diff options
author | Dimitry Andric <dim@FreeBSD.org> | 2016-02-13 14:58:13 +0000 |
---|---|---|
committer | Dimitry Andric <dim@FreeBSD.org> | 2016-02-13 14:58:13 +0000 |
commit | d4aec3a22f5b4c987be1c3815fdadbac72c6de5b (patch) | |
tree | a57d2c8ef3c5eaafb94bf394eb3dacdfeb6df08b /lib/CodeGen | |
parent | c2f1760e15d4e9e7a9eeadc99e8575fa1b2f7389 (diff) |
Notes
Diffstat (limited to 'lib/CodeGen')
-rw-r--r-- | lib/CodeGen/Address.h | 20 | ||||
-rw-r--r-- | lib/CodeGen/CGOpenMPRuntime.cpp | 17 | ||||
-rw-r--r-- | lib/CodeGen/CGStmtOpenMP.cpp | 226 |
3 files changed, 110 insertions, 153 deletions
diff --git a/lib/CodeGen/Address.h b/lib/CodeGen/Address.h index 9d145fa26b5f8..334308081ff3c 100644 --- a/lib/CodeGen/Address.h +++ b/lib/CodeGen/Address.h @@ -104,23 +104,15 @@ public: }; } -} -namespace llvm { - // Present a minimal LLVM-like casting interface. - template <class U> inline U cast(clang::CodeGen::Address addr) { - return U::castImpl(addr); - } - template <class U> inline bool isa(clang::CodeGen::Address addr) { - return U::isaImpl(addr); - } +// Present a minimal LLVM-like casting interface. +template <class U> inline U cast(CodeGen::Address addr) { + return U::castImpl(addr); +} +template <class U> inline bool isa(CodeGen::Address addr) { + return U::isaImpl(addr); } -namespace clang { - // Make our custom isa and cast available in namespace clang, to mirror - // what we do for LLVM's versions in Basic/LLVM.h. - using llvm::isa; - using llvm::cast; } #endif diff --git a/lib/CodeGen/CGOpenMPRuntime.cpp b/lib/CodeGen/CGOpenMPRuntime.cpp index 015a7396ffbe3..5cfacacbe01a2 100644 --- a/lib/CodeGen/CGOpenMPRuntime.cpp +++ b/lib/CodeGen/CGOpenMPRuntime.cpp @@ -483,7 +483,7 @@ llvm::Value *CGOpenMPRuntime::getThreadID(CodeGenFunction &CGF, if (ThreadID != nullptr) return ThreadID; } - if (auto OMPRegionInfo = + if (auto *OMPRegionInfo = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { if (OMPRegionInfo->getThreadIDVariable()) { // Check if this an outlined function with thread id passed as argument. @@ -1356,7 +1356,7 @@ void CGOpenMPRuntime::emitParallelCall(CodeGenFunction &CGF, SourceLocation Loc, // return the address of that temp. Address CGOpenMPRuntime::emitThreadIDAddress(CodeGenFunction &CGF, SourceLocation Loc) { - if (auto OMPRegionInfo = + if (auto *OMPRegionInfo = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) if (OMPRegionInfo->getThreadIDVariable()) return OMPRegionInfo->getThreadIDVariableLValue(CGF).getAddress(); @@ -1717,15 +1717,10 @@ void CGOpenMPRuntime::emitBarrierCall(CodeGenFunction &CGF, SourceLocation Loc, } // Build call __kmpc_cancel_barrier(loc, thread_id) or __kmpc_barrier(loc, // thread_id); - auto *OMPRegionInfo = - dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo); - // Do not emit barrier call in the single directive emitted in some rare cases - // for sections directives. - if (OMPRegionInfo && OMPRegionInfo->getDirectiveKind() == OMPD_single) - return; llvm::Value *Args[] = {emitUpdateLocation(CGF, Loc, Flags), getThreadID(CGF, Loc)}; - if (OMPRegionInfo) { + if (auto *OMPRegionInfo = + dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { if (!ForceSimpleCall && OMPRegionInfo->hasCancel()) { auto *Result = CGF.EmitRuntimeCall( createRuntimeFunction(OMPRTL__kmpc_cancel_barrier), Args); @@ -3649,8 +3644,6 @@ void CGOpenMPRuntime::emitCancellationPointCall( // global_tid, kmp_int32 cncl_kind); if (auto *OMPRegionInfo = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { - if (OMPRegionInfo->getDirectiveKind() == OMPD_single) - return; if (OMPRegionInfo->hasCancel()) { llvm::Value *Args[] = { emitUpdateLocation(CGF, Loc), getThreadID(CGF, Loc), @@ -3687,8 +3680,6 @@ void CGOpenMPRuntime::emitCancelCall(CodeGenFunction &CGF, SourceLocation