https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120236
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |ASSIGNED
Ever confirmed|0 |1
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
OK, so the issue is then LIM hoisting
<bb 2> [local count: 118111600]:
# PT = null
_2 = c;
which becomes available as CSE target in PRE and we then eventually elide
the
c.2_6 = c;
*c.2_6 = 1;
sequence because a store to 'null' is unnecessary.
That said, -fno-move-loop-stores makes it easier to look at what LIM does.
<bb 2> [local count: 118111600]:
+ # PT = null
+ _2 = c;
+ b.0_4 = b;
goto <bb 5>; [100.00%]
<bb 3> [local count: 955630224]:
*d_12(D) = &c;
- # PT = null
- _2 = c;
_3 = *_2;
a = _3;
- b.0_4 = b;
if (b.0_4 != 0)
goto <bb 4>; [0.00%]
else
The issue then still is that points-to analysis computed
*d_13(D) = &c;
# PT = null
_1 = *d_13(D);
# PT = null
_2 = *_1;
_3 = *_2;
a = _3;
b.0_4 = b;
and DOM (aka CSE) made that
*d_12(D) = &c;
# PT = null
_1 = &c;
# PT = null
_2 = c;
_3 = *_2;
a = _3;
b.0_4 = b;
so basically the points-to info is UB manifested but with CSE the code
no longer has UB (but the points-to info is still there) and we hoist
the code, making the UB target to CSE.
Without DOM there's the dependence of the *d_13(D) load on the previous
store, so nothing is hoisted.
I don't see how to "fix" this. A hack would be to try fixing up things
at the point we CSE the PT = null pointer with the later pointer with
sensible points-to info. The most optimistic handling would intersect
both sets, but the latter is
# PT = nonlocal unit-escaped null { D.2832 } (nonlocal, escaped)
c.2_6 = c;
so we'd end up with 'null' this way.
We could start cleaning points-to info whenever we clear flow-sensitive
info. For the reason that, as seen in this case, we inject flow-sensitive
info into the otherwise flow-insensitive solution by handling SSA default
defs specially:
/* For undefined SSA names return nothing. */
else if (!ssa_defined_default_def_p (t))
{
cexpr.var = nothing_id;
cexpr.type = SCALAR;
cexpr.offset = 0;
results->safe_push (cexpr);
return;
short of using pt_anything here and polluting all PHI induced default defs
this way there's no solution.
