[OpenMP][AMDGPU] Enable use of abs labs and llabs math functions in C code

[doru1004]

Summary

Enable use of abs labs and llabs math functions in C code. Before this patch these math functions could only be used in OpenMP target regions in C++.

[jdoerfert]

I'm not 100% sure if this was excluded on purpose or not. FWIW, in C, these functions are not defined in math.h (https://en.cppreference.com/w/c/numeric/math/abs), but in C++ they are (https://en.cppreference.com/w/cpp/numeric/math/abs).

Once we enable libm_gpu it might not make a difference. For now, I'm OK with this if it doesn't break peoples code.

[arsenm]

In D139723#3984846, @jdoerfert wrote:

I'm not 100% sure if this was excluded on purpose or not. FWIW, in C, these functions are not defined in math.h (https://en.cppreference.com/w/c/numeric/math/abs), but in C++ they are (https://en.cppreference.com/w/cpp/numeric/math/abs).

Once we enable libm_gpu it might not make a difference. For now, I'm OK with this if it doesn't break peoples code.

Should there be a separate builtin header corresponding to stdlib.h?

[jdoerfert]

In D139723#3984880, @arsenm wrote:

In D139723#3984846, @jdoerfert wrote:

I'm not 100% sure if this was excluded on purpose or not. FWIW, in C, these functions are not defined in math.h (https://en.cppreference.com/w/c/numeric/math/abs), but in C++ they are (https://en.cppreference.com/w/cpp/numeric/math/abs).

Once we enable libm_gpu it might not make a difference. For now, I'm OK with this if it doesn't break peoples code.

Should there be a separate builtin header corresponding to stdlib.h?

Probably. And we want to finish the libc_gpu/libm_gpu work to provide the definitions late (in LTO mode)

[doru1004]

In D139723#3984846, @jdoerfert wrote:

I'm not 100% sure if this was excluded on purpose or not. FWIW, in C, these functions are not defined in math.h (https://en.cppreference.com/w/c/numeric/math/abs), but in C++ they are (https://en.cppreference.com/w/cpp/numeric/math/abs).

Once we enable libm_gpu it might not make a difference. For now, I'm OK with this if it doesn't break peoples code.

This fixes the case where we have a C file with offloading OpenMP in it that calls these functions. I'm not sure I see another way to do this in the current setup but I'm open to other solutions if this one breaks things unexpectedly.

[arsenm]

clang/lib/Headers/__clang_hip_math.h

145 ↗

(On Diff #481738)

Why not introduce __clang_hip_stdlib.h now?

[doru1004]

clang/lib/Headers/__clang_hip_math.h

145 ↗

(On Diff #481738)

Do we want to do this in this patch or as a separate patch? Is the work on libc_gpu/libm_gpu going to build up on these headers or will these headers be replaced in the process?

[arsenm]

clang/lib/Headers/__clang_hip_math.h

145 ↗

(On Diff #481738)

Yes. I think this kind of thing is hard to untangle once people start using it

[doru1004]

@arsenm I have factored out the abs/labs/llabs functions in a separate __clang_hip_stdlib.h file which could be included by C sources.

[doru1004]

Commit: 07ff3c5ccce68aed6c1a270b3f89ea14de7aa250

[mgorny]

In D139723#4004536, @doru1004 wrote:

Commit: 07ff3c5ccce68aed6c1a270b3f89ea14de7aa250

The added test fails on 32-bit platforms:

