#include "nir.h"
#include "nir_builder.h"
+#include "nir_deref.h"
+#include "nir_phi_builder.h"
#include "nir_vla.h"
bool lower_to_ssa;
/* Only valid for things that end up in the direct list.
- * Note that multiple nir_deref_vars may correspond to this node, but they
- * will all be equivalent, so any is as good as the other.
+ * Note that multiple nir_deref_instrs may correspond to this node, but
+ * they will all be equivalent, so any is as good as the other.
*/
- nir_deref_var *deref;
+ nir_deref_path path;
struct exec_node direct_derefs_link;
struct set *loads;
struct set *stores;
struct set *copies;
- nir_ssa_def **def_stack;
- nir_ssa_def **def_stack_tail;
+ struct nir_phi_builder_value *pb_value;
+
+ /* True if this node is fully direct. If set, it must be in the children
+ * array of its parent.
+ */
+ bool is_direct;
+
+ /* Set on a root node for a variable to indicate that variable is used by a
+ * cast or passed through some other sequence of instructions that are not
+ * derefs.
+ */
+ bool has_complex_use;
struct deref_node *wildcard;
struct deref_node *indirect;
struct deref_node *children[0];
};
+#define UNDEF_NODE ((struct deref_node *)(uintptr_t)1)
+
struct lower_variables_state {
nir_shader *shader;
void *dead_ctx;
*/
bool add_to_direct_deref_nodes;
- /* A hash table mapping phi nodes to deref_state data */
- struct hash_table *phi_table;
+ struct nir_phi_builder *phi_builder;
};
static struct deref_node *
deref_node_create(struct deref_node *parent,
- const struct glsl_type *type, nir_shader *shader)
+ const struct glsl_type *type,
+ bool is_direct, void *mem_ctx)
{
size_t size = sizeof(struct deref_node) +
glsl_get_length(type) * sizeof(struct deref_node *);
- struct deref_node *node = rzalloc_size(shader, size);
+ struct deref_node *node = rzalloc_size(mem_ctx, size);
node->type = type;
node->parent = parent;
- node->deref = NULL;
exec_node_init(&node->direct_derefs_link);
+ node->is_direct = is_direct;
return node;
}
if (var_entry) {
return var_entry->data;
} else {
- node = deref_node_create(NULL, var->type, state->dead_ctx);
+ node = deref_node_create(NULL, var->type, true, state->dead_ctx);
_mesa_hash_table_insert(state->deref_var_nodes, var, node);
return node;
}
* table of of fully-qualified direct derefs.
*/
static struct deref_node *
-get_deref_node(nir_deref_var *deref, struct lower_variables_state *state)
+get_deref_node_recur(nir_deref_instr *deref,
+ struct lower_variables_state *state)
{
- bool is_direct = true;
+ if (deref->deref_type == nir_deref_type_var)
+ return get_deref_node_for_var(deref->var, state);
- /* Start at the base of the chain. */
- struct deref_node *node = get_deref_node_for_var(deref->var, state);
- assert(deref->deref.type == node->type);
+ if (deref->deref_type == nir_deref_type_cast)
+ return NULL;
- for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) {
- switch (tail->deref_type) {
- case nir_deref_type_struct: {
- nir_deref_struct *deref_struct = nir_deref_as_struct(tail);
+ struct deref_node *parent =
+ get_deref_node_recur(nir_deref_instr_parent(deref), state);
+ if (parent == NULL)
+ return NULL;
- assert(deref_struct->index < glsl_get_length(node->type));
+ if (parent == UNDEF_NODE)
+ return UNDEF_NODE;
- if (node->children[deref_struct->index] == NULL)
- node->children[deref_struct->index] =
- deref_node_create(node, tail->type, state->dead_ctx);
+ switch (deref->deref_type) {
+ case nir_deref_type_struct:
+ assert(glsl_type_is_struct_or_ifc(parent->type));
+ assert(deref->strct.index < glsl_get_length(parent->type));
- node = node->children[deref_struct->index];
- break;
+ if (parent->children[deref->strct.index] == NULL) {
+ parent->children[deref->strct.index] =
+ deref_node_create(parent, deref->type, parent->is_direct,
+ state->dead_ctx);
}
- case nir_deref_type_array: {
- nir_deref_array *arr = nir_deref_as_array(tail);
-
- switch (arr->deref_array_type) {
- case nir_deref_array_type_direct:
- /* This is possible if a loop unrolls and generates an
- * out-of-bounds offset. We need to handle this at least
- * somewhat gracefully.
