From 53bf2508fe3d5bd7889ccf4817fcd2bcc531d6ae Mon Sep 17 00:00:00 2001 From: marha Date: Mon, 7 Feb 2011 07:14:23 +0000 Subject: mesa git update 7 Feb 2011 --- mesalib/src/glsl/Makefile | 1 + mesalib/src/glsl/SConscript | 1 + mesalib/src/glsl/glsl_parser_extras.cpp | 1 + mesalib/src/glsl/ir_optimization.h | 145 +++---- mesalib/src/glsl/linker.cpp | 14 + mesalib/src/glsl/opt_constant_propagation.cpp | 7 +- mesalib/src/glsl/opt_constant_variable.cpp | 7 +- mesalib/src/glsl/opt_copy_propagation.cpp | 9 +- mesalib/src/glsl/opt_copy_propagation_elements.cpp | 458 +++++++++++++++++++++ mesalib/src/glsl/opt_dead_code_local.cpp | 9 +- 10 files changed, 556 insertions(+), 96 deletions(-) create mode 100644 mesalib/src/glsl/opt_copy_propagation_elements.cpp (limited to 'mesalib/src/glsl') diff --git a/mesalib/src/glsl/Makefile b/mesalib/src/glsl/Makefile index 4f307422c..ec11c8a7b 100644 --- a/mesalib/src/glsl/Makefile +++ b/mesalib/src/glsl/Makefile @@ -69,6 +69,7 @@ CXX_SOURCES = \ opt_constant_propagation.cpp \ opt_constant_variable.cpp \ opt_copy_propagation.cpp \ + opt_copy_propagation_elements.cpp \ opt_dead_code.cpp \ opt_dead_code_local.cpp \ opt_dead_functions.cpp \ diff --git a/mesalib/src/glsl/SConscript b/mesalib/src/glsl/SConscript index 2cc7b0eda..f55d7de74 100644 --- a/mesalib/src/glsl/SConscript +++ b/mesalib/src/glsl/SConscript @@ -64,6 +64,7 @@ sources = [ 'opt_constant_propagation.cpp', 'opt_constant_variable.cpp', 'opt_copy_propagation.cpp', + 'opt_copy_propagation_elements.cpp', 'opt_dead_code.cpp', 'opt_dead_code_local.cpp', 'opt_dead_functions.cpp', diff --git a/mesalib/src/glsl/glsl_parser_extras.cpp b/mesalib/src/glsl/glsl_parser_extras.cpp index 1fa209c03..c2bb59b9c 100644 --- a/mesalib/src/glsl/glsl_parser_extras.cpp +++ b/mesalib/src/glsl/glsl_parser_extras.cpp @@ -764,6 +764,7 @@ do_common_optimization(exec_list *ir, bool linked, unsigned max_unroll_iteration progress = do_if_simplification(ir) || progress; progress = do_discard_simplification(ir) || progress; progress = do_copy_propagation(ir) || progress; + progress = do_copy_propagation_elements(ir) || progress; if (linked) progress = do_dead_code(ir) || progress; else diff --git a/mesalib/src/glsl/ir_optimization.h b/mesalib/src/glsl/ir_optimization.h index 609092e76..dd265673c 100644 --- a/mesalib/src/glsl/ir_optimization.h +++ b/mesalib/src/glsl/ir_optimization.h @@ -1,72 +1,73 @@ -/* - * Copyright © 2010 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. - */ - - -/** - * \file ir_optimization.h - * - * Prototypes for optimization passes to be called by the compiler and drivers. - */ - -/* Operations for lower_instructions() */ -#define SUB_TO_ADD_NEG 0x01 -#define DIV_TO_MUL_RCP 0x02 -#define EXP_TO_EXP2 0x04 -#define POW_TO_EXP2 0x08 -#define LOG_TO_LOG2 0x10 -#define MOD_TO_FRACT 0x20 - -bool do_common_optimization(exec_list *ir, bool linked, unsigned max_unroll_iterations); - -bool do_algebraic(exec_list *instructions); -bool do_constant_folding(exec_list *instructions); -bool do_constant_variable(exec_list *instructions); -bool do_constant_variable_unlinked(exec_list *instructions); -bool do_copy_propagation(exec_list *instructions); -bool do_constant_propagation(exec_list *instructions); -bool do_dead_code(exec_list *instructions); -bool do_dead_code_local(exec_list *instructions); -bool do_dead_code_unlinked(exec_list *instructions); -bool do_dead_functions(exec_list *instructions); -bool do_function_inlining(exec_list *instructions); -bool do_lower_jumps(exec_list *instructions, bool pull_out_jumps = true, bool lower_sub_return = true, bool lower_main_return = false, bool lower_continue = false, bool lower_break = false); -bool do_lower_texture_projection(exec_list *instructions); -bool do_if_simplification(exec_list *instructions); -bool do_discard_simplification(exec_list *instructions); -bool lower_if_to_cond_assign(exec_list *instructions, unsigned max_depth = 0); -bool do_mat_op_to_vec(exec_list *instructions); -bool do_mod_to_fract(exec_list *instructions); -bool do_noop_swizzle(exec_list *instructions); -bool do_structure_splitting(exec_list *instructions); -bool do_sub_to_add_neg(exec_list *instructions); -bool do_swizzle_swizzle(exec_list *instructions); -bool do_tree_grafting(exec_list *instructions); -bool do_vec_index_to_cond_assign(exec_list *instructions); -bool do_vec_index_to_swizzle(exec_list *instructions); -bool lower_discard(exec_list *instructions); -bool lower_instructions(exec_list *instructions, unsigned what_to_lower); -bool lower_noise(exec_list *instructions); -bool lower_variable_index_to_cond_assign(exec_list *instructions, - bool lower_input, bool lower_output, bool lower_temp, bool lower_uniform); -bool lower_quadop_vector(exec_list *instructions, bool dont_lower_swz); -bool optimize_redundant_jumps(exec_list *instructions); +/* + * Copyright © 2010 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. + */ + + +/** + * \file ir_optimization.h + * + * Prototypes for optimization passes to be called by the compiler and drivers. + */ + +/* Operations for lower_instructions() */ +#define SUB_TO_ADD_NEG 0x01 +#define DIV_TO_MUL_RCP 0x02 +#define EXP_TO_EXP2 0x04 +#define POW_TO_EXP2 0x08 +#define LOG_TO_LOG2 0x10 +#define MOD_TO_FRACT 0x20 + +bool do_common_optimization(exec_list *ir, bool linked, unsigned max_unroll_iterations); + +bool do_algebraic(exec_list *instructions); +bool do_constant_folding(exec_list *instructions); +bool do_constant_variable(exec_list *instructions); +bool do_constant_variable_unlinked(exec_list *instructions); +bool do_copy_propagation(exec_list *instructions); +bool do_copy_propagation_elements(exec_list *instructions); +bool do_constant_propagation(exec_list *instructions); +bool do_dead_code(exec_list *instructions); +bool do_dead_code_local(exec_list *instructions); +bool do_dead_code_unlinked(exec_list *instructions); +bool do_dead_functions(exec_list *instructions); +bool do_function_inlining(exec_list *instructions); +bool do_lower_jumps(exec_list *instructions, bool pull_out_jumps = true, bool lower_sub_return = true, bool lower_main_return = false, bool lower_continue = false, bool lower_break = false); +bool do_lower_texture_projection(exec_list *instructions); +bool do_if_simplification(exec_list *instructions); +bool do_discard_simplification(exec_list *instructions); +bool lower_if_to_cond_assign(exec_list *instructions, unsigned max_depth = 0); +bool do_mat_op_to_vec(exec_list *instructions); +bool do_mod_to_fract(exec_list *instructions); +bool do_noop_swizzle(exec_list *instructions); +bool do_structure_splitting(exec_list *instructions); +bool do_sub_to_add_neg(exec_list *instructions); +bool do_swizzle_swizzle(exec_list *instructions); +bool do_tree_grafting(exec_list *instructions); +bool do_vec_index_to_cond_assign(exec_list *instructions); +bool do_vec_index_to_swizzle(exec_list *instructions); +bool lower_discard(exec_list *instructions); +bool lower_instructions(exec_list *instructions, unsigned what_to_lower); +bool lower_noise(exec_list *instructions); +bool lower_variable_index_to_cond_assign(exec_list *instructions, + bool lower_input, bool lower_output, bool lower_temp, bool lower_uniform); +bool lower_quadop_vector(exec_list *instructions, bool dont_lower_swz); +bool optimize_redundant_jumps(exec_list *instructions); diff --git a/mesalib/src/glsl/linker.cpp b/mesalib/src/glsl/linker.cpp index c7fb62437..46cd1950c 100644 --- a/mesalib/src/glsl/linker.cpp +++ b/mesalib/src/glsl/linker.cpp @@ -1682,6 +1682,20 @@ link_shaders(struct gl_context *ctx, struct gl_shader_program *prog) demote_shader_inputs_and_outputs(sh, ir_var_in); } + /* OpenGL