[InstCombine] canonicalize shift+logic+shift to reduce dependency chain

Authored by spatel on Nov 7 2019, 9:08 AM.


[InstCombine] canonicalize shift+logic+shift to reduce dependency chain

shift (logic (shift X, C0), Y), C1 --> logic (shift X, C0+C1), (shift Y, C1)

This is an IR translation of an existing SDAG transform added here:

So we again have 9 possible patterns with a commuted IR variant of each pattern:

Part of the motivation is to allow easier recognition and subsequent
canonicalization of bswap patterns as discussed in PR43146:

We had to delay this transform because it used to allow the SLP vectorizer
to create awful reductions out of simple load-combines.
That problem was fixed with:
(we'll bring back load combining in IR someday...)

The backend is also better equipped to deal with these patterns now
using hooks like TLI.getShiftAmountThreshold().

The only remaining potential controversy is that the -reassociate pass
tends to reverse this kind of pattern (to help GVN?). But since -reassociate
doesn't do anything with these specific patterns, there is no conflict currently.

Finally, there's a new pass proposal at D67383 for general tree-height-reduction
reassociation, and it could use a cost model to decide how to optimally rearrange
these kinds of ops for a target. That patch appears to be stalled.

Differential Revision: https://reviews.llvm.org/D69842