1 /* Analysis Utilities for Loop Vectorization.
2 Copyright (C) 2006-2013 Free Software Foundation, Inc.
3 Contributed by Dorit Nuzman <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
27 #include "stor-layout.h"
29 #include "basic-block.h"
30 #include "gimple-pretty-print.h"
33 #include "gimple-iterator.h"
34 #include "gimple-ssa.h"
35 #include "tree-phinodes.h"
36 #include "ssa-iterators.h"
37 #include "stringpool.h"
38 #include "tree-ssanames.h"
43 #include "tree-data-ref.h"
44 #include "tree-vectorizer.h"
45 #include "recog.h" /* FIXME: for insn_data */
46 #include "diagnostic-core.h"
49 /* Pattern recognition functions */
50 static gimple
vect_recog_widen_sum_pattern (vec
<gimple
> *, tree
*,
52 static gimple
vect_recog_widen_mult_pattern (vec
<gimple
> *, tree
*,
54 static gimple
vect_recog_dot_prod_pattern (vec
<gimple
> *, tree
*,
56 static gimple
vect_recog_pow_pattern (vec
<gimple
> *, tree
*, tree
*);
57 static gimple
vect_recog_over_widening_pattern (vec
<gimple
> *, tree
*,
59 static gimple
vect_recog_widen_shift_pattern (vec
<gimple
> *,
61 static gimple
vect_recog_rotate_pattern (vec
<gimple
> *, tree
*, tree
*);
62 static gimple
vect_recog_vector_vector_shift_pattern (vec
<gimple
> *,
64 static gimple
vect_recog_divmod_pattern (vec
<gimple
> *,
66 static gimple
vect_recog_mixed_size_cond_pattern (vec
<gimple
> *,
68 static gimple
vect_recog_bool_pattern (vec
<gimple
> *, tree
*, tree
*);
69 static vect_recog_func_ptr vect_vect_recog_func_ptrs
[NUM_PATTERNS
] = {
70 vect_recog_widen_mult_pattern
,
71 vect_recog_widen_sum_pattern
,
72 vect_recog_dot_prod_pattern
,
73 vect_recog_pow_pattern
,
74 vect_recog_widen_shift_pattern
,
75 vect_recog_over_widening_pattern
,
76 vect_recog_rotate_pattern
,
77 vect_recog_vector_vector_shift_pattern
,
78 vect_recog_divmod_pattern
,
79 vect_recog_mixed_size_cond_pattern
,
80 vect_recog_bool_pattern
};
83 append_pattern_def_seq (stmt_vec_info stmt_info
, gimple stmt
)
85 gimple_seq_add_stmt_without_update (&STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
),
90 new_pattern_def_seq (stmt_vec_info stmt_info
, gimple stmt
)
92 STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
) = NULL
;
93 append_pattern_def_seq (stmt_info
, stmt
);
96 /* Check whether STMT2 is in the same loop or basic block as STMT1.
97 Which of the two applies depends on whether we're currently doing
98 loop-based or basic-block-based vectorization, as determined by
99 the vinfo_for_stmt for STMT1 (which must be defined).
101 If this returns true, vinfo_for_stmt for STMT2 is guaranteed
102 to be defined as well. */
105 vect_same_loop_or_bb_p (gimple stmt1
, gimple stmt2
)
107 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt1
);
108 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
109 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
111 if (!gimple_bb (stmt2
))
116 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
117 if (!flow_bb_inside_loop_p (loop
, gimple_bb (stmt2
)))
122 if (gimple_bb (stmt2
) != BB_VINFO_BB (bb_vinfo
)
123 || gimple_code (stmt2
) == GIMPLE_PHI
)
127 gcc_assert (vinfo_for_stmt (stmt2
));
131 /* If the LHS of DEF_STMT has a single use, and that statement is
132 in the same loop or basic block, return it. */
135 vect_single_imm_use (gimple def_stmt
)
137 tree lhs
= gimple_assign_lhs (def_stmt
);
141 if (!single_imm_use (lhs
, &use_p
, &use_stmt
))
144 if (!vect_same_loop_or_bb_p (def_stmt
, use_stmt
))
150 /* Check whether NAME, an ssa-name used in USE_STMT,
151 is a result of a type promotion or demotion, such that:
152 DEF_STMT: NAME = NOP (name0)
153 where the type of name0 (ORIG_TYPE) is smaller/bigger than the type of NAME.
154 If CHECK_SIGN is TRUE, check that either both types are signed or both are
158 type_conversion_p (tree name
, gimple use_stmt
, bool check_sign
,
159 tree
*orig_type
, gimple
*def_stmt
, bool *promotion
)
163 loop_vec_info loop_vinfo
;
164 stmt_vec_info stmt_vinfo
;
165 tree type
= TREE_TYPE (name
);
167 enum vect_def_type dt
;
169 bb_vec_info bb_vinfo
;
171 stmt_vinfo
= vinfo_for_stmt (use_stmt
);
172 loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
173 bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
174 if (!vect_is_simple_use (name
, use_stmt
, loop_vinfo
, bb_vinfo
, def_stmt
,
178 if (dt
!= vect_internal_def
179 && dt
!= vect_external_def
&& dt
!= vect_constant_def
)
185 if (!is_gimple_assign (*def_stmt
))
188 if (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (*def_stmt
)))
191 oprnd0
= gimple_assign_rhs1 (*def_stmt
);
193 *orig_type
= TREE_TYPE (oprnd0
);
194 if (!INTEGRAL_TYPE_P (type
) || !INTEGRAL_TYPE_P (*orig_type
)
195 || ((TYPE_UNSIGNED (type
) != TYPE_UNSIGNED (*orig_type
)) && check_sign
))
198 if (TYPE_PRECISION (type
) >= (TYPE_PRECISION (*orig_type
) * 2))
200 else if (TYPE_PRECISION (*orig_type
) >= (TYPE_PRECISION (type
) * 2))
205 if (!vect_is_simple_use (oprnd0
, *def_stmt
, loop_vinfo
,
206 bb_vinfo
, &dummy_gimple
, &dummy
, &dt
))
212 /* Helper to return a new temporary for pattern of TYPE for STMT. If STMT
213 is NULL, the caller must set SSA_NAME_DEF_STMT for the returned SSA var. */
216 vect_recog_temp_ssa_var (tree type
, gimple stmt
)
218 return make_temp_ssa_name (type
, stmt
, "patt");
221 /* Function vect_recog_dot_prod_pattern
223 Try to find the following pattern:
229 sum_0 = phi <init, sum_1>
232 S3 x_T = (TYPE1) x_t;
233 S4 y_T = (TYPE1) y_t;
235 [S6 prod = (TYPE2) prod; #optional]
236 S7 sum_1 = prod + sum_0;
238 where 'TYPE1' is exactly double the size of type 'type', and 'TYPE2' is the
239 same size of 'TYPE1' or bigger. This is a special case of a reduction
244 * STMTS: Contains a stmt from which the pattern search begins. In the
245 example, when this function is called with S7, the pattern {S3,S4,S5,S6,S7}
250 * TYPE_IN: The type of the input arguments to the pattern.
252 * TYPE_OUT: The type of the output of this pattern.
254 * Return value: A new stmt that will be used to replace the sequence of
255 stmts that constitute the pattern. In this case it will be:
256 WIDEN_DOT_PRODUCT <x_t, y_t, sum_0>
258 Note: The dot-prod idiom is a widening reduction pattern that is
259 vectorized without preserving all the intermediate results. It
260 produces only N/2 (widened) results (by summing up pairs of
261 intermediate results) rather than all N results. Therefore, we
262 cannot allow this pattern when we want to get all the results and in
263 the correct order (as is the case when this computation is in an
264 inner-loop nested in an outer-loop that us being vectorized). */
267 vect_recog_dot_prod_pattern (vec
<gimple
> *stmts
, tree
*type_in
,
270 gimple stmt
, last_stmt
= (*stmts
)[0];
272 tree oprnd00
, oprnd01
;
273 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
274 tree type
, half_type
;
277 loop_vec_info loop_info
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
285 loop
= LOOP_VINFO_LOOP (loop_info
);
287 if (!is_gimple_assign (last_stmt
))
290 type
= gimple_expr_type (last_stmt
);
292 /* Look for the following pattern
296 DDPROD = (TYPE2) DPROD;
297 sum_1 = DDPROD + sum_0;
299 - DX is double the size of X
300 - DY is double the size of Y
301 - DX, DY, DPROD all have the same type
302 - sum is the same size of DPROD or bigger
303 - sum has been recognized as a reduction variable.
305 This is equivalent to:
306 DPROD = X w* Y; #widen mult
307 sum_1 = DPROD w+ sum_0; #widen summation
309 DPROD = X w* Y; #widen mult
310 sum_1 = DPROD + sum_0; #summation
313 /* Starting from LAST_STMT, follow the defs of its uses in search
314 of the above pattern. */
316 if (gimple_assign_rhs_code (last_stmt
) != PLUS_EXPR
)
319 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
321 /* Has been detected as widening-summation? */
323 stmt
= STMT_VINFO_RELATED_STMT (stmt_vinfo
);
324 type
= gimple_expr_type (stmt
);
325 if (gimple_assign_rhs_code (stmt
) != WIDEN_SUM_EXPR
)
327 oprnd0
= gimple_assign_rhs1 (stmt
);
328 oprnd1
= gimple_assign_rhs2 (stmt
);
329 half_type
= TREE_TYPE (oprnd0
);
335 if (STMT_VINFO_DEF_TYPE (stmt_vinfo
) != vect_reduction_def
)
337 oprnd0
= gimple_assign_rhs1 (last_stmt
);
338 oprnd1
= gimple_assign_rhs2 (last_stmt
);
339 if (!types_compatible_p (TREE_TYPE (oprnd0
), type
)
340 || !types_compatible_p (TREE_TYPE (oprnd1
), type
))
344 if (type_conversion_p (oprnd0
, stmt
, true, &half_type
, &def_stmt
,
349 oprnd0
= gimple_assign_rhs1 (stmt
);
355 /* So far so good. Since last_stmt was detected as a (summation) reduction,
356 we know that oprnd1 is the reduction variable (defined by a loop-header
357 phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
358 Left to check that oprnd0 is defined by a (widen_)mult_expr */
359 if (TREE_CODE (oprnd0
) != SSA_NAME
)
362 prod_type
= half_type
;
363 stmt
= SSA_NAME_DEF_STMT (oprnd0
);
365 /* It could not be the dot_prod pattern if the stmt is outside the loop. */
366 if (!gimple_bb (stmt
) || !flow_bb_inside_loop_p (loop
, gimple_bb (stmt
)))
369 /* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
370 inside the loop (in case we are analyzing an outer-loop). */
371 if (!is_gimple_assign (stmt
))
373 stmt_vinfo
= vinfo_for_stmt (stmt
);
374 gcc_assert (stmt_vinfo
);
375 if (STMT_VINFO_DEF_TYPE (stmt_vinfo
) != vect_internal_def
)
377 if (gimple_assign_rhs_code (stmt
) != MULT_EXPR
)
379 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
381 /* Has been detected as a widening multiplication? */
383 stmt
= STMT_VINFO_RELATED_STMT (stmt_vinfo
);
384 if (gimple_assign_rhs_code (stmt
) != WIDEN_MULT_EXPR
)
386 stmt_vinfo
= vinfo_for_stmt (stmt
);
387 gcc_assert (stmt_vinfo
);
388 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_internal_def
);
389 oprnd00
= gimple_assign_rhs1 (stmt
);
390 oprnd01
= gimple_assign_rhs2 (stmt
);
394 tree half_type0
, half_type1
;
398 oprnd0
= gimple_assign_rhs1 (stmt
);
399 oprnd1
= gimple_assign_rhs2 (stmt
);
400 if (!types_compatible_p (TREE_TYPE (oprnd0
), prod_type
)
401 || !types_compatible_p (TREE_TYPE (oprnd1
), prod_type
))
403 if (!type_conversion_p (oprnd0
, stmt
, true, &half_type0
, &def_stmt
,
407 oprnd00
= gimple_assign_rhs1 (def_stmt
);
408 if (!type_conversion_p (oprnd1
, stmt
, true, &half_type1
, &def_stmt
,
412 oprnd01
= gimple_assign_rhs1 (def_stmt
);
413 if (!types_compatible_p (half_type0
, half_type1
))
415 if (TYPE_PRECISION (prod_type
) != TYPE_PRECISION (half_type0
) * 2)
419 half_type
= TREE_TYPE (oprnd00
);
420 *type_in
= half_type
;
423 /* Pattern detected. Create a stmt to be used to replace the pattern: */
424 var
= vect_recog_temp_ssa_var (type
, NULL
);
425 pattern_stmt
= gimple_build_assign_with_ops (DOT_PROD_EXPR
, var
,
426 oprnd00
, oprnd01
, oprnd1
);
428 if (dump_enabled_p ())
430 dump_printf_loc (MSG_NOTE
, vect_location
,
431 "vect_recog_dot_prod_pattern: detected: ");
432 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
433 dump_printf (MSG_NOTE
, "\n");
436 /* We don't allow changing the order of the computation in the inner-loop
437 when doing outer-loop vectorization. */
438 gcc_assert (!nested_in_vect_loop_p (loop
, last_stmt
));
444 /* Handle widening operation by a constant. At the moment we support MULT_EXPR
447 For MULT_EXPR we check that CONST_OPRND fits HALF_TYPE, and for LSHIFT_EXPR
448 we check that CONST_OPRND is less or equal to the size of HALF_TYPE.
