You can subscribe to this list here.
| 2002 |
Jan
|
Feb
|
Mar
|
Apr
|
May
|
Jun
|
Jul
|
Aug
|
Sep
(1) |
Oct
(122) |
Nov
(152) |
Dec
(69) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2003 |
Jan
(6) |
Feb
(25) |
Mar
(73) |
Apr
(82) |
May
(24) |
Jun
(25) |
Jul
(10) |
Aug
(11) |
Sep
(10) |
Oct
(54) |
Nov
(203) |
Dec
(182) |
| 2004 |
Jan
(307) |
Feb
(305) |
Mar
(430) |
Apr
(312) |
May
(187) |
Jun
(342) |
Jul
(487) |
Aug
(637) |
Sep
(336) |
Oct
(373) |
Nov
(441) |
Dec
(210) |
| 2005 |
Jan
(385) |
Feb
(480) |
Mar
(636) |
Apr
(544) |
May
(679) |
Jun
(625) |
Jul
(810) |
Aug
(838) |
Sep
(634) |
Oct
(521) |
Nov
(965) |
Dec
(543) |
| 2006 |
Jan
(494) |
Feb
(431) |
Mar
(546) |
Apr
(411) |
May
(406) |
Jun
(322) |
Jul
(256) |
Aug
(401) |
Sep
(345) |
Oct
(542) |
Nov
(308) |
Dec
(481) |
| 2007 |
Jan
(427) |
Feb
(326) |
Mar
(367) |
Apr
(255) |
May
(244) |
Jun
(204) |
Jul
(223) |
Aug
(231) |
Sep
(354) |
Oct
(374) |
Nov
(497) |
Dec
(362) |
| 2008 |
Jan
(322) |
Feb
(482) |
Mar
(658) |
Apr
(422) |
May
(476) |
Jun
(396) |
Jul
(455) |
Aug
(267) |
Sep
(280) |
Oct
(253) |
Nov
(232) |
Dec
(304) |
| 2009 |
Jan
(486) |
Feb
(470) |
Mar
(458) |
Apr
(423) |
May
(696) |
Jun
(461) |
Jul
(551) |
Aug
(575) |
Sep
(134) |
Oct
(110) |
Nov
(157) |
Dec
(102) |
| 2010 |
Jan
(226) |
Feb
(86) |
Mar
(147) |
Apr
(117) |
May
(107) |
Jun
(203) |
Jul
(193) |
Aug
(238) |
Sep
(300) |
Oct
(246) |
Nov
(23) |
Dec
(75) |
| 2011 |
Jan
(133) |
Feb
(195) |
Mar
(315) |
Apr
(200) |
May
(267) |
Jun
(293) |
Jul
(353) |
Aug
(237) |
Sep
(278) |
Oct
(611) |
Nov
(274) |
Dec
(260) |
| 2012 |
Jan
(303) |
Feb
(391) |
Mar
(417) |
Apr
(441) |
May
(488) |
Jun
(655) |
Jul
(590) |
Aug
(610) |
Sep
(526) |
Oct
(478) |
Nov
(359) |
Dec
(372) |
| 2013 |
Jan
(467) |
Feb
(226) |
Mar
(391) |
Apr
(281) |
May
(299) |
Jun
(252) |
Jul
(311) |
Aug
(352) |
Sep
(481) |
Oct
(571) |
Nov
(222) |
Dec
(231) |
| 2014 |
Jan
(185) |
Feb
(329) |
Mar
(245) |
Apr
(238) |
May
(281) |
Jun
(399) |
Jul
(382) |
Aug
(500) |
Sep
(579) |
Oct
(435) |
Nov
(487) |
Dec
(256) |
| 2015 |
Jan
(338) |
Feb
(357) |
Mar
(330) |
Apr
(294) |
May
(191) |
Jun
(108) |
Jul
(142) |
Aug
(261) |
Sep
(190) |
Oct
(54) |
Nov
(83) |
Dec
(22) |
| 2016 |
Jan
(49) |
Feb
(89) |
Mar
(33) |
Apr
(50) |
May
(27) |
Jun
(34) |
Jul
(53) |
Aug
(53) |
Sep
(98) |
Oct
(206) |
Nov
(93) |
Dec
(53) |
| 2017 |
Jan
(65) |
Feb
(82) |
Mar
(102) |
Apr
(86) |
May
(187) |
Jun
(67) |
Jul
(23) |
Aug
(93) |
Sep
(65) |
Oct
(45) |
Nov
(35) |
Dec
(17) |
| 2018 |
Jan
(26) |
Feb
(35) |
Mar
(38) |
Apr
(32) |
May
(8) |
Jun
(43) |
Jul
(27) |
Aug
(30) |
Sep
(43) |
Oct
(42) |
Nov
(38) |
Dec
(67) |
| 2019 |
Jan
(32) |
Feb
(37) |
Mar
(53) |
Apr
(64) |
May
(49) |
Jun
(18) |
Jul
(14) |
Aug
(53) |
Sep
(25) |
Oct
(30) |
Nov
(49) |
Dec
(31) |
| 2020 |
Jan
(87) |
Feb
(45) |
Mar
(37) |
Apr
(51) |
May
(99) |
Jun
(36) |
Jul
(11) |
Aug
(14) |
Sep
(20) |
Oct
(24) |
Nov
(40) |
Dec
(23) |
| 2021 |
Jan
(14) |
Feb
(53) |
Mar
(85) |
Apr
(15) |
May
(19) |
Jun
(3) |
Jul
(14) |
Aug
(1) |
Sep
(57) |
Oct
(73) |
Nov
(56) |
Dec
