nir: Add a pass to combine store_derefs to same vector
[mesa.git] / src / compiler / nir / nir_liveness.c
1 /*
2 * Copyright © 2014 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 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 */
26
27 #include "nir.h"
28 #include "nir_worklist.h"
29 #include "nir_vla.h"
30
31 /*
32 * Basic liveness analysis. This works only in SSA form.
33 *
34 * This liveness pass treats phi nodes as being melded to the space between
35 * blocks so that the destinations of a phi are in the livein of the block
36 * in which it resides and the sources are in the liveout of the
37 * corresponding block. By formulating the liveness information in this
38 * way, we ensure that the definition of any variable dominates its entire
39 * live range. This is true because the only way that the definition of an
40 * SSA value may not dominate a use is if the use is in a phi node and the
41 * uses in phi no are in the live-out of the corresponding predecessor
42 * block but not in the live-in of the block containing the phi node.
43 */
44
45 struct live_ssa_defs_state {
46 unsigned num_ssa_defs;
47 unsigned bitset_words;
48
49 nir_block_worklist worklist;
50 };
51
52 static bool
53 index_ssa_def(nir_ssa_def *def, void *void_state)
54 {
55 struct live_ssa_defs_state *state = void_state;
56
57 if (def->parent_instr->type == nir_instr_type_ssa_undef)
58 def->live_index = 0;
59 else
60 def->live_index = state->num_ssa_defs++;
61
62 return true;
63 }
64
65 /* Initialize the liveness data to zero and add the given block to the
66 * worklist.
67 */
68 static bool
69 init_liveness_block(nir_block *block,
70 struct live_ssa_defs_state *state)
71 {
72 block->live_in = reralloc(block, block->live_in, BITSET_WORD,
73 state->bitset_words);
74 memset(block->live_in, 0, state->bitset_words * sizeof(BITSET_WORD));
75
76 block->live_out = reralloc(block, block->live_out, BITSET_WORD,
77 state->bitset_words);
78 memset(block->live_out, 0, state->bitset_words * sizeof(BITSET_WORD));
79
80 nir_block_worklist_push_head(&state->worklist, block);
81
82 return true;
83 }
84
85 static bool
86 set_src_live(nir_src *src, void *void_live)
87 {
88 BITSET_WORD *live = void_live;
89
90 if (!src->is_ssa)
91 return true;
92
93 if (src->ssa->live_index == 0)
94 return true; /* undefined variables are never live */
95
96 BITSET_SET(live, src->ssa->live_index);
97
98 return true;
99 }
100
101 static bool
102 set_ssa_def_dead(nir_ssa_def *def, void *void_live)
103 {
104 BITSET_WORD *live = void_live;
105
106 BITSET_CLEAR(live, def->live_index);
107
108 return true;
109 }
110
111 /** Propagates the live in of succ across the edge to the live out of pred
112 *
113 * Phi nodes exist "between" blocks and all the phi nodes at the start of a
114 * block act "in parallel". When we propagate from the live_in of one
115 * block to the live out of the other, we have to kill any writes from phis
116 * and make live any sources.
117 *
118 * Returns true if updating live out of pred added anything
119 */
120 static bool
121 propagate_across_edge(nir_block *pred, nir_block *succ,
122 struct live_ssa_defs_state *state)
123 {
124 NIR_VLA(BITSET_WORD, live, state->bitset_words);
125 memcpy(live, succ->live_in, state->bitset_words * sizeof *live);
126
127 nir_foreach_instr(instr, succ) {
128 if (instr->type != nir_instr_type_phi)
129 break;
130 nir_phi_instr *phi = nir_instr_as_phi(instr);
131
132 assert(phi->dest.is_ssa);
133 set_ssa_def_dead(&phi->dest.ssa, live);
134 }
135
136 nir_foreach_instr(instr, succ) {
137 if (instr->type != nir_instr_type_phi)
138 break;
139 nir_phi_instr *phi = nir_instr_as_phi(instr);
140
141 nir_foreach_phi_src(src, phi) {
142 if (src->pred == pred) {
143 set_src_live(&src->src, live);
144 break;
145 }
146 }
147 }
148
149 BITSET_WORD progress = 0;
150 for (unsigned i = 0; i < state->bitset_words; ++i) {
151 progress |= live[i] & ~pred->live_out[i];
152 pred->live_out[i] |= live[i];
153 }
154 return progress != 0;
155 }
156
157 void
158 nir_live_ssa_defs_impl(nir_function_impl *impl)
159 {
160 struct live_ssa_defs_state state;
161
162 /* We start at 1 because we reserve the index value of 0 for ssa_undef
163 * instructions. Those are never live, so their liveness information
164 * can be compacted into a single bit.