- */
- if (arr->base_offset >= glsl_get_length(node->type))
- return NULL;
+ return parent->children[deref->strct.index];
- if (node->children[arr->base_offset] == NULL)
- node->children[arr->base_offset] =
- deref_node_create(node, tail->type, state->dead_ctx);
+ case nir_deref_type_array: {
+ if (nir_src_is_const(deref->arr.index)) {
+ uint32_t index = nir_src_as_uint(deref->arr.index);
+ /* This is possible if a loop unrolls and generates an
+ * out-of-bounds offset. We need to handle this at least
+ * somewhat gracefully.
+ */
+ if (index >= glsl_get_length(parent->type))
+ return UNDEF_NODE;
- node = node->children[arr->base_offset];
- break;
+ if (parent->children[index] == NULL) {
+ parent->children[index] =
+ deref_node_create(parent, deref->type, parent->is_direct,
+ state->dead_ctx);
+ }
- case nir_deref_array_type_indirect:
- if (node->indirect == NULL)
- node->indirect = deref_node_create(node, tail->type,
- state->dead_ctx);
+ return parent->children[index];
+ } else {
+ if (parent->indirect == NULL) {
+ parent->indirect =
+ deref_node_create(parent, deref->type, false, state->dead_ctx);
+ }
- node = node->indirect;
- is_direct = false;
- break;
+ return parent->indirect;
+ }
+ break;
+ }
- case nir_deref_array_type_wildcard:
- if (node->wildcard == NULL)
- node->wildcard = deref_node_create(node, tail->type,
- state->dead_ctx);
+ case nir_deref_type_array_wildcard:
+ if (parent->wildcard == NULL) {
+ parent->wildcard =
+ deref_node_create(parent, deref->type, false, state->dead_ctx);
+ }
- node = node->wildcard;
- is_direct = false;
- break;
+ return parent->wildcard;
- default:
- unreachable("Invalid array deref type");
- }
- break;
- }
- default:
- unreachable("Invalid deref type");
- }
+ default:
+ unreachable("Invalid deref type");
}
+}
- assert(node);
+static struct deref_node *
+get_deref_node(nir_deref_instr *deref, struct lower_variables_state *state)
+{
+ /* This pass only works on local variables. Just ignore any derefs with
+ * a non-local mode.
+ */
+ if (deref->mode != nir_var_function_temp)
+ return NULL;
+
+ struct deref_node *node = get_deref_node_recur(deref, state);
+ if (!node)
+ return NULL;
- /* Only insert if it isn't already in the list. */
- if (is_direct && state->add_to_direct_deref_nodes &&
+ /* Insert the node in the direct derefs list. We only do this if it's not
+ * already in the list and we only bother for deref nodes which are used
+ * directly in a load or store.
+ */
+ if (node != UNDEF_NODE && node->is_direct &&
+ state->add_to_direct_deref_nodes &&
node->direct_derefs_link.next == NULL) {
- node->deref = deref;
+ nir_deref_path_init(&node->path, deref, state->dead_ctx);
assert(deref->var != NULL);
exec_list_push_tail(&state->direct_deref_nodes,
&node->direct_derefs_link);
}
/* \sa foreach_deref_node_match */
-static bool
-foreach_deref_node_worker(struct deref_node *node, nir_deref *deref,
- bool (* cb)(struct deref_node *node,
+static void
+foreach_deref_node_worker(struct deref_node *node, nir_deref_instr **path,
+ void (* cb)(struct deref_node *node,
struct lower_variables_state *state),
struct lower_variables_state *state)
{
- if (deref->child == NULL) {
- return cb(node, state);
- } else {
- switch (deref->child->deref_type) {
- case nir_deref_type_array: {
- nir_deref_array *arr = nir_deref_as_array(deref->child);
- assert(arr->deref_array_type == nir_deref_array_type_direct);
- if (node->children[arr->base_offset] &&
- !foreach_deref_node_worker(node->children[arr->base_offset],
- deref->child, cb, state))
- return false;
-
- if (node->wildcard &&
- !foreach_deref_node_worker(node->wildcard,
- deref->child, cb, state))
- return false;
+ if (*path == NULL) {
+ cb(node, state);
+ return;
+ }
- return true;
+ switch ((*path)->deref_type) {
+ case nir_deref_type_struct:
+ if (node->children[(*path)->strct.index]) {
+ foreach_deref_node_worker(node->children[(*path)->strct.index],
+ path + 1, cb, state);
}
+ return;
+
+ case nir_deref_type_array: {
+ uint32_t index = nir_src_as_uint((*path)->arr.index);
- case nir_deref_type_struct: {
- nir_deref_struct *str = nir_deref_as_struct(deref->child);
- return foreach_deref_node_worker(node->children[str->index],
- deref->child, cb, state);
+ if (node->children[index]) {
+ foreach_deref_node_worker(node->children[index],
+ path + 1, cb, state);
}
- default:
- unreachable("Invalid deref child type");
+ if (node->wildcard) {
+ foreach_deref_node_worker(node->wildcard,
+ path + 1, cb, state);
}
+ return;
+ }
+
+ default:
+ unreachable("Unsupported deref type");
}
}
* The given deref must be a full-length and fully qualified (no wildcards
* or indirects) deref chain.
