Re: serious bugs in Altivec implementation of gcc

[Eyal Bagon <eyal at mobileye dot com>]

hello
The max patch is working - thanks.
The problem with vec_msum still exist, the same problem occurs when I 
use vec_vmsummbm instead of vec_msum - could the problem be somewhere in 
the implementation of __builtin_altivec_vmsummbm ? ( by the way the 
altivec.h file you send me is the same as I had )

another thing that I noticed when comparing assembler code of 2 
corresponding files it looks like the new gcc code is less optimised
( I attached the testMax_2.9.2.cpp and testMax_3.2.cpp and their 
corresponding assembler files.

I used : g++ -S -Wall -pedantic -ansi -Wno-long-long -DNDEBUG -O6 
-DALTIVEC -fvec testMax_2.9.2.cpp

and : g++ -S -Wall -pedantic -ansi -Wno-long-long -DNDEBUG -O6 -DALTIVEC 
-maltivec testMax_3.2.cpp

Thanks
Eyal

Aldy Hernandez wrote:
> On Mon, Aug 12, 2002 at 02:53:42PM +0300, Eyal Bagon wrote:
>
> > serious bugs in Altivec implementation of gcc
> >
> > hello
> >
> > I'm trying to use gcc 3.2 to compile C code for Altivec and I found some 
> > serious problems in the compiler:
> > 1) vec_max is converted to the assembler function of MIN ( not MAX )
> > 	you can compile the attached file testMax.cpp
> > 	g++ -S -Wall -pedantic -ansi -Wno-long-long -DNDEBUG -O6 	
> > 	-DALTIVEC -maltivec testMax.cpp
> >
> > 	you can see in the assembler file that you get vminuw instead of
> > 	vmax...
>
> See:
>
> http://gcc.gnu.org/ml/gcc-patches/2002-05/msg00630.html
>
>
> > 2) vec_msum : when vec_msum is used with signed char as the first 		
> > 	parameter the assembler code will be vmsumubm -> as for 	
> > 	UNsigned.
> >
> > 		you can enjoy this yourself - just compile testMsum.cpp
> >
> > Is there a patch that sole these problems ?
> > Is there a reliable version of gcc for Altivec ?
>
> I can't recall, but this may be fixed on current sources.  Use the
> altivec.h from CVS.  I'm attaching it just in case.  If it doesn't fix
> it, ping me again and CC the gcc@gcc.gnu.org.
>
> A few buglets were fixed from 3.2 to the CVS sources, but most of
> bugfixes were in <altivec.h> so you may get away with using this new
> file I'm attaching.  (except in the first bug you mention.  That
> requires a patch to the sources).
>
> Aldy
//#include <altivec.h>

__vector unsigned int testMax()
{
  __vector unsigned int vss1 = (__vector unsigned int)(1,2,3,4);
  __vector unsigned int vss2 = (__vector unsigned int)(8,7,6,5);
    
  __vector unsigned int vss3;
  
  vss3 = vec_max(vss1,vss2);

  return vss3;
    
}
  
#include <altivec.h>

__vector unsigned int testMax()
{
  __vector unsigned int vss1 = {1,2,3,4};
  __vector unsigned int vss2 = {8,7,6,5};
    
  __vector unsigned int vss3;
  
  vss3 = vec_max(vss1,vss2);

  return vss3;
    
}
	.file	"testMax_2.9.2.cpp"
gcc2_compiled.:
	.section	".rodata"
	.align 4
.LC2:
	.long 1
	.long 2
	.long 3
	.long 4
	.align 4
.LC3:
	.long 8
	.long 7
	.long 6
	.long 5
	.section	".text"
	.align 2
	.globl testMax__Fv
	.type	 testMax__Fv,@function
testMax__Fv:
.LFB1:
	lis 9,.LC2@ha
	lis 11,.LC3@ha
	la 9,.LC2@l(9)
	la 11,.LC3@l(11)
	lvx 0,0,9
	lvx 2,0,11
	vmaxuw 2,0,2
	blr
.LFE1:
.Lfe1:
	.size	 testMax__Fv,.Lfe1-testMax__Fv

	.section	".eh_frame","aw"
__FRAME_BEGIN__:
	.4byte	.LLCIE1
.LSCIE1:
	.4byte	0x0
	.byte	0x1
	.byte	0x0
	.byte	0x1
	.byte	0x7c
	.byte	0x6e
	.byte	0xc
	.byte	0x1
	.byte	0x0
	.byte	0x9
	.byte	0x6e
	.byte	0x41
	.align 2
.LECIE1:
	.set	.LLCIE1,.LECIE1-.LSCIE1
	.4byte	.LLFDE1
.LSFDE1:
	.4byte	.LSFDE1-__FRAME_BEGIN__
	.4byte	.LFB1

	.4byte	.LFE1-.LFB1
	.align 2
.LEFDE1:
	.set	.LLFDE1,.LEFDE1-.LSFDE1
	.ident	"GCC: (GNU) 2.95.2 19991024 (moto-1.7 release)"
	.file	"testMax_3.2.cpp"
	.section	.rodata
	.align 4
.LC0:
	.long	1
	.long	2
	.long	3
	.long	4
	.align 4
.LC1:
	.long	8
	.long	7
	.long	6
	.long	5
	.section	".text"
	.align 2
	.globl _Z7testMaxv
	.type	_Z7testMaxv,@function
_Z7testMaxv:
.LFB1:
	mfvrsave 8
	stwu 1,-32(1)
.LCFI0:
	oris 7,8,0xdfff
	ori 0,7,61440
	stw 8,28(1)
	mtvrsave 0
	lis 4,.LC1@ha
	lis 6,.LC0@ha
	la 3,.LC1@l(4)
	la 5,.LC0@l(6)
	lvx 3,0,3
	lvx 1,0,5
	lwz 0,28(1)
	vmaxuw 2,1,3
	mtvrsave 0
	addi 1,1,32
	blr
.LFE1:
.Lfe1:
	.size	_Z7testMaxv,.Lfe1-_Z7testMaxv
	.ident	"GCC: (GNU) 3.2 20020811 (prerelease)"