Add more FP-rounding LLVM intrinsics

[kinke]

Resolves #4754.

[kinke]

Oof, once again buggy intrinsic semantics by the looks of it (definitely not 'does what the libm function would do, except for NOT setting errno'), at least when LLVM has to resort to x87 (32-bit x86 not defaulting to SSE etc.). The Win64 std.mathspecial failure is probably related to its 64-bit real.

[kinke]

Turns out the LLVM docs are wrong; there is a difference between llvm.lround.i64 and llvm.llround.i64 after all. In the -mtriple=x86_64-windows-msvc case, calling the lround vs. llround MSVCRT function. This fixed the 32-bit x86 issues too.

[JohanEngelen]

Turns out the LLVM docs are wrong; there is a difference between llvm.lround.i64 and llvm.llround.i64 after all.

Out of interest, what exactly is wrong in the LLVM docs?

[kinke]

The syntax section uses llvm.lround.i64 exclusively.