788476 - IonMonkey: Fold x & -1 to x

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Jan de Mooij [:jandem]]

Attachment: Patch

As Alon said in bug 788166 comment 2, x & -1 is a common way to convert values to int32. This patch folds them just like we do with x|0.

(Note that there was already an unused foldIfNegOne method).

Comment 1

[Marty Rosenberg [:mjrosenb]]

My bad, I half-added this code many moons ago, but didn't actually add it, since it was just one more place to fail, and none of our benchmarks actually use &-1.

Comment 2

[David Mandelin [:dmandelin]]

(In reply to Marty Rosenberg [:mjrosenb] from comment #1)
> My bad, I half-added this code many moons ago, but didn't actually add it,
> since it was just one more place to fail, and none of our benchmarks
> actually use &-1.

Do we know of anything that does? I have no objection to the optimization--I know Mandreel uses this pattern--but it would be fun to see the effect on something.

Comment 3

[Alon Zakai (:azakai)]

Mandreel used to use it everywhere, but it looks like they stopped, it isn't in the octane-mandreel benchmark. I'm not aware of another compiler or VM using this pattern (but I think it's nice to optimize it for consistency just so if some new project uses &-1 there isn't a surprising slowdown compared to |0).

Comment 4

[Jan de Mooij [:jandem]]

The interpreter roc wrote (bug 608733) does something like this a lot:

    regs[xx] = (regs[xx] + regs[yy]) & 0xFFFFFFFF;

0xFFFFFFFF (double) is folded to -1 (int32).

If I change the call to testCompiler to testInterpreter in the shell testcase in bug 608733, the time drops from 1711 to 1560 ms with the patch. Strangely enough I see only a very small win on the (slightly modified) testcase I added to awfy-assorted, not sure why (iteration count is a bit lower but still high (50000000) and the numbers are comparable otherwise).

Comment 5

[David Mandelin [:dmandelin]]

Cool. I learned something by asking a question, even if it wasn't what I expected. :-) And 8% is a big win on the interpreter for such a small thing.

Comment 6

[Jan de Mooij [:jandem]]

https://hg.mozilla.org/projects/ionmonkey/rev/54711415fb53