Page MenuHomePhabricator

[SLP] Add a check before skipping inserting a trunc after z|sext vectorization tree root
AbandonedPublic

Authored by haicheng on Mar 16 2018, 7:18 AM.

Details

Reviewers
mssimpso
ABataev
Summary

If the vectorization tree root operations are integer cast operations, D41948 does not consider them as candidates for truncation. However, sometimes the trunc is needed by instcombine to trigger type shrinking or prevent type extension.

One example is like this, when vectorizing

  %800 = zext <16 x i16> %799 to <16 x i32>
  %803 = zext <16 x i16> %802 to <16 x i32>
  %804 = sub nsw <16 x i32> %800, %803
  %805 = extractelement <16 x i32> %804, i32 0
  %806 = sext i32 %805 to i64
  %808 = extractelement <16 x i32> %804, i32 1
  %809 = sext i32 %808 to i64
...
  %850 = extractelement <16 x i32> %804, i32 15
  %851 = sext i32 %850 to i64

to

  %800 = zext <16 x i16> %799 to <16 x i32>
  %803 = zext <16 x i16> %802 to <16 x i32>
  %804 = sub nsw <16 x i32> %800, %803
  %805 = sext <16 x i32> %804 to <16 x i64>
  %806 = trunc <16 x i64> %805 to <16 x i32>
  %807 = extractelement <16 x i32> %806, i32 0
  %808 = extractelement <16 x i32> %806, i32 1
...
  %837 = extractelement <16 x i32> %806, i32 15

A trunc after the vectorized sext is needed to prevent the types of the first three IRs (2 zext and 1 sub) from being expanded to i64.

This patch checks if the sources of the vectorized cast operations can be changed to evaluate in different types. If so, the trunc is still added.

Diff Detail

Repository
rL LLVM

Event Timeline

haicheng created this revision.Mar 16 2018, 7:18 AM

Not sure that this should be handled in SLPVectorizer. It is the problem of the instcombiner that it is too optimistic sometimes, not the problem of the SLPVectorizer. InstCombiner should check that this kind of the expansion is good for the current platform

Matt, what do you think?

haicheng abandoned this revision.Mar 24 2018, 9:54 AM

Abandon in favor of D44868.