• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2018 Red Hat
3  * Copyright © 2019 Valve Corporation
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22  * IN THE SOFTWARE.
23  *
24  * Authors:
25  *    Rob Clark (robdclark@gmail.com>
26  *    Daniel Schürmann (daniel.schuermann@campus.tu-berlin.de)
27  *    Rhys Perry (pendingchaos02@gmail.com)
28  *
29  */
30 
31 #include "nir.h"
32 
33 
34 /*
35  * A simple pass that moves some instructions into the least common
36  * anscestor of consuming instructions.
37  */
38 
39 bool
nir_can_move_instr(nir_instr * instr,nir_move_options options)40 nir_can_move_instr(nir_instr *instr, nir_move_options options)
41 {
42    switch (instr->type) {
43    case nir_instr_type_load_const:
44    case nir_instr_type_ssa_undef: {
45       return options & nir_move_const_undef;
46    }
47    case nir_instr_type_alu: {
48       if (nir_op_is_vec(nir_instr_as_alu(instr)->op) ||
49           nir_instr_as_alu(instr)->op == nir_op_b2i32)
50          return options & nir_move_copies;
51       if (nir_alu_instr_is_comparison(nir_instr_as_alu(instr)))
52          return options & nir_move_comparisons;
53       return false;
54    }
55    case nir_instr_type_intrinsic: {
56       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
57       switch (intrin->intrinsic) {
58       case nir_intrinsic_load_ubo:
59       case nir_intrinsic_load_ubo_vec4:
60          return options & nir_move_load_ubo;
61       case nir_intrinsic_load_ssbo:
62          return (options & nir_move_load_ssbo) && nir_intrinsic_can_reorder(intrin);
63       case nir_intrinsic_load_input:
64       case nir_intrinsic_load_interpolated_input:
65       case nir_intrinsic_load_per_vertex_input:
66          return options & nir_move_load_input;
67       case nir_intrinsic_load_uniform:
68          return options & nir_move_load_uniform;
69       default:
70          return false;
71       }
72    }
73    default:
74       return false;
75    }
76 }
77 
78 static nir_loop *
get_innermost_loop(nir_cf_node * node)79 get_innermost_loop(nir_cf_node *node)
80 {
81    for (; node != NULL; node = node->parent) {
82       if (node->type == nir_cf_node_loop)
83          return (nir_loop*)node;
84    }
85    return NULL;
86 }
87 
88 static bool
loop_contains_block(nir_loop * loop,nir_block * block)89 loop_contains_block(nir_loop *loop, nir_block *block)
90 {
91    nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node));
92    nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node));
93 
94    return block->index > before->index && block->index < after->index;
95 }
96 
97 /* Given the LCA of all uses and the definition, find a block on the path
98  * between them in the dominance tree that is outside of as many loops as
99  * possible. If "sink_out_of_loops" is false, then we disallow sinking the
100  * definition outside of the loop it's defined in (if any).
101  */
102 
103 static nir_block *
adjust_block_for_loops(nir_block * use_block,nir_block * def_block,bool sink_out_of_loops)104 adjust_block_for_loops(nir_block *use_block, nir_block *def_block,
105                        bool sink_out_of_loops)
106 {
107    nir_loop *def_loop = NULL;
108    if (!sink_out_of_loops)
109       def_loop = get_innermost_loop(&def_block->cf_node);
110 
111    for (nir_block *cur_block = use_block; cur_block != def_block->imm_dom;
112         cur_block = cur_block->imm_dom) {
113       if (!sink_out_of_loops && def_loop &&
114           !loop_contains_block(def_loop, use_block)) {
115          use_block = cur_block;
116          continue;
117       }
118 
119       nir_cf_node *next = nir_cf_node_next(&cur_block->cf_node);
120       if (next && next->type == nir_cf_node_loop) {
121          nir_loop *following_loop = nir_cf_node_as_loop(next);
122          if (loop_contains_block(following_loop, use_block)) {
123              use_block = cur_block;
124              continue;
125          }
126       }
127    }
128 
129    return use_block;
130 }
131 
132 /* iterate a ssa def's use's and try to find a more optimal block to
133  * move it to, using the dominance tree.  In short, if all of the uses
134  * are contained in a single block, the load will be moved there,
135  * otherwise it will be move to the least common ancestor block of all
136  * the uses
137  */
138 static nir_block *
get_preferred_block(nir_ssa_def * def,bool sink_out_of_loops)139 get_preferred_block(nir_ssa_def *def, bool sink_out_of_loops)
140 {
141    nir_block *lca = NULL;
142 
143    nir_foreach_use(use, def) {
144       nir_instr *instr = use->parent_instr;
145       nir_block *use_block = instr->block;
146 
147       /*
148        * Kind of an ugly special-case, but phi instructions
149        * need to appear first in the block, so by definition
150        * we can't move an instruction into a block where it is
151        * consumed by a phi instruction.  We could conceivably
152        * move it into a dominator block.
