Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-vect-loop-manip.c @ 158:494b0b89df80 default tip
...
author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 25 May 2020 18:13:55 +0900 |
parents | 1830386684a0 |
children |
rev | line source |
---|---|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1 /* Vectorizer Specific Loop Manipulations |
145 | 2 Copyright (C) 2003-2020 Free Software Foundation, Inc. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3 Contributed by Dorit Naishlos <dorit@il.ibm.com> |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
4 and Ira Rosen <irar@il.ibm.com> |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
5 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
6 This file is part of GCC. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
7 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
8 GCC is free software; you can redistribute it and/or modify it under |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
9 the terms of the GNU General Public License as published by the Free |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
10 Software Foundation; either version 3, or (at your option) any later |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
11 version. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
12 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
16 for more details. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
17 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
18 You should have received a copy of the GNU General Public License |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
19 along with GCC; see the file COPYING3. If not see |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
20 <http://www.gnu.org/licenses/>. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
21 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
22 #include "config.h" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
23 #include "system.h" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
24 #include "coretypes.h" |
111 | 25 #include "backend.h" |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
26 #include "tree.h" |
111 | 27 #include "gimple.h" |
28 #include "cfghooks.h" | |
29 #include "tree-pass.h" | |
30 #include "ssa.h" | |
31 #include "fold-const.h" | |
32 #include "cfganal.h" | |
33 #include "gimplify.h" | |
34 #include "gimple-iterator.h" | |
35 #include "gimplify-me.h" | |
36 #include "tree-cfg.h" | |
37 #include "tree-ssa-loop-manip.h" | |
38 #include "tree-into-ssa.h" | |
39 #include "tree-ssa.h" | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
40 #include "cfgloop.h" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
41 #include "tree-scalar-evolution.h" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
42 #include "tree-vectorizer.h" |
111 | 43 #include "tree-ssa-loop-ivopts.h" |
131 | 44 #include "gimple-fold.h" |
45 #include "tree-ssa-loop-niter.h" | |
46 #include "internal-fn.h" | |
47 #include "stor-layout.h" | |
48 #include "optabs-query.h" | |
49 #include "vec-perm-indices.h" | |
145 | 50 #include "insn-config.h" |
51 #include "rtl.h" | |
52 #include "recog.h" | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
53 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
54 /************************************************************************* |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
55 Simple Loop Peeling Utilities |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
56 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
57 Utilities to support loop peeling for vectorization purposes. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
58 *************************************************************************/ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
59 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
60 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
61 /* Renames the use *OP_P. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
62 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
63 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
64 rename_use_op (use_operand_p op_p) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
65 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
66 tree new_name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
67 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
68 if (TREE_CODE (USE_FROM_PTR (op_p)) != SSA_NAME) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
69 return; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
70 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
71 new_name = get_current_def (USE_FROM_PTR (op_p)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
72 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
73 /* Something defined outside of the loop. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
74 if (!new_name) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
75 return; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
76 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
77 /* An ordinary ssa name defined in the loop. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
78 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
79 SET_USE (op_p, new_name); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
80 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
81 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
82 |
111 | 83 /* Renames the variables in basic block BB. Allow renaming of PHI arguments |
84 on edges incoming from outer-block header if RENAME_FROM_OUTER_LOOP is | |
85 true. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
86 |
111 | 87 static void |
88 rename_variables_in_bb (basic_block bb, bool rename_from_outer_loop) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
89 { |
111 | 90 gimple *stmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
91 use_operand_p use_p; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
92 ssa_op_iter iter; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
93 edge e; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
94 edge_iterator ei; |
145 | 95 class loop *loop = bb->loop_father; |
96 class loop *outer_loop = NULL; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
97 |
111 | 98 if (rename_from_outer_loop) |
99 { | |
100 gcc_assert (loop); | |
101 outer_loop = loop_outer (loop); | |
102 } | |
103 | |
104 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); | |
105 gsi_next (&gsi)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
106 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
107 stmt = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
108 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
109 rename_use_op (use_p); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
110 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
111 |
111 | 112 FOR_EACH_EDGE (e, ei, bb->preds) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
113 { |
111 | 114 if (!flow_bb_inside_loop_p (loop, e->src)) |
115 { | |
116 if (!rename_from_outer_loop) | |
117 continue; | |
118 if (e->src != outer_loop->header) | |
119 { | |
120 if (outer_loop->inner->next) | |
121 { | |
122 /* If outer_loop has 2 inner loops, allow there to | |
123 be an extra basic block which decides which of the | |
124 two loops to use using LOOP_VECTORIZED. */ | |
125 if (!single_pred_p (e->src) | |
126 || single_pred (e->src) != outer_loop->header) | |
127 continue; | |
128 } | |
129 } | |
130 } | |
131 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi); | |
132 gsi_next (&gsi)) | |
133 rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi.phi (), e)); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
134 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
135 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
136 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
137 |
111 | 138 struct adjust_info |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
139 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
140 tree from, to; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
141 basic_block bb; |
111 | 142 }; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
143 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
144 /* A stack of values to be adjusted in debug stmts. We have to |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
145 process them LIFO, so that the closest substitution applies. If we |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
146 processed them FIFO, without the stack, we might substitute uses |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
147 with a PHI DEF that would soon become non-dominant, and when we got |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
148 to the suitable one, it wouldn't have anything to substitute any |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
149 more. */ |
111 | 150 static vec<adjust_info, va_heap> adjust_vec; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
151 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
152 /* Adjust any debug stmts that referenced AI->from values to use the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
153 loop-closed AI->to, if the references are dominated by AI->bb and |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
154 not by the definition of AI->from. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
155 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
156 static void |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
157 adjust_debug_stmts_now (adjust_info *ai) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
158 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
159 basic_block bbphi = ai->bb; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
160 tree orig_def = ai->from; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
161 tree new_def = ai->to; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
162 imm_use_iterator imm_iter; |
111 | 163 gimple *stmt; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
164 basic_block bbdef = gimple_bb (SSA_NAME_DEF_STMT (orig_def)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
165 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
166 gcc_assert (dom_info_available_p (CDI_DOMINATORS)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
167 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
168 /* Adjust any debug stmts that held onto non-loop-closed |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
169 references. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
170 FOR_EACH_IMM_USE_STMT (stmt, imm_iter, orig_def) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
171 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
172 use_operand_p use_p; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
173 basic_block bbuse; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
174 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
175 if (!is_gimple_debug (stmt)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
176 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
177 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
178 gcc_assert (gimple_debug_bind_p (stmt)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
179 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
180 bbuse = gimple_bb (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
181 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
182 if ((bbuse == bbphi |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
183 || dominated_by_p (CDI_DOMINATORS, bbuse, bbphi)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
184 && !(bbuse == bbdef |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
185 || dominated_by_p (CDI_DOMINATORS, bbuse, bbdef))) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
186 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
187 if (new_def) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
188 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
189 SET_USE (use_p, new_def); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
190 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
191 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
192 gimple_debug_bind_reset_value (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
193 update_stmt (stmt); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
194 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
195 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
196 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
197 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
198 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
199 /* Adjust debug stmts as scheduled before. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
200 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
201 static void |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
202 adjust_vec_debug_stmts (void) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
203 { |
131 | 204 if (!MAY_HAVE_DEBUG_BIND_STMTS) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
205 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
206 |
111 | 207 gcc_assert (adjust_vec.exists ()); |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
208 |
111 | 209 while (!adjust_vec.is_empty ()) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
210 { |
111 | 211 adjust_debug_stmts_now (&adjust_vec.last ()); |
212 adjust_vec.pop (); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
213 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
214 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
215 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
216 /* Adjust any debug stmts that referenced FROM values to use the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
217 loop-closed TO, if the references are dominated by BB and not by |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
218 the definition of FROM. If adjust_vec is non-NULL, adjustments |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
219 will be postponed until adjust_vec_debug_stmts is called. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
220 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
221 static void |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
222 adjust_debug_stmts (tree from, tree to, basic_block bb) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
223 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
224 adjust_info ai; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
225 |
131 | 226 if (MAY_HAVE_DEBUG_BIND_STMTS |
111 | 227 && TREE_CODE (from) == SSA_NAME |
228 && ! SSA_NAME_IS_DEFAULT_DEF (from) | |
229 && ! virtual_operand_p (from)) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
230 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
231 ai.from = from; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
232 ai.to = to; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
233 ai.bb = bb; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
234 |
111 | 235 if (adjust_vec.exists ()) |
236 adjust_vec.safe_push (ai); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
237 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
238 adjust_debug_stmts_now (&ai); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
239 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
240 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
241 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
242 /* Change E's phi arg in UPDATE_PHI to NEW_DEF, and record information |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
243 to adjust any debug stmts that referenced the old phi arg, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
244 presumably non-loop-closed references left over from other |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
245 transformations. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
246 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
247 static void |
111 | 248 adjust_phi_and_debug_stmts (gimple *update_phi, edge e, tree new_def) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
249 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
250 tree orig_def = PHI_ARG_DEF_FROM_EDGE (update_phi, e); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
251 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
252 SET_PHI_ARG_DEF (update_phi, e->dest_idx, new_def); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
253 |
131 | 254 if (MAY_HAVE_DEBUG_BIND_STMTS) |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
255 adjust_debug_stmts (orig_def, PHI_RESULT (update_phi), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
256 gimple_bb (update_phi)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
257 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
258 |
131 | 259 /* Define one loop mask MASK from loop LOOP. INIT_MASK is the value that |
260 the mask should have during the first iteration and NEXT_MASK is the | |
261 value that it should have on subsequent iterations. */ | |
262 | |
263 static void | |
145 | 264 vect_set_loop_mask (class loop *loop, tree mask, tree init_mask, |
131 | 265 tree next_mask) |
266 { | |
267 gphi *phi = create_phi_node (mask, loop->header); | |
268 add_phi_arg (phi, init_mask, loop_preheader_edge (loop), UNKNOWN_LOCATION); | |
269 add_phi_arg (phi, next_mask, loop_latch_edge (loop), UNKNOWN_LOCATION); | |
270 } | |
271 | |
272 /* Add SEQ to the end of LOOP's preheader block. */ | |
273 | |
274 static void | |
145 | 275 add_preheader_seq (class loop *loop, gimple_seq seq) |
131 | 276 { |
277 if (seq) | |
278 { | |
279 edge pe = loop_preheader_edge (loop); | |
280 basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
281 gcc_assert (!new_bb); | |
282 } | |
283 } | |
284 | |
285 /* Add SEQ to the beginning of LOOP's header block. */ | |
286 | |
287 static void | |
145 | 288 add_header_seq (class loop *loop, gimple_seq seq) |
131 | 289 { |
290 if (seq) | |
291 { | |
292 gimple_stmt_iterator gsi = gsi_after_labels (loop->header); | |
293 gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
294 } | |
295 } | |
296 | |
297 /* Return true if the target can interleave elements of two vectors. | |
298 OFFSET is 0 if the first half of the vectors should be interleaved | |
299 or 1 if the second half should. When returning true, store the | |
300 associated permutation in INDICES. */ | |
301 | |
302 static bool | |
303 interleave_supported_p (vec_perm_indices *indices, tree vectype, | |
304 unsigned int offset) | |
305 { | |
306 poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (vectype); | |
307 poly_uint64 base = exact_div (nelts, 2) * offset; | |
308 vec_perm_builder sel (nelts, 2, 3); | |
309 for (unsigned int i = 0; i < 3; ++i) | |
310 { | |
311 sel.quick_push (base + i); | |
312 sel.quick_push (base + i + nelts); | |
313 } | |
314 indices->new_vector (sel, 2, nelts); | |
315 return can_vec_perm_const_p (TYPE_MODE (vectype), *indices); | |
316 } | |
317 | |
318 /* Try to use permutes to define the masks in DEST_RGM using the masks | |
319 in SRC_RGM, given that the former has twice as many masks as the | |
320 latter. Return true on success, adding any new statements to SEQ. */ | |
321 | |
322 static bool | |
323 vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_masks *dest_rgm, | |
324 rgroup_masks *src_rgm) | |
325 { | |
326 tree src_masktype = src_rgm->mask_type; | |
327 tree dest_masktype = dest_rgm->mask_type; | |
328 machine_mode src_mode = TYPE_MODE (src_masktype); | |
145 | 329 insn_code icode1, icode2; |
131 | 330 if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter |
145 | 331 && (icode1 = optab_handler (vec_unpacku_hi_optab, |
332 src_mode)) != CODE_FOR_nothing | |
333 && (icode2 = optab_handler (vec_unpacku_lo_optab, | |
334 src_mode)) != CODE_FOR_nothing) | |
131 | 335 { |
336 /* Unpacking the source masks gives at least as many mask bits as | |
337 we need. We can then VIEW_CONVERT any excess bits away. */ | |
145 | 338 machine_mode dest_mode = insn_data[icode1].operand[0].mode; |
339 gcc_assert (dest_mode == insn_data[icode2].operand[0].mode); | |
340 tree unpack_masktype = vect_halve_mask_nunits (src_masktype, dest_mode); | |
131 | 341 for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i) |
342 { | |
343 tree src = src_rgm->masks[i / 2]; | |
344 tree dest = dest_rgm->masks[i]; | |
345 tree_code code = ((i & 1) == (BYTES_BIG_ENDIAN ? 0 : 1) | |
346 ? VEC_UNPACK_HI_EXPR | |
347 : VEC_UNPACK_LO_EXPR); | |
348 gassign *stmt; | |
349 if (dest_masktype == unpack_masktype) | |
350 stmt = gimple_build_assign (dest, code, src); | |
351 else | |
352 { | |
353 tree temp = make_ssa_name (unpack_masktype); | |
354 stmt = gimple_build_assign (temp, code, src); | |
355 gimple_seq_add_stmt (seq, stmt); | |
356 stmt = gimple_build_assign (dest, VIEW_CONVERT_EXPR, | |
357 build1 (VIEW_CONVERT_EXPR, | |
358 dest_masktype, temp)); | |
359 } | |
360 gimple_seq_add_stmt (seq, stmt); | |
361 } | |
362 return true; | |
363 } | |
364 vec_perm_indices indices[2]; | |
365 if (dest_masktype == src_masktype | |
366 && interleave_supported_p (&indices[0], src_masktype, 0) | |
367 && interleave_supported_p (&indices[1], src_masktype, 1)) | |
368 { | |
369 /* The destination requires twice as many mask bits as the source, so | |
370 we can use interleaving permutes to double up the number of bits. */ | |
371 tree masks[2]; | |
372 for (unsigned int i = 0; i < 2; ++i) | |
373 masks[i] = vect_gen_perm_mask_checked (src_masktype, indices[i]); | |
374 for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i) | |
375 { | |
376 tree src = src_rgm->masks[i / 2]; | |
377 tree dest = dest_rgm->masks[i]; | |
