1b8e80941Smrg/* 2b8e80941Smrg * Copyright © 2015 Red Hat 3b8e80941Smrg * 4b8e80941Smrg * Permission is hereby granted, free of charge, to any person obtaining a 5b8e80941Smrg * copy of this software and associated documentation files (the "Software"), 6b8e80941Smrg * to deal in the Software without restriction, including without limitation 7b8e80941Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8b8e80941Smrg * and/or sell copies of the Software, and to permit persons to whom the 9b8e80941Smrg * Software is furnished to do so, subject to the following conditions: 10b8e80941Smrg * 11b8e80941Smrg * The above copyright notice and this permission notice (including the next 12b8e80941Smrg * paragraph) shall be included in all copies or substantial portions of the 13b8e80941Smrg * Software. 14b8e80941Smrg * 15b8e80941Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16b8e80941Smrg * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17b8e80941Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18b8e80941Smrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19b8e80941Smrg * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20b8e80941Smrg * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 21b8e80941Smrg * IN THE SOFTWARE. 22b8e80941Smrg */ 23b8e80941Smrg 24b8e80941Smrg#include "nir.h" 25b8e80941Smrg#include "nir_control_flow.h" 26b8e80941Smrg 27b8e80941Smrg/* Secret Decoder Ring: 28b8e80941Smrg * clone_foo(): 29b8e80941Smrg * Allocate and clone a foo. 30b8e80941Smrg * __clone_foo(): 31b8e80941Smrg * Clone body of foo (ie. parent class, embedded struct, etc) 32b8e80941Smrg */ 33b8e80941Smrg 34b8e80941Smrgtypedef struct { 35b8e80941Smrg /* True if we are cloning an entire shader. */ 36b8e80941Smrg bool global_clone; 37b8e80941Smrg 38b8e80941Smrg /* If true allows the clone operation to fall back to the original pointer 39b8e80941Smrg * if no clone pointer is found in the remap table. This allows us to 40b8e80941Smrg * clone a loop body without having to add srcs from outside the loop to 41b8e80941Smrg * the remap table. This is useful for loop unrolling. 42b8e80941Smrg */ 43b8e80941Smrg bool allow_remap_fallback; 44b8e80941Smrg 45b8e80941Smrg /* maps orig ptr -> cloned ptr: */ 46b8e80941Smrg struct hash_table *remap_table; 47b8e80941Smrg 48b8e80941Smrg /* List of phi sources. */ 49b8e80941Smrg struct list_head phi_srcs; 50b8e80941Smrg 51b8e80941Smrg /* new shader object, used as memctx for just about everything else: */ 52b8e80941Smrg nir_shader *ns; 53b8e80941Smrg} clone_state; 54b8e80941Smrg 55b8e80941Smrgstatic void 56b8e80941Smrginit_clone_state(clone_state *state, struct hash_table *remap_table, 57b8e80941Smrg bool global, bool allow_remap_fallback) 58b8e80941Smrg{ 59b8e80941Smrg state->global_clone = global; 60b8e80941Smrg state->allow_remap_fallback = allow_remap_fallback; 61b8e80941Smrg 62b8e80941Smrg if (remap_table) { 63b8e80941Smrg state->remap_table = remap_table; 64b8e80941Smrg } else { 65b8e80941Smrg state->remap_table = _mesa_pointer_hash_table_create(NULL); 66b8e80941Smrg } 67b8e80941Smrg 68b8e80941Smrg list_inithead(&state->phi_srcs); 69b8e80941Smrg} 70b8e80941Smrg 71b8e80941Smrgstatic void 72b8e80941Smrgfree_clone_state(clone_state *state) 73b8e80941Smrg{ 74b8e80941Smrg _mesa_hash_table_destroy(state->remap_table, NULL); 75b8e80941Smrg} 76b8e80941Smrg 77b8e80941Smrgstatic inline void * 78b8e80941Smrg_lookup_ptr(clone_state *state, const void *ptr, bool global) 79b8e80941Smrg{ 80b8e80941Smrg struct hash_entry *entry; 81b8e80941Smrg 82b8e80941Smrg if (!ptr) 83b8e80941Smrg return NULL; 84b8e80941Smrg 85b8e80941Smrg if (!state->global_clone && global) 86b8e80941Smrg return (void *)ptr; 87b8e80941Smrg 88b8e80941Smrg entry = _mesa_hash_table_search(state->remap_table, ptr); 89b8e80941Smrg if (!entry) { 90b8e80941Smrg assert(state->allow_remap_fallback); 91b8e80941Smrg return (void *)ptr; 92b8e80941Smrg } 93b8e80941Smrg 94b8e80941Smrg return entry->data; 95b8e80941Smrg} 96b8e80941Smrg 97b8e80941Smrgstatic void 98b8e80941Smrgadd_remap(clone_state *state, void *nptr, const void *ptr) 99b8e80941Smrg{ 100b8e80941Smrg _mesa_hash_table_insert(state->remap_table, ptr, nptr); 101b8e80941Smrg} 102b8e80941Smrg 103b8e80941Smrgstatic void * 104b8e80941Smrgremap_local(clone_state *state, const void *ptr) 105b8e80941Smrg{ 106b8e80941Smrg return _lookup_ptr(state, ptr, false); 107b8e80941Smrg} 108b8e80941Smrg 109b8e80941Smrgstatic void * 110b8e80941Smrgremap_global(clone_state *state, const void *ptr) 111b8e80941Smrg{ 112b8e80941Smrg return _lookup_ptr(state, ptr, true); 113b8e80941Smrg} 114b8e80941Smrg 115b8e80941Smrgstatic nir_register * 116b8e80941Smrgremap_reg(clone_state *state, const nir_register *reg) 117b8e80941Smrg{ 118b8e80941Smrg return _lookup_ptr(state, reg, false); 119b8e80941Smrg} 120b8e80941Smrg 121b8e80941Smrgstatic nir_variable * 122b8e80941Smrgremap_var(clone_state *state, const nir_variable *var) 123b8e80941Smrg{ 124b8e80941Smrg return _lookup_ptr(state, var, nir_variable_is_global(var)); 125b8e80941Smrg} 126b8e80941Smrg 127b8e80941Smrgnir_constant * 128b8e80941Smrgnir_constant_clone(const nir_constant *c, nir_variable *nvar) 129b8e80941Smrg{ 130b8e80941Smrg nir_constant *nc = ralloc(nvar, nir_constant); 131b8e80941Smrg 132b8e80941Smrg memcpy(nc->values, c->values, sizeof(nc->values)); 133b8e80941Smrg nc->num_elements = c->num_elements; 134b8e80941Smrg nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements); 135b8e80941Smrg for (unsigned i = 0; i < c->num_elements; i++) { 136b8e80941Smrg nc->elements[i] = nir_constant_clone(c->elements[i], nvar); 137b8e80941Smrg } 138b8e80941Smrg 139b8e80941Smrg return nc; 140b8e80941Smrg} 141b8e80941Smrg 142b8e80941Smrg/* NOTE: for cloning nir_variables, bypass nir_variable_create to avoid 143b8e80941Smrg * having to deal with locals and globals separately: 144b8e80941Smrg */ 145b8e80941Smrgnir_variable * 146b8e80941Smrgnir_variable_clone(const nir_variable *var, nir_shader *shader) 147b8e80941Smrg{ 148b8e80941Smrg nir_variable *nvar = rzalloc(shader, nir_variable); 149b8e80941Smrg 150b8e80941Smrg nvar->type = var->type; 151b8e80941Smrg nvar->name = ralloc_strdup(nvar, var->name); 152b8e80941Smrg nvar->data = var->data; 153b8e80941Smrg nvar->num_state_slots = var->num_state_slots; 154b8e80941Smrg if (var->num_state_slots) { 155b8e80941Smrg nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots); 156b8e80941Smrg memcpy(nvar->state_slots, var->state_slots, 157b8e80941Smrg var->num_state_slots * sizeof(nir_state_slot)); 158b8e80941Smrg } 159b8e80941Smrg if (var->constant_initializer) { 160b8e80941Smrg nvar->constant_initializer = 161b8e80941Smrg nir_constant_clone(var->constant_initializer, nvar); 162b8e80941Smrg } 163b8e80941Smrg nvar->interface_type = var->interface_type; 164b8e80941Smrg 165b8e80941Smrg nvar->num_members = var->num_members; 166b8e80941Smrg if (var->num_members) { 167b8e80941Smrg nvar->members = ralloc_array(nvar, struct nir_variable_data, 168b8e80941Smrg var->num_members); 169b8e80941Smrg memcpy(nvar->members, var->members, 170b8e80941Smrg var->num_members * sizeof(*var->members)); 171b8e80941Smrg } 172b8e80941Smrg 173b8e80941Smrg return nvar; 174b8e80941Smrg} 175b8e80941Smrg 176b8e80941Smrgstatic nir_variable * 177b8e80941Smrgclone_variable(clone_state *state, const nir_variable *var) 178b8e80941Smrg{ 179b8e80941Smrg nir_variable *nvar = nir_variable_clone(var, state->ns); 180b8e80941Smrg add_remap(state, nvar, var); 181b8e80941Smrg 182b8e80941Smrg return nvar; 183b8e80941Smrg} 184b8e80941Smrg 185b8e80941Smrg/* clone list of nir_variable: */ 186b8e80941Smrgstatic void 187b8e80941Smrgclone_var_list(clone_state *state, struct exec_list *dst, 188b8e80941Smrg const struct exec_list *list) 189b8e80941Smrg{ 190b8e80941Smrg exec_list_make_empty(dst); 191b8e80941Smrg foreach_list_typed(nir_variable, var, node, list) { 192b8e80941Smrg nir_variable *nvar = clone_variable(state, var); 193b8e80941Smrg exec_list_push_tail(dst, &nvar->node); 194b8e80941Smrg } 195b8e80941Smrg} 196b8e80941Smrg 197b8e80941Smrg/* NOTE: for cloning nir_registers, bypass