******************** TEST 'Clang :: Headers/amdgcn_openmp_device_math_c.c' FAILED ********************
Script:
--
: 'RUN: at line 2';   /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/clang -cc1 -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/../../../../lib/clang/16/include -nostdsysteminc -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/Inputs/include -x c -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -D__OFFLOAD_ARCH_gfx90a__ -emit-llvm-bc /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/amdgcn_openmp_device_math_c.c -o /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/test/Headers/Output/amdgcn_openmp_device_math_c.c.tmp-host.bc
: 'RUN: at line 3';   /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/clang -cc1 -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/../../../../lib/clang/16/include -nostdsysteminc -include __clang_hip_runtime_wrapper.h -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/../../lib/Headers/openmp_wrappers -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/Inputs/include -x c -fopenmp -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown -fopenmp-targets=amdgcn-amd-amdhsa -emit-llvm /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/amdgcn_openmp_device_math_c.c -fopenmp-is-device -fopenmp-host-ir-file-path /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/test/Headers/Output/amdgcn_openmp_device_math_c.c.tmp-host.bc -o - | /usr/lib/llvm/16/bin/FileCheck /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/amdgcn_openmp_device_math_c.c --check-prefixes=CHECK
--
Exit Code: 1

Command Output (stderr):
--
+ : 'RUN: at line 2'
+ /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/clang -cc1 -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/bin/../../../../lib/clang/16/include -nostdsysteminc -internal-isystem /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/Inputs/include -x c -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -D__OFFLOAD_ARCH_gfx90a__ -emit-llvm-bc /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/clang/test/Headers/amdgcn_openmp_device_math_c.c -o /var/tmp/portage/sys-devel/clang-16.0.0_pre20221225/work/x/y/clang-abi_x86_32.x86/test/Headers/Output/amdgcn_openmp_device_math_c.c.tmp-host.bc
error: OpenMP target architecture 'amdgcn-amd-amdhsa' pointer size is incompatible with host 'i686-pc-linux-gnu'

--

********************

Please fix or revert.

[mgorny]

Also, please link commits to diffs using Differential Revision: trailer in the future.

[mgorny]

I've pushed a fix in dab67c66932b9149842f7c8431e951f952125fc0, based on @jhuber6's fix from the other diff.

[doru1004]

In D139723#4016685, @mgorny wrote:

I've pushed a fix in dab67c66932b9149842f7c8431e951f952125fc0, based on @jhuber6's fix from the other diff.

Thank you for pushing a fix! :)

clang/lib/Headers/CMakeLists.txt

set(hexagon_files
   hexagon_types.h
   hvx_hexagon_protos.h
   )
 
 set(hip_files
   __clang_hip_libdevice_declares.h
   __clang_hip_cmath.h
   __clang_hip_math.h
+  __clang_hip_stdlib.h
   __clang_hip_runtime_wrapper.h
   )
 
 set(hlsl_h
   hlsl.h
   )
 set(hlsl_subdir_files
   hlsl/hlsl_basic_types.h

clang/lib/Headers/__clang_hip_runtime_wrapper.h

 #endif // HIP version check
 
 #ifdef __cplusplus
 } // extern "C"
 #endif //__cplusplus
 
 #include <__clang_hip_libdevice_declares.h>
 #include <__clang_hip_math.h>
+#include <__clang_hip_stdlib.h>
 
 #if defined(__HIPCC_RTC__)
 #include <__clang_hip_cmath.h>
 #else
 #include <__clang_cuda_math_forward_declares.h>
 #include <__clang_hip_cmath.h>
 #include <__clang_cuda_complex_builtins.h>
 #include <algorithm>

clang/lib/Headers/__clang_hip_stdlib.h

+/*===---- __clang_hip_stdlib.h - Device-side HIP math support --------------===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ *===-----------------------------------------------------------------------===
+ */
+#ifndef __CLANG_HIP_STDLIB_H__
+
+#if !defined(__HIP__) && !defined(__OPENMP_AMDGCN__)
+#error "This file is for HIP and OpenMP AMDGCN device compilation only."
+#endif
+
+#if !defined(__cplusplus)
+
+#include <limits.h>
+
+#ifdef __OPENMP_AMDGCN__
+#define __DEVICE__ static inline __attribute__((always_inline, nothrow))
+#else
+#define __DEVICE__ static __device__ inline __attribute__((always_inline))
+#endif
+
+__DEVICE__
+int abs(int __x) {
+  int __sgn = __x >> (sizeof(int) * CHAR_BIT - 1);
+  return (__x ^ __sgn) - __sgn;
+}
+__DEVICE__
+long labs(long __x) {
+  long __sgn = __x >> (sizeof(long) * CHAR_BIT - 1);
+  return (__x ^ __sgn) - __sgn;
+}
+__DEVICE__
+long long llabs(long long __x) {
+  long long __sgn = __x >> (sizeof(long long) * CHAR_BIT - 1);
+  return (__x ^ __sgn) - __sgn;
+}
+
+#endif // !defined(__cplusplus)
+
+#endif // #define __CLANG_HIP_STDLIB_H__

