1825791 - checkAndUpdateView is slower in SM than V8

Status: NEW

(Core :: JavaScript Engine: JIT, enhancement, P2)

Description

[Jeff Muizelaar [:jrmuizel]]

function benchmark() {
    var ViewAction = {};
    ViewAction.CreateViewNodes = 0;
    ViewAction.CheckNoChanges = 1;
    ViewAction.CheckNoChangesProjectedViews = 2;
    ViewAction.CheckAndUpdate = 3;
    ViewAction.CheckAndUpdateProjectedViews = 4;
    ViewAction.Destroy = 5;
    var Services = {
         updateDirectives(view, x) {},
         updateRenderer(view, x) {}
    }
    function markProjectedViewsForCheck(view) {}
    function execEmbeddedViewsAction(view, x) {}
    function execQueriesAction(view, x, y, z) {}
    function callLifecycleHooksChildrenFirst(view, x) {}
    function execComponentViewsAction(view, x) {}
    function callLifecycleHooksChildrenFirst(view, x) {}
    function checkAndUpdateView(view) {
        if (view.state & 1) {
            view.state &= ~1;
            view.state |= 2;
        } else {
            view.state &= ~2;
        }
        markProjectedViewsForCheck(view);
        Services.updateDirectives(view, 0);
        execEmbeddedViewsAction(view, ViewAction.CheckAndUpdate);
        execQueriesAction(view, 67108864, 536870912, 0);
        callLifecycleHooksChildrenFirst(view, 2097152 | (view.state & 2 ? 1048576 : 0));
        Services.updateRenderer(view, 0);
        execComponentViewsAction(view, ViewAction.CheckAndUpdate);
        execQueriesAction(view, 134217728, 536870912, 0);
        callLifecycleHooksChildrenFirst(view, 8388608 | (view.state & 2 ? 4194304 : 0));
        if (view.def.flags & 2) {
            view.state &= ~8;
        }
        view.state &= ~(64 | 32);
    }

    function run() {
        var start = performance.now();
        var view = { state: 0, def: { flags: 2} }
        for (var i = 0; i < 10000000; i++) {
            checkAndUpdateView(view);
        }
        var end = performance.now();
        console.log(`took ${end - start}ms`)
    }
    run()
}
benchmark()

Running the benchmark above with inlining disabled gives:
d8 --no-turbo-inlining: took 338.722ms
js --only-inline-selfhosted: took 418ms
jsc --maximumInliningDepth=1: took 402ms

We do quite bad with inlining enabled as well:
d8: took 39ms
js: took 173ms
jsc: took 69ms

Comment 1

[Jan de Mooij [:jandem]]

The checkAndUpdateView function is too large for us to inline (bytecode is 431 bytes). With --small-function-length=500 the shell test improves from 155 ms to 48 ms.

Inlining that function lets the compiler hoist a lot of code out of the loop.

The second issue is that our codegen for StoreFixedSlotT isn't great because we check if we need a pre-barrier etc even though that's unlikely when storing an int32 value. We might be able to figure that out at compile time with better MIR analysis.