]> git.ipfire.org Git - thirdparty/gcc.git/commit
gimple, internal-fn: Add IFN_TRAP and use it for __builtin_unreachable [PR106099]
authorJakub Jelinek <jakub@redhat.com>
Thu, 28 Jul 2022 10:42:14 +0000 (12:42 +0200)
committerJakub Jelinek <jakub@redhat.com>
Thu, 28 Jul 2022 10:43:13 +0000 (12:43 +0200)
commitf64eb636677d714781b4543f111b1c9239328db6
treeb365d582214c402540c809b66b405d97cdd05d28
parent0652087375e7813f06140f1bb5e18375f45d28ff
gimple, internal-fn: Add IFN_TRAP and use it for __builtin_unreachable [PR106099]

__builtin_unreachable and __ubsan_handle_builtin_unreachable don't
use vops, they are marked const/leaf/noreturn/nothrow/cold.
But __builtin_trap uses vops, isn't const, just leaf/noreturn/nothrow/cold.
This is I believe so that when users explicitly use __builtin_trap in their
sources they get stores visible at the trap side.
-fsanitize=unreachable -fsanitize-undefined-trap-on-error used to transform
__builtin_unreachable to __builtin_trap even in the past, but the sanopt pass
has TODO_update_ssa, so it worked fine.

Now that gimple_build_builtin_unreachable can build a __builtin_trap call
right away, we can run into problems that whenever we need it we would need
to either manually or through TODO_update* ensure the vops being updated.

Though, as it is originally __builtin_unreachable which is just implemented
as trap, I think for this case it is fine to avoid vops.  For this the
patch introduces IFN_TRAP, which has ECF_* flags like __builtin_unreachable
and is expanded as __builtin_trap.

2022-07-28  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/106099
* internal-fn.def (TRAP): New internal fn.
* internal-fn.h (expand_TRAP): Declare.
* internal-fn.cc (expand_TRAP): Define.
* gimple.cc (gimple_build_builtin_unreachable): For BUILT_IN_TRAP,
use internal fn rather than builtin.

* gcc.dg/ubsan/pr106099.c: New test.
gcc/gimple.cc
gcc/internal-fn.cc
gcc/internal-fn.def
gcc/internal-fn.h
gcc/testsuite/gcc.dg/ubsan/pr106099.c [new file with mode: 0644]