378 gimple *stmt = gimple_build_assign (dest, VEC_PERM_EXPR, | |
379 src, src, masks[i & 1]); | |
380 gimple_seq_add_stmt (seq, stmt); | |
381 } | |
382 return true; | |
383 } | |
384 return false; | |
385 } | |
386 | |
387 /* Helper for vect_set_loop_condition_masked. Generate definitions for | |
388 all the masks in RGM and return a mask that is nonzero when the loop | |
389 needs to iterate. Add any new preheader statements to PREHEADER_SEQ. | |
390 Use LOOP_COND_GSI to insert code before the exit gcond. | |
391 | |
392 RGM belongs to loop LOOP. The loop originally iterated NITERS | |
145 | 393 times and has been vectorized according to LOOP_VINFO. |
131 | 394 |
395 If NITERS_SKIP is nonnull, the first iteration of the vectorized loop | |
396 starts with NITERS_SKIP dummy iterations of the scalar loop before | |
397 the real work starts. The mask elements for these dummy iterations | |
398 must be 0, to ensure that the extra iterations do not have an effect. | |
399 | |
400 It is known that: | |
401 | |
402 NITERS * RGM->max_nscalars_per_iter | |
403 | |
404 does not overflow. However, MIGHT_WRAP_P says whether an induction | |
405 variable that starts at 0 and has step: | |
406 | |
407 VF * RGM->max_nscalars_per_iter | |
408 | |
409 might overflow before hitting a value above: | |
410 | |
411 (NITERS + NITERS_SKIP) * RGM->max_nscalars_per_iter | |
412 | |
413 This means that we cannot guarantee that such an induction variable | |
414 would ever hit a value that produces a set of all-false masks for RGM. */ | |
415 | |
416 static tree | |
145 | 417 vect_set_loop_masks_directly (class loop *loop, loop_vec_info loop_vinfo, |
131 | 418 gimple_seq *preheader_seq, |
419 gimple_stmt_iterator loop_cond_gsi, | |
145 | 420 rgroup_masks *rgm, tree niters, tree niters_skip, |
131 | 421 bool might_wrap_p) |
422 { | |
423 tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
145 | 424 tree iv_type = LOOP_VINFO_MASK_IV_TYPE (loop_vinfo); |
131 | 425 tree mask_type = rgm->mask_type; |
426 unsigned int nscalars_per_iter = rgm->max_nscalars_per_iter; | |
427 poly_uint64 nscalars_per_mask = TYPE_VECTOR_SUBPARTS (mask_type); | |
145 | 428 poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); |
131 | 429 |
430 /* Calculate the maximum number of scalar values that the rgroup | |
431 handles in total, the number that it handles for each iteration | |
432 of the vector loop, and the number that it should skip during the | |
433 first iteration of the vector loop. */ | |
434 tree nscalars_total = niters; | |
145 | 435 tree nscalars_step = build_int_cst (iv_type, vf); |
131 | 436 tree nscalars_skip = niters_skip; |
437 if (nscalars_per_iter != 1) | |
438 { | |
439 /* We checked before choosing to use a fully-masked loop that these | |
440 multiplications don't overflow. */ | |
145 | 441 tree compare_factor = build_int_cst (compare_type, nscalars_per_iter); |
442 tree iv_factor = build_int_cst (iv_type, nscalars_per_iter); | |
131 | 443 nscalars_total = gimple_build (preheader_seq, MULT_EXPR, compare_type, |
145 | 444 nscalars_total, compare_factor); |
445 nscalars_step = gimple_build (preheader_seq, MULT_EXPR, iv_type, | |
446 nscalars_step, iv_factor); | |
131 | 447 if (nscalars_skip) |
448 nscalars_skip = gimple_build (preheader_seq, MULT_EXPR, compare_type, | |
145 | 449 nscalars_skip, compare_factor); |
131 | 450 } |
451 | |
452 /* Create an induction variable that counts the number of scalars | |
453 processed. */ | |
454 tree index_before_incr, index_after_incr; | |
455 gimple_stmt_iterator incr_gsi; | |
456 bool insert_after; | |
145 | 457 standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
458 create_iv (build_int_cst (iv_type, 0), nscalars_step, NULL_TREE, loop, | |
459 &incr_gsi, insert_after, &index_before_incr, &index_after_incr); | |
460 | |
131 | 461 tree zero_index = build_int_cst (compare_type, 0); |
462 tree test_index, test_limit, first_limit; | |
463 gimple_stmt_iterator *test_gsi; | |
464 if (might_wrap_p) | |
465 { | |
466 /* In principle the loop should stop iterating once the incremented | |
467 IV reaches a value greater than or equal to: | |
468 | |
469 NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP | |
470 | |
471 However, there's no guarantee that this addition doesn't overflow | |
472 the comparison type, or that the IV hits a value above it before | |
473 wrapping around. We therefore adjust the limit down by one | |
474 IV step: | |
475 | |
476 (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP) | |
477 -[infinite-prec] NSCALARS_STEP | |
478 | |
479 and compare the IV against this limit _before_ incrementing it. | |
480 Since the comparison type is unsigned, we actually want the | |
481 subtraction to saturate at zero: | |
482 | |
483 (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP) | |
484 -[sat] NSCALARS_STEP | |
485 | |
486 And since NSCALARS_SKIP < NSCALARS_STEP, we can reassociate this as: | |
487 | |
488 NSCALARS_TOTAL -[sat] (NSCALARS_STEP - NSCALARS_SKIP) | |
489 | |
490 where the rightmost subtraction can be done directly in | |
491 COMPARE_TYPE. */ | |
492 test_index = index_before_incr; | |
145 | 493 tree adjust = gimple_convert (preheader_seq, compare_type, |
494 nscalars_step); | |
131 | 495 if (nscalars_skip) |
496 adjust = gimple_build (preheader_seq, MINUS_EXPR, compare_type, | |
497 adjust, nscalars_skip); | |
498 test_limit = gimple_build (preheader_seq, MAX_EXPR, compare_type, | |
499 nscalars_total, adjust); | |
500 test_limit = gimple_build (preheader_seq, MINUS_EXPR, compare_type, | |
501 test_limit, adjust); | |
502 test_gsi = &incr_gsi; | |
503 | |
504 /* Get a safe limit for the first iteration. */ | |
505 if (nscalars_skip) | |
506 { | |
507 /* The first vector iteration can handle at most NSCALARS_STEP | |
508 scalars. NSCALARS_STEP <= CONST_LIMIT, and adding | |
509 NSCALARS_SKIP to that cannot overflow. */ | |
510 tree const_limit = build_int_cst (compare_type, | |
511 LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
512 * nscalars_per_iter); | |
513 first_limit = gimple_build (preheader_seq, MIN_EXPR, compare_type, | |
514 nscalars_total, const_limit); | |
515 first_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type, | |
516 first_limit, nscalars_skip); | |
517 } | |
518 else | |
519 /* For the first iteration it doesn't matter whether the IV hits | |
520 a value above NSCALARS_TOTAL. That only matters for the latch | |
521 condition. */ | |
522 first_limit = nscalars_total; | |
523 } | |
524 else | |
525 { | |
526 /* Test the incremented IV, which will always hit a value above | |
527 the bound before wrapping. */ | |
528 test_index = index_after_incr; | |
529 test_limit = nscalars_total; | |
530 if (nscalars_skip) | |
531 test_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type, | |
532 test_limit, nscalars_skip); | |
533 test_gsi = &loop_cond_gsi; | |
534 | |
535 first_limit = test_limit; | |
536 } | |
537 | |
145 | 538 /* Convert the IV value to the comparison type (either a no-op or |
539 a demotion). */ | |
540 gimple_seq test_seq = NULL; | |
541 test_index = gimple_convert (&test_seq, compare_type, test_index); | |
542 gsi_insert_seq_before (test_gsi, test_seq, GSI_SAME_STMT); | |
543 | |
131 | 544 /* Provide a definition of each mask in the group. */ |
545 tree next_mask = NULL_TREE; | |
546 tree mask; | |
547 unsigned int i; | |
548 FOR_EACH_VEC_ELT_REVERSE (rgm->masks, i, mask) | |
549 { | |
550 /* Previous masks will cover BIAS scalars. This mask covers the | |
551 next batch. */ | |
552 poly_uint64 bias = nscalars_per_mask * i; | |
553 tree bias_tree = build_int_cst (compare_type, bias); | |
554 gimple *tmp_stmt; | |
555 | |
556 /* See whether the first iteration of the vector loop is known | |
557 to have a full mask. */ | |
558 poly_uint64 const_limit; | |
559 bool first_iteration_full | |
560 = (poly_int_tree_p (first_limit, &const_limit) | |
561 && known_ge (const_limit, (i + 1) * nscalars_per_mask)); | |
562 | |
563 /* Rather than have a new IV that starts at BIAS and goes up to | |
564 TEST_LIMIT, prefer to use the same 0-based IV for each mask | |
565 and adjust the bound down by BIAS. */ | |
566 tree this_test_limit = test_limit; | |
567 if (i != 0) | |
568 { | |
569 this_test_limit = gimple_build (preheader_seq, MAX_EXPR, | |
570 compare_type, this_test_limit, | |
571 bias_tree); | |
572 this_test_limit = gimple_build (preheader_seq, MINUS_EXPR, | |
573 compare_type, this_test_limit, | |
574 bias_tree); | |
575 } | |
576 | |
577 /* Create the initial mask. First include all scalars that | |
578 are within the loop limit. */ | |
579 tree init_mask = NULL_TREE; | |
580 if (!first_iteration_full) | |
581 { | |
582 tree start, end; | |
583 if (first_limit == test_limit) | |
584 { | |
585 /* Use a natural test between zero (the initial IV value) | |
586 and the loop limit. The "else" block would be valid too, | |
587 but this choice can avoid the need to load BIAS_TREE into | |
588 a register. */ | |
589 start = zero_index; | |
590 end = this_test_limit; | |
591 } | |
592 else | |
593 { | |
594 /* FIRST_LIMIT is the maximum number of scalars handled by the | |
595 first iteration of the vector loop. Test the portion | |
596 associated with this mask. */ | |
597 start = bias_tree; | |
598 end = first_limit; | |
599 } | |
600 | |
601 init_mask = make_temp_ssa_name (mask_type, NULL, "max_mask"); | |
602 tmp_stmt = vect_gen_while (init_mask, start, end); | |
603 gimple_seq_add_stmt (preheader_seq, tmp_stmt); | |
604 } | |
605 | |
606 /* Now AND out the bits that are within the number of skipped | |
607 scalars. */ | |
608 poly_uint64 const_skip; | |
609 if (nscalars_skip | |
610 && !(poly_int_tree_p (nscalars_skip, &const_skip) | |
611 && known_le (const_skip, bias))) | |
612 { | |
613 tree unskipped_mask = vect_gen_while_not (preheader_seq, mask_type, | |
614 bias_tree, nscalars_skip); | |
615 if (init_mask) | |
616 init_mask = gimple_build (preheader_seq, BIT_AND_EXPR, mask_type, | |
617 init_mask, unskipped_mask); | |
618 else | |
619 init_mask = unskipped_mask; | |
620 } | |
621 | |
622 if (!init_mask) | |
623 /* First iteration is full. */ | |
624 init_mask = build_minus_one_cst (mask_type); | |
625 | |
626 /* Get the mask value for the next iteration of the loop. */ | |
627 next_mask = make_temp_ssa_name (mask_type, NULL, "next_mask"); | |
628 gcall *call = vect_gen_while (next_mask, test_index, this_test_limit); | |
629 gsi_insert_before (test_gsi, call, GSI_SAME_STMT); | |
630 | |
631 vect_set_loop_mask (loop, mask, init_mask, next_mask); | |
632 } | |
633 return next_mask; | |
634 } | |
635 | |
636 /* Make LOOP iterate NITERS times using masking and WHILE_ULT calls. | |
637 LOOP_VINFO describes the vectorization of LOOP. NITERS is the | |
638 number of iterations of the original scalar loop that should be | |
639 handled by the vector loop. NITERS_MAYBE_ZERO and FINAL_IV are | |
640 as for vect_set_loop_condition. | |
641 | |
642 Insert the branch-back condition before LOOP_COND_GSI and return the | |
643 final gcond. */ | |
644 | |
645 static gcond * | |
145 | 646 vect_set_loop_condition_masked (class loop *loop, loop_vec_info loop_vinfo, |
131 | 647 tree niters, tree final_iv, |
648 bool niters_maybe_zero, | |
649 gimple_stmt_iterator loop_cond_gsi) | |
650 { | |
651 gimple_seq preheader_seq = NULL; | |
652 gimple_seq header_seq = NULL; | |
653 | |
654 tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
655 unsigned int compare_precision = TYPE_PRECISION (compare_type); | |
656 tree orig_niters = niters; | |
657 | |
658 /* Type of the initial value of NITERS. */ | |
659 tree ni_actual_type = TREE_TYPE (niters); | |
660 unsigned int ni_actual_precision = TYPE_PRECISION (ni_actual_type); | |
145 | 661 tree niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo); |
131 | 662 |
663 /* Convert NITERS to the same size as the compare. */ | |
664 if (compare_precision > ni_actual_precision | |
665 && niters_maybe_zero) | |
666 { | |
667 /* We know that there is always at least one iteration, so if the | |
668 count is zero then it must have wrapped. Cope with this by | |
669 subtracting 1 before the conversion and adding 1 to the result. */ | |
670 gcc_assert (TYPE_UNSIGNED (ni_actual_type)); | |
671 niters = gimple_build (&preheader_seq, PLUS_EXPR, ni_actual_type, | |
672 niters, build_minus_one_cst (ni_actual_type)); | |
673 niters = gimple_convert (&preheader_seq, compare_type, niters); | |
674 niters = gimple_build (&preheader_seq, PLUS_EXPR, compare_type, | |
675 niters, build_one_cst (compare_type)); | |
676 } | |
677 else | |
678 niters = gimple_convert (&preheader_seq, compare_type, niters); | |
679 | |
145 | 680 widest_int iv_limit = vect_iv_limit_for_full_masking (loop_vinfo); |
131 | 681 |
682 /* Iterate over all the rgroups and fill in their masks. We could use | |
683 the first mask from any rgroup for the loop condition; here we | |
684 arbitrarily pick the last. */ | |
685 tree test_mask = NULL_TREE; | |
686 rgroup_masks *rgm; | |
687 unsigned int i; | |
688 vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo); | |
689 FOR_EACH_VEC_ELT (*masks, i, rgm) | |
690 if (!rgm->masks.is_empty ()) | |
691 { | |
692 /* First try using permutes. This adds a single vector | |
693 instruction to the loop for each mask, but needs no extra | |
694 loop invariants or IVs. */ | |
695 unsigned int nmasks = i + 1; | |
696 if ((nmasks & 1) == 0) | |
697 { | |
698 rgroup_masks *half_rgm = &(*masks)[nmasks / 2 - 1]; | |
699 if (!half_rgm->masks.is_empty () | |
700 && vect_maybe_permute_loop_masks (&header_seq, rgm, half_rgm)) | |
701 continue; | |
702 } | |
703 | |
704 /* See whether zero-based IV would ever generate all-false masks | |
705 before wrapping around. */ | |
706 bool might_wrap_p | |
145 | 707 = (iv_limit == -1 |
131 | 708 || (wi::min_precision (iv_limit * rgm->max_nscalars_per_iter, |
709 UNSIGNED) | |
710 > compare_precision)); | |
711 | |
712 /* Set up all masks for this group. */ | |
713 test_mask = vect_set_loop_masks_directly (loop, loop_vinfo, | |
714 &preheader_seq, | |
145 | 715 loop_cond_gsi, rgm, |
131 | 716 niters, niters_skip, |
717 might_wrap_p); | |
718 } | |
719 | |
720 /* Emit all accumulated statements. */ | |
721 add_preheader_seq (loop, preheader_seq); | |
722 add_header_seq (loop, header_seq); | |
723 | |
724 /* Get a boolean result that tells us whether to iterate. */ | |
725 edge exit_edge = single_exit (loop); | |
726 tree_code code = (exit_edge->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR; | |
727 tree zero_mask = build_zero_cst (TREE_TYPE (test_mask)); | |
728 gcond *cond_stmt = gimple_build_cond (code, test_mask, zero_mask, | |
729 NULL_TREE, NULL_TREE); | |
730 gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); | |
731 | |
732 /* The loop iterates (NITERS - 1) / VF + 1 times. | |
733 Subtract one from this to get the latch count. */ | |
734 tree step = build_int_cst (compare_type, | |
735 LOOP_VINFO_VECT_FACTOR (loop_vinfo)); | |
736 tree niters_minus_one = fold_build2 (PLUS_EXPR, compare_type, niters, | |
737 build_minus_one_cst (compare_type)); | |
738 loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, compare_type, | |
739 niters_minus_one, step); | |
740 | |
741 if (final_iv) | |
742 { | |
743 gassign *assign = gimple_build_assign (final_iv, orig_niters); | |
744 gsi_insert_on_edge_immediate (single_exit (loop), assign); | |
745 } | |
746 | |
747 return cond_stmt; | |
748 } | |
749 | |
750 /* Like vect_set_loop_condition, but handle the case in which there | |
751 are no loop masks. */ | |
752 | |
753 static gcond * | |
145 | 754 vect_set_loop_condition_unmasked (class loop *loop, tree niters, |
131 | 755 tree step, tree final_iv, |
756 bool niters_maybe_zero, | |
757 gimple_stmt_iterator loop_cond_gsi) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
758 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
759 tree indx_before_incr, indx_after_incr; |
111 | 760 gcond *cond_stmt; |
761 gcond *orig_cond; | |
131 | 762 edge pe = loop_preheader_edge (loop); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
763 edge exit_edge = single_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
764 gimple_stmt_iterator incr_gsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
765 bool insert_after; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
766 enum tree_code code; |
131 | 767 tree niters_type = TREE_TYPE (niters); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
768 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
769 orig_cond = get_loop_exit_condition (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
770 gcc_assert (orig_cond); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
771 loop_cond_gsi = gsi_for_stmt (orig_cond); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
772 |
131 | 773 tree init, limit; |
774 if (!niters_maybe_zero && integer_onep (step)) | |
775 { | |
776 /* In this case we can use a simple 0-based IV: | |
777 | |
778 A: | |
779 x = 0; | |
780 do | |
781 { | |
782 ... | |
783 x += 1; | |
784 } | |
785 while (x < NITERS); */ | |
786 code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
787 init = build_zero_cst (niters_type); | |
788 limit = niters; | |
789 } | |
790 else | |
791 { | |
792 /* The following works for all values of NITERS except 0: | |
793 | |
794 B: | |
795 x = 0; | |
796 do | |
797 { | |
798 ... | |
799 x += STEP; | |
800 } | |
801 while (x <= NITERS - STEP); | |
802 | |
803 so that the loop continues to iterate if x + STEP - 1 < NITERS | |
804 but stops if x + STEP - 1 >= NITERS. | |
805 | |
806 However, if NITERS is zero, x never hits a value above NITERS - STEP | |
807 before wrapping around. There are two obvious ways of dealing with | |
808 this: | |
809 | |
810 - start at STEP - 1 and compare x before incrementing it | |
811 - start at -1 and compare x after incrementing it | |
812 | |
813 The latter is simpler and is what we use. The loop in this case | |
814 looks like: | |
815 | |
816 C: | |
817 x = -1; | |
818 do | |
819 { | |
820 ... | |
821 x += STEP; | |
822 } | |
823 while (x < NITERS - STEP); | |
824 | |
825 In both cases the loop limit is NITERS - STEP. */ | |
826 gimple_seq seq = NULL; | |
827 limit = force_gimple_operand (niters, &seq, true, NULL_TREE); | |
828 limit = gimple_build (&seq, MINUS_EXPR, TREE_TYPE (limit), limit, step); | |
829 if (seq) | |
830 { | |
831 basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq); | |
832 gcc_assert (!new_bb); | |
833 } | |
834 if (niters_maybe_zero) | |
835 { | |
836 /* Case C. */ | |
837 code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR; | |
838 init = build_all_ones_cst (niters_type); | |
839 } | |
840 else | |
841 { | |
842 /* Case B. */ | |
843 code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GT_EXPR : LE_EXPR; | |
844 init = build_zero_cst (niters_type); | |
845 } | |
846 } | |
847 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
848 standard_iv_increment_position (loop, &incr_gsi, &insert_after); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
849 create_iv (init, step, NULL_TREE, loop, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
850 &incr_gsi, insert_after, &indx_before_incr, &indx_after_incr); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
851 indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
852 true, NULL_TREE, true, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
853 GSI_SAME_STMT); |
131 | 854 limit = force_gimple_operand_gsi (&loop_cond_gsi, limit, true, NULL_TREE, |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
855 true, GSI_SAME_STMT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
856 |
131 | 857 cond_stmt = gimple_build_cond (code, indx_after_incr, limit, NULL_TREE, |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
858 NULL_TREE); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
859 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
860 gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
861 |
131 | 862 /* Record the number of latch iterations. */ |
863 if (limit == niters) | |
864 /* Case A: the loop iterates NITERS times. Subtract one to get the | |
865 latch count. */ | |
866 loop->nb_iterations = fold_build2 (MINUS_EXPR, niters_type, niters, | |
867 build_int_cst (niters_type, 1)); | |
868 else | |
869 /* Case B or C: the loop iterates (NITERS - STEP) / STEP + 1 times. | |
870 Subtract one from this to get the latch count. */ | |
871 loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, niters_type, | |
872 limit, step); | |
873 | |
874 if (final_iv) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
875 { |
131 | 876 gassign *assign = gimple_build_assign (final_iv, MINUS_EXPR, |
877 indx_after_incr, init); | |
878 gsi_insert_on_edge_immediate (single_exit (loop), assign); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
879 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
880 |
131 | 881 return cond_stmt; |
882 } | |
883 | |
884 /* If we're using fully-masked loops, make LOOP iterate: | |
885 | |
886 N == (NITERS - 1) / STEP + 1 | |
887 | |
888 times. When NITERS is zero, this is equivalent to making the loop | |
889 execute (1 << M) / STEP times, where M is the precision of NITERS. | |
890 NITERS_MAYBE_ZERO is true if this last case might occur. | |
891 | |
892 If we're not using fully-masked loops, make LOOP iterate: | |
893 | |
894 N == (NITERS - STEP) / STEP + 1 | |
895 | |
896 times, where NITERS is known to be outside the range [1, STEP - 1]. | |
897 This is equivalent to making the loop execute NITERS / STEP times | |
898 when NITERS is nonzero and (1 << M) / STEP times otherwise. | |
