[PATCH] PR tree-optimization/68413 : Only check for integer cond reduction on analysis stage

[Alan Hayward]

On 20/11/2015 13:47, "Richard Biener" <richard.guenther@gmail.com> wrote:

>On Fri, Nov 20, 2015 at 1:33 PM, Alan Hayward <alan.hayward@arm.com>
>wrote:
>>
>>
>>On 20/11/2015 11:00, "Richard Biener" <richard.guenther@gmail.com> wrote:
>>
>>>On Fri, Nov 20, 2015 at 10:24 AM, Alan Hayward <alan.hayward@arm.com>
>>>wrote:
>>>>When vectorising a integer induction condition reduction,
>>>>is_nonwrapping_integer_induction ends up with different values for base
>>>>during the analysis and build phases. In the first it is an
>>>>INTEGER_CST,
>>>>in the second the loop has been vectorised out and the base is now a
>>>>variable.
>>>>
>>>>This results in the analysis and build stage detecting the
>>>>STMT_VINFO_VEC_REDUCTION_TYPE as different types.
>>>>
>>>>The easiest way to fix this is to only check for integer induction
>>>>conditions on the analysis stage.
>>>
>>>I don't like this.  For the evolution part we have added
>>>STMT_VINFO_LOOP_PHI_EVOLUTION_PART.  If you now need
>>>the original initial value as well then just save it.
>>>
>>>Or if you really want to go with the hack then please do not call
>>>is_nonwrapping_integer_induction with vec_stmt != NULL but
>>>initialize cond_expr_is_nonwrapping_integer_induction from
>>>STMT_VINFO_VEC_REDUCTION_TYPE (stmt_info)
>>>
>>>The hack also lacks a comment.
>>>
>>
>>Ok. I've gone for a combination of both:
>>
>>I now cache the base in STMT_VINFO_LOOP_PHI_EVOLUTION_BASE.
>>
>>I've removed the vec_stmt != NULL checks.
>>
>>I've moved the call to is_nonwrapping_integer_induction until after the
>>vect_is_simple_reduction check. I never liked that I had
>>is_nonwrapping_integer_induction early in the function, and think this
>>looks better.
>
>It looks better but the comment for loop_phi_evolution_base is wrong.
>The value is _not_ "correct" after prologue peeling (unless you
>update it there to a non-constant expr).  It is conservatively "correct"
>for is_nonwrapping_integer_induction though.  Which is why I'd
>probably rename it to STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED
>to reflect this (and similar to STMT_VINFO_NITERS_UNCHANGED).
>
>Ok with that change.

Updated as requested and submitted.

Alan.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: analysisonlycondcheck3.patch
Type: application/octet-stream
Size: 5362 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20151120/9bae56da/attachment.obj>