Fixes#151786
The original `ceilf` expansion lowers to `fptosi`, which produces poison
for Inf, and any subsequent use leads to undefined behavior. This patch
adds a safe path, similar to the existing `round` expansion, for large
or special inputs and avoids the UB.
The `math.clampf` op is semantically incorrect when compared to both the
CUDA reference implementation and the SPIRV spec, both of which have a
clamp op.
- Fix the definition of `math.clampf` to agree with CUDA and SPIRV
- Explicitly state when `math.clampf` produces `ub.poison`
- Update the ExpandOps pass to reflect the corrected semantics
This patch adds the `clampf` operation to the math dialect. The
semantics op are defined as:
```
clampf(x, min_v, max_v) = max(min(x, min_v), max_v)
```
The reasoning behind adding this operation is that some GPU vendors
offer specialized intrinsics for this operation, or subsets of this
operation. For example,
[__saturatef](https://docs.nvidia.com/cuda/cuda-math-api/cuda_math_api/group__CUDA__MATH__INTRINSIC__SINGLE.html#group__cuda__math__intrinsic__single_1ga2c84f08e0db7117a14509d21c3aec04e)
in NVIDIA GPUs, or `__builtin_amdgcn_fmed3f` in AMD GPUs.
This patch also removes `test-expand-math` in favor of
`math-expand-ops`.
Finally, it removes individual expansion population API calls like
`populateExpandCoshPattern` in favor of:
```C++
void populateExpansionPatterns(RewritePatternSet &patterns,
ArrayRef<StringRef> opMnemonics = {});
```
