Merge branch 'wip/nir-vtn' into vulkan
[mesa.git] / src / gallium / drivers / freedreno / ir3 / ir3_group.c
1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3 /*
4 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 * Rob Clark <robclark@freedesktop.org>
27 */
28
29 #include "freedreno_util.h"
30
31 #include "ir3.h"
32
33 /*
34 * Find/group instruction neighbors:
35 */
36
37 /* bleh.. we need to do the same group_n() thing for both inputs/outputs
38 * (where we have a simple instr[] array), and fanin nodes (where we have
39 * an extra indirection via reg->instr).
40 */
41 struct group_ops {
42 struct ir3_instruction *(*get)(void *arr, int idx);
43 void (*insert_mov)(void *arr, int idx, struct ir3_instruction *instr);
44 };
45
46 static struct ir3_instruction *arr_get(void *arr, int idx)
47 {
48 return ((struct ir3_instruction **)arr)[idx];
49 }
50 static void arr_insert_mov_out(void *arr, int idx, struct ir3_instruction *instr)
51 {
52 ((struct ir3_instruction **)arr)[idx] =
53 ir3_MOV(instr->block, instr, TYPE_F32);
54 }
55 static void arr_insert_mov_in(void *arr, int idx, struct ir3_instruction *instr)
56 {
57 /* so, we can't insert a mov in front of a meta:in.. and the downstream
58 * instruction already has a pointer to 'instr'. So we cheat a bit and
59 * morph the meta:in instruction into a mov and insert a new meta:in
60 * in front.
61 */
62 struct ir3_instruction *in;
63
64 debug_assert(instr->regs_count == 1);
65
66 in = ir3_instr_create(instr->block, -1, OPC_META_INPUT);
67 in->inout.block = instr->block;
68 ir3_reg_create(in, instr->regs[0]->num, 0);
69
70 /* create src reg for meta:in and fixup to now be a mov: */
71 ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = in;
72 instr->category = 1;
73 instr->opc = 0;
74 instr->cat1.src_type = TYPE_F32;
75 instr->cat1.dst_type = TYPE_F32;
76
77 ((struct ir3_instruction **)arr)[idx] = in;
78 }
79 static struct group_ops arr_ops_out = { arr_get, arr_insert_mov_out };
80 static struct group_ops arr_ops_in = { arr_get, arr_insert_mov_in };
81
82 static struct ir3_instruction *instr_get(void *arr, int idx)
83 {
84 return ssa(((struct ir3_instruction *)arr)->regs[idx+1]);
85 }
86 static void
87 instr_insert_mov(void *arr, int idx, struct ir3_instruction *instr)
88 {
89 ((struct ir3_instruction *)arr)->regs[idx+1]->instr =
90 ir3_MOV(instr->block, instr, TYPE_F32);
91 }
92 static struct group_ops instr_ops = { instr_get, instr_insert_mov };
93
94
95 static void
96 group_n(struct group_ops *ops, void *arr, unsigned n)
97 {
98 unsigned i, j;
99
100 /* first pass, figure out what has conflicts and needs a mov
101 * inserted. Do this up front, before starting to setup
102 * left/right neighbor pointers. Trying to do it in a single
103 * pass could result in a situation where we can't even setup
104 * the mov's right neighbor ptr if the next instr also needs
105 * a mov.
106 */
107 restart:
108 for (i = 0; i < n; i++) {
109 struct ir3_instruction *instr = ops->get(arr, i);
110 if (instr) {
111 struct ir3_instruction *left = (i > 0) ? ops->get(arr, i - 1) : NULL;
112 struct ir3_instruction *right = (i < (n-1)) ? ops->get(arr, i + 1) : NULL;
113 bool conflict;
114
115 /* check for left/right neighbor conflicts: */
116 conflict = conflicts(instr->cp.left, left) ||
117 conflicts(instr->cp.right, right);
118
119 /* RA can't yet deal very well w/ group'd phi's: */
120 if (is_meta(instr) && (instr->opc == OPC_META_PHI))
121 conflict = true;
122
123 /* we also can't have an instr twice in the group: */
124 for (j = i + 1; (j < n) && !conflict; j++)
125 if (ops->get(arr, j) == instr)
126 conflict = true;
127
128 if (conflict) {
129 ops->insert_mov(arr, i, instr);
130 /* inserting the mov may have caused a conflict
131 * against the previous:
132 */
133 goto restart;
134 }
135 }
136 }
137
138 /* second pass, now that we've inserted mov's, fixup left/right
139 * neighbors. This is guaranteed to succeed, since by definition
140 * the newly inserted mov's cannot conflict with anything.
