* to keep track of and update phis along the way which gets tricky and
* doesn't add much value over converting to regs.
*
- * The loop may have a continue instruction at the end of the loop which does
+ * The loop may have a jump instruction at the end of the loop which does
* nothing. Once we're out of SSA, we can safely delete it so we don't have
* to deal with it later.
*/
nir_lower_phis_to_regs_block(block_after_loop);
- /* Remove continue if its the last instruction in the loop */
+ /* Remove jump if it's the last instruction in the loop */
nir_instr *last_instr = nir_block_last_instr(nir_loop_last_block(loop));
if (last_instr && last_instr->type == nir_instr_type_jump) {
nir_instr_remove(last_instr);
nir_cf_extract(&loop_body, nir_after_cf_node(&limiting_term->nif->cf_node),
nir_after_block(nir_loop_last_block(loop)));
- struct hash_table *remap_table =
- _mesa_hash_table_create(NULL, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ struct hash_table *remap_table = _mesa_pointer_hash_table_create(NULL);
/* Clone the loop header and insert before the loop */
nir_cf_list_clone_and_reinsert(&lp_header, loop->cf_node.parent,
num_times_to_clone = loop->info->max_trip_count;
}
- struct hash_table *remap_table =
- _mesa_hash_table_create(NULL, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ struct hash_table *remap_table = _mesa_pointer_hash_table_create(NULL);
nir_cf_list lp_body;
nir_cf_node *unroll_loc =
_mesa_hash_table_destroy(remap_table, NULL);
}
+/**
+ * Unroll loops where we only have a single terminator but the exact trip
+ * count is unknown. For example:
+ *
+ * for (int i = 0; i < imin(x, 4); i++)
+ * ...
+ */
+static void
+complex_unroll_single_terminator(nir_loop *loop)
+{
+ assert(list_length(&loop->info->loop_terminator_list) == 1);
+ assert(loop->info->limiting_terminator);
+ assert(nir_is_trivial_loop_if(loop->info->limiting_terminator->nif,
+ loop->info->limiting_terminator->break_block));
+
+ nir_loop_terminator *terminator = loop->info->limiting_terminator;
+
+ loop_prepare_for_unroll(loop);
+
+ /* Pluck out the loop header */
+ nir_cf_list lp_header;
+ nir_cf_extract(&lp_header, nir_before_block(nir_loop_first_block(loop)),
+ nir_before_cf_node(&terminator->nif->cf_node));
+
+ struct hash_table *remap_table =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ /* We need to clone the loop one extra time in order to clone the lcssa
+ * vars for the last iteration (they are inside the following ifs break
+ * branch). We leave other passes to clean up this redundant if.
+ */
+ unsigned num_times_to_clone = loop->info->max_trip_count + 1;
+
+ nir_cf_list lp_body;
+ UNUSED nir_cf_node *unroll_loc =
+ complex_unroll_loop_body(loop, terminator, &lp_header, &lp_body,
+ remap_table, num_times_to_clone);
+
+ /* Delete the original loop header and body */
+ nir_cf_delete(&lp_header);
+ nir_cf_delete(&lp_body);
+
+ /* The original loop has been replaced so remove it. */
+ nir_cf_node_remove(&loop->cf_node);
+
+ _mesa_hash_table_destroy(remap_table, NULL);
+}
+
/* Unrolls the classic wrapper loops e.g
*
* do {
static bool
wrapper_unroll(nir_loop *loop)
{
- if (!list_empty(&loop->info->loop_terminator_list)) {
+ if (!list_is_empty(&loop->info->loop_terminator_list)) {
/* Unrolling a loop with a large number of exits can result in a
* large inrease in register pressure. For now we just skip
nir_after_block(nir_if_last_else_block(terminator->nif));
}
} else {
- nir_block *blk_after_loop =
- nir_cursor_current_block(nir_after_cf_node(&loop->cf_node));
-
- /* There may still be some single src phis following the loop that
- * have not yet been cleaned up by another pass. Tidy those up
- * before unrolling the loop.
