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 "CodeGenPGO.h"
18#include "TargetInfo.h"
19#include "clang/AST/Attr.h"
20#include "clang/AST/Expr.h"
21#include "clang/AST/Stmt.h"
22#include "clang/AST/StmtVisitor.h"
23#include "clang/Basic/Builtins.h"
24#include "clang/Basic/DiagnosticSema.h"
25#include "clang/Basic/PrettyStackTrace.h"
26#include "clang/Basic/SourceManager.h"
27#include "clang/Basic/TargetInfo.h"
28#include "llvm/ADT/ArrayRef.h"
29#include "llvm/ADT/DenseMap.h"
30#include "llvm/ADT/SmallSet.h"
31#include "llvm/ADT/StringExtras.h"
32#include "llvm/IR/Assumptions.h"
33#include "llvm/IR/DataLayout.h"
34#include "llvm/IR/InlineAsm.h"
35#include "llvm/IR/Intrinsics.h"
36#include "llvm/IR/MDBuilder.h"
37#include "llvm/Support/SaveAndRestore.h"
38#include <optional>
39
40using namespace clang;
41using namespace CodeGen;
42
43//===----------------------------------------------------------------------===//
44// Statement Emission
45//===----------------------------------------------------------------------===//
46
47void CodeGenFunction::EmitStopPoint(const Stmt *S) {
48 if (CGDebugInfo *DI = getDebugInfo()) {
49 SourceLocation Loc;
50 Loc = S->getBeginLoc();
51 DI->EmitLocation(Builder, Loc);
52
53 LastStopPoint = Loc;
54 }
55}
56
57void CodeGenFunction::EmitStmt(const Stmt *S, ArrayRef<const Attr *> Attrs) {
58 assert(S && "Null statement?");
59 PGO->setCurrentStmt(S);
60
61 // These statements have their own debug info handling.
62 if (EmitSimpleStmt(S, Attrs))
63 return;
64
65 // Check if we are generating unreachable code.
66 if (!HaveInsertPoint()) {
67 // If so, and the statement doesn't contain a label, then we do not need to
68 // generate actual code. This is safe because (1) the current point is
69 // unreachable, so we don't need to execute the code, and (2) we've already
70 // handled the statements which update internal data structures (like the
71 // local variable map) which could be used by subsequent statements.
72 if (!ContainsLabel(S)) {
73 // Verify that any decl statements were handled as simple, they may be in
74 // scope of subsequent reachable statements.
75 assert(!isa<DeclStmt>(*S) && "Unexpected DeclStmt!");
76 PGO->markStmtMaybeUsed(S);
77 return;
78 }
79
80 // Otherwise, make a new block to hold the code.
81 EnsureInsertPoint();
82 }
83
84 // Generate a stoppoint if we are emitting debug info.
85 EmitStopPoint(S);
86
87 // Ignore all OpenMP directives except for simd if OpenMP with Simd is
88 // enabled.
89 if (getLangOpts().OpenMP && getLangOpts().OpenMPSimd) {
90 if (const auto *D = dyn_cast<OMPExecutableDirective>(Val: S)) {
91 EmitSimpleOMPExecutableDirective(D: *D);
92 return;
93 }
94 }
95
96 switch (S->getStmtClass()) {
97 case Stmt::NoStmtClass:
98 case Stmt::CXXCatchStmtClass:
99 case Stmt::SEHExceptStmtClass:
100 case Stmt::SEHFinallyStmtClass:
101 case Stmt::MSDependentExistsStmtClass:
102 llvm_unreachable("invalid statement class to emit generically");
103 case Stmt::NullStmtClass:
104 case Stmt::CompoundStmtClass:
105 case Stmt::DeclStmtClass:
106 case Stmt::LabelStmtClass:
107 case Stmt::AttributedStmtClass:
108 case Stmt::GotoStmtClass:
109 case Stmt::BreakStmtClass:
110 case Stmt::ContinueStmtClass:
111 case Stmt::DefaultStmtClass:
112 case Stmt::CaseStmtClass:
113 case Stmt::DeferStmtClass:
114 case Stmt::SEHLeaveStmtClass:
115 case Stmt::SYCLKernelCallStmtClass:
116 llvm_unreachable("should have emitted these statements as simple");
117
118#define STMT(Type, Base)
119#define ABSTRACT_STMT(Op)
120#define EXPR(Type, Base) \
121 case Stmt::Type##Class:
122#include "clang/AST/StmtNodes.inc"
123 {
124 // Remember the block we came in on.
125 llvm::BasicBlock *incoming = Builder.GetInsertBlock();
126 assert(incoming && "expression emission must have an insertion point");
127
128 EmitIgnoredExpr(E: cast<Expr>(Val: S));
129
130 llvm::BasicBlock *outgoing = Builder.GetInsertBlock();
131 assert(outgoing && "expression emission cleared block!");
132
133 // The expression emitters assume (reasonably!) that the insertion
134 // point is always set. To maintain that, the call-emission code
135 // for noreturn functions has to enter a new block with no
136 // predecessors. We want to kill that block and mark the current
137 // insertion point unreachable in the common case of a call like
138 // "exit();". Since expression emission doesn't otherwise create
139 // blocks with no predecessors, we can just test for that.
140 // However, we must be careful not to do this to our incoming
141 // block, because *statement* emission does sometimes create
142 // reachable blocks which will have no predecessors until later in
143 // the function. This occurs with, e.g., labels that are not
144 // reachable by fallthrough.
145 if (incoming != outgoing && outgoing->use_empty()) {
146 outgoing->eraseFromParent();
147 Builder.ClearInsertionPoint();
148 }
149 break;
150 }
151
152 case Stmt::IndirectGotoStmtClass:
153 EmitIndirectGotoStmt(S: cast<IndirectGotoStmt>(Val: *S)); break;
154
155 case Stmt::IfStmtClass: EmitIfStmt(S: cast<IfStmt>(Val: *S)); break;
156 case Stmt::WhileStmtClass: EmitWhileStmt(S: cast<WhileStmt>(Val: *S), Attrs); break;
157 case Stmt::DoStmtClass: EmitDoStmt(S: cast<DoStmt>(Val: *S), Attrs); break;
158 case Stmt::ForStmtClass: EmitForStmt(S: cast<ForStmt>(Val: *S), Attrs); break;
159
160 case Stmt::ReturnStmtClass: EmitReturnStmt(S: cast<ReturnStmt>(Val: *S)); break;
161
162 case Stmt::SwitchStmtClass: EmitSwitchStmt(S: cast<SwitchStmt>(Val: *S)); break;
163 case Stmt::GCCAsmStmtClass: // Intentional fall-through.
164 case Stmt::MSAsmStmtClass: EmitAsmStmt(S: cast<AsmStmt>(Val: *S)); break;
165 case Stmt::CoroutineBodyStmtClass:
166 EmitCoroutineBody(S: cast<CoroutineBodyStmt>(Val: *S));
167 break;
168 case Stmt::CoreturnStmtClass:
169 EmitCoreturnStmt(S: cast<CoreturnStmt>(Val: *S));
170 break;
171 case Stmt::CapturedStmtClass: {
172 const CapturedStmt *CS = cast<CapturedStmt>(Val: S);
173 EmitCapturedStmt(S: *CS, K: CS->getCapturedRegionKind());
174 }
175 break;
176 case Stmt::ObjCAtTryStmtClass:
177 EmitObjCAtTryStmt(S: cast<ObjCAtTryStmt>(Val: *S));
178 break;
179 case Stmt::ObjCAtCatchStmtClass:
180 llvm_unreachable(
181 "@catch statements should be handled by EmitObjCAtTryStmt");
182 case Stmt::ObjCAtFinallyStmtClass:
183 llvm_unreachable(
184 "@finally statements should be handled by EmitObjCAtTryStmt");
185 case Stmt::ObjCAtThrowStmtClass:
186 EmitObjCAtThrowStmt(S: cast<ObjCAtThrowStmt>(Val: *S));
187 break;
188 case Stmt::ObjCAtSynchronizedStmtClass:
189 EmitObjCAtSynchronizedStmt(S: cast<ObjCAtSynchronizedStmt>(Val: *S));
190 break;
191 case Stmt::ObjCForCollectionStmtClass:
192 EmitObjCForCollectionStmt(S: cast<ObjCForCollectionStmt>(Val: *S));
193 break;
194 case Stmt::ObjCAutoreleasePoolStmtClass:
195 EmitObjCAutoreleasePoolStmt(S: cast<ObjCAutoreleasePoolStmt>(Val: *S));
196 break;
197
198 case Stmt::CXXTryStmtClass:
199 EmitCXXTryStmt(S: cast<CXXTryStmt>(Val: *S));
200 break;
201 case Stmt::CXXForRangeStmtClass:
202 EmitCXXForRangeStmt(S: cast<CXXForRangeStmt>(Val: *S), Attrs);
203 break;
204 case Stmt::SEHTryStmtClass:
205 EmitSEHTryStmt(S: cast<SEHTryStmt>(Val: *S));
206 break;
207 case Stmt::OMPMetaDirectiveClass:
208 EmitOMPMetaDirective(S: cast<OMPMetaDirective>(Val: *S));
209 break;
210 case Stmt::OMPCanonicalLoopClass:
211 EmitOMPCanonicalLoop(S: cast<OMPCanonicalLoop>(Val: S));
212 break;
213 case Stmt::OMPParallelDirectiveClass:
214 EmitOMPParallelDirective(S: cast<OMPParallelDirective>(Val: *S));
215 break;
216 case Stmt::OMPSimdDirectiveClass:
217 EmitOMPSimdDirective(S: cast<OMPSimdDirective>(Val: *S));
218 break;
219 case Stmt::OMPTileDirectiveClass:
220 EmitOMPTileDirective(S: cast<OMPTileDirective>(Val: *S));
221 break;
222 case Stmt::OMPStripeDirectiveClass:
223 EmitOMPStripeDirective(S: cast<OMPStripeDirective>(Val: *S));
224 break;
225 case Stmt::OMPUnrollDirectiveClass:
226 EmitOMPUnrollDirective(S: cast<OMPUnrollDirective>(Val: *S));
227 break;
228 case Stmt::OMPReverseDirectiveClass:
229 EmitOMPReverseDirective(S: cast<OMPReverseDirective>(Val: *S));
230 break;
231 case Stmt::OMPInterchangeDirectiveClass:
232 EmitOMPInterchangeDirective(S: cast<OMPInterchangeDirective>(Val: *S));
233 break;
234 case Stmt::OMPFuseDirectiveClass:
235 EmitOMPFuseDirective(S: cast<OMPFuseDirective>(Val: *S));
236 break;
237 case Stmt::OMPForDirectiveClass:
238 EmitOMPForDirective(S: cast<OMPForDirective>(Val: *S));
239 break;
240 case Stmt::OMPForSimdDirectiveClass:
241 EmitOMPForSimdDirective(S: cast<OMPForSimdDirective>(Val: *S));
242 break;
243 case Stmt::OMPSectionsDirectiveClass:
244 EmitOMPSectionsDirective(S: cast<OMPSectionsDirective>(Val: *S));
245 break;
246 case Stmt::OMPSectionDirectiveClass:
247 EmitOMPSectionDirective(S: cast<OMPSectionDirective>(Val: *S));
248 break;
249 case Stmt::OMPSingleDirectiveClass:
250 EmitOMPSingleDirective(S: cast<OMPSingleDirective>(Val: *S));
251 break;
252 case Stmt::OMPMasterDirectiveClass:
253 EmitOMPMasterDirective(S: cast<OMPMasterDirective>(Val: *S));
254 break;
255 case Stmt::OMPCriticalDirectiveClass:
256 EmitOMPCriticalDirective(S: cast<OMPCriticalDirective>(Val: *S));
257 break;
258 case Stmt::OMPParallelForDirectiveClass:
259 EmitOMPParallelForDirective(S: cast<OMPParallelForDirective>(Val: *S));
260 break;
261 case Stmt::OMPParallelForSimdDirectiveClass:
262 EmitOMPParallelForSimdDirective(S: cast<OMPParallelForSimdDirective>(Val: *S));
263 break;
264 case Stmt::OMPParallelMasterDirectiveClass:
265 EmitOMPParallelMasterDirective(S: cast<OMPParallelMasterDirective>(Val: *S));
266 break;
267 case Stmt::OMPParallelSectionsDirectiveClass:
268 EmitOMPParallelSectionsDirective(S: cast<OMPParallelSectionsDirective>(Val: *S));
269 break;
270 case Stmt::OMPTaskDirectiveClass:
271 EmitOMPTaskDirective(S: cast<OMPTaskDirective>(Val: *S));
272 break;
273 case Stmt::OMPTaskyieldDirectiveClass:
274 EmitOMPTaskyieldDirective(S: cast<OMPTaskyieldDirective>(Val: *S));
275 break;
276 case Stmt::OMPErrorDirectiveClass:
277 EmitOMPErrorDirective(S: cast<OMPErrorDirective>(Val: *S));
278 break;
279 case Stmt::OMPBarrierDirectiveClass:
280 EmitOMPBarrierDirective(S: cast<OMPBarrierDirective>(Val: *S));
281 break;
282 case Stmt::OMPTaskwaitDirectiveClass:
283 EmitOMPTaskwaitDirective(S: cast<OMPTaskwaitDirective>(Val: *S));
284 break;
285 case Stmt::OMPTaskgroupDirectiveClass:
286 EmitOMPTaskgroupDirective(S: cast<OMPTaskgroupDirective>(Val: *S));
287 break;
288 case Stmt::OMPFlushDirectiveClass:
289 EmitOMPFlushDirective(S: cast<OMPFlushDirective>(Val: *S));
290 break;
291 case Stmt::OMPDepobjDirectiveClass:
292 EmitOMPDepobjDirective(S: cast<OMPDepobjDirective>(Val: *S));
293 break;
294 case Stmt::OMPScanDirectiveClass:
295 EmitOMPScanDirective(S: cast<OMPScanDirective>(Val: *S));
296 break;
297 case Stmt::OMPOrderedDirectiveClass:
298 EmitOMPOrderedDirective(S: cast<OMPOrderedDirective>(Val: *S));
299 break;
300 case Stmt::OMPAtomicDirectiveClass:
301 EmitOMPAtomicDirective(S: cast<OMPAtomicDirective>(Val: *S));
302 break;
303 case Stmt::OMPTargetDirectiveClass:
304 EmitOMPTargetDirective(S: cast<OMPTargetDirective>(Val: *S));
305 break;
306 case Stmt::OMPTeamsDirectiveClass:
307 EmitOMPTeamsDirective(S: cast<OMPTeamsDirective>(Val: *S));
308 break;
309 case Stmt::OMPCancellationPointDirectiveClass:
310 EmitOMPCancellationPointDirective(S: cast<OMPCancellationPointDirective>(Val: *S));
311 break;
312 case Stmt::OMPCancelDirectiveClass:
313 EmitOMPCancelDirective(S: cast<OMPCancelDirective>(Val: *S));
314 break;
315 case Stmt::OMPTargetDataDirectiveClass:
316 EmitOMPTargetDataDirective(S: cast<OMPTargetDataDirective>(Val: *S));
317 break;
318 case Stmt::OMPTargetEnterDataDirectiveClass:
319 EmitOMPTargetEnterDataDirective(S: cast<OMPTargetEnterDataDirective>(Val: *S));
320 break;
321 case Stmt::OMPTargetExitDataDirectiveClass:
322 EmitOMPTargetExitDataDirective(S: cast<OMPTargetExitDataDirective>(Val: *S));
323 break;
324 case Stmt::OMPTargetParallelDirectiveClass:
325 EmitOMPTargetParallelDirective(S: cast<OMPTargetParallelDirective>(Val: *S));
326 break;
327 case Stmt::OMPTargetParallelForDirectiveClass:
328 EmitOMPTargetParallelForDirective(S: cast<OMPTargetParallelForDirective>(Val: *S));
329 break;
330 case Stmt::OMPTaskLoopDirectiveClass:
331 EmitOMPTaskLoopDirective(S: cast<OMPTaskLoopDirective>(Val: *S));
332 break;
333 case Stmt::OMPTaskLoopSimdDirectiveClass:
334 EmitOMPTaskLoopSimdDirective(S: cast<OMPTaskLoopSimdDirective>(Val: *S));
335 break;
336 case Stmt::OMPMasterTaskLoopDirectiveClass:
337 EmitOMPMasterTaskLoopDirective(S: cast<OMPMasterTaskLoopDirective>(Val: *S));
338 break;
339 case Stmt::OMPMaskedTaskLoopDirectiveClass:
340 EmitOMPMaskedTaskLoopDirective(S: cast<OMPMaskedTaskLoopDirective>(Val: *S));
341 break;
342 case Stmt::OMPMasterTaskLoopSimdDirectiveClass:
343 EmitOMPMasterTaskLoopSimdDirective(
344 S: cast<OMPMasterTaskLoopSimdDirective>(Val: *S));
345 break;
346 case Stmt::OMPMaskedTaskLoopSimdDirectiveClass:
347 EmitOMPMaskedTaskLoopSimdDirective(
348 S: cast<OMPMaskedTaskLoopSimdDirective>(Val: *S));
349 break;
350 case Stmt::OMPParallelMasterTaskLoopDirectiveClass:
351 EmitOMPParallelMasterTaskLoopDirective(
352 S: cast<OMPParallelMasterTaskLoopDirective>(Val: *S));
353 break;
354 case Stmt::OMPParallelMaskedTaskLoopDirectiveClass:
355 EmitOMPParallelMaskedTaskLoopDirective(
356 S: cast<OMPParallelMaskedTaskLoopDirective>(Val: *S));
357 break;
358 case Stmt::OMPParallelMasterTaskLoopSimdDirectiveClass:
359 EmitOMPParallelMasterTaskLoopSimdDirective(
360 S: cast<OMPParallelMasterTaskLoopSimdDirective>(Val: *S));
361 break;
362 case Stmt::OMPParallelMaskedTaskLoopSimdDirectiveClass:
363 EmitOMPParallelMaskedTaskLoopSimdDirective(
364 S: cast<OMPParallelMaskedTaskLoopSimdDirective>(Val: *S));
365 break;
366 case Stmt::OMPDistributeDirectiveClass:
367 EmitOMPDistributeDirective(S: cast<OMPDistributeDirective>(Val: *S));
368 break;
369 case Stmt::OMPTargetUpdateDirectiveClass:
370 EmitOMPTargetUpdateDirective(S: cast<OMPTargetUpdateDirective>(Val: *S));
371 break;
372 case Stmt::OMPDistributeParallelForDirectiveClass:
373 EmitOMPDistributeParallelForDirective(
374 S: cast<OMPDistributeParallelForDirective>(Val: *S));
375 break;
376 case Stmt::OMPDistributeParallelForSimdDirectiveClass:
377 EmitOMPDistributeParallelForSimdDirective(
378 S: cast<OMPDistributeParallelForSimdDirective>(Val: *S));
379 break;
380 case Stmt::OMPDistributeSimdDirectiveClass:
381 EmitOMPDistributeSimdDirective(S: cast<OMPDistributeSimdDirective>(Val: *S));
382 break;
383 case Stmt::OMPTargetParallelForSimdDirectiveClass:
384 EmitOMPTargetParallelForSimdDirective(
385 S: cast<OMPTargetParallelForSimdDirective>(Val: *S));
386 break;
387 case Stmt::OMPTargetSimdDirectiveClass:
388 EmitOMPTargetSimdDirective(S: cast<OMPTargetSimdDirective>(Val: *S));
389 break;
390 case Stmt::OMPTeamsDistributeDirectiveClass:
391 EmitOMPTeamsDistributeDirective(S: cast<OMPTeamsDistributeDirective>(Val: *S));
392 break;
393 case Stmt::OMPTeamsDistributeSimdDirectiveClass:
394 EmitOMPTeamsDistributeSimdDirective(
395 S: cast<OMPTeamsDistributeSimdDirective>(Val: *S));
396 break;
397 case Stmt::OMPTeamsDistributeParallelForSimdDirectiveClass:
398 EmitOMPTeamsDistributeParallelForSimdDirective(
399 S: cast<OMPTeamsDistributeParallelForSimdDirective>(Val: *S));
400 break;
401 case Stmt::OMPTeamsDistributeParallelForDirectiveClass:
402 EmitOMPTeamsDistributeParallelForDirective(
403 S: cast<OMPTeamsDistributeParallelForDirective>(Val: *S));
404 break;
405 case Stmt::OMPTargetTeamsDirectiveClass:
406 EmitOMPTargetTeamsDirective(S: cast<OMPTargetTeamsDirective>(Val: *S));
407 break;
408 case Stmt::OMPTargetTeamsDistributeDirectiveClass:
409 EmitOMPTargetTeamsDistributeDirective(
410 S: cast<OMPTargetTeamsDistributeDirective>(Val: *S));
411 break;
412 case Stmt::OMPTargetTeamsDistributeParallelForDirectiveClass:
413 EmitOMPTargetTeamsDistributeParallelForDirective(
414 S: cast<OMPTargetTeamsDistributeParallelForDirective>(Val: *S));
415 break;
416 case Stmt::OMPTargetTeamsDistributeParallelForSimdDirectiveClass:
417 EmitOMPTargetTeamsDistributeParallelForSimdDirective(
418 S: cast<OMPTargetTeamsDistributeParallelForSimdDirective>(Val: *S));
419 break;
420 case Stmt::OMPTargetTeamsDistributeSimdDirectiveClass:
421 EmitOMPTargetTeamsDistributeSimdDirective(
422 S: cast<OMPTargetTeamsDistributeSimdDirective>(Val: *S));
423 break;
424 case Stmt::OMPInteropDirectiveClass:
425 EmitOMPInteropDirective(S: cast<OMPInteropDirective>(Val: *S));
426 break;
427 case Stmt::OMPDispatchDirectiveClass:
428 CGM.ErrorUnsupported(S, Type: "OpenMP dispatch directive");
429 break;
430 case Stmt::OMPScopeDirectiveClass:
431 EmitOMPScopeDirective(S: cast<OMPScopeDirective>(Val: *S));
432 break;
433 case Stmt::OMPMaskedDirectiveClass:
434 EmitOMPMaskedDirective(S: cast<OMPMaskedDirective>(Val: *S));
435 break;
