Re: [PATCH] Find more cases for AUTO_INC_DEC

[Adam Nemet <anemet at lnxw dot com>]

Richard Henderson writes:
> Have a look at Joern's large auto-increment patch.

OK, I did.  It does not fix the basic testcase.  It seems to me that
unless recognize_related_for_insn creates an entry in rel_base_list
nothing gets optimized in optmize_relate_values_0.  In the testcase,
insns that set up potentially related registers are hard regs and
recognize_related_for_insn ignores those.  It is not clear if this is
just a temporary limitation but this prevents any related-value
optmization on the testcase.

But in my opinion even if this pass could deal with all auto-increment
opportunities the real question is whether it will supersede the basic
auto-increment optimization in flow.  I assume it would not.  All the
required information to do the tranformation there is readily
avaiable: reg_next_use is used to set up LOG_LINKs not just for
AUTO_INC.  And if I am correct so far then my patch just takes another
look at the same reg_next_use information in terms of the RHS of the
addition: (r = p + 1; *(p + 1)), if P's next use is the address
expression then we can substitute R there and then do the basic
auto-increment optimization assuming R is not used or set in between.
This results in 52% more auto-increment optimizations during a
ppc-darwin bootstrap.

In the meantime, I also tested arm-elf on the simulator with no
regression.  On ARM the basic testcase behaves similarly after the
patch:

Before:

  stwu4:
  	  str     r1, [r0, #4]
  	  add     r0, r0, #4
  	  bx      lr


After:

  stwu4:
  	  str     r1, [r0, #4]!
  	  bx      lr


Adam