nir_global/local_reg_create() 198b8e80941Smrg * to avoid having to deal with locals and globals separately: 199b8e80941Smrg */ 200b8e80941Smrgstatic nir_register * 201b8e80941Smrgclone_register(clone_state *state, const nir_register *reg) 202b8e80941Smrg{ 203b8e80941Smrg nir_register *nreg = rzalloc(state->ns, nir_register); 204b8e80941Smrg add_remap(state, nreg, reg); 205b8e80941Smrg 206b8e80941Smrg nreg->num_components = reg->num_components; 207b8e80941Smrg nreg->bit_size = reg->bit_size; 208b8e80941Smrg nreg->num_array_elems = reg->num_array_elems; 209b8e80941Smrg nreg->index = reg->index; 210b8e80941Smrg nreg->name = ralloc_strdup(nreg, reg->name); 211b8e80941Smrg 212b8e80941Smrg /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */ 213b8e80941Smrg list_inithead(&nreg->uses); 214b8e80941Smrg list_inithead(&nreg->defs); 215b8e80941Smrg list_inithead(&nreg->if_uses); 216b8e80941Smrg 217b8e80941Smrg return nreg; 218b8e80941Smrg} 219b8e80941Smrg 220b8e80941Smrg/* clone list of nir_register: */ 221b8e80941Smrgstatic void 222b8e80941Smrgclone_reg_list(clone_state *state, struct exec_list *dst, 223b8e80941Smrg const struct exec_list *list) 224b8e80941Smrg{ 225b8e80941Smrg exec_list_make_empty(dst); 226b8e80941Smrg foreach_list_typed(nir_register, reg, node, list) { 227b8e80941Smrg nir_register *nreg = clone_register(state, reg); 228b8e80941Smrg exec_list_push_tail(dst, &nreg->node); 229b8e80941Smrg } 230b8e80941Smrg} 231b8e80941Smrg 232b8e80941Smrgstatic void 233b8e80941Smrg__clone_src(clone_state *state, void *ninstr_or_if, 234b8e80941Smrg nir_src *nsrc, const nir_src *src) 235b8e80941Smrg{ 236b8e80941Smrg nsrc->is_ssa = src->is_ssa; 237b8e80941Smrg if (src->is_ssa) { 238b8e80941Smrg nsrc->ssa = remap_local(state, src->ssa); 239b8e80941Smrg } else { 240b8e80941Smrg nsrc->reg.reg = remap_reg(state, src->reg.reg); 241b8e80941Smrg if (src->reg.indirect) { 242b8e80941Smrg nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src); 243b8e80941Smrg __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect); 244b8e80941Smrg } 245b8e80941Smrg nsrc->reg.base_offset = src->reg.base_offset; 246b8e80941Smrg } 247b8e80941Smrg} 248b8e80941Smrg 249b8e80941Smrgstatic void 250b8e80941Smrg__clone_dst(clone_state *state, nir_instr *ninstr, 251b8e80941Smrg nir_dest *ndst, const nir_dest *dst) 252b8e80941Smrg{ 253b8e80941Smrg ndst->is_ssa = dst->is_ssa; 254b8e80941Smrg if (dst->is_ssa) { 255b8e80941Smrg nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components, 256b8e80941Smrg dst->ssa.bit_size, dst->ssa.name); 257b8e80941Smrg add_remap(state, &ndst->ssa, &dst->ssa); 258b8e80941Smrg } else { 259b8e80941Smrg ndst->reg.reg = remap_reg(state, dst->reg.reg); 260b8e80941Smrg if (dst->reg.indirect) { 261b8e80941Smrg ndst->reg.indirect = ralloc(ninstr, nir_src); 262b8e80941Smrg __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect); 263b8e80941Smrg } 264b8e80941Smrg ndst->reg.base_offset = dst->reg.base_offset; 265b8e80941Smrg } 266b8e80941Smrg} 267b8e80941Smrg 268b8e80941Smrgstatic nir_alu_instr * 269b8e80941Smrgclone_alu(clone_state *state, const nir_alu_instr *alu) 270b8e80941Smrg{ 271b8e80941Smrg nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op); 272b8e80941Smrg nalu->exact = alu->exact; 273b8e80941Smrg 274b8e80941Smrg __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest); 275b8e80941Smrg nalu->dest.saturate = alu->dest.saturate; 276b8e80941Smrg nalu->dest.write_mask = alu->dest.write_mask; 277b8e80941Smrg 278b8e80941Smrg for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) { 279b8e80941Smrg __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src); 280b8e80941Smrg nalu->src[i].negate = alu->src[i].negate; 281b8e80941Smrg nalu->src[i].abs = alu->src[i].abs; 282b8e80941Smrg memcpy(nalu->src[i].swizzle, alu->src[i].swizzle, 283b8e80941Smrg sizeof(nalu->src[i].swizzle)); 284b8e80941Smrg } 285b8e80941Smrg 286b8e80941Smrg return nalu; 287b8e80941Smrg} 288b8e80941Smrg 289b8e80941Smrgstatic nir_deref_instr * 290b8e80941Smrgclone_deref_instr(clone_state *state, const