436 case Stmt::OMPGenericLoopDirectiveClass:
437 EmitOMPGenericLoopDirective(S: cast<OMPGenericLoopDirective>(Val: *S));
438 break;
439 case Stmt::OMPTeamsGenericLoopDirectiveClass:
440 EmitOMPTeamsGenericLoopDirective(S: cast<OMPTeamsGenericLoopDirective>(Val: *S));
441 break;
442 case Stmt::OMPTargetTeamsGenericLoopDirectiveClass:
443 EmitOMPTargetTeamsGenericLoopDirective(
444 S: cast<OMPTargetTeamsGenericLoopDirective>(Val: *S));
445 break;
446 case Stmt::OMPParallelGenericLoopDirectiveClass:
447 EmitOMPParallelGenericLoopDirective(
448 S: cast<OMPParallelGenericLoopDirective>(Val: *S));
449 break;
450 case Stmt::OMPTargetParallelGenericLoopDirectiveClass:
451 EmitOMPTargetParallelGenericLoopDirective(
452 S: cast<OMPTargetParallelGenericLoopDirective>(Val: *S));
453 break;
454 case Stmt::OMPParallelMaskedDirectiveClass:
455 EmitOMPParallelMaskedDirective(S: cast<OMPParallelMaskedDirective>(Val: *S));
456 break;
457 case Stmt::OMPAssumeDirectiveClass:
458 EmitOMPAssumeDirective(S: cast<OMPAssumeDirective>(Val: *S));
459 break;
460 case Stmt::OpenACCComputeConstructClass:
461 EmitOpenACCComputeConstruct(S: cast<OpenACCComputeConstruct>(Val: *S));
462 break;
463 case Stmt::OpenACCLoopConstructClass:
464 EmitOpenACCLoopConstruct(S: cast<OpenACCLoopConstruct>(Val: *S));
465 break;
466 case Stmt::OpenACCCombinedConstructClass:
467 EmitOpenACCCombinedConstruct(S: cast<OpenACCCombinedConstruct>(Val: *S));
468 break;
469 case Stmt::OpenACCDataConstructClass:
470 EmitOpenACCDataConstruct(S: cast<OpenACCDataConstruct>(Val: *S));
471 break;
472 case Stmt::OpenACCEnterDataConstructClass:
473 EmitOpenACCEnterDataConstruct(S: cast<OpenACCEnterDataConstruct>(Val: *S));
474 break;
475 case Stmt::OpenACCExitDataConstructClass:
476 EmitOpenACCExitDataConstruct(S: cast<OpenACCExitDataConstruct>(Val: *S));
477 break;
478 case Stmt::OpenACCHostDataConstructClass:
479 EmitOpenACCHostDataConstruct(S: cast<OpenACCHostDataConstruct>(Val: *S));
480 break;
481 case Stmt::OpenACCWaitConstructClass:
482 EmitOpenACCWaitConstruct(S: cast<OpenACCWaitConstruct>(Val: *S));
483 break;
484 case Stmt::OpenACCInitConstructClass:
485 EmitOpenACCInitConstruct(S: cast<OpenACCInitConstruct>(Val: *S));
486 break;
487 case Stmt::OpenACCShutdownConstructClass:
488 EmitOpenACCShutdownConstruct(S: cast<OpenACCShutdownConstruct>(Val: *S));
489 break;
490 case Stmt::OpenACCSetConstructClass:
491 EmitOpenACCSetConstruct(S: cast<OpenACCSetConstruct>(Val: *S));
492 break;
493 case Stmt::OpenACCUpdateConstructClass:
494 EmitOpenACCUpdateConstruct(S: cast<OpenACCUpdateConstruct>(Val: *S));
495 break;
496 case Stmt::OpenACCAtomicConstructClass:
497 EmitOpenACCAtomicConstruct(S: cast<OpenACCAtomicConstruct>(Val: *S));
498 break;
499 case Stmt::OpenACCCacheConstructClass:
500 EmitOpenACCCacheConstruct(S: cast<OpenACCCacheConstruct>(Val: *S));
501 break;
502 }
503}
504
505bool CodeGenFunction::EmitSimpleStmt(const Stmt *S,
506 ArrayRef<const Attr *> Attrs) {
507 switch (S->getStmtClass()) {
508 default:
509 return false;
510 case Stmt::NullStmtClass:
511 break;
512 case Stmt::CompoundStmtClass:
513 EmitCompoundStmt(S: cast<CompoundStmt>(Val: *S));
514 break;
515 case Stmt::DeclStmtClass:
516 EmitDeclStmt(S: cast<DeclStmt>(Val: *S));
517 break;
518 case Stmt::LabelStmtClass:
519 EmitLabelStmt(S: cast<LabelStmt>(Val: *S));
520 break;
521 case Stmt::AttributedStmtClass:
522 EmitAttributedStmt(S: cast<AttributedStmt>(Val: *S));
523 break;
524 case Stmt::GotoStmtClass:
525 EmitGotoStmt(S: cast<GotoStmt>(Val: *S));
526 break;
527 case Stmt::BreakStmtClass:
528 EmitBreakStmt(S: cast<BreakStmt>(Val: *S));
529 break;
530 case Stmt::ContinueStmtClass:
531 EmitContinueStmt(S: cast<ContinueStmt>(Val: *S));
532 break;
533 case Stmt::DefaultStmtClass:
534 EmitDefaultStmt(S: cast<DefaultStmt>(Val: *S), Attrs);
535 break;
536 case Stmt::CaseStmtClass:
537 EmitCaseStmt(S: cast<CaseStmt>(Val: *S), Attrs);
538 break;
539 case Stmt::DeferStmtClass:
540 EmitDeferStmt(S: cast<DeferStmt>(Val: *S));
541 break;
542 case Stmt::SEHLeaveStmtClass:
543 EmitSEHLeaveStmt(S: cast<SEHLeaveStmt>(Val: *S));
544 break;
545 case Stmt::SYCLKernelCallStmtClass:
546 // SYCL kernel call statements are generated as wrappers around the body
547 // of functions declared with the sycl_kernel_entry_point attribute. Such
548 // functions are used to specify how a SYCL kernel (a function object) is
549 // to be invoked; the SYCL kernel call statement contains a transformed
550 // variation of the function body and is used to generate a SYCL kernel
551 // caller function; a function that serves as the device side entry point
552 // used to execute the SYCL kernel. The sycl_kernel_entry_point attributed
553 // function is invoked by host code in order to trigger emission of the
554 // device side SYCL kernel caller function and to generate metadata needed
555 // by SYCL run-time library implementations; the function is otherwise
556 // intended to have no effect. As such, the function body is not evaluated
557 // as part of the invocation during host compilation (and the function
558 // should not be called or emitted during device compilation); the SYCL
559 // kernel call statement is thus handled as a null statement for the
560 // purpose of code generation.
561 break;
562 }
563 return true;
564}
565
566/// EmitCompoundStmt - Emit a compound statement {..} node. If GetLast is true,
567/// this captures the expression result of the last sub-statement and returns it
568/// (for use by the statement expression extension).
569Address CodeGenFunction::EmitCompoundStmt(const CompoundStmt &S, bool GetLast,
570 AggValueSlot AggSlot) {
571 PrettyStackTraceLoc CrashInfo(getContext().getSourceManager(),S.getLBracLoc(),
572 "LLVM IR generation of compound statement ('{}')");
573
574 // Keep track of the current cleanup stack depth, including debug scopes.
575 LexicalScope Scope(*this, S.getSourceRange());
576
577 return EmitCompoundStmtWithoutScope(S, GetLast, AVS: AggSlot);
578}
579
580Address
581CodeGenFunction::EmitCompoundStmtWithoutScope(const CompoundStmt &S,
582 bool GetLast,
583 AggValueSlot AggSlot) {
584
585 for (CompoundStmt::const_body_iterator I = S.body_begin(),
586 E = S.body_end() - GetLast;
587 I != E; ++I)
588 EmitStmt(S: *I);
589
590 Address RetAlloca = Address::invalid();
591 if (GetLast) {
592 // We have to special case labels here. They are statements, but when put
593 // at the end of a statement expression, they yield the value of their
594 // subexpression. Handle this by walking through all labels we encounter,
595 // emitting them before we evaluate the subexpr.
596 // Similar issues arise for attributed statements.
597 const Stmt *LastStmt = S.body_back();
598 while (!isa<Expr>(Val: LastStmt)) {
599 if (const auto *LS = dyn_cast<LabelStmt>(Val: LastStmt)) {
600 EmitLabel(D: LS->getDecl());
601 LastStmt = LS->getSubStmt();
602 } else if (const auto *AS = dyn_cast<AttributedStmt>(Val: LastStmt)) {
603 // FIXME: Update this if we ever have attributes that affect the
604 // semantics of an expression.
605 LastStmt = AS->getSubStmt();
606 } else {
607 llvm_unreachable("unknown value statement");
608 }
609 }
610
611 EnsureInsertPoint();
612
613 const Expr *E = cast<Expr>(Val: LastStmt);
614 QualType ExprTy = E->getType();
615 if (hasAggregateEvaluationKind(T: ExprTy)) {
616 EmitAggExpr(E, AS: AggSlot);
617 } else {
618 // We can't return an RValue here because there might be cleanups at
619 // the end of the StmtExpr. Because of that, we have to emit the result
620 // here into a temporary alloca.
621 RetAlloca = CreateMemTemp(T: ExprTy);
622 EmitAnyExprToMem(E, Location: RetAlloca, Quals: Qualifiers(),
623 /*IsInit*/ IsInitializer: false);
624 }
625 }
626
627 return RetAlloca;
628}
629
630void CodeGenFunction::SimplifyForwardingBlocks(llvm::BasicBlock *BB) {
631 llvm::BranchInst *BI = dyn_cast<llvm::BranchInst>(Val: BB->getTerminator());
632
633 // If there is a cleanup stack, then we it isn't worth trying to
634 // simplify this block (we would need to remove it from the scope map
635 // and cleanup entry).
636 if (!EHStack.empty())
637 return;
638
639 // Can only simplify direct branches.
640 if (!BI || !BI->isUnconditional())
641 return;
642
643 // Can only simplify empty blocks.
644 if (BI->getIterator() != BB->begin())
645 return;
646
647 BB->replaceAllUsesWith(V: BI->getSuccessor(i: 0));
648 BI->eraseFromParent();
649 BB->eraseFromParent();
650}
651
652void CodeGenFunction::EmitBlock(llvm::BasicBlock *BB, bool IsFinished) {
653 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
654
655 // Fall out of the current block (if necessary).
656 EmitBranch(Block: BB);
657
658 if (IsFinished && BB->use_empty()) {
659 delete BB;
660 return;
661 }
662
663 // Place the block after the current block, if possible, or else at
664 // the end of the function.
665 if (CurBB && CurBB->getParent())
666 CurFn->insert(Position: std::next(x: CurBB->getIterator()), BB);
667 else
668 CurFn->insert(Position: CurFn->end(), BB);
669 Builder.SetInsertPoint(BB);
670}
671
672void CodeGenFunction::EmitBranch(llvm::BasicBlock *Target) {
673 // Emit a branch from the current block to the target one if this
674 // was a real block. If this was just a fall-through block after a
675 // terminator, don't emit it.
676 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
677
678 if (!CurBB || CurBB->getTerminator()) {
679 // If there is no insert point or the previous block is already
680 // terminated, don't touch it.
681 } else {
682 // Otherwise, create a fall-through branch.
683 Builder.CreateBr(Dest: Target);
684 }
685
686 Builder.ClearInsertionPoint();
687}
688
689void CodeGenFunction::EmitBlockAfterUses(llvm::BasicBlock *block) {
690 bool inserted = false;
691 for (llvm::User *u : block->users()) {
692 if (llvm::Instruction *insn = dyn_cast<llvm::Instruction>(Val: u)) {
693 CurFn->insert(Position: std::next(x: insn->getParent()->getIterator()), BB: block);
694 inserted = true;
695 break;
696 }
697 }
698
699 if (!inserted)
700 CurFn->insert(Position: CurFn->end(), BB: block);
701
702 Builder.SetInsertPoint(block);
703}
704
705CodeGenFunction::JumpDest
706CodeGenFunction::getJumpDestForLabel(const LabelDecl *D) {
707 JumpDest &Dest = LabelMap[D];
708 if (Dest.isValid()) return Dest;
709
710 // Create, but don't insert, the new block.
711 Dest = JumpDest(createBasicBlock(name: D->getName()),
712 EHScopeStack::stable_iterator::invalid(),
713 NextCleanupDestIndex++);
714 return Dest;
715}
716
717void CodeGenFunction::EmitLabel(const LabelDecl *D) {
718 // Add this label to the current lexical scope if we're within any
719 // normal cleanups. Jumps "in" to this label --- when permitted by
720 // the language --- may need to be routed around such cleanups.
721 if (EHStack.hasNormalCleanups() && CurLexicalScope)
722 CurLexicalScope->addLabel(label: D);
723
724 JumpDest &Dest = LabelMap[D];
725
726 // If we didn't need a forward reference to this label, just go
727 // ahead and create a destination at the current scope.
728 if (!Dest.isValid()) {
729 Dest = getJumpDestInCurrentScope(Name: D->getName());
730
731 // Otherwise, we need to give this label a target depth and remove
732 // it from the branch-fixups list.
733 } else {
734 assert(!Dest.getScopeDepth().isValid() && "already emitted label!");
735 Dest.setScopeDepth(EHStack.stable_begin());
736 ResolveBranchFixups(Target: Dest.getBlock());
737 }
738
739 EmitBlock(BB: Dest.getBlock());
740
741 // Emit debug info for labels.
742 if (CGDebugInfo *DI = getDebugInfo()) {
743 if (CGM.getCodeGenOpts().hasReducedDebugInfo()) {
744 DI->setLocation(D->getLocation());
745 DI->EmitLabel(D, Builder);
746 }
747 }
748
749 incrementProfileCounter(S: D->getStmt());
750}
751
752/// Change the cleanup scope of the labels in this lexical scope to
753/// match the scope of the enclosing context.
754void CodeGenFunction::LexicalScope::rescopeLabels() {
755 assert(!Labels.empty());
756 EHScopeStack::stable_iterator innermostScope
757 = CGF.EHStack.getInnermostNormalCleanup();
758
759 // Change the scope depth of all the labels.
760 for (const LabelDecl *Label : Labels) {
761 assert(CGF.LabelMap.count(Label));
762 JumpDest &dest = CGF.LabelMap.find(Val: Label)->second;
763 assert(dest.getScopeDepth().isValid());
764 assert(innermostScope.encloses(dest.getScopeDepth()));
765 dest.setScopeDepth(innermostScope);
766 }
767
768 // Reparent the labels if the new scope also has cleanups.
769 if (innermostScope != EHScopeStack::stable_end() && ParentScope) {
770 ParentScope->Labels.append(in_start: Labels.begin(), in_end: Labels.end());
771 }
772}
773
774
775void CodeGenFunction::EmitLabelStmt(const LabelStmt &S) {
776 EmitLabel(D: S.getDecl());
777
778 // IsEHa - emit eha.scope.begin if it's a side entry of a scope
779 if (getLangOpts().EHAsynch && S.isSideEntry())
780 EmitSehCppScopeBegin();
781
782 EmitStmt(S: S.getSubStmt());
783}
784
785void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
786 bool nomerge = false;
787 bool noinline = false;
788 bool alwaysinline = false;
789 bool noconvergent = false;
790 HLSLControlFlowHintAttr::Spelling flattenOrBranch =
791 HLSLControlFlowHintAttr::SpellingNotCalculated;
792 const CallExpr *musttail = nullptr;
793 const AtomicAttr *AA = nullptr;
794
795 for (const auto *A : S.getAttrs()) {
796 switch (A->getKind()) {
797 default:
798 break;
799 case attr::NoMerge:
800 nomerge = true;
801 break;
802 case attr::NoInline:
803 noinline = true;
804 break;
805 case attr::AlwaysInline:
806 alwaysinline = true;
807 break;
808 case attr::NoConvergent:
809 noconvergent = true;
810 break;
811 case attr::MustTail: {
812 const Stmt *Sub = S.getSubStmt();
813 const ReturnStmt *R = cast<ReturnStmt>(Val: Sub);
814 musttail = cast<CallExpr>(Val: R->getRetValue()->IgnoreParens());
815 } break;
816 case attr::CXXAssume: {
817 const Expr *Assumption = cast<CXXAssumeAttr>(Val: A)->getAssumption();
818 if (getLangOpts().CXXAssumptions && Builder.GetInsertBlock() &&
819 !Assumption->HasSideEffects(Ctx: getContext())) {
820 llvm::Value *AssumptionVal = EmitCheckedArgForAssume(E: Assumption);
821 Builder.CreateAssumption(Cond: AssumptionVal);
822 }
823 } break;
824 case attr::Atomic:
825 AA = cast<AtomicAttr>(Val: A);
826 break;
827 case attr::HLSLControlFlowHint: {
828 flattenOrBranch = cast<HLSLControlFlowHintAttr>(Val: A)->getSemanticSpelling();
829 } break;
830 }
831 }
832 SaveAndRestore save_nomerge(InNoMergeAttributedStmt, nomerge);
833 SaveAndRestore save_noinline(InNoInlineAttributedStmt, noinline);
834 SaveAndRestore save_alwaysinline(InAlwaysInlineAttributedStmt, alwaysinline);
835 SaveAndRestore save_noconvergent(InNoConvergentAttributedStmt, noconvergent);
836 SaveAndRestore save_musttail(MustTailCall, musttail);
837 SaveAndRestore save_flattenOrBranch(HLSLControlFlowAttr, flattenOrBranch);
838 CGAtomicOptionsRAII AORAII(CGM, AA);
839 EmitStmt(S: S.getSubStmt(), Attrs: S.getAttrs());
840}
841
842void CodeGenFunction::EmitGotoStmt(const GotoStmt &S) {
843 // If this code is reachable then emit a stop point (if generating
844 // debug info). We have to do this ourselves because we are on the
845 // "simple" statement path.
846 if (HaveInsertPoint())
847 EmitStopPoint(S: &S);
848
849 ApplyAtomGroup Grp(getDebugInfo());
850 EmitBranchThroughCleanup(Dest: getJumpDestForLabel(D: S.getLabel()));
851}
852
853
854void CodeGenFunction::EmitIndirectGotoStmt(const IndirectGotoStmt &S) {
855 ApplyAtomGroup Grp(getDebugInfo());
856 if (const LabelDecl *Target = S.getConstantTarget()) {
857 EmitBranchThroughCleanup(Dest: getJumpDestForLabel(D: Target));
858 return;
859 }
860
861 // Ensure that we have an i8* for our PHI node.
862 llvm::Value *V = Builder.CreateBitCast(V: EmitScalarExpr(E: S.getTarget()),
863 DestTy: Int8PtrTy, Name: "addr");
864 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
865
866 // Get the basic block for the indirect goto.
867 llvm::BasicBlock *IndGotoBB = GetIndirectGotoBlock();
868
869 // The first instruction in the block has to be the PHI for the switch dest,
870 // add an entry for this branch.
871 cast<llvm::PHINode>(Val: IndGotoBB->begin())->addIncoming(V, BB: CurBB);
872
873 EmitBranch(Target: IndGotoBB);
874 if (CurBB && CurBB->getTerminator())
875 addInstToCurrentSourceAtom(KeyInstruction: CurBB->getTerminator(), Backup: nullptr);
876}
877
878void CodeGenFunction::EmitIfStmt(const IfStmt &S) {
879 const Stmt *Else = S.getElse();
880
881 // The else branch of a consteval if statement is always the only branch that
882 // can be runtime evaluated.
883 if (S.isConsteval()) {
884 const Stmt *Executed = S.isNegatedConsteval() ? S.getThen() : Else;
885 if (Executed) {
886 RunCleanupsScope ExecutedScope(*this);
887 EmitStmt(S: Executed);
888 }
889 return;
890 }
891
892 // C99 6.8.4.1: The first substatement is executed if the expression compares
893 // unequal to 0. The condition must be a scalar type.
