On Mon, Dec 9, 2013 at 12:27 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> On Fri, Dec 06, 2013 at 01:49:50PM +0100, Richard Biener wrote:
>> > basic_block bb = ifc_bbs[i];
>> > gimple_seq stmts;
>> >
>> > - if (!is_predicated (bb))
>> > + if (!is_predicated (bb)
>> > + || dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
>>
>> isn't that redundant now?
>
> After IRC discussion, moved this dominated_by_p call and another one
> from predicate_bbs to add_to_predicate_list, so that we don't change
> an always true predicate to something else on a bb that dominates
> loop->latch and therefore will be executed unconditionally inside of
> the loop body.
>
>> > + if (TREE_CODE (addr) == SSA_NAME && !SSA_NAME_PTR_INFO (addr))
>> > + copy_ref_info (build2 (MEM_REF, TREE_TYPE (ref), addr, ptr),
>> > + ref);
>>
>> Eh - can you split out a copy_ref_info_to_addr so you can avoid
>> creating the MEM_REF?
>
> Haven't changed this one, because copy_ref_info uses the new_ref in
> several places (in addition to the initial few ones).
>
>> > static bool
>> > +version_loop_for_if_conversion (struct loop *loop, bool *do_outer)
>> > +{
>>
>> What's the do_outer parameter?
>
> Outer loops aren't versioned anymore.
>
>> Please add a comment before this. Seems you match what outer loop
>> vectorization handles? Thus, best factor out a predicate in
>> tree-loop-vect.c that you can use in both places?
>
> So this went away completely.
>
>> This needs a comment with explaining what code you create.
>
> Ditto (and several other spots).
>
>> Btw I hate that we do update_ssa multiple times per pass per
>> function. That makes us possibly worse than O(N^2) as update_ssa computes
>> the IDF of the whole function.
>>
>> This is something your patch introduces (it's only rewriting the
>> virtuals, not the incremental SSA update by BB copying).
>
> This too.
>
>> See above. And factor this out into a function. Also move this
>> to the cleanup loop below.
>
> Not moved to the cleanup loop, because there is no easy way to find
> then if a loop has been vectorized or not. But it is in a separate helper
> etc.
>
> The rest should be in this new version of the patch, bootstrapped/regtested
> on x86_64-linux and i686-linux, ok for trunk?
>
> 2013-12-09 Jakub Jelinek <ja...@redhat.com>
>
> * tree-vectorizer.h (struct _loop_vec_info): Add scalar_loop field.
> (LOOP_VINFO_SCALAR_LOOP): Define.
> (slpeel_tree_duplicate_loop_to_edge_cfg): Add scalar_loop argument.
> * config/i386/sse.md (maskload<mode>, maskstore<mode>): New expanders.
> * tree-data-ref.c (get_references_in_stmt): Handle MASK_LOAD and
> MASK_STORE.
> * internal-fn.def (LOOP_VECTORIZED, MASK_LOAD, MASK_STORE): New
> internal fns.
> * tree-if-conv.c: Include expr.h, optabs.h, tree-ssa-loop-ivopts.h and
> tree-ssa-address.h.
> (release_bb_predicate): New function.
> (free_bb_predicate): Use it.
> (reset_bb_predicate): Likewise. Don't unallocate bb->aux
> just to immediately allocate it again.
> (add_to_predicate_list): Add loop argument. If basic blocks that
> dominate loop->latch don't insert any predicate.
> (add_to_dst_predicate_list): Adjust caller.
> (if_convertible_phi_p): Add any_mask_load_store argument, if true,
> handle it like flag_tree_loop_if_convert_stores.
> (insert_gimplified_predicates): Likewise.
> (ifcvt_can_use_mask_load_store): New function.
> (if_convertible_gimple_assign_stmt_p): Add any_mask_load_store
> argument, check if some conditional loads or stores can't be
> converted into MASK_LOAD or MASK_STORE.
> (if_convertible_stmt_p): Add any_mask_load_store argument,
> pass it down to if_convertible_gimple_assign_stmt_p.
> (predicate_bbs): Don't return bool, only check if the last stmt
> of a basic block is GIMPLE_COND and handle that. Adjust
> add_to_predicate_list caller.
