[Bug fortran/95690] [11 Regression] ICE in set_mem_attributes_minus_bitpos, at emit-rtl.c:2092

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95690

--- Comment #6 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Richard Biener
<rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:7d919c33fbd29a996326840dae3b5e093c3190f4

commit r10-8355-g7d919c33fbd29a996326840dae3b5e093c3190f4
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Jun 4 13:44:58 2020 +0200

    middle-end/95493 - bogus MEM_ATTRS for variable array access

    The following patch avoids keeping the inherited MEM_ATTRS when
    set_mem_attributes_minus_bitpos is called with a variable ARRAY_REF.
    The inherited ones may not reflect the correct offset and neither
    does the updated alias-set match the inherited MEM_EXPR.  This all
    ends up confusing path-based alias-analysis, causing wrong-code.

    The fix is to stop not adopting a MEM_EXPR for certain kinds of
    expressions and instead handle everything we can.  There's still
    the constant kind trees case which I'm too lazy to look into right
    now.  I did refrain from adding SSA_NAME there and instead avoided
    calling set_mem_attributes_minus_bitpos when debug expression
    expansion ended up expanding a SSA definition RHS which should
    already have taken care of setting the appropriate MEM_ATTRS.

    It also avoids calling set_mem_attributes on the
    DECL_INITIAL of a CONST_DECL which seems pointless since there
    cannot be a sensible MEM_EXPR derived from that.  We're overwriting
    both other possibly useful info, alias-set and alignment immediately
    so the following patch simply removes the call instead of making
    the function deal with even more (unexpected) trees that are not
    memory accesses.

    2020-06-23  Richard Biener  <rguenther@suse.de>

            PR middle-end/95493
            PR middle-end/95690
            * cfgexpand.c (expand_debug_expr): Avoid calling
            set_mem_attributes_minus_bitpos when we were expanding
            an SSA name.
            * emit-rtl.c (set_mem_attributes_minus_bitpos): Remove
            ARRAY_REF special-casing, add CONSTRUCTOR to the set of
            special-cases we do not want MEM_EXPRs for.  Assert
            we end up with reasonable MEM_EXPRs.
            * varasm.c (build_constant_desc): Remove set_mem_attributes call.

            * g++.dg/torture/pr95493.C: New testcase.
            * g++.dg/torture/pr95493-1.C: Likewise.
            * gfortran.dg/pr95690.f90: Likewise.