ES requires that a vertex shader and a fragment shader both be + * present in a linked program. By checking for use of shading language + * version 1.00, we also catch the GL_ARB_ES2_compatibility case. + */ + if (ctx->API == API_OPENGLES2 || prog->Version == 100) { + if (prog->_LinkedShaders[MESA_SHADER_VERTEX] == NULL) { + linker_error_printf(prog, "program lacks a vertex shader\n"); + prog->LinkStatus = false; + } else if (prog->_LinkedShaders[MESA_SHADER_FRAGMENT] == NULL) { + linker_error_printf(prog, "program lacks a fragment shader\n"); + prog->LinkStatus = false; + } + } + /* FINISHME: Assign fragment shader output locations. */ done: diff --git a/mesalib/src/glsl/opt_constant_propagation.cpp b/mesalib/src/glsl/opt_constant_propagation.cpp index 1fff71753..814f6e61f 100644 --- a/mesalib/src/glsl/opt_constant_propagation.cpp +++ b/mesalib/src/glsl/opt_constant_propagation.cpp @@ -400,11 +400,8 @@ ir_constant_propagation_visitor::add_constant(ir_assignment *ir) { acp_entry *entry; - if (ir->condition) { - ir_constant *condition = ir->condition->as_constant(); - if (!condition || !condition->value.b[0]) - return; - } + if (ir->condition) + return; if (!ir->write_mask) return; diff --git a/mesalib/src/glsl/opt_constant_variable.cpp b/mesalib/src/glsl/opt_constant_variable.cpp index 1f51da6e4..9dca0ff8e 100644 --- a/mesalib/src/glsl/opt_constant_variable.cpp +++ b/mesalib/src/glsl/opt_constant_variable.cpp @@ -107,11 +107,8 @@ ir_constant_variable_visitor::visit_enter(ir_assignment *ir) /* OK, now find if we actually have all the right conditions for * this to be a constant value assigned to the var. */ - if (ir->condition) { - constval = ir->condition->constant_expression_value(); - if (!constval || !constval->value.b[0]) - return visit_continue; - } + if (ir->condition) + return visit_continue; ir_variable *var = ir->whole_variable_written(); if (!var) diff --git a/mesalib/src/glsl/opt_copy_propagation.cpp b/mesalib/src/glsl/opt_copy_propagation.cpp index 4ab23bf9b..efa3afda3 100644 --- a/mesalib/src/glsl/opt_copy_propagation.cpp +++ b/mesalib/src/glsl/opt_copy_propagation.cpp @@ -191,7 +191,7 @@ ir_copy_propagation_visitor::visit_enter(ir_call *ir) sig_param_iter.next(); } - /* Since we're unlinked, we don't (necssarily) know the side effects of + /* Since we're unlinked, we don't (necessarily) know the side effects of * this call. So kill all copies. */ acp->make_empty(); @@ -309,11 +309,8 @@ ir_copy_propagation_visitor::add_copy(ir_assignment *ir) { acp_entry *entry; - if (ir->condition) { - ir_constant *condition = ir->condition->as_constant(); - if (!condition || !condition->value.b[0]) - return; - } + if (ir->condition) + return; ir_variable *lhs_var = ir->whole_variable_written(); ir_variable *rhs_var = ir->rhs->whole_variable_referenced(); diff --git a/mesalib/src/glsl/opt_copy_propagation_elements.cpp b/mesalib/src/glsl/opt_copy_propagation_elements.cpp new file mode 100644 index 000000000..8541d9a8e --- /dev/null +++ b/mesalib/src/glsl/opt_copy_propagation_elements.cpp @@ -0,0 +1,458 @@ +/* + * Copyright © 2010 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. + */ + +/** + * \file opt_copy_propagation_elements.cpp + * + * Replaces usage of recently-copied components of variables with the + * previous copy of the variable. + * + * This pass can be compared with opt_copy_propagation, which operands + * on arbitrary whole-variable copies. However, in order to handle + * the copy propagation of swizzled variables or writemasked writes, + * we want to track things on a channel-wise basis. I found that + * trying to mix the swizzled/writemasked support here with the + * whole-variable stuff in opt_copy_propagation.cpp just made a mess, + * so this is separate despite the ACP handling being somewhat + * similar. + * + * This