https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112398
--- Comment #8 from Jeffrey A. Law <law at gcc dot gnu.org> ---
So Alexey's patch helps the first case, generating the more efficient lbu+xori
sequence.
I suspect it's not helping the 2nd case because the constant is going to be out
of range. Given the 2nd case also suffers from a lack of HI->DI zero
extension, there is some budget to improve that code. If someone wanted to
tackle that, be my guest. But be aware that with Zbb enabled the code as it
stands is probably optimal.
Hints for anyone wanting to tackle the second testcase. You'll need a
define_insn_and_split for this construct:
(set (reg:DI 141)
(xor:DI (reg:DI 140)
(const_int 65535 [0xffff])))
It has to be a define_insn_and_split because it would be a 2->2 split which
isn't generally profitable. It would need to be conditional on !Zbb and only
apply before register allocation.
The 3rd case can't be optimized in this way because the result of the NOT is
used as a source operand in multiple instructions.
