[clang] [clang] Lower non-builtin sincos[f|l] calls to llvm.sincos.* when -fno-math-errno is set (PR #121763)

[Benjamin Maxwell via cfe-commits]

MacDue wrote:

> > > > @RKSimon do you know if there is a reason CodeGen AA is not enabled for x86?
> > > > I'd otherwise be happy to approve this patch, lowering to the intrinsic seems like the right thing to do.
> > > 
> > > 
> > > I don't remember any reason for alias analysis to not be enabled on x86 - @phoebewang @topperc do you know?
> > 
> > 
> > I don't remember either.
> 
> Me neither. @clin111 do you happen to know it?

I've created a patch enabling it here: https://github.com/llvm/llvm-project/pull/123787 (maybe you'll spot why it's disabled there :sweat_smile:). I've updated the generated tests, and gone over the handwritten ones more carefully, and updated them myself.

The handwritten tests that changed were:
```
llvm/test/CodeGen/X86/cfguard-x86-vectorcall.ll
llvm/test/CodeGen/X86/fixup-bw-inst.ll
llvm/test/CodeGen/X86/misched-matrix.ll
llvm/test/CodeGen/X86/regalloc-advanced-split-cost.ll
llvm/test/CodeGen/X86/vectorcall.ll
llvm/test/CodeGen/X86/win32-eh.ll
``` 

Someone more experienced with x86(_64) should double check the changes make sense. I am fairly confident the underlying transforms should be correct though as this feature is enabled for other targets (at least AArch64, RISC-V, off the top of my head).

https://github.com/llvm/llvm-project/pull/121763