1f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/* 2f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Copyright © 2010 Intel Corporation 3f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 4f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Permission is hereby granted, free of charge, to any person obtaining a 5f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * copy of this software and associated documentation files (the "Software"), 6f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * to deal in the Software without restriction, including without limitation 7f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * and/or sell copies of the Software, and to permit persons to whom the 9f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Software is furnished to do so, subject to the following conditions: 10f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 11f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * The above copyright notice and this permission notice (including the next 12f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * paragraph) shall be included in all copies or substantial portions of the 13f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Software. 14f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 15f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 21f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * DEALINGS IN THE SOFTWARE. 22f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 23f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 24f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/** 25f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * \file opt_tree_grafting.cpp 26f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 27f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Takes assignments to variables that are dereferenced only once and 28f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * pastes the RHS expression into where the variable is dereferenced. 29f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 30f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * In the process of various operations like function inlining and 31f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * tertiary op handling, we'll end up with our expression trees having 32f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * been chopped up into a series of assignments of short expressions 33f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * to temps. Other passes like ir_algebraic.cpp would prefer to see 34f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * the deepest expression trees they can to try to optimize them. 35f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 36f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * This is a lot like copy propagaton. In comparison, copy 37f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * propagation only acts on plain copies, not arbitrary expressions on 38f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * the RHS. Generally, we wouldn't want to go pasting some 39f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * complicated expression everywhere it got used, though, so we don't 40f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * handle expressions in that pass. 41f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 42f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * The hard part is making sure we don't move an expression across 43f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * some other assignments that would change the value of the 44f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * expression. So we split this into two passes: First, find the 45f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * variables in our scope which are written to once and read once, and 46f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * then go through basic blocks seeing if we find an opportunity to 47f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * move those expressions safely. 48f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 49f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 50f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "ir.h" 51f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "ir_visitor.h" 52f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "ir_variable_refcount.h" 53f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "ir_basic_block.h" 54f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "ir_optimization.h" 55f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "glsl_types.h" 56f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 57f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgnamespace { 58f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 59f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic bool debug = false; 60f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 61f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgclass ir_tree_grafting_visitor : public ir_hierarchical_visitor { 62f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgpublic: 63f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_tree_grafting_visitor(ir_assignment *graft_assign, 64f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *graft_var) 65f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org { 66f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->progress = false; 67f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->graft_assign = graft_assign; 68f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->graft_var = graft_var; 69f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 70f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 71f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_leave(class ir_assignment *); 72f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_call *); 73f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_expression *); 74f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_function *); 75f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_function_signature *); 76f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_if *); 77f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_loop *); 78f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_swizzle *); 79f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org virtual ir_visitor_status visit_enter(class ir_texture *); 80f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 81f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_visitor_status check_graft(ir_instruction *ir, ir_variable *var); 82f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 83f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool do_graft(ir_rvalue **rvalue); 84f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 85f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool progress; 86f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *graft_var; 87f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_assignment *graft_assign; 88f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 89f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 90f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstruct find_deref_info { 91f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *var; 92f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool found; 93f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 94f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 95f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgvoid 96f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgdereferences_variable_callback(ir_instruction *ir, void *data) 97f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 98f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org struct find_deref_info *info = (struct find_deref_info *)data; 99f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_dereference_variable *deref = ir->as_dereference_variable(); 100f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 101f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (deref && deref->var == info->var) 102f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info->found = true; 103f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 104f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 105f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic bool 106f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgdereferences_variable(ir_instruction *ir, ir_variable *var) 107f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 108f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org struct find_deref_info info; 109f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 110f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info.var = var; 111f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info.found = false; 112f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 113f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org visit_tree(ir, dereferences_variable_callback, &info); 114f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 115f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return info.found; 116f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 117f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 118f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 119f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::do_graft(ir_rvalue **rvalue) 120f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 121f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!*rvalue) 122f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 123f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 124f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_dereference_variable *deref = (*rvalue)->as_dereference_variable(); 125f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 126f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!deref || deref->var != this->graft_var) 127f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 128f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 129f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (debug) { 130f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("GRAFTING:\n"); 131f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->graft_assign->print(); 132f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("\n"); 133f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("TO:\n"); 134f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (*rvalue)->print(); 135f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("\n"); 136f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 137f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 138f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->graft_assign->remove(); 139f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org *rvalue = this->graft_assign->rhs; 140f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 141f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org this->progress = true; 142f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 143f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 144f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 145f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 146f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_loop *ir) 147f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 148f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (void)ir; 149f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* Do not traverse into the body of the loop since that is a 150f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * different basic block. 151f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 152f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 153f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 154f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 155f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/** 156f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Check if we can continue grafting after writing to a variable. If the 157f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * expression we're trying to graft references the variable, we must stop. 158f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 159f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * \param ir An instruction that writes to a variable. 160f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * \param var The variable being updated. 161f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 162f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 163f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::check_graft(ir_instruction *ir, ir_variable *var) 164f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 165f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (dereferences_variable(this->graft_assign->rhs, var)) { 166f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (debug) { 167f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("graft killed by: "); 168f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir->print(); 169f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("\n"); 170f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 171f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 172f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 173f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 174f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue; 175f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 176f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 177f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 178f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_leave(ir_assignment *ir) 179f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 180f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->rhs) || 181f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org do_graft(&ir->condition)) 182f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 183f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 184f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* If this assignment updates a variable used in the assignment 185f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * we're trying to graft, then we're done. 186f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 187f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return check_graft(ir, ir->lhs->variable_referenced()); 188f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 189f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 190f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 191f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_function *ir) 192f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 193f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (void) ir; 