- */
- nir_foreach_instr_safe(instr, blk_after_loop) {
- if (instr->type != nir_instr_type_phi)
- break;
-
- nir_phi_instr *phi = nir_instr_as_phi(instr);
- assert(exec_list_length(&phi->srcs) == 1);
-
- nir_phi_src *phi_src =
- exec_node_data(nir_phi_src, exec_list_get_head(&phi->srcs), node);
-
- nir_ssa_def_rewrite_uses(&phi->dest.ssa, phi_src->src);
- nir_instr_remove(instr);
- }
-
- /* Remove break at end of the loop */
- nir_block *last_loop_blk = nir_loop_last_block(loop);
- nir_instr *break_instr = nir_block_last_instr(last_loop_blk);
- nir_instr_remove(break_instr);
+ loop_prepare_for_unroll(loop);
}
/* Pluck out the loop body. */
}
static bool
-is_loop_small_enough_to_unroll(nir_shader *shader, nir_loop_info *li)
+is_access_out_of_bounds(nir_loop_terminator *term, nir_deref_instr *deref,
+ unsigned trip_count)
+{
+ for (nir_deref_instr *d = deref; d; d = nir_deref_instr_parent(d)) {
+ if (d->deref_type != nir_deref_type_array)
+ continue;
+
+ nir_alu_instr *alu = nir_instr_as_alu(term->conditional_instr);
+ nir_src src = term->induction_rhs ? alu->src[1].src : alu->src[0].src;
+ if (!nir_srcs_equal(d->arr.index, src))
+ continue;
+
+ nir_deref_instr *parent = nir_deref_instr_parent(d);
+ assert(glsl_type_is_array(parent->type) ||
+ glsl_type_is_matrix(parent->type));
+
+ /* We have already unrolled the loop and the new one will be imbedded in
+ * the innermost continue branch. So unless the array is greater than
+ * the trip count any iteration over the loop will be an out of bounds
+ * access of the array.
+ */
+ return glsl_get_length(parent->type) <= trip_count;
+ }
+
+ return false;
+}
+
+/* If we know an array access is going to be out of bounds remove or replace
+ * the access with an undef. This can later result in the entire loop being
+ * removed by nir_opt_dead_cf().
+ */
+static void
+remove_out_of_bounds_induction_use(nir_shader *shader, nir_loop *loop,
+ nir_loop_terminator *term,
+ nir_cf_list *lp_header,
+ nir_cf_list *lp_body,
+ unsigned trip_count)
+{
+ if (!loop->info->guessed_trip_count)
+ return;
+
+ /* Temporarily recreate the original loop so we can alter it */
+ nir_cf_reinsert(lp_header, nir_after_block(nir_loop_last_block(loop)));
+ nir_cf_reinsert(lp_body, nir_after_block(nir_loop_last_block(loop)));
+
+ nir_builder b;
+ nir_builder_init(&b, nir_cf_node_get_function(&loop->cf_node));
+
+ nir_foreach_block_in_cf_node(block, &loop->cf_node) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
+
+ /* Check for arrays variably-indexed by a loop induction variable.
+ * If this access is out of bounds remove the instruction or replace
+ * its use with an undefined instruction.
+ * If the loop is no longer useful we leave it for the appropriate
+ * pass to clean it up for us.
+ */
+ if (intrin->intrinsic == nir_intrinsic_load_deref ||
+ intrin->intrinsic == nir_intrinsic_store_deref ||
+ intrin->intrinsic == nir_intrinsic_copy_deref) {
+
+ if (is_access_out_of_bounds(term, nir_src_as_deref(intrin->src[0]),
+ trip_count)) {
+ if (intrin->intrinsic == nir_intrinsic_load_deref) {
+ nir_ssa_def *undef =
+ nir_ssa_undef(&b, intrin->dest.ssa.num_components,
+ intrin->dest.ssa.bit_size);
+ nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
+ nir_src_for_ssa(undef));
+ } else {
+ nir_instr_remove(instr);
+ continue;
+ }
+ }
+
+ if (intrin->intrinsic == nir_intrinsic_copy_deref &&
+ is_access_out_of_bounds(term, nir_src_as_deref(intrin->src[1]),
+ trip_count)) {
+ nir_instr_remove(instr);
+ }
+ }
+ }
+ }
+
+ /* Now that we are done extract the loop header and body again */
+ nir_cf_extract(lp_header, nir_before_block(nir_loop_first_block(loop)),
+ nir_before_cf_node(&term->nif->cf_node));
+ nir_cf_extract(lp_body, nir_before_block(nir_loop_first_block(loop)),
+ nir_after_block(nir_loop_last_block(loop)));
+}
+
+/* Partially unrolls loops that don't have a known trip count.
