood;
1797 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
1798
1799 // See if we can constant fold the condition of the switch and therefore only
1800 // emit the live case statement (if any) of the switch.
1801 llvm::APSInt ConstantCondValue;
1802 if (ConstantFoldsToSimpleInteger(S.getCond(), ConstantCondValue)) {
1803 SmallVector<const Stmt*, 4> CaseStmts;
1804 const SwitchCase *Case = nullptr;
1805 if (FindCaseStatementsForValue(S, ConstantCondValue, CaseStmts,
1806 getContext(), Case)) {
1807 if (Case)
1808 incrementProfileCounter(Case);
1809 RunCleanupsScope ExecutedScope(*this);
1810
1811 if (S.getInit())
1812 EmitStmt(S.getInit());
1813
1814 // Emit the condition variable if needed inside the entire cleanup scope
1815 // used by this special case for constant folded switches.
1816 if (S.getConditionVariable())
1817 EmitDecl(*S.getConditionVariable());
1818
1819 // At this point, we are no longer "within" a switch instance, so
1820 // we can temporarily enforce this to ensure that any embedded case
1821 // statements are not emitted.
1822 SwitchInsn = nullptr;
1823
1824 // Okay, we can dead code eliminate everything except this case. Emit the
1825 // specified series of statements and we're good.
1826 for (unsigned i = 0, e = CaseStmts.size(); i != e; ++i)
1827 EmitStmt(CaseStmts[i]);
1828 incrementProfileCounter(&S);
1829
1830 // Now we want to restore the saved switch instance so that nested
1831 // switches continue to function properly
1832 SwitchInsn = SavedSwitchInsn;
1833
1834 return;
1835 }
1836 }
1837
1838 JumpDest SwitchExit = getJumpDestInCurrentScope("sw.epilog");
1839
1840 RunCleanupsScope ConditionScope(*this);
1841
1842 if (S.getInit())
1843 EmitStmt(S.getInit());
1844
1845 if (S.getConditionVariable())
1846 EmitDecl(*S.getConditionVariable());
1847 llvm::Value *CondV = EmitScalarExpr(S.getCond());
1848
1849 // Create basic block to hold stuff that comes after switch
1850 // statement. We also need to create a default block now so that
1851 // explicit case ranges tests can have a place to jump to on
1852 // failure.
1853 llvm::BasicBlock *DefaultBlock = createBasicBlock("sw.default");
1854 SwitchInsn = Builder.CreateSwitch(CondV, DefaultBlock);
1855 if (PGO.haveRegionCounts()) {
1856 // Walk the SwitchCase list to find how many there are.
1857 uint64_t DefaultCount = 0;
1858 unsigned NumCases = 0;
1859 for (const SwitchCase *Case = S.getSwitchCaseList();
1860 Case;
1861 Case = Case->getNextSwitchCase()) {
1862 if (isa<DefaultStmt>(Case))
1863 DefaultCount = getProfileCount(Case);
1864 NumCases += 1;
1865 }
1866 SwitchWeights = new SmallVector<uint64_t, 16>();
1867 SwitchWeights->reserve(NumCases);
1868 // The default needs to be first. We store the edge count, so we already
1869 // know the right weight.
1870 SwitchWeights->push_back(DefaultCount);
1871 } else if (CGM.getCodeGenOpts().OptimizationLevel) {
1872 SwitchLikelihood = new SmallVector<Stmt::Likelihood, 16>();
1873 // Initialize the default case.
1874 SwitchLikelihood->push_back(Stmt::LH_None);
1875 }
1876
1877 CaseRangeBlock = DefaultBlock;
1878
1879 // Clear the insertion point to indicate we are in unreachable code.
1880 Builder.ClearInsertionPoint();
1881
1882 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
1883 // then reuse last ContinueBlock.
1884 JumpDest OuterContinue;
1885 if (!BreakContinueStack.empty())
1886 OuterContinue = BreakContinueStack.back().ContinueBlock;
1887
1888 BreakContinueStack.push_back(BreakContinue(SwitchExit, OuterContinue));
1889
1890 // Emit switch body.
1891 EmitStmt(S.getBody());
1892
1893 BreakContinueStack.pop_back();
1894
1895 // Update the default block in case explicit case range tests have
1896 // been chained on top.
