PR 17949 (unaligned memory accesses due to ivopts) question

PR 17949 (unaligned memory accesses due to ivopts) question

[Zdenek Dvorak]

Hello,

strength reduction in ivopts may turn references to "packed" fields
in structures to accesses via pointers.  The adress obtained this
way may be unaligned, which causes crash when STRICT_ALIGNMENT = 1.

This problem however does not seem to be limited to ivopts.  For example
the testcase below probably (I do not have access to the hardware
just now, so I have checked only assembler, and I cannot read ia64 asm
very well, so I am not entirely sure) fails in the same way.

It seems to me that fixing this in ivopts is not the right place.
Maybe setting DECL_NONADDRESSABLE_P for unaligned fields of structures
when STRICT_ALIGNMENT = 1 would be more correct?

Zdenek

typedef struct
{
  short i __attribute__ ((packed));
  int f[2] __attribute__ ((packed));
} A;

A a;

void foo(int *);

int main(void)
{
  foo (a.f);
  return 0;
}

void foo(int *x)
{
  *x = 0;
}

Re: PR 17949 (unaligned memory accesses due to ivopts) question

[Daniel Jacobowitz]

On Sat, Oct 16, 2004 at 09:35:34PM +0200, Zdenek Dvorak wrote:
> This problem however does not seem to be limited to ivopts.  For example
> the testcase below probably (I do not have access to the hardware
> just now, so I have checked only assembler, and I cannot read ia64 asm
> very well, so I am not entirely sure) fails in the same way.
> 
> It seems to me that fixing this in ivopts is not the right place.
> Maybe setting DECL_NONADDRESSABLE_P for unaligned fields of structures
> when STRICT_ALIGNMENT = 1 would be more correct?

> typedef struct
> {
>   short i __attribute__ ((packed));
>   int f[2] __attribute__ ((packed));
> } A;
> 
> A a;
> 
> void foo(int *);
> 
> int main(void)
> {
>   foo (a.f);
>   return 0;
> }
> 
> void foo(int *x)
> {
>   *x = 0;
> }

Well, a.f ought to be a pointer to unaligned integer.  Don't you get
some kind of warning for this?

-- 
Daniel Jacobowitz

Re: PR 17949 (unaligned memory accesses due to ivopts) question

[Zdenek Dvorak]

Hello,

> Well, a.f ought to be a pointer to unaligned integer.  Don't you get
> some kind of warning for this?

no.  There is -Wcast-align (disabled by default), but it does not
trigger in this case.  -Wpacked (also disabled by default) announces
that "packed attribute causes inefficient alignment for 'i'"
(but does not mention f; umm...).  And that's it, these are all warnings
that are triggered by STRICT_ALIGNMENT.

Zdenek