899 NITERS_MAYBE_ZERO again indicates whether this last case might occur. | |
900 | |
901 If FINAL_IV is nonnull, it is an SSA name that should be set to | |
902 N * STEP on exit from the loop. | |
903 | |
904 Assumption: the exit-condition of LOOP is the last stmt in the loop. */ | |
905 | |
906 void | |
145 | 907 vect_set_loop_condition (class loop *loop, loop_vec_info loop_vinfo, |
131 | 908 tree niters, tree step, tree final_iv, |
909 bool niters_maybe_zero) | |
910 { | |
911 gcond *cond_stmt; | |
912 gcond *orig_cond = get_loop_exit_condition (loop); | |
913 gimple_stmt_iterator loop_cond_gsi = gsi_for_stmt (orig_cond); | |
914 | |
915 if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) | |
916 cond_stmt = vect_set_loop_condition_masked (loop, loop_vinfo, niters, | |
917 final_iv, niters_maybe_zero, | |
918 loop_cond_gsi); | |
919 else | |
920 cond_stmt = vect_set_loop_condition_unmasked (loop, niters, step, | |
921 final_iv, niters_maybe_zero, | |
922 loop_cond_gsi); | |
923 | |
924 /* Remove old loop exit test. */ | |
925 stmt_vec_info orig_cond_info; | |
926 if (loop_vinfo | |
927 && (orig_cond_info = loop_vinfo->lookup_stmt (orig_cond))) | |
928 loop_vinfo->remove_stmt (orig_cond_info); | |
929 else | |
930 gsi_remove (&loop_cond_gsi, true); | |
931 | |
932 if (dump_enabled_p ()) | |
933 dump_printf_loc (MSG_NOTE, vect_location, "New loop exit condition: %G", | |
934 cond_stmt); | |
111 | 935 } |
936 | |
937 /* Helper routine of slpeel_tree_duplicate_loop_to_edge_cfg. | |
938 For all PHI arguments in FROM->dest and TO->dest from those | |
939 edges ensure that TO->dest PHI arguments have current_def | |
940 to that in from. */ | |
941 | |
942 static void | |
943 slpeel_duplicate_current_defs_from_edges (edge from, edge to) | |
944 { | |
945 gimple_stmt_iterator gsi_from, gsi_to; | |
946 | |
947 for (gsi_from = gsi_start_phis (from->dest), | |
948 gsi_to = gsi_start_phis (to->dest); | |
949 !gsi_end_p (gsi_from) && !gsi_end_p (gsi_to);) | |
950 { | |
951 gimple *from_phi = gsi_stmt (gsi_from); | |
952 gimple *to_phi = gsi_stmt (gsi_to); | |
953 tree from_arg = PHI_ARG_DEF_FROM_EDGE (from_phi, from); | |
954 tree to_arg = PHI_ARG_DEF_FROM_EDGE (to_phi, to); | |
955 if (virtual_operand_p (from_arg)) | |
956 { | |
957 gsi_next (&gsi_from); | |
958 continue; | |
959 } | |
960 if (virtual_operand_p (to_arg)) | |
961 { | |
962 gsi_next (&gsi_to); | |
963 continue; | |
964 } | |
965 if (TREE_CODE (from_arg) != SSA_NAME) | |
966 gcc_assert (operand_equal_p (from_arg, to_arg, 0)); | |
145 | 967 else if (TREE_CODE (to_arg) == SSA_NAME |
968 && from_arg != to_arg) | |
111 | 969 { |
970 if (get_current_def (to_arg) == NULL_TREE) | |
145 | 971 { |
972 gcc_assert (types_compatible_p (TREE_TYPE (to_arg), | |
973 TREE_TYPE (get_current_def | |
974 (from_arg)))); | |
975 set_current_def (to_arg, get_current_def (from_arg)); | |
976 } | |
111 | 977 } |
978 gsi_next (&gsi_from); | |
979 gsi_next (&gsi_to); | |
980 } | |
981 | |
982 gphi *from_phi = get_virtual_phi (from->dest); | |
983 gphi *to_phi = get_virtual_phi (to->dest); | |
984 if (from_phi) | |
985 set_current_def (PHI_ARG_DEF_FROM_EDGE (to_phi, to), | |
986 get_current_def (PHI_ARG_DEF_FROM_EDGE (from_phi, from))); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
987 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
988 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
989 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
990 /* Given LOOP this function generates a new copy of it and puts it |
111 | 991 on E which is either the entry or exit of LOOP. If SCALAR_LOOP is |
992 non-NULL, assume LOOP and SCALAR_LOOP are equivalent and copy the | |
993 basic blocks from SCALAR_LOOP instead of LOOP, but to either the | |
994 entry or exit of LOOP. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
995 |
145 | 996 class loop * |
997 slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop, | |
998 class loop *scalar_loop, edge e) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
999 { |
145 | 1000 class loop *new_loop; |
111 | 1001 basic_block *new_bbs, *bbs, *pbbs; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1002 bool at_exit; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1003 bool was_imm_dom; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1004 basic_block exit_dest; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1005 edge exit, new_exit; |
111 | 1006 bool duplicate_outer_loop = false; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1007 |
111 | 1008 exit = single_exit (loop); |
1009 at_exit = (e == exit); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1010 if (!at_exit && e != loop_preheader_edge (loop)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1011 return NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1012 |
111 | 1013 if (scalar_loop == NULL) |
1014 scalar_loop = loop; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1015 |
111 | 1016 bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); |
1017 pbbs = bbs + 1; | |
1018 get_loop_body_with_size (scalar_loop, pbbs, scalar_loop->num_nodes); | |
1019 /* Allow duplication of outer loops. */ | |
1020 if (scalar_loop->inner) | |
1021 duplicate_outer_loop = true; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1022 /* Check whether duplication is possible. */ |
111 | 1023 if (!can_copy_bbs_p (pbbs, scalar_loop->num_nodes)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1024 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1025 free (bbs); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1026 return NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1027 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1028 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1029 /* Generate new loop structure. */ |
111 | 1030 new_loop = duplicate_loop (scalar_loop, loop_outer (scalar_loop)); |
1031 duplicate_subloops (scalar_loop, new_loop); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1032 |
111 | 1033 exit_dest = exit->dest; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1034 was_imm_dom = (get_immediate_dominator (CDI_DOMINATORS, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1035 exit_dest) == loop->header ? |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1036 true : false); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1037 |
111 | 1038 /* Also copy the pre-header, this avoids jumping through hoops to |
1039 duplicate the loop entry PHI arguments. Create an empty | |
1040 pre-header unconditionally for this. */ | |
1041 basic_block preheader = split_edge (loop_preheader_edge (scalar_loop)); | |
1042 edge entry_e = single_pred_edge (preheader); | |
1043 bbs[0] = preheader; | |
1044 new_bbs = XNEWVEC (basic_block, scalar_loop->num_nodes + 1); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1045 |
111 | 1046 exit = single_exit (scalar_loop); |
1047 copy_bbs (bbs, scalar_loop->num_nodes + 1, new_bbs, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1048 &exit, 1, &new_exit, NULL, |
111 | 1049 at_exit ? loop->latch : e->src, true); |
1050 exit = single_exit (loop); | |
1051 basic_block new_preheader = new_bbs[0]; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1052 |
111 | 1053 add_phi_args_after_copy (new_bbs, scalar_loop->num_nodes + 1, NULL); |
1054 | |
1055 if (scalar_loop != loop) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1056 { |
111 | 1057 /* If we copied from SCALAR_LOOP rather than LOOP, SSA_NAMEs from |
1058 SCALAR_LOOP will have current_def set to SSA_NAMEs in the new_loop, | |
1059 but LOOP will not. slpeel_update_phi_nodes_for_guard{1,2} expects | |
1060 the LOOP SSA_NAMEs (on the exit edge and edge from latch to | |
1061 header) to have current_def set, so copy them over. */ | |
1062 slpeel_duplicate_current_defs_from_edges (single_exit (scalar_loop), | |
1063 exit); | |
1064 slpeel_duplicate_current_defs_from_edges (EDGE_SUCC (scalar_loop->latch, | |
1065 0), | |
1066 EDGE_SUCC (loop->latch, 0)); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1067 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1068 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1069 if (at_exit) /* Add the loop copy at exit. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1070 { |
111 | 1071 if (scalar_loop != loop) |
1072 { | |
1073 gphi_iterator gsi; | |
1074 new_exit = redirect_edge_and_branch (new_exit, exit_dest); | |
1075 | |
1076 for (gsi = gsi_start_phis (exit_dest); !gsi_end_p (gsi); | |
1077 gsi_next (&gsi)) | |
1078 { | |
1079 gphi *phi = gsi.phi (); | |
1080 tree orig_arg = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
1081 location_t orig_locus | |
1082 = gimple_phi_arg_location_from_edge (phi, e); | |
1083 | |
1084 add_phi_arg (phi, orig_arg, new_exit, orig_locus); | |
1085 } | |
1086 } | |
1087 redirect_edge_and_branch_force (e, new_preheader); | |
1088 flush_pending_stmts (e); | |
1089 set_immediate_dominator (CDI_DOMINATORS, new_preheader, e->src); | |
1090 if (was_imm_dom || duplicate_outer_loop) | |
1091 set_immediate_dominator (CDI_DOMINATORS, exit_dest, new_exit->src); | |
1092 | |
1093 /* And remove the non-necessary forwarder again. Keep the other | |
1094 one so we have a proper pre-header for the loop at the exit edge. */ | |
1095 redirect_edge_pred (single_succ_edge (preheader), | |
1096 single_pred (preheader)); | |
1097 delete_basic_block (preheader); | |
1098 set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, | |
1099 loop_preheader_edge (scalar_loop)->src); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1100 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1101 else /* Add the copy at entry. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1102 { |
111 | 1103 if (scalar_loop != loop) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1104 { |
111 | 1105 /* Remove the non-necessary forwarder of scalar_loop again. */ |
1106 redirect_edge_pred (single_succ_edge (preheader), | |
1107 single_pred (preheader)); | |
1108 delete_basic_block (preheader); | |
1109 set_immediate_dominator (CDI_DOMINATORS, scalar_loop->header, | |
1110 loop_preheader_edge (scalar_loop)->src); | |
1111 preheader = split_edge (loop_preheader_edge (loop)); | |
1112 entry_e = single_pred_edge (preheader); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1113 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1114 |
111 | 1115 redirect_edge_and_branch_force (entry_e, new_preheader); |
1116 flush_pending_stmts (entry_e); | |
1117 set_immediate_dominator (CDI_DOMINATORS, new_preheader, entry_e->src); | |
1118 | |
1119 redirect_edge_and_branch_force (new_exit, preheader); | |
1120 flush_pending_stmts (new_exit); | |
1121 set_immediate_dominator (CDI_DOMINATORS, preheader, new_exit->src); | |
1122 | |
1123 /* And remove the non-necessary forwarder again. Keep the other | |
1124 one so we have a proper pre-header for the loop at the exit edge. */ | |
1125 redirect_edge_pred (single_succ_edge (new_preheader), | |
1126 single_pred (new_preheader)); | |
1127 delete_basic_block (new_preheader); | |
1128 set_immediate_dominator (CDI_DOMINATORS, new_loop->header, | |
1129 loop_preheader_edge (new_loop)->src); | |
1130 } | |
1131 | |
1132 /* Skip new preheader since it's deleted if copy loop is added at entry. */ | |
1133 for (unsigned i = (at_exit ? 0 : 1); i < scalar_loop->num_nodes + 1; i++) | |
1134 rename_variables_in_bb (new_bbs[i], duplicate_outer_loop); | |
1135 | |
1136 if (scalar_loop != loop) | |
1137 { | |
1138 /* Update new_loop->header PHIs, so that on the preheader | |
1139 edge they are the ones from loop rather than scalar_loop. */ | |
1140 gphi_iterator gsi_orig, gsi_new; | |
1141 edge orig_e = loop_preheader_edge (loop); | |
1142 edge new_e = loop_preheader_edge (new_loop); | |
1143 | |
1144 for (gsi_orig = gsi_start_phis (loop->header), | |
1145 gsi_new = gsi_start_phis (new_loop->header); | |
1146 !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_new); | |
1147 gsi_next (&gsi_orig), gsi_next (&gsi_new)) | |
1148 { | |
1149 gphi *orig_phi = gsi_orig.phi (); | |
1150 gphi *new_phi = gsi_new.phi (); | |
1151 tree orig_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, orig_e); | |
1152 location_t orig_locus | |
1153 = gimple_phi_arg_location_from_edge (orig_phi, orig_e); | |
1154 | |
1155 add_phi_arg (new_phi, orig_arg, new_e, orig_locus); | |
1156 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1157 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1158 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1159 free (new_bbs); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1160 free (bbs); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1161 |
111 | 1162 checking_verify_dominators (CDI_DOMINATORS); |
1163 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1164 return new_loop; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1165 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1166 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1167 |
111 | 1168 /* Given the condition expression COND, put it as the last statement of |
1169 GUARD_BB; set both edges' probability; set dominator of GUARD_TO to | |
1170 DOM_BB; return the skip edge. GUARD_TO is the target basic block to | |
1171 skip the loop. PROBABILITY is the skip edge's probability. Mark the | |
1172 new edge as irreducible if IRREDUCIBLE_P is true. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1173 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1174 static edge |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1175 slpeel_add_loop_guard (basic_block guard_bb, tree cond, |
111 | 1176 basic_block guard_to, basic_block dom_bb, |
1177 profile_probability probability, bool irreducible_p) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1178 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1179 gimple_stmt_iterator gsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1180 edge new_e, enter_e; |
111 | 1181 gcond *cond_stmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1182 gimple_seq gimplify_stmt_list = NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1183 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1184 enter_e = EDGE_SUCC (guard_bb, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1185 enter_e->flags &= ~EDGE_FALLTHRU; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1186 enter_e->flags |= EDGE_FALSE_VALUE; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1187 gsi = gsi_last_bb (guard_bb); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1188 |
111 | 1189 cond = force_gimple_operand_1 (cond, &gimplify_stmt_list, is_gimple_condexpr, |
1190 NULL_TREE); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1191 if (gimplify_stmt_list) |
111 | 1192 gsi_insert_seq_after (&gsi, gimplify_stmt_list, GSI_NEW_STMT); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1193 |
111 | 1194 cond_stmt = gimple_build_cond_from_tree (cond, NULL_TREE, NULL_TREE); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1195 gsi = gsi_last_bb (guard_bb); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1196 gsi_insert_after (&gsi, cond_stmt, GSI_NEW_STMT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1197 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1198 /* Add new edge to connect guard block to the merge/loop-exit block. */ |
111 | 1199 new_e = make_edge (guard_bb, guard_to, EDGE_TRUE_VALUE); |
1200 | |
1201 new_e->probability = probability; | |
1202 if (irreducible_p) | |
1203 new_e->flags |= EDGE_IRREDUCIBLE_LOOP; | |
1204 | |
1205 enter_e->probability = probability.invert (); | |
1206 set_immediate_dominator (CDI_DOMINATORS, guard_to, dom_bb); | |
1207 | |
1208 /* Split enter_e to preserve LOOPS_HAVE_PREHEADERS. */ | |
1209 if (enter_e->dest->loop_father->header == enter_e->dest) | |
1210 split_edge (enter_e); | |
1211 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1212 return new_e; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1213 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1214 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1215 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1216 /* This function verifies that the following restrictions apply to LOOP: |
111 | 1217 (1) it consists of exactly 2 basic blocks - header, and an empty latch |
1218 for innermost loop and 5 basic blocks for outer-loop. | |
1219 (2) it is single entry, single exit | |
1220 (3) its exit condition is the last stmt in the header | |
1221 (4) E is the entry/exit edge of LOOP. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1222 */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1223 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1224 bool |
145 | 1225 slpeel_can_duplicate_loop_p (const class loop *loop, const_edge e) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1226 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1227 edge exit_e = single_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1228 edge entry_e = loop_preheader_edge (loop); |
111 | 1229 gcond *orig_cond = get_loop_exit_condition (loop); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1230 gimple_stmt_iterator loop_exit_gsi = gsi_last_bb (exit_e->src); |
111 | 1231 unsigned int num_bb = loop->inner? 5 : 2; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1232 |
111 | 1233 /* All loops have an outer scope; the only case loop->outer is NULL is for |
1234 the function itself. */ | |
1235 if (!loop_outer (loop) | |
1236 || loop->num_nodes != num_bb | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1237 || !empty_block_p (loop->latch) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1238 || !single_exit (loop) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1239 /* Verify that new loop exit condition can be trivially modified. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1240 || (!orig_cond || orig_cond != gsi_stmt (loop_exit_gsi)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1241 || (e != exit_e && e != entry_e)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1242 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1243 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1244 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1245 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1246 |
111 | 1247 /* If the loop has a virtual PHI, but exit bb doesn't, create a virtual PHI |
1248 in the exit bb and rename all the uses after the loop. This simplifies | |
1249 the *guard[12] routines, which assume loop closed SSA form for all PHIs | |
1250 (but normally loop closed SSA form doesn't require virtual PHIs to be | |
1251 in the same form). Doing this early simplifies the checking what | |
145 | 1252 uses should be renamed. |
1253 | |
1254 If we create a new phi after the loop, return the definition that | |
1255 applies on entry to the loop, otherwise return null. */ | |
1256 | |
1257 static tree | |
1258 create_lcssa_for_virtual_phi (class loop *loop) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1259 { |
111 | 1260 gphi_iterator gsi; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1261 edge exit_e = single_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1262 |
111 | 1263 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
1264 if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) | |
1265 { | |
1266 gphi *phi = gsi.phi (); | |
1267 for (gsi = gsi_start_phis (exit_e->dest); | |
1268 !gsi_end_p (gsi); gsi_next (&gsi)) | |
1269 if (virtual_operand_p (gimple_phi_result (gsi_stmt (gsi)))) | |
1270 break; | |
1271 if (gsi_end_p (gsi)) | |
1272 { | |
1273 tree new_vop = copy_ssa_name (PHI_RESULT (phi)); | |
1274 gphi *new_phi = create_phi_node (new_vop, exit_e->dest); | |
1275 tree vop = PHI_ARG_DEF_FROM_EDGE (phi, EDGE_SUCC (loop->latch, 0)); | |
1276 imm_use_iterator imm_iter; | |
1277 gimple *stmt; | |
1278 use_operand_p use_p; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1279 |
131 | 1280 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new_vop) |
1281 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (vop); | |
111 | 1282 add_phi_arg (new_phi, vop, exit_e, UNKNOWN_LOCATION); |
1283 gimple_phi_set_result (new_phi, new_vop); | |
1284 FOR_EACH_IMM_USE_STMT (stmt, imm_iter, vop) | |
1285 if (stmt != new_phi | |
1286 && !flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
1287 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1288 SET_USE (use_p, new_vop); | |
145 | 1289 |
1290 return PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
111 | 1291 } |
1292 break; | |
1293 } | |
145 | 1294 return NULL_TREE; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1295 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1296 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1297 /* Function vect_get_loop_location. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1298 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1299 Extract the location of the loop in the source code. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1300 If the loop is not well formed for vectorization, an estimated |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1301 location is calculated. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1302 Return the loop location if succeed and NULL if not. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1303 |
131 | 1304 dump_user_location_t |
145 | 1305 find_loop_location (class loop *loop) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1306 { |
111 | 1307 gimple *stmt = NULL; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1308 basic_block bb; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1309 gimple_stmt_iterator si; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1310 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1311 if (!loop) |
