[clang] [clang][OpenMP] Add OpenMP GPU optimization flag framework (PR #178914)
via cfe-commits
cfe-commits at lists.llvm.org
Fri Jan 30 08:32:22 PST 2026
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-clang-driver
Author: Nicole Aschenbrenner (nicebert)
<details>
<summary>Changes</summary>
This patch series adds a framework of OpenMP GPU optimization flags to enable more efficient code generation for GPU offloading. The series consists of three patches:
1. Add negative flag variants (-fno-*) for existing assume options to allow explicit disabling of optimizations.
2. Add -fopenmp-target-ignore-env-vars flag to indicate that OpenMP runtime can ignore environment variables during code generation, enabling optimizations like skipping runtime checks and eliminating conditional branches.
3. Add -fopenmp-target-fast meta-flag that implies the above optimization flags. This convenience flag is automatically enabled by -Ofast and provides a simple interface for aggressive GPU optimizations.
The flags benefit all GPU targets (NVPTX, AMDGPU, Intel GPU) by providing a standard way to enable common GPU optimization patterns. Individual flags can be selectively overridden while keeping others enabled.
---
Full diff: https://github.com/llvm/llvm-project/pull/178914.diff
4 Files Affected:
- (modified) clang/include/clang/Basic/LangOptions.def (+1)
- (modified) clang/include/clang/Options/Options.td (+19)
- (modified) clang/lib/Driver/ToolChains/Clang.cpp (+48-2)
- (added) clang/test/Driver/openmp-target-fast-flag.c (+48)
``````````diff
diff --git a/clang/include/clang/Basic/LangOptions.def b/clang/include/clang/Basic/LangOptions.def
index ba12e522f331f..a2cc47a6528c4 100644
--- a/clang/include/clang/Basic/LangOptions.def
+++ b/clang/include/clang/Basic/LangOptions.def
@@ -239,6 +239,7 @@ LANGOPT(OpenMPThreadSubscription , 1, 0, NotCompatible, "Assume work-shared loo
LANGOPT(OpenMPTeamSubscription , 1, 0, NotCompatible, "Assume distributed loops do not have more iterations than participating teams.")
LANGOPT(OpenMPNoThreadState , 1, 0, NotCompatible, "Assume that no thread in a parallel region will modify an ICV.")
LANGOPT(OpenMPNoNestedParallelism , 1, 0, NotCompatible, "Assume that no thread in a parallel region will encounter a parallel region")
+LANGOPT(OpenMPTargetIgnoreEnvVars, 1, 0, NotCompatible, "Assume that the OpenMP runtime can ignore environment variables during code generation for GPU offload")
LANGOPT(OpenMPOffloadMandatory , 1, 0, NotCompatible, "Assert that offloading is mandatory and do not create a host fallback.")
LANGOPT(OpenMPForceUSM , 1, 0, NotCompatible, "Enable OpenMP unified shared memory mode via compiler.")
LANGOPT(NoGPULib , 1, 0, NotCompatible, "Indicate a build without the standard GPU libraries.")
diff --git a/clang/include/clang/Options/Options.td b/clang/include/clang/Options/Options.td
index 421208a812bbc..2080cfa5fc091 100644
--- a/clang/include/clang/Options/Options.td
+++ b/clang/include/clang/Options/Options.td
@@ -3986,9 +3986,13 @@ def fno_openmp_assume_threads_oversubscription : Flag<["-"], "fno-openmp-assume-
def fopenmp_assume_no_thread_state : Flag<["-"], "fopenmp-assume-no-thread-state">,
HelpText<"Assert no thread in a parallel region modifies an ICV">,
MarshallingInfoFlag<LangOpts<"OpenMPNoThreadState">>;
+def fno_openmp_assume_no_thread_state : Flag<["-"], "fno-openmp-assume-no-thread-state">,
+ HelpText<"Assert that a thread in a parallel region may modify an ICV">;
def fopenmp_assume_no_nested_parallelism : Flag<["-"], "fopenmp-assume-no-nested-parallelism">,
HelpText<"Assert no nested parallel regions in the GPU">,
MarshallingInfoFlag<LangOpts<"OpenMPNoNestedParallelism">>;
