https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121711
Andrew Macleod <amacleod at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amacleod at redhat dot com
--- Comment #1 from Andrew Macleod <amacleod at redhat dot com> ---
t_5 = b_3(D) - a_4(D);
if (t_5 == 1)
goto <bb 3>; [INV]
else
goto <bb 5>; [INV]
<bb 3> :
foo (1);
_1 = a_4(D) + 1;
if (_1 != b_3(D))
goto <bb 4>; [INV]
else
goto <bb 5>; [INV]
This sort of algebraic reasoning is beyond what we can do with just ranges and
relations. We can determine basic relations, ie that in BB3 we could
determine that b_3 > a_4, but not that b_3 is a_4 + 1
What exactly do we have for expression adjustments and comparisons in fold or
match and compare, or elsewhere?
it could be possible to create a new kind of relation, maybe VREL_EXPR, that
could be generated by operator_minus (and other operations) when the LHS is a
singleton. On edge 2->3 is could create a relation VREL_EXPR between b_3 and
a_4 with an expressions something like b_3 = a_4 + LHS , which would come
out as b_3 - a_4 + 1
The dependency info is available to determine that _1 uses a_4, so we can
figure out to look for that relation when we evaluate _1 != b_3 . we'd then
need to be able to "fold" that expression b_3s relation ompared to _1's
expression . ie a_4 + 1 != a_4 + 1
There may be alternative ways to use the info if it was created, but As things
currently stand, VRP cannot determine this algebraic simplification.
