https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63184
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
--- Comment #16 from Richard Biener <rguenth at gcc dot gnu.org> ---
Now to answer Jeff my original idea was to change how VN handles addresses
to use tree-affine in there. But that takes some time...
Given that we lower non-invariant addresses now another idea would be to
recognize
_4 = _3 + 4;
_5 = &a[1] + _4;
and reassociate that to combine the + 4 with the &a[1] (which is &a + 4).
Now the reassoc pass associates the constant last which makes this
unreliable to detect in a match.pd pattern (the constant may be far away)
but the reassoc pass itself could, when associating a chain, look for
whether the SSA name we start from is (single-)used in a POINTER_PLUS_EXPR
and the constant element in the reassoc ops[] array can be combined with
a constant offset in the address operand.
I'll see if that works out.
