[PATCH] D35340: [x86] use more shift or LEA for select-of-constants

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

In https://reviews.llvm.org/D35340#826403, @spatel wrote:

> Sometimes we use movzbl and other times we use xor, and I don't know what causes the difference. IMO the improvements stand independent of that question/problem, but I plan to look at that next.


I stepped through a couple of the memcmp examples in 'X86 Fixup SetCC'. The difference is that we only transform to xor if we can find the instruction that def'd the flags in the same BB. In the memcmp examples, if there's a 'cmp' in the res_block, we'll get the xor transform, but if the flags are defined by a sub in a preceding block, we'll see movzbl instead. I think either way avoids partial-reg problems in these examples. I'm still looking at memcmp expansion improvements, so these differences may disappear for memcmp.


https://reviews.llvm.org/D35340