k_tanf.c but with different details.
The polynomial is odd with degree 13 for tanf() and odd with degree
9 for sinf(), so the details are not very different for sinf() -- the
term with the x**11 and x**13 coefficients goes awaym and (mysteriously)
it helps to do the evaluation of w = z*z early although moving it later
was a key optimization for tanf(). The details are different but simpler
for cosf() because the polynomial is even and of lower degree.
On Athlons, for uniformly distributed args in [-2pi, 2pi], this gives
an optimization of about 4 cycles (10%) in most cases (13% for sinf()
on AXP, but 0% for cosf() with gcc-3.3 -O1 on AXP). The best case
(sinf() with gcc-3.4 -O1 -fcaller-saves on A64) now takes 33-39 cycles
(was 37-45 cycles). Hardware sinf takes 74-129 cycles. Despite
being fine tuned for Athlons, the optimization is even larger on
some other arches (about 15% on ia64 (pluto2) and 20% on alpha (beast)
with gcc -O2 -fomit-frame-pointer).
cosf(x) is supposed to return something like x when x is a NaN, and
we actually fairly consistently return x-x which is normally very like
x (on i386 and and it is x if x is a quiet NaN and x with the quiet bit
set if x is a signaling NaN. Rev.1.10 broke this by normalising x to
fabsf(x). It's not clear if fabsf(x) is should preserve x if x is a NaN,
but it actually clears the sign bit, and other parts of the code depended
on this.
The bugs can be fixed by saving x before normalizing it, and using the
saved x only for NaNs, and using uint32_t instead of int32_t for ix
so that negative NaNs are not misclassified even if fabsf() doesn't
clear their sign bit, but gcc pessimizes the saving very well, especially
on Athlon XPs (it generates extra loads and stores, and mixes use of
the SSE and i387, and this somehow messes up pipelines). Normalizing
x is not a very good optimization anyway, so stop doing it. (It adds
latency to the FPU pipelines, but in previous versions it helped except
for |x| <= 3pi/4 by simplifying the integer pipelines.) Use the same
organization as in s_sinf.c and s_tanf.c with some branches reordered.
These changes combined recover most of the performance of the unfixed
version on A64 but still lose 10% on AXP with gcc-3.4 -O1 but not with
gcc-3.3 -O1.
that was used doesn't work normally here, since we want to be able to
multiply `hi' by the exponent of x _exactly_, and the exponent of x has
more than 7 significant bits for most denormal x's, so the multiplication
was not always exact despite a cloned comment claiming that it was. (The
comment is correct in the double precision case -- with the normal 33+53
bit decomposition the exponent can have 20 significant bits and the extra
bit for denormals is only the 11th.)
Fixing this had little or no effect for denormals (I think because
more precision is inherently lost for denormals than is lost by roundoff
errors in the multiplication).
The fix is to reduce the precision of the decomposition to 16+24 bits.
Due to 2 bugs in the old deomposition and numerical accidents, reducing
the precision actually increased the precision of hi+lo. The old hi+lo
had about 39 bits instead of at least 41 like it should have had.
There were off-by-1-bit errors in each of hi and lo, apparently due
to mistranslation from the double precision hi and lo. The correct
16 bit hi happens to give about 19 bits of precision, so the correct
hi+lo gives about 43 bits instead of at least 40. The end result is
that expf() is now perfectly rounded (to nearest) except in 52561 cases
instead of except in 67027 cases, and the maximum error is 0.5013 ulps
instead of 0.5023 ulps.
sosend(). Robert accidentally changed the snderr() macro to jump to the
out label which assumes the lock is already released rather than the
release label which drops the lock in his previous change to sosend().
This should fix the recent panics about returning from write(2) with the
socket lock held and the most recent LOR on current@.
much later than before, and it is now after we do a mkdir ../compile/FILE.
As a result, if you do 'config DOESNOTEXIST', it now creates the directory
../config/DOESNOTEXIST. It did not do that before. If DEFAULTS does not
exist, it still fails early before any permanent changes.
