[llvm] r195496 - X86: Perform integer comparisons at i32 or larger.

[Rafael Espíndola]

> I would expect conclusive benchmark results on a large corpus of code tested
> on a broad spectrum of microarchitectures from both Intel and AMD before
> making a change like this. Also, the code size impact needs to be
> quantified; having to break out to a whole new instruction for what would
> otherwise be an immediate has the potential for insane code-size increase
> (clang has something like 30,000 cmpb's and it seems like almost all of them
> have immediate memory operands). Also, not being able to fold a load into
> the comparison will likely increase register pressure.

I tested this with a .bc from a LTO build of clang. The size of .text
goes from 35468517 bytes to 35589941 bytes, so 1.0034 times larger,
which seems a pretty small code growth.

Cheers,
Rafael