+def fno_openmp_assume_no_nested_parallelism : Flag<["-"], "fno-openmp-assume-no-nested-parallelism">,
+ HelpText<"Assert that a nested parallel region may be used in the GPU">;
} // let Group = f_Group
} // let Visibility = [ClangOption, CC1Option, FC1Option]
@@ -4012,6 +4016,21 @@ def fopenmp_target_new_runtime : Flag<["-"], "fopenmp-target-new-runtime">,
Group<f_Group>, Flags<[HelpHidden]>, Visibility<[ClangOption, CC1Option]>;
def fno_openmp_target_new_runtime : Flag<["-"], "fno-openmp-target-new-runtime">,
Group<f_Group>, Flags<[HelpHidden]>, Visibility<[ClangOption, CC1Option]>;
+def fopenmp_target_ignore_env_vars : Flag<["-"], "fopenmp-target-ignore-env-vars">,
+ Group<f_Group>, Flags<[NoArgumentUnused, HelpHidden]>,
+ Visibility<[ClangOption, CC1Option]>,
+ HelpText<"Assume that the OpenMP runtime can ignore environment variables during code generation for GPU offload">,
+ MarshallingInfoFlag<LangOpts<"OpenMPTargetIgnoreEnvVars">>;
+def fno_openmp_target_ignore_env_vars : Flag<["-"], "fno-openmp-target-ignore-env-vars">,
+ Group<f_Group>, Flags<[NoArgumentUnused, HelpHidden]>,
+ Visibility<[ClangOption, CC1Option]>;
+def fopenmp_target_fast : Flag<["-"], "fopenmp-target-fast">,
+ Group<f_Group>, Flags<[NoArgumentUnused, HelpHidden]>,
+ Visibility<[ClangOption, CC1Option]>,
+ HelpText<"Convenience flag to enable aggressive OpenMP GPU optimizations">;
+def fno_openmp_target_fast : Flag<["-"], "fno-openmp-target-fast">,
+ Group<f_Group>, Flags<[NoArgumentUnused, HelpHidden]>,
+ Visibility<[ClangOption, CC1Option]>;
defm openmp_optimistic_collapse : BoolFOption<"openmp-optimistic-collapse",
LangOpts<"OpenMPOptimisticCollapse">, DefaultFalse,
PosFlag<SetTrue, [], [ClangOption, CC1Option]>,
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index 0293b04217673..4c90d7cd33d4e 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -6675,6 +6675,11 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
options::OPT_fno_offload_via_llvm, false) &&
(JA.isDeviceOffloading(Action::OFK_None) ||
JA.isDeviceOffloading(Action::OFK_OpenMP))) {
+
+ // Determine if target-fast optimizations should be enabled
+ bool TargetFastUsed = Args.hasFlag(options::OPT_fopenmp_target_fast,
+ options::OPT_fno_openmp_target_fast,
+ OFastEnabled);
switch (D.getOpenMPRuntime(Args)) {
case Driver::OMPRT_OMP:
case Driver::OMPRT_IOMP5:
@@ -6725,10 +6730,51 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
options::OPT_fno_openmp_assume_threads_oversubscription,
/*Default=*/false))
CmdArgs.push_back("-fopenmp-assume-threads-oversubscription");
- if (Args.hasArg(options::OPT_fopenmp_assume_no_thread_state))
+
+ // Handle -fopenmp-target-fast
+ if (Arg *A = Args.getLastArg(options::OPT_fopenmp_target_fast,
+ options::OPT_fno_openmp_target_fast)) {
+ if (A->getOption().matches(options::OPT_fopenmp_target_fast))
+ CmdArgs.push_back("-fopenmp-target-fast");
+ else
+ CmdArgs.push_back("-fno-openmp-target-fast");
+ } else if (OFastEnabled) {
+ CmdArgs.push_back("-fopenmp-target-fast");
+ }
+
+ // Handle -fopenmp-target-ignore-env-vars (implied by target-fast)
+ if (Arg *A = Args.getLastArg(options::OPT_fopenmp_target_ignore_env_vars,
+ options::OPT_fno_openmp_target_ignore_env_vars)) {
+ if (A->getOption().matches(options::OPT_fopenmp_target_ignore_env_vars))
+ CmdArgs.push_back("-fopenmp-target-ignore-env-vars");
+ else
+ CmdArgs.push_back("-fno-openmp-target-ignore-env-vars");
+ } else if (TargetFastUsed) {
+ CmdArgs.push_back("-fopenmp-target-ignore-env-vars");
+ }
+
+ // Handle -fopenmp-assume-no-thread-state (implied by target-fast)
+ if (Arg *A = Args.getLastArg(options::OPT_fopenmp_assume_no_thread_state,
+ options::OPT_fno_openmp_assume_no_thread_state)) {
