1 /* 2 * Copyright 2015 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 * Authors: 24 * Jason Ekstrand (jason (at) jlekstrand.net) 25 * 26 */ 27 28 #include "nir.h" 29 30 /* 31 * Implements a pass that lowers vector phi nodes to scalar phi nodes when 32 * we don't think it will hurt anything. 33 */ 34 35 struct lower_phis_to_scalar_state { 36 void *mem_ctx; 37 void *dead_ctx; 38 39 /* Hash table marking which phi nodes are scalarizable. The key is 40 * pointers to phi instructions and the entry is either NULL for not 41 * scalarizable or non-null for scalarizable. 42 */ 43 struct hash_table *phi_table; 44 }; 45 46 static bool 47 should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state); 48 49 static bool 50 is_phi_src_scalarizable(nir_phi_src *src, 51 struct lower_phis_to_scalar_state *state) 52 { 53 /* Don't know what to do with non-ssa sources */ 54 if (!src->src.is_ssa) 55 return false; 56 57 nir_instr *src_instr = src->src.ssa->parent_instr; 58 switch (src_instr->type) { 59 case nir_instr_type_alu: { 60 nir_alu_instr *src_alu = nir_instr_as_alu(src_instr); 61 62 /* ALU operations with output_size == 0 should be scalarized. We 63 * will also see a bunch of vecN operations from scalarizing ALU 64 * operations and, since they can easily be copy-propagated, they 65 * are ok too. 66 */ 67 return nir_op_infos[src_alu->op].output_size == 0 || 68 src_alu->op == nir_op_vec2 || 69 src_alu->op == nir_op_vec3 || 70 src_alu->op == nir_op_vec4; 71 } 72 73 case nir_instr_type_phi: 74 /* A phi is scalarizable if we're going to lower it */ 75 return should_lower_phi(nir_instr_as_phi(src_instr), state); 76 77 case nir_instr_type_load_const: 78 case nir_instr_type_ssa_undef: 79 /* These are trivially scalarizable */ 80 return true; 81 82 case nir_instr_type_intrinsic: { 83 nir_intrinsic_instr *src_intrin = nir_instr_as_intrinsic(src_instr); 84 85 switch (src_intrin->intrinsic) { 86 case nir_intrinsic_load_var: 87 return src_intrin->variables[0]->var->data.mode == nir_var_shader_in || 88 src_intrin->variables[0]->var->data.mode == nir_var_uniform; 89 90 case nir_intrinsic_interp_var_at_centroid: 91 case nir_intrinsic_interp_var_at_sample: 92 case nir_intrinsic_interp_var_at_offset: 93 case nir_intrinsic_load_uniform: 94 case nir_intrinsic_load_ubo: 95 case nir_intrinsic_load_ssbo: 96 case nir_intrinsic_load_input: 97 return true; 98 default: 99 break; 100 } 101 } 102 103 default: 104 /* We can't scalarize this type of instruction */ 105 return false; 106 } 107 } 108 109 /** 110 * Determines if the given phi node should be lowered. The only phi nodes 111 * we will scalarize at the moment are those where all of the sources are 112 * scalarizable. 113 * 114 * The reason for this comes down to coalescing. Since phi sources can't 115 * swizzle, swizzles on phis have to be resolved by inserting a mov right 116 * before the phi. The choice then becomes between movs to pick off 117 * components for a scalar phi or potentially movs to recombine components 118 * for a vector phi. The problem is that the movs generated to pick off 119 * the components are almost uncoalescable. We can't coalesce them in NIR 120 * because we need them to pick off components and we can't coalesce them 121 * in the backend because the source register is a vector and the 122 * destination is a scalar that may be used at other places in the program. 123 * On the other hand, if we have a bunch of scalars going into a vector 124 * phi, the situation is much better. In this case, if the SSA def is 125 * generated in the predecessor block to the corresponding phi source, the 126 * backend code will be an ALU op into a temporary and then a mov into the 127 * given vector component; this move can almost certainly be coalesced 128 * away. 129 */ 130 static bool 131 should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state) 132 { 133 /* Already scalar */ 134 if (phi->dest.ssa.num_components == 1) 135 return false; 136 137 struct hash_entry *entry = _mesa_hash_table_search(state->phi_table, phi); 138 if (entry) 139 return entry->data != NULL; 140 141 /* Insert an entry and mark it as scalarizable for now. That way 142 * we don't recurse forever and a cycle in the dependence graph 143 * won't automatically make us fail to scalarize. 