Compiler Issue on PA-RISC when Accessing Unaligned Packed Structures

[Dekun Cao]

Compiler Issue on PA-RISC when Accessing Unaligned Packed Structures

Background: we are attempting to minimize memory usage in areas where 10s of millions of items may exist. Using #pragma pack(1) to pack structures as much as possible and then these structures are being packed (along with other data) into large blocks of malloc'ed memory. Due to the variable nature of the other data, the structures may be aligned on any byte boundary, which should be ok as the pragma implies alignment of 1 as well.

On PA-RISC machines, we have been seeing "Bus error (core dumped)" messages and have narrowed it down to alignment when doubly de-referencing pointers to that memory. i.e. if we align the structures it works, or if we copy the pointer and then de-reference it works too!

Output of "uname -a" is: HP-UX dh-holly B.11.11 U 9000/800 3314646695 unlimited-user license
Compiler version is: HP ANSI C++ B3910B A.03.65

The same source code works OK on HP Itanium machine with
Output of "uname -a" is: HP-UX dh-ida B.11.23 U ia64 2988385317 unlimited-user license
Compiler version is: HP aC++/ANSI C B3910B A.06.05 [Jul 25 2005]

Build and Run Instruction: Build it with aCC and run it.

/********** Source code *************/
#include
#include

typedef struct
{
int i;
} REF_LEVEL_T;

#pragma pack 1
typedef struct
{
REF_LEVEL_T *pRefLevel;
char c; // doesn't core if this is moved above pRefLevel
} REF_MEMBER_T;
#pragma pack 0

typedef struct
{
REF_MEMBER_T *pRefMember;
} MEMBER_T;

int main()
{
REF_LEVEL_T RefLevel;
MEMBER_T Member;

RefLevel.i = 0;

char *p = (char *)malloc( 1000 );

for (int unalignBytes = 0; unalignBytes < 4; unalignBytes++)
{
fprintf( stderr, "trying unalignBytes %d\n", unalignBytes );

REF_MEMBER_T *pRefMember = (REF_MEMBER_T *)(p+unalignBytes);

pRefMember->pRefLevel = &RefLevel;

Member.pRefMember = pRefMember;

REF_MEMBER_T *pTestRefMember = Member.pRefMember;

int x = pTestRefMember->pRefLevel->i;

fprintf( stderr, "passed first test\n" );

int y = (Member.pRefMember)->pRefLevel->i;

fprintf( stderr, "passed second test\n" );
}
}

[Steven E. Protter]

Shalom,

You've followed the guidelines for making code work on Itanium and PA-RISC.

Oracle couldn't do it, they maintain different code bases for Itanium versus PA-RISC

SEP

[Sandman!]

Certainly looks more like a compiler issue. The code compiles fine with the ANSI C compiler but core dumps using the aC++ compiler.

[Sandman!]

Correction to my earlier post. When using aC++ supply the "+u1" switch on the command line for pointers to access non-natively aligned data.

# aCC +u1 your_prog.c

# ./a.out
trying unalignBytes 0
passed first test
passed second test
trying unalignBytes 1
passed first test
passed second test
trying unalignBytes 2
passed first test
passed second test
trying unalignBytes 3
passed first test
passed second test

~cheers

[rick jones]

Is it a safe assumption that the performance of this application isn't terribly important?

[Dekun Cao]

Performance of this application is very important. I know the solution I got will have performance penalty. Any other suggestions?

Thanks.

[Sandman! · Accepted solution]

If the performance of this application is so very important then it makes no sense to go against the grain as the machine architecture is optimized for doing loads and stores on a word boundary. Better to waste memory than consume CPU cycles.

[Dekun Cao]

I think I got all the information I need.

Thank you all very much for the help.

Dekun

[Dekun Cao]

I have got the solution.

Thanks for the help