Arrays and Alignment

Arrays and Alignment

[Michael N. Moran]

I am declaring an array of an aligned type. However
it would appear that when used in an array, the alignment
is not maintained as I would expect. Here's a test-case.

#include <stdio.h>

typedef unsigned char achar __attribute__ ((__aligned__(32)));

achar           a[2];

int main(int argc,char* argv[]){
    printf("__alignof__(achar):%u\n",__alignof__(achar));
    printf("__alignof__(a[1]):%u\n",__alignof__(a[1]));
    printf("&a: %8.8lX\n",(unsigned long)&a);
    printf("sizeof(a):%u\n",sizeof(a));
    printf("sizeof(a[0]):%u\n",sizeof(a[0]));
    printf("&a[0]: %8.8lX\n",(unsigned long)&a[0]);
    printf("&a[1]: %8.8lX\n",(unsigned long)&a[1]);
    return 0;
    }

% gcc -o main main.c ; main

__alignof__(achar):32
__alignof__(a[1]):1
&a: 08049740
sizeof(a):32
sizeof(a[0]):1
&a[0]: 08049740
&a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760

% gcc --version
gcc (GCC) 3.2.2 20030222 (Red Hat Linux 3.2.2-5)

I expected the second element of the array to be aligned
as well as the first, since the alignment of the type is 32.

The Beatles were wrong: 1 & 1 & 1 is 1

Re: Arrays and Alignment

[Andreas Schwab]

"Michael N. Moran" <mnmoran@bellsouth.net> writes:

> I am declaring an array of an aligned type. However
> it would appear that when used in an array, the alignment
> is not maintained as I would expect. Here's a test-case.
>
> #include <stdio.h>
>
> typedef unsigned char achar __attribute__ ((__aligned__(32)));
>
> achar           a[2];
>
> int main(int argc,char* argv[]){
>     printf("__alignof__(achar):%u\n",__alignof__(achar));
>     printf("__alignof__(a[1]):%u\n",__alignof__(a[1]));
>     printf("&a: %8.8lX\n",(unsigned long)&a);
>     printf("sizeof(a):%u\n",sizeof(a));
>     printf("sizeof(a[0]):%u\n",sizeof(a[0]));
>     printf("&a[0]: %8.8lX\n",(unsigned long)&a[0]);
>     printf("&a[1]: %8.8lX\n",(unsigned long)&a[1]);
>     return 0;
>     }
>
> % gcc -o main main.c ; main
>
> __alignof__(achar):32
> __alignof__(a[1]):1
> &a: 08049740
> sizeof(a):32
> sizeof(a[0]):1
> &a[0]: 08049740
> &a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760

You can't have both.  That would require to put padding between array
elements, which is not allowed (unlike structure members).

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE Linux AG, Deutschherrnstr. 15-19, D-90429 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: Arrays and Alignment

[Michael N. Moran]

Andreas Schwab wrote:

>"Michael N. Moran" <mnmoran@bellsouth.net> writes:
>
>  
>
>>I am declaring an array of an aligned type. However
>>it would appear that when used in an array, the alignment
>>is not maintained as I would expect. Here's a test-case.
>>
>>#include <stdio.h>
>>
>>typedef unsigned char achar __attribute__ ((__aligned__(32)));
>>
>>achar           a[2];
>>
>>int main(int argc,char* argv[]){
>>    printf("__alignof__(achar):%u\n",__alignof__(achar));
>>    printf("__alignof__(a[1]):%u\n",__alignof__(a[1]));
>>    printf("&a: %8.8lX\n",(unsigned long)&a);
>>    printf("sizeof(a):%u\n",sizeof(a));
>>    printf("sizeof(a[0]):%u\n",sizeof(a[0]));
>>    printf("&a[0]: %8.8lX\n",(unsigned long)&a[0]);
>>    printf("&a[1]: %8.8lX\n",(unsigned long)&a[1]);
>>    return 0;
>>    }
>>
>>% gcc -o main main.c ; main
>>
>>__alignof__(achar):32
>>__alignof__(a[1]):1
>>&a: 08049740
>>sizeof(a):32
>>sizeof(a[0]):1
>>&a[0]: 08049740
>>&a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760
>>    
>>
>
>You can't have both.  That would require to put padding between array
>elements, which is not allowed (unlike structure members).
>
>Andreas.
>  
>
Andreas,