1897 SwitchInsn->setDefaultDest(CaseRangeBlock);
1898
1899 // If a default was never emitted:
1900 if (!DefaultBlock->getParent()) {
1901 // If we have cleanups, emit the default block so that there's a
1902 // place to jump through the cleanups from.
1903 if (ConditionScope.requiresCleanups()) {
1904 EmitBlock(DefaultBlock);
1905
1906 // Otherwise, just forward the default block to the switch end.
1907 } else {
1908 DefaultBlock->replaceAllUsesWith(SwitchExit.getBlock());
1909 delete DefaultBlock;
1910 }
1911 }
1912
1913 ConditionScope.ForceCleanup();
1914
1915 // Emit continuation.
1916 EmitBlock(SwitchExit.getBlock(), true);
1917 incrementProfileCounter(&S);
1918
1919 // If the switch has a condition wrapped by __builtin_unpredictable,
1920 // create metadata that specifies that the switch is unpredictable.
1921 // Don't bother if not optimizing because that metadata would not be used.
1922 auto *Call = dyn_cast<CallExpr>(S.getCond());
1923 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
1924 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
1925 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
1926 llvm::MDBuilder MDHelper(getLLVMContext());
1927 SwitchInsn->setMetadata(llvm::LLVMContext::MD_unpredictable,
1928 MDHelper.createUnpredictable());
1929 }
1930 }
1931
1932 if (SwitchWeights) {
1933 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
1934 "switch weights do not match switch cases");
1935 // If there's only one jump destination there's no sense weighting it.
1936 if (SwitchWeights->size() > 1)
1937 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1938 createProfileWeights(*SwitchWeights));
1939 delete SwitchWeights;
1940 } else if (SwitchLikelihood) {
1941 assert(SwitchLikelihood->size() == 1 + SwitchInsn->getNumCases() &&
1942 "switch likelihoods do not match switch cases");
1943 Optional<SmallVector<uint64_t, 16>> LHW =
1944 getLikelihoodWeights(*SwitchLikelihood);
1945 if (LHW) {
1946 llvm::MDBuilder MDHelper(CGM.getLLVMContext());
1947 SwitchInsn->setMetadata(llvm::LLVMContext::MD_prof,
1948 createProfileWeights(*LHW));
1949 }
1950 delete SwitchLikelihood;
1951 }
1952 SwitchInsn = SavedSwitchInsn;
1953 SwitchWeights = SavedSwitchWeights;
1954 SwitchLikelihood = SavedSwitchLikelihood;
1955 CaseRangeBlock = SavedCRBlock;
1956}
1957
1958static std::string
1959SimplifyConstraint(const char *Constraint, const TargetInfo &Target,
1960 SmallVectorImpl<TargetInfo::ConstraintInfo> *OutCons=nullptr) {
1961 std::string Result;
1962
1963 while (*Constraint) {
1964 switch (*Constraint) {
1965 default:
1966 Result += Target.convertConstraint(Constraint);
1967 break;
1968 // Ignore these
1969 case '*':
1970 case '?':
1971 case '!':
1972 case '=': // Will see this and the following in mult-alt constraints.
1973 case '+':
1974 break;
1975 case '#': // Ignore the rest of the constraint alternative.
1976 while (Constraint[1] && Constraint[1] != ',')
1977 Constraint++;
1978 break;
1979 case '&':
1980 case '%':
1981 Result += *Constraint;
1982 while (Constraint[1] && Constraint[1] == *Constraint)
1983 Constraint++;
1984 break;
1985 case ',':
1986 Result += "|";
1987 break;
1988 case 'g':
1989 Result += "imr";
1990 break;
1991 case '[': {
1992 assert(OutCons &&
1993 "Must pass output names to constraints with a symbolic name");
1994 unsigned Index;
1995 bool result = Target.resolveSymbolicName(Constraint, *OutCons, Index);
1996 assert(result && "Could not resolve symbolic name"); (void)result;
1997 Result += llvm::utostr(Index);
1998 break;
1999 }
2000 }
2001
2002 Constraint++;
2003 }
2004
2005 return Result;
2006}
2007
2008/// AddVariableConstraints - Look at AsmExpr and if it is a variable declared
2009/// as using a particular register add that as a constraint that will be used
2010/// in this asm stmt.
