2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software
4 Contributed by Dorit Naishlos <dorit@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* Loop and basic block vectorizer.
24 This file contains drivers for the three vectorizers:
25 (1) loop vectorizer (inter-iteration parallelism),
26 (2) loop-aware SLP (intra-iteration parallelism) (invoked by the loop
28 (3) BB vectorizer (out-of-loops), aka SLP
30 The rest of the vectorizer's code is organized as follows:
31 - tree-vect-loop.c - loop specific parts such as reductions, etc. These are
32 used by drivers (1) and (2).
33 - tree-vect-loop-manip.c - vectorizer's loop control-flow utilities, used by
35 - tree-vect-slp.c - BB vectorization specific analysis and transformation,
36 used by drivers (2) and (3).
37 - tree-vect-stmts.c - statements analysis and transformation (used by all).
38 - tree-vect-data-refs.c - vectorizer specific data-refs analysis and
39 manipulations (used by all).
40 - tree-vect-patterns.c - vectorizable code patterns detector (used by all)
42 Here's a poor attempt at illustrating that:
45 loop_vect() loop_aware_slp() slp_vect()
48 tree-vect-loop.c tree-vect-slp.c
53 tree-vect-stmts.c tree-vect-data-refs.c
60 #include "coretypes.h"
64 #include "diagnostic.h"
65 #include "tree-flow.h"
66 #include "tree-dump.h"
68 #include "cfglayout.h"
69 #include "tree-vectorizer.h"
70 #include "tree-pass.h"
72 /* vect_dump will be set to stderr or dump_file if exist. */
75 /* vect_verbosity_level set to an invalid value
76 to mark that it's uninitialized. */
77 static enum verbosity_levels vect_verbosity_level
= MAX_VERBOSITY_LEVEL
;
82 /* Bitmap of virtual variables to be renamed. */
83 bitmap vect_memsyms_to_rename
;
85 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
86 VEC(vec_void_p
,heap
) *stmt_vec_info_vec
;
90 /* Function vect_set_verbosity_level.
92 Called from opts.c upon detection of the
93 -ftree-vectorizer-verbose=N option. */
96 vect_set_verbosity_level (const char *val
)
101 if (vl
< MAX_VERBOSITY_LEVEL
)
102 vect_verbosity_level
= (enum verbosity_levels
) vl
;
104 vect_verbosity_level
= (enum verbosity_levels
) (MAX_VERBOSITY_LEVEL
- 1);
108 /* Function vect_set_dump_settings.
110 Fix the verbosity level of the vectorizer if the
111 requested level was not set explicitly using the flag
112 -ftree-vectorizer-verbose=N.
113 Decide where to print the debugging information (dump_file/stderr).
114 If the user defined the verbosity level, but there is no dump file,
115 print to stderr, otherwise print to the dump file. */
118 vect_set_dump_settings (void)
120 vect_dump
= dump_file
;
122 /* Check if the verbosity level was defined by the user: */
123 if (vect_verbosity_level
!= MAX_VERBOSITY_LEVEL
)
125 /* If there is no dump file, print to stderr. */
131 /* User didn't specify verbosity level: */
132 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
133 vect_verbosity_level
= REPORT_DETAILS
;
134 else if (dump_file
&& (dump_flags
& TDF_STATS
))
135 vect_verbosity_level
= REPORT_UNVECTORIZED_LOCATIONS
;
137 vect_verbosity_level
= REPORT_NONE
;
139 gcc_assert (dump_file
|| vect_verbosity_level
== REPORT_NONE
);
143 /* Function debug_loop_details.
145 For vectorization debug dumps. */
148 vect_print_dump_info (enum verbosity_levels vl
)
150 if (vl
> vect_verbosity_level
)
153 if (!current_function_decl
|| !vect_dump
)
156 if (vect_location
== UNKNOWN_LOC
)
157 fprintf (vect_dump
, "\n%s:%d: note: ",
158 DECL_SOURCE_FILE (current_function_decl
),
159 DECL_SOURCE_LINE (current_function_decl
));
161 fprintf (vect_dump
, "\n%s:%d: note: ",
162 LOC_FILE (vect_location
), LOC_LINE (vect_location
));
168 /* Function vectorize_loops.