clang/lib/Headers/openmp_wrappers/stdlib.h

+/*===---- openmp_wrapper/stdlib.h ------ OpenMP math.h intercept ----- c++ -===
+ *
+ * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+ * See https://llvm.org/LICENSE.txt for license information.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ *===-----------------------------------------------------------------------===
+ */
+
+#ifndef __CLANG_OPENMP_STDLIB_H__
+#define __CLANG_OPENMP_STDLIB_H__
+
+#ifndef _OPENMP
+#error "This file is for OpenMP compilation only."
+#endif
+
+#include_next <stdlib.h>
+
+#ifdef __AMDGCN__
+#pragma omp begin declare variant match(device = {arch(amdgcn)})
+
+#define __OPENMP_AMDGCN__
+#include <__clang_hip_stdlib.h>
+#undef __OPENMP_AMDGCN__
+
+#pragma omp end declare variant
+#endif
+
+#endif // __CLANG_OPENMP_STDLIB_H__

clang/test/Headers/Inputs/include/stdlib.h

 #pragma once
 typedef __SIZE_TYPE__ size_t;
 
 void *malloc(size_t);
 void free(void*);
 
 #ifndef __cplusplus
 extern int abs(int __x) __attribute__((__const__));
+extern long labs(long __x) __attribute__((__const__));
+extern long long llabs(long long __x) __attribute__((__const__));
 #endif

