This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: PowerPC's math_ldbl.h
- From: Adhemerval Zanella <azanella at linux dot vnet dot ibm dot com>
- To: libc-alpha at sourceware dot org
- Date: Sun, 17 Mar 2013 19:44:13 -0300
- Subject: Re: PowerPC's math_ldbl.h
- References: <87ehg9akpn.fsf@kepler.schwinge.homeip.net> <512626E1.4090801@linux.vnet.ibm.com> <5126DF54.4020806@linux.vnet.ibm.com> <Pine.LNX.4.64.1302221545410.6146@digraph.polyomino.org.uk> <5127C2DD.4090406@linux.vnet.ibm.com> <Pine.LNX.4.64.1302222119210.10719@digraph.polyomino.org.uk> <512835E4.70108@linux.vnet.ibm.com> <Pine.LNX.4.64.1303160028510.6194@digraph.polyomino.org.uk> <5143DE75.9040605@linux.vnet.ibm.com>
On 03/15/2013 11:52 PM, Adhemerval Zanella wrote:
> On 03/15/2013 09:34 PM, Joseph S. Myers wrote:
>> On Sat, 23 Feb 2013, Adhemerval Zanella wrote:
>>
>>> Also, the big change for fpu related code selection is
>>> sysdeps/powerpc/fpu now up to sysdeps/ieee754/ldlb-128ibm and
>>> sysdeps/ieee754/ldlb-opt (which also made sense). This currently does
>>> not change anything, because sysdeps/powerpc/fpu does not have any
>>> specific implementation for long double.
>> In a powerpc64 build, I see ABI test failures because sqrtl, version
>> GLIBC_2.3 (the compat alias to sqrt) is no longer exported from libm.
>> sysdeps/powerpc/fpu/w_sqrt.c is used, where I think
>> sysdeps/ieee754/ldbl-opt/w_sqrt.c would have been used before to provide
>> that compat symbol.
>>
>> Could you please fix this ABI failure if it was indeed caused by your
>> change and, if so, review the sysdeps orderings again to identify any
>> other cases of files that are now being found from a different directory
>> from before that your logic above may have missed?
>>
>> (Note: I haven't done tests before and after your commit to confirm it's
>> responsible, but it certainly looks likely to be.)
>>
> I believe my patch in fact uncover an issue with sysdeps/powerpc/fpu/w_sqrt.c:
> different from sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c it does *not* have:
>
> #if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
> compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0);
> #endif
>
> Which is basically what sysdeps/ieee754/ldbl-opt/w_sqrt.c does. Also, I think there
> is not need to neither sysdeps/powerpc/fpu/w_sqrt.c or
> ./sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c anymore: the latter was added to
> inline the ISA 2.03 fsqrt instruction and math_private.h already has an inline version
> that is used on math/w_sqrt.c.
>
> I'll recheck it tomorrow, but I'm sure we can delete both sysdeps/powerpc/fpu/w_sqrt.c
> and sysdeps/powerpc/powerpc64/power4/fpu/w_sqrt.c and use the default math/w_sqrt.c
> instead. And the fsqrt inline optimization will be include since it is defined at
> math_private.h.
>
> Thanks for spotting it, I usually just test build for power4+.
>
This patch fixes the ABI issue. I though about make a cleanup of the multiple w_sqrt
version on PPC, but I add in a TODO list. Ok to apply?
--
2013-03-17 Adhemerval Zanella <azanella@linux.vnet.ibm.com>
* sysdeps/powerpc/fpu/w_sqrt.c (compat_symbol): Add sqrtl compat symbol.
--
diff --git a/sysdeps/powerpc/fpu/w_sqrt.c b/sysdeps/powerpc/fpu/w_sqrt.c
index 2488ad9..70f28dd 100644
--- a/sysdeps/powerpc/fpu/w_sqrt.c
+++ b/sysdeps/powerpc/fpu/w_sqrt.c
@@ -19,6 +19,7 @@
#include <math.h>
#include <math_private.h>
#include <fenv_libc.h>
+#include <math_ldbl_opt.h>
double
__sqrt (double x) /* wrapper sqrt */
@@ -42,3 +43,6 @@ weak_alias (__sqrt, sqrt)
#ifdef NO_LONG_DOUBLE
strong_alias (__sqrt, __sqrtl) weak_alias (__sqrt, sqrtl)
#endif
+#if LONG_DOUBLE_COMPAT(libm, GLIBC_2_0)
+compat_symbol (libm, __sqrt, sqrtl, GLIBC_2_0);
+#endif
--
1.7.1