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
28#include "nir.h"
29#include "nir_control_flow.h"
30
31/*
32 * Implements a small peephole optimization that looks for
33 *
34 * if (cond) {
35 *    <then SSA defs>
36 * } else {
37 *    <else SSA defs>
38 * }
39 * phi
40 * ...
41 * phi
42 *
43 * and replaces it with:
44 *
45 * <then SSA defs>
46 * <else SSA defs>
47 * bcsel
48 * ...
49 * bcsel
50 *
51 * where the SSA defs are ALU operations or other cheap instructions (not
52 * texturing, for example).
53 *
54 * If the number of ALU operations in the branches is greater than the limit
55 * parameter, then the optimization is skipped.  In limit=0 mode, the SSA defs
56 * must only be MOVs which we expect to get copy-propagated away once they're
57 * out of the inner blocks.
58 */
59
60static bool
61block_check_for_allowed_instrs(nir_block *block, unsigned *count, bool alu_ok)
62{
63   nir_foreach_instr(instr, block) {
64      switch (instr->type) {
65      case nir_instr_type_intrinsic: {
66         nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
67
68         switch (intrin->intrinsic) {
69         case nir_intrinsic_load_var:
70            switch (intrin->variables[0]->var->data.mode) {
71            case nir_var_shader_in:
72            case nir_var_uniform:
73               break;
74
75            default:
76               return false;
77            }
78            break;
79
80         case nir_intrinsic_load_uniform:
81            if (!alu_ok)
82               return false;
83            break;
84
85         default:
86            return false;
87         }
88
89         break;
90      }
91
92      case nir_instr_type_load_const:
93         break;
94
95      case nir_instr_type_alu: {
96         nir_alu_instr *mov = nir_instr_as_alu(instr);
97         switch (mov->op) {
98         case nir_op_fmov:
99         case nir_op_imov:
100         case nir_op_fneg:
101         case nir_op_ineg:
102         case nir_op_fabs:
103         case nir_op_iabs:
104         case nir_op_vec2:
105         case nir_op_vec3:
106         case nir_op_vec4:
107            break;
108         default:
109            if (!alu_ok) {
110               /* It must be a move-like operation. */
111               return false;
112            }
113            break;
114         }
115
116         /* It must be SSA */
117         if (!mov->dest.dest.is_ssa)
118            return false;
119
120         if (alu_ok) {
121            (*count)++;
122         } else {
123            /* Can't handle saturate */
124            if (mov->dest.saturate)
125               return false;
126
127            /* It cannot have any if-uses */
128            if (!list_empty(&mov->dest.dest.ssa.if_uses))
129               return false;
130
131            /* The only uses of this definition must be phi's in the successor */
132            nir_foreach_use(use, &mov->dest.dest.ssa) {
133               if (use->parent_instr->type != nir_instr_type_phi ||
134                   use->parent_instr->block != block->successors[0])
135                  return false;
136            }
137         }
138         break;
139      }
140
141      default:
142         return false;
143      }
144   }
145
146   return true;
147}
148
149static bool
150nir_opt_peephole_select_block(nir_block *block, nir_shader *shader,
151                              unsigned limit)
152{
153   if (nir_cf_node_is_first(&block->cf_node))
154      return false;
155
156   nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node);
157   if (prev_node->type != nir_cf_node_if)
158      return false;
159
160   nir_if *if_stmt = nir_cf_node_as_if(prev_node);
161   nir_block *then_block = nir_if_first_then_block(if_stmt);
162   nir_block *else_block = nir_if_first_else_block(if_stmt);
163
164   /* We can only have one block in each side ... */
165   if (nir_if_last_then_block(if_stmt) != then_block ||
166       nir_if_last_else_block(if_stmt) != else_block)
167      return false;
168
169   /* ... and those blocks must only contain "allowed" instructions. */
170   unsigned count = 0;
171   if (!block_check_for_allowed_instrs(then_block, &count, limit != 0) ||
172       !block_check_for_allowed_instrs(else_block, &count, limit != 0))
173      return false;
174
175   if (count > limit)
176      return false;
177
178   /* At this point, we know that the previous CFG node is an if-then
179    * statement containing only moves to phi nodes in this block.  We can
180    * just remove that entire CF node and replace all of the phi nodes with
181    * selects.
182    */
183
184   nir_block *prev_block = nir_cf_node_as_block(nir_cf_node_prev(prev_node));
185
186   /* First, we move the remaining instructions from the blocks to the
187    * block before.  We have already guaranteed that this is safe by
188    * calling block_check_for_allowed_instrs()
189    */
190   nir_foreach_instr_safe(instr, then_block) {
191      exec_node_remove(&instr->node);
192      instr->block = prev_block;
193      exec_list_push_tail(&prev_block->instr_list, &instr->node);
194   }
195
196   nir_foreach_instr_safe(instr, else_block) {
197      exec_node_remove(&instr->node);
198      instr->block = prev_block;
199      exec_list_push_tail(&prev_block->instr_list, &instr->node);
200   }
201
202   nir_foreach_instr_safe(instr, block) {
203      if (instr->type != nir_instr_type_phi)
204         break;
205
206      nir_phi_instr *phi = nir_instr_as_phi(instr);
207      nir_alu_instr *sel = nir_alu_instr_create(shader, nir_op_bcsel);
208      nir_src_copy(&sel->src[0].src, &if_stmt->condition, sel);
209      /* Splat the condition to all channels */
210      memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle);
211
212      assert(exec_list_length(&phi->srcs) == 2);
213      nir_foreach_phi_src(src, phi) {
214         assert(src->pred == then_block || src->pred == else_block);
215         assert(src->src.is_ssa);
216
217         unsigned idx = src->pred == then_block ? 1 : 2;
218         nir_src_copy(&sel->src[idx].src, &src->src, sel);
219      }
220
221      nir_ssa_dest_init(&sel->instr, &sel->dest.dest,
222                        phi->dest.ssa.num_components,
223                        phi->dest.ssa.bit_size, phi->dest.ssa.name);
224      sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1;
225
226      nir_ssa_def_rewrite_uses(&phi->dest.ssa,
227                               nir_src_for_ssa(&sel->dest.dest.ssa));
228
229      nir_instr_insert_before(&phi->instr, &sel->instr);
230      nir_instr_remove(&phi->instr);
231   }
232
233   nir_cf_node_remove(&if_stmt->cf_node);
234   return true;
235}
236
237static bool
238nir_opt_peephole_select_impl(nir_function_impl *impl, unsigned limit)
239{
240   nir_shader *shader = impl->function->shader;
241   bool progress = false;
242
243   nir_foreach_block_safe(block, impl) {
244      progress |= nir_opt_peephole_select_block(block, shader, limit);
245   }
246
247   if (progress)
248      nir_metadata_preserve(impl, nir_metadata_none);
249
250   return progress;
251}
252
253bool
254nir_opt_peephole_select(nir_shader *shader, unsigned limit)
255{
256   bool progress = false;
257
258   nir_foreach_function(function, shader) {
259      if (function->impl)
260         progress |= nir_opt_peephole_select_impl(function->impl, limit);
261   }
262
263   return progress;
264}
265