450 Otherwise, if the type of the result (TYPE) is at least 4 times bigger than
451 HALF_TYPE, and there is an intermediate type (2 times smaller than TYPE)
452 that satisfies the above restrictions, we can perform a widening opeartion
453 from the intermediate type to TYPE and replace a_T = (TYPE) a_t;
454 with a_it = (interm_type) a_t; */
457 vect_handle_widen_op_by_const (gimple stmt
, enum tree_code code
,
458 tree const_oprnd
, tree
*oprnd
,
459 vec
<gimple
> *stmts
, tree type
,
460 tree
*half_type
, gimple def_stmt
)
462 tree new_type
, new_oprnd
;
465 if (code
!= MULT_EXPR
&& code
!= LSHIFT_EXPR
)
468 if (((code
== MULT_EXPR
&& int_fits_type_p (const_oprnd
, *half_type
))
469 || (code
== LSHIFT_EXPR
470 && compare_tree_int (const_oprnd
, TYPE_PRECISION (*half_type
))
472 && TYPE_PRECISION (type
) == (TYPE_PRECISION (*half_type
) * 2))
474 /* CONST_OPRND is a constant of HALF_TYPE. */
475 *oprnd
= gimple_assign_rhs1 (def_stmt
);
479 if (TYPE_PRECISION (type
) < (TYPE_PRECISION (*half_type
) * 4))
482 if (!vect_same_loop_or_bb_p (stmt
, def_stmt
))
485 /* TYPE is 4 times bigger than HALF_TYPE, try widening operation for
486 a type 2 times bigger than HALF_TYPE. */
487 new_type
= build_nonstandard_integer_type (TYPE_PRECISION (type
) / 2,
488 TYPE_UNSIGNED (type
));
489 if ((code
== MULT_EXPR
&& !int_fits_type_p (const_oprnd
, new_type
))
490 || (code
== LSHIFT_EXPR
491 && compare_tree_int (const_oprnd
, TYPE_PRECISION (new_type
)) == 1))
494 /* Use NEW_TYPE for widening operation. */
495 if (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
)))
497 new_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
));
498 /* Check if the already created pattern stmt is what we need. */
499 if (!is_gimple_assign (new_stmt
)
500 || gimple_assign_rhs_code (new_stmt
) != NOP_EXPR
501 || TREE_TYPE (gimple_assign_lhs (new_stmt
)) != new_type
)
504 stmts
->safe_push (def_stmt
);
505 *oprnd
= gimple_assign_lhs (new_stmt
);
509 /* Create a_T = (NEW_TYPE) a_t; */
510 *oprnd
= gimple_assign_rhs1 (def_stmt
);
511 new_oprnd
= make_ssa_name (new_type
, NULL
);
512 new_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, new_oprnd
, *oprnd
,
514 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
)) = new_stmt
;
515 stmts
->safe_push (def_stmt
);
519 *half_type
= new_type
;
524 /* Function vect_recog_widen_mult_pattern
526 Try to find the following pattern:
529 TYPE a_T, b_T, prod_T;
535 S5 prod_T = a_T * b_T;
537 where type 'TYPE' is at least double the size of type 'type'.
539 Also detect unsigned cases:
541 unsigned type a_t, b_t;
542 unsigned TYPE u_prod_T;
543 TYPE a_T, b_T, prod_T;
549 S5 prod_T = a_T * b_T;
550 S6 u_prod_T = (unsigned TYPE) prod_T;
552 and multiplication by constants:
559 S5 prod_T = a_T * CONST;
561 A special case of multiplication by constants is when 'TYPE' is 4 times
562 bigger than 'type', but CONST fits an intermediate type 2 times smaller
563 than 'TYPE'. In that case we create an additional pattern stmt for S3
564 to create a variable of the intermediate type, and perform widen-mult
565 on the intermediate type as well:
569 TYPE a_T, prod_T, prod_T';
573 '--> a_it = (interm_type) a_t;
574 S5 prod_T = a_T * CONST;
575 '--> prod_T' = a_it w* CONST;
579 * STMTS: Contains a stmt from which the pattern search begins. In the
580 example, when this function is called with S5, the pattern {S3,S4,S5,(S6)}
581 is detected. In case of unsigned widen-mult, the original stmt (S5) is
582 replaced with S6 in STMTS. In case of multiplication by a constant
583 of an intermediate type (the last case above), STMTS also contains S3
584 (inserted before S5).
588 * TYPE_IN: The type of the input arguments to the pattern.
590 * TYPE_OUT: The type of the output of this pattern.
592 * Return value: A new stmt that will be used to replace the sequence of
593 stmts that constitute the pattern. In this case it will be:
594 WIDEN_MULT <a_t, b_t>
598 vect_recog_widen_mult_pattern (vec
<gimple
> *stmts
,
599 tree
*type_in
, tree
*type_out
)
601 gimple last_stmt
= stmts
->pop ();
602 gimple def_stmt0
, def_stmt1
;
604 tree type
, half_type0
, half_type1
;
606 tree vectype
, vectype_out
= NULL_TREE
;
608 enum tree_code dummy_code
;
614 if (!is_gimple_assign (last_stmt
))
617 type
= gimple_expr_type (last_stmt
);
619 /* Starting from LAST_STMT, follow the defs of its uses in search
620 of the above pattern. */
622 if (gimple_assign_rhs_code (last_stmt
) != MULT_EXPR
)
625 oprnd0
= gimple_assign_rhs1 (last_stmt
);
626 oprnd1
= gimple_assign_rhs2 (last_stmt
);
627 if (!types_compatible_p (TREE_TYPE (oprnd0
), type
)
628 || !types_compatible_p (TREE_TYPE (oprnd1
), type
))
631 /* Check argument 0. */
632 if (!type_conversion_p (oprnd0
, last_stmt
, false, &half_type0
, &def_stmt0
,
636 /* Check argument 1. */
637 op1_ok
= type_conversion_p (oprnd1
, last_stmt
, false, &half_type1
,
638 &def_stmt1
, &promotion
);
640 if (op1_ok
&& promotion
)
642 oprnd0
= gimple_assign_rhs1 (def_stmt0
);
643 oprnd1
= gimple_assign_rhs1 (def_stmt1
);
647 if (TREE_CODE (oprnd1
) == INTEGER_CST
648 && TREE_CODE (half_type0
) == INTEGER_TYPE
649 && vect_handle_widen_op_by_const (last_stmt
, MULT_EXPR
, oprnd1
,
650 &oprnd0
, stmts
, type
,
651 &half_type0
, def_stmt0
))
653 half_type1
= half_type0
;
654 oprnd1
= fold_convert (half_type1
, oprnd1
);
660 /* Handle unsigned case. Look for
661 S6 u_prod_T = (unsigned TYPE) prod_T;
662 Use unsigned TYPE as the type for WIDEN_MULT_EXPR. */
663 if (TYPE_UNSIGNED (type
) != TYPE_UNSIGNED (half_type0
))
669 if (TYPE_UNSIGNED (type
) == TYPE_UNSIGNED (half_type1
))
672 use_stmt
= vect_single_imm_use (last_stmt
);
673 if (!use_stmt
|| !is_gimple_assign (use_stmt
)
674 || gimple_assign_rhs_code (use_stmt
) != NOP_EXPR
)
677 use_lhs
= gimple_assign_lhs (use_stmt
);
678 use_type
= TREE_TYPE (use_lhs
);
679 if (!INTEGRAL_TYPE_P (use_type
)
680 || (TYPE_UNSIGNED (type
) == TYPE_UNSIGNED (use_type
))
681 || (TYPE_PRECISION (type
) != TYPE_PRECISION (use_type
)))
685 last_stmt
= use_stmt
;
688 if (!types_compatible_p (half_type0
, half_type1
))
691 /* Pattern detected. */
692 if (dump_enabled_p ())
693 dump_printf_loc (MSG_NOTE
, vect_location
,
694 "vect_recog_widen_mult_pattern: detected:\n");
696 /* Check target support */
697 vectype
= get_vectype_for_scalar_type (half_type0
);
698 vectype_out
= get_vectype_for_scalar_type (type
);
701 || !supportable_widening_operation (WIDEN_MULT_EXPR
, last_stmt
,
702 vectype_out
, vectype
,
703 &dummy_code
, &dummy_code
,
704 &dummy_int
, &dummy_vec
))
708 *type_out
= vectype_out
;
710 /* Pattern supported. Create a stmt to be used to replace the pattern: */
711 var
= vect_recog_temp_ssa_var (type
, NULL
);
712 pattern_stmt
= gimple_build_assign_with_ops (WIDEN_MULT_EXPR
, var
, oprnd0
,
715 if (dump_enabled_p ())
716 dump_gimple_stmt_loc (MSG_NOTE
, vect_location
, TDF_SLIM
, pattern_stmt
, 0);
718 stmts
->safe_push (last_stmt
);
723 /* Function vect_recog_pow_pattern
725 Try to find the following pattern:
729 with POW being one of pow, powf, powi, powif and N being
734 * LAST_STMT: A stmt from which the pattern search begins.
738 * TYPE_IN: The type of the input arguments to the pattern.
740 * TYPE_OUT: The type of the output of this pattern.
742 * Return value: A new stmt that will be used to replace the sequence of
743 stmts that constitute the pattern. In this case it will be:
750 vect_recog_pow_pattern (vec
<gimple
> *stmts
, tree
*type_in
,
753 gimple last_stmt
= (*stmts
)[0];
754 tree fn
, base
, exp
= NULL
;
758 if (!is_gimple_call (last_stmt
) || gimple_call_lhs (last_stmt
) == NULL
)
761 fn
= gimple_call_fndecl (last_stmt
);
762 if (fn
== NULL_TREE
|| DECL_BUILT_IN_CLASS (fn
) != BUILT_IN_NORMAL
)
765 switch (DECL_FUNCTION_CODE (fn
))
771 base
= gimple_call_arg (last_stmt
, 0);
772 exp
= gimple_call_arg (last_stmt
, 1);
773 if (TREE_CODE (exp
) != REAL_CST
774 && TREE_CODE (exp
) != INTEGER_CST
)
782 /* We now have a pow or powi builtin function call with a constant
785 *type_out
= NULL_TREE
;
787 /* Catch squaring. */
788 if ((tree_fits_shwi_p (exp
)
789 && tree_to_shwi (exp
) == 2)
790 || (TREE_CODE (exp
) == REAL_CST
791 && REAL_VALUES_EQUAL (TREE_REAL_CST (exp
), dconst2
)))
793 *type_in
= TREE_TYPE (base
);
795 var
= vect_recog_temp_ssa_var (TREE_TYPE (base
), NULL
);
796 stmt
= gimple_build_assign_with_ops (MULT_EXPR
, var
, base
, base
);
800 /* Catch square root. */
801 if (TREE_CODE (exp
) == REAL_CST
802 && REAL_VALUES_EQUAL (TREE_REAL_CST (exp
), dconsthalf
))
804 tree newfn
= mathfn_built_in (TREE_TYPE (base
), BUILT_IN_SQRT
);
805 *type_in
= get_vectype_for_scalar_type (TREE_TYPE (base
));
808 gimple stmt
= gimple_build_call (newfn
, 1, base
);
809 if (vectorizable_function (stmt
, *type_in
, *type_in
)
812 var
= vect_recog_temp_ssa_var (TREE_TYPE (base
), stmt
);
813 gimple_call_set_lhs (stmt
, var
);
823 /* Function vect_recog_widen_sum_pattern
825 Try to find the following pattern:
828 TYPE x_T, sum = init;
830 sum_0 = phi <init, sum_1>
833 S3 sum_1 = x_T + sum_0;
835 where type 'TYPE' is at least double the size of type 'type', i.e - we're
836 summing elements of type 'type' into an accumulator of type 'TYPE'. This is
837 a special case of a reduction computation.
841 * LAST_STMT: A stmt from which the pattern search begins. In the example,
842 when this function is called with S3, the pattern {S2,S3} will be detected.
846 * TYPE_IN: The type of the input arguments to the pattern.
848 * TYPE_OUT: The type of the output of this pattern.
850 * Return value: A new stmt that will be used to replace the sequence of
851 stmts that constitute the pattern. In this case it will be:
852 WIDEN_SUM <x_t, sum_0>
854 Note: The widening-sum idiom is a widening reduction pattern that is
855 vectorized without preserving all the intermediate results. It
856 produces only N/2 (widened) results (by summing up pairs of
857 intermediate results) rather than all N results. Therefore, we
858 cannot allow this pattern when we want to get all the results and in
859 the correct order (as is the case when this computation is in an
860 inner-loop nested in an outer-loop that us being vectorized). */
863 vect_recog_widen_sum_pattern (vec
<gimple
> *stmts
, tree
*type_in
,
866 gimple stmt
, last_stmt
= (*stmts
)[0];
868 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
869 tree type
, half_type
;
871 loop_vec_info loop_info
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
879 loop
= LOOP_VINFO_LOOP (loop_info
);
881 if (!is_gimple_assign (last_stmt
))
884 type
= gimple_expr_type (last_stmt
);
886 /* Look for the following pattern
889 In which DX is at least double the size of X, and sum_1 has been
890 recognized as a reduction variable.