165 */
166 state.num_ssa_defs = 1;
167 nir_foreach_block(block, impl) {
168 nir_foreach_instr(instr, block)
169 nir_foreach_ssa_def(instr, index_ssa_def, &state);
170 }
171
172 nir_block_worklist_init(&state.worklist, impl->num_blocks, NULL);
173
174 /* We now know how many unique ssa definitions we have and we can go
175 * ahead and allocate live_in and live_out sets and add all of the
176 * blocks to the worklist.
177 */
178 state.bitset_words = BITSET_WORDS(state.num_ssa_defs);
179 nir_foreach_block(block, impl) {
180 init_liveness_block(block, &state);
181 }
182
183
184 /* We're now ready to work through the worklist and update the liveness
185 * sets of each of the blocks. By the time we get to this point, every
186 * block in the function implementation has been pushed onto the
187 * worklist in reverse order. As long as we keep the worklist
188 * up-to-date as we go, everything will get covered.
189 */
190 while (!nir_block_worklist_is_empty(&state.worklist)) {
191 /* We pop them off in the reverse order we pushed them on. This way
192 * the first walk of the instructions is backwards so we only walk
193 * once in the case of no control flow.
194 */
195 nir_block *block = nir_block_worklist_pop_head(&state.worklist);
196
197 memcpy(block->live_in, block->live_out,
198 state.bitset_words * sizeof(BITSET_WORD));
199
200 nir_if *following_if = nir_block_get_following_if(block);
201 if (following_if)
202 set_src_live(&following_if->condition, block->live_in);
203
204 nir_foreach_instr_reverse(instr, block) {
205 /* Phi nodes are handled seperately so we want to skip them. Since
206 * we are going backwards and they are at the beginning, we can just
207 * break as soon as we see one.
208 */
209 if (instr->type == nir_instr_type_phi)
210 break;
211
212 nir_foreach_ssa_def(instr, set_ssa_def_dead, block->live_in);
213 nir_foreach_src(instr, set_src_live, block->live_in);
214 }
215
216 /* Walk over all of the predecessors of the current block updating
217 * their live in with the live out of this one. If anything has
218 * changed, add the predecessor to the work list so that we ensure
219 * that the new information is used.
220 */
221 set_foreach(block->predecessors, entry) {
222 nir_block *pred = (nir_block *)entry->key;
223 if (propagate_across_edge(pred, block, &state))
224 nir_block_worklist_push_tail(&state.worklist, pred);
225 }
226 }
227
228 nir_block_worklist_fini(&state.worklist);
229 }
230
231 static bool
232 src_does_not_use_def(nir_src *src, void *def)
233 {
234 return !src->is_ssa || src->ssa != (nir_ssa_def *)def;
235 }
236
237 static bool
238 search_for_use_after_instr(nir_instr *start, nir_ssa_def *def)
239 {
240 /* Only look for a use strictly after the given instruction */
241 struct exec_node *node = start->node.next;
242 while (!exec_node_is_tail_sentinel(node)) {
243 nir_instr *instr = exec_node_data(nir_instr, node, node);
244 if (!nir_foreach_src(instr, src_does_not_use_def, def))
245 return true;
246 node = node->next;
247 }
248 return false;
249 }
250
251 /* Returns true if def is live at instr assuming that def comes before
252 * instr in a pre DFS search of the dominance tree.
253 */
254 static bool
255 nir_ssa_def_is_live_at(nir_ssa_def *def, nir_instr *instr)
256 {
257 if (BITSET_TEST(instr->block->live_out, def->live_index)) {
258 /* Since def dominates instr, if def is in the liveout of the block,
259 * it's live at instr
260 */
261 return true;
262 } else {
263 if (BITSET_TEST(instr->block->live_in, def->live_index) ||
264 def->parent_instr->block == instr->block) {
265 /* In this case it is either live coming into instr's block or it
266 * is defined in the same block. In this case, we simply need to
267 * see if it is used after instr.
268 */
269 return search_for_use_after_instr(instr, def);
270 } else {
271 return false;
272 }
273 }
274 }
275
276 bool
277 nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b)
278 {
279 if (a->parent_instr == b->parent_instr) {
280 /* Two variables defined at the same time interfere assuming at
281 * least one isn't dead.
282 */
283 return true;
284 } else if (a->live_index == 0 || b->live_index == 0) {
285 /* If either variable is an ssa_undef, then there's no interference */
286 return false;
287 } else if (a->live_index < b->live_index) {
288 return nir_ssa_def_is_live_at(a, b->parent_instr);
289 } else {
290 return nir_ssa_def_is_live_at(b, a->parent_instr);
291 }
292 }