153        */
154       if (instr->type == nir_instr_type_phi) {
155          nir_phi_instr *phi = nir_instr_as_phi(instr);
156          nir_block *phi_lca = NULL;
157          nir_foreach_phi_src(src, phi) {
158             if (&src->src == use)
159                phi_lca = nir_dominance_lca(phi_lca, src->pred);
160          }
161          use_block = phi_lca;
162       }
163 
164       lca = nir_dominance_lca(lca, use_block);
165    }
166 
167    nir_foreach_if_use(use, def) {
168       nir_block *use_block =
169          nir_cf_node_as_block(nir_cf_node_prev(&use->parent_if->cf_node));
170 
171       lca = nir_dominance_lca(lca, use_block);
172    }
173 
174    /* return in case, we didn't find a reachable user */
175    if (!lca)
176       return NULL;
177 
178    /* We don't sink any instructions into loops to avoid repeated executions
179     * This might occasionally increase register pressure, but seems overall
180     * the better choice.
181     */
182    lca = adjust_block_for_loops(lca, def->parent_instr->block,
183                                 sink_out_of_loops);
184    assert(nir_block_dominates(def->parent_instr->block, lca));
185 
186    return lca;
187 }
188 
189 static bool
can_sink_out_of_loop(nir_intrinsic_instr * intrin)190 can_sink_out_of_loop(nir_intrinsic_instr *intrin)
191 {
192    /* Don't sink buffer loads out of loops because that can make its
193     * resource divergent and break code like that which is generated
194     * by nir_lower_non_uniform_access.
195     */
196    return intrin->intrinsic != nir_intrinsic_load_ubo &&
197           intrin->intrinsic != nir_intrinsic_load_ssbo;
198 }
199 
200 bool
nir_opt_sink(nir_shader * shader,nir_move_options options)201 nir_opt_sink(nir_shader *shader, nir_move_options options)
202 {
203    bool progress = false;
204 
205    nir_foreach_function(function, shader) {
206       if (!function->impl)
207          continue;
208 
209       nir_metadata_require(function->impl,
210                            nir_metadata_block_index | nir_metadata_dominance);
211 
212       nir_foreach_block_reverse(block, function->impl) {
213          nir_foreach_instr_reverse_safe(instr, block) {
214             if (!nir_can_move_instr(instr, options))
215                continue;
216 
217             nir_ssa_def *def = nir_instr_ssa_def(instr);
218 
219             bool sink_out_of_loops =
220                instr->type != nir_instr_type_intrinsic ||
221                can_sink_out_of_loop(nir_instr_as_intrinsic(instr));
222             nir_block *use_block =
223                   get_preferred_block(def, sink_out_of_loops);
224 
225             if (!use_block || use_block == instr->block)
226                continue;
227 
228             nir_instr_remove(instr);
229             nir_instr_insert(nir_after_phis(use_block), instr);
230 
231             progress = true;
232          }
233       }
234 
235       nir_metadata_preserve(function->impl,
236                             nir_metadata_block_index | nir_metadata_dominance);
237    }
238 
239    return progress;
240 }
241