[Bug c++/16381] New: Structure layout bug

[Bug c++/16381] New: Structure layout bug

[zhangjie at magima dot com dot cn]

GCC incorrectly calculated the structure layout in the following case.

#include <stdio.h>

typedef int __m128 __attribute__ ((__mode__(__V4SF__)));

struct Quad
{
	Quad() { }
	float x __attribute__((aligned(16)));
	float y;
	float z;
	float w;
};

struct Base1
{
	Quad q;
	int i;
};

struct Derived1 : public Base1
{
	int i2;
};

struct Base2
{
	__m128 m;
	int i;
};

struct Derived2 : public Base2
{
	int i2;
};

#define offset(C,M) (int((&((C*)(16))->M))-16)

int main()
{
	printf("__m128 size %i, align %i\n", sizeof(__m128), __alignof__(__m128));
	printf("Quad   size %i, align %i\n", sizeof(Quad), __alignof__(Quad));
	printf("Base1 %i Derived1 %i Offset %i\n", sizeof(Base1), sizeof(Derived1),
offset(Derived1, i2) );
	printf("Base2 %i Derived2 %i Offset %i\n", sizeof(Base2), sizeof(Derived2),
offset(Derived2, i2) );
	return 0;
}

The results on cygwin:

g++ (GCC) 3.3.1 (cygming special)

__m128 size 16, align 16
Quad   size 16, align 16
Base1 24 Derived1 24 Offset 20
Base2 24 Derived2 24 Offset 20


g++ (GCC) 3.5.0 20040613 (experimental)

__m128 size 16, align 16
Quad   size 16, align 16
Base1 32 Derived1 32 Offset 20
Base2 24 Derived2 32 Offset 24

Neither looks correct. 

This test case was given in http://gcc.gnu.org/ml/gcc/2004-06/msg01490.html.
Discussion of this bug began from this mail
http://gcc.gnu.org/ml/gcc/2004-06/msg00622.html.

-- 
           Summary: Structure layout bug
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: zhangjie at magima dot com dot cn
                CC: gcc-bugs at gcc dot gnu dot org,nathan at codesourcery
                    dot com,stephen dot kennedy at havok dot com
  GCC host triplet: i686-pc-cygwin


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] Structure layout bug

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-07-06 15:42 -------
Related to 10179.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/3.4/3.5 Regression] Structure layout bug

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-08-16 10:47 -------
3.4.0 gives this:

__m128 size 16, align 16
Quad   size 16, align 16
Base1 32 Derived1 32 Offset 20
Base2 24 Derived2 24 Offset 20


3.5.0 20040626 gives this:

__m128 size 16, align 16
Quad   size 16, align 16
Base1 32 Derived1 32 Offset 20
Base2 32 Derived2 48 Offset 32



Jie's expected result (from a mail of him) is:

__m128 size 16, align 16
Quad   size 16, align 4
Base1 20 Derived1 24 Offset 20
Base2 32 Derived2 48 Offset 32


Given the importance of this, and the inconsistencies between 3.3/3.4/3.5, I'll 
confirm this bug as a regression. ABI gurus can revisit this, and check which 
version is correct (if any), and what needs fixing.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |giovannibajo at libero dot
                   |                            |it
           Severity|normal                      |critical
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |ABI, wrong-code
      Known to fail|                            |3.3.4 3.4.0 3.5.0
   Last reconfirmed|0000-00-00 00:00:00         |2004-08-16 10:47:41
               date|                            |
            Summary|Structure layout bug        |[3.3/3.4/3.5 Regression]
                   |                            |Structure layout bug
   Target Milestone|---                         |3.4.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/3.5 Regression] Structure layout bug

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-08-17 06:39 -------
It's my opinion that GCC 3.4.x is correct, and 3.5 is wrong.

Jie's suggestion that Quad have alignment 4 does not make sense to me, given the
"__attribute__ ((aligned (16)))" on its member "x". 

The distinction between GCC 3.4 and GCC 3.5 is that GCC 3.4 is treating Base2 as
a POD, where as GCC 3.5 is not.  I see no reason why a struct containing
"__m128" should disqualify a type from being a POD, although clearly that is not
specified by the standard.

