#include "ir_basic_block.h"
#include "ir_optimization.h"
#include "glsl_types.h"
+#include "util/hash_table.h"
+
+namespace {
class acp_entry : public exec_node
{
this->var = var;
this->write_mask = write_mask;
this->constant = constant;
+ this->initial_values = write_mask;
+ }
+
+ acp_entry(const acp_entry *src)
+ {
+ this->var = src->var;
+ this->write_mask = src->write_mask;
+ this->constant = src->constant;
+ this->initial_values = src->initial_values;
}
ir_variable *var;
ir_constant *constant;
unsigned write_mask;
+
+ /** Mask of values initially available in the constant. */
+ unsigned initial_values;
};
ir_constant_propagation_visitor()
{
progress = false;
- mem_ctx = talloc_new(0);
+ killed_all = false;
+ mem_ctx = ralloc_context(0);
this->acp = new(mem_ctx) exec_list;
- this->kills = new(mem_ctx) exec_list;
+ this->kills = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
}
~ir_constant_propagation_visitor()
{
- talloc_free(mem_ctx);
+ ralloc_free(mem_ctx);
}
virtual ir_visitor_status visit_enter(class ir_loop *);
virtual ir_visitor_status visit_enter(class ir_if *);
void add_constant(ir_assignment *ir);
+ void constant_folding(ir_rvalue **rvalue);
+ void constant_propagation(ir_rvalue **rvalue);
void kill(ir_variable *ir, unsigned write_mask);
void handle_if_block(exec_list *instructions);
void handle_rvalue(ir_rvalue **rvalue);
* List of kill_entry: The masks of variables whose values were
* killed in this block.
*/
- exec_list *kills;
+ hash_table *kills;
bool progress;
void
-ir_constant_propagation_visitor::handle_rvalue(ir_rvalue **rvalue)
-{
+ir_constant_propagation_visitor::constant_folding(ir_rvalue **rvalue) {
+
+ if (*rvalue == NULL || (*rvalue)->ir_type == ir_type_constant)
+ return;
+
+ /* Note that we visit rvalues one leaving. So if an expression has a
+ * non-constant operand, no need to go looking down it to find if it's
+ * constant. This cuts the time of this pass down drastically.
+ */
+ ir_expression *expr = (*rvalue)->as_expression();
+ if (expr) {
+ for (unsigned int i = 0; i < expr->get_num_operands(); i++) {
+ if (!expr->operands[i]->as_constant())
+ return;
+ }
+ }
+
+ /* Ditto for swizzles. */
+ ir_swizzle *swiz = (*rvalue)->as_swizzle();
+ if (swiz && !swiz->val->as_constant())
+ return;
+
+ ir_constant *constant = (*rvalue)->constant_expression_value();
+ if (constant) {
+ *rvalue = constant;
+ this->progress = true;
+ }
+}
+
+void
+ir_constant_propagation_visitor::constant_propagation(ir_rvalue **rvalue) {
+
if (this->in_assignee || !*rvalue)
return;
channel = i;
}
- foreach_iter(exec_list_iterator, iter, *this->acp) {
- acp_entry *entry = (acp_entry *)iter.get();
+ foreach_in_list(acp_entry, entry, this->acp) {
if (entry->var == deref->var && entry->write_mask & (1 << channel)) {
found = entry;
break;
for (int j = 0; j < 4; j++) {
if (j == channel)
break;
- if (found->write_mask & (1 << j))
+ if (found->initial_values & (1 << j))
rhs_channel++;
}
case GLSL_TYPE_FLOAT:
data.f[i] = found->constant->value.f[rhs_channel];
break;
+ case GLSL_TYPE_DOUBLE:
+ data.d[i] = found->constant->value.d[rhs_channel];
+ break;
case GLSL_TYPE_INT:
data.i[i] = found->constant->value.i[rhs_channel];
break;
}
}
- *rvalue = new(talloc_parent(deref)) ir_constant(type, &data);
+ *rvalue = new(ralloc_parent(deref)) ir_constant(type, &data);
this->progress = true;
}
+void
+ir_constant_propagation_visitor::handle_rvalue(ir_rvalue **rvalue)
+{
+ constant_propagation(rvalue);
+ constant_folding(rvalue);
+}
+
ir_visitor_status
ir_constant_propagation_visitor::visit_enter(ir_function_signature *ir)
{
* main() at link time, so they're irrelevant to us.
*/
exec_list *orig_acp = this->acp;
- exec_list *orig_kills = this->kills;
+ hash_table *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->kills = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
this->killed_all = false;
visit_list_elements(this, &ir->body);
ir_visitor_status
ir_constant_propagation_visitor::visit_leave(ir_assignment *ir)
{
+ constant_folding(&ir->rhs);
+
if (this->in_assignee)
return visit_continue;
- kill(ir->lhs->variable_referenced(), ir->write_mask);
+ unsigned kill_mask = ir->write_mask;
+ if (ir->lhs->as_dereference_array()) {
+ /* The LHS of the assignment uses an array indexing operator (e.g. v[i]
+ * = ...;). Since we only try to constant propagate vectors and
+ * scalars, this means that either (a) array indexing is being used to
+ * select a vector component, or (b) the variable in question is neither
+ * a scalar or a vector, so we don't care about it. In the former case,
+ * we want to kill the whole vector, since in general we can't predict
+ * which vector component will be selected by array indexing. In the
+ * latter case, it doesn't matter what we do, so go ahead and kill the
+ * whole variable anyway.
