[OFFLOAD] Interop fixes for Windows #162652
Conversation
|
@llvm/pr-subscribers-offload Author: Alex Duran (adurang) ChangesOn Windows, for a reason I don't fully understand boolean bits get extra padding (even when asking for packed structures) in the structures that messes the offsets between the compiler and the runtime. Also, "weak" works differently on Windows than Linux (i.e., the "local" routine has preference) which causes it to crash as we don't really have an alternate implementation of __kmpc_omp_wait_deps. Given this, it doesn't either make sense to mark it as "weak" for Linux either. Full diff: https://github.com/llvm/llvm-project/pull/162652.diff 2 Files Affected:
diff --git a/offload/include/OpenMP/InteropAPI.h b/offload/include/OpenMP/InteropAPI.h
index 53ac4be2e2e98..8c06ba36fc3f3 100644
--- a/offload/include/OpenMP/InteropAPI.h
+++ b/offload/include/OpenMP/InteropAPI.h
@@ -25,8 +25,8 @@ typedef enum kmp_interop_type_t {
} kmp_interop_type_t;
struct interop_attrs_t {
- bool inorder : 1;
- int reserved : 31;
+ uint32_t inorder : 1;
+ uint32_t reserved : 31;
/// Check if the supported attributes are compatible with the current
/// attributes. Only if an attribute is supported can the value be true,
@@ -44,15 +44,15 @@ struct interop_spec_t {
};
struct interop_flags_t {
- bool implicit : 1; // dispatch (true) or interop (false)
- bool nowait : 1; // has nowait flag
- int reserved : 30;
+ uint32_t implicit : 1; // dispatch (true) or interop (false)
+ uint32_t nowait : 1; // has nowait flag
+ uint32_t reserved : 30;
};
struct interop_ctx_t {
uint32_t version; // version of the interface (current is 0)
interop_flags_t flags;
- int gtid;
+ int32_t gtid;
};
struct dep_pack_t {
diff --git a/offload/libomptarget/OpenMP/InteropAPI.cpp b/offload/libomptarget/OpenMP/InteropAPI.cpp
index c55ef2c2e672c..d6ef17c06355c 100644
--- a/offload/libomptarget/OpenMP/InteropAPI.cpp
+++ b/offload/libomptarget/OpenMP/InteropAPI.cpp
@@ -22,8 +22,7 @@ extern "C" {
void __kmpc_omp_wait_deps(ident_t *loc_ref, int32_t gtid, int32_t ndeps,
kmp_depend_info_t *dep_list, int32_t ndeps_noalias,
- kmp_depend_info_t *noalias_dep_list)
- __attribute__((weak));
+ kmp_depend_info_t *noalias_dep_list);
} // extern "C"
|
| void __kmpc_omp_wait_deps(ident_t *loc_ref, int32_t gtid, int32_t ndeps, | ||
| kmp_depend_info_t *dep_list, int32_t ndeps_noalias, | ||
| kmp_depend_info_t *noalias_dep_list) | ||
| __attribute__((weak)); |
There was a problem hiding this comment.
So it was marked as weak but didn't provide any definition in libomptarget (apart from the one in libomp)?
There was a problem hiding this comment.
It seems so. The ones in other places (like InternalTypes.h) also have "weak" but no definition. I have no idea who decided to do it like this but I think all "weak" attributes should probably removed.
There was a problem hiding this comment.
I think so too. I don't see why these attributes are needed without a definition in libomptarget. If it's not linked to libomp, or another lib providing this symbol, there will be a linking error anyway.
There was a problem hiding this comment.
Do you prefer to do it on this PR or a sepparate one?
There was a problem hiding this comment.
I'd like to know other opinions before, to confirm this is not necessary.
CC @jhuber6
There was a problem hiding this comment.
Okay, so weak extern symbols are defined as nullptr and not a linker error in ELF. This might be trying to prevent that? I don't really know for sure what it's trying to do. There's no check on this function so if it wasn't defined this code would segfault. I'm just going to assume it's not needed, since we should always link liboimp w/ libomptarget` anyways.
There was a problem hiding this comment.
Thanks for the comment!
Maybe it's better to remove these weak attributes in a different PR.
There was a problem hiding this comment.
Yes, I can do a follow-up for other functions specified as weak. Also, to gather them in one place, for some reason they're scattered here and there...
| void __kmpc_omp_wait_deps(ident_t *loc_ref, int32_t gtid, int32_t ndeps, | ||
| kmp_depend_info_t *dep_list, int32_t ndeps_noalias, | ||
| kmp_depend_info_t *noalias_dep_list) | ||
| __attribute__((weak)); |
There was a problem hiding this comment.
Okay, so weak extern symbols are defined as nullptr and not a linker error in ELF. This might be trying to prevent that? I don't really know for sure what it's trying to do. There's no check on this function so if it wasn't defined this code would segfault. I'm just going to assume it's not needed, since we should always link liboimp w/ libomptarget` anyways.
…er (#164613) Follow-up from #162652 --------- Co-authored-by: Michael Klemm <[email protected]>
4 participants
On Windows, for a reason I don't fully understand boolean bits get extra padding (even when asking for packed structures) in the structures that messes the offsets between the compiler and the runtime.
Also, "weak" works differently on Windows than Linux (i.e., the "local" routine has preference) which causes it to crash as we don't really have an alternate implementation of __kmpc_omp_wait_deps. Given this, it doesn't make sense to mark it as "weak" for Linux either.