Re: [COMMITTED] testsuite/gcc.target/cris/pr93372-2.c: Handle xpass from combine improvement

[Segher Boessenkool <segher@kernel.crashing.org>]

Hi!

On Fri, Apr 05, 2024 at 04:06:01AM +0200, Hans-Peter Nilsson wrote:
> The xpassing change in generated code was as follows, at
> r14-9788-gb7bd2ec73d66f7 (where I locally applied a revert
> to verify that this suspect was the cause).  That was so
> much of an improvement that I had to share it!  Worth the
> testsuite churn anyway. :)
> 
> Segher, if you end up reverting r14-9692-g839bc42772ba7a (as
> unfortunately seems not unlikely), then please also revert this
> commit: r14-9799-g4c8b3600c4856f7915281ae3ff4d97271c83a540.

I won't revert it, it fixes an actual bug.  Not a regression no, but a
very serious longstanding problem.

We have accidentally done a limited version of a feature requested for
more than 20 years now, "UNCSE".  I'll do this for just combine (instead
of as a separate pass, lots of issues there with pass ordering, results
could be better though) in GCC 15.  This really is a stage 1 thing
though!

Any testcase that relies on something that combine does not promise and
that can not reasonably be expected to always hold is *buggy*.

The combine-2-2 testcase (that I wrote myself) isn't very good, and
should be replaced by something that is much more clearly a 2->2
combination, instead of 1->1 with context.

All (target-specific) new testsuite failures are just like that: bad
testcases!

So no, no reversion.

> That's the only test that's improved to the point of
> affecting test-patterns.  E.g. pr93372-5.c (which references
> pr93372-2.c) is also improved, though it retains a redundant
> compare insn.  (PR 93372 was about regressions from the cc0
> representation; not further improvement like here, thus it's
> not tagged.  Though, I did not double-check whether this
> actually *was* a regression from cc0.)

Interesting that this improved tests for you.  Huh.  Do you have an
explanation how this happened?  I suspect that as uaual it is just a
side effect of random factors: combine is opportunistic, always does the
first change it thinks good, not considering what this then does for
other possible combinations; it is greedy.  It would be nice to see
written out what happens in this example though :-)


Segher