Jan de Mooij [:jandem] Assignee
	Description • 13 years ago

Currently, we don't emit GNAME ops in strict mode. jsemit.cpp explains why:

 * This conversion is not made if we are in strict mode.  In eval code nested
 * within (strict mode) eval code, access to an undeclared "global" might
 * merely be to a binding local to that outer eval:
 *
 *   "use strict";
 *   var x = "global";
 *   eval('var x = "eval"; eval("x");'); // 'eval', not 'global'

The attached testcase measures global variable access, strict / non-strict:

d8       :  129 / 128
js -j    :  227 / 256
js -m -n :  427 /  84
js -m    :  514 / 511

Without TI there's not much of a difference but this really hurts TI+JM.

I don't know the emitter very well but it should be possible somehow to do this NAME -> GNAME optimization outside eval'ed code, right?

This affects scripts generated by Emscripten: it seems to use strict mode now and uses globals for the heap etc.

	Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)
	Comment 1 • 12 years ago

Is there any performance benefit to having emscripten use strict mode?  If so this would be a higher priority for us, but if it's just for strict-ness, then we can just not emit use strict..

	Alon Zakai (:azakai)
	Comment 2 • 12 years ago

In theory strict mode could be faster than normal mode, but there is no real benefit now except for debugging.

	Jan de Mooij [:jandem] Assignee
	Comment 3 • 11 years ago

Instead of never optimizing *NAME => *GNAME ops inside strict-mode code, this patch changes things so that we only stop doing this inside eval code. I now get the same numbers for the testcase in comment 0.

I didn't find a way to check for "eval or nested inside eval", so this patch adds a "bool insideEval" to BytecodeEmitter. Initially I checked whether "HandleScript evalCaller" is non-NULL, but this breaks the Debugger's evalInGlobal() (it passes a NULL evalCaller).

When I just comment out bce->sc->strict, multiple jit-tests start failing, so I don't think we need more tests for this.

	Jan de Mooij [:jandem] Assignee
	Comment 4 • 11 years ago

https://hg.mozilla.org/integration/mozilla-inbound/rev/39ff12be624b

	Ryan VanderMeulen [:RyanVM]
	Comment 5 • 11 years ago

https://hg.mozilla.org/mozilla-central/rev/39ff12be624b