[PATCH v2] powerpc: Update ULPs and output for j0 with ibm128

[Matheus Castanho]

On 9/10/20 2:08 PM, Carlos O'Donell wrote:
> On 9/10/20 9:40 AM, Tulio Magno Quites Machado Filho wrote:
>> Joseph Myers <joseph@codesourcery.com> writes:
>>
>>> On Wed, 9 Sep 2020, Matheus Castanho wrote:
>>>
>>>> Instead of updating the ldouble ULPs for j0 with an upstream GCC,
>>>> constantly increasing their values, this patch regenerates them with a
>>>> GCC compiled with the patch mentioned above. This way we have the actual
>>>> precise ULPs listed in libm-test-ulps.  Of course, when compiling with
>>>> an upstream compiler some tests will fail as the calculated ULPs will be
>>>> higher than the expected ones.  For such tests, we mark the
>>>> corresponding entries in math/auto-libm-test-in with
>>>> xfail-rounding:ibm128-libgcc.
>>>
>>> xfail-rounding:ibm128-libgcc is intended for the sort of edge cases where 
>>> upstream GCC produces results that are wildly off.  It's not intended for 
>>> these sorts of cases where it gives slightly bigger ulps that are still 
>>> within the accepted bounds.
>>
>> Carlos, Joseph,
>>
>> I'm afraid that Matheus is either in a deadlock or we need a clearer
>> explanation of what is acceptable for ibm128.
>>
>> Notice that Matheus' first patch was rejected because results were greater
>> than 9.
> 
> It was later shown to me that >9 ULPs was acceptable for ibm128, my apologies
> for not being clearer that I was withdrawing my objection.
>  
>> With that said, would both of you accept the first version of this patch?
>> https://patchwork.sourceware.org/project/glibc/patch/20200820183700.115087-1-msc@linux.ibm.com/
> 
> Yes.
> 
> Patsy Griffin from my team also suggested this on September 2nd, she
> is seeing these failures in our own testing. It would be good to have
> them resolved.
> 

Thank you all for the feedback and clarification.

In this case, I have a slightly updated version of that initial patch,
which also updates the ULPs for float128 to fix other failing tests.

Just need someone to push it for me.

Thanks,
Matheus Castanho

--- 8< ---

>From c55d5ab5f386a85f04dcc3b9b71b92ac6523e321 Mon Sep 17 00:00:00 2001
From: Matheus Castanho <msc@linux.ibm.com>
Date: Thu, 20 Aug 2020 15:37:00 -0300
Subject: [PATCH] Update powerpc libm-test-ulps

Before this patch, the following tests were failing:

ppc and ppc64:
    FAIL: math/test-ldouble-j0

ppc64le:
    FAIL: math/test-float128-j0
    FAIL: math/test-float64x-j0
    FAIL: math/test-ibm128-j0
    FAIL: math/test-ldouble-j0
---
 sysdeps/powerpc/fpu/libm-test-ulps | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/sysdeps/powerpc/fpu/libm-test-ulps
b/sysdeps/powerpc/fpu/libm-test-ulps
index cd2a5fed45..fa2555f22b 100644
--- a/sysdeps/powerpc/fpu/libm-test-ulps
+++ b/sysdeps/powerpc/fpu/libm-test-ulps
@@ -1317,13 +1317,13 @@ Function: "j0_downward":
 double: 2
 float: 4
 float128: 4
-ldouble: 11
+ldouble: 12

 Function: "j0_towardzero":
 double: 5
 float: 6
-float128: 2
-ldouble: 8
+float128: 4
+ldouble: 16

 Function: "j0_upward":
 double: 4
-- 
2.26.2