🚀 go-pugleaf

Compilers can do well, but I have found it trivial to consistently beat the
latest and greatest in local optimizations.

If you are not overly concerned with portability, I would still suggest
dynamic recompilation. Even inefficient native code will execute an order of
magnitude faster than emulation.

-Matt

"Vis Mike" <visionary25@_nospam_hotmail.com> wrote in message
news:b2d6f7$fne$1@bob.news.rcn.net...
> Actually, I get worse performance the more aligned the jumps! (8, 16
bytes).
> I give up, I'll leave the details to the person who knows best - the
> compiler writer :)
>
> Mike
>
> "Vis Mike" <visionary25@_nospam_hotmail.com> wrote in message
> news:b2d1sj$8j4$1@bob.news.rcn.net...
> > I found a nice feature in gcc - -malign-jumps=n
> >
> > > wcc -O3 -malign-jumps=3 vmtest2.cpp
> > > a.exe
> > Testing stack4...
> > &&addi = 0x401180 +4 +8 +16
> >
> > 0x401180 is the address of the first label, followed by the offset of
the
> > rest of the labels.  So if I can squeeze the code in, all I need is a
"jmp
> > (opcode*4)".  This is sooo unportable :)  I'll just use conditional
> > compilation.
> >
> > I don't quite understand gcc's inline assembly syntax, any help for the
> jmp?
> > Or am I just going overboard?
> >
> > Cheers!
> > Mike
> >
> > "James Van Buskirk" <not_valid@attbi.com> wrote in message
> > news:dCk2a.80786$HN5.307889@rwcrnsc51.ops.asp.att.net...
> > >
> > > "Matt Taylor" <para@tampabay.rr.com> wrote in message
> > > news:h2j2a.61691$AQ3.981480@twister.tampabay.rr.com...
> > >
> > > > Any good compiler will convert switch/case to a jump table for you.
VC
> > > does
> > > > this; I'm quite sure that GCC does too. Unless you're writing in
> > assembly,
> > > > don't try to pad things out. The compiler should do this for you. I
> > don't
> > > > know when specifically VC/GCC do and when they don't.
> > >
> > > I think I have a different interpretation of Mike's question than
> > > you: my take on it is that he has code where all branches are just
> > > little snippets no longer than some nice easy number of bytes to
> > > multiply by, e.g. if all his branches were less than 32 bytes
> > > long and he had 16 possible cases numbered 0:15 he could, assuming
> > > the branch number were in EAX do something like:
> > >
> > > lea EAX, [4*EAX]
> > > jmp branch_implementation_start[8*EAX]
> > >
> > > In this way he could avoid waiting for the address of a previous
> > > write to memory to be computed before prefetching the code
> > > following the branch.  Maybe ALIGN directives or ORG
> > > directives could tell the assembler where to insert padding
> > > without having to maintain it by hand every time the code
> > > is modified?
> > >
> > >
> > >
> >
> >
> >
>
>
>
>