https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86707
Andrew Macleod <amacleod at redhat dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amacleod at redhat dot com
--- Comment #2 from Andrew Macleod <amacleod at redhat dot com> ---
This should get taken care of when ranger tracks must-be-one bits.
_1 = x_7(D) & 1;
if (_1 == 0)
goto <bb 3>; [INV]
else
goto <bb 4>; [INV]
_1 : unsigned int [0, 1]
2->3 (T) _1 : unsigned int [0, 0]
2->3 (T) x_7(D) : unsigned int [0, 4294967294]
2->4 (F) _1 : unsigned int [1, 1]
2->4 (F) x_7(D) : unsigned int [1, +INF]
<..>
=========== BB 4 ============
x_7(D) unsigned int [1, +INF]
<bb 4> :
_2 = x_7(D) & 3;
if (_2 == 0)
goto <bb 5>; [INV]
Ideally, taking edge 2->4 will instead produce something like:
x_7(D) unsigned int [1, +INF], bits set 0x0001
This would come from operator_bitwise_and::op1_range() with the LHS being
subsituted back from the edge:
[1,1] = x_7 & 1 in order for that to be true, the mask will have to be
must-be-one
and then the calculation of
_2 = x_7(D) & 3; would produce a non-zero range when operator_bitwise_and sees
the must-be-one bit, and propagates it since that bit is in the mask. so we'd
get something like:
_2 = [1, 3] bits set 0x0001
which would then allow _2 == 0 to be folded as always false.
Planned for next release.
PR 83073 also touched on this, but not in as much detail nor with as good an
exmaple.