+ */
+static void
+partial_unroll(nir_shader *shader, nir_loop *loop, unsigned trip_count)
{
+ assert(list_length(&loop->info->loop_terminator_list) == 1);
+
+ nir_loop_terminator *terminator =
+ list_first_entry(&loop->info->loop_terminator_list,
+ nir_loop_terminator, loop_terminator_link);
+
+ assert(nir_is_trivial_loop_if(terminator->nif, terminator->break_block));
+
+ loop_prepare_for_unroll(loop);
+
+ /* Pluck out the loop header */
+ nir_cf_list lp_header;
+ nir_cf_extract(&lp_header, nir_before_block(nir_loop_first_block(loop)),
+ nir_before_cf_node(&terminator->nif->cf_node));
+
+ struct hash_table *remap_table =
+ _mesa_hash_table_create(NULL, _mesa_hash_pointer,
+ _mesa_key_pointer_equal);
+
+ nir_cf_list lp_body;
+ nir_cf_node *unroll_loc =
+ complex_unroll_loop_body(loop, terminator, &lp_header, &lp_body,
+ remap_table, trip_count);
+
+ /* Attempt to remove out of bounds array access */
+ remove_out_of_bounds_induction_use(shader, loop, terminator, &lp_header,
+ &lp_body, trip_count);
+
+ nir_cursor cursor =
+ get_complex_unroll_insert_location(unroll_loc,
+ terminator->continue_from_then);
+
+ /* Reinsert the loop in the innermost nested continue branch of the unrolled
+ * loop.
+ */
+ nir_loop *new_loop = nir_loop_create(shader);
+ nir_cf_node_insert(cursor, &new_loop->cf_node);
+ new_loop->partially_unrolled = true;
+
+ /* Clone loop header and insert into new loop */
+ nir_cf_list_clone_and_reinsert(&lp_header, loop->cf_node.parent,
+ nir_after_cf_list(&new_loop->body),
+ remap_table);
+
+ /* Clone loop body and insert into new loop */
+ nir_cf_list_clone_and_reinsert(&lp_body, loop->cf_node.parent,
+ nir_after_cf_list(&new_loop->body),
+ remap_table);
+
+ /* Insert break back into terminator */
+ nir_jump_instr *brk = nir_jump_instr_create(shader, nir_jump_break);
+ nir_if *nif = nir_block_get_following_if(nir_loop_first_block(new_loop));
+ if (terminator->continue_from_then) {
+ nir_instr_insert_after_block(nir_if_last_else_block(nif), &brk->instr);
+ } else {
+ nir_instr_insert_after_block(nir_if_last_then_block(nif), &brk->instr);
+ }
+
+ /* Delete the original loop header and body */
+ nir_cf_delete(&lp_header);
+ nir_cf_delete(&lp_body);
+
+ /* The original loop has been replaced so remove it. */
+ nir_cf_node_remove(&loop->cf_node);
+
+ _mesa_hash_table_destroy(remap_table, NULL);
+}
+
+/*
+ * Returns true if we should unroll the loop, otherwise false.
+ */
+static bool
+check_unrolling_restrictions(nir_shader *shader, nir_loop *loop)
+{
+ if (loop->control == nir_loop_control_unroll)
+ return true;
+
+ if (loop->control == nir_loop_control_dont_unroll)
+ return false;
+
+ nir_loop_info *li = loop->info;
unsigned max_iter = shader->options->max_unroll_iterations;
+ unsigned trip_count =
+ li->max_trip_count ? li->max_trip_count : li->guessed_trip_count;
- if (li->max_trip_count > max_iter)
+ if (trip_count > max_iter)
return false;
- if (li->force_unroll)
+ if (li->force_unroll && !li->guessed_trip_count)
return true;
bool loop_not_too_large =
- li->num_instructions * li->max_trip_count <= max_iter * LOOP_UNROLL_LIMIT;
+ li->instr_cost * trip_count <= max_iter * LOOP_UNROLL_LIMIT;
return loop_not_too_large;
}
static bool
-process_loops(nir_shader *sh, nir_cf_node *cf_node, bool *has_nested_loop_out)
+process_loops(nir_shader *sh, nir_cf_node *cf_node, bool *has_nested_loop_out,
+ bool *unrolled_this_block);
+
+static bool
+process_loops_in_block(nir_shader *sh, struct exec_list *block,
+ bool *has_nested_loop_out)
+{
+ /* We try to unroll as many loops in one pass as possible.
+ * E.g. we can safely unroll both loops in this block:
+ *
+ * if (...) {
+ * loop {...}
+ * }
+ *
+ * if (...) {
+ * loop {...}
+ * }
+ *
+ * Unrolling one loop doesn't affect the other one.
+ *
+ * On the other hand for block with:
+ *
+ * loop {...}
+ * ...
+ * loop {...}
+ *
+ * It is unsafe to unroll both loops in one pass without taking
+ * complicating precautions, since the structure of the block would
+ * change after unrolling the first loop. So in such a case we leave
+ * the second loop for the next iteration of unrolling to handle.
+ */
+
+ bool progress = false;
+ bool unrolled_this_block = false;
+
+ foreach_list_typed(nir_cf_node, nested_node, node, block) {
+ if (process_loops(sh, nested_node,
+ has_nested_loop_out, &unrolled_this_block)) {
+ progress = true;
+
+ /* If current node is unrolled we could not safely continue
+ * our iteration since we don't know the next node
+ * and it's hard to guarantee that we won't end up unrolling
+ * inner loop of the currently unrolled one, if such exists.
+ */
+ if (unrolled_this_block) {
+ break;
+ }
+ }
+ }
+
+ return progress;
+}
+
+static bool
+process_loops(nir_shader *sh, nir_cf_node *cf_node, bool *has_nested_loop_out,
+ bool *unrolled_this_block)
{
bool progress = false;
bool has_nested_loop = false;
return progress;
case nir_cf_node_if: {
nir_if *if_stmt = nir_cf_node_as_if(cf_node);
- foreach_list_typed_safe(nir_cf_node, nested_node, node, &if_stmt->then_list)
- progress |= process_loops(sh, nested_node, has_nested_loop_out);
- foreach_list_typed_safe(nir_cf_node, nested_node, node, &if_stmt->else_list)
- progress |= process_loops(sh, nested_node, has_nested_loop_out);
+ progress |= process_loops_in_block(sh, &if_stmt->then_list,
+ has_nested_loop_out);
+ progress |= process_loops_in_block(sh, &if_stmt->else_list,
+ has_nested_loop_out);
return progress;
}
case nir_cf_node_loop: {
loop = nir_cf_node_as_loop(cf_node);
- foreach_list_typed_safe(nir_cf_node, nested_node, node, &loop->body)
- progress |= process_loops(sh, nested_node, &has_nested_loop);
+ progress |= process_loops_in_block(sh, &loop->body, &has_nested_loop);
break;
}
unreachable("unknown cf node type");
}
+ const bool unrolled_child_block = progress;
+
/* Don't attempt to unroll a second inner loop in this pass, wait until the
* next pass as we have altered the cf.
*/
- if (!progress) {
+ if (!progress && loop->control != nir_loop_control_dont_unroll) {
/* Check for the classic
*
!loop->info->complex_loop) {
nir_block *last_loop_blk = nir_loop_last_block(loop);
- if (!nir_block_ends_in_break(last_loop_blk))
+ if (nir_block_ends_in_break(last_loop_blk)) {
+ progress = wrapper_unroll(loop);
goto exit;
+ }
- progress = wrapper_unroll(loop);
-
- goto exit;
+ /* If we were able to guess the loop iteration based on array access
+ * then do a partial unroll.
+ */
+ unsigned num_lt = list_length(&loop->info->loop_terminator_list);
+ if (!has_nested_loop && num_lt == 1 && !loop->partially_unrolled &&
+ loop->info->guessed_trip_count &&
+ check_unrolling_restrictions(sh, loop)) {
+ partial_unroll(sh, loop, loop->info->guessed_trip_count);
+ progress = true;
+ }
}
- if (has_nested_loop || loop->info->limiting_terminator == NULL)
+ if (has_nested_loop || !loop->info->limiting_terminator)
goto exit;
- if (!is_loop_small_enough_to_unroll(sh, loop->info))
+ if (!check_unrolling_restrictions(sh, loop))
goto exit;
if (loop->info->exact_trip_count_known) {
} else {
/* Attempt to unroll loops with two terminators. */
unsigned num_lt = list_length(&loop->info->loop_terminator_list);
- if (num_lt == 2) {
+ if (num_lt == 2 &&
+ !loop->info->limiting_terminator->exact_trip_count_unknown) {
bool limiting_term_second = true;
nir_loop_terminator *terminator =
list_first_entry(&loop->info->loop_terminator_list,
}
progress = true;
}
+
+ if (num_lt == 1) {
+ assert(loop->info->limiting_terminator->exact_trip_count_unknown);
+ complex_unroll_single_terminator(loop);
+ progress = true;
+ }
}
}
exit:
*has_nested_loop_out = true;
+ if (progress && !unrolled_child_block)
+ *unrolled_this_block = true;
+
return progress;
}
nir_metadata_require(impl, nir_metadata_loop_analysis, indirect_mask);
nir_metadata_require(impl, nir_metadata_block_index);
- foreach_list_typed_safe(nir_cf_node, node, node, &impl->body) {
- bool has_nested_loop = false;
- progress |= process_loops(impl->function->shader, node,
- &has_nested_loop);
- }
+ bool has_nested_loop = false;
+ progress |= process_loops_in_block(impl->function->shader, &impl->body,
+ &has_nested_loop);
- if (progress)
+ if (progress) {
+ nir_metadata_preserve(impl, nir_metadata_none);
nir_lower_regs_to_ssa_impl(impl);
+ } else {
+ nir_metadata_preserve(impl, nir_metadata_all);
+ }
return progress;
}
projects / mesa.git / blobdiff