2011static std::string
2012AddVariableConstraints(const std::string &Constraint, const Expr &AsmExpr,
2013 const TargetInfo &Target, CodeGenModule &CGM,
2014 const AsmStmt &Stmt, const bool EarlyClobber,
2015 std::string *GCCReg = nullptr) {
2016 const DeclRefExpr *AsmDeclRef = dyn_cast<DeclRefExpr>(&AsmExpr);
2017 if (!AsmDeclRef)
2018 return Constraint;
2019 const ValueDecl &Value = *AsmDeclRef->getDecl();
2020 const VarDecl *Variable = dyn_cast<VarDecl>(&Value);
2021 if (!Variable)
2022 return Constraint;
2023 if (Variable->getStorageClass() != SC_Register)
2024 return Constraint;
2025 AsmLabelAttr *Attr = Variable->getAttr<AsmLabelAttr>();
2026 if (!Attr)
2027 return Constraint;
2028 StringRef Register = Attr->getLabel();
2029 assert(Target.isValidGCCRegisterName(Register));
2030 // We're using validateOutputConstraint here because we only care if
2031 // this is a register constraint.
2032 TargetInfo::ConstraintInfo Info(Constraint, "");
2033 if (Target.validateOutputConstraint(Info) &&
2034 !Info.allowsRegister()) {
2035 CGM.ErrorUnsupported(&Stmt, "__asm__");
2036 return Constraint;
2037 }
2038 // Canonicalize the register here before returning it.
2039 Register = Target.getNormalizedGCCRegisterName(Register);
2040 if (GCCReg != nullptr)
2041 *GCCReg = Register.str();
2042 return (EarlyClobber ? "&{" : "{") + Register.str() + "}";
2043}
2044
2045llvm::Value*
2046CodeGenFunction::EmitAsmInputLValue(const TargetInfo::ConstraintInfo &Info,
2047 LValue InputValue, QualType InputType,
2048 std::string &ConstraintStr,
2049 SourceLocation Loc) {
2050 llvm::Value *Arg;
2051 if (Info.allowsRegister() || !Info.allowsMemory()) {
2052 if (CodeGenFunction::hasScalarEvaluationKind(InputType)) {
2053 Arg = EmitLoadOfLValue(InputValue, Loc).getScalarVal();
2054 } else {
2055 llvm::Type *Ty = ConvertType(InputType);
2056 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
2057 if (Size <= 64 && llvm::isPowerOf2_64(Size)) {
2058 Ty = llvm::IntegerType::get(getLLVMContext(), Size);
2059 Ty = llvm::PointerType::getUnqual(Ty);
2060
2061 Arg = Builder.CreateLoad(
2062 Builder.CreateBitCast(InputValue.getAddress(*this), Ty));
2063 } else {
2064 Arg = InputValue.getPointer(*this);
2065 ConstraintStr += '*';
2066 }
2067 }
2068 } else {
2069 Arg = InputValue.getPointer(*this);
2070 ConstraintStr += '*';
2071 }
2072
2073 return Arg;
2074}
2075
2076llvm::Value* CodeGenFunction::EmitAsmInput(
2077 const TargetInfo::ConstraintInfo &Info,
2078 const Expr *InputExpr,
2079 std::string &ConstraintStr) {
2080 // If this can't be a register or memory, i.e., has to be a constant
2081 // (immediate or symbolic), try to emit it as such.
2082 if (!Info.allowsRegister() && !Info.allowsMemory()) {
2083 if (Info.requiresImmediateConstant()) {
2084 Expr::EvalResult EVResult;
2085 InputExpr->EvaluateAsRValue(EVResult, getContext(), true);
2086
2087 llvm::APSInt IntResult;
2088 if (EVResult.Val.toIntegralConstant(IntResult, InputExpr->getType(),
2089 getContext()))
2090 return llvm::ConstantInt::get(getLLVMContext(), IntResult);
2091 }
2092
2093 Expr::EvalResult Result;
2094 if (InputExpr->EvaluateAsInt(Result, getContext()))
2095 return llvm::ConstantInt::get(getLLVMContext(), Result.Val.getInt());
2096 }
2097
2098 if (Info.allowsRegister() || !Info.allowsMemory())
2099 if (CodeGenFunction::hasScalarEvaluationKind(InputExpr->getType()))
2100 return EmitScalarExpr(InputExpr);
2101 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
2102 return EmitScalarExpr(InputExpr);
2103 InputExpr = InputExpr->IgnoreParenNoopCasts(getContext());
2104 LValue Dest = EmitLValue(InputExpr);
2105 return EmitAsmInputLValue(Info, Dest, InputExpr->getType(), ConstraintStr,
2106 InputExpr->getExprLoc());
2107}
2108
2109/// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
2110/// asm call instruction. The !srcloc MDNode contains a list of constant
2111/// integers which are the source locations of the start of each line in the
2112/// asm.
