--- Comment #10 from hjl dot tools at gmail dot com 2009-02-13 15:57
---
Fixed. Gcc 4.4.0 revision 144128 generates:
foo2:
movd%edi, %xmm0
movd%esi, %xmm1
movd%edx, %xmm2
punpckldq %xmm0, %xmm1
movd%ecx, %xmm0
--- Comment #8 from uweigand at gcc dot gnu dot org 2008-05-18 15:58
---
That special case in find_reloads is really about a different situation.
We do not have a simple move here.
The problem also is not really related to vector instruction in particular;
reload doesn't at all care
--- Comment #9 from rguenth at gcc dot gnu dot org 2008-05-18 16:58 ---
Did you investigate whether IRA fixes this issue?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36222
--- Comment #6 from ubizjak at gmail dot com 2008-05-16 19:09 ---
There is still underlying RA problem left. Compiling the testcase from Comment
0 on 32bit i686 (-O2 -msse2), we get:
_mm_set_epi32:
pushl %ebp
movl%esp, %ebp
movd8(%ebp), %xmm0
--- Comment #7 from hjl dot tools at gmail dot com 2008-05-16 21:54 ---
find_reloads in reload.c has
/* Special case a simple move with an input reload and a
destination of a hard reg, if the hard reg is ok, use it. */
for (i = 0; i n_reloads; i++)
if
--- Comment #4 from hjl dot tools at gmail dot com 2008-05-13 14:09 ---
It looks like reload doesn't check any vector instructions. I guess there may
be many missed optimizations with vector instructions.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36222
--- Comment #3 from hjl dot tools at gmail dot com 2008-05-13 13:58 ---
This bug may be related to PR 30961. Another example:
bash-3.2$ cat d.c
#include emmintrin.h
__m128i
foo2 (long long x1, long long x2)
{
return _mm_set_epi64x (x1, x2);
}
bash-3.2$
--- Comment #5 from uros at gcc dot gnu dot org 2008-05-13 21:34 ---
Subject: Bug 36222
Author: uros
Date: Tue May 13 21:33:40 2008
New Revision: 135275
URL: http://gcc.gnu.org/viewcvs?root=gccview=revrev=135275
Log:
PR target/36222
* config/i386/i386.c
--- Comment #1 from hjl dot tools at gmail dot com 2008-05-12 15:57 ---
Also do we need movq%xmm1, %xmm2?
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36222
--- Comment #2 from ubizjak at gmail dot com 2008-05-12 19:23 ---
(In reply to comment #1)
Also do we need movq%xmm1, %xmm2?
We can help RA a bit by emitting RTL sequence that requires less pseudos.
Index: i386.c
===
10 matches
Mail list logo
