nir_var_mem_shared;
break;
+ case nir_intrinsic_scoped_memory_barrier:
+ if (nir_intrinsic_memory_semantics(intrin) & NIR_MEMORY_ACQUIRE)
+ written->modes |= nir_intrinsic_memory_modes(intrin);
+ break;
+
case nir_intrinsic_emit_vertex:
case nir_intrinsic_emit_vertex_with_counter:
written->modes = nir_var_shader_out;
*entry = util_dynarray_pop(copies, struct copy_entry);
}
+static bool
+is_array_deref_of_vector(nir_deref_instr *deref)
+{
+ if (deref->deref_type != nir_deref_type_array)
+ return false;
+ nir_deref_instr *parent = nir_deref_instr_parent(deref);
+ return glsl_type_is_vector(parent->type);
+}
+
static struct copy_entry *
lookup_entry_for_deref(struct util_dynarray *copies,
nir_deref_instr *deref,
nir_deref_compare_result allowed_comparisons)
{
+ struct copy_entry *entry = NULL;
util_dynarray_foreach(copies, struct copy_entry, iter) {
- if (nir_compare_derefs(iter->dst, deref) & allowed_comparisons)
- return iter;
+ nir_deref_compare_result result = nir_compare_derefs(iter->dst, deref);
+ if (result & allowed_comparisons) {
+ entry = iter;
+ if (result & nir_derefs_equal_bit)
+ break;
+ /* Keep looking in case we have an equal match later in the array. */
+ }
}
-
- return NULL;
+ return entry;
}
static struct copy_entry *
static void
value_set_from_value(struct value *value, const struct value *from,
- unsigned write_mask)
+ unsigned base_index, unsigned write_mask)
{
+ /* We can't have non-zero indexes with non-trivial write masks */
+ assert(base_index == 0 || write_mask == 1);
+
if (from->is_ssa) {
/* Clear value if it was being used as non-SSA. */
if (!value->is_ssa)
/* Only overwrite the written components */
for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
if (write_mask & (1 << i)) {
- value->ssa.def[i] = from->ssa.def[i];
- value->ssa.component[i] = from->ssa.component[i];
+ value->ssa.def[base_index + i] = from->ssa.def[i];
+ value->ssa.component[base_index + i] = from->ssa.component[i];
}
}
} else {
}
}
+/* Try to load a single element of a vector from the copy_entry. If the data
+ * isn't available, just let the original intrinsic do the work.
+ */
+static bool
+load_element_from_ssa_entry_value(struct copy_prop_var_state *state,
+ struct copy_entry *entry,
+ nir_builder *b, nir_intrinsic_instr *intrin,
+ struct value *value, unsigned index)
+{
+ assert(index < glsl_get_vector_elements(entry->dst->type));
+
+ /* We don't have the element available, so let the instruction do the work. */
+ if (!entry->src.ssa.def[index])
+ return false;
+
+ b->cursor = nir_instr_remove(&intrin->instr);
+ intrin->instr.block = NULL;
+
+ assert(entry->src.ssa.component[index] <
+ entry->src.ssa.def[index]->num_components);
+ nir_ssa_def *def = nir_channel(b, entry->src.ssa.def[index],
+ entry->src.ssa.component[index]);
+
+ *value = (struct value) {
+ .is_ssa = true,
+ {
+ .ssa = {
+ .def = { def },
+ .component = { 0 },
+ },
+ }
+ };
+
+ return true;
+}
+
/* Do a "load" from an SSA-based entry return it in "value" as a value with a
* single SSA def. Because an entry could reference multiple different SSA
* defs, a vecN operation may be inserted to combine them into a single SSA
load_from_ssa_entry_value(struct copy_prop_var_state *state,
struct copy_entry *entry,
nir_builder *b, nir_intrinsic_instr *intrin,
- struct value *value)
+ nir_deref_instr *src, struct value *value)
{
+ if (is_array_deref_of_vector(src)) {
+ if (nir_src_is_const(src->arr.index)) {
+ return load_element_from_ssa_entry_value(state, entry, b, intrin, value,
+ nir_src_as_uint(src->arr.index));