131 | 1312 return dump_user_location_t (); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1313 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1314 stmt = get_loop_exit_condition (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1315 |
111 | 1316 if (stmt |
1317 && LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) | |
131 | 1318 return stmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1319 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1320 /* If we got here the loop is probably not "well formed", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1321 try to estimate the loop location */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1322 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1323 if (!loop->header) |
131 | 1324 return dump_user_location_t (); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1325 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1326 bb = loop->header; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1327 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1328 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1329 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1330 stmt = gsi_stmt (si); |
111 | 1331 if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION) |
131 | 1332 return stmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1333 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1334 |
131 | 1335 return dump_user_location_t (); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1336 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1337 |
131 | 1338 /* Return true if the phi described by STMT_INFO defines an IV of the |
1339 loop to be vectorized. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1340 |
111 | 1341 static bool |
131 | 1342 iv_phi_p (stmt_vec_info stmt_info) |
111 | 1343 { |
131 | 1344 gphi *phi = as_a <gphi *> (stmt_info->stmt); |
111 | 1345 if (virtual_operand_p (PHI_RESULT (phi))) |
1346 return false; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1347 |
111 | 1348 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def |
1349 || STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def) | |
1350 return false; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1351 |
111 | 1352 return true; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1353 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1354 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1355 /* Function vect_can_advance_ivs_p |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1356 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1357 In case the number of iterations that LOOP iterates is unknown at compile |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1358 time, an epilog loop will be generated, and the loop induction variables |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1359 (IVs) will be "advanced" to the value they are supposed to take just before |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1360 the epilog loop. Here we check that the access function of the loop IVs |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1361 and the expression that represents the loop bound are simple enough. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1362 These restrictions will be relaxed in the future. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1363 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1364 bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1365 vect_can_advance_ivs_p (loop_vec_info loop_vinfo) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1366 { |
145 | 1367 class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1368 basic_block bb = loop->header; |
111 | 1369 gphi_iterator gsi; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1370 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1371 /* Analyze phi functions of the loop header. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1372 |
111 | 1373 if (dump_enabled_p ()) |
1374 dump_printf_loc (MSG_NOTE, vect_location, "vect_can_advance_ivs_p:\n"); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1375 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1376 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1377 tree evolution_part; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1378 |
111 | 1379 gphi *phi = gsi.phi (); |
131 | 1380 stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi); |
111 | 1381 if (dump_enabled_p ()) |
131 | 1382 dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G", |
1383 phi_info->stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1384 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1385 /* Skip virtual phi's. The data dependences that are associated with |
111 | 1386 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1387 |
111 | 1388 Skip reduction phis. */ |
131 | 1389 if (!iv_phi_p (phi_info)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1390 { |
111 | 1391 if (dump_enabled_p ()) |
1392 dump_printf_loc (MSG_NOTE, vect_location, | |
1393 "reduc or virtual phi. skip.\n"); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1394 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1395 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1396 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1397 /* Analyze the evolution function. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1398 |
131 | 1399 evolution_part = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info); |
111 | 1400 if (evolution_part == NULL_TREE) |
1401 { | |
1402 if (dump_enabled_p ()) | |
1403 dump_printf (MSG_MISSED_OPTIMIZATION, | |
1404 "No access function or evolution.\n"); | |
1405 return false; | |
1406 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1407 |
111 | 1408 /* FORNOW: We do not transform initial conditions of IVs |
1409 which evolution functions are not invariants in the loop. */ | |
1410 | |
1411 if (!expr_invariant_in_loop_p (loop, evolution_part)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1412 { |
111 | 1413 if (dump_enabled_p ()) |
1414 dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1415 "evolution not invariant in loop.\n"); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1416 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1417 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1418 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1419 /* FORNOW: We do not transform initial conditions of IVs |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1420 which evolution functions are a polynomial of degree >= 2. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1421 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1422 if (tree_is_chrec (evolution_part)) |
111 | 1423 { |
1424 if (dump_enabled_p ()) | |
1425 dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
1426 "evolution is chrec.\n"); | |
1427 return false; | |
1428 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1429 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1430 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1431 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1432 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1433 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1434 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1435 /* Function vect_update_ivs_after_vectorizer. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1436 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1437 "Advance" the induction variables of LOOP to the value they should take |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1438 after the execution of LOOP. This is currently necessary because the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1439 vectorizer does not handle induction variables that are used after the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1440 loop. Such a situation occurs when the last iterations of LOOP are |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1441 peeled, because: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1442 1. We introduced new uses after LOOP for IVs that were not originally used |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1443 after LOOP: the IVs of LOOP are now used by an epilog loop. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1444 2. LOOP is going to be vectorized; this means that it will iterate N/VF |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1445 times, whereas the loop IVs should be bumped N times. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1446 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1447 Input: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1448 - LOOP - a loop that is going to be vectorized. The last few iterations |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1449 of LOOP were peeled. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1450 - NITERS - the number of iterations that LOOP executes (before it is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1451 vectorized). i.e, the number of times the ivs should be bumped. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1452 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1453 coming out from LOOP on which there are uses of the LOOP ivs |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1454 (this is the path from LOOP->exit to epilog_loop->preheader). |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1455 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1456 The new definitions of the ivs are placed in LOOP->exit. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1457 The phi args associated with the edge UPDATE_E in the bb |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1458 UPDATE_E->dest are updated accordingly. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1459 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1460 Assumption 1: Like the rest of the vectorizer, this function assumes |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1461 a single loop exit that has a single predecessor. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1462 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1463 Assumption 2: The phi nodes in the LOOP header and in update_bb are |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1464 organized in the same order. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1465 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1466 Assumption 3: The access function of the ivs is simple enough (see |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1467 vect_can_advance_ivs_p). This assumption will be relaxed in the future. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1468 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1469 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1470 coming out of LOOP on which the ivs of LOOP are used (this is the path |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1471 that leads to the epilog loop; other paths skip the epilog loop). This |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1472 path starts with the edge UPDATE_E, and its destination (denoted update_bb) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1473 needs to have its phis updated. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1474 */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1475 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1476 static void |
111 | 1477 vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, |
1478 tree niters, edge update_e) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1479 { |
111 | 1480 gphi_iterator gsi, gsi1; |
145 | 1481 class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
111 | 1482 basic_block update_bb = update_e->dest; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1483 basic_block exit_bb = single_exit (loop)->dest; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1484 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1485 /* Make sure there exists a single-predecessor exit bb: */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1486 gcc_assert (single_pred_p (exit_bb)); |
111 | 1487 gcc_assert (single_succ_edge (exit_bb) == update_e); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1488 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1489 for (gsi = gsi_start_phis (loop->header), gsi1 = gsi_start_phis (update_bb); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1490 !gsi_end_p (gsi) && !gsi_end_p (gsi1); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1491 gsi_next (&gsi), gsi_next (&gsi1)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1492 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1493 tree init_expr; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1494 tree step_expr, off; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1495 tree type; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1496 tree var, ni, ni_name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1497 gimple_stmt_iterator last_gsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1498 |
111 | 1499 gphi *phi = gsi.phi (); |
1500 gphi *phi1 = gsi1.phi (); | |
131 | 1501 stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi); |
111 | 1502 if (dump_enabled_p ()) |
131 | 1503 dump_printf_loc (MSG_NOTE, vect_location, |
1504 "vect_update_ivs_after_vectorizer: phi: %G", phi); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1505 |
111 | 1506 /* Skip reduction and virtual phis. */ |
131 | 1507 if (!iv_phi_p (phi_info)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1508 { |
111 | 1509 if (dump_enabled_p ()) |
1510 dump_printf_loc (MSG_NOTE, vect_location, | |
1511 "reduc or virtual phi. skip.\n"); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1512 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1513 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1514 |
111 | 1515 type = TREE_TYPE (gimple_phi_result (phi)); |
131 | 1516 step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info); |
111 | 1517 step_expr = unshare_expr (step_expr); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1518 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1519 /* FORNOW: We do not support IVs whose evolution function is a polynomial |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1520 of degree >= 2 or exponential. */ |
111 | 1521 gcc_assert (!tree_is_chrec (step_expr)); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1522 |
111 | 1523 init_expr = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1524 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1525 off = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1526 fold_convert (TREE_TYPE (step_expr), niters), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1527 step_expr); |
111 | 1528 if (POINTER_TYPE_P (type)) |
1529 ni = fold_build_pointer_plus (init_expr, off); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1530 else |
111 | 1531 ni = fold_build2 (PLUS_EXPR, type, |
1532 init_expr, fold_convert (type, off)); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1533 |
111 | 1534 var = create_tmp_var (type, "tmp"); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1535 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1536 last_gsi = gsi_last_bb (exit_bb); |
111 | 1537 gimple_seq new_stmts = NULL; |
1538 ni_name = force_gimple_operand (ni, &new_stmts, false, var); | |
1539 /* Exit_bb shouldn't be empty. */ | |
1540 if (!gsi_end_p (last_gsi)) | |
1541 gsi_insert_seq_after (&last_gsi, new_stmts, GSI_SAME_STMT); | |
1542 else | |
1543 gsi_insert_seq_before (&last_gsi, new_stmts, GSI_SAME_STMT); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1544 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1545 /* Fix phi expressions in the successor bb. */ |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1546 adjust_phi_and_debug_stmts (phi1, update_e, ni_name); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1547 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1548 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1549 |
131 | 1550 /* Return a gimple value containing the misalignment (measured in vector |
1551 elements) for the loop described by LOOP_VINFO, i.e. how many elements | |
1552 it is away from a perfectly aligned address. Add any new statements | |
1553 to SEQ. */ | |
1554 | |
1555 static tree | |
1556 get_misalign_in_elems (gimple **seq, loop_vec_info loop_vinfo) | |
1557 { | |
1558 dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
1559 stmt_vec_info stmt_info = dr_info->stmt; | |
1560 tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1561 | |
145 | 1562 poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info); |
1563 unsigned HOST_WIDE_INT target_align_c; | |
1564 tree target_align_minus_1; | |
131 | 1565 |
1566 bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr), | |
1567 size_zero_node) < 0; | |
1568 tree offset = (negative | |
1569 ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) | |
1570 : size_zero_node); | |
1571 tree start_addr = vect_create_addr_base_for_vector_ref (stmt_info, seq, | |
1572 offset); | |
1573 tree type = unsigned_type_for (TREE_TYPE (start_addr)); | |
145 | 1574 if (target_align.is_constant (&target_align_c)) |
1575 target_align_minus_1 = build_int_cst (type, target_align_c - 1); | |
1576 else | |
1577 { | |
1578 tree vla = build_int_cst (type, target_align); | |
1579 tree vla_align = fold_build2 (BIT_AND_EXPR, type, vla, | |
1580 fold_build2 (MINUS_EXPR, type, | |
1581 build_int_cst (type, 0), vla)); | |
1582 target_align_minus_1 = fold_build2 (MINUS_EXPR, type, vla_align, | |
1583 build_int_cst (type, 1)); | |
1584 } | |
1585 | |
131 | 1586 HOST_WIDE_INT elem_size |
1587 = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
1588 tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); | |
1589 | |
1590 /* Create: misalign_in_bytes = addr & (target_align - 1). */ | |
1591 tree int_start_addr = fold_convert (type, start_addr); | |
1592 tree misalign_in_bytes = fold_build2 (BIT_AND_EXPR, type, int_start_addr, | |
1593 target_align_minus_1); | |
1594 | |
1595 /* Create: misalign_in_elems = misalign_in_bytes / element_size. */ | |
1596 tree misalign_in_elems = fold_build2 (RSHIFT_EXPR, type, misalign_in_bytes, | |
1597 elem_size_log); | |
1598 | |
1599 return misalign_in_elems; | |
1600 } | |
1601 | |
111 | 1602 /* Function vect_gen_prolog_loop_niters |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1603 |
111 | 1604 Generate the number of iterations which should be peeled as prolog for the |
1605 loop represented by LOOP_VINFO. It is calculated as the misalignment of | |
1606 DR - the data reference recorded in LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). | |
1607 As a result, after the execution of this loop, the data reference DR will | |
1608 refer to an aligned location. The following computation is generated: | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1609 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1610 If the misalignment of DR is known at compile time: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1611 addr_mis = int mis = DR_MISALIGNMENT (dr); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1612 Else, compute address misalignment in bytes: |
131 | 1613 addr_mis = addr & (target_align - 1) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1614 |
111 | 1615 prolog_niters = ((VF - addr_mis/elem_size)&(VF-1))/step |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1616 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1617 (elem_size = element type size; an element is the scalar element whose type |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1618 is the inner type of the vectype) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1619 |
111 | 1620 The computations will be emitted at the end of BB. We also compute and |
1621 store upper bound (included) of the result in BOUND. | |
1622 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1623 When the step of the data-ref in the loop is not 1 (as in interleaved data |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1624 and SLP), the number of iterations of the prolog must be divided by the step |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1625 (which is equal to the size of interleaved group). |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1626 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1627 The above formulas assume that VF == number of elements in the vector. This |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1628 may not hold when there are multiple-types in the loop. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1629 In this case, for some data-references in the loop the VF does not represent |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1630 the number of elements that fit in the vector. Therefore, instead of VF we |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1631 use TYPE_VECTOR_SUBPARTS. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1632 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1633 static tree |
111 | 1634 vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo, |
1635 basic_block bb, int *bound) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1636 { |
131 | 1637 dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1638 tree var; |
111 | 1639 tree niters_type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)); |
1640 gimple_seq stmts = NULL, new_stmts = NULL; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1641 tree iters, iters_name; |
