esolved, albeit not optimally, in the current version.
--
Peter Gibbs
EmKel Systems
Index: include/parrot/interpreter.h
===
RCS file: /home/perlcvs/parrot/include/parrot/interpreter.h,v
retrieving revision 1.45
diff -u -r1.45 interpreter.h
-
# New Ticket Created by "Peter Gibbs"
# Please include the string: [netlabs #628]
# in the subject line of all future correspondence about this issue.
# http://bugs6.perl.org/rt2/Ticket/Display.html?id=628 >
Following patch adds dependencies entry for hash.c to Makefile.
Stop
# New Ticket Created by "Peter Gibbs"
# Please include the string: [netlabs #629]
# in the subject line of all future correspondence about this issue.
# http://bugs6.perl.org/rt2/Ticket/Display.html?id=629 >
Attached patch does some fairly radical things to the memory manag
# New Ticket Created by "Peter Gibbs"
# Please include the string: [netlabs #691]
# in the subject line of all future correspondence about this issue.
# http://bugs6.perl.org/rt2/Ticket/Display.html?id=691 >
Index: RESPON
aster machine
and see what happens.
--
Peter Gibbs
EmKel Systems
doesn't want to use them due to the problems that occur when we
> start using exceptions (and longjmp, etc).
If performance has to halve in order to implement such features, I hope
somebody plans to write Parrot::Lite!
--
Peter Gibbs
EmKel Systems
> Finallythe unification of buffers and PMCs means that buffers can now
> point to things of their own accord, without requiring that they be
> surrounded by an accompanying PMC type.
How about the other way round? If the one-size-fits-all PMCs were to be
replaced by custom structures, then everything could be a PMC, and
buffer headers as a separate resource could just disappear!
--
Peter Gibbs
EmKel Systems
ed block, first allocate 16K, discard the amount before the alignment
point, and dish out the rest as 15 (or 16 if you're really lucky) 1K aligned
pages. I seriously considered this when I changed my buffer memory to be
paged instead of a single allocation per memory pool; but I haven't act
and perlstring that I also left out originally.
--
Peter Gibbs
EmKel Systems
ontext and act
accordingly. If context was an enumeration, a simple switch statement
would suffice. In this case, it might be interesting to benchmark the
consequences of replacing some of the other get_X vtable functions.
e.g. get_integer(pmc) -> get_value(pmc, CONTEXT_INTEGER)
Comments, any
Peter Gibbs wrote:
> vtable method get_scalar(pmc)
> vtable method get_value(pmc, context)
Having broken the rule of posting before drinking coffee in the morning,
this is obviously nonsense - neither of these will work, because we don't
know the return type of these fu
t) {
case CONTEXT_SCALAR: return BASETYPE_INTVAL;
}
}
void set_pmc(PMC* value)
{
switch (value->get_preferred_type(INTERP, value, CONTEXT_SCALAR)) {
case BASETYPE_INTVAL:
set_integer(INTERP, SELF, value);
break;
}
}
--
Peter Gibbs
EmKel Systems
C in any way. The future 'assign Px, Py' will call a vtable function
(set_pmc); however, in the quoted example, the pure register level
behaviour is all that is intended.
--
Peter Gibbs
EmKel Systems
PerlInt - in future, the perl compiler
will have to create a PMC with the correct type initially, and this
morphing will be removed for typed variables.
--
Peter Gibbs
EmKel Systems
to somewhere close to
that again.
--
Peter Gibbs
EmKel Systems
to the Buffer header structure
arose in a thread started by Dan himself, with contributions from Mike
Lambert and myself:
http:[EMAIL PROTECTED]/msg11424.html
The specific message from Dan prohibiting it was:
http:[EMAIL PROTECTED]/msg11570.html
--
Peter Gibbs
EmKel Systems
both fields to determine if transcoding is required. I do not
believe that the two fields are orthogonal, and therefore the
number of charsets would be less than #chartypes * #encodings.
b) some alterations to the single vtable thus created, in
particular the addition of a find_substring method.
--
sed on point 2. I will be sending
my patch to Mike so he can see if there is anything he can use.
--
Peter Gibbs
EmKel Systems
to do a diff) is an experimental version of string_repeat that
tries to reduce the number of calls to memcpy. I can't find my notes
at the moment as to what benefit it gave - perhaps you might like
to try it sometime. (Note that the string_make call will need to be
changed to the split chartype/encodin
an improvement on Hanoi of about 15% using your
COW patch, and your COW is better on both tests than mine.
