https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87104
--- Comment #9 from pipcet at gmail dot com ---
(In reply to Richard Biener from comment #6)
> So on GIMPLE the following are not canonicalized:
>
> <bb 2> [local count: 1073741825]:
> _1 = i_4(D) & 7;
> _8 = (int) i_4(D);
> if (_1 == 6)
> goto <bb 3>; [20.97%]
> else
> goto <bb 4>; [79.03%]
>
> vs.
>
> <bb 2> [local count: 1073741825]:
> _1 = i_5(D) + 18446744073709551610;
> _2 = _1 & 7;
> _9 = (int) i_5(D);
> if (_2 == 0)
> goto <bb 3>; [34.00%]
> else
> goto <bb 4>; [66.00%]
>
> where I'd call the former better. Thus for some unknown constraint
> on @1, @2 and @3
>
> (simplify
> (eq (convert? (bit_and (plus @0 INTEGER_CST@3) @2)) @1)
> (eq (convert (bit_and @0 @2)) { ... }))
I suggest that the constraint be that @2 is of the form
000...000111...111000...000 and @3 is of the form
???...??????...???000...000 and @1 is zero.
So your plan is to canonicalize to (X & MASK) == VALUE first, then do something
target-dependent to emit (X - VALUE) & MASK == 0 instead? How would the target
realize that? I tried adding a peephole2 rule but that's apparently too late
and doesn't match the insn sequences in Paul's test.