894 LexicalScope ConditionScope(*this, S.getCond()->getSourceRange());
895 ApplyDebugLocation DL(*this, S.getCond());
896
897 if (S.getInit())
898 EmitStmt(S: S.getInit());
899
900 if (S.getConditionVariable())
901 EmitDecl(D: *S.getConditionVariable());
902
903 // If the condition constant folds and can be elided, try to avoid emitting
904 // the condition and the dead arm of the if/else.
905 bool CondConstant;
906 if (ConstantFoldsToSimpleInteger(Cond: S.getCond(), Result&: CondConstant,
907 AllowLabels: S.isConstexpr())) {
908 // Figure out which block (then or else) is executed.
909 const Stmt *Executed = S.getThen();
910 const Stmt *Skipped = Else;
911 if (!CondConstant) // Condition false?
912 std::swap(a&: Executed, b&: Skipped);
913
914 // If the skipped block has no labels in it, just emit the executed block.
915 // This avoids emitting dead code and simplifies the CFG substantially.
916 if (S.isConstexpr() || !ContainsLabel(S: Skipped)) {
917 incrementProfileCounter(ExecSkip: CondConstant ? UseExecPath : UseSkipPath, S: &S,
918 /*UseBoth=*/true);
919 if (Executed) {
920 MaybeEmitDeferredVarDeclInit(var: S.getConditionVariable());
921 RunCleanupsScope ExecutedScope(*this);
922 EmitStmt(S: Executed);
923 }
924 PGO->markStmtMaybeUsed(S: Skipped);
925 return;
926 }
927 }
928
929 auto HasSkip = hasSkipCounter(S: &S);
930
931 // Otherwise, the condition did not fold, or we couldn't elide it. Just emit
932 // the conditional branch.
933 llvm::BasicBlock *ThenBlock = createBasicBlock(name: "if.then");
934 llvm::BasicBlock *ContBlock = createBasicBlock(name: "if.end");
935 llvm::BasicBlock *ElseBlock =
936 (Else || HasSkip ? createBasicBlock(name: "if.else") : ContBlock);
937 // Prefer the PGO based weights over the likelihood attribute.
938 // When the build isn't optimized the metadata isn't used, so don't generate
939 // it.
940 // Also, differentiate between disabled PGO and a never executed branch with
941 // PGO. Assuming PGO is in use:
942 // - we want to ignore the [[likely]] attribute if the branch is never
943 // executed,
944 // - assuming the profile is poor, preserving the attribute may still be
945 // beneficial.
946 // As an approximation, preserve the attribute only if both the branch and the
947 // parent context were not executed.
948 Stmt::Likelihood LH = Stmt::LH_None;
949 uint64_t ThenCount = getProfileCount(S: S.getThen());
950 if (!ThenCount && !getCurrentProfileCount() &&
951 CGM.getCodeGenOpts().OptimizationLevel)
952 LH = Stmt::getLikelihood(Then: S.getThen(), Else);
953
954 // When measuring MC/DC, always fully evaluate the condition up front using
955 // EvaluateExprAsBool() so that the test vector bitmap can be updated prior to
956 // executing the body of the if.then or if.else. This is useful for when
957 // there is a 'return' within the body, but this is particularly beneficial
958 // when one if-stmt is nested within another if-stmt so that all of the MC/DC
959 // updates are kept linear and consistent.
960 if (!CGM.getCodeGenOpts().MCDCCoverage) {
961 EmitBranchOnBoolExpr(Cond: S.getCond(), TrueBlock: ThenBlock, FalseBlock: ElseBlock, TrueCount: ThenCount, LH,
962 /*ConditionalOp=*/nullptr,
963 /*ConditionalDecl=*/S.getConditionVariable());
964 } else {
965 llvm::Value *BoolCondVal = EvaluateExprAsBool(E: S.getCond());
966 MaybeEmitDeferredVarDeclInit(var: S.getConditionVariable());
967 Builder.CreateCondBr(Cond: BoolCondVal, True: ThenBlock, False: ElseBlock);
968 }
969
970 // Emit the 'then' code.
971 EmitBlock(BB: ThenBlock);
972 incrementProfileCounter(ExecSkip: UseExecPath, S: &S);
973 {
974 RunCleanupsScope ThenScope(*this);
975 EmitStmt(S: S.getThen());
976 }
977 EmitBranch(Target: ContBlock);
978
979 // Emit the 'else' code if present.
980 if (Else) {
981 {
982 // There is no need to emit line number for an unconditional branch.
983 auto NL = ApplyDebugLocation::CreateEmpty(CGF&: *this);
984 EmitBlock(BB: ElseBlock);
985 }
986 // Add a counter to else block unless it has CounterExpr.
987 if (HasSkip)
988 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S);
989 {
990 RunCleanupsScope ElseScope(*this);
991 EmitStmt(S: Else);
992 }
993 {
994 // There is no need to emit line number for an unconditional branch.
995 auto NL = ApplyDebugLocation::CreateEmpty(CGF&: *this);
996 EmitBranch(Target: ContBlock);
997 }
998 } else if (HasSkip) {
999 EmitBlock(BB: ElseBlock);
1000 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S);
1001 EmitBranch(Target: ContBlock);
1002 }
1003
1004 // Emit the continuation block for code after the if.
1005 EmitBlock(BB: ContBlock, IsFinished: true);
1006}
1007
1008bool CodeGenFunction::checkIfLoopMustProgress(const Expr *ControllingExpression,
1009 bool HasEmptyBody) {
1010 if (CGM.getCodeGenOpts().getFiniteLoops() ==
1011 CodeGenOptions::FiniteLoopsKind::Never)
1012 return false;
1013
1014 // Now apply rules for plain C (see 6.8.5.6 in C11).
1015 // Loops with constant conditions do not have to make progress in any C
1016 // version.
1017 // As an extension, we consisider loops whose constant expression
1018 // can be constant-folded.
1019 Expr::EvalResult Result;
1020 bool CondIsConstInt =
1021 !ControllingExpression ||
1022 (ControllingExpression->EvaluateAsInt(Result, Ctx: getContext()) &&
1023 Result.Val.isInt());
1024
1025 bool CondIsTrue = CondIsConstInt && (!ControllingExpression ||
1026 Result.Val.getInt().getBoolValue());
1027
1028 // Loops with non-constant conditions must make progress in C11 and later.
1029 if (getLangOpts().C11 && !CondIsConstInt)
1030 return true;
1031
1032 // [C++26][intro.progress] (DR)
1033 // The implementation may assume that any thread will eventually do one of the
1034 // following:
1035 // [...]
1036 // - continue execution of a trivial infinite loop ([stmt.iter.general]).
1037 if (CGM.getCodeGenOpts().getFiniteLoops() ==
1038 CodeGenOptions::FiniteLoopsKind::Always ||
1039 getLangOpts().CPlusPlus11) {
1040 if (HasEmptyBody && CondIsTrue) {
1041 CurFn->removeFnAttr(Kind: llvm::Attribute::MustProgress);
1042 return false;
1043 }
1044 return true;
1045 }
1046 return false;
1047}
1048
1049// [C++26][stmt.iter.general] (DR)
1050// A trivially empty iteration statement is an iteration statement matching one
1051// of the following forms:
1052// - while ( expression ) ;
1053// - while ( expression ) { }
1054// - do ; while ( expression ) ;
1055// - do { } while ( expression ) ;
1056// - for ( init-statement expression(opt); ) ;
1057// - for ( init-statement expression(opt); ) { }
1058template <typename LoopStmt> static bool hasEmptyLoopBody(const LoopStmt &S) {
1059 if constexpr (std::is_same_v<LoopStmt, ForStmt>) {
1060 if (S.getInc())
1061 return false;
1062 }
1063 const Stmt *Body = S.getBody();
1064 if (!Body || isa<NullStmt>(Val: Body))
1065 return true;
1066 if (const CompoundStmt *Compound = dyn_cast<CompoundStmt>(Val: Body))
1067 return Compound->body_empty();
1068 return false;
1069}
1070
1071void CodeGenFunction::EmitWhileStmt(const WhileStmt &S,
1072 ArrayRef<const Attr *> WhileAttrs) {
1073 // Emit the header for the loop, which will also become
1074 // the continue target.
1075 JumpDest LoopHeader = getJumpDestInCurrentScope(Name: "while.cond");
1076 EmitBlock(BB: LoopHeader.getBlock());
1077
1078 if (CGM.shouldEmitConvergenceTokens())
1079 ConvergenceTokenStack.push_back(
1080 Elt: emitConvergenceLoopToken(BB: LoopHeader.getBlock()));
1081
1082 // Create an exit block for when the condition fails, which will
1083 // also become the break target.
1084 JumpDest LoopExit = getJumpDestInCurrentScope(Name: "while.end");
1085
1086 // Store the blocks to use for break and continue.
1087 BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, LoopHeader));
1088
1089 // C++ [stmt.while]p2:
1090 // When the condition of a while statement is a declaration, the
1091 // scope of the variable that is declared extends from its point
1092 // of declaration (3.3.2) to the end of the while statement.
1093 // [...]
1094 // The object created in a condition is destroyed and created
1095 // with each iteration of the loop.
1096 RunCleanupsScope ConditionScope(*this);
1097
1098 if (S.getConditionVariable())
1099 EmitDecl(D: *S.getConditionVariable());
1100
1101 // Evaluate the conditional in the while header. C99 6.8.5.1: The
1102 // evaluation of the controlling expression takes place before each
1103 // execution of the loop body.
1104 llvm::Value *BoolCondVal = EvaluateExprAsBool(E: S.getCond());
1105
1106 MaybeEmitDeferredVarDeclInit(var: S.getConditionVariable());
1107
1108 // while(1) is common, avoid extra exit blocks. Be sure
1109 // to correctly handle break/continue though.
1110 llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(Val: BoolCondVal);
1111 bool EmitBoolCondBranch = !C || !C->isOne();
1112 const SourceRange &R = S.getSourceRange();
1113 LoopStack.push(Header: LoopHeader.getBlock(), Ctx&: CGM.getContext(), CGOpts: CGM.getCodeGenOpts(),
1114 Attrs: WhileAttrs, StartLoc: SourceLocToDebugLoc(Location: R.getBegin()),
1115 EndLoc: SourceLocToDebugLoc(Location: R.getEnd()),
1116 MustProgress: checkIfLoopMustProgress(ControllingExpression: S.getCond(), HasEmptyBody: hasEmptyLoopBody(S)));
1117
1118 // As long as the condition is true, go to the loop body.
1119 llvm::BasicBlock *LoopBody = createBasicBlock(name: "while.body");
1120 if (EmitBoolCondBranch) {
1121 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1122 if (hasSkipCounter(S: &S) || ConditionScope.requiresCleanups())
1123 ExitBlock = createBasicBlock(name: "while.exit");
1124 llvm::MDNode *Weights =
1125 createProfileWeightsForLoop(Cond: S.getCond(), LoopCount: getProfileCount(S: S.getBody()));
1126 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
1127 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
1128 Cond: BoolCondVal, LH: Stmt::getLikelihood(S: S.getBody()));
1129 auto *I = Builder.CreateCondBr(Cond: BoolCondVal, True: LoopBody, False: ExitBlock, BranchWeights: Weights);
1130 // Key Instructions: Emit the condition and branch as separate source
1131 // location atoms otherwise we may omit a step onto the loop condition in
1132 // favour of the `while` keyword.
1133 // FIXME: We could have the branch as the backup location for the condition,
1134 // which would probably be a better experience. Explore this later.
1135 if (auto *CondI = dyn_cast<llvm::Instruction>(Val: BoolCondVal))
1136 addInstToNewSourceAtom(KeyInstruction: CondI, Backup: nullptr);
1137 addInstToNewSourceAtom(KeyInstruction: I, Backup: nullptr);
1138
1139 if (ExitBlock != LoopExit.getBlock()) {
1140 EmitBlock(BB: ExitBlock);
1141 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S);
1142 EmitBranchThroughCleanup(Dest: LoopExit);
1143 }
1144 } else if (const Attr *A = Stmt::getLikelihoodAttr(S: S.getBody())) {
1145 CGM.getDiags().Report(Loc: A->getLocation(),
1146 DiagID: diag::warn_attribute_has_no_effect_on_infinite_loop)
1147 << A << A->getRange();
1148 CGM.getDiags().Report(
1149 Loc: S.getWhileLoc(),
1150 DiagID: diag::note_attribute_has_no_effect_on_infinite_loop_here)
1151 << SourceRange(S.getWhileLoc(), S.getRParenLoc());
1152 }
1153
1154 // Emit the loop body. We have to emit this in a cleanup scope
1155 // because it might be a singleton DeclStmt.
1156 {
1157 RunCleanupsScope BodyScope(*this);
1158 EmitBlock(BB: LoopBody);
1159 incrementProfileCounter(ExecSkip: UseExecPath, S: &S);
1160 EmitStmt(S: S.getBody());
1161 }
1162
1163 BreakContinueStack.pop_back();
1164
1165 // Immediately force cleanup.
1166 ConditionScope.ForceCleanup();
1167
1168 EmitStopPoint(S: &S);
1169 // Branch to the loop header again.
1170 EmitBranch(Target: LoopHeader.getBlock());
1171
1172 LoopStack.pop();
1173
1174 // Emit the exit block.
1175 EmitBlock(BB: LoopExit.getBlock(), IsFinished: true);
1176
1177 // The LoopHeader typically is just a branch if we skipped emitting
1178 // a branch, try to erase it.
1179 if (!EmitBoolCondBranch) {
1180 SimplifyForwardingBlocks(BB: LoopHeader.getBlock());
1181 PGO->markStmtAsUsed(Skipped: true, S: &S);
1182 }
1183
1184 if (CGM.shouldEmitConvergenceTokens())
1185 ConvergenceTokenStack.pop_back();
1186}
1187
1188void CodeGenFunction::EmitDoStmt(const DoStmt &S,
1189 ArrayRef<const Attr *> DoAttrs) {
1190 JumpDest LoopExit = getJumpDestInCurrentScope(Name: "do.end");
1191 JumpDest LoopCond = getJumpDestInCurrentScope(Name: "do.cond");
1192
1193 uint64_t ParentCount = getCurrentProfileCount();
1194
1195 // Store the blocks to use for break and continue.
1196 BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, LoopCond));
1197
1198 // Emit the body of the loop.
1199 llvm::BasicBlock *LoopBody = createBasicBlock(name: "do.body");
1200
1201 EmitBlockWithFallThrough(BB: LoopBody, S: &S);
1202
1203 if (CGM.shouldEmitConvergenceTokens())
1204 ConvergenceTokenStack.push_back(Elt: emitConvergenceLoopToken(BB: LoopBody));
1205
1206 {
1207 RunCleanupsScope BodyScope(*this);
1208 EmitStmt(S: S.getBody());
1209 }
1210
1211 EmitBlock(BB: LoopCond.getBlock());
1212
1213 // C99 6.8.5.2: "The evaluation of the controlling expression takes place
1214 // after each execution of the loop body."
1215
1216 // Evaluate the conditional in the while header.
1217 // C99 6.8.5p2/p4: The first substatement is executed if the expression
1218 // compares unequal to 0. The condition must be a scalar type.
1219 llvm::Value *BoolCondVal = EvaluateExprAsBool(E: S.getCond());
1220
1221 BreakContinueStack.pop_back();
1222
1223 // "do {} while (0)" is common in macros, avoid extra blocks. Be sure
1224 // to correctly handle break/continue though.
1225 llvm::ConstantInt *C = dyn_cast<llvm::ConstantInt>(Val: BoolCondVal);
1226 bool EmitBoolCondBranch = !C || !C->isZero();
1227
1228 const SourceRange &R = S.getSourceRange();
1229 LoopStack.push(Header: LoopBody, Ctx&: CGM.getContext(), CGOpts: CGM.getCodeGenOpts(), Attrs: DoAttrs,
1230 StartLoc: SourceLocToDebugLoc(Location: R.getBegin()),
1231 EndLoc: SourceLocToDebugLoc(Location: R.getEnd()),
1232 MustProgress: checkIfLoopMustProgress(ControllingExpression: S.getCond(), HasEmptyBody: hasEmptyLoopBody(S)));
1233
1234 auto *LoopFalse = (hasSkipCounter(S: &S) ? createBasicBlock(name: "do.loopfalse")
1235 : LoopExit.getBlock());
1236
1237 // As long as the condition is true, iterate the loop.
1238 if (EmitBoolCondBranch) {
1239 uint64_t BackedgeCount = getProfileCount(S: S.getBody()) - ParentCount;
1240 auto *I = Builder.CreateCondBr(
1241 Cond: BoolCondVal, True: LoopBody, False: LoopFalse,
1242 BranchWeights: createProfileWeightsForLoop(Cond: S.getCond(), LoopCount: BackedgeCount));
1243
1244 // Key Instructions: Emit the condition and branch as separate source
1245 // location atoms otherwise we may omit a step onto the loop condition in
1246 // favour of the closing brace.
1247 // FIXME: We could have the branch as the backup location for the condition,
1248 // which would probably be a better experience (no jumping to the brace).
1249 if (auto *CondI = dyn_cast<llvm::Instruction>(Val: BoolCondVal))
1250 addInstToNewSourceAtom(KeyInstruction: CondI, Backup: nullptr);
1251 addInstToNewSourceAtom(KeyInstruction: I, Backup: nullptr);
1252 }
1253
1254 LoopStack.pop();
1255
1256 if (LoopFalse != LoopExit.getBlock()) {
1257 EmitBlock(BB: LoopFalse);
1258 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S, /*UseBoth=*/true);
1259 }
1260
1261 // Emit the exit block.
1262 EmitBlock(BB: LoopExit.getBlock());
1263
1264 // The DoCond block typically is just a branch if we skipped
1265 // emitting a branch, try to erase it.
1266 if (!EmitBoolCondBranch)
1267 SimplifyForwardingBlocks(BB: LoopCond.getBlock());
1268
1269 if (CGM.shouldEmitConvergenceTokens())
1270 ConvergenceTokenStack.pop_back();
1271}
1272
1273void CodeGenFunction::EmitForStmt(const ForStmt &S,
1274 ArrayRef<const Attr *> ForAttrs) {
1275 JumpDest LoopExit = getJumpDestInCurrentScope(Name: "for.end");
1276
1277 std::optional<LexicalScope> ForScope;
1278 if (getLangOpts().C99 || getLangOpts().CPlusPlus)
1279 ForScope.emplace(args&: *this, args: S.getSourceRange());
1280
1281 // Evaluate the first part before the loop.
1282 if (S.getInit())
1283 EmitStmt(S: S.getInit());
1284
1285 // Start the loop with a block that tests the condition.
1286 // If there's an increment, the continue scope will be overwritten
1287 // later.
1288 JumpDest CondDest = getJumpDestInCurrentScope(Name: "for.cond");
1289 llvm::BasicBlock *CondBlock = CondDest.getBlock();
1290 EmitBlock(BB: CondBlock);
1291
1292 if (CGM.shouldEmitConvergenceTokens())
1293 ConvergenceTokenStack.push_back(Elt: emitConvergenceLoopToken(BB: CondBlock));
1294
1295 const SourceRange &R = S.getSourceRange();
1296 LoopStack.push(Header: CondBlock, Ctx&: CGM.getContext(), CGOpts: CGM.getCodeGenOpts(), Attrs: ForAttrs,
1297 StartLoc: SourceLocToDebugLoc(Location: R.getBegin()),
1298 EndLoc: SourceLocToDebugLoc(Location: R.getEnd()),
1299 MustProgress: checkIfLoopMustProgress(ControllingExpression: S.getCond(), HasEmptyBody: hasEmptyLoopBody(S)));
1300
1301 // Create a cleanup scope for the condition variable cleanups.
1302 LexicalScope ConditionScope(*this, S.getSourceRange());
1303
1304 // If the for loop doesn't have an increment we can just use the condition as
1305 // the continue block. Otherwise, if there is no condition variable, we can
1306 // form the continue block now. If there is a condition variable, we can't
1307 // form the continue block until after we've emitted the condition, because
1308 // the condition is in scope in the increment, but Sema's jump diagnostics
1309 // ensure that there are no continues from the condition variable that jump
1310 // to the loop increment.
1311 JumpDest Continue;
1312 if (!S.getInc())
1313 Continue = CondDest;
1314 else if (!S.getConditionVariable())
1315 Continue = getJumpDestInCurrentScope(Name: "for.inc");
1316 BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, Continue));
1317
1318 if (S.getCond()) {
1319 // If the for statement has a condition scope, emit the local variable
1320 // declaration.
1321 if (S.getConditionVariable()) {
1322 EmitDecl(D: *S.getConditionVariable());
1323
1324 // We have entered the condition variable's scope, so we're now able to
1325 // jump to the continue block.
1326 Continue = S.getInc() ? getJumpDestInCurrentScope(Name: "for.inc") : CondDest;
1327 BreakContinueStack.back().ContinueBlock = Continue;
1328 }
1329
1330 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1331 // If there are any cleanups between here and the loop-exit scope,
1332 // create a block to stage a loop exit along.
