Run DCE after a LoopFlatten test to reduce spurious output [nfc]
[llvm-project.git] / libc / startup / gpu / amdgpu / start.cpp
blob89b0be208d2e7fd00451f535549ad7afb0923c9b
1 //===-- Implementation of crt for amdgpu ----------------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
9 #include "src/__support/GPU/utils.h"
10 #include "src/__support/RPC/rpc_client.h"
11 #include "src/stdlib/atexit.h"
12 #include "src/stdlib/exit.h"
14 extern "C" int main(int argc, char **argv, char **envp);
16 namespace LIBC_NAMESPACE {
18 // The AMDGPU architecture provides a fixed frequency clock used for obtaining
19 // real time. However, the frequency of this clock varies between cards and can
20 // only be obtained via the driver. The loader will set this so we can use it.
21 extern "C" [[gnu::visibility("protected")]] uint64_t
22 [[clang::address_space(4)]] __llvm_libc_clock_freq = 0;
24 extern "C" uintptr_t __init_array_start[];
25 extern "C" uintptr_t __init_array_end[];
26 extern "C" uintptr_t __fini_array_start[];
27 extern "C" uintptr_t __fini_array_end[];
29 using InitCallback = void(int, char **, char **);
30 using FiniCallback = void(void);
32 static void call_init_array_callbacks(int argc, char **argv, char **env) {
33 size_t init_array_size = __init_array_end - __init_array_start;
34 for (size_t i = 0; i < init_array_size; ++i)
35 reinterpret_cast<InitCallback *>(__init_array_start[i])(argc, argv, env);
38 static void call_fini_array_callbacks() {
39 size_t fini_array_size = __fini_array_end - __fini_array_start;
40 for (size_t i = 0; i < fini_array_size; ++i)
41 reinterpret_cast<FiniCallback *>(__fini_array_start[i])();
44 } // namespace LIBC_NAMESPACE
46 extern "C" [[gnu::visibility("protected"), clang::amdgpu_kernel]] void
47 _begin(int argc, char **argv, char **env) {
48 // We want the fini array callbacks to be run after other atexit
49 // callbacks are run. So, we register them before running the init
50 // array callbacks as they can potentially register their own atexit
51 // callbacks.
52 LIBC_NAMESPACE::atexit(&LIBC_NAMESPACE::call_fini_array_callbacks);
53 LIBC_NAMESPACE::call_init_array_callbacks(argc, argv, env);
56 extern "C" [[gnu::visibility("protected"), clang::amdgpu_kernel]] void
57 _start(int argc, char **argv, char **envp, int *ret) {
58 // Invoke the 'main' function with every active thread that the user launched
59 // the _start kernel with.
60 __atomic_fetch_or(ret, main(argc, argv, envp), __ATOMIC_RELAXED);
63 extern "C" [[gnu::visibility("protected"), clang::amdgpu_kernel]] void
64 _end(int retval) {
65 // Only a single thread should call `exit` here, the rest should gracefully
66 // return from the kernel. This is so only one thread calls the destructors
67 // registred with 'atexit' above.
68 LIBC_NAMESPACE::exit(retval);