clang/test/Headers/amdgcn_openmp_device_math_c.c

+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --function-signature --include-generated-funcs --replace-value-regex "__omp_offloading_[0-9a-z]+_[0-9a-z]+" "reduction_size[.].+[.]" "pl_cond[.].+[.|,]"
+// RUN: %clang_cc1 -internal-isystem %S/Inputs/include -x c -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa -D__OFFLOAD_ARCH_gfx90a__ -emit-llvm-bc %s -o %t-host.bc
+// RUN: %clang_cc1 -include __clang_hip_runtime_wrapper.h -internal-isystem %S/../../lib/Headers/openmp_wrappers -include __clang_openmp_device_functions.h -internal-isystem %S/../../lib/Headers/openmp_wrappers -internal-isystem %S/Inputs/include -x c -fopenmp -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown -fopenmp-targets=amdgcn-amd-amdhsa -emit-llvm %s -fopenmp-is-device -fopenmp-host-ir-file-path %t-host.bc -o - | FileCheck %s --check-prefixes=CHECK
+// REQUIRES: amdgpu-registered-target
+
+#include <stdlib.h>
+
+void test_math_int(int x) {
+#pragma omp target
+  {
+    int l1 = abs(x);
+  }
+}
+
+void test_math_long(long x) {
+#pragma omp target
+  {
+    long l1 = labs(x);
+  }
+}
+
+void test_math_long_long(long long x) {
+#pragma omp target
+  {
+    long long l1 = llabs(x);
+  }
+}
+// CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_test_math_int_l9
+// CHECK-SAME: (i64 noundef [[X:%.*]]) #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[RETVAL_I:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[__X_ADDR_I:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[__SGN_I:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[L1:%.*]] = alloca i32, align 4, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[X_ADDR]] to ptr
+// CHECK-NEXT:    [[L1_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[L1]] to ptr
+// CHECK-NEXT:    store i64 [[X]], ptr [[X_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr addrspacecast (ptr addrspace(1) @[[GLOB1:[0-9]+]] to ptr), i8 1, i1 true, i1 true)
+// CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
+// CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
+// CHECK:       user_code.entry:
+// CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[X_ADDR_ASCAST]], align 4
+// CHECK-NEXT:    [[RETVAL_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[RETVAL_I]] to ptr
+// CHECK-NEXT:    [[__X_ADDR_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__X_ADDR_I]] to ptr
+// CHECK-NEXT:    [[__SGN_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__SGN_I]] to ptr
+// CHECK-NEXT:    store i32 [[TMP1]], ptr [[__X_ADDR_ASCAST_I]], align 4
+// CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr [[__X_ADDR_ASCAST_I]], align 4
+// CHECK-NEXT:    [[SHR_I:%.*]] = ashr i32 [[TMP2]], 31
+// CHECK-NEXT:    store i32 [[SHR_I]], ptr [[__SGN_ASCAST_I]], align 4
+// CHECK-NEXT:    [[TMP3:%.*]] = load i32, ptr [[__X_ADDR_ASCAST_I]], align 4
+// CHECK-NEXT:    [[TMP4:%.*]] = load i32, ptr [[__SGN_ASCAST_I]], align 4
+// CHECK-NEXT:    [[XOR_I:%.*]] = xor i32 [[TMP3]], [[TMP4]]
+// CHECK-NEXT:    [[TMP5:%.*]] = load i32, ptr [[__SGN_ASCAST_I]], align 4
+// CHECK-NEXT:    [[SUB_I:%.*]] = sub nsw i32 [[XOR_I]], [[TMP5]]
+// CHECK-NEXT:    store i32 [[SUB_I]], ptr [[L1_ASCAST]], align 4
+// CHECK-NEXT:    call void @__kmpc_target_deinit(ptr addrspacecast (ptr addrspace(1) @[[GLOB1]] to ptr), i8 1, i1 true)
+// CHECK-NEXT:    ret void
+// CHECK:       worker.exit:
+// CHECK-NEXT:    ret void
+//
+//
+// CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_test_math_long_l16
+// CHECK-SAME: (i64 noundef [[X:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[RETVAL_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[__X_ADDR_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[__SGN_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[L1:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[X_ADDR]] to ptr
+// CHECK-NEXT:    [[L1_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[L1]] to ptr
+// CHECK-NEXT:    store i64 [[X]], ptr [[X_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr addrspacecast (ptr addrspace(1) @[[GLOB1]] to ptr), i8 1, i1 true, i1 true)
+// CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
+// CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
+// CHECK:       user_code.entry:
+// CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr [[X_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[RETVAL_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[RETVAL_I]] to ptr
+// CHECK-NEXT:    [[__X_ADDR_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__X_ADDR_I]] to ptr
+// CHECK-NEXT:    [[__SGN_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__SGN_I]] to ptr
+// CHECK-NEXT:    store i64 [[TMP1]], ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP2:%.*]] = load i64, ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[SHR_I:%.*]] = ashr i64 [[TMP2]], 63
+// CHECK-NEXT:    store i64 [[SHR_I]], ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP3:%.*]] = load i64, ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[XOR_I:%.*]] = xor i64 [[TMP3]], [[TMP4]]
+// CHECK-NEXT:    [[TMP5:%.*]] = load i64, ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[SUB_I:%.*]] = sub nsw i64 [[XOR_I]], [[TMP5]]
+// CHECK-NEXT:    store i64 [[SUB_I]], ptr [[L1_ASCAST]], align 8
+// CHECK-NEXT:    call void @__kmpc_target_deinit(ptr addrspacecast (ptr addrspace(1) @[[GLOB1]] to ptr), i8 1, i1 true)
+// CHECK-NEXT:    ret void
+// CHECK:       worker.exit:
+// CHECK-NEXT:    ret void
+//
+//
+// CHECK-LABEL: define {{[^@]+}}@{{__omp_offloading_[0-9a-z]+_[0-9a-z]+}}_test_math_long_long_l23
+// CHECK-SAME: (i64 noundef [[X:%.*]]) #[[ATTR0]] {
+// CHECK-NEXT:  entry:
+// CHECK-NEXT:    [[RETVAL_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[__X_ADDR_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[__SGN_I:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[L1:%.*]] = alloca i64, align 8, addrspace(5)
+// CHECK-NEXT:    [[X_ADDR_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[X_ADDR]] to ptr
+// CHECK-NEXT:    [[L1_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[L1]] to ptr
+// CHECK-NEXT:    store i64 [[X]], ptr [[X_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[TMP0:%.*]] = call i32 @__kmpc_target_init(ptr addrspacecast (ptr addrspace(1) @[[GLOB1]] to ptr), i8 1, i1 true, i1 true)
+// CHECK-NEXT:    [[EXEC_USER_CODE:%.*]] = icmp eq i32 [[TMP0]], -1
+// CHECK-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
+// CHECK:       user_code.entry:
+// CHECK-NEXT:    [[TMP1:%.*]] = load i64, ptr [[X_ADDR_ASCAST]], align 8
+// CHECK-NEXT:    [[RETVAL_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[RETVAL_I]] to ptr
+// CHECK-NEXT:    [[__X_ADDR_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__X_ADDR_I]] to ptr
+// CHECK-NEXT:    [[__SGN_ASCAST_I:%.*]] = addrspacecast ptr addrspace(5) [[__SGN_I]] to ptr
+// CHECK-NEXT:    store i64 [[TMP1]], ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP2:%.*]] = load i64, ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[SHR_I:%.*]] = ashr i64 [[TMP2]], 63
+// CHECK-NEXT:    store i64 [[SHR_I]], ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP3:%.*]] = load i64, ptr [[__X_ADDR_ASCAST_I]], align 8
+// CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[XOR_I:%.*]] = xor i64 [[TMP3]], [[TMP4]]
+// CHECK-NEXT:    [[TMP5:%.*]] = load i64, ptr [[__SGN_ASCAST_I]], align 8
+// CHECK-NEXT:    [[SUB_I:%.*]] = sub nsw i64 [[XOR_I]], [[TMP5]]
+// CHECK-NEXT:    store i64 [[SUB_I]], ptr [[L1_ASCAST]], align 8
+// CHECK-NEXT:    call void @__kmpc_target_deinit(ptr addrspacecast (ptr addrspace(1) @[[GLOB1]] to ptr), i8 1, i1 true)
+// CHECK-NEXT:    ret void
+// CHECK:       worker.exit:
+// CHECK-NEXT:    ret void
+//

