[PATCH] D45867: [InstCombine] Unfold masked merge with constant mask

[Sanjay Patel via Phabricator via llvm-commits]

spatel accepted this revision.
spatel added a comment.
This revision is now accepted and ready to land.

LGTM - see inline comments.



================
Comment at: lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:2421-2423
+///   If A has one use, we want to canonicalize it to non-inverted mask:
 ///     ((x ^ y) & M) ^ x
+/// * If the M is constant, and A has one use, we want to unfold it:
----------------
We should state our reasoning here - because I know I won't remember for long. :)

So something like:
We can canonicalize by swapping the final xor operand to eliminate the 'not' of the mask.
If M is a constant, we transform to 'and' / 'or' ops because that shortens the dependency chain and improves analysis.


================
Comment at: lib/Transforms/InstCombine/InstCombineAndOrXor.cpp:2447
+    Value *LHS = Builder.CreateAnd(X, C);
+    Value *iC = Builder.CreateNot(C);
+    Value *RHS = Builder.CreateAnd(B, iC);
----------------
iC -> NotC


Repository:
  rL LLVM

https://reviews.llvm.org/D45867