(22) |
| 2022 |
Jan
(3) |
Feb
(22) |
Mar
(6) |
Apr
(55) |
May
(46) |
Jun
(39) |
Jul
(15) |
Aug
(9) |
Sep
(11) |
Oct
(34) |
Nov
(20) |
Dec
(36) |
| 2023 |
Jan
(79) |
Feb
(41) |
Mar
(99) |
Apr
(169) |
May
(48) |
Jun
(16) |
Jul
(16) |
Aug
(57) |
Sep
(19) |
Oct
|
Nov
|
Dec
|
| S | M | T | W | T | F | S |
|---|---|---|---|---|---|---|
|
|
|
1
(1) |
2
(8) |
3
(7) |
4
(16) |
5
|
|
6
(3) |
7
(4) |
8
(1) |
9
(1) |
10
(4) |
11
(5) |
12
(1) |
|
13
|
14
(4) |
15
(2) |
16
|
17
(2) |
18
(9) |
19
(5) |
|
20
(9) |
21
(7) |
22
(9) |
23
(5) |
24
|
25
(1) |
26
|
|
27
|
28
(1) |
29
(11) |
30
(6) |
31
|
|
|
|
From: Jeremy F. <je...@go...> - 2002-10-04 23:02:59
|
On Fri, 2002-10-04 at 13:55, Julian Seward wrote: > Hmm, not bad. What's the numbers for --skin=memcheck and --skin=addrcheck ? > I know the improvement factor will be a lot less, but I'd still like to know > what it is. baseline lazy fp addrcheck 77.33 41.10 memcheck 82.98 44.76 J |
|
From: Julian S. <js...@ac...> - 2002-10-04 20:48:32
|
On Friday 04 October 2002 9:43 pm, Jeremy Fitzhardinge wrote: > On Fri, 2002-10-04 at 03:51, Julian Seward wrote: > > How much faster is "significantly faster" ? > > OK, I've quantified this now. > > Using the attached test program (a matrix multiply extracted from Mesa), > I'm getting the following timings (fptest compiled with -O; 600MHz PIII > laptop; --skin=none): > > native execution: 0.38s > baseline valgrind: 65.35s > lazy-fp valgrind: 4.05s > > In other words, Valgrind is currently 172 times slower than running > native for FP-intensive code; the lazy save-restore improves this by a > factor of 16 or so to make the valgrind overhead only about 11 times > slower than native. Hmm, not bad. What's the numbers for --skin=memcheck and --skin=addrcheck ? I know the improvement factor will be a lot less, but I'd still like to know what it is. /me is suitably impressed, just in case you were getting any other impression, btw. J |
|
From: Jeremy F. <je...@go...> - 2002-10-04 20:42:41
|
On Fri, 2002-10-04 at 03:51, Julian Seward wrote: > How much faster is "significantly faster" ? OK, I've quantified this now. Using the attached test program (a matrix multiply extracted from Mesa), I'm getting the following timings (fptest compiled with -O; 600MHz PIII laptop; --skin=none): native execution: 0.38s baseline valgrind: 65.35s lazy-fp valgrind: 4.05s In other words, Valgrind is currently 172 times slower than running native for FP-intensive code; the lazy save-restore improves this by a factor of 16 or so to make the valgrind overhead only about 11 times slower than native. I'd say that's significant. I'm attaching my current diff against HEAD; the previous one left out saving before JIFZ. J |
|
From: Jeremy F. <je...@go...> - 2002-10-04 20:29:25
|
On Fri, 2002-10-04 at 11:24, Josef Weidendorfer wrote:
> > index % time self children called name
> > [1] 100.0 0.03 0.23 1+2 <cycle 1 as a whole> [1]
> > 0.02 0.23 2 A <cycle 1> [3]
>
> Why isn't "B <cycle1>" shown as child of <cycle1> ?
