On Tue, Mar 16, 2021 at 4:02 PM Thomas Schwinge <tho...@codesourcery.com> wrote:
>
> Hi!
>
> Richard, thanks for your answer. I'll need to look into this more; two
> questions already:
>
> On 2021-03-15T20:17:17+0100, Richard Biener via Gcc <gcc@gcc.gnu.org> wrote:
> > On March 15, 2021 7:31:46 PM GMT+01:00, Thomas Schwinge
> > <tho...@codesourcery.com> wrote:
> >>First time I'm using this API -- so the error certainly may be on my
> >>side. ;-)
> >>
> >>What I'm doing, is a 'walk_gimple_seq', and in that one's
> >>'callback_stmt', call 'walk_stmt_load_store_addr_ops', to collect
> >>variable load/store/address-taken instances. This did seem quite
> >>straight-forward, given the description; 'gcc/gimple-walk.c':
> >>
> >>/* For the statement STMT call the callbacks VISIT_LOAD, VISIT_STORE
> >>and
> >> VISIT_ADDR if non-NULL on loads, store and address-taken operands
> >>passing the STMT, the base of the operand, the operand itself
> >>containing
> >> the base and DATA to it. The base will be either a decl, an indirect
> >> reference (including TARGET_MEM_REF) or the argument of an address
> >> expression.
> >> Returns the results of these callbacks or'ed. */
> >>
> >> bool
> >> walk_stmt_load_store_addr_ops (gimple *stmt, void *data,
> >> walk_stmt_load_store_addr_fn visit_load,
> >> walk_stmt_load_store_addr_fn visit_store,
> >> walk_stmt_load_store_addr_fn visit_addr)
> >> { [...] }
> >>
> >>Indeed, given (Fortran) 'zzz = 1', we produce GIMPLE:
> >>
> >> gimple_assign <real_cst, zzz, 1.0e+0, NULL, NULL>
> >>
> >>..., and calling 'walk_stmt_load_store_addr_ops' on that, I see, as
> >>expected, the 'visit_store' callback invoked, with 'rhs' and 'arg':
> >>'<var_decl zzz>'.
> >>
> >>However, given (Fortran) 'zzz = r + r2', we produce GIMPLE:
> >>
> >> gimple_assign <plus_expr, zzz, r, r2, NULL>
> >>
> >>..., and calling 'walk_stmt_load_store_addr_ops' on that, I see,
> >>unexpectedly, no callback at all invoked: neither 'visit_load', nor
> >>'visit_store' (nor 'visit_address', obviously).
> >
> > The variables involved are registers. You only get called on memory
> > operands.
>
> How would I have told that from the 'walk_stmt_load_store_addr_ops'
> function description? (How to improve that one "to reflect relatity"?)
Hmm, from the function name which mentions 'load' and 'store'? ;)
For example we have
static inline bool
gimple_store_p (const gimple *gs)
{
tree lhs = gimple_get_lhs (gs);
return lhs && !is_gimple_reg (lhs);
}
> But 'zzz' surely is the same in 'zzz = 1' vs. 'zzz = r + r2' -- for the
> former I *do* see the 'visit_store' callback invoked, for the latter I
> don't?
>
> >>From a quick look at 'gcc/gimple-walk.c:walk_stmt_load_store_addr_ops',
> >>this seems to intentionally be implemented in this way -- but I don't
> >>understand the rationale?
> >>
> >>
> >>Instead of 'walk_gimple_seq' -> 'callback_stmt' ->
> >>'walk_stmt_load_store_addr_ops', do I need to use 'walk_gimple_seq' ->
> >>'callback_op' -> "something"?
> >
> > Yes, if you want to visit register sets and uses as a well. Note you'll
> > also see constants that way.
>
> I'll look into that; in particular to figure out "something" for what I
> need: load/store/address-taken.
Well, indeed. The question is what you are doing the exercise for ;)
(and thus what you actually need visited)
Richard.
>
> Grüße
> Thomas
> -----------------
> Mentor Graphics (Deutschland) GmbH, Arnulfstrasse 201, 80634 München
> Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Frank
> Thürauf
