This is the mail archive of the
glibc-cvs@sourceware.org
mailing list for the glibc project.
GNU C Library master sources branch master updated. glibc-2.24-204-gf280fa6
- From: jsm28 at sourceware dot org
- To: glibc-cvs at sourceware dot org
- Date: 30 Sep 2016 15:50:51 -0000
- Subject: GNU C Library master sources branch master updated. glibc-2.24-204-gf280fa6
This is an automated email from the git hooks/post-receive script. It was
generated because a ref change was pushed to the repository containing
the project "GNU C Library master sources".
The branch, master has been updated
via f280fa6d171c4d3414c005ad2a7529e0d1d9ee0c (commit)
from 8278d50ce7f48dc5af68fac850192c7666661cd4 (commit)
Those revisions listed above that are new to this repository have
not appeared on any other notification email; so we list those
revisions in full, below.
- Log -----------------------------------------------------------------
http://sourceware.org/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=f280fa6d171c4d3414c005ad2a7529e0d1d9ee0c
commit f280fa6d171c4d3414c005ad2a7529e0d1d9ee0c
Author: Joseph Myers <joseph@codesourcery.com>
Date: Fri Sep 30 15:49:51 2016 +0000
Use __builtin_fma more in dbl-64 code.
sysdeps/ieee754/dbl-64/dla.h can use a macro DLA_FMS for more
efficient double-width operations when fused multiply-subtract is
supported. However, this macro is only defined for x86_64,
conditional on architecture-specific __FMA4__. This patch makes the
code use __builtin_fma conditional on __FP_FAST_FMA, as used elsewhere
in glibc.
Tested for x86_64, x86 and powerpc. On powerpc (where this is causing
fused operations to be used where they weren't previously) I see an
increase from 1ulp to 2ulp in the imaginary part of clog10:
testing double (without inline functions)
Failure: Test: Imaginary part of: clog10 (0x1.7a858p+0 - 0x6.d940dp-4 i)
Result:
is: -1.2237865208199886e-01 -0x1.f5435146bb61ap-4
should be: -1.2237865208199888e-01 -0x1.f5435146bb61cp-4
difference: 2.7755575615628914e-17 0x1.0000000000000p-55
ulp : 2.0000
max.ulp : 1.0000
Maximal error of real part of: clog10
is : 3 ulp
accepted: 3 ulp
Maximal error of imaginary part of: clog10
is : 2 ulp
accepted: 1 ulp
This is actually resulting from atan2 becoming *more* accurate (atan2
(-0x6.d940dp-4, 0x1.7a858p+0) should ideally be -0x1.208cd6e841554p-2
but was -0x1.208cd6e841555p-2 from a powerpc libm built before this
change, and is -0x1.208cd6e841554p-2 from a powerpc libm built after
this change). Since these functions are not expected to be correctly
rounding by glibc's accuracy goals, neither result is a problem, but
this does imply that some of this code, although designed to be
correctly rounding, is not in fact correctly rounding (possibly
because of GCC creating fused operations where the code does not
expect it, something we've only disabled for specific functions where
it was found to cause large errors). (Of course as previously
discussed I think we should remove the slow cases where an error
analysis shows this wouldn't increase the errors much above 0.5ulp;
it's only functions such as cratan2 that are expected to be correctly
rounding, not atan2.)
* sysdeps/ieee754/dbl-64/dla.h [__FP_FAST_FMA] (DLA_FMS): Define
macro to use __builtin_fma.
* sysdeps/x86_64/fpu/dla.h: Remove file.
diff --git a/ChangeLog b/ChangeLog
index c61990d..ec66770 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,9 @@
2016-09-30 Joseph Myers <joseph@codesourcery.com>
+ * sysdeps/ieee754/dbl-64/dla.h [__FP_FAST_FMA] (DLA_FMS): Define
+ macro to use __builtin_fma.
+ * sysdeps/x86_64/fpu/dla.h: Remove file.
+
* sysdeps/ieee754/ldbl-128ibm/bits/iscanonical.h
[__NO_LONG_DOUBLE_MATH] (__iscanonicall): Do not declare.
[__NO_LONG_DOUBLE_MATH] (iscanonical): Define to evaluate to 1.
diff --git a/sysdeps/ieee754/dbl-64/dla.h b/sysdeps/ieee754/dbl-64/dla.h
index d21c47a..d64e630 100644
--- a/sysdeps/ieee754/dbl-64/dla.h
+++ b/sysdeps/ieee754/dbl-64/dla.h
@@ -57,6 +57,10 @@
z=(x)-(y); zz=(fabs(x)>fabs(y)) ? (((x)-(z))-(y)) : ((x)-((y)+(z)));
+#ifdef __FP_FAST_FMA
+# define DLA_FMS(x, y, z) __builtin_fma (x, y, -(z))
+#endif
+
/* Exact multiplication of two single-length floating point numbers, */
/* Veltkamp. The macro produces a double-length number (z,zz) that */
/* satisfies z+zz = x*y exactly. p,hx,tx,hy,ty are temporary */
diff --git a/sysdeps/x86_64/fpu/dla.h b/sysdeps/x86_64/fpu/dla.h
deleted file mode 100644
index 688efa0..0000000
--- a/sysdeps/x86_64/fpu/dla.h
+++ /dev/null
@@ -1,8 +0,0 @@
-#include <features.h>
-
-#ifdef __FMA4__
-# define DLA_FMS(x,y,z) \
- __builtin_fma (x, y, -(z))
-#endif
-
-#include "sysdeps/ieee754/dbl-64/dla.h"
-----------------------------------------------------------------------
Summary of changes:
ChangeLog | 4 ++++
sysdeps/ieee754/dbl-64/dla.h | 4 ++++
sysdeps/x86_64/fpu/dla.h | 8 --------
3 files changed, 8 insertions(+), 8 deletions(-)
delete mode 100644 sysdeps/x86_64/fpu/dla.h
hooks/post-receive
--
GNU C Library master sources