Because [1] is the entry for "cycle 1 as a whole"; A is considered to be
the head of the cycle (which is then described in [3]).
[2] is the linear part of the call graph to A through main.
> > [Handwave mode:] For the purposes of displaying a profile as a TreeMap,
> > it seems to me that you need to break the graph up into strongly
> > connected subgraphs, and display each of those as a node in your tree.
>
> The subgraphs being the cycles with its children?
Yes. "Strongly connected graph" is graph-theory talk for a graph with
internal cycles. For example:
A A'
| --> |
v v
B <=> C BC'
B <=> C is a simple cycle, and so is a strongly connected subgraph. You
can rewrite the overall graph so that strongly connected subgraphs are
collapsed into a single node.
> I think I'm doing exactly this: Stop drawing at recursions...
> And the problem is: I don't know what the cost of the recursive function
> should be. I don't have the gprof results, but only my own logging.
> I think I will have to dive into gprof algorithms a little bit...
> As I understand at the moment, gprof never talks about recursive calls, but
> makes "cycles" with subobjects.
http://citeseer.nj.nec.com/graham82gprof.html should explain the
algorithm.
J
|
|
From: Josef W. <Jos...@gm...> - 2002-10-04 18:41:17
|
On Friday 04 October 2002 18:17, Jeremy Fitzhardinge wrote: > On Fri, 2002-10-04 at 03:10, Josef Weidendorfer wrote: > > Suppose a call chain starting from A: A calls B and C; C calls A a= gain. > > I don't think you should calculate anything at runtime; just record > enough so you can do it all afterwards. There shouldn't be a huge > problem in working all this out: gprof does it, after all. Perhaps I'm already logging to much with the "cumulative" costs of calls. But I don't understand how to calculate this with self costs and call tre= e=20 information alone, especially for functions taking part in multiple cycle= s... > Does this do what your example describes? > ... Yes :-) > gprof generates this flat profile: > ... Seems quite correct :-) > ... > index % time self children called name > [1] 100.0 0.03 0.23 1+2 <cycle 1 as a whole> [1] > 0.02 0.23 2 A <cycle 1> [3] Why isn't "B <cycle1>" shown as child of <cycle1> ? > ... > [Handwave mode:] For the purposes of displaying a profile as a TreeMap, > it seems to me that you need to break the graph up into strongly > connected subgraphs, and display each of those as a node in your tree. The subgraphs being the cycles with its children? > When you want to descend into that node (ie display its sub-parts), you > remove all the backwards edges (recursive calls) and treat it as a plai= n > tree. When you remove the backwards edge, you can recompute the cost o= f > the subgraph without the recursion; the difference between the original > cost and the new cost is how much you should attribute to the function > which is making the recursive call. I think I'm doing exactly this: Stop drawing at recursions... And the problem is: I don't know what the cost of the recursive function=20 should be. I don't have the gprof results, but only my own logging. I think I will have to dive into gprof algorithms a little bit... As I understand at the moment, gprof never talks about recursive calls, b= ut makes "cycles" with subobjects. Thanks anyway! Josef |
|
From: Josef W. <Jos...@gm...> - 2002-10-04 18:41:16
|
Forgot to send to the list...
On Friday 04 October 2002 12:48, Nicholas Nethercote wrote:
> On Fri, 4 Oct 2002, Josef Weidendorfer wrote:
> > The only problem I saw was that I need a valgrind version of the LIBC
> > "unlink", which I already mailed to Nick...
>
> I just added VG_(unlink) to head; it's untested, hopefully I got it
> right.
Thanks!
> > Regarding the valgrind skin architecture: Shouldn't it be possible to
> > "stack" skins? At the moment, for my skin I have to include all the
> > cachegrind code again. And if the cachegrind skin decides to simulate=
a
> > 3rd level cache, I have to copy it.
