This is an archive of the discontinued LLVM Phabricator instance.

[ValueTracking] Compute known bits for add/sub with less temporary APInts
AbandonedPublic

Authored by craig.topper on Mar 24 2017, 12:10 PM.

Details

Summary

The previous operation used quite a few temporary objects which can allocate memory for larger than 64-bit types and made no use of the storage already allocated storage for KnownZeroes/Ones.

This patch tries to make more intelligent reuse of the objects. Computing the overall known bits is still kind of bad as the OR operation for LHS and RHS both create a temporary to hold the OR result.

Diff Detail

Event Timeline

craig.topper created this revision.Mar 24 2017, 12:10 PM
craig.topper abandoned this revision.Mar 25 2017, 9:09 PM