2113static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
2114 CodeGenFunction &CGF) {
2115 SmallVector<llvm::Metadata *, 8> Locs;
2116 // Add the location of the first line to the MDNode.
2117 Locs.push_back(llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
2118 CGF.Int32Ty, Str->getBeginLoc().getRawEncoding())));
2119 StringRef StrVal = Str->getString();
2120 if (!StrVal.empty()) {
2121 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
2122 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
2123 unsigned StartToken = 0;
2124 unsigned ByteOffset = 0;
2125
2126 // Add the location of the start of each subsequent line of the asm to the
2127 // MDNode.
2128 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
2129 if (StrVal[i] != '\n') continue;
2130 SourceLocation LineLoc = Str->getLocationOfByte(
2131 i + 1, SM, LangOpts, CGF.getTarget(), &StartToken, &ByteOffset);
2132 Locs.push_back(llvm::ConstantAsMetadata::get(
2133 llvm::ConstantInt::get(CGF.Int32Ty, LineLoc.getRawEncoding())));
2134 }
2135 }
2136
2137 return llvm::MDNode::get(CGF.getLLVMContext(), Locs);
2138}
2139
2140static void UpdateAsmCallInst(llvm::CallBase &Result, bool HasSideEffect,
2141 bool ReadOnly, bool ReadNone, bool NoMerge,
2142 const AsmStmt &S,
2143 const std::vector<llvm::Type *> &ResultRegTypes,
2144 CodeGenFunction &CGF,
2145 std::vector<llvm::Value *> &RegResults) {
2146 Result.addAttribute(llvm::AttributeList::FunctionIndex,
2147 llvm::Attribute::NoUnwind);
2148 if (NoMerge)
2149 Result.addAttribute(llvm::AttributeList::FunctionIndex,
2150 llvm::Attribute::NoMerge);
2151 // Attach readnone and readonly attributes.
2152 if (!HasSideEffect) {
2153 if (ReadNone)
2154 Result.addAttribute(llvm::AttributeList::FunctionIndex,
2155 llvm::Attribute::ReadNone);
2156 else if (ReadOnly)
2157 Result.addAttribute(llvm::AttributeList::FunctionIndex,
2158 llvm::Attribute::ReadOnly);
2159 }
2160
2161 // Slap the source location of the inline asm into a !srcloc metadata on the
2162 // call.
2163 if (const auto *gccAsmStmt = dyn_cast<GCCAsmStmt>(&S))
2164 Result.setMetadata("srcloc",
2165 getAsmSrcLocInfo(gccAsmStmt->getAsmString(), CGF));
2166 else {
2167 // At least put the line number on MS inline asm blobs.
2168 llvm::Constant *Loc = llvm::ConstantInt::get(CGF.Int32Ty,
2169 S.getAsmLoc().getRawEncoding());
2170 Result.setMetadata("srcloc",
2171 llvm::MDNode::get(CGF.getLLVMContext(),
2172 llvm::ConstantAsMetadata::get(Loc)));
2173 }
2174
2175 if (CGF.getLangOpts().assumeFunctionsAreConvergent())
2176 // Conservatively, mark all inline asm blocks in CUDA or OpenCL as
2177 // convergent (meaning, they may call an intrinsically convergent op, such
2178 // as bar.sync, and so can't have certain optimizations applied around
2179 // them).
2180 Result.addAttribute(llvm::AttributeList::FunctionIndex,
2181 llvm::Attribute::Convergent);
2182 // Extract all of the register value results from the asm.
2183 if (ResultRegTypes.size() == 1) {
2184 RegResults.push_back(&Result);
2185 } else {
2186 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2187 llvm::Value *Tmp = CGF.Builder.CreateExtractValue(&Result, i, "asmresult");
2188 RegResults.push_back(Tmp);
2189 }
2190 }
2191}
2192
2193void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
2194 // Assemble the final asm string.
