Optimize comparisons of adjusted values #5008

Activity

1. kripken commented on Sep 2, 2022

   

   kripken

   on Sep 2, 2022

   MemberAuthor

   More actions

   Another example from rule #10:

   (i32.lt_s
    (i32.shr_s
     (local.get $0)
     (i32.const 1)
    )
    (i32.const 12)
   )
    =->
   (i32.lt_s
    (local.get $0)
    (i32.const 24)
   )
2. kripken commented on Sep 2, 2022

   

   kripken

   on Sep 2, 2022

   MemberAuthor

   More actions

   This also happens with floats, though there we may need fast-math, e.g.

   (f32.ge
    (f32.add
     (local.get $0)
     (f32.const -9.99999993922529e-09)
    )
    (f32.const 0)
   )
    =->
   (f32.gt
    (local.get $0)
    (f32.const 9.99999905104687e-09)
   )

   (rule #21)
3. kripken commented on Sep 6, 2022

   

   kripken

   on Sep 6, 2022

   Last edited by kripken

   MemberAuthor

   More actions

   Another example:

    (i32.gt_u
     (i32.add
      (i32.shr_u
       (local.get $0)
       (i32.const 1)
      )
      (i32.const 8)
     )
     (i32.const 65535)
    )

   The add's left operand is less than 32 bits, so it cannot overflow even with an added 8, and so x + 8 >= 65535 can be turned into x >= 65527.

   edit: This and related things seem to be the top superoptimizer finding for Java and Dart, followed by #5008 (comment)
4. kripken

   self-assigned this

   on Sep 9, 2022
5. kripken

   added a commit that references this issue on Sep 9, 2022

   OptimizeInstructions: Optimize comparisons with an added offset (#5025)

   ...

   1e8eb59
6. kripken commented on Sep 13, 2022

   

   kripken

   on Sep 13, 2022

   MemberAuthor

   More actions

   All comparisons <, >=, etc should be handled with +. That leaves things like shifts, and float operations, as mentioned above.