>
> Hmm, the LD_PRELOADing of two shared objects (skin.so + core.so) is
> already a bit fragile, having multiple .so's feels like a bad idea to
> me... Anyway, aren't Cachegrind and your patched version dissimilar
> enough that it wouldn't be easy to "stack" them in a sensible way? A
> better way might be to factor out the common code which gets included i=
n
> both skin .so's, if you see what I mean. This should be done with
> addrcheck and memcheck at some stage because they share a lot of identi=
cal
> code.
Yes. It seems to be the simplest way.
Two skins stacked on each other seems strange: The 1st does instrumentati=
on,=20
and the UCode outcome is instrumentated by the 2nd. And so on...
Regarding the LD_PRELOADing: can't this made be explicit by runtime-loadi=
ng of=20
the skins from valgrind.so (i.e. a plugin architecture)?
But a general "cost center API" would be nice to have for skins counting=20
events, linked as library to any such skin.
We need cost types (can be an array of subcost types), and cost center ta=
rget=20
types (e.g. instruction, basic block, call, function, ELF object, memory=20
access etc..) using these cost types. For this, we need register_XXX=20
functions, supplying call backs for zeroing/adding/writing ASCII version/=
=2E..
For each cost center a skin creates, it registers it and links it either =
to=20
some other cost center target ("parent") or to an existing object of the=20
valgrind core (e.g. basic block, memory area, ...).
Dumping out the profile could be fully done in a generic way: We need a "=
cost=20
center position" structure and go through all available positions of=20
registered cost centers (using e.g. the "parent" relation).
I'm sure the file format of dumps would change to be a lot more generic.
Support for per-thread cost centers could be generic, too.
I still need some time to think about it.
> Thinking longer term, your version of Cachegrind could entirely replace
> the original Cachegrind one day, since AFAICT your Cachegrind's
> functionality is a strict superset of my Cachegrind's.
Yes :-)
But it's still your baby, I only extended it. I can make small patches fo=
r=20
independent features separately (e.g. jump cost centers, recursion detect=
ion,=20
shared lib support [alias hash/_dl_resolve_runtime], "compressed" profile=
=20
format, threading support), and you decide if some modification is needed=
or=20
we can put it in as it is...
> > Perhaps you have some suggestions for my problem with recursive calls=
:
> >
> > Suppose a call chain starting from A: A calls B and C; C calls A agai=
n.
> > [...]
> > Suggestions?
>
> My brain is melting. Do you know how gprof handles it?
If I understand correctly:
gprof makes an additional virtual function, calling in "cycle" out of the
"A=3D>C=3D>A" chain, with subcycle objects A and C.
I would draw this cycle as area, splitting it up totally among A and C as
subareas. Unfortunately this doesn't show the function where the cycle wa=
s=20
entered from the outside. So I can rename the cycle back to the function=20
where the cycle was entered, and I'm back to my original proposal :-)
Still the question is how gprof calculates its results.
(see reply to the mail of Jeremy, too)
|
|
From: Jeremy F. <je...@go...> - 2002-10-04 16:50:11
|
On Fri, 2002-10-04 at 09:22, Julian Seward wrote:
The important change is to do with thread-specific data. New scheme on
x86s at least is that the linux kernel creates a GDT entry which points
at a thread's local storage; the thread can then access that merely
by using a suitable segment override prefix on any old memory referencing
insn. Kernel swizzles the GDT entry when scheduling threads, and
there is a new syscall by which a thread can tell the kernel
where it's TLD is [i assume because GDT can't be swizzled from user-space]
and because the kernel needs to know this anyway.
But if all this is just implementation detail of libpthreads and you're
replacing all of libpthreads, then surely you wouldn't have to deal with
it anyway?
- possibly your patches of 25 Sept if they don't cause build problems
on any test platform I have
Oh, you mean the poll/select name issue and the msgsnd/msgrcv
implementation?
J
|
|
From: Jeremy F. <je...@go...> - 2002-10-04 16:47:14
|
On Fri, 2002-10-04 at 09:17, Jeremy Fitzhardinge wrote:
On Fri, 2002-10-04 at 03:10, Josef Weidendorfer wrote:
Regarding the valgrind skin architecture: Shouldn't it be possible to "stack"
skins?