*/
-static bool
-foreach_deref_node_match(nir_deref_var *deref,
- bool (* cb)(struct deref_node *node,
+static void
+foreach_deref_node_match(nir_deref_path *path,
+ void (* cb)(struct deref_node *node,
struct lower_variables_state *state),
struct lower_variables_state *state)
{
- nir_deref_var var_deref = *deref;
- var_deref.deref.child = NULL;
- struct deref_node *node = get_deref_node(&var_deref, state);
+ assert(path->path[0]->deref_type == nir_deref_type_var);
+ struct deref_node *node = get_deref_node_for_var(path->path[0]->var, state);
if (node == NULL)
- return false;
+ return;
- return foreach_deref_node_worker(node, &deref->deref, cb, state);
+ foreach_deref_node_worker(node, &path->path[1], cb, state);
}
/* \sa deref_may_be_aliased */
static bool
-deref_may_be_aliased_node(struct deref_node *node, nir_deref *deref,
- struct lower_variables_state *state)
+path_may_be_aliased_node(struct deref_node *node, nir_deref_instr **path,
+ struct lower_variables_state *state)
{
- if (deref->child == NULL) {
+ if (*path == NULL)
return false;
- } else {
- switch (deref->child->deref_type) {
- case nir_deref_type_array: {
- nir_deref_array *arr = nir_deref_as_array(deref->child);
- if (arr->deref_array_type == nir_deref_array_type_indirect)
- return true;
- /* If there is an indirect at this level, we're aliased. */
- if (node->indirect)
- return true;
+ switch ((*path)->deref_type) {
+ case nir_deref_type_struct:
+ if (node->children[(*path)->strct.index]) {
+ return path_may_be_aliased_node(node->children[(*path)->strct.index],
+ path + 1, state);
+ } else {
+ return false;
+ }
- assert(arr->deref_array_type == nir_deref_array_type_direct);
+ case nir_deref_type_array: {
+ if (!nir_src_is_const((*path)->arr.index))
+ return true;
- if (node->children[arr->base_offset] &&
- deref_may_be_aliased_node(node->children[arr->base_offset],
- deref->child, state))
- return true;
+ uint32_t index = nir_src_as_uint((*path)->arr.index);
- if (node->wildcard &&
- deref_may_be_aliased_node(node->wildcard, deref->child, state))
- return true;
+ /* If there is an indirect at this level, we're aliased. */
+ if (node->indirect)
+ return true;
- return false;
- }
+ if (node->children[index] &&
+ path_may_be_aliased_node(node->children[index],
+ path + 1, state))
+ return true;
- case nir_deref_type_struct: {
- nir_deref_struct *str = nir_deref_as_struct(deref->child);
- if (node->children[str->index]) {
- return deref_may_be_aliased_node(node->children[str->index],
- deref->child, state);
- } else {
- return false;
- }
- }
+ if (node->wildcard &&
+ path_may_be_aliased_node(node->wildcard, path + 1, state))
+ return true;
- default:
- unreachable("Invalid nir_deref child type");
- }
+ return false;
+ }
+
+ default:
+ unreachable("Unsupported deref type");
}
}
* references.
*/
static bool
-deref_may_be_aliased(nir_deref_var *deref,
+path_may_be_aliased(nir_deref_path *path,
+ struct lower_variables_state *state)
+{
+ assert(path->path[0]->deref_type == nir_deref_type_var);
+ nir_variable *var = path->path[0]->var;
+ struct deref_node *var_node = get_deref_node_for_var(var, state);
+
+ /* First see if this variable is ever used by anything other than a
+ * load/store. If there's even so much as a cast in the way, we have to
+ * assume aliasing and bail.
