[Clang] Add __CLANG_GPU_DISABLE_MATH_WRAPPERS macro for offloading math (llvm#98234)

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.

Author: jhuber6

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__

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__

clang/test/Headers/gpu_disabled_math.cpp

@@ -0,0 +1,41 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// RUN: %clang_cc1 -include __clang_hip_runtime_wrapper.h      \
+// RUN:   -internal-isystem %S/../../lib/Headers/cuda_wrappers \
+// RUN:   -internal-isystem %S/Inputs/include \
+// RUN:   -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown \
+// RUN:   -target-cpu gfx906 -emit-llvm %s -fcuda-is-device -o - \
+// RUN:   -D __CLANG_GPU_DISABLE_MATH_WRAPPERS | FileCheck -check-prefix=AMDGPU %s
+
+// RUN: %clang_cc1 -include __clang_cuda_runtime_wrapper.h      \
+// RUN:   -internal-isystem %S/../../lib/Headers/cuda_wrappers \
+// RUN:   -internal-isystem %S/Inputs/include \
+// RUN:   -triple nvptx64-nvidia-cuda -aux-triple x86_64-unknown-unknown \
+// RUN:   -target-cpu sm_90 -emit-llvm %s -fcuda-is-device -o - \
+// RUN:   -D __CLANG_GPU_DISABLE_MATH_WRAPPERS | FileCheck -check-prefix=NVPTX %s
+
+extern "C" double sin(double x);
+
+// AMDGPU-LABEL: define dso_local noundef double @_Z3food(
+// AMDGPU-SAME: double noundef [[X:%.*]]) #[[ATTR0:[0-9]+]] {
+// AMDGPU-NEXT:  [[ENTRY:.*:]]
+// AMDGPU-NEXT:    [[RETVAL:%.*]] = alloca double, align 8, addrspace(5)
+// AMDGPU-NEXT:    [[X_ADDR:%.*]] = alloca double, align 8, addrspace(5)
+// AMDGPU-NEXT:    [[RETVAL_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[RETVAL]] to ptr
+// AMDGPU-NEXT:    [[X_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[X_ADDR]] to ptr
+// AMDGPU-NEXT:    store double [[X]], ptr [[X_ADDR_ASCAST]], align 8
+// AMDGPU-NEXT:    [[TMP0:%.*]] = load double, ptr [[X_ADDR_ASCAST]], align 8
+// AMDGPU-NEXT:    [[TMP1:%.*]] = call double @llvm.sin.f64(double [[TMP0]])
+// AMDGPU-NEXT:    ret double [[TMP1]]
+//
+// NVPTX-LABEL: define dso_local noundef double @_Z3food(
+// NVPTX-SAME: double noundef [[X:%.*]]) #[[ATTR0:[0-9]+]] {
+// NVPTX-NEXT:  [[ENTRY:.*:]]
+// NVPTX-NEXT:    [[X_ADDR:%.*]] = alloca double, align 8
+// NVPTX-NEXT:    store double [[X]], ptr [[X_ADDR]], align 8
+// NVPTX-NEXT:    [[TMP0:%.*]] = load double, ptr [[X_ADDR]], align 8
+// NVPTX-NEXT:    [[TMP1:%.*]] = call double @llvm.sin.f64(double [[TMP0]])
+// NVPTX-NEXT:    ret double [[TMP1]]
+//
+double foo(double x) {
+  return sin(x);
+}