1333 if (hasSkipCounter(S: &S) || (ForScope && ForScope->requiresCleanups()))
1334 ExitBlock = createBasicBlock(name: "for.cond.cleanup");
1335
1336 // As long as the condition is true, iterate the loop.
1337 llvm::BasicBlock *ForBody = createBasicBlock(name: "for.body");
1338
1339 // C99 6.8.5p2/p4: The first substatement is executed if the expression
1340 // compares unequal to 0. The condition must be a scalar type.
1341 llvm::Value *BoolCondVal = EvaluateExprAsBool(E: S.getCond());
1342
1343 MaybeEmitDeferredVarDeclInit(var: S.getConditionVariable());
1344
1345 llvm::MDNode *Weights =
1346 createProfileWeightsForLoop(Cond: S.getCond(), LoopCount: getProfileCount(S: S.getBody()));
1347 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
1348 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
1349 Cond: BoolCondVal, LH: Stmt::getLikelihood(S: S.getBody()));
1350
1351 auto *I = Builder.CreateCondBr(Cond: BoolCondVal, True: ForBody, False: ExitBlock, BranchWeights: Weights);
1352 // Key Instructions: Emit the condition and branch as separate atoms to
1353 // match existing loop stepping behaviour. FIXME: We could have the branch
1354 // as the backup location for the condition, which would probably be a
1355 // better experience (no jumping to the brace).
1356 if (auto *CondI = dyn_cast<llvm::Instruction>(Val: BoolCondVal))
1357 addInstToNewSourceAtom(KeyInstruction: CondI, Backup: nullptr);
1358 addInstToNewSourceAtom(KeyInstruction: I, Backup: nullptr);
1359
1360 if (ExitBlock != LoopExit.getBlock()) {
1361 EmitBlock(BB: ExitBlock);
1362 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S);
1363 EmitBranchThroughCleanup(Dest: LoopExit);
1364 }
1365
1366 EmitBlock(BB: ForBody);
1367 } else {
1368 // Treat it as a non-zero constant. Don't even create a new block for the
1369 // body, just fall into it.
1370 PGO->markStmtAsUsed(Skipped: true, S: &S);
1371 }
1372
1373 incrementProfileCounter(ExecSkip: UseExecPath, S: &S);
1374
1375 {
1376 // Create a separate cleanup scope for the body, in case it is not
1377 // a compound statement.
1378 RunCleanupsScope BodyScope(*this);
1379 EmitStmt(S: S.getBody());
1380 }
1381
1382 // The last block in the loop's body (which unconditionally branches to the
1383 // `inc` block if there is one).
1384 auto *FinalBodyBB = Builder.GetInsertBlock();
1385
1386 // If there is an increment, emit it next.
1387 if (S.getInc()) {
1388 EmitBlock(BB: Continue.getBlock());
1389 EmitStmt(S: S.getInc());
1390 }
1391
1392 BreakContinueStack.pop_back();
1393
1394 ConditionScope.ForceCleanup();
1395
1396 EmitStopPoint(S: &S);
1397 EmitBranch(Target: CondBlock);
1398
1399 if (ForScope)
1400 ForScope->ForceCleanup();
1401
1402 LoopStack.pop();
1403
1404 // Emit the fall-through block.
1405 EmitBlock(BB: LoopExit.getBlock(), IsFinished: true);
1406
1407 if (CGM.shouldEmitConvergenceTokens())
1408 ConvergenceTokenStack.pop_back();
1409
1410 if (FinalBodyBB) {
1411 // Key Instructions: We want the for closing brace to be step-able on to
1412 // match existing behaviour.
1413 addInstToNewSourceAtom(KeyInstruction: FinalBodyBB->getTerminator(), Backup: nullptr);
1414 }
1415}
1416
1417void
1418CodeGenFunction::EmitCXXForRangeStmt(const CXXForRangeStmt &S,
1419 ArrayRef<const Attr *> ForAttrs) {
1420 JumpDest LoopExit = getJumpDestInCurrentScope(Name: "for.end");
1421
1422 LexicalScope ForScope(*this, S.getSourceRange());
1423
1424 // Evaluate the first pieces before the loop.
1425 if (S.getInit())
1426 EmitStmt(S: S.getInit());
1427 EmitStmt(S: S.getRangeStmt());
1428 EmitStmt(S: S.getBeginStmt());
1429 EmitStmt(S: S.getEndStmt());
1430
1431 // Start the loop with a block that tests the condition.
1432 // If there's an increment, the continue scope will be overwritten
1433 // later.
1434 llvm::BasicBlock *CondBlock = createBasicBlock(name: "for.cond");
1435 EmitBlock(BB: CondBlock);
1436
1437 if (CGM.shouldEmitConvergenceTokens())
1438 ConvergenceTokenStack.push_back(Elt: emitConvergenceLoopToken(BB: CondBlock));
1439
1440 const SourceRange &R = S.getSourceRange();
1441 LoopStack.push(Header: CondBlock, Ctx&: CGM.getContext(), CGOpts: CGM.getCodeGenOpts(), Attrs: ForAttrs,
1442 StartLoc: SourceLocToDebugLoc(Location: R.getBegin()),
1443 EndLoc: SourceLocToDebugLoc(Location: R.getEnd()));
1444
1445 // If there are any cleanups between here and the loop-exit scope,
1446 // create a block to stage a loop exit along.
1447 llvm::BasicBlock *ExitBlock = LoopExit.getBlock();
1448 if (hasSkipCounter(S: &S) || ForScope.requiresCleanups())
1449 ExitBlock = createBasicBlock(name: "for.cond.cleanup");
1450
1451 // The loop body, consisting of the specified body and the loop variable.
1452 llvm::BasicBlock *ForBody = createBasicBlock(name: "for.body");
1453
1454 // The body is executed if the expression, contextually converted
1455 // to bool, is true.
1456 llvm::Value *BoolCondVal = EvaluateExprAsBool(E: S.getCond());
1457 llvm::MDNode *Weights =
1458 createProfileWeightsForLoop(Cond: S.getCond(), LoopCount: getProfileCount(S: S.getBody()));
1459 if (!Weights && CGM.getCodeGenOpts().OptimizationLevel)
1460 BoolCondVal = emitCondLikelihoodViaExpectIntrinsic(
1461 Cond: BoolCondVal, LH: Stmt::getLikelihood(S: S.getBody()));
1462 auto *I = Builder.CreateCondBr(Cond: BoolCondVal, True: ForBody, False: ExitBlock, BranchWeights: Weights);
1463 // Key Instructions: Emit the condition and branch as separate atoms to
1464 // match existing loop stepping behaviour. FIXME: We could have the branch as
1465 // the backup location for the condition, which would probably be a better
1466 // experience.
1467 if (auto *CondI = dyn_cast<llvm::Instruction>(Val: BoolCondVal))
1468 addInstToNewSourceAtom(KeyInstruction: CondI, Backup: nullptr);
1469 addInstToNewSourceAtom(KeyInstruction: I, Backup: nullptr);
1470
1471 if (ExitBlock != LoopExit.getBlock()) {
1472 EmitBlock(BB: ExitBlock);
1473 incrementProfileCounter(ExecSkip: UseSkipPath, S: &S);
1474 EmitBranchThroughCleanup(Dest: LoopExit);
1475 }
1476
1477 EmitBlock(BB: ForBody);
1478 incrementProfileCounter(ExecSkip: UseExecPath, S: &S);
1479
1480 // Create a block for the increment. In case of a 'continue', we jump there.
1481 JumpDest Continue = getJumpDestInCurrentScope(Name: "for.inc");
1482
1483 // Store the blocks to use for break and continue.
1484 BreakContinueStack.push_back(Elt: BreakContinue(S, LoopExit, Continue));
1485
1486 {
1487 // Create a separate cleanup scope for the loop variable and body.
1488 LexicalScope BodyScope(*this, S.getSourceRange());
1489 EmitStmt(S: S.getLoopVarStmt());
1490 EmitStmt(S: S.getBody());
1491 }
1492 // The last block in the loop's body (which unconditionally branches to the
1493 // `inc` block if there is one).
1494 auto *FinalBodyBB = Builder.GetInsertBlock();
1495
1496 EmitStopPoint(S: &S);
1497 // If there is an increment, emit it next.
1498 EmitBlock(BB: Continue.getBlock());
1499 EmitStmt(S: S.getInc());
1500
1501 BreakContinueStack.pop_back();
1502
1503 EmitBranch(Target: CondBlock);
1504
1505 ForScope.ForceCleanup();
1506
1507 LoopStack.pop();
1508
1509 // Emit the fall-through block.
1510 EmitBlock(BB: LoopExit.getBlock(), IsFinished: true);
1511
1512 if (CGM.shouldEmitConvergenceTokens())
1513 ConvergenceTokenStack.pop_back();
1514
1515 if (FinalBodyBB) {
1516 // We want the for closing brace to be step-able on to match existing
1517 // behaviour.
1518 addInstToNewSourceAtom(KeyInstruction: FinalBodyBB->getTerminator(), Backup: nullptr);
1519 }
1520}
1521
1522void CodeGenFunction::EmitReturnOfRValue(RValue RV, QualType Ty) {
1523 if (RV.isScalar()) {
1524 Builder.CreateStore(Val: RV.getScalarVal(), Addr: ReturnValue);
1525 } else if (RV.isAggregate()) {
1526 LValue Dest = MakeAddrLValue(Addr: ReturnValue, T: Ty);
1527 LValue Src = MakeAddrLValue(Addr: RV.getAggregateAddress(), T: Ty);
1528 EmitAggregateCopy(Dest, Src, EltTy: Ty, MayOverlap: getOverlapForReturnValue());
1529 } else {
1530 EmitStoreOfComplex(V: RV.getComplexVal(), dest: MakeAddrLValue(Addr: ReturnValue, T: Ty),
1531 /*init*/ isInit: true);
1532 }
1533 EmitBranchThroughCleanup(Dest: ReturnBlock);
1534}
1535
1536namespace {
1537// RAII struct used to save and restore a return statment's result expression.
1538struct SaveRetExprRAII {
1539 SaveRetExprRAII(const Expr *RetExpr, CodeGenFunction &CGF)
1540 : OldRetExpr(CGF.RetExpr), CGF(CGF) {
1541 CGF.RetExpr = RetExpr;
1542 }
1543 ~SaveRetExprRAII() { CGF.RetExpr = OldRetExpr; }
1544 const Expr *OldRetExpr;
1545 CodeGenFunction &CGF;
1546};
1547} // namespace
1548
1549/// Determine if the given call uses the swiftasync calling convention.
1550static bool isSwiftAsyncCallee(const CallExpr *CE) {
1551 auto calleeQualType = CE->getCallee()->getType();
1552 const FunctionType *calleeType = nullptr;
1553 if (calleeQualType->isFunctionPointerType() ||
1554 calleeQualType->isFunctionReferenceType() ||
1555 calleeQualType->isBlockPointerType() ||
1556 calleeQualType->isMemberFunctionPointerType()) {
1557 calleeType = calleeQualType->getPointeeType()->castAs<FunctionType>();
1558 } else if (auto *ty = dyn_cast<FunctionType>(Val&: calleeQualType)) {
1559 calleeType = ty;
1560 } else if (auto CMCE = dyn_cast<CXXMemberCallExpr>(Val: CE)) {
1561 if (auto methodDecl = CMCE->getMethodDecl()) {
1562 // getMethodDecl() doesn't handle member pointers at the moment.
1563 calleeType = methodDecl->getType()->castAs<FunctionType>();
1564 } else {
1565 return false;
1566 }
1567 } else {
1568 return false;
1569 }
1570 return calleeType->getCallConv() == CallingConv::CC_SwiftAsync;
1571}
1572
1573/// EmitReturnStmt - Note that due to GCC extensions, this can have an operand
1574/// if the function returns void, or may be missing one if the function returns
1575/// non-void. Fun stuff :).
1576void CodeGenFunction::EmitReturnStmt(const ReturnStmt &S) {
1577 ApplyAtomGroup Grp(getDebugInfo());
1578 if (requiresReturnValueCheck()) {
1579 llvm::Constant *SLoc = EmitCheckSourceLocation(Loc: S.getBeginLoc());
1580 auto *SLocPtr =
1581 new llvm::GlobalVariable(CGM.getModule(), SLoc->getType(), false,
1582 llvm::GlobalVariable::PrivateLinkage, SLoc);
1583 SLocPtr->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
1584 CGM.getSanitizerMetadata()->disableSanitizerForGlobal(GV: SLocPtr);
1585 assert(ReturnLocation.isValid() && "No valid return location");
1586 Builder.CreateStore(Val: SLocPtr, Addr: ReturnLocation);
1587 }
1588
1589 // Returning from an outlined SEH helper is UB, and we already warn on it.
1590 if (IsOutlinedSEHHelper) {
1591 Builder.CreateUnreachable();
1592 Builder.ClearInsertionPoint();
1593 }
1594
1595 // Emit the result value, even if unused, to evaluate the side effects.
1596 const Expr *RV = S.getRetValue();
1597
1598 // Record the result expression of the return statement. The recorded
1599 // expression is used to determine whether a block capture's lifetime should
1600 // end at the end of the full expression as opposed to the end of the scope
1601 // enclosing the block expression.
1602 //
1603 // This permits a small, easily-implemented exception to our over-conservative
1604 // rules about not jumping to statements following block literals with
1605 // non-trivial cleanups.
1606 SaveRetExprRAII SaveRetExpr(RV, *this);
1607
1608 RunCleanupsScope cleanupScope(*this);
1609 if (const auto *EWC = dyn_cast_or_null<ExprWithCleanups>(Val: RV))
1610 RV = EWC->getSubExpr();
1611
1612 // If we're in a swiftasynccall function, and the return expression is a
1613 // call to a swiftasynccall function, mark the call as the musttail call.
1614 std::optional<llvm::SaveAndRestore<const CallExpr *>> SaveMustTail;
1615 if (RV && CurFnInfo &&
1616 CurFnInfo->getASTCallingConvention() == CallingConv::CC_SwiftAsync) {
1617 if (auto CE = dyn_cast<CallExpr>(Val: RV)) {
1618 if (isSwiftAsyncCallee(CE)) {
1619 SaveMustTail.emplace(args&: MustTailCall, args&: CE);
1620 }
1621 }
1622 }
1623
1624 // FIXME: Clean this up by using an LValue for ReturnTemp,
1625 // EmitStoreThroughLValue, and EmitAnyExpr.
1626 // Check if the NRVO candidate was not globalized in OpenMP mode.
1627 if (getLangOpts().ElideConstructors && S.getNRVOCandidate() &&
1628 S.getNRVOCandidate()->isNRVOVariable() &&
1629 (!getLangOpts().OpenMP ||
1630 !CGM.getOpenMPRuntime()
1631 .getAddressOfLocalVariable(CGF&: *this, VD: S.getNRVOCandidate())
1632 .isValid())) {
1633 // Apply the named return value optimization for this return statement,
1634 // which means doing nothing: the appropriate result has already been
1635 // constructed into the NRVO variable.
1636
1637 // If there is an NRVO flag for this variable, set it to 1 into indicate
1638 // that the cleanup code should not destroy the variable.
1639 if (llvm::Value *NRVOFlag = NRVOFlags[S.getNRVOCandidate()])
1640 Builder.CreateFlagStore(Value: Builder.getTrue(), Addr: NRVOFlag);
1641 } else if (!ReturnValue.isValid() || (RV && RV->getType()->isVoidType())) {
1642 // Make sure not to return anything, but evaluate the expression
1643 // for side effects.
1644 if (RV) {
1645 EmitAnyExpr(E: RV);
1646 }
1647 } else if (!RV) {
1648 // Do nothing (return value is left uninitialized)
1649 } else if (FnRetTy->isReferenceType()) {
1650 // If this function returns a reference, take the address of the expression
1651 // rather than the value.
1652 RValue Result = EmitReferenceBindingToExpr(E: RV);
1653 auto *I = Builder.CreateStore(Val: Result.getScalarVal(), Addr: ReturnValue);
1654 addInstToCurrentSourceAtom(KeyInstruction: I, Backup: I->getValueOperand());
1655 } else {
1656 switch (getEvaluationKind(T: RV->getType())) {
1657 case TEK_Scalar: {
1658 llvm::Value *Ret = EmitScalarExpr(E: RV);
1659 if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
1660 EmitStoreOfScalar(value: Ret, lvalue: MakeAddrLValue(Addr: ReturnValue, T: RV->getType()),
1661 /*isInit*/ true);
1662 } else {
1663 auto *I = Builder.CreateStore(Val: Ret, Addr: ReturnValue);
1664 addInstToCurrentSourceAtom(KeyInstruction: I, Backup: I->getValueOperand());
1665 }
1666 break;
1667 }
1668 case TEK_Complex:
1669 EmitComplexExprIntoLValue(E: RV, dest: MakeAddrLValue(Addr: ReturnValue, T: RV->getType()),
1670 /*isInit*/ true);
1671 break;
1672 case TEK_Aggregate:
1673 EmitAggExpr(E: RV, AS: AggValueSlot::forAddr(
1674 addr: ReturnValue, quals: Qualifiers(),
1675 isDestructed: AggValueSlot::IsDestructed,
1676 needsGC: AggValueSlot::DoesNotNeedGCBarriers,
1677 isAliased: AggValueSlot::IsNotAliased,
1678 mayOverlap: getOverlapForReturnValue()));
1679 break;
1680 }
1681 }
1682
1683 ++NumReturnExprs;
1684 if (!RV || RV->isEvaluatable(Ctx: getContext()))
1685 ++NumSimpleReturnExprs;
1686
1687 cleanupScope.ForceCleanup();
1688 EmitBranchThroughCleanup(Dest: ReturnBlock);
1689}
1690
1691void CodeGenFunction::EmitDeclStmt(const DeclStmt &S) {
1692 // As long as debug info is modeled with instructions, we have to ensure we
1693 // have a place to insert here and write the stop point here.
1694 if (HaveInsertPoint())
1695 EmitStopPoint(S: &S);
1696
1697 for (const auto *I : S.decls())
1698 EmitDecl(D: *I, /*EvaluateConditionDecl=*/true);
1699}
1700
1701auto CodeGenFunction::GetDestForLoopControlStmt(const LoopControlStmt &S)
1702 -> const BreakContinue * {
1703 if (!S.hasLabelTarget())
1704 return &BreakContinueStack.back();
1705
1706 const Stmt *LoopOrSwitch = S.getNamedLoopOrSwitch();
1707 assert(LoopOrSwitch && "break/continue target not set?");
1708 for (const BreakContinue &BC : llvm::reverse(C&: BreakContinueStack))
1709 if (BC.LoopOrSwitch == LoopOrSwitch)
1710 return &BC;
1711
1712 llvm_unreachable("break/continue target not found");
1713}
1714
1715void CodeGenFunction::EmitBreakStmt(const BreakStmt &S) {
1716 assert(!BreakContinueStack.empty() && "break stmt not in a loop or switch!");
1717
1718 // If this code is reachable then emit a stop point (if generating
1719 // debug info). We have to do this ourselves because we are on the
1720 // "simple" statement path.
1721 if (HaveInsertPoint())
1722 EmitStopPoint(S: &S);
1723
1724 ApplyAtomGroup Grp(getDebugInfo());
1725 EmitBranchThroughCleanup(Dest: GetDestForLoopControlStmt(S)->BreakBlock);
1726}
1727
1728void CodeGenFunction::EmitContinueStmt(const ContinueStmt &S) {
1729 assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
1730
1731 // If this code is reachable then emit a stop point (if generating
1732 // debug info). We have to do this ourselves because we are on the
1733 // "simple" statement path.
1734 if (HaveInsertPoint())
1735 EmitStopPoint(S: &S);
1736
1737 ApplyAtomGroup Grp(getDebugInfo());
1738 EmitBranchThroughCleanup(Dest: GetDestForLoopControlStmt(S)->ContinueBlock);
1739}
1740
1741/// EmitCaseStmtRange - If case statement range is not too big then
1742/// add multiple cases to switch instruction, one for each value within
1743/// the range. If range is too big then emit "if" condition check.
1744void CodeGenFunction::EmitCaseStmtRange(const CaseStmt &S,
1745 ArrayRef<const Attr *> Attrs) {
1746 assert(S.getRHS() && "Expected RHS value in CaseStmt");
1747
1748 llvm::APSInt LHS = S.getLHS()->EvaluateKnownConstInt(Ctx: getContext());
1749 llvm::APSInt RHS = S.getRHS()->EvaluateKnownConstInt(Ctx: getContext());
1750
1751 // Emit the code for this case. We do this first to make sure it is
1752 // properly chained from our predecessor before generating the
1753 // switch machinery to enter this block.
1754 llvm::BasicBlock *CaseDest = createBasicBlock(name: "sw.bb");
1755 EmitBlockWithFallThrough(BB: CaseDest, S: &S);
1756 EmitStmt(S: S.getSubStmt());
1757
1758 // If range is empty, do nothing.
