On Thu, Jun 28, 2012 at 04:05:58PM +0200, Jakub Jelinek wrote:
>On Thu, Jun 28, 2012@09:17:55AM +0200, Jakub Jelinek wrote:
>> I'll look@using MULT_HIGHPART_EXPR in the pattern recognizer and
>> vectorizing it as either of the sequences next.
>
>And here is corresponding pattern recognizer and vectorizer patch.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
>Unfortunately the addition of the builtin_mul_widen_* hooks on i?86 seems
>to pessimize the generated code for gcc.dg/vect/pr51581-3.c
>testcase (at least with -O3 -mavx) compared to when the hooks aren't
>present, because i?86 has more natural support for widen mult lo/hi
>compoared to widen mult even/odd, but I assume that on powerpc it is the
>other way around. So, how should I find out if both VEC_WIDEN_MULT_*_EXPR
>and builtin_mul_widen_* are possible for the particular vectype which one
>will be cheaper?
>
>2012-06-28 Jakub Jelinek <ja...@redhat.com>
>
> PR tree-optimization/51581
> * tree-vect-stmts.c (permute_vec_elements): Add forward decl.
> (vectorizable_operation): Handle vectorization of MULT_HIGHPART_EXPR
> also using VEC_WIDEN_MULT_*_EXPR or builtin_mul_widen_* plus
> VEC_PERM_EXPR if vector MULT_HIGHPART_EXPR isn't supported.
> * tree-vect-patterns.c (vect_recog_divmod_pattern): Use
> MULT_HIGHPART_EXPR instead of VEC_WIDEN_MULT_*_EXPR and shifts.
>
> * gcc.dg/vect/pr51581-4.c: New test.
>
>--- gcc/tree-vect-stmts.c.jj 2012-06-26 11:38:28.000000000 +0200
>+++ gcc/tree-vect-stmts.c 2012-06-28 13:27:50.475158271 +0200
>@@ -3300,17 +3304,18 @@ static bool
>+ icode = optab ? (int) optab_handler (optab, vec_mode) : CODE_FOR_nothing;
>+
>+ if (icode == CODE_FOR_nothing
>+ && code == MULT_HIGHPART_EXPR
>+ && VECTOR_MODE_P (vec_mode)
>+ && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN)
>+ {
>+ /* If MULT_HIGHPART_EXPR isn't supported by the backend, see
>+ if we can emit VEC_WIDEN_MULT_{LO,HI}_EXPR followed by VEC_PERM_EXPR
>+ or builtin_mul_widen_{even,odd} followed by VEC_PERM_EXPR. */
>+ unsigned int prec = TYPE_PRECISION (TREE_TYPE (scalar_dest));
>+ unsigned int unsignedp = TYPE_UNSIGNED (TREE_TYPE (scalar_dest));
>+ tree wide_type
>+ = build_nonstandard_integer_type (prec * 2, unsignedp);
>+ wide_vectype
>+ = get_same_sized_vectype (wide_type, vectype);
>+
>+ sel = XALLOCAVEC (unsigned char, nunits_in);
>+ if (VECTOR_MODE_P (TYPE_MODE (wide_vectype))
>+ && GET_MODE_SIZE (TYPE_MODE (wide_vectype))
>+ == GET_MODE_SIZE (vec_mode))
>+ {
>+ if (targetm.vectorize.builtin_mul_widen_even
>+ && (decl1 = targetm.vectorize.builtin_mul_widen_even (vectype))
>+ && targetm.vectorize.builtin_mul_widen_odd
>+ && (decl2 = targetm.vectorize.builtin_mul_widen_odd (vectype))
>+ && TYPE_MODE (TREE_TYPE (TREE_TYPE (decl1)))
>+ == TYPE_MODE (wide_vectype))
>+ {
>+ for (i = 0; i < nunits_in; i++)
>+ sel[i] = !BYTES_BIG_ENDIAN + (i & ~1)
>+ + ((i & 1) ? nunits_in : 0);
>+ if (0 && can_vec_perm_p (vec_mode, false, sel))
>+ icode = 0;
>+ }
>+ if (icode == CODE_FOR_nothing)
>+ {
>+ decl1 = NULL_TREE;
>+ decl2 = NULL_TREE;
>+ optab = optab_for_tree_code (VEC_WIDEN_MULT_HI_EXPR,
>+ vectype, optab_default);
>+ optab2 = optab_for_tree_code (VEC_WIDEN_MULT_HI_EXPR,
>+ vectype, optab_default);
Really both HI? If so optab2 could be removed from that fn altogether..
>+ if (optab != NULL
>+ && optab2 != NULL
>+ && optab_handler (optab, vec_mode) != CODE_FOR_nothing
>+ && optab_handler (optab2, vec_mode) != CODE_FOR_nothing)
>+ {
>+ for (i = 0; i < nunits_in; i++)
>+ sel[i] = !BYTES_BIG_ENDIAN + 2 * i;
>+ if (can_vec_perm_p (vec_mode, false, sel))
>+ icode = optab_handler (optab, vec_mode);
>+ }
>+ }
>+ }
>+ if (icode == CODE_FOR_nothing)
>+ {
>+ if (optab_for_tree_code (code, vectype, optab_default) == NULL)
>+ {
>+ if (vect_print_dump_info (REPORT_DETAILS))
>+ fprintf (vect_dump, "no optab.");
>+ return false;
>+ }
>+ wide_vectype = NULL_TREE;
>+ optab2 = NULL;
>+ }
>+ }
>+
