https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89463
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
<bb 6> :
# i_5 = PHI <_11(5), 0(8)>
# DEBUG i => i_5
if (i_5 != 6)
goto <bb 5>; [INV]
else
goto <bb 7>; [100.00%]
<bb 5> :
# DEBUG BEGIN_STMT
# DEBUG i => i_5 + 1
So it goes like so the following. First we remove the iterator update in
BB5 (unrelated). Then we remove the control stmt, removing the edge 6 -> 5.
This leaves us with
# i_5 = PHI <0(8)>
# DEBUG i => i_5
then we remove the PHI node which inserts the bogus debug stmt.
Stmt and control-flow removal in DCE is somewhat a twisted maze but the fix
has to be there, avoiding the above situation.
For simple IVs we could compute the final value and insert a debug stmt
after the loop. But for IVs we cannot analyze a similar situation will
arise and thus we have to somehow compute a # DEBUG i => NULL stmt from
the above.
But that means we have to queue edges we want to remove (plus eventually
blocks we can avoid looking at for stmt removal) and perform CFG pruning
only afterwards (like leave it to CFG cleanup by changing conditionals
to if (true/false)). Doing that (just for conditions) gives us the
expected
@@ -24,22 +52,10 @@
if (a.1_3 <= 9)
goto <bb 3>; [INV]
else
- goto <bb 6>; [INV]
-
- <bb 5> :
- # DEBUG BEGIN_STMT
- i_11 = i_5 + 1;
- # DEBUG i => i_11
-
- <bb 6> :
- # i_5 = PHI <i_11(5), 0(4)>
- # DEBUG i => i_5
- if (i_5 != 6)
goto <bb 5>; [INV]
- else
- goto <bb 7>; [INV]
- <bb 7> :
+ <bb 5> :
+ # DEBUG i => NULL
# DEBUG BEGIN_STMT
optimize_me_not ();
return 0;
but that's not because of CFG cleanup but because we do not remove the
edge so at PHI removal time we still have both values.
A non-loopy testcase may also go wrong I guess but it's usually optimized
earlier and not by DCE.
I have a patch.
