This is an archive of the discontinued LLVM Phabricator instance.

[InstCombine] Improve the expansion in SimplifyUsingDistributiveLaws to handle cases where one side doesn't simplify, but the other side resolves to an identity value
ClosedPublic

Authored by craig.topper on Jul 15 2017, 12:05 AM.

Details

Summary

If one side simplifies to the identity value for inner opcode, we can replace the value with just the operation that can't be simplified.

I've removed a couple now unneeded special cases in visitAnd and visitOr. There are probably other cases I missed.

@spatel I believe this fixes your hidden not test cases with compares. If you send over your test cases I can add them to this patch.

Diff Detail

Repository
rL LLVM

Event Timeline

craig.topper created this revision.Jul 15 2017, 12:05 AM
spatel edited edge metadata.Jul 15 2017, 6:48 AM

Thanks. I hadn't gotten to the first pattern that you're fixing/deleting, but these tests should cover the 2nd and 3rd.

grandinj added inline comments.
lib/Transforms/InstCombine/InstructionCombining.cpp
653 ↗(On Diff #106755)

the do -> they do

662 ↗(On Diff #106755)

the do -> they do

690 ↗(On Diff #106755)

ditto

699 ↗(On Diff #106755)

ditto

Added Sanjay's compare test cases.

spatel accepted this revision.Jul 15 2017, 1:12 PM

There's an open question about whether InstCombine should be in the refactoring/distribution optimization business in the first place. However, given that this patch both improves that folding and reduces specialized matchers, it will make it easier to lift this code into a different pass when we decide to move this.

LGTM.

lib/Transforms/InstCombine/InstructionCombining.cpp
651 ↗(On Diff #106776)

"Do" -> "Does" (and same for the comments below)

653 ↗(On Diff #106776)

"They do!" -> "It does!" (and repeat below)

test/Transforms/InstCombine/and.ll
698 ↗(On Diff #106776)

Should be "Commute the 'and':"
Also reverse the other similar comments; these got inverted from the 'or' tests.

753 ↗(On Diff #106776)

This is a copy of my mistake in the original comment, so it's repeated 8 times I think: the result is actually (~X & Y).

This revision is now accepted and ready to land.Jul 15 2017, 1:12 PM
This revision was automatically updated to reflect the committed changes.