[PATCH 8/8] math: Use math_opt_barrier on underflow/overflow

[Adhemerval Zanella Netto]

On 18/12/25 20:42, H.J. Lu wrote:
> On Thu, Dec 18, 2025 at 8:46 PM Adhemerval Zanella Netto
> <adhemerval.zanella@linaro.org> wrote:
>>
>>
>>
>> On 17/12/25 21:05, H.J. Lu wrote:
>>> On Thu, Dec 18, 2025 at 1:56 AM Adhemerval Zanella
>>> <adhemerval.zanella@linaro.org> wrote:
>>>>
>>>> This avoids compiler to move the operation before the 'iy' compare.
>>>>
>>>> It fixes math/test-float128-pow regreesions when building with clang:
>>>>
>>>> Failure: pow (-0x1.000002p+0, 0xf.ffffffffffff8p+1020): Exception "Underflow" set
>>>> Failure: pow (-0x1.000002p+0, 0xf.ffffffffffffbffffffffffffcp+1020): Exception "Underflow" set
>>>> Failure: pow (-0x1.000002p+0, 0xf.fffffffffffffffffffffffffff8p+16380): Exception "Underflow" set
>>>> Failure: pow (-0x1.000002p+0, 0xf.fffffffffffffffp+16380): Exception "Underflow" set
>>>> Failure: pow (-0x1.000002p+0, 0xf.fffffp+124): Exception "Underflow" set
>>>> Failure: pow (-0x1.00000ep+0, 0xf.ffffffffffff8p+1020): Exception "Underflow" set
>>>> Failure: pow (-0x1.00000ep+0, 0xf.ffffffffffffbffffffffffffcp+1020): Exception "Underflow" set
>>>> Failure: pow (-0x1.00000ep+0, 0xf.fffffffffffffffffffffffffff8p+16380): Exception "Underflow" set
>>>> Failure: pow (-0x1.00000ep+0, 0xf.fffffffffffffffp+16380): Exception "Underflow" set
>>>> Failure: pow (-0x2p+0, -0xf.ffffffffffff8p+1020): Exception "Overflow" set
>>>> Failure: pow (-0x2p+0, -0xf.ffffffffffffbffffffffffffcp+1020): Exception "Overflow" set
>>>> Failure: pow (-0x2p+0, -0xf.fffffffffffffffffffffffffff8p+16380): Exception "Overflow" set
>>>> Failure: pow (-0x2p+0, -0xf.fffffffffffffffp+16380): Exception "Overflow" set
>>>> Failure: pow (-0x2p+0, -0xf.fffffp+124): Exception "Overflow" set
>>>> Failure: pow (-0x2p+0, 0xf.ffffffffffff8p+1020): Exception "Underflow" set
>>>> Failure: pow (-0x2p+0, 0xf.ffffffffffffbffffffffffffcp+1020): Exception "Underflow" set
>>>> Failure: pow (-0x2p+0, 0xf.fffffffffffffffffffffffffff8p+16380): Exception "Underflow" set
>>>> Failure: pow (-0x2p+0, 0xf.fffffffffffffffp+16380): Exception "Underflow" set
>>>> Failure: pow (-0x2p+0, 0xf.fffffp+124): Exception "Underflow" set
>>>> [...]
>>>>
>>>> Checked on x86_64-linux-gnu and aarch64-linux-gnu with gcc-15 and
>>>> clang-18.
>>>> ---
>>>>  sysdeps/ieee754/ldbl-128/e_powl.c | 6 ++++--
>>>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/sysdeps/ieee754/ldbl-128/e_powl.c b/sysdeps/ieee754/ldbl-128/e_powl.c
>>>> index 4e20616705..9ff5a43573 100644
>>>> --- a/sysdeps/ieee754/ldbl-128/e_powl.c
>>>> +++ b/sysdeps/ieee754/ldbl-128/e_powl.c
>>>> @@ -279,9 +279,11 @@ __ieee754_powl (_Float128 x, _Float128 y)
>>>>        if (iy > 0x407d654b)
>>>>         {
>>>>           if (ix <= 0x3ffeffff)
>>>> -           return (hy < 0) ? huge * huge : tiny * tiny;
>>>> +           return (hy < 0) ? math_opt_barrier (huge * huge)
>>>> +             : math_opt_barrier (tiny * tiny);
>>>>           if (ix >= 0x3fff0000)
>>>> -           return (hy > 0) ? huge * huge : tiny * tiny;
>>>> +           return (hy > 0) ? math_opt_barrier (huge * huge)
>>>> +             : math_opt_barrier (tiny * tiny);
>>>>         }
>>>
>>> Does it generate less optimized codes for GCC?
>>
>> I do not think this really matter here, these are for corner cases where
>> the function underflow/overflow.  And I think this are really a issue
>> with clang with -frounding-math, since it can not assume that the
>> multiplication will not generate underflow/overflow exceptions (so it can
>> move out the comparison).
> 
> Is this a clang bug? I think this info should be mentioned in the commit log.

I think so, although I don't think I will have time to pinpoint exactly the
problem or create a contrive example.