194f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue_with_parent; 195f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 196f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 197f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 198f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_function_signature *ir) 199f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 200f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (void)ir; 201f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue_with_parent; 202f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 203f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 204f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 205f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_call *ir) 206f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 207f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org exec_list_iterator sig_iter = ir->callee->parameters.iterator(); 208f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* Reminder: iterating ir_call iterates its parameters. */ 209f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org foreach_iter(exec_list_iterator, iter, *ir) { 210f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *sig_param = (ir_variable *)sig_iter.get(); 211f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_rvalue *ir = (ir_rvalue *)iter.get(); 212f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_rvalue *new_ir = ir; 213f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 214f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (sig_param->mode != ir_var_in && sig_param->mode != ir_var_const_in) { 215f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (check_graft(ir, sig_param) == visit_stop) 216f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 217f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org continue; 218f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 219f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 220f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&new_ir)) { 221f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir->replace_with(new_ir); 222f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 223f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 224f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org sig_iter.next(); 225f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 226f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 227f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (ir->return_deref && check_graft(ir, ir->return_deref->var) == visit_stop) 228f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 229f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 230f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue; 231f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 232f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 233f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 234f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_expression *ir) 235f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 236f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (unsigned int i = 0; i < ir->get_num_operands(); i++) { 237f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->operands[i])) 238f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 239f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 240f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 241f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue; 242f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 243f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 244f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 245f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_if *ir) 246f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 247f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->condition)) 248f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 249f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 250f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* Do not traverse into the body of the if-statement since that is a 251f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * different basic block. 252f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 253f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue_with_parent; 254f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 255f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 256f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 257f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_swizzle *ir) 258f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 259f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->val)) 260f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 261f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 262f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue; 263f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 264f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 265f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_visitor_status 266f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgir_tree_grafting_visitor::visit_enter(ir_texture *ir) 267f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 268f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->coordinate) || 269f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org do_graft(&ir->projector) || 270f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org do_graft(&ir->offset) || 271f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org do_graft(&ir->shadow_comparitor)) 272f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 273f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 274f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org switch (ir->op) { 275f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_tex: 276f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 277f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_txb: 278f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->lod_info.bias)) 279f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 280f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 281f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_txf: 282f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_txl: 283f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_txs: 284f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->lod_info.lod)) 285f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 286f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 287f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org case ir_txd: 288f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (do_graft(&ir->lod_info.grad.dPdx) || 289f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org do_graft(&ir->lod_info.grad.dPdy)) 290f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_stop; 291f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org break; 292f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 293f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 294f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return visit_continue; 295f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 296f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 297f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstruct tree_grafting_info { 298f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable_refcount_visitor *refs; 299f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool progress; 300f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 301f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 302f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic bool 303f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgtry_tree_grafting(ir_assignment *start, 304f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *lhs_var, 305f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_instruction *bb_last) 306f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 307f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_tree_grafting_visitor v(start, lhs_var); 308f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 309f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (debug) { 310f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("trying to graft: "); 311f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org lhs_var->print(); 312f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("\n"); 313f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 314f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 315f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (ir_instruction *ir = (ir_instruction *)start->next; 316f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir != bb_last->next; 317f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir = (ir_instruction *)ir->next) { 318f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 319f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (debug) { 320f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("- "); 321f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir->print(); 322f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org printf("\n"); 323f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 324f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 325f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_visitor_status s = ir->accept(&v); 326f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (s == visit_stop) 327f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return v.progress; 328f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 329f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 330f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 331f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 332f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 333f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstatic void 334f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgtree_grafting_basic_block(ir_instruction *bb_first, 335f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_instruction *bb_last, 336f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org void *data) 337f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 338f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org struct tree_grafting_info *info = (struct tree_grafting_info *)data; 339f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_instruction *ir, *next; 340f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 341f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (ir = bb_first, next = (ir_instruction *)ir->next; 342f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir != bb_last->next; 343f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir = next, next = (ir_instruction *)ir->next) { 344f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_assignment *assign = ir->as_assignment(); 345f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 346f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!assign) 347f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org continue; 348f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 349f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable *lhs_var = assign->whole_variable_written(); 350f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!lhs_var) 351f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org continue; 352f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 353f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (lhs_var->mode == ir_var_out || 354f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org lhs_var->mode == ir_var_inout) 355f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org continue; 356f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 357f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable_refcount_entry *entry = info->refs->get_variable_entry(lhs_var); 358f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 359f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!entry->declaration || 360f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->assigned_count != 1 || 361f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->referenced_count != 2) 362f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org continue; 363f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 364f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org assert(assign == entry->assign); 365f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 366f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* Found a possibly graftable assignment. Now, walk through the 367f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * rest of the BB seeing if the deref is here, and if nothing interfered with 368f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * pasting its expression's values in between. 369f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 370f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info->progress |= try_tree_grafting(assign, lhs_var, bb_last); 371f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 372f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 373f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 374f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} /* unnamed namespace */ 375f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 376f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/** 377f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Does a copy propagation pass on the code present in the instruction stream. 378f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 379f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 380f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgdo_tree_grafting(exec_list *instructions) 381f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 382f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ir_variable_refcount_visitor refs; 383f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org struct tree_grafting_info info; 384f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 385f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info.progress = false; 386f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org info.refs = &refs; 387f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 388f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org visit_list_elements(info.refs, instructions); 389f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 390f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org call_for_basic_blocks(instructions, tree_grafting_basic_block, &info); 391f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 392f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return info.progress; 393f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 394