131 | 1642 stmt_vec_info stmt_info = dr_info->stmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1643 tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
145 | 1644 poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1645 |
111 | 1646 if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1647 { |
111 | 1648 int npeel = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1649 |
111 | 1650 if (dump_enabled_p ()) |
1651 dump_printf_loc (MSG_NOTE, vect_location, | |
1652 "known peeling = %d.\n", npeel); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1653 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1654 iters = build_int_cst (niters_type, npeel); |
111 | 1655 *bound = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1656 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1657 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1658 { |
131 | 1659 tree misalign_in_elems = get_misalign_in_elems (&stmts, loop_vinfo); |
1660 tree type = TREE_TYPE (misalign_in_elems); | |
111 | 1661 HOST_WIDE_INT elem_size |
1662 = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); | |
145 | 1663 /* We only do prolog peeling if the target alignment is known at compile |
1664 time. */ | |
1665 poly_uint64 align_in_elems = | |
1666 exact_div (target_align, elem_size); | |
1667 tree align_in_elems_minus_1 = | |
1668 build_int_cst (type, align_in_elems - 1); | |
111 | 1669 tree align_in_elems_tree = build_int_cst (type, align_in_elems); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1670 |
111 | 1671 /* Create: (niters_type) ((align_in_elems - misalign_in_elems) |
1672 & (align_in_elems - 1)). */ | |
131 | 1673 bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr), |
1674 size_zero_node) < 0; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1675 if (negative) |
111 | 1676 iters = fold_build2 (MINUS_EXPR, type, misalign_in_elems, |
1677 align_in_elems_tree); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1678 else |
111 | 1679 iters = fold_build2 (MINUS_EXPR, type, align_in_elems_tree, |
1680 misalign_in_elems); | |
1681 iters = fold_build2 (BIT_AND_EXPR, type, iters, align_in_elems_minus_1); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1682 iters = fold_convert (niters_type, iters); |
145 | 1683 unsigned HOST_WIDE_INT align_in_elems_c; |
1684 if (align_in_elems.is_constant (&align_in_elems_c)) | |
1685 *bound = align_in_elems_c - 1; | |
1686 else | |
1687 *bound = -1; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1688 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1689 |
111 | 1690 if (dump_enabled_p ()) |
131 | 1691 dump_printf_loc (MSG_NOTE, vect_location, |
1692 "niters for prolog loop: %T\n", iters); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1693 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1694 var = create_tmp_var (niters_type, "prolog_loop_niters"); |
111 | 1695 iters_name = force_gimple_operand (iters, &new_stmts, false, var); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1696 |
111 | 1697 if (new_stmts) |
1698 gimple_seq_add_seq (&stmts, new_stmts); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1699 if (stmts) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1700 { |
111 | 1701 gcc_assert (single_succ_p (bb)); |
1702 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
1703 if (gsi_end_p (gsi)) | |
1704 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT); | |
1705 else | |
1706 gsi_insert_seq_after (&gsi, stmts, GSI_SAME_STMT); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1707 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1708 return iters_name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1709 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1710 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1711 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1712 /* Function vect_update_init_of_dr |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1713 |
131 | 1714 If CODE is PLUS, the vector loop starts NITERS iterations after the |
1715 scalar one, otherwise CODE is MINUS and the vector loop starts NITERS | |
1716 iterations before the scalar one (using masking to skip inactive | |
1717 elements). This function updates the information recorded in DR to | |
1718 account for the difference. Specifically, it updates the OFFSET | |
145 | 1719 field of DR_INFO. */ |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1720 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1721 static void |
145 | 1722 vect_update_init_of_dr (dr_vec_info *dr_info, tree niters, tree_code code) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1723 { |
145 | 1724 struct data_reference *dr = dr_info->dr; |
1725 tree offset = dr_info->offset; | |
1726 if (!offset) | |
1727 offset = build_zero_cst (sizetype); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1728 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1729 niters = fold_build2 (MULT_EXPR, sizetype, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1730 fold_convert (sizetype, niters), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1731 fold_convert (sizetype, DR_STEP (dr))); |
131 | 1732 offset = fold_build2 (code, sizetype, |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1733 fold_convert (sizetype, offset), niters); |
145 | 1734 dr_info->offset = offset; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1735 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1736 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1737 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1738 /* Function vect_update_inits_of_drs |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1739 |
131 | 1740 Apply vect_update_inits_of_dr to all accesses in LOOP_VINFO. |
1741 CODE and NITERS are as for vect_update_inits_of_dr. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1742 |
145 | 1743 void |
131 | 1744 vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters, |
1745 tree_code code) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1746 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1747 unsigned int i; |
111 | 1748 vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1749 struct data_reference *dr; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1750 |
131 | 1751 DUMP_VECT_SCOPE ("vect_update_inits_of_dr"); |
111 | 1752 |
145 | 1753 /* Adjust niters to sizetype. We used to insert the stmts on loop preheader |
1754 here, but since we might use these niters to update the epilogues niters | |
1755 and data references we can't insert them here as this definition might not | |
1756 always dominate its uses. */ | |
111 | 1757 if (!types_compatible_p (sizetype, TREE_TYPE (niters))) |
145 | 1758 niters = fold_convert (sizetype, niters); |
111 | 1759 |
1760 FOR_EACH_VEC_ELT (datarefs, i, dr) | |
131 | 1761 { |
1762 dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr); | |
1763 if (!STMT_VINFO_GATHER_SCATTER_P (dr_info->stmt)) | |
145 | 1764 vect_update_init_of_dr (dr_info, niters, code); |
131 | 1765 } |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1766 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1767 |
131 | 1768 /* For the information recorded in LOOP_VINFO prepare the loop for peeling |
1769 by masking. This involves calculating the number of iterations to | |
1770 be peeled and then aligning all memory references appropriately. */ | |
1771 | |
1772 void | |
1773 vect_prepare_for_masked_peels (loop_vec_info loop_vinfo) | |
1774 { | |
1775 tree misalign_in_elems; | |
1776 tree type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo); | |
1777 | |
1778 gcc_assert (vect_use_loop_mask_for_alignment_p (loop_vinfo)); | |
1779 | |
1780 /* From the information recorded in LOOP_VINFO get the number of iterations | |
1781 that need to be skipped via masking. */ | |
1782 if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) | |
1783 { | |
1784 poly_int64 misalign = (LOOP_VINFO_VECT_FACTOR (loop_vinfo) | |
1785 - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)); | |
1786 misalign_in_elems = build_int_cst (type, misalign); | |
1787 } | |
1788 else | |
1789 { | |
1790 gimple_seq seq1 = NULL, seq2 = NULL; | |
1791 misalign_in_elems = get_misalign_in_elems (&seq1, loop_vinfo); | |
1792 misalign_in_elems = fold_convert (type, misalign_in_elems); | |
1793 misalign_in_elems = force_gimple_operand (misalign_in_elems, | |
1794 &seq2, true, NULL_TREE); | |
1795 gimple_seq_add_seq (&seq1, seq2); | |
1796 if (seq1) | |
1797 { | |
1798 edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); | |
1799 basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq1); | |
1800 gcc_assert (!new_bb); | |
1801 } | |
1802 } | |
1803 | |
1804 if (dump_enabled_p ()) | |
1805 dump_printf_loc (MSG_NOTE, vect_location, | |
1806 "misalignment for fully-masked loop: %T\n", | |
1807 misalign_in_elems); | |
1808 | |
1809 LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo) = misalign_in_elems; | |
1810 | |
1811 vect_update_inits_of_drs (loop_vinfo, misalign_in_elems, MINUS_EXPR); | |
1812 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1813 |
111 | 1814 /* This function builds ni_name = number of iterations. Statements |
1815 are emitted on the loop preheader edge. If NEW_VAR_P is not NULL, set | |
1816 it to TRUE if new ssa_var is generated. */ | |
1817 | |
1818 tree | |
1819 vect_build_loop_niters (loop_vec_info loop_vinfo, bool *new_var_p) | |
1820 { | |
1821 tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo)); | |
1822 if (TREE_CODE (ni) == INTEGER_CST) | |
1823 return ni; | |
1824 else | |
1825 { | |
1826 tree ni_name, var; | |
1827 gimple_seq stmts = NULL; | |
1828 edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); | |
1829 | |
1830 var = create_tmp_var (TREE_TYPE (ni), "niters"); | |
1831 ni_name = force_gimple_operand (ni, &stmts, false, var); | |
1832 if (stmts) | |
1833 { | |
1834 gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1835 if (new_var_p != NULL) | |
1836 *new_var_p = true; | |
1837 } | |
1838 | |
1839 return ni_name; | |
1840 } | |
1841 } | |
1842 | |
1843 /* Calculate the number of iterations above which vectorized loop will be | |
1844 preferred than scalar loop. NITERS_PROLOG is the number of iterations | |
1845 of prolog loop. If it's integer const, the integer number is also passed | |
131 | 1846 in INT_NITERS_PROLOG. BOUND_PROLOG is the upper bound (inclusive) of the |
1847 number of iterations of the prolog loop. BOUND_EPILOG is the corresponding | |
1848 value for the epilog loop. If CHECK_PROFITABILITY is true, TH is the | |
1849 threshold below which the scalar (rather than vectorized) loop will be | |
1850 executed. This function stores the upper bound (inclusive) of the result | |
1851 in BOUND_SCALAR. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1852 |
111 | 1853 static tree |
1854 vect_gen_scalar_loop_niters (tree niters_prolog, int int_niters_prolog, | |
131 | 1855 int bound_prolog, poly_int64 bound_epilog, int th, |
1856 poly_uint64 *bound_scalar, | |
1857 bool check_profitability) | |
111 | 1858 { |
1859 tree type = TREE_TYPE (niters_prolog); | |
1860 tree niters = fold_build2 (PLUS_EXPR, type, niters_prolog, | |
131 | 1861 build_int_cst (type, bound_epilog)); |
1862 | |
1863 *bound_scalar = bound_prolog + bound_epilog; | |
111 | 1864 if (check_profitability) |
1865 { | |
1866 /* TH indicates the minimum niters of vectorized loop, while we | |
1867 compute the maximum niters of scalar loop. */ | |
1868 th--; | |
1869 /* Peeling for constant times. */ | |
1870 if (int_niters_prolog >= 0) | |
1871 { | |
131 | 1872 *bound_scalar = upper_bound (int_niters_prolog + bound_epilog, th); |
111 | 1873 return build_int_cst (type, *bound_scalar); |
1874 } | |
131 | 1875 /* Peeling an unknown number of times. Note that both BOUND_PROLOG |
1876 and BOUND_EPILOG are inclusive upper bounds. */ | |
1877 if (known_ge (th, bound_prolog + bound_epilog)) | |
111 | 1878 { |
1879 *bound_scalar = th; | |
1880 return build_int_cst (type, th); | |
1881 } | |
131 | 1882 /* Need to do runtime comparison. */ |
1883 else if (maybe_gt (th, bound_epilog)) | |
1884 { | |
1885 *bound_scalar = upper_bound (*bound_scalar, th); | |
1886 return fold_build2 (MAX_EXPR, type, | |
1887 build_int_cst (type, th), niters); | |
1888 } | |
111 | 1889 } |
1890 return niters; | |
1891 } | |
1892 | |
131 | 1893 /* NITERS is the number of times that the original scalar loop executes |
1894 after peeling. Work out the maximum number of iterations N that can | |
1895 be handled by the vectorized form of the loop and then either: | |
1896 | |
1897 a) set *STEP_VECTOR_PTR to the vectorization factor and generate: | |
1898 | |
1899 niters_vector = N | |
1900 | |
1901 b) set *STEP_VECTOR_PTR to one and generate: | |
1902 | |
1903 niters_vector = N / vf | |
1904 | |
1905 In both cases, store niters_vector in *NITERS_VECTOR_PTR and add | |
1906 any new statements on the loop preheader edge. NITERS_NO_OVERFLOW | |
1907 is true if NITERS doesn't overflow (i.e. if NITERS is always nonzero). */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1908 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1909 void |
111 | 1910 vect_gen_vector_loop_niters (loop_vec_info loop_vinfo, tree niters, |
131 | 1911 tree *niters_vector_ptr, tree *step_vector_ptr, |
1912 bool niters_no_overflow) | |
111 | 1913 { |
1914 tree ni_minus_gap, var; | |
131 | 1915 tree niters_vector, step_vector, type = TREE_TYPE (niters); |
1916 poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
111 | 1917 edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo)); |
131 | 1918 tree log_vf = NULL_TREE; |
111 | 1919 |
1920 /* If epilogue loop is required because of data accesses with gaps, we | |
1921 subtract one iteration from the total number of iterations here for | |
1922 correct calculation of RATIO. */ | |
1923 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
1924 { | |
1925 ni_minus_gap = fold_build2 (MINUS_EXPR, type, niters, | |
1926 build_one_cst (type)); | |
1927 if (!is_gimple_val (ni_minus_gap)) | |
1928 { | |
1929 var = create_tmp_var (type, "ni_gap"); | |
1930 gimple *stmts = NULL; | |
1931 ni_minus_gap = force_gimple_operand (ni_minus_gap, &stmts, | |
1932 true, var); | |
1933 gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1934 } | |
1935 } | |
1936 else | |
1937 ni_minus_gap = niters; | |
1938 | |
131 | 1939 unsigned HOST_WIDE_INT const_vf; |
1940 if (vf.is_constant (&const_vf) | |
1941 && !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)) | |
1942 { | |
1943 /* Create: niters >> log2(vf) */ | |
1944 /* If it's known that niters == number of latch executions + 1 doesn't | |
1945 overflow, we can generate niters >> log2(vf); otherwise we generate | |
1946 (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio | |
1947 will be at least one. */ | |
1948 log_vf = build_int_cst (type, exact_log2 (const_vf)); | |
1949 if (niters_no_overflow) | |
1950 niters_vector = fold_build2 (RSHIFT_EXPR, type, ni_minus_gap, log_vf); | |
1951 else | |
1952 niters_vector | |
1953 = fold_build2 (PLUS_EXPR, type, | |
1954 fold_build2 (RSHIFT_EXPR, type, | |
1955 fold_build2 (MINUS_EXPR, type, | |
1956 ni_minus_gap, | |
1957 build_int_cst (type, vf)), | |
1958 log_vf), | |
1959 build_int_cst (type, 1)); | |
1960 step_vector = build_one_cst (type); | |
1961 } | |
111 | 1962 else |
131 | 1963 { |
1964 niters_vector = ni_minus_gap; | |
1965 step_vector = build_int_cst (type, vf); | |
1966 } | |
111 | 1967 |
1968 if (!is_gimple_val (niters_vector)) | |
1969 { | |
1970 var = create_tmp_var (type, "bnd"); | |
1971 gimple_seq stmts = NULL; | |
1972 niters_vector = force_gimple_operand (niters_vector, &stmts, true, var); | |
1973 gsi_insert_seq_on_edge_immediate (pe, stmts); | |
1974 /* Peeling algorithm guarantees that vector loop bound is at least ONE, | |
1975 we set range information to make niters analyzer's life easier. */ | |
131 | 1976 if (stmts != NULL && log_vf) |
111 | 1977 set_range_info (niters_vector, VR_RANGE, |
1978 wi::to_wide (build_int_cst (type, 1)), | |
1979 wi::to_wide (fold_build2 (RSHIFT_EXPR, type, | |
1980 TYPE_MAX_VALUE (type), | |
1981 log_vf))); | |
1982 } | |
1983 *niters_vector_ptr = niters_vector; | |
131 | 1984 *step_vector_ptr = step_vector; |
111 | 1985 |
1986 return; | |
1987 } | |
1988 | |
1989 /* Given NITERS_VECTOR which is the number of iterations for vectorized | |
1990 loop specified by LOOP_VINFO after vectorization, compute the number | |
1991 of iterations before vectorization (niters_vector * vf) and store it | |
1992 to NITERS_VECTOR_MULT_VF_PTR. */ | |
1993 | |
1994 static void | |
1995 vect_gen_vector_loop_niters_mult_vf (loop_vec_info loop_vinfo, | |
1996 tree niters_vector, | |
1997 tree *niters_vector_mult_vf_ptr) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
1998 { |
131 | 1999 /* We should be using a step_vector of VF if VF is variable. */ |
2000 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant (); | |
145 | 2001 class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
111 | 2002 tree type = TREE_TYPE (niters_vector); |
2003 tree log_vf = build_int_cst (type, exact_log2 (vf)); | |
2004 basic_block exit_bb = single_exit (loop)->dest; | |
2005 | |
2006 gcc_assert (niters_vector_mult_vf_ptr != NULL); | |
2007 tree niters_vector_mult_vf = fold_build2 (LSHIFT_EXPR, type, | |
2008 niters_vector, log_vf); | |
2009 if (!is_gimple_val (niters_vector_mult_vf)) | |
2010 { | |
2011 tree var = create_tmp_var (type, "niters_vector_mult_vf"); | |
2012 gimple_seq stmts = NULL; | |
2013 niters_vector_mult_vf = force_gimple_operand (niters_vector_mult_vf, | |
2014 &stmts, true, var); | |
2015 gimple_stmt_iterator gsi = gsi_start_bb (exit_bb); | |
2016 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT); | |
2017 } | |
2018 *niters_vector_mult_vf_ptr = niters_vector_mult_vf; | |
2019 } | |
2020 | |
145 | 2021 /* LCSSA_PHI is a lcssa phi of EPILOG loop which is copied from LOOP, |
2022 this function searches for the corresponding lcssa phi node in exit | |
2023 bb of LOOP. If it is found, return the phi result; otherwise return | |
2024 NULL. */ | |
2025 | |
2026 static tree | |
2027 find_guard_arg (class loop *loop, class loop *epilog ATTRIBUTE_UNUSED, | |
2028 gphi *lcssa_phi) | |
2029 { | |
2030 gphi_iterator gsi; | |
2031 edge e = single_exit (loop); | |
2032 | |
2033 gcc_assert (single_pred_p (e->dest)); | |
2034 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2035 { | |
2036 gphi *phi = gsi.phi (); | |
2037 if (operand_equal_p (PHI_ARG_DEF (phi, 0), | |
2038 PHI_ARG_DEF (lcssa_phi, 0), 0)) | |
2039 return PHI_RESULT (phi); | |
2040 } | |
2041 return NULL_TREE; | |
2042 } | |
2043 | |
111 | 2044 /* Function slpeel_tree_duplicate_loop_to_edge_cfg duplciates FIRST/SECOND |
2045 from SECOND/FIRST and puts it at the original loop's preheader/exit | |
2046 edge, the two loops are arranged as below: | |
2047 | |
2048 preheader_a: | |
2049 first_loop: | |
2050 header_a: | |
2051 i_1 = PHI<i_0, i_2>; | |
2052 ... | |
2053 i_2 = i_1 + 1; | |
2054 if (cond_a) | |
2055 goto latch_a; | |
2056 else | |
2057 goto between_bb; | |
2058 latch_a: | |
2059 goto header_a; | |
2060 | |
2061 between_bb: | |
2062 ;; i_x = PHI<i_2>; ;; LCSSA phi node to be created for FIRST, | |
2063 | |
2064 second_loop: | |
2065 header_b: | |
2066 i_3 = PHI<i_0, i_4>; ;; Use of i_0 to be replaced with i_x, | |
2067 or with i_2 if no LCSSA phi is created | |
2068 under condition of CREATE_LCSSA_FOR_IV_PHIS. | |
2069 ... | |
2070 i_4 = i_3 + 1; | |
2071 if (cond_b) | |
2072 goto latch_b; | |
2073 else | |
2074 goto exit_bb; | |
2075 latch_b: | |
2076 goto header_b; | |
2077 | |
2078 exit_bb: | |
2079 | |
2080 This function creates loop closed SSA for the first loop; update the | |
2081 second loop's PHI nodes by replacing argument on incoming edge with the | |
2082 result of newly created lcssa PHI nodes. IF CREATE_LCSSA_FOR_IV_PHIS | |
2083 is false, Loop closed ssa phis will only be created for non-iv phis for | |
2084 the first loop. | |
2085 | |
2086 This function assumes exit bb of the first loop is preheader bb of the | |
2087 second loop, i.e, between_bb in the example code. With PHIs updated, | |
2088 the second loop will execute rest iterations of the first. */ | |
2089 | |
2090 static void | |
2091 slpeel_update_phi_nodes_for_loops (loop_vec_info loop_vinfo, | |
145 | 2092 class loop *first, class loop *second, |
111 | 2093 bool create_lcssa_for_iv_phis) |
2094 { | |
2095 gphi_iterator gsi_update, gsi_orig; | |
145 | 2096 class loop *loop = LOOP_VINFO_LOOP (loop_vinfo); |
111 | 2097 |
2098 edge first_latch_e = EDGE_SUCC (first->latch, 0); | |
2099 edge second_preheader_e = loop_preheader_edge (second); | |
2100 basic_block between_bb = single_exit (first)->dest; | |
2101 | |
2102 gcc_assert (between_bb == second_preheader_e->src); | |
2103 gcc_assert (single_pred_p (between_bb) && single_succ_p (between_bb)); | |
2104 /* Either the first loop or the second is the loop to be vectorized. */ | |
2105 gcc_assert (loop == first || loop == second); | |
2106 | |
2107 for (gsi_orig = gsi_start_phis (first->header), | |
2108 gsi_update = gsi_start_phis (second->header); | |
2109 !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
2110 gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
2111 { | |
2112 gphi *orig_phi = gsi_orig.phi (); | |
2113 gphi *update_phi = gsi_update.phi (); | |
2114 | |
2115 tree arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, first_latch_e); | |
2116 /* Generate lcssa PHI node for the first loop. */ | |
2117 gphi *vect_phi = (loop == first) ? orig_phi : update_phi; | |