nir_deref_instr *deref) 291b8e80941Smrg{ 292b8e80941Smrg nir_deref_instr *nderef = 293b8e80941Smrg nir_deref_instr_create(state->ns, deref->deref_type); 294b8e80941Smrg 295b8e80941Smrg __clone_dst(state, &nderef->instr, &nderef->dest, &deref->dest); 296b8e80941Smrg 297b8e80941Smrg nderef->mode = deref->mode; 298b8e80941Smrg nderef->type = deref->type; 299b8e80941Smrg 300b8e80941Smrg if (deref->deref_type == nir_deref_type_var) { 301b8e80941Smrg nderef->var = remap_var(state, deref->var); 302b8e80941Smrg return nderef; 303b8e80941Smrg } 304b8e80941Smrg 305b8e80941Smrg __clone_src(state, &nderef->instr, &nderef->parent, &deref->parent); 306b8e80941Smrg 307b8e80941Smrg switch (deref->deref_type) { 308b8e80941Smrg case nir_deref_type_struct: 309b8e80941Smrg nderef->strct.index = deref->strct.index; 310b8e80941Smrg break; 311b8e80941Smrg 312b8e80941Smrg case nir_deref_type_array: 313b8e80941Smrg case nir_deref_type_ptr_as_array: 314b8e80941Smrg __clone_src(state, &nderef->instr, 315b8e80941Smrg &nderef->arr.index, &deref->arr.index); 316b8e80941Smrg break; 317b8e80941Smrg 318b8e80941Smrg case nir_deref_type_array_wildcard: 319b8e80941Smrg /* Nothing to do */ 320b8e80941Smrg break; 321b8e80941Smrg 322b8e80941Smrg case nir_deref_type_cast: 323b8e80941Smrg nderef->cast.ptr_stride = deref->cast.ptr_stride; 324b8e80941Smrg break; 325b8e80941Smrg 326b8e80941Smrg default: 327b8e80941Smrg unreachable("Invalid instruction deref type"); 328b8e80941Smrg } 329b8e80941Smrg 330b8e80941Smrg return nderef; 331b8e80941Smrg} 332b8e80941Smrg 333b8e80941Smrgstatic nir_intrinsic_instr * 334b8e80941Smrgclone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr) 335b8e80941Smrg{ 336b8e80941Smrg nir_intrinsic_instr *nitr = 337b8e80941Smrg nir_intrinsic_instr_create(state->ns, itr->intrinsic); 338b8e80941Smrg 339b8e80941Smrg unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs; 340b8e80941Smrg 341b8e80941Smrg if (nir_intrinsic_infos[itr->intrinsic].has_dest) 342b8e80941Smrg __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest); 343b8e80941Smrg 344b8e80941Smrg nitr->num_components = itr->num_components; 345b8e80941Smrg memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index)); 346b8e80941Smrg 347b8e80941Smrg for (unsigned i = 0; i < num_srcs; i++) 348b8e80941Smrg __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]); 349b8e80941Smrg 350b8e80941Smrg return nitr; 351b8e80941Smrg} 352b8e80941Smrg 353b8e80941Smrgstatic nir_load_const_instr * 354b8e80941Smrgclone_load_const(clone_state *state, const nir_load_const_instr *lc) 355b8e80941Smrg{ 356b8e80941Smrg nir_load_const_instr *nlc = 357b8e80941Smrg nir_load_const_instr_create(state->ns, lc->def.num_components, 358b8e80941Smrg lc->def.bit_size); 359b8e80941Smrg 360b8e80941Smrg memcpy(&nlc->value, &lc->value, sizeof(*nlc->value) * lc->def.num_components); 361b8e80941Smrg 362b8e80941Smrg add_remap(state, &nlc->def, &lc->def); 363b8e80941Smrg 364b8e80941Smrg return nlc; 365b8e80941Smrg} 366b8e80941Smrg 367b8e80941Smrgstatic nir_ssa_undef_instr * 368b8e80941Smrgclone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa) 369b8e80941Smrg{ 370b8e80941Smrg nir_ssa_undef_instr *nsa = 371b8e80941Smrg nir_ssa_undef_instr_create(state->ns, sa->def.num_components, 372b8e80941Smrg sa->def.bit_size); 373b8e80941Smrg 374b8e80941Smrg add_remap(state, &nsa->def, &sa->def); 375b8e80941Smrg 376b8e80941Smrg return nsa; 377b8e80941Smrg} 378b8e80941Smrg 379b8e80941Smrgstatic nir_tex_instr * 380b8e80941Smrgclone_tex(clone_state *state, const nir_tex_instr *tex) 381b8e80941Smrg{ 382b8e80941Smrg nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs); 383b8e80941Smrg 384b8e80941Smrg ntex->sampler_dim = tex->sampler_dim; 385b8e80941Smrg ntex->dest_type = tex->dest_type; 386b8e80941Smrg ntex->op = tex->op; 387b8e80941Smrg __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest); 388b8e80941Smrg for (unsigned i = 0; i < ntex->num_srcs; i++) { 389b8e80941Smrg ntex->src[i].src_type = tex->src[i].src_type; 390b8e80941Smrg __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src); 