Thanks for the instantaneous reply :-)

I must confess, however, that I don't understand what you mean
by "both". Do you mean (1) array alignment and (2) array element
alignment? In this case it would seem that the type "achar"
requires an alignment of 32bytes. Why would that not extend
to the elements of the array?

mike

Re: Arrays and Alignment

[Neil Booth]

Michael N. Moran wrote:-

> >>% gcc -o main main.c ; main
> >>
> >>__alignof__(achar):32
> >>__alignof__(a[1]):1
> >>&a: 08049740
> >>sizeof(a):32
> >>sizeof(a[0]):1
> >>&a[0]: 08049740
> >>&a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760
> >>   
> >>
> >
> >You can't have both.  That would require to put padding between array
> >elements, which is not allowed (unlike structure members).
> >
> >Andreas.
> > 
> >
> Andreas,
> 
> Thanks for the instantaneous reply :-)
> 
> I must confess, however, that I don't understand what you mean
> by "both". Do you mean (1) array alignment and (2) array element

He means, I think, that &a[n] == &a[0] + n * sizeof (a[0]) is an
inescapable fact of life.

Neil.

Re: Arrays and Alignment

[Gabriel Dos Reis]

Neil Booth <neil@daikokuya.co.uk> writes:

| Michael N. Moran wrote:-
| 
| > >>% gcc -o main main.c ; main
| > >>
| > >>__alignof__(achar):32
| > >>__alignof__(a[1]):1
| > >>&a: 08049740
| > >>sizeof(a):32
| > >>sizeof(a[0]):1
| > >>&a[0]: 08049740
| > >>&a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760
| > >>   
| > >>
| > >
| > >You can't have both.  That would require to put padding between array
| > >elements, which is not allowed (unlike structure members).
| > >
| > >Andreas.
| > > 
| > >
| > Andreas,
| > 
| > Thanks for the instantaneous reply :-)
| > 
| > I must confess, however, that I don't understand what you mean
| > by "both". Do you mean (1) array alignment and (2) array element
| 
| He means, I think, that &a[n] == &a[0] + n * sizeof (a[0]) is an
| inescapable fact of life.

Certainly, but alignment and sizeof are not unrelated attributes.
Since achar is not standard char, there is no requirement that
sizeof(achar)  be 1.  That should not be.  sizeof(achar) should be
determined by iyts alignment.

The following output
32
32
as I would expect

   #include <iostream>

   struct __attribute__((__aligned__(32))) A {
      char c;
   };

   int main()
   {
      A a[30]
      std::cout << __alignof(a[0]) << std::endl
                << sizeof(a[0]) << std::endl;
   }

There should be no difference with Michael's example.

-- Gaby

Re: Arrays and Alignment

[Neil Booth]

Gabriel Dos Reis wrote:-

> | > Thanks for the instantaneous reply :-)
> | > 
> | > I must confess, however, that I don't understand what you mean
> | > by "both". Do you mean (1) array alignment and (2) array element
> | 
> | He means, I think, that &a[n] == &a[0] + n * sizeof (a[0]) is an
> | inescapable fact of life.
> 
> Certainly, but alignment and sizeof are not unrelated attributes.

That may be true, but I stand by my statement, which I believe you
agree with, and I interpret that as what Andreas meant.

Neil.

Re: Arrays and Alignment

[Gabriel Dos Reis]

Neil Booth <neil@daikokuya.co.uk> writes:

| Gabriel Dos Reis wrote:-
| 
| > | > Thanks for the instantaneous reply :-)
| > | > 
| > | > I must confess, however, that I don't understand what you mean
| > | > by "both". Do you mean (1) array alignment and (2) array element
| > | 
| > | He means, I think, that &a[n] == &a[0] + n * sizeof (a[0]) is an
| > | inescapable fact of life.
| > 
| > Certainly, but alignment and sizeof are not unrelated attributes.
| 
| That may be true, but I stand by my statement, which I believe you
| agree with,

More precisely, I do not agree with the statement that one can't
have both.  But I agree with the above relation.

| and I interpret that as what Andreas meant.


