Re: [ros-svn] [ion] 13130: Bug fixes against uninitizlied variables and support for tree-wide optimization (do not try yet, crashes in bootvid).

[Alex Ionescu]

ion(a)svn.reactos.com wrote:

>  Bug fixes against uninitizlied variables and support
> for tree-wide optimization (do not try yet, crashes in bootvid).
>
>
> Updated files:
> trunk/reactos/config
> trunk/reactos/drivers/lib/ip/transport/datagram/datagram.c
> trunk/reactos/drivers/net/afd/afd/connect.c
> trunk/reactos/drivers/net/afd/afd/write.c
> trunk/reactos/drivers/net/tcpip/tcpip/info.c
> trunk/reactos/hal/halx86/generic/dma.c
> trunk/reactos/lib/epsapi/enum/drivers.c
> trunk/reactos/lib/msafd/misc/dllmain.c
> trunk/reactos/lib/shellext/slayer/slayer.c
> trunk/reactos/lib/ws2_32/misc/ns.c
> trunk/reactos/ntoskrnl/Makefile
> trunk/reactos/regtests/shared/regtests.h
> trunk/reactos/subsys/csrss/win32csr/guiconsole.c
> trunk/reactos/subsys/win32k/makefile
> trunk/reactos/tools/helper.mk
>   
>  
If  some of you would like to turn OPTIMIZED = 1 and help me with some 
of the current problems, I'd appreciate it greatly:

1) Bootvid crashes during boot logo intialization.

Booting with NOGUIBOOT gets ros to the installation page, where

2) Default selections for radio buttons and combobox do not appear, they 
must now be selected manually

Apart from this glitch, driver installation and second boot were 
perfect, minus the following:

3) Debug message about unhandled exception, followed by an explorer 
crash related to a header file.
4) Browsing some registry keys in regedit causes regedit to crash.

I haven't noted any other problems for now, but I haven't done any deep 
testing.

On the up side, this brings my little syscall benchmark to 170ms, an 
additional 15% improvement. In total, this is an 80% improvement over 
int2e builds. Furthermore, all of ReactOS is extremly fast. Directory 
listing in cmd.exe is instant, explorer navigation as well. Menus 
re-draw instantly and the whole experience is smooth. It really feels 
like Windows, you have to try it to believe it.

There's been some talk on activating OPTIMIZED = 1 if (K)DBG != 1 and 
all the developers on IRC were in agreement. But before such a change is 
made, the bugs above need to be fixed, as well as any other possible 
things which might crop up.

Best regards,
Alex Ionescu

[Alex Ionescu]

Alex Ionescu wrote:

>   If  some of you would like to turn OPTIMIZED = 1 and
> help me with some 
>  of the current problems, I'd appreciate it greatly:
>
>  1) Bootvid crashes during boot logo intialization.
>
>  Booting with NOGUIBOOT gets ros to the installation page, where
>
>  2) Default selections for radio buttons and combobox do not appear, 
>  they must now be selected manually 
This seems to be fixed now.

>  
>  Apart from this glitch, driver installation and second boot were 
>  perfect, minus the following:
>
>  3) Debug message about unhandled exception, followed by an explorer 
>  crash related to a header file. 
This is now fixed.

>   4) Browsing some registry keys in regedit causes
> regedit to crash. 
I am now able to get a trace:

(KERNEL32:except/except.c:159) Unhandled exception
(KERNEL32:except/except.c:160) Address:
(KERNEL32:except/except.c:163)    7617a022   
C:\ReactOS\system32\comctl32.dll
CS:EIP 1b:7617a022
DS 23 ES 23 FS 3b GS 23
EAX: 0067a408   EBX: 00000000   ECX: 0066e7d8
EDX: 000001cc   EBP: 0066e808   ESI: 00000000   ESP: 0066e7cc
EDI: 00000001   EFLAGS: 00000202
(KERNEL32:except/except.c:166) Frames:
(KERNEL32:except/except.c:174)    761674ca   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    7617bb55   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    7617f791   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    77e6f944   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)    77e703ba   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)      405a99   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)      401d31   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)    77e6f906   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)    77e70283   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)    761ad1f6   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    761ad3cc   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    761afd0a   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    761b0963   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    761b22ac   
C:\ReactOS\system32\comctl32.dll
(KERNEL32:except/except.c:174)    77e6f971   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)    77e6fb2a   C:\ReactOS\system32\user32.dll
(KERNEL32:except/except.c:174)      406020   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)      40a4fb   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)      40116a   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)      401038   C:\ReactOS\regedit.exe
(KERNEL32:except/except.c:174)    7c8209d0   
C:\ReactOS\system32\kernel32.dll

The bad news is that PSEH is really broken in optimized builds and 
RtlUnwind needs some fixing.

Best regards,
Alex Ionescu

[Alex Ionescu]

Alex Ionescu wrote:

>   4) Browsing some registry keys in regedit causes
> regedit to crash.
>
>
>  
Err, this is because GCC is extremly stupid. It loads a stack structure 
into ebx, and then the callback code becomes jmp(%ebx).
The problem is that it restores protected registers before doing the 
jmp, so ebx is reset to it's value before the function is called...

I guess this needs to be made __volatile, but I'm afraid that this sort 
of bug might happen in other DLLs. Does anyone know if gcc team is aware 
of this?

Best regards,
Alex Ionescu

[Mike Nordell]

Alex Ionescu wrote:

> >   4) Browsing
> > some registry keys in regedit causes regedit to crash.
> >  
>  Err, this is because GCC is extremly stupid. It loads a stack structure
>  into ebx, and then the callback code becomes jmp(%ebx). 
Could you elaborate. I didn't quite follow in what function/context it loads
ebx, and what function it jumps from/to, and where ebx wasn't what it was
assumed to be.

>   The problem is that it restores protected registers
> before doing the
>  jmp, so ebx is reset to it's value before the function is called...
>
>  I guess this needs to be made __volatile, 
It's just "volatile" (no leading underscores since it's a C keyword, not
a
language extension).

/Mike