From: Kyrylo Tkachov <[email protected]>
nodes_overlap_p returns true for any two DT_OPERAND nodes whose
operand cases overlap. decision_tree::find_node uses this to set
overlap_node, which prevents a later node from merging into an earlier
equal sibling across the blocking node.
Of the 163 barrier refusals on trunk, 135 (83%) are between the GIMPLE
form and the GENERIC form that lower_cond gives every COND_EXPR
condition. gen_kids_1 dispatches the GIMPLE form under the SSA_NAME
case and the GENERIC form by the operand's own TREE_CODE. An input
cannot satisfy both dispatch conditions simultaneously, so these 135
refusals protect no ordering-sensitive case.
Add a check in nodes_overlap_p: return false when the two expressions
differ in is_generic. This eliminates the spurious GIMPLE/GENERIC
overlap refusals while keeping every DT_MATCH and overlapping
operator-set refusal. The 7 remaining order-sensitive pattern pairs
all require degenerate inputs that other patterns fold first, and both
of the patterns in each pair produce the same result value.
On aarch64 at ten GIMPLE parts, GIMPLE lines fall by about 0.6%.
GENERIC output is unchanged.
Bootstrapped and tested on aarch64-none-linux-gnu.
Ok for trunk?
Thanks,
Kyrill
gcc/ChangeLog:
* genmatch.cc (nodes_overlap_p): Return false when the two
expressions have different is_generic, as they cannot match the
same input.
Signed-off-by: Kyrylo Tkachov <[email protected]>
---
gcc/genmatch.cc | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/gcc/genmatch.cc b/gcc/genmatch.cc
index e6065a039ef..905ff420d6a 100644
--- a/gcc/genmatch.cc
+++ b/gcc/genmatch.cc
@@ -2847,8 +2847,16 @@ nodes_overlap_p (dt_node *n1, dt_node *n2)
if (n1->type == dt_node::DT_MATCH || n2->type == dt_node::DT_MATCH)
return true;
- return operand_cases_overlap_p (as_a <dt_operand *> (n1)->op,
- as_a <dt_operand *> (n2)->op);
+ operand *op1 = as_a <dt_operand *> (n1)->op;
+ operand *op2 = as_a <dt_operand *> (n2)->op;
+ /* A GIMPLE expression and a GENERIC expression are dispatched through
+ separate paths (SSA_NAME case vs. operand TREE_CODE) and cannot
+ match the same input. */
+ expr *e1 = op1 ? dyn_cast <expr *> (op1) : NULL;
+ expr *e2 = op2 ? dyn_cast <expr *> (op2) : NULL;
+ if (e1 && e2 && e1->is_generic != e2->is_generic)
+ return false;
+ return operand_cases_overlap_p (op1, op2);
}
/* Search OPS for a decision tree node like P and return it if found. */
--
2.50.1 (Apple Git-155)