I've been thinking about that too. If the skins have non-overlapping
functionality and instrument the code in non-overlapping ways, then
maybe. It seems to me that you'd want to avoid having multiple skins
instrumenting each other's instrumentation. I guess you could add a
couple of new core UInstrs which limit CALLM_[SE], except for delimiting
s/limit/are like/
J
|
|
From: Jeremy F. <je...@go...> - 2002-10-04 16:17:41
|
On Fri, 2002-10-04 at 03:10, Josef Weidendorfer wrote:
Regarding the valgrind skin architecture: Shouldn't it be possible to "stack"
skins?
I've been thinking about that too. If the skins have non-overlapping
functionality and instrument the code in non-overlapping ways, then
maybe. It seems to me that you'd want to avoid having multiple skins
instrumenting each other's instrumentation. I guess you could add a
couple of new core UInstrs which limit CALLM_[SE], except for delimiting
a skin's instrumentation so other skins can avoid each other, but
there's then the problem of making sure the generated code is actually
correct...
But I still like the idea because my gprof skin is very non-intrusive
(it inserts just one call per basic block and a memory store per jump),
and could easily be composed with memcheck.
Suppose a call chain starting from A: A calls B and C; C calls A again. The
recursively called A only calls B. Say the cost of B is always 100, the self
cost of each A and C are 10. So the cumulative cost of C will be 120
(C=>A=>B), and the one of the first call to A will be 230. I log (cumulative)
costs for the call A=>B only once, so this call gets cost 200.
The problem: The callee list of A shows a call to B with cost 200 and a call
to C with cost 120, but A itself only has cumulative cost 230 !?! This is
confusing to the user, and really makes problems drawing the TreeMap...
The real problem is that KCachegrind can't see that the cost of A=>B is
included in the call cost A=>C and thus shown twice in the callee list of A.
And in the Treemap, I simple stop drawing recursive calls: This leaves empty
space where there should be none and it looks like a performance problem for
the user where there is none !!
The real solution (without the ad-hoc one) would be:
The callee list of A shows a call to B with cost 200. This is correct: B is
called twice from A, leading to cost 200 for calls to B. But the call A=>C
should be 20 only, skipping costs from any recursive A inside (perhaps
stating that cost 100 is already included in other calls). And this would
make the Treemap drawing fine again.
So the only question I have:
HOW to calculate this value (20) in the general case ?!?
I suppose I can't calulate it at post-processing time, but have to log it in
Cachegrind somehow (that is, the skipped cost of 100 in the example above).
I don't think you should calculate anything at runtime; just record
enough so you can do it all afterwards. There shouldn't be a huge
problem in working all this out: gprof does it, after all.
Does this do what your example describes?
#define COST(N) { int i; for(i = 0; i < N*1000000; i++); }
void A(int);
void B(void);
void C(void);
int main()
{
A(0);
return 0;
}
void A(int a)
{
COST(10);
B();
if (!a)
C();
}
void B(void)
{
COST(100);
}
void C(void)
{
COST(10);
A(1);
}
gprof generates this flat profile:
Each sample counts as 0.01 seconds.
% cumulative self self total
time seconds seconds calls us/call us/call name
88.46 0.23 0.23 2 115000.00 115000.00 B
7.69 0.25 0.02 2 10000.00 125000.00 A
3.85 0.26 0.01 1 10000.00 10000.00 C
and this call graph profile:
index % time self children called name
[1] 100.0 0.03 0.23 1+2 <cycle 1 as a whole> [1]
0.02 0.23 2 A <cycle 1> [3]
-----------------------------------------------
<spontaneous>
[2] 100.0 0.00 0.26 main [2]
0.03 0.23 1/1 A <cycle 1> [3]
-----------------------------------------------
1 C <cycle 1> [5]
0.03 0.23 1/1 main [2]
[3] 96.2 0.02 0.23 2 A <cycle 1> [3]
0.23 0.00 2/2 B [4]
1 C <cycle 1> [5]
-----------------------------------------------
0.23 0.00 2/2 A <cycle 1> [3]
[4] 88.5 0.23 0.00 2 B [4]
-----------------------------------------------
1 A <cycle 1> [3]
[5] 3.8 0.01 0.00 1 C <cycle 1> [5]
1 A <cycle 1> [3]
-----------------------------------------------
So basically this splits call graph into 3 pieces (indices 3, 4, 5):
3 is A seen as part of the cycle through A->C
4 is A calling B
5 is C seen as part of the cycle through C->A
[Handwave mode:] For the purposes of displaying a profile as a TreeMap,
it seems to me that you need to break the graph up into strongly
connected subgraphs, and display each of those as a node in your tree.
