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 attempts to flatten if-statements to conditional assignments for
  28  * GPUs with limited or no flow control support.
  29  *
  30  * It can't handle other control flow being inside of its block, such
  31  * as calls or loops.  Hopefully loop unrolling and inlining will take
  32  * care of those.
  33  *
  34  * Drivers for GPUs with no control flow support should simply call
  35  *
  36  *    lower_if_to_cond_assign(instructions)
  37  *
  38  * to attempt to flatten all if-statements.
  39  *
  40  * Some GPUs (such as i965 prior to gen6) do support control flow, but have a
  41  * maximum nesting depth N.  Drivers for such hardware can call
  42  *
  43  *    lower_if_to_cond_assign(instructions, N)
  44  *
  45  * to attempt to flatten any if-statements appearing at depth > N.
  46  */
  47
  48 #include "compiler/glsl_types.h"
  49 #include "ir.h"
  50 #include "util/set.h"
  51 #include "util/hash_table.h" /* Needed for the hashing functions */
  52
  53 namespace {
  54
  55 class ir_if_to_cond_assign_visitor : public ir_hierarchical_visitor {
  56 public:
  57    ir_if_to_cond_assign_visitor(unsigned max_depth)
  58    {
  59       this->progress = false;
  60       this->max_depth = max_depth;
  61       this->depth = 0;
  62
  63       this->condition_variables =
  64             _mesa_set_create(NULL, _mesa_hash_pointer,
  65                                     _mesa_key_pointer_equal);
  66    }
  67
  68    ~ir_if_to_cond_assign_visitor()
  69    {
  70       _mesa_set_destroy(this->condition_variables, NULL);
  71    }
  72
  73    ir_visitor_status visit_enter(ir_if *);
  74    ir_visitor_status visit_leave(ir_if *);
  75
  76    bool found_unsupported_op;
  77    bool progress;
  78    unsigned max_depth;
  79    unsigned depth;
  80
  81    struct set *condition_variables;
  82 };
  83
  84 } /* anonymous namespace */
  85
  86 bool
  87 lower_if_to_cond_assign(exec_list *instructions, unsigned max_depth)
  88 {
  89    if (max_depth == UINT_MAX)
  90       return false;
  91
  92    ir_if_to_cond_assign_visitor v(max_depth);
  93
  94    visit_list_elements(&v, instructions);
  95
  96    return v.progress;
  97 }
  98
  99 void
 100 check_ir_node(ir_instruction *ir, void *data)
 101 {
 102    ir_if_to_cond_assign_visitor *v = (ir_if_to_cond_assign_visitor *)data;
 103
 104    switch (ir->ir_type) {
 105    case ir_type_call:
 106    case ir_type_discard:
 107    case ir_type_loop:
 108    case ir_type_loop_jump:
 109    case ir_type_return:
 110    case ir_type_emit_vertex:
 111    case ir_type_end_primitive:
 112    case ir_type_barrier:
 113       v->found_unsupported_op = true;
 114       break;
 115    default:
 116       break;
 117    }
 118 }
 119
 120 void
 121 move_block_to_cond_assign(void *mem_ctx,
 122                           ir_if *if_ir, ir_rvalue *cond_expr,
 123                           exec_list *instructions,
 124                           struct set *set)
 125 {
 126    foreach_in_list_safe(ir_instruction, ir, instructions) {
 127       if (ir->ir_type == ir_type_assignment) {
 128          ir_assignment *assign = (ir_assignment *)ir;
 129
 130          if (_mesa_set_search(set, assign) == NULL) {
 131             _mesa_set_add(set, assign);
 132
 133             /* If the LHS of the assignment is a condition variable that was
 134              * previously added, insert an additional assignment of false to
 135              * the variable.
