[Vm-dev] VM Maker: VMMaker.oscog-eem.2602.mcz

[Eliot Miranda]

On Fri, Jan 10, 2020 at 2:10 PM Henrik Sperre Johansen <
henrik.s.johansen at veloxit.no> wrote:

>
> Henrik Sperre Johansen wrote
> >  I have a harder time seeing a use for the smaller element sizes, but I
> > guess I'm just narrow-minded :)
> >  (... or I'm missing something, again)
>
> Like, does ARM have packed arithmetic operations ala SSE?
> That would seem a natural fit, at least...
>

It does.  There's a 7,900 page manual.  Section 7 covers the SIMD
processing model.  It is a very powerful processor. 32 64 bit registers.
 32 128 bit FP registers, accessible (I *think*) as 32-bit FP, 4 32-bit FP,
64-bit FP or 2 64-bit FP, and a 16-bit FP format I haven't looked into
[blush].

https://developer.arm.com/docs/ddi0487/latest/arm-architecture-reference-manual-armv8-for-armv8-a-architecture-profile

I also have the PDF armv8_arch_ref_man_DDI0487E_a.pdf which takes over 10
seconds to load on my rather powerful last year MacBook Pro (gulp).

Instruction density is about 20% worse than x86-64 but it provides much
more power (three address instructions, twice as much register state, more
SIMD instructions, etc).  It doesn't;t have the raws compute power either,
being a low power design, but it is simpler and more regular than x86-64.
I admire it very much. Unlike ARM32 which, with its PC reg and naked delay
slots, never really appealed, =he ARM64 set seems carefully designed to
deliver excellent processing/power ratios.  Apart from the asymmetry in the
SP (a field of 31 means the SP reg in some instructions and the zero
register in others) and the trickyness of its immediate bit pattern
specification it is a pleasure to compile to.

Cheers,
> Henry
>

_,,,^..^,,,_
best, Eliot
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20200110/c08aedd2/attachment.html>