When you want to descend into that node (ie display its sub-parts), you
remove all the backwards edges (recursive calls) and treat it as a plain
tree. When you remove the backwards edge, you can recompute the cost of
the subgraph without the recursion; the difference between the original
cost and the new cost is how much you should attribute to the function
which is making the recursive call.
Does this help at all?
J
|
|
From: Julian S. <js...@ac...> - 2002-10-04 16:15:47
|
On Friday 04 October 2002 4:52 pm, Jeremy Fitzhardinge wrote: > On Fri, 2002-10-04 at 03:21, Julian Seward wrote: > Good news. After peering at weird segfaults on Red Hat Null (8.0 beta) > last night, I can see that it might be possible to assemble enough > hacks so that the stable branch will work on R H 8. Assuming that they > haven't changed the threading model used in the transition between the > "null" final beta and 8.0 itself, which doesn't sound likely. I thought > that 8.0 would use glibc-2.3, but apparently it is only at 2.2.93, so we > don't have to deal yet with big threading changes. > > How has threading changed in RH8 and/or glibc 2.3? Have they dropped > LinuxThreads? The important change is to do with thread-specific data. New scheme on x86s at least is that the linux kernel creates a GDT entry which points at a thread's local storage; the thread can then access that merely by using a suitable segment override prefix on any old memory referencing insn. Kernel swizzles the GDT entry when scheduling threads, and there is a new syscall by which a thread can tell the kernel where it's TLD is [i assume because GDT can't be swizzled from user-space] and because the kernel needs to know this anyway. This means accesses to TLD can be done in a single insn, which is dozens of times faster and more convenient doing pthread_key_get or whatever it's really called. However, this stuff isn't in R H Null, and I would be surprised if something that fundamental was changed in the transition from Null (the final RH8 beta) to RH8 itself. I'm downloading RH8 now but it takes ~ 4 days on a 64kbit/sec cable modem. LinuxThreads will be replaced by "Native Posix Threads" (I think) in due course although that will be binary-compatible with current libpthread.so . > I had expected only to be able to support R H 8 on the head, using the > LDT/GDT support, but it seems that might not be necessary. > > Vague plan therefore is to assemble this and various other bugfixes > > What fixes do you intend putting in the 1.0.X branch? For 1.0.4: - Support for RH8 if practical; mostly this means some (old-style) TLD improvments to do with thread-specific locales - add in the gcc-3.2 support patch available on the web page - someone sent a helpful patch to use dynamic symbols in elf executables/so's; this improves backtraces sometimes - the usual round of minor unimplemented opcodes, syscalls and ioctls - possibly your patches of 25 Sept if they don't cause build problems on any test platform I have No new functionality; that stuff is for the head. Far more important that the stable branch is stable. J |
|
From: Jeremy F. <je...@go...> - 2002-10-04 15:52:24
|
On Fri, 2002-10-04 at 03:21, Julian Seward wrote:
Good news. After peering at weird segfaults on Red Hat Null (8.0 beta)
last night, I can see that it might be possible to assemble enough hacks
so that the stable branch will work on R H 8. Assuming that they haven't
changed the threading model used in the transition between the "null"
final beta and 8.0 itself, which doesn't sound likely. I thought that
8.0 would use glibc-2.3, but apparently it is only at 2.2.93, so we
don't have to deal yet with big threading changes.
How has threading changed in RH8 and/or glibc 2.3? Have they dropped
LinuxThreads?
I had expected only to be able to support R H 8 on the head, using the
LDT/GDT support, but it seems that might not be necessary.
Vague plan therefore is to assemble this and various other bugfixes
What fixes do you intend putting in the 1.0.X branch?
J
|
|
From: Jeremy F. <je...@go...> - 2002-10-04 15:44:12
|
On Fri, 2002-10-04 at 03:51, Julian Seward wrote:
How much faster is "significantly faster" ?
I haven't measured it in detail, but the frame rate increased from about
1100ms/frame to 800-900ms/frame. I'll so some more scientific
measurements soon.
So, my main point. I think this patch is unsafe and will lead to hard
to find problems down the line. The difficulty is that it allows the
simulated FPU state to hang around in the real FPU for long periods,
up to a whole basic block's worth of execution (if I understand it
write).