+ */
+ if (var_node->has_complex_use)
+ return true;
+
+ return path_may_be_aliased_node(var_node, &path->path[1], state);
+}
+
+static void
+register_complex_use(nir_deref_instr *deref,
struct lower_variables_state *state)
{
- return deref_may_be_aliased_node(get_deref_node_for_var(deref->var, state),
- &deref->deref, state);
+ assert(deref->deref_type == nir_deref_type_var);
+ struct deref_node *node = get_deref_node_for_var(deref->var, state);
+ if (node == NULL)
+ return;
+
+ node->has_complex_use = true;
}
static void
register_load_instr(nir_intrinsic_instr *load_instr,
struct lower_variables_state *state)
{
- struct deref_node *node = get_deref_node(load_instr->variables[0], state);
- if (node == NULL)
+ nir_deref_instr *deref = nir_src_as_deref(load_instr->src[0]);
+ struct deref_node *node = get_deref_node(deref, state);
+ if (node == NULL || node == UNDEF_NODE)
return;
if (node->loads == NULL)
- node->loads = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ node->loads = _mesa_pointer_set_create(state->dead_ctx);
_mesa_set_add(node->loads, load_instr);
}
register_store_instr(nir_intrinsic_instr *store_instr,
struct lower_variables_state *state)
{
- struct deref_node *node = get_deref_node(store_instr->variables[0], state);
- if (node == NULL)
+ nir_deref_instr *deref = nir_src_as_deref(store_instr->src[0]);
+ struct deref_node *node = get_deref_node(deref, state);
+ if (node == NULL || node == UNDEF_NODE)
return;
if (node->stores == NULL)
- node->stores = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ node->stores = _mesa_pointer_set_create(state->dead_ctx);
_mesa_set_add(node->stores, store_instr);
}
struct lower_variables_state *state)
{
for (unsigned idx = 0; idx < 2; idx++) {
- struct deref_node *node =
- get_deref_node(copy_instr->variables[idx], state);
-
- if (node == NULL)
+ nir_deref_instr *deref = nir_src_as_deref(copy_instr->src[idx]);
+ struct deref_node *node = get_deref_node(deref, state);
+ if (node == NULL || node == UNDEF_NODE)
continue;
if (node->copies == NULL)
- node->copies = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ node->copies = _mesa_pointer_set_create(state->dead_ctx);
_mesa_set_add(node->copies, copy_instr);
}
}
-/* Registers all variable uses in the given block. */
-static bool
-register_variable_uses_block(nir_block *block, void *void_state)
+static void
+register_variable_uses(nir_function_impl *impl,
+ struct lower_variables_state *state)
{
- struct lower_variables_state *state = void_state;
+ nir_foreach_block(block, impl) {
+ nir_foreach_instr_safe(instr, block) {
+ switch (instr->type) {
+ case nir_instr_type_deref: {
+ nir_deref_instr *deref = nir_instr_as_deref(instr);
- nir_foreach_instr_safe(block, instr) {
- if (instr->type != nir_instr_type_intrinsic)
- continue;
+ if (deref->deref_type == nir_deref_type_var &&
+ nir_deref_instr_has_complex_use(deref))
+ register_complex_use(deref, state);
+
+ break;
+ }
- nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+ case nir_instr_type_intrinsic: {
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
- switch (intrin->intrinsic) {
- case nir_intrinsic_load_var:
- register_load_instr(intrin, state);
- break;
+ switch (intrin->intrinsic) {
+ case nir_intrinsic_load_deref:
+ register_load_instr(intrin, state);
+ break;
- case nir_intrinsic_store_var:
- register_store_instr(intrin, state);
- break;
+ case nir_intrinsic_store_deref:
+ register_store_instr(intrin, state);
+ break;
- case nir_intrinsic_copy_var:
- register_copy_instr(intrin, state);
- break;
+ case nir_intrinsic_copy_deref:
+ register_copy_instr(intrin, state);
+ break;
- default:
- continue;
+ default:
+ continue;
+ }
+ break;
+ }
+
+ default:
+ break;
+ }
}
}
-
- return true;
}
/* Walks over all of the copy instructions to or from the given deref_node
* and lowers them to load/store intrinsics.
*/
-static bool
+static void
lower_copies_to_load_store(struct deref_node *node,
struct lower_variables_state *state)
{
if (!node->copies)
- return true;
+ return;
+
+ nir_builder b;
+ nir_builder_init(&b, state->impl);
- struct set_entry *copy_entry;
set_foreach(node->copies, copy_entry) {
nir_intrinsic_instr *copy = (void *)copy_entry->key;
- nir_lower_var_copy_instr(copy, state->shader);
+ nir_lower_deref_copy_instr(&b, copy);
for (unsigned i = 0; i < 2; ++i) {
- struct deref_node *arg_node =
- get_deref_node(copy->variables[i], state);
+ nir_deref_instr *arg_deref = nir_src_as_deref(copy->src[i]);
+ struct deref_node *arg_node = get_deref_node(arg_deref, state);
/* Only bother removing copy entries for other nodes */
if (arg_node == NULL || arg_node == node)
struct set_entry *arg_entry = _mesa_set_search(arg_node->copies, copy);
assert(arg_entry);
- _mesa_set_remove(node->copies, arg_entry);
+ _mesa_set_remove(arg_node->copies, arg_entry);
}
nir_instr_remove(©->instr);
}
node->copies = NULL;
-
- return true;
}
-/** Pushes an SSA def onto the def stack for the given node
- *
- * Each node is potentially associated with a stack of SSA definitions.