144 */ 145 entry = _mesa_hash_table_insert(state->phi_table, phi, (void *)(intptr_t)1); 146 147 bool scalarizable = true; 148 149 nir_foreach_phi_src(src, phi) { 150 scalarizable = is_phi_src_scalarizable(src, state); 151 if (!scalarizable) 152 break; 153 } 154 155 /* The hash table entry for 'phi' may have changed while recursing the 156 * dependence graph, so we need to reset it */ 157 entry = _mesa_hash_table_search(state->phi_table, phi); 158 assert(entry); 159 160 entry->data = (void *)(intptr_t)scalarizable; 161 162 return scalarizable; 163 } 164 165 static bool 166 lower_phis_to_scalar_block(nir_block *block, 167 struct lower_phis_to_scalar_state *state) 168 { 169 bool progress = false; 170 171 /* Find the last phi node in the block */ 172 nir_phi_instr *last_phi = NULL; 173 nir_foreach_instr(instr, block) { 174 if (instr->type != nir_instr_type_phi) 175 break; 176 177 last_phi = nir_instr_as_phi(instr); 178 } 179 180 /* We have to handle the phi nodes in their own pass due to the way 181 * we're modifying the linked list of instructions. 182 */ 183 nir_foreach_instr_safe(instr, block) { 184 if (instr->type != nir_instr_type_phi) 185 break; 186 187 nir_phi_instr *phi = nir_instr_as_phi(instr); 188 189 if (!should_lower_phi(phi, state)) 190 continue; 191 192 unsigned bit_size = phi->dest.ssa.bit_size; 193 194 /* Create a vecN operation to combine the results. Most of these 195 * will be redundant, but copy propagation should clean them up for 196 * us. No need to add the complexity here. 197 */ 198 nir_op vec_op; 199 switch (phi->dest.ssa.num_components) { 200 case 2: vec_op = nir_op_vec2; break; 201 case 3: vec_op = nir_op_vec3; break; 202 case 4: vec_op = nir_op_vec4; break; 203 default: unreachable("Invalid number of components"); 204 } 205 206 nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, vec_op); 207 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, 208 phi->dest.ssa.num_components, 209 bit_size, NULL); 210 vec->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1; 211 212 for (unsigned i = 0; i < phi->dest.ssa.num_components; i++) { 213 nir_phi_instr *new_phi = nir_phi_instr_create(state->mem_ctx); 214 nir_ssa_dest_init(&new_phi->instr, &new_phi->dest, 1, 215 phi->dest.ssa.bit_size, NULL); 216 217 vec->src[i].src = nir_src_for_ssa(&new_phi->dest.ssa); 218 219 nir_foreach_phi_src(src, phi) { 220 /* We need to insert a mov to grab the i'th component of src */ 221 nir_alu_instr *mov = nir_alu_instr_create(state->mem_ctx, 222 nir_op_imov); 223 nir_ssa_dest_init(&mov->instr, &mov->dest.dest, 1, bit_size, NULL); 224 mov->dest.write_mask = 1; 225 nir_src_copy(&mov->src[0].src, &src->src, state->mem_ctx); 226 mov->src[0].swizzle[0] = i; 227 228 /* Insert at the end of the predecessor but before the jump */ 229 nir_instr *pred_last_instr = nir_block_last_instr(src->pred); 230 if (pred_last_instr && pred_last_instr->type == nir_instr_type_jump) 231 nir_instr_insert_before(pred_last_instr, &mov->instr); 232 else 233 nir_instr_insert_after_block(src->pred, &mov->instr); 234 235 nir_phi_src *new_src = ralloc(new_phi, nir_phi_src); 236 new_src->pred = src->pred; 237 new_src->src = nir_src_for_ssa(&mov->dest.dest.ssa); 238 239 exec_list_push_tail(&new_phi->srcs, &new_src->node); 240 } 241 242 nir_instr_insert_before(&phi->instr, &new_phi->instr); 243 } 244 245 nir_instr_insert_after(&last_phi->instr, &vec->instr); 246 247 nir_ssa_def_rewrite_uses(&phi->dest.ssa, 248 nir_src_for_ssa(&vec->dest.dest.ssa)); 249 250 ralloc_steal(state->dead_ctx, phi); 251 nir_instr_remove(&phi->instr); 252 253 progress = true; 254 255 /* We're using the safe iterator and inserting all the newly 256 * scalarized phi nodes before their non-scalarized version so that's 257 * ok. However, we are also inserting vec operations after all of 258 * the last phi node so once we get here, we can't trust even the 259 * safe iterator to stop properly. We have to break manually. 260 */ 261 if (instr == &last_phi->instr) 262 break; 263 } 264 265 return progress; 266 } 267 268 static bool 269 lower_phis_to_scalar_impl(nir_function_impl *impl) 270 { 271 struct lower_phis_to_scalar_state state; 272 bool progress = false; 273 274 state.mem_ctx = ralloc_parent(impl); 275 state.dead_ctx = ralloc_context(NULL); 276 state.phi_table = _mesa_hash_table_create(state.dead_ctx, _mesa_hash_pointer, 277 _mesa_key_pointer_equal); 278 279 nir_foreach_block(block, impl) { 280 progress = lower_phis_to_scalar_block(block, &state) || progress; 281 } 282 283 nir_metadata_preserve(impl, nir_metadata_block_index | 284 nir_metadata_dominance); 285 286 ralloc_free(state.dead_ctx); 287 return progress; 288 } 289 290 /** A pass that lowers vector phi nodes to scalar 291 * 292 * This pass loops through the blocks and lowers looks for vector phi nodes 293 * it can lower to scalar phi nodes. Not all phi nodes are lowered. For 294 * instance, if one of the sources is a non-scalarizable vector, then we 295 * don't bother lowering because that would generate hard-to-coalesce movs. 296 */ 297 bool 298 nir_lower_phis_to_scalar(nir_shader *shader) 299 { 300 bool progress = false; 301 302 nir_foreach_function(function, shader) { 303 if (function->impl) 304 progress = lower_phis_to_scalar_impl(function->impl) || progress; 305 } 306 307 return progress; 308 } 309