893 /* Starting from LAST_STMT, follow the defs of its uses in search
894 of the above pattern. */
896 if (gimple_assign_rhs_code (last_stmt
) != PLUS_EXPR
)
899 if (STMT_VINFO_DEF_TYPE (stmt_vinfo
) != vect_reduction_def
)
902 oprnd0
= gimple_assign_rhs1 (last_stmt
);
903 oprnd1
= gimple_assign_rhs2 (last_stmt
);
904 if (!types_compatible_p (TREE_TYPE (oprnd0
), type
)
905 || !types_compatible_p (TREE_TYPE (oprnd1
), type
))
908 /* So far so good. Since last_stmt was detected as a (summation) reduction,
909 we know that oprnd1 is the reduction variable (defined by a loop-header
910 phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
911 Left to check that oprnd0 is defined by a cast from type 'type' to type
914 if (!type_conversion_p (oprnd0
, last_stmt
, true, &half_type
, &stmt
,
919 oprnd0
= gimple_assign_rhs1 (stmt
);
920 *type_in
= half_type
;
923 /* Pattern detected. Create a stmt to be used to replace the pattern: */
924 var
= vect_recog_temp_ssa_var (type
, NULL
);
925 pattern_stmt
= gimple_build_assign_with_ops (WIDEN_SUM_EXPR
, var
,
928 if (dump_enabled_p ())
930 dump_printf_loc (MSG_NOTE
, vect_location
,
931 "vect_recog_widen_sum_pattern: detected: ");
932 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
933 dump_printf (MSG_NOTE
, "\n");
936 /* We don't allow changing the order of the computation in the inner-loop
937 when doing outer-loop vectorization. */
938 gcc_assert (!nested_in_vect_loop_p (loop
, last_stmt
));
944 /* Return TRUE if the operation in STMT can be performed on a smaller type.
947 STMT - a statement to check.
948 DEF - we support operations with two operands, one of which is constant.
949 The other operand can be defined by a demotion operation, or by a
950 previous statement in a sequence of over-promoted operations. In the
951 later case DEF is used to replace that operand. (It is defined by a
952 pattern statement we created for the previous statement in the
956 NEW_TYPE - Output: a smaller type that we are trying to use. Input: if not
957 NULL, it's the type of DEF.
958 STMTS - additional pattern statements. If a pattern statement (type
959 conversion) is created in this function, its original statement is
963 OP0, OP1 - if the operation fits a smaller type, OP0 and OP1 are the new
964 operands to use in the new pattern statement for STMT (will be created
965 in vect_recog_over_widening_pattern ()).
966 NEW_DEF_STMT - in case DEF has to be promoted, we create two pattern
967 statements for STMT: the first one is a type promotion and the second
968 one is the operation itself. We return the type promotion statement
969 in NEW_DEF_STMT and further store it in STMT_VINFO_PATTERN_DEF_SEQ of
970 the second pattern statement. */
973 vect_operation_fits_smaller_type (gimple stmt
, tree def
, tree
*new_type
,
974 tree
*op0
, tree
*op1
, gimple
*new_def_stmt
,
978 tree const_oprnd
, oprnd
;
979 tree interm_type
= NULL_TREE
, half_type
, new_oprnd
, type
;
980 gimple def_stmt
, new_stmt
;
986 *new_def_stmt
= NULL
;
988 if (!is_gimple_assign (stmt
))
991 code
= gimple_assign_rhs_code (stmt
);
992 if (code
!= LSHIFT_EXPR
&& code
!= RSHIFT_EXPR
993 && code
!= BIT_IOR_EXPR
&& code
!= BIT_XOR_EXPR
&& code
!= BIT_AND_EXPR
)
996 oprnd
= gimple_assign_rhs1 (stmt
);
997 const_oprnd
= gimple_assign_rhs2 (stmt
);
998 type
= gimple_expr_type (stmt
);
1000 if (TREE_CODE (oprnd
) != SSA_NAME
1001 || TREE_CODE (const_oprnd
) != INTEGER_CST
)
1004 /* If oprnd has other uses besides that in stmt we cannot mark it
1005 as being part of a pattern only. */
1006 if (!has_single_use (oprnd
))
1009 /* If we are in the middle of a sequence, we use DEF from a previous
1010 statement. Otherwise, OPRND has to be a result of type promotion. */
1013 half_type
= *new_type
;
1019 if (!type_conversion_p (oprnd
, stmt
, false, &half_type
, &def_stmt
,
1022 || !vect_same_loop_or_bb_p (stmt
, def_stmt
))
1026 /* Can we perform the operation on a smaller type? */
1032 if (!int_fits_type_p (const_oprnd
, half_type
))
1034 /* HALF_TYPE is not enough. Try a bigger type if possible. */
1035 if (TYPE_PRECISION (type
) < (TYPE_PRECISION (half_type
) * 4))
1038 interm_type
= build_nonstandard_integer_type (
1039 TYPE_PRECISION (half_type
) * 2, TYPE_UNSIGNED (type
));
1040 if (!int_fits_type_p (const_oprnd
, interm_type
))
1047 /* Try intermediate type - HALF_TYPE is not enough for sure. */
1048 if (TYPE_PRECISION (type
) < (TYPE_PRECISION (half_type
) * 4))
1051 /* Check that HALF_TYPE size + shift amount <= INTERM_TYPE size.
1052 (e.g., if the original value was char, the shift amount is at most 8
1053 if we want to use short). */
1054 if (compare_tree_int (const_oprnd
, TYPE_PRECISION (half_type
)) == 1)
1057 interm_type
= build_nonstandard_integer_type (
1058 TYPE_PRECISION (half_type
) * 2, TYPE_UNSIGNED (type
));
1060 if (!vect_supportable_shift (code
, interm_type
))
1066 if (vect_supportable_shift (code
, half_type
))
1069 /* Try intermediate type - HALF_TYPE is not supported. */
1070 if (TYPE_PRECISION (type
) < (TYPE_PRECISION (half_type
) * 4))
1073 interm_type
= build_nonstandard_integer_type (
1074 TYPE_PRECISION (half_type
) * 2, TYPE_UNSIGNED (type
));
1076 if (!vect_supportable_shift (code
, interm_type
))
1085 /* There are four possible cases:
1086 1. OPRND is defined by a type promotion (in that case FIRST is TRUE, it's
1087 the first statement in the sequence)
1088 a. The original, HALF_TYPE, is not enough - we replace the promotion
1089 from HALF_TYPE to TYPE with a promotion to INTERM_TYPE.
1090 b. HALF_TYPE is sufficient, OPRND is set as the RHS of the original
1092 2. OPRND is defined by a pattern statement we created.
1093 a. Its type is not sufficient for the operation, we create a new stmt:
1094 a type conversion for OPRND from HALF_TYPE to INTERM_TYPE. We store
1095 this statement in NEW_DEF_STMT, and it is later put in
1096 STMT_VINFO_PATTERN_DEF_SEQ of the pattern statement for STMT.
1097 b. OPRND is good to use in the new statement. */
1102 /* Replace the original type conversion HALF_TYPE->TYPE with
1103 HALF_TYPE->INTERM_TYPE. */
1104 if (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
)))
1106 new_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
));
1107 /* Check if the already created pattern stmt is what we need. */
1108 if (!is_gimple_assign (new_stmt
)
1109 || gimple_assign_rhs_code (new_stmt
) != NOP_EXPR
1110 || TREE_TYPE (gimple_assign_lhs (new_stmt
)) != interm_type
)
1113 stmts
->safe_push (def_stmt
);
1114 oprnd
= gimple_assign_lhs (new_stmt
);
1118 /* Create NEW_OPRND = (INTERM_TYPE) OPRND. */
1119 oprnd
= gimple_assign_rhs1 (def_stmt
);
1120 new_oprnd
= make_ssa_name (interm_type
, NULL
);
1121 new_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, new_oprnd
,
1123 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt
)) = new_stmt
;
1124 stmts
->safe_push (def_stmt
);
1130 /* Retrieve the operand before the type promotion. */
1131 oprnd
= gimple_assign_rhs1 (def_stmt
);
1138 /* Create a type conversion HALF_TYPE->INTERM_TYPE. */
1139 new_oprnd
= make_ssa_name (interm_type
, NULL
);
1140 new_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, new_oprnd
,
1143 *new_def_stmt
= new_stmt
;
1146 /* Otherwise, OPRND is already set. */
1150 *new_type
= interm_type
;
1152 *new_type
= half_type
;
1155 *op1
= fold_convert (*new_type
, const_oprnd
);
1161 /* Try to find a statement or a sequence of statements that can be performed
1165 TYPE x_T, res0_T, res1_T;
1168 S2 x_T = (TYPE) x_t;
1169 S3 res0_T = op (x_T, C0);
1170 S4 res1_T = op (res0_T, C1);
1171 S5 ... = () res1_T; - type demotion
1173 where type 'TYPE' is at least double the size of type 'type', C0 and C1 are
1175 Check if S3 and S4 can be done on a smaller type than 'TYPE', it can either
1176 be 'type' or some intermediate type. For now, we expect S5 to be a type
1177 demotion operation. We also check that S3 and S4 have only one use. */
1180 vect_recog_over_widening_pattern (vec
<gimple
> *stmts
,
1181 tree
*type_in
, tree
*type_out
)
1183 gimple stmt
= stmts
->pop ();
1184 gimple pattern_stmt
= NULL
, new_def_stmt
, prev_stmt
= NULL
, use_stmt
= NULL
;
1185 tree op0
, op1
, vectype
= NULL_TREE
, use_lhs
, use_type
;
1186 tree var
= NULL_TREE
, new_type
= NULL_TREE
, new_oprnd
;
1193 if (!vinfo_for_stmt (stmt
)
1194 || STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (stmt
)))
1197 new_def_stmt
= NULL
;
1198 if (!vect_operation_fits_smaller_type (stmt
, var
, &new_type
,
1199 &op0
, &op1
, &new_def_stmt
,
1208 /* STMT can be performed on a smaller type. Check its uses. */
1209 use_stmt
= vect_single_imm_use (stmt
);
1210 if (!use_stmt
|| !is_gimple_assign (use_stmt
))
1213 /* Create pattern statement for STMT. */
1214 vectype
= get_vectype_for_scalar_type (new_type
);
1218 /* We want to collect all the statements for which we create pattern
1219 statetments, except for the case when the last statement in the
1220 sequence doesn't have a corresponding pattern statement. In such
1221 case we associate the last pattern statement with the last statement
1222 in the sequence. Therefore, we only add the original statement to
1223 the list if we know that it is not the last. */
1225 stmts
->safe_push (prev_stmt
);
1227 var
= vect_recog_temp_ssa_var (new_type
, NULL
);
1229 = gimple_build_assign_with_ops (gimple_assign_rhs_code (stmt
), var
,
1231 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt
)) = pattern_stmt
;
1232 new_pattern_def_seq (vinfo_for_stmt (stmt
), new_def_stmt
);
1234 if (dump_enabled_p ())
1236 dump_printf_loc (MSG_NOTE
, vect_location
,
1237 "created pattern stmt: ");
1238 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
1239 dump_printf (MSG_NOTE
, "\n");
1242 type
= gimple_expr_type (stmt
);
1249 /* We got a sequence. We expect it to end with a type demotion operation.
1250 Otherwise, we quit (for now). There are three possible cases: the
1251 conversion is to NEW_TYPE (we don't do anything), the conversion is to
1252 a type bigger than NEW_TYPE and/or the signedness of USE_TYPE and
1253 NEW_TYPE differs (we create a new conversion statement). */
1254 if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (use_stmt
)))
1256 use_lhs
= gimple_assign_lhs (use_stmt
);
1257 use_type
= TREE_TYPE (use_lhs
);
1258 /* Support only type demotion or signedess change. */
1259 if (!INTEGRAL_TYPE_P (use_type
)
1260 || TYPE_PRECISION (type
) <= TYPE_PRECISION (use_type
))
1263 /* Check that NEW_TYPE is not bigger than the conversion result. */
1264 if (TYPE_PRECISION (new_type
) > TYPE_PRECISION (use_type
))
1267 if (TYPE_UNSIGNED (new_type
) != TYPE_UNSIGNED (use_type
)
1268 || TYPE_PRECISION (new_type
) != TYPE_PRECISION (use_type
))
1270 /* Create NEW_TYPE->USE_TYPE conversion. */
1271 new_oprnd
= make_ssa_name (use_type
, NULL
);
1272 pattern_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, new_oprnd
,
1274 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use_stmt
)) = pattern_stmt
;
1276 *type_in
= get_vectype_for_scalar_type (new_type
);
1277 *type_out
= get_vectype_for_scalar_type (use_type
);
1279 /* We created a pattern statement for the last statement in the
1280 sequence, so we don't need to associate it with the pattern
1281 statement created for PREV_STMT. Therefore, we add PREV_STMT
1282 to the list in order to mark it later in vect_pattern_recog_1. */
1284 stmts
->safe_push (prev_stmt
);
1289 STMT_VINFO_PATTERN_DEF_SEQ (vinfo_for_stmt (use_stmt
))
1290 = STMT_VINFO_PATTERN_DEF_SEQ (vinfo_for_stmt (prev_stmt
));
1293 *type_out
= NULL_TREE
;
1296 stmts
->safe_push (use_stmt
);
1299 /* TODO: support general case, create a conversion to the correct type. */
1302 /* Pattern detected. */
1303 if (dump_enabled_p ())
1305 dump_printf_loc (MSG_NOTE
, vect_location
,
1306 "vect_recog_over_widening_pattern: detected: ");
1307 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
1308 dump_printf (MSG_NOTE
, "\n");
1311 return pattern_stmt
;
1314 /* Detect widening shift pattern:
1320 S2 a_T = (TYPE) a_t;
1321 S3 res_T = a_T << CONST;
1323 where type 'TYPE' is at least double the size of type 'type'.
1325 Also detect cases where the shift result is immediately converted
1326 to another type 'result_type' that is no larger in size than 'TYPE'.
1327 In those cases we perform a widen-shift that directly results in
1328 'result_type', to avoid a possible over-widening situation:
1332 result_type res_result;
1335 S2 a_T = (TYPE) a_t;
1336 S3 res_T = a_T << CONST;
1337 S4 res_result = (result_type) res_T;
1338 '--> res_result' = a_t w<< CONST;
1340 And a case when 'TYPE' is 4 times bigger than 'type'. In that case we
1341 create an additional pattern stmt for S2 to create a variable of an
1342 intermediate type, and perform widen-shift on the intermediate type:
1346 TYPE a_T, res_T, res_T';
1349 S2 a_T = (TYPE) a_t;
1350 '--> a_it = (interm_type) a_t;
1351 S3 res_T = a_T << CONST;
1352 '--> res_T' = a_it <<* CONST;
1356 * STMTS: Contains a stmt from which the pattern search begins.
