1/*
2 * Copyright © 2010 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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24/**
25 * \file lower_if_to_cond_assign.cpp
26 *
27 * This flattens if-statements to conditional assignments if:
28 *
29 * - the GPU has limited or no flow control support
30 *   (controlled by max_depth)
31 *
32 * - small conditional branches are more expensive than conditional assignments
33 *   (controlled by min_branch_cost, that's the cost for a branch to be
34 *    preserved)
35 *
36 * It can't handle other control flow being inside of its block, such
37 * as calls or loops.  Hopefully loop unrolling and inlining will take
38 * care of those.
39 *
40 * Drivers for GPUs with no control flow support should simply call
41 *
42 *    lower_if_to_cond_assign(instructions)
43 *
44 * to attempt to flatten all if-statements.
45 *
46 * Some GPUs (such as i965 prior to gen6) do support control flow, but have a
47 * maximum nesting depth N.  Drivers for such hardware can call
48 *
49 *    lower_if_to_cond_assign(instructions, N)
50 *
51 * to attempt to flatten any if-statements appearing at depth > N.
52 */
53
54#include "compiler/glsl_types.h"
55#include "ir.h"
56#include "util/set.h"
57#include "util/hash_table.h" /* Needed for the hashing functions */
58#include "main/macros.h" /* for MAX2 */
59
60namespace {
61
62class ir_if_to_cond_assign_visitor : public ir_hierarchical_visitor {
63public:
64   ir_if_to_cond_assign_visitor(gl_shader_stage stage,
65                                unsigned max_depth,
66                                unsigned min_branch_cost)
67   {
68      this->found_unsupported_op = false;
69      this->found_expensive_op = false;
70      this->found_dynamic_arrayref = false;
71      this->is_then = false;
72      this->progress = false;
73      this->stage = stage;
74      this->then_cost = 0;
75      this->else_cost = 0;
76      this->max_depth = max_depth;
77      this->min_branch_cost = min_branch_cost;
78      this->depth = 0;
79
80      this->condition_variables = _mesa_pointer_set_create(NULL);
81   }
82
83   ~ir_if_to_cond_assign_visitor()
84   {
85      _mesa_set_destroy(this->condition_variables, NULL);
86   }
87
88   ir_visitor_status visit_enter(ir_if *);
89   ir_visitor_status visit_leave(ir_if *);
90
91   bool found_unsupported_op;
92   bool found_expensive_op;
93   bool found_dynamic_arrayref;
94   bool is_then;
95   bool progress;
96   gl_shader_stage stage;
97   unsigned then_cost;
98   unsigned else_cost;
99   unsigned min_branch_cost;
100   unsigned max_depth;
101   unsigned depth;
102
103   struct set *condition_variables;
104};
105
106} /* anonymous namespace */
107
108bool
109lower_if_to_cond_assign(gl_shader_stage stage, exec_list *instructions,
110                        unsigned max_depth, unsigned min_branch_cost)
111{
112   if (max_depth == UINT_MAX)
113      return false;
114
115   ir_if_to_cond_assign_visitor v(stage, max_depth, min_branch_cost);
116
117   visit_list_elements(&v, instructions);
118
119   return v.progress;
120}
121
122static void
123check_ir_node(ir_instruction *ir, void *data)
124{
125   ir_if_to_cond_assign_visitor *v = (ir_if_to_cond_assign_visitor *)data;
126
127   switch (ir->ir_type) {
128   case ir_type_call:
129   case ir_type_discard:
130   case ir_type_loop:
131   case ir_type_loop_jump:
132   case ir_type_return:
133   case ir_type_emit_vertex:
134   case ir_type_end_primitive:
135   case ir_type_barrier:
136      v->found_unsupported_op = true;
137      break;
138
139   case ir_type_dereference_variable: {
140      ir_variable *var = ir->as_dereference_variable()->variable_referenced();
141
142      /* Lowering branches with TCS output accesses breaks many piglit tests,
143       * so don't touch them for now.
144       */
145      if (v->stage == MESA_SHADER_TESS_CTRL &&
146          var->data.mode == ir_var_shader_out)
147         v->found_unsupported_op = true;
148      break;
149   }
150
151   /* SSBO, images, atomic counters are handled by ir_type_call */
152   case ir_type_texture:
153      v->found_expensive_op = true;
154      break;
155
156   case ir_type_dereference_array: {
157      ir_dereference_array *deref = ir->as_dereference_array();
158
159      if (deref->array_index->ir_type != ir_type_constant)
160         v->found_dynamic_arrayref = true;
161   } FALLTHROUGH;
162   case ir_type_expression:
163   case ir_type_dereference_record:
164      if (v->is_then)
165         v->then_cost++;
166      else
167         v->else_cost++;
168      break;
169
170   default:
171      break;
172   }
173}
174
175static void
176move_block_to_cond_assign(void *mem_ctx,
177                          ir_if *if_ir, ir_rvalue *cond_expr,
178                          exec_list *instructions,
179                          struct set *set)
180{
181   foreach_in_list_safe(ir_instruction, ir, instructions) {
182      if (ir->ir_type == ir_type_assignment) {
183         ir_assignment *assign = (ir_assignment *)ir;
184
185         if (_mesa_set_search(set, assign) == NULL) {
186            _mesa_set_add(set, assign);
187
188            /* If the LHS of the assignment is a condition variable that was
189             * previously added, insert an additional assignment of false to
190             * the variable.