 136              */
 137             const bool assign_to_cv =
 138                   _mesa_set_search(
 139                         set, assign->lhs->variable_referenced()) != NULL;
 140
 141             if (!assign->condition) {
 142           if (assign_to_cv) {
 143              assign->rhs =
 144                           new(mem_ctx) ir_expression(ir_binop_logic_and,
 145                                            glsl_type::bool_type,
 146                                            cond_expr->clone(mem_ctx, NULL),
 147                                            assign->rhs);
 148           } else {
 149              assign->condition = cond_expr->clone(mem_ctx, NULL);
 150                }
 151             } else {
 152                assign->condition =
 153              new(mem_ctx) ir_expression(ir_binop_logic_and,
 154                                         glsl_type::bool_type,
 155                                         cond_expr->clone(mem_ctx, NULL),
 156                                         assign->condition);
 157             }
 158          }
 159       }
 160
 161       /* Now, move from the if block to the block surrounding it. */
 162       ir->remove();
 163       if_ir->insert_before(ir);
 164    }
 165 }
 166
 167 ir_visitor_status
 168 ir_if_to_cond_assign_visitor::visit_enter(ir_if *ir)
 169 {
 170    (void) ir;
 171    this->depth++;
 172
 173    return visit_continue;
 174 }
 175
 176 ir_visitor_status
 177 ir_if_to_cond_assign_visitor::visit_leave(ir_if *ir)
 178 {
 179    /* Only flatten when beyond the GPU's maximum supported nesting depth. */
 180    if (this->depth-- <= this->max_depth)
 181       return visit_continue;
 182
 183    this->found_unsupported_op = false;
 184
 185    ir_assignment *assign;
 186
 187    /* Check that both blocks don't contain anything we can't support. */
 188    foreach_in_list(ir_instruction, then_ir, &ir->then_instructions) {
 189       visit_tree(then_ir, check_ir_node, this);
 190    }
 191    foreach_in_list(ir_instruction, else_ir, &ir->else_instructions) {
 192       visit_tree(else_ir, check_ir_node, this);
 193    }
 194    if (this->found_unsupported_op)
 195       return visit_continue; /* can't handle inner unsupported opcodes */
 196
 197    void *mem_ctx = ralloc_parent(ir);
 198
 199    /* Store the condition to a variable.  Move all of the instructions from
 200     * the then-clause of the if-statement.  Use the condition variable as a
 201     * condition for all assignments.
 202     */
 203    ir_variable *const then_var =
 204       new(mem_ctx) ir_variable(glsl_type::bool_type,
 205                                "if_to_cond_assign_then",
 206                                ir_var_temporary);
 207    ir->insert_before(then_var);
 208
 209    ir_dereference_variable *then_cond =
 210       new(mem_ctx) ir_dereference_variable(then_var);
 211
 212    assign = new(mem_ctx) ir_assignment(then_cond, ir->condition);
 213    ir->insert_before(assign);
 214
 215    move_block_to_cond_assign(mem_ctx, ir, then_cond,
 216                              &ir->then_instructions,
 217                              this->condition_variables);
 218
 219    /* Add the new condition variable to the hash table.  This allows us to
 220     * find this variable when lowering other (enclosing) if-statements.
 221     */
 222    _mesa_set_add(this->condition_variables, then_var);
 223
 224    /* If there are instructions in the else-clause, store the inverse of the
 225     * condition to a variable.  Move all of the instructions from the
 226     * else-clause if the if-statement.  Use the (inverse) condition variable
 227     * as a condition for all assignments.
 228     */
 229    if (!ir->else_instructions.is_empty()) {
 230       ir_variable *const else_var =
 231          new(mem_ctx) ir_variable(glsl_type::bool_type,
 232                                   "if_to_cond_assign_else",
 233                                   ir_var_temporary);
 234       ir->insert_before(else_var);
 235
 236       ir_dereference_variable *else_cond =
 237          new(mem_ctx) ir_dereference_variable(else_var);
 238
 239       ir_rvalue *inverse =
 240          new(mem_ctx) ir_expression(ir_unop_logic_not,
 241                                     then_cond->clone(mem_ctx, NULL));
 242
 243       assign = new(mem_ctx) ir_assignment(else_cond, inverse);
 244       ir->insert_before(assign);
 245
 246       move_block_to_cond_assign(mem_ctx, ir, else_cond,
 247                                 &ir->else_instructions,
 248                                 this->condition_variables);
 249
 250       /* Add the new condition variable to the hash table.  This allows us to
 251        * find this variable when lowering other (enclosing) if-statements.
 252        */
 253       _mesa_set_add(this->condition_variables, else_var);
 254    }
 255
 256    ir->remove();
 257
 258    this->progress = true;
 259
 260    return visit_continue;
 261 }