Re: [Qemu-devel] [PATCH 1/2] linux-user: Define target alignment size

[Laurent Vivier]

Le dimanche 13 février 2011 à 10:24 +0200, Blue Swirl a écrit :
> On Sun, Feb 13, 2011 at 4:22 AM, Laurent Vivier <address@hidden> wrote:
> > Datatype alignment can be found using following application:
> >
> > int main(void)
> > {
> >        printf("alignof(short) %ld\n", __alignof__(short));
> >        printf("alignof(int) %ld\n", __alignof__(int));
> >        printf("alignof(long) %ld\n", __alignof__(long));
> >        printf("alignof(long long) %ld\n", __alignof__(long long));
> > }
> >
> > This patch includes following alignments:
> >
> > i386
> >
> >   alignof(short) 2
> >   alignof(int) 4
> >   alignof(long) 4
> >   alignof(long long) 8
> >
> >  x86_64
> >
> >   alignof(short) 2
> >   alignof(int) 4
> >   alignof(long) 8
> >   alignof(long long) 8
> >
> >  arm
> >
> >   alignof(short) 2
> >   alignof(int) 4
> >   alignof(long) 4
> >   alignof(long long) 4
> >
> >  m68k (680x0)
> >
> >   alignof(short) 2
> >   alignof(int) 2
> >   alignof(long) 2
> >   alignof(long long) 2
> >
> >  mips
> >
> >   alignof(short) 2
> >   alignof(int) 4
> >   alignof(long) 4
> >   alignof(long long) 8
> >
> >  ppc
> >
> >   alignof(short) 2
> >   alignof(int) 4
> >   alignof(long) 4
> >   alignof(long long) 8
> >
> > for other targets, use by default (2,4,4,8).
> >
> > Please, update for your favorite target...
> 
> For Sparc32 (I think also sparc32plus), the default is OK.
> 
> For Sparc64, please use 2, 4, 8, 8. I'd guess other 64 bit platforms
> (Alpha, MIPS64, PPC64 etc) should use the same.

OK, I update my patch.

> Does GCC produce correct code using the attributes on strictly aligned
> host, when the target is less strictly aligned?

It seems it is OK. I did some tests into a mips-linux-user chroot (sparc
one is broken ;-) ) :

mips is a strictly aligned host, from gcc/config/mips/mips.h:

#define STRICT_ALIGNMENT 1

aligments are:

   alignof(short) 2
   alignof(int) 4
   alignof(long) 4
   alignof(long long) 8

We try to align int on 2.

#include <stdio.h>

typedef int target_int __attribute__((aligned(2)));

struct {
        char Z;
        target_int A;
} B;

int main(void)
{
        B.A = 0xdeadbeaf;
        printf("%d: %x\n", __alignof__(B.A), B.A);
}

./test_align
2: deadbeaf

disass:

        lw      $2,%got(B)($28)
        li      $3,-559087616                   # 0xffffffffdead0000
        ori     $3,$3,0xbeaf
        swl     $3,2($2)
        swr     $3,5($2)

normal case:

        lw      $3,%got(B)($28)
        li      $2,-559087616                   # 0xffffffffdead0000
        ori     $2,$2,0xbeaf
        sw      $2,4($3)

So, gcc seems smart enough to split the memory access in several ones
compatible with the strict alignment rules.

> Should the alignment of floating point variables be specified as well?

At the moment it seems useless.

> The strict alignment required for doubles is 4, but recommended
> alignment is 8, I'm not sure which one is used for structures
> containing doubles.

if necessary, some tests will be helpfull.

Thank you for your comments.

Regards,
Laurent 

-- 
--------------------- address@hidden ----------------------
"Tout ce qui est impossible reste à accomplir"    Jules Verne
"Things are only impossible until they're not" Jean-Luc Picard