[autovect] [patch] vectorize dot-product

[Dorit Naishlos <DORIT at il dot ibm dot com>]

Added recognition and vectorization of the dot-product pattern,
along with a new tree-code - DOT_PROD_EXPR - and optabs.
Also added tree-codes for WIDEN_MULT_EXPR and MULT_HI_EXPR, but haven't yet
implemented their vectorization.

Bootstrapped with vectorization enabled on powerpc-linux, tested on the
vectorizer testcases. Committed to the branch.

dorit

        * expr.c (expand_expr_real_1): Add DOT_PROD_EXPR case. Tentatively
add
         cases for WIDEN_MULT_EXPR and MULT_HI_EXPR.
        * genopinit.c (sdot_prod_optab, udot_prod_optab): Initialize new
optabs.
        * optabs.c (optab_for_tree_code): Add cases for DOT_PROD_EXPR,
        WIDEN_MULT_EXPR and MULT_HI_EXPR.
        (expand_widen_pattern_expr): Extend to handled widening patterns
with
        more than two arguments.
        (init_optabs): Initalize new optabs sdot_prod_optab,
udot_prod_optab.
        * optabs.h (OTI_sdot_prod, OTI_udot_prod): New optabs indexes.
        (expand_widen_pattern_expr): Add argument to function declaration,
        * tree-inline.c (estimate_num_insns_1): Add cases for
DOT_PROD_EXPR,
        WIDEN_MULT_EXPR and MULT_HI_EXPR.
        * tree-pretty-print.c (dump_generic_node, op_prio): Add cases for
        DOT_PROD_EXPR, WIDEN_MULT_EXPR and MULT_HI_EXPR.
        (op_symbol): Added cases for WIDEN_MULT_EXPR and MULT_HI_EXPR.
        * tree-ssa-operands (get_expr_operands): Add case for
DOT_PROD_EXPR.
        * tree.def (WIDEN_MULT_EXPR, MULT_HI_EXPR, DOT_PROD_EXPR): New
        tree-codes.

        * tree-vect-analyze.c (widened_name_p): New function.
        (vect_pattern_recog_funcs): Add new pattern recognition functions
to
        the initialization.
        (vect_recog_mult_hi_pattern): New function (not yet implemented).
        (vect_recog_dot_prod_pattern): New function.
        (vect_recog_widen_mult_pattern): New function.
        (vect_recog_widen_sum_pattern): Call widened_name_p. Fix type
        comparison. Do not consider stmts that have already been recognized
        as part of a pattern.
        (vect_recog_unsigned_subsat_pattern): Do not consider stmts that
have
        already been recognized as part of a pattern.
        * tree-vect-transform.c (get_initial_def_for_reduction): Add case
for
        DOT_PROD_EXPR.
        (vect_create_epilog_for_reduction): Add assert.
        (vectorizable_reduction): Support reductions with more than two
        arguments.
        * tree-vectorizer.h (ternary_op): New operation_type.
        (NUM_PATTERNS): Set to 5.
        (vect_recog_widen_mult_pattern): New function declaration.
        (vect_recog_mult_hi_pattern): New function declaration.
        (vect_recog_dot_prod_pattern): New function declaration.

        * config/rs6000/altivec.md (altivec_vmsumshm, altivec_vmsumuhs):
        (altivec_vmsumshs, altivec_vmsumu<VI_char>m,
altivec_vmsumm<VI_char>m):
        (widen_usum<mode>3, widen_ssumv16qi3, widen_ssumv8hi3): Expose the
PLUS.
        (udot_prod<mode>, sdot_prodv16qi, sdot_prodv8hi): New patterns.

(See attached file: sept23.dot.txt)

Attachment: sept23.dot.txt
Description: Text document