Re: [tree-ssa] Branch prediction infrastructure

[Diego Novillo <dnovillo at redhat dot com>]

On Tue, 2003-12-23 at 07:03, Jan Hubicka wrote:

> 	* basic-block.h: Include predict.h
> 	(tree_predicted_by_p, rtl_predicted_by_p,  rtl_predict_edge,
> 	predict_edge_def): Declare.
> 	* cfghooks.h (cfg_hooks): add predict_edge and predicted_by_p
> 	(predict_edge, predicted_by_p): New macros.
> 	* cfgrtl (rtl_cfg_hooks, cfg_layout_rtl_cfg_hook): Add new hooks.
> 	* cse.c (struct cse_basic_block_data): Rename enum values to not
> 	conflict with profile.h; update all uses.
> 	* predict.c:  Include tree-flow.h, ggc.h, tree-dump.h
> 	(predicted_by_p): Rename to ...
> 	(rtl_predicted_by_p): .. this one; make global
> 	(tree_predicted_by_p): New.
> 	(dump_prediction): Add FILE argument.
> 	(predict_edge): Rename to ...
> 	(rtl_predict_edge): .. this one.
> 	(tree_predict_edge): New.
> 	(combine_predictions_for_insn): Update calls of predict_edge.
> 	(predict_loops): Break out from ...
> 	(estimate_probability): ... here; update comments; move updating
> 	of unknown probabilities from ...
> 	(estimate_bb_frequencies): ... here.
> 	(combine_predictions_for_bb): New.
> 	(tree_predict_by_opcode): New.
> 	(tree_estimate_probability): New.
> 	* predict.def (PRED_TREE_POINTER, PRED_TREE_OPCODE_POSITIVE,
> 	PRED_TREE_OPCODE_NONEQUAL, PRED_TREE_FPOPCODE): New predictors.
> 	* predict.h: Add include guard.
> 	(predict_edge, predict_edge_def): Move prototypes to basic_block.h
> 	* tree-cfg.c (tree_cfg_hooks): Add prediction hooks.
> 	* tree-dump.c (dump_files): Add profile.
> 	* tree-flow.h (struct edge_prediction): New structure.
> 	(struct bb_ann_d): Add field predictions.
> 	(tree_estimate_probability): Declare.
> 	* tree-optimize.c (optimize_function_tree): Call tree_estimate_probability.
> 	* tree.h (tree_dump_index): Add TDI_profile.
>
Looks OK, I guess (minor changes below).

Now, what's the plan wrt exporting the predictors to RTL?  Have we
settled on that already?


Diego.

> *************** predict_edge (edge e, enum br_predictor 
> *** 211,216 ****
> --- 226,244 ----
>     predict_insn (last_insn, predictor, probability);
>   }
>   
> + /* Same as PREDICT_EDGE but using tree presentation.  */
>
Change to something more meaningful like 'Predict edge E with the given
PROBABILITY.'

> ! 
> !   /* When there is no succestor or only one choice, prediction is easy. 
> ! 
s/succestor/successor/

> !      We are lazy for now and predict only basic blocks with two outgoing
> !      edges.  It is possible to predict generic case too, but we have to
> !      ignore first match heursitics and do more involved combining.  Implement
>
s/heursitics/heuristics/


Diego.