performance: Peephole optimise (128 : Word8 & x == 0) resp. (128 : Word8 >= x)

[ggreif]

to bitShiftRight 31

These happen frequently in the Random.mo code.

Note: this issue should be in motoko proper, but it ended up here by mistake. I cannot move it there.

[ggreif]

To give a concrete example, this Motoko

func upperMask(a : Word8) : Bool = a & (0x80 : Word8) != (0 : Word8);

compiles currently to this Wasm:

  (func $upperMask (type 5) (param $clos i32) (param $a i32) (result i32)
    local.get $a
    i32.const -2147483648 // 0x80000000
    i32.and
    i32.eqz
    i32.eqz
    i32.const 1
    i32.shl)

We could have (unverified)

  (func $upperMask (type 5) (param $clos i32) (param $a i32) (result i32)
    local.get $a
    i32.const 31
    i32.shr
    i32.const 1
    i32.shl)

[ggreif]

Something like

func upperBitGe(a : Word8) : Bool = a >= (0x80 : Word8);
func upperBitGt(a : Word8) : Bool = a > (0x7F : Word8);

with upperBitGt will likely not work out of the box:

  (func $upperBitGt (type 5) (param $clos i32) (param $a i32) (result i32)
    local.get $a
    i32.const 2130706432 // 0x7f000000
    i32.gt_u
    i32.const 1
    i32.shl)
  (func $upperBitGe (type 5) (param $clos i32) (param $a i32) (result i32)
    local.get $a
    i32.const -2147483648 // 0x80000000
    i32.ge_u
    i32.const 1
    i32.shl)

unless we insert redundant hint instructions into the Wasm:

  (func $upperBitGt (type 5) (param $clos i32) (param $a i32) (result i32)
    local.get $a
    hint (i32.const 0xFF000000; i32.and) -- adds some info about the TOS: 24 lower bits are zeroed
    i32.const 2130706432 // 0x7f000000
    i32.gt_u
    i32.const 1
    i32.shl)

With such hints we can perform the i32.const 31; i32.shr trick.

I envision the hint meta-instruction (never emitted) to operate on top-of-stack asserting a semantic nop. This part is probably not beginner-friendly, though.

[ggreif]

eqz; eqz; if --> if this is already performing: double swapping of the if legs.

[ggreif]

Also note that after a 1-bit masking shr 31; shl 1 is just shr 30. Free tagging! Of course that is a different optimisation, since it needs the and to be preserved.

[ggreif]

See motoko/gabor/peephole-177 for an implementation sketch (very WIP).

[ggreif]

Another reduction path:

    block (result i32)  ;; label = @1
      i32.const 1
    end
    drop

block --> const --> eliminated by drop