Re: Steps toward aarch64 Cog

Made the mistake of running the FFI unit tests in Cuis, which directly led me down the rabbit hole.

Think I have made some progress in understanding, but don't yet see how to teach VMMaker to get code from register.

Here is crux of one of my investigations:

==============================

===== C source ======sqFFITestFuncs.c

/* test passing and returning doubles */
EXPORT(double) ffiTestDoubles(double d1, double d2) {
printf("The two floats are %f and %f\n", (float)d1, (float)d2);
return d1+d2;
}

===== ASM === from objectdump -d
0000000000000ac0 <ffiTestDoubles>:
ac0: a9be7bfd stp x29, x30, [sp, #-32]!
ac4: 90000000 adrp x0, 0 <ffiTestChars>
ac8: 91000000 add x0, x0, #0x0
acc: 910003fd mov x29, sp
ad0: 6d0127e8 stp d8, d9, [sp, #16]
ad4: 1e604009 fmov d9, d0
ad8: 1e604028 fmov d8, d1
adc: 1e624000 fcvt s0, d0
ae0: 1e624021 fcvt s1, d1
ae4: 1e22c000 fcvt d0, s0
ae8: 1e22c021 fcvt d1, s1
aec: 94000000 bl 0 <printf>
af0: 1e682920 fadd d0, d9, d8
af4: 6d4127e8 ldp d8, d9, [sp, #16]
af8: a8c27bfd ldp x29, x30, [sp], #32
afc: d65f03c0 ret
// result returned in d0 (NB: _not_ x0)
=================Cuis Workspace (after FFI unit test)===========
FFITestLibrary ffiTestDoubles: 30.0 with: 65.0.
"--> 00.00558872584952 -- should be 95.0"
=================Terminal Output===================================
The two floats are 30.000000 and 65.000000
===================================================================
FFIPlugin>>ffiCreateReturnOop: retVal
...

^interpreterProxy floatObjectOf: (self ffiReturnFloatValue).
====================================================================
My theory: return value taken from x0, not d0.==============================================Is this a sensible interpretation?Any hints from Heloise?
Thanks much,-KenDPS: Merry Christmas! (I would use Xmas, but who knows what is lurking under the X? 8^)

From the rabbit hole down into the mine shaft.

Gee, sure is dark down here.  Hope it's shallow!  ;^)

-KenD

See, if you don't give me VMMaker write access, you won't get a certain class of dumb mistakes!

Thanks,

-KenD

----- Original Message -----

From:

"Eliot Miranda" <[hidden email]>

To:

"Ken.Dickey" <[hidden email]>

Cc:

<[hidden email]>

Sent:

Sun, 23 Dec 2018 11:20:16 -0800

Subject:

Re: Steps toward aarch64 Cog [FFI]

Hi Ken,

    I have a suggestion.  Instead of initializing WordSize from Smalltalk wordSize I would create two subclasses with a wordSize method, hardcoded as ^4 and ^8 respectively.  We want to be able to simulate any of the processors on any platform, hence we can simulate x64 on the 32-bit version and x86 on the 64-bit version as well as 32-bit processors on 32-bit versions and 64-bit processors on 64-bit versions.  So structuring the ARM FFI as an abstract subclass with a subclass responsibility for wordSize, along with two subclasses for 64 & 32 bits will give us that flexibility.

Where a wordSize define would make sense would be in platforms/Cross/plugins/IA32ABI/*ARM*, although you might want to review the code first and see if #ifdef or splitting the files into 32 and 64-bit versions is cleaner.

On Sun, Dec 23, 2018 at 8:06 AM <[hidden email]> wrote:

Rats!  

Can't even use this webmail thingie...

[the _other_ changes file]

Cheers,

-KenD

--

_,,,^..^,,,_

best, Eliot