Round-to-zero float32 conversion not supported

[int3]

Small repro case:

import triton
import triton.language as tl
import torch


@triton.jit
def type_convert_triton(src, dst, rounding: tl.constexpr):
    range = tl.arange(0, 128)
    x = tl.load(src + range)
    y = x.to(dst.dtype.element_ty, fp_downcast_rounding=rounding)
    tl.store(dst + range, y)


src = torch.zeros([128], dtype=torch.float32)
dst = torch.empty(src.shape, dtype=torch.float16)
type_convert_triton[(1, )](src, dst, rounding='rtz')

[ienkovich]

I tried to enable this one some time ago and I thought I hit a LLVM bug but it appeared that arith::TruncFOp lowering doesn't guarantee us a proper rounding mode. It generates llvm.experimental.constrained.fptrunc but this intrinsic doesn't control rounding, it only hints the compiler about runtime rounding settings to be used: llvm/llvm-project#96815
I think we should lower it directly to vcvtps2ph intrinsic calls with explicit rounding to make it work.