768-bit multiplication implemented in wasm

mul768.wat

;; C ABI: void mul768(uint64_t* product, const uint64_t* a, const uint64_t* b)

(module
  (type (;0;) (func (param i32 i32 i32)))
  (import "env" "memory" (memory (;0;) 256 256))
  (func (;0;) (type 0) (param i32 i32 i32)
    (local i32 i32 i32 i32 i32 i32 i64 i64 i64 i64 i64 i64 i64 i64)
    i32.const 11
    local.set 5
    loop  ;; label = @1
      local.get 3
      i32.const 3
      i32.shl
      local.get 2
      i32.add
      local.set 6
      i32.const 12
      local.get 3
      i32.sub
      local.tee 7
      i32.const 1
      i32.gt_u
      if  ;; label = @2
        i32.const 0
        local.set 4
        i64.const 0
        local.set 11
        loop  ;; label = @3
          local.get 4
          i32.const 3
          i32.shl
          local.get 1
          i32.add
          i64.load
          local.tee 9
          i64.const 4294967295
          i64.and
          local.set 10
          local.get 6
          i64.load
          local.tee 12
          i64.const 4294967295
          i64.and
          local.tee 13
          local.get 9
          i64.const 32
          i64.shr_u
          local.tee 14
          i64.mul
          local.get 10
          local.get 13
          i64.mul
          local.tee 9
          i64.const 32
          i64.shr_u
          i64.add
          local.tee 13
          i64.const 4294967295
          i64.and
          local.get 10
          local.get 12
          i64.const 32
          i64.shr_u
          local.tee 15
          i64.mul
          i64.add
          local.set 10
          local.get 9
          i64.const 4294967295
          i64.and
          local.get 10
          i64.const 32
          i64.shl
          i64.or
          local.tee 16
          local.get 3
          local.get 4
          i32.add
          i32.const 3
          i32.shl
          local.get 0
          i32.add
          local.tee 8
          i64.load
          i64.add
          local.tee 9
          local.get 11
          i64.add
          local.set 12
          local.get 9
          local.get 12
          i64.gt_u
          i64.extend_i32_u
          local.get 16
          local.get 9
          i64.gt_u
          i64.extend_i32_u
          local.get 14
          local.get 15
          i64.mul
          local.get 13
          i64.const 32
          i64.shr_u
          i64.add
          local.get 10
          i64.const 32
          i64.shr_u
          i64.add
          i64.add
          i64.add
          local.set 11
          local.get 8
          local.get 12
          i64.store
          local.get 5
          local.get 4
          i32.const 1
          i32.add
          local.tee 4
          i32.ne
          br_if 0 (;@3;)
        end
      else
        i64.const 0
        local.set 11
      end
      local.get 0
      local.get 0
      i64.load offset=88
      local.get 11
      local.get 7
      i32.const -1
      i32.add
      i32.const 3
      i32.shl
      local.get 1
      i32.add
      i64.load
      local.get 6
      i64.load
      i64.mul
      i64.add
      i64.add
      i64.store offset=88
      local.get 5
      i32.const -1
      i32.add
      local.set 5
      local.get 3
      i32.const 1
      i32.add
      local.tee 3
      i32.const 12
      i32.ne
      br_if 0 (;@1;)
    end)
  (export "mul768" (func 0)))