[Bug c/15941] New: [regression 3.4.1] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[Bug c/15941] New: [regression 3.4.1] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[danglin at gcc dot gnu dot org]

The following new fails occured
Executing on host: /xxx/gnu/gcc-3.4/objdir/gcc/xgcc -B/xxx/gnu/gcc-3.4/objdir/gc
c/ c_compat_main_tst.o c_compat_x_tst.o c_compat_y_tst.o     -lm   -o gcc-dg-com
pat-struct-by-value-11-01    (timeout = 300)
PASS: gcc.dg/compat/struct-by-value-11 c_compat_x_tst.o-c_compat_y_tst.o link
Setting LD_LIBRARY_PATH to :/usr/openwin/lib
FAIL: gcc.dg/compat/struct-by-value-11 c_compat_x_tst.o-c_compat_y_tst.o execute

Executing on host: /xxx/gnu/gcc-3.4/objdir/gcc/xgcc -B/xxx/gnu/gcc-3.4/objdir/gc
c/ c_compat_main_tst.o c_compat_x_tst.o c_compat_y_tst.o     -lm   -o gcc-dg-com
pat-struct-by-value-1-01    (timeout = 300)
PASS: gcc.dg/compat/struct-by-value-1 c_compat_x_tst.o-c_compat_y_tst.o link
Setting LD_LIBRARY_PATH to :/usr/openwin/lib
FAIL: gcc.dg/compat/struct-by-value-1 c_compat_x_tst.o-c_compat_y_tst.o execute

This is for

-bash-2.05b$ ../xgcc -B../ -v
Reading specs from ../specs
Configured with: ../gcc/configure --with-gnu-as --with-as=/opt/gnu/bin/as
--enable-shared --with-local-prefix=/opt/gnu --prefix=/opt/gnu/gcc/gcc-3.4.1
--enable-debug=no --disable-nls
Thread model: single
gcc version 3.4.1 20040609 (prerelease)

-- 
           Summary: [regression 3.4.1] new fails gcc.dg/compat/struct-by-
                    value-11 and gcc.dg/compat/struct-by-value-1
           Product: gcc
           Version: 3.4.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: danglin at gcc dot gnu dot org
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: hppa2.0-hp-hpux11.00
  GCC host triplet: hppa2.0-hp-hpux11.00
GCC target triplet: hppa2.0-hp-hpux11.00


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |critical
          Component|c                           |target
           Keywords|                            |ABI, wrong-code
            Summary|[regression 3.4.1] new fails|[3.4 regression] new fails
                   |gcc.dg/compat/struct-by-    |gcc.dg/compat/struct-by-
                   |value-11 and                |value-11 and
                   |gcc.dg/compat/struct-by-    |gcc.dg/compat/struct-by-
                   |value-1                     |value-1
   Target Milestone|---                         |3.4.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2004-06-11 16:45 -------
Mark, an ABI bug on a primary platform should really be targeted at 3.4.1. Do 
you agree?

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-06-11 16:59 -------
Yes, I agree.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|3.4.2                       |3.4.1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[dave at hiauly1 dot hia dot nrc dot ca]

------- Additional Comments From dave at hiauly1 dot hia dot nrc dot ca  2004-06-11 20:52 -------
Subject: Re:  New: [regression 3.4.1] new fails gcc.dg/compat/str

>  GCC build triplet: hppa2.0-hp-hpux11.00
>   GCC host triplet: hppa2.0-hp-hpux11.00
> GCC target triplet: hppa2.0-hp-hpux11.00

I'm also seeing this regression on the main on hppa2.0-hp-hpux11.11.

It looks like the only patch that could have introduced this regression is:

2004-06-09  Richard Henderson  <rth@redhat.com>
 
	PR middle-end/15228
	* function.c (assign_parms): Always set_mem_align with the computed
	FUNCTION_ARG_BOUNDARY.  Don't clear stack_parm if !STRICT_ALIGNMENT.

