Re: [MinGW-dvlpr] Re: mthread patch

[<dan...@ya...>]

 --- Thomas Pfaff <tp...@gm...> wrote: > 
> 
> On Wed, 28 Nov 2001, Danny Smith wrote:
> 
> > Your patch does indeed fix a memory leak with 2.95.3.  However, there
> does
> > not appear to be any leak (using your testcase) on 3.1, with or without
> > your patch.  I'm guessing that EH  code cleans up after itself better
> in
> > GCC 3.1.
> 
> If you can give me an ftp access to your gcc-3.1 snapshot i will take a
> look at it.
> 

Sorry gcc core and g++ src about 12MB gzipped.  My connection is very slow.
 I won't be uploading 3.1 srcs and binaries anywhere until I'm sure I won't
have to do it again soon.  I can send gcc/gthr-win32.h and
gcc/unwind-sjlj.c (where libgcc's generic exception code lives) or you can
grab those from gcc CVS

> >
> > Is patch necessary or beneficial for other reasons? Can you comment on
> > performance costs of your patch versus Mumit's version?
> >
> 
> This patch was only meant to create TLS keys similar to
> pthread_key_create, where you can give a dtor function that will be
> called
> on thread termination. Look at Mumits comments in mthr.c
> Unfortunately his implementation was not correct and has lead to memory
> leaks. Therefore it does not matter which code is faster (i guess Mumits
> is because free() is only called once ;-) ).
> 
> If my patch is not neccessary with gcc-3.1 you can forget Mumits version
> as well. You can easily try this if you link the program without
> libmingthrd.a. If there is no differende than the mthr code is simply not
> needed (at least whith thread exception cleanup).
> 

I have tried 3.1 without -mthreads switch and  these are results

calls   free    delta   
        mem
0	101912	0
1000	101872	-40
2000	101924	12
3000	101924	12
4000	101916	4
5000	101908	-4
6000	101936	24
7000	101928	16
8000	101928	16
9000	101928	16
10000	101936	24
11000	101936	24
12000	101928	16
13000	101928	16
14000	101876	-36
15000	101876	-36
16000	101848	-64
17000	101856	-56
18000	101856	-56
19000	101828	-84
20000	101828	-84
21000	101864	-48
22000	101864	-48
23000	101856	-56
24000	101876	-36
25000	101868	-44
26000	101868	-44
27000	101868	-44
28000	101884	-28
29000	101884	-28

No apparent leak

I need to study unwind-sjlj.c some more to follow what is actually
happening

I won't push the patch to gthr-win32.h for gcc-3.1 foe awhile.  Feel free
to do so yourself if you can answer the questions that will come back from
that list.  Please understand that I am not criticising your efforts.  I am
just having trouble with what is for me a very steep learning curve.

Danny

> 
> _______________________________________________
> MinGW-dvlpr mailing list
> Min...@li...
> https://lists.sourceforge.net/lists/listinfo/mingw-dvlpr 

http://shopping.yahoo.com.au - Yahoo! Shopping
- Get organised for Christmas early this year!