• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
60 namespace {
61 
62 class ir_if_to_cond_assign_visitor : public ir_hierarchical_visitor {
63 public:
ir_if_to_cond_assign_visitor(gl_shader_stage stage,unsigned max_depth,unsigned min_branch_cost)64    ir_if_to_cond_assign_visitor(gl_shader_stage stage,
65                                 unsigned max_depth,
66                                 unsigned min_branch_cost)
67    {
68       this->progress = false;
69       this->stage = stage;
70       this->max_depth = max_depth;
71       this->min_branch_cost = min_branch_cost;
72       this->depth = 0;
73 
74       this->condition_variables =
75             _mesa_set_create(NULL, _mesa_hash_pointer,
76                                     _mesa_key_pointer_equal);
77    }
78 
~ir_if_to_cond_assign_visitor()79    ~ir_if_to_cond_assign_visitor()
80    {
81       _mesa_set_destroy(this->condition_variables, NULL);
82    }
83 
84    ir_visitor_status visit_enter(ir_if *);
85    ir_visitor_status visit_leave(ir_if *);
86 
87    bool found_unsupported_op;
88    bool found_expensive_op;
89    bool found_dynamic_arrayref;
90    bool is_then;
91    bool progress;
92    gl_shader_stage stage;
93    unsigned then_cost;
94    unsigned else_cost;
95    unsigned min_branch_cost;
96    unsigned max_depth;
97    unsigned depth;
98 
99    struct set *condition_variables;
100 };
101 
102 } /* anonymous namespace */
103 
104 bool
lower_if_to_cond_assign(gl_shader_stage stage,exec_list * instructions,unsigned max_depth,unsigned min_branch_cost)105 lower_if_to_cond_assign(gl_shader_stage stage, exec_list *instructions,
106                         unsigned max_depth, unsigned min_branch_cost)
107 {
108    if (max_depth == UINT_MAX)
109       return false;
110 
111    ir_if_to_cond_assign_visitor v(stage, max_depth, min_branch_cost);
112 
113    visit_list_elements(&v, instructions);
114 
115    return v.progress;
116 }
117 
118 void
check_ir_node(ir_instruction * ir,void * data)119 check_ir_node(ir_instruction *ir, void *data)
120 {
121    ir_if_to_cond_assign_visitor *v = (ir_if_to_cond_assign_visitor *)data;
122 
123    switch (ir->ir_type) {
124    case ir_type_call:
125    case ir_type_discard:
126    case ir_type_loop:
127    case ir_type_loop_jump:
128    case ir_type_return:
129    case ir_type_emit_vertex:
130    case ir_type_end_primitive:
131    case ir_type_barrier:
132       v->found_unsupported_op = true;
133       break;
134 
135    case ir_type_dereference_variable: {
136       ir_variable *var = ir->as_dereference_variable()->variable_referenced();
137 
138       /* Lowering branches with TCS output accesses breaks many piglit tests,
139        * so don't touch them for now.
140        */
141       if (v->stage == MESA_SHADER_TESS_CTRL &&
142           var->data.mode == ir_var_shader_out)
143          v->found_unsupported_op = true;
144       break;
145    }
146 
147    /* SSBO, images, atomic counters are handled by ir_type_call */
148    case ir_type_texture:
149       v->found_expensive_op = true;
150       break;
151 
152    case ir_type_dereference_array: {
153       ir_dereference_array *deref = ir->as_dereference_array();
154 
155       if (deref->array_index->ir_type != ir_type_constant)
156          v->found_dynamic_arrayref = true;
157    } /* fall-through */
158    case ir_type_expression:
159    case ir_type_dereference_record:
160       if (v->is_then)
161          v->then_cost++;
162       else
163          v->else_cost++;
164       break;
165 
166    default:
167       break;
168    }
169 }
170 
171 void
move_block_to_cond_assign(void * mem_ctx,ir_if * if_ir,ir_rvalue * cond_expr,exec_list * instructions,struct set * set)172 move_block_to_cond_assign(void *mem_ctx,
173 			  ir_if *if_ir, ir_rvalue *cond_expr,
174 			  exec_list *instructions,
175 			  struct set *set)
176 {
177    foreach_in_list_safe(ir_instruction, ir, instructions) {
178       if (ir->ir_type == ir_type_assignment) {
179 	 ir_assignment *assign = (ir_assignment *)ir;
180 
181 	 if (_mesa_set_search(set, assign) == NULL) {
182 	    _mesa_set_add(set, assign);
183 
184 	    /* If the LHS of the assignment is a condition variable that was
185 	     * previously added, insert an additional assignment of false to
186 	     * the variable.
