[llvm] [CodeGen] [AMDGPU] Attempt DAGCombine for fmul with select to ldexp (PR #111109)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I updated with new test cases as well as the old ones that needs changes, but for some reason, the only below testcase
> 
> > llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-pow-codegen.ll
> 
> It is giving invalid bitcast error **"from i16 to i32"** while trying to update with **update_llc_tests_check.py** script. But, bizarelly when running the command sequence of the test in terminal its working fine, generating assembly as supposed to be with dagcombine happening. I am perplexed!!

I don't know how you could hit this bitcast issue, but should fix however it is getting it.

This test is odd in that it has an llc line with an opt invocation, which update_llc_test_checks does not understand. In any case, there should be no path to hitting an invalid bitcast 

https://github.com/llvm/llvm-project/pull/111109