https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126904
Bug ID: 126904
Summary: backward threader: no path-sensitive load CSE (value
numbering?)
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: aldyh at gcc dot gnu.org
CC: amacleod at redhat dot com, law at gcc dot gnu.org,
rguenth at gcc dot gnu.org
Blocks: 126010
Target Milestone: ---
This is one case where FRE/PRE cannot currently rescue the backward threader
from what is currently being done by DOM’s avail_exprs.
[Full disclosure, as mentioned earlier I used Claude Fable to reduce the
missing optimizations I’m auditing for the backwards threader. Is it useful to
stick an Assisted-by tag here in the PR? I’ve reduced and verified myself, but
the paragraph analyzing the cause is NOT mine. For the record, I’m not using
AI to do a report-and-run here… I’m more than happy to do some of the leg work,
answer questions, and reproduce, but I know this is a sensitive subject.]
/* { dg-do compile } */
/* { dg-options "-O2 --param=dom-jump-threading=0
-fdump-tree-threadfull1-details" } */
/* The two conditionals test two different SSA names that are loads
of the same location *p. On the path where the first is true
there is no intervening store, so the second load equals the first
and the second conditional is false. DOM's avail-exprs sees this
and threads the path; the path ranger has no memory value
numbering, sees the second load as VARYING, and rejects. FRE/PRE
cannot pre-empt the problem: foo () on the other arm clobbers *p
(no full redundancy), and the possibly non-returning spin ()
before the may-trap load keeps PRE from making it anticipable. */
void foo (void);
static void __attribute__((noinline))
spin (int n)
{
while (n)
;
}
void
f (int *p, int k)
{
if (*p == k)
;
else
foo ();
spin (k);
if (*p != k)
foo ();
}
/* { dg-final { scan-tree-dump-times "Registering jump thread" 1 "threadfull1"
} } */
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126010
[Bug 126010] [meta-bug] Remove DOM