+ if (A->getOption().matches(options::OPT_fopenmp_assume_no_thread_state))
+ CmdArgs.push_back("-fopenmp-assume-no-thread-state");
+ else
+ CmdArgs.push_back("-fno-openmp-assume-no-thread-state");
+ } else if (TargetFastUsed) {
CmdArgs.push_back("-fopenmp-assume-no-thread-state");
- if (Args.hasArg(options::OPT_fopenmp_assume_no_nested_parallelism))
+ }
+
+ // Handle -fopenmp-assume-no-nested-parallelism (implied by target-fast)
+ if (Arg *A = Args.getLastArg(options::OPT_fopenmp_assume_no_nested_parallelism,
+ options::OPT_fno_openmp_assume_no_nested_parallelism)) {
+ if (A->getOption().matches(options::OPT_fopenmp_assume_no_nested_parallelism))
+ CmdArgs.push_back("-fopenmp-assume-no-nested-parallelism");
+ else
+ CmdArgs.push_back("-fno-openmp-assume-no-nested-parallelism");
+ } else if (TargetFastUsed) {
CmdArgs.push_back("-fopenmp-assume-no-nested-parallelism");
+ }
+
if (Args.hasArg(options::OPT_fopenmp_offload_mandatory))
CmdArgs.push_back("-fopenmp-offload-mandatory");
if (Args.hasArg(options::OPT_fopenmp_force_usm))
diff --git a/clang/test/Driver/openmp-target-fast-flag.c b/clang/test/Driver/openmp-target-fast-flag.c
new file mode 100644
index 0000000000000..703df0efaeedd
--- /dev/null
+++ b/clang/test/Driver/openmp-target-fast-flag.c
@@ -0,0 +1,48 @@
+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a %s -O0 2>&1 \
+// RUN: | FileCheck -check-prefixes=DefaultTFast,DefaultEnV,DefaultTState,DefaultNoNestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -O0 -fopenmp-target-fast %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=TFast,EnV,TState,NestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -O3 %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=O3,DefaultTFast,DefaultEnV,DefaultTState,DefaultNoNestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -O3 -fno-openmp-target-fast %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=O3,NoTFast,DefaultEnV,DefaultTState,DefaultNoNestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -Ofast %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=OFast,TFast,EnV,TState,NestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -Ofast -fno-openmp-target-fast %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=OFast,NoTFast,DefaultEnV,DefaultTState,DefaultNoNestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -fopenmp-target-fast -fno-openmp-target-ignore-env-vars %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=TFast,NoEnV,TState,NestParallel %s
+
+// RUN: %clang -### -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx90a -O0 -fno-openmp-target-fast -fopenmp-target-fast %s 2>&1 \
+// RUN: | FileCheck -check-prefixes=TFast,EnV,TState,NestParallel %s
+
+// O3: -O3
+// OFast: -Ofast
+
+// TFast: "-fopenmp-target-fast"
+// TFast-NOT: "-fno-openmp-target-fast"
+// NoTFast: "-fno-openmp-target-fast"
+// NoTFast-NOT: "-fopenmp-target-fast"
+// DefaultTFast-NOT: {{"-f(no-)?openmp-target-fast"}}
+
+// EnV: "-fopenmp-target-ignore-env-vars"
+// EnV-NOT: "-fno-openmp-target-ignore-env-vars"
+// NoEnV: "-fno-openmp-target-ignore-env-vars"
+// NoEnV-NOT: "-fopenmp-target-ignore-env-vars"
+// DefaultEnV-NOT: {{"-f(no-)?openmp-target-ignore-env-vars"}}
+
+// TState: "-fopenmp-assume-no-thread-state"
+// TState-NOT: "-fno-openmp-assume-no-thread-state"
+// DefaultTState-NOT: {{"-f(no-)?openmp-assume-no-thread-state"}}
+
+// NestParallel: "-fopenmp-assume-no-nested-parallelism"
+// NestParallel-NOT: "-fno-openmp-assume-no-nested-parallelism"
+// DefaultNoNestParallel-NOT: {{"-f(-no-)?openmp-assume-no-nested-parallelism"}}
``````````
</details>
https://github.com/llvm/llvm-project/pull/178914
More information about the cfe-commits
mailing list