[Bug target/55634] New: ARM: gcc vector extensions: storing vector to unaligned memory location does not use VST1.8 NEON instruction

[Bug target/55634] New: ARM: gcc vector extensions: storing vector to unaligned memory location does not use VST1.8 NEON instruction

[siarhei.siamashka at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55634

             Bug #: 55634
           Summary: ARM: gcc vector extensions: storing vector to
                    unaligned memory location does not use VST1.8 NEON
                    instruction
    Classification: Unclassified
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: siarhei.siamashka@gmail.com


The following test program tries to use GCC vector extensions to add two
vectors together and store the result to unaligned memory location in a
"portable" way with memcpy:

/***********************************************/

#include <string.h>

typedef unsigned int T __attribute__ ((vector_size (16)));

void foo (void *result, T *a, T *b)
{
  T tmp = *a + *b;
  memcpy (result, &tmp, sizeof(tmp));
}

/***********************************************/

Compiling with gcc 4.7.2:

$ arm-none-linux-gnueabi-gcc -O2 -mcpu=cortex-a8 -mfpu=neon -c test.c
$ objdump -d test.o

00000000 <foo>:
   0:    e52d4004     push    {r4}        ; (str r4, [sp, #-4]!)
   4:    ecd12b04     vldmia    r1, {d18-d19}
   8:    e24dd014     sub    sp, sp, #20
   c:    ecd20b04     vldmia    r2, {d16-d17}
  10:    e28dc010     add    ip, sp, #16
  14:    f26208e0     vadd.i32    q8, q9, q8
  18:    ed6c0b04     vstmdb    ip!, {d16-d17}
  1c:    e1a0c00d     mov    ip, sp
  20:    e1a04000     mov    r4, r0
  24:    e8bc000f     ldm    ip!, {r0, r1, r2, r3}
  28:    e5840000     str    r0, [r4]
  2c:    e5841004     str    r1, [r4, #4]
  30:    e5842008     str    r2, [r4, #8]
  34:    e584300c     str    r3, [r4, #12]
  38:    e28dd014     add    sp, sp, #20
  3c:    e8bd0010     pop    {r4}
  40:    e12fff1e     bx    lr

The same test program results in the following code if compiled for x86-64:

0000000000000000 <foo>:
   0:    66 0f 6f 06              movdqa (%rsi),%xmm0
   4:    66 0f fe 02              paddd  (%rdx),%xmm0
   8:    f3 0f 7f 07              movdqu %xmm0,(%rdi)
   c:    c3                       retq   

So x86-64 target is able to use MOVDQU instruction. Hence ARM target should be
able to use VST1.8 as well.

[Bug target/55634] ARM: gcc vector extensions: storing vector to unaligned memory location does not use VST1.8 NEON instruction

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55634

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2012-12-10 09:40:59 UTC ---
Issue with the movmisalign instructions.  They seem to take only a limited
set of modes, in particular not the mode that is requested.

[Bug target/55634] ARM: gcc vector extensions: storing vector to unaligned memory location does not use VST1.8 NEON instruction

[ramana at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55634

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2013-08-05
                 CC|                            |ramana at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #2 from Ramana Radhakrishnan <ramana at gcc dot gnu.org> ---
On AArch64 with no strict alignment we end up generating with .002t.original

trunk 


{
  T tmp = *a + *b;
  extern void * memcpy (void *, const void *, long unsigned int);

    T tmp = *a + *b;
  MEM[(char * {ref-all})result] = MEM[(char * {ref-all})&tmp];, result;
}



On A32 or indeed AArch64 with -mstrict-align we end up generating 

{
  T tmp = *a + *b;
  extern void * memcpy (void *, const void *, long unsigned int);

    T tmp = *a + *b;
  memcpy (result, (const void *) &tmp, 16);
}


Where do you expect the memcpy to have been made redundant or a use of the
appropriate movmisalign insn - richi ? 

Ramana

[Bug target/55634] ARM: gcc vector extensions: storing vector to unaligned memory location does not use VST1.8 NEON instruction

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55634

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
         Resolution|---                         |FIXED
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=102125
   Target Milestone|---                         |12.0
             Status|NEW                         |RESOLVED

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed for GCC 12.1.0 via the patches which fixed PR 102125 .