2195 std::string AsmString = S.generateAsmString(getContext());
2196
2197 // Get all the output and input constraints together.
2198 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
2199 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
2200
2201 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
2202 StringRef Name;
2203 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
2204 Name = GAS->getOutputName(i);
2205 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
2206 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
2207 assert(IsValid && "Failed to parse output constraint");
2208 OutputConstraintInfos.push_back(Info);
2209 }
2210
2211 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
2212 StringRef Name;
2213 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(&S))
2214 Name = GAS->getInputName(i);
2215 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
2216 bool IsValid =
2217 getTarget().validateInputConstraint(OutputConstraintInfos, Info);
2218 assert(IsValid && "Failed to parse input constraint"); (void)IsValid;
2219 InputConstraintInfos.push_back(Info);
2220 }
2221
2222 std::string Constraints;
2223
2224 std::vector<LValue> ResultRegDests;
2225 std::vector<QualType> ResultRegQualTys;
2226 std::vector<llvm::Type *> ResultRegTypes;
2227 std::vector<llvm::Type *> ResultTruncRegTypes;
2228 std::vector<llvm::Type *> ArgTypes;
2229 std::vector<llvm::Value*> Args;
2230 llvm::BitVector ResultTypeRequiresCast;
2231
2232 // Keep track of inout constraints.
2233 std::string InOutConstraints;
2234 std::vector<llvm::Value*> InOutArgs;
2235 std::vector<llvm::Type*> InOutArgTypes;
2236
2237 // Keep track of out constraints for tied input operand.
2238 std::vector<std::string> OutputConstraints;
2239
2240 // Keep track of defined physregs.
2241 llvm::SmallSet<std::string, 8> PhysRegOutputs;
2242
2243 // An inline asm can be marked readonly if it meets the following conditions:
2244 // - it doesn't have any sideeffects
2245 // - it doesn't clobber memory
2246 // - it doesn't return a value by-reference
2247 // It can be marked readnone if it doesn't have any input memory constraints
2248 // in addition to meeting the conditions listed above.
2249 bool ReadOnly = true, ReadNone = true;
2250
2251 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
2252 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
2253
2254 // Simplify the output constraint.
2255 std::string OutputConstraint(S.getOutputConstraint(i));
2256 OutputConstraint = SimplifyConstraint(OutputConstraint.c_str() + 1,
2257 getTarget(), &OutputConstraintInfos);
2258
2259 const Expr *OutExpr = S.getOutputExpr(i);
2260 OutExpr = OutExpr->IgnoreParenNoopCasts(getContext());
2261
2262 std::string GCCReg;
2263 OutputConstraint = AddVariableConstraints(OutputConstraint, *OutExpr,
2264 getTarget(), CGM, S,
2265 Info.earlyClobber(),
2266 &GCCReg);
2267 // Give an error on multiple outputs to same physreg.
2268 if (!GCCReg.empty() && !PhysRegOutputs.insert(GCCReg).second)
2269 CGM.Error(S.getAsmLoc(), "multiple outputs to hard register: " + GCCReg);
2270
2271 OutputConstraints.push_back(OutputConstraint);
2272 LValue Dest = EmitLValue(OutExpr);
2273 if (!Constraints.empty())
2274 Constraints += ',';
2275
2276 // If this is a register output, then make the inline asm return it
2277 // by-value. If this is a memory result, return the value by-reference.
2278 bool isScalarizableAggregate =
2279 hasAggregateEvaluationKind(OutExpr->getType());
2280 if (!Info.allowsMemory() && (hasScalarEvaluationKind(OutExpr->getType()) ||
2281 isScalarizableAggregate)) {
2282 Constraints += "=" + OutputConstraint;
2283 ResultRegQualTys.push_back(OutExpr->getType());
2284 ResultRegDests.push_back(Dest);
2285 ResultTruncRegTypes.push_back(ConvertTypeForMem(OutExpr->getType()));
2286 if (Info.allowsRegister() && isScalarizableAggregate) {
2287 ResultTypeRequiresCast.push_back(true);
2288 unsigned Size = getContext().getTypeSize(OutExpr->getType());
2289 llvm::Type *ConvTy = llvm::IntegerType::get(getLLVMContext(), Size);
2290 ResultRegTypes.push_back(ConvTy);
2291 } else {
2292 ResultTypeRequiresCast.push_back(false);
2293 ResultRegTypes.push_back(ResultTruncRegTypes.back());
2294 }
2295 // If this output is tied to an input, and if the input is larger, then
2296 // we need to set the actual result type of the inline asm node to be the
2297 // same as the input type.
