[Bug lto/122620] LTO ICE while compiling firefox-145.0 with LTO+PGO

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122620

--- Comment #14 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-13 branch has been updated by Jakub Jelinek
<jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:15e4879c87de6e0a8679e135a19bb635600d2bc9

commit r13-10171-g15e4879c87de6e0a8679e135a19bb635600d2bc9
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Tue Nov 11 08:29:22 2025 +0100

    gimplify-me: Fix regimplification of gimple-reg-type clobbers [PR122620]

    Since r11-2238-ge443d8213864ac337c29092d4767224f280d2062 the C++ FE
    emits clobbers like *_1 = {CLOBBER}; where *_1 MEM_REF has some scalar
    type like int for -flifetime-dse={1,2} and most of the compiler manages
    to cope with that.
    If we are very unlucky, we trigger an ICE while trying to regimplify it
    (at least during inlining), as happens with GCC 15.2 on firefox-145.0
    built with LTO+PGO.
    I haven't managed to reduce that to a small testcase that would ICE though,
    the clobber certainly appears in code like
    template <typename T>
    struct S {
      T *p;
      union { char a; T b; };
      static S foo (T *x) { S s; s.p = x; s.b.~T (); return s; }
      ~S ();
    };

    void
    bar ()
    {
      int i = 42;
      S <int> s = S <int>::foo (&i);
    }
    but convincing inliner that it should id->regimplify = true; on exactly
    that stmt has been difficult.

    The ICE is because we try (in two spots) to regimplify the rhs of the
    gimple_clobber_p stmt if gimple-reg-type type (i.e. the TREE_CLOBBER),
    because it doesn't satisfy the is_gimple_mem_rhs_or_call predicate
    returned by rhs_predicate_for for the MEM_REF lhs.  And regimplify it
    by trying to gimplify SSA_NAME = {CLOBBER}; INIT_EXPR and in there reach
    a special case which stores that freshly made SSA_NAME into memory and
    loads it from memory, so uses a SSA_NAME without SSA_NAME_DEF_STMT.

    Fixed thusly by saying clobbers are ok even for the gimple-reg-types.

    2025-11-11  Jakub Jelinek  <jakub@redhat.com>

            PR lto/122620
            * gimplify-me.cc (gimple_regimplify_operands): Don't try to
regimplify
            TREE_CLOBBER on rhs of gimple_clobber_p if it has gimple_reg_type.

    (cherry picked from commit 8f3242ce5c03c30f806f54ceaff2a15d6e5b5ee6)