On 18/08/14 10:19, Richard Earnshaw wrote:
On 14/08/14 09:45, Kyrill Tkachov wrote:
On 13/08/14 18:32, Segher Boessenkool wrote:
On Wed, Aug 13, 2014 at 03:57:31PM +0100, Richard Earnshaw wrote:
The problem with the frankenmonster patterns is that they tend to
proliferate into the machine description, and before you know where you
are the back-end is full of them.
Furthermore, they are very sensitive to the greedy first-match nature of
combine: a better, later, combination is missed because a less good,
earlier, optimization matched. If the first insn in the sequence is
merged into an earlier instruction then you can end up with a junk
sequence that completely fails to simplify. That ends up with
super-frankenmonster patterns to deal with all the subcases and the
problems grow exponentially from there.
Right. Of course, combine should be fixed, yadda yadda.
I really do think that the best solution would be to try and catch this
during expand if possible and generate the right pattern from the start;
then you don't risk combine failing to come to the rescue after several
intermediate transformations have taken place.
I think ssa-phiopt should simply not do this obfuscation at all. Without
it, RTL ifcvt picks it up just fine on targets with conditional assignment
instructions. I agree on targets without expand should do a better job
(also for more generic conditional assignment).
That particular transformation was added to tree-ssa-phiopt.c for PR
45685, the problem it was trying to solve was a missed vectorisation
opportunity and transforming it made it into straightline code that was
more amenable to vectorisation, that's why I'm somewhat reluctant to
completely disable it.
Hmm... I noticed in the midend we guard some optimisations on
HAVE_conditional_move. Maybe we can guard this one on something like
!HAVE_conditional_negation ?
Can't we just guard it on HAVE_conditional_move? With such an
instruction expand would then generate
t1 = -a
r = <cond> ? b : t1
and combine will do the rest.
That was my first idea, but then it disables this transformation for
x86, for which it was added
specifically to solve PR45685...
Kyrill
R.
Kyrill
Instruction selection belongs in RTL land.
Segher