- * This stack is used for determining what SSA definition reaches a given
- * point in the program for variable renaming. The stack is always kept in
- * dominance-order with at most one SSA def per block. If the SSA
- * definition on the top of the stack is in the same block as the one being
- * pushed, the top element is replaced.
- */
-static void
-def_stack_push(struct deref_node *node, nir_ssa_def *def,
- struct lower_variables_state *state)
-{
- if (node->def_stack == NULL) {
- node->def_stack = ralloc_array(state->dead_ctx, nir_ssa_def *,
- state->impl->num_blocks);
- node->def_stack_tail = node->def_stack - 1;
- }
-
- if (node->def_stack_tail >= node->def_stack) {
- nir_ssa_def *top_def = *node->def_stack_tail;
-
- if (def->parent_instr->block == top_def->parent_instr->block) {
- /* They're in the same block, just replace the top */
- *node->def_stack_tail = def;
- return;
- }
- }
-
- *(++node->def_stack_tail) = def;
-}
-
-/* Pop the top of the def stack if it's in the given block */
-static void
-def_stack_pop_if_in_block(struct deref_node *node, nir_block *block)
-{
- /* If we're popping, then we have presumably pushed at some time in the
- * past so this should exist.
- */
- assert(node->def_stack != NULL);
-
- /* The stack is already empty. Do nothing. */
- if (node->def_stack_tail < node->def_stack)
- return;
-
- nir_ssa_def *def = *node->def_stack_tail;
- if (def->parent_instr->block == block)
- node->def_stack_tail--;
-}
-
-/** Retrieves the SSA definition on the top of the stack for the given
- * node, if one exists. If the stack is empty, then we return the constant
- * initializer (if it exists) or an SSA undef.
- */
-static nir_ssa_def *
-get_ssa_def_for_block(struct deref_node *node, nir_block *block,
- struct lower_variables_state *state)
-{
- /* If we have something on the stack, go ahead and return it. We're
- * assuming that the top of the stack dominates the given block.
- */
- if (node->def_stack && node->def_stack_tail >= node->def_stack)
- return *node->def_stack_tail;
-
- /* If we got here then we don't have a definition that dominates the
- * given block. This means that we need to add an undef and use that.
- */
- nir_ssa_undef_instr *undef =
- nir_ssa_undef_instr_create(state->shader,
- glsl_get_vector_elements(node->type));
- nir_instr_insert_before_cf_list(&state->impl->body, &undef->instr);
- def_stack_push(node, &undef->def, state);
- return &undef->def;
-}
-
-/* Given a block and one of its predecessors, this function fills in the
- * souces of the phi nodes to take SSA defs from the given predecessor.
- * This function must be called exactly once per block/predecessor pair.
- */
-static void
-add_phi_sources(nir_block *block, nir_block *pred,
- struct lower_variables_state *state)
-{
- nir_foreach_instr(block, instr) {
- if (instr->type != nir_instr_type_phi)
- break;
-
- nir_phi_instr *phi = nir_instr_as_phi(instr);
-
- struct hash_entry *entry =
- _mesa_hash_table_search(state->phi_table, phi);
- if (!entry)
- continue;
-
- struct deref_node *node = entry->data;
-
- nir_phi_src *src = ralloc(phi, nir_phi_src);
- src->pred = pred;
- src->src.parent_instr = &phi->instr;
- src->src.is_ssa = true;
- src->src.ssa = get_ssa_def_for_block(node, pred, state);
-
- list_addtail(&src->src.use_link, &src->src.ssa->uses);
-
- exec_list_push_tail(&phi->srcs, &src->node);
- }
-}
-
-/* Performs variable renaming by doing a DFS of the dominance tree
+/* Performs variable renaming
*
* This algorithm is very similar to the one outlined in "Efficiently
* Computing Static Single Assignment Form and the Control Dependence
- * Graph" by Cytron et. al. The primary difference is that we only put one
+ * Graph" by Cytron et al. The primary difference is that we only put one
* SSA def on the stack per block.
*/
static bool
-rename_variables_block(nir_block *block, struct lower_variables_state *state)
+rename_variables(struct lower_variables_state *state)
{
nir_builder b;
nir_builder_init(&b, state->impl);
- nir_foreach_instr_safe(block, instr) {
- if (instr->type == nir_instr_type_phi) {
- nir_phi_instr *phi = nir_instr_as_phi(instr);
-
- struct hash_entry *entry =
- _mesa_hash_table_search(state->phi_table, phi);
-
- /* This can happen if we already have phi nodes in the program
- * that were not created in this pass.