Dave


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[3.4 regression] new fails  |[3.4/3.5 regression] new
                   |gcc.dg/compat/struct-by-    |fails gcc.dg/compat/struct-
                   |value-11 and                |by-value-11 and
                   |gcc.dg/compat/struct-by-    |gcc.dg/compat/struct-by-
                   |value-1                     |value-1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[dave at hiauly1 dot hia dot nrc dot ca]

------- Additional Comments From dave at hiauly1 dot hia dot nrc dot ca  2004-06-11 21:26 -------
Subject: Re:  [3.4/3.5 regression] new fails gcc.dg/compat/s

Looking at the failure on the trunk, I see that we have two
unaligned accesses in test5 in gcc-dg-compat-struct-by-value-1-01.

0x0000bcb0 <test5+296>: ldw -35(,r3),r21
...
0x0000bcb0 <test5+296>: ldw -35(,r3),r21

Same in gcc-dg-compat-struct-by-value-11-01.

0x0001f3a4 <testScc3+324>:      ldw -36(,r3),r21

The offsets are in hex.

Dave


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[danglin at gcc dot gnu dot org]

------- Additional Comments From danglin at gcc dot gnu dot org  2004-06-12 16:13 -------
Reverting Richard's patch to function.c eliminates the regression.

The regression also affects hppa-unknown-linux if I disable the kernel
unaligned handler.

The 32-bit PA ABI has rather special requirements for passing 5-7 byte
structs.  See the comment in function_arg in pa.c.  I believe the fail
in test5 is caused by trying to pass a 5 byte struct.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rth at redhat dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[rth at redhat dot com]

------- Additional Comments From rth at redhat dot com  2004-06-12 19:01 -------
Subject: Re:  [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

On Sat, Jun 12, 2004 at 04:13:30PM -0000, danglin at gcc dot gnu dot org wrote:
> The regression also affects hppa-unknown-linux if I disable the kernel
> unaligned handler.
> 
> The 32-bit PA ABI has rather special requirements for passing 5-7 byte
> structs.  See the comment in function_arg in pa.c.  I believe the fail
> in test5 is caused by trying to pass a 5 byte struct.

So is the PA FUNCTION_ARG_BOUNDARY function lying about the alignment?


r~


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[dave at hiauly1 dot hia dot nrc dot ca]

------- Additional Comments From dave at hiauly1 dot hia dot nrc dot ca  2004-06-12 20:43 -------
Subject: Re:  [3.4/3.5 regression] new fails gcc.dg/compat/s

> > The 32-bit PA ABI has rather special requirements for passing 5-7 byte
> > structs.  See the comment in function_arg in pa.c.  I believe the fail
> > in test5 is caused by trying to pass a 5 byte struct.
> 
> So is the PA FUNCTION_ARG_BOUNDARY function lying about the alignment?

Well, it's not defined so how could we be lying ...

The default is clearly wrong.  We seem to have been depending on
MAX_PARM_BOUNDARY which appears to be a relic.

Let me see if defining FUNCTION_ARG_BOUNDARY helps.

Dave


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[dave at hiauly1 dot hia dot nrc dot ca]

------- Additional Comments From dave at hiauly1 dot hia dot nrc dot ca  2004-06-12 21:37 -------
Subject: Re:  [3.4/3.5 regression] new fails gcc.dg/comp

> > > The 32-bit PA ABI has rather special requirements for passing 5-7 byte
> > > structs.  See the comment in function_arg in pa.c.  I believe the fail
> > > in test5 is caused by trying to pass a 5 byte struct.
> > 
> > So is the PA FUNCTION_ARG_BOUNDARY function lying about the alignment?
> 
> Well, it's not defined so how could we be lying ...

Sorry, the last mail was wrong.  FUNCTION_ARG_BOUNDARY does the
right thing as far as I can tell.  It returns either PARM_BOUNDARY
or MAX_PARM_BOUNDARY (32 or 64 on the 32-bit port).

We need 64 bit alignment for a 5 byte struct on the 32-bit port.  It
has to be padded downward.  Achieving the required padding for structs
larger than a word was difficult and was an outstanding ABI bug for
years.

