C66x alignment problem

By declaring the pointer as "struct t_test1 *pointer", you have told the compiler that it is a pointer to an aligned instance of "struct t_test1."  If you want that pointer to point to something with a different alignment, you should use "void *pointer" or "char *pointer".

Hello pf

Thank you for your reply. I changed as following program, but it could not resolve.

struct t_test2 {
 short vs;
 void *pointer;
} str2;

p1 = &((struct t_test1 *)str2.pointer)->c1;
p2 = &((struct t_test1 *)str2.pointer)->c2;

I also found that this problem depends on CGT version.

  NG : v7.4.x
  OK : v7.3.x (e.g. v7.3.8)

My customer told me that there are a lot of potentials by this problem in customer program.
They are implementing for porting from existing CPU to C66x. They want to use v7.4.x.
Please advise me.

Regards

This appears to be a bug in the compiler.  In the case you indicate, it is using LDW to read the "pointer" structure member when LDNW should be used instead.  I filed SDSCM00047774 in the SDOWP system to have this addressed.  Please feel free to follow it with the SDOWP link below in my structure.

KAZUHITO KONDO said:

  OK : v7.3.x (e.g. v7.3.8)

That's because #pragma pack is not supported in version 7.3.x.  That feature is introduced in version 7.4.0.

Thanks and regards,

-George

Hello pf,

Please see an sample code below.

This example can be built by both of C6000 CGT 7.3.x and 7.4.x with GNU extension.
This example runs correctly when it is built by 7.3.8-7.3.12,
but does not run correctly when it is built by 7.3.7 or earlier, or 7.4.x.
I guess this issue has been fixed on 7.3.8.


#include <stdio.h>

struct __attribute__((__packed__)) t_test2 {
    short vs;    // Padding
    struct t_test1 *pointer;
} str2;

struct __attribute__((__packed__)) t_test1 {
    char c1;
    char c2;
} str1;

void main(void) {
    char *p1;
    char *p2;

    str2.pointer = &str1;

    p1 = &str2.pointer->c1;
    printf ("p1=0x%08x\n", p1);

    p2 = &str2.pointer->c2;
    printf ("p2=0x%08x\n", p2);
}


Best Regards,
Nobuhiro Masaoka

I think I misunderstood you at the very beginning.

The problem is not that you want the pointer in the field "pointer" to point at an unaligned instance of t_test1, it's that the field "pointer" itself is not aligned to a 32-bit boundary.  Correct?

If so, could you please check what happens when t_test1 has more than two fields?  Is the problem only with the first field?  Only with the second and third (and subsequent)?

Thanks.

Hello pf,

Yes, the problem happens when CPU reads the field "pointer" that is not aligned to a 32-bit boundary.
And this happens when the "pointer" pointed only first element of the instance "t_test1".
This does not happen on second field and subsequents.

Best Regards,
Nobuhiro Masaoka

Hello pf,

This problem looks like fixed at 7.3.8 in 7.3 series, as I mentioned above.
Is it correct?

But I couldn't find its description in the release note...

I have to use 7.4 series CGT because I need "#pragma pack" function.

Best Regards,
Nobuhiro Masaoka