---
src/glsl/nir/nir_split_var_copies.c | 96 ++++++++++++++++++++++++++++++-------
1 file changed, 78 insertions(+), 18 deletions(-)
diff --git a/src/glsl/nir/nir_split_var_copies.c
b/src/glsl/nir/nir_split_var_copies.c
index 70ce30a..4d663b5 100644
--- a/src/glsl/nir/nir_split_var_copies.c
+++ b/src/glsl/nir/nir_split_var_copies.c
@@ -74,40 +74,81 @@ get_deref_tail(nir_deref *deref)
return deref;
}
+/* Recursively constructs deref chains to split a copy instruction into
+ * multiple (if needed) copy instructions with full-length deref chains.
+ * External callers of this function should pass the tail and head of the
+ * deref chains found as the source and destination of the copy instruction
+ * into this function.
+ *
+ * \param old_copy The copy instruction we are splitting
+ * \param dest_head The head of the destination deref chain we are building
+ * \param src_head The head of the source deref chain we are building
+ * \param dest_tail The tail of the destination deref chain we are building
+ * \param src_tail The tail of the source deref chain we are building
+ * \param state The current split_var_copies_state object
+ */
static void
-nir_split_var_copy_instr(nir_intrinsic_instr *old_copy,
- nir_deref *dest_head, nir_deref *src_head,
- nir_deref *dest_tail, nir_deref *src_tail,
- struct split_var_copies_state *state)
+split_var_copy_instr(nir_intrinsic_instr *old_copy,
+ nir_deref *dest_head, nir_deref *src_head,
+ nir_deref *dest_tail, nir_deref *src_tail,
+ struct split_var_copies_state *state)
{
assert(src_tail->type == dest_tail->type);
+ /* Make sure these really are the tails of the deref chains */
+ assert(dest_tail->child == NULL);
+ assert(src_tail->child == NULL);
+
switch (glsl_get_base_type(src_tail->type)) {
case GLSL_TYPE_ARRAY: {
+ /* Make a wildcard dereference */
nir_deref_array *deref = nir_deref_array_create(state->dead_ctx);
deref->deref.type = glsl_get_array_element(src_tail->type);
deref->deref_array_type = nir_deref_array_type_wildcard;
+ /* Set the tail of both as the newly created wildcard deref. It is
+ * safe to use the same wildcard in both places because a) we will be
+ * copying it before we put it in an actual instruction and b)
+ * everything that will potentially add another link in the deref
+ * chain will also add the same thing to both chains.
+ */
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
- nir_split_var_copy_instr(old_copy, dest_head, src_head,
- dest_tail->child, src_tail->child, state);
+
+ split_var_copy_instr(old_copy, dest_head, src_head,
+ dest_tail->child, src_tail->child, state);
+
+ /* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
break;
}
case GLSL_TYPE_STRUCT:
+ /* This is the only part that actually does any interesting
+ * splitting. For array types, we just use wildcards and resolve
+ * them later. For structure types, we need to emit one copy
+ * instruction for every structure element. Because we may have
+ * structs inside structs, we just recurse and let the next level
+ * take care of any additional structures.
+ */
for (unsigned i = 0; i < glsl_get_length(src_tail->type); i++) {
nir_deref_struct *deref = nir_deref_struct_create(state->dead_ctx, i);
deref->deref.type = glsl_get_struct_field(src_tail->type, i);
+ /* Set the tail of both as the newly created structure deref. It
+ * is safe to use the same wildcard in both places because a) we
+ * will be copying it before we put it in an actual instruction
+ * and b) everything that will potentially add another link in the
+ * deref chain will also add the same thing to both chains.
+ */
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
- nir_split_var_copy_instr(old_copy, dest_head, src_head,
- dest_tail->child, src_tail->child, state);
+ split_var_copy_instr(old_copy, dest_head, src_head,
+ dest_tail->child, src_tail->child, state);
}
+ /* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
break;
@@ -121,22 +162,41 @@ nir_split_var_copy_instr(nir_intrinsic_instr *old_copy,
deref->deref.type = glsl_get_column_type(src_tail->type);
deref->deref_array_type = nir_deref_array_type_wildcard;
+ /* Set the tail of both as the newly created wildcard deref. It
+ * is safe to use the same wildcard in both places because a) we
+ * will be copying it before we put it in an actual instruction
+ * and b) everything that will potentially add another link in the
+ * deref chain will also add the same thing to both chains.
