sparc32 : Signed integer division overflow

I wanted to test an integer divider for SPARC32, and tried the stress 
test program on QEMU, which choked on the division
-0x8000_0000_0000_0000 / -1 (QEMU compiled on x86_64).

Excerpt from the test program :
-------------------------------------------------------------------
#include <stdio.h>

typedef unsigned uint32;
typedef unsigned long long uint64;
typedef signed int32;
typedef signed long long int64;

void sdiv(int64 x, int32 y, int32 *q, uint32 *ov)
{
   uint32 a,b,c,d,e;
   a=x>>32;
   b=x;
   c=y;

   __asm__ __volatile__("wr %2,%%y\n\t"
                        "nop\n\t"
                        "nop\n\t"
                        "nop\n\t"
                        "sdivcc %3,%4,%0\n\t"
                        "or %%r0,0,%1\n\t"
                        "bvc xxa\n\t"
                        "nop\n\t"
                        "add %1,1,%1\n\t"
                        "xxa:bpos xxb\n\t"
                        "nop\n\t"
                        "add %1,2,%1\n\t"
                        "xxb:bne xxc\n\t"
                        "nop\n\t"
                        "add %1,4,%1\n\t"
                "xxc:nop\n\t"
                : "=r"(d),"=r"(e):"r"(a),"r"(b),"r"(c)
                );
   *q=d;
   *ov=e & 1;
}

void main(void)
{
   uint64 x;
   uint32 y;
   int32 q;
   int32 ov;
   x = 1LL << 63;
   y = -1;
   sdiv(x, y, &q, &ov);
}

Here is a patch for handling this corner case on SPARC32.
SPARC64 division already checks this in helper_sdivx(), some other 
architectures
seem to do the same (for example, target-arm/helper.c: HELPER(sdiv))

          env->cc_src2 = overflow;
-------------------------------------------------------------------

On 12/03/14 21:26, Olivier Danet wrote:

Hi Olivier,

> Here is a patch for handling this corner case on SPARC32.
> SPARC64 division already checks this in helper_sdivx(), some other
> architectures
> seem to do the same (for example, target-arm/helper.c: HELPER(sdiv))
>
> ===================================================================
> The integer division 0x8000_0000_0000_0000 / -1 must be handled separately
> to avoid overflows on the QEMU host.
>
> Signed-off-by: Olivier Danet <odanet@caramail.com>
>
> -------------------------------------------------------------------
> diff --git a/target-sparc/helper.c b/target-sparc/helper.c
> index 57c20af..b6b5937 100644
> --- a/target-sparc/helper.c
> +++ b/target-sparc/helper.c
> @@ -116,14 +116,16 @@ static target_ulong
> helper_sdiv_common(CPUSPARCState *env, target_ulong a,
> if (x1 == 0) {
> cpu_restore_state(env, GETPC());
> helper_raise_exception(env, TT_DIV_ZERO);
> - }
> -
> - x0 = x0 / x1;
> - if ((int32_t) x0 != x0) {
> - x0 = x0 < 0 ? 0x80000000 : 0x7fffffff;
> + } else if (x1 == -1 && x0 == 0x8000000000000000) {
> + x0 = 0x7fffffff;
> overflow = 1;

Thanks for the patch! I think based upon Peter's recent series that the 
sign constant would need a ULL suffix in order to function correctly on 
32-bit platforms.

My personal preference would be for (1ULL << 63) unless Peter (CC added) 
can think of a reason to leave the hex constant in its current form?

That said, I've tested the patch on a Debian etch Linux image and it 
works for me.

> + } else {
> + x0 = x0 / x1;
> + if ((int32_t) x0 != x0) {
> + x0 = x0 < 0 ? 0x80000000 : 0x7fffffff;
> + overflow = 1;
> + }
> }
> -

Looks like a whitespace change accidentally made it into this patch too.

> if (cc) {
> env->cc_dst = x0;
> env->cc_src2 = overflow;
> -------------------------------------------------------------------


ATB,

Mark.