We only need a skin to call out to a helper function which modifies
the real FPU state on some obscure path, and we're hosed. Since we don't
have any control over what skins people might plug in, this seems like
and unsafe modification to the core.
The modification I had in mind for a while was a lot more conservative,
and more along the lines of a peephole optimisation. Essentially
if we see a FPU-no-mem op followed by another FPU-no-mem op we can
skip the save at the end of the first and the restore at the start of
the second.
What I'm doing is not conceptually different from caching an ArchReg in
a RealReg for the lifetime of a basic block. The general idea is that
the FP state is pulled in just before the first FPU/FPU_[RW]
instruction, and saved again just before:
- JMP
- CCALL
- any skin UInstr
I can't see how a skin can introduce any instrumentation which would be
able to catch the FP state unsaved (is there any way for a skin to do
instrumentation or call a C function without using either CCALL or its
own UInstr?).
Your idea is basically the same, except we add a fourth saving
condition:
- any non FPU instruction
This would only be necessary if you imagine a non-FPU instruction which
can inspect the architectural state of the FPU (in other words, is a
memory access offset into the baseBlock: something which skins can't
generate directly).
In summary, I think this is actually pretty conservative, simple and
safe.
J
|
|
From: Nicholas N. <nj...@ca...> - 2002-10-04 10:48:27
|
On Fri, 4 Oct 2002, Josef Weidendorfer wrote: > The only problem I saw was that I need a valgrind version of the LIBC > "unlink", which I already mailed to Nick... I just added VG_(unlink) to head; it's untested, hopefully I got it right. > Regarding the valgrind skin architecture: Shouldn't it be possible to "stack" > skins? At the moment, for my skin I have to include all the cachegrind code > again. And if the cachegrind skin decides to simulate a 3rd level cache, I > have to copy it. Hmm, the LD_PRELOADing of two shared objects (skin.so + core.so) is already a bit fragile, having multiple .so's feels like a bad idea to me... Anyway, aren't Cachegrind and your patched version dissimilar enough that it wouldn't be easy to "stack" them in a sensible way? A better way might be to factor out the common code which gets included in both skin .so's, if you see what I mean. This should be done with addrcheck and memcheck at some stage because they share a lot of identical code. Thinking longer term, your version of Cachegrind could entirely replace the original Cachegrind one day, since AFAICT your Cachegrind's functionality is a strict superset of my Cachegrind's. > In fact: Call tree logging should be totally orthogonal to event > logging. Shouldn't we have some general support for expandable cost > centers in the core? Maybe, I thought about this but didn't get any further. If would help if you could give some specific suggestions as to the form it might take. > Perhaps you have some suggestions for my problem with recursive calls: > > Suppose a call chain starting from A: A calls B and C; C calls A again. > [...] > Suggestions? My brain is melting. Do you know how gprof handles it? N |
|
From: Julian S. <js...@ac...> - 2002-10-04 10:44:56
|
Cobbling together a response to this from the archives, since I didn't get it via the normal routes. > This patch makes FPU state changes lazy, so there should only be one > save/restore pair per basic block. With this change in place, > FPU-intensive programs (in my case, some 3D code using OpenGL) are > significantly faster. Interesting. This is something I'd wondered about doing at the time I did the FPU stuff in the first place. How much faster is "significantly faster" ? So, my main point. I think this patch is unsafe and will lead to hard to find problems down the line. The difficulty is that it allows the simulated FPU state to hang around in the real FPU for long periods, up to a whole basic block's worth of execution (if I understand it write). We only need a skin to call out to a helper function which modifies the real FPU state on some obscure path, and we're hosed. Since we don't have any control over what skins people might plug in, this seems like and unsafe modification to the core. The modification I had in mind for a while was a lot more conservative, and more along the lines of a peephole optimisation. Essentially if we see a FPU-no-mem op followed by another FPU-no-mem op we can skip the save at the end of the first and the restore at the start of the second. Looking at the stable branch vg_from_ucode.c and the codegen cases for FPU, FPU_R and FPU_W it's clear we can also do the same for FPU_R/W followed by FPU since there is no calls to helpers in the gap between these two. Or am I missing something? It would definitely be good to speed up the FPU stuff a bit, but I need to be convinced that you've got this 100% tied down in a not-too-complex way, in the face of arbitrary actions carried out by skins-not-invented-yet. J |
