On 28 June 2018 at 15:12, Peter Maydell <[email protected]> wrote:
> On 28 June 2018 at 12:30, Alex Bennée <[email protected]> wrote:
>> loading test image
>> sve-all-short-v8.3+sve@vq3/insn_sdiv_z_p_zz___INC.risu.bin...
>> starting apprentice image at 0x4000801000
>> starting image
>> fish: “../qemu.git/aarch64-linux-user/…” terminated by signal SIGFPE
>> (Floating point exception)
>
> Do you have the insn that it's barfing on? In particular,
> I'm guessing from the test name that this is for something
> covered by one of the SDIV_zpzz lines in sve.decode, which
> is already in master rather than in this test series.
> If that's true, then it shouldn't block applying this set...
Further discussion on IRC suggests that this is failing on
MININT idiv -1, which is an annoying special case that x86
happens to generate SIGFPE for. Compare our HELPER(sdiv) code:
int32_t HELPER(sdiv)(int32_t num, int32_t den)
{
if (den == 0)
return 0;
if (num == INT_MIN && den == -1)
return INT_MIN;
return num / den;
}
with what we do for SVE:
#define DO_DIV(N, M) (M ? N / M : 0)
This is OK for unsigned division, but signed division needs to
special case INT_MIN / -1.
In any case, this is in an existing insn, so I'm going to apply
this series to target-arm.next (fixing up the patch 5 comment
typo).
thanks
-- PMM
