/* Flags used in the instr->pass_flags field for various instruction states */
enum {
- GCM_INSTR_PINNED = (1 << 0),
- GCM_INSTR_SCHEDULED_EARLY = (1 << 1),
- GCM_INSTR_SCHEDULED_LATE = (1 << 2),
- GCM_INSTR_PLACED = (1 << 3),
+ GCM_INSTR_PINNED = (1 << 0),
+ GCM_INSTR_SCHEDULE_EARLIER_ONLY = (1 << 1),
+ GCM_INSTR_SCHEDULED_EARLY = (1 << 2),
+ GCM_INSTR_SCHEDULED_LATE = (1 << 3),
+ GCM_INSTR_PLACED = (1 << 4),
};
struct gcm_state {
}
}
+static bool
+is_src_scalarizable(nir_src *src)
+{
+ assert(src->is_ssa);
+
+ nir_instr *src_instr = src->ssa->parent_instr;
+ switch (src_instr->type) {
+ case nir_instr_type_alu: {
+ nir_alu_instr *src_alu = nir_instr_as_alu(src_instr);
+
+ /* ALU operations with output_size == 0 should be scalarized. We
+ * will also see a bunch of vecN operations from scalarizing ALU
+ * operations and, since they can easily be copy-propagated, they
+ * are ok too.
+ */
+ return nir_op_infos[src_alu->op].output_size == 0 ||
+ src_alu->op == nir_op_vec2 ||
+ src_alu->op == nir_op_vec3 ||
+ src_alu->op == nir_op_vec4;
+ }
+
+ case nir_instr_type_load_const:
+ /* These are trivially scalarizable */
+ return true;
+
+ case nir_instr_type_ssa_undef:
+ return true;
+
+ case nir_instr_type_intrinsic: {
+ nir_intrinsic_instr *src_intrin = nir_instr_as_intrinsic(src_instr);
+
+ switch (src_intrin->intrinsic) {
+ case nir_intrinsic_load_deref: {
+ nir_deref_instr *deref = nir_src_as_deref(src_intrin->src[0]);
+ return deref->mode == nir_var_shader_in ||
+ deref->mode == nir_var_uniform ||
+ deref->mode == nir_var_mem_ubo ||
+ deref->mode == nir_var_mem_ssbo ||
+ deref->mode == nir_var_mem_global;
+ }
+
+ case nir_intrinsic_interp_deref_at_centroid:
+ case nir_intrinsic_interp_deref_at_sample:
+ case nir_intrinsic_interp_deref_at_offset:
+ case nir_intrinsic_load_uniform:
+ case nir_intrinsic_load_ubo:
+ case nir_intrinsic_load_ssbo:
+ case nir_intrinsic_load_global:
+ case nir_intrinsic_load_global_constant:
+ case nir_intrinsic_load_input:
+ return true;
+ default:
+ break;
+ }
+
+ return false;
+ }
+
+ default:
+ /* We can't scalarize this type of instruction */
+ return false;
+ }
+}
+
/* Walks the instruction list and marks immovable instructions as pinned
*
* This function also serves to initialize the instr->pass_flags field.
case nir_op_fddy_fine:
case nir_op_fddx_coarse:
case nir_op_fddy_coarse:
- /* These can only go in uniform control flow; pin them for now */
- instr->pass_flags = GCM_INSTR_PINNED;
+ /* These can only go in uniform control flow */
+ instr->pass_flags = GCM_INSTR_SCHEDULE_EARLIER_ONLY;
break;
+ case nir_op_mov:
+ if (!is_src_scalarizable(&(nir_instr_as_alu(instr)->src[0].src))) {
+ instr->pass_flags = GCM_INSTR_PINNED;
+ break;
+ }
+ /* fallthrough */
+
default:
instr->pass_flags = 0;
break;
case nir_instr_type_tex:
if (nir_tex_instr_has_implicit_derivative(nir_instr_as_tex(instr)))
- instr->pass_flags = GCM_INSTR_PINNED;
+ instr->pass_flags = GCM_INSTR_SCHEDULE_EARLIER_ONLY;
break;
case nir_instr_type_deref:
nir_block *best = late_block;
for (nir_block *block = late_block; block != NULL; block = block->imm_dom) {
+ /* Being too aggressive with how we pull instructions out of loops can
+ * result in extra register pressure and spilling. For example its fairly
+ * common for loops in compute shaders to calculate SSBO offsets using
+ * the workgroup id, subgroup id and subgroup invocation, pulling all
+ * these calculations outside the loop causes register pressure.
+ *
+ * To work around these issues for now we only allow constant and texture
+ * instructions to be moved outside their original loops.
+ *
+ * TODO: figure out some heuristics to allow more to be moved out of loops.
+ */
if (state->blocks[block->index].loop_depth <
- state->blocks[best->index].loop_depth)
+ state->blocks[best->index].loop_depth &&
+ (nir_block_dominates(instr->block, block) ||
+ instr->type == nir_instr_type_load_const ||
+ instr->type == nir_instr_type_tex))
+ best = block;
+ else if (block == instr->block)
best = block;
if (block == early_block)
nir_block *early_block =
state->instr_infos[def->parent_instr->index].early_block;
- /* Some instructions may never be used. We'll just schedule them early and
- * let dead code clean them up.
+ /* Some instructions may never be used. Flag them and the instruction
+ * placement code will get rid of them for us.
*/
if (lca == NULL) {
- def->parent_instr->block = early_block;
+ def->parent_instr->block = NULL;
return true;
}
+ if (def->parent_instr->pass_flags & GCM_INSTR_SCHEDULE_EARLIER_ONLY &&
+ lca != def->parent_instr->block &&
+ nir_block_dominates(def->parent_instr->block, lca)) {
+ lca = def->parent_instr->block;
+ }
+
/* We now have the LCA of all of the uses. If our invariants hold,
* this is dominated by the block that we chose when scheduling early.
* We now walk up the dominance tree and pick the lowest block that is
return false;
}
+static bool
+gcm_replace_def_with_undef(nir_ssa_def *def, void *void_state)
+{
+ struct gcm_state *state = void_state;
+
+ if (list_is_empty(&def->uses) && list_is_empty(&def->if_uses))
+ return true;
+
+ nir_ssa_undef_instr *undef =
+ nir_ssa_undef_instr_create(state->impl->function->shader,
+ def->num_components, def->bit_size);
+ nir_instr_insert(nir_before_cf_list(&state->impl->body), &undef->instr);
+ nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(&undef->def));
+
+ return true;
+}
+
/** Places an instrution back into the program
*
* The earlier passes of GCM simply choose blocks for each instruction and
instr->pass_flags |= GCM_INSTR_PLACED;
+ if (instr->block == NULL) {
+ nir_foreach_ssa_def(instr, gcm_replace_def_with_undef, state);
+ nir_instr_remove(instr);
+ return;
+ }
+
/* Phi nodes are our once source of back-edges. Since right now we are
* only doing scheduling within blocks, we don't need to worry about
* them since they are always at the top. Just skip them completely.