|
From: Josef W. <Jos...@gm...> - 2002-10-04 10:18:46
|
Hi, On Friday 04 October 2002 03:01, Jeremy Fitzhardinge wrote: > Hi, > > Do you have a patch for the current CVS version of valgrind? I finally > got enough of KDE installed on my laptop to compile kcachegrind, so I'm > keen to try it out. > > Thanks, > J Sorry: I don't have much time at the moment... Last wednesday I looked for the first time at valgrind-HEAD. It seems to be= =20 quite easy to port my patch to a skin. The only problem I saw was that I ne= ed=20 a valgrind version of the LIBC "unlink", which I already mailed to Nick...= =20 Regarding the valgrind skin architecture: Shouldn't it be possible to "stac= k"=20 skins? At the moment, for my skin I have to include all the cachegrind code= =20 again. And if the cachegrind skin decides to simulate a 3rd level cache, I= =20 have to copy it. In fact: Call tree logging should be totally orthogonal to= =20 event logging. Shouldn't we have some general support for expandable cost=20 centers in the core? Then I could use these to add/subtract costs without=20 even knowing which ones are logged... To be honest, I didn't thought much about this idea yet... Regarding KCachegrind: I still have a problem with visualizing recursive=20 calls. This seems to involve changes in the Cachegrind patch, too. So I first have to solve this one before I'm making any new patch/release... Aside from that: I switched to GCC 3.2 with an update to Suse 8.1, and now I have a lot of problems with lost debugging info :-( Question: What are the exact problems with GCC 3.x, that it's not officiall= y=20 supported in Valgrind ? Perhaps you have some suggestions for my problem with recursive calls: Suppose a call chain starting from A: A calls B and C; C calls A again. The= =20 recursively called A only calls B. Say the cost of B is always 100, the sel= f=20 cost of each A and C are 10. So the cumulative cost of C will be 120=20 (C=3D>A=3D>B), and the one of the first call to A will be 230. I log (cumul= ative)=20 costs for the call A=3D>B only once, so this call gets cost 200. The problem: The callee list of A shows a call to B with cost 200 and a cal= l=20 to C with cost 120, but A itself only has cumulative cost 230 !?! This is=20 confusing to the user, and really makes problems drawing the TreeMap... The real problem is that KCachegrind can't see that the cost of A=3D>B is=20 included in the call cost A=3D>C and thus shown twice in the callee list of= A. And in the Treemap, I simple stop drawing recursive calls: This leaves empty space where there should be none and it looks like a performance problem for the user where there is none !! The first ad-hoc solution was to distinguish among calls from recursively=20 called functions, i.e. the 2 calls A=3D>B will be logged independent from e= ach=20 other: this makes the example looking quite fine again. But this makes the real problem disappear for a few simple examples (as the= =20 above one) only; it's still there for deeper recursions, and cumulative cos= ts=20 of calls always include ALL recursion costs inside of this call: I log the= =20 cost counter difference at entering/leaving the function. The real solution (without the ad-hoc one) would be: The callee list of A shows a call to B with cost 200. This is correct: B is= =20 called twice from A, leading to cost 200 for calls to B. But the call A=3D>= C=20 should be 20 only, skipping costs from any recursive A inside (perhaps=20 stating that cost 100 is already included in other calls). And this would=20 make the Treemap drawing fine again. So the only question I have: HOW to calculate this value (20) in the general case ?!? I suppose I can't calulate it at post-processing time, but have to log it i= n=20 Cachegrind somehow (that is, the skipped cost of 100 in the example above). Suggestions? Josef |
|
From: Julian S. <js...@ac...> - 2002-10-04 10:14:55
|
Good news. After peering at weird segfaults on Red Hat Null (8.0 beta) last night, I can see that it might be possible to assemble enough hacks so that the stable branch will work on R H 8. Assuming that they haven't changed the threading model used in the transition between the "null" final beta and 8.0 itself, which doesn't sound likely. I thought that 8.0 would use glibc-2.3, but apparently it is only at 2.2.93, so we don't have to deal yet with big threading changes. I had expected only to be able to support R H 8 on the head, using the LDT/GDT support, but it seems that might not be necessary. Vague plan therefore is to assemble this and various other bugfixes into 1.0.4 and release that within about a week. I still plan to make a snapshot release of the head as 1.1.0 in the next couple of days, for the more adventurous. J |