Dave


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[danglin at gcc dot gnu dot org]

------- Additional Comments From danglin at gcc dot gnu dot org  2004-06-18 02:48 -------
Created an attachment (id=6552)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=6552&action=view)
Simplified test case.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[danglin at gcc dot gnu dot org]

------- Additional Comments From danglin at gcc dot gnu dot org  2004-06-18 04:05 -------
The file yyy.c demonstrates the problem introduced by Richard's change.
The first hunk now always sets the memory alignment for stack_parm to
that specified by FUNCTION_ARG_BOUNDARY.  For a 5-byte struct, this is
64 bits.  Previously, the maximum memory alignment for stack_parm was
PARM_BOUNDARY.  PARM_BOUNDARY is 32 bits on this target.

This is the code that 3.4.1 generates to load the 5-byte struct passed
on the stack for the call to check5:

        ldw -117(%r30),%r21
        ldb -113(%r30),%r19
        zdep %r21,23,24,%r20
        extru %r21,0+8-1,8,%r25
        .CALL ARGW0=GR,ARGW1=GR,ARGW2=GR
        bl check5,%r2
        or %r19,%r20,%r26

Except for the fact that the ldw insn is not aligned correctly, the code
is correct.  On hpux, the unaligned load causes a segmentation fault.

This is the code that 3.3 generates:

        ldb -117(%r30),%r25
        ldb -116(%r30),%r21
        ldb -115(%r30),%r20
        zdep %r21,7,8,%r21
        zdep %r20,15,16,%r20
        ldb -114(%r30),%r19
        ldb -113(%r30),%r22
        or %r20,%r21,%r20
        zdep %r19,23,24,%r19
        or %r19,%r20,%r19
        .CALL ARGW0=GR,ARGW1=GR,ARGW2=GR
        bl check5,%r2
        or %r22,%r19,%r26

It's significantly less efficient but has no unaligned accesses.  For
comparison, this is what the HP compiler generates to load the arguments:

        LDB     -117(%r30),%r25 ;offset 0x18
        LDW     -116(%r30),%r26 ;offset 0x1c

I believe what has happened is that Richard's change has enabled the
the access optimization in emit_group_load, and we are not correctly
setting the alignment of the source data when it is in a stack slot.
This is a result of using downward padding in stack slots on this target.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[rth at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rth at gcc dot gnu dot org
                   |dot org                     |
             Status|UNCONFIRMED                 |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[mmitchel at gcc dot gnu dot org]

------- Additional Comments From mmitchel at gcc dot gnu dot org  2004-06-19 17:49 -------
Dave, if your final test results come out OK, would you please commit this patch
to mainline and 3.4.1?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-06-19 19:25 -------
Subject: Bug 15941

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	danglin@gcc.gnu.org	2004-06-19 19:24:56

Modified files:
	gcc            : ChangeLog function.c 

Log message:
	PR target/15941
	* function.c (assign_parms): If not padding upward or intentionally
	forcing upward padding, take offset_rtx into account when determining
	the alignment for stack_parm.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.4044&r2=2.4045
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/function.c.diff?cvsroot=gcc&r1=1.526&r2=1.527



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-06-19 19:49 -------
Subject: Bug 15941

CVSROOT:	/cvs/gcc
Module name:	gcc
Branch: 	gcc-3_4-branch
Changes by:	danglin@gcc.gnu.org	2004-06-19 19:49:46

Modified files:
	gcc            : ChangeLog function.c 

Log message:
	PR target/15941
	* function.c (assign_parms): If not padding upward or intentionally
	forcing upward padding, take offset_rtx into account when determining
	the alignment for stack_parm.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&only_with_tag=gcc-3_4-branch&r1=2.2326.2.514&r2=2.2326.2.515
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/function.c.diff?cvsroot=gcc&only_with_tag=gcc-3_4-branch&r1=1.483.4.14&r2=1.483.4.15



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941

[Bug target/15941] [3.4/3.5 regression] new fails gcc.dg/compat/struct-by-value-11 and gcc.dg/compat/struct-by-value-1

[danglin at gcc dot gnu dot org]

------- Additional Comments From danglin at gcc dot gnu dot org  2004-06-19 19:50 -------
Patch applied.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15941