match.pd: fold a round up spelled with inclusive or

Linux spells a power-of-two round up as:

  ((x - 1) | (2^k - 1)) + 1

For wrapping arithmetic, this is:

  (x + (2^k - 1)) & -(2^k)

The latter form needs one fewer operation and exposes the zero low bits.
Apply the identity to scalar and vector integral types.  Use :s on the
inclusive-or expression so the fold does not add work when that value
remains live.

  unsigned long round_up (unsigned long x)
  {
    return ((x - 1) | 4095) + 1;
  }

aarch64 -O2:

before:

	round_up:
		sub	x0, x0, #1
		orr	x0, x0, 4095
		add	x0, x0, 1
		ret

after:

	round_up:
		add	x0, x0, 4095
		and	x0, x0, -4096
		ret

Bootstrapped and tested on aarch64-none-linux-gnu.

gcc/ChangeLog:

	* match.pd (((X - 1) | C) + 1): New simplification.

gcc/testsuite/ChangeLog:

	* gcc.dg/tree-ssa/round-up-ior-1.c: New test.
	* gcc.dg/tree-ssa/round-up-ior-2.c: Likewise.

Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>