+ }
+
+ /* An SSA copy_entry for the vector won't help indirect load. */
+ if (glsl_type_is_vector(entry->dst->type)) {
+ assert(entry->dst->type == nir_deref_instr_parent(src)->type);
+ /* TODO: If all SSA entries are there, try an if-ladder. */
+ return false;
+ }
+ }
+
*value = entry->src;
assert(value->is_ssa);
return false;
if (entry->src.is_ssa) {
- return load_from_ssa_entry_value(state, entry, b, intrin, value);
+ return load_from_ssa_entry_value(state, entry, b, intrin, src, value);
} else {
return load_from_deref_entry_value(state, entry, b, intrin, src, value);
}
}
}
-static bool
-is_array_deref_of_vector(nir_deref_instr *deref)
-{
- if (deref->deref_type != nir_deref_type_array)
- return false;
- nir_deref_instr *parent = nir_deref_instr_parent(deref);
- return glsl_type_is_vector(parent->type);
-}
-
static void
print_value(struct value *value, unsigned num_components)
{
nir_var_mem_shared);
break;
+ case nir_intrinsic_scoped_memory_barrier:
+ if (debug) dump_instr(instr);
+
+ if (nir_intrinsic_memory_semantics(intrin) & NIR_MEMORY_ACQUIRE)
+ apply_barrier_for_modes(copies, nir_intrinsic_memory_modes(intrin));
+ break;
+
case nir_intrinsic_emit_vertex:
case nir_intrinsic_emit_vertex_with_counter:
if (debug) dump_instr(instr);
case nir_intrinsic_load_deref: {
if (debug) dump_instr(instr);
+ if (nir_intrinsic_access(intrin) & ACCESS_VOLATILE)
+ break;
+
nir_deref_instr *src = nir_src_as_deref(intrin->src[0]);
- if (is_array_deref_of_vector(src)) {
- /* Not handled yet. This load won't invalidate existing copies. */
- break;
+ /* Direct array_derefs of vectors operate on the vectors (the parent
+ * deref). Indirects will be handled like other derefs.
+ */
+ int vec_index = 0;
+ nir_deref_instr *vec_src = src;
+ if (is_array_deref_of_vector(src) && nir_src_is_const(src->arr.index)) {
+ vec_src = nir_deref_instr_parent(src);
+ unsigned vec_comps = glsl_get_vector_elements(vec_src->type);
+ vec_index = nir_src_as_uint(src->arr.index);
+
+ /* Loading from an invalid index yields an undef */
+ if (vec_index >= vec_comps) {
+ b->cursor = nir_instr_remove(instr);
+ nir_ssa_def *u = nir_ssa_undef(b, 1, intrin->dest.ssa.bit_size);
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(u));
+ break;
+ }
}
struct copy_entry *src_entry =
if (value.is_ssa) {
/* lookup_load has already ensured that we get a single SSA
* value that has all of the channels. We just have to do the
- * rewrite operation.
+ * rewrite operation. Note for array derefs of vectors, the
+ * channel 0 is used.
*/
if (intrin->instr.block) {
/* The lookup left our instruction in-place. This means it
* contains what we're looking for.
*/
struct copy_entry *entry =
- lookup_entry_for_deref(copies, src, nir_derefs_equal_bit);
+ lookup_entry_for_deref(copies, vec_src, nir_derefs_equal_bit);
if (!entry)
- entry = copy_entry_create(copies, src);
+ entry = copy_entry_create(copies, vec_src);
/* Update the entry with the value of the load. This way
* we can potentially remove subsequent loads.
*/
- value_set_from_value(&entry->src, &value,
+ value_set_from_value(&entry->src, &value, vec_index,
(1 << intrin->num_components) - 1);
break;
}
case nir_intrinsic_store_deref: {
if (debug) dump_instr(instr);
+ if (nir_intrinsic_access(intrin) & ACCESS_VOLATILE)
+ break;
+
nir_deref_instr *dst = nir_src_as_deref(intrin->src[0]);
+ assert(glsl_type_is_vector_or_scalar(dst->type));
+
+ /* Direct array_derefs of vectors operate on the vectors (the parent
+ * deref). Indirects will be handled like other derefs.
