[Bug sanitizer/112709] [13/14 Regression] address sanitize and returns_twice causes an ICE

["cvs-commit at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112709

--- Comment #7 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:ad860cc27b3312f9119c7fecb8638a7c1f6d77c9

commit r14-9438-gad860cc27b3312f9119c7fecb8638a7c1f6d77c9
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Mar 12 11:34:50 2024 +0100

    asan: Instrument <retval> stores in callees rather than callers [PR112709]

    asan currently instruments since PR69276 r6-6758 fix calls which store
    the return value into memory on the caller side, before the call it
    verifies the memory is writable.
    Now PR112709 where we ICE on trying to instrument such calls made me
    think about whether that is what we want to do.

    There are 3 different cases.

    One is when a function returns an aggregate which is passed e.g. in
    registers, say like struct S { int a[4]; }; returning on x86_64.
    That would be ideally instrumented in between the actual call and
    storing of the aggregate into memory, but asan currently mostly
    works as a GIMPLE pass and arranging for the instrumentation to happen
    at that spot would be really hard.  We could diagnose after the call
    but generally asan attempts to diagnose stuff before something is
    overwritten rather than after, or keep the current behavior (that is
    what this patch does, which has the disadvantage that it can complain
    about UB even for functions which never return and so never actually store,
    and doesn't check whether the memory wasn't e.g. poisoned during the call)
    or could e.g. instrument both before and after the call (that would have
    the disadvantage the current state has but at least would check post-factum
    the store again afterwards).

    Another case is when a function returns an aggregate through a hidden
    reference, struct T { int a[128]; }; on x86_64 or even the above struct S
    on ia32 as example.  In the actual program such stores happen when storing
    something to <retval> or its parts in the callee, because <retval> there
    expands to *hidden_retval.  So, IMHO we should instrument those in the
    callee rather than caller, that is where the writes are and we can do that
    easily.  This is what the patch below does.

    And the last case is for builtins/internal functions.  Usually those don't
    return aggregates, but in case they'd do and can be expanded inline, it is
    better to instrument them in the caller (as before) rather than not
    instrumenting the return stores at all.

    I had to tweak the expected output on the PR69276 testcase, because
    with the patch it keeps previous behavior on x86_64 (structure returned
    in registers, stored in the caller, so reported as UB in A::A()),
    while on i686 it changed the behavior and is reported as UB in the
    vnull::operator vec which stores the structure, A::A() is then a frame
    above it in the backtrace.

    2024-03-12  Jakub Jelinek  <jakub@redhat.com>

            PR sanitizer/112709
            * asan.cc (has_stmt_been_instrumented_p): Don't instrument call
            stores on the caller side unless it is a call to a builtin or
            internal function or function doesn't return by hidden reference.
            (maybe_instrument_call): Likewise.
            (instrument_derefs): Instrument stores to RESULT_DECL if
            returning by hidden reference.

            * gcc.dg/asan/pr112709-1.c: New test.
            * g++.dg/asan/pr69276.C: Adjust expected output for some targets.