[llvm] [InstCombine] Eliminate fptrunc/fpext if fast math flags allow it (PR #115027)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

After looking into this some more I don't think instcombine is the right place to fix this. Different targets handle fp16 differently in clang, so e.g. with aarch64 we have "fadd half" generated by clang directly so there's no fptrunc/fpext to eliminate. And instcombine also has a transform to convert "fpext, fadd float, fptrunc" info "fadd half", which means this instcombine transformation doesn't even work on the original test case I was looking at.

Instead I've created #131345 to do this in dagcombine.

https://github.com/llvm/llvm-project/pull/115027