[llvm] AMDGPU: Replace sqrt OpenCL libcalls with llvm.sqrt (PR #74197)

[Stanislav Mekhanoshin via llvm-commits]

rampitec wrote:

Speaking of the the "unsafe-fp-math" attribute, clang will produce it for the following options:

-cl-fast-relaxed-math
-cl-unsafe-math-optimizations
-ffast-math
-funsafe-math-optimizations AND -ffp-contract=fast (but not with -ffp-contract=on or off)

Then if we are not relying on the attribute anymore, either afn or !fpmath >= 2.0 shall be set on a call site with the same list of options. I am not sure if and why a call to @_Z4sqrtf would have any of these.

https://github.com/llvm/llvm-project/pull/74197