[Bug target/49094] New: ARM aligned(1) attribute is sometimes dropped

[Bug target/49094] New: ARM aligned(1) attribute is sometimes dropped

[will_lentz at trimble dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

           Summary: ARM aligned(1) attribute is sometimes dropped
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: will_lentz@trimble.com


Created attachment 24313
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24313
simple test case to demonstrate problem

GCC 4.6.0 sometimes produces unaligned accesses for structures with
__attribute__((aligned(1),packed)).  GCC 4.4.1 seems fine.

See the attached "tst.c" file.  GCC 4.6.0 will access an unaligned ip->ip_dst
with "ldr" instead of several "ldrb" instructions.

The file can be compiled and examined with:
  arm-none-eabi-gcc -c  -g -Os tst.c
  arm-none-eabi-objdump -S tst.o

*********
GCC 4.6.0 produces:
Disassembly of section .text:

00000000 <test_alignment>:
{
  struct ip *ip = (struct ip *)m;
  struct in_addr pkt_dst;
  pkt_dst = ip->ip_dst ;
  if( pkt_dst.s_addr == 0 )
    return 1;
   0:   e5900007        ldr     r0, [r0, #7]
  else
    return 0;
}
   4:   e2700001        rsbs    r0, r0, #1
   8:   33a00000        movcc   r0, #0
   c:   e12fff1e        bx      lr


*********
GCC 4.4.1 produces:
Disassembly of section .text:

00000000 <test_alignment>:
} __attribute__ ((aligned(1), packed));

struct ip *ip_fw_fwd_addr;

int test_alignment( char *m )
{
   0:   e5d02008        ldrb    r2, [r0, #8]
   4:   e5d03007        ldrb    r3, [r0, #7]
   8:   e5d01009        ldrb    r1, [r0, #9]
   c:   e1833402        orr     r3, r3, r2, lsl #8
  10:   e5d0200a        ldrb    r2, [r0, #10]
  14:   e1833801        orr     r3, r3, r1, lsl #16
  18:   e1932c02        orrs    r2, r3, r2, lsl #24
  pkt_dst = ip->ip_dst ;
  if( pkt_dst.s_addr == 0 )
    return 1;
  else
    return 0;
}
  1c:   13a00000        movne   r0, #0  ; 0x0
  20:   03a00001        moveq   r0, #1  ; 0x1
  24:   e12fff1e        bx      lr

[Bug target/49094] ARM aligned(1) attribute is sometimes dropped

[ramana at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

Ramana Radhakrishnan <ramana at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.05.31 22:04:11
                 CC|                            |ramana at gcc dot gnu.org
     Ever Confirmed|0                           |1
      Known to fail|                            |4.6.0, 4.7.0

--- Comment #1 from Ramana Radhakrishnan <ramana at gcc dot gnu.org> 2011-05-31 22:04:11 UTC ---
Confirmed and copyrename appears to generate this which is where things appear
to start going pear-shaped. 

test_alignment (char * m)
{
  unsigned int pkt_dst$s_addr;
  int D.2017;

<bb 2>:
  pkt_dst$s_addr_8 = MEM[(struct ip *)m_2(D) + 7B].s_addr;
  if (pkt_dst$s_addr_8 == 0)
    goto <bb 4>;
  else
    goto <bb 3>;

<bb 3>:

<bb 4>:
  # D.2017_1 = PHI <1(2), 0(3)>
  return D.2017_1;
}

[Bug target/49094] ARM aligned(1) attribute is sometimes dropped

[mikpe at it dot uu.se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

Mikael Pettersson <mikpe at it dot uu.se> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jamborm at gcc dot gnu.org,
                   |                            |mikpe at it dot uu.se

--- Comment #2 from Mikael Pettersson <mikpe at it dot uu.se> 2011-06-05 21:51:42 UTC ---
It's caused by r164136:

Author: jamborm
Date: Thu Sep  9 23:38:23 2010
New Revision: 164136

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=164136
Log:
2010-09-10  Martin Jambor  <mjambor@suse.cz>

    PR tree-optimization/44972
    * tree-sra.c: Include toplev.h.
    (build_ref_for_offset): Entirely reimplemented.
...

sparc64-linux regresses similarly at this revision.

[Bug target/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #3 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-06-11 19:12:16 UTC ---
Just for the record, I am aware of this bug, I have managed to
reproduce it on the sparc64 in the compile farm and have had a look at
it a few times already (albeit so far only briefly).

