https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81897
--- Comment #8 from Jeffrey A. Law <law at redhat dot com> ---
This really looks like something tree-ssa-uninit.c ought to be handling too.
[ Just to be clear we should fix both tree-ssa-uninit.c the cfgcleanup. ]
We have a PHI with a default definition on the RHS:
<bb 4> [local count: 182536114]:
# dst_32 = PHI <dst_7(D)(18), dst_9(20)>
__warned = 1;
if (fibmatch_6 != 0)
goto <bb 13>; [0.00%]
else
goto <bb 14>; [100.00%]
So if you dump the CFG and analyze the path you'll find that the only way to
traverse the edge 18->4 has a series of predicates, *all* of which must be true
to ever reach 18->4. Yet the predicate analysis bits only analyze the last
predicate in the path:
[WORKLIST]: add to initial list: dst_32 = PHI <dst_7(D)(18), dst_9(20)>
[CHECK]: examining phi: dst_32 = PHI <dst_7(D)(18), dst_9(20)>
[BEFORE SIMPLICATION -- [DEF]:
is guarded by :
_19 == 0
[BEFORE NORMALIZATION --[DEF]:
is guarded by :
_19 == 0
[AFTER NORMALIZATION -- [DEF]:
is guarded by :
_19 == 0
I think that's because our choice for where to start the predicate analysis for
the def of a PHI arg is terrible. We look at the domiantor of e->src:
/* Build the control dependency chain for uninit operand `i'... */
uninit_preds = vNULL;
if (!compute_control_dep_chain (find_dom (e->src),
e->src, dep_chains, &num_chains,
&cur_chain, &num_calls))
I'm not sure why it's done that way. THere's no reason we can't go all the way
back to the entry block here (note there's some lameness in
compute_control_dep_chain that needs to be worked around -- you can't just slam
ENTRY_BLOCK_PTR in there).
But fixing that still doesn't resolve the problem, but gives us a shot at
actually seeing a full predicate chain -- which would include the predicate
from bb2:
<bb 2> [local count: 1073741825]:
fibmatch_6 = f ();
if (fibmatch_6 == 0)
goto <bb 3>; [33.00%]
else
goto <bb 7>; [67.00%]
We care about the 2-->7 edge here. ie, the *only* way to traverse the 18->4
edge is to first have traversed the 2->7 edge which occurs when fibmatch_6 !=
0.
That's the def side. Now looking at the use side.
<bb 4> [local count: 182536114]:
# dst_32 = PHI <dst_7(D)(18), dst_9(20)>
__warned = 1;
if (fibmatch_6 != 0)
goto <bb 13>; [0.00%]
else
goto <bb 14>; [100.00%]
<bb 14> [local count: 182536114]:
goto <bb 6>; [100.00%]
<bb 13> [local count: 0]:
<bb 5> [local count: 536870912]:
<bb 6> [local count: 1073741825]:
# dst_2 = PHI <dst_9(16), 0(5), dst_9(12), dst_32(14), dst_9(21)>
return dst_2;
So given that the only way we ever traverse 18->4 is when fibmatch_6 != 0 which
implies that we always traverse 4->13->5->6 and on that path we never use
dst_32.