[OPENMP][NVPTX]Run combined constructs with if clause in SPMD mode.
All target-parallel-based constructs can be run in SPMD mode from now
on. Even if num_threads clauses or if clauses are used, such constructs
can be executed in SPMD mode.
llvm-svn: 358595
diff --git a/clang/lib/CodeGen/CGOpenMPRuntime.cpp b/clang/lib/CodeGen/CGOpenMPRuntime.cpp
index da4e355..1e1251e 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntime.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntime.cpp
@@ -6657,6 +6657,47 @@
CGF.getContext(), CS->getCapturedStmt());
if (const auto *Dir = dyn_cast_or_null<OMPExecutableDirective>(Child)) {
if (isOpenMPParallelDirective(Dir->getDirectiveKind())) {
+ llvm::Value *NumThreads = nullptr;
+ llvm::Value *CondVal = nullptr;
+ // Handle if clause. If if clause present, the number of threads is
+ // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1.
+ if (Dir->hasClausesOfKind<OMPIfClause>()) {
+ CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
+ CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
+ const OMPIfClause *IfClause = nullptr;
+ for (const auto *C : Dir->getClausesOfKind<OMPIfClause>()) {
+ if (C->getNameModifier() == OMPD_unknown ||
+ C->getNameModifier() == OMPD_parallel) {
+ IfClause = C;
+ break;
+ }
+ }
+ if (IfClause) {
+ const Expr *Cond = IfClause->getCondition();
+ bool Result;
+ if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) {
+ if (!Result)
+ return CGF.Builder.getInt32(1);
+ } else {
+ CodeGenFunction::LexicalScope Scope(CGF, Cond->getSourceRange());
+ if (const auto *PreInit =
+ cast_or_null<DeclStmt>(IfClause->getPreInitStmt())) {
+ for (const auto *I : PreInit->decls()) {
+ if (!I->hasAttr<OMPCaptureNoInitAttr>()) {
+ CGF.EmitVarDecl(cast<VarDecl>(*I));
+ } else {
+ CodeGenFunction::AutoVarEmission Emission =
+ CGF.EmitAutoVarAlloca(cast<VarDecl>(*I));
+ CGF.EmitAutoVarCleanups(Emission);
+ }
+ }
+ }
+ CondVal = CGF.EvaluateExprAsBool(Cond);
+ }
+ }
+ }
+ // Check the value of num_threads clause iff if clause was not specified
+ // or is not evaluated to false.
if (Dir->hasClausesOfKind<OMPNumThreadsClause>()) {
CGOpenMPInnerExprInfo CGInfo(CGF, *CS);
CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(CGF, &CGInfo);
@@ -6676,19 +6717,23 @@
}
}
}
- llvm::Value *NumThreads =
- CGF.EmitScalarExpr(NumThreadsClause->getNumThreads());
+ NumThreads = CGF.EmitScalarExpr(NumThreadsClause->getNumThreads());
NumThreads = CGF.Builder.CreateIntCast(NumThreads, CGF.Int32Ty,
- /*IsSigned=*/true);
- return DefaultThreadLimitVal
- ? CGF.Builder.CreateSelect(
- CGF.Builder.CreateICmpULT(DefaultThreadLimitVal,
- NumThreads),
- DefaultThreadLimitVal, NumThreads)
- : NumThreads;
+ /*IsSigned=*/false);
+ if (DefaultThreadLimitVal)
+ NumThreads = CGF.Builder.CreateSelect(
+ CGF.Builder.CreateICmpULT(DefaultThreadLimitVal, NumThreads),
+ DefaultThreadLimitVal, NumThreads);
+ } else {
+ NumThreads = DefaultThreadLimitVal ? DefaultThreadLimitVal
+ : CGF.Builder.getInt32(0);
}
- return DefaultThreadLimitVal ? DefaultThreadLimitVal
- : CGF.Builder.getInt32(0);
+ // Process condition of the if clause.
+ if (CondVal) {
+ NumThreads = CGF.Builder.CreateSelect(CondVal, NumThreads,
+ CGF.Builder.getInt32(1));
+ }
+ return NumThreads;
}
if (isOpenMPSimdDirective(Dir->getDirectiveKind()))
return CGF.Builder.getInt32(1);
@@ -6748,7 +6793,7 @@
llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
ThreadLimitVal =
- Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/true);
+ Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false);
}
if (isOpenMPTeamsDirective(Dir->getDirectiveKind()) &&
!isOpenMPDistributeDirective(Dir->getDirectiveKind())) {
@@ -6775,7 +6820,7 @@
llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
ThreadLimitVal =
- Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/true);
+ Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false);
}
const CapturedStmt *CS = D.getInnermostCapturedStmt();
if (llvm::Value *NumThreads = getNumThreads(CGF, CS, ThreadLimitVal))
@@ -6798,21 +6843,45 @@
llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
ThreadLimitVal =
- Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/true);
+ Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false);
}
return getNumThreads(CGF, D.getInnermostCapturedStmt(), ThreadLimitVal);
case OMPD_target_parallel:
case OMPD_target_parallel_for:
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
- case OMPD_target_teams_distribute_parallel_for_simd:
+ case OMPD_target_teams_distribute_parallel_for_simd: {
+ llvm::Value *CondVal = nullptr;
+ // Handle if clause. If if clause present, the number of threads is
+ // calculated as <cond> ? (<numthreads> ? <numthreads> : 0 ) : 1.
