__atomic_test_and_set: Fall back to library, not non-atomic code
commit8e6757b30d0f3f13d47d0f842801a751ba6293c2
authorHans-Peter Nilsson <hp@axis.com>
Sat, 23 Sep 2023 03:06:52 +0000 (23 05:06 +0200)
committerHans-Peter Nilsson <hp@bitrange.com>
Tue, 26 Sep 2023 21:39:46 +0000 (26 23:39 +0200)
treea56aa786ab9b2835e64c7ba278331aad0f4d424a
parentdd0c42cd37dab62f9f61ac834333ec899305dee4
__atomic_test_and_set: Fall back to library, not non-atomic code

Make __atomic_test_and_set consistent with other __atomic_ and __sync_
builtins: call a matching library function instead of emitting
non-atomic code when the target has no direct insn support.

There's special-case code handling targetm.atomic_test_and_set_trueval
!= 1 trying a modified maybe_emit_sync_lock_test_and_set.  Previously,
if that worked but its matching emit_store_flag_force returned NULL,
we'd segfault later on.  Now that the caller handles NULL, gcc_assert
here instead.

While the referenced PR:s are ARM-specific, the issue is general.

PR target/107567
PR target/109166
* builtins.cc (expand_builtin) <case BUILT_IN_ATOMIC_TEST_AND_SET>:
Handle failure from expand_builtin_atomic_test_and_set.
* optabs.cc (expand_atomic_test_and_set): When all attempts fail to
generate atomic code through target support, return NULL
instead of emitting non-atomic code.  Also, for code handling
targetm.atomic_test_and_set_trueval != 1, gcc_assert result
from calling emit_store_flag_force instead of returning NULL.
gcc/builtins.cc
gcc/optabs.cc