2298 if (Info.hasMatchingInput()) {
2299 unsigned InputNo;
2300 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
2301 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
2302 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
2303 break;
2304 }
2305 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
2306
2307 QualType InputTy = S.getInputExpr(InputNo)->getType();
2308 QualType OutputType = OutExpr->getType();
2309
2310 uint64_t InputSize = getContext().getTypeSize(InputTy);
2311 if (getContext().getTypeSize(OutputType) < InputSize) {
2312 // Form the asm to return the value as a larger integer or fp type.
2313 ResultRegTypes.back() = ConvertType(InputTy);
2314 }
2315 }
2316 if (llvm::Type* AdjTy =
2317 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
2318 ResultRegTypes.back()))
2319 ResultRegTypes.back() = AdjTy;
2320 else {
2321 CGM.getDiags().Report(S.getAsmLoc(),
2322 diag::err_asm_invalid_type_in_input)
2323 << OutExpr->getType() << OutputConstraint;
2324 }
2325
2326 // Update largest vector width for any vector types.
2327 if (auto *VT = dyn_cast<llvm::VectorType>(ResultRegTypes.back()))
2328 LargestVectorWidth =
2329 std::max((uint64_t)LargestVectorWidth,
2330 VT->getPrimitiveSizeInBits().getKnownMinSize());
2331 } else {
2332 llvm::Type *DestAddrTy = Dest.getAddress(*this).getType();
2333 llvm::Value *DestPtr = Dest.getPointer(*this);
2334 // Matrix types in memory are represented by arrays, but accessed through
2335 // vector pointers, with the alignment specified on the access operation.
2336 // For inline assembly, update pointer arguments to use vector pointers.
2337 // Otherwise there will be a mis-match if the matrix is also an
2338 // input-argument which is represented as vector.
2339 if (isa<MatrixType>(OutExpr->getType().getCanonicalType())) {
2340 DestAddrTy = llvm::PointerType::get(
2341 ConvertType(OutExpr->getType()),
2342 cast<llvm::PointerType>(DestAddrTy)->getAddressSpace());
2343 DestPtr = Builder.CreateBitCast(DestPtr, DestAddrTy);
2344 }
2345 ArgTypes.push_back(DestAddrTy);
2346 Args.push_back(DestPtr);
2347 Constraints += "=*";
2348 Constraints += OutputConstraint;
2349 ReadOnly = ReadNone = false;
2350 }
2351
2352 if (Info.isReadWrite()) {
2353 InOutConstraints += ',';
2354
2355 const Expr *InputExpr = S.getOutputExpr(i);
2356 llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
2357 InOutConstraints,
2358 InputExpr->getExprLoc());
2359
2360 if (llvm::Type* AdjTy =
2361 getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
2362 Arg->getType()))
2363 Arg = Builder.CreateBitCast(Arg, AdjTy);
2364
2365 // Update largest vector width for any vector types.
2366 if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType()))
2367 LargestVectorWidth =
2368 std::max((uint64_t)LargestVectorWidth,
2369 VT->getPrimitiveSizeInBits().getKnownMinSize());
2370 // Only tie earlyclobber physregs.
2371 if (Info.allowsRegister() && (GCCReg.empty() || Info.earlyClobber()))
2372 InOutConstraints += llvm::utostr(i);
2373 else
2374 InOutConstraints += OutputConstraint;
2375
2376 InOutArgTypes.push_back(Arg->getType());
2377 InOutArgs.push_back(Arg);
2378 }
2379 }
2380
2381 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
2382 // to the return value slot. Only do this when returning in registers.
2383 if (isa<MSAsmStmt>(&S)) {
2384 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
2385 if (RetAI.isDirect() || RetAI.isExtend()) {
2386 // Make a fake lvalue for the return value slot.
2387 LValue ReturnSlot =