A few mechanisms within SRA come into play as we generate the integer
load statement which is wrong:

  pkt_dst$s_addr_8 = MEM[(struct ip *)ip_3 + 7B].s_addr;

I will continue to try to figure out what to do about this but it is
not very straightforward.

[Bug target/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #4 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-06-11 23:08:59 UTC ---
Created attachment 24494
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24494
Fix - a first draft

OK, a fix is probably going to look very much like this, at least
roughly.  Perhaps it would be better to employ this new
disqualification of SRA candidates only if the target is strict
aligned.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #5 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-06-26 18:25:16 UTC ---
I have submitted a fix for this bug to the mailing list:

http://gcc.gnu.org/ml/gcc-patches/2011-06/msg01941.html

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #6 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-06-30 13:24:23 UTC ---
Author: jamborm
Date: Thu Jun 30 13:24:19 2011
New Revision: 175703

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175703
Log:
2011-06-30  Martin Jambor  <mjambor@suse.cz>

    PR tree-optimization/49094
    * tree-sra.c (tree_non_mode_aligned_mem_p): New function.
    (build_accesses_from_assign): Use it.

    * testsuite/gcc.dg/tree-ssa/pr49094.c: New test.


Added:
    trunk/gcc/testsuite/gcc.dg/tree-ssa/pr49094.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-sra.c

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[will_lentz at trimble dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #7 from will_lentz at trimble dot com 2011-07-11 16:29:52 UTC ---
(In reply to comment #6)
> Author: jamborm
> Date: Thu Jun 30 13:24:19 2011
> New Revision: 175703
> 

Thanks for fixing this!  Do you know if this fix will get into a 4.6.x release?

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #8 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-07-11 16:54:46 UTC ---
Author: jamborm
Date: Mon Jul 11 16:54:42 2011
New Revision: 176166

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=176166
Log:
2011-07-11  Martin Jambor  <mjambor@suse.cz>

    PR tree-optimization/49094
    * tree-sra.c (tree_non_mode_aligned_mem_p): New function.
    (build_accesses_from_assign): Use it.

    * testsuite/gcc.dg/tree-ssa/pr49094.c: New test.


Added:
    branches/gcc-4_6-branch/gcc/testsuite/gcc.dg/tree-ssa/pr49094.c
Modified:
    branches/gcc-4_6-branch/gcc/ChangeLog
    branches/gcc-4_6-branch/gcc/testsuite/ChangeLog
    branches/gcc-4_6-branch/gcc/tree-sra.c

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

Martin Jambor <jamborm at gcc dot gnu.org> changed:

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

--- Comment #9 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-07-11 17:36:19 UTC ---
(In reply to comment #7)
> (In reply to comment #6)
> > Author: jamborm
> > Date: Thu Jun 30 13:24:19 2011
> > New Revision: 175703
> > 
> 
> Thanks for fixing this!  Do you know if this fix will get into a 4.6.x release?

Yeah, sorry it took so long, I almost forgot.  The patch for 4.6 has
been sent to the mailing list in
http://gcc.gnu.org/ml/gcc-patches/2011-07/msg00827.html

This is now fixed on both the trunk and the 4.6 branch.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[sgh at sgh dot dk]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #10 from Søren Holm <sgh at sgh dot dk> 2011-07-30 12:10:57 UTC ---
Created attachment 24869
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24869
Another testcase that fails. Even with current 4.6

With "static" on line 33 the code generated is wrong. Removing "static" the
code generated is correct.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[sgh at sgh dot dk]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

Søren Holm <sgh at sgh dot dk> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |sgh at sgh dot dk

--- Comment #11 from Søren Holm <sgh at sgh dot dk> 2011-07-30 12:13:56 UTC ---
tst2.c also fails with an alignment related problem. I'm unsue if this should
be reported in another bug.

With "static" on line 33 the code generated is this :

<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
unsigned int aostk_font_strwidth(const struct aostk_font* font, const char*
str) {
        const struct aostk_glyph* g = aostk_get_glyph(font, 0);
        return g->advance.x;
   0:   e5903003        ldr     r3, [r0, #3]
}
   4:   e5d30005        ldrb    r0, [r3, #5]
   8:   e12fff1e        bx      lr
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>

The first ldr-instruction is wrong afaik sinct struct aostk_font is packed.

