[AIE2P] add missing Vextr Bcst Pattern and harmonize G_ASSERT_[S/Z]EXT removal. #677
Conversation
F-Stuckmann
requested review from
abhinay-anubola,
abnikant,
andcarminati,
katerynamuts,
khallouh,
konstantinschwarz,
martien-de-jong,
niwinanto,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
| %483:_(s32) = COPY $r6 | ||
| %484:_(s32) = G_AIE_ZEXT_EXTRACT_VECTOR_ELT %4(<16 x s32>), %483(s32) | ||
| %485:_(s32) = G_ASSERT_SEXT %484, 16 | ||
| %486:_(s32) = G_ASSERT_ZEXT %485, 16 |
There was a problem hiding this comment.
Are we really produces this IR? Is looks illegal: the G_AIE_ZEXT_EXTRACT_VECTOR_ELT guarantees to zero-extend the vector element it extracts (or doesn't do any extension if the element type size matches the destination type size).
Having two asserts that claim the same value is both sign and zero extended looks wrong:
%485:_(s32) = G_ASSERT_SEXT %484, 16
%486:_(s32) = G_ASSERT_ZEXT %485, 16
Note that G_ASSERT_SEXT/G_ASSERT_ZEXT does not perform any operation. It simply tells the optimizers that they can remove redundant extensions
There was a problem hiding this comment.
Yes we produce it unfortunately.
The second unit test (VEXTBCST_SEXT_CHAIN) is constructed to test the symmetry of the G_ASSERT_[S/Z]EXT removal.
There was a problem hiding this comment.
Can you investigate how it ends up in this situation? I don't think this is correct
Minor Performance improvements: -0.03%