tree-optimization: Improve spaceship_replacement [PR94589]

tree-optimization: Improve spaceship_replacement [PR94589]

On Wed, May 19, 2021 at 01:30:31PM -0400, Jason Merrill via Gcc-patches wrote:
> Here, when genericizing lexicographical_compare_three_way, we haven't yet
> walked the operands, so (a == a) still sees ADDR_EXPR <a>, but this is after
> we've changed the type of a to REFERENCE_TYPE.  When we try to fold (a == a)
> by constexpr evaluation, the constexpr code doesn't understand trying to
> take the address of a reference, and we end up crashing.
>
> Fixed by avoiding constexpr evaluation in genericize_spaceship, by using
> fold_build2 instead of build_new_op on scalar operands.  Class operands
> should have been expanded during parsing.

Unfortunately this slightly changed the IL and spaceship_replacement no
longer pattern matches it.

Here are 3 improvements that make it match:

1) as mentioned in the comment above spaceship_replacement, for
   strong_ordering, we are pattern matching something like:
   x == y ? 0 : x < y ? -1 : 1;
   and for partial_ordering
   x == y ? 0 : x < y ? -1 : x > y ? 1 : 2;
   but given the == comparison done first and the other comparisons only
   if == was false, we actually don't care if the other comparisons
   are < vs. <= (or > vs. >=), provided the operands of the comparison
   are the same; we know == is false when doing those and < vs. <= or
   > vs. >= have the same behavior for NaNs too
2) when y is an integral constant, we should treat x < 5 equivalently
   to x <= 4 etc.
3) the code punted if cond2_phi_edge wasn't a EDGE_TRUE_VALUE edge, but
   as the new IL shows, that isn't really needed; given 1) that
   > and >= are equivalent in the code, any of swapping the comparison
   operands, changing L[TE]_EXPR to G[TE]_EXPR or vice versa or
   swapping the EDGE_TRUE_VALUE / EDGE_FALSE_VALUE bits on the edges
   reverses one of the two comparisons

2021-05-21  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/94589
	* tree-ssa-phiopt.c (spaceship_replacement): For integral rhs1 and
	rhs2, treat x <= 4 equivalently to x < 5 etc.  In cmp1 and cmp2 (if
	not the same as cmp3) treat <= the same as < and >= the same as >.
	Don't require that cond2_phi_edge is true edge, instead take
	false/true edges into account based on cmp1/cmp2 comparison kinds.