Loc, // kmp_int32 cncl_kind); if (auto *OMPRegionInfo = dyn_cast_or_null<CGOpenMPRegionInfo>(CGF.CapturedStmtInfo)) { - if (OMPRegionInfo->getDirectiveKind() == OMPD_single) - return; auto &&ThenGen = [this, Loc, CancelRegion, OMPRegionInfo](CodeGenFunction &CGF) { llvm::Value *Args[] = { diff --git a/lib/CodeGen/CGStmtOpenMP.cpp b/lib/CodeGen/CGStmtOpenMP.cpp index 68555128ea015..59821a8d0330b 100644 --- a/lib/CodeGen/CGStmtOpenMP.cpp +++ b/lib/CodeGen/CGStmtOpenMP.cpp @@ -1657,50 +1657,51 @@ OpenMPDirectiveKind CodeGenFunction::EmitSections(const OMPExecutableDirective &S) { auto *Stmt = cast<CapturedStmt>(S.getAssociatedStmt())->getCapturedStmt(); auto *CS = dyn_cast<CompoundStmt>(Stmt); - if (CS && CS->size() > 1) { - bool HasLastprivates = false; - auto &&CodeGen = [&S, CS, &HasLastprivates](CodeGenFunction &CGF) { - auto &C = CGF.CGM.getContext(); - auto KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); - // Emit helper vars inits. - LValue LB = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.lb.", - CGF.Builder.getInt32(0)); - auto *GlobalUBVal = CGF.Builder.getInt32(CS->size() - 1); - LValue UB = - createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.ub.", GlobalUBVal); - LValue ST = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.st.", - CGF.Builder.getInt32(1)); - LValue IL = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.il.", - CGF.Builder.getInt32(0)); - // Loop counter. - LValue IV = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.iv."); - OpaqueValueExpr IVRefExpr(S.getLocStart(), KmpInt32Ty, VK_LValue); - CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV); - OpaqueValueExpr UBRefExpr(S.getLocStart(), KmpInt32Ty, VK_LValue); - CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB); - // Generate condition for loop. - BinaryOperator Cond(&IVRefExpr, &UBRefExpr, BO_LE, C.BoolTy, VK_RValue, - OK_Ordinary, S.getLocStart(), - /*fpContractable=*/false); - // Increment for loop counter. - UnaryOperator Inc(&IVRefExpr, UO_PreInc, KmpInt32Ty, VK_RValue, - OK_Ordinary, S.getLocStart()); - auto BodyGen = [CS, &S, &IV](CodeGenFunction &CGF) { - // Iterate through all sections and emit a switch construct: - // switch (IV) { - // case 0: - // <SectionStmt[0]>; - // break; - // ... - // case <NumSection> - 1: - // <SectionStmt[<NumSection> - 1]>; - // break; - // } - // .omp.sections.exit: - auto *ExitBB = CGF.createBasicBlock(".omp.sections.exit"); - auto *SwitchStmt = CGF.Builder.CreateSwitch( - CGF.EmitLoadOfLValue(IV, S.getLocStart()).getScalarVal(), ExitBB, - CS->size()); + bool HasLastprivates = false; + auto &&CodeGen = [&S, Stmt, CS, &HasLastprivates](CodeGenFunction &CGF) { + auto &C = CGF.CGM.getContext(); + auto KmpInt32Ty = C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); + // Emit helper vars inits. + LValue LB = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.lb.", + CGF.Builder.getInt32(0)); + auto *GlobalUBVal = CS != nullptr ? CGF.Builder.getInt32(CS->size() - 1) + : CGF.Builder.getInt32(0); + LValue UB = + createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.ub.", GlobalUBVal); + LValue ST = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.st.", + CGF.Builder.getInt32(1)); + LValue IL = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.il.", + CGF.Builder.getInt32(0)); + // Loop counter. + LValue IV = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.iv."); + OpaqueValueExpr IVRefExpr(S.getLocStart(), KmpInt32Ty, VK_LValue); + CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV); + OpaqueValueExpr UBRefExpr(S.getLocStart(), KmpInt32Ty, VK_LValue); + CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB); + // Generate condition for loop. + BinaryOperator