src/compiler/glsl/lower_if_to_cond_assign.cpp

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