- */
- if (!entry)
+ nir_foreach_block(block, state->impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
continue;
- struct deref_node *node = entry->data;
-
- def_stack_push(node, &phi->dest.ssa, state);
- } else if (instr->type == nir_instr_type_intrinsic) {
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
switch (intrin->intrinsic) {
- case nir_intrinsic_load_var: {
- struct deref_node *node =
- get_deref_node(intrin->variables[0], state);
+ case nir_intrinsic_load_deref: {
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ if (deref->mode != nir_var_function_temp)
+ continue;
+
+ struct deref_node *node = get_deref_node(deref, state);
+ if (node == NULL)
+ continue;
- if (node == NULL) {
+ if (node == UNDEF_NODE) {
/* If we hit this path then we are referencing an invalid
* value. Most likely, we unrolled something and are
* reading past the end of some array. In any case, this
*/
nir_ssa_undef_instr *undef =
nir_ssa_undef_instr_create(state->shader,
- intrin->num_components);
+ intrin->num_components,
+ intrin->dest.ssa.bit_size);
nir_instr_insert_before(&intrin->instr, &undef->instr);
nir_instr_remove(&intrin->instr);
continue;
nir_alu_instr *mov = nir_alu_instr_create(state->shader,
- nir_op_imov);
- mov->src[0].src.is_ssa = true;
- mov->src[0].src.ssa = get_ssa_def_for_block(node, block, state);
- for (unsigned i = intrin->num_components; i < 4; i++)
+ nir_op_mov);
+ mov->src[0].src = nir_src_for_ssa(
+ nir_phi_builder_value_get_block_def(node->pb_value, block));
+ for (unsigned i = intrin->num_components; i < NIR_MAX_VEC_COMPONENTS; i++)
mov->src[0].swizzle[i] = 0;
assert(intrin->dest.is_ssa);
break;
}
- case nir_intrinsic_store_var: {
- struct deref_node *node =
- get_deref_node(intrin->variables[0], state);
+ case nir_intrinsic_store_deref: {
+ nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
+ if (deref->mode != nir_var_function_temp)
+ continue;
+
+ struct deref_node *node = get_deref_node(deref, state);
+ if (node == NULL)
+ continue;
+
+ assert(intrin->src[1].is_ssa);
+ nir_ssa_def *value = intrin->src[1].ssa;
- if (node == NULL) {
+ if (node == UNDEF_NODE) {
/* Probably an out-of-bounds array store. That should be a
* no-op. */
nir_instr_remove(&intrin->instr);
assert(intrin->num_components ==
glsl_get_vector_elements(node->type));
- assert(intrin->src[0].is_ssa);
-
nir_ssa_def *new_def;
b.cursor = nir_before_instr(&intrin->instr);
unsigned wrmask = nir_intrinsic_write_mask(intrin);
if (wrmask == (1 << intrin->num_components) - 1) {
/* Whole variable store - just copy the source. Note that
- * intrin->num_components and intrin->src[0].ssa->num_components
+ * intrin->num_components and value->num_components
* may differ.
*/
- unsigned swiz[4];
- for (unsigned i = 0; i < 4; i++)
+ unsigned swiz[NIR_MAX_VEC_COMPONENTS];
+ for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++)
swiz[i] = i < intrin->num_components ? i : 0;
- new_def = nir_swizzle(&b, intrin->src[0].ssa, swiz,
- intrin->num_components, false);
+ new_def = nir_swizzle(&b, value, swiz,
+ intrin->num_components);
} else {
- nir_ssa_def *old_def = get_ssa_def_for_block(node, block, state);
+ nir_ssa_def *old_def =
+ nir_phi_builder_value_get_block_def(node->pb_value, block);
/* For writemasked store_var intrinsics, we combine the newly
* written values with the existing contents of unwritten
* channels, creating a new SSA value for the whole vector.
*/
- nir_ssa_def *srcs[4];
+ nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS];
for (unsigned i = 0; i < intrin->num_components; i++) {
if (wrmask & (1 << i)) {
- srcs[i] = nir_channel(&b, intrin->src[0].ssa, i);
+ srcs[i] = nir_channel(&b, value, i);
} else {
srcs[i] = nir_channel(&b, old_def, i);
}
assert(new_def->num_components == intrin->num_components);
- def_stack_push(node, new_def, state);
-
- /* We'll wait to remove the instruction until the next pass
- * where we pop the node we just pushed back off the stack.