1357 In case of unsigned widen-shift, the original stmt (S3) is replaced with S4
1358 in STMTS. When an intermediate type is used and a pattern statement is
1359 created for S2, we also put S2 here (before S3).
1363 * TYPE_IN: The type of the input arguments to the pattern.
1365 * TYPE_OUT: The type of the output of this pattern.
1367 * Return value: A new stmt that will be used to replace the sequence of
1368 stmts that constitute the pattern. In this case it will be:
1369 WIDEN_LSHIFT_EXPR <a_t, CONST>. */
1372 vect_recog_widen_shift_pattern (vec
<gimple
> *stmts
,
1373 tree
*type_in
, tree
*type_out
)
1375 gimple last_stmt
= stmts
->pop ();
1377 tree oprnd0
, oprnd1
;
1378 tree type
, half_type0
;
1379 gimple pattern_stmt
;
1380 tree vectype
, vectype_out
= NULL_TREE
;
1382 enum tree_code dummy_code
;
1384 vec
<tree
> dummy_vec
;
1388 if (!is_gimple_assign (last_stmt
) || !vinfo_for_stmt (last_stmt
))
1391 if (STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (last_stmt
)))
1394 if (gimple_assign_rhs_code (last_stmt
) != LSHIFT_EXPR
)
1397 oprnd0
= gimple_assign_rhs1 (last_stmt
);
1398 oprnd1
= gimple_assign_rhs2 (last_stmt
);
1399 if (TREE_CODE (oprnd0
) != SSA_NAME
|| TREE_CODE (oprnd1
) != INTEGER_CST
)
1402 /* Check operand 0: it has to be defined by a type promotion. */
1403 if (!type_conversion_p (oprnd0
, last_stmt
, false, &half_type0
, &def_stmt0
,
1408 /* Check operand 1: has to be positive. We check that it fits the type
1409 in vect_handle_widen_op_by_const (). */
1410 if (tree_int_cst_compare (oprnd1
, size_zero_node
) <= 0)
1413 oprnd0
= gimple_assign_rhs1 (def_stmt0
);
1414 type
= gimple_expr_type (last_stmt
);
1416 /* Check for subsequent conversion to another type. */
1417 use_stmt
= vect_single_imm_use (last_stmt
);
1418 if (use_stmt
&& is_gimple_assign (use_stmt
)
1419 && CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (use_stmt
))
1420 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt
)))
1422 tree use_lhs
= gimple_assign_lhs (use_stmt
);
1423 tree use_type
= TREE_TYPE (use_lhs
);
1425 if (INTEGRAL_TYPE_P (use_type
)
1426 && TYPE_PRECISION (use_type
) <= TYPE_PRECISION (type
))
1428 last_stmt
= use_stmt
;
1433 /* Check if this a widening operation. */
1434 if (!vect_handle_widen_op_by_const (last_stmt
, LSHIFT_EXPR
, oprnd1
,
1436 type
, &half_type0
, def_stmt0
))
1439 /* Pattern detected. */
1440 if (dump_enabled_p ())
1441 dump_printf_loc (MSG_NOTE
, vect_location
,
1442 "vect_recog_widen_shift_pattern: detected:\n");
1444 /* Check target support. */
1445 vectype
= get_vectype_for_scalar_type (half_type0
);
1446 vectype_out
= get_vectype_for_scalar_type (type
);
1450 || !supportable_widening_operation (WIDEN_LSHIFT_EXPR
, last_stmt
,
1451 vectype_out
, vectype
,
1452 &dummy_code
, &dummy_code
,
1453 &dummy_int
, &dummy_vec
))
1457 *type_out
= vectype_out
;
1459 /* Pattern supported. Create a stmt to be used to replace the pattern. */
1460 var
= vect_recog_temp_ssa_var (type
, NULL
);
1462 gimple_build_assign_with_ops (WIDEN_LSHIFT_EXPR
, var
, oprnd0
, oprnd1
);
1464 if (dump_enabled_p ())
1465 dump_gimple_stmt_loc (MSG_NOTE
, vect_location
, TDF_SLIM
, pattern_stmt
, 0);
1467 stmts
->safe_push (last_stmt
);
1468 return pattern_stmt
;
1471 /* Detect a rotate pattern wouldn't be otherwise vectorized:
1475 S0 a_t = b_t r<< c_t;
1479 * STMTS: Contains a stmt from which the pattern search begins,
1480 i.e. the shift/rotate stmt. The original stmt (S0) is replaced
1484 S2 e_t = d_t & (B - 1);
1485 S3 f_t = b_t << c_t;
1486 S4 g_t = b_t >> e_t;
1489 where B is element bitsize of type.
1493 * TYPE_IN: The type of the input arguments to the pattern.
1495 * TYPE_OUT: The type of the output of this pattern.
1497 * Return value: A new stmt that will be used to replace the rotate
1501 vect_recog_rotate_pattern (vec
<gimple
> *stmts
, tree
*type_in
, tree
*type_out
)
1503 gimple last_stmt
= stmts
->pop ();
1504 tree oprnd0
, oprnd1
, lhs
, var
, var1
, var2
, vectype
, type
, stype
, def
, def2
;
1505 gimple pattern_stmt
, def_stmt
;
1506 enum tree_code rhs_code
;
1507 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
1508 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
1509 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
1510 enum vect_def_type dt
;
1511 optab optab1
, optab2
;
1512 edge ext_def
= NULL
;
1514 if (!is_gimple_assign (last_stmt
))
1517 rhs_code
= gimple_assign_rhs_code (last_stmt
);
1527 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
1530 lhs
= gimple_assign_lhs (last_stmt
);
1531 oprnd0
= gimple_assign_rhs1 (last_stmt
);
1532 type
= TREE_TYPE (oprnd0
);
1533 oprnd1
= gimple_assign_rhs2 (last_stmt
);
1534 if (TREE_CODE (oprnd0
) != SSA_NAME
1535 || TYPE_PRECISION (TREE_TYPE (lhs
)) != TYPE_PRECISION (type
)
1536 || !INTEGRAL_TYPE_P (type
)
1537 || !TYPE_UNSIGNED (type
))
1540 if (!vect_is_simple_use (oprnd1
, last_stmt
, loop_vinfo
, bb_vinfo
, &def_stmt
,
1544 if (dt
!= vect_internal_def
1545 && dt
!= vect_constant_def
1546 && dt
!= vect_external_def
)
1549 vectype
= get_vectype_for_scalar_type (type
);
1550 if (vectype
== NULL_TREE
)
1553 /* If vector/vector or vector/scalar rotate is supported by the target,
1554 don't do anything here. */
1555 optab1
= optab_for_tree_code (rhs_code
, vectype
, optab_vector
);
1557 && optab_handler (optab1
, TYPE_MODE (vectype
)) != CODE_FOR_nothing
)
1560 if (bb_vinfo
!= NULL
|| dt
!= vect_internal_def
)
1562 optab2
= optab_for_tree_code (rhs_code
, vectype
, optab_scalar
);
1564 && optab_handler (optab2
, TYPE_MODE (vectype
)) != CODE_FOR_nothing
)
1568 /* If vector/vector or vector/scalar shifts aren't supported by the target,
1569 don't do anything here either. */
1570 optab1
= optab_for_tree_code (LSHIFT_EXPR
, vectype
, optab_vector
);
1571 optab2
= optab_for_tree_code (RSHIFT_EXPR
, vectype
, optab_vector
);
1573 || optab_handler (optab1
, TYPE_MODE (vectype
)) == CODE_FOR_nothing
1575 || optab_handler (optab2
, TYPE_MODE (vectype
)) == CODE_FOR_nothing
)
1577 if (bb_vinfo
== NULL
&& dt
== vect_internal_def
)
1579 optab1
= optab_for_tree_code (LSHIFT_EXPR
, vectype
, optab_scalar
);
1580 optab2
= optab_for_tree_code (RSHIFT_EXPR
, vectype
, optab_scalar
);
1582 || optab_handler (optab1
, TYPE_MODE (vectype
)) == CODE_FOR_nothing
1584 || optab_handler (optab2
, TYPE_MODE (vectype
)) == CODE_FOR_nothing
)
1589 *type_out
= vectype
;
1590 if (*type_in
== NULL_TREE
)
1593 if (dt
== vect_external_def
1594 && TREE_CODE (oprnd1
) == SSA_NAME
1597 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1598 ext_def
= loop_preheader_edge (loop
);
1599 if (!SSA_NAME_IS_DEFAULT_DEF (oprnd1
))
1601 basic_block bb
= gimple_bb (SSA_NAME_DEF_STMT (oprnd1
));
1603 || !dominated_by_p (CDI_DOMINATORS
, ext_def
->dest
, bb
))
1609 if (TREE_CODE (oprnd1
) == INTEGER_CST
1610 || TYPE_MODE (TREE_TYPE (oprnd1
)) == TYPE_MODE (type
))
1612 else if (def_stmt
&& gimple_assign_cast_p (def_stmt
))
1614 tree rhs1
= gimple_assign_rhs1 (def_stmt
);
1615 if (TYPE_MODE (TREE_TYPE (rhs1
)) == TYPE_MODE (type
)
1616 && TYPE_PRECISION (TREE_TYPE (rhs1
))
1617 == TYPE_PRECISION (type
))
1621 STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo
) = NULL
;
1622 if (def
== NULL_TREE
)
1624 def
= vect_recog_temp_ssa_var (type
, NULL
);
1625 def_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, def
, oprnd1
,
1630 = gsi_insert_on_edge_immediate (ext_def
, def_stmt
);
1631 gcc_assert (!new_bb
);
1634 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1636 stype
= TREE_TYPE (def
);
1638 if (TREE_CODE (def
) == INTEGER_CST
)
1640 if (!tree_fits_uhwi_p (def
)
1641 || tree_to_uhwi (def
) >= GET_MODE_PRECISION (TYPE_MODE (type
))
1642 || integer_zerop (def
))
1644 def2
= build_int_cst (stype
,
1645 GET_MODE_PRECISION (TYPE_MODE (type
))
1646 - tree_to_uhwi (def
));
1650 tree vecstype
= get_vectype_for_scalar_type (stype
);
1651 stmt_vec_info def_stmt_vinfo
;
1653 if (vecstype
== NULL_TREE
)
1655 def2
= vect_recog_temp_ssa_var (stype
, NULL
);
1656 def_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, def2
, def
,
1661 = gsi_insert_on_edge_immediate (ext_def
, def_stmt
);
1662 gcc_assert (!new_bb
);
1666 def_stmt_vinfo
= new_stmt_vec_info (def_stmt
, loop_vinfo
, bb_vinfo
);
1667 set_vinfo_for_stmt (def_stmt
, def_stmt_vinfo
);
1668 STMT_VINFO_VECTYPE (def_stmt_vinfo
) = vecstype
;
1669 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1672 def2
= vect_recog_temp_ssa_var (stype
, NULL
);
1674 = build_int_cst (stype
, GET_MODE_PRECISION (TYPE_MODE (stype
)) - 1);
1675 def_stmt
= gimple_build_assign_with_ops (BIT_AND_EXPR
, def2
,
1676 gimple_assign_lhs (def_stmt
),
1681 = gsi_insert_on_edge_immediate (ext_def
, def_stmt
);
1682 gcc_assert (!new_bb
);
1686 def_stmt_vinfo
= new_stmt_vec_info (def_stmt
, loop_vinfo
, bb_vinfo
);
1687 set_vinfo_for_stmt (def_stmt
, def_stmt_vinfo
);
1688 STMT_VINFO_VECTYPE (def_stmt_vinfo
) = vecstype
;
1689 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1693 var1
= vect_recog_temp_ssa_var (type
, NULL
);
1694 def_stmt
= gimple_build_assign_with_ops (rhs_code
== LROTATE_EXPR
1695 ? LSHIFT_EXPR
: RSHIFT_EXPR
,
1697 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1699 var2
= vect_recog_temp_ssa_var (type
, NULL
);
1700 def_stmt
= gimple_build_assign_with_ops (rhs_code
== LROTATE_EXPR
1701 ? RSHIFT_EXPR
: LSHIFT_EXPR
,
1702 var2
, oprnd0
, def2
);
1703 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1705 /* Pattern detected. */
1706 if (dump_enabled_p ())
1707 dump_printf_loc (MSG_NOTE
, vect_location
,
1708 "vect_recog_rotate_pattern: detected:\n");
1710 /* Pattern supported. Create a stmt to be used to replace the pattern. */
1711 var
= vect_recog_temp_ssa_var (type
, NULL
);
1712 pattern_stmt
= gimple_build_assign_with_ops (BIT_IOR_EXPR
, var
, var1
, var2
);
1714 if (dump_enabled_p ())
1715 dump_gimple_stmt_loc (MSG_NOTE
, vect_location
, TDF_SLIM
, pattern_stmt
, 0);
1717 stmts
->safe_push (last_stmt
);
1718 return pattern_stmt
;
1721 /* Detect a vector by vector shift pattern that wouldn't be otherwise
1729 S3 res_T = b_T op a_t;
1731 where type 'TYPE' is a type with different size than 'type',
1732 and op is <<, >> or rotate.
1737 TYPE b_T, c_T, res_T;
1740 S1 a_t = (type) c_T;
1742 S3 res_T = b_T op a_t;
1746 * STMTS: Contains a stmt from which the pattern search begins,
1747 i.e. the shift/rotate stmt. The original stmt (S3) is replaced
1748 with a shift/rotate which has same type on both operands, in the
1749 second case just b_T op c_T, in the first case with added cast
1750 from a_t to c_T in STMT_VINFO_PATTERN_DEF_SEQ.