--
Peter Gibbs
EmKel Systems
of the
loop. (patch attached, but not fully tested)
On my machine, that changes 5000 lives from 168 seconds to 133
(tested for one run each only)
I'm sure there is more that can be done, but that may help for now.
--
Peter Gibbs
EmKel Systems
dod.patch
Description: Binary data
(and hence PARROT_PTR_ALIGNMENT = 1), but IIRC a push
or pop will always change the SP by 2 or 4, therefore the stack walk code
can do the same. In a quick test, using a hardcoded value of 2 in the
stack walk code, 5000 lives dropped down to 112 seconds.
--
Peter Gibbs
EmKel Systems
ns somewhere.
I am still trying to track it down further - there is definitely
something very strange happening here somewhere.
I suspect a combination of slower and fatter may make it
hard to persuade Dan that we need cows?
--
Peter Gibbs
EmKel Systems
ocation here, rather than excess usage?
If so, grey will fix it in the next release with paged memory
allocation; and I'm sure you'll think of a solution also.
--
Peter Gibbs
EmKel Systems
with the addition of strstart also.
--
Peter Gibbs
EmKel Systems
overhead.
The next item on the list for grey was paged memory allocation - this
may be usable to some extent without the buffer linked lists; so I will
probably give that a spin anyway.
--
Peter Gibbs
EmKel Systems
main()
{
void * a;
void * b;
char c;
void * d;
struct {
void * a;
s, and strlen as being in encoding-defined characters.
Did something change when I wasn't looking, or is this a bug
just waiting for somebody to actually implement Unicode?
--
Peter Gibbs
EmKel Systems
zen ideas that
caused performance to deteriorate and were discarded.
Only changes that improved performance (or had a very
small detrimental effect but were required for subsequent
changes) survived.
--
Peter Gibbs
EmKel Systems
All tests pass with GC_DEBUG=0; three fail with GC_DEBUG=1,
due to bugs in the index op in core.ops (string_to_cstring returns
highly volatile pointers)
I have not done a full audit for cycle counter wrap problems yet,
so long-running programs may experience difficulties.
--
Peter Gibbs
EmKel Sy
d doesn't output quux.
Patch below should fix the problem. This is not an optimal solution,
as the unmake_COW is probably not required if the string_grow is
going to happen anyway, but it seems to follow the general spirit of
the current code.
ll fix the problem,
but I have been leaving those for Mike to look at.
In the meantime, a temporary quick-fix patch is included below.
--
Peter Gibbs
EmKel Systems
Index: string.c
===
RCS file: /cvs/public/parrot/string.c,v
retr
d pass, and hence string_grow not be called, but then
unmake_COW could shrink the destination buffer before the memcopy
happens.
--
Peter Gibbs
EmKel Systems
scrap selfpoolptr, and instead
set the external flag where you currently clear selfpoolptr.
--
Peter Gibbs
EmKel Systems
is not adjusted to
the desired alignment.
--
Peter Gibbs
EmKel Systems
structs
The 'total bytes allocated' for grey refers to the aligned page
allocator, and therefore now includes headers.
72k patch not attached.
--
Peter Gibbs
EmKel Systems
that could have been collected, or some sort of statistic that
> indicates how often the paged collections kick in?
I'll take a look at this tomorrow.
Thanks for your interest
--
Peter Gibbs
EmKel Systems
doesn't work when I try to change S0 from the
test
> set. But when I try to put a
> set S0, "v"
>
> just before using the ord instruction, then I get the good behavior.
> Of course, I can't use the set, since the char in S0 can be anything in a
> real
string Sy per type Iz
Currently supported types are:
0 = 8-bit
1 = 16-bit little-endian
2 = 16-bit big-endian
3 = 32-bit little-endian
4 = 32-bit big-endian
--
Peter Gibbs
EmKel Systems
convert.t
Description: Binary data
convert.h
Description: Binary data
t the platform's size and byteorder.
If the concept gets accepted, then we can start extending.
--
Peter Gibbs
EmKel Systems
ell with the current CVS DOD code; the above
tests were done with appropriate fixes.
--
Peter Gibbs
EmKel Systems
dy to check.
