diff options
author | marha <marha@users.sourceforge.net> | 2011-04-14 08:39:30 +0000 |
---|---|---|
committer | marha <marha@users.sourceforge.net> | 2011-04-14 08:39:30 +0000 |
commit | 898081f31f99dc35a1602a607a07d1aaff49ac40 (patch) | |
tree | b2adbc0fd699cc5dc2af141df26130e7874f9369 /mesalib/src/glsl | |
parent | c1e91b66cbcf91645f65b9d63f115dcb5a441406 (diff) | |
parent | 019fc27ce6dc2a1809107be10d4deb80e0fa436b (diff) | |
download | vcxsrv-898081f31f99dc35a1602a607a07d1aaff49ac40.tar.gz vcxsrv-898081f31f99dc35a1602a607a07d1aaff49ac40.tar.bz2 vcxsrv-898081f31f99dc35a1602a607a07d1aaff49ac40.zip |
svn merge ^/branches/released .
Diffstat (limited to 'mesalib/src/glsl')
-rw-r--r-- | mesalib/src/glsl/ast_to_hir.cpp | 146 | ||||
-rw-r--r-- | mesalib/src/glsl/glsl_parser_extras.cpp | 2 | ||||
-rw-r--r-- | mesalib/src/glsl/opt_copy_propagation_elements.cpp | 925 |
3 files changed, 532 insertions, 541 deletions
diff --git a/mesalib/src/glsl/ast_to_hir.cpp b/mesalib/src/glsl/ast_to_hir.cpp index 5e1851c11..d7e54d823 100644 --- a/mesalib/src/glsl/ast_to_hir.cpp +++ b/mesalib/src/glsl/ast_to_hir.cpp @@ -848,6 +848,36 @@ do_comparison(void *mem_ctx, int operation, ir_rvalue *op0, ir_rvalue *op1) return cmp;
}
+/* For logical operations, we want to ensure that the operands are
+ * scalar booleans. If it isn't, emit an error and return a constant
+ * boolean to avoid triggering cascading error messages.
+ */
+ir_rvalue *
+get_scalar_boolean_operand(exec_list *instructions,
+ struct _mesa_glsl_parse_state *state,
+ ast_expression *parent_expr,
+ int operand,
+ const char *operand_name,
+ bool *error_emitted)
+{
+ ast_expression *expr = parent_expr->subexpressions[operand];
+ void *ctx = state;
+ ir_rvalue *val = expr->hir(instructions, state);
+
+ if (val->type->is_boolean() && val->type->is_scalar())
+ return val;
+
+ if (!*error_emitted) {
+ YYLTYPE loc = expr->get_location();
+ _mesa_glsl_error(&loc, state, "%s of `%s' must be scalar boolean",
+ operand_name,
+ parent_expr->operator_string(parent_expr->oper));
+ *error_emitted = true;
+ }
+
+ return new(ctx) ir_constant(true);
+}
+
ir_rvalue *
ast_expression::hir(exec_list *instructions,
struct _mesa_glsl_parse_state *state)
@@ -1043,10 +1073,14 @@ ast_expression::hir(exec_list *instructions, error_emitted = true;
}
- result = do_comparison(ctx, operations[this->oper], op[0], op[1]);
- type = glsl_type::bool_type;
+ if (error_emitted) {
+ result = new(ctx) ir_constant(false);
+ } else {
+ result = do_comparison(ctx, operations[this->oper], op[0], op[1]);
+ assert(result->type == glsl_type::bool_type);
+ type = glsl_type::bool_type;
+ }
- assert(error_emitted || (result->type == glsl_type::bool_type));
break;
case ast_bit_and:
@@ -1079,29 +1113,16 @@ ast_expression::hir(exec_list *instructions, break;
case ast_logic_and: {
- op[0] = this->subexpressions[0]->hir(instructions, state);
-
- if (!op[0]->type->is_boolean() || !op[0]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[0]->get_location();
-
- _mesa_glsl_error(& loc, state, "LHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
+ exec_list rhs_instructions;
+ op[0] = get_scalar_boolean_operand(instructions, state, this, 0,
+ "LHS", &error_emitted);
+ op[1] = get_scalar_boolean_operand(&rhs_instructions, state, this, 1,
+ "RHS", &error_emitted);
ir_constant *op0_const = op[0]->constant_expression_value();
if (op0_const) {
if (op0_const->value.b[0]) {
- op[1] = this->subexpressions[1]->hir(instructions, state);
-
- if (!op[1]->type->is_boolean() || !op[1]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[1]->get_location();
-
- _mesa_glsl_error(& loc, state,
- "RHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
+ instructions->append_list(&rhs_instructions);
result = op[1];
} else {
result = op0_const;
@@ -1116,17 +1137,7 @@ ast_expression::hir(exec_list *instructions, ir_if *const stmt = new(ctx) ir_if(op[0]);
instructions->push_tail(stmt);
- op[1] = this->subexpressions[1]->hir(&stmt->then_instructions, state);
-
- if (!op[1]->type->is_boolean() || !op[1]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[1]->get_location();
-
- _mesa_glsl_error(& loc, state,
- "RHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
-
+ stmt->then_instructions.append_list(&rhs_instructions);
ir_dereference *const then_deref = new(ctx) ir_dereference_variable(tmp);
ir_assignment *const then_assign =
new(ctx) ir_assignment(then_deref, op[1], NULL);
@@ -1144,31 +1155,17 @@ ast_expression::hir(exec_list *instructions, }
case ast_logic_or: {
- op[0] = this->subexpressions[0]->hir(instructions, state);
-
- if (!op[0]->type->is_boolean() || !op[0]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[0]->get_location();
-
- _mesa_glsl_error(& loc, state, "LHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
+ exec_list rhs_instructions;
+ op[0] = get_scalar_boolean_operand(instructions, state, this, 0,