1754 * TYPE_IN: The type of the input arguments to the pattern.
1756 * TYPE_OUT: The type of the output of this pattern.
1758 * Return value: A new stmt that will be used to replace the shift/rotate
1762 vect_recog_vector_vector_shift_pattern (vec
<gimple
> *stmts
,
1763 tree
*type_in
, tree
*type_out
)
1765 gimple last_stmt
= stmts
->pop ();
1766 tree oprnd0
, oprnd1
, lhs
, var
;
1767 gimple pattern_stmt
, def_stmt
;
1768 enum tree_code rhs_code
;
1769 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
1770 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
1771 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
1772 enum vect_def_type dt
;
1775 if (!is_gimple_assign (last_stmt
))
1778 rhs_code
= gimple_assign_rhs_code (last_stmt
);
1790 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
1793 lhs
= gimple_assign_lhs (last_stmt
);
1794 oprnd0
= gimple_assign_rhs1 (last_stmt
);
1795 oprnd1
= gimple_assign_rhs2 (last_stmt
);
1796 if (TREE_CODE (oprnd0
) != SSA_NAME
1797 || TREE_CODE (oprnd1
) != SSA_NAME
1798 || TYPE_MODE (TREE_TYPE (oprnd0
)) == TYPE_MODE (TREE_TYPE (oprnd1
))
1799 || TYPE_PRECISION (TREE_TYPE (oprnd1
))
1800 != GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (oprnd1
)))
1801 || TYPE_PRECISION (TREE_TYPE (lhs
))
1802 != TYPE_PRECISION (TREE_TYPE (oprnd0
)))
1805 if (!vect_is_simple_use (oprnd1
, last_stmt
, loop_vinfo
, bb_vinfo
, &def_stmt
,
1809 if (dt
!= vect_internal_def
)
1812 *type_in
= get_vectype_for_scalar_type (TREE_TYPE (oprnd0
));
1813 *type_out
= *type_in
;
1814 if (*type_in
== NULL_TREE
)
1818 if (gimple_assign_cast_p (def_stmt
))
1820 tree rhs1
= gimple_assign_rhs1 (def_stmt
);
1821 if (TYPE_MODE (TREE_TYPE (rhs1
)) == TYPE_MODE (TREE_TYPE (oprnd0
))
1822 && TYPE_PRECISION (TREE_TYPE (rhs1
))
1823 == TYPE_PRECISION (TREE_TYPE (oprnd0
)))
1827 if (def
== NULL_TREE
)
1829 def
= vect_recog_temp_ssa_var (TREE_TYPE (oprnd0
), NULL
);
1830 def_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, def
, oprnd1
,
1832 new_pattern_def_seq (stmt_vinfo
, def_stmt
);
1835 /* Pattern detected. */
1836 if (dump_enabled_p ())
1837 dump_printf_loc (MSG_NOTE
, vect_location
,
1838 "vect_recog_vector_vector_shift_pattern: detected:\n");
1840 /* Pattern supported. Create a stmt to be used to replace the pattern. */
1841 var
= vect_recog_temp_ssa_var (TREE_TYPE (oprnd0
), NULL
);
1842 pattern_stmt
= gimple_build_assign_with_ops (rhs_code
, var
, oprnd0
, def
);
1844 if (dump_enabled_p ())
1845 dump_gimple_stmt_loc (MSG_NOTE
, vect_location
, TDF_SLIM
, pattern_stmt
, 0);
1847 stmts
->safe_push (last_stmt
);
1848 return pattern_stmt
;
1851 /* Detect a signed division by a constant that wouldn't be
1852 otherwise vectorized:
1858 where type 'type' is an integral type and N is a constant.
1860 Similarly handle modulo by a constant:
1866 * STMTS: Contains a stmt from which the pattern search begins,
1867 i.e. the division stmt. S1 is replaced by if N is a power
1868 of two constant and type is signed:
1869 S3 y_t = b_t < 0 ? N - 1 : 0;
1871 S1' a_t = x_t >> log2 (N);
1873 S4 is replaced if N is a power of two constant and
1874 type is signed by (where *_T temporaries have unsigned type):
1875 S9 y_T = b_t < 0 ? -1U : 0U;
1876 S8 z_T = y_T >> (sizeof (type_t) * CHAR_BIT - log2 (N));
1877 S7 z_t = (type) z_T;
1879 S5 x_t = w_t & (N - 1);
1880 S4' a_t = x_t - z_t;
1884 * TYPE_IN: The type of the input arguments to the pattern.
1886 * TYPE_OUT: The type of the output of this pattern.
1888 * Return value: A new stmt that will be used to replace the division
1889 S1 or modulo S4 stmt. */
1892 vect_recog_divmod_pattern (vec
<gimple
> *stmts
,
1893 tree
*type_in
, tree
*type_out
)
1895 gimple last_stmt
= stmts
->pop ();
1896 tree oprnd0
, oprnd1
, vectype
, itype
, cond
;
1897 gimple pattern_stmt
, def_stmt
;
1898 enum tree_code rhs_code
;
1899 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
1900 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
1901 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
1904 int dummy_int
, prec
;
1905 stmt_vec_info def_stmt_vinfo
;
1907 if (!is_gimple_assign (last_stmt
))
1910 rhs_code
= gimple_assign_rhs_code (last_stmt
);
1913 case TRUNC_DIV_EXPR
:
1914 case TRUNC_MOD_EXPR
:
1920 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
1923 oprnd0
= gimple_assign_rhs1 (last_stmt
);
1924 oprnd1
= gimple_assign_rhs2 (last_stmt
);
1925 itype
= TREE_TYPE (oprnd0
);
1926 if (TREE_CODE (oprnd0
) != SSA_NAME
1927 || TREE_CODE (oprnd1
) != INTEGER_CST
1928 || TREE_CODE (itype
) != INTEGER_TYPE
1929 || TYPE_PRECISION (itype
) != GET_MODE_PRECISION (TYPE_MODE (itype
)))
1932 vectype
= get_vectype_for_scalar_type (itype
);
1933 if (vectype
== NULL_TREE
)
1936 /* If the target can handle vectorized division or modulo natively,
1937 don't attempt to optimize this. */
1938 optab
= optab_for_tree_code (rhs_code
, vectype
, optab_default
);
1939 if (optab
!= unknown_optab
)
1941 enum machine_mode vec_mode
= TYPE_MODE (vectype
);
1942 int icode
= (int) optab_handler (optab
, vec_mode
);
1943 if (icode
!= CODE_FOR_nothing
)
1947 prec
= TYPE_PRECISION (itype
);
1948 if (integer_pow2p (oprnd1
))
1950 if (TYPE_UNSIGNED (itype
) || tree_int_cst_sgn (oprnd1
) != 1)
1953 /* Pattern detected. */
1954 if (dump_enabled_p ())
1955 dump_printf_loc (MSG_NOTE
, vect_location
,
1956 "vect_recog_divmod_pattern: detected:\n");
1958 cond
= build2 (LT_EXPR
, boolean_type_node
, oprnd0
,
1959 build_int_cst (itype
, 0));
1960 if (rhs_code
== TRUNC_DIV_EXPR
)
1962 tree var
= vect_recog_temp_ssa_var (itype
, NULL
);
1965 = gimple_build_assign_with_ops (COND_EXPR
, var
, cond
,
1966 fold_build2 (MINUS_EXPR
, itype
,
1968 build_int_cst (itype
,
1970 build_int_cst (itype
, 0));
1971 new_pattern_def_seq (stmt_vinfo
, def_stmt
);
1972 var
= vect_recog_temp_ssa_var (itype
, NULL
);
1974 = gimple_build_assign_with_ops (PLUS_EXPR
, var
, oprnd0
,
1975 gimple_assign_lhs (def_stmt
));
1976 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
1978 shift
= build_int_cst (itype
, tree_log2 (oprnd1
));
1980 = gimple_build_assign_with_ops (RSHIFT_EXPR
,
1981 vect_recog_temp_ssa_var (itype
,
1988 STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo
) = NULL
;
1989 if (compare_tree_int (oprnd1
, 2) == 0)
1991 signmask
= vect_recog_temp_ssa_var (itype
, NULL
);
1993 = gimple_build_assign_with_ops (COND_EXPR
, signmask
, cond
,
1994 build_int_cst (itype
, 1),
1995 build_int_cst (itype
, 0));
1996 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2001 = build_nonstandard_integer_type (prec
, 1);
2002 tree vecutype
= get_vectype_for_scalar_type (utype
);
2004 = build_int_cst (utype
, GET_MODE_BITSIZE (TYPE_MODE (itype
))
2005 - tree_log2 (oprnd1
));
2006 tree var
= vect_recog_temp_ssa_var (utype
, NULL
);
2009 = gimple_build_assign_with_ops (COND_EXPR
, var
, cond
,
2010 build_int_cst (utype
, -1),
2011 build_int_cst (utype
, 0));
2013 = new_stmt_vec_info (def_stmt
, loop_vinfo
, bb_vinfo
);
2014 set_vinfo_for_stmt (def_stmt
, def_stmt_vinfo
);
2015 STMT_VINFO_VECTYPE (def_stmt_vinfo
) = vecutype
;
2016 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2017 var
= vect_recog_temp_ssa_var (utype
, NULL
);
2019 = gimple_build_assign_with_ops (RSHIFT_EXPR
, var
,
2020 gimple_assign_lhs (def_stmt
),
2023 = new_stmt_vec_info (def_stmt
, loop_vinfo
, bb_vinfo
);
2024 set_vinfo_for_stmt (def_stmt
, def_stmt_vinfo
);
2025 STMT_VINFO_VECTYPE (def_stmt_vinfo
) = vecutype
;
2026 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2027 signmask
= vect_recog_temp_ssa_var (itype
, NULL
);
2029 = gimple_build_assign_with_ops (NOP_EXPR
, signmask
, var
,
2031 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2034 = gimple_build_assign_with_ops (PLUS_EXPR
,
2035 vect_recog_temp_ssa_var (itype
,
2038 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2040 = gimple_build_assign_with_ops (BIT_AND_EXPR
,
2041 vect_recog_temp_ssa_var (itype
,
2043 gimple_assign_lhs (def_stmt
),
2044 fold_build2 (MINUS_EXPR
, itype
,
2046 build_int_cst (itype
,
2048 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2051 = gimple_build_assign_with_ops (MINUS_EXPR
,
2052 vect_recog_temp_ssa_var (itype
,
2054 gimple_assign_lhs (def_stmt
),
2058 if (dump_enabled_p ())
2059 dump_gimple_stmt_loc (MSG_NOTE
, vect_location
, TDF_SLIM
, pattern_stmt
,
2062 stmts
->safe_push (last_stmt
);
2065 *type_out
= vectype
;
2066 return pattern_stmt
;
2069 if (prec
> HOST_BITS_PER_WIDE_INT
2070 || integer_zerop (oprnd1
))
2073 if (!can_mult_highpart_p (TYPE_MODE (vectype
), TYPE_UNSIGNED (itype
)))
2076 STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo
) = NULL
;
2078 if (TYPE_UNSIGNED (itype
))
2080 unsigned HOST_WIDE_INT mh
, ml
;
2081 int pre_shift
, post_shift
;
2082 unsigned HOST_WIDE_INT d
= (TREE_INT_CST_LOW (oprnd1
)
2083 & GET_MODE_MASK (TYPE_MODE (itype
)));
2084 tree t1
, t2
, t3
, t4
;
2086 if (d
>= ((unsigned HOST_WIDE_INT
) 1 << (prec
- 1)))
2087 /* FIXME: Can transform this into oprnd0 >= oprnd1 ? 1 : 0. */
2090 /* Find a suitable multiplier and right shift count
2091 instead of multiplying with D. */
2092 mh
= choose_multiplier (d
, prec
, prec
, &ml
, &post_shift
, &dummy_int
);
2094 /* If the suggested multiplier is more than SIZE bits, we can do better
2095 for even divisors, using an initial right shift. */
2096 if (mh
!= 0 && (d
& 1) == 0)
2098 pre_shift
= floor_log2 (d
& -d
);
2099 mh
= choose_multiplier (d
>> pre_shift
, prec
, prec
- pre_shift
,
2100 &ml
, &post_shift
, &dummy_int
);
2108 if (post_shift
- 1 >= prec
)
2111 /* t1 = oprnd0 h* ml;
2115 q = t4 >> (post_shift - 1); */
2116 t1
= vect_recog_temp_ssa_var (itype
, NULL
);
2118 = gimple_build_assign_with_ops (MULT_HIGHPART_EXPR
, t1
, oprnd0
,
2119 build_int_cst (itype
, ml
));
2120 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2122 t2
= vect_recog_temp_ssa_var (itype
, NULL
);
2124 = gimple_build_assign_with_ops (MINUS_EXPR
, t2
, oprnd0
, t1
);
2125 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2127 t3
= vect_recog_temp_ssa_var (itype
, NULL
);
2129 = gimple_build_assign_with_ops (RSHIFT_EXPR
, t3
, t2
,
2131 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2133 t4
= vect_recog_temp_ssa_var (itype
, NULL
);
2135 = gimple_build_assign_with_ops (PLUS_EXPR
, t4
, t1
, t3
);
2137 if (post_shift
!= 1)
2139 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2141 q
= vect_recog_temp_ssa_var (itype
, NULL
);
2143 = gimple_build_assign_with_ops (RSHIFT_EXPR
, q
, t4
,
2144 build_int_cst (itype
,
2151 pattern_stmt
= def_stmt
;
2156 if (pre_shift
>= prec
|| post_shift
>= prec
)
2159 /* t1 = oprnd0 >> pre_shift;
2161 q = t2 >> post_shift; */
2164 t1
= vect_recog_temp_ssa_var (itype
, NULL
);
2166 = gimple_build_assign_with_ops (RSHIFT_EXPR
, t1
, oprnd0
,
2167 build_int_cst (NULL
,
2169 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2174 t2
= vect_recog_temp_ssa_var (itype
, NULL
);
2176 = gimple_build_assign_with_ops (MULT_HIGHPART_EXPR
, t2
, t1
,
2177 build_int_cst (itype
, ml
));
2181 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2183 q
= vect_recog_temp_ssa_var (itype
, NULL
);
2185 = gimple_build_assign_with_ops (RSHIFT_EXPR
, q
, t2
,
2186 build_int_cst (itype
,
2192 pattern_stmt
= def_stmt
;
2197 unsigned HOST_WIDE_INT ml
;
2199 HOST_WIDE_INT d
= TREE_INT_CST_LOW (oprnd1
);
2200 unsigned HOST_WIDE_INT abs_d
;
2202 tree t1
, t2
, t3
, t4
;
2204 /* Give up for -1. */
2208 /* Since d might be INT_MIN, we have to cast to
2209 unsigned HOST_WIDE_INT before negating to avoid
2210 undefined signed overflow. */
2212 ? (unsigned HOST_WIDE_INT
) d
2213 : - (unsigned HOST_WIDE_INT
) d
);
2215 /* n rem d = n rem -d */
2216 if (rhs_code
== TRUNC_MOD_EXPR
&& d
< 0)
2219 oprnd1
= build_int_cst (itype
, abs_d
);
2221 else if (HOST_BITS_PER_WIDE_INT
>= prec
2222 && abs_d
== (unsigned HOST_WIDE_INT
) 1 << (prec
- 1))
2223 /* This case is not handled correctly below. */
