The following removes quadraticness when visiting each predecessor
of a large CFG merge with dominated_by_p_w_unex.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
PR tree-optimization/119389
* tree-ssa-sccvn.cc (dominated_by_p_w_unex): Limit the number
of predecessors of a CFG merge we try to skip.
---
gcc/tree-ssa-sccvn.cc | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/gcc/tree-ssa-sccvn.cc b/gcc/tree-ssa-sccvn.cc
index 40c38fa020a..481ab8b243d 100644
--- a/gcc/tree-ssa-sccvn.cc
+++ b/gcc/tree-ssa-sccvn.cc
@@ -5172,7 +5172,11 @@ dominated_by_p_w_unex (basic_block bb1, basic_block bb2,
bool allow_back)
/* Iterate to the single successor of bb2 with only a single executable
incoming edge. */
else if (EDGE_COUNT (bb2->succs) == 1
- && EDGE_COUNT (single_succ (bb2)->preds) > 1)
+ && EDGE_COUNT (single_succ (bb2)->preds) > 1
+ /* Limit the number of edges we check, we should bring in
+ context from the iteration and compute the single
+ executable incoming edge when visiting a block. */
+ && EDGE_COUNT (single_succ (bb2)->preds) < 8)
{
edge prede = NULL;
FOR_EACH_EDGE (e, ei, single_succ (bb2)->preds)
--
2.43.0