should reduce the number of MOV instructions in the generated + * programs unless copy propagation is also done on the LIR, and may + * help anyway by triggering other optimizations that live in the HIR. + */ + +#include "ir.h" +#include "ir_rvalue_visitor.h" +#include "ir_basic_block.h" +#include "ir_optimization.h" +#include "glsl_types.h" + +static bool debug = false; + +class acp_entry : public exec_node +{ +public: + acp_entry(ir_variable *lhs, ir_variable *rhs, int write_mask, int swizzle[4]) + { + this->lhs = lhs; + this->rhs = rhs; + this->write_mask = write_mask; + memcpy(this->swizzle, swizzle, sizeof(this->swizzle)); + } + + acp_entry(acp_entry *a) + { + this->lhs = a->lhs; + this->rhs = a->rhs; + this->write_mask = a->write_mask; + memcpy(this->swizzle, a->swizzle, sizeof(this->swizzle)); + } + + ir_variable *lhs; + ir_variable *rhs; + unsigned int write_mask; + int swizzle[4]; +}; + + +class kill_entry : public exec_node +{ +public: + kill_entry(ir_variable *var, int write_mask) + { + this->var = var; + this->write_mask = write_mask; + } + + ir_variable *var; + unsigned int write_mask; +}; + +class ir_copy_propagation_elements_visitor : public ir_rvalue_visitor { +public: + ir_copy_propagation_elements_visitor() + { + this->progress = false; + this->mem_ctx = ralloc_context(NULL); + this->shader_mem_ctx = NULL; + this->acp = new(mem_ctx) exec_list; + this->kills = new(mem_ctx) exec_list; + } + ~ir_copy_propagation_elements_visitor() + { + ralloc_free(mem_ctx); + } + + virtual ir_visitor_status visit_enter(class ir_loop *); + virtual ir_visitor_status visit_enter(class ir_function_signature *); + virtual ir_visitor_status visit_leave(class ir_assignment *); + virtual ir_visitor_status visit_enter(class ir_call *); + virtual ir_visitor_status visit_enter(class ir_if *); + + void handle_rvalue(ir_rvalue **rvalue); + + void add_copy(ir_assignment *ir); + void kill(kill_entry *k); + void handle_if_block(exec_list *instructions); + + /** List of acp_entry: The available copies to propagate */ + exec_list *acp; + /** + * List of kill_entry: The variables whose values were killed in this + * block. + */ + exec_list *kills; + + bool progress; + + bool killed_all; + + /* Context for our local data structures. */ + void *mem_ctx; + /* Context for allocating new shader nodes. */ + void *shader_mem_ctx; +}; + +ir_visitor_status +ir_copy_propagation_elements_visitor::visit_enter(ir_function_signature *ir) +{ + /* Treat entry into a function signature as a completely separate + * block. Any instructions at global scope will be shuffled into + * main() at link time, so they're irrelevant to us. + */ + exec_list *orig_acp = this->acp; + exec_list *orig_kills = this->kills; + bool orig_killed_all = this->killed_all; + + this->acp = new(mem_ctx) exec_list; + this->kills = new(mem_ctx) exec_list; + this->killed_all = false; + + visit_list_elements(this, &ir->body); + + this->kills = orig_kills; + this->acp = orig_acp; + this->killed_all = orig_killed_all; + + return visit_continue_with_parent; +} + +ir_visitor_status +ir_copy_propagation_elements_visitor::visit_leave(ir_assignment *ir) +{ + ir_dereference_variable *lhs = ir->lhs->as_dereference_variable(); + + if (lhs && (lhs->type->is_scalar() || lhs->type->is_vector())) { + kill_entry *k = new(mem_ctx) kill_entry(lhs->var, ir->write_mask); + kill(k); + } + + add_copy(ir); + + return visit_continue; +} + +/** + * Replaces dereferences of ACP RHS variables with ACP LHS variables. + * + * This is where the actual copy propagation occurs. Note that the + * rewriting of ir_dereference means that the ir_dereference instance + * must not be shared by multiple IR operations! + */ +void +ir_copy_propagation_elements_visitor::handle_rvalue(ir_rvalue **ir) +{ + int swizzle_chan[4]; + ir_dereference_variable *deref_var; + ir_variable *source[4] = {NULL, NULL, NULL, NULL}; + int source_chan[4]; + int chans; + + if (!