+ "LHS", &error_emitted);
+ op[1] = get_scalar_boolean_operand(&rhs_instructions, state, this, 1,
+ "RHS", &error_emitted);
ir_constant *op0_const = op[0]->constant_expression_value();
if (op0_const) {
if (op0_const->value.b[0]) {
result = op0_const;
} else {
- op[1] = this->subexpressions[1]->hir(instructions, state);
-
- if (!op[1]->type->is_boolean() || !op[1]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[1]->get_location();
-
- _mesa_glsl_error(& loc, state,
- "RHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
result = op[1];
}
type = glsl_type::bool_type;
@@ -1181,21 +1178,12 @@ ast_expression::hir(exec_list *instructions, ir_if *const stmt = new(ctx) ir_if(op[0]);
instructions->push_tail(stmt);
- op[1] = this->subexpressions[1]->hir(&stmt->else_instructions, state);
-
- if (!op[1]->type->is_boolean() || !op[1]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[1]->get_location();
-
- _mesa_glsl_error(& loc, state, "RHS of `%s' must be scalar boolean",
- operator_string(this->oper));
- error_emitted = true;
- }
-
ir_dereference *const then_deref = new(ctx) ir_dereference_variable(tmp);
ir_assignment *const then_assign =
new(ctx) ir_assignment(then_deref, new(ctx) ir_constant(true), NULL);
stmt->then_instructions.push_tail(then_assign);
+ stmt->else_instructions.append_list(&rhs_instructions);
ir_dereference *const else_deref = new(ctx) ir_dereference_variable(tmp);
ir_assignment *const else_assign =
new(ctx) ir_assignment(else_deref, op[1], NULL);
@@ -1208,9 +1196,16 @@ ast_expression::hir(exec_list *instructions, }
case ast_logic_xor:
- op[0] = this->subexpressions[0]->hir(instructions, state);
- op[1] = this->subexpressions[1]->hir(instructions, state);
-
+ /* From page 33 (page 39 of the PDF) of the GLSL 1.10 spec:
+ *
+ * "The logical binary operators and (&&), or ( | | ), and
+ * exclusive or (^^). They operate only on two Boolean
+ * expressions and result in a Boolean expression."
+ */
+ op[0] = get_scalar_boolean_operand(instructions, state, this, 0, "LHS",
+ &error_emitted);
+ op[1] = get_scalar_boolean_operand(instructions, state, this, 1, "RHS",
+ &error_emitted);
result = new(ctx) ir_expression(operations[this->oper], glsl_type::bool_type,
op[0], op[1]);
@@ -1218,15 +1213,8 @@ ast_expression::hir(exec_list *instructions, break;
case ast_logic_not:
- op[0] = this->subexpressions[0]->hir(instructions, state);
-
- if (!op[0]->type->is_boolean() || !op[0]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[0]->get_location();
-
- _mesa_glsl_error(& loc, state,
- "operand of `!' must be scalar boolean");
- error_emitted = true;
- }
+ op[0] = get_scalar_boolean_operand(instructions, state, this, 0,
+ "operand", &error_emitted);
result = new(ctx) ir_expression(operations[this->oper], glsl_type::bool_type,
op[0], NULL);
@@ -1313,20 +1301,14 @@ ast_expression::hir(exec_list *instructions, }
case ast_conditional: {
- op[0] = this->subexpressions[0]->hir(instructions, state);
-
/* From page 59 (page 65 of the PDF) of the GLSL 1.50 spec:
*
* "The ternary selection operator (?:). It operates on three
* expressions (exp1 ? exp2 : exp3). This operator evaluates the
* first expression, which must result in a scalar Boolean."
*/
- if (!op[0]->type->is_boolean() || !op[0]->type->is_scalar()) {
- YYLTYPE loc = this->subexpressions[0]->get_location();
-
- _mesa_glsl_error(& loc, state, "?: condition must be scalar boolean");
- error_emitted = true;
- }
+ op[0] = get_scalar_boolean_operand(instructions, state, this, 0,
+ "condition", &error_emitted);
/* The :? operator is implemented by generating an anonymous temporary
* followed by an if-statement. The last instruction in each branch of
diff --git a/mesalib/src/glsl/glsl_parser_extras.cpp b/mesalib/src/glsl/glsl_parser_extras.cpp index 1e8123491..823e75167 100644 --- a/mesalib/src/glsl/glsl_parser_extras.cpp +++ b/mesalib/src/glsl/glsl_parser_extras.cpp @@ -768,7 +768,7 @@ 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;*/
+ progress = do_copy_propagation_elements(ir) || progress;
if (linked)
progress = do_dead_code(ir) || progress;
else
diff --git a/mesalib/src/glsl/opt_copy_propagation_elements.cpp b/mesalib/src/glsl/opt_copy_propagation_elements.cpp index 8541d9a8e..580c10f27 100644 --- a/mesalib/src/glsl/opt_copy_propagation_elements.cpp +++ b/mesalib/src/glsl/opt_copy_propagation_elements.cpp @@ -1,458 +1,467 @@ -/* - * 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; -} +/*
+ * 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();
+ ir_variable *var = ir->lhs->variable_referenced();
+
+ if (var->type->is_scalar() || var->type->is_vector()) {
+ kill_entry *k;
+
+ if (lhs)
+ k = new(mem_ctx) kill_entry(var, ir->write_mask);
+ else
+ k = new(mem_ctx) kill_entry(var, ~0);
+
+ 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();
+ continue;
+ }
+ }
+ 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;
+}
|