GCC 4.8 enables -ftree-loop-distribute-patterns at -O3 by default and
this optimization may transform loops into memset/memmove calls. Without
proper handling this may generate unexpected PLT calls on GLIBC.
This patch fixes by create memset/memmove alias to internal GLIBC
__GI_memset/__GI_memmove symbols.
The most common use case of math functions is with default rounding
mode, i.e. rounding to nearest. Setting and restoring rounding mode
is an unnecessary overhead for this, so I've added support for a
context, which does the set/restore only if the FP status needs a
change. The code is written such that only x86 uses these. Other
architectures should be unaffected by it, but would definitely benefit
if the set/restore has as much overhead relative to the rest of the
code, as the x86 bits do.
Here's a summary of the performance improvement due to these
improvements; I've only mentioned functions that use the set/restore
and have benchmark inputs for x86_64:
Before:
cos(): ITERS:4.69335e+08: TOTAL:28884.6Mcy, MAX:4080.28cy, MIN:57.562cy, 16248.6 calls/Mcy
exp(): ITERS:4.47604e+08: TOTAL:28796.2Mcy, MAX:207.721cy, MIN:62.385cy, 15543.9 calls/Mcy
pow(): ITERS:1.63485e+08: TOTAL:28879.9Mcy, MAX:362.255cy, MIN:172.469cy, 5660.86 calls/Mcy
sin(): ITERS:3.89578e+08: TOTAL:28900Mcy, MAX:704.859cy, MIN:47.583cy, 13480.2 calls/Mcy
tan(): ITERS:7.0971e+07: TOTAL:28902.2Mcy, MAX:1357.79cy, MIN:388.58cy, 2455.55 calls/Mcy
After:
cos(): ITERS:6.0014e+08: TOTAL:28875.9Mcy, MAX:364.283cy, MIN:45.716cy, 20783.4 calls/Mcy
exp(): ITERS:5.48578e+08: TOTAL:28764.9Mcy, MAX:191.617cy, MIN:51.011cy, 19071.1 calls/Mcy
pow(): ITERS:1.70013e+08: TOTAL:28873.6Mcy, MAX:689.522cy, MIN:163.989cy, 5888.18 calls/Mcy
sin(): ITERS:4.64079e+08: TOTAL:28891.5Mcy, MAX:6959.3cy, MIN:36.189cy, 16062.8 calls/Mcy
tan(): ITERS:7.2354e+07: TOTAL:28898.9Mcy, MAX:1295.57cy, MIN:380.698cy, 2503.7 calls/Mcy
So the improvements are:
cos: 27.9089%
exp: 22.6919%
pow: 4.01564%
sin: 19.1585%
tan: 1.96086%
The downside of the change is that it will have an adverse performance
impact on non-default rounding modes, but I think the tradeoff is
justified.
This is the initial support for string function performance tests,
along with copying tests for memcpy and memcpy-ifunc as proof of
concept. The string function benchmarks perform operations at
different alignments and for different sizes and compare performance
between plain operations and the optimized string operations. Due to
this their output is incompatible with the function benchmarks where
we're interested in fastest time, throughput, etc.
In future, the correctness checks in the benchmark tests can be
removed. Same goes for the performance measurements in the
string/test-*.
__clock_gettime and other __clock_* functions could result in an extra
PLT reference within libc.so if it actually gets used. None of the
code currently uses them, which is why this probably went unnoticed.
It is very very possible that the futex syscall returns an
error and that the caller of lll_futex_wake may want to
look at that error and propagate the failure.
This patch allows a caller to see the syscall error.
There are no users of the syscall error at present, but
future cleanups are now be able to check for the error.
--
nplt/
2013-06-10 Carlos O'Donell <carlos@redhat.com>
* sysdeps/unix/sysv/linux/i386/lowlevellock.h
(lll_futex_wake): Return syscall error.
* sysdeps/unix/sysv/linux/x86_64/lowlevellock.h
(lll_futex_wake): Return syscall error.
When setting BENCH_DURATION in CPPFLAGS-nonlib, append to the variable
instead of assigning to it, to avoid overwriting earlier set flags,
notably the -DNOT_IN_libc=1 flag.
In 128-bit IBM long double the precision of the type
decreases as you approach subnormal numbers, equaling
that of a double for subnormal numbers. Therefore
adjust the computation in ulp to use 2^(MIN_EXP - MANT_DIG)
which is correct for FP_SUBNORMAL for all types.
