586232 - IonMonkey: Investigate an optimal solution for branchTruncateDoubleToInt32 for ARM.

Status: NEW

(Core :: JavaScript Engine, enhancement, P3)

Description

[Jacob Bramley [:jbramley]]

Refer to bug 585918 for a short discussion.

In jsnum.h's DoubleToECMAInt32 function, a decent performance improvement could be gained by using a soft-float routine to cover all the corner-cases of ECMA's "ToInt32" function. (VFP can't do the whole operation by itself.) branchTruncateDoubleToInt32 might behave differently. Bug 585918 added a VFP-based conversion routine which falls back to a slow path in case of overflow, but this might not be the best solution.

Comment 1

[Till Schneidereit [:till]]

Still relevant for IM?

Comment 2

[Marty Rosenberg [:mjrosenb]]

IM still needs to convert doubles to int32s.  I suspect we call into the interpreter a lot less, but it still happens. Also, I have spent about 0 time optimizing the IM code that is generated for these operations.  I'm sure we can get a small gain by actually getting that code closer to optimal.

Comment 3

[Till Schneidereit [:till]]

Ok, thanks for the info. Updating description and resetting assignee. @jbramley, please don't let that keep you from working on this, of course. ;)

Comment 4

[Matthew Gaudet (he/him) [:mgaudet]]

Appears we're still pretty much using the implementation from bug 585918