391b8e80941Smrg } 392b8e80941Smrg ntex->coord_components = tex->coord_components; 393b8e80941Smrg ntex->is_array = tex->is_array; 394b8e80941Smrg ntex->is_shadow = tex->is_shadow; 395b8e80941Smrg ntex->is_new_style_shadow = tex->is_new_style_shadow; 396b8e80941Smrg ntex->component = tex->component; 397b8e80941Smrg memcpy(ntex->tg4_offsets, tex->tg4_offsets, sizeof(tex->tg4_offsets)); 398b8e80941Smrg 399b8e80941Smrg ntex->texture_index = tex->texture_index; 400b8e80941Smrg ntex->texture_array_size = tex->texture_array_size; 401b8e80941Smrg ntex->sampler_index = tex->sampler_index; 402b8e80941Smrg 403b8e80941Smrg return ntex; 404b8e80941Smrg} 405b8e80941Smrg 406b8e80941Smrgstatic nir_phi_instr * 407b8e80941Smrgclone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk) 408b8e80941Smrg{ 409b8e80941Smrg nir_phi_instr *nphi = nir_phi_instr_create(state->ns); 410b8e80941Smrg 411b8e80941Smrg __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest); 412b8e80941Smrg 413b8e80941Smrg /* Cloning a phi node is a bit different from other instructions. The 414b8e80941Smrg * sources of phi instructions are the only time where we can use an SSA 415b8e80941Smrg * def before it is defined. In order to handle this, we just copy over 416b8e80941Smrg * the sources from the old phi instruction directly and then fix them up 417b8e80941Smrg * in a second pass once all the instrutions in the function have been 418b8e80941Smrg * properly cloned. 419b8e80941Smrg * 420b8e80941Smrg * In order to ensure that the copied sources (which are the same as the 421b8e80941Smrg * old phi instruction's sources for now) don't get inserted into the old 422b8e80941Smrg * shader's use-def lists, we have to add the phi instruction *before* we 423b8e80941Smrg * set up its sources. 424b8e80941Smrg */ 425b8e80941Smrg nir_instr_insert_after_block(nblk, &nphi->instr); 426b8e80941Smrg 427b8e80941Smrg foreach_list_typed(nir_phi_src, src, node, &phi->srcs) { 428b8e80941Smrg nir_phi_src *nsrc = ralloc(nphi, nir_phi_src); 429b8e80941Smrg 430b8e80941Smrg /* Just copy the old source for now. */ 431b8e80941Smrg memcpy(nsrc, src, sizeof(*src)); 432b8e80941Smrg 433b8e80941Smrg /* Since we're not letting nir_insert_instr handle use/def stuff for us, 434b8e80941Smrg * we have to set the parent_instr manually. It doesn't really matter 435b8e80941Smrg * when we do it, so we might as well do it here. 436b8e80941Smrg */ 437b8e80941Smrg nsrc->src.parent_instr = &nphi->instr; 438b8e80941Smrg 439b8e80941Smrg /* Stash it in the list of phi sources. We'll walk this list and fix up 440b8e80941Smrg * sources at the very end of clone_function_impl. 441b8e80941Smrg */ 442b8e80941Smrg list_add(&nsrc->src.use_link, &state->phi_srcs); 443b8e80941Smrg 444b8e80941Smrg exec_list_push_tail(&nphi->srcs, &nsrc->node); 445b8e80941Smrg } 446b8e80941Smrg 447b8e80941Smrg return nphi; 448b8e80941Smrg} 449b8e80941Smrg 450b8e80941Smrgstatic nir_jump_instr * 451b8e80941Smrgclone_jump(clone_state *state, const nir_jump_instr *jmp) 452b8e80941Smrg{ 453b8e80941Smrg nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type); 454b8e80941Smrg 455b8e80941Smrg return njmp; 456b8e80941Smrg} 457b8e80941Smrg 458b8e80941Smrgstatic nir_call_instr * 459b8e80941Smrgclone_call(clone_state *state, const nir_call_instr *call) 460b8e80941Smrg{ 461b8e80941Smrg nir_function *ncallee = remap_global(state, call->callee); 462b8e80941Smrg nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee); 463b8e80941Smrg 464b8e80941Smrg for (unsigned i = 0; i < ncall->num_params; i++) 465b8e80941Smrg __clone_src(state, ncall, &ncall->params[i], &call->params[i]); 466b8e80941Smrg 467b8e80941Smrg return ncall; 468b8e80941Smrg} 469b8e80941Smrg 470b8e80941Smrgstatic nir_instr * 471b8e80941Smrgclone_instr(clone_state *state, const nir_instr *instr) 472b8e80941Smrg{ 473b8e80941Smrg switch (instr->type) { 474b8e80941Smrg case nir_instr_type_alu: 475b8e80941Smrg return &clone_alu(state, nir_instr_as_alu(instr))->instr; 476b8e80941Smrg case nir_instr_type_deref: 477b8e80941Smrg