Experimental Physics and
Industrial Control System

Till Straumann <[email protected]> · Sat, 20 Apr 2002 09:24:58 -0700

Jeff Hill wrote:
Till,
After closer inspection, I see that there were two entries in the CVS
log related to this issue. Attached is the one that applies to
preserving the alignment returned by ::new().
Yeah -that's what I did also. I just was surprised that I could not
find any report of this problem in the tech-talk archive and yet,
the distribution Harvey downloaded not too long ago does not
contain the fix yet...
The change reported in my
previous message will have no impact because it is only rearranging
fields in a union.
Right - this one looks much better :-),
thanks.
-- Till.
Sorry about the confusion. I added "diff -c" to my
.cvsrc file and I now receive context diffs from CVS.
Revision : 1.22
Date : 2002/1/15 16:46:11
Author : 'jhill'
State : 'Exp'
Lines : +2 -2
Description :
fixed to preserve 8 byte alignment
cvs diff -r 1.22 -r 1.21 tsFreeList.h (in directory

D:\users\hill\epicsDvl\epics\base\src\libCom\cxxTemplates\)

Index: tsFreeList.h

===================================================================

RCS file:

/net/phoebus/epicsmgr/cvsroot/epics/base/src/libCom/cxxTemplates/tsFreeL

ist.h,v

retrieving revision 1.22

retrieving revision 1.21

diff -c -r1.22 -r1.21

*** tsFreeList.h 2002/01/15 16:46:11 1.22

--- tsFreeList.h 2001/12/14 00:36:25 1.21

***************

*** 1,5 ****

! /* $Id: tsFreeList.h,v 1.22 2002/01/15 16:46:11 jhill Exp $

*

* L O S A L A M O S

* Los Alamos National Laboratory

--- 1,5 ----

! /* $Id: tsFreeList.h,v 1.21 2001/12/14 00:36:25 jhill Exp $

*

* L O S A L A M O S

* Los Alamos National Laboratory

***************

*** 78,85 ****

template < class T, unsigned N = 0x400, unsigned DEBUG_LEVEL = 0u >

struct tsFreeListChunk {

- tsFreeListItem < T, DEBUG_LEVEL > items [N];

tsFreeListChunk < T, N, DEBUG_LEVEL > *pNext;

};

template < class T, unsigned N = 0x400, unsigned DEBUG_LEVEL = 0u >

--- 78,85 ----

template < class T, unsigned N = 0x400, unsigned DEBUG_LEVEL = 0u >

struct tsFreeListChunk {

tsFreeListChunk < T, N, DEBUG_LEVEL > *pNext;

+ tsFreeListItem < T, DEBUG_LEVEL > items [N];

};

template < class T, unsigned N = 0x400, unsigned DEBUG_LEVEL = 0u >
*****CVS exited normally with code 1*****

Jeff
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Till Straumann
Sent: Tuesday, April 09, 2002 6:35 PM
To: [email protected]
Cc: [email protected]
Subject: Re: ca_test bus error on solaris-sparc-gnu + fix
Jeff Hill wrote:
Till,
I think that we already eradicated that one along the way to the
next
R3.14 release. See attached. Nevertheless, thanks for the heads up.
Jeff
Jeff.
From the attached diff it is not obvious to me what your fix is

- would you mind giving me a hint, please?
I am not sure if 8byte alignment is enough
(e.g. optimized code on PPC/altivec may need 16 bytes.)
IMHO, it would be best to preserve the alignment
returned by malloc()
-- Till

----------------------------
Revision : 1.23
Date : 2002/1/15 20:2:45
Author : 'jhill'
State : 'Exp'
Lines : +2 -2
Description :
fixed to preserve 8 byte alignment
cvs diff -r 1.23 -r 1.22 tsFreeList.h (in directory
D:\users\hill\epicsDvl\epics\base\src\libCom\cxxTemplates\)
Index: tsFreeList.h
===================================================================
RCS file:
/net/phoebus/epicsmgr/cvsroot/epics/base/src/libCom/cxxTemplates/tsFreeL
ist.h,v
retrieving revision 1.23
retrieving revision 1.22
diff -r1.23 -r1.22
2c2
< /*  $Id: tsFreeList.h,v 1.23 2002/01/15 20:02:45 jhill Exp $
---
/* $Id: tsFreeList.h,v 1.22 2002/01/15 16:46:11 jhill Exp $
75d74
<     char pad[ sizeof ( T ) ];
76a76
 char pad[ sizeof ( T ) ];
*****CVS exited normally with code 1*****
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Till
Straumann
Sent: Thursday, April 04, 2002 9:04 PM
To: [email protected]
Cc: [email protected]
Subject: ca_test bus error on solaris-sparc-gnu + fix
Hi all.
We got a SIGBUS (due to an alignment violation)
running ca_test (and other apps)
EPICS:      3.14.0.beta1
system:    solaris-sparc-gnu
compiler: egcs-2.91.66
I found that
void *
tsFreeList<T,N,DEBUG_LEVEL>::allocate(size_t)
may return improperly aligned memory:
Because 'items' are declared
union {
  tsFreeListItem<T,...> *pNext;
  char                           pad[sizeof(T)];
}
the compiler only makes sure 'items' fulfill
the alignment requirements of pNext.
However, 'items' may need stricter alignment
(which goes unnoticed because the 'item' pointer
is later cast to (void*)): e.g. it may contain a double
which on sparc needs to be 8byte aligned...
The proposed fix is to put the 'items' into a
chunk first, so they get the chunk's alignment.
Unfortunately, this fix breaks binary compatibility
- everything needs to be rebuilt.
Regards,
-- Till

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System