[negative result] B3 should have generalized path specialization

[Filip Pizlo]

Patch forthcoming.

I don't know if this will be profitable yet.  Reading IR in Octane, it seems like it might be.  Only one way to find out!

[Filip Pizlo]

Created attachment 338939 [details]
it's a start

[Filip Pizlo]

Created attachment 338952 [details]
a bit more

[Filip Pizlo]

Created attachment 338954 [details]
this phase is nuts

[Filip Pizlo]

Created attachment 338965 [details]
I'm slowly putting this phase together

[Filip Pizlo]

Created attachment 339038 [details]
close to written

[Filip Pizlo]

Created attachment 339050 [details]
it might be written

[Filip Pizlo]

Created attachment 339075 [details]
it's starting to compile

[Filip Pizlo]

Created attachment 339087 [details]
it specialized something without crashing or emitting broken code

It was a small function, too small to cause interesting compiler problems.

[Filip Pizlo]

It's a 14% speed-up on the following benchmark. Note that we're duplicating a CheckAdd (which has a full OSR thingy). Note that we're not (yet) unswitching the loop, because the structure load is not hoisted out of the loop.

function foo(o) {
    var result = 0;
    for (var i = 0; i < 100; ++i) {
        result += o.f;
        o.f = i;
    }
    return result;
}

noInline(foo);

var before = preciseTime();
for (var i = 0; i < 100000; ++i) {
    var result = foo(i & 1 ? {f:42} : {e:41, f:42});
    if (result != (99 * 98) / 2 + 42)
        throw "Error: bad result: " + result;
}
var after = preciseTime();
print("That took " + (after - before) * 1000 + " ms");

[Filip Pizlo]

Created attachment 339104 [details]
it's starting to come to life

[Filip Pizlo]

Current perf data suggests that this is not a speed-up.  Will verify again once I fix some bugs.  But this is looking like a negative result, mostly because taildup is too darn good.

[Filip Pizlo]

Created attachment 339163 [details]
mostly works, not a speed-up

Still has test failures.  But it's definitely not a speed-up, so I'm going to stop working on it.