1f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/* 2f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Copyright © 2012 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 DEALINGS 21f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * IN THE SOFTWARE. 22f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 23f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 24f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "brw_fs.h" 25f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org#include "brw_fs_cfg.h" 26f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 27f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgnamespace { /* avoid conflict with opt_copy_propagation_elements */ 28f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgstruct acp_entry : public exec_node { 29f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org fs_reg dst; 30f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org fs_reg src; 31f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org}; 32f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 33f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 34f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 35f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgfs_visitor::try_copy_propagate(fs_inst *inst, int arg, acp_entry *entry) 36f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 37f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (inst->src[arg].file != entry->dst.file || 38f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].reg != entry->dst.reg || 39f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].reg_offset != entry->dst.reg_offset) { 40f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 41f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 42f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 43f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* See resolve_ud_negate() and comment in brw_fs_emit.cpp. */ 44f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (inst->conditional_mod && 45f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].type == BRW_REGISTER_TYPE_UD && 46f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->src.negate) 47f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 48f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 49f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool has_source_modifiers = entry->src.abs || entry->src.negate; 50f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 51f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (intel->gen == 6 && inst->is_math() && 52f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (has_source_modifiers || entry->src.file == UNIFORM)) 53f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return false; 54f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 55f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].file = entry->src.file; 56f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].reg = entry->src.reg; 57f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].reg_offset = entry->src.reg_offset; 58f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 59f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (!inst->src[arg].abs) { 60f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].abs = entry->src.abs; 61f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[arg].negate ^= entry->src.negate; 62f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 63f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 64f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return true; 65f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 66f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 67f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/** @file brw_fs_copy_propagation.cpp 68f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 69f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * Support for local copy propagation by walking the list of instructions 70f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * and maintaining the ACP table of available copies for propagation. 71f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 72f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * See Muchnik's Advanced Compiler Design and Implementation, section 73f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * 12.5 (p356). 74f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 75f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 76f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org/* Walks a basic block and does copy propagation on it using the acp 77f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org * list. 78f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org */ 79f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 80f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgfs_visitor::opt_copy_propagate_local(void *mem_ctx, 81f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org fs_bblock *block, exec_list *acp) 82f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 83f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool progress = false; 84f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 85f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (fs_inst *inst = block->start; 86f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst != block->end->next; 87f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst = (fs_inst *)inst->next) { 88f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 89f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* Try propagating into this instruction. */ 90f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org foreach_list(entry_node, acp) { 91f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org acp_entry *entry = (acp_entry *)entry_node; 92f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 93f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int i = 0; i < 3; i++) { 94f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (try_copy_propagate(inst, i, entry)) 95f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org progress = true; 96f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 97f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 98f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 99f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* kill the destination from the ACP */ 100f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (inst->dst.file == GRF) { 101f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org foreach_list_safe(entry_node, acp) { 102f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org acp_entry *entry = (acp_entry *)entry_node; 103f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 104f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (inst->overwrites_reg(entry->dst) || 105f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->overwrites_reg(entry->src)) { 106f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->remove(); 107f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 108f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 109f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 110f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 111f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org /* If this instruction is a raw copy, add it to the ACP. */ 112f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (inst->opcode == BRW_OPCODE_MOV && 113f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->dst.file == GRF && 114f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ((inst->src[0].file == GRF && 115f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org (inst->src[0].reg != inst->dst.reg || 116f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[0].reg_offset != inst->dst.reg_offset)) || 117f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[0].file == UNIFORM) && 118f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[0].type == inst->dst.type && 119f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org !inst->saturate && 120f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org !inst->predicated && 121f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org !inst->force_uncompressed && 122f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org !inst->force_sechalf && 123f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org inst->src[0].smear == -1) { 124f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org acp_entry *entry = ralloc(mem_ctx, acp_entry); 125f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->dst = inst->dst; 126f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org entry->src = inst->src[0]; 127f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org acp->push_tail(entry); 128f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 129f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 130f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 131f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return progress; 132f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 133f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 134f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgbool 135f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.orgfs_visitor::opt_copy_propagate() 136f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org{ 137f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org bool progress = false; 138f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org void *mem_ctx = ralloc_context(this->mem_ctx); 139f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 140f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org fs_cfg cfg(this); 141f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 142f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org for (int b = 0; b < cfg.num_blocks; b++) { 143f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org fs_bblock *block = cfg.blocks[b]; 144f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org exec_list acp; 145f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 146f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org progress = opt_copy_propagate_local(mem_ctx, block, &acp) || progress; 147f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org } 148f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 149f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org ralloc_free(mem_ctx); 150f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 151f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org if (progress) 152f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org live_intervals_valid = false; 153f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org 154f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org return progress; 155f2ba7591b1407a7ee9209f842c50696914dc2dedkbr@chromium.org} 156