1759 if (LHS.isSigned() ? RHS.slt(RHS: LHS) : RHS.ult(RHS: LHS))
1760 return;
1761
1762 Stmt::Likelihood LH = Stmt::getLikelihood(Attrs);
1763 llvm::APInt Range = RHS - LHS;
1764 // FIXME: parameters such as this should not be hardcoded.
1765 if (Range.ult(RHS: llvm::APInt(Range.getBitWidth(), 64))) {
1766 // Range is small enough to add multiple switch instruction cases.
1767 uint64_t Total = getProfileCount(S: &S);
1768 unsigned NCases = Range.getZExtValue() + 1;
1769 // We only have one region counter for the entire set of cases here, so we
1770 // need to divide the weights evenly between the generated cases, ensuring
1771 // that the total weight is preserved. E.g., a weight of 5 over three cases
1772 // will be distributed as weights of 2, 2, and 1.
1773 uint64_t Weight = Total / NCases, Rem = Total % NCases;
1774 for (unsigned I = 0; I != NCases; ++I) {
1775 if (SwitchWeights)
1776 SwitchWeights->push_back(Elt: Weight + (Rem ? 1 : 0));
1777 else if (SwitchLikelihood)
1778 SwitchLikelihood->push_back(Elt: LH);
1779
1780 if (Rem)
1781 Rem--;
1782 SwitchInsn->addCase(OnVal: Builder.getInt(AI: LHS), Dest: CaseDest);
1783 ++LHS;
1784 }
1785 return;
1786 }
1787
1788 // The range is too big. Emit "if" condition into a new block,
1789 // making sure to save and restore the current insertion point.
1790 llvm::BasicBlock *RestoreBB = Builder.GetInsertBlock();
1791
1792 // Push this test onto the chain of range checks (which terminates
1793 // in the default basic block). The switch's default will be changed
1794 // to the top of this chain after switch emission is complete.
1795 llvm::BasicBlock *FalseDest = CaseRangeBlock;
1796 CaseRangeBlock = createBasicBlock(name: "sw.caserange");
1797
1798 CurFn->insert(Position: CurFn->end(), BB: CaseRangeBlock);
1799 Builder.SetInsertPoint(CaseRangeBlock);
1800
1801 // Emit range check.
1802 llvm::Value *Diff =
1803 Builder.CreateSub(LHS: SwitchInsn->getCondition(), RHS: Builder.getInt(AI: LHS));
1804 llvm::Value *Cond =
1805 Builder.CreateICmpULE(LHS: Diff, RHS: Builder.getInt(AI: Range), Name: "inbounds");
1806
1807 llvm::MDNode *Weights = nullptr;
1808 if (SwitchWeights) {
1809 uint64_t ThisCount = getProfileCount(S: &S);
1810 uint64_t DefaultCount = (*SwitchWeights)[0];
1811 Weights = createProfileWeights(TrueCount: ThisCount, FalseCount: DefaultCount);
1812
1813 // Since we're chaining the switch default through each large case range, we
1814 // need to update the weight for the default, ie, the first case, to include
1815 // this case.
1816 (*SwitchWeights)[0] += ThisCount;
1817 } else if (SwitchLikelihood)
1818 Cond = emitCondLikelihoodViaExpectIntrinsic(Cond, LH);
1819
1820 Builder.CreateCondBr(Cond, True: CaseDest, False: FalseDest, BranchWeights: Weights);
1821
1822 // Restore the appropriate insertion point.
1823 if (RestoreBB)
1824 Builder.SetInsertPoint(RestoreBB);
1825 else
1826 Builder.ClearInsertionPoint();
1827}
1828
1829void CodeGenFunction::EmitCaseStmt(const CaseStmt &S,
1830 ArrayRef<const Attr *> Attrs) {
1831 // If there is no enclosing switch instance that we're aware of, then this
1832 // case statement and its block can be elided. This situation only happens
1833 // when we've constant-folded the switch, are emitting the constant case,
1834 // and part of the constant case includes another case statement. For
1835 // instance: switch (4) { case 4: do { case 5: } while (1); }
1836 if (!SwitchInsn) {
1837 EmitStmt(S: S.getSubStmt());
1838 return;
1839 }
1840
1841 // Handle case ranges.
1842 if (S.getRHS()) {
1843 EmitCaseStmtRange(S, Attrs);
1844 return;
1845 }
1846
1847 llvm::ConstantInt *CaseVal =
1848 Builder.getInt(AI: S.getLHS()->EvaluateKnownConstInt(Ctx: getContext()));
1849
1850 // Emit debuginfo for the case value if it is an enum value.
1851 const ConstantExpr *CE;
1852 if (auto ICE = dyn_cast<ImplicitCastExpr>(Val: S.getLHS()))
1853 CE = dyn_cast<ConstantExpr>(Val: ICE->getSubExpr());
1854 else
1855 CE = dyn_cast<ConstantExpr>(Val: S.getLHS());
1856 if (CE) {
1857 if (auto DE = dyn_cast<DeclRefExpr>(Val: CE->getSubExpr()))
1858 if (CGDebugInfo *Dbg = getDebugInfo())
1859 if (CGM.getCodeGenOpts().hasReducedDebugInfo())
1860 Dbg->EmitGlobalVariable(VD: DE->getDecl(),
1861 Init: APValue(llvm::APSInt(CaseVal->getValue())));
1862 }
1863
1864 if (SwitchLikelihood)
1865 SwitchLikelihood->push_back(Elt: Stmt::getLikelihood(Attrs));
1866
1867 // If the body of the case is just a 'break', try to not emit an empty block.
1868 // If we're profiling or we're not optimizing, leave the block in for better
1869 // debug and coverage analysis.
1870 if (!CGM.getCodeGenOpts().hasProfileClangInstr() &&
1871 CGM.getCodeGenOpts().OptimizationLevel > 0 &&
1872 isa<BreakStmt>(Val: S.getSubStmt())) {
1873 JumpDest Block = BreakContinueStack.back().BreakBlock;
1874
1875 // Only do this optimization if there are no cleanups that need emitting.
1876 if (isObviouslyBranchWithoutCleanups(Dest: Block)) {
1877 if (SwitchWeights)
1878 SwitchWeights->push_back(Elt: getProfileCount(S: &S));
1879 SwitchInsn->addCase(OnVal: CaseVal, Dest: Block.getBlock());
1880
1881 // If there was a fallthrough into this case, make sure to redirect it to
1882 // the end of the switch as well.
1883 if (Builder.GetInsertBlock()) {
1884 Builder.CreateBr(Dest: Block.getBlock());
1885 Builder.ClearInsertionPoint();
1886 }
1887 return;
1888 }
1889 }
1890
1891 llvm::BasicBlock *CaseDest = createBasicBlock(name: "sw.bb");
1892 EmitBlockWithFallThrough(BB: CaseDest, S: &S);
1893 if (SwitchWeights)
1894 SwitchWeights->push_back(Elt: getProfileCount(S: &S));
1895 SwitchInsn->addCase(OnVal: CaseVal, Dest: CaseDest);
1896
1897 // Recursively emitting the statement is acceptable, but is not wonderful for
1898 // code where we have many case statements nested together, i.e.:
1899 // case 1:
1900 // case 2:
1901 // case 3: etc.
1902 // Handling this recursively will create a new block for each case statement
1903 // that falls through to the next case which is IR intensive. It also causes
1904 // deep recursion which can run into stack depth limitations. Handle
1905 // sequential non-range case statements specially.
1906 //
1907 // TODO When the next case has a likelihood attribute the code returns to the
1908 // recursive algorithm. Maybe improve this case if it becomes common practice
1909 // to use a lot of attributes.
1910 const CaseStmt *CurCase = &S;
1911 const CaseStmt *NextCase = dyn_cast<CaseStmt>(Val: S.getSubStmt());
1912
1913 // Otherwise, iteratively add consecutive cases to this switch stmt.
1914 while (NextCase && NextCase->getRHS() == nullptr) {
1915 CurCase = NextCase;
1916 llvm::ConstantInt *CaseVal =
1917 Builder.getInt(AI: CurCase->getLHS()->EvaluateKnownConstInt(Ctx: getContext()));
1918
1919 if (SwitchWeights)
1920 SwitchWeights->push_back(Elt: getProfileCount(S: NextCase));
1921 if (CGM.getCodeGenOpts().hasProfileClangInstr()) {
1922 CaseDest = createBasicBlock(name: "sw.bb");
1923 EmitBlockWithFallThrough(BB: CaseDest, S: CurCase);
1924 }
1925 // Since this loop is only executed when the CaseStmt has no attributes
1926 // use a hard-coded value.
1927 if (SwitchLikelihood)
1928 SwitchLikelihood->push_back(Elt: Stmt::LH_None);
1929
1930 SwitchInsn->addCase(OnVal: CaseVal, Dest: CaseDest);
1931 NextCase = dyn_cast<CaseStmt>(Val: CurCase->getSubStmt());
1932 }
1933
1934 // Generate a stop point for debug info if the case statement is
1935 // followed by a default statement. A fallthrough case before a
1936 // default case gets its own branch target.
1937 if (CurCase->getSubStmt()->getStmtClass() == Stmt::DefaultStmtClass)
1938 EmitStopPoint(S: CurCase);
1939
1940 // Normal default recursion for non-cases.
1941 EmitStmt(S: CurCase->getSubStmt());
1942}
1943
1944void CodeGenFunction::EmitDefaultStmt(const DefaultStmt &S,
1945 ArrayRef<const Attr *> Attrs) {
1946 // If there is no enclosing switch instance that we're aware of, then this
1947 // default statement can be elided. This situation only happens when we've
1948 // constant-folded the switch.
1949 if (!SwitchInsn) {
1950 EmitStmt(S: S.getSubStmt());
1951 return;
1952 }
1953
1954 llvm::BasicBlock *DefaultBlock = SwitchInsn->getDefaultDest();
1955 assert(DefaultBlock->empty() &&
1956 "EmitDefaultStmt: Default block already defined?");
1957
1958 if (SwitchLikelihood)
1959 SwitchLikelihood->front() = Stmt::getLikelihood(Attrs);
1960
1961 EmitBlockWithFallThrough(BB: DefaultBlock, S: &S);
1962
1963 EmitStmt(S: S.getSubStmt());
1964}
1965
1966namespace {
1967struct EmitDeferredStatement final : EHScopeStack::Cleanup {
1968 const DeferStmt &Stmt;
1969 EmitDeferredStatement(const DeferStmt *Stmt) : Stmt(*Stmt) {}
1970
1971 void Emit(CodeGenFunction &CGF, Flags) override {
1972 // Take care that any cleanups pushed by the body of a '_Defer' statement
1973 // don't clobber the current cleanup slot value.
1974 //
1975 // Assume we have a scope that pushes a cleanup; when that scope is exited,
1976 // we need to run that cleanup; this is accomplished by emitting the cleanup
1977 // into a separate block and then branching to that block at scope exit.
1978 //
1979 // Where this gets complicated is if we exit the scope in multiple different
1980 // ways; e.g. in a 'for' loop, we may exit the scope of its body by falling
1981 // off the end (in which case we need to run the cleanup and then branch to
1982 // the increment), or by 'break'ing out of the loop (in which case we need
1983 // to run the cleanup and then branch to the loop exit block); in both cases
1984 // we first branch to the cleanup block to run the cleanup, but the block we
1985 // need to jump to *after* running the cleanup is different.
1986 //
1987 // This is accomplished using a local integer variable called the 'cleanup
1988 // slot': before branching to the cleanup block, we store a value into that
1989 // slot. Then, in the cleanup block, after running the cleanup, we load the
1990 // value of that variable and 'switch' on it to branch to the appropriate
1991 // continuation block.
1992 //
1993 // The problem that arises once '_Defer' statements are involved is that the
1994 // body of a '_Defer' is an arbitrary statement which itself can create more
1995 // cleanups. This means we may end up overwriting the cleanup slot before we
1996 // ever have a chance to 'switch' on it, which means that once we *do* get
1997 // to the 'switch', we end up in whatever block the cleanup code happened to
1998 // pick as the default 'switch' exit label!
1999 //
2000 // That is, what is normally supposed to happen is something like:
2001 //
2002 // 1. Store 'X' to cleanup slot.
2003 // 2. Branch to cleanup block.
2004 // 3. Execute cleanup.
2005 // 4. Read value from cleanup slot.
2006 // 5. Branch to the block associated with 'X'.
2007 //
2008 // But if we encounter a _Defer' statement that contains a cleanup, then
2009 // what might instead happen is:
2010 //
2011 // 1. Store 'X' to cleanup slot.
2012 // 2. Branch to cleanup block.
2013 // 3. Execute cleanup; this ends up pushing another cleanup, so:
2014 // 3a. Store 'Y' to cleanup slot.
2015 // 3b. Run steps 2–5 recursively.
2016 // 4. Read value from cleanup slot, which is now 'Y' instead of 'X'.
2017 // 5. Branch to the block associated with 'Y'... which doesn't even
2018 // exist because the value 'Y' is only meaningful for the inner
2019 // cleanup. The result is we just branch 'somewhere random'.
2020 //
2021 // The rest of the cleanup code simply isn't prepared to handle this case
2022 // because most other cleanups can't push more cleanups, and thus, emitting
2023 // other cleanups generally cannot clobber the cleanup slot.
2024 //
2025 // To prevent this from happening, save the current cleanup slot value and
2026 // restore it after emitting the '_Defer' statement.
2027 llvm::Value *SavedCleanupDest = nullptr;
2028 if (CGF.NormalCleanupDest.isValid())
2029 SavedCleanupDest =
2030 CGF.Builder.CreateLoad(Addr: CGF.NormalCleanupDest, Name: "cleanup.dest.saved");
2031
2032 CGF.EmitStmt(S: Stmt.getBody());
2033
2034 if (SavedCleanupDest && CGF.HaveInsertPoint())
2035 CGF.Builder.CreateStore(Val: SavedCleanupDest, Addr: CGF.NormalCleanupDest);
2036
2037 // Cleanups must end with an insert point.
2038 CGF.EnsureInsertPoint();
2039 }
2040};
2041} // namespace
2042
2043void CodeGenFunction::EmitDeferStmt(const DeferStmt &S) {
2044 EHStack.pushCleanup<EmitDeferredStatement>(Kind: NormalAndEHCleanup, A: &S);
2045}
2046
2047/// CollectStatementsForCase - Given the body of a 'switch' statement and a
2048/// constant value that is being switched on, see if we can dead code eliminate
2049/// the body of the switch to a simple series of statements to emit. Basically,
2050/// on a switch (5) we want to find these statements:
2051/// case 5:
2052/// printf(...); <--
2053/// ++i; <--
2054/// break;
2055///
2056/// and add them to the ResultStmts vector. If it is unsafe to do this
2057/// transformation (for example, one of the elided statements contains a label
2058/// that might be jumped to), return CSFC_Failure. If we handled it and 'S'
2059/// should include statements after it (e.g. the printf() line is a substmt of
2060/// the case) then return CSFC_FallThrough. If we handled it and found a break
2061/// statement, then return CSFC_Success.
2062///
2063/// If Case is non-null, then we are looking for the specified case, checking
2064/// that nothing we jump over contains labels. If Case is null, then we found
2065/// the case and are looking for the break.
2066///
2067/// If the recursive walk actually finds our Case, then we set FoundCase to
2068/// true.
2069///
2070enum CSFC_Result { CSFC_Failure, CSFC_FallThrough, CSFC_Success };
2071static CSFC_Result CollectStatementsForCase(const Stmt *S,
2072 const SwitchCase *Case,
2073 bool &FoundCase,
2074 SmallVectorImpl<const Stmt*> &ResultStmts) {
2075 // If this is a null statement, just succeed.
2076 if (!S)
2077 return Case ? CSFC_Success : CSFC_FallThrough;
2078
2079 // If this is the switchcase (case 4: or default) that we're looking for, then
2080 // we're in business. Just add the substatement.
2081 if (const SwitchCase *SC = dyn_cast<SwitchCase>(Val: S)) {
2082 if (S == Case) {
2083 FoundCase = true;
2084 return CollectStatementsForCase(S: SC->getSubStmt(), Case: nullptr, FoundCase,
2085 ResultStmts);
2086 }
2087
2088 // Otherwise, this is some other case or default statement, just ignore it.
2089 return CollectStatementsForCase(S: SC->getSubStmt(), Case, FoundCase,
2090 ResultStmts);
2091 }
2092
2093 // If we are in the live part of the code and we found our break statement,
2094 // return a success!
2095 if (!Case && isa<BreakStmt>(Val: S))
2096 return CSFC_Success;
2097
2098 // If this is a switch statement, then it might contain the SwitchCase, the
2099 // break, or neither.
2100 if (const CompoundStmt *CS = dyn_cast<CompoundStmt>(Val: S)) {
2101 // Handle this as two cases: we might be looking for the SwitchCase (if so
2102 // the skipped statements must be skippable) or we might already have it.
2103 CompoundStmt::const_body_iterator I = CS->body_begin(), E = CS->body_end();
2104 bool StartedInLiveCode = FoundCase;
2105 unsigned StartSize = ResultStmts.size();
2106
2107 // If we've not found the case yet, scan through looking for it.
2108 if (Case) {
2109 // Keep track of whether we see a skipped declaration. The code could be
2110 // using the declaration even if it is skipped, so we can't optimize out
2111 // the decl if the kept statements might refer to it.
2112 bool HadSkippedDecl = false;
2113
2114 // If we're looking for the case, just see if we can skip each of the
2115 // substatements.
2116 for (; Case && I != E; ++I) {
2117 HadSkippedDecl |= CodeGenFunction::mightAddDeclToScope(S: *I);
2118
2119 switch (CollectStatementsForCase(S: *I, Case, FoundCase, ResultStmts)) {
2120 case CSFC_Failure: return CSFC_Failure;
2121 case CSFC_Success:
2122 // A successful result means that either 1) that the statement doesn't
2123 // have the case and is skippable, or 2) does contain the case value
2124 // and also contains the break to exit the switch. In the later case,
2125 // we just verify the rest of the statements are elidable.
2126 if (FoundCase) {
2127 // If we found the case and skipped declarations, we can't do the
2128 // optimization.
2129 if (HadSkippedDecl)
2130 return CSFC_Failure;
2131
2132 for (++I; I != E; ++I)
2133 if (CodeGenFunction::ContainsLabel(S: *I, IgnoreCaseStmts: true))
2134 return CSFC_Failure;
2135 return CSFC_Success;
2136 }
2137 break;
2138 case CSFC_FallThrough:
2139 // If we have a fallthrough condition, then we must have found the
2140 // case started to include statements. Consider the rest of the
2141 // statements in the compound statement as candidates for inclusion.
2142 assert(FoundCase && "Didn't find case but returned fallthrough?");
2143 // We recursively found Case, so we're not looking for it anymore.
2144 Case = nullptr;
2145
2146 // If we found the case and skipped declarations, we can't do the
2147 // optimization.
2148 if (HadSkippedDecl)
2149 return CSFC_Failure;
2150 break;
2151 }
2152 }
2153
2154 if (!FoundCase)
2155 return CSFC_Success;
2156
2157 assert(!HadSkippedDecl && "fallthrough after skipping decl");
2158 }
2159
2160 // If we have statements in our range, then we know that the statements are
2161 // live and need to be added to the set of statements we're tracking.
2162 bool AnyDecls = false;
2163 for (; I != E; ++I) {
2164 AnyDecls |= CodeGenFunction::mightAddDeclToScope(S: *I);
2165
2166 switch (CollectStatementsForCase(S: *I, Case: nullptr, FoundCase, ResultStmts)) {
2167 case CSFC_Failure: return CSFC_Failure;
2168 case CSFC_FallThrough:
2169 // A fallthrough result means that the statement was simple and just
2170 // included in ResultStmt, keep adding them afterwards.
2171 break;
2172 case CSFC_Success:
2173 // A successful result means that we found the break statement and
2174 // stopped statement inclusion. We just ensure that any leftover stmts
2175 // are skippable and return success ourselves.
2176 for (++I; I != E; ++I)
2177 if (CodeGenFunction::ContainsLabel(S: *I, IgnoreCaseStmts: true))
2178 return CSFC_Failure;
2179 return CSFC_Success;
2180 }
2181 }
2182
2183 // If we're about to fall out of a scope without hitting a 'break;', we
2184 // can't perform the optimization if there were any decls in that scope
2185 // (we'd lose their end-of-lifetime).
2186 if (AnyDecls) {
2187 // If the entire compound statement was live, there's one more thing we
2188 // can try before giving up: emit the whole thing as a single statement.
2189 // We can do that unless the statement contains a 'break;'.
2190 // FIXME: Such a break must be at the end of a construct within this one.
2191 // We could emit this by just ignoring the BreakStmts entirely.
2192 if (StartedInLiveCode && !CodeGenFunction::containsBreak(S)) {
2193 ResultStmts.resize(N: StartSize);
2194 ResultStmts.push_back(Elt: S);
2195 } else {
2196 return CSFC_Failure;
2197 }
2198 }
2199
2200 return CSFC_FallThrough;
2201 }
2202
2203 // Okay, this is some other statement that we don't handle explicitly, like a
2204 // for statement or increment etc. If we are skipping over this statement,
2205 // just verify it doesn't have labels, which would make it invalid to elide.
