[clang] [CIR] Only emit FP math intrinsics when precision/errno settings allow it (PR #169424)

[Andy Kaylor via cfe-commits]

https://github.com/andykaylor commented:

This needs a test. The existing tests are using the default cc1 settings, which means math errno is disabled. You need a test that passes `-fmath-errno` on the RUN line and also has a separate RUN line without it. Calling one function would be sufficient. Also, note that the normal libm declarations go through this path as if they were builtins. So something like this would be good:

```
float cosf(float);

float test_normal(float f) {
  return cosf(f);
  // CIR-ERRNO: cir.call @cosf
  // CIR-NO-ERRNO: cir.cos
  // LLVM-ERRNO: call float @cosf
  // LLVM-NO-ERRNO: call float @llvm.cos.f32
  // OGCG-ERRNO: call float @cosf
  // OGCG-NO-ERRNO: call float @llvm.cos.f32
}

float test_precise(float f) {
#pragma float_control(precise, on)
  // Should never produce an intrinsic
  return cosf(f);
  // CIR-ERRNO: cir.call @cosf
  // CIR-NO-ERRNO: cir.call @cosf
  // LLVM-ERRNO: call float @cosf
  // LLVM-NO-ERRNO: cir.call @cosf
  // OGCG-ERRNO: call float @cosf
  // OGCG-NO-ERRNO: cir.call @cosf
}

float test_fast(float f) {
#pragma float_control(precise, off)
  // Should produce an intrinsic at -O1
  return cosf(f);
  // CIR-ERRNO-O1: cir.cos
  // CIR-NO-ERRNO-O1: cir.cos
  // LLVM-ERRNO-O1: call float @llvm.cos.f32
  // LLVM-NO-ERRNO-O1: call float @llvm.cos.f32
  // OGCG-ERRNO-O1: call float @llvm.cos.f32
  // OGCG-NO-ERRNO-O1: call float @llvm.cos.f32
}
```
Note that you'll need RUN lines that include `-O1` to get the fast pragma to have an effect.

https://github.com/llvm/llvm-project/pull/169424