187 	     */
188 	    const bool assign_to_cv =
189 	          _mesa_set_search(
190 	                set, assign->lhs->variable_referenced()) != NULL;
191 
192 	    if (!assign->condition) {
193           if (assign_to_cv) {
194              assign->rhs =
195 		          new(mem_ctx) ir_expression(ir_binop_logic_and,
196                                            glsl_type::bool_type,
197                                            cond_expr->clone(mem_ctx, NULL),
198                                            assign->rhs);
199           } else {
200              assign->condition = cond_expr->clone(mem_ctx, NULL);
201 	       }
202 	    } else {
203 	       assign->condition =
204              new(mem_ctx) ir_expression(ir_binop_logic_and,
205                                         glsl_type::bool_type,
206                                         cond_expr->clone(mem_ctx, NULL),
207                                         assign->condition);
208 	    }
209 	 }
210       }
211 
212       /* Now, move from the if block to the block surrounding it. */
213       ir->remove();
214       if_ir->insert_before(ir);
215    }
216 }
217 
218 ir_visitor_status
visit_enter(ir_if * ir)219 ir_if_to_cond_assign_visitor::visit_enter(ir_if *ir)
220 {
221    (void) ir;
222    this->depth++;
223 
224    return visit_continue;
225 }
226 
227 ir_visitor_status
visit_leave(ir_if * ir)228 ir_if_to_cond_assign_visitor::visit_leave(ir_if *ir)
229 {
230    bool must_lower = this->depth-- > this->max_depth;
231 
232    /* Only flatten when beyond the GPU's maximum supported nesting depth. */
233    if (!must_lower && this->min_branch_cost == 0)
234       return visit_continue;
235 
236    this->found_unsupported_op = false;
237    this->found_expensive_op = false;
238    this->found_dynamic_arrayref = false;
239    this->then_cost = 0;
240    this->else_cost = 0;
241 
242    ir_assignment *assign;
243 
244    /* Check that both blocks don't contain anything we can't support. */
245    this->is_then = true;
246    foreach_in_list(ir_instruction, then_ir, &ir->then_instructions) {
247       visit_tree(then_ir, check_ir_node, this);
248    }
249 
250    this->is_then = false;
251    foreach_in_list(ir_instruction, else_ir, &ir->else_instructions) {
252       visit_tree(else_ir, check_ir_node, this);
253    }
254 
255    if (this->found_unsupported_op)
256       return visit_continue; /* can't handle inner unsupported opcodes */
257 
258    /* Skip if the branch cost is high enough or if there's an expensive op.
259     *
260     * Also skip if non-constant array indices were encountered, since those
261     * can be out-of-bounds for a not-taken branch, and so generating an
262     * assignment would be incorrect. In the case of must_lower, it's up to the
263     * backend to deal with any potential fall-out (perhaps by translating the
264     * assignments to hardware-predicated moves).
265     */
266    if (!must_lower &&
267        (this->found_expensive_op ||
268         this->found_dynamic_arrayref ||
269         MAX2(this->then_cost, this->else_cost) >= this->min_branch_cost))
270       return visit_continue;
271 
272    void *mem_ctx = ralloc_parent(ir);
273 
274    /* Store the condition to a variable.  Move all of the instructions from
275     * the then-clause of the if-statement.  Use the condition variable as a
276     * condition for all assignments.
277     */
278    ir_variable *const then_var =
279       new(mem_ctx) ir_variable(glsl_type::bool_type,
280 			       "if_to_cond_assign_then",
281 			       ir_var_temporary);
282    ir->insert_before(then_var);
283 
284    ir_dereference_variable *then_cond =
285       new(mem_ctx) ir_dereference_variable(then_var);
286 
287    assign = new(mem_ctx) ir_assignment(then_cond, ir->condition);
288    ir->insert_before(assign);
289 
290    move_block_to_cond_assign(mem_ctx, ir, then_cond,
291 			     &ir->then_instructions,
292 			     this->condition_variables);
293 
294    /* Add the new condition variable to the hash table.  This allows us to
295     * find this variable when lowering other (enclosing) if-statements.
296     */
297    _mesa_set_add(this->condition_variables, then_var);
298 
299    /* If there are instructions in the else-clause, store the inverse of the
300     * condition to a variable.  Move all of the instructions from the
301     * else-clause if the if-statement.  Use the (inverse) condition variable
302     * as a condition for all assignments.
303     */
304    if (!ir->else_instructions.is_empty()) {
305       ir_variable *const else_var =
306 	 new(mem_ctx) ir_variable(glsl_type::bool_type,
307 				  "if_to_cond_assign_else",
308 				  ir_var_temporary);
309       ir->insert_before(else_var);
310 
311       ir_dereference_variable *else_cond =
312 	 new(mem_ctx) ir_dereference_variable(else_var);
313 
314       ir_rvalue *inverse =
315 	 new(mem_ctx) ir_expression(ir_unop_logic_not,
316 				    then_cond->clone(mem_ctx, NULL));
317 
318       assign = new(mem_ctx) ir_assignment(else_cond, inverse);
319       ir->insert_before(assign);
320 
321       move_block_to_cond_assign(mem_ctx, ir, else_cond,
322 				&ir->else_instructions,
323 				this->condition_variables);
324 
325       /* Add the new condition variable to the hash table.  This allows us to
326        * find this variable when lowering other (enclosing) if-statements.
327        */
328       _mesa_set_add(this->condition_variables, else_var);
329    }
330 
331    ir->remove();
332 
333    this->progress = true;
334 
335    return visit_continue;
336 }
337