1 /*
2 * Copyright © 2013 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include "brw_fs.h"
25 #include "brw_cfg.h"
26
27 /** @file brw_fs_sel_peephole.cpp
28 *
29 * This file contains the opt_peephole_sel() optimization pass that replaces
30 * MOV instructions to the same destination in the "then" and "else" bodies of
31 * an if statement with SEL instructions.
32 */
33
34 /* Four MOVs seems to be pretty typical, so I picked the next power of two in
35 * the hopes that it would handle almost anything possible in a single
36 * pass.
37 */
38 #define MAX_MOVS 8 /**< The maximum number of MOVs to attempt to match. */
39
40 using namespace brw;
41
42 /**
43 * Scans forwards from an IF counting consecutive MOV instructions in the
44 * "then" and "else" blocks of the if statement.
45 *
46 * A pointer to the bblock_t following the IF is passed as the <then_block>
47 * argument. The function stores pointers to the MOV instructions in the
48 * <then_mov> and <else_mov> arrays.
49 *
50 * \return the minimum number of MOVs found in the two branches or zero if
51 * an error occurred.
52 *
53 * E.g.:
54 * IF ...
55 * then_mov[0] = MOV g4, ...
56 * then_mov[1] = MOV g5, ...
57 * then_mov[2] = MOV g6, ...
58 * ELSE ...
59 * else_mov[0] = MOV g4, ...
60 * else_mov[1] = MOV g5, ...
61 * else_mov[2] = MOV g7, ...
62 * ENDIF
63 * returns 3.
64 */
65 static int
count_movs_from_if(const intel_device_info * devinfo,fs_inst * then_mov[MAX_MOVS],fs_inst * else_mov[MAX_MOVS],bblock_t * then_block,bblock_t * else_block)66 count_movs_from_if(const intel_device_info *devinfo,
67 fs_inst *then_mov[MAX_MOVS], fs_inst *else_mov[MAX_MOVS],
68 bblock_t *then_block, bblock_t *else_block)
69 {
70 int then_movs = 0;
71 foreach_inst_in_block(fs_inst, inst, then_block) {
72 if (then_movs == MAX_MOVS || inst->opcode != BRW_OPCODE_MOV ||
73 inst->flags_written(devinfo))
74 break;
75
76 then_mov[then_movs] = inst;
77 then_movs++;
78 }
79
80 int else_movs = 0;
81 foreach_inst_in_block(fs_inst, inst, else_block) {
82 if (else_movs == MAX_MOVS || inst->opcode != BRW_OPCODE_MOV ||
83 inst->flags_written(devinfo))
84 break;
85
86 else_mov[else_movs] = inst;
87 else_movs++;
88 }
89
90 return MIN2(then_movs, else_movs);
91 }
92
93 /**
94 * Try to replace IF/MOV+/ELSE/MOV+/ENDIF with SEL.
95 *
96 * Many GLSL shaders contain the following pattern:
97 *
98 * x = condition ? foo : bar
99 *
100 * or
101 *
102 * if (...) a.xyzw = foo.xyzw;
103 * else a.xyzw = bar.xyzw;
104 *
105 * The compiler emits an ir_if tree for this, since each subexpression might be
106 * a complex tree that could have side-effects or short-circuit logic.
107 *
108 * However, the common case is to simply select one of two constants or
109 * variable values---which is exactly what SEL is for. In this case, the
110 * assembly looks like:
111 *
112 * (+f0) IF
113 * MOV dst src0
114 * ...
115 * ELSE
116 * MOV dst src1
117 * ...
118 * ENDIF
119 *
120 * where each pair of MOVs to a common destination and can be easily translated
121 * into
122 *
123 * (+f0) SEL dst src0 src1
124 *
125 * If src0 is an immediate value, we promote it to a temporary GRF.
126 */
127 bool
opt_peephole_sel()128 fs_visitor::opt_peephole_sel()
129 {
130 bool progress = false;
131
132 foreach_block (block, cfg) {
133 /* IF instructions, by definition, can only be found at the ends of
134 * basic blocks.
135 */
136 fs_inst *if_inst = (fs_inst *)block->end();
137 if (if_inst->opcode != BRW_OPCODE_IF)
138 continue;
139
140 fs_inst *else_mov[MAX_MOVS] = { NULL };
141 fs_inst *then_mov[MAX_MOVS] = { NULL };
142
143 bblock_t *then_block = block->next();
144 bblock_t *else_block = NULL;
145 foreach_list_typed(bblock_link, child, link, &block->children) {
146 if (child->block != then_block) {
147 if (child->block->prev()->end()->opcode == BRW_OPCODE_ELSE) {
148 else_block = child->block;
149 }
150 break;
151 }
152 }
153 if (else_block == NULL)
154 continue;
155
156 int movs = count_movs_from_if(devinfo, then_mov, else_mov, then_block, else_block);
157
158 if (movs == 0)
159 continue;
160
161 /* Generate SEL instructions for pairs of MOVs to a common destination. */
162 for (int i = 0; i < movs; i++) {
163 if (!then_mov[i] || !else_mov[i])
164 break;
165
166 /* Check that the MOVs are the right form. */
167 if (!then_mov[i]->dst.equals(else_mov[i]->dst) ||
168 then_mov[i]->exec_size != else_mov[i]->exec_size ||
169 then_mov[i]->group != else_mov[i]->group ||
170 then_mov[i]->force_writemask_all != else_mov[i]->force_writemask_all ||
171 then_mov[i]->is_partial_write() ||
172 else_mov[i]->is_partial_write() ||
173 then_mov[i]->conditional_mod != BRW_CONDITIONAL_NONE ||
174 else_mov[i]->conditional_mod != BRW_CONDITIONAL_NONE) {
175 movs = i;
176 break;
177 }
178
179 /* Check that source types for mov operations match. */
180 if (then_mov[i]->src[0].type != else_mov[i]->src[0].type) {
181 movs = i;
182 break;
183 }
184 }
185
186 if (movs == 0)
187 continue;
188
189 for (int i = 0; i < movs; i++) {
190 const fs_builder ibld = fs_builder(this, then_block, then_mov[i])
191 .at(block, if_inst);
192
193 if (then_mov[i]->src[0].equals(else_mov[i]->src[0])) {
194 ibld.MOV(then_mov[i]->dst, then_mov[i]->src[0]);
195 } else {
196 /* Only the last source register can be a constant, so if the MOV
197 * in the "then" clause uses a constant, we need to put it in a
198 * temporary.
199 */
200 fs_reg src0(then_mov[i]->src[0]);
201 if (src0.file == IMM) {
202 src0 = ibld.vgrf(then_mov[i]->src[0].type);
203 ibld.MOV(src0, then_mov[i]->src[0]);
204 }
205
206 /* 64-bit immediates can't be placed in src1. */
207 fs_reg src1(else_mov[i]->src[0]);
208 if (src1.file == IMM && type_sz(src1.type) == 8) {
209 src1 = ibld.vgrf(else_mov[i]->src[0].type);
210 ibld.MOV(src1, else_mov[i]->src[0]);
211 }
212
213 set_predicate_inv(if_inst->predicate, if_inst->predicate_inverse,
214 ibld.SEL(then_mov[i]->dst, src0, src1));
215 }
216
217 then_mov[i]->remove(then_block);
218 else_mov[i]->remove(else_block);
219 }
220
221 progress = true;
222 }
223
224 if (progress)
225 invalidate_analysis(DEPENDENCY_INSTRUCTIONS | DEPENDENCY_VARIABLES);
226
227 return progress;
228 }
229