[Bug target/25413] wrong alignment or incorrect address computation in vectorized code on Pentium 4 SSE

["dirtyepic at gentoo dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #14 from dirtyepic at gentoo dot org  2007-07-24 19:37 -------
Created an attachment (id=13965)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13965&action=view)
zlib testcase

> A fix for PR25413 was committed to mainline.
> Ryan, could you please check if it solves the zlib miscompilation?
> Andrew, would you plase check if it solves the libgcc miscompilation that you
> are seeing?

Unfortunately it doesn't.  I'm still getting unaligned movdqa instructions with
both mainline and the 4.2 patch.  Both testcases on this bug now work however,
so maybe the problem lies elsewhere.

I'm attaching a (badly) reduced testcase from inftrees.c in zlib which i
believe shows the bug.  Compile with -O -msse2 -ftree-vectorize.  On i686 i'm
getting this:

inftrees.lo:     file format elf32-i386

Disassembly of section .text:

00000000 <inflate_table>:
   0:   55                      push   %ebp
   1:   89 e5                   mov    %esp,%ebp
   3:   53                      push   %ebx
   4:   83 ec 24                sub    $0x24,%esp
   7:   8b 5d 0c                mov    0xc(%ebp),%ebx
   a:   8b 4d 10                mov    0x10(%ebp),%ecx
   d:   8d 45 d8                lea    -0x28(%ebp),%eax
  10:   66 0f ef c0             pxor   %xmm0,%xmm0
  14:   66 0f 7f 00             movdqa %xmm0,(%eax)
  18:   66 0f 7f 40 10          movdqa %xmm0,0x10(%eax)
  1d:   85 c9                   test   %ecx,%ecx
  1f:   74 16                   je     37 <inflate_table+0x37>
  21:   ba 00 00 00 00          mov    $0x0,%edx
  26:   0f b7 04 53             movzwl (%ebx,%edx,2),%eax
  2a:   66 83 44 45 d8 01       addw   $0x1,-0x28(%ebp,%eax,2)
  30:   83 c2 01                add    $0x1,%edx
  33:   39 ca                   cmp    %ecx,%edx
  35:   75 ef                   jne    26 <inflate_table+0x26>
  37:   b8 0f 00 00 00          mov    $0xf,%eax
  3c:   8d 55 d8                lea    -0x28(%ebp),%edx
  3f:   66 83 3c 42 00          cmpw   $0x0,(%edx,%eax,2)
  44:   75 05                   jne    4b <inflate_table+0x4b>
  46:   83 e8 01                sub    $0x1,%eax
  49:   75 f4                   jne    3f <inflate_table+0x3f>
  4b:   83 c4 24                add    $0x24,%esp
  4e:   5b                      pop    %ebx
  4f:   5d                      pop    %ebp
  50:   c3                      ret


And without the vectorizer:

inftrees.lo:     file format elf32-i386

Disassembly of section .text:

00000000 <inflate_table>:
   0:   55                      push   %ebp
   1:   89 e5                   mov    %esp,%ebp
   3:   53                      push   %ebx
   4:   83 ec 20                sub    $0x20,%esp
   7:   8b 5d 0c                mov    0xc(%ebp),%ebx
   a:   8b 4d 10                mov    0x10(%ebp),%ecx
   d:   b8 00 00 00 00          mov    $0x0,%eax
  12:   8d 55 dc                lea    -0x24(%ebp),%edx
  15:   66 c7 04 42 00 00       movw   $0x0,(%edx,%eax,2)
  1b:   83 c0 01                add    $0x1,%eax
  1e:   83 f8 10                cmp    $0x10,%eax
  21:   75 f2                   jne    15 <inflate_table+0x15>
  23:   85 c9                   test   %ecx,%ecx
  25:   74 16                   je     3d <inflate_table+0x3d>
  27:   ba 00 00 00 00          mov    $0x0,%edx
  2c:   0f b7 04 53             movzwl (%ebx,%edx,2),%eax
  30:   66 83 44 45 dc 01       addw   $0x1,-0x24(%ebp,%eax,2)
  36:   83 c2 01                add    $0x1,%edx
  39:   39 ca                   cmp    %ecx,%edx
  3b:   75 ef                   jne    2c <inflate_table+0x2c>
  3d:   b8 0f 00 00 00          mov    $0xf,%eax
  42:   8d 55 dc                lea    -0x24(%ebp),%edx
  45:   66 83 3c 42 00          cmpw   $0x0,(%edx,%eax,2)
  4a:   75 05                   jne    51 <inflate_table+0x51>
  4c:   83 e8 01                sub    $0x1,%eax
  4f:   75 f4                   jne    45 <inflate_table+0x45>
  51:   83 c4 20                add    $0x20,%esp
  54:   5b                      pop    %ebx
  55:   5d                      pop    %ebp
  56:   c3                      ret


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=25413