https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101673
Bug ID: 101673
Summary: shorter unprofitable jump thread path inhibits
threading of larger path
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: aldyh at gcc dot gnu.org
Target Milestone: ---
This is tree-ssa/ranger-threader-5.c which is now failing after the rewrite of
the backwards threader with a ranger based threader (commit
2e96b5f14e4025691b57d2301d71aa6092ed44bc).
I am not sure this is worth fixing, but am adding a PR so we don't loose track
of the regression.
Here are my notes from the test:
/* This tests that we can thread BB4->BB999 coming in through the
following path:
latch many insns
| |
V V
6 -> 7 -> 3 -> 4 -> 999
The ranger based threader cannot thread this because BB4 has too
many instructions so it gives up looking back. However, if we were
able to looking further, we would notice that a profitable path
passing through the loop latch (BB7) exists.
That is, 3->4->N in isolation is not profitable, but 6->7->3->4->N is.
It is not clear whether handling this case in the backwards
threader is profitable, as it would increase the search space
considerably. The test is being added to note a regression from
the old backward threader code.
This test has been distilled from libphobos/src/std/net/isemail.d.
The ranger threader stops at the 3->4 subpath with: "did not thread
around loop and would copy too many statements". */