--
Peter Gibbs
EmKel Systems
y is only freed when the header
is freed, and such embedded headers are not
known to the dod system, memory allocated to
these embedded headers would stick around
forever??
--
Peter Gibbs
EmKel Systems
ey = 30 seconds
Doing any more iterations than that will take far too long,
we already know that CVS will always win when access
happens a lot more often than creation.
--
Peter Gibbs
EmKel Systems
Leopold Toetsch wrote:
> Peter Gibbs wrote:
> Did you look at indexed access, shift, unshift, splice ;-)
I have simply modified classes/array.pmc, which does not yet
implement shift or unshift; grey aims to provide alternative
implementations, rather than add new functionality. However,
value of a PMC was being updated,
rather than the PMC register being changed.
There are a lot of changes needed to existing code (e.g.
most of the tests) before the 'set' versions can be deleted.
--
Peter Gibbs
EmKel Systems
7;assign P0, P1' which would, somehow,
cause the existing PMC referenced by P0 to acquire the value
of the PMC referenced by P1 (via the vtable method set_pmc).
How to achieve this has never been resolved. The changes of
the other 'set P0, xxx' opcodes to 'assign' are basically for
consistency.
--
Peter Gibbs
EmKel Systems
wo can be fixed by the patch below.
The last one looks like a fundamental problem in MultiArray.
The line
b->cell_buffer = new_buffer_header(interpreter);
in function new_marray is creating a new buffer header, overwriting
the new_bufferlike_header created earlier.
--
Peter Gibbs
EmKel Sy
do is try to keep the overall noise
level as low as possible.
--
Peter Gibbs
EmKel Systems
;t (and I can expect it to vanish eventually)?
This is simply a long-standing GC bug. Try the attached patch and see if
it helps.
--
Peter Gibbs
EmKel Systems
dod.patch
Description: Binary data
ybody else, but with a potential
performance cost.
--
Peter Gibbs
EmKel Systems
set S1, S0
set S2, "another"
substr S0, S2, 1, 3
print S1
print "\n"
end
--
Peter Gibbs
EmKel Systems
revious example follows, as it points
out some more inconsistent behaviour.
--
Peter Gibbs
EmKel Systems
set S0, "test"
set S1, S0
set S2, "another"
substr S0, S2, 1, 3
print S1
print "\n"
set S0, "test"
new P1, .PerlString
set P1, S0
set S2, &q
nt *length);
};
--
--
Peter Gibbs
EmKel Systems
ever seems to have cancelled my privileges :-)
--
Peter Gibbs
EmKel Systems
Index: include/parrot/unicode.h
===
RCS file: /cvs/public/parrot/include/parrot/unicode.h,v
retrieving revision 1.5
diff -u -r1.5 unicode.h
--- include/parrot
I I
040 i s t h e s a m e a s U
060 T F 8 \n
[EMAIL PROTECTED] parrot]$ uname -a
Linux linux01.emkel.co.za 2.2.23 #2 Sun Apr 20 12:00:21 SAST 2003 i586
unknown
--
Peter Gibbs
EmKel Systems
turn the concatenation of the elements.
Iterator code has been amended to iterate into the elements.
Work is ongoing to handle substr, etc.
Comments welcome.
--
Peter Gibbs
EmKel Systems
values
to digit values
3) Add a pointer to the above struct to the CHARTYPE structure
Any comments on the above before I go ahead?
--
Peter Gibbs
EmKel Systems
or the less simple ones.
The current methods for both digit handling and transcoding are
context-free,
which I suspect may become a problem later; if so, some form of iterator
with
context information will be required.
--
Peter Gibbs
EmKel Systems
Leopold Toetsch wrote:
> Peter Gibbs <[EMAIL PROTECTED]> wrote:
> > +if (!chartype_array)
> >chartype_init();
> > +if (!encoding_array)
> >encoding_init();
>
> T think, both should be done at interpreter startup, called
her development.
All feedback welcome
--
Peter Gibbs
EmKel Systems
e sequence 'DIGIT ONE' 'DIGIT TWO' 'TAMIL DIGIT THREE' result
in twelve or one hundred and twenty three?
3) What should 'DIGIT ONE' 'VULGAR FRACTION ONE QUARTER' produce?
Any thoughts regarding signs, decimal points, exponent indicators etc. are
also welcome
--
Peter Gibbs
EmKel Systems
_unicode, /* from_unicode() */
Incorrect prototype for this function, now fixed - please try again.
