https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123475

--- Comment #6 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <[email protected]>:

https://gcc.gnu.org/g:2c83c3522e563ad0d85d50b39dcfe24336cf419a

commit r16-6661-g2c83c3522e563ad0d85d50b39dcfe24336cf419a
Author: Jakub Jelinek <[email protected]>
Date:   Fri Jan 9 20:46:39 2026 +0100

    c: Optimize TARGET_EXPRs for _Atomic loads [PR123475]

    On the following testcase we emit a false positive warning that
    a temporary (TARGET_EXPR slot) is used uninitialized, from the early_uninit
    pass.

    This regressed with my change to instrument for
    -ftrivial-auto-var-init={zero,pattern} not just DECL_EXPRs, but also
    TARGET_EXPR initializations if the TARGET_EXPR_INITIALIZER has void type.
    Those cases are where the initializer doesn't necessarily have to
initialize
    the whole TARGET_EXPR slot, or might use parts or the whole slot before
    those are initialized; this is how e.g. various C++ temporary objects are
    constructed.

    The problem is in pass interaction.  The FE creates a TARGET_EXPR with
    void type initializer because the initializer is originally
    __atomic_load (&expr, &tmp, SEQ_CST); but it is folded instantly into
    (void) (tmp = (type) __atomic_load_N (&expr, SEQ_CST)).  The FE also
    marks the TARGET_EXPR slot as TREE_ADDRESSABLE, because it would be
    if it will use libatomic, but nothing in the IL then takes its address.
    Now, since my r16-4212 change which was for mainly C++26 compliance
    we see the TARGET_EXPR and because it has void type
TARGET_EXPR_INITIALIZER,
    we start with tmp = .DEFERRED_INIT (...); just in case the initialization
    would attempt to use the slot before initialization or not initialize
fully.
    Because tmp is TREE_ADDRESSABLE and has gimple reg type, it is actually not
    gimplified as tmp = .DEFERRED_INIT (...); but as _1 = .DEFERRED_INIT (...);
    tmp = _1; but because it is not actually address taken in the IL, already
    the ssa pass turns it into SSA_NAME (dead one), so we have
    _1 = .DEFERRED_INIT (...); _2 = _1; and _2 is unused.  Next comes
    early_uninit and warns on the dead SSA_NAME copy that it uses uninitialized
    var.

    The following patch attempts to fix that by checking if
    c_build_function_call_vec has optimized the call right away into pure
    assignment to the TARGET_EXPR slot without the slot being used anywhere
    else in the expression and 1) clearing again TREE_ADDRESSABLE on the slot,
    because it isn't really addressable 2) optimizing the TARGET_EXPR, so that
    it doesn't have void type TARGET_EXPR_INITIALIZER by changing it to the rhs
    of the MODIFY_EXPR.  That way gimplifier doesn't bother creating
    .DEFERRED_INIT for it at all.

    Or should something like this be done instead in the TARGET_EXPR
    gimplification?  I mean not the TREE_ADDRESSABLE clearing, that can't be
    done without knowing what we know in the FE, but the rest, generally
    TARGET_EXPR with initializer (void) (TARGET_EXPR_SLOT = something)
    where something doesn't refer to TARGET_EXPR_SLOT can be optimized into
    just something TARGET_EXPR_INITIALIZER.

    2026-01-09  Jakub Jelinek  <[email protected]>

            PR c/123475
            * c-typeck.cc (c_find_var_r): New function.
            (convert_lvalue_to_rvalue): If c_build_function_call_vec
            folded __atomic_load (&expr, &tmp, SEQ_CST); into
            (void) (tmp = __atomic_load_<N> (&expr, SEQ_CST)), drop
            TREE_ADDRESSABLE flag from tmp and set TARGET_EXPR
            initializer just to the rhs of the MODIFY_EXPR.

            * gcc.dg/pr123475.c: New test.

Reply via email to