Therefore, I've retargeted this at GCC 3.5.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|3.3.4 3.4.0 3.5.0           |3.3.4 3.5.0
            Summary|[3.3/3.4/3.5 Regression]    |[3.3/3.5 Regression]
                   |Structure layout bug        |Structure layout bug
   Target Milestone|3.4.2                       |3.5.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/3.5 Regression] Structure layout bug

[zhangjie at magima dot com dot cn]

------- Additional Comments From zhangjie at magima dot com dot cn  2004-08-17 07:30 -------
It seems that Giovanni refered this email from me
<http://gcc.gnu.org/ml/gcc/2004-06/msg00667.html>. But I made a mistake in it as
indicated by this email <http://gcc.gnu.org/ml/gcc/2004-06/msg00669.html>. And
the results in these emails came from a slightly different case
<http://gcc.gnu.org/ml/gcc/2004-06/msg00622.html>, in which Quad had no
constructor. Sorry, I should have written more clearly when I reported this bug.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/3.5 Regression] Structure layout bug

[zhangjie at magima dot com dot cn]

------- Additional Comments From zhangjie at magima dot com dot cn  2004-08-17 09:53 -------
hi Giovanni, did you run the test on Linux? The result with gcc-3.5-20040815 on
cygwin is as same as the bug was reported. 

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/3.5 Regression] Structure layout bug

[zhangjie at magima dot com dot cn]

------- Additional Comments From zhangjie at magima dot com dot cn  2004-08-17 10:04 -------
(In reply to comment #3)
> 
> The distinction between GCC 3.4 and GCC 3.5 is that GCC 3.4 is treating Base2 as
> a POD, where as GCC 3.5 is not.  I see no reason why a struct containing
> "__m128" should disqualify a type from being a POD, although clearly that is not
> specified by the standard.
> 
I think "__m128" would better be treated as plain data type like int and long.
So I prefer the result of "3.5.0 20040626" in comment #2.

> Therefore, I've retargeted this at GCC 3.5.

Although I have not tested, I guess GCC 3.4.1 on cygwin will also give the wrong
result, whether "__m128" is treated as POD or not. So if we want to fix it on
cygwin, we'd better retarget it at GCC 3.4.2.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/4.0 Regression] Structure layout bug

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-10-28 06:00 -------
On i686-pc-linux-gnu, with both GCC 3.4.2 and GCC 4.0, I see :

__m128 size 16, align 16
Quad   size 16, align 16
Base1 32 Derived1 32 Offset 20
Base2 32 Derived2 48 Offset 32

I am building a Cygwin cross compiler to check there.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
      Known to fail|3.3.4 4.0                   |3.3.4 4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/4.0 Regression] Structure layout bug

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-10-28 06:18 -------
On Cygwin, I still get "24 32 24" for Base 2.

The reason turns out to have nothing to do with PODs.  Rather, it is due to the
fact that on Cygwin, BIGGEST_FIELD_ALIGNMENT is set to 8 bytes.  So, even though
the __m128 type has 16-byte alignment, Base 2 has only 8 byte alignment, which
results in a size of 24 bytes.

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381

[Bug c++/16381] [3.3/4.0 Regression] Structure layout bug

[dannysmith at users dot sourceforge dot net]

------- Additional Comments From dannysmith at users dot sourceforge dot net  2004-10-28 07:14 -------
Isn't this a bug in cygming.h then?  Would something like this
still retain  MSVC compatibility while allowing logical alignment
for SSE[2] types?

/* This is only needed in libobjc. There we use 8-byte alignment to be
   consistent with  TARGET_ALIGN_DOUBLE default. Otherwise the
   i386 version of ADJUST_FIELD_ALIGN  will do the right thing (i.e
   not force 4-byte alignment on doubles and long longs) since 
   TARGET_ALIGN_DOUBLE is set. */
#ifdef IN_TARGET_LIBS
#undef	BIGGEST_FIELD_ALIGNMENT
#define BIGGEST_FIELD_ALIGNMENT 64
#endif

Danny


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16381