Re: cacosh/csin failures on s390x since CORE-MATH integration

[Mark Wielaard <mark@klomp.org>]

Hi Adhemerval,

On Thu, 2024-11-07 at 13:40 +0000, Adhemerval Zanella Netto wrote:
> It is expected [1], I had to adjust it on x86, arm, and powerpc as well.
> The __expm1 is exact for the input range, but due how csin/cacoshf
> handles the expm1 result there is addition cancellation errors.
> 
> In this case someone will need to run 'make update-ulps' on a real
> hardware and update the libm-test-ulps. 

Aha, found https://sourceware.org/glibc/wiki/Regeneration#Math_files
Running make regen-ulps on an s390x machine produced:

Automatic regeneration of ULPs complete.
Difference between the current baseline and the new baseline is:
--- ../sysdeps/s390/fpu/libm-test-ulps	2024-11-07 16:53:35.036318696 +0000
+++ /home/mjw/glibc-obj/math/NewUlps	2024-11-07 17:09:45.376836323 +0000
@@ -228,7 +228,7 @@
 
 Function: Real part of "cacosh_upward":
 double: 4
-float: 3
+float: 4
 ldouble: 6
 
 Function: Imaginary part of "cacosh_upward":
@@ -720,6 +720,7 @@
 ldouble: 1
 
 Function: Imaginary part of "csin":
+float: 1
 ldouble: 1
 
 Function: Real part of "csin_downward":
Copy /home/mjw/glibc-obj/math/NewUlps to ../sysdeps/s390/fpu/libm-test-ulps (relative to source).

Will sent a patch after another make check run to verify the results.

Thanks,

Mark

> [1] https://sourceware.org/pipermail/libc-alpha/2024-October/161131.html