#include "ir_basic_block.h"
#include "ir_optimization.h"
#include "compiler/glsl_types.h"
+#include "util/hash_table.h"
static bool debug = false;
namespace {
+class acp_entry;
+
+/* Class that refers to acp_entry in another exec_list. Used
+ * when making removals based on rhs.
+ */
+class acp_ref : public exec_node
+{
+public:
+ acp_ref(acp_entry *e)
+ {
+ entry = e;
+ }
+ acp_entry *entry;
+};
+
class acp_entry : public exec_node
{
public:
+ /* override operator new from exec_node */
+ DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(acp_entry)
+
acp_entry(ir_variable *lhs, ir_variable *rhs, int write_mask, int swizzle[4])
+ : rhs_node(this)
{
this->lhs = lhs;
this->rhs = rhs;
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];
+ acp_ref rhs_node;
};
+class copy_propagation_state {
+public:
+ DECLARE_RZALLOC_CXX_OPERATORS(copy_propagation_state);
+
+ copy_propagation_state(void *mem_ctx, void *lin_ctx)
+ {
+ /* Use 'this' as context for the tables, no explicit destruction
+ * needed later.
+ */
+ lhs_ht = _mesa_hash_table_create(this, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+ rhs_ht = _mesa_hash_table_create(this, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+ this->mem_ctx = mem_ctx;
+ this->lin_ctx = lin_ctx;
+ }
+
+ copy_propagation_state* clone()
+ {
+ return new (ralloc_parent(this)) copy_propagation_state(this);
+ }
+
+ void erase_all()
+ {
+ _mesa_hash_table_clear(lhs_ht, NULL);
+ _mesa_hash_table_clear(rhs_ht, NULL);
+ }
+
+ void erase(ir_variable *var, unsigned write_mask)
+ {
+ /* removal of lhs entries */
+ hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, var);
+ if (ht_entry) {
+ exec_list *lhs_list = (exec_list *) ht_entry->data;
+ foreach_in_list_safe(acp_entry, entry, lhs_list) {
+ entry->write_mask = entry->write_mask & ~write_mask;
+ if (entry->write_mask == 0) {
+ entry->remove();
+ continue;
+ }
+ }
+ }
+
+ /* removal of rhs entries */
+ ht_entry = _mesa_hash_table_search(rhs_ht, var);
+ if (ht_entry) {
+ exec_list *rhs_list = (exec_list *) ht_entry->data;
+ acp_ref *ref;
+
+ while ((ref = (acp_ref *) rhs_list->pop_head()) != NULL) {
+ acp_entry *entry = ref->entry;
+
+ /* If entry is still in a list (not already removed by lhs entry
+ * removal above), remove it.
+ */
+ if (entry->prev || entry->next)
+ entry->remove();
+ }
+ }
+ }
+
+ exec_list *read(ir_variable *var)
+ {
+ hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, var);
+ if (ht_entry)
+ return (exec_list *) ht_entry->data;
+ return NULL;
+ }
+
+ void write(ir_variable *lhs, ir_variable *rhs, unsigned write_mask, int swizzle[4])
+ {
+ acp_entry *entry = new(this->lin_ctx) acp_entry(lhs, rhs, write_mask, swizzle);
+
+ /* lhs hash, hash of lhs -> acp_entry lists */
+ hash_entry *ht_entry = _mesa_hash_table_search(lhs_ht, lhs);
+ if (ht_entry) {
+ exec_list *lhs_list = (exec_list *) ht_entry->data;
+ lhs_list->push_tail(entry);
+ } else {
+ exec_list *lhs_list = new(mem_ctx) exec_list;
+ lhs_list->push_tail(entry);
+ _mesa_hash_table_insert(lhs_ht, lhs, lhs_list);
+ }
+
+ /* rhs hash, hash of rhs -> acp_entry pointers to lhs lists */
+ ht_entry = _mesa_hash_table_search(rhs_ht, rhs);
+ if (ht_entry) {
+ exec_list *rhs_list = (exec_list *) ht_entry->data;
+ rhs_list->push_tail(&entry->rhs_node);
+ } else {
+ exec_list *rhs_list = new(mem_ctx) exec_list;
+ rhs_list->push_tail(&entry->rhs_node);
+ _mesa_hash_table_insert(rhs_ht, rhs, rhs_list);
+ }
+ }
+
+private:
+ explicit copy_propagation_state(copy_propagation_state *original)
+ {
+ lhs_ht = _mesa_hash_table_clone(original->lhs_ht, this);
+ rhs_ht = _mesa_hash_table_clone(original->rhs_ht, this);
+ lin_ctx = original->lin_ctx;
+ mem_ctx = original->mem_ctx;
+ }
+
+ /** Hash of acp_entry: The available copies to propagate */
+ hash_table *lhs_ht;
+
+ /** Reverse index of the lhs_ht, to optimize finding uses of a certain variable. */
+ hash_table *rhs_ht;
+
+ void *mem_ctx;
+ void *lin_ctx;
+};
class kill_entry : public exec_node
{
public:
+ /* override operator new from exec_node */
+ DECLARE_LINEAR_ZALLOC_CXX_OPERATORS(kill_entry)
+
kill_entry(ir_variable *var, int write_mask)
{
this->var = var;
this->progress = false;
this->killed_all = false;
this->mem_ctx = ralloc_context(NULL);
+ this->lin_ctx = linear_alloc_parent(this->mem_ctx, 0);
this->shader_mem_ctx = NULL;
- this->acp = new(mem_ctx) exec_list;
this->kills = new(mem_ctx) exec_list;
+ this->state = new(mem_ctx) copy_propagation_state(mem_ctx, lin_ctx);
}
~ir_copy_propagation_elements_visitor()
{
ralloc_free(mem_ctx);
}
+ void handle_loop(ir_loop *, bool keep_acp);
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 *);
void kill(kill_entry *k);
void handle_if_block(exec_list *instructions);
- /** List of acp_entry: The available copies to propagate */
- exec_list *acp;
+ copy_propagation_state *state;
+
/**
* List of kill_entry: The variables whose values were killed in this
* block.