2226 choose_multiplier (abs_d
, prec
, prec
- 1, &ml
, &post_shift
, &dummy_int
);
2227 if (ml
>= (unsigned HOST_WIDE_INT
) 1 << (prec
- 1))
2230 ml
|= (~(unsigned HOST_WIDE_INT
) 0) << (prec
- 1);
2232 if (post_shift
>= prec
)
2235 /* t1 = oprnd0 h* ml; */
2236 t1
= vect_recog_temp_ssa_var (itype
, NULL
);
2238 = gimple_build_assign_with_ops (MULT_HIGHPART_EXPR
, t1
, oprnd0
,
2239 build_int_cst (itype
, ml
));
2243 /* t2 = t1 + oprnd0; */
2244 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2245 t2
= vect_recog_temp_ssa_var (itype
, NULL
);
2247 = gimple_build_assign_with_ops (PLUS_EXPR
, t2
, t1
, oprnd0
);
2254 /* t3 = t2 >> post_shift; */
2255 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2256 t3
= vect_recog_temp_ssa_var (itype
, NULL
);
2258 = gimple_build_assign_with_ops (RSHIFT_EXPR
, t3
, t2
,
2259 build_int_cst (itype
, post_shift
));
2264 double_int oprnd0_min
, oprnd0_max
;
2266 if (get_range_info (oprnd0
, &oprnd0_min
, &oprnd0_max
) == VR_RANGE
)
2268 if (!oprnd0_min
.is_negative ())
2270 else if (oprnd0_max
.is_negative ())
2274 if (msb
== 0 && d
>= 0)
2278 pattern_stmt
= def_stmt
;
2282 /* t4 = oprnd0 >> (prec - 1);
2283 or if we know from VRP that oprnd0 >= 0
2285 or if we know from VRP that oprnd0 < 0
2287 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2288 t4
= vect_recog_temp_ssa_var (itype
, NULL
);
2291 = gimple_build_assign_with_ops (INTEGER_CST
,
2292 t4
, build_int_cst (itype
, msb
),
2296 = gimple_build_assign_with_ops (RSHIFT_EXPR
, t4
, oprnd0
,
2297 build_int_cst (itype
, prec
- 1));
2298 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2300 /* q = t3 - t4; or q = t4 - t3; */
2301 q
= vect_recog_temp_ssa_var (itype
, NULL
);
2303 = gimple_build_assign_with_ops (MINUS_EXPR
, q
, d
< 0 ? t4
: t3
,
2308 if (rhs_code
== TRUNC_MOD_EXPR
)
2312 /* We divided. Now finish by:
2315 append_pattern_def_seq (stmt_vinfo
, pattern_stmt
);
2317 t1
= vect_recog_temp_ssa_var (itype
, NULL
);
2319 = gimple_build_assign_with_ops (MULT_EXPR
, t1
, q
, oprnd1
);
2320 append_pattern_def_seq (stmt_vinfo
, def_stmt
);
2322 r
= vect_recog_temp_ssa_var (itype
, NULL
);
2324 = gimple_build_assign_with_ops (MINUS_EXPR
, r
, oprnd0
, t1
);
2327 /* Pattern detected. */
2328 if (dump_enabled_p ())
2330 dump_printf_loc (MSG_NOTE
, vect_location
,
2331 "vect_recog_divmod_pattern: detected: ");
2332 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
2333 dump_printf (MSG_NOTE
, "\n");
2336 stmts
->safe_push (last_stmt
);
2339 *type_out
= vectype
;
2340 return pattern_stmt
;
2343 /* Function vect_recog_mixed_size_cond_pattern
2345 Try to find the following pattern:
2350 S1 a_T = x_t CMP y_t ? b_T : c_T;
2352 where type 'TYPE' is an integral type which has different size
2353 from 'type'. b_T and c_T are either constants (and if 'TYPE' is wider
2354 than 'type', the constants need to fit into an integer type
2355 with the same width as 'type') or results of conversion from 'type'.
2359 * LAST_STMT: A stmt from which the pattern search begins.
2363 * TYPE_IN: The type of the input arguments to the pattern.
2365 * TYPE_OUT: The type of the output of this pattern.
2367 * Return value: A new stmt that will be used to replace the pattern.
2368 Additionally a def_stmt is added.
2370 a_it = x_t CMP y_t ? b_it : c_it;
2371 a_T = (TYPE) a_it; */
2374 vect_recog_mixed_size_cond_pattern (vec
<gimple
> *stmts
, tree
*type_in
,
2377 gimple last_stmt
= (*stmts
)[0];
2378 tree cond_expr
, then_clause
, else_clause
;
2379 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
), def_stmt_info
;
2380 tree type
, vectype
, comp_vectype
, itype
= NULL_TREE
, vecitype
;
2381 enum machine_mode cmpmode
;
2382 gimple pattern_stmt
, def_stmt
;
2383 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2384 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
2385 tree orig_type0
= NULL_TREE
, orig_type1
= NULL_TREE
;
2386 gimple def_stmt0
= NULL
, def_stmt1
= NULL
;
2388 tree comp_scalar_type
;
2390 if (!is_gimple_assign (last_stmt
)
2391 || gimple_assign_rhs_code (last_stmt
) != COND_EXPR
2392 || STMT_VINFO_DEF_TYPE (stmt_vinfo
) != vect_internal_def
)
2395 cond_expr
= gimple_assign_rhs1 (last_stmt
);
2396 then_clause
= gimple_assign_rhs2 (last_stmt
);
2397 else_clause
= gimple_assign_rhs3 (last_stmt
);
2399 if (!COMPARISON_CLASS_P (cond_expr
))
2402 comp_scalar_type
= TREE_TYPE (TREE_OPERAND (cond_expr
, 0));
2403 comp_vectype
= get_vectype_for_scalar_type (comp_scalar_type
);
2404 if (comp_vectype
== NULL_TREE
)
2407 type
= gimple_expr_type (last_stmt
);
2408 if (types_compatible_p (type
, comp_scalar_type
)
2409 || ((TREE_CODE (then_clause
) != INTEGER_CST
2410 || TREE_CODE (else_clause
) != INTEGER_CST
)
2411 && !INTEGRAL_TYPE_P (comp_scalar_type
))
2412 || !INTEGRAL_TYPE_P (type
))
2415 if ((TREE_CODE (then_clause
) != INTEGER_CST
2416 && !type_conversion_p (then_clause
, last_stmt
, false, &orig_type0
,
2417 &def_stmt0
, &promotion
))
2418 || (TREE_CODE (else_clause
) != INTEGER_CST
2419 && !type_conversion_p (else_clause
, last_stmt
, false, &orig_type1
,
2420 &def_stmt1
, &promotion
)))
2423 if (orig_type0
&& orig_type1
2424 && !types_compatible_p (orig_type0
, orig_type1
))
2429 if (!types_compatible_p (orig_type0
, comp_scalar_type
))
2431 then_clause
= gimple_assign_rhs1 (def_stmt0
);
2437 if (!types_compatible_p (orig_type1
, comp_scalar_type
))
2439 else_clause
= gimple_assign_rhs1 (def_stmt1
);
2443 cmpmode
= GET_MODE_INNER (TYPE_MODE (comp_vectype
));
2445 if (GET_MODE_BITSIZE (TYPE_MODE (type
)) == GET_MODE_BITSIZE (cmpmode
))
2448 vectype
= get_vectype_for_scalar_type (type
);
2449 if (vectype
== NULL_TREE
)
2452 if (expand_vec_cond_expr_p (vectype
, comp_vectype
))
2455 if (itype
== NULL_TREE
)
2456 itype
= build_nonstandard_integer_type (GET_MODE_BITSIZE (cmpmode
),
2457 TYPE_UNSIGNED (type
));
2459 if (itype
== NULL_TREE
2460 || GET_MODE_BITSIZE (TYPE_MODE (itype
)) != GET_MODE_BITSIZE (cmpmode
))
2463 vecitype
= get_vectype_for_scalar_type (itype
);
2464 if (vecitype
== NULL_TREE
)
2467 if (!expand_vec_cond_expr_p (vecitype
, comp_vectype
))
2470 if (GET_MODE_BITSIZE (TYPE_MODE (type
)) > GET_MODE_BITSIZE (cmpmode
))
2472 if ((TREE_CODE (then_clause
) == INTEGER_CST
2473 && !int_fits_type_p (then_clause
, itype
))
2474 || (TREE_CODE (else_clause
) == INTEGER_CST
2475 && !int_fits_type_p (else_clause
, itype
)))
2480 = gimple_build_assign_with_ops (COND_EXPR
,
2481 vect_recog_temp_ssa_var (itype
, NULL
),
2482 unshare_expr (cond_expr
),
2483 fold_convert (itype
, then_clause
),
2484 fold_convert (itype
, else_clause
));
2486 = gimple_build_assign_with_ops (NOP_EXPR
,
2487 vect_recog_temp_ssa_var (type
, NULL
),
2488 gimple_assign_lhs (def_stmt
), NULL_TREE
);
2490 new_pattern_def_seq (stmt_vinfo
, def_stmt
);
2491 def_stmt_info
= new_stmt_vec_info (def_stmt
, loop_vinfo
, bb_vinfo
);
2492 set_vinfo_for_stmt (def_stmt
, def_stmt_info
);
2493 STMT_VINFO_VECTYPE (def_stmt_info
) = vecitype
;
2494 *type_in
= vecitype
;
2495 *type_out
= vectype
;
2497 if (dump_enabled_p ())
2498 dump_printf_loc (MSG_NOTE
, vect_location
,
2499 "vect_recog_mixed_size_cond_pattern: detected:\n");
2501 return pattern_stmt
;
2505 /* Helper function of vect_recog_bool_pattern. Called recursively, return
2506 true if bool VAR can be optimized that way. */
2509 check_bool_pattern (tree var
, loop_vec_info loop_vinfo
, bb_vec_info bb_vinfo
)
2512 enum vect_def_type dt
;
2514 enum tree_code rhs_code
;
2516 if (!vect_is_simple_use (var
, NULL
, loop_vinfo
, bb_vinfo
, &def_stmt
, &def
,
2520 if (dt
!= vect_internal_def
)
2523 if (!is_gimple_assign (def_stmt
))
2526 if (!has_single_use (def
))
2529 rhs1
= gimple_assign_rhs1 (def_stmt
);
2530 rhs_code
= gimple_assign_rhs_code (def_stmt
);
2534 return check_bool_pattern (rhs1
, loop_vinfo
, bb_vinfo
);
2537 if ((TYPE_PRECISION (TREE_TYPE (rhs1
)) != 1
2538 || !TYPE_UNSIGNED (TREE_TYPE (rhs1
)))
2539 && TREE_CODE (TREE_TYPE (rhs1
)) != BOOLEAN_TYPE
)
2541 return check_bool_pattern (rhs1
, loop_vinfo
, bb_vinfo
);
2544 return check_bool_pattern (rhs1
, loop_vinfo
, bb_vinfo
);
2549 if (!check_bool_pattern (rhs1
, loop_vinfo
, bb_vinfo
))
2551 return check_bool_pattern (gimple_assign_rhs2 (def_stmt
), loop_vinfo
,
2555 if (TREE_CODE_CLASS (rhs_code
) == tcc_comparison
)
2557 tree vecitype
, comp_vectype
;
2559 /* If the comparison can throw, then is_gimple_condexpr will be
2560 false and we can't make a COND_EXPR/VEC_COND_EXPR out of it. */
2561 if (stmt_could_throw_p (def_stmt
))
2564 comp_vectype
= get_vectype_for_scalar_type (TREE_TYPE (rhs1
));
2565 if (comp_vectype
== NULL_TREE
)
2568 if (TREE_CODE (TREE_TYPE (rhs1
)) != INTEGER_TYPE
)
2570 enum machine_mode mode
= TYPE_MODE (TREE_TYPE (rhs1
));
2572 = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode
), 1);
2573 vecitype
= get_vectype_for_scalar_type (itype
);
2574 if (vecitype
== NULL_TREE
)
2578 vecitype
= comp_vectype
;
2579 return expand_vec_cond_expr_p (vecitype
, comp_vectype
);
2586 /* Helper function of adjust_bool_pattern. Add a cast to TYPE to a previous
2587 stmt (SSA_NAME_DEF_STMT of VAR) by moving the COND_EXPR from RELATED_STMT
2588 to PATTERN_DEF_SEQ and adding a cast as RELATED_STMT. */
2591 adjust_bool_pattern_cast (tree type
, tree var
)
2593 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (SSA_NAME_DEF_STMT (var
));
2594 gimple cast_stmt
, pattern_stmt
;
2596 gcc_assert (!STMT_VINFO_PATTERN_DEF_SEQ (stmt_vinfo
));
2597 pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_vinfo
);
2598 new_pattern_def_seq (stmt_vinfo
, pattern_stmt
);
2600 = gimple_build_assign_with_ops (NOP_EXPR
,
2601 vect_recog_temp_ssa_var (type
, NULL
),
2602 gimple_assign_lhs (pattern_stmt
),
2604 STMT_VINFO_RELATED_STMT (stmt_vinfo
) = cast_stmt
;
2605 return gimple_assign_lhs (cast_stmt
);
2609 /* Helper function of vect_recog_bool_pattern. Do the actual transformations,
2610 recursively. VAR is an SSA_NAME that should be transformed from bool
2611 to a wider integer type, OUT_TYPE is the desired final integer type of
2612 the whole pattern, TRUEVAL should be NULL unless optimizing
2613 BIT_AND_EXPR into a COND_EXPR with one integer from one of the operands
2614 in the then_clause, STMTS is where statements with added pattern stmts
2615 should be pushed to. */
2618 adjust_bool_pattern (tree var
, tree out_type
, tree trueval
,
2621 gimple stmt
= SSA_NAME_DEF_STMT (var
);
2622 enum tree_code rhs_code
, def_rhs_code
;
2623 tree itype
, cond_expr
, rhs1
, rhs2
, irhs1
, irhs2
;
2625 gimple pattern_stmt
, def_stmt
;
2627 rhs1
= gimple_assign_rhs1 (stmt
);
2628 rhs2
= gimple_assign_rhs2 (stmt
);
2629 rhs_code
= gimple_assign_rhs_code (stmt
);
2630 loc
= gimple_location (stmt
);
2635 irhs1
= adjust_bool_pattern (rhs1
, out_type
, NULL_TREE
, stmts
);
2636 itype
= TREE_TYPE (irhs1
);
2638 = gimple_build_assign_with_ops (SSA_NAME
,
2639 vect_recog_temp_ssa_var (itype
, NULL
),
2644 irhs1
= adjust_bool_pattern (rhs1
, out_type
, NULL_TREE
, stmts
);
2645 itype
= TREE_TYPE (irhs1
);
2647 = gimple_build_assign_with_ops (BIT_XOR_EXPR
,
2648 vect_recog_temp_ssa_var (itype
, NULL
),
2649 irhs1
, build_int_cst (itype
, 1));
2653 /* Try to optimize x = y & (a < b ? 1 : 0); into
2654 x = (a < b ? y : 0);
2660 S1 a_b = x1 CMP1 y1;
2661 S2 b_b = x2 CMP2 y2;
2663 S4 d_T = (TYPE) c_b;
2665 we would normally emit:
2667 S1' a_T = x1 CMP1 y1 ? 1 : 0;
2668 S2' b_T = x2 CMP2 y2 ? 1 : 0;
2669 S3' c_T = a_T & b_T;
2672 but we can save one stmt by using the
2673 result of one of the COND_EXPRs in the other COND_EXPR and leave
2674 BIT_AND_EXPR stmt out:
2676 S1' a_T = x1 CMP1 y1 ? 1 : 0;
2677 S3' c_T = x2 CMP2 y2 ? a_T : 0;
2680 At least when VEC_COND_EXPR is implemented using masks
2681 cond ? 1 : 0 is as expensive as cond ? var : 0, in both cases it
2682 computes the comparison masks and ands it, in one case with
2683 all ones vector, in the other case with a vector register.