return &clone_deref_instr(state, nir_instr_as_deref(instr))->instr; 478b8e80941Smrg case nir_instr_type_intrinsic: 479b8e80941Smrg return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr; 480b8e80941Smrg case nir_instr_type_load_const: 481b8e80941Smrg return &clone_load_const(state, nir_instr_as_load_const(instr))->instr; 482b8e80941Smrg case nir_instr_type_ssa_undef: 483b8e80941Smrg return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr; 484b8e80941Smrg case nir_instr_type_tex: 485b8e80941Smrg return &clone_tex(state, nir_instr_as_tex(instr))->instr; 486b8e80941Smrg case nir_instr_type_phi: 487b8e80941Smrg unreachable("Cannot clone phis with clone_instr"); 488b8e80941Smrg case nir_instr_type_jump: 489b8e80941Smrg return &clone_jump(state, nir_instr_as_jump(instr))->instr; 490b8e80941Smrg case nir_instr_type_call: 491b8e80941Smrg return &clone_call(state, nir_instr_as_call(instr))->instr; 492b8e80941Smrg case nir_instr_type_parallel_copy: 493b8e80941Smrg unreachable("Cannot clone parallel copies"); 494b8e80941Smrg default: 495b8e80941Smrg unreachable("bad instr type"); 496b8e80941Smrg return NULL; 497b8e80941Smrg } 498b8e80941Smrg} 499b8e80941Smrg 500b8e80941Smrgstatic nir_block * 501b8e80941Smrgclone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk) 502b8e80941Smrg{ 503b8e80941Smrg /* Don't actually create a new block. Just use the one from the tail of 504b8e80941Smrg * the list. NIR guarantees that the tail of the list is a block and that 505b8e80941Smrg * no two blocks are side-by-side in the IR; It should be empty. 506b8e80941Smrg */ 507b8e80941Smrg nir_block *nblk = 508b8e80941Smrg exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node); 509b8e80941Smrg assert(nblk->cf_node.type == nir_cf_node_block); 510b8e80941Smrg assert(exec_list_is_empty(&nblk->instr_list)); 511b8e80941Smrg 512b8e80941Smrg /* We need this for phi sources */ 513b8e80941Smrg add_remap(state, nblk, blk); 514b8e80941Smrg 515b8e80941Smrg nir_foreach_instr(instr, blk) { 516b8e80941Smrg if (instr->type == nir_instr_type_phi) { 517b8e80941Smrg /* Phi instructions are a bit of a special case when cloning because 518b8e80941Smrg * we don't want inserting the instruction to automatically handle 519b8e80941Smrg * use/defs for us. Instead, we need to wait until all the 520b8e80941Smrg * blocks/instructions are in so that we can set their sources up. 521b8e80941Smrg */ 522b8e80941Smrg clone_phi(state, nir_instr_as_phi(instr), nblk); 523b8e80941Smrg } else { 524b8e80941Smrg nir_instr *ninstr = clone_instr(state, instr); 525b8e80941Smrg nir_instr_insert_after_block(nblk, ninstr); 526b8e80941Smrg } 527b8e80941Smrg } 528b8e80941Smrg 529b8e80941Smrg return nblk; 530b8e80941Smrg} 531b8e80941Smrg 532b8e80941Smrgstatic void 533b8e80941Smrgclone_cf_list(clone_state *state, struct exec_list *dst, 534b8e80941Smrg const struct exec_list *list); 535b8e80941Smrg 536b8e80941Smrgstatic nir_if * 537b8e80941Smrgclone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i) 538b8e80941Smrg{ 539b8e80941Smrg nir_if *ni = nir_if_create(state->ns); 540b8e80941Smrg ni->control = i->control; 541b8e80941Smrg 542b8e80941Smrg __clone_src(state, ni, &ni->condition, &i->condition); 543b8e80941Smrg 544b8e80941Smrg nir_cf_node_insert_end(cf_list, &ni->cf_node); 545b8e80941Smrg 546b8e80941Smrg clone_cf_list(state, &ni->then_list, &i->then_list); 547b8e80941Smrg clone_cf_list(state, &ni->else_list, &i->else_list); 548b8e80941Smrg 549b8e80941Smrg return ni; 550b8e80941Smrg} 551b8e80941Smrg 552b8e80941Smrgstatic nir_loop * 553b8e80941Smrgclone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop) 554b8e80941Smrg{ 555b8e80941Smrg nir_loop *nloop = nir_loop_create(state->ns); 556b8e80941Smrg nloop->control = loop->control; 557b8e80941Smrg nloop->partially_unrolled = loop->partially_unrolled; 558b8e80941Smrg 559b8e80941Smrg nir_cf_node_insert_end(cf_list, &nloop->cf_node); 560b8e80941Smrg 561b8e80941Smrg clone_cf_list(state, &nloop->body, &loop->body); 562b8e80941Smrg 563b8e80941Smrg return nloop; 564b8e80941Smrg} 