/* * Copyright © 2015 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. */ #include "vtn_private.h" #include "nir/nir_vla.h" static bool vtn_cfg_handle_prepass_instruction(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { switch (opcode) { case SpvOpFunction: { assert(b->func == NULL); b->func = rzalloc(b, struct vtn_function); list_inithead(&b->func->body); b->func->control = w[3]; MAYBE_UNUSED const struct glsl_type *result_type = vtn_value(b, w[1], vtn_value_type_type)->type->type; struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_function); val->func = b->func; const struct vtn_type *func_type = vtn_value(b, w[4], vtn_value_type_type)->type; assert(func_type->return_type->type == result_type); nir_function *func = nir_function_create(b->shader, ralloc_strdup(b->shader, val->name)); func->num_params = func_type->length; func->params = ralloc_array(b->shader, nir_parameter, func->num_params); for (unsigned i = 0; i < func->num_params; i++) { if (func_type->params[i]->base_type == vtn_base_type_pointer) { func->params[i].type = func_type->params[i]->deref->type; } else { func->params[i].type = func_type->params[i]->type; } /* TODO: We could do something smarter here. */ func->params[i].param_type = nir_parameter_inout; } func->return_type = func_type->return_type->type; b->func->impl = nir_function_impl_create(func); b->nb.cursor = nir_before_cf_list(&b->func->impl->body); b->func_param_idx = 0; break; } case SpvOpFunctionEnd: b->func->end = w; b->func = NULL; break; case SpvOpFunctionParameter: { struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; assert(b->func_param_idx < b->func->impl->num_params); nir_variable *param = b->func->impl->params[b->func_param_idx++]; if (type->base_type == vtn_base_type_pointer) { struct vtn_variable *vtn_var = rzalloc(b, struct vtn_variable); vtn_var->type = type->deref; vtn_var->var = param; assert(vtn_var->type->type == param->type); struct vtn_type *without_array = vtn_var->type; while(glsl_type_is_array(without_array->type)) without_array = without_array->array_element; if (glsl_type_is_image(without_array->type)) { vtn_var->mode = vtn_variable_mode_image; param->interface_type = without_array->type; } else if (glsl_type_is_sampler(without_array->type)) { vtn_var->mode = vtn_variable_mode_sampler; param->interface_type = without_array->type; } else { vtn_var->mode = vtn_variable_mode_param; } struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_pointer); /* Name the parameter so it shows up nicely in NIR */ param->name = ralloc_strdup(param, val->name); val->pointer = vtn_pointer_for_variable(b, vtn_var, type); } else { /* We're a regular SSA value. */ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); /* Name the parameter so it shows up nicely in NIR */ param->name = ralloc_strdup(param, val->name); val->ssa = vtn_local_load(b, nir_deref_var_create(b, param)); } break; } case SpvOpLabel: { assert(b->block == NULL); b->block = rzalloc(b, struct vtn_block); b->block->node.type = vtn_cf_node_type_block; b->block->label = w; vtn_push_value(b, w[1], vtn_value_type_block)->block = b->block; if (b->func->start_block == NULL) { /* This is the first block encountered for this function. In this * case, we set the start block and add it to the list of * implemented functions that we'll walk later. */ b->func->start_block = b->block; exec_list_push_tail(&b->functions, &b->func->node); } break; } case SpvOpSelectionMerge: case SpvOpLoopMerge: assert(b->block && b->block->merge == NULL); b->block->merge = w; break; case SpvOpBranch: case SpvOpBranchConditional: case SpvOpSwitch: case SpvOpKill: case SpvOpReturn: case SpvOpReturnValue: case SpvOpUnreachable: assert(b->block && b->block->branch == NULL); b->block->branch = w; b->block = NULL; break; default: /* Continue on as per normal */ return true; } return true; } static void vtn_add_case(struct vtn_builder *b, struct vtn_switch *swtch, struct vtn_block *break_block, uint32_t block_id, uint32_t val, bool is_default) { struct vtn_block *case_block = vtn_value(b, block_id, vtn_value_type_block)->block; /* Don't create dummy cases that just break */ if (case_block == break_block) return; if (case_block->switch_case == NULL) { struct vtn_case *c = ralloc(b, struct vtn_case); list_inithead(&c->body); c->start_block = case_block; c->fallthrough = NULL; util_dynarray_init(&c->values, b); c->is_default = false; c->visited = false; list_addtail(&c->link, &swtch->cases); case_block->switch_case = c; } if (is_default) { case_block->switch_case->is_default = true; } else { util_dynarray_append(&case_block->switch_case->values, uint32_t, val); } } /* This function performs a depth-first search of the cases and puts them * in fall-through order. */ static void vtn_order_case(struct vtn_switch *swtch, struct vtn_case *cse) { if (cse->visited) return; cse->visited = true; list_del(&cse->link); if (cse->fallthrough) { vtn_order_case(swtch, cse->fallthrough); /* If we have a fall-through, place this case right before the case it * falls through to. This ensures that fallthroughs come one after * the other. These two can never get separated because that would * imply something else falling through to the same case. Also, this * can't break ordering because the DFS ensures that this case is * visited before anything that falls through to it. */ list_addtail(&cse->link, &cse->fallthrough->link); } else { list_add(&cse->link, &swtch->cases); } } static enum vtn_branch_type vtn_get_branch_type(struct vtn_block *block, struct vtn_case *swcase, struct vtn_block *switch_break, struct vtn_block *loop_break, struct vtn_block *loop_cont) { if (block->switch_case) { /* This branch is actually a fallthrough */ assert(swcase->fallthrough == NULL || swcase->fallthrough == block->switch_case); swcase->fallthrough = block->switch_case; return vtn_branch_type_switch_fallthrough; } else if (block == loop_break) { return vtn_branch_type_loop_break; } else if (block == loop_cont) { return vtn_branch_type_loop_continue; } else if (block == switch_break) { return vtn_branch_type_switch_break; } else { return vtn_branch_type_none; } } static void vtn_cfg_walk_blocks(struct vtn_builder *b, struct list_head *cf_list, struct vtn_block *start, struct vtn_case *switch_case, struct vtn_block *switch_break, struct vtn_block *loop_break, struct vtn_block *loop_cont, struct vtn_block *end) { struct vtn_block *block = start; while (block != end) { if (block->merge && (*block->merge & SpvOpCodeMask) == SpvOpLoopMerge && !block->loop) { struct vtn_loop *loop = ralloc(b, struct vtn_loop); loop->node.type = vtn_cf_node_type_loop; list_inithead(&loop->body); list_inithead(&loop->cont_body); loop->control = block->merge[3]; list_addtail(&loop->node.link, cf_list); block->loop = loop; struct vtn_block *new_loop_break = vtn_value(b, block->merge[1], vtn_value_type_block)->block; struct vtn_block *new_loop_cont = vtn_value(b, block->merge[2], vtn_value_type_block)->block; /* Note: This recursive call will start with the current block as * its start block. If we weren't careful, we would get here * again and end up in infinite recursion. This is why we set * block->loop above and check for it before creating one. This * way, we only create the loop once and the second call that * tries to handle this loop goes to the cases below and gets * handled as a regular block. * * Note: When we make the recursive walk calls, we pass NULL for * the switch break since you have to break out of the loop first. * We do, however, still pass the current switch case because it's * possible that the merge block for the loop is the start of * another case. */ vtn_cfg_walk_blocks(b, &loop->body, block, switch_case, NULL, new_loop_break, new_loop_cont, NULL ); vtn_cfg_walk_blocks(b, &loop->cont_body, new_loop_cont, NULL, NULL, new_loop_break, NULL, block); block = new_loop_break; continue; } assert(block->node.link.next == NULL); list_addtail(&block->node.link, cf_list); switch (*block->branch & SpvOpCodeMask) { case SpvOpBranch: { struct vtn_block *branch_block = vtn_value(b, block->branch[1], vtn_value_type_block)->block; block->branch_type = vtn_get_branch_type(branch_block, switch_case, switch_break, loop_break, loop_cont); if (block->branch_type != vtn_branch_type_none) return; block = branch_block; continue; } case SpvOpReturn: case SpvOpReturnValue: block->branch_type = vtn_branch_type_return; return; case SpvOpKill: block->branch_type = vtn_branch_type_discard; return; case