[Bug c/18287] New: Unaligned access to fields inside packed records

[Bug c/18287] New: Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

The testcase I'm about to attach generates a bug error at runtime on the SPARC
(and very likely on any other STRICT_ALIGNMENT target).  Verified with 2.95.3,
3.3.x, 3.4.x and 4.0pre.

I think the front-end wrongly turns ADDR_EXPR<COMPONENT_REF<INDIRECT_REF>> into
PLUS_EXPR whatever the alignment of the field.  For reference, the Ada front-end
doesn't do the simplification and lets the middle-end generate the temporary if
required.

-- 
           Summary: Unaligned access to fields inside packed records
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ebotcazou at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: *-*-*
  GCC host triplet: *-*-*
GCC target triplet: *-*-*


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-11-03 17:53 -------
Created an attachment (id=7467)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=7467&action=view)
Testcase.

To be compiled with -O0.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-03 18:02 -------
Are you sure that 4.0.0 does not work, as it no longer does the transformation for &(a->b) into a + 
offsetof(a,b).
>From the last tree dump for 4.0.0:
;; Function Assign_Hour_Of (Assign_Hour_Of)

Assign_Hour_Of (t)
{
<bb 0>:
  t->Hour = 44;
  return;

}



;; Function Clobber_Hour_Of (Clobber_Hour_Of)

Clobber_Hour_Of (dt)
{
  struct Time_T * D.1128;

<bb 0>:
  D.1128 = &dt->Time;
  Assign_Hour_Of (D.1128);
  return;

}



;; Function main (main)

main ()
{
  struct Date_And_Time_T dt;
  int D.1135;

<bb 0>:
  dt.Time.Hour = 22;
  Clobber_Hour_Of (&dt);
  D.1135 = 0;
  return D.1135;

}


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-03 18:03 -------
  t->Hour = 44;  /* unaligned word access */
is unaligned word access if t is not word aligned in the first place, correct which is where the problem 
comes in right?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2004-11-03 18:17 -------
>   t->Hour = 44;  /* unaligned word access */
> is unaligned word access if t is not word aligned in the first place, correct
> which is where the problem comes in right?

I confirm that the bus error (not bug error!) happens with 4.0.0 20041103.  But,
you're right, probably another problem there.

The compiler should not pass the adress of &dt->Time to Assign_Hour_Of if it
can't prove that the field is at least as aligned as its type.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-03 19:57 -------
There is another example of this in PR 17949 where the tree-optimizations cause something werid to 
happen only because the middle-end does not know that they are unaligned loads.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[wintermute101 at wp dot pl]

------- Additional Comments From wintermute101 at wp dot pl  2004-11-04 16:54 -------
I write here cause it's simmilar situation as reported here but I have no data
about other gcc versions than 3.3.2.

I have following:

/////////////////

define _aligned(n) __attribute__((aligned(n), packed))

struct CzazPort {
	unsigned short number;
	unsigned long  speed;
	unsigned long  protocol;
	unsigned long  mode;
	unsigned short timeout;
	unsigned short loglevel;
	unsigned short timerid;
} _aligned(2);

struct CzazPort tmp;
sscanf( data, "%lu", &tmp.speed );

/////////////////////

data="9600"

before sscanf (content of tmp)
00:01:00:00:96:00:52:54:55:00:38:4e:31:00:03:e8:00:04:2b:67
after sscanf
00:00:25:80:96:00:52:54:55:00:38:4e:31:00:03:e8:00:04:2b:67

9600 in hex -> 0x2580

it was compiled for MOXA UC7420

the same program compiled for i686 works fine

this MOXA has XScale Intel prcessor working in big-endian. I guess this is the
problem (big-endian).

sscanf is just example from source. The same occurs when I relpace sscanf with 
        (*(&tmp.speed)) = 9600; 
but
tmp.speed=9600;
works fine.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2005-01-02 10:57 -------
Reconfirmed with gcc version 4.0.0 20050101 (experimental) at -O0 on SPARC 32-bit.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2005-01-02 10:57:40
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2005-01-02 11:03 -------
Clobber_Hour_Of:
	save	%sp, -112, %sp
	st	%i0, [%fp+68]
	ld	[%fp+68], %g1
	add	%g1, 5, %g1
	mov	%g1, %o0
	call	Assign_Hour_Of, 0
	 nop

is wrong: you can't merely add 5 to the address on STRICT_ALIGNMENT platforms.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2005-01-02 11:11 -------
.t03.generic is already wrong:

Clobber_Hour_Of (dt)
{
  struct Time_T * D.1122;

  D.1122 = &dt->Time;
  Assign_Hour_Of (D.1122);
}

struct Time_T is 32-bit aligned so D.1122 must be a multiple of 4; as such, it
cannot be assigned &dt->Time which is not a multiple of 4.  A properly aligned
temporary should be created instead and its address assigned to D.1122.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2005-01-02 16:45 -------
Are we really forced to support ADDR_EXPR on unaligned fields? How would you 
fix it with a temporary? Are you going to generate a FINALLY_EXPR to copy the 
contents of the temporary back into the original field?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2005-01-02 17:20 -------
> Are we really forced to support ADDR_EXPR on unaligned fields? How would you 
> fix it with a temporary? Are you going to generate a FINALLY_EXPR to copy the 
> contents of the temporary back into the original field?

Yes and yes.  The Ada compiler has been doing this for ages.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[rth at gcc dot gnu dot org]

------- Additional Comments From rth at gcc dot gnu dot org  2005-01-03 23:16 -------
The Ada compiler may do it, but the C compiler does not.  This test case,
written in C, is invalid.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[ebotcazou at gcc dot gnu dot org]

------- Additional Comments From ebotcazou at gcc dot gnu dot org  2005-01-04 05:17 -------
> The Ada compiler may do it, but the C compiler does not.  This test case,
> written in C, is invalid.

Could you tell on what grounds?  AFAICS all fields are still addressable.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rth at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287

[Bug c/18287] Unaligned access to fields inside packed records

[rth at gcc dot gnu dot org]

------- Additional Comments From rth at gcc dot gnu dot org  2005-01-04 09:14 -------
Subject: Re:  Unaligned access to fields inside packed records

> Could you tell on what grounds?  AFAICS all fields are still addressable.

Not really, they aren't.  I've argued in the past that we should either
make them non-addressable, or (more usefully) implement pointers to
unaligned data.  But having done neither, the result of the addressof
operator is undefined.


r~


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18287