https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90663
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2019-05-29
Component|tree-optimization |middle-end
Target Milestone|--- |7.5
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
&a[2] &MEM[(void *)&a + 2B]
that's two equivalent ways of writing the same address but they are not
equal as far as operand_equal_p is concerned. More "precise" equality
can be determined by using get_addr_base_and_unit_offset (for addresses
that are known not to be variable) or get_inner_reference (for the rest).
Note the latter eventually relies on similar foldings / canonicalizations
being applied when the variable offset tree is built. Even more precise
equivalence may be obtained by using tree-affine but that's even more
expensive. What we lack is a get_inner_reference_aff that doesn't
first use get_inner_reference and thus avoids building/folding of the
GENERIC offset tree.
operand_equal_p is (more-or-less) checking for syntactical equivalence
while here we want value equivalence.
Anyhow, I suggest to use get_addr_base_and_unit_offset in the gimple
folder since there variable addresses will not appear literally but
only in separate SSA defs.