+ */
+ int vec_index = 0;
+ nir_deref_instr *vec_dst = dst;
+ if (is_array_deref_of_vector(dst) && nir_src_is_const(dst->arr.index)) {
+ vec_dst = nir_deref_instr_parent(dst);
+ unsigned vec_comps = glsl_get_vector_elements(vec_dst->type);
+
+ vec_index = nir_src_as_uint(dst->arr.index);
+
+ /* Storing to an invalid index is a no-op. */
+ if (vec_index >= vec_comps) {
+ nir_instr_remove(instr);
+ break;
+ }
+ }
+
struct copy_entry *entry =
lookup_entry_for_deref(copies, dst, nir_derefs_equal_bit);
if (entry && value_equals_store_src(&entry->src, intrin)) {
* store is redundant so remove it.
*/
nir_instr_remove(instr);
- } else if (is_array_deref_of_vector(dst)) {
- /* Not handled yet. Writing into an element of 'dst' invalidates
- * any related entries in copies.
- */
- nir_deref_instr *vector = nir_deref_instr_parent(dst);
- unsigned vector_components = glsl_get_vector_elements(vector->type);
- kill_aliases(copies, vector, (1 << vector_components) - 1);
} else {
struct value value = {0};
value_set_ssa_components(&value, intrin->src[1].ssa,
intrin->num_components);
unsigned wrmask = nir_intrinsic_write_mask(intrin);
struct copy_entry *entry =
- get_entry_and_kill_aliases(copies, dst, wrmask);
- value_set_from_value(&entry->src, &value, wrmask);
+ get_entry_and_kill_aliases(copies, vec_dst, wrmask);
+ value_set_from_value(&entry->src, &value, vec_index, wrmask);
}
break;
case nir_intrinsic_copy_deref: {
if (debug) dump_instr(instr);
+ if ((nir_intrinsic_src_access(intrin) & ACCESS_VOLATILE) ||
+ (nir_intrinsic_dst_access(intrin) & ACCESS_VOLATILE))
+ break;
+
nir_deref_instr *dst = nir_src_as_deref(intrin->src[0]);
nir_deref_instr *src = nir_src_as_deref(intrin->src[1]);
unsigned num_components = glsl_get_vector_elements(dst->type);
unsigned full_mask = (1 << num_components) - 1;
- if (is_array_deref_of_vector(src) || is_array_deref_of_vector(dst)) {
- /* Cases not handled yet. Writing into an element of 'dst'
- * invalidates any related entries in copies. Reading from 'src'
- * doesn't invalidate anything, so no action needed for it.
- */
+ /* Copy of direct array derefs of vectors are not handled. Just
+ * invalidate what's written and bail.
+ */
+ if ((is_array_deref_of_vector(src) && nir_src_is_const(src->arr.index)) ||
+ (is_array_deref_of_vector(dst) && nir_src_is_const(dst->arr.index))) {
kill_aliases(copies, dst, full_mask);
break;
}
};
}
+ nir_variable *src_var = nir_deref_instr_get_variable(src);
+ if (src_var && src_var->data.cannot_coalesce) {
+ /* The source cannot be coaleseced, which means we can't propagate
+ * this copy.
+ */
+ break;
+ }
+
struct copy_entry *dst_entry =
get_entry_and_kill_aliases(copies, dst, full_mask);
- value_set_from_value(&dst_entry->src, &value, full_mask);
+ value_set_from_value(&dst_entry->src, &value, 0, full_mask);
break;
}
case nir_intrinsic_deref_atomic_comp_swap:
if (debug) dump_instr(instr);
+ if (nir_intrinsic_access(intrin) & ACCESS_VOLATILE)
+ break;
+
nir_deref_instr *dst = nir_src_as_deref(intrin->src[0]);
unsigned num_components = glsl_get_vector_elements(dst->type);
unsigned full_mask = (1 << num_components) - 1;
projects / mesa.git / blobdiff