141 */
142 for (i = 0; i < n; i++) {
143 struct ir3_instruction *instr = ops->get(arr, i);
144 if (instr) {
145 struct ir3_instruction *left = (i > 0) ? ops->get(arr, i - 1) : NULL;
146 struct ir3_instruction *right = (i < (n-1)) ? ops->get(arr, i + 1) : NULL;
147
148 debug_assert(!conflicts(instr->cp.left, left));
149 if (left) {
150 instr->cp.left_cnt++;
151 instr->cp.left = left;
152 }
153
154 debug_assert(!conflicts(instr->cp.right, right));
155 if (right) {
156 instr->cp.right_cnt++;
157 instr->cp.right = right;
158 }
159 }
160 }
161 }
162
163 static void
164 instr_find_neighbors(struct ir3_instruction *instr)
165 {
166 struct ir3_instruction *src;
167
168 if (ir3_instr_check_mark(instr))
169 return;
170
171 if (is_meta(instr) && (instr->opc == OPC_META_FI))
172 group_n(&instr_ops, instr, instr->regs_count - 1);
173
174 foreach_ssa_src(src, instr)
175 instr_find_neighbors(src);
176 }
177
178 /* a bit of sadness.. we can't have "holes" in inputs from PoV of
179 * register assignment, they still need to be grouped together. So
180 * we need to insert dummy/padding instruction for grouping, and
181 * then take it back out again before anyone notices.
182 */
183 static void
184 pad_and_group_input(struct ir3_instruction **input, unsigned n)
185 {
186 int i, mask = 0;
187 struct ir3_block *block = NULL;
188
189 for (i = n - 1; i >= 0; i--) {
190 struct ir3_instruction *instr = input[i];
191 if (instr) {
192 block = instr->block;
193 } else if (block) {
194 instr = ir3_NOP(block);
195 ir3_reg_create(instr, 0, IR3_REG_SSA); /* dummy dst */
196 input[i] = instr;
197 mask |= (1 << i);
198 }
199 }
200
201 group_n(&arr_ops_in, input, n);
202
203 for (i = 0; i < n; i++) {
204 if (mask & (1 << i))
205 input[i] = NULL;
206 }
207 }
208
209 static void
210 find_neighbors(struct ir3 *ir)
211 {
212 unsigned i;
213
214 /* shader inputs/outputs themselves must be contiguous as well:
215 *
216 * NOTE: group inputs first, since we only insert mov's
217 * *before* the conflicted instr (and that would go badly
218 * for inputs). By doing inputs first, we should never
219 * have a conflict on inputs.. pushing any conflict to
220 * resolve to the outputs, for stuff like:
221 *
222 * MOV OUT[n], IN[m].wzyx
223 *
224 * NOTE: we assume here inputs/outputs are grouped in vec4.
225 * This logic won't quite cut it if we don't align smaller
226 * on vec4 boundaries
227 */
228 for (i = 0; i < ir->ninputs; i += 4)
229 pad_and_group_input(&ir->inputs[i], 4);
230 for (i = 0; i < ir->noutputs; i += 4)
231 group_n(&arr_ops_out, &ir->outputs[i], 4);
232
233 for (i = 0; i < ir->noutputs; i++) {
234 if (ir->outputs[i]) {
235 struct ir3_instruction *instr = ir->outputs[i];
236 instr_find_neighbors(instr);
237 }
238 }
239 }
240
241 void
242 ir3_group(struct ir3 *ir)
243 {
244 ir3_clear_mark(ir);
245 find_neighbors(ir);
246 }