[Bug tree-optimization/125597] [16/17 Regression] ICE: SIGSEGV in analyze_scalar_evolution_1 (tree-scalar-evolution.cc:2026) with -O3 -mcpu=c1-ultra -mautovec-preference=sve-only

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125597

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 9 Jun 2026, tnfchris at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125597
> 
> --- Comment #5 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
> So this isn't an early break bug, but a generic masking one (the bitint code
> has multiple loops).
> 
> Since masked loops were introduced during peeling we'd create the dummy
> variable which is then filled in when vect_set_loop_condition is called.
> 
> This particular issue started with
> 
> commit g:4ac080e384cd30d37fac069d791b813100e6524a
> Author: Andrew MacLeod <amacleod@redhat.com>
> Date:   Mon Jan 19 13:44:25 2026 -0500
> 
>     Do not trap on a stmt with no basic block.
> 
>     WHen calculating ranges for statements not in the IL, avoid looking for
>     range on entry values.
> 
>             PR tree-optimization/123314
>             gcc/
>             * gimple-range.cc (gimple_ranger::range_on_entry): Do not check
>             ranger cache for an SSA_NAME with no BB.
>             (gimple_ranger::prefill_stmt_dependencies): Stop filling
>             dependencies when an out-of IL name is encountered.
> 
>             gcc/testsuite/
>             * gcc.dg/pr123314.c: New.
> 
> Because Ranger would now no longer check the cache for out of IL statements and
> instead forces a re-analysis.
> Eventually it ends up in SCEV which requires that the statement be in IL.
> 
> So it's not the fact that it's a GIMPLE_NOP that's the problem but that it's
> out of IL and using .VARYING still keeps it out of IL so that didn't fix it.

Sure, if you put that in the IL it would.

> the question is why does SCEV need it in IL? given that the only reason for
> this is to determine if the variable is in the loop or not.  logically
> speaking, if it's not in IL it can't be in the loop.

Because all GIMPLE defs are supposed to be in the IL, apart from
SSA_NAME_IS_DEFAULT_DEF ones.  It's simply invalid/unexpected IL and
passes should not have workarounds for that.

> 
> diff --git a/gcc/tree-scalar-evolution.cc b/gcc/tree-scalar-evolution.cc
> index 279ff46474e..2954ff7a55a 100644
> --- a/gcc/tree-scalar-evolution.cc
> +++ b/gcc/tree-scalar-evolution.cc
> @@ -2023,6 +2023,14 @@ analyze_scalar_evolution_1 (class loop *loop, tree var)
> 
>    def = SSA_NAME_DEF_STMT (var);
>    bb = gimple_bb (def);
> +  if (!bb)
> +    {
> +      /* If we don't know where the variable is, just keep the symbolic
> +        form.  */
> +      res = chrec_dont_know;
> +      goto set_and_end;
> +    }
> +
>    def_loop = bb->loop_father;
> 
>    if (!flow_bb_inside_loop_p (loop, bb))
> 
> Fixes it, and allows the various forms in the vectorizer where we fold out of
> IL statements before inserting to continue working.
> 
> Or do you want me to force these temporary placeholders into the IL Richi? My
> concern here is that we can create a use-before-def issue because any new
> statements emitted must be in the right order whereas now we just build up seq
> and flush at the gcond GSI.

Yes, I want us to avoid even intermediate invalid GIMPLE because that can
(as we see here) lead to all sorts of issues.

> The above SCEV fix seems less hacky to me.

Not to me.