1 /* 2 * Copyright 2016 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 DEALINGS 21 * IN THE SOFTWARE. 22 */ 23 24 #include "nir.h" 25 #include "nir_builder.h" 26 27 static void 28 emit_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr, 29 nir_deref_var *deref, nir_deref *tail, 30 nir_ssa_def **dest, nir_ssa_def *src); 31 32 static void 33 emit_indirect_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr, 34 nir_deref_var *deref, nir_deref *arr_parent, 35 int start, int end, 36 nir_ssa_def **dest, nir_ssa_def *src) 37 { 38 nir_deref_array *arr = nir_deref_as_array(arr_parent->child); 39 assert(arr->deref_array_type == nir_deref_array_type_indirect); 40 assert(arr->indirect.is_ssa); 41 42 assert(start < end); 43 if (start == end - 1) { 44 /* Base case. Just emit the load/store op */ 45 nir_deref_array direct = *arr; 46 direct.deref_array_type = nir_deref_array_type_direct; 47 direct.base_offset += start; 48 direct.indirect = NIR_SRC_INIT; 49 50 arr_parent->child = &direct.deref; 51 emit_load_store(b, orig_instr, deref, &direct.deref, dest, src); 52 arr_parent->child = &arr->deref; 53 } else { 54 int mid = start + (end - start) / 2; 55 56 nir_ssa_def *then_dest, *else_dest; 57 58 nir_if *if_stmt = nir_if_create(b->shader); 59 if_stmt->condition = nir_src_for_ssa(nir_ilt(b, arr->indirect.ssa, 60 nir_imm_int(b, mid))); 61 nir_cf_node_insert(b->cursor, &if_stmt->cf_node); 62 63 b->cursor = nir_after_cf_list(&if_stmt->then_list); 64 emit_indirect_load_store(b, orig_instr, deref, arr_parent, 65 start, mid, &then_dest, src); 66 67 b->cursor = nir_after_cf_list(&if_stmt->else_list); 68 emit_indirect_load_store(b, orig_instr, deref, arr_parent, 69 mid, end, &else_dest, src); 70 71 b->cursor = nir_after_cf_node(&if_stmt->cf_node); 72 73 if (src == NULL) { 74 /* We're a load. We need to insert a phi node */ 75 nir_phi_instr *phi = nir_phi_instr_create(b->shader); 76 unsigned bit_size = then_dest->bit_size; 77 nir_ssa_dest_init(&phi->instr, &phi->dest, 78 then_dest->num_components, bit_size, NULL); 79 80 nir_phi_src *src0 = ralloc(phi, nir_phi_src); 81 src0->pred = nir_if_last_then_block(if_stmt); 82 src0->src = nir_src_for_ssa(then_dest); 83 exec_list_push_tail(&phi->srcs, &src0->node); 84 85 nir_phi_src *src1 = ralloc(phi, nir_phi_src); 86 src1->pred = nir_if_last_else_block(if_stmt); 87 src1->src = nir_src_for_ssa(else_dest); 88 exec_list_push_tail(&phi->srcs, &src1->node); 89 90 nir_builder_instr_insert(b, &phi->instr); 91 *dest = &phi->dest.ssa; 92 } 93 } 94 } 95 96 static void 97 emit_load_store(nir_builder *b, nir_intrinsic_instr *orig_instr, 98 nir_deref_var *deref, nir_deref *tail, 99 nir_ssa_def **dest, nir_ssa_def *src) 100 { 101 for (; tail->child; tail = tail->child) { 102 if (tail->child->deref_type != nir_deref_type_array) 103 continue; 104 105 nir_deref_array *arr = nir_deref_as_array(tail->child); 106 if (arr->deref_array_type != nir_deref_array_type_indirect) 107 continue; 108 109 int length = glsl_get_length(tail->type); 110 111 emit_indirect_load_store(b, orig_instr, deref, tail, -arr->base_offset, 112 length - arr->base_offset, dest, src); 113 return; 114 } 115 116 assert(tail && tail->child == NULL); 117 118 /* We reached the end of the deref chain. Emit the instruction */ 119 120 if (src == NULL) { 121 /* This is a load instruction */ 122 nir_intrinsic_instr *load = 123 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var); 124 load->num_components = orig_instr->num_components; 125 load->variables[0] = nir_deref_var_clone(deref, load); 126 unsigned bit_size = orig_instr->dest.ssa.bit_size; 127 nir_ssa_dest_init(&load->instr, &load->dest, 128 load->num_components, bit_size, NULL); 129 nir_builder_instr_insert(b, &load->instr); 130 *dest = &load->dest.ssa; 131 } else { 132 /* This is a store instruction */ 133 nir_intrinsic_instr *store = 134 nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var); 135 store->num_components = orig_instr->num_components; 136 nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(orig_instr)); 137 store->variables[0] = nir_deref_var_clone(deref, store); 138 store->src[0] = nir_src_for_ssa(src); 139 nir_builder_instr_insert(b, &store->instr); 140 } 141 } 142 143 static bool 144 deref_has_indirect(nir_deref_var *deref) 145 { 146 for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) { 147 if (tail->deref_type != nir_deref_type_array) 148 continue; 149 150 nir_deref_array *arr = nir_deref_as_array(tail); 151 if (arr->deref_array_type == nir_deref_array_type_indirect) 152 return true; 153 } 154 155 return false; 156 } 157 158 static bool 159 lower_indirect_block(nir_block *block, nir_builder *b, 160 nir_variable_mode modes) 161 { 162 bool progress = false; 163 164 nir_foreach_instr_safe(instr, block) { 165 if (instr->type != nir_instr_type_intrinsic) 166 continue; 167 168 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); 169 if (intrin->intrinsic != nir_intrinsic_load_var && 170 intrin->intrinsic != nir_intrinsic_store_var) 171 continue; 172 173 if (!deref_has_indirect(intrin->variables[0])) 174 continue; 175 176 /* Only lower variables whose mode is in the mask, or compact 177 * array variables. (We can't handle indirects on tightly packed 178 * scalar arrays, so we need to lower them regardless.) 179 */ 180 if (!(modes & intrin->variables[0]->var->data.mode) && 181 !intrin->variables[0]->var->data.compact) 182 continue; 183 184 b->cursor = nir_before_instr(&intrin->instr); 185 186 if (intrin->intrinsic == nir_intrinsic_load_var) { 187 nir_ssa_def *result; 188 emit_load_store(b, intrin, intrin->variables[0], 189 &intrin->variables[0]->deref, &result, NULL); 190 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(result)); 191 } else { 192 assert(intrin->src[0].is_ssa); 193 emit_load_store(b, intrin, intrin->variables[0], 194 &intrin->variables[0]->deref, NULL, intrin->src[0].ssa); 195 } 196 nir_instr_remove(&intrin->instr); 197 progress = true; 198 } 199 200 return progress; 201 } 202 203 static bool 204 lower_indirects_impl(nir_function_impl *impl, nir_variable_mode modes) 205 { 206 nir_builder builder; 207 nir_builder_init(&builder, impl); 208 bool progress = false; 209 210 nir_foreach_block_safe(block, impl) { 211 progress |= lower_indirect_block(block, &builder, modes); 212 } 213 214 if (progress) 215 nir_metadata_preserve(impl, nir_metadata_none); 216 217 return progress; 218 } 219 220 /** Lowers indirect variable loads/stores to direct loads/stores. 221 * 222 * The pass works by replacing any indirect load or store with an if-ladder 223 * that does a binary search on the array index. 224 */ 225 bool 226 nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes) 227 { 228 bool progress = false; 229 230 nir_foreach_function(function, shader) { 231 if (function->impl) 232 progress = lower_indirects_impl(function->impl, modes) || progress; 233 } 234 235 return progress; 236 } 237