-- Gaby

Re: Arrays and Alignment

[Andreas Schwab]

Neil Booth <neil@daikokuya.co.uk> writes:

> Gabriel Dos Reis wrote:-
>
>> | > Thanks for the instantaneous reply :-)
>> | > 
>> | > I must confess, however, that I don't understand what you mean
>> | > by "both". Do you mean (1) array alignment and (2) array element
>> | 
>> | He means, I think, that &a[n] == &a[0] + n * sizeof (a[0]) is an
>> | inescapable fact of life.
>> 
>> Certainly, but alignment and sizeof are not unrelated attributes.
>
> That may be true, but I stand by my statement, which I believe you
> agree with, and I interpret that as what Andreas meant.

Yes, this is what I meant.  The only option we have is to increase sizeof,
as Gaby wrote.

Andreas.

-- 
Andreas Schwab, SuSE Labs, schwab@suse.de
SuSE Linux AG, Deutschherrnstr. 15-19, D-90429 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: Arrays and Alignment

[Gabriel Dos Reis]

Andreas Schwab <schwab@suse.de> writes:

| "Michael N. Moran" <mnmoran@bellsouth.net> writes:
| 
| > I am declaring an array of an aligned type. However
| > it would appear that when used in an array, the alignment
| > is not maintained as I would expect. Here's a test-case.
| >
| > #include <stdio.h>
| >
| > typedef unsigned char achar __attribute__ ((__aligned__(32)));
| >
| > achar           a[2];
| >
| > int main(int argc,char* argv[]){
| >     printf("__alignof__(achar):%u\n",__alignof__(achar));
| >     printf("__alignof__(a[1]):%u\n",__alignof__(a[1]));
| >     printf("&a: %8.8lX\n",(unsigned long)&a);
| >     printf("sizeof(a):%u\n",sizeof(a));
| >     printf("sizeof(a[0]):%u\n",sizeof(a[0]));
| >     printf("&a[0]: %8.8lX\n",(unsigned long)&a[0]);
| >     printf("&a[1]: %8.8lX\n",(unsigned long)&a[1]);
| >     return 0;
| >     }
| >
| > % gcc -o main main.c ; main
| >
| > __alignof__(achar):32
| > __alignof__(a[1]):1
| > &a: 08049740
| > sizeof(a):32
| > sizeof(a[0]):1
| > &a[0]: 08049740
| > &a[1]: 08049741         <<<<<<<<<<<< I expected this to be 08049760
| 
| You can't have both.  That would require to put padding between array
| elements, which is not allowed (unlike structure members).

then that is a hole in GCC's type system.  A hole that needs fixing.
There seems to be two issues being confused here:
  (1) type alignement; and
  (2) object alignment.

If a type T is declared to have an alignment N, then given  a
pointer p  to a T,  ++p should be a multiple of N.  That is possible
only it you make sizeof(T) a multiple of N.  Which implies some padding.

Note however that an object may be defined to have a user-supplied
alignment without implying that its type has the same alignment.  

I remember Kenner argued for that distinction (a while ago).  I can't
recall what the conclusion of the debate was; but at any rate
Michael's expectations are rasonable, and I think they should be met,
i.e. we have a bug.

-- Gaby

Re: Arrays and Alignment

[Richard Kenner]

| > typedef unsigned char achar __attribute__ ((__aligned__(32)));
| > achar           a[2];

| > __alignof__(a[1]):1

I think this is wrong: it should be 32.

    Note however that an object may be defined to have a user-supplied
    alignment without implying that its type has the same alignment.  

    I remember Kenner argued for that distinction (a while ago).  

Yes, but not in this direction!  An object cannot be *less* aligned than
it's type.  I argued for being able to represent an object that was *more*
aligned than its type.

Re: Arrays and Alignment

[Gabriel Dos Reis]

kenner@vlsi1.ultra.nyu.edu (Richard Kenner) writes:

| | > typedef unsigned char achar __attribute__ ((__aligned__(32)));
| | > achar           a[2];
| 
| | > __alignof__(a[1]):1
| 
| I think this is wrong: it should be 32.

seconded.

|     Note however that an object may be defined to have a user-supplied
|     alignment without implying that its type has the same alignment.  
| 
|     I remember Kenner argued for that distinction (a while ago).  
| 
| Yes, but not in this direction!

I stand corrected.

|  An object cannot be *less* aligned than
| it's type.  I argued for being able to represent an object that was *more*
| aligned than its type.

Thanks for the precision.

-- Gaby