[PATCH] D136015: [InstCombine] Fold series of instructions into mull

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

We're not creating a new multiply that is wider than we started with, so I'm assuming codegen can't be worse.
As mentioned earlier, the code should be generalized to handle any even bitwidth; we don't want highly type-specific transforms in IR canonicalization.
https://alive2.llvm.org/ce/z/2BqKLt

The commutative pattern matching doesn't look correct at first glance, so we need tests that exercise all of those possible patterns. The instructions with constants will always have the constant as operand 1, so you don't need to worry about those. But the 3 muls and 2 adds can all be commuted, so that's 16 potential patterns?

Since we are only creating a single new instruction, there's no need to check for m_OneUse on any of the existing values (but we should include at least one test with extra uses to show that works as expected).


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136015/new/

https://reviews.llvm.org/D136015