+ if (D.hasClausesOfKind<OMPIfClause>()) {
+ const OMPIfClause *IfClause = nullptr;
+ for (const auto *C : D.getClausesOfKind<OMPIfClause>()) {
+ if (C->getNameModifier() == OMPD_unknown ||
+ C->getNameModifier() == OMPD_parallel) {
+ IfClause = C;
+ break;
+ }
+ }
+ if (IfClause) {
+ const Expr *Cond = IfClause->getCondition();
+ bool Result;
+ if (Cond->EvaluateAsBooleanCondition(Result, CGF.getContext())) {
+ if (!Result)
+ return Bld.getInt32(1);
+ } else {
+ CodeGenFunction::RunCleanupsScope Scope(CGF);
+ CondVal = CGF.EvaluateExprAsBool(Cond);
+ }
+ }
+ }
if (D.hasClausesOfKind<OMPThreadLimitClause>()) {
CodeGenFunction::RunCleanupsScope ThreadLimitScope(CGF);
const auto *ThreadLimitClause = D.getSingleClause<OMPThreadLimitClause>();
llvm::Value *ThreadLimit = CGF.EmitScalarExpr(
ThreadLimitClause->getThreadLimit(), /*IgnoreResultAssign=*/true);
ThreadLimitVal =
- Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/true);
+ Bld.CreateIntCast(ThreadLimit, CGF.Int32Ty, /*IsSigned=*/false);
}
if (D.hasClausesOfKind<OMPNumThreadsClause>()) {
CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF);
@@ -6820,14 +6889,19 @@
llvm::Value *NumThreads = CGF.EmitScalarExpr(
NumThreadsClause->getNumThreads(), /*IgnoreResultAssign=*/true);
NumThreadsVal =
- Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*IsSigned=*/true);
+ Bld.CreateIntCast(NumThreads, CGF.Int32Ty, /*IsSigned=*/false);
ThreadLimitVal = ThreadLimitVal
? Bld.CreateSelect(Bld.CreateICmpULT(NumThreadsVal,
ThreadLimitVal),
NumThreadsVal, ThreadLimitVal)
: NumThreadsVal;
}
- return ThreadLimitVal ? ThreadLimitVal : Bld.getInt32(0);
+ if (!ThreadLimitVal)
+ ThreadLimitVal = Bld.getInt32(0);
+ if (CondVal)
+ return Bld.CreateSelect(CondVal, ThreadLimitVal, Bld.getInt32(1));
+ return ThreadLimitVal;
+ }
case OMPD_target_teams_distribute_simd:
case OMPD_target_simd:
return Bld.getInt32(1);
diff --git a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
index 0085b01..16f6c12 100644
--- a/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
+++ b/clang/lib/CodeGen/CGOpenMPRuntimeNVPTX.cpp
@@ -714,24 +714,6 @@
: CGOpenMPRuntimeNVPTX::Generic;
}
-/// Check if the parallel directive has an 'if' clause with non-constant or
-/// false condition.
-static bool hasParallelIfClause(ASTContext &Ctx,
- const OMPExecutableDirective &D,
- bool StandaloneParallel) {
- for (const auto *C : D.getClausesOfKind<OMPIfClause>()) {
- OpenMPDirectiveKind NameModifier = C->getNameModifier();
- if (NameModifier != OMPD_parallel &&
- (!StandaloneParallel || NameModifier != OMPD_unknown))
- continue;
- const Expr *Cond = C->getCondition();
- bool Result;
- if (!Cond->EvaluateAsBooleanCondition(Result, Ctx) || !Result)
- return true;
- }
- return false;
-}
-
/// Check for inner (nested) SPMD construct, if any
static bool hasNestedSPMDDirective(ASTContext &Ctx,
const OMPExecutableDirective &D) {
@@ -745,8 +727,7 @@
OpenMPDirectiveKind DKind = NestedDir->getDirectiveKind();
switch (D.getDirectiveKind()) {
case OMPD_target:
- if (isOpenMPParallelDirective(DKind) &&
- !hasParallelIfClause(Ctx, *NestedDir, /*StandaloneParallel=*/true))
+ if (isOpenMPParallelDirective(DKind))
return true;
if (DKind == OMPD_teams) {
Body = NestedDir->getInnermostCapturedStmt()->IgnoreContainers(
@@ -757,15 +738,13 @@
if (const auto *NND =
dyn_cast_or_null<OMPExecutableDirective>(ChildStmt)) {
DKind = NND->getDirectiveKind();
- if (isOpenMPParallelDirective(DKind) &&
- !hasParallelIfClause(Ctx, *NND, /*StandaloneParallel=*/true))
+ if (isOpenMPParallelDirective(DKind))
return true;
}
}
return false;
case OMPD_target_teams:
- return isOpenMPParallelDirective(DKind) &&
- !hasParallelIfClause(Ctx, *NestedDir, /*StandaloneParallel=*/true);
+ return isOpenMPParallelDirective(DKind);
case OMPD_target_simd:
case OMPD_target_parallel:
case OMPD_target_parallel_for:
@@ -839,10 +818,10 @@
case OMPD_target_parallel_for_simd:
case OMPD_target_teams_distribute_parallel_for:
case OMPD_target_teams_distribute_parallel_for_simd:
- return !hasParallelIfClause(Ctx, D, /*StandaloneParallel=*/false);
case OMPD_target_simd:
- case OMPD_target_teams_distribute:
case OMPD_target_teams_distribute_simd:
+ return true;
+ case OMPD_target_teams_distribute:
return false;
case OMPD_parallel:
case OMPD_for: