DFS-based backedge algorithm can be slow for pathological cases

cc @sklam @quasiben @philippjfr @AjayThorve

See https://github.com/holoviz/datashader/issues/956

No Numba-only reproducer as yet, but the DFS-based algorithm increases the runtime of the reproducer in the Datashader issue from 10 seconds to 34 minutes. Going back to the old algorithm with:

diff --git a/numba/core/controlflow.py b/numba/core/controlflow.py
index f68572031..f37e9ae3d 100644
--- a/numba/core/controlflow.py
+++ b/numba/core/controlflow.py
@@ -572,40 +572,14 @@ class CFGraph(object):
         Find back edges.  An edge (src, dest) is a back edge if and
         only if *dest* dominates *src*.
         """
-
-        # Uses a simple DFS to find back-edges.
-        # The new algorithm is faster than the the previous dominator based
-        # algorithm.
         back_edges = set()
-        # stack: keeps track of the traversal path
-        stack = []
-        # succs_state: keep track of unvisited successors of a node
-        succs_state = {}
-        entry_point = self.entry_point()
-
-        def push_state(node):
-            stack.append(node)
-            succs_state[node] = [dest for dest in self._succs[node]]
-
-        push_state(entry_point)

-        while stack:
-            tos = stack[-1]
-            tos_succs = succs_state[tos]
-            # Are there successors not checked?
-            if tos_succs:
-                # Check the next successor
-                cur_node = tos_succs.pop()
-                # Is it in our traversal path?
-                if cur_node in stack:
-                    # Yes, it's a backedge
-                    back_edges.add((tos, cur_node))
-                else:
-                    # Push
-                    push_state(cur_node)
-            else:
-                # Checked all successors. Pop
-                stack.pop()
+        for src, succs in self._succs.items():
+            back = self._doms[src] & succs
+            # In CPython bytecode, at most one back edge can flow from a
+            # given block.
+            assert len(back) <= 1
+            back_edges.update((src, dest) for dest in back)

         return back_edges

Brings the runtime back down again:

$ time python repro.py
x_range: (1285891200000000000, 1291891140000000000) y_range: (-104.8094774985483, 100.09816883332905)

real    0m9.732s
user    0m9.601s
sys     0m0.113s

[gmarkall]

Seems that #6187 could be a potential fix, but we could still do with a Numba-only reproducer.

[jakirkham]

Thanks all! 😄