The CheckedInt32Mod lowering in the EffectControlLinearizer wasn't
playing well with subsequent optimizations in the MachineOperatorReducer
especially due to the use of Int32Mod, which introduces another (floating)
diamond in the MachineOperatorReducer. Switching to Uint32Mod and explicit
sign handling fixes the problem, plus we also do the mask trick in the
case where the left hand side is negative now.

With this change the performance on the benchmark mentioned in the bug
report goes from

  console.timeEnd: binary, 1872.346000
  console.timeEnd: modulo, 5967.464000
  console.timeEnd: binary, 6006.789000
  console.timeEnd: modulo, 6293.496000
  console.timeEnd: binary, 5969.264000
  console.timeEnd: modulo, 6291.874000

to

  console.timeEnd: binary, 1876.464000
  console.timeEnd: modulo, 5846.643000
  console.timeEnd: binary, 5962.545000
  console.timeEnd: modulo, 5972.639000
  console.timeEnd: binary, 5958.221000
  console.timeEnd: modulo, 5973.171000

so even the peak performance of the modulus is now mostly the same as
the binary bitwise and.

Bug: v8:8069
Change-Id: Iaf3828fc0f6c53352367e8bf6c42534f8b13bfb3
Reviewed-on: https://chromium-review.googlesource.com/1180971Reviewed-by: Jaroslav Sevcik <jarin@chromium.org>
Commit-Queue: Benedikt Meurer <bmeurer@chromium.org>
Cr-Commit-Position: refs/heads/master@{#55211}