hoist_adjacent_loads moves the two loads from under the condition of
the diamond into its head, but keeps the range info computed for them
under the condition. Clear it.
A follow-up patch to range_from_readonly_load takes the range of a
load from a constant aggregate from its index:
struct S { int a, b; };
static const struct S sa[4] = { { 1, 2 }, { 3, 4 }, { 5, 6 }, { 7, 8 } };
if (i < 2)
r = sa[i].a; /* [1, 3] here, [1, 7] before the if */
else
r = sa[i].b;
if (sa[i].a > 3) /* folds to false */
...
Nothing on trunk gives a load a range that depends on where it sits, so
the wrong code cannot be exercised on its own. The testcase comes with
the follow-on patch that provides this functionality, which generates
wrong code without this patch: chicken and egg :).
Tested on ppc64le Linux.
OK?
gcc/ChangeLog:
* tree-ssa-phiopt.cc (hoist_adjacent_loads): Clear the range
info of the hoisted loads.
---
gcc/tree-ssa-phiopt.cc | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/gcc/tree-ssa-phiopt.cc b/gcc/tree-ssa-phiopt.cc
index 6e17d21b58c..7203f4d29c8 100644
--- a/gcc/tree-ssa-phiopt.cc
+++ b/gcc/tree-ssa-phiopt.cc
@@ -4569,6 +4569,10 @@ hoist_adjacent_loads (basic_block bb0, basic_block bb1,
gsi_move_to_bb_end (&gsi2, bb0);
gsi2 = gsi_for_stmt (def2);
gsi_move_to_bb_end (&gsi2, bb0);
+ /* Clear range info from the defs we've moved from under the
+ condition. */
+ reset_flow_sensitive_info (gimple_assign_lhs (def1));
+ reset_flow_sensitive_info (gimple_assign_lhs (def2));
statistics_counter_event (cfun, "hoisted loads", 1);
if (dump_file && (dump_flags & TDF_DETAILS))
--
2.53.0