2206 if (Case) {
2207 if (CodeGenFunction::ContainsLabel(S, IgnoreCaseStmts: true))
2208 return CSFC_Failure;
2209 return CSFC_Success;
2210 }
2211
2212 // Otherwise, we want to include this statement. Everything is cool with that
2213 // so long as it doesn't contain a break out of the switch we're in.
2214 if (CodeGenFunction::containsBreak(S)) return CSFC_Failure;
2215
2216 // Otherwise, everything is great. Include the statement and tell the caller
2217 // that we fall through and include the next statement as well.
2218 ResultStmts.push_back(Elt: S);
2219 return CSFC_FallThrough;
2220}
2221
2222/// FindCaseStatementsForValue - Find the case statement being jumped to and
2223/// then invoke CollectStatementsForCase to find the list of statements to emit
2224/// for a switch on constant. See the comment above CollectStatementsForCase
2225/// for more details.
2226static bool FindCaseStatementsForValue(const SwitchStmt &S,
2227 const llvm::APSInt &ConstantCondValue,
2228 SmallVectorImpl<const Stmt*> &ResultStmts,
2229 ASTContext &C,
2230 const SwitchCase *&ResultCase) {
2231 // First step, find the switch case that is being branched to. We can do this
2232 // efficiently by scanning the SwitchCase list.
2233 const SwitchCase *Case = S.getSwitchCaseList();
2234 const DefaultStmt *DefaultCase = nullptr;
2235
2236 for (; Case; Case = Case->getNextSwitchCase()) {
2237 // It's either a default or case. Just remember the default statement in
2238 // case we're not jumping to any numbered cases.
2239 if (const DefaultStmt *DS = dyn_cast<DefaultStmt>(Val: Case)) {
2240 DefaultCase = DS;
2241 continue;
2242 }
2243
2244 // Check to see if this case is the one we're looking for.
2245 const CaseStmt *CS = cast<CaseStmt>(Val: Case);
2246 // Don't handle case ranges yet.
2247 if (CS->getRHS()) return false;
2248
2249 // If we found our case, remember it as 'case'.
2250 if (CS->getLHS()->EvaluateKnownConstInt(Ctx: C) == ConstantCondValue)
2251 break;
2252 }
2253
2254 // If we didn't find a matching case, we use a default if it exists, or we
2255 // elide the whole switch body!
2256 if (!Case) {
2257 // It is safe to elide the body of the switch if it doesn't contain labels
2258 // etc. If it is safe, return successfully with an empty ResultStmts list.
2259 if (!DefaultCase)
2260 return !CodeGenFunction::ContainsLabel(S: &S);
2261 Case = DefaultCase;
2262 }
2263
2264 // Ok, we know which case is being jumped to, try to collect all the
2265 // statements that follow it. This can fail for a variety of reasons. Also,
2266 // check to see that the recursive walk actually found our case statement.
2267 // Insane cases like this can fail to find it in the recursive walk since we
2268 // don't handle every stmt kind:
2269 // switch (4) {
2270 // while (1) {
2271 // case 4: ...
2272 bool FoundCase = false;
2273 ResultCase = Case;
2274 return CollectStatementsForCase(S: S.getBody(), Case, FoundCase,
2275 ResultStmts) != CSFC_Failure &&
2276 FoundCase;
2277}
2278
2279static std::optional<SmallVector<uint64_t, 16>>
2280getLikelihoodWeights(ArrayRef<Stmt::Likelihood> Likelihoods) {
2281 // Are there enough branches to weight them?
2282 if (Likelihoods.size() <= 1)
2283 return std::nullopt;
2284
2285 uint64_t NumUnlikely = 0;
2286 uint64_t NumNone = 0;
2287 uint64_t NumLikely = 0;
2288 for (const auto LH : Likelihoods) {
2289 switch (LH) {
2290 case Stmt::LH_Unlikely:
2291 ++NumUnlikely;
2292 break;
2293 case Stmt::LH_None:
2294 ++NumNone;
2295 break;
2296 case Stmt::LH_Likely:
2297 ++NumLikely;
2298 break;
2299 }
2300 }
2301
2302 // Is there a likelihood attribute used?
2303 if (NumUnlikely == 0 && NumLikely == 0)
2304 return std::nullopt;
2305
2306 // When multiple cases share the same code they can be combined during
2307 // optimization. In that case the weights of the branch will be the sum of
2308 // the individual weights. Make sure the combined sum of all neutral cases
2309 // doesn't exceed the value of a single likely attribute.
2310 // The additions both avoid divisions by 0 and make sure the weights of None
2311 // don't exceed the weight of Likely.
2312 const uint64_t Likely = INT32_MAX / (NumLikely + 2);
2313 const uint64_t None = Likely / (NumNone + 1);
2314 const uint64_t Unlikely = 0;
2315
2316 SmallVector<uint64_t, 16> Result;
2317 Result.reserve(N: Likelihoods.size());
2318 for (const auto LH : Likelihoods) {
2319 switch (LH) {
2320 case Stmt::LH_Unlikely:
2321 Result.push_back(Elt: Unlikely);
2322 break;
2323 case Stmt::LH_None:
2324 Result.push_back(Elt: None);
2325 break;
2326 case Stmt::LH_Likely:
2327 Result.push_back(Elt: Likely);
2328 break;
2329 }
2330 }
2331
2332 return Result;
2333}
2334
2335void CodeGenFunction::EmitSwitchStmt(const SwitchStmt &S) {
2336 // Handle nested switch statements.
2337 llvm::SwitchInst *SavedSwitchInsn = SwitchInsn;
2338 SmallVector<uint64_t, 16> *SavedSwitchWeights = SwitchWeights;
2339 SmallVector<Stmt::Likelihood, 16> *SavedSwitchLikelihood = SwitchLikelihood;
2340 llvm::BasicBlock *SavedCRBlock = CaseRangeBlock;
2341
2342 // See if we can constant fold the condition of the switch and therefore only
2343 // emit the live case statement (if any) of the switch.
2344 llvm::APSInt ConstantCondValue;
2345 if (ConstantFoldsToSimpleInteger(Cond: S.getCond(), Result&: ConstantCondValue)) {
2346 SmallVector<const Stmt*, 4> CaseStmts;
2347 const SwitchCase *Case = nullptr;
2348 if (FindCaseStatementsForValue(S, ConstantCondValue, ResultStmts&: CaseStmts,
2349 C&: getContext(), ResultCase&: Case)) {
2350 if (Case)
2351 incrementProfileCounter(S: Case);
2352 RunCleanupsScope ExecutedScope(*this);
2353
2354 if (S.getInit())
2355 EmitStmt(S: S.getInit());
2356
2357 // Emit the condition variable if needed inside the entire cleanup scope
2358 // used by this special case for constant folded switches.
2359 if (S.getConditionVariable())
2360 EmitDecl(D: *S.getConditionVariable(), /*EvaluateConditionDecl=*/true);
2361
2362 // At this point, we are no longer "within" a switch instance, so
2363 // we can temporarily enforce this to ensure that any embedded case
2364 // statements are not emitted.
2365 SwitchInsn = nullptr;
2366
2367 // Okay, we can dead code eliminate everything except this case. Emit the
2368 // specified series of statements and we're good.
2369 for (const Stmt *CaseStmt : CaseStmts)
2370 EmitStmt(S: CaseStmt);
2371 incrementProfileCounter(S: &S);
2372 PGO->markStmtMaybeUsed(S: S.getBody());
2373
2374 // Now we want to restore the saved switch instance so that nested
2375 // switches continue to function properly
2376 SwitchInsn = SavedSwitchInsn;
2377
2378 return;
2379 }
2380 }
2381
2382 JumpDest SwitchExit = getJumpDestInCurrentScope(Name: "sw.epilog");
2383
2384 RunCleanupsScope ConditionScope(*this);
2385
2386 if (S.getInit())
2387 EmitStmt(S: S.getInit());
2388
2389 if (S.getConditionVariable())
2390 EmitDecl(D: *S.getConditionVariable());
2391 llvm::Value *CondV = EmitScalarExpr(E: S.getCond());
2392 MaybeEmitDeferredVarDeclInit(var: S.getConditionVariable());
2393
2394 // Create basic block to hold stuff that comes after switch
2395 // statement. We also need to create a default block now so that
2396 // explicit case ranges tests can have a place to jump to on
2397 // failure.
2398 llvm::BasicBlock *DefaultBlock = createBasicBlock(name: "sw.default");
2399 SwitchInsn = Builder.CreateSwitch(V: CondV, Dest: DefaultBlock);
2400 addInstToNewSourceAtom(KeyInstruction: SwitchInsn, Backup: CondV);
2401
2402 if (HLSLControlFlowAttr != HLSLControlFlowHintAttr::SpellingNotCalculated) {
2403 llvm::MDBuilder MDHelper(CGM.getLLVMContext());
2404 llvm::ConstantInt *BranchHintConstant =
2405 HLSLControlFlowAttr ==
2406 HLSLControlFlowHintAttr::Spelling::Microsoft_branch
2407 ? llvm::ConstantInt::get(Ty: CGM.Int32Ty, V: 1)
2408 : llvm::ConstantInt::get(Ty: CGM.Int32Ty, V: 2);
2409 llvm::Metadata *Vals[] = {MDHelper.createString(Str: "hlsl.controlflow.hint"),
2410 MDHelper.createConstant(C: BranchHintConstant)};
2411 SwitchInsn->setMetadata(Kind: "hlsl.controlflow.hint",
2412 Node: llvm::MDNode::get(Context&: CGM.getLLVMContext(), MDs: Vals));
2413 }
2414
2415 if (PGO->haveRegionCounts()) {
2416 // Walk the SwitchCase list to find how many there are.
2417 uint64_t DefaultCount = 0;
2418 unsigned NumCases = 0;
2419 for (const SwitchCase *Case = S.getSwitchCaseList();
2420 Case;
2421 Case = Case->getNextSwitchCase()) {
2422 if (isa<DefaultStmt>(Val: Case))
2423 DefaultCount = getProfileCount(S: Case);
2424 NumCases += 1;
2425 }
2426 SwitchWeights = new SmallVector<uint64_t, 16>();
2427 SwitchWeights->reserve(N: NumCases);
2428 // The default needs to be first. We store the edge count, so we already
2429 // know the right weight.
2430 SwitchWeights->push_back(Elt: DefaultCount);
2431 } else if (CGM.getCodeGenOpts().OptimizationLevel) {
2432 SwitchLikelihood = new SmallVector<Stmt::Likelihood, 16>();
2433 // Initialize the default case.
2434 SwitchLikelihood->push_back(Elt: Stmt::LH_None);
2435 }
2436
2437 CaseRangeBlock = DefaultBlock;
2438
2439 // Clear the insertion point to indicate we are in unreachable code.
2440 Builder.ClearInsertionPoint();
2441
2442 // All break statements jump to NextBlock. If BreakContinueStack is non-empty
2443 // then reuse last ContinueBlock.
2444 JumpDest OuterContinue;
2445 if (!BreakContinueStack.empty())
2446 OuterContinue = BreakContinueStack.back().ContinueBlock;
2447
2448 BreakContinueStack.push_back(Elt: BreakContinue(S, SwitchExit, OuterContinue));
2449
2450 // Emit switch body.
2451 EmitStmt(S: S.getBody());
2452
2453 BreakContinueStack.pop_back();
2454
2455 // Update the default block in case explicit case range tests have
2456 // been chained on top.
2457 SwitchInsn->setDefaultDest(CaseRangeBlock);
2458
2459 // If a default was never emitted:
2460 if (!DefaultBlock->getParent()) {
2461 // If we have cleanups, emit the default block so that there's a
2462 // place to jump through the cleanups from.
2463 if (ConditionScope.requiresCleanups()) {
2464 EmitBlock(BB: DefaultBlock);
2465
2466 // Otherwise, just forward the default block to the switch end.
2467 } else {
2468 DefaultBlock->replaceAllUsesWith(V: SwitchExit.getBlock());
2469 delete DefaultBlock;
2470 }
2471 }
2472
2473 ConditionScope.ForceCleanup();
2474
2475 // Close the last case (or DefaultBlock).
2476 EmitBranch(Target: SwitchExit.getBlock());
2477
2478 // Insert a False Counter if SwitchStmt doesn't have DefaultStmt.
2479 if (hasSkipCounter(S: S.getCond())) {
2480 auto *ImplicitDefaultBlock = createBasicBlock(name: "sw.false");
2481 EmitBlock(BB: ImplicitDefaultBlock);
2482 incrementProfileCounter(ExecSkip: UseSkipPath, S: S.getCond());
2483 Builder.CreateBr(Dest: SwitchInsn->getDefaultDest());
2484 SwitchInsn->setDefaultDest(ImplicitDefaultBlock);
2485 }
2486
2487 // Emit continuation.
2488 EmitBlock(BB: SwitchExit.getBlock(), IsFinished: true);
2489 incrementProfileCounter(S: &S);
2490
2491 // If the switch has a condition wrapped by __builtin_unpredictable,
2492 // create metadata that specifies that the switch is unpredictable.
2493 // Don't bother if not optimizing because that metadata would not be used.
2494 auto *Call = dyn_cast<CallExpr>(Val: S.getCond());
2495 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
2496 auto *FD = dyn_cast_or_null<FunctionDecl>(Val: Call->getCalleeDecl());
2497 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
2498 llvm::MDBuilder MDHelper(getLLVMContext());
2499 SwitchInsn->setMetadata(KindID: llvm::LLVMContext::MD_unpredictable,
2500 Node: MDHelper.createUnpredictable());
2501 }
2502 }
2503
2504 if (SwitchWeights) {
2505 assert(SwitchWeights->size() == 1 + SwitchInsn->getNumCases() &&
2506 "switch weights do not match switch cases");
2507 // If there's only one jump destination there's no sense weighting it.
2508 if (SwitchWeights->size() > 1)
2509 SwitchInsn->setMetadata(KindID: llvm::LLVMContext::MD_prof,
2510 Node: createProfileWeights(Weights: *SwitchWeights));
2511 delete SwitchWeights;
2512 } else if (SwitchLikelihood) {
2513 assert(SwitchLikelihood->size() == 1 + SwitchInsn->getNumCases() &&
2514 "switch likelihoods do not match switch cases");
2515 std::optional<SmallVector<uint64_t, 16>> LHW =
2516 getLikelihoodWeights(Likelihoods: *SwitchLikelihood);
2517 if (LHW) {
2518 llvm::MDBuilder MDHelper(CGM.getLLVMContext());
2519 SwitchInsn->setMetadata(KindID: llvm::LLVMContext::MD_prof,
2520 Node: createProfileWeights(Weights: *LHW));
2521 }
2522 delete SwitchLikelihood;
2523 }
2524 SwitchInsn = SavedSwitchInsn;
2525 SwitchWeights = SavedSwitchWeights;
2526 SwitchLikelihood = SavedSwitchLikelihood;
2527 CaseRangeBlock = SavedCRBlock;
2528}
2529
2530std::pair<llvm::Value*, llvm::Type *> CodeGenFunction::EmitAsmInputLValue(
2531 const TargetInfo::ConstraintInfo &Info, LValue InputValue,
2532 QualType InputType, std::string &ConstraintStr, SourceLocation Loc) {
2533 if (Info.allowsRegister() || !Info.allowsMemory()) {
2534 if (CodeGenFunction::hasScalarEvaluationKind(T: InputType))
2535 return {EmitLoadOfLValue(V: InputValue, Loc).getScalarVal(), nullptr};
2536
2537 llvm::Type *Ty = ConvertType(T: InputType);
2538 uint64_t Size = CGM.getDataLayout().getTypeSizeInBits(Ty);
2539 if ((Size <= 64 && llvm::isPowerOf2_64(Value: Size)) ||
2540 getTargetHooks().isScalarizableAsmOperand(CGF&: *this, Ty)) {
2541 Ty = llvm::IntegerType::get(C&: getLLVMContext(), NumBits: Size);
2542
2543 return {Builder.CreateLoad(Addr: InputValue.getAddress().withElementType(ElemTy: Ty)),
2544 nullptr};
2545 }
2546 }
2547
2548 Address Addr = InputValue.getAddress();
2549 ConstraintStr += '*';
2550 return {InputValue.getPointer(CGF&: *this), Addr.getElementType()};
2551}
2552std::pair<llvm::Value *, llvm::Type *>
2553CodeGenFunction::EmitAsmInput(const TargetInfo::ConstraintInfo &Info,
2554 const Expr *InputExpr,
2555 std::string &ConstraintStr) {
2556 // If this can't be a register or memory, i.e., has to be a constant
2557 // (immediate or symbolic), try to emit it as such.
2558 if (!Info.allowsRegister() && !Info.allowsMemory()) {
2559 if (Info.requiresImmediateConstant()) {
2560 Expr::EvalResult EVResult;
2561 InputExpr->EvaluateAsRValue(Result&: EVResult, Ctx: getContext(), InConstantContext: true);
2562
2563 llvm::APSInt IntResult;
2564 if (EVResult.Val.toIntegralConstant(Result&: IntResult, SrcTy: InputExpr->getType(),
2565 Ctx: getContext()))
2566 return {llvm::ConstantInt::get(Context&: getLLVMContext(), V: IntResult), nullptr};
2567 }
2568
2569 Expr::EvalResult Result;
2570 if (InputExpr->EvaluateAsInt(Result, Ctx: getContext()))
2571 return {llvm::ConstantInt::get(Context&: getLLVMContext(), V: Result.Val.getInt()),
2572 nullptr};
2573 }
2574
2575 if (Info.allowsRegister() || !Info.allowsMemory())
2576 if (CodeGenFunction::hasScalarEvaluationKind(T: InputExpr->getType()))
2577 return {EmitScalarExpr(E: InputExpr), nullptr};
2578 if (InputExpr->getStmtClass() == Expr::CXXThisExprClass)
2579 return {EmitScalarExpr(E: InputExpr), nullptr};
2580 InputExpr = InputExpr->IgnoreParenNoopCasts(Ctx: getContext());
2581 LValue Dest = EmitLValue(E: InputExpr);
2582 return EmitAsmInputLValue(Info, InputValue: Dest, InputType: InputExpr->getType(), ConstraintStr,
2583 Loc: InputExpr->getExprLoc());
2584}
2585
2586/// getAsmSrcLocInfo - Return the !srcloc metadata node to attach to an inline
2587/// asm call instruction. The !srcloc MDNode contains a list of constant
2588/// integers which are the source locations of the start of each line in the
2589/// asm.
2590static llvm::MDNode *getAsmSrcLocInfo(const StringLiteral *Str,
2591 CodeGenFunction &CGF) {
2592 SmallVector<llvm::Metadata *, 8> Locs;
2593 // Add the location of the first line to the MDNode.
2594 Locs.push_back(Elt: llvm::ConstantAsMetadata::get(C: llvm::ConstantInt::get(
2595 Ty: CGF.Int64Ty, V: Str->getBeginLoc().getRawEncoding())));
2596 StringRef StrVal = Str->getString();
2597 if (!StrVal.empty()) {
2598 const SourceManager &SM = CGF.CGM.getContext().getSourceManager();
2599 const LangOptions &LangOpts = CGF.CGM.getLangOpts();
2600 unsigned StartToken = 0;
2601 unsigned ByteOffset = 0;
2602
2603 // Add the location of the start of each subsequent line of the asm to the
2604 // MDNode.
2605 for (unsigned i = 0, e = StrVal.size() - 1; i != e; ++i) {
2606 if (StrVal[i] != '\n') continue;
2607 SourceLocation LineLoc = Str->getLocationOfByte(
2608 ByteNo: i + 1, SM, Features: LangOpts, Target: CGF.getTarget(), StartToken: &StartToken, StartTokenByteOffset: &ByteOffset);
2609 Locs.push_back(Elt: llvm::ConstantAsMetadata::get(
2610 C: llvm::ConstantInt::get(Ty: CGF.Int64Ty, V: LineLoc.getRawEncoding())));
2611 }
2612 }
2613
2614 return llvm::MDNode::get(Context&: CGF.getLLVMContext(), MDs: Locs);
2615}
2616
2617static void UpdateAsmCallInst(llvm::CallBase &Result, bool HasSideEffect,
2618 bool HasUnwindClobber, bool ReadOnly,
2619 bool ReadNone, bool NoMerge, bool NoConvergent,
2620 const AsmStmt &S,
2621 const std::vector<llvm::Type *> &ResultRegTypes,
2622 const std::vector<llvm::Type *> &ArgElemTypes,
2623 CodeGenFunction &CGF,
2624 std::vector<llvm::Value *> &RegResults) {
2625 if (!HasUnwindClobber)
2626 Result.addFnAttr(Kind: llvm::Attribute::NoUnwind);
2627
2628 if (NoMerge)
2629 Result.addFnAttr(Kind: llvm::Attribute::NoMerge);
2630 // Attach readnone and readonly attributes.
2631 if (!HasSideEffect) {
2632 if (ReadNone)
2633 Result.setDoesNotAccessMemory();
2634 else if (ReadOnly)
2635 Result.setOnlyReadsMemory();
2636 }
2637
2638 // Add elementtype attribute for indirect constraints.
