1 /*
2 * Copyright (c) 2019 Connor Abbott
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, sub license,
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
12 * next paragraph) shall be included in all copies or substantial portions
13 * of the 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 NON-INFRINGEMENT. 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
21 * DEALINGS IN THE SOFTWARE.
22 *
23 */
24
25 #include "gpir.h"
26
27 /* Here we perform a few optimizations that can't currently be done in NIR:
28 *
29 * - Optimize the result of a conditional break/continue. In NIR something
30 * like:
31 *
32 * loop {
33 * ...
34 * if (cond)
35 * continue;
36 *
37 * would get lowered to:
38 *
39 * block_0:
40 * ...
41 * block_1:
42 * branch_cond !cond block_3
43 * block_2:
44 * branch_uncond block_0
45 * block_3:
46 * ...
47 *
48 * We recognize the conditional branch skipping over the unconditional
49 * branch, and turn it into:
50 *
51 * block_0:
52 * ...
53 * block_1:
54 * branch_cond cond block_0
55 * block_2:
56 * block_3:
57 * ...
58 *
59 * - Optimize away nots of comparisons as produced by lowering ifs to
60 * branches, and nots of nots produced by the above optimization.
61 *
62 * - DCE
63 */
64
65 static void
optimize_branches(gpir_compiler * comp)66 optimize_branches(gpir_compiler *comp)
67 {
68 list_for_each_entry(gpir_block, block, &comp->block_list, list) {
69 /* Look for a block with a single unconditional branch. */
70 if (!list_is_singular(&block->node_list))
71 continue;
72
73 gpir_node *node = list_first_entry(&block->node_list, gpir_node, list);
74 if (node->op != gpir_op_branch_uncond)
75 continue;
76
77 gpir_block *target = gpir_node_to_branch(node)->dest;
78
79 /* Find the previous block */
80 if (block->list.prev == &comp->block_list)
81 continue;
82
83 gpir_block *prev_block = LIST_ENTRY(gpir_block, block->list.prev, list);
84 if (list_is_empty(&prev_block->node_list))
85 continue;
86
87 /* Previous block must end with a conditional branch */
88 gpir_node *prev_block_last =
89 list_last_entry(&prev_block->node_list, gpir_node, list);
90 if (prev_block_last->op != gpir_op_branch_cond)
91 continue;
92
93 /* That branch must branch to the block after this */
94 gpir_branch_node *prev_branch = gpir_node_to_branch(prev_block_last);
95 gpir_block *prev_target = prev_branch->dest;
96
97 if (&prev_target->list != block->list.next)
98 continue;
99
100 /* Hooray! Invert the conditional branch and change the target */
101 gpir_alu_node *cond = gpir_node_create(prev_block, gpir_op_not);
102 cond->children[0] = prev_branch->cond;
103 cond->num_child = 1;
104 gpir_node_add_dep(&cond->node, cond->children[0], GPIR_DEP_INPUT);
105 list_addtail(&cond->node.list, &prev_block_last->list);
106 gpir_node_insert_child(prev_block_last, prev_branch->cond, &cond->node);
107 prev_branch->dest = target;
108 prev_block->successors[1] = target;
109
110 /* Delete the branch */
111 list_del(&node->list);
112 block->successors[0] = LIST_ENTRY(gpir_block, block->list.next, list);
113 }
114 }
115
116 static void
optimize_not(gpir_compiler * comp)117 optimize_not(gpir_compiler *comp)
118 {
119 list_for_each_entry(gpir_block, block, &comp->block_list, list) {
120 list_for_each_entry_rev(gpir_node, node, &block->node_list, list) {
121 if (node->op != gpir_op_not)
122 continue;
123
124 gpir_alu_node *alu = gpir_node_to_alu(node);
125
126 gpir_node *replace = NULL;
127 if (alu->children[0]->op == gpir_op_not) {
128 /* (not (not a)) -> a */
129 gpir_alu_node *child = gpir_node_to_alu(alu->children[0]);
130 replace = child->children[0];
131 } else if (alu->children[0]->op == gpir_op_ge ||
132 alu->children[0]->op == gpir_op_lt) {
133 /* (not (ge a, b)) -> (lt a, b) and
134 * (not (lt a, b)) -> (ge a, b)
135 */
136 gpir_alu_node *child = gpir_node_to_alu(alu->children[0]);
137 gpir_op op = alu->children[0]->op == gpir_op_ge ?
138 gpir_op_lt : gpir_op_ge;
139 gpir_alu_node *new = gpir_node_create(block, op);
140 new->children[0] = child->children[0];
141 new->children[1] = child->children[1];
142 new->num_child = 2;
143 gpir_node_add_dep(&new->node, new->children[0], GPIR_DEP_INPUT);
144 gpir_node_add_dep(&new->node, new->children[1], GPIR_DEP_INPUT);
145 list_addtail(&new->node.list, &alu->children[0]->list);
146 replace = &new->node;
147 }
148
149 if (replace) {
150 gpir_node_replace_succ(replace, node);
151 }
152 }
153 }
154 }
155
156 /* Does what it says. In addition to removing unused nodes from the not
157 * optimization above, we need this to remove unused load_const nodes which
158 * were created from store_output intrinsics in NIR, since we ignore the
159 * offset.
160 */
161
162 static void
dead_code_eliminate(gpir_compiler * comp)163 dead_code_eliminate(gpir_compiler *comp)
164 {
165 list_for_each_entry(gpir_block, block, &comp->block_list, list) {
166 list_for_each_entry_safe_rev(gpir_node, node, &block->node_list, list) {
167 if (node->type != gpir_node_type_store &&
168 node->type != gpir_node_type_branch &&
169 list_is_empty(&node->succ_list)) {
170 gpir_node_delete(node);
171 }
172 }
173 }
174
175 /* Kill all the writes to regs that are never read. All the known
176 * instances of these are coming from the cycle-breaking register
177 * created in out-of-SSA. See resolve_parallel_copy() in nir_from_ssa.c
178 * Since we kill redundant movs when we translate nir into gpir, it
179 * results in this reg being written, but never read.
180 */
181 BITSET_WORD *regs = rzalloc_array(comp, BITSET_WORD, comp->cur_reg);
182 list_for_each_entry(gpir_block, block, &comp->block_list, list) {
183 list_for_each_entry(gpir_node, node, &block->node_list, list) {
184 if (node->op != gpir_op_load_reg)
185 continue;
186 gpir_load_node *load = gpir_node_to_load(node);
187 BITSET_SET(regs, load->reg->index);
188 }
189 }
190
191 list_for_each_entry(gpir_block, block, &comp->block_list, list) {
192 list_for_each_entry_safe(gpir_node, node, &block->node_list, list) {
193 if (node->op != gpir_op_store_reg)
194 continue;
195 gpir_store_node *store = gpir_node_to_store(node);
196 if (!BITSET_TEST(regs, store->reg->index))
197 gpir_node_delete(node);
198 }
199 }
200
201 ralloc_free(regs);
202 }
203
204 bool
gpir_optimize(gpir_compiler * comp)205 gpir_optimize(gpir_compiler *comp)
206 {
207 optimize_branches(comp);
208 optimize_not(comp);
209 dead_code_eliminate(comp);
210
211 gpir_debug("after optimization\n");
212 gpir_node_print_prog_seq(comp);
213
214 return true;
215 }
216
217