https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65851
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so it goes like
Visiting PHI node: f_12 = PHI <0B(2), f_8(5)>
Argument #0 (2 -> 6 executable)
0B Value: CONSTANT 0B
Argument #1 (5 -> 6 not executable)
PHI node value: CONSTANT 0B
...
Visiting statement:
_15 = (long unsigned int) _14;
which is likely CONSTANT
Lattice value changed to CONSTANT 0x0 (0x0fffffffffffffff0). Adding SSA edges
to worklist.
Visiting statement:
_3 = f_12 + _15;
which is likely CONSTANT
Applying pattern match.pd:54, gimple-match.c:9714
Applying pattern match.pd:54, generic-match.c:4594
Lattice value changed to CONSTANT 0x0 (0x0fffffffffffffff0). Adding SSA edges
to worklist.
(should have simplified to _15 ...? Ah, no - to (ptrtype)_15!)
...
Visiting PHI node: f_12 = PHI <0B(2), f_8(5)>
Argument #0 (2 -> 6 executable)
0B Value: CONSTANT 0B
Argument #1 (5 -> 6 executable)
f_8 Value: VARYING
PHI node value: VARYING
...
Visiting statement:
_15 = (long unsigned int) _14;
which is likely CONSTANT
Lattice value changed to VARYING. Adding SSA edges to worklist.
Now we can look at both defs when combining stmts during simplification:
Visiting statement:
_3 = f_12 + _15;
which is likely CONSTANT
Applying pattern match.pd:413, gimple-match.c:9792
Match-and-simplified f_12 + _15 to &e
ICE.
e.4_11 = (long int) &e;
f.5_13 = (long int) f_12;
_14 = e.4_11 - f.5_13;
_15 = (long unsigned int) _14;
_3 = f_12 + _15;
and yes - that combines to &e. The interesting thing here is that our
safety net for tuning down combining stmts is too conservative (again),
because we can very well look at stmt definitions as long as we don't
valueize anything for stmts we want to re-visit again. That's of course
not so trivial to implement ... (ideally we'd have applied the simplification
earlier - this is after all "unfolded" IL)