191             */
192            const bool assign_to_cv =
193               _mesa_set_search(
194                  set, assign->lhs->variable_referenced()) != NULL;
195
196            if (!assign->condition) {
197               if (assign_to_cv) {
198                  assign->rhs =
199                     new(mem_ctx) ir_expression(ir_binop_logic_and,
200                                                glsl_type::bool_type,
201                                                cond_expr->clone(mem_ctx, NULL),
202                                                assign->rhs);
203               } else {
204                  assign->condition = cond_expr->clone(mem_ctx, NULL);
205               }
206            } else {
207               assign->condition =
208                  new(mem_ctx) ir_expression(ir_binop_logic_and,
209                                             glsl_type::bool_type,
210                                             cond_expr->clone(mem_ctx, NULL),
211                                             assign->condition);
212            }
213         }
214      }
215
216      /* Now, move from the if block to the block surrounding it. */
217      ir->remove();
218      if_ir->insert_before(ir);
219   }
220}
221
222ir_visitor_status
223ir_if_to_cond_assign_visitor::visit_enter(ir_if *)
224{
225   this->depth++;
226
227   return visit_continue;
228}
229
230ir_visitor_status
231ir_if_to_cond_assign_visitor::visit_leave(ir_if *ir)
232{
233   bool must_lower = this->depth-- > this->max_depth;
234
235   /* Only flatten when beyond the GPU's maximum supported nesting depth. */
236   if (!must_lower && this->min_branch_cost == 0)
237      return visit_continue;
238
239   this->found_unsupported_op = false;
240   this->found_expensive_op = false;
241   this->found_dynamic_arrayref = false;
242   this->then_cost = 0;
243   this->else_cost = 0;
244
245   ir_assignment *assign;
246
247   /* Check that both blocks don't contain anything we can't support. */
248   this->is_then = true;
249   foreach_in_list(ir_instruction, then_ir, &ir->then_instructions) {
250      visit_tree(then_ir, check_ir_node, this);
251   }
252
253   this->is_then = false;
254   foreach_in_list(ir_instruction, else_ir, &ir->else_instructions) {
255      visit_tree(else_ir, check_ir_node, this);
256   }
257
258   if (this->found_unsupported_op)
259      return visit_continue; /* can't handle inner unsupported opcodes */
260
261   /* Skip if the branch cost is high enough or if there's an expensive op.
262    *
263    * Also skip if non-constant array indices were encountered, since those
264    * can be out-of-bounds for a not-taken branch, and so generating an
265    * assignment would be incorrect. In the case of must_lower, it's up to the
266    * backend to deal with any potential fall-out (perhaps by translating the
267    * assignments to hardware-predicated moves).
268    */
269   if (!must_lower &&
270       (this->found_expensive_op ||
271        this->found_dynamic_arrayref ||
272        MAX2(this->then_cost, this->else_cost) >= this->min_branch_cost))
273      return visit_continue;
274
275   void *mem_ctx = ralloc_parent(ir);
276
277   /* Store the condition to a variable.  Move all of the instructions from
278    * the then-clause of the if-statement.  Use the condition variable as a
279    * condition for all assignments.
280    */
281   ir_variable *const then_var =
282      new(mem_ctx) ir_variable(glsl_type::bool_type,
283                               "if_to_cond_assign_then",
284                               ir_var_temporary);
285   ir->insert_before(then_var);
286
287   ir_dereference_variable *then_cond =
288      new(mem_ctx) ir_dereference_variable(then_var);
289
290   assign = new(mem_ctx) ir_assignment(then_cond, ir->condition);
291   ir->insert_before(assign);
292
293   move_block_to_cond_assign(mem_ctx, ir, then_cond,
294                             &ir->then_instructions,
295                             this->condition_variables);
296
297   /* Add the new condition variable to the hash table.  This allows us to
298    * find this variable when lowering other (enclosing) if-statements.
299    */
300   _mesa_set_add(this->condition_variables, then_var);
301
302   /* If there are instructions in the else-clause, store the inverse of the
303    * condition to a variable.  Move all of the instructions from the
304    * else-clause if the if-statement.  Use the (inverse) condition variable
305    * as a condition for all assignments.
306    */
307   if (!ir->else_instructions.is_empty()) {
308      ir_variable *const else_var =
309         new(mem_ctx) ir_variable(glsl_type::bool_type,
310                                  "if_to_cond_assign_else",
311                                  ir_var_temporary);
312      ir->insert_before(else_var);
313
314      ir_dereference_variable *else_cond =
315         new(mem_ctx) ir_dereference_variable(else_var);
316
317      ir_rvalue *inverse =
318         new(mem_ctx) ir_expression(ir_unop_logic_not,
319                                    then_cond->clone(mem_ctx, NULL));
320
321      assign = new(mem_ctx) ir_assignment(else_cond, inverse);
322      ir->insert_before(assign);
323
324      move_block_to_cond_assign(mem_ctx, ir, else_cond,
325                                &ir->else_instructions,
326                                this->condition_variables);
327
328      /* Add the new condition variable to the hash table.  This allows us to
329       * find this variable when lowering other (enclosing) if-statements.
330       */
331      _mesa_set_add(this->condition_variables, else_var);
332   }
333
334   ir->remove();
335
336   this->progress = true;
337
338   return visit_continue;
339}
340