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 * Connor Abbott (cwabbott0@gmail.com)
25 *
26 */
27
28 #include "nir.h"
29 #include <main/imports.h>
30
31 /**
32 * SSA-based copy propagation
33 */
34
is_move(nir_alu_instr * instr)35 static bool is_move(nir_alu_instr *instr)
36 {
37 if (instr->op != nir_op_fmov &&
38 instr->op != nir_op_imov)
39 return false;
40
41 if (instr->dest.saturate)
42 return false;
43
44 /* we handle modifiers in a separate pass */
45
46 if (instr->src[0].abs || instr->src[0].negate)
47 return false;
48
49 if (!instr->src[0].src.is_ssa)
50 return false;
51
52 return true;
53
54 }
55
is_vec(nir_alu_instr * instr)56 static bool is_vec(nir_alu_instr *instr)
57 {
58 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
59 if (!instr->src[i].src.is_ssa)
60 return false;
61
62 /* we handle modifiers in a separate pass */
63 if (instr->src[i].abs || instr->src[i].negate)
64 return false;
65 }
66
67 return instr->op == nir_op_vec2 ||
68 instr->op == nir_op_vec3 ||
69 instr->op == nir_op_vec4;
70 }
71
72 static bool
is_swizzleless_move(nir_alu_instr * instr)73 is_swizzleless_move(nir_alu_instr *instr)
74 {
75 if (is_move(instr)) {
76 for (unsigned i = 0; i < 4; i++) {
77 if (!((instr->dest.write_mask >> i) & 1))
78 break;
79 if (instr->src[0].swizzle[i] != i)
80 return false;
81 }
82 return true;
83 } else if (is_vec(instr)) {
84 nir_ssa_def *def = NULL;
85 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
86 if (instr->src[i].swizzle[0] != i)
87 return false;
88
89 if (def == NULL) {
90 def = instr->src[i].src.ssa;
91 } else if (instr->src[i].src.ssa != def) {
92 return false;
93 }
94 }
95 return true;
96 } else {
97 return false;
98 }
99 }
100
101 static bool
copy_prop_src(nir_src * src,nir_instr * parent_instr,nir_if * parent_if,unsigned num_components)102 copy_prop_src(nir_src *src, nir_instr *parent_instr, nir_if *parent_if,
103 unsigned num_components)
104 {
105 if (!src->is_ssa) {
106 if (src->reg.indirect)
107 return copy_prop_src(src->reg.indirect, parent_instr, parent_if, 1);
108 return false;
109 }
110
111 nir_instr *src_instr = src->ssa->parent_instr;
112 if (src_instr->type != nir_instr_type_alu)
113 return false;
114
115 nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
116 if (!is_swizzleless_move(alu_instr))
117 return false;
118
119 if (alu_instr->src[0].src.ssa->num_components != num_components)
120 return false;
121
122 if (parent_instr) {
123 nir_instr_rewrite_src(parent_instr, src,
124 nir_src_for_ssa(alu_instr->src[0].src.ssa));
125 } else {
126 assert(src == &parent_if->condition);
127 nir_if_rewrite_condition(parent_if,
128 nir_src_for_ssa(alu_instr->src[0].src.ssa));
129 }
130
131 return true;
132 }
133
134 static bool
copy_prop_alu_src(nir_alu_instr * parent_alu_instr,unsigned index)135 copy_prop_alu_src(nir_alu_instr *parent_alu_instr, unsigned index)
136 {
137 nir_alu_src *src = &parent_alu_instr->src[index];
138 if (!src->src.is_ssa) {
139 if (src->src.reg.indirect)
140 return copy_prop_src(src->src.reg.indirect, &parent_alu_instr->instr,
141 NULL, 1);
142 return false;
143 }
144
145 nir_instr *src_instr = src->src.ssa->parent_instr;
146 if (src_instr->type != nir_instr_type_alu)
147 return false;
148
149 nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr);
150 if (!is_move(alu_instr) && !is_vec(alu_instr))
151 return false;
152
153 nir_ssa_def *def;
154 unsigned new_swizzle[4] = {0, 0, 0, 0};
155
156 if (alu_instr->op == nir_op_fmov ||
157 alu_instr->op == nir_op_imov) {
158 for (unsigned i = 0; i < 4; i++)
159 new_swizzle[i] = alu_instr->src[0].swizzle[src->swizzle[i]];
160 def = alu_instr->src[0].src.ssa;
161 } else {
162 def = NULL;
163
164 for (unsigned i = 0; i < 4; i++) {
165 if (!nir_alu_instr_channel_used(parent_alu_instr, index, i))
166 continue;
167
168 nir_ssa_def *new_def = alu_instr->src[src->swizzle[i]].src.ssa;
169 if (def == NULL)
170 def = new_def;
171 else {
172 if (def != new_def)
173 return false;
174 }
175 new_swizzle[i] = alu_instr->src[src->swizzle[i]].swizzle[0];
176 }
177 }
178
179 for (unsigned i = 0; i < 4; i++)
180 src->swizzle[i] = new_swizzle[i];
181
182 nir_instr_rewrite_src(&parent_alu_instr->instr, &src->src,
