[PATCH] D17288: [CodeGenPrepare] Do select to branch transform when cmp's operand is expensive.

[Mehdi Amini via llvm-commits]

> On Feb 17, 2016, at 9:05 AM, Sanjay Patel <spatel at rotateright.com> wrote:
> 
> spatel added a comment.
> 
> In http://reviews.llvm.org/D17288#354233, @flyingforyou wrote:
> 
>> Thanks Sanjay for comment.
>> 
>> We discussed about load-cmp heuristic in http://reviews.llvm.org/D16836. This is something likes load-cmp heuristic.
> 
> 
> Yes - I can see that this would be a similar heuristic. The question we have is whether the load-cmp heuristic itself can be improved by moving it lower or using more data to make it smarter. Therefore, I don't think we should add another heuristic-based transform here right now.
> 
> As an example, I don't think this transform or the load transform would help an in-order CPU of any architecture. But we're still doing the transform for those subtargets.
> 
>> Even if branch prediction is failed, we may not lost anything. Because fdiv float/double takes more cycles than branch prediction miss penalty.
> 
>> And if branch prediction is correct, we can hide fdiv's execution cycles.
> 
>> 
> 
>> I think this is logically correct. But we need to test more. (test-suite, spec, commertial benchmarks...)
> 
> 
> I think I understand your point, and certainly it looks logically correct. But unless there is some good benchmark evidence to support the heuristic, I don't think we should add it here. If others disagree or if I'm misunderstanding, please let me know.

FWIW I have the same opinion as Sanjay on this topic.

-- 
Mehdi