2684 Don't do this for BIT_IOR_EXPR, because cond ? 1 : var; is
2685 often more expensive. */
2686 def_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2687 def_rhs_code
= gimple_assign_rhs_code (def_stmt
);
2688 if (TREE_CODE_CLASS (def_rhs_code
) == tcc_comparison
)
2690 tree def_rhs1
= gimple_assign_rhs1 (def_stmt
);
2691 irhs1
= adjust_bool_pattern (rhs1
, out_type
, NULL_TREE
, stmts
);
2692 if (TYPE_PRECISION (TREE_TYPE (irhs1
))
2693 == GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (def_rhs1
))))
2696 stmt_vec_info stmt_def_vinfo
= vinfo_for_stmt (def_stmt
);
2697 irhs2
= adjust_bool_pattern (rhs2
, out_type
, irhs1
, stmts
);
2698 tstmt
= stmts
->pop ();
2699 gcc_assert (tstmt
== def_stmt
);
2700 stmts
->quick_push (stmt
);
2701 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt
))
2702 = STMT_VINFO_RELATED_STMT (stmt_def_vinfo
);
2703 gcc_assert (!STMT_VINFO_PATTERN_DEF_SEQ (stmt_def_vinfo
));
2704 STMT_VINFO_RELATED_STMT (stmt_def_vinfo
) = NULL
;
2708 irhs2
= adjust_bool_pattern (rhs2
, out_type
, NULL_TREE
, stmts
);
2711 def_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2712 def_rhs_code
= gimple_assign_rhs_code (def_stmt
);
2713 if (TREE_CODE_CLASS (def_rhs_code
) == tcc_comparison
)
2715 tree def_rhs1
= gimple_assign_rhs1 (def_stmt
);
2716 irhs2
= adjust_bool_pattern (rhs2
, out_type
, NULL_TREE
, stmts
);
2717 if (TYPE_PRECISION (TREE_TYPE (irhs2
))
2718 == GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (def_rhs1
))))
2721 stmt_vec_info stmt_def_vinfo
= vinfo_for_stmt (def_stmt
);
2722 irhs1
= adjust_bool_pattern (rhs1
, out_type
, irhs2
, stmts
);
2723 tstmt
= stmts
->pop ();
2724 gcc_assert (tstmt
== def_stmt
);
2725 stmts
->quick_push (stmt
);
2726 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt
))
2727 = STMT_VINFO_RELATED_STMT (stmt_def_vinfo
);
2728 gcc_assert (!STMT_VINFO_PATTERN_DEF_SEQ (stmt_def_vinfo
));
2729 STMT_VINFO_RELATED_STMT (stmt_def_vinfo
) = NULL
;
2733 irhs1
= adjust_bool_pattern (rhs1
, out_type
, NULL_TREE
, stmts
);
2739 irhs1
= adjust_bool_pattern (rhs1
, out_type
, NULL_TREE
, stmts
);
2740 irhs2
= adjust_bool_pattern (rhs2
, out_type
, NULL_TREE
, stmts
);
2742 if (TYPE_PRECISION (TREE_TYPE (irhs1
))
2743 != TYPE_PRECISION (TREE_TYPE (irhs2
)))
2745 int prec1
= TYPE_PRECISION (TREE_TYPE (irhs1
));
2746 int prec2
= TYPE_PRECISION (TREE_TYPE (irhs2
));
2747 int out_prec
= TYPE_PRECISION (out_type
);
2748 if (absu_hwi (out_prec
- prec1
) < absu_hwi (out_prec
- prec2
))
2749 irhs2
= adjust_bool_pattern_cast (TREE_TYPE (irhs1
), rhs2
);
2750 else if (absu_hwi (out_prec
- prec1
) > absu_hwi (out_prec
- prec2
))
2751 irhs1
= adjust_bool_pattern_cast (TREE_TYPE (irhs2
), rhs1
);
2754 irhs1
= adjust_bool_pattern_cast (out_type
, rhs1
);
2755 irhs2
= adjust_bool_pattern_cast (out_type
, rhs2
);
2758 itype
= TREE_TYPE (irhs1
);
2760 = gimple_build_assign_with_ops (rhs_code
,
2761 vect_recog_temp_ssa_var (itype
, NULL
),
2766 gcc_assert (TREE_CODE_CLASS (rhs_code
) == tcc_comparison
);
2767 if (TREE_CODE (TREE_TYPE (rhs1
)) != INTEGER_TYPE
2768 || !TYPE_UNSIGNED (TREE_TYPE (rhs1
))
2769 || (TYPE_PRECISION (TREE_TYPE (rhs1
))
2770 != GET_MODE_BITSIZE (TYPE_MODE (TREE_TYPE (rhs1
)))))
2772 enum machine_mode mode
= TYPE_MODE (TREE_TYPE (rhs1
));
2774 = build_nonstandard_integer_type (GET_MODE_BITSIZE (mode
), 1);
2777 itype
= TREE_TYPE (rhs1
);
2778 cond_expr
= build2_loc (loc
, rhs_code
, itype
, rhs1
, rhs2
);
2779 if (trueval
== NULL_TREE
)
2780 trueval
= build_int_cst (itype
, 1);
2782 gcc_checking_assert (useless_type_conversion_p (itype
,
2783 TREE_TYPE (trueval
)));
2785 = gimple_build_assign_with_ops (COND_EXPR
,
2786 vect_recog_temp_ssa_var (itype
, NULL
),
2788 build_int_cst (itype
, 0));
2792 stmts
->safe_push (stmt
);
2793 gimple_set_location (pattern_stmt
, loc
);
2794 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt
)) = pattern_stmt
;
2795 return gimple_assign_lhs (pattern_stmt
);
2799 /* Function vect_recog_bool_pattern
2801 Try to find pattern like following:
2803 bool a_b, b_b, c_b, d_b, e_b;
2806 S1 a_b = x1 CMP1 y1;
2807 S2 b_b = x2 CMP2 y2;
2809 S4 d_b = x3 CMP3 y3;
2811 S6 f_T = (TYPE) e_b;
2813 where type 'TYPE' is an integral type.
2817 * LAST_STMT: A stmt at the end from which the pattern
2818 search begins, i.e. cast of a bool to
2823 * TYPE_IN: The type of the input arguments to the pattern.
2825 * TYPE_OUT: The type of the output of this pattern.
2827 * Return value: A new stmt that will be used to replace the pattern.
