/**
* \file lower_if_to_cond_assign.cpp
*
- * This attempts to flatten if-statements to conditional assignments for
- * GPUs with limited or no flow control support.
+ * This flattens if-statements to conditional assignments if:
+ *
+ * - the GPU has limited or no flow control support
+ * (controlled by max_depth)
+ *
+ * - small conditional branches are more expensive than conditional assignments
+ * (controlled by min_branch_cost, that's the cost for a branch to be
+ * preserved)
*
* It can't handle other control flow being inside of its block, such
* as calls or loops. Hopefully loop unrolling and inlining will take
#include "compiler/glsl_types.h"
#include "ir.h"
-#include "program/hash_table.h"
+#include "util/set.h"
+#include "util/hash_table.h" /* Needed for the hashing functions */
+#include "main/macros.h" /* for MAX2 */
namespace {
class ir_if_to_cond_assign_visitor : public ir_hierarchical_visitor {
public:
- ir_if_to_cond_assign_visitor(unsigned max_depth)
+ ir_if_to_cond_assign_visitor(gl_shader_stage stage,
+ unsigned max_depth,
+ unsigned min_branch_cost)
{
this->progress = false;
+ this->stage = stage;
this->max_depth = max_depth;
+ this->min_branch_cost = min_branch_cost;
this->depth = 0;
- this->condition_variables = hash_table_ctor(0, hash_table_pointer_hash,
- hash_table_pointer_compare);
+ this->condition_variables =
+ _mesa_set_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
}
~ir_if_to_cond_assign_visitor()
{
- hash_table_dtor(this->condition_variables);
+ _mesa_set_destroy(this->condition_variables, NULL);
}
ir_visitor_status visit_enter(ir_if *);
ir_visitor_status visit_leave(ir_if *);
+ bool found_unsupported_op;
+ bool found_expensive_op;
+ bool found_dynamic_arrayref;
+ bool is_then;
bool progress;
+ gl_shader_stage stage;
+ unsigned then_cost;
+ unsigned else_cost;
+ unsigned min_branch_cost;
unsigned max_depth;
unsigned depth;
- struct hash_table *condition_variables;
+ struct set *condition_variables;
};
} /* anonymous namespace */
bool
-lower_if_to_cond_assign(exec_list *instructions, unsigned max_depth)
+lower_if_to_cond_assign(gl_shader_stage stage, exec_list *instructions,
+ unsigned max_depth, unsigned min_branch_cost)
{
if (max_depth == UINT_MAX)
return false;
- ir_if_to_cond_assign_visitor v(max_depth);
+ ir_if_to_cond_assign_visitor v(stage, max_depth, min_branch_cost);
visit_list_elements(&v, instructions);
}
void
-check_control_flow(ir_instruction *ir, void *data)
+check_ir_node(ir_instruction *ir, void *data)
{
- bool *found_control_flow = (bool *)data;
+ ir_if_to_cond_assign_visitor *v = (ir_if_to_cond_assign_visitor *)data;
+
switch (ir->ir_type) {
case ir_type_call:
case ir_type_discard:
case ir_type_loop:
case ir_type_loop_jump:
case ir_type_return:
- *found_control_flow = true;
+ case ir_type_emit_vertex:
+ case ir_type_end_primitive:
+ case ir_type_barrier:
+ v->found_unsupported_op = true;
+ break;
+
+ case ir_type_dereference_variable: {
+ ir_variable *var = ir->as_dereference_variable()->variable_referenced();
+
+ /* Lowering branches with TCS output accesses breaks many piglit tests,
+ * so don't touch them for now.
+ */
+ if (v->stage == MESA_SHADER_TESS_CTRL &&
+ var->data.mode == ir_var_shader_out)
+ v->found_unsupported_op = true;
+ break;
+ }
+
+ /* SSBO, images, atomic counters are handled by ir_type_call */
+ case ir_type_texture:
+ v->found_expensive_op = true;
break;
+
+ case ir_type_dereference_array: {
+ ir_dereference_array *deref = ir->as_dereference_array();
+
+ if (deref->array_index->ir_type != ir_type_constant)
+ v->found_dynamic_arrayref = true;
+ } /* fall-through */
+ case ir_type_expression:
+ case ir_type_dereference_record:
+ if (v->is_then)
+ v->then_cost++;
+ else
+ v->else_cost++;
+ break;
+
default:
break;
}
move_block_to_cond_assign(void *mem_ctx,
ir_if *if_ir, ir_rvalue *cond_expr,
exec_list *instructions,
- struct hash_table *ht)
+ struct set *set)
{
foreach_in_list_safe(ir_instruction, ir, instructions) {
if (ir->ir_type == ir_type_assignment) {
ir_assignment *assign = (ir_assignment *)ir;
- if (hash_table_find(ht, assign) == NULL) {
- hash_table_insert(ht, assign, assign);
+ if (_mesa_set_search(set, assign) == NULL) {
+ _mesa_set_add(set, assign);
/* If the LHS of the assignment is a condition variable that was
* previously added, insert an additional assignment of false to
* the variable.
