Re: PPC compiler splits up ints & shorts?

[Dirk Zimoch <dirk.zimoch@psi.ch>]

Hi Michael,

have vou tried this:

static void copy4 (int *p, volatile int *q, int n)

I always declare any pointer to hardware volatile to prevent the 
compliler from trying funny things.

BTW: I can't reproduce you assember code with neither gcc-2.8.1 on 
Tornado 1.0.1 nor with gcc-2.96 on Tornado 2.2. I always get word 
instructions, not byte instructions.

Dirk


Laznovsky, Michael wrote:
> Hi- does anyone know why our power-pc compiler likes to turn short & int pointer
> dereferences into byte moves, and how to get it to not to?  Some hardware may not
> like this, as suggested by bus analyzer.
>
> For instance, this code:
>
> 	static void copy4 (int *p, int *q, int n) {
> 	  while (n-- > 0) *q++ = *p++;
> 	}
>
> generates this:
>
>         .align 2
>         .type    copy4,@function
> copy4:
>         mr 0,5
>         cmpwi 1,0,0
>         mr 7,3
>         mr 6,4
>         addi 5,5,-1
>         bclr 4,5
>         addi 4,4,3
>         addi 3,3,3
> .L16:
>         lbz 9,-2(3)	<---
>         lbz 11,-1(3)	<---
>         lbz 10,0(3)	<---
>         lbz 0,0(7)	<---
>         addi 3,3,4
>         addi 7,7,4
>         mr 8,5
>         addi 5,5,-1
>         cmpwi 1,8,0
>         slwi 0,0,24
>         slwi 9,9,16
>         or 9,9,0
>         slwi 11,11,8
>         or 11,11,9
>         or 10,10,11
>         srwi 0,10,24
>         stb 0,0(6)	<---
>         srwi 0,10,16
>         stb 0,-2(4)	<---
>         srwi 0,10,8
>         stb 0,-1(4)	<---
>         stb 10,0(4)	<---
>         addi 4,4,4
>         addi 6,6,4
>         bc 12,5,.L16
>         blr
> .Lfe4:
>         .size    copy4,.Lfe4-copy4
>
> Switches which may be relevant:
>     .../vxworks/devel/tor-2.0.2/cd/host/sun4-solaris2/bin/ccppc
>         -B.../vxworks/devel/tor-2.0.2/cd/host/sun4-solaris2/lib/gcc-lib/
>         -nostdinc
>         -ansi
>         -pedantic
>         -DCPU=PPC604
>         -D_GNU_TOOL
>         -DTRUE=1
>         -DNIVXI
>         -O2
>         -fstrength
>         -reduce
>         -Wall
>         -DVXI
>         -mcpu=604
>         --no-builtin
>         -mstrict-align
>         -mlongcall
>         -D_REENTRANT
>         -DvxWorks
>         -DV5_vxWorks
>         -fno-builtin
>
> "ccppc --version" yields:
> 	cygnus-2.7.2-960126 egcs-971225 tornado 2.0
>
> I thought maybe #pragma pack() in the source had something to do with it, but that
> made no difference.
>
> I took a look in the tech-talk archives, but couldn't find anything related.  Could
> be just poor search skills....
>
> Thanks for any help!
>
> Mike

--
Dr. Dirk Zimoch
Swiss Light Source
Paul Scherrer Institut
Computing and Controls
phone +41 56 310 5182
fax   +41 56 310 4413