[clang] [llvm] [ValueTracking] use KnownBits to compute fpclass from bitcast (PR #97762)

[Yingwei Zheng via cfe-commits]

dtcxzyw wrote:

> This patch is motivated by an internal benchmark where there were some cases where this helped, though even that case is in some sense artificial.

It is ok to optimize some pattern in proxy apps/benchmarks :)

> Is this a necessary criteria for landing this change?

As recursively computing known bits/fpclass is expensive, it is not worth adding additional complexity unless it is useful for real-world applications.

> I believe we already handle float to int in KnownBits

I posted https://github.com/llvm/llvm-project/pull/86409 because I found it enabled more optimizations in [my ir dataset] (https://github.com/dtcxzyw/llvm-opt-benchmark/pull/451/files).
 
> and adding the inverse in KnownFPClass seems like a correct and reasonable extension of the logic, ***even if there are not many cases where it is used.***

Don't do this. It is a trap of code coverage. We need to strike a balance between optimizations, compilation time and maintainability.

For more details, see https://llvm.org/docs/InstCombineContributorGuide.html#real-world-usefulness.


https://github.com/llvm/llvm-project/pull/97762