2639 for (auto Pair : llvm::enumerate(First: ArgElemTypes)) {
2640 if (Pair.value()) {
2641 auto Attr = llvm::Attribute::get(
2642 Context&: CGF.getLLVMContext(), Kind: llvm::Attribute::ElementType, Ty: Pair.value());
2643 Result.addParamAttr(ArgNo: Pair.index(), Attr);
2644 }
2645 }
2646
2647 // Slap the source location of the inline asm into a !srcloc metadata on the
2648 // call.
2649 const StringLiteral *SL;
2650 if (const auto *gccAsmStmt = dyn_cast<GCCAsmStmt>(Val: &S);
2651 gccAsmStmt &&
2652 (SL = dyn_cast<StringLiteral>(Val: gccAsmStmt->getAsmStringExpr()))) {
2653 Result.setMetadata(Kind: "srcloc", Node: getAsmSrcLocInfo(Str: SL, CGF));
2654 } else {
2655 // At least put the line number on MS inline asm blobs and GCC asm constexpr
2656 // strings.
2657 llvm::Constant *Loc =
2658 llvm::ConstantInt::get(Ty: CGF.Int64Ty, V: S.getAsmLoc().getRawEncoding());
2659 Result.setMetadata(Kind: "srcloc",
2660 Node: llvm::MDNode::get(Context&: CGF.getLLVMContext(),
2661 MDs: llvm::ConstantAsMetadata::get(C: Loc)));
2662 }
2663
2664 // Make inline-asm calls Key for the debug info feature Key Instructions.
2665 CGF.addInstToNewSourceAtom(KeyInstruction: &Result, Backup: nullptr);
2666
2667 if (!NoConvergent && CGF.getLangOpts().assumeFunctionsAreConvergent())
2668 // Conservatively, mark all inline asm blocks in CUDA or OpenCL as
2669 // convergent (meaning, they may call an intrinsically convergent op, such
2670 // as bar.sync, and so can't have certain optimizations applied around
2671 // them) unless it's explicitly marked 'noconvergent'.
2672 Result.addFnAttr(Kind: llvm::Attribute::Convergent);
2673 // Extract all of the register value results from the asm.
2674 if (ResultRegTypes.size() == 1) {
2675 RegResults.push_back(x: &Result);
2676 } else {
2677 for (unsigned i = 0, e = ResultRegTypes.size(); i != e; ++i) {
2678 llvm::Value *Tmp = CGF.Builder.CreateExtractValue(Agg: &Result, Idxs: i, Name: "asmresult");
2679 RegResults.push_back(x: Tmp);
2680 }
2681 }
2682}
2683
2684static void
2685EmitAsmStores(CodeGenFunction &CGF, const AsmStmt &S,
2686 const llvm::ArrayRef<llvm::Value *> RegResults,
2687 const llvm::ArrayRef<llvm::Type *> ResultRegTypes,
2688 const llvm::ArrayRef<llvm::Type *> ResultTruncRegTypes,
2689 const llvm::ArrayRef<LValue> ResultRegDests,
2690 const llvm::ArrayRef<QualType> ResultRegQualTys,
2691 const llvm::BitVector &ResultTypeRequiresCast,
2692 const std::vector<std::optional<std::pair<unsigned, unsigned>>>
2693 &ResultBounds) {
2694 CGBuilderTy &Builder = CGF.Builder;
2695 CodeGenModule &CGM = CGF.CGM;
2696 llvm::LLVMContext &CTX = CGF.getLLVMContext();
2697
2698 assert(RegResults.size() == ResultRegTypes.size());
2699 assert(RegResults.size() == ResultTruncRegTypes.size());
2700 assert(RegResults.size() == ResultRegDests.size());
2701 // ResultRegDests can be also populated by addReturnRegisterOutputs() above,
2702 // in which case its size may grow.
2703 assert(ResultTypeRequiresCast.size() <= ResultRegDests.size());
2704 assert(ResultBounds.size() <= ResultRegDests.size());
2705
2706 for (unsigned i = 0, e = RegResults.size(); i != e; ++i) {
2707 llvm::Value *Tmp = RegResults[i];
2708 llvm::Type *TruncTy = ResultTruncRegTypes[i];
2709
2710 if ((i < ResultBounds.size()) && ResultBounds[i].has_value()) {
2711 const auto [LowerBound, UpperBound] = ResultBounds[i].value();
2712 // FIXME: Support for nonzero lower bounds not yet implemented.
2713 assert(LowerBound == 0 && "Output operand lower bound is not zero.");
2714 llvm::Constant *UpperBoundConst =
2715 llvm::ConstantInt::get(Ty: Tmp->getType(), V: UpperBound);
2716 llvm::Value *IsBooleanValue =
2717 Builder.CreateCmp(Pred: llvm::CmpInst::ICMP_ULT, LHS: Tmp, RHS: UpperBoundConst);
2718 llvm::Function *FnAssume = CGM.getIntrinsic(IID: llvm::Intrinsic::assume);
2719 Builder.CreateCall(Callee: FnAssume, Args: IsBooleanValue);
2720 }
2721
2722 // If the result type of the LLVM IR asm doesn't match the result type of
2723 // the expression, do the conversion.
2724 if (ResultRegTypes[i] != TruncTy) {
2725
2726 // Truncate the integer result to the right size, note that TruncTy can be
2727 // a pointer.
2728 if (TruncTy->isFloatingPointTy())
2729 Tmp = Builder.CreateFPTrunc(V: Tmp, DestTy: TruncTy);
2730 else if (TruncTy->isPointerTy() && Tmp->getType()->isIntegerTy()) {
2731 uint64_t ResSize = CGM.getDataLayout().getTypeSizeInBits(Ty: TruncTy);
2732 Tmp = Builder.CreateTrunc(
2733 V: Tmp, DestTy: llvm::IntegerType::get(C&: CTX, NumBits: (unsigned)ResSize));
2734 Tmp = Builder.CreateIntToPtr(V: Tmp, DestTy: TruncTy);
2735 } else if (Tmp->getType()->isPointerTy() && TruncTy->isIntegerTy()) {
2736 uint64_t TmpSize =
2737 CGM.getDataLayout().getTypeSizeInBits(Ty: Tmp->getType());
2738 Tmp = Builder.CreatePtrToInt(
2739 V: Tmp, DestTy: llvm::IntegerType::get(C&: CTX, NumBits: (unsigned)TmpSize));
2740 Tmp = Builder.CreateTrunc(V: Tmp, DestTy: TruncTy);
2741 } else if (Tmp->getType()->isIntegerTy() && TruncTy->isIntegerTy()) {
2742 Tmp = Builder.CreateZExtOrTrunc(V: Tmp, DestTy: TruncTy);
2743 } else if (Tmp->getType()->isVectorTy() || TruncTy->isVectorTy()) {
2744 Tmp = Builder.CreateBitCast(V: Tmp, DestTy: TruncTy);
2745 }
2746 }
2747
2748 ApplyAtomGroup Grp(CGF.getDebugInfo());
2749 LValue Dest = ResultRegDests[i];
2750 // ResultTypeRequiresCast elements correspond to the first
2751 // ResultTypeRequiresCast.size() elements of RegResults.
2752 if ((i < ResultTypeRequiresCast.size()) && ResultTypeRequiresCast[i]) {
2753 unsigned Size = CGF.getContext().getTypeSize(T: ResultRegQualTys[i]);
2754 Address A = Dest.getAddress().withElementType(ElemTy: ResultRegTypes[i]);
2755 if (CGF.getTargetHooks().isScalarizableAsmOperand(CGF, Ty: TruncTy)) {
2756 llvm::StoreInst *S = Builder.CreateStore(Val: Tmp, Addr: A);
2757 CGF.addInstToCurrentSourceAtom(KeyInstruction: S, Backup: S->getValueOperand());
2758 continue;
2759 }
2760
2761 QualType Ty =
2762 CGF.getContext().getIntTypeForBitwidth(DestWidth: Size, /*Signed=*/false);
2763 if (Ty.isNull()) {
2764 const Expr *OutExpr = S.getOutputExpr(i);
2765 CGM.getDiags().Report(Loc: OutExpr->getExprLoc(),
2766 DiagID: diag::err_store_value_to_reg);
2767 return;
2768 }
2769 Dest = CGF.MakeAddrLValue(Addr: A, T: Ty);
2770 }
2771 CGF.EmitStoreThroughLValue(Src: RValue::get(V: Tmp), Dst: Dest);
2772 }
2773}
2774
2775static void EmitHipStdParUnsupportedAsm(CodeGenFunction *CGF,
2776 const AsmStmt &S) {
2777 constexpr auto Name = "__ASM__hipstdpar_unsupported";
2778
2779 std::string Asm;
2780 if (auto GCCAsm = dyn_cast<GCCAsmStmt>(Val: &S))
2781 Asm = GCCAsm->getAsmString();
2782
2783 auto &Ctx = CGF->CGM.getLLVMContext();
2784
2785 auto StrTy = llvm::ConstantDataArray::getString(Context&: Ctx, Initializer: Asm);
2786 auto FnTy = llvm::FunctionType::get(Result: llvm::Type::getVoidTy(C&: Ctx),
2787 Params: {StrTy->getType()}, isVarArg: false);
2788 auto UBF = CGF->CGM.getModule().getOrInsertFunction(Name, T: FnTy);
2789
2790 CGF->Builder.CreateCall(Callee: UBF, Args: {StrTy});
2791}
2792
2793void CodeGenFunction::EmitAsmStmt(const AsmStmt &S) {
2794 // Pop all cleanup blocks at the end of the asm statement.
2795 CodeGenFunction::RunCleanupsScope Cleanups(*this);
2796
2797 // Assemble the final asm string.
2798 std::string AsmString = S.generateAsmString(C: getContext());
2799
2800 // Get all the output and input constraints together.
2801 SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
2802 SmallVector<TargetInfo::ConstraintInfo, 4> InputConstraintInfos;
2803
2804 bool IsHipStdPar = getLangOpts().HIPStdPar && getLangOpts().CUDAIsDevice;
2805 bool IsValidTargetAsm = true;
2806 for (unsigned i = 0, e = S.getNumOutputs(); i != e && IsValidTargetAsm; i++) {
2807 StringRef Name;
2808 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(Val: &S))
2809 Name = GAS->getOutputName(i);
2810 TargetInfo::ConstraintInfo Info(S.getOutputConstraint(i), Name);
2811 bool IsValid = getTarget().validateOutputConstraint(Info); (void)IsValid;
2812 if (IsHipStdPar && !IsValid)
2813 IsValidTargetAsm = false;
2814 else
2815 assert(IsValid && "Failed to parse output constraint");
2816 OutputConstraintInfos.push_back(Elt: Info);
2817 }
2818
2819 for (unsigned i = 0, e = S.getNumInputs(); i != e && IsValidTargetAsm; i++) {
2820 StringRef Name;
2821 if (const GCCAsmStmt *GAS = dyn_cast<GCCAsmStmt>(Val: &S))
2822 Name = GAS->getInputName(i);
2823 TargetInfo::ConstraintInfo Info(S.getInputConstraint(i), Name);
2824 bool IsValid =
2825 getTarget().validateInputConstraint(OutputConstraints: OutputConstraintInfos, info&: Info);
2826 if (IsHipStdPar && !IsValid)
2827 IsValidTargetAsm = false;
2828 else
2829 assert(IsValid && "Failed to parse input constraint");
2830 InputConstraintInfos.push_back(Elt: Info);
2831 }
2832
2833 if (!IsValidTargetAsm)
2834 return EmitHipStdParUnsupportedAsm(CGF: this, S);
2835
2836 std::string Constraints;
2837
2838 std::vector<LValue> ResultRegDests;
2839 std::vector<QualType> ResultRegQualTys;
2840 std::vector<llvm::Type *> ResultRegTypes;
2841 std::vector<llvm::Type *> ResultTruncRegTypes;
2842 std::vector<llvm::Type *> ArgTypes;
2843 std::vector<llvm::Type *> ArgElemTypes;
2844 std::vector<llvm::Value*> Args;
2845 llvm::BitVector ResultTypeRequiresCast;
2846 std::vector<std::optional<std::pair<unsigned, unsigned>>> ResultBounds;
2847
2848 // Keep track of inout constraints.
2849 std::string InOutConstraints;
2850 std::vector<llvm::Value*> InOutArgs;
2851 std::vector<llvm::Type*> InOutArgTypes;
2852 std::vector<llvm::Type*> InOutArgElemTypes;
2853
2854 // Keep track of out constraints for tied input operand.
2855 std::vector<std::string> OutputConstraints;
2856
2857 // Keep track of defined physregs.
2858 llvm::SmallSet<std::string, 8> PhysRegOutputs;
2859
2860 // An inline asm can be marked readonly if it meets the following conditions:
2861 // - it doesn't have any sideeffects
2862 // - it doesn't clobber memory
2863 // - it doesn't return a value by-reference
2864 // It can be marked readnone if it doesn't have any input memory constraints
2865 // in addition to meeting the conditions listed above.
2866 bool ReadOnly = true, ReadNone = true;
2867
2868 for (unsigned i = 0, e = S.getNumOutputs(); i != e; i++) {
2869 TargetInfo::ConstraintInfo &Info = OutputConstraintInfos[i];
2870
2871 // Simplify the output constraint.
2872 std::string OutputConstraint(S.getOutputConstraint(i));
2873 OutputConstraint = getTarget().simplifyConstraint(
2874 Constraint: StringRef(OutputConstraint).substr(Start: 1), OutCons: &OutputConstraintInfos);
2875
2876 const Expr *OutExpr = S.getOutputExpr(i);
2877 OutExpr = OutExpr->IgnoreParenNoopCasts(Ctx: getContext());
2878
2879 std::string GCCReg;
2880 OutputConstraint = S.addVariableConstraints(
2881 Constraint: OutputConstraint, AsmExpr: *OutExpr, Target: getTarget(), EarlyClobber: Info.earlyClobber(),
2882 UnsupportedCB: [&](const Stmt *UnspStmt, StringRef Msg) {
2883 CGM.ErrorUnsupported(S: UnspStmt, Type: Msg);
2884 },
2885 GCCReg: &GCCReg);
2886 // Give an error on multiple outputs to same physreg.
2887 if (!GCCReg.empty() && !PhysRegOutputs.insert(V: GCCReg).second)
2888 CGM.Error(loc: S.getAsmLoc(), error: "multiple outputs to hard register: " + GCCReg);
2889
2890 OutputConstraints.push_back(x: OutputConstraint);
2891 LValue Dest = EmitLValue(E: OutExpr);
2892 if (!Constraints.empty())
2893 Constraints += ',';
2894
2895 // If this is a register output, then make the inline asm return it
2896 // by-value. If this is a memory result, return the value by-reference.
2897 QualType QTy = OutExpr->getType();
2898 const bool IsScalarOrAggregate = hasScalarEvaluationKind(T: QTy) ||
2899 hasAggregateEvaluationKind(T: QTy);
2900 if (!Info.allowsMemory() && IsScalarOrAggregate) {
2901
2902 Constraints += "=" + OutputConstraint;
2903 ResultRegQualTys.push_back(x: QTy);
2904 ResultRegDests.push_back(x: Dest);
2905
2906 ResultBounds.emplace_back(args: Info.getOutputOperandBounds());
2907
2908 llvm::Type *Ty = ConvertTypeForMem(T: QTy);
2909 const bool RequiresCast = Info.allowsRegister() &&
2910 (getTargetHooks().isScalarizableAsmOperand(CGF&: *this, Ty) ||
2911 Ty->isAggregateType());
2912
2913 ResultTruncRegTypes.push_back(x: Ty);
2914 ResultTypeRequiresCast.push_back(Val: RequiresCast);
2915
2916 if (RequiresCast) {
2917 unsigned Size = getContext().getTypeSize(T: QTy);
2918 if (Size)
2919 Ty = llvm::IntegerType::get(C&: getLLVMContext(), NumBits: Size);
2920 else
2921 CGM.Error(loc: OutExpr->getExprLoc(), error: "output size should not be zero");
2922 }
2923 ResultRegTypes.push_back(x: Ty);
2924 // If this output is tied to an input, and if the input is larger, then
2925 // we need to set the actual result type of the inline asm node to be the
2926 // same as the input type.
2927 if (Info.hasMatchingInput()) {
2928 unsigned InputNo;
2929 for (InputNo = 0; InputNo != S.getNumInputs(); ++InputNo) {
2930 TargetInfo::ConstraintInfo &Input = InputConstraintInfos[InputNo];
2931 if (Input.hasTiedOperand() && Input.getTiedOperand() == i)
2932 break;
2933 }
2934 assert(InputNo != S.getNumInputs() && "Didn't find matching input!");
2935
2936 QualType InputTy = S.getInputExpr(i: InputNo)->getType();
2937 QualType OutputType = OutExpr->getType();
2938
2939 uint64_t InputSize = getContext().getTypeSize(T: InputTy);
2940 if (getContext().getTypeSize(T: OutputType) < InputSize) {
2941 // Form the asm to return the value as a larger integer or fp type.
2942 ResultRegTypes.back() = ConvertType(T: InputTy);
2943 }
2944 }
2945 if (llvm::Type* AdjTy =
2946 getTargetHooks().adjustInlineAsmType(CGF&: *this, Constraint: OutputConstraint,
2947 Ty: ResultRegTypes.back()))
2948 ResultRegTypes.back() = AdjTy;
2949 else {
2950 CGM.getDiags().Report(Loc: S.getAsmLoc(),
2951 DiagID: diag::err_asm_invalid_type_in_input)
2952 << OutExpr->getType() << OutputConstraint;
2953 }
2954
2955 // Update largest vector width for any vector types.
2956 if (auto *VT = dyn_cast<llvm::VectorType>(Val: ResultRegTypes.back()))
2957 LargestVectorWidth =
2958 std::max(a: (uint64_t)LargestVectorWidth,
2959 b: VT->getPrimitiveSizeInBits().getKnownMinValue());
2960 } else {
2961 Address DestAddr = Dest.getAddress();
2962 // Matrix types in memory are represented by arrays, but accessed through
2963 // vector pointers, with the alignment specified on the access operation.
2964 // For inline assembly, update pointer arguments to use vector pointers.
2965 // Otherwise there will be a mis-match if the matrix is also an
2966 // input-argument which is represented as vector.
2967 if (isa<MatrixType>(Val: OutExpr->getType().getCanonicalType()))
2968 DestAddr = DestAddr.withElementType(ElemTy: ConvertType(T: OutExpr->getType()));
2969
2970 ArgTypes.push_back(x: DestAddr.getType());
2971 ArgElemTypes.push_back(x: DestAddr.getElementType());
2972 Args.push_back(x: DestAddr.emitRawPointer(CGF&: *this));
2973 Constraints += "=*";
2974 Constraints += OutputConstraint;
2975 ReadOnly = ReadNone = false;
2976 }
2977
2978 if (Info.isReadWrite()) {
2979 InOutConstraints += ',';
2980
2981 const Expr *InputExpr = S.getOutputExpr(i);
2982 llvm::Value *Arg;
2983 llvm::Type *ArgElemType;
2984 std::tie(args&: Arg, args&: ArgElemType) = EmitAsmInputLValue(
2985 Info, InputValue: Dest, InputType: InputExpr->getType(), ConstraintStr&: InOutConstraints,
2986 Loc: InputExpr->getExprLoc());
2987
2988 if (llvm::Type* AdjTy =
2989 getTargetHooks().adjustInlineAsmType(CGF&: *this, Constraint: OutputConstraint,
2990 Ty: Arg->getType()))
2991 Arg = Builder.CreateBitCast(V: Arg, DestTy: AdjTy);
2992
2993 // Update largest vector width for any vector types.
2994 if (auto *VT = dyn_cast<llvm::VectorType>(Val: Arg->getType()))
2995 LargestVectorWidth =
2996 std::max(a: (uint64_t)LargestVectorWidth,
2997 b: VT->getPrimitiveSizeInBits().getKnownMinValue());
2998 // Only tie earlyclobber physregs.
2999 if (Info.allowsRegister() && (GCCReg.empty() || Info.earlyClobber()))
3000 InOutConstraints += llvm::utostr(X: i);
3001 else
3002 InOutConstraints += OutputConstraint;
3003
3004 InOutArgTypes.push_back(x: Arg->getType());
3005 InOutArgElemTypes.push_back(x: ArgElemType);
3006 InOutArgs.push_back(x: Arg);
3007 }
3008 }
3009
3010 // If this is a Microsoft-style asm blob, store the return registers (EAX:EDX)
3011 // to the return value slot. Only do this when returning in registers.
3012 if (isa<MSAsmStmt>(Val: &S)) {
3013 const ABIArgInfo &RetAI = CurFnInfo->getReturnInfo();
3014 if (RetAI.isDirect() || RetAI.isExtend()) {
3015 // Make a fake lvalue for the return value slot.