170 Entry point to loop vectorization phase. */
173 vectorize_loops (void)
176 unsigned int num_vectorized_loops
= 0;
177 unsigned int vect_loops_num
;
181 vect_loops_num
= number_of_loops ();
183 /* Bail out if there are no loops. */
184 if (vect_loops_num
<= 1)
187 /* Fix the verbosity level if not defined explicitly by the user. */
188 vect_set_dump_settings ();
190 /* Allocate the bitmap that records which virtual variables
191 need to be renamed. */
192 vect_memsyms_to_rename
= BITMAP_ALLOC (NULL
);
194 init_stmt_vec_info_vec ();
196 /* ----------- Analyze loops. ----------- */
198 /* If some loop was duplicated, it gets bigger number
199 than all previously defined loops. This fact allows us to run
200 only over initial loops skipping newly generated ones. */
201 FOR_EACH_LOOP (li
, loop
, 0)
202 if (optimize_loop_nest_for_speed_p (loop
))
204 loop_vec_info loop_vinfo
;
206 vect_location
= find_loop_location (loop
);
207 loop_vinfo
= vect_analyze_loop (loop
);
208 loop
->aux
= loop_vinfo
;
210 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
213 vect_transform_loop (loop_vinfo
);
214 num_vectorized_loops
++;
217 vect_location
= UNKNOWN_LOC
;
219 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
220 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
)
221 || (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS
)
222 && num_vectorized_loops
> 0))
223 fprintf (vect_dump
, "vectorized %u loops in function.\n",
224 num_vectorized_loops
);
226 /* ----------- Finalize. ----------- */
228 BITMAP_FREE (vect_memsyms_to_rename
);
230 for (i
= 1; i
< vect_loops_num
; i
++)
232 loop_vec_info loop_vinfo
;
237 loop_vinfo
= (loop_vec_info
) loop
->aux
;
238 destroy_loop_vec_info (loop_vinfo
, true);
242 free_stmt_vec_info_vec ();
244 return num_vectorized_loops
> 0 ? TODO_cleanup_cfg
: 0;
248 /* Increase alignment of global arrays to improve vectorization potential.
250 - Consider also structs that have an array field.
251 - Use ipa analysis to prune arrays that can't be vectorized?
252 This should involve global alignment analysis and in the future also
256 increase_alignment (void)
258 struct varpool_node
*vnode
;
260 /* Increase the alignment of all global arrays for vectorization. */
261 for (vnode
= varpool_nodes_queue
;
263 vnode
= vnode
->next_needed
)
265 tree vectype
, decl
= vnode
->decl
;
266 unsigned int alignment
;
268 if (TREE_CODE (TREE_TYPE (decl
)) != ARRAY_TYPE
)
270 vectype
= get_vectype_for_scalar_type (TREE_TYPE (TREE_TYPE (decl
)));
273 alignment
= TYPE_ALIGN (vectype
);
274 if (DECL_ALIGN (decl
) >= alignment
)
277 if (vect_can_force_dr_alignment_p (decl
, alignment
))
279 DECL_ALIGN (decl
) = TYPE_ALIGN (vectype
);
280 DECL_USER_ALIGN (decl
) = 1;
283 fprintf (dump_file
, "Increasing alignment of decl: ");
284 print_generic_expr (dump_file
, decl
, TDF_SLIM
);
293 gate_increase_alignment (void)
295 return flag_section_anchors
&& flag_tree_vectorize
;
299 struct simple_ipa_opt_pass pass_ipa_increase_alignment
=
303 "increase_alignment", /* name */
304 gate_increase_alignment
, /* gate */
305 increase_alignment
, /* execute */
308 0, /* static_pass_number */
310 0, /* properties_required */
311 0, /* properties_provided */
312 0, /* properties_destroyed */
313 0, /* todo_flags_start */
314 0 /* todo_flags_finish */