jhuber6 added a comment.
In D123471#3443612 <https://reviews.llvm.org/D123471#3443612>, @yaxunl wrote:
> Is OpenMP runtime able to find these entries without registering them through
> some API functions? If so, do you have a pointer to the code doing that?
Yes, the linker will define `__start/__stop` symbols for any sections found
with a section name that is a valid C-identifier. If you compile the following
file with any OpenMP offloading code you should be able to print out all the
symbol names that will be registered when the runtime is initialized.
#include <stdint.h>
#include <stdio.h>
struct __tgt_offload_entry {
void *addr; // Pointer to the offload entry info.
// (function or global)
char *name; // Name of the function or global.
size_t size; // Size of the entry info (0 if it a function).
int32_t flags;
int32_t reserved;
};
extern struct __tgt_offload_entry __start_omp_offloading_entries;
extern struct __tgt_offload_entry __stop_omp_offloading_entries;
__attribute__((constructor)) void print() {
struct __tgt_offload_entry *iter = &__start_omp_offloading_entries;
for (; iter != &__stop_omp_offloading_entries; ++iter)
printf("%s\n", iter->name);
}
And then compile like
$ clang input.c -fopenmp -fopenmp-targets=nvptx64 -c
$ clang print.c -c
$ clang input.o print.o -fopenmp -fopenmp-targets=nvptx64
$ ./a.out
x
__omp_offloading_fd02_605785f3_main_l8
> most CUDA/HIP programs assume -fno-gpu-rdc mode, which have multiple sections
> containing these entries merged by linker, with gaps between them. How do
> runtime identify such gaps and skip them?
I'm making the executive decision to always enable `fgpu-rdc` when using this
new driver in the future. The above is handled by the linker so there shouldn't
be any gaps.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D123471/new/
https://reviews.llvm.org/D123471
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