*ir) + return; + + ir_swizzle *swizzle = (*ir)->as_swizzle(); + if (swizzle) { + deref_var = swizzle->val->as_dereference_variable(); + if (!deref_var) + return; + + swizzle_chan[0] = swizzle->mask.x; + swizzle_chan[1] = swizzle->mask.y; + swizzle_chan[2] = swizzle->mask.z; + swizzle_chan[3] = swizzle->mask.w; + chans = swizzle->type->vector_elements; + } else { + deref_var = (*ir)->as_dereference_variable(); + if (!deref_var) + return; + + swizzle_chan[0] = 0; + swizzle_chan[1] = 1; + swizzle_chan[2] = 2; + swizzle_chan[3] = 3; + chans = deref_var->type->vector_elements; + } + + if (this->in_assignee) + return; + + ir_variable *var = deref_var->var; + + /* Try to find ACP entries covering swizzle_chan[], hoping they're + * the same source variable. + */ + foreach_iter(exec_list_iterator, iter, *this->acp) { + acp_entry *entry = (acp_entry *)iter.get(); + + if (var == entry->lhs) { + for (int c = 0; c < chans; c++) { + if (entry->write_mask & (1 << swizzle_chan[c])) { + source[c] = entry->rhs; + source_chan[c] = entry->swizzle[swizzle_chan[c]]; + } + } + } + } + + /* Make sure all channels are copying from the same source variable. */ + if (!source[0]) + return; + for (int c = 1; c < chans; c++) { + if (source[c] != source[0]) + return; + } + + if (!shader_mem_ctx) + shader_mem_ctx = ralloc_parent(deref_var); + + if (debug) { + printf("Copy propagation from:\n"); + (*ir)->print(); + } + + deref_var = new(shader_mem_ctx) ir_dereference_variable(source[0]); + *ir = new(shader_mem_ctx) ir_swizzle(deref_var, + source_chan[0], + source_chan[1], + source_chan[2], + source_chan[3], + chans); + + if (debug) { + printf("to:\n"); + (*ir)->print(); + printf("\n"); + } +} + + +ir_visitor_status +ir_copy_propagation_elements_visitor::visit_enter(ir_call *ir) +{ + /* Do copy propagation on call parameters, but skip any out params */ + exec_list_iterator sig_param_iter = ir->get_callee()->parameters.iterator(); + foreach_iter(exec_list_iterator, iter, ir->actual_parameters) { + ir_variable *sig_param = (ir_variable *)sig_param_iter.get(); + ir_instruction *ir = (ir_instruction *)iter.get(); + if (sig_param->mode != ir_var_out && sig_param->mode != ir_var_inout) { + ir->accept(this); + } + sig_param_iter.next(); + } + + /* Since we're unlinked, we don't (necessarily) know the side effects of + * this call. So kill all copies. + */ + acp->make_empty(); + this->killed_all = true; + + return visit_continue_with_parent; +} + +void +ir_copy_propagation_elements_visitor::handle_if_block(exec_list *instructions) +{ + exec_list *orig_acp = this->acp; + exec_list *orig_kills = this->kills; + bool orig_killed_all = this->killed_all; + + this->acp = new(mem_ctx) exec_list; + this->kills = new(mem_ctx) exec_list; + this->killed_all = false; + + /* Populate the initial acp with a copy of the original */ + foreach_iter(exec_list_iterator, iter, *orig_acp) { + acp_entry *a = (acp_entry *)iter.get(); + this->acp->push_tail(new(this->mem_ctx) acp_entry(a)); + } + + visit_list_elements(this, instructions); + + if (this->killed_all) { + orig_acp->make_empty(); + } + + exec_list *new_kills = this->kills; + this->kills = orig_kills; + this->acp = orig_acp; + this->killed_all = this->killed_all || orig_killed_all; + + /* Move the new kills into the parent block's list, removing them + * from the parent's ACP list in the process. + */ + foreach_list_safe(node, new_kills) { + kill_entry *k = (kill_entry *)node; + kill(k); + } +} + +ir_visitor_status +ir_copy_propagation_elements_visitor::visit_enter(ir_if *ir) +{ + ir->condition->accept(this); + + handle_if_block(&ir->then_instructions); + handle_if_block(&ir->else_instructions); + + /* handle_if_block() already descended into the children. */ + return