This shameless hack restores the old behavior of ensuring the config file
actually exists before mkdiring its counterpart directory.
Now I can rmdir ../compile/D and it will stay dead, after my fingers keep
sabotaging me with 'config D<tab><enter>'. (Some of my kernel names
started with D, which used to be 1-character unique and my fingers knew
this very well...)
now that all that stuff has been abstracted out of the disk drivers with
GEOM.
- Reference bsdlabel(8) rather than disklabel(8).
Ok'd by: phk, scottl (1)
Submitted by: Björn König (2)
IPv6 support was committed:
- Stop treating `ip' and `ipv6' as special in `proto' option as they
conflict with /etc/protocols.
- Disuse `ipv4' in `proto' option as it is corresponding to `ipv6'.
- When protocol is specified as numeric, treat it as it is even it is
41 (ipv6).
- Allow zero for protocol as it is valid number of `ip'.
Still, we cannot specify an IPv6 over an IPv4 tunnel like before such
as:
pass ipv6 from any to any
But, now, you can specify it like:
pass ip4 from any to any proto ipv6
PR: kern/89472
Reported by: Ga l Roualland <gael.roualland__at__dial.oleane.com>
MFC after: 1 week
The sys/sys/stddef.h is here for some time now to fulfil the
kernel needs. It also was not reliable due to the exists(@)
check: in an empty module directory, "make depend; mv .depend
.depend~; make depend" ran mkdep(1) with different arguments.
o Do not use ipfw_insn_pipe->pipe_ptr in locate_flowset(). The
_ipfw_insn_pipe isn't touched by this commit to preserve ABI
compatibility.
o To optimize the lookup of the pipe/flowset in locate_flowset()
introduce hashes for pipes and queues:
- To preserve ABI compatibility utilize the place of global list
pointer for SLIST_ENTRY.
- Introduce locate_flowset(queue nr) and locate_pipe(pipe nr).
o Rework all the dummynet code to deal with the hashes, not global
lists. Also did some style(9) changes in the code blocks that were
touched by this sweep:
- Be conservative about flowset and pipe variable names on stack,
use "fs" and "pipe" everywhere.
- Cleanup whitespaces.
- Sort variables.
- Give variables more meaningful names.
- Uppercase and dots in comments.
- ENOMEM when malloc(9) failed.
- S3 Savage driver ported.
- Added support for ATI_fragment_shader registers for r200.
- Improved r300 support, needed for latest r300 DRI driver.
- (possibly) r300 PCIE support, needs X.Org server from CVS.
- Added support for PCI Matrox cards.
- Software fallbacks fixed for Rage 128, which used to render badly or hang.
- Some issues reported by WITNESS are fixed.
- i915 module Makefile added, as the driver may now be working, but is untested.
- Added scripts for copying and preprocessing DRM CVS for inclusion in the
kernel. Thanks to Daniel Stone for getting me started on that.
rare case of a stray interrupt to an unregistered source (such as a stray
interrupt from the 8259As when using APIC), this could result in a page
fault when it tried to walk the list of interrupt handlers to execute
INTR_FAST handlers. This bug was introduced with the intr_event changes,
so it's not present in 5.x or 6.x.
Submitted by: Mark Tinguely tinguely at casselton dot net
process as over the limit when its time is >= to the limit rather than >
the limit. Technically, if p->p_rux.rux_runtime.sec == p->p_pcpulimit
and p->p_rux.rux_runtime.frac == 0, the process hasn't exceeded the limit
yet. However, having the fraction exactly equal to 0 is rather rare, and
it is not worth the overhead to handle that edge case. With just the >
comparison, the process would have to exceed its limit by almost a second
before it was killed.
PR: kern/83192
Submitted by: Maciej Zawadzinski mzawadzinski at gmail dot com
Reviewed by: bde
MFC after: 1 week