Cond(&IVRefExpr, &UBRefExpr, BO_LE, C.BoolTy, VK_RValue, + OK_Ordinary, S.getLocStart(), + /*fpContractable=*/false); + // Increment for loop counter. + UnaryOperator Inc(&IVRefExpr, UO_PreInc, KmpInt32Ty, VK_RValue, OK_Ordinary, + S.getLocStart()); + auto BodyGen = [Stmt, CS, &S, &IV](CodeGenFunction &CGF) { + // Iterate through all sections and emit a switch construct: + // switch (IV) { + // case 0: + // <SectionStmt[0]>; + // break; + // ... + // case <NumSection> - 1: + // <SectionStmt[<NumSection> - 1]>; + // break; + // } + // .omp.sections.exit: + auto *ExitBB = CGF.createBasicBlock(".omp.sections.exit"); + auto *SwitchStmt = CGF.Builder.CreateSwitch( + CGF.EmitLoadOfLValue(IV, S.getLocStart()).getScalarVal(), ExitBB, + CS == nullptr ? 1 : CS->size()); + if (CS) { unsigned CaseNumber = 0; for (auto *SubStmt : CS->children()) { auto CaseBB = CGF.createBasicBlock(".omp.sections.case"); @@ -1710,99 +1711,72 @@ CodeGenFunction::EmitSections(const OMPExecutableDirective &S) { CGF.EmitBranch(ExitBB); ++CaseNumber; } - CGF.EmitBlock(ExitBB, /*IsFinished=*/true); - }; - - CodeGenFunction::OMPPrivateScope LoopScope(CGF); - if (CGF.EmitOMPFirstprivateClause(S, LoopScope)) { - // Emit implicit barrier to synchronize threads and avoid data races on - // initialization of firstprivate variables. - CGF.CGM.getOpenMPRuntime().emitBarrierCall( - CGF, S.getLocStart(), OMPD_unknown, /*EmitChecks=*/false, - /*ForceSimpleCall=*/true); + } else { + auto CaseBB = CGF.createBasicBlock(".omp.sections.case"); + CGF.EmitBlock(CaseBB); + SwitchStmt->addCase(CGF.Builder.getInt32(0), CaseBB); + CGF.EmitStmt(Stmt); + CGF.EmitBranch(ExitBB); } - CGF.EmitOMPPrivateClause(S, LoopScope); - HasLastprivates = CGF.EmitOMPLastprivateClauseInit(S, LoopScope); - CGF.EmitOMPReductionClauseInit(S, LoopScope); - (void)LoopScope.Privatize(); - - // Emit static non-chunked loop. - CGF.CGM.getOpenMPRuntime().emitForStaticInit( - CGF, S.getLocStart(), OMPC_SCHEDULE_static, /*IVSize=*/32, - /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(), - LB.getAddress(), UB.getAddress(), ST.getAddress()); - // UB = min(UB, GlobalUB); - auto *UBVal = CGF.EmitLoadOfScalar(UB, S.getLocStart()); - auto *MinUBGlobalUB = CGF.Builder.CreateSelect( - CGF.Builder.CreateICmpSLT(UBVal, GlobalUBVal), UBVal, GlobalUBVal); - CGF.EmitStoreOfScalar(MinUBGlobalUB, UB); - // IV = LB; - CGF.EmitStoreOfScalar(CGF.EmitLoadOfScalar(LB, S.getLocStart()), IV); - // while (idx <= UB) { BODY; ++idx; } - CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, &Cond, &Inc, BodyGen, - [](CodeGenFunction &) {}); - // Tell the runtime we are done. - CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getLocStart()); - CGF.EmitOMPReductionClauseFinal(S); - - // Emit final copy of the lastprivate variables if IsLastIter != 0. - if (HasLastprivates) - CGF.EmitOMPLastprivateClauseFinal( - S, CGF.Builder.CreateIsNotNull( - CGF.EmitLoadOfScalar(IL, S.getLocStart()))); + CGF.EmitBlock(ExitBB, /*IsFinished=*/true); }; - bool HasCancel = false; - if (auto *OSD = dyn_cast<OMPSectionsDirective>(&S)) - HasCancel = OSD->hasCancel(); - else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&S)) - HasCancel = OPSD->hasCancel(); - CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_sections, CodeGen, - HasCancel); - // Emit barrier for lastprivates only if 'sections' directive has 'nowait' - // clause. Otherwise the barrier will be generated by the codegen for the - // directive. - if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) { + CodeGenFunction::OMPPrivateScope LoopScope(CGF); + if (CGF.EmitOMPFirstprivateClause(S, LoopScope)) { // Emit implicit barrier to synchronize threads and avoid data races on // initialization of firstprivate