*/
const bool assign_to_cv =
- hash_table_find(ht, assign->lhs->variable_referenced()) != NULL;
+ _mesa_set_search(
+ set, assign->lhs->variable_referenced()) != NULL;
if (!assign->condition) {
- if (assign_to_cv) {
- assign->rhs =
- new(mem_ctx) ir_expression(ir_binop_logic_and,
- glsl_type::bool_type,
- cond_expr->clone(mem_ctx, NULL),
- assign->rhs);
- } else {
- assign->condition = cond_expr->clone(mem_ctx, NULL);
+ if (assign_to_cv) {
+ assign->rhs =
+ new(mem_ctx) ir_expression(ir_binop_logic_and,
+ glsl_type::bool_type,
+ cond_expr->clone(mem_ctx, NULL),
+ assign->rhs);
+ } else {
+ assign->condition = cond_expr->clone(mem_ctx, NULL);
}
} else {
assign->condition =
- new(mem_ctx) ir_expression(ir_binop_logic_and,
- glsl_type::bool_type,
- cond_expr->clone(mem_ctx, NULL),
- assign->condition);
+ new(mem_ctx) ir_expression(ir_binop_logic_and,
+ glsl_type::bool_type,
+ cond_expr->clone(mem_ctx, NULL),
+ assign->condition);
}
}
}
ir_visitor_status
ir_if_to_cond_assign_visitor::visit_leave(ir_if *ir)
{
+ bool must_lower = this->depth-- > this->max_depth;
+
/* Only flatten when beyond the GPU's maximum supported nesting depth. */
- if (this->depth-- <= this->max_depth)
+ if (!must_lower && this->min_branch_cost == 0)
return visit_continue;
- bool found_control_flow = false;
+ this->found_unsupported_op = false;
+ this->found_expensive_op = false;
+ this->found_dynamic_arrayref = false;
+ this->then_cost = 0;
+ this->else_cost = 0;
+
ir_assignment *assign;
/* Check that both blocks don't contain anything we can't support. */
+ this->is_then = true;
foreach_in_list(ir_instruction, then_ir, &ir->then_instructions) {
- visit_tree(then_ir, check_control_flow, &found_control_flow);
+ visit_tree(then_ir, check_ir_node, this);
}
+
+ this->is_then = false;
foreach_in_list(ir_instruction, else_ir, &ir->else_instructions) {
- visit_tree(else_ir, check_control_flow, &found_control_flow);
+ visit_tree(else_ir, check_ir_node, this);
}
- if (found_control_flow)
+
+ if (this->found_unsupported_op)
+ return visit_continue; /* can't handle inner unsupported opcodes */
+
+ /* Skip if the branch cost is high enough or if there's an expensive op.
+ *
+ * Also skip if non-constant array indices were encountered, since those
+ * can be out-of-bounds for a not-taken branch, and so generating an
+ * assignment would be incorrect. In the case of must_lower, it's up to the
+ * backend to deal with any potential fall-out (perhaps by translating the
+ * assignments to hardware-predicated moves).
+ */
+ if (!must_lower &&
+ (this->found_expensive_op ||
+ this->found_dynamic_arrayref ||
+ MAX2(this->then_cost, this->else_cost) >= this->min_branch_cost))
return visit_continue;
void *mem_ctx = ralloc_parent(ir);
/* Add the new condition variable to the hash table. This allows us to
* find this variable when lowering other (enclosing) if-statements.
*/
- hash_table_insert(this->condition_variables, then_var, then_var);
+ _mesa_set_add(this->condition_variables, then_var);
/* If there are instructions in the else-clause, store the inverse of the
* condition to a variable. Move all of the instructions from the
/* Add the new condition variable to the hash table. This allows us to
* find this variable when lowering other (enclosing) if-statements.
*/
- hash_table_insert(this->condition_variables, else_var, else_var);
+ _mesa_set_add(this->condition_variables, else_var);
}
ir->remove();
projects / mesa.git / blobdiff