183 nir_src_for_ssa(def));
184
185 return true;
186 }
187
188 static bool
copy_prop_dest(nir_dest * dest,nir_instr * instr)189 copy_prop_dest(nir_dest *dest, nir_instr *instr)
190 {
191 if (!dest->is_ssa && dest->reg.indirect)
192 return copy_prop_src(dest->reg.indirect, instr, NULL, 1);
193
194 return false;
195 }
196
197 static bool
copy_prop_deref_var(nir_instr * instr,nir_deref_var * deref_var)198 copy_prop_deref_var(nir_instr *instr, nir_deref_var *deref_var)
199 {
200 if (!deref_var)
201 return false;
202
203 bool progress = false;
204 for (nir_deref *deref = deref_var->deref.child;
205 deref; deref = deref->child) {
206 if (deref->deref_type != nir_deref_type_array)
207 continue;
208
209 nir_deref_array *arr = nir_deref_as_array(deref);
210 if (arr->deref_array_type != nir_deref_array_type_indirect)
211 continue;
212
213 while (copy_prop_src(&arr->indirect, instr, NULL, 1))
214 progress = true;
215 }
216 return progress;
217 }
218
219 static bool
copy_prop_instr(nir_instr * instr)220 copy_prop_instr(nir_instr *instr)
221 {
222 bool progress = false;
223 switch (instr->type) {
224 case nir_instr_type_alu: {
225 nir_alu_instr *alu_instr = nir_instr_as_alu(instr);
226
227 for (unsigned i = 0; i < nir_op_infos[alu_instr->op].num_inputs; i++)
228 while (copy_prop_alu_src(alu_instr, i))
229 progress = true;
230
231 while (copy_prop_dest(&alu_instr->dest.dest, instr))
232 progress = true;
233
234 return progress;
235 }
236
237 case nir_instr_type_tex: {
238 nir_tex_instr *tex = nir_instr_as_tex(instr);
239 for (unsigned i = 0; i < tex->num_srcs; i++) {
240 unsigned num_components = nir_tex_instr_src_size(tex, i);
241 while (copy_prop_src(&tex->src[i].src, instr, NULL, num_components))
242 progress = true;
243 }
244
245 if (copy_prop_deref_var(instr, tex->texture))
246 progress = true;
247 if (copy_prop_deref_var(instr, tex->sampler))
248 progress = true;
249
250 while (copy_prop_dest(&tex->dest, instr))
251 progress = true;
252
253 return progress;
254 }
255
256 case nir_instr_type_intrinsic: {
257 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
258 for (unsigned i = 0;
259 i < nir_intrinsic_infos[intrin->intrinsic].num_srcs; i++) {
260 unsigned num_components =
261 nir_intrinsic_infos[intrin->intrinsic].src_components[i];
262 if (!num_components)
263 num_components = intrin->num_components;
264
265 while (copy_prop_src(&intrin->src[i], instr, NULL, num_components))
266 progress = true;
267 }
268
269 for (unsigned i = 0;
270 i < nir_intrinsic_infos[intrin->intrinsic].num_variables; i++) {
271 if (copy_prop_deref_var(instr, intrin->variables[i]))
272 progress = true;
273 }
274
275 if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
276 while (copy_prop_dest(&intrin->dest, instr))
277 progress = true;
278 }
279
280 return progress;
281 }
282
283 case nir_instr_type_phi: {
284 nir_phi_instr *phi = nir_instr_as_phi(instr);
285 assert(phi->dest.is_ssa);
286 unsigned num_components = phi->dest.ssa.num_components;
287 nir_foreach_phi_src(src, phi) {
288 while (copy_prop_src(&src->src, instr, NULL, num_components))
289 progress = true;
290 }
291
292 return progress;
293 }
294
295 default:
296 return false;
297 }
298 }
299
300 static bool
copy_prop_if(nir_if * if_stmt)301 copy_prop_if(nir_if *if_stmt)
302 {
303 return copy_prop_src(&if_stmt->condition, NULL, if_stmt, 1);
304 }
305
306 static bool
nir_copy_prop_impl(nir_function_impl * impl)307 nir_copy_prop_impl(nir_function_impl *impl)
308 {
309 bool progress = false;
310
311 nir_foreach_block(block, impl) {
312 nir_foreach_instr(instr, block) {
313 if (copy_prop_instr(instr))
314 progress = true;
315 }
316
317 nir_if *if_stmt = nir_block_get_following_if(block);
318 if (if_stmt && copy_prop_if(if_stmt))
319 progress = true;
320 }
321
322 if (progress) {
323 nir_metadata_preserve(impl, nir_metadata_block_index |
324 nir_metadata_dominance);
325 }
326
327 return progress;
328 }
329
330 bool
nir_copy_prop(nir_shader * shader)331 nir_copy_prop(nir_shader *shader)
332 {
333 bool progress = false;
334
335 nir_foreach_function(function, shader) {
336 if (function->impl && nir_copy_prop_impl(function->impl))
337 progress = true;
338 }
339
340 return progress;
341 }
342