[Clang] Fix __cpuidex conflict with CUDA #152556
Conversation
The landing of llvm#126324 made it so that __has_builtin returns false for aux triple builtins. CUDA offloading can sometimes compile where the host is in the aux triple (ie x86_64). This patch explicitly carves out NVPTX so that we do not run into redefinition errors.
llvmbot
added
clang
|
@llvm/pr-subscribers-backend-x86 @llvm/pr-subscribers-clang Author: Aiden Grossman (boomanaiden154) ChangesThe landing of #126324 made it so that __has_builtin returns false for aux triple builtins. CUDA offloading can sometimes compile where the host is in the aux triple (ie x86_64). This patch explicitly carves out NVPTX so that we do not run into redefinition errors. Full diff: https://github.com/llvm/llvm-project/pull/152556.diff 2 Files Affected:
diff --git a/clang/lib/Headers/cpuid.h b/clang/lib/Headers/cpuid.h
index 52addb7bfa856..ce8c79e77dc18 100644
--- a/clang/lib/Headers/cpuid.h
+++ b/clang/lib/Headers/cpuid.h
@@ -345,10 +345,15 @@ static __inline int __get_cpuid_count (unsigned int __leaf,
// In some configurations, __cpuidex is defined as a builtin (primarily
// -fms-extensions) which will conflict with the __cpuidex definition below.
#if !(__has_builtin(__cpuidex))
+// In some cases, offloading will set the host as the aux triple and define the
+// builtin. Given __has_builtin does not detect builtins on aux triples, we need
+// to explicitly check for some offloading cases.
+#ifndef __NVPTX__
static __inline void __cpuidex(int __cpu_info[4], int __leaf, int __subleaf) {
__cpuid_count(__leaf, __subleaf, __cpu_info[0], __cpu_info[1], __cpu_info[2],
__cpu_info[3]);
}
#endif
+#endif
#endif /* __CPUID_H */
diff --git a/clang/test/Headers/__cpuidex_conflict.c b/clang/test/Headers/__cpuidex_conflict.c
index 74f45327de2bb..d14ef293e586d 100644
--- a/clang/test/Headers/__cpuidex_conflict.c
+++ b/clang/test/Headers/__cpuidex_conflict.c
@@ -5,6 +5,7 @@
// Ensure that we do not run into conflicts when offloading.
// RUN: %clang_cc1 %s -DIS_STATIC=static -ffreestanding -fopenmp -fopenmp-is-target-device -aux-triple x86_64-unknown-linux-gnu
+// RUN: %clang_cc1 -DIS_STATIC="" -triple nvptx64-nvidia-cuda -aux-triple x86_64-unknown-linux-gnu -aux-target-cpu x86-64 -fcuda-is-device -internal-isystem /home/gha/llvm-project/build/lib/clang/22/include -x cuda %s -o -
typedef __SIZE_TYPE__ size_t;
|
|
This is a specific carve-out which I'm not a huge fan of, but it doesn't seem like there's any systematic way to identify builtins tied to the aux triple, so I think this makes sense for now. |
| // In some cases, offloading will set the host as the aux triple and define the | ||
| // builtin. Given __has_builtin does not detect builtins on aux triples, we need | ||
| // to explicitly check for some offloading cases. | ||
| #ifndef __NVPTX__ |
There was a problem hiding this comment.
Thanks so much for the quick investigation!
I'm not really familiar with CUDA offloading but if the cc1 command is cc1 -triple nvptx64-nvidia-cuda -aux-triple x86_64-unknown-linux-gnu, __has_builtin(__cpuidex) returning false seems correct, so I would think the actual root cause is that cpuidex is getting defined for CUDA with -triple nvptx64-nvidia-cuda -aux-triple x86_64-unknown-linux-gnu.
If so, I'm happy to have this workaround for now but we might want to file a tracker somewhere for the root cause to be investigated.
There was a problem hiding this comment.
Good point. I've filed #152558 to track this.
The landing of #126324 made it so that __has_builtin returns false for aux triple builtins. CUDA offloading can sometimes compile where the host is in the aux triple (ie x86_64). This patch explicitly carves out NVPTX so that we do not run into redefinition errors.