variables. - CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getLocStart(), - OMPD_unknown); + CGF.CGM.getOpenMPRuntime().emitBarrierCall( + CGF, S.getLocStart(), OMPD_unknown, /*EmitChecks=*/false, + /*ForceSimpleCall=*/true); } - return OMPD_sections; - } - // If only one section is found - no need to generate loop, emit as a single - // region. - bool HasFirstprivates; - // No need to generate reductions for sections with single section region, we - // can use original shared variables for all operations. - bool HasReductions = S.hasClausesOfKind<OMPReductionClause>(); - // No need to generate lastprivates for sections with single section region, - // we can use original shared variable for all calculations with barrier at - // the end of the sections. - bool HasLastprivates = S.hasClausesOfKind<OMPLastprivateClause>(); - auto &&CodeGen = [Stmt, &S, &HasFirstprivates](CodeGenFunction &CGF) { - CodeGenFunction::OMPPrivateScope SingleScope(CGF); - HasFirstprivates = CGF.EmitOMPFirstprivateClause(S, SingleScope); - CGF.EmitOMPPrivateClause(S, SingleScope); - (void)SingleScope.Privatize(); + CGF.EmitOMPPrivateClause(S, LoopScope); + HasLastprivates = CGF.EmitOMPLastprivateClauseInit(S, LoopScope); + CGF.EmitOMPReductionClauseInit(S, LoopScope); + (void)LoopScope.Privatize(); + + // Emit static non-chunked loop. + CGF.CGM.getOpenMPRuntime().emitForStaticInit( + CGF, S.getLocStart(), OMPC_SCHEDULE_static, /*IVSize=*/32, + /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(), LB.getAddress(), + UB.getAddress(), ST.getAddress()); + // UB = min(UB, GlobalUB); + auto *UBVal = CGF.EmitLoadOfScalar(UB, S.getLocStart()); + auto *MinUBGlobalUB = CGF.Builder.CreateSelect( + CGF.Builder.CreateICmpSLT(UBVal, GlobalUBVal), UBVal, GlobalUBVal); + CGF.EmitStoreOfScalar(MinUBGlobalUB, UB); + // IV = LB; + CGF.EmitStoreOfScalar(CGF.EmitLoadOfScalar(LB, S.getLocStart()), IV); + // while (idx <= UB) { BODY; ++idx; } + CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, &Cond, &Inc, BodyGen, + [](CodeGenFunction &) {}); + // Tell the runtime we are done. + CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getLocStart()); + CGF.EmitOMPReductionClauseFinal(S); - CGF.EmitStmt(Stmt); + // Emit final copy of the lastprivate variables if IsLastIter != 0. + if (HasLastprivates) + CGF.EmitOMPLastprivateClauseFinal( + S, CGF.Builder.CreateIsNotNull( + CGF.EmitLoadOfScalar(IL, S.getLocStart()))); }; - CGM.getOpenMPRuntime().emitSingleRegion(*this, CodeGen, S.getLocStart(), - llvm::None, llvm::None, llvm::None, - llvm::None); - // Emit barrier for firstprivates, lastprivates or reductions only if - // 'sections' directive has 'nowait' clause. Otherwise the barrier will be - // generated by the codegen for the directive. - if ((HasFirstprivates || HasLastprivates || HasReductions) && - S.getSingleClause<OMPNowaitClause>()) { + + bool HasCancel = false; + if (auto *OSD = dyn_cast<OMPSectionsDirective>(&S)) + HasCancel = OSD->hasCancel(); + else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&S)) + HasCancel = OPSD->hasCancel(); + CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_sections, CodeGen, + HasCancel); + // Emit barrier for lastprivates only if 'sections' directive has 'nowait' + // clause. Otherwise the barrier will be generated by the codegen for the + // directive. + if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) { // Emit implicit barrier to synchronize threads and avoid data races on // initialization of firstprivate variables. - CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getLocStart(), OMPD_unknown, - /*EmitChecks=*/false, - /*ForceSimpleCall=*/true); + CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getLocStart(), + OMPD_unknown); } - return OMPD_single; + return OMPD_sections; } void CodeGenFunction::EmitOMPSectionsDirective(const OMPSectionsDirective &S) { |