> (if_convertible_loop_p_1): Only call predicate_bbs if
> flag_tree_loop_if_convert_stores and free_bb_predicate in that case
> afterwards, check gimple_code of stmts here. Replace is_predicated
> check with dominance check. Add any_mask_load_store argument,
> pass it down to if_convertible_stmt_p and if_convertible_phi_p,
> call if_convertible_phi_p only after all if_convertible_stmt_p
> calls.
> (if_convertible_loop_p): Add any_mask_load_store argument,
> pass it down to if_convertible_loop_p_1.
> (predicate_mem_writes): Emit MASK_LOAD and/or MASK_STORE calls.
> (combine_blocks): Add any_mask_load_store argument, pass
> it down to insert_gimplified_predicates and call predicate_mem_writes
> if it is set. Call predicate_bbs.
> (version_loop_for_if_conversion): New function.
> (tree_if_conversion): Adjust if_convertible_loop_p and combine_blocks
> calls. Return todo flags instead of bool, call
> version_loop_for_if_conversion if if-conversion should be just
> for the vectorized loops and nothing else.
> (main_tree_if_conversion): Adjust caller. Don't call
> tree_if_conversion for dont_vectorize loops if if-conversion
> isn't explicitly enabled.
> * tree-vect-data-refs.c (vect_check_gather): Handle
> MASK_LOAD/MASK_STORE.
> (vect_analyze_data_refs, vect_supportable_dr_alignment): Likewise.
> * gimple.h (gimple_expr_type): Handle MASK_STORE.
> * internal-fn.c (expand_LOOP_VECTORIZED, expand_MASK_LOAD,
> expand_MASK_STORE): New functions.
> * tree-vectorizer.c: Include tree-cfg.h and gimple-fold.h.
> (vect_loop_vectorized_call, fold_loop_vectorized_call): New functions.
> (vectorize_loops): Don't try to vectorize loops with
> loop->dont_vectorize set. Set LOOP_VINFO_SCALAR_LOOP for if-converted
> loops, fold LOOP_VECTORIZED internal call depending on if loop
> has been vectorized or not.
> * tree-vect-loop-manip.c (slpeel_duplicate_current_defs_from_edges):
> New function.
> (slpeel_tree_duplicate_loop_to_edge_cfg): Add scalar_loop argument.
> If non-NULL, copy basic blocks from scalar_loop instead of loop, but
> still to loop's entry or exit edge.
> (slpeel_tree_peel_loop_to_edge): Add scalar_loop argument, pass it
> down to slpeel_tree_duplicate_loop_to_edge_cfg.
> (vect_do_peeling_for_loop_bound, vect_do_peeling_for_loop_alignment):
> Adjust callers.
> (vect_loop_versioning): If LOOP_VINFO_SCALAR_LOOP, perform loop
> versioning from that loop instead of LOOP_VINFO_LOOP, move it to the
> right place in the CFG afterwards.
> * tree-vect-loop.c (vect_determine_vectorization_factor): Handle
> MASK_STORE.
> * cfgloop.h (struct loop): Add dont_vectorize field.
> * tree-loop-distribution.c (copy_loop_before): Adjust
> slpeel_tree_duplicate_loop_to_edge_cfg caller.
> * optabs.def (maskload_optab, maskstore_optab): New optabs.
> * passes.def: Add a note that pass_vectorize must immediately follow
> pass_if_conversion.
> * tree-predcom.c (split_data_refs_to_components): Give up if
> DR_STMT is a call.
> * tree-vect-stmts.c (vect_mark_relevant): Don't crash if lhs
> is NULL.
> (exist_non_indexing_operands_for_use_p): Handle MASK_LOAD
> and MASK_STORE.
> (vectorizable_mask_load_store): New function.
> (vectorizable_call): Call it for MASK_LOAD or MASK_STORE.
> (vect_transform_stmt): Handle MASK_STORE.
> * tree-ssa-phiopt.c (cond_if_else_store_replacement): Ignore
> DR_STMT where lhs is NULL.
> * optabs.h (can_vec_perm_p): Fix up comment typo.
> (can_vec_mask_load_store_p): New prototype.
> * optabs.c (can_vec_mask_load_store_p): New function.
>
This caused:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59523.
H.J.