visit_continue_with_parent; +} + +ir_visitor_status +ir_copy_propagation_elements_visitor::visit_enter(ir_loop *ir) +{ + exec_list *orig_acp = this->acp; + exec_list *orig_kills = this->kills; + bool orig_killed_all = this->killed_all; + + /* FINISHME: For now, the initial acp for loops is totally empty. + * We could go through once, then go through again with the acp + * cloned minus the killed entries after the first run through. + */ + this->acp = new(mem_ctx) exec_list; + this->kills = new(mem_ctx) exec_list; + this->killed_all = false; + + visit_list_elements(this, &ir->body_instructions); + + if (this->killed_all) { + orig_acp->make_empty(); + } + + exec_list *new_kills = this->kills; + this->kills = orig_kills; + this->acp = orig_acp; + this->killed_all = this->killed_all || orig_killed_all; + + foreach_list_safe(node, new_kills) { + kill_entry *k = (kill_entry *)node; + kill(k); + } + + /* already descended into the children. */ + return visit_continue_with_parent; +} + +/* Remove any entries currently in the ACP for this kill. */ +void +ir_copy_propagation_elements_visitor::kill(kill_entry *k) +{ + foreach_list_safe(node, acp) { + acp_entry *entry = (acp_entry *)node; + + if (entry->lhs == k->var) { + entry->write_mask = entry->write_mask & ~k->write_mask; + if (entry->write_mask == 0) + entry->remove(); + } + if (entry->rhs == k->var) { + entry->remove(); + } + } + + /* If we were on a list, remove ourselves before inserting */ + if (k->next) + k->remove(); + + this->kills->push_tail(k); +} + +/** + * Adds directly-copied channels between vector variables to the available + * copy propagation list. + */ +void +ir_copy_propagation_elements_visitor::add_copy(ir_assignment *ir) +{ + acp_entry *entry; + int orig_swizzle[4] = {0, 1, 2, 3}; + int swizzle[4]; + + if (ir->condition) + return; + + ir_dereference_variable *lhs = ir->lhs->as_dereference_variable(); + if (!lhs || !(lhs->type->is_scalar() || lhs->type->is_vector())) + return; + + ir_dereference_variable *rhs = ir->rhs->as_dereference_variable(); + if (!rhs) { + ir_swizzle *swiz = ir->rhs->as_swizzle(); + if (!swiz) + return; + + rhs = swiz->val->as_dereference_variable(); + if (!rhs) + return; + + orig_swizzle[0] = swiz->mask.x; + orig_swizzle[1] = swiz->mask.y; + orig_swizzle[2] = swiz->mask.z; + orig_swizzle[3] = swiz->mask.w; + } + + /* Move the swizzle channels out to the positions they match in the + * destination. We don't want to have to rewrite the swizzle[] + * array every time we clear a bit of the write_mask. + */ + int j = 0; + for (int i = 0; i < 4; i++) { + if (ir->write_mask & (1 << i)) + swizzle[i] = orig_swizzle[j++]; + } + + entry = new(this->mem_ctx) acp_entry(lhs->var, rhs->var, ir->write_mask, + swizzle); + this->acp->push_tail(entry); +} + +bool +do_copy_propagation_elements(exec_list *instructions) +{ + ir_copy_propagation_elements_visitor v; + + visit_list_elements(&v, instructions); + + return v.progress; +} diff --git a/mesalib/src/glsl/opt_dead_code_local.cpp b/mesalib/src/glsl/opt_dead_code_local.cpp index 88dcdc251..f67b46337 100644 --- a/mesalib/src/glsl/opt_dead_code_local.cpp +++ b/mesalib/src/glsl/opt_dead_code_local.cpp @@ -131,15 +131,8 @@ process_assignment(void *ctx, ir_assignment *ir, exec_list *assignments) var = ir->lhs->variable_referenced(); assert(var); - bool always_assign = true; - if (ir->condition) { - ir_constant *condition = ir->condition->as_constant(); - if (!condition || !condition->value.b[0]) - always_assign = false; - } - /* Now, check if we did a whole-variable assignment. */ - if (always_assign && (ir->whole_variable_written() != NULL)) { + if (!ir->condition && (ir->whole_variable_written() != NULL)) { /* We did a whole-variable assignment. So, any instruction in * the assignment list with the same LHS is dead. */ -- cgit v1.2.3