- */
+ nir_phi_builder_value_set_block_def(node->pb_value, block, new_def);
+ nir_instr_remove(&intrin->instr);
break;
}
}
}
- if (block->successors[0])
- add_phi_sources(block->successors[0], block, state);
- if (block->successors[1])
- add_phi_sources(block->successors[1], block, state);
-
- for (unsigned i = 0; i < block->num_dom_children; ++i)
- rename_variables_block(block->dom_children[i], state);
-
- /* Now we iterate over the instructions and pop off any SSA defs that we
- * pushed in the first loop.
- */
- nir_foreach_instr_safe(block, instr) {
- if (instr->type == nir_instr_type_phi) {
- nir_phi_instr *phi = nir_instr_as_phi(instr);
-
- struct hash_entry *entry =
- _mesa_hash_table_search(state->phi_table, phi);
-
- /* This can happen if we already have phi nodes in the program
- * that were not created in this pass.
- */
- if (!entry)
- continue;
-
- struct deref_node *node = entry->data;
-
- def_stack_pop_if_in_block(node, block);
- } else if (instr->type == nir_instr_type_intrinsic) {
- nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
-
- if (intrin->intrinsic != nir_intrinsic_store_var)
- continue;
-
- struct deref_node *node = get_deref_node(intrin->variables[0], state);
- if (!node)
- continue;
-
- if (!node->lower_to_ssa)
- continue;
-
- def_stack_pop_if_in_block(node, block);
- nir_instr_remove(&intrin->instr);
- }
- }
-
return true;
}
-/* Inserts phi nodes for all variables marked lower_to_ssa
- *
- * This is the same algorithm as presented in "Efficiently Computing Static
- * Single Assignment Form and the Control Dependence Graph" by Cytron et.
- * al.
- */
-static void
-insert_phi_nodes(struct lower_variables_state *state)
-{
- NIR_VLA_ZERO(unsigned, work, state->impl->num_blocks);
- NIR_VLA_ZERO(unsigned, has_already, state->impl->num_blocks);
-
- /*
- * Since the work flags already prevent us from inserting a node that has
- * ever been inserted into W, we don't need to use a set to represent W.
- * Also, since no block can ever be inserted into W more than once, we know
- * that the maximum size of W is the number of basic blocks in the
- * function. So all we need to handle W is an array and a pointer to the
- * next element to be inserted and the next element to be removed.
- */
- NIR_VLA(nir_block *, W, state->impl->num_blocks);
-
- unsigned w_start, w_end;
- unsigned iter_count = 0;
-
- foreach_list_typed(struct deref_node, node, direct_derefs_link,
- &state->direct_deref_nodes) {
- if (node->stores == NULL)
- continue;
-
- if (!node->lower_to_ssa)
- continue;
-
- unsigned bit_size = glsl_get_bit_size(glsl_get_base_type(node->type));
-
- w_start = w_end = 0;
- iter_count++;
-
- struct set_entry *store_entry;
- set_foreach(node->stores, store_entry) {
- nir_intrinsic_instr *store = (nir_intrinsic_instr *)store_entry->key;
- if (work[store->instr.block->index] < iter_count)
- W[w_end++] = store->instr.block;
- work[store->instr.block->index] = iter_count;
- }
-
- while (w_start != w_end) {
- nir_block *cur = W[w_start++];
- struct set_entry *dom_entry;
- set_foreach(cur->dom_frontier, dom_entry) {
- nir_block *next = (nir_block *) dom_entry->key;
-
- /*
- * If there's more than one return statement, then the end block
- * can be a join point for some definitions. However, there are
- * no instructions in the end block, so nothing would use those
- * phi nodes. Of course, we couldn't place those phi nodes
- * anyways due to the restriction of having no instructions in the
- * end block...
- */
- if (next == state->impl->end_block)
- continue;
-
- if (has_already[next->index] < iter_count) {
- nir_phi_instr *phi = nir_phi_instr_create(state->shader);
- nir_ssa_dest_init(&phi->instr, &phi->dest,
- glsl_get_vector_elements(node->type),
- bit_size, NULL);
- nir_instr_insert_before_block(next, &phi->instr);
-
- _mesa_hash_table_insert(state->phi_table, phi, node);
-
- has_already[next->index] = iter_count;
- if (work[next->index] < iter_count) {
- work[next->index] = iter_count;
- W[w_end++] = next;
- }
- }
- }
- }
- }
-}
-
-
/** Implements a pass to lower variable uses to SSA values
*
* This path walks the list of instructions and tries to lower as many
* fully-direct references we see and store them in the
* direct_deref_nodes hash table.
*
- * 2) Walk over the the list of fully-qualified direct derefs generated in
+ * 2) Walk over the list of fully-qualified direct derefs generated in
* the previous pass. For each deref, we determine if it can ever be
* aliased, i.e. if there is an indirect reference anywhere that may
* refer to it. If it cannot be aliased, we mark it for lowering to an
* SSA value. At this point, we lower any var_copy instructions that
- * use the given deref to load/store operations and, if the deref has a
- * constant initializer, we go ahead and add a load_const value at the
- * beginning of the function with the initialized value.
