[PATCH 3/4] Use libc_fe* macros in ldbl-128/e_expl.c.
Stefan Liebler
stli@linux.ibm.com
Tue Mar 31 07:39:32 GMT 2020
On 3/30/20 8:12 PM, Adhemerval Zanella via Libc-alpha wrote:
>
>
> On 27/03/2020 11:23, Stefan Liebler via Libc-alpha wrote:
>> On 3/26/20 3:53 PM, Adhemerval Zanella via Libc-alpha wrote:
>>>
>>>
>>> On 26/03/2020 06:08, Stefan Liebler via Libc-alpha wrote:
>>>> On 3/25/20 4:42 PM, Adhemerval Zanella via Libc-alpha wrote:
>>>>>
>>>>>
>>>>> On 25/03/2020 12:07, Adhemerval Zanella wrote:
>>>>>>
>>>>>>
>>>>>> On 25/03/2020 12:00, Adhemerval Zanella wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 25/03/2020 07:13, Stefan Liebler via Libc-alpha wrote:
>>>>>>>> Unfortunately, this patch is responsible for testfails on x86_64:
>>>>>>>>
>>>>>>>> math/test-float128-exp.out:
>>>>>>>> Failure: exp (-0x1p-10000): Exception "Underflow" set
>>>>>>>> Failure: exp (-0x2p-16384): Exception "Underflow" set
>>>>>>>> ...
>>>>>>>>
>>>>>>>> math/test-float128-cexp.out:
>>>>>>>> Failure: Real part of: cexp (0x2p-16384 - 0x4p-1076 i): Exception "Underflow" set
>>>>>>>> Failure: Real part of: cexp (0x2p-16384 - 0x8p-152 i): Exception "Underflow" set
>>>>>>>
>>>>>>> The sysdeps/x86/fpu/fenv_private.h states:
>>>>>>>
>>>>>>> 296 #ifdef __x86_64__
>>>>>>> 297 /* The SSE rounding mode is used by soft-fp (libgcc and glibc) on
>>>>>>> 298 x86_64, so that must be set for float128 computations. */
>>>>>>> 299 # define SET_RESTORE_ROUNDF128(RM) \
>>>>>>> 300 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_sse, libc_feresetround_sse)
>>>>>>>
>>>>>>> So
>>>>>>>
>>>>>>>>> diff --git a/sysdeps/ieee754/ldbl-128/e_expl.c b/sysdeps/ieee754/ldbl-128/e_expl.c
>>>>>>>>> index 37c1538c08..104ace1690 100644
>>>>>>>>> --- a/sysdeps/ieee754/ldbl-128/e_expl.c
>>>>>>>>> +++ b/sysdeps/ieee754/ldbl-128/e_expl.c
>>>>>>>>> @@ -66,6 +66,7 @@
>>>>>>>>> #include <inttypes.h>
>>>>>>>>> #include <math-barriers.h>
>>>>>>>>> #include <math_private.h>
>>>>>>>>> +#include <fenv_private.h>
>>>>>>>>> #include <math-underflow.h>
>>>>>>>>> #include <stdlib.h>
>>>>>>>>> #include "t_expl.h"
>>>>>>>>> @@ -146,9 +147,10 @@ __ieee754_expl (_Float128 x)
>>>>>>>>> union ieee854_long_double ex2_u, scale_u;
>>>>>>>>> fenv_t oldenv;
>>>>>>>>> - feholdexcept (&oldenv);
>>>>>>>>> #ifdef FE_TONEAREST
>>>>>>>>> - fesetround (FE_TONEAREST);
>>>>>>>>> + libc_feholdexcept_setroundl (&oldenv, FE_TONEAREST);
>>>>>>>
>>>>>>> Should be libc_feholdexcept_setroundf128.
>>>>>>
>>>>>> But it does not see to help here, so I don't know what is failing as well.
>>>>>
>>>>> Ok, so what is happening __sfp_handle_exceptions always use 387 exception
>>>>> mode for FP_EX_OVERFLOW and FP_EX_UNDERFLOW:
>>>>>
>>>>> config/i386/sfp-exceptions.c
>>>>>
>>>>> 79 if (_fex & FP_EX_OVERFLOW)
>>>>> 80 {
>>>>> 81 struct fenv temp;
>>>>> 82 asm volatile ("fnstenv\t%0" : "=m" (temp));
>>>>> 83 temp.__status_word |= FP_EX_OVERFLOW;
>>>>> 84 asm volatile ("fldenv\t%0" : : "m" (temp));
>>>>> 85 asm volatile ("fwait");
>>>>> 86 }
>>>>> 87 if (_fex & FP_EX_UNDERFLOW)
>>>>> 88 {
>>>>> 89 struct fenv temp;
>>>>> 90 asm volatile ("fnstenv\t%0" : "=m" (temp));
>>>>> 91 temp.__status_word |= FP_EX_UNDERFLOW;
>>>>> 92 asm volatile ("fldenv\t%0" : : "m" (temp));
>>>>> 93 asm volatile ("fwait");
>>>>> 94 }
>>>>> Yes this looks like the mentioned disassembly.
