[Bug rtl-optimization/26190] combine misses some distributivity

["wilson at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26190

--- Comment #3 from Jim Wilson <wilson at gcc dot gnu.org> ---
This ior/shift/xor optimization be done during combine with a simplify-rtx.c
patch.  I wrote a prototype and tested it.  Combine canonicalizes shift/logical
as logical/shift, so we actually have to look for ior/xor/shift.  I see that
the optimization does not happen on arm and aarch64 because the 0x4002 value
does not fit into the immediate range for logical ops, gets loaded into a
register, pulled out of the loop, and hence is not available to combine.  We
would have to perform the optimization earlier for it to work for arm/aarch64. 
I tried MIPS, and see that because MIPS promotes HImode to SImode we don't have
enough info to prove that the opt is safe.  We need type info to make this
work.  This prototype does work for the x86 target which has both 16-bit
immediates and HImode instructions.

Next I wrote a prototype for match-and-simplify.  This one works for all 4
targets.  We have to handle ior/shift/xor as match-and-simplify does not
canonicalize logical/shift.  I noticed that if I change the xor to an ior, then
it gets optimized in combine because of the canonicalization for mips and x86,
but not for arm/aarch64 again because the constant was pulled out of the loop. 
So it seems that match-and-simplify should canonicalize shift/logical to
logical/shift too.  That would also reduce the number of patterns we need to
match when performing this ior/shift/xor optimization.

These are prototype patches that need a bit more work to handle more cases, and
to prove that they test all conditions necessary to make them safe.  This is my
first attempt to use the wi:: functions, so there might be a better way to do
this.