Thanks
Peter Gibbs
Peter Gibbs
EmKel Systems
switch statement,
versus individual wrappers for each class.
I prefer the single function approach, so that is what I will start
implementing if there are no timeous objections.
Regards
Peter Gibbs
EmKel Systems
"Michael Scott" <[EMAIL PROTECTED]> wrote:
> Since the enum will specify what you yourself call character classes
> can't we call the function is_charclass() instead?
The isascii etc macros have been defined in a header called ctype.h for
some time, and glibc actually has a macro 'isctype' whic
check to see if I've broken anything too badly.
Regards
Peter Gibbs
EmKel Systems
eturn 1;
}
return 0;
For the hash_utf8 benchmark with the current code I get numbers like:
3.758691
5.535916
With the above iterator code (and the UTF8 decode_and_advance function
implemented) I get:
3.757812
4.844776
Does anybody think this is worth implementing?
Regards
Peter Gibbs
EmKel Systems
benchmark went from approx. 4.01/5.88 seconds to 3.97/5.31
seconds on my system.
The behaviour of the decode_and_advance functions in the various
encodings still needs proper testing; this will follow during the
weekend.
Regards
Peter Gibbs
EmKel Systems
2.4.22-1.2115.nptlsmp #1 SMP Wed Oct 29 15:30:09 EST 2003
i686 i686 i386 GNU/Linux
gcc --version
gcc (GCC) 3.3.2 20031022 (Red Hat Linux 3.3.2-1)
Please let me know if there is any further testing that would help.
Regards
Peter Gibbs
EmKel Systems
alized function, e.g.:
> - mem_alloc_executable
> - mem_realloc_executable
> - mem_flush_executable
> - mem_free_executable
Note that JIT already fails under Fedora Core 1 because of 'Exec-Shield'
, which is the Linux equivalent of the functionality described by
Jonathan.
Regards
Peter Gibbs
EmKel Systems
hether exec-shield is active, only that
there is an mprotect function with the appropriate attribute constants.
Regards
Peter Gibbs
EmKel Systems
n32 is now just plain broken,
as I do not have facilities to test it)
Regards
Peter Gibbs
EmKel Systems
> +if (p && *p != '#') {
Patch applied.
Many thanks
Peter Gibbs
EmKel Systems
David
In your last change (splitting buffer allocation from string) I assume you
also intended to shorten the initial allocation.
Peter Gibbs
EmKel Systems
Index: string.c
===
RCS file: /home/perlcvs/parrot/string.c,v
retrieving
free()
calls should go in free_string() rather than one in string_destroy() but I
don't know exactly what you and/or Dan intended here.
Peter Gibbs
EmKel Systems
Index: resources.c
===
RCS file: /home/perlcvs/parrot/res
to COW strings - would the buffer-related parameters (eg
bufstart, buflen, bufused) then not be linked to the buffer rather than the
string? ie should we have two structs (STRING & BUFFER) rather than one
combined one, with STRING pointing to BUFFER rather than directly to memory?
Peter Gibbs
I suspect that the length of the output buffer for string_transcode should
be based on the output encoding, not on the input encoding.
Peter Gibbs
EmKel Systems
Index: string.c
===
RCS file: /home/perlcvs/parrot/string.c,v
Another correction to string_transcode; this function now seems to work okay
(tested using a dummy 'encode' op added to my local copy of core.ops)
Peter Gibbs
EmKel Systems
Index: string.c
===
RCS file: /home/perl
udes buffer re-use (copy-on-write) for
substrings. On the other hand, passing strings to native C library routines
requires the termination. One compromise would be for 'unique' strings to be
terminated, and do 'copy-on-read' if a non-unique string (i.e. one borrowing
a buffer)
Jeff
If I understand the STRING struct correctly, buflen is the physical
allocation of the buffer, and bufused is the number of bytes actually used
by the string; therefore the latter would be correct for key_hash?
Peter Gibbs
EmKel Systems
Index: key.c
ion will be happening
on fixed boundaries, it may make sense to round the buffer length up
to match.
--
Peter Gibbs
EmKel Systems
Index: include/parrot/string.h
===
RCS file: /home/perlcvs/parrot/include/parrot/string.h,v
retrieving
t" to life.pasm in the
middle of the stats output lines, and on my linux x86 system this gives a
segfault, which seems to confirm that something has gone awry somewhere.
