Re: Continue stmt branch prediction

["H.J. Lu" <hjl dot tools at gmail dot com>]

Hi,

This patch may have caused

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35636


H.J.
On Sat, Feb 9, 2008 at 12:48 PM, Jan Hubicka <jh@suse.cz> wrote:
>
>  Hi,
>  this patch adds a continue branch predictor that was dropped by introduction of
>  gimplifier (that replaced continue statement by goto).
>
>  The idea is that continue statements forms a loops that are usually not looping,
>  as proper loops are written via the loop construct.  On SPECint2000 average loop
>  tends to loop 7 times, while average loop constructed via continue loops less
>  than once.
>
>  The prediction is implemented via high level predictors that used also to predict,
>  for example, that explicit goto statements are not taken (this is helping kernel)
>  and can be used for other stuff in other languages too.
>  I will re-introduce the other predictors by followups later, this patch was
>  developed while looking into the gzip regression as predicting continue formed
>  loop to loop a lot is one of problems there:
>
>  statement
>   if (test1 || test2 || test3 || test4)
>     continue;
>
>  is in mainline handled as 4 nested loops, each predicted to loop roughly 4 times,
>  making all the rest of function cold.
>
>  Since the patch itself is not solving the regression fully (it makes improvement
>  on prelbmk 64bit peak, 920->980, and gzip 64bit, 750->770) but irronically it causes
>  regression at -O3 score of gzip 32bit by exposing other problems.  So I am holding
>  it for 4.4.
>
>  The branch prediction bits I can approve myself, but I would welcome comment on
>  introducing PREDICT_EXPR.  This was discussed while ago on GCC summit.  (the
>  original patch was done at CFG expansion times, but since we was about to run
>  out of tree codes I holded it and then forgot about the issue completely).
>
>  Bootstrapped/regtested i686-linux.
>
>  Honza
>
>         * tree-pretty-print.c: Include predict.h.
>         (dump_generic_node): Dump predictor.
>         * tree.h (PREDICT_EXPR_OUTCOME, PREDICT_EXPR_PREDICTION): Update.
>         * tree-gimple.c (is_gimple_stmt): Add PREDICT_EXPR.
>         * gimple-low.c (lower_stmt): Likewise.
>         * expr.c (expand_expr_real): Likewise.
>         * predict.c (tree_bb_level_predictions): Use PREDICT_EXPRs and remove
>         them.
>         (build_predict_expr, build_predict_expr): New.
>         * predict.h (predictor_name, build_predict_expr): Update.
>         * c-typeck.c (c_finish_bc_stmt): Add prediction.
>         * gimplify.c (gimplify_expr): Add PREDICT_EXPR.
>         * predict.def (PRED_CONTINUE): Update hitrate.
>         * tree.def (PREDICT_EXPR): Define.
>         * tree-ssa-dce.c (mark_stmt_if_obviously_necessary): Mark PREDICT_EXPR;
>         do not handle BIND_EXPR.
>         * tree-inline.c (estimate_num_insns_1): PREDICT_EXPR is free.
>         * tree-cfg.c (verify_gimple_stmt): PREDICT_EXPR is valid.
>         * tree-ssa-operands.c (get_expr_operands): PREDICT_EXPR takes no operands.
>
>