SpvOpBranchConditional: { struct vtn_block *then_block = vtn_value(b, block->branch[2], vtn_value_type_block)->block; struct vtn_block *else_block = vtn_value(b, block->branch[3], vtn_value_type_block)->block; struct vtn_if *if_stmt = ralloc(b, struct vtn_if); if_stmt->node.type = vtn_cf_node_type_if; if_stmt->condition = block->branch[1]; list_inithead(&if_stmt->then_body); list_inithead(&if_stmt->else_body); list_addtail(&if_stmt->node.link, cf_list); if (block->merge && (*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge) { if_stmt->control = block->merge[2]; } if_stmt->then_type = vtn_get_branch_type(then_block, switch_case, switch_break, loop_break, loop_cont); if_stmt->else_type = vtn_get_branch_type(else_block, switch_case, switch_break, loop_break, loop_cont); if (if_stmt->then_type == vtn_branch_type_none && if_stmt->else_type == vtn_branch_type_none) { /* Neither side of the if is something we can short-circuit. */ assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge); struct vtn_block *merge_block = vtn_value(b, block->merge[1], vtn_value_type_block)->block; vtn_cfg_walk_blocks(b, &if_stmt->then_body, then_block, switch_case, switch_break, loop_break, loop_cont, merge_block); vtn_cfg_walk_blocks(b, &if_stmt->else_body, else_block, switch_case, switch_break, loop_break, loop_cont, merge_block); enum vtn_branch_type merge_type = vtn_get_branch_type(merge_block, switch_case, switch_break, loop_break, loop_cont); if (merge_type == vtn_branch_type_none) { block = merge_block; continue; } else { return; } } else if (if_stmt->then_type != vtn_branch_type_none && if_stmt->else_type != vtn_branch_type_none) { /* Both sides were short-circuited. We're done here. */ return; } else { /* Exeactly one side of the branch could be short-circuited. * We set the branch up as a predicated break/continue and we * continue on with the other side as if it were what comes * after the if. */ if (if_stmt->then_type == vtn_branch_type_none) { block = then_block; } else { block = else_block; } continue; } unreachable("Should have returned or continued"); } case SpvOpSwitch: { assert((*block->merge & SpvOpCodeMask) == SpvOpSelectionMerge); struct vtn_block *break_block = vtn_value(b, block->merge[1], vtn_value_type_block)->block; struct vtn_switch *swtch = ralloc(b, struct vtn_switch); swtch->node.type = vtn_cf_node_type_switch; swtch->selector = block->branch[1]; list_inithead(&swtch->cases); list_addtail(&swtch->node.link, cf_list); /* First, we go through and record all of the cases. */ const uint32_t *branch_end = block->branch + (block->branch[0] >> SpvWordCountShift); vtn_add_case(b, swtch, break_block, block->branch[2], 0, true); for (const uint32_t *w = block->branch + 3; w < branch_end; w += 2) vtn_add_case(b, swtch, break_block, w[1], w[0], false); /* Now, we go through and walk the blocks. While we walk through * the blocks, we also gather the much-needed fall-through * information. */ list_for_each_entry(struct vtn_case, cse, &swtch->cases, link) { assert(cse->start_block != break_block); vtn_cfg_walk_blocks(b, &cse->body, cse->start_block, cse, break_block, NULL, loop_cont, NULL); } /* Finally, we walk over all of the cases one more time and put * them in fall-through order. */ for (const uint32_t *w = block->branch + 2; w < branch_end; w += 2) { struct vtn_block *case_block = vtn_value(b, *w, vtn_value_type_block)->block; if (case_block == break_block) continue; assert(case_block->switch_case); vtn_order_case(swtch, case_block->switch_case); } enum vtn_branch_type branch_type = vtn_get_branch_type(break_block, switch_case, NULL, loop_break, loop_cont); if (branch_type != vtn_branch_type_none) { /* It is possible that the break is actually the continue block * for the containing loop. In this case, we need to bail and let * the loop parsing code handle the continue properly. */ assert(branch_type == vtn_branch_type_loop_continue); return; } block = break_block; continue; } case SpvOpUnreachable: return; default: unreachable("Unhandled opcode"); } } } void vtn_build_cfg(struct vtn_builder *b, const uint32_t *words, const uint32_t *end) { vtn_foreach_instruction(b, words, end, vtn_cfg_handle_prepass_instruction); foreach_list_typed(struct