+ *
+ * Note that if the array index is constant (e.g. v[2] = ...;), we could
+ * in principle be smarter, but we don't need to, because a future
+ * optimization pass will convert it to a simple assignment with the
+ * correct mask.
+ */
+ kill_mask = ~0;
+ }
+ kill(ir->lhs->variable_referenced(), kill_mask);
add_constant(ir);
ir_constant_propagation_visitor::visit_enter(ir_call *ir)
{
/* Do constant 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_rvalue *param = (ir_rvalue *)iter.get();
- if (sig_param->mode != ir_var_out && sig_param->mode != ir_var_inout) {
+ foreach_two_lists(formal_node, &ir->callee->parameters,
+ actual_node, &ir->actual_parameters) {
+ ir_variable *sig_param = (ir_variable *) formal_node;
+ ir_rvalue *param = (ir_rvalue *) actual_node;
+ if (sig_param->data.mode != ir_var_function_out
+ && sig_param->data.mode != ir_var_function_inout) {
ir_rvalue *new_param = param;
handle_rvalue(&new_param);
if (new_param != param)
else
param->accept(this);
}
- sig_param_iter.next();
}
/* Since we're unlinked, we don't (necssarily) know the side effects of
ir_constant_propagation_visitor::handle_if_block(exec_list *instructions)
{
exec_list *orig_acp = this->acp;
- exec_list *orig_kills = this->kills;
+ hash_table *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->kills = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
this->killed_all = false;
/* Populate the initial acp with a constant 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->var, a->write_mask,
- a->constant));
+ foreach_in_list(acp_entry, a, orig_acp) {
+ this->acp->push_tail(new(this->mem_ctx) acp_entry(a));
}
visit_list_elements(this, instructions);
orig_acp->make_empty();
}
- exec_list *new_kills = this->kills;
+ hash_table *new_kills = this->kills;
this->kills = orig_kills;
this->acp = orig_acp;
this->killed_all = this->killed_all || orig_killed_all;
- foreach_iter(exec_list_iterator, iter, *new_kills) {
- kill_entry *k = (kill_entry *)iter.get();
+ hash_entry *htk;
+ hash_table_foreach(new_kills, htk) {
+ kill_entry *k = (kill_entry *) htk->data;
kill(k->var, k->write_mask);
}
}
ir_constant_propagation_visitor::visit_enter(ir_loop *ir)
{
exec_list *orig_acp = this->acp;
- exec_list *orig_kills = this->kills;
+ hash_table *orig_kills = this->kills;
bool orig_killed_all = this->killed_all;
/* FINISHME: For now, the initial acp for loops is totally empty.
* 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->kills = _mesa_hash_table_create(mem_ctx, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
this->killed_all = false;
visit_list_elements(this, &ir->body_instructions);
orig_acp->make_empty();
}
- exec_list *new_kills = this->kills;
+ hash_table *new_kills = this->kills;
this->kills = orig_kills;
this->acp = orig_acp;
this->killed_all = this->killed_all || orig_killed_all;
- foreach_iter(exec_list_iterator, iter, *new_kills) {
- kill_entry *k = (kill_entry *)iter.get();
+ hash_entry *htk;
+ hash_table_foreach(new_kills, htk) {
+ kill_entry *k = (kill_entry *) htk->data;
kill(k->var, k->write_mask);
}
return;
/* Remove any entries currently in the ACP for this kill. */
- foreach_iter(exec_list_iterator, iter, *this->acp) {
- acp_entry *entry = (acp_entry *)iter.get();
-
+ foreach_in_list_safe(acp_entry, entry, this->acp) {
if (entry->var == var) {
entry->write_mask &= ~write_mask;
if (entry->write_mask == 0)
/* Add this writemask of the variable to the list of killed
* variables in this block.
*/
- foreach_iter(exec_list_iterator, iter, *this->kills) {
- kill_entry *entry = (kill_entry *)iter.get();
-
- if (entry->var == var) {
- entry->write_mask |= write_mask;
- return;
- }
+ hash_entry *kill_hash_entry = _mesa_hash_table_search(this->kills, var);
+ if (kill_hash_entry) {
+ kill_entry *entry = (kill_entry *) kill_hash_entry->data;
+ entry->write_mask |= write_mask;
+ return;
}
/* Not already in the list. Make new entry. */
- this->kills->push_tail(new(this->mem_ctx) kill_entry(var, write_mask));
+ _mesa_hash_table_insert(this->kills, var,
+ new(this->mem_ctx) kill_entry(var, write_mask));
}
/**
{
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;
if (!deref->var->type->is_vector() && !deref->var->type->is_scalar())
return;
+ /* We can't do copy propagation on buffer variables, since the underlying
+ * memory storage is shared across multiple threads we can't be sure that
+ * the variable value isn't modified between this assignment and the next
+ * instruction where its value is read.
+ */
+ if (deref->var->data.mode == ir_var_shader_storage ||
+ deref->var->data.mode == ir_var_shader_shared)
+ return;
+
entry = new(this->mem_ctx) acp_entry(deref->var, ir->write_mask, constant);
this->acp->push_tail(entry);
}
+} /* unnamed namespace */
+
/**
* Does a constant propagation pass on the code present in the instruction stream.
*/