--
Peter Gibbs
EmKel Systems
Dan
I assume you meant pmc_pool not buffer_header_pool ??
--
Peter Gibbs
EmKel Systems
Index: resources.c
===
RCS file: /home/perlcvs/parrot/resources.c,v
retrieving revision 1.30
diff -c -r1.30 resources.c
*** resources.c 18 Mar
low fixes it by clearing the buffer pointer on allocation of a new
header; it could alternatively be done when the header is added to the free
list.
--
Peter Gibbs
EmKel Systems
Index: resources.c
===
RCS file: /home/perlcvs/parrot/r
nce both these problems involve design considerations, I am leaving it to
Dan to fix them.
Quick hacks to bypass the problems show that the pushp/popp test then works
fine.
--
Peter Gibbs
EmKel Systems
Clinton
The following patch seems to fix both these problems. It makes some slight
changes to the logic, so should be considered a temporary fix until Dan has
time to take a look at the code.
--
Peter Gibbs
EmKel Systems
Index: resources.c
again. Since the free pools are not
compressed by go_collect, perhaps they should be allocated independently and
not copied around all the time??
After fixing the above, the test program still abends, this time with
"subend somehow is less than substart" - I have not yet followed
reallocate.
A patch to create such a function and use it within add_pmc_to_free and
add_header_to_free follows; since all uses of mem_realloc are potentially at
risk, it might be a good idea to change any such places also.
--
Peter Gibbs
EmKel Systems
Index: incl
bc3, as the reference stays intact following
the re-assignment to $foo. Assuming that this will stay the same for perl 6,
we must either keep the original PMC address, or find and correct all
references to it.
--
Peter Gibbs
EmKel Systems
n the same place
as only creating a new string if the final parameter is NULL; the
implementations again differ.
I agree that the whole issue of garbage collection relating to newborns and
temporaries is problematic. One option might be to always flag newly-created
objects in some way, then have a single function call to release them; at
least this means just one function call at the end of each procedure. This
still leaves a problem with nested procedures.
--
Peter Gibbs
EmKel Systems
as discussed above, so "once a cow, always a cow", which will
have some negative impact.
--
Peter Gibbs
EmKel Systems
f useless blocks, or should we just let them be created and
discarded?
--
Peter Gibbs
EmKel Systems
epeat; this could be modified to do a simple
string_copy if the repeat count was one, and to act in-place if possible
Since then, string_replace has been added; for now, I will just change this
to do a straight un-cow (i.e. make a copy of the buffer if it is marked as
COW)
> Mike Lambert
>
--
Peter Gibbs
EmKel Systems
>> On Tuesday 02 April 2002 04:36, Peter Gibbs wrote:
>"Bryan C. Warnock" <[EMAIL PROTECTED]> wrote:
> buflen should be 16, not 0. bufused should be 0.
Currently, string_make sets the value of buflen. It does not know about
Parrot_allocate's rounding rule,
on; since this is only for a mode string, the
chances are very small, but not zero. Why are we passing a C-style string
around inside parrot anyway?
--
Peter Gibbs
EmKel Systems
uch), which would then
be called by both Parrot_allocate and Parrot_reallocate. The
alignment/padding rules would then be implemented in _allocate and
_reallocate, but nowhere else.
Of course, putting the 'correct' buflen into the string header messes up my
evil use of the last byte for
, after the DOD run, there is still less than N headers
available, allocate more even though we can satisfy the immediate
requirement. This would improve performance by reducing the number of DOD
runs, but at the cost of additional memory - a classic tradeoff!
--
Peter Gibbs
EmKel Systems
2412 collection runs were made
Copying a total of 0 bytes
There are 81 active Buffer structs
There are 256 total Buffer structs
--
Peter Gibbs
EmKel Systems
The string_grow function (currently used only by string_replace) does not
allocate a new buffer if there are no bytes to be copied from old buffer to
new buffer. Patch below fixes this.
--
Peter Gibbs
EmKel Systems
Index: string.c
al of 57801936 bytes
so a 97% decrease in the number of collection runs only turns into an 11%
improvement in total performance.
[166MHz Pentium; linux 2.2.18]
--
Peter Gibbs
EmKel Systems
1 - 100 of 159 matches
Mail list logo
