summaryrefslogtreecommitdiffstats
path: root/src/glsl/ir_algebraic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/glsl/ir_algebraic.cpp')
-rw-r--r--src/glsl/ir_algebraic.cpp392
1 files changed, 392 insertions, 0 deletions
diff --git a/src/glsl/ir_algebraic.cpp b/src/glsl/ir_algebraic.cpp
new file mode 100644
index 00000000000..a66c820a278
--- /dev/null
+++ b/src/glsl/ir_algebraic.cpp
@@ -0,0 +1,392 @@
+/*
+ * 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_algebraic.cpp
+ *
+ * Takes advantage of association, commutivity, and other algebraic
+ * properties to simplify expressions.
+ */
+
+#include "ir.h"
+#include "ir_visitor.h"
+#include "ir_rvalue_visitor.h"
+#include "ir_optimization.h"
+#include "glsl_types.h"
+
+/**
+ * Visitor class for replacing expressions with ir_constant values.
+ */
+
+class ir_algebraic_visitor : public ir_rvalue_visitor {
+public:
+ ir_algebraic_visitor()
+ {
+ this->progress = false;
+ }
+
+ virtual ~ir_algebraic_visitor()
+ {
+ }
+
+ ir_rvalue *handle_expression(ir_expression *ir);
+ void handle_rvalue(ir_rvalue **rvalue);
+ bool reassociate_constant(ir_expression *ir1,
+ int const_index,
+ ir_constant *constant,
+ ir_expression *ir2);
+ void reassociate_operands(ir_expression *ir1,
+ int op1,
+ ir_expression *ir2,
+ int op2);
+ bool progress;
+};
+
+static bool
+is_vec_zero(ir_constant *ir)
+{
+ int c;
+
+ if (!ir)
+ return false;
+ if (!ir->type->is_scalar() &&
+ !ir->type->is_vector())
+ return false;
+
+ for (c = 0; c < ir->type->vector_elements; c++) {
+ switch (ir->type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ if (ir->value.f[c] != 0.0)
+ return false;
+ break;
+ case GLSL_TYPE_INT:
+ if (ir->value.i[c] != 0)
+ return false;
+ break;
+ case GLSL_TYPE_UINT:
+ if (ir->value.u[c] != 0)
+ return false;
+ break;
+ case GLSL_TYPE_BOOL:
+ if (ir->value.b[c] != false)
+ return false;
+ break;
+ default:
+ assert(!"bad base type");
+ return false;
+ }
+ }
+
+ return true;
+}
+
+static bool
+is_vec_one(ir_constant *ir)
+{
+ int c;
+
+ if (!ir)
+ return false;
+ if (!ir->type->is_scalar() &&
+ !ir->type->is_vector())
+ return false;
+
+ for (c = 0; c < ir->type->vector_elements; c++) {
+ switch (ir->type->base_type) {
+ case GLSL_TYPE_FLOAT:
+ if (ir->value.f[c] != 1.0)
+ return false;
+ break;
+ case GLSL_TYPE_INT:
+ if (ir->value.i[c] != 1)
+ return false;
+ break;
+ case GLSL_TYPE_UINT:
+ if (ir->value.u[c] != 1)
+ return false;
+ break;
+ case GLSL_TYPE_BOOL:
+ if (ir->value.b[c] != true)
+ return false;
+ break;
+ default:
+ assert(!"bad base type");
+ return false;
+ }
+ }
+
+ return true;
+}
+
+static void
+update_type(ir_expression *ir)
+{
+ if (ir->operands[0]->type->is_vector())
+ ir->type = ir->operands[0]->type;
+ else
+ ir->type = ir->operands[1]->type;
+}
+
+void
+ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
+ int op1,
+ ir_expression *ir2,
+ int op2)
+{
+ ir_rvalue *temp = ir2->operands[op2];
+ ir2->operands[op2] = ir1->operands[op1];
+ ir1->operands[op1] = temp;
+
+ /* Update the type of ir2. The type of ir1 won't have changed --
+ * base types matched, and at least one of the operands of the 2
+ * binops is still a vector if any of them were.
+ */
+ update_type(ir2);
+
+ this->progress = true;
+}
+
+/**
+ * Reassociates a constant down a tree of adds or multiplies.
+ *
+ * Consider (2 * (a * (b * 0.5))). We want to send up with a * b.
