As a first step, this patch arranges for the result of an ABI call to be transferred from the x87 stack to the xmm0 register (if the result is a double, and we're on x86, the x64 ABI already puts the result in xmm0 by default). Before I continue, can somebody take a look at this patch and see if it makes sense?

Comment on attachment 587910 [details] [diff] [review] Move double result to xmm0 on x86 Review of attachment 587910 [details] [diff] [review]: ----------------------------------------------------------------- ::: js/src/ion/x86/MacroAssembler-x86.h @@ +74,5 @@ > // callee. It returns the space which has to be allocated for calling the > // function. > // > // arg Number of arguments of the function. > + uint32 setupABICall(uint32 arg, bool resultIsDouble = false); Two suggestions - could this be moved to callWithABI? then you don't have to save it locally, and don't need to change stack computation (you'd just reserve/free in the same place). It'd also be nice to have an enum input instead of a boolean (it's hard to tell what a boolean means). Like Result_Normal, Result_Double or something.

What is the status here?

(In reply to Paul Wright from comment #3) > What is the status here? I'm working on this bug. It's currently blocked on bug 709423: we want to use the fmod function to implement the floating-point version of JSOP_MOD, but for this to work, the callWithABI API needs to be able to pass double arguments to a function, and return a double result. I just filed a patch that does exactly that for x86 and x64. Once it gets r+'d, I intend to finish this the implementation of JSOP_MOD for those platforms. The plan is to open a separate bug for the ARM version. Hopes that answers your question.

Yes, it does. Thank you!

Pretty straight-forward now that we can use doubles with callWithAbi. With --no-jm this wins about 5-10% on v8-raytrace.

Now marked as call instruction..