/* Context for our local data structures. */
void *mem_ctx;
+ void *lin_ctx;
/* Context for allocating new shader nodes. */
void *shader_mem_ctx;
};
* 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;
+ copy_propagation_state *orig_state = state;
+ this->state = new(mem_ctx) copy_propagation_state(mem_ctx, lin_ctx);
+
visit_list_elements(this, &ir->body);
- ralloc_free(this->acp);
- ralloc_free(this->kills);
+ delete this->state;
+ this->state = orig_state;
+ ralloc_free(this->kills);
this->kills = orig_kills;
- this->acp = orig_acp;
this->killed_all = orig_killed_all;
return visit_continue_with_parent;
kill_entry *k;
if (lhs)
- k = new(this->kills) kill_entry(var, ir->write_mask);
+ k = new(this->lin_ctx) kill_entry(var, ir->write_mask);
else
- k = new(this->kills) kill_entry(var, ~0);
+ k = new(this->lin_ctx) kill_entry(var, ~0);
kill(k);
}
/* Try to find ACP entries covering swizzle_chan[], hoping they're
* the same source variable.
*/
- foreach_in_list(acp_entry, entry, this->acp) {
- 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]];
+ exec_list *ht_list = state->read(var);
+ if (ht_list) {
+ foreach_in_list(acp_entry, entry, ht_list) {
+ 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]];
if (source_chan[c] != swizzle_chan[c])
noop_swizzle = false;
- }
- }
+ }
+ }
}
}
/* Since we're unlinked, we don't (necessarily) know the side effects of
* this call. So kill all copies.
*/
- acp->make_empty();
+ this->state->erase_all();
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_in_list(acp_entry, a, orig_acp) {
- this->acp->push_tail(new(this->acp) acp_entry(a));
- }
+ copy_propagation_state *orig_state = state;
+ this->state = orig_state->clone();
visit_list_elements(this, instructions);
- if (this->killed_all) {
- orig_acp->make_empty();
- }
+ delete this->state;
+ this->state = orig_state;
+
+ if (this->killed_all)
+ this->state->erase_all();
exec_list *new_kills = this->kills;
this->kills = orig_kills;
- ralloc_free(this->acp);
- 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
return visit_continue_with_parent;
}
-ir_visitor_status
-ir_copy_propagation_elements_visitor::visit_enter(ir_loop *ir)
+void
+ir_copy_propagation_elements_visitor::handle_loop(ir_loop *ir, bool keep_acp)
{
- 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);
+ copy_propagation_state *orig_state = state;
- if (this->killed_all) {
- orig_acp->make_empty();
+ if (keep_acp) {
+ /* Populate the initial acp with a copy of the original */
+ this->state = orig_state->clone();
+ } else {
+ this->state = new(mem_ctx) copy_propagation_state(mem_ctx, lin_ctx);
}
+ visit_list_elements(this, &ir->body_instructions);
+
+ delete this->state;
+ this->state = orig_state;
+
+ if (this->killed_all)
+ this->state->erase_all();
+
exec_list *new_kills = this->kills;
this->kills = orig_kills;
- ralloc_free(this->acp);
- this->acp = orig_acp;
this->killed_all = this->killed_all || orig_killed_all;
foreach_in_list_safe(kill_entry, k, new_kills) {
}
ralloc_free(new_kills);
+}
+
+ir_visitor_status
+ir_copy_propagation_elements_visitor::visit_enter(ir_loop *ir)
+{
+ handle_loop(ir, false);
+ handle_loop(ir, true);
/* already descended into the children. */
return visit_continue_with_parent;
void
ir_copy_propagation_elements_visitor::kill(kill_entry *k)
{
- foreach_in_list_safe(acp_entry, entry, acp) {
- 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();
- }
- }
+ state->erase(k->var, k->write_mask);
/* If we were on a list, remove ourselves before inserting */
if (k->next)
k->remove();
- ralloc_steal(this->kills, k);
this->kills->push_tail(k);
}
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 (!lhs || !(lhs->type->is_scalar() || lhs->type->is_vector()))
return;
+ if (lhs->var->data.mode == ir_var_shader_storage ||
+ lhs->var->data.mode == ir_var_shader_shared)
+ return;
+
ir_dereference_variable *rhs = ir->rhs->as_dereference_variable();
if (!rhs) {
ir_swizzle *swiz = ir->rhs->as_swizzle();
orig_swizzle[3] = swiz->mask.w;
}
+ if (rhs->var->data.mode == ir_var_shader_storage ||
+ rhs->var->data.mode == ir_var_shader_shared)
+ return;
+
/* 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.
}
}
- entry = new(this->mem_ctx) acp_entry(lhs->var, rhs->var, write_mask,
- swizzle);
- this->acp->push_tail(entry);
+ if (lhs->var->data.precise != rhs->var->data.precise)
+ return;
+
+ state->write(lhs->var, rhs->var, write_mask, swizzle);
}
bool
projects / mesa.git / blobdiff