Mixed 68k / PPC code on VBCC.

[Cowcat]

Hello. This is my first post.

So far I've been testing some 68k C code that sends data (array with integers) to a PPC C code function and returns the result: Compiled each object file with their respective cpu directives and then linked both finally as 68k code. It works.

The problem I face is when my data is based arround float point: The PPC code needs the link library -lm to work correctly, but at linking time I can only use 68k link libraries.

Being a half begginer coder I don't know if it is possible such a thing, or if I need some kind of stub (fd,etc) with mixed math code from both cpu's.

Any idea?
"2013 and playing with my old Commodore A 1200-40/PPC "

[Bamiga2002]

I don't know the answers to your coding questions, but I'm always after members to join our Blizzard PPC club

Welcome to EAB

[Cowcat]

Thanks!

[phx]

Originally Posted by Cowcat

The problem I face is when my data is based arround float point: The PPC code needs the link library -lm to work correctly, but at linking time I can only use 68k link libraries.

Are you trying to build a WarpOS executable?

You have to resolve all PPC symbol references, before doing the final linking step with the 68k startup code, objects and libraries. So you may try the following with all the PPC objects (from memory, assuming WarpOS EHF object file format):

vlink -bamigaehf -r -o ppc.o ppcobj1.o ppcobj2.o ... -Lvlibwos: -lm

-r means that you don't want to create an executable, but an object file, suitable for another linker pass. This should resolve all PPC references to floating point functions from m.lib.

In the final linking step you just link with ppc.o.

[Cowcat]

Thanks a lot phx. It's a warpos object. I tried something like that but the result it's a warpos exception. I'll post code later.

[Cowcat]

code1.c

#include <stdio.h>
#ifdef __VBCC__
#pragma amiga-align
#endif
#include <proto/exec.h>
#include <powerpc/powerpc.h>
#include <proto/powerpc.h>
#ifdef __VBCC__
#pragma default-align
#endif
#include "st.h"
extern void PPC_add();
int i;
vertex v[] = {
{-0.5,-0.5, 1.0},
{-0.5, 0.5, 1.0},
{ 0.5,-0.5, 1.0},
{ 0.5, 0.5, 1.0}
};

int main()
{

struct PPCArgs pa;
pa.PP_Code =(ULONG *)&PPC_add;
pa.PP_Offset = 0;
pa.PP_Flags = PPF_LINEAR;
pa.PP_Stack = 0;
pa.PP_StackSize=0;
pa.PP_Regs[0]=(ULONG) &v;

printf("Going PPC\n");
RunPPC(&pa);
for(i=0;i<4;i++){
printf("Result X %f\n",v[i].x);
printf("Result Y %f\n",v[i].y);
printf("Result Z %f\n",v[i].z);
printf("\n");
}
return(0);
}

##############################################
code2.c

#include <math.h>
#include "st.h"

void PPC_add (__reg("r3") vertex *p) {
int i;
for(i=0;i<4;i++){
p[i].x = p[i].x +0.5;
p[i].y = p[i].y +0.5;
p[i].z = p[i].z +0.5;
}
}

##########################
st.h

typedef struct{
float x,y,z;
}vertex;

Then....
vc +aos68k -c99 -c code1.c
vc +warpos -c99 -c code2.c
vlink -bamigaehf -r -o ppc.o code2.o -Lvlibwos: -lm
vc +aos68k -o crap code1.o ppc.o -lamiga -lmieee -lauto

Something is wrong in my code (crap) : Results are weird (sometimes same original values, others unexpected float values)....... Same happens without linking ppc object with -lm.
No problem if I modify array and code to work only with integers.

[phx]

I guess your problem is that the TOC pointer in r2 is not initialized for the PPC code, so static data references will access random memory locations.

The TOC pointer in r2 is the same as the small data pointer in a4 for the M68k. You either have to pass it via PP_Regs (from LinkerDB symbol) or use the attribute __saveds for PPC_add(), so the function will restore the TOC register itself. I would recommend

__saveds void PPC_add (__reg("r3") vertex *p) {

for a first test.

When it still doesn't work then I need some time to reproduce it on a real WarpOS environment at home.

Originally Posted by Cowcat

#ifdef __VBCC__
#pragma amiga-align
#endif

BTW, those #ifdefs are not needed, because by definition a #pragma directive must be ignored by the compiler, when it is unknown.

Results are weird (sometimes same original values, others unexpected float values).......

That's because of the random TOC register. The 0.5 float constant has to be loaded from the data section, which fails. It works with integers, because integers can be loaded directly into a register. No data access needed.

Same happens without linking ppc object with -lm.

This was not even needed. I didn't see any reference to m.lib from your PPC code.

But now, when using __saveds, you have to link with the WarpOS -lvc here, to resolve a reference to the __setr2 function, which restores the TOC pointer.

[Cowcat]

Awesome. My initial thought about linking m.lib, was that compiling for only warpos a similiar code, and without linking that lib, the printf output is a "%f" string (??) More brain storming, thank you again.

[Cowcat]

Perfect. It works (__saveds & linking -lvc ) !!! ,)

[phx]

Originally Posted by Cowcat

My initial thought about linking m.lib, was that compiling for only warpos a similiar code, and without linking that lib, the printf output is a "%f" string (??)

True. When the PPC code would call printf() to output float, then you need the math-extended version of it, so you have to link with m.lib in this case. But in your PPC_add() was nothing like that.

Perfect. It works (__saveds & linking -lvc ) !!! ,)

Great!

[Cowcat]

I owe you two beers.