On Wed, 18 Jan 2023, Jan Hubicka wrote:
> > On Tue, 17 Jan 2023, Jan Hubicka wrote:
> >
> > > > > We don't use same argumentation about other control flow statements.
> > > > > The following:
> > > > >
> > > > > fn()
> > > > > {
> > > > > try {
> > > > > i_read_no_global_memory ();
> > > > > } catch (...)
> > > > > {
> > > > > reutrn 1;
> > > > > }
> > > > > return 0;
> > > > > }
> > > > >
> > > > > should be detected as const. Marking throw pure would make fn pure
> > > > > too.
> > > >
> > > > I suppose i_read_no_global_memory is const here. Not sure why that
> > > Suppose we have:
> > >
> > > void
> > > i_read_no_global_memory ()
> > > {
> > > throw(0);
> > > }
> > >
> > > If cxa_throw itself was annotated as 'p' rahter than 'c' ipa-modref will
> > > believe that cxa_throw will read any global memory and will propagate it
> > > to all callers. So fn() will be also marked as reading all global
> > > memory.
> >
> > Sure - but for the purpose of local optimizations in
> > i_read_no_global_memory cxa_throw has to appear to read memory.
>
> Yes, I think every stmt that can throw externally need VUSE (just like
> return_stmt needs it). Even if throw(0) was replaced by a=b/c with
> -fnon-call-exceptions. It is still not clear to me why this should
> imply that we need 'p' instead of 'c' in fnspecs.
>
> So I think we should try to make the following to work:
>
> diff --git a/gcc/tree-ssa-operands.cc b/gcc/tree-ssa-operands.cc
> index 57e393ae164..d24f1721eb2 100644
> --- a/gcc/tree-ssa-operands.cc
> +++ b/gcc/tree-ssa-operands.cc
> @@ -951,6 +951,9 @@ operands_scanner::parse_ssa_operands ()
> enum gimple_code code = gimple_code (stmt);
> size_t i, n, start = 0;
>
> + if (stmt_can_throw_external (fn, stmt))
> + append_vuse (gimple_vop (fn));
> +
> switch (code)
> {
> case GIMPLE_ASM:
It's going to be a bit tricky since in many places we use
gimple_vuse () != NULL to check whether an assignment is a
load/store. But yes, the above is sort-of what we'd need to do.
> > Having a VUSE there dependent on whether the function performs any
> > load or store would be quite ugly. Instead modref could special-case
> > cxa_throw and not treat it as reading memory (like it already does
> > for the return stmt I suppose - that also has a VUSE).
>
> modref looks into statements with VUSEs on them and checks what
> reads/stores are done. So return statement with VUSE is walked and no
> load is recorded because no actual load is found.
> Similarly that would happen with __cxa_throw if it was 'c'.
> With 'p' it has nothing to analyze so it would trust the fact that
> cxa_throw itself reads some global state.
I see. But does __cxa_throw stmt_can_throw_external ()? Otherwise
the operand scanner elides VUSE on const function calls.
> >
> > The problem is IIRC GIMPLE_RESX which doesn't derive from
> > gimple_statement_with_memory_ops_base. There's a bugzilla I can't find
> > right now refering to this issue.
>
> I never tried to play with gimple hiearchy. It is hard to fix resx? I
> wonder if we have other cases. I guess for a=b/c we are luck just
> because gimple_assign can also be load or store so it has memory_ops...
Fixing resx would come at the cost of deriving from _with_ops, but not
sure if that waste of space is too important.
Richard.
> Thanks,
> Honza
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
