libc-test on aarch64

libc-test on aarch64

[Sung-Taek Kim]

[-- Attachment #1: Type: text/plain, Size: 645 bytes --]

Hello, 



I'm a long time lurker and first time poster.



I was able to run Szabolcs' libc-test on aarch64 & x86_64, and aarch64 test produced noticeably more "bad fp exception".

Being aware of musl's experimental status on aarch64, I'd just like to make sure this is expected.



Here's test environment.





lib & testsuite version

- musl : 1.1.7

- libc-test : commit defcb8 (2017/06/22)



aarch64 (odroid c2)

- gcc : (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.5) 5.4.0 20160609

- kernel : Linux odroid 3.14.79-102



x86_64 

- gcc : (Ubuntu 5.4.0-6ubuntu1~16.04.5) 5.4.0 20160609

- kernel : Linux pc-core 4.4.0-59-generic









[-- Attachment #2: Type: text/html, Size: 1238 bytes --]

Re: libc-test on aarch64

[Szabolcs Nagy]

* Sung-Taek Kim <stkim1@pocketcluster.io> [2017-10-30 21:03:09 -0700]:
> I was able to run Szabolcs' libc-test on aarch64 &amp; x86_64, and aarch64 test produced noticeably more "bad fp exception".
> 
> Being aware of musl's experimental status on aarch64, I'd just like to make sure this is expected.
> 

those tests are wrong and i should fix them.

the current standard allows rounding functions to either
consistently raise inexact or not raise exceptions for
non-integer inputs.  the c code in musl was written so
it raises inexact (following fdlibm convention), and the
tests try to check for that.

but the other way is correct too and ieee754-2008 made
that a requirement and future c fp standard (ts 18661)
will require that too, so eventually both the c code in
musl and the test system should be fixed to check for that.

on aarch64 recently the rounding functions were changed
to use single instruction asm implementation and those
follow the ieee754-2008 requirement.

(fixing this in the c code can be expensive: either int
arithmetics has to be used or the inexact flag has to be
suppressed by modifying the fenv status, both of which
can be significantly slower than the current implementation
on some targets)

that said aarch64 has one outstanding issue: some long
double functions are broken for the 128bit float format.
other than this aarch64 port should work well.

> lib &amp; testsuite version
> 
> - musl : 1.1.7
> 

i assume this is 1.1.17

> - libc-test : commit defcb8 (2017/06/22)
> 
> 
> 
> aarch64 (odroid c2)
> 
> - gcc : (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.5) 5.4.0 20160609
> 
> - kernel : Linux odroid 3.14.79-102
> 

Re: libc-test on aarch64

[Sung-Taek Kim]

---- On Tue, 31 Oct 2017 04:50:32 -0700 Szabolcs Nagy <nsz@port70.net> wrote ----
 
 > the current standard allows rounding functions to either 
 > consistently raise inexact or not raise exceptions for 
 > non-integer inputs.  the c code in musl was written so 
 > it raises inexact (following fdlibm convention), and the 
 > tests try to check for that. 
 >  
 > but the other way is correct too and ieee754-2008 made 
 > that a requirement and future c fp standard (ts 18661) 
 > will require that too, so eventually both the c code in 
 > musl and the test system should be fixed to check for that. 
 >  

In that case,  may I start looking into updating the test cases? 

Could you be kind enough to point out an entry point case with which I can work out from there on?