[Bug tree-optimization/106138] Inefficient code generation: logical AND of disjoint booleans from equal and bitwise AND not optimized to constant false

[amacleod at redhat dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106138

--- Comment #4 from Andrew Macleod <amacleod at redhat dot com> ---
(In reply to Richard Biener from comment #1)
> So we're seeing
> 
>   b1_8 = x_7(D) == 4;
>   # RANGE [0, 3] NONZERO 3
>   _3 = x_7(D) & 3;
>   b2_9 = _3 != 0;
>   _5 = b1_8 & b2_9;
> 
> and fail to optimize that.  It somewhat looks like "relations" (on x_7), but
> I'm not sure if suitable to handle for ranger.
> 
> Adding something in match.pd might be possible - the key is of course that
> both b1 and b2 are derived from the same value.  I think we already have
> a pattern looking at not intersecting nonzero bits.

Its not suitable yet anyway. theres an order of indirection in that any
relation is based on whether b1_8 and b2_9 are both 1 or both 0...  If there
was  a branch we start to pick up information on the edges...  but without an
edge using _5, we don't yet try to paw back the same way.

perhaps we could optimize &/&& and |/|| to check if one of the operands doesnt
matter..

ie, if b1_8 is 1, then check if b2_9 is always 1 with the range of x_7([4,4]), 
likewise if b1_8 is 0, then b2_9 doesnt matter.  so we can simplify this to _5
= b1_8

it seems like something we could do with simplify_using_ranges quite easily by
utilizing gori perhaps?  easy enough to pick up a range for b2_9 based on the
range of x_7 from b1_8 == [1,1]