vtn_function, func, node, &b->functions) { vtn_cfg_walk_blocks(b, &func->body, func->start_block, NULL, NULL, NULL, NULL, NULL); } } static bool vtn_handle_phis_first_pass(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { if (opcode == SpvOpLabel) return true; /* Nothing to do */ /* If this isn't a phi node, stop. */ if (opcode != SpvOpPhi) return false; /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot. * For each phi, we create a variable with the appropreate type and * do a load from that variable. Then, in a second pass, we add * stores to that variable to each of the predecessor blocks. * * We could do something more intelligent here. However, in order to * handle loops and things properly, we really need dominance * information. It would end up basically being the into-SSA * algorithm all over again. It's easier if we just let * lower_vars_to_ssa do that for us instead of repeating it here. */ struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa); struct vtn_type *type = vtn_value(b, w[1], vtn_value_type_type)->type; nir_variable *phi_var = nir_local_variable_create(b->nb.impl, type->type, "phi"); _mesa_hash_table_insert(b->phi_table, w, phi_var); val->ssa = vtn_local_load(b, nir_deref_var_create(b, phi_var)); return true; } static bool vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode, const uint32_t *w, unsigned count) { if (opcode != SpvOpPhi) return true; struct hash_entry *phi_entry = _mesa_hash_table_search(b->phi_table, w); assert(phi_entry); nir_variable *phi_var = phi_entry->data; for (unsigned i = 3; i < count; i += 2) { struct vtn_block *pred = vtn_value(b, w[i + 1], vtn_value_type_block)->block; b->nb.cursor = nir_after_instr(&pred->end_nop->instr); struct vtn_ssa_value *src = vtn_ssa_value(b, w[i]); vtn_local_store(b, src, nir_deref_var_create(b, phi_var)); } return true; } static void vtn_emit_branch(struct vtn_builder *b, enum vtn_branch_type branch_type, nir_variable *switch_fall_var, bool *has_switch_break) { switch (branch_type) { case vtn_branch_type_switch_break: nir_store_var(&b->nb, switch_fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1); *has_switch_break = true; break; case vtn_branch_type_switch_fallthrough: break; /* Nothing to do */ case vtn_branch_type_loop_break: nir_jump(&b->nb, nir_jump_break); break; case vtn_branch_type_loop_continue: nir_jump(&b->nb, nir_jump_continue); break; case vtn_branch_type_return: nir_jump(&b->nb, nir_jump_return); break; case vtn_branch_type_discard: { nir_intrinsic_instr *discard = nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_discard); nir_builder_instr_insert(&b->nb, &discard->instr); break; } default: unreachable("Invalid branch type"); } } static void vtn_emit_cf_list(struct vtn_builder *b, struct list_head *cf_list, nir_variable *switch_fall_var, bool *has_switch_break, vtn_instruction_handler handler) { list_for_each_entry(struct vtn_cf_node, node, cf_list, link) { switch (node->type) { case vtn_cf_node_type_block: { struct vtn_block *block = (struct vtn_block *)node; const uint32_t *block_start = block->label; const uint32_t *block_end = block->merge ? block->merge : block->branch; block_start = vtn_foreach_instruction(b, block_start, block_end, vtn_handle_phis_first_pass); vtn_foreach_instruction(b, block_start, block_end, handler); block->end_nop = nir_intrinsic_instr_create(b->nb.shader, nir_intrinsic_nop); nir_builder_instr_insert(&b->nb, &block->end_nop->instr); if ((*block->branch & SpvOpCodeMask) == SpvOpReturnValue) { struct vtn_ssa_value *src = vtn_ssa_value(b, block->branch[1]); vtn_local_store(b, src, nir_deref_var_create(b, b->impl->return_var)); } if (block->branch_type != vtn_branch_type_none) { vtn_emit_branch(b, block->branch_type, switch_fall_var, has_switch_break); } break; } case vtn_cf_node_type_if: { struct vtn_if *vtn_if = (struct vtn_if *)node; bool sw_break = false; nir_if *nif = nir_push_if(&b->nb, vtn_ssa_value(b, vtn_if->condition)->def); if (vtn_if->then_type == vtn_branch_type_none) { vtn_emit_cf_list(b, &vtn_if->then_body, switch_fall_var, &sw_break, handler); } else { vtn_emit_branch(b, vtn_if->then_type, switch_fall_var, &sw_break); } nir_push_else(&b->nb, nif); if (vtn_if->else_type == vtn_branch_type_none) { vtn_emit_cf_list(b, &vtn_if->else_body, switch_fall_var, &sw_break, handler); } else { vtn_emit_branch(b, vtn_if->else_type, switch_fall_var, &sw_break); } nir_pop_if(&b->nb, nif); /* If we encountered a switch break somewhere inside of the if, * then it would have been handled correctly by calling * emit_cf_list or emit_branch for the interrior. However, we * need to predicate everything following on wether or not we're * still going. */ if (sw_break) { *has_switch_break = true; nir_push_if(&b->nb, nir_load_var(&b->nb, switch_fall_var)); } break; } case vtn_cf_node_type_loop: { struct vtn_loop *vtn_loop = (struct vtn_loop *)node; nir_loop *loop = nir_push_loop(&b->nb); vtn_emit_cf_list(b, &vtn_loop->body, NULL, NULL, handler); if (!list_empty(&vtn_loop->cont_body)) { /* If we have a non-trivial continue body then we need to put * it at the beginning of the loop with a flag to ensure that * it doesn't get executed in the first iteration. */ nir_variable *do_cont = nir_local_variable_create(b->nb.impl, glsl_bool_type(), "cont"); b->nb.cursor = nir_before_cf_node(&loop->cf_node); nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_FALSE), 1); b->nb.cursor = nir_before_cf_list(&loop->body); nir_if *cont_if = nir_push_if(&b->nb, nir_load_var(&b->nb, do_cont)); vtn_emit_cf_list(b, &vtn_loop->cont_body, NULL, NULL, handler); nir_pop_if(&b->nb, cont_if); nir_store_var(&b->nb, do_cont, nir_imm_int(&b->nb, NIR_TRUE), 1); b->has_loop_continue = true; } nir_pop_loop(&b->nb, loop); break; } case vtn_cf_node_type_switch: { struct vtn_switch *vtn_switch = (struct vtn_switch *)node; /* First, we create a variable to keep track of whether or not the * switch is still going at any given point. Any switch breaks * will set this variable to false. */ nir_variable *fall_var = nir_local_variable_create(b->nb.impl, glsl_bool_type(), "fall"); nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_FALSE), 1); /* Next, we gather up all of the conditions. We have to do this * up-front because we also need to build an "any" condition so * that we can use !any for default. */ const int num_cases = list_length(&vtn_switch->cases); NIR_VLA(nir_ssa_def *, conditions, num_cases); nir_ssa_def *sel = vtn_ssa_value(b, vtn_switch->selector)->def; /* An accumulation of all conditions. Used for the default */ nir_ssa_def *any = NULL; int i = 0; list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) { if (cse->is_default) { conditions[i++] = NULL; continue; } nir_ssa_def *cond = NULL; util_dynarray_foreach(&cse->values, uint32_t, val) { nir_ssa_def *is_val = nir_ieq(&b->nb, sel, nir_imm_int(&b->nb, *val)); cond = cond ? nir_ior(&b->nb, cond, is_val) : is_val; } any = any ? nir_ior(&b->nb, any, cond) : cond; conditions[i++] = cond; } assert(i == num_cases); /* Now we can walk the list of cases and actually emit code */ i = 0; list_for_each_entry(struct vtn_case, cse, &vtn_switch->cases, link) { /* Figure out the condition */ nir_ssa_def *cond = conditions[i++]; if (cse->is_default) { assert(cond == NULL); cond = nir_inot(&b->nb, any); } /* Take fallthrough into account */ cond = nir_ior(&b->nb, cond, nir_load_var(&b->nb, fall_var)); nir_if *case_if = nir_push_if(&b->nb, cond); bool has_break = false; nir_store_var(&b->nb, fall_var, nir_imm_int(&b->nb, NIR_TRUE), 1); vtn_emit_cf_list(b, &cse->body, fall_var, &has_break, handler); (void)has_break; /* We don't care */ nir_pop_if(&b->nb, case_if); } assert(i == num_cases); break; } default: unreachable("Invalid CF node type"); } } } void vtn_function_emit(struct vtn_builder *b, struct vtn_function *func, vtn_instruction_handler instruction_handler) { nir_builder_init(&b->nb, func->impl); b->nb.cursor = nir_after_cf_list(&func->impl->body); b->has_loop_continue = false; b->phi_table = _mesa_hash_table_create(b, _mesa_hash_pointer, _mesa_key_pointer_equal); vtn_emit_cf_list(b, &func->body, NULL, NULL, instruction_handler); vtn_foreach_instruction(b, func->start_block->label, func->end, vtn_handle_phi_second_pass); /* Continue blocks for loops get inserted before the body of the loop * but instructions in the continue may use SSA defs in the loop body. * Therefore, we need to repair SSA to insert the needed phi nodes. */ if (b->has_loop_continue) nir_repair_ssa_impl(func->impl); }