This is the mail archive of the
mailing list for the GCC project.
Re: Match-and-simplify and COND_EXPR
- From: Richard Biener <rguenther at suse dot de>
- To: Andrew Pinski <pinskia at gmail dot com>
- Cc: GCC Mailing List <gcc at gcc dot gnu dot org>
- Date: Fri, 07 Nov 2014 08:24:42 +0100
- Subject: Re: Match-and-simplify and COND_EXPR
- Authentication-results: sourceware.org; auth=none
- References: <CA+=Sn1=ZrkDn8FS-5qFvh-ndF2isFD9xBQmE_gLvA6qAVmE02w at mail dot gmail dot com> <alpine dot LSU dot 2 dot 11 dot 1411061126240 dot 27850 at zhemvz dot fhfr dot qr> <alpine dot LSU dot 2 dot 11 dot 1411061138230 dot 27850 at zhemvz dot fhfr dot qr> <CA+=Sn1=3tTc54x87WFQM3+GZeUF9UK35zbWVdY2ymcfqNEMC6A at mail dot gmail dot com>
On November 7, 2014 5:03:19 AM CET, Andrew Pinski <firstname.lastname@example.org> wrote:
>On Thu, Nov 6, 2014 at 2:40 AM, Richard Biener <email@example.com>
>> On Thu, 6 Nov 2014, Richard Biener wrote:
>>> On Wed, 5 Nov 2014, Andrew Pinski wrote:
>>> > Hi,
>>> > I was trying to hook up tree-ssa-phiopt to match-and-simplify
>>> > either gimple_build (or rather using gimple_simplify depending on
>>> > we want to produce cond_expr for conditional move). I ran into a
>>> > problem.
>>> > With the pattern below:
>>> > /* a ? 0 : 1 -> a if 0 and 1 are integral types. */
>>> > (simplify
>>> > (cond_expr @0 integer_zerop integer_onep)
>>> > (if (INTEGRAL_TYPE_P (type))
>>> > (convert @0)))
>>> Ok, so you are capturing a GENERIC expr here but nothing knows that.
>>> It would work if you'd do (ugh)
>>> (for op (lt le eq ne ge gt)
>>> (cond_expr (op @0 @1) integer_zerop integer_onep)
>>> (if (INTEGRAL_TYPE_P (type))
>>> (convert (op @0 @1)))))
>>> (cond_expr SSA_NAME@0 integer_zerop integer_onep)
>>> (if (INTEGRAL_TYPE_P (type))
>>> (convert @0))))
>>> as a workaround. To make your version work will require (quite)
>>> some special-casing in the code generator or maybe the resimplify
>>> helper. Let me see if I can cook up a "simple" fix.
>> Sth like below (for the real fix this has to be replicated in
>> all gimple_resimplifyN functions). I'm missing a testcase
>> where the pattern would apply (and not be already folded by fold),
>> so I didn't check if it actually works.
>You do need to check if seq is NULL though as gimple_build depends on
>seq not being NULL. But otherwise yes this works for me.
>> Bah, of course we should fix COND_EXPRs to not embed a GENERIC
>Yes totally agree. For my changes to tree-ssa-phiopt, I no longer
>embed it. Though we need to change loop ifconvert still.
Istr expansion or code quality does not like us to cse the condition of two cobd_exprs either. After all I had a patch set at some point doing that conversion (though as well for gimple_conds).
>> Index: gcc/gimple-match-head.c
>> --- gcc/gimple-match-head.c (revision 217035)
>> +++ gcc/gimple-match-head.c (working copy)
>> @@ -90,6 +90,13 @@ gimple_resimplify1 (gimple_seq *seq,
>> code_helper *res_code, tree type, tree *res_ops,
>> tree (*valueize)(tree))
>> + /* ??? Stupid tcc_comparison GENERIC trees in COND_EXPRs. */
>> + if (COMPARISON_CLASS_P (res_ops))
>> + res_ops = gimple_build (seq,
>> + TREE_CODE (res_ops), TREE_TYPE
>> + TREE_OPERAND (res_ops, 0),
>> + TREE_OPERAND (res_ops, 1));
>> if (constant_for_folding (res_ops))
>> tree tem = NULL_TREE;