AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-11-08 14:20:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use __builtin_fma more in dbl-64 code.

Browse Source 
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.
This commit is contained in:
Joseph Myers 2016-09-30 15:49:51 +00:00
parent 8278d50ce7
commit f280fa6d17
3 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ChangeLog
View File

@ -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.

4
sysdeps/ieee754/dbl-64/dla.h
View File

@ -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 */

8
sysdeps/x86_64/fpu/dla.h
View File

@ -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"