Re: A bug in IA32 assembler

["H . J . Lu" <hjl at lucon dot org>]

On Fri, Jun 29, 2001 at 10:21:20AM +0930, Alan Modra wrote:
> > > > 
> > > > Which one is valid,
> > > > 
> > > > 	movaps (A.3),%xmm0
> > > > 
> > > > or
> > > > 	movaps ($A.3),%xmm0
> > > 
> > > Both, I suppose, although the second one is trickery to get a
> > 
> > Are they the same? FYI, gcc seems to generate both from asm statements.
> 
> No.  They reference different symbols.

Compile the code with gcc 2.96 from RedHat 7.1,

        asm("movaps (%0),%%xmm0"                                //SSE
        :
        :"g"(A));

generates

	movaps ($A.0),%xmm0

But
        asm("movaps %%xmm0,(%0)":"=g"(C));                        //SSE

gets

	movaps %xmm0,(C.2)

> 
> > > '$' into a label.
> > > 
> > > 	movaps A.3,%xmm0
> > > 
> > > is really what you want;  Your added parentheses just act as they
> > > normally do in arithmetic expressions.
> > 
> > Right now, gas treats $A.3 as a label in
> > 
> > 	movaps ($A.3),%xmm0
> > 
> > That is the relocation is against `$A.3'. Also it doesn't allow
> > 
> > 	movaps $A.3,%xmm0
> > 
> > Do you have a fix for it?
> 
> I don't understand what you want a "fix" for.  "$A.3" in your last
> example is an immediate with value of the address of "A.3".  movaps
> doesn't accept immediates, only xmm registers or memory.
> 
> It's not at all normal to have `$' in labels on x86, as `$' is the
> immediate prefix.  Do you think your trickery with `($A.3)' should
> not be allowed?

Since

	movaps (A),%xmm0

generates the same binary as


	movaps A,%xmm0

shouldn't we treat

	movaps ($A.3),%xmm0
	
the same as

	movaps $A.3,%xmm0


H.J.
------
/* test program for SSE by kaz @ kobe1995.net
        % cc -g bug.c && ./a.out
  But it fails. So,
        % cc -g -S bug.c
  And change bug.s to avoid bug of gcc (may be):
         #APP
        -       movaps ($A.15),%xmm0
        -       movaps ($B.16),%xmm1
        +       movaps A.15,%xmm0
        +       movaps B.16,%xmm1
         #NO_APP
        % c++ loop.s && ./a.out
*/

#include <stdio.h>
#include <sys/time.h>
#include <unistd.h>
#include <stdlib.h>
#include <limits.h>

main(){
static float __attribute__((aligned(16)))
        A[4]={1.,1.,1.,1.},
        B[4]={0.,1.,2.,3.},
        C[4];

        int i;

        printf("float A=(");
        for(i=0;i<4;i++)        printf("%f,",A[i]);
        printf("\b)\n");

        printf("float B=(");
        for(i=0;i<4;i++)        printf("%f,",B[i]);
        printf("\b)\n");

        asm("movaps (%0),%%xmm0"                                //SSE
        :
        :"g"(A));
        asm("movaps (%0),%%xmm1"                                //SSE
        :
        :"g"(B));

        asm("addps %xmm1,%xmm0");                               //SSE
//      asm("mulps %xmm1,%xmm0");                               //SSE
        asm("movaps %%xmm0,(%0)":"=g"(C));                        //SSE
        printf("float C=(");
        for(i=0;i<4;i++)        printf("%f,",C[i]);
        printf("\b)\n");
}