131 | 2118 stmt_vec_info vect_phi_info = loop_vinfo->lookup_stmt (vect_phi); |
2119 if (create_lcssa_for_iv_phis || !iv_phi_p (vect_phi_info)) | |
111 | 2120 { |
2121 tree new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
2122 gphi *lcssa_phi = create_phi_node (new_res, between_bb); | |
2123 add_phi_arg (lcssa_phi, arg, single_exit (first), UNKNOWN_LOCATION); | |
2124 arg = new_res; | |
2125 } | |
2126 | |
2127 /* Update PHI node in the second loop by replacing arg on the loop's | |
2128 incoming edge. */ | |
2129 adjust_phi_and_debug_stmts (update_phi, second_preheader_e, arg); | |
2130 } | |
145 | 2131 |
2132 /* For epilogue peeling we have to make sure to copy all LC PHIs | |
2133 for correct vectorization of live stmts. */ | |
2134 if (loop == first) | |
2135 { | |
2136 basic_block orig_exit = single_exit (second)->dest; | |
2137 for (gsi_orig = gsi_start_phis (orig_exit); | |
2138 !gsi_end_p (gsi_orig); gsi_next (&gsi_orig)) | |
2139 { | |
2140 gphi *orig_phi = gsi_orig.phi (); | |
2141 tree orig_arg = PHI_ARG_DEF (orig_phi, 0); | |
2142 if (TREE_CODE (orig_arg) != SSA_NAME || virtual_operand_p (orig_arg)) | |
2143 continue; | |
2144 | |
2145 /* Already created in the above loop. */ | |
2146 if (find_guard_arg (first, second, orig_phi)) | |
2147 continue; | |
2148 | |
2149 tree new_res = copy_ssa_name (orig_arg); | |
2150 gphi *lcphi = create_phi_node (new_res, between_bb); | |
2151 add_phi_arg (lcphi, orig_arg, single_exit (first), UNKNOWN_LOCATION); | |
2152 } | |
2153 } | |
111 | 2154 } |
2155 | |
2156 /* Function slpeel_add_loop_guard adds guard skipping from the beginning | |
2157 of SKIP_LOOP to the beginning of UPDATE_LOOP. GUARD_EDGE and MERGE_EDGE | |
2158 are two pred edges of the merge point before UPDATE_LOOP. The two loops | |
2159 appear like below: | |
2160 | |
2161 guard_bb: | |
2162 if (cond) | |
2163 goto merge_bb; | |
2164 else | |
2165 goto skip_loop; | |
2166 | |
2167 skip_loop: | |
2168 header_a: | |
2169 i_1 = PHI<i_0, i_2>; | |
2170 ... | |
2171 i_2 = i_1 + 1; | |
2172 if (cond_a) | |
2173 goto latch_a; | |
2174 else | |
2175 goto exit_a; | |
2176 latch_a: | |
2177 goto header_a; | |
2178 | |
2179 exit_a: | |
2180 i_5 = PHI<i_2>; | |
2181 | |
2182 merge_bb: | |
2183 ;; PHI (i_x = PHI<i_0, i_5>) to be created at merge point. | |
2184 | |
2185 update_loop: | |
2186 header_b: | |
2187 i_3 = PHI<i_5, i_4>; ;; Use of i_5 to be replaced with i_x. | |
2188 ... | |
2189 i_4 = i_3 + 1; | |
2190 if (cond_b) | |
2191 goto latch_b; | |
2192 else | |
2193 goto exit_bb; | |
2194 latch_b: | |
2195 goto header_b; | |
2196 | |
2197 exit_bb: | |
2198 | |
2199 This function creates PHI nodes at merge_bb and replaces the use of i_5 | |
2200 in the update_loop's PHI node with the result of new PHI result. */ | |
2201 | |
2202 static void | |
145 | 2203 slpeel_update_phi_nodes_for_guard1 (class loop *skip_loop, |
2204 class loop *update_loop, | |
111 | 2205 edge guard_edge, edge merge_edge) |
2206 { | |
145 | 2207 location_t merge_loc, guard_loc; |
111 | 2208 edge orig_e = loop_preheader_edge (skip_loop); |
2209 edge update_e = loop_preheader_edge (update_loop); | |
2210 gphi_iterator gsi_orig, gsi_update; | |
2211 | |
2212 for ((gsi_orig = gsi_start_phis (skip_loop->header), | |
2213 gsi_update = gsi_start_phis (update_loop->header)); | |
2214 !gsi_end_p (gsi_orig) && !gsi_end_p (gsi_update); | |
2215 gsi_next (&gsi_orig), gsi_next (&gsi_update)) | |
2216 { | |
2217 gphi *orig_phi = gsi_orig.phi (); | |
2218 gphi *update_phi = gsi_update.phi (); | |
2219 | |
2220 /* Generate new phi node at merge bb of the guard. */ | |
2221 tree new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
2222 gphi *new_phi = create_phi_node (new_res, guard_edge->dest); | |
2223 | |
2224 /* Merge bb has two incoming edges: GUARD_EDGE and MERGE_EDGE. Set the | |
2225 args in NEW_PHI for these edges. */ | |
2226 tree merge_arg = PHI_ARG_DEF_FROM_EDGE (update_phi, update_e); | |
2227 tree guard_arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, orig_e); | |
2228 merge_loc = gimple_phi_arg_location_from_edge (update_phi, update_e); | |
2229 guard_loc = gimple_phi_arg_location_from_edge (orig_phi, orig_e); | |
2230 add_phi_arg (new_phi, merge_arg, merge_edge, merge_loc); | |
2231 add_phi_arg (new_phi, guard_arg, guard_edge, guard_loc); | |
2232 | |
2233 /* Update phi in UPDATE_PHI. */ | |
2234 adjust_phi_and_debug_stmts (update_phi, update_e, new_res); | |
2235 } | |
2236 } | |
2237 | |
2238 /* LOOP and EPILOG are two consecutive loops in CFG and EPILOG is copied | |
2239 from LOOP. Function slpeel_add_loop_guard adds guard skipping from a | |
2240 point between the two loops to the end of EPILOG. Edges GUARD_EDGE | |
2241 and MERGE_EDGE are the two pred edges of merge_bb at the end of EPILOG. | |
2242 The CFG looks like: | |
2243 | |
2244 loop: | |
2245 header_a: | |
2246 i_1 = PHI<i_0, i_2>; | |
2247 ... | |
2248 i_2 = i_1 + 1; | |
2249 if (cond_a) | |
2250 goto latch_a; | |
2251 else | |
2252 goto exit_a; | |
2253 latch_a: | |
2254 goto header_a; | |
2255 | |
2256 exit_a: | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2257 |
111 | 2258 guard_bb: |
2259 if (cond) | |
2260 goto merge_bb; | |
2261 else | |
2262 goto epilog_loop; | |
2263 | |
2264 ;; fall_through_bb | |
2265 | |
2266 epilog_loop: | |
2267 header_b: | |
2268 i_3 = PHI<i_2, i_4>; | |
2269 ... | |
2270 i_4 = i_3 + 1; | |
2271 if (cond_b) | |
2272 goto latch_b; | |
2273 else | |
2274 goto merge_bb; | |
2275 latch_b: | |
2276 goto header_b; | |
2277 | |
2278 merge_bb: | |
2279 ; PHI node (i_y = PHI<i_2, i_4>) to be created at merge point. | |
2280 | |
2281 exit_bb: | |
2282 i_x = PHI<i_4>; ;Use of i_4 to be replaced with i_y in merge_bb. | |
2283 | |
2284 For each name used out side EPILOG (i.e - for each name that has a lcssa | |
2285 phi in exit_bb) we create a new PHI in merge_bb. The new PHI has two | |
2286 args corresponding to GUARD_EDGE and MERGE_EDGE. Arg for MERGE_EDGE is | |
2287 the arg of the original PHI in exit_bb, arg for GUARD_EDGE is defined | |
2288 by LOOP and is found in the exit bb of LOOP. Arg of the original PHI | |
2289 in exit_bb will also be updated. */ | |
2290 | |
2291 static void | |
145 | 2292 slpeel_update_phi_nodes_for_guard2 (class loop *loop, class loop *epilog, |
111 | 2293 edge guard_edge, edge merge_edge) |
2294 { | |
2295 gphi_iterator gsi; | |
2296 basic_block merge_bb = guard_edge->dest; | |
2297 | |
2298 gcc_assert (single_succ_p (merge_bb)); | |
2299 edge e = single_succ_edge (merge_bb); | |
2300 basic_block exit_bb = e->dest; | |
2301 gcc_assert (single_pred_p (exit_bb)); | |
2302 gcc_assert (single_pred (exit_bb) == single_exit (epilog)->dest); | |
2303 | |
2304 for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2305 { | |
2306 gphi *update_phi = gsi.phi (); | |
2307 tree old_arg = PHI_ARG_DEF (update_phi, 0); | |
145 | 2308 |
2309 tree merge_arg = NULL_TREE; | |
2310 | |
2311 /* If the old argument is a SSA_NAME use its current_def. */ | |
2312 if (TREE_CODE (old_arg) == SSA_NAME) | |
2313 merge_arg = get_current_def (old_arg); | |
2314 /* If it's a constant or doesn't have a current_def, just use the old | |
2315 argument. */ | |
111 | 2316 if (!merge_arg) |
2317 merge_arg = old_arg; | |
2318 | |
2319 tree guard_arg = find_guard_arg (loop, epilog, update_phi); | |
2320 /* If the var is live after loop but not a reduction, we simply | |
2321 use the old arg. */ | |
2322 if (!guard_arg) | |
2323 guard_arg = old_arg; | |
2324 | |
2325 /* Create new phi node in MERGE_BB: */ | |
2326 tree new_res = copy_ssa_name (PHI_RESULT (update_phi)); | |
2327 gphi *merge_phi = create_phi_node (new_res, merge_bb); | |
2328 | |
2329 /* MERGE_BB has two incoming edges: GUARD_EDGE and MERGE_EDGE, Set | |
2330 the two PHI args in merge_phi for these edges. */ | |
2331 add_phi_arg (merge_phi, merge_arg, merge_edge, UNKNOWN_LOCATION); | |
2332 add_phi_arg (merge_phi, guard_arg, guard_edge, UNKNOWN_LOCATION); | |
2333 | |
2334 /* Update the original phi in exit_bb. */ | |
2335 adjust_phi_and_debug_stmts (update_phi, e, new_res); | |
2336 } | |
2337 } | |
2338 | |
2339 /* EPILOG loop is duplicated from the original loop for vectorizing, | |
2340 the arg of its loop closed ssa PHI needs to be updated. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2341 |
111 | 2342 static void |
145 | 2343 slpeel_update_phi_nodes_for_lcssa (class loop *epilog) |
111 | 2344 { |
2345 gphi_iterator gsi; | |
2346 basic_block exit_bb = single_exit (epilog)->dest; | |
2347 | |
2348 gcc_assert (single_pred_p (exit_bb)); | |
2349 edge e = EDGE_PRED (exit_bb, 0); | |
2350 for (gsi = gsi_start_phis (exit_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2351 rename_use_op (PHI_ARG_DEF_PTR_FROM_EDGE (gsi.phi (), e)); | |
2352 } | |
2353 | |
2354 /* Function vect_do_peeling. | |
2355 | |
2356 Input: | |
2357 - LOOP_VINFO: Represent a loop to be vectorized, which looks like: | |
2358 | |
2359 preheader: | |
2360 LOOP: | |
2361 header_bb: | |
2362 loop_body | |
2363 if (exit_loop_cond) goto exit_bb | |
2364 else goto header_bb | |
2365 exit_bb: | |
2366 | |
2367 - NITERS: The number of iterations of the loop. | |
2368 - NITERSM1: The number of iterations of the loop's latch. | |
2369 - NITERS_NO_OVERFLOW: No overflow in computing NITERS. | |
2370 - TH, CHECK_PROFITABILITY: Threshold of niters to vectorize loop if | |
2371 CHECK_PROFITABILITY is true. | |
2372 Output: | |
131 | 2373 - *NITERS_VECTOR and *STEP_VECTOR describe how the main loop should |
2374 iterate after vectorization; see vect_set_loop_condition for details. | |
2375 - *NITERS_VECTOR_MULT_VF_VAR is either null or an SSA name that | |
2376 should be set to the number of scalar iterations handled by the | |
2377 vector loop. The SSA name is only used on exit from the loop. | |
111 | 2378 |
2379 This function peels prolog and epilog from the loop, adds guards skipping | |
2380 PROLOG and EPILOG for various conditions. As a result, the changed CFG | |
2381 would look like: | |
2382 | |
2383 guard_bb_1: | |
2384 if (prefer_scalar_loop) goto merge_bb_1 | |
2385 else goto guard_bb_2 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2386 |
111 | 2387 guard_bb_2: |
2388 if (skip_prolog) goto merge_bb_2 | |
2389 else goto prolog_preheader | |
2390 | |
2391 prolog_preheader: | |
2392 PROLOG: | |
2393 prolog_header_bb: | |
2394 prolog_body | |
2395 if (exit_prolog_cond) goto prolog_exit_bb | |
2396 else goto prolog_header_bb | |
2397 prolog_exit_bb: | |
2398 | |
2399 merge_bb_2: | |
2400 | |
2401 vector_preheader: | |
2402 VECTOR LOOP: | |
2403 vector_header_bb: | |
2404 vector_body | |
2405 if (exit_vector_cond) goto vector_exit_bb | |
2406 else goto vector_header_bb | |
2407 vector_exit_bb: | |
2408 | |
2409 guard_bb_3: | |
2410 if (skip_epilog) goto merge_bb_3 | |
2411 else goto epilog_preheader | |
2412 | |
2413 merge_bb_1: | |
2414 | |
2415 epilog_preheader: | |
2416 EPILOG: | |
2417 epilog_header_bb: | |
2418 epilog_body | |
2419 if (exit_epilog_cond) goto merge_bb_3 | |
2420 else goto epilog_header_bb | |
2421 | |
2422 merge_bb_3: | |
2423 | |
2424 Note this function peels prolog and epilog only if it's necessary, | |
2425 as well as guards. | |
145 | 2426 This function returns the epilogue loop if a decision was made to vectorize |
2427 it, otherwise NULL. | |
2428 | |
2429 The analysis resulting in this epilogue loop's loop_vec_info was performed | |
2430 in the same vect_analyze_loop call as the main loop's. At that time | |
2431 vect_analyze_loop constructs a list of accepted loop_vec_info's for lower | |
2432 vectorization factors than the main loop. This list is stored in the main | |
2433 loop's loop_vec_info in the 'epilogue_vinfos' member. Everytime we decide to | |
2434 vectorize the epilogue loop for a lower vectorization factor, the | |
2435 loop_vec_info sitting at the top of the epilogue_vinfos list is removed, | |
2436 updated and linked to the epilogue loop. This is later used to vectorize | |
2437 the epilogue. The reason the loop_vec_info needs updating is that it was | |
2438 constructed based on the original main loop, and the epilogue loop is a | |
2439 copy of this loop, so all links pointing to statements in the original loop | |
2440 need updating. Furthermore, these loop_vec_infos share the | |
2441 data_reference's records, which will also need to be updated. | |
111 | 2442 |
2443 TODO: Guard for prefer_scalar_loop should be emitted along with | |
2444 versioning conditions if loop versioning is needed. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2445 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2446 |
145 | 2447 class loop * |
111 | 2448 vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1, |
131 | 2449 tree *niters_vector, tree *step_vector, |
2450 tree *niters_vector_mult_vf_var, int th, | |
145 | 2451 bool check_profitability, bool niters_no_overflow, |
2452 tree *advance) | |
111 | 2453 { |
2454 edge e, guard_e; | |
2455 tree type = TREE_TYPE (niters), guard_cond; | |
2456 basic_block guard_bb, guard_to; | |
2457 profile_probability prob_prolog, prob_vector, prob_epilog; | |
131 | 2458 int estimated_vf; |
2459 int prolog_peeling = 0; | |
145 | 2460 bool vect_epilogues = loop_vinfo->epilogue_vinfos.length () > 0; |
2461 /* We currently do not support prolog peeling if the target alignment is not | |
2462 known at compile time. 'vect_gen_prolog_loop_niters' depends on the | |
2463 target alignment being constant. */ | |
2464 dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo); | |
2465 if (dr_info && !DR_TARGET_ALIGNMENT (dr_info).is_constant ()) | |
2466 return NULL; | |
2467 | |
131 | 2468 if (!vect_use_loop_mask_for_alignment_p (loop_vinfo)) |
2469 prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo); | |
2470 | |
2471 poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
2472 poly_uint64 bound_epilog = 0; | |
2473 if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) | |
2474 && LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo)) | |
2475 bound_epilog += vf - 1; | |
2476 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
2477 bound_epilog += 1; | |
2478 bool epilog_peeling = maybe_ne (bound_epilog, 0U); | |
2479 poly_uint64 bound_scalar = bound_epilog; | |
111 | 2480 |
2481 if (!prolog_peeling && !epilog_peeling) | |
2482 return NULL; | |
2483 | |
2484 prob_vector = profile_probability::guessed_always ().apply_scale (9, 10); | |
131 | 2485 estimated_vf = vect_vf_for_cost (loop_vinfo); |
2486 if (estimated_vf == 2) | |
2487 estimated_vf = 3; | |
111 | 2488 prob_prolog = prob_epilog = profile_probability::guessed_always () |
131 | 2489 .apply_scale (estimated_vf - 1, estimated_vf); |
111 | 2490 |
145 | 2491 class loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo); |
2492 class loop *first_loop = loop; | |
111 | 2493 bool irred_flag = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP; |
145 | 2494 |
2495 /* We might have a queued need to update virtual SSA form. As we | |
2496 delete the update SSA machinery below after doing a regular | |
2497 incremental SSA update during loop copying make sure we don't | |
2498 lose that fact. | |
2499 ??? Needing to update virtual SSA form by renaming is unfortunate | |
2500 but not all of the vectorizer code inserting new loads / stores | |
2501 properly assigns virtual operands to those statements. */ | |
2502 update_ssa (TODO_update_ssa_only_virtuals); | |
2503 | |
111 | 2504 create_lcssa_for_virtual_phi (loop); |
145 | 2505 |
2506 /* If we're vectorizing an epilogue loop, the update_ssa above will | |
2507 have ensured that the virtual operand is in SSA form throughout the | |
2508 vectorized main loop. Normally it is possible to trace the updated | |
2509 vector-stmt vdefs back to scalar-stmt vdefs and vector-stmt vuses | |
2510 back to scalar-stmt vuses, meaning that the effect of the SSA update | |
2511 remains local to the main loop. However, there are rare cases in | |
2512 which the vectorized loop has vdefs even when the original scalar | |
2513 loop didn't. For example, vectorizing a load with IFN_LOAD_LANES | |
2514 introduces clobbers of the temporary vector array, which in turn | |
2515 needs new vdefs. If the scalar loop doesn't write to memory, these | |
2516 new vdefs will be the only ones in the vector loop. | |
2517 | |
2518 In that case, update_ssa will have added a new virtual phi to the | |
2519 main loop, which previously didn't need one. Ensure that we (locally) | |
2520 maintain LCSSA form for the virtual operand, just as we would have | |
2521 done if the virtual phi had existed from the outset. This makes it | |
2522 easier to duplicate the scalar epilogue loop below. */ | |
2523 tree vop_to_rename = NULL_TREE; | |
2524 if (loop_vec_info orig_loop_vinfo = LOOP_VINFO_ORIG_LOOP_INFO (loop_vinfo)) | |
2525 { | |
2526 class loop *orig_loop = LOOP_VINFO_LOOP (orig_loop_vinfo); | |
2527 vop_to_rename = create_lcssa_for_virtual_phi (orig_loop); | |
2528 } | |
111 | 2529 |
131 | 2530 if (MAY_HAVE_DEBUG_BIND_STMTS) |
111 | 2531 { |
2532 gcc_assert (!adjust_vec.exists ()); | |
2533 adjust_vec.create (32); | |
2534 } | |
2535 initialize_original_copy_tables (); | |
2536 | |
131 | 2537 /* Record the anchor bb at which the guard should be placed if the scalar |
2538 loop might be preferred. */ | |
2539 basic_block anchor = loop_preheader_edge (loop)->src; | |
2540 | |
2541 /* Generate the number of iterations for the prolog loop. We do this here | |
2542 so that we can also get the upper bound on the number of iterations. */ | |
2543 tree niters_prolog; | |
2544 int bound_prolog = 0; | |
2545 if (prolog_peeling) | |
2546 niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor, | |
145 | 2547 &bound_prolog); |
131 | 2548 else |
2549 niters_prolog = build_int_cst (type, 0); | |
2550 | |
145 | 2551 loop_vec_info epilogue_vinfo = NULL; |
2552 if (vect_epilogues) | |
2553 { | |
2554 epilogue_vinfo = loop_vinfo->epilogue_vinfos[0]; | |
2555 loop_vinfo->epilogue_vinfos.ordered_remove (0); | |
2556 } | |
2557 | |
2558 tree niters_vector_mult_vf = NULL_TREE; | |
2559 /* Saving NITERs before the loop, as this may be changed by prologue. */ | |
2560 tree before_loop_niters = LOOP_VINFO_NITERS (loop_vinfo); | |
2561 edge update_e = NULL, skip_e = NULL; | |
2562 unsigned int lowest_vf = constant_lower_bound (vf); | |
2563 /* If we know the number of scalar iterations for the main loop we should | |
2564 check whether after the main loop there are enough iterations left over | |
2565 for the epilogue. */ | |
2566 if (vect_epilogues | |
2567 && LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) | |
2568 && prolog_peeling >= 0 | |
2569 && known_eq (vf, lowest_vf)) | |
2570 { | |
2571 unsigned HOST_WIDE_INT eiters | |
2572 = (LOOP_VINFO_INT_NITERS (loop_vinfo) | |
2573 - LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)); | |
2574 | |
2575 eiters -= prolog_peeling; | |
2576 eiters | |
2577 = eiters % lowest_vf + LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo); | |
2578 | |
2579 unsigned int ratio; | |
2580 unsigned int epilogue_gaps | |
2581 = LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo); | |
2582 while (!(constant_multiple_p | |
2583 (GET_MODE_SIZE (loop_vinfo->vector_mode), | |
2584 GET_MODE_SIZE (epilogue_vinfo->vector_mode), &ratio) | |
2585 && eiters >= lowest_vf / ratio + epilogue_gaps)) | |
2586 { | |
2587 delete epilogue_vinfo; | |
2588 epilogue_vinfo = NULL; | |
2589 if (loop_vinfo->epilogue_vinfos.length () == 0) | |
2590 { | |
2591 vect_epilogues = false; | |
2592 break; | |
2593 } | |
2594 epilogue_vinfo = loop_vinfo->epilogue_vinfos[0]; | |
2595 loop_vinfo->epilogue_vinfos.ordered_remove (0); | |
2596 epilogue_gaps = LOOP_VINFO_PEELING_FOR_GAPS (epilogue_vinfo); | |
2597 } | |
2598 } | |
111 | 2599 /* Prolog loop may be skipped. */ |
2600 bool skip_prolog = (prolog_peeling != 0); | |
145 | 2601 /* Skip this loop to epilog when there are not enough iterations to enter this |
2602 vectorized loop. If true we should perform runtime checks on the NITERS | |
2603 to check whether we should skip the current vectorized loop. If we know | |
2604 the number of scalar iterations we may choose to add a runtime check if | |
2605 this number "maybe" smaller than the number of iterations required | |
2606 when we know the number of scalar iterations may potentially | |
2607 be smaller than the number of iterations required to enter this loop, for | |
2608 this we use the upper bounds on the prolog and epilog peeling. When we | |
2609 don't know the number of iterations and don't require versioning it is | |
2610 because we have asserted that there are enough scalar iterations to enter | |
2611 the main loop, so this skip is not necessary. When we are versioning then | |
2612 we only add such a skip if we have chosen to vectorize the epilogue. */ | |
131 | 2613 bool skip_vector = (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) |
2614 ? maybe_lt (LOOP_VINFO_INT_NITERS (loop_vinfo), | |
2615 bound_prolog + bound_epilog) | |
145 | 2616 : (!LOOP_REQUIRES_VERSIONING (loop_vinfo) |
2617 || vect_epilogues)); | |
111 | 2618 /* Epilog loop must be executed if the number of iterations for epilog |
2619 loop is known at compile time, otherwise we need to add a check at | |
2620 the end of vector loop and skip to the end of epilog loop. */ | |
2621 bool skip_epilog = (prolog_peeling < 0 | |
131 | 2622 || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) |
2623 || !vf.is_constant ()); | |
111 | 2624 /* PEELING_FOR_GAPS is special because epilog loop must be executed. */ |
2625 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) | |