565b8e80941Smrg 566b8e80941Smrg/* clone list of nir_cf_node: */ 567b8e80941Smrgstatic void 568b8e80941Smrgclone_cf_list(clone_state *state, struct exec_list *dst, 569b8e80941Smrg const struct exec_list *list) 570b8e80941Smrg{ 571b8e80941Smrg foreach_list_typed(nir_cf_node, cf, node, list) { 572b8e80941Smrg switch (cf->type) { 573b8e80941Smrg case nir_cf_node_block: 574b8e80941Smrg clone_block(state, dst, nir_cf_node_as_block(cf)); 575b8e80941Smrg break; 576b8e80941Smrg case nir_cf_node_if: 577b8e80941Smrg clone_if(state, dst, nir_cf_node_as_if(cf)); 578b8e80941Smrg break; 579b8e80941Smrg case nir_cf_node_loop: 580b8e80941Smrg clone_loop(state, dst, nir_cf_node_as_loop(cf)); 581b8e80941Smrg break; 582b8e80941Smrg default: 583b8e80941Smrg unreachable("bad cf type"); 584b8e80941Smrg } 585b8e80941Smrg } 586b8e80941Smrg} 587b8e80941Smrg 588b8e80941Smrg/* After we've cloned almost everything, we have to walk the list of phi 589b8e80941Smrg * sources and fix them up. Thanks to loops, the block and SSA value for a 590b8e80941Smrg * phi source may not be defined when we first encounter it. Instead, we 591b8e80941Smrg * add it to the phi_srcs list and we fix it up here. 592b8e80941Smrg */ 593b8e80941Smrgstatic void 594b8e80941Smrgfixup_phi_srcs(clone_state *state) 595b8e80941Smrg{ 596b8e80941Smrg list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) { 597b8e80941Smrg src->pred = remap_local(state, src->pred); 598b8e80941Smrg 599b8e80941Smrg /* Remove from this list */ 600b8e80941Smrg list_del(&src->src.use_link); 601b8e80941Smrg 602b8e80941Smrg if (src->src.is_ssa) { 603b8e80941Smrg src->src.ssa = remap_local(state, src->src.ssa); 604b8e80941Smrg list_addtail(&src->src.use_link, &src->src.ssa->uses); 605b8e80941Smrg } else { 606b8e80941Smrg src->src.reg.reg = remap_reg(state, src->src.reg.reg); 607b8e80941Smrg list_addtail(&src->src.use_link, &src->src.reg.reg->uses); 608b8e80941Smrg } 609b8e80941Smrg } 610b8e80941Smrg assert(list_empty(&state->phi_srcs)); 611b8e80941Smrg} 612b8e80941Smrg 613b8e80941Smrgvoid 614b8e80941Smrgnir_cf_list_clone(nir_cf_list *dst, nir_cf_list *src, nir_cf_node *parent, 615b8e80941Smrg struct hash_table *remap_table) 616b8e80941Smrg{ 617b8e80941Smrg exec_list_make_empty(&dst->list); 618b8e80941Smrg dst->impl = src->impl; 619b8e80941Smrg 620b8e80941Smrg if (exec_list_is_empty(&src->list)) 621b8e80941Smrg return; 622b8e80941Smrg 623b8e80941Smrg clone_state state; 624b8e80941Smrg init_clone_state(&state, remap_table, false, true); 625b8e80941Smrg 626b8e80941Smrg /* We use the same shader */ 627b8e80941Smrg state.ns = src->impl->function->shader; 628b8e80941Smrg 629b8e80941Smrg /* The control-flow code assumes that the list of cf_nodes always starts 630b8e80941Smrg * and ends with a block. We start by adding an empty block. 631b8e80941Smrg */ 632b8e80941Smrg nir_block *nblk = nir_block_create(state.ns); 633b8e80941Smrg nblk->cf_node.parent = parent; 634b8e80941Smrg exec_list_push_tail(&dst->list, &nblk->cf_node.node); 635b8e80941Smrg 636b8e80941Smrg clone_cf_list(&state, &dst->list, &src->list); 637b8e80941Smrg 638b8e80941Smrg fixup_phi_srcs(&state); 639b8e80941Smrg} 640b8e80941Smrg 641b8e80941Smrgstatic nir_function_impl * 642b8e80941Smrgclone_function_impl(clone_state *state, const nir_function_impl *fi) 643b8e80941Smrg{ 644b8e80941Smrg nir_function_impl *nfi = nir_function_impl_create_bare(state->ns); 645b8e80941Smrg 646b8e80941Smrg clone_var_list(state, &nfi->locals, &fi->locals); 647b8e80941Smrg clone_reg_list(state, &nfi->registers, &fi->registers); 648b8e80941Smrg nfi->reg_alloc = fi->reg_alloc; 649b8e80941Smrg 650b8e80941Smrg assert(list_empty(&state->phi_srcs)); 651b8e80941Smrg 652b8e80941Smrg clone_cf_list(state, &nfi->body, &fi->body); 653b8e80941Smrg 654b8e80941Smrg fixup_phi_srcs(state); 655b8e80941Smrg 656b8e80941Smrg /* All metadata is invalidated in the cloning process */ 657b8e80941Smrg nfi->valid_metadata = 0; 658b8e80941Smrg 659b8e80941Smrg return nfi; 660b8e80941Smrg} 661b8e80941Smrg 662b8e80941Smrgnir_function_impl * 