+ */
src_tail->child = &deref->deref;
dest_tail->child = &deref->deref;
- nir_split_var_copy_instr(old_copy, dest_head, src_head,
- dest_tail->child, src_tail->child, state);
+
+ split_var_copy_instr(old_copy, dest_head, src_head,
+ dest_tail->child, src_tail->child, state);
+
+ /* Set it back to the way we found it */
src_tail->child = NULL;
dest_tail->child = NULL;
} else {
+ /* At this point, we have fully built our deref chains and can
+ * actually add the new copy instruction.
+ */
nir_intrinsic_instr *new_copy =
nir_intrinsic_instr_create(state->mem_ctx, nir_intrinsic_copy_var);
+ /* We need to make copies because a) this deref chain actually
+ * belongs to the copy instruction and b) the deref chains may
+ * have some of the same links due to the way we constructed them
+ */
nir_deref *src = nir_copy_deref(state->mem_ctx, src_head);
nir_deref *dest = nir_copy_deref(state->mem_ctx, dest_head);
new_copy->variables[0] = nir_deref_as_var(dest);
new_copy->variables[1] = nir_deref_as_var(src);
+ /* Emit the copy instruction after the old instruction. We'll
+ * remove the old one later.
+ */
nir_instr_insert_after(&old_copy->instr, &new_copy->instr);
}
break;
@@ -151,7 +211,7 @@ nir_split_var_copy_instr(nir_intrinsic_instr *old_copy,
}
static bool
-nir_split_var_copies_block(nir_block *block, void *void_state)
+split_var_copies_block(nir_block *block, void *void_state)
{
struct split_var_copies_state *state = void_state;
@@ -171,8 +231,8 @@ nir_split_var_copies_block(nir_block *block, void
*void_state)
switch (glsl_get_base_type(src_tail->type)) {
case GLSL_TYPE_ARRAY:
case GLSL_TYPE_STRUCT:
- nir_split_var_copy_instr(intrinsic, dest_head, src_head,
- dest_tail, src_tail, state);
+ split_var_copy_instr(intrinsic, dest_head, src_head,
+ dest_tail, src_tail, state);
nir_instr_remove(&intrinsic->instr);
ralloc_steal(state->dead_ctx, instr);
break;
@@ -181,8 +241,8 @@ nir_split_var_copies_block(nir_block *block, void
*void_state)
case GLSL_TYPE_UINT:
case GLSL_TYPE_BOOL:
if (glsl_type_is_matrix(src_tail->type)) {
- nir_split_var_copy_instr(intrinsic, dest_head, src_head,
- dest_tail, src_tail, state);
+ split_var_copy_instr(intrinsic, dest_head, src_head,
+ dest_tail, src_tail, state);
nir_instr_remove(&intrinsic->instr);
ralloc_steal(state->dead_ctx, instr);
}
@@ -197,14 +257,14 @@ nir_split_var_copies_block(nir_block *block, void
*void_state)
}
static void
-nir_split_var_copies_impl(nir_function_impl *impl)
+split_var_copies_impl(nir_function_impl *impl)
{
struct split_var_copies_state state;
state.mem_ctx = ralloc_parent(impl);
state.dead_ctx = ralloc_context(NULL);
- nir_foreach_block(impl, nir_split_var_copies_block, &state);
+ nir_foreach_block(impl, split_var_copies_block, &state);
ralloc_free(state.dead_ctx);
}
@@ -214,6 +274,6 @@ nir_split_var_copies(nir_shader *shader)
{
nir_foreach_overload(shader, overload) {
if (overload->impl)
- nir_split_var_copies_impl(overload->impl);
+ split_var_copies_impl(overload->impl);
}
}
--
2.2.1
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