+ * use the given deref to load/store operations.
*
* 3) Walk over the list of derefs we plan to lower to SSA values and
* insert phi nodes as needed.
state.dead_ctx = ralloc_context(state.shader);
state.impl = impl;
- state.deref_var_nodes = _mesa_hash_table_create(state.dead_ctx,
- _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ state.deref_var_nodes = _mesa_pointer_hash_table_create(state.dead_ctx);
exec_list_make_empty(&state.direct_deref_nodes);
- state.phi_table = _mesa_hash_table_create(state.dead_ctx,
- _mesa_hash_pointer,
- _mesa_key_pointer_equal);
/* Build the initial deref structures and direct_deref_nodes table */
state.add_to_direct_deref_nodes = true;
- nir_foreach_block(impl, register_variable_uses_block, &state);
+
+ register_variable_uses(impl, &state);
bool progress = false;
foreach_list_typed_safe(struct deref_node, node, direct_derefs_link,
&state.direct_deref_nodes) {
- nir_deref_var *deref = node->deref;
+ nir_deref_path *path = &node->path;
- if (deref->var->data.mode != nir_var_local) {
- exec_node_remove(&node->direct_derefs_link);
- continue;
- }
+ assert(path->path[0]->deref_type == nir_deref_type_var);
- if (deref_may_be_aliased(deref, &state)) {
+ /* We don't build deref nodes for non-local variables */
+ assert(path->path[0]->var->data.mode == nir_var_function_temp);
+
+ if (path_may_be_aliased(path, &state)) {
exec_node_remove(&node->direct_derefs_link);
continue;
}
node->lower_to_ssa = true;
progress = true;
- if (deref->var->constant_initializer) {
- nir_load_const_instr *load =
- nir_deref_get_const_initializer_load(state.shader, deref);
- nir_ssa_def_init(&load->instr, &load->def,
- glsl_get_vector_elements(node->type),
- glsl_get_bit_size(glsl_get_base_type(node->type)),
- NULL);
- nir_instr_insert_before_cf_list(&impl->body, &load->instr);
- def_stack_push(node, &load->def, &state);
- }
-
- foreach_deref_node_match(deref, lower_copies_to_load_store, &state);
+ foreach_deref_node_match(path, lower_copies_to_load_store, &state);
}
- if (!progress)
+ if (!progress) {
+ nir_metadata_preserve(impl, nir_metadata_all);
return false;
+ }
nir_metadata_require(impl, nir_metadata_dominance);
* added load/store instructions are registered. We need this
* information for phi node insertion below.
*/
- nir_foreach_block(impl, register_variable_uses_block, &state);
+ register_variable_uses(impl, &state);
+
+ state.phi_builder = nir_phi_builder_create(state.impl);
+
+ BITSET_WORD *store_blocks =
+ ralloc_array(state.dead_ctx, BITSET_WORD,
+ BITSET_WORDS(state.impl->num_blocks));
+ foreach_list_typed(struct deref_node, node, direct_derefs_link,
+ &state.direct_deref_nodes) {
+ if (!node->lower_to_ssa)
+ continue;
+
+ memset(store_blocks, 0,
+ BITSET_WORDS(state.impl->num_blocks) * sizeof(*store_blocks));
+
+ assert(node->path.path[0]->var->constant_initializer == NULL &&
+ node->path.path[0]->var->pointer_initializer == NULL);
+
+ if (node->stores) {
+ set_foreach(node->stores, store_entry) {
+ nir_intrinsic_instr *store =
+ (nir_intrinsic_instr *)store_entry->key;
+ BITSET_SET(store_blocks, store->instr.block->index);
+ }
+ }
- insert_phi_nodes(&state);
- rename_variables_block(nir_start_block(impl), &state);
+ node->pb_value =
+ nir_phi_builder_add_value(state.phi_builder,
+ glsl_get_vector_elements(node->type),
+ glsl_get_bit_size(node->type),
+ store_blocks);
+ }
+
+ rename_variables(&state);
+
+ nir_phi_builder_finish(state.phi_builder);
nir_metadata_preserve(impl, nir_metadata_block_index |
nir_metadata_dominance);
return progress;
}
-void
+bool
nir_lower_vars_to_ssa(nir_shader *shader)
{
- nir_foreach_function(shader, function) {
+ bool progress = false;
+
+ nir_foreach_function(function, shader) {
if (function->impl)
- nir_lower_vars_to_ssa_impl(function->impl);
+ progress |= nir_lower_vars_to_ssa_impl(function->impl);
}
+
+ return progress;
}