2626 skip_epilog = false; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2627 |
111 | 2628 if (skip_vector) |
2629 { | |
2630 split_edge (loop_preheader_edge (loop)); | |
2631 | |
2632 /* Due to the order in which we peel prolog and epilog, we first | |
2633 propagate probability to the whole loop. The purpose is to | |
2634 avoid adjusting probabilities of both prolog and vector loops | |
2635 separately. Note in this case, the probability of epilog loop | |
2636 needs to be scaled back later. */ | |
2637 basic_block bb_before_loop = loop_preheader_edge (loop)->src; | |
2638 if (prob_vector.initialized_p ()) | |
131 | 2639 { |
2640 scale_bbs_frequencies (&bb_before_loop, 1, prob_vector); | |
2641 scale_loop_profile (loop, prob_vector, 0); | |
2642 } | |
111 | 2643 } |
2644 | |
131 | 2645 dump_user_location_t loop_loc = find_loop_location (loop); |
145 | 2646 class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); |
2647 if (vect_epilogues) | |
2648 /* Make sure to set the epilogue's epilogue scalar loop, such that we can | |
2649 use the original scalar loop as remaining epilogue if necessary. */ | |
2650 LOOP_VINFO_SCALAR_LOOP (epilogue_vinfo) | |
2651 = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); | |
2652 | |
111 | 2653 if (prolog_peeling) |
2654 { | |
2655 e = loop_preheader_edge (loop); | |
2656 if (!slpeel_can_duplicate_loop_p (loop, e)) | |
2657 { | |
2658 dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2659 "loop can't be duplicated to preheader edge.\n"); | |
2660 gcc_unreachable (); | |
2661 } | |
2662 /* Peel prolog and put it on preheader edge of loop. */ | |
2663 prolog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, scalar_loop, e); | |
2664 if (!prolog) | |
2665 { | |
2666 dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2667 "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n"); | |
2668 gcc_unreachable (); | |
2669 } | |
145 | 2670 prolog->force_vectorize = false; |
111 | 2671 slpeel_update_phi_nodes_for_loops (loop_vinfo, prolog, loop, true); |
2672 first_loop = prolog; | |
2673 reset_original_copy_tables (); | |
2674 | |
131 | 2675 /* Update the number of iterations for prolog loop. */ |
2676 tree step_prolog = build_one_cst (TREE_TYPE (niters_prolog)); | |
2677 vect_set_loop_condition (prolog, NULL, niters_prolog, | |
2678 step_prolog, NULL_TREE, false); | |
111 | 2679 |
2680 /* Skip the prolog loop. */ | |
2681 if (skip_prolog) | |
2682 { | |
2683 guard_cond = fold_build2 (EQ_EXPR, boolean_type_node, | |
2684 niters_prolog, build_int_cst (type, 0)); | |
2685 guard_bb = loop_preheader_edge (prolog)->src; | |
2686 basic_block bb_after_prolog = loop_preheader_edge (loop)->src; | |
2687 guard_to = split_edge (loop_preheader_edge (loop)); | |
2688 guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, | |
2689 guard_to, guard_bb, | |
2690 prob_prolog.invert (), | |
2691 irred_flag); | |
2692 e = EDGE_PRED (guard_to, 0); | |
2693 e = (e != guard_e ? e : EDGE_PRED (guard_to, 1)); | |
2694 slpeel_update_phi_nodes_for_guard1 (prolog, loop, guard_e, e); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2695 |
111 | 2696 scale_bbs_frequencies (&bb_after_prolog, 1, prob_prolog); |
2697 scale_loop_profile (prolog, prob_prolog, bound_prolog); | |
2698 } | |
145 | 2699 |
111 | 2700 /* Update init address of DRs. */ |
131 | 2701 vect_update_inits_of_drs (loop_vinfo, niters_prolog, PLUS_EXPR); |
111 | 2702 /* Update niters for vector loop. */ |
2703 LOOP_VINFO_NITERS (loop_vinfo) | |
2704 = fold_build2 (MINUS_EXPR, type, niters, niters_prolog); | |
2705 LOOP_VINFO_NITERSM1 (loop_vinfo) | |
2706 = fold_build2 (MINUS_EXPR, type, | |
2707 LOOP_VINFO_NITERSM1 (loop_vinfo), niters_prolog); | |
2708 bool new_var_p = false; | |
2709 niters = vect_build_loop_niters (loop_vinfo, &new_var_p); | |
2710 /* It's guaranteed that vector loop bound before vectorization is at | |
2711 least VF, so set range information for newly generated var. */ | |
2712 if (new_var_p) | |
2713 set_range_info (niters, VR_RANGE, | |
2714 wi::to_wide (build_int_cst (type, vf)), | |
2715 wi::to_wide (TYPE_MAX_VALUE (type))); | |
2716 | |
2717 /* Prolog iterates at most bound_prolog times, latch iterates at | |
2718 most bound_prolog - 1 times. */ | |
2719 record_niter_bound (prolog, bound_prolog - 1, false, true); | |
2720 delete_update_ssa (); | |
2721 adjust_vec_debug_stmts (); | |
2722 scev_reset (); | |
2723 } | |
2724 | |
2725 if (epilog_peeling) | |
2726 { | |
2727 e = single_exit (loop); | |
2728 if (!slpeel_can_duplicate_loop_p (loop, e)) | |
2729 { | |
2730 dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2731 "loop can't be duplicated to exit edge.\n"); | |
2732 gcc_unreachable (); | |
2733 } | |
145 | 2734 /* Peel epilog and put it on exit edge of loop. If we are vectorizing |
2735 said epilog then we should use a copy of the main loop as a starting | |
2736 point. This loop may have already had some preliminary transformations | |
2737 to allow for more optimal vectorization, for example if-conversion. | |
2738 If we are not vectorizing the epilog then we should use the scalar loop | |
2739 as the transformations mentioned above make less or no sense when not | |
2740 vectorizing. */ | |
2741 epilog = vect_epilogues ? get_loop_copy (loop) : scalar_loop; | |
2742 if (vop_to_rename) | |
2743 { | |
2744 /* Vectorizing the main loop can sometimes introduce a vdef to | |
2745 a loop that previously didn't have one; see the comment above | |
2746 the definition of VOP_TO_RENAME for details. The definition | |
2747 D that holds on E will then be different from the definition | |
2748 VOP_TO_RENAME that holds during SCALAR_LOOP, so we need to | |
2749 rename VOP_TO_RENAME to D when copying the loop. | |
2750 | |
2751 The virtual operand is in LCSSA form for the main loop, | |
2752 and no stmt between the main loop and E needs a vdef, | |
2753 so we know that D is provided by a phi rather than by a | |
2754 vdef on a normal gimple stmt. */ | |
2755 basic_block vdef_bb = e->src; | |
2756 gphi *vphi; | |
2757 while (!(vphi = get_virtual_phi (vdef_bb))) | |
2758 vdef_bb = get_immediate_dominator (CDI_DOMINATORS, vdef_bb); | |
2759 gcc_assert (vop_to_rename != gimple_phi_result (vphi)); | |
2760 set_current_def (vop_to_rename, gimple_phi_result (vphi)); | |
2761 } | |
2762 epilog = slpeel_tree_duplicate_loop_to_edge_cfg (loop, epilog, e); | |
111 | 2763 if (!epilog) |
2764 { | |
2765 dump_printf_loc (MSG_MISSED_OPTIMIZATION, loop_loc, | |
2766 "slpeel_tree_duplicate_loop_to_edge_cfg failed.\n"); | |
2767 gcc_unreachable (); | |
2768 } | |
145 | 2769 epilog->force_vectorize = false; |
111 | 2770 slpeel_update_phi_nodes_for_loops (loop_vinfo, loop, epilog, false); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2771 |
111 | 2772 /* Scalar version loop may be preferred. In this case, add guard |
2773 and skip to epilog. Note this only happens when the number of | |
2774 iterations of loop is unknown at compile time, otherwise this | |
2775 won't be vectorized. */ | |
2776 if (skip_vector) | |
2777 { | |
2778 /* Additional epilogue iteration is peeled if gap exists. */ | |
2779 tree t = vect_gen_scalar_loop_niters (niters_prolog, prolog_peeling, | |
131 | 2780 bound_prolog, bound_epilog, |
111 | 2781 th, &bound_scalar, |
2782 check_profitability); | |
2783 /* Build guard against NITERSM1 since NITERS may overflow. */ | |
2784 guard_cond = fold_build2 (LT_EXPR, boolean_type_node, nitersm1, t); | |
2785 guard_bb = anchor; | |
2786 guard_to = split_edge (loop_preheader_edge (epilog)); | |
2787 guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, | |
2788 guard_to, guard_bb, | |
2789 prob_vector.invert (), | |
2790 irred_flag); | |
145 | 2791 skip_e = guard_e; |
111 | 2792 e = EDGE_PRED (guard_to, 0); |
2793 e = (e != guard_e ? e : EDGE_PRED (guard_to, 1)); | |
2794 slpeel_update_phi_nodes_for_guard1 (first_loop, epilog, guard_e, e); | |
2795 | |
2796 /* Simply propagate profile info from guard_bb to guard_to which is | |
2797 a merge point of control flow. */ | |
2798 guard_to->count = guard_bb->count; | |
131 | 2799 |
111 | 2800 /* Scale probability of epilog loop back. |
2801 FIXME: We should avoid scaling down and back up. Profile may | |
2802 get lost if we scale down to 0. */ | |
2803 basic_block *bbs = get_loop_body (epilog); | |
131 | 2804 for (unsigned int i = 0; i < epilog->num_nodes; i++) |
2805 bbs[i]->count = bbs[i]->count.apply_scale | |
2806 (bbs[i]->count, | |
2807 bbs[i]->count.apply_probability | |
2808 (prob_vector)); | |
111 | 2809 free (bbs); |
2810 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2811 |
111 | 2812 basic_block bb_before_epilog = loop_preheader_edge (epilog)->src; |
2813 /* If loop is peeled for non-zero constant times, now niters refers to | |
2814 orig_niters - prolog_peeling, it won't overflow even the orig_niters | |
2815 overflows. */ | |
2816 niters_no_overflow |= (prolog_peeling > 0); | |
2817 vect_gen_vector_loop_niters (loop_vinfo, niters, | |
131 | 2818 niters_vector, step_vector, |
2819 niters_no_overflow); | |
2820 if (!integer_onep (*step_vector)) | |
2821 { | |
2822 /* On exit from the loop we will have an easy way of calcalating | |
2823 NITERS_VECTOR / STEP * STEP. Install a dummy definition | |
2824 until then. */ | |
2825 niters_vector_mult_vf = make_ssa_name (TREE_TYPE (*niters_vector)); | |
2826 SSA_NAME_DEF_STMT (niters_vector_mult_vf) = gimple_build_nop (); | |
2827 *niters_vector_mult_vf_var = niters_vector_mult_vf; | |
2828 } | |
2829 else | |
2830 vect_gen_vector_loop_niters_mult_vf (loop_vinfo, *niters_vector, | |
2831 &niters_vector_mult_vf); | |
111 | 2832 /* Update IVs of original loop as if they were advanced by |
2833 niters_vector_mult_vf steps. */ | |
2834 gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo)); | |
145 | 2835 update_e = skip_vector ? e : loop_preheader_edge (epilog); |
111 | 2836 vect_update_ivs_after_vectorizer (loop_vinfo, niters_vector_mult_vf, |
2837 update_e); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2838 |
111 | 2839 if (skip_epilog) |
2840 { | |
2841 guard_cond = fold_build2 (EQ_EXPR, boolean_type_node, | |
2842 niters, niters_vector_mult_vf); | |
2843 guard_bb = single_exit (loop)->dest; | |
2844 guard_to = split_edge (single_exit (epilog)); | |
2845 guard_e = slpeel_add_loop_guard (guard_bb, guard_cond, guard_to, | |
2846 skip_vector ? anchor : guard_bb, | |
2847 prob_epilog.invert (), | |
2848 irred_flag); | |
2849 slpeel_update_phi_nodes_for_guard2 (loop, epilog, guard_e, | |
2850 single_exit (epilog)); | |
2851 /* Only need to handle basic block before epilog loop if it's not | |
2852 the guard_bb, which is the case when skip_vector is true. */ | |
2853 if (guard_bb != bb_before_epilog) | |
2854 { | |
2855 prob_epilog = prob_vector * prob_epilog + prob_vector.invert (); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2856 |
111 | 2857 scale_bbs_frequencies (&bb_before_epilog, 1, prob_epilog); |
2858 } | |
131 | 2859 scale_loop_profile (epilog, prob_epilog, 0); |
111 | 2860 } |
2861 else | |
2862 slpeel_update_phi_nodes_for_lcssa (epilog); | |
2863 | |
131 | 2864 unsigned HOST_WIDE_INT bound; |
2865 if (bound_scalar.is_constant (&bound)) | |
2866 { | |
2867 gcc_assert (bound != 0); | |
2868 /* -1 to convert loop iterations to latch iterations. */ | |
2869 record_niter_bound (epilog, bound - 1, false, true); | |
2870 } | |
111 | 2871 |
2872 delete_update_ssa (); | |
2873 adjust_vec_debug_stmts (); | |
2874 scev_reset (); | |
2875 } | |
145 | 2876 |
2877 if (vect_epilogues) | |
2878 { | |
2879 epilog->aux = epilogue_vinfo; | |
2880 LOOP_VINFO_LOOP (epilogue_vinfo) = epilog; | |
2881 | |
2882 loop_constraint_clear (epilog, LOOP_C_INFINITE); | |
2883 | |
2884 /* We now must calculate the number of NITERS performed by the previous | |
2885 loop and EPILOGUE_NITERS to be performed by the epilogue. */ | |
2886 tree niters = fold_build2 (PLUS_EXPR, TREE_TYPE (niters_vector_mult_vf), | |
2887 niters_prolog, niters_vector_mult_vf); | |
2888 | |
2889 /* If skip_vector we may skip the previous loop, we insert a phi-node to | |
2890 determine whether we are coming from the previous vectorized loop | |
2891 using the update_e edge or the skip_vector basic block using the | |
2892 skip_e edge. */ | |
2893 if (skip_vector) | |
2894 { | |
2895 gcc_assert (update_e != NULL && skip_e != NULL); | |
2896 gphi *new_phi = create_phi_node (make_ssa_name (TREE_TYPE (niters)), | |
2897 update_e->dest); | |
2898 tree new_ssa = make_ssa_name (TREE_TYPE (niters)); | |
2899 gimple *stmt = gimple_build_assign (new_ssa, niters); | |
2900 gimple_stmt_iterator gsi; | |
2901 if (TREE_CODE (niters_vector_mult_vf) == SSA_NAME | |
2902 && SSA_NAME_DEF_STMT (niters_vector_mult_vf)->bb != NULL) | |
2903 { | |
2904 gsi = gsi_for_stmt (SSA_NAME_DEF_STMT (niters_vector_mult_vf)); | |
2905 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
2906 } | |
2907 else | |
2908 { | |
2909 gsi = gsi_last_bb (update_e->src); | |
2910 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2911 } | |
2912 | |
2913 niters = new_ssa; | |
2914 add_phi_arg (new_phi, niters, update_e, UNKNOWN_LOCATION); | |
2915 add_phi_arg (new_phi, build_zero_cst (TREE_TYPE (niters)), skip_e, | |
2916 UNKNOWN_LOCATION); | |
2917 niters = PHI_RESULT (new_phi); | |
2918 } | |
2919 | |
2920 /* Subtract the number of iterations performed by the vectorized loop | |
2921 from the number of total iterations. */ | |
2922 tree epilogue_niters = fold_build2 (MINUS_EXPR, TREE_TYPE (niters), | |
2923 before_loop_niters, | |
2924 niters); | |
2925 | |
2926 LOOP_VINFO_NITERS (epilogue_vinfo) = epilogue_niters; | |
2927 LOOP_VINFO_NITERSM1 (epilogue_vinfo) | |
2928 = fold_build2 (MINUS_EXPR, TREE_TYPE (epilogue_niters), | |
2929 epilogue_niters, | |
2930 build_one_cst (TREE_TYPE (epilogue_niters))); | |
2931 | |
2932 /* Set ADVANCE to the number of iterations performed by the previous | |
2933 loop and its prologue. */ | |
2934 *advance = niters; | |
2935 | |
2936 /* Redo the peeling for niter analysis as the NITERs and alignment | |
2937 may have been updated to take the main loop into account. */ | |
2938 determine_peel_for_niter (epilogue_vinfo); | |
2939 } | |
2940 | |
111 | 2941 adjust_vec.release (); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2942 free_original_copy_tables (); |
111 | 2943 |
145 | 2944 return vect_epilogues ? epilog : NULL; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2945 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2946 |
111 | 2947 /* Function vect_create_cond_for_niters_checks. |
2948 | |
2949 Create a conditional expression that represents the run-time checks for | |
131 | 2950 loop's niter. The loop is guaranteed to terminate if the run-time |
111 | 2951 checks hold. |
2952 | |
2953 Input: | |
2954 COND_EXPR - input conditional expression. New conditions will be chained | |
2955 with logical AND operation. If it is NULL, then the function | |
2956 is used to return the number of alias checks. | |
2957 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs | |
2958 to be checked. | |
2959 | |
2960 Output: | |
2961 COND_EXPR - conditional expression. | |
2962 | |
2963 The returned COND_EXPR is the conditional expression to be used in the | |
2964 if statement that controls which version of the loop gets executed at | |
2965 runtime. */ | |
2966 | |
2967 static void | |
2968 vect_create_cond_for_niters_checks (loop_vec_info loop_vinfo, tree *cond_expr) | |
2969 { | |
2970 tree part_cond_expr = LOOP_VINFO_NITERS_ASSUMPTIONS (loop_vinfo); | |
2971 | |
2972 if (*cond_expr) | |
2973 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2974 *cond_expr, part_cond_expr); | |
2975 else | |
2976 *cond_expr = part_cond_expr; | |
2977 } | |
2978 | |
2979 /* Set *COND_EXPR to a tree that is true when both the original *COND_EXPR | |
2980 and PART_COND_EXPR are true. Treat a null *COND_EXPR as "true". */ | |
2981 | |
2982 static void | |
2983 chain_cond_expr (tree *cond_expr, tree part_cond_expr) | |
2984 { | |
2985 if (*cond_expr) | |
2986 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2987 *cond_expr, part_cond_expr); | |
2988 else | |
2989 *cond_expr = part_cond_expr; | |
2990 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2991 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2992 /* Function vect_create_cond_for_align_checks. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2993 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2994 Create a conditional expression that represents the alignment checks for |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2995 all of data references (array element references) whose alignment must be |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2996 checked at runtime. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2997 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2998 Input: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
2999 COND_EXPR - input conditional expression. New conditions will be chained |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3000 with logical AND operation. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3001 LOOP_VINFO - two fields of the loop information are used. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3002 LOOP_VINFO_PTR_MASK is the mask used to check the alignment. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3003 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3004 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3005 Output: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3006 COND_EXPR_STMT_LIST - statements needed to construct the conditional |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3007 expression. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3008 The returned value is the conditional expression to be used in the if |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3009 statement that controls which version of the loop gets executed at runtime. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3010 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3011 The algorithm makes two assumptions: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3012 1) The number of bytes "n" in a vector is a power of 2. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3013 2) An address "a" is aligned if a%n is zero and that this |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3014 test can be done as a&(n-1) == 0. For example, for 16 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3015 byte vectors the test is a&0xf == 0. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3016 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3017 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3018 vect_create_cond_for_align_checks (loop_vec_info loop_vinfo, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3019 tree *cond_expr, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3020 gimple_seq *cond_expr_stmt_list) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3021 { |
131 | 3022 vec<stmt_vec_info> may_misalign_stmts |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3023 = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); |
131 | 3024 stmt_vec_info stmt_info; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3025 int mask = LOOP_VINFO_PTR_MASK (loop_vinfo); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3026 tree mask_cst; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3027 unsigned int i; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3028 tree int_ptrsize_type; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3029 char tmp_name[20]; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3030 tree or_tmp_name = NULL_TREE; |
111 | 3031 tree and_tmp_name; |
3032 gimple *and_stmt; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3033 tree ptrsize_zero; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3034 tree part_cond_expr; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3035 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3036 /* Check that mask is one less than a power of 2, i.e., mask is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3037 all zeros followed by all ones. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3038 gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3039 |
111 | 3040 int_ptrsize_type = signed_type_for (ptr_type_node); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3041 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3042 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3043 of the first vector of the i'th data reference. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3044 |
131 | 3045 FOR_EACH_VEC_ELT (may_misalign_stmts, i, stmt_info) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3046 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3047 gimple_seq new_stmt_list = NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3048 tree addr_base; |
111 | 3049 tree addr_tmp_name; |
3050 tree new_or_tmp_name; | |
3051 gimple *addr_stmt, *or_stmt; | |
131 | 3052 tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3053 bool negative = tree_int_cst_compare |
131 | 3054 (DR_STEP (STMT_VINFO_DATA_REF (stmt_info)), size_zero_node) < 0; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3055 tree offset = negative |
111 | 3056 ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3057 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3058 /* create: addr_tmp = (int)(address_of_first_vector) */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3059 addr_base = |
131 | 3060 vect_create_addr_base_for_vector_ref (stmt_info, &new_stmt_list, |
