On Mon, Dec 12, 2011 at 9:54 PM, Ira Rosen <i...@il.ibm.com> wrote:
>
>
> gcc-patches-ow...@gcc.gnu.org wrote on 13/12/2011 04:05:57 AM:
>
>> On core2, unaligned vector load/store using movdqu is a very slow
> operation.
>> Experiments show it is six times slower than movdqa (aligned) and this is
>> irrespective of whether the resulting data happens to be aligned or not.
>> For Corei7, there is no performance difference between the two and on
> AMDs,
>> movdqu is only about 10% slower.
>>
>> This patch does not vectorize loops that need to generate the slow
> unaligned
>> memory load/stores on core2.
>>
>>
>> Do not vectorize loops on Core2 that need to use unaligned
>> vector load/stores.
>> * tree-vect-stmts.c (is_slow_vect_unaligned_load_store): New function.
>> (vect_analyze_stmt): Check if the vectorizable load/store is slow.
>> * target.def (TARGET_SLOW_UNALIGNED_VECTOR_MEMOP): New target hook.
>> * doc/m.texi.in: Document new target hook:
>> TARGET_SLOW_UNALIGNED_VECTOR_MEMOP
>> * doc/m.texi: Regenerate.
>> * config/i386/i386.c (ix86_slow_unaligned_vector_memop): New function.
>> (TARGET_SLOW_UNALIGNED_VECTOR_MEMOP): New macro.
>>
>
>
>> @@ -5065,27 +5112,43 @@ vect_analyze_stmt (gimple stmt, bool
> *need_to_vect
>> if (!bb_vinfo
>> && (STMT_VINFO_RELEVANT_P (stmt_info)
>> || STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def))
>> + {
>> ok = (vectorizable_type_promotion (stmt, NULL, NULL, NULL)
>> || vectorizable_type_demotion (stmt, NULL, NULL, NULL)
>> || vectorizable_conversion (stmt, NULL, NULL, NULL)
>> || vectorizable_shift (stmt, NULL, NULL, NULL)
>> || vectorizable_operation (stmt, NULL, NULL, NULL)
>> || vectorizable_assignment (stmt, NULL, NULL, NULL)
>> - || vectorizable_load (stmt, NULL, NULL, NULL, NULL)
>> || vectorizable_call (stmt, NULL, NULL)
>> - || vectorizable_store (stmt, NULL, NULL, NULL)
>> - || vectorizable_reduction (stmt, NULL, NULL, NULL)
>> + || vectorizable_reduction (stmt, NULL, NULL, NULL)
>> || vectorizable_condition (stmt, NULL, NULL, NULL, 0));
>> +
>> + if (!ok)
>> + {
>> + ok = (vectorizable_load (stmt, NULL, NULL, NULL, NULL)
>> + || vectorizable_store (stmt, NULL, NULL, NULL));
>> +
>> + if (ok && is_slow_vect_unaligned_load_store (stmt))
>> + ok = false;
>
> Why not call is_slow_vect_unaligned_load_store from
> vectorizable_load/store?
Yes, I should have done that. Somehow, I missed that!
>
> Ira
>
>
>> + }
>> + }
>> else
>> {
>> if (bb_vinfo)
>> - ok = (vectorizable_type_promotion (stmt, NULL, NULL, node)
>> - || vectorizable_type_demotion (stmt, NULL, NULL, node)
>> - || vectorizable_shift (stmt, NULL, NULL, node)
>> - || vectorizable_operation (stmt, NULL, NULL, node)
>> - || vectorizable_assignment (stmt, NULL, NULL, node)
>> - || vectorizable_load (stmt, NULL, NULL, node, NULL)
>> - || vectorizable_store (stmt, NULL, NULL, node));
>> + {
>> + ok = (vectorizable_type_promotion (stmt, NULL, NULL, node)
>> + || vectorizable_type_demotion (stmt, NULL, NULL, node)
>> + || vectorizable_shift (stmt, NULL, NULL, node)
>> + || vectorizable_operation (stmt, NULL, NULL, node)
>> + || vectorizable_assignment (stmt, NULL, NULL, node));
>> + if (!ok)
>> + {
>> + ok = (vectorizable_load (stmt, NULL, NULL, node, NULL)
>> + || vectorizable_store (stmt, NULL, NULL, node));
>> + if (ok && is_slow_vect_unaligned_load_store (stmt))
>> + ok = false;
>> + }
>> + }
>> }
>