>>>>> Different that FP_EX_INEXACT, for instance, where __SSE_MATH__ sets
>>>>> whether SSE is used or not.
>>>>>
>>>>> So I think it is not safe to use the SSE variants for libc_*_testf128,
>>>>> as for i387 we should use the default_* instead.
>>>>>
>>>> I've just switched to default_* in sysdeps/x86/fpu/fenv_private.h:
>>>> -#ifdef __x86_64__
>>>> +#if 0
>>>> /* The SSE rounding mode is used by soft-fp (libgcc and glibc) on
>>>> x86_64, so that must be set for float128 computations. */
>>>> # define SET_RESTORE_ROUNDF128(RM) \
>>>>
>>>> But now there are 7 testfails. For some of them, the max.ulp goes up (over 10), but there are even worse cases. Here are shortened excerpts of the out files:
>>>> - math/test-float128-clog.out:
>>>> Failure: Test: Real part of: clog_towardzero (0x2.82b795e420b281a934c6dd315cb2p-4 + 0xf.cd42a15bf9a361243a89663e81e8p-4 i)
>>>> ulp : 162259276829213363391578010288127.0000
>>>> max.ulp : 3.0000
>>>> Failure: Test: Real part of: clog_upward (0x2.82b795e420b281a934c6dd315cb2p-4 + 0xf.cd42a15bf9a361243a89663e81e8p-4 i)
>>>> ulp : 162259276829213363391578010288128.0000
>>>> max.ulp : 4.0000
>>>>
>>>> - math/test-float128-clog10.out:
>>>> Failure: Test: Real part of: clog10_downward (0x3.bea2bd62e35p-4 + 0xf.8e3d619a8d11bfd30b038eep-4 i)
>>>> ulp : 4.0000
>>>> max.ulp : 3.0000
>>>> Failure: Test: Real part of: clog10_towardzero (0x2.82b795e420b281a934c6dd315cb2p-4 + 0xf.cd42a15bf9a361243a89663e81e8p-4 i)
>>>> ulp : 140936617129079063283494433422698.0000
>>>> max.ulp : 4.0000
>>>> Failure: Test: Real part of: clog10_upward (0x2.82b795e420b281a934c6dd315cb2p-4 + 0xf.cd42a15bf9a361243a89663e81e8p-4 i)
>>>> ulp : 140936617129079063283494433422698.0000
>>>> max.ulp : 4.0000
>>>>
>>>> - math/test-float128-jn.out
>>>> - math/test-float128-lgamma.out
>>>> - math/test-float128-tgamma.out:
>>>> something like:
>>>> ulp : 12.0000
>>>> max.ulp : 4.0000
>>>>
>>>> Failure: tgamma_upward (-0x6.ec00000000000008p+8): errno set to 0, expected 34 (ERANGE)
>>>>
>>>> - math/test-float128-y1.out:
>>>> Failure: Test: y1_downward (0x2p+0)
>>>> ulp : 13.0000
>>>> max.ulp : 4.0000
>>>> Failure: Test: y1_towardzero (0x2p+0)
>>>> ulp : 6.0000
>>>> max.ulp : 2.0000
>>>> Failure: Test: y1_upward (0x2p+0)
>>>> ulp : 10.0000
>>>> max.ulp : 5.0000
>>>>
>>>> - math/test-float128-yn.out
>>>>
>>>
>>> So it seems the issue is the mix on how libm fenv function, the internal
>>> libc_fe*, and libgcc handles the exception register. The exported fenv
>>> operates on both i387 and SSE (since it should work on long double as well),
>>> and the internal libc_fe* will set either SSE for float, double, and float128
>>> and i387 for long double (as expected).
>>>
>>> The libgcc, however, will set either SEE or i387 depending of the exception.
>>> This broke the assumption of libc_fe* for float128 where either SSE or i387
>>> will be used.
>>>
>>> One option might be to force libgcc to not use its __sfp_handle_exceptions
>>> on x86_64 and provide one that uses only SEE operations since libgcc does
>>> not use 'long double' on float128 operations. The patch below does it
>>> and applied on top your patches shows no regressions.
>>
>> Great news. Thanks Adhemerval.
>> I've also successfully build and run the testsuite with your patch on top of mine and with only your patch without mine.
>>
>> As e.g. __multf3 or __addtf3 is used in various f128 functions, can you please first commit your patch? Then I will add a reference to this commit id in the commit-message.
>
> I will send a RFC for this patch, we need to check with x86 maintainers
> if this the desirable direction and if I got everything right.
>
>>
>> One other question: Why are the soft-fp functions (for add / multiply) called at all. Are the corresponding hardware instructions not available on all x86_64 machines? Or do we miss a compiler flag?
>
> The float128 on gcc/x86_64 is implemented by soft-fp library in libgcc [1]
> and its ABI passes arguments through SSE register [2].
>
> [1] https://stackoverflow.com/questions/26639477/what-exactly-is-a-float128-if-im-using-gcc-4-9-on-x86-64
> [2] https://github.com/hjl-tools/x86-psABI/wiki/X86-psABI
>
Thanks for the info and for working on this.
More information about the Libc-alpha
mailing list