663b8e80941Smrgnir_function_impl_clone(nir_shader *shader, const nir_function_impl *fi) 664b8e80941Smrg{ 665b8e80941Smrg clone_state state; 666b8e80941Smrg init_clone_state(&state, NULL, false, false); 667b8e80941Smrg 668b8e80941Smrg state.ns = shader; 669b8e80941Smrg 670b8e80941Smrg nir_function_impl *nfi = clone_function_impl(&state, fi); 671b8e80941Smrg 672b8e80941Smrg free_clone_state(&state); 673b8e80941Smrg 674b8e80941Smrg return nfi; 675b8e80941Smrg} 676b8e80941Smrg 677b8e80941Smrgstatic nir_function * 678b8e80941Smrgclone_function(clone_state *state, const nir_function *fxn, nir_shader *ns) 679b8e80941Smrg{ 680b8e80941Smrg assert(ns == state->ns); 681b8e80941Smrg nir_function *nfxn = nir_function_create(ns, fxn->name); 682b8e80941Smrg 683b8e80941Smrg /* Needed for call instructions */ 684b8e80941Smrg add_remap(state, nfxn, fxn); 685b8e80941Smrg 686b8e80941Smrg nfxn->num_params = fxn->num_params; 687b8e80941Smrg nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params); 688b8e80941Smrg memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params); 689b8e80941Smrg nfxn->is_entrypoint = fxn->is_entrypoint; 690b8e80941Smrg 691b8e80941Smrg /* At first glance, it looks like we should clone the function_impl here. 692b8e80941Smrg * However, call instructions need to be able to reference at least the 693b8e80941Smrg * function and those will get processed as we clone the function_impls. 694b8e80941Smrg * We stop here and do function_impls as a second pass. 695b8e80941Smrg */ 696b8e80941Smrg 697b8e80941Smrg return nfxn; 698b8e80941Smrg} 699b8e80941Smrg 700b8e80941Smrgnir_shader * 701b8e80941Smrgnir_shader_clone(void *mem_ctx, const nir_shader *s) 702b8e80941Smrg{ 703b8e80941Smrg clone_state state; 704b8e80941Smrg init_clone_state(&state, NULL, true, false); 705b8e80941Smrg 706b8e80941Smrg nir_shader *ns = nir_shader_create(mem_ctx, s->info.stage, s->options, NULL); 707b8e80941Smrg state.ns = ns; 708b8e80941Smrg 709b8e80941Smrg clone_var_list(&state, &ns->uniforms, &s->uniforms); 710b8e80941Smrg clone_var_list(&state, &ns->inputs, &s->inputs); 711b8e80941Smrg clone_var_list(&state, &ns->outputs, &s->outputs); 712b8e80941Smrg clone_var_list(&state, &ns->shared, &s->shared); 713b8e80941Smrg clone_var_list(&state, &ns->globals, &s->globals); 714b8e80941Smrg clone_var_list(&state, &ns->system_values, &s->system_values); 715b8e80941Smrg 716b8e80941Smrg /* Go through and clone functions */ 717b8e80941Smrg foreach_list_typed(nir_function, fxn, node, &s->functions) 718b8e80941Smrg clone_function(&state, fxn, ns); 719b8e80941Smrg 720b8e80941Smrg /* Only after all functions are cloned can we clone the actual function 721b8e80941Smrg * implementations. This is because nir_call_instrs need to reference the 722b8e80941Smrg * functions of other functions and we don't know what order the functions 723b8e80941Smrg * will have in the list. 724b8e80941Smrg */ 725b8e80941Smrg nir_foreach_function(fxn, s) { 726b8e80941Smrg nir_function *nfxn = remap_global(&state, fxn); 727b8e80941Smrg nfxn->impl = clone_function_impl(&state, fxn->impl); 728b8e80941Smrg nfxn->impl->function = nfxn; 729b8e80941Smrg } 730b8e80941Smrg 731b8e80941Smrg ns->info = s->info; 732b8e80941Smrg ns->info.name = ralloc_strdup(ns, ns->info.name); 733b8e80941Smrg if (ns->info.label) 734b8e80941Smrg ns->info.label = ralloc_strdup(ns, ns->info.label); 735b8e80941Smrg 736b8e80941Smrg ns->num_inputs = s->num_inputs; 737b8e80941Smrg ns->num_uniforms = s->num_uniforms; 738b8e80941Smrg ns->num_outputs = s->num_outputs; 739b8e80941Smrg ns->num_shared = s->num_shared; 740b8e80941Smrg ns->scratch_size = s->scratch_size; 741b8e80941Smrg 742b8e80941Smrg ns->constant_data_size = s->constant_data_size; 743b8e80941Smrg if (s->constant_data_size > 0) { 744b8e80941Smrg ns->constant_data = ralloc_size(ns, s->constant_data_size); 745b8e80941Smrg memcpy(ns->constant_data, s->constant_data, s->constant_data_size); 746b8e80941Smrg } 747b8e80941Smrg 748b8e80941Smrg free_clone_state(&state); 749b8e80941Smrg 750b8e80941Smrg return ns; 751b8e80941Smrg} 752