for review (M): 6812831 factor megamorphic invokeinterface (for 6655638)

[John Rose]

On Mar 6, 2009, at 11:11 AM, Vladimir Kozlov wrote:

> vtableStubs_sparc.cpp:
> Why you changed 2* to 6* for CompressedOops? There is only 1  
> load_klass
> and with CompressedOops it need only 2 additional instructions  
> (shift;add)
> to decode it.
>
> -      // save, ld, ld, sll, and, add, add, ld, cmp, br, add, ld,  
> add, ld, ld, jmp, restore, sethi, jmpl, restore
> -      const int basic = (20 LP64_ONLY(+ 6)) * BytesPerInstWord +
> -                        // shift;add for load_klass
> -                        (UseCompressedOops ? 2*BytesPerInstWord : 0);
> +      const int basic = (28 LP64_ONLY(+ 6)) * BytesPerInstWord +
> +                        // shift;add for load_klass, decode, and  
> encode
> +                        (UseCompressedOops ? 6*BytesPerInstWord : 0);

You are right.  Frankly, I don't remember how that diff got in there;  
maybe just an excess of caution.  I'll nuke it.

> vtableStubs_x86_32.cpp:
> Did your changes increase the size or it was wrong before?
> Or it is upper/rounded limit on the size?
>
> -    return (DebugVtables ? 144 : 64) + (CountCompiledCalls ? 6 : 0);
> +    return (DebugVtables ? 256 : 66) + (CountCompiledCalls ? 6 : 0);

My changes increased the size slightly, because I peeled the loop a  
little bit for speed.

I found those numbers (like 28, 66) by running with PrintMisc.  The  
new print logic tells exactly how many instruction bytes are wasted.   
I aimed at a slop of 4 bytes, so as not to overflow when really large  
itable offsets are encountered.  I'll note this in the comments before  
I push!

Thanks,

-- John