https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110132
Bug ID: 110132
Summary: aarch64: Bogus -Wbuiltin-declaration-mismatch with
ls64 builtins
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: acoplan at gcc dot gnu.org
Target Milestone: ---
If one subverts arm_acle.h and tries to create a minimal preprocessed test case
using ls64 builtins:
$ cat ls64.i
#pragma GCC aarch64 "arm_acle.h"
typedef __arm_data512_t data512_t;
void f(void *p, data512_t d) {
__builtin_aarch64_st64b(p, d);
}
./xgcc -B . -c ls64.i -S -o /dev/null -march=armv8.7-a
ls64.i: In function ‘f’:
ls64.i:4:3: warning: implicit declaration of function ‘__builtin_aarch64_st64b’
[-Wimplicit-function-declaration]
4 | __builtin_aarch64_st64b(p, d);
| ^~~~~~~~~~~~~~~~~~~~~~~
ls64.i:4:3: warning: incompatible implicit declaration of built-in function
‘__builtin_aarch64_st64b’ [-Wbuiltin-declaration-mismatch]
we get these bogus warnings. The normal flow for e.g. Advanced SIMD builtins
is:
- Backend calls add_builtin_function.
- c_builtin_function adds the builtin to the list of visible_builtins.
- push_file_scope gets called and builtins in visible_builtins are made
visible.
The problem here is that aarch64_init_ls64_builtins is called when processing
the pragma for arm_acle.h, which is *after* push_file_scope gets called, so the
builtins never get made visible by the C FE.
I think it might be easiest to make the ls64 handling more SVE-like and use the
simulate_builtin_function_decl langhook (which doesn't have this deferred
visibility situation), dropping the wrapper functions from arm_acle.h
altogether.