[PATCH] D36858: [x86] Teach the cmov converter to aggressively convert cmovs with memory operands into control flow.

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a comment.

In https://reviews.llvm.org/D36858#845378, @aaboud wrote:

> Thanks Chandler for preparing the patch, the implementation looks elegant, however, it overlooked a case where the memory registers are a result of a previous CMOV instructions.


Doh, of course. I'll add a remapping step to generating unfolded load. Should be easy because we know that the load is always on one side so can just remap to one of the inputs.

> This is a small reproducer that result in bad MIR:

Sweet!

> 
> 
>   target datalayout = "e-m:w-i64:64-f80:128-n8:16:32:64-S128"
>   target triple = "x86_64-unknown-unknown"
>   
>   define i32 @bar(i32* %a, i32* %b, i32 %n1, i32 %n2, i32 %d) #0 {
>   entry:
>     %cmp = icmp sgt i32 %n1, %n2
>     %s = select i1 %cmp, i32* %a, i32* %b
>     %p = getelementptr inbounds i32, i32* %s, i64 1
>     %load = load i32, i32* %p, align 4
>     %res = select i1 %cmp, i32 %d, i32 %load
>     
>     ret i32 %res
>   }
>   
>   attributes #0 = {"target-cpu"="skylake"}
>   




https://reviews.llvm.org/D36858