[llvm-branch-commits] [clang] clang/AMDGPU: Remove driver restriction on --gpu-max-threads-per-block (PR #204864)

[Matt Arsenault via llvm-branch-commits]

https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/204864

>From 143f364ed8df9697e08458507f32bf6d5cc68255 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Fri, 19 Jun 2026 17:01:05 +0200
Subject: [PATCH] clang/AMDGPU: Remove driver restriction on
 --gpu-max-threads-per-block

Previously this flag was only handled for HIP, and would produce an unused
argument warning. There is a custom warning produced by cc1 that the
argument isn't supported, but practically speaking that was unreachable
due to not forwarding the argument. Also add a test for the untested warning.
Also use a simpler method for forwarding the flag to cc1.
---
 clang/lib/Driver/ToolChains/AMDGPU.cpp             | 10 ++--------
 .../amdgpu-openmp-gpu-max-threads-per-block.c      |  6 ++++++
 .../openmp-warn-gpu-max-threads-per-block.c        | 14 ++++++++++++++
 3 files changed, 22 insertions(+), 8 deletions(-)
 create mode 100644 clang/test/Driver/amdgpu-openmp-gpu-max-threads-per-block.c
 create mode 100644 clang/test/Frontend/openmp-warn-gpu-max-threads-per-block.c

diff --git a/clang/lib/Driver/ToolChains/AMDGPU.cpp b/clang/lib/Driver/ToolChains/AMDGPU.cpp
index 64c64e68ee8ea..5f8b1b8d8446d 100644
--- a/clang/lib/Driver/ToolChains/AMDGPU.cpp
+++ b/clang/lib/Driver/ToolChains/AMDGPU.cpp
@@ -878,16 +878,10 @@ void AMDGPUToolChain::addClangTargetOptions(
       if (DriverArgs.hasArgNoClaim(options::OPT_hipstdpar))
         CC1Args.append({"-mllvm", "-amdgpu-enable-hipstdpar"});
     }
-
-    // FIXME: This should not depend on HIP
-    StringRef MaxThreadsPerBlock =
-        DriverArgs.getLastArgValue(options::OPT_gpu_max_threads_per_block_EQ);
-    if (!MaxThreadsPerBlock.empty()) {
-      CC1Args.push_back(DriverArgs.MakeArgString(
-          Twine("--gpu-max-threads-per-block=") + MaxThreadsPerBlock));
-    }
   }
 
+  DriverArgs.AddLastArg(CC1Args, options::OPT_gpu_max_threads_per_block_EQ);
+
   // Default to "hidden" visibility, as object level linking will not be
   // supported for the foreseeable future.
   // TODO: remove the SPIR-V bypass once it can encode (hidden) visibility.
diff --git a/clang/test/Driver/amdgpu-openmp-gpu-max-threads-per-block.c b/clang/test/Driver/amdgpu-openmp-gpu-max-threads-per-block.c
new file mode 100644
index 0000000000000..ba5dd4c24a0bd
--- /dev/null
+++ b/clang/test/Driver/amdgpu-openmp-gpu-max-threads-per-block.c
@@ -0,0 +1,6 @@
+// Test that --gpu-max-threads-per-block is not ignored by openmp.
+// RUN: %clang -### --target=x86_64-unknown-linux-gnu -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa \
+// RUN:   --offload-arch=gfx906 -nogpulib --gpu-max-threads-per-block=256 %s 2>&1 | FileCheck %s
+
+// CHECK: "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "--gpu-max-threads-per-block=256"
diff --git a/clang/test/Frontend/openmp-warn-gpu-max-threads-per-block.c b/clang/test/Frontend/openmp-warn-gpu-max-threads-per-block.c
new file mode 100644
index 0000000000000..3efdc9ac2b5a0
--- /dev/null
+++ b/clang/test/Frontend/openmp-warn-gpu-max-threads-per-block.c
@@ -0,0 +1,14 @@
+// REQUIRES: amdgpu-registered-target
+
+// Test that --gpu-max-threads-per-block warns when used outside of HIP.
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa --gpu-max-threads-per-block=1024 \
+// RUN:     -fsyntax-only -verify=warn %s
+
+// Test that no warning is emitted for HIP.
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -x hip -fcuda-is-device \
+// RUN:     --gpu-max-threads-per-block=1024 -fsyntax-only -verify=hip %s
+
+// warn-warning@*{{'--gpu-max-threads-per-block=1024' is ignored since it is only supported for HIP}}
+// hip-no-diagnostics
+
+void f(void) {}