llvm/utils/gn/secondary/clang/lib/Headers/BUILD.gn

sources = [
     "__clang_cuda_libdevice_declares.h",
     "__clang_cuda_math.h",
     "__clang_cuda_math_forward_declares.h",
     "__clang_cuda_runtime_wrapper.h",
     "__clang_cuda_texture_intrinsics.h",
     "__clang_hip_cmath.h",
     "__clang_hip_libdevice_declares.h",
     "__clang_hip_math.h",
+    "__clang_hip_stdlib.h",
     "__clang_hip_runtime_wrapper.h",
     "__stddef_max_align_t.h",
     "__wmmintrin_aes.h",
     "__wmmintrin_pclmul.h",
     "adxintrin.h",
     "altivec.h",
     "ammintrin.h",
     "amxfp16intrin.h",
sources = [
     "nmmintrin.h",
     "opencl-c-base.h",
     "opencl-c.h",
     "openmp_wrappers/__clang_openmp_device_functions.h",
     "openmp_wrappers/cmath",
     "openmp_wrappers/complex.h",
     "openmp_wrappers/complex_cmath.h",
     "openmp_wrappers/math.h",
+    "openmp_wrappers/stdlib.h",
     "pconfigintrin.h",
     "pkuintrin.h",
     "pmmintrin.h",
     "popcntintrin.h",
     "ppc_wrappers/bmi2intrin.h",
     "ppc_wrappers/bmiintrin.h",
     "ppc_wrappers/emmintrin.h",
     "ppc_wrappers/immintrin.h",