Removing "static" fomr line 33 gives this code :

<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
const struct aostk_glyph* aostk_get_glyph(const struct aostk_font* f, unsigned
int c) {
        return &f->glyphs[0];
   0:   e5d02004        ldrb    r2, [r0, #4]
   4:   e5d03003        ldrb    r3, [r0, #3]
   8:   e1833402        orr     r3, r3, r2, lsl #8
   c:   e5d02005        ldrb    r2, [r0, #5]
  10:   e5d00006        ldrb    r0, [r0, #6]
  14:   e1833802        orr     r3, r3, r2, lsl #16
}
  18:   e1830c00        orr     r0, r3, r0, lsl #24
  1c:   e12fff1e        bx      lr

00000020 <aostk_font_strwidth>:
        const struct aostk_glyph* glyphs;
};


const struct aostk_glyph* aostk_get_glyph(const struct aostk_font* f, unsigned
int c) {
        return &f->glyphs[0];
  20:   e5d02004        ldrb    r2, [r0, #4]
  24:   e5d03003        ldrb    r3, [r0, #3]
  28:   e1833402        orr     r3, r3, r2, lsl #8
  2c:   e5d02005        ldrb    r2, [r0, #5]
  30:   e1833802        orr     r3, r3, r2, lsl #16
  34:   e5d02006        ldrb    r2, [r0, #6]
  38:   e1833c02        orr     r3, r3, r2, lsl #24


unsigned int aostk_font_strwidth(const struct aostk_font* font, const char*
str) {
        const struct aostk_glyph* g = aostk_get_glyph(font, 0);
        return g->advance.x;
}
  3c:   e5d30005        ldrb    r0, [r3, #5]
  40:   e12fff1e        bx      lr
>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>

Notice in aostk_get_glyph that the ldrb-instructions are used instead of ldr.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[sgh at sgh dot dk]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #12 from Søren Holm <sgh at sgh dot dk> 2011-07-30 12:30:25 UTC ---
Oh... I just realized that the failing case does not fail because of the first
ldr-instruction. It fails becasue of something else. Maybe you guys have a
better eye for ARM-assembler to be able to verify that the code.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[sgh at sgh dot dk]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #13 from Søren Holm <sgh at sgh dot dk> 2011-07-30 12:37:41 UTC ---
gcc 4.4.6 generates the following code with the "static" keyword

unsigned int aostk_font_strwidth(const struct aostk_font* font, const char*
str) {
   0:   e5d02004        ldrb    r2, [r0, #4]
   4:   e5d03003        ldrb    r3, [r0, #3]
   8:   e5d01005        ldrb    r1, [r0, #5]
   c:   e1833402        orr     r3, r3, r2, lsl #8
  10:   e5d02006        ldrb    r2, [r0, #6]
  14:   e1833801        orr     r3, r3, r1, lsl #16
  18:   e1833c02        orr     r3, r3, r2, lsl #24
        const struct aostk_glyph* g = aostk_get_glyph(font, 0);
        return g->advance.x;
}
  1c:   e5d30005        ldrb    r0, [r3, #5]
  20:   e12fff1e        bx      lr

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[jamborm at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #14 from Martin Jambor <jamborm at gcc dot gnu.org> 2011-07-30 17:15:12 UTC ---
(In reply to comment #11)
> tst2.c also fails with an alignment related problem. I'm unsue if this should
> be reported in another bug.
> 

It it fails with a current checkout of the trunk or the 4.6 branch it
is most probably a different issue and should be reported in a
separate bug.

So please file a new bug, post its number here, and, if the testcase
does not fail with -fno-tree-sra, CC me.  Thanks.

[Bug tree-optimization/49094] ARM aligned(1) attribute is sometimes dropped

[sgh at sgh dot dk]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49094

--- Comment #15 from Søren Holm <sgh at sgh dot dk> 2011-07-31 15:18:19 UTC ---
I have reported the bug as #49923 

-fno-tree-sra does not help.

I have bisected it to commit trunk@164135

2010-09-10  Martin Jambor  <mjambor@suse.cz>

        PR tree-optimization/44972                                              
        * ipa-prop.c (ipa_modify_call_arguments): Build MEM_REF instead of      
        calling build_ref_for_offset.

        * testsuite/g++.dg/torture/pr34850.C: Remove expected warning.