[clang] [Clang] Add `__CLANG_GPU_DISABLE_MATH_WRAPPERS` macro for offloading math (PR #98234)

[via cfe-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-clang

Author: Joseph Huber (jhuber6)

<details>
<summary>Changes</summary>

Summary:
Currently we replace all math calls with vendor specific ones. This
patch introduces a macro `__CLANG_GPU_DISABLE_MATH_WRAPPERS` that when
defined will disable this.

I went this route instead of a flag for two reasons. One, I think we
have too many flags as is, and we already have `-nogpuinc` to cover
disabling these wrappers entirely, so this would be a really specific
subset of that. Second, these math headers aren't easily decoupled by
simply not including a single header from the clang driver layer.
There's the cmath and the regular math forward declares it would disable
as well.

Note, this currently causes errors because the GPU `libm` doesn't have
`powi`, that's an NVIDIA extension I'll add to LLVM libm.


---
Full diff: https://github.com/llvm/llvm-project/pull/98234.diff


2 Files Affected:

- (modified) clang/lib/Headers/__clang_cuda_math.h (+5) 
- (modified) clang/lib/Headers/__clang_hip_math.h (+5) 


``````````diff
diff --git a/clang/lib/Headers/__clang_cuda_math.h b/clang/lib/Headers/__clang_cuda_math.h
index 0401916506866..44c6e9a4e48d1 100644
--- a/clang/lib/Headers/__clang_cuda_math.h
+++ b/clang/lib/Headers/__clang_cuda_math.h
@@ -12,6 +12,10 @@
 #error "This file is for CUDA compilation only."
 #endif
 
+// The __CLANG_GPU_DISABLE_MATH_WRAPPERS macro provides a way to let standard
+// libcalls reach the link step instead of being eagerly replaced.
+#ifndef __CLANG_GPU_DISABLE_MATH_WRAPPERS
+
 #ifndef __OPENMP_NVPTX__
 #if CUDA_VERSION < 9000
 #error This file is intended to be used with CUDA-9+ only.
@@ -345,4 +349,5 @@ __DEVICE__ float ynf(int __a, float __b) { return __nv_ynf(__a, __b); }
 #pragma pop_macro("__DEVICE_VOID__")
 #pragma pop_macro("__FAST_OR_SLOW")
 
+#endif // __CLANG_GPU_DISABLE_MATH_WRAPPERS
 #endif // __CLANG_CUDA_MATH_H__
diff --git a/clang/lib/Headers/__clang_hip_math.h b/clang/lib/Headers/__clang_hip_math.h
index 11e1e7d032586..8468751d9de26 100644
--- a/clang/lib/Headers/__clang_hip_math.h
+++ b/clang/lib/Headers/__clang_hip_math.h
@@ -13,6 +13,10 @@
 #error "This file is for HIP and OpenMP AMDGCN device compilation only."
 #endif
 
+// The __CLANG_GPU_DISABLE_MATH_WRAPPERS macro provides a way to let standard
+// libcalls reach the link step instead of being eagerly replaced.
+#ifndef __CLANG_GPU_DISABLE_MATH_WRAPPERS
+
 #if !defined(__HIPCC_RTC__)
 #include <limits.h>
 #include <stdint.h>
@@ -1321,4 +1325,5 @@ __host__ inline static int max(int __arg1, int __arg2) {
 #pragma pop_macro("__RETURN_TYPE")
 #pragma pop_macro("__FAST_OR_SLOW")
 
+#endif // __CLANG_GPU_DISABLE_MATH_WRAPPERS
 #endif // __CLANG_HIP_MATH_H__

``````````

</details>


https://github.com/llvm/llvm-project/pull/98234