2829 Assuming size of TYPE is the same as size of all comparisons
2830 (otherwise some casts would be added where needed), the above
2831 sequence we create related pattern stmts:
2832 S1' a_T = x1 CMP1 y1 ? 1 : 0;
2833 S3' c_T = x2 CMP2 y2 ? a_T : 0;
2834 S4' d_T = x3 CMP3 y3 ? 1 : 0;
2835 S5' e_T = c_T | d_T;
2838 Instead of the above S3' we could emit:
2839 S2' b_T = x2 CMP2 y2 ? 1 : 0;
2840 S3' c_T = a_T | b_T;
2841 but the above is more efficient. */
2844 vect_recog_bool_pattern (vec
<gimple
> *stmts
, tree
*type_in
,
2847 gimple last_stmt
= stmts
->pop ();
2848 enum tree_code rhs_code
;
2849 tree var
, lhs
, rhs
, vectype
;
2850 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (last_stmt
);
2851 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2852 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (stmt_vinfo
);
2853 gimple pattern_stmt
;
2855 if (!is_gimple_assign (last_stmt
))
2858 var
= gimple_assign_rhs1 (last_stmt
);
2859 lhs
= gimple_assign_lhs (last_stmt
);
2861 if ((TYPE_PRECISION (TREE_TYPE (var
)) != 1
2862 || !TYPE_UNSIGNED (TREE_TYPE (var
)))
2863 && TREE_CODE (TREE_TYPE (var
)) != BOOLEAN_TYPE
)
2866 rhs_code
= gimple_assign_rhs_code (last_stmt
);
2867 if (CONVERT_EXPR_CODE_P (rhs_code
))
2869 if (TREE_CODE (TREE_TYPE (lhs
)) != INTEGER_TYPE
2870 || TYPE_PRECISION (TREE_TYPE (lhs
)) == 1)
2872 vectype
= get_vectype_for_scalar_type (TREE_TYPE (lhs
));
2873 if (vectype
== NULL_TREE
)
2876 if (!check_bool_pattern (var
, loop_vinfo
, bb_vinfo
))
2879 rhs
= adjust_bool_pattern (var
, TREE_TYPE (lhs
), NULL_TREE
, stmts
);
2880 lhs
= vect_recog_temp_ssa_var (TREE_TYPE (lhs
), NULL
);
2881 if (useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2883 = gimple_build_assign_with_ops (SSA_NAME
, lhs
, rhs
, NULL_TREE
);
2886 = gimple_build_assign_with_ops (NOP_EXPR
, lhs
, rhs
, NULL_TREE
);
2887 *type_out
= vectype
;
2889 stmts
->safe_push (last_stmt
);
2890 if (dump_enabled_p ())
2891 dump_printf_loc (MSG_NOTE
, vect_location
,
2892 "vect_recog_bool_pattern: detected:\n");
2894 return pattern_stmt
;
2896 else if (rhs_code
== SSA_NAME
2897 && STMT_VINFO_DATA_REF (stmt_vinfo
))
2899 stmt_vec_info pattern_stmt_info
;
2900 vectype
= STMT_VINFO_VECTYPE (stmt_vinfo
);
2901 gcc_assert (vectype
!= NULL_TREE
);
2902 if (!VECTOR_MODE_P (TYPE_MODE (vectype
)))
2904 if (!check_bool_pattern (var
, loop_vinfo
, bb_vinfo
))
2907 rhs
= adjust_bool_pattern (var
, TREE_TYPE (vectype
), NULL_TREE
, stmts
);
2908 lhs
= build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (vectype
), lhs
);
2909 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2911 tree rhs2
= vect_recog_temp_ssa_var (TREE_TYPE (lhs
), NULL
);
2913 = gimple_build_assign_with_ops (NOP_EXPR
, rhs2
, rhs
, NULL_TREE
);
2914 new_pattern_def_seq (stmt_vinfo
, cast_stmt
);
2918 = gimple_build_assign_with_ops (SSA_NAME
, lhs
, rhs
, NULL_TREE
);
2919 pattern_stmt_info
= new_stmt_vec_info (pattern_stmt
, loop_vinfo
,
2921 set_vinfo_for_stmt (pattern_stmt
, pattern_stmt_info
);
2922 STMT_VINFO_DATA_REF (pattern_stmt_info
)
2923 = STMT_VINFO_DATA_REF (stmt_vinfo
);
2924 STMT_VINFO_DR_BASE_ADDRESS (pattern_stmt_info
)
2925 = STMT_VINFO_DR_BASE_ADDRESS (stmt_vinfo
);
2926 STMT_VINFO_DR_INIT (pattern_stmt_info
) = STMT_VINFO_DR_INIT (stmt_vinfo
);
2927 STMT_VINFO_DR_OFFSET (pattern_stmt_info
)
2928 = STMT_VINFO_DR_OFFSET (stmt_vinfo
);
2929 STMT_VINFO_DR_STEP (pattern_stmt_info
) = STMT_VINFO_DR_STEP (stmt_vinfo
);
2930 STMT_VINFO_DR_ALIGNED_TO (pattern_stmt_info
)
2931 = STMT_VINFO_DR_ALIGNED_TO (stmt_vinfo
);
2932 DR_STMT (STMT_VINFO_DATA_REF (stmt_vinfo
)) = pattern_stmt
;
2933 *type_out
= vectype
;
2935 stmts
->safe_push (last_stmt
);
2936 if (dump_enabled_p ())
2937 dump_printf_loc (MSG_NOTE
, vect_location
,
2938 "vect_recog_bool_pattern: detected:\n");
2939 return pattern_stmt
;
2946 /* Mark statements that are involved in a pattern. */
2949 vect_mark_pattern_stmts (gimple orig_stmt
, gimple pattern_stmt
,
2950 tree pattern_vectype
)
2952 stmt_vec_info pattern_stmt_info
, def_stmt_info
;
2953 stmt_vec_info orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
2954 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (orig_stmt_info
);
2955 bb_vec_info bb_vinfo
= STMT_VINFO_BB_VINFO (orig_stmt_info
);
2958 pattern_stmt_info
= vinfo_for_stmt (pattern_stmt
);
2959 if (pattern_stmt_info
== NULL
)
2961 pattern_stmt_info
= new_stmt_vec_info (pattern_stmt
, loop_vinfo
,
2963 set_vinfo_for_stmt (pattern_stmt
, pattern_stmt_info
);
2965 gimple_set_bb (pattern_stmt
, gimple_bb (orig_stmt
));
2967 STMT_VINFO_RELATED_STMT (pattern_stmt_info
) = orig_stmt
;
2968 STMT_VINFO_DEF_TYPE (pattern_stmt_info
)
2969 = STMT_VINFO_DEF_TYPE (orig_stmt_info
);
2970 STMT_VINFO_VECTYPE (pattern_stmt_info
) = pattern_vectype
;
2971 STMT_VINFO_IN_PATTERN_P (orig_stmt_info
) = true;
2972 STMT_VINFO_RELATED_STMT (orig_stmt_info
) = pattern_stmt
;
2973 STMT_VINFO_PATTERN_DEF_SEQ (pattern_stmt_info
)
2974 = STMT_VINFO_PATTERN_DEF_SEQ (orig_stmt_info
);
2975 if (STMT_VINFO_PATTERN_DEF_SEQ (pattern_stmt_info
))
2977 gimple_stmt_iterator si
;
2978 for (si
= gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (pattern_stmt_info
));
2979 !gsi_end_p (si
); gsi_next (&si
))
2981 def_stmt
= gsi_stmt (si
);
2982 def_stmt_info
= vinfo_for_stmt (def_stmt
);
2983 if (def_stmt_info
== NULL
)
2985 def_stmt_info
= new_stmt_vec_info (def_stmt
, loop_vinfo
,
2987 set_vinfo_for_stmt (def_stmt
, def_stmt_info
);
2989 gimple_set_bb (def_stmt
, gimple_bb (orig_stmt
));
2990 STMT_VINFO_RELATED_STMT (def_stmt_info
) = orig_stmt
;
2991 STMT_VINFO_DEF_TYPE (def_stmt_info
)
2992 = STMT_VINFO_DEF_TYPE (orig_stmt_info
);
2993 if (STMT_VINFO_VECTYPE (def_stmt_info
) == NULL_TREE
)
2994 STMT_VINFO_VECTYPE (def_stmt_info
) = pattern_vectype
;
2999 /* Function vect_pattern_recog_1
3002 PATTERN_RECOG_FUNC: A pointer to a function that detects a certain
3003 computation pattern.
3004 STMT: A stmt from which the pattern search should start.
3006 If PATTERN_RECOG_FUNC successfully detected the pattern, it creates an
3007 expression that computes the same functionality and can be used to
3008 replace the sequence of stmts that are involved in the pattern.
3011 This function checks if the expression returned by PATTERN_RECOG_FUNC is
3012 supported in vector form by the target. We use 'TYPE_IN' to obtain the
3013 relevant vector type. If 'TYPE_IN' is already a vector type, then this
3014 indicates that target support had already been checked by PATTERN_RECOG_FUNC.
3015 If 'TYPE_OUT' is also returned by PATTERN_RECOG_FUNC, we check that it fits
3016 to the available target pattern.
3018 This function also does some bookkeeping, as explained in the documentation
3019 for vect_recog_pattern. */
3022 vect_pattern_recog_1 (vect_recog_func_ptr vect_recog_func
,
3023 gimple_stmt_iterator si
,
3024 vec
<gimple
> *stmts_to_replace
)
3026 gimple stmt
= gsi_stmt (si
), pattern_stmt
;
3027 stmt_vec_info stmt_info
;
3028 loop_vec_info loop_vinfo
;
3029 tree pattern_vectype
;
3030 tree type_in
, type_out
;
3031 enum tree_code code
;
3035 stmts_to_replace
->truncate (0);
3036 stmts_to_replace
->quick_push (stmt
);
3037 pattern_stmt
= (* vect_recog_func
) (stmts_to_replace
, &type_in
, &type_out
);
3041 stmt
= stmts_to_replace
->last ();
3042 stmt_info
= vinfo_for_stmt (stmt
);
3043 loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3045 if (VECTOR_MODE_P (TYPE_MODE (type_in
)))
3047 /* No need to check target support (already checked by the pattern
3048 recognition function). */
3049 pattern_vectype
= type_out
? type_out
: type_in
;
3053 enum machine_mode vec_mode
;
3054 enum insn_code icode
;
3057 /* Check target support */
3058 type_in
= get_vectype_for_scalar_type (type_in
);
3062 type_out
= get_vectype_for_scalar_type (type_out
);
3067 pattern_vectype
= type_out
;
3069 if (is_gimple_assign (pattern_stmt
))
3070 code
= gimple_assign_rhs_code (pattern_stmt
);
3073 gcc_assert (is_gimple_call (pattern_stmt
));
3077 optab
= optab_for_tree_code (code
, type_in
, optab_default
);
3078 vec_mode
= TYPE_MODE (type_in
);
3080 || (icode
= optab_handler (optab
, vec_mode
)) == CODE_FOR_nothing
3081 || (insn_data
[icode
].operand
[0].mode
!= TYPE_MODE (type_out
)))
3085 /* Found a vectorizable pattern. */
3086 if (dump_enabled_p ())
3088 dump_printf_loc (MSG_NOTE
, vect_location
,
3089 "pattern recognized: ");
3090 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
3091 dump_printf (MSG_NOTE
, "\n");
3094 /* Mark the stmts that are involved in the pattern. */
3095 vect_mark_pattern_stmts (stmt
, pattern_stmt
, pattern_vectype
);
3097 /* Patterns cannot be vectorized using SLP, because they change the order of
3100 FOR_EACH_VEC_ELT (LOOP_VINFO_REDUCTIONS (loop_vinfo
), i
, next
)
3102 LOOP_VINFO_REDUCTIONS (loop_vinfo
).ordered_remove (i
);
3104 /* It is possible that additional pattern stmts are created and inserted in
3105 STMTS_TO_REPLACE. We create a stmt_info for each of them, and mark the
3106 relevant statements. */
3107 for (i
= 0; stmts_to_replace
->iterate (i
, &stmt
)
3108 && (unsigned) i
< (stmts_to_replace
->length () - 1);
3111 stmt_info
= vinfo_for_stmt (stmt
);
3112 pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
3113 if (dump_enabled_p ())
3115 dump_printf_loc (MSG_NOTE
, vect_location
,
3116 "additional pattern stmt: ");
3117 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, pattern_stmt
, 0);
3118 dump_printf (MSG_NOTE
, "\n");
3121 vect_mark_pattern_stmts (stmt
, pattern_stmt
, NULL_TREE
);
3126 /* Function vect_pattern_recog
3129 LOOP_VINFO - a struct_loop_info of a loop in which we want to look for
3132 Output - for each computation idiom that is detected we create a new stmt
3133 that provides the same functionality and that can be vectorized. We
3134 also record some information in the struct_stmt_info of the relevant
3135 stmts, as explained below:
3137 At the entry to this function we have the following stmts, with the
3138 following initial value in the STMT_VINFO fields:
3140 stmt in_pattern_p related_stmt vec_stmt
3141 S1: a_i = .... - - -
3142 S2: a_2 = ..use(a_i).. - - -
3143 S3: a_1 = ..use(a_2).. - - -
3144 S4: a_0 = ..use(a_1).. - - -
3145 S5: ... = ..use(a_0).. - - -
3147 Say the sequence {S1,S2,S3,S4} was detected as a pattern that can be
3148 represented by a single stmt. We then:
3149 - create a new stmt S6 equivalent to the pattern (the stmt is not
3150 inserted into the code)
3151 - fill in the STMT_VINFO fields as follows:
3153 in_pattern_p related_stmt vec_stmt
3154 S1: a_i = .... - - -
3155 S2: a_2 = ..use(a_i).. - - -
3156 S3: a_1 = ..use(a_2).. - - -
3157 S4: a_0 = ..use(a_1).. true S6 -
3158 '---> S6: a_new = .... - S4 -
3159 S5: ... = ..use(a_0).. - - -
3161 (the last stmt in the pattern (S4) and the new pattern stmt (S6) point
3162 to each other through the RELATED_STMT field).
3164 S6 will be marked as relevant in vect_mark_stmts_to_be_vectorized instead
3165 of S4 because it will replace all its uses. Stmts {S1,S2,S3} will
3166 remain irrelevant unless used by stmts other than S4.
3168 If vectorization succeeds, vect_transform_stmt will skip over {S1,S2,S3}
3169 (because they are marked as irrelevant). It will vectorize S6, and record
3170 a pointer to the new vector stmt VS6 from S6 (as usual).
3171 S4 will be skipped, and S5 will be vectorized as usual:
3173 in_pattern_p related_stmt vec_stmt
3174 S1: a_i = .... - - -
3175 S2: a_2 = ..use(a_i).. - - -
3176 S3: a_1 = ..use(a_2).. - - -
3177 > VS6: va_new = .... - - -
3178 S4: a_0 = ..use(a_1).. true S6 VS6
3179 '---> S6: a_new = .... - S4 VS6
3180 > VS5: ... = ..vuse(va_new).. - - -
3181 S5: ... = ..use(a_0).. - - -
3183 DCE could then get rid of {S1,S2,S3,S4,S5} (if their defs are not used
3184 elsewhere), and we'll end up with:
3187 VS5: ... = ..vuse(va_new)..
3189 In case of more than one pattern statements, e.g., widen-mult with
3193 S2 a_T = (TYPE) a_t;
3194 '--> S3: a_it = (interm_type) a_t;
3195 S4 prod_T = a_T * CONST;
3196 '--> S5: prod_T' = a_it w* CONST;
3198 there may be other users of a_T outside the pattern. In that case S2 will
3199 be marked as relevant (as well as S3), and both S2 and S3 will be analyzed
3200 and vectorized. The vector stmt VS2 will be recorded in S2, and VS3 will
3201 be recorded in S3. */
3204 vect_pattern_recog (loop_vec_info loop_vinfo
, bb_vec_info bb_vinfo
)
3209 gimple_stmt_iterator si
;
3211 vect_recog_func_ptr vect_recog_func
;
3212 stack_vec
<gimple
, 1> stmts_to_replace
;
3215 if (dump_enabled_p ())
3216 dump_printf_loc (MSG_NOTE
, vect_location
,
3217 "=== vect_pattern_recog ===\n");
3221 loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3222 bbs
= LOOP_VINFO_BBS (loop_vinfo
);
3223 nbbs
= loop
->num_nodes
;
3227 bbs
= &BB_VINFO_BB (bb_vinfo
);
3231 /* Scan through the loop stmts, applying the pattern recognition
3232 functions starting at each stmt visited: */
3233 for (i
= 0; i
< nbbs
; i
++)
3235 basic_block bb
= bbs
[i
];
3236 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
3238 if (bb_vinfo
&& (stmt
= gsi_stmt (si
))
3239 && vinfo_for_stmt (stmt
)
3240 && !STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (stmt
)))
3243 /* Scan over all generic vect_recog_xxx_pattern functions. */
3244 for (j
= 0; j
< NUM_PATTERNS
; j
++)
3246 vect_recog_func
= vect_vect_recog_func_ptrs
[j
];
3247 vect_pattern_recog_1 (vect_recog_func
, si
,