[llvm] [X86] Faster truncf and roundf on x86 SSE2 (PR #226513)

[Divyansh Yadav via llvm-commits]

schizophrenicmaniac wrote:

@RunDevelopment @lntue Great points. LLVM doesn't have a mechanism for "flexible constants", we have to pick a concrete value at lowering time. Given that any value in [2^23, 2^31] is correct for float and [2^52, 2^63] for double, I'll switch to 2^31 / 2^63 since it maximizes the range where known-range optimizations can eliminate the branch. The correctness is identical either way. Let me know if I am wrong otherwise.

https://github.com/llvm/llvm-project/pull/226513