111 | 3061 offset); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3062 if (new_stmt_list != NULL) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3063 gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3064 |
111 | 3065 sprintf (tmp_name, "addr2int%d", i); |
3066 addr_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
3067 addr_stmt = gimple_build_assign (addr_tmp_name, NOP_EXPR, addr_base); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3068 gimple_seq_add_stmt (cond_expr_stmt_list, addr_stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3069 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3070 /* The addresses are OR together. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3071 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3072 if (or_tmp_name != NULL_TREE) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3073 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3074 /* create: or_tmp = or_tmp | addr_tmp */ |
111 | 3075 sprintf (tmp_name, "orptrs%d", i); |
3076 new_or_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, tmp_name); | |
3077 or_stmt = gimple_build_assign (new_or_tmp_name, BIT_IOR_EXPR, | |
3078 or_tmp_name, addr_tmp_name); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3079 gimple_seq_add_stmt (cond_expr_stmt_list, or_stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3080 or_tmp_name = new_or_tmp_name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3081 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3082 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3083 or_tmp_name = addr_tmp_name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3084 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3085 } /* end for i */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3086 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3087 mask_cst = build_int_cst (int_ptrsize_type, mask); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3088 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3089 /* create: and_tmp = or_tmp & mask */ |
111 | 3090 and_tmp_name = make_temp_ssa_name (int_ptrsize_type, NULL, "andmask"); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3091 |
111 | 3092 and_stmt = gimple_build_assign (and_tmp_name, BIT_AND_EXPR, |
3093 or_tmp_name, mask_cst); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3094 gimple_seq_add_stmt (cond_expr_stmt_list, and_stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3095 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3096 /* Make and_tmp the left operand of the conditional test against zero. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3097 if and_tmp has a nonzero bit then some address is unaligned. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3098 ptrsize_zero = build_int_cst (int_ptrsize_type, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3099 part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3100 and_tmp_name, ptrsize_zero); |
111 | 3101 chain_cond_expr (cond_expr, part_cond_expr); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3102 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3103 |
111 | 3104 /* If LOOP_VINFO_CHECK_UNEQUAL_ADDRS contains <A1, B1>, ..., <An, Bn>, |
3105 create a tree representation of: (&A1 != &B1) && ... && (&An != &Bn). | |
3106 Set *COND_EXPR to a tree that is true when both the original *COND_EXPR | |
3107 and this new condition are true. Treat a null *COND_EXPR as "true". */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3108 |
111 | 3109 static void |
3110 vect_create_cond_for_unequal_addrs (loop_vec_info loop_vinfo, tree *cond_expr) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3111 { |
111 | 3112 vec<vec_object_pair> pairs = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo); |
3113 unsigned int i; | |
3114 vec_object_pair *pair; | |
3115 FOR_EACH_VEC_ELT (pairs, i, pair) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3116 { |
111 | 3117 tree addr1 = build_fold_addr_expr (pair->first); |
3118 tree addr2 = build_fold_addr_expr (pair->second); | |
3119 tree part_cond_expr = fold_build2 (NE_EXPR, boolean_type_node, | |
3120 addr1, addr2); | |
3121 chain_cond_expr (cond_expr, part_cond_expr); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3122 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3123 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3124 |
131 | 3125 /* Create an expression that is true when all lower-bound conditions for |
3126 the vectorized loop are met. Chain this condition with *COND_EXPR. */ | |
3127 | |
3128 static void | |
3129 vect_create_cond_for_lower_bounds (loop_vec_info loop_vinfo, tree *cond_expr) | |
3130 { | |
3131 vec<vec_lower_bound> lower_bounds = LOOP_VINFO_LOWER_BOUNDS (loop_vinfo); | |
3132 for (unsigned int i = 0; i < lower_bounds.length (); ++i) | |
3133 { | |
3134 tree expr = lower_bounds[i].expr; | |
3135 tree type = unsigned_type_for (TREE_TYPE (expr)); | |
3136 expr = fold_convert (type, expr); | |
3137 poly_uint64 bound = lower_bounds[i].min_value; | |
3138 if (!lower_bounds[i].unsigned_p) | |
3139 { | |
3140 expr = fold_build2 (PLUS_EXPR, type, expr, | |
3141 build_int_cstu (type, bound - 1)); | |
3142 bound += bound - 1; | |
3143 } | |
3144 tree part_cond_expr = fold_build2 (GE_EXPR, boolean_type_node, expr, | |
3145 build_int_cstu (type, bound)); | |
3146 chain_cond_expr (cond_expr, part_cond_expr); | |
3147 } | |
3148 } | |
3149 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3150 /* Function vect_create_cond_for_alias_checks. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3151 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3152 Create a conditional expression that represents the run-time checks for |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3153 overlapping of address ranges represented by a list of data references |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3154 relations passed as input. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3155 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3156 Input: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3157 COND_EXPR - input conditional expression. New conditions will be chained |
111 | 3158 with logical AND operation. If it is NULL, then the function |
3159 is used to return the number of alias checks. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3160 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3161 to be checked. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3162 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3163 Output: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3164 COND_EXPR - conditional expression. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3165 |
111 | 3166 The returned COND_EXPR is the conditional expression to be used in the if |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3167 statement that controls which version of the loop gets executed at runtime. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3168 */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3169 |
111 | 3170 void |
3171 vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo, tree * cond_expr) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3172 { |
111 | 3173 vec<dr_with_seg_len_pair_t> comp_alias_ddrs = |
3174 LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3175 |
111 | 3176 if (comp_alias_ddrs.is_empty ()) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3177 return; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3178 |
111 | 3179 create_runtime_alias_checks (LOOP_VINFO_LOOP (loop_vinfo), |
3180 &comp_alias_ddrs, cond_expr); | |
3181 if (dump_enabled_p ()) | |
3182 dump_printf_loc (MSG_NOTE, vect_location, | |
3183 "created %u versioning for alias checks.\n", | |
3184 comp_alias_ddrs.length ()); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3185 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3186 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3187 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3188 /* Function vect_loop_versioning. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3189 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3190 If the loop has data references that may or may not be aligned or/and |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3191 has data reference relations whose independence was not proven then |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3192 two versions of the loop need to be generated, one which is vectorized |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3193 and one which isn't. A test is then generated to control which of the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3194 loops is executed. The test checks for the alignment of all of the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3195 data references that may or may not be aligned. An additional |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3196 sequence of runtime tests is generated for each pairs of DDRs whose |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3197 independence was not proven. The vectorized version of loop is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3198 executed only if both alias and alignment tests are passed. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3199 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3200 The test generated to check which version of loop is executed |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3201 is modified to also check for profitability as indicated by the |
111 | 3202 cost model threshold TH. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3203 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3204 The versioning precondition(s) are placed in *COND_EXPR and |
111 | 3205 *COND_EXPR_STMT_LIST. */ |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3206 |
145 | 3207 class loop * |
111 | 3208 vect_loop_versioning (loop_vec_info loop_vinfo, |
145 | 3209 gimple *loop_vectorized_call) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3210 { |
145 | 3211 class loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop; |
3212 class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3213 basic_block condition_bb; |
111 | 3214 gphi_iterator gsi; |
3215 gimple_stmt_iterator cond_exp_gsi; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3216 basic_block merge_bb; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3217 basic_block new_exit_bb; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3218 edge new_exit_e, e; |
111 | 3219 gphi *orig_phi, *new_phi; |
3220 tree cond_expr = NULL_TREE; | |
3221 gimple_seq cond_expr_stmt_list = NULL; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3222 tree arg; |
111 | 3223 profile_probability prob = profile_probability::likely (); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3224 gimple_seq gimplify_stmt_list = NULL; |
111 | 3225 tree scalar_loop_iters = LOOP_VINFO_NITERSM1 (loop_vinfo); |
3226 bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo); | |
3227 bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); | |
3228 bool version_niter = LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo); | |
145 | 3229 poly_uint64 versioning_threshold |
3230 = LOOP_VINFO_VERSIONING_THRESHOLD (loop_vinfo); | |
3231 tree version_simd_if_cond | |
3232 = LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (loop_vinfo); | |
3233 unsigned th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); | |
3234 | |
3235 if (vect_apply_runtime_profitability_check_p (loop_vinfo) | |
3236 && !ordered_p (th, versioning_threshold)) | |
111 | 3237 cond_expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters, |
3238 build_int_cst (TREE_TYPE (scalar_loop_iters), | |
3239 th - 1)); | |
131 | 3240 if (maybe_ne (versioning_threshold, 0U)) |
3241 { | |
3242 tree expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters, | |
3243 build_int_cst (TREE_TYPE (scalar_loop_iters), | |
3244 versioning_threshold - 1)); | |
3245 if (cond_expr) | |
3246 cond_expr = fold_build2 (BIT_AND_EXPR, boolean_type_node, | |
3247 expr, cond_expr); | |
3248 else | |
3249 cond_expr = expr; | |
3250 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3251 |
111 | 3252 if (version_niter) |
3253 vect_create_cond_for_niters_checks (loop_vinfo, &cond_expr); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3254 |
111 | 3255 if (cond_expr) |
145 | 3256 cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr), |
3257 &cond_expr_stmt_list, | |
111 | 3258 is_gimple_condexpr, NULL_TREE); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3259 |
111 | 3260 if (version_align) |
3261 vect_create_cond_for_align_checks (loop_vinfo, &cond_expr, | |
3262 &cond_expr_stmt_list); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3263 |
111 | 3264 if (version_alias) |
3265 { | |
3266 vect_create_cond_for_unequal_addrs (loop_vinfo, &cond_expr); | |
131 | 3267 vect_create_cond_for_lower_bounds (loop_vinfo, &cond_expr); |
111 | 3268 vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr); |
3269 } | |
3270 | |
145 | 3271 if (version_simd_if_cond) |
3272 { | |
3273 gcc_assert (dom_info_available_p (CDI_DOMINATORS)); | |
3274 if (flag_checking) | |
3275 if (basic_block bb | |
3276 = gimple_bb (SSA_NAME_DEF_STMT (version_simd_if_cond))) | |
3277 gcc_assert (bb != loop->header | |
3278 && dominated_by_p (CDI_DOMINATORS, loop->header, bb) | |
3279 && (scalar_loop == NULL | |
3280 || (bb != scalar_loop->header | |
3281 && dominated_by_p (CDI_DOMINATORS, | |
3282 scalar_loop->header, bb)))); | |
3283 tree zero = build_zero_cst (TREE_TYPE (version_simd_if_cond)); | |
3284 tree c = fold_build2 (NE_EXPR, boolean_type_node, | |
3285 version_simd_if_cond, zero); | |
3286 if (cond_expr) | |
3287 cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
3288 c, cond_expr); | |
3289 else | |
3290 cond_expr = c; | |
3291 if (dump_enabled_p ()) | |
3292 dump_printf_loc (MSG_NOTE, vect_location, | |
3293 "created versioning for simd if condition check.\n"); | |
3294 } | |
3295 | |
131 | 3296 cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr), |
3297 &gimplify_stmt_list, | |
111 | 3298 is_gimple_condexpr, NULL_TREE); |
3299 gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3300 |
145 | 3301 /* Compute the outermost loop cond_expr and cond_expr_stmt_list are |
3302 invariant in. */ | |
3303 class loop *outermost = outermost_invariant_loop_for_expr (loop, cond_expr); | |
3304 for (gimple_stmt_iterator gsi = gsi_start (cond_expr_stmt_list); | |
3305 !gsi_end_p (gsi); gsi_next (&gsi)) | |
111 | 3306 { |
145 | 3307 gimple *stmt = gsi_stmt (gsi); |
3308 update_stmt (stmt); | |
3309 ssa_op_iter iter; | |
3310 use_operand_p use_p; | |
3311 basic_block def_bb; | |
3312 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE) | |
3313 if ((def_bb = gimple_bb (SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p)))) | |
3314 && flow_bb_inside_loop_p (outermost, def_bb)) | |
3315 outermost = superloop_at_depth (loop, bb_loop_depth (def_bb) + 1); | |
3316 } | |
3317 | |
3318 /* Search for the outermost loop we can version. Avoid versioning of | |
3319 non-perfect nests but allow if-conversion versioned loops inside. */ | |
3320 class loop *loop_to_version = loop; | |
3321 if (flow_loop_nested_p (outermost, loop)) | |
3322 { | |
3323 if (dump_enabled_p ()) | |
3324 dump_printf_loc (MSG_NOTE, vect_location, | |
3325 "trying to apply versioning to outer loop %d\n", | |
3326 outermost->num); | |
3327 if (outermost->num == 0) | |
3328 outermost = superloop_at_depth (loop, 1); | |
3329 /* And avoid applying versioning on non-perfect nests. */ | |
3330 while (loop_to_version != outermost | |
3331 && single_exit (loop_outer (loop_to_version)) | |
3332 && (!loop_outer (loop_to_version)->inner->next | |
3333 || vect_loop_vectorized_call (loop_to_version)) | |
3334 && (!loop_outer (loop_to_version)->inner->next | |
3335 || !loop_outer (loop_to_version)->inner->next->next)) | |
3336 loop_to_version = loop_outer (loop_to_version); | |
3337 } | |
3338 | |
3339 /* Apply versioning. If there is already a scalar version created by | |
3340 if-conversion re-use that. Note we cannot re-use the copy of | |
3341 an if-converted outer-loop when vectorizing the inner loop only. */ | |
3342 gcond *cond; | |
3343 if ((!loop_to_version->inner || loop == loop_to_version) | |
3344 && loop_vectorized_call) | |
3345 { | |
3346 gcc_assert (scalar_loop); | |
3347 condition_bb = gimple_bb (loop_vectorized_call); | |
3348 cond = as_a <gcond *> (last_stmt (condition_bb)); | |
3349 gimple_cond_set_condition_from_tree (cond, cond_expr); | |
3350 update_stmt (cond); | |
3351 | |
3352 if (cond_expr_stmt_list) | |
3353 { | |
3354 cond_exp_gsi = gsi_for_stmt (loop_vectorized_call); | |
3355 gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, | |
3356 GSI_SAME_STMT); | |
3357 } | |
3358 | |
3359 /* if-conversion uses profile_probability::always () for both paths, | |
3360 reset the paths probabilities appropriately. */ | |
3361 edge te, fe; | |
3362 extract_true_false_edges_from_block (condition_bb, &te, &fe); | |
3363 te->probability = prob; | |
3364 fe->probability = prob.invert (); | |
3365 /* We can scale loops counts immediately but have to postpone | |
3366 scaling the scalar loop because we re-use it during peeling. */ | |
3367 scale_loop_frequencies (loop_to_version, te->probability); | |
3368 LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo) = fe->probability; | |
3369 | |
3370 nloop = scalar_loop; | |
3371 if (dump_enabled_p ()) | |
3372 dump_printf_loc (MSG_NOTE, vect_location, | |
3373 "reusing %sloop version created by if conversion\n", | |
3374 loop_to_version != loop ? "outer " : ""); | |
111 | 3375 } |
3376 else | |
145 | 3377 { |
3378 if (loop_to_version != loop | |
3379 && dump_enabled_p ()) | |
3380 dump_printf_loc (MSG_NOTE, vect_location, | |
3381 "applying loop versioning to outer loop %d\n", | |
3382 loop_to_version->num); | |
3383 | |
3384 initialize_original_copy_tables (); | |
3385 nloop = loop_version (loop_to_version, cond_expr, &condition_bb, | |
3386 prob, prob.invert (), prob, prob.invert (), true); | |
3387 gcc_assert (nloop); | |
3388 nloop = get_loop_copy (loop); | |
3389 | |
3390 /* Kill off IFN_LOOP_VECTORIZED_CALL in the copy, nobody will | |
3391 reap those otherwise; they also refer to the original | |
3392 loops. */ | |
3393 class loop *l = loop; | |
3394 while (gimple *call = vect_loop_vectorized_call (l)) | |
3395 { | |
3396 call = SSA_NAME_DEF_STMT (get_current_def (gimple_call_lhs (call))); | |
3397 fold_loop_internal_call (call, boolean_false_node); | |
3398 l = loop_outer (l); | |
3399 } | |
3400 free_original_copy_tables (); | |
3401 | |
3402 if (cond_expr_stmt_list) | |
3403 { | |
3404 cond_exp_gsi = gsi_last_bb (condition_bb); | |
3405 gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list, | |
3406 GSI_SAME_STMT); | |
3407 } | |
3408 | |
3409 /* Loop versioning violates an assumption we try to maintain during | |
3410 vectorization - that the loop exit block has a single predecessor. | |
3411 After versioning, the exit block of both loop versions is the same | |
3412 basic block (i.e. it has two predecessors). Just in order to simplify | |
3413 following transformations in the vectorizer, we fix this situation | |
3414 here by adding a new (empty) block on the exit-edge of the loop, | |
3415 with the proper loop-exit phis to maintain loop-closed-form. | |
3416 If loop versioning wasn't done from loop, but scalar_loop instead, | |
3417 merge_bb will have already just a single successor. */ | |
3418 | |
3419 merge_bb = single_exit (loop_to_version)->dest; | |
3420 if (EDGE_COUNT (merge_bb->preds) >= 2) | |
3421 { | |
3422 gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2); | |
3423 new_exit_bb = split_edge (single_exit (loop_to_version)); | |
3424 new_exit_e = single_exit (loop_to_version); | |
3425 e = EDGE_SUCC (new_exit_bb, 0); | |
3426 | |
3427 for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); | |
3428 gsi_next (&gsi)) | |
3429 { | |
3430 tree new_res; | |
3431 orig_phi = gsi.phi (); | |
3432 new_res = copy_ssa_name (PHI_RESULT (orig_phi)); | |
3433 new_phi = create_phi_node (new_res, new_exit_bb); | |
3434 arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e); | |
3435 add_phi_arg (new_phi, arg, new_exit_e, | |
3436 gimple_phi_arg_location_from_edge (orig_phi, e)); | |
3437 adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi)); | |
3438 } | |
3439 } | |
3440 | |
3441 update_ssa (TODO_update_ssa); | |
3442 } | |
111 | 3443 |
3444 if (version_niter) | |
3445 { | |
3446 /* The versioned loop could be infinite, we need to clear existing | |
3447 niter information which is copied from the original loop. */ | |
3448 gcc_assert (loop_constraint_set_p (loop, LOOP_C_FINITE)); | |
3449 vect_free_loop_info_assumptions (nloop); | |
3450 /* And set constraint LOOP_C_INFINITE for niter analyzer. */ | |
3451 loop_constraint_set (loop, LOOP_C_INFINITE); | |
3452 } | |
3453 | |
131 | 3454 if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION |
111 | 3455 && dump_enabled_p ()) |
3456 { | |
3457 if (version_alias) | |
131 | 3458 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING, |
3459 vect_location, | |
111 | 3460 "loop versioned for vectorization because of " |
3461 "possible aliasing\n"); | |
3462 if (version_align) | |
131 | 3463 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING, |
3464 vect_location, | |
111 | 3465 "loop versioned for vectorization to enhance " |
3466 "alignment\n"); | |
3467 | |
3468 } | |
145 | 3469 |
3470 return nloop; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3471 } |