Bug 16525 – Possible to optimize out for(var x = 0; x < 500; x++) in JSC?

• Status: RESOLVED
• Resolution: INVALID
• Component: JavaScriptCore

[Eric Seidel (no email)]

Possible to optimize out for(var x = 0; x < 500; x++) in JSC?

I wonder if it wouldn't be possible to completely optimize out the variable handling for "x" in the above statement in JavaScriptCore.  Meaning, pull x completely out of JS and make it be only on c-stack.

This would require two things:
1.  Walk the for subtree and make sure x is never referenced.
2.  Assure that "x < 500" is a statement involving only x and constants.

Requirement #1 could be relaxed if we actually exposed x in the scope chain.

This sort of optimization might be way overly complicated at this point.  But I saw LessNode::evaluateToNumber() as 1.2% as well as LocalVarAccessNode::evaluateToNumber() as 1.6% on the sample tonight, and started to wonder.

[Eric Seidel (no email)]

Along this same vein: our time spent in LocalVarAccessNode ::evaluateToNumber() in pushing the float value onto the stack for return.  We could avoid calling evaluateToNumber() entirely if we could know that evaluateToInt32() would be sufficient.  Yay x68 calling conventions!

[Mark Rowe (bdash)]

How would you account for x being referenced inside an eval?

[Eric Seidel (no email)]

Any use of "eval" would count as a reference to "x" and would break this optimization.  This would obviously require our planned eval tainting.

[Maciej Stachowiak]

Note that a reference to x could be anywhere inside the same function or in nested function declarations or expressions.

[Eric Seidel (no email)]

Which would mean we'd need to find the enclosing ScopeNode and walk the entire subtree.  ick.

[Cameron Zwarich (cpst)]

This is no longer relevant after SquirrelFish.