Implement core.checkedint in terms of LLVM intrinsics.

[redstar]

No description provided.

[dnadlinger]

I'm not sure whether |= is the best choice here. Whether a branch or load/or is worse pretty much depends on whether the overflow flag is in a register after inlining. See Walter's original PR for a discussion on this.

[redstar]

Thanks for the hint - I take a look at Walter's PR. I think the code generated by LLVM is pretty bad but it is difficult to create a good branch-free version. With the parameters in rax and rbx my spontaneous approach is:

mov    rdx, 1
add    rax, rbx
cmovnc rdx, overflow
mov    overflow, rdx

In the non-overflow case this creates a stall because of memory latency. I can avoid this stall by always reading memory:

mov    rcx, 1
mov    rdx, overflow
add    rax, rbx
cmovc  rdx, rcx
mov    overflow, rdx

In contrast the code with a branch is obvious:

add    rax, rbx
jnc    .L1
mov    overflow, 1
.L1:

[dnadlinger]

Yeah, and since overflow is probably rare in most applications, the branch should predict really well. All in all, I favored going with a branch-y version (that might be cleaned up by instruction selection anyway?) as the code is more straightforward and it seemed to be preferred by some microbenchmarks. In any case, I guess we'll see how it turns out once somebody actually uses these in a hot loop and we have some real use case to optimize for.