On 12/2/25 20:09, Andrew Pinski wrote:
On Tue, Dec 2, 2025 at 1:37 PM Andrew MacLeod <[email protected]> wrote:
The record() and register_relation() routines in the relation oracle
and folding routines currently return void. They are called when there
is a relation to register and there has not been a need for feedback on
whether one was registered or not.
The patch which caused this PR recognized that and on-demand calculation
of something in the middle of the IL may not have access to relations
which are not yet registered. When that relation is added later
during the DOM walk, and then the statement encountered again, the
original cached version is used and any benefit of the now known
relation is lost.
This was resolved by updating the timestamp of the 2 ssa-names in the
relation when it is registered. When a statement using those names is
re-examined, the use timestamp will be newer than the definition and the
statement will be recalculated with the relation available.
The problem shown here is a cycle of PHIs feeding each other... and each
one adding a relation potentially consumed by the other. This forces the
other to be recalculated. This was done whether a relation was actually
added or not. If no new relation is added, we should not update the
timestamps as there is no need to recalculate.
This patch changes the record() and register_relation() methods in the
both the relation oracle and fold_using_range classes . These routines
now return TRUE if a new relation was added, and FALSE otherwise. The
timestamps are only updated now if TRUE is returned.
In order for the debugging output to make sense, its been tweaked to
only print that a relation is registered, not that an attempt was being
made.
Bootstraps on x86_64-pc-linux-gnu with no regressions. OK?
One minor suggestion. Mention in comments before register_relation's
that false means nothing changed.
Likewise for add_partial_equiv and equiv_oracle::record and the others.
Oh yeah, forgot to add a comment didn't I ?? doh!
I will do that before committing
Andrew
