[frontend] IIFEs could be made cheaper?

[modulovalue]

This issue contains a small benchmark that attempts to show that the performance of IIFEs could be improved.

In short, It looks like IIFEs are not being converted to Kernel BlockExpressions (which have no equivalent on the Dart level), and it looks like they could be converted to BlockExpressions as an optimization.

Consider the following:

void main() {
  const iterations = 1000000;
  for (int i = 0; i < 20; i++) {
    print("=" * 80);
    measure(
      fn: () {
        int total = 0;
        for (int i = 0; i < iterations; i++) {
          {
            total += i;
          }
        }
        print(" - total: " + total.toString());
      },
      name: "   => in a block statement",
    );
    measure(
      fn: () {
        int total = 0;
        for (int i = 0; i < iterations; i++) {
          (){
            total += id_fn(i);
          }();
        }
        print(" - total: " + total.toString());
      },
      name: "   => in an IIFE",
    );
    print("=" * 80);
  }
}

int id_fn(
  final int i,
) =>
    i;

void measure({
  required final void Function() fn,
  required final String name,
}) {
  final s = Stopwatch();
  s.start();
  fn();
  s.stop();
  print(
    name + " took: " + (s.elapsedMicroseconds / 1000).toString() + "ms",
  );
}

AOT:

 - total: 499999500000
   => in a block statement took: 0.48ms
 - total: 499999500000
   => in an IIFE took: 4.417ms

JIT:

 - total: 499999500000
   => in a block statement took: 0.52ms
 - total: 499999500000
   => in an IIFE took: 1.59ms

IIFEs are useful for keeping code close to where it is being used which reduces the effort needed to reason about code. I think that making IIFEs cheaper would be great.

[ds84182]

I don't actually understand why you'd want an IIFE there (or at all). To me it doesn't seem to provide anything of value as blocks are also close to where its used.

[modulovalue]

The code that I've provided in the issue description is only meant to demonstrate the difference in performance.

This dart-lang/site-www#4613 issue contains more about the why aspect of IIFEs.

Edit: In short, IIFEs are expressions and can be used to "localize" lists of statements where only expressions are being expected.

[johnniwinther]

cc @jensjoha @alexmarkov

[jensjoha]

Isn't this more the VM that should inline it but apparently doesn't?
Or maybe, looking at the flowgraph, inlines, but still keeps all the closure stuff, e.g.:

 28:     v19 <- AllocateUninitializedContext:42(