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