+ */
+bool
+ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
+ ir_constant *constant,
+ ir_expression *ir2)
+{
+ if (!ir2 || ir1->operation != ir2->operation)
+ return false;
+
+ /* Don't want to even think about matrices. */
+ if (ir1->operands[0]->type->is_matrix() ||
+ ir1->operands[0]->type->is_matrix() ||
+ ir2->operands[1]->type->is_matrix() ||
+ ir2->operands[1]->type->is_matrix())
+ return false;
+
+ ir_constant *ir2_const[2];
+ ir2_const[0] = ir2->operands[0]->constant_expression_value();
+ ir2_const[1] = ir2->operands[1]->constant_expression_value();
+
+ if (ir2_const[0] && ir2_const[1])
+ return false;
+
+ if (ir2_const[0]) {
+ reassociate_operands(ir1, const_index, ir2, 1);
+ return true;
+ } else if (ir2_const[1]) {
+ reassociate_operands(ir1, const_index, ir2, 0);
+ return true;
+ }
+
+ if (reassociate_constant(ir1, const_index, constant,
+ ir2->operands[0]->as_expression())) {
+ update_type(ir2);
+ return true;
+ }
+
+ if (reassociate_constant(ir1, const_index, constant,
+ ir2->operands[1]->as_expression())) {
+ update_type(ir2);
+ return true;
+ }
+
+ return false;
+}
+
+ir_rvalue *
+ir_algebraic_visitor::handle_expression(ir_expression *ir)
+{
+ ir_constant *op_const[2] = {NULL, NULL};
+ ir_expression *op_expr[2] = {NULL, NULL};
+ unsigned int i;
+
+ for (i = 0; i < ir->get_num_operands(); i++) {
+ if (ir->operands[i]->type->is_matrix())
+ return ir;
+
+ op_const[i] = ir->operands[i]->constant_expression_value();
+ op_expr[i] = ir->operands[i]->as_expression();
+ }
+
+ switch (ir->operation) {
+ case ir_unop_logic_not: {
+ enum ir_expression_operation new_op = ir_unop_logic_not;
+
+ if (op_expr[0] == NULL)
+ break;
+
+ switch (op_expr[0]->operation) {
+ case ir_binop_less: new_op = ir_binop_gequal; break;
+ case ir_binop_greater: new_op = ir_binop_lequal; break;
+ case ir_binop_lequal: new_op = ir_binop_greater; break;
+ case ir_binop_gequal: new_op = ir_binop_less; break;
+ case ir_binop_equal: new_op = ir_binop_nequal; break;
+ case ir_binop_nequal: new_op = ir_binop_equal; break;
+
+ default:
+ /* The default case handler is here to silence a warning from GCC.
+ */
+ break;
+ }
+
+ if (new_op != ir_unop_logic_not) {
+ this->progress = true;
+ return new(ir) ir_expression(new_op,
+ ir->type,
+ op_expr[0]->operands[0],
+ op_expr[0]->operands[1]);
+ }
+
+ break;
+ }
+
+ case ir_binop_add:
+ if (is_vec_zero(op_const[0])) {
+ this->progress = true;
+ return ir->operands[1];
+ }
+ if (is_vec_zero(op_const[1])) {
+ this->progress = true;
+ return ir->operands[0];
+ }
+
+ /* Reassociate addition of constants so that we can do constant
+ * folding.
+ */
+ if (op_const[0] && !op_const[1])
+ reassociate_constant(ir, 0, op_const[0],
+ ir->operands[1]->as_expression());
+ if (op_const[1] && !op_const[0])
+ reassociate_constant(ir, 1, op_const[1],
+ ir->operands[0]->as_expression());
+ break;
+
+ case ir_binop_sub:
+ if (is_vec_zero(op_const[0])) {
+ this->progress = true;
+ return new(ir) ir_expression(ir_unop_neg,
+ ir->type,
+ ir->operands[1],
+ NULL);
+ }
+ if (is_vec_zero(op_const[1])) {
+ this->progress = true;
+ return ir->operands[0];
+ }
+ break;
+
+ case ir_binop_mul:
+ if (is_vec_one(op_const[0])) {
+ this->progress = true;
+ return ir->operands[1];
+ }
+ if (is_vec_one(op_const[1])) {
+ this->progress = true;
+ return ir->operands[0];
+ }
+
+ if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
+ this->progress = true;
+ return ir_constant::zero(ir, ir->type);
+ }
+
+ /* Reassociate multiplication of constants so that we can do
+ * constant folding.
+ */
+ if (op_const[0] && !op_const[1])
+ reassociate_constant(ir, 0, op_const[0],
+ ir->operands[1]->as_expression());
+ if (op_const[1] && !op_const[0])
+ reassociate_constant(ir, 1, op_const[1],
+ ir->operands[0]->as_expression());
+
+ break;
+
+ case ir_binop_div:
+ if (is_vec_one(op_const[0]) && ir->type->base_type == GLSL_TYPE_FLOAT) {
+ this->progress = true;
+ return new(ir) ir_expression(ir_unop_rcp,
+ ir->type,
+ ir->operands[1],
+ NULL);
+ }
+ if (is_vec_one(op_const[1])) {
+ this->progress = true;
+ return ir->operands[0];
+ }
+ break;
+
+ case ir_unop_rcp:
+ if (op_expr[0] && op_expr[0]->operation == ir_unop_rcp) {
+ this->progress = true;
+ return op_expr[0]->operands[0];
+ }
+
+ /* FINISHME: We should do rcp(rsq(x)) -> sqrt(x) for some
+ * backends, except that some backends will have done sqrt ->
+ * rcp(rsq(x)) and we don't want to undo it for them.
+ */
+
+ /* As far as we know, all backends are OK with rsq. */
+ if (op_expr[0] && op_expr[0]->operation == ir_unop_sqrt) {
+ this->progress = true;
+ return new(ir) ir_expression(ir_unop_rsq,
+ ir->type,
+ op_expr[0]->operands[0],
+ NULL);
+ }
+
+ break;
+
+ default:
+ break;
+ }
+
+ return ir;
+}
+
+void
+ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
+{
+ if (!*rvalue)
+ return;
+
+ ir_expression *expr = (*rvalue)->as_expression();
+ if (!expr)
+ return;
+
+ *rvalue = handle_expression(expr);
+}
+
+bool
+do_algebraic(exec_list *instructions)
+{
+ ir_algebraic_visitor v;
+
+ visit_list_elements(&v, instructions);
+
+ return v.progress;
+}