[Bug tree-optimization/106594] [13 Regression] sign-extensions no longer merged into addressing mode

["roger at nextmovesoftware dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106594

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2022-08-12
             Status|UNCONFIRMED                 |NEW
                 CC|                            |roger at nextmovesoftware dot com
     Ever confirmed|0                           |1

--- Comment #3 from Roger Sayle <roger at nextmovesoftware dot com> ---
Ah interesting.  Because index is a char, the tree-level optimizers realize
that the shift by 4 can be an 8-bit shift instead of an int-sized shift. 
What's interesting is that because of (x & 3) << 4, is used, the optimizers
realize that because index can never be negative, that in the array memory
reference expression constellation_64qam[index], when the 8-bit index is being
sign extended, it is effectively being zero-extended.

I think that the aarch64 backend needs to be taught that in this case (because
of the AND), the zero extension is the same as (can be implemented using) a
sign-extension, i.e. restoring the original code generation.  Practically, the
sxtw;ldr[..lsl 2] above can legitimately be optimized to ldr[..sxtw 2] (or
ldr[..uxtw 2] like LLVM) in this case [cases like this].

(sign_extend:DI (and:SI x (const_int 3))) and
(zero_extend:DI (and:SI x (const_int 3))) should (ideally) produce the exact
same code [the more efficient if two implementations are possible].