I am working on an internal data structure overhaul in the vectorizer to
support future changes. The first task is to remove the non-SLP
representation which will simplify code and improve maintainability.
This first task is actually the most difficult and time-consuming one.
I have pushed the current state of my work as a series of commits
onto the rguenth/forceslp git branch. The first commits in the series
are improvements/fixes for trunk I plan to commit once stage1 opens
again. The real changes done sofar are
- allow group_size == 1 SLP instances
- force vectorization fail when non-SLP stmts are left as debugging
tool to see what remains to be done
- make the SLP graph cyclic to handle cycles in inner loops with
outer loop vectorization
- start to rewrite reduction handling - I've split out nested
cycles handling sofar which now is much simpler and handles
almost arbitrary inner loops
vectorizer testsuite summary on x86_64 is
=== gcc Summary ===
# of expected passes 3103
# of unexpected failures 85
# of unexpected successes 10
# of expected failures 52
# of unsupported tests 30
=== g++ Summary ===
# of expected passes 255
# of expected failures 6
# of unsupported tests 3
=== gfortran Summary ===
# of expected passes 131
# of unexpected failures 8
# of unsupported tests 1
where a PASS of course doesn't mean the generated code has the same
quality than before. The main remaining issues are reduction handling
(esp. condition reduction and other cases that were never implemented
for SLP) plus interleaving cases that cannot be handled within the
current SLP permutation framework (require more than two source vectors).
For the interleaving issue I have to tackle part two of the data structure
overhaul and rework permutation handling to be represented more explicit
in the SLP graph and make sure we do not ever need to throw away / dissolve
DR groups.
There are various vectorizations that lack SLP support on current trunk
and you are free to contribute missing pieces there. I hope to not get
too many extra road-blocks from changes done on trunk, so please do not
add new features w/o considering SLP support ;)
Thanks,
Richard.