3016 LValue ReturnSlot = MakeAddrLValueWithoutTBAA(Addr: ReturnValue, T: FnRetTy);
3017 CGM.getTargetCodeGenInfo().addReturnRegisterOutputs(
3018 CGF&: *this, ReturnValue: ReturnSlot, Constraints, ResultRegTypes, ResultTruncRegTypes,
3019 ResultRegDests, AsmString, NumOutputs: S.getNumOutputs());
3020 SawAsmBlock = true;
3021 }
3022 }
3023
3024 for (unsigned i = 0, e = S.getNumInputs(); i != e; i++) {
3025 const Expr *InputExpr = S.getInputExpr(i);
3026
3027 TargetInfo::ConstraintInfo &Info = InputConstraintInfos[i];
3028
3029 if (Info.allowsMemory())
3030 ReadNone = false;
3031
3032 if (!Constraints.empty())
3033 Constraints += ',';
3034
3035 // Simplify the input constraint.
3036 std::string InputConstraint(S.getInputConstraint(i));
3037 InputConstraint =
3038 getTarget().simplifyConstraint(Constraint: InputConstraint, OutCons: &OutputConstraintInfos);
3039
3040 InputConstraint = S.addVariableConstraints(
3041 Constraint: InputConstraint, AsmExpr: *InputExpr->IgnoreParenNoopCasts(Ctx: getContext()),
3042 Target: getTarget(), EarlyClobber: false /* No EarlyClobber */,
3043 UnsupportedCB: [&](const Stmt *UnspStmt, std::string_view Msg) {
3044 CGM.ErrorUnsupported(S: UnspStmt, Type: Msg);
3045 });
3046
3047 std::string ReplaceConstraint (InputConstraint);
3048 llvm::Value *Arg;
3049 llvm::Type *ArgElemType;
3050 std::tie(args&: Arg, args&: ArgElemType) = EmitAsmInput(Info, InputExpr, ConstraintStr&: Constraints);
3051
3052 // If this input argument is tied to a larger output result, extend the
3053 // input to be the same size as the output. The LLVM backend wants to see
3054 // the input and output of a matching constraint be the same size. Note
3055 // that GCC does not define what the top bits are here. We use zext because
3056 // that is usually cheaper, but LLVM IR should really get an anyext someday.
3057 if (Info.hasTiedOperand()) {
3058 unsigned Output = Info.getTiedOperand();
3059 QualType OutputType = S.getOutputExpr(i: Output)->getType();
3060 QualType InputTy = InputExpr->getType();
3061
3062 if (getContext().getTypeSize(T: OutputType) >
3063 getContext().getTypeSize(T: InputTy)) {
3064 // Use ptrtoint as appropriate so that we can do our extension.
3065 if (isa<llvm::PointerType>(Val: Arg->getType()))
3066 Arg = Builder.CreatePtrToInt(V: Arg, DestTy: IntPtrTy);
3067 llvm::Type *OutputTy = ConvertType(T: OutputType);
3068 if (isa<llvm::IntegerType>(Val: OutputTy))
3069 Arg = Builder.CreateZExt(V: Arg, DestTy: OutputTy);
3070 else if (isa<llvm::PointerType>(Val: OutputTy))
3071 Arg = Builder.CreateZExt(V: Arg, DestTy: IntPtrTy);
3072 else if (OutputTy->isFloatingPointTy())
3073 Arg = Builder.CreateFPExt(V: Arg, DestTy: OutputTy);
3074 }
3075 // Deal with the tied operands' constraint code in adjustInlineAsmType.
3076 ReplaceConstraint = OutputConstraints[Output];
3077 }
3078 if (llvm::Type* AdjTy =
3079 getTargetHooks().adjustInlineAsmType(CGF&: *this, Constraint: ReplaceConstraint,
3080 Ty: Arg->getType()))
3081 Arg = Builder.CreateBitCast(V: Arg, DestTy: AdjTy);
3082 else
3083 CGM.getDiags().Report(Loc: S.getAsmLoc(), DiagID: diag::err_asm_invalid_type_in_input)
3084 << InputExpr->getType() << InputConstraint;
3085
3086 // Update largest vector width for any vector types.
3087 if (auto *VT = dyn_cast<llvm::VectorType>(Val: Arg->getType()))
3088 LargestVectorWidth =
3089 std::max(a: (uint64_t)LargestVectorWidth,
3090 b: VT->getPrimitiveSizeInBits().getKnownMinValue());
3091
3092 ArgTypes.push_back(x: Arg->getType());
3093 ArgElemTypes.push_back(x: ArgElemType);
3094 Args.push_back(x: Arg);
3095 Constraints += InputConstraint;
3096 }
3097
3098 // Append the "input" part of inout constraints.
3099 for (unsigned i = 0, e = InOutArgs.size(); i != e; i++) {
3100 ArgTypes.push_back(x: InOutArgTypes[i]);
3101 ArgElemTypes.push_back(x: InOutArgElemTypes[i]);
3102 Args.push_back(x: InOutArgs[i]);
3103 }
3104 Constraints += InOutConstraints;
3105
3106 // Labels
3107 SmallVector<llvm::BasicBlock *, 16> Transfer;
3108 llvm::BasicBlock *Fallthrough = nullptr;
3109 bool IsGCCAsmGoto = false;
3110 if (const auto *GS = dyn_cast<GCCAsmStmt>(Val: &S)) {
3111 IsGCCAsmGoto = GS->isAsmGoto();
3112 if (IsGCCAsmGoto) {
3113 for (const auto *E : GS->labels()) {
3114 JumpDest Dest = getJumpDestForLabel(D: E->getLabel());
3115 Transfer.push_back(Elt: Dest.getBlock());
3116 if (!Constraints.empty())
3117 Constraints += ',';
3118 Constraints += "!i";
3119 }
3120 Fallthrough = createBasicBlock(name: "asm.fallthrough");
3121 }
3122 }
3123
3124 bool HasUnwindClobber = false;
3125
3126 // Clobbers
3127 for (unsigned i = 0, e = S.getNumClobbers(); i != e; i++) {
3128 std::string Clobber = S.getClobber(i);
3129
3130 if (Clobber == "memory")
3131 ReadOnly = ReadNone = false;
3132 else if (Clobber == "unwind") {
3133 HasUnwindClobber = true;
3134 continue;
3135 } else if (Clobber != "cc") {
3136 Clobber = getTarget().getNormalizedGCCRegisterName(Name: Clobber);
3137 if (CGM.getCodeGenOpts().StackClashProtector &&
3138 getTarget().isSPRegName(Clobber)) {
3139 CGM.getDiags().Report(Loc: S.getAsmLoc(),
3140 DiagID: diag::warn_stack_clash_protection_inline_asm);
3141 }
3142 }
3143
3144 if (isa<MSAsmStmt>(Val: &S)) {
3145 if (Clobber == "eax" || Clobber == "edx") {
3146 if (Constraints.find(s: "=&A") != std::string::npos)
3147 continue;
3148 std::string::size_type position1 =
3149 Constraints.find(str: "={" + Clobber + "}");
3150 if (position1 != std::string::npos) {
3151 Constraints.insert(pos: position1 + 1, s: "&");
3152 continue;
3153 }
3154 std::string::size_type position2 = Constraints.find(s: "=A");
3155 if (position2 != std::string::npos) {
3156 Constraints.insert(pos: position2 + 1, s: "&");
3157 continue;
3158 }
3159 }
3160 }
3161 if (!Constraints.empty())
3162 Constraints += ',';
3163
3164 Constraints += "~{";
3165 Constraints += Clobber;
3166 Constraints += '}';
3167 }
3168
3169 assert(!(HasUnwindClobber && IsGCCAsmGoto) &&
3170 "unwind clobber can't be used with asm goto");
3171
3172 // Add machine specific clobbers
3173 std::string_view MachineClobbers = getTarget().getClobbers();
3174 if (!MachineClobbers.empty()) {
3175 if (!Constraints.empty())
3176 Constraints += ',';
3177 Constraints += MachineClobbers;
3178 }
3179
3180 llvm::Type *ResultType;
3181 if (ResultRegTypes.empty())
3182 ResultType = VoidTy;
3183 else if (ResultRegTypes.size() == 1)
3184 ResultType = ResultRegTypes[0];
3185 else
3186 ResultType = llvm::StructType::get(Context&: getLLVMContext(), Elements: ResultRegTypes);
3187
3188 llvm::FunctionType *FTy =
3189 llvm::FunctionType::get(Result: ResultType, Params: ArgTypes, isVarArg: false);
3190
3191 bool HasSideEffect = S.isVolatile() || S.getNumOutputs() == 0;
3192
3193 llvm::InlineAsm::AsmDialect GnuAsmDialect =
3194 CGM.getCodeGenOpts().getInlineAsmDialect() == CodeGenOptions::IAD_ATT
3195 ? llvm::InlineAsm::AD_ATT
3196 : llvm::InlineAsm::AD_Intel;
3197 llvm::InlineAsm::AsmDialect AsmDialect = isa<MSAsmStmt>(Val: &S) ?
3198 llvm::InlineAsm::AD_Intel : GnuAsmDialect;
3199
3200 llvm::InlineAsm *IA = llvm::InlineAsm::get(
3201 Ty: FTy, AsmString, Constraints, hasSideEffects: HasSideEffect,
3202 /* IsAlignStack */ isAlignStack: false, asmDialect: AsmDialect, canThrow: HasUnwindClobber);
3203 std::vector<llvm::Value*> RegResults;
3204 llvm::CallBrInst *CBR;
3205 llvm::DenseMap<llvm::BasicBlock *, SmallVector<llvm::Value *, 4>>
3206 CBRRegResults;
3207 if (IsGCCAsmGoto) {
3208 CBR = Builder.CreateCallBr(Callee: IA, DefaultDest: Fallthrough, IndirectDests: Transfer, Args);
3209 EmitBlock(BB: Fallthrough);
3210 UpdateAsmCallInst(Result&: *CBR, HasSideEffect, /*HasUnwindClobber=*/false, ReadOnly,
3211 ReadNone, NoMerge: InNoMergeAttributedStmt,
3212 NoConvergent: InNoConvergentAttributedStmt, S, ResultRegTypes,
3213 ArgElemTypes, CGF&: *this, RegResults);
3214 // Because we are emitting code top to bottom, we don't have enough
3215 // information at this point to know precisely whether we have a critical
3216 // edge. If we have outputs, split all indirect destinations.
3217 if (!RegResults.empty()) {
3218 unsigned i = 0;
3219 for (llvm::BasicBlock *Dest : CBR->getIndirectDests()) {
3220 llvm::Twine SynthName = Dest->getName() + ".split";
3221 llvm::BasicBlock *SynthBB = createBasicBlock(name: SynthName);
3222 llvm::IRBuilderBase::InsertPointGuard IPG(Builder);
3223 Builder.SetInsertPoint(SynthBB);
3224
3225 if (ResultRegTypes.size() == 1) {
3226 CBRRegResults[SynthBB].push_back(Elt: CBR);
3227 } else {
3228 for (unsigned j = 0, e = ResultRegTypes.size(); j != e; ++j) {
3229 llvm::Value *Tmp = Builder.CreateExtractValue(Agg: CBR, Idxs: j, Name: "asmresult");
3230 CBRRegResults[SynthBB].push_back(Elt: Tmp);
3231 }
3232 }
3233
3234 EmitBranch(Target: Dest);
3235 EmitBlock(BB: SynthBB);
3236 CBR->setIndirectDest(i: i++, B: SynthBB);
3237 }
3238 }
3239 } else if (HasUnwindClobber) {
3240 llvm::CallBase *Result = EmitCallOrInvoke(Callee: IA, Args, Name: "");
3241 UpdateAsmCallInst(Result&: *Result, HasSideEffect, /*HasUnwindClobber=*/true,
3242 ReadOnly, ReadNone, NoMerge: InNoMergeAttributedStmt,
3243 NoConvergent: InNoConvergentAttributedStmt, S, ResultRegTypes,
3244 ArgElemTypes, CGF&: *this, RegResults);
3245 } else {
3246 llvm::CallInst *Result =
3247 Builder.CreateCall(Callee: IA, Args, OpBundles: getBundlesForFunclet(Callee: IA));
3248 UpdateAsmCallInst(Result&: *Result, HasSideEffect, /*HasUnwindClobber=*/false,
3249 ReadOnly, ReadNone, NoMerge: InNoMergeAttributedStmt,
3250 NoConvergent: InNoConvergentAttributedStmt, S, ResultRegTypes,
3251 ArgElemTypes, CGF&: *this, RegResults);
3252 }
3253
3254 EmitAsmStores(CGF&: *this, S, RegResults, ResultRegTypes, ResultTruncRegTypes,
3255 ResultRegDests, ResultRegQualTys, ResultTypeRequiresCast,
3256 ResultBounds);
3257
3258 // If this is an asm goto with outputs, repeat EmitAsmStores, but with a
3259 // different insertion point; one for each indirect destination and with
3260 // CBRRegResults rather than RegResults.
3261 if (IsGCCAsmGoto && !CBRRegResults.empty()) {
3262 for (llvm::BasicBlock *Succ : CBR->getIndirectDests()) {
3263 llvm::IRBuilderBase::InsertPointGuard IPG(Builder);
3264 Builder.SetInsertPoint(TheBB: Succ, IP: --(Succ->end()));
3265 EmitAsmStores(CGF&: *this, S, RegResults: CBRRegResults[Succ], ResultRegTypes,
3266 ResultTruncRegTypes, ResultRegDests, ResultRegQualTys,
3267 ResultTypeRequiresCast, ResultBounds);
3268 }
3269 }
3270}
3271
3272LValue CodeGenFunction::InitCapturedStruct(const CapturedStmt &S) {
3273 const RecordDecl *RD = S.getCapturedRecordDecl();
3274 CanQualType RecordTy = getContext().getCanonicalTagType(TD: RD);
3275
3276 // Initialize the captured struct.
3277 LValue SlotLV =
3278 MakeAddrLValue(Addr: CreateMemTemp(T: RecordTy, Name: "agg.captured"), T: RecordTy);
3279
3280 RecordDecl::field_iterator CurField = RD->field_begin();
3281 for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(),
3282 E = S.capture_init_end();
3283 I != E; ++I, ++CurField) {
3284 LValue LV = EmitLValueForFieldInitialization(Base: SlotLV, Field: *CurField);
3285 if (CurField->hasCapturedVLAType()) {
3286 EmitLambdaVLACapture(VAT: CurField->getCapturedVLAType(), LV);
3287 } else {
3288 EmitInitializerForField(Field: *CurField, LHS: LV, Init: *I);
3289 }
3290 }
3291
3292 return SlotLV;
3293}
3294
3295/// Generate an outlined function for the body of a CapturedStmt, store any
3296/// captured variables into the captured struct, and call the outlined function.
3297llvm::Function *
3298CodeGenFunction::EmitCapturedStmt(const CapturedStmt &S, CapturedRegionKind K) {
3299 LValue CapStruct = InitCapturedStruct(S);
3300
3301 // Emit the CapturedDecl
3302 CodeGenFunction CGF(CGM, true);
3303 CGCapturedStmtRAII CapInfoRAII(CGF, new CGCapturedStmtInfo(S, K));
3304 llvm::Function *F = CGF.GenerateCapturedStmtFunction(S);
3305 delete CGF.CapturedStmtInfo;
3306
3307 // Emit call to the helper function.
3308 EmitCallOrInvoke(Callee: F, Args: CapStruct.getPointer(CGF&: *this));
3309
3310 return F;
3311}
3312
3313Address CodeGenFunction::GenerateCapturedStmtArgument(const CapturedStmt &S) {
3314 LValue CapStruct = InitCapturedStruct(S);
3315 return CapStruct.getAddress();
3316}
3317
3318/// Creates the outlined function for a CapturedStmt.
3319llvm::Function *
3320CodeGenFunction::GenerateCapturedStmtFunction(const CapturedStmt &S) {
3321 assert(CapturedStmtInfo &&
3322 "CapturedStmtInfo should be set when generating the captured function");
3323 const CapturedDecl *CD = S.getCapturedDecl();
3324 const RecordDecl *RD = S.getCapturedRecordDecl();
3325 SourceLocation Loc = S.getBeginLoc();
3326 assert(CD->hasBody() && "missing CapturedDecl body");
3327
3328 // Build the argument list.
3329 ASTContext &Ctx = CGM.getContext();
3330 FunctionArgList Args;
3331 Args.append(in_start: CD->param_begin(), in_end: CD->param_end());
3332
3333 // Create the function declaration.
3334 const CGFunctionInfo &FuncInfo =
3335 CGM.getTypes().arrangeBuiltinFunctionDeclaration(resultType: Ctx.VoidTy, args: Args);
3336 llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(Info: FuncInfo);
3337
3338 llvm::Function *F =
3339 llvm::Function::Create(Ty: FuncLLVMTy, Linkage: llvm::GlobalValue::InternalLinkage,
3340 N: CapturedStmtInfo->getHelperName(), M: &CGM.getModule());
3341 CGM.SetInternalFunctionAttributes(GD: CD, F, FI: FuncInfo);
3342 if (!CGM.getCodeGenOpts().SampleProfileFile.empty())
3343 F->addFnAttr(Kind: "sample-profile-suffix-elision-policy", Val: "selected");
3344 if (CD->isNothrow())
3345 F->addFnAttr(Kind: llvm::Attribute::NoUnwind);
3346
3347 // Generate the function.
3348 StartFunction(GD: CD, RetTy: Ctx.VoidTy, Fn: F, FnInfo: FuncInfo, Args, Loc: CD->getLocation(),
3349 StartLoc: CD->getBody()->getBeginLoc());
3350 // Set the context parameter in CapturedStmtInfo.
3351 Address DeclPtr = GetAddrOfLocalVar(VD: CD->getContextParam());
3352 CapturedStmtInfo->setContextValue(Builder.CreateLoad(Addr: DeclPtr));
3353
3354 // Initialize variable-length arrays.
3355 LValue Base = MakeNaturalAlignRawAddrLValue(
3356 V: CapturedStmtInfo->getContextValue(), T: Ctx.getCanonicalTagType(TD: RD));
3357 for (auto *FD : RD->fields()) {
3358 if (FD->hasCapturedVLAType()) {
3359 auto *ExprArg =
3360 EmitLoadOfLValue(V: EmitLValueForField(Base, Field: FD), Loc: S.getBeginLoc())
3361 .getScalarVal();
3362 auto VAT = FD->getCapturedVLAType();
3363 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
3364 }
3365 }
3366
3367 // If 'this' is captured, load it into CXXThisValue.
3368 if (CapturedStmtInfo->isCXXThisExprCaptured()) {
3369 FieldDecl *FD = CapturedStmtInfo->getThisFieldDecl();
3370 LValue ThisLValue = EmitLValueForField(Base, Field: FD);
3371 CXXThisValue = EmitLoadOfLValue(V: ThisLValue, Loc).getScalarVal();
3372 }
3373
3374 PGO->assignRegionCounters(GD: GlobalDecl(CD), Fn: F);
3375 CapturedStmtInfo->EmitBody(CGF&: *this, S: CD->getBody());
3376 FinishFunction(EndLoc: CD->getBodyRBrace());
3377
3378 return F;
3379}
3380
3381// Returns the first convergence entry/loop/anchor instruction found in |BB|.
3382// std::nullptr otherwise.
3383static llvm::ConvergenceControlInst *getConvergenceToken(llvm::BasicBlock *BB) {
3384 for (auto &I : *BB) {
3385 if (auto *CI = dyn_cast<llvm::ConvergenceControlInst>(Val: &I))
3386 return CI;
3387 }
3388 return nullptr;
3389}
3390
3391llvm::CallBase *
3392CodeGenFunction::addConvergenceControlToken(llvm::CallBase *Input) {
3393 llvm::ConvergenceControlInst *ParentToken = ConvergenceTokenStack.back();
3394 assert(ParentToken);
3395
3396 llvm::Value *bundleArgs[] = {ParentToken};
3397 llvm::OperandBundleDef OB("convergencectrl", bundleArgs);
3398 auto *Output = llvm::CallBase::addOperandBundle(
3399 CB: Input, ID: llvm::LLVMContext::OB_convergencectrl, OB, InsertPt: Input->getIterator());
3400 Input->replaceAllUsesWith(V: Output);
3401 Input->eraseFromParent();
3402 return Output;
3403}
3404
3405llvm::ConvergenceControlInst *
3406CodeGenFunction::emitConvergenceLoopToken(llvm::BasicBlock *BB) {
3407 llvm::ConvergenceControlInst *ParentToken = ConvergenceTokenStack.back();
3408 assert(ParentToken);
3409 return llvm::ConvergenceControlInst::CreateLoop(BB&: *BB, Parent: ParentToken);
3410}
3411
3412llvm::ConvergenceControlInst *
3413CodeGenFunction::getOrEmitConvergenceEntryToken(llvm::Function *F) {
3414 llvm::BasicBlock *BB = &F->getEntryBlock();
3415 llvm::ConvergenceControlInst *Token = getConvergenceToken(BB);
3416 if (Token)
3417 return Token;
3418
3419 // Adding a convergence token requires the function to be marked as
3420 // convergent.
3421 F->setConvergent();
3422 return llvm::ConvergenceControlInst::CreateEntry(BB&: *BB);
3423}
3424