summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorTimothy Arceri <tarceri@itsqueeze.com>2018-11-20 13:45:58 +1100
committerTimothy Arceri <tarceri@itsqueeze.com>2019-03-12 00:52:30 +0000
commit68ce0ec22244726c1b61d91936a6b79ac20ab77a (patch)
tree4916d51a4433d6743c5ddeb5931fcd1a1bc1d214
parente8a8937a04fd9531069616c63c099ba276296112 (diff)
nir: calculate trip count for more loops
This adds support to loop analysis for loops where the induction variable is compared to the result of min(variable, constant). For example: for (int i = 0; i < imin(x, 4); i++) ... We add a new bool to the loop terminator struct in order to differentiate terminators with this exit condition. Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
-rw-r--r--src/compiler/nir/nir.h11
-rw-r--r--src/compiler/nir/nir_loop_analyze.c41
-rw-r--r--src/compiler/nir/nir_opt_loop_unroll.c3
3 files changed, 49 insertions, 6 deletions
diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index aac9ac448b8..19e69c50f8f 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -1912,6 +1912,17 @@ typedef struct {
bool continue_from_then;
bool induction_rhs;
+ /* This is true if the terminators exact trip count is unknown. For
+ * example:
+ *
+ * for (int i = 0; i < imin(x, 4); i++)
+ * ...
+ *
+ * Here loop analysis would have set a max_trip_count of 4 however we dont
+ * know for sure that this is the exact trip count.
+ */
+ bool exact_trip_count_unknown;
+
struct list_head loop_terminator_link;
} nir_loop_terminator;
diff --git a/src/compiler/nir/nir_loop_analyze.c b/src/compiler/nir/nir_loop_analyze.c
index 9cff670051e..4c8028a666f 100644
--- a/src/compiler/nir/nir_loop_analyze.c
+++ b/src/compiler/nir/nir_loop_analyze.c
@@ -539,6 +539,35 @@ guess_loop_limit(loop_info_state *state, nir_const_value *limit_val,
return false;
}
+static bool
+try_find_limit_of_alu(nir_loop_variable *limit, nir_const_value *limit_val,
+ nir_loop_terminator *terminator, loop_info_state *state)
+{
+ if(!is_var_alu(limit))
+ return false;
+
+ nir_alu_instr *limit_alu = nir_instr_as_alu(limit->def->parent_instr);
+
+ if (limit_alu->op == nir_op_imin ||
+ limit_alu->op == nir_op_fmin) {
+ limit = get_loop_var(limit_alu->src[0].src.ssa, state);
+
+ if (!is_var_constant(limit))
+ limit = get_loop_var(limit_alu->src[1].src.ssa, state);
+
+ if (!is_var_constant(limit))
+ return false;
+
+ *limit_val = nir_instr_as_load_const(limit->def->parent_instr)->value;
+
+ terminator->exact_trip_count_unknown = true;
+
+ return true;
+ }
+
+ return false;
+}
+
static int32_t
get_iteration(nir_op cond_op, nir_const_value *initial, nir_const_value *step,
nir_const_value *limit)
@@ -770,12 +799,14 @@ find_trip_count(loop_info_state *state)
} else {
trip_count_known = false;
- /* Guess loop limit based on array access */
- if (!guess_loop_limit(state, &limit_val, basic_ind)) {
- continue;
- }
+ if (!try_find_limit_of_alu(limit, &limit_val, terminator, state)) {
+ /* Guess loop limit based on array access */
+ if (!guess_loop_limit(state, &limit_val, basic_ind)) {
+ continue;
+ }
- guessed_trip_count = true;
+ guessed_trip_count = true;
+ }
}
/* We have determined that we have the following constants:
diff --git a/src/compiler/nir/nir_opt_loop_unroll.c b/src/compiler/nir/nir_opt_loop_unroll.c
index c700e6704da..a5a5ca3deee 100644
--- a/src/compiler/nir/nir_opt_loop_unroll.c
+++ b/src/compiler/nir/nir_opt_loop_unroll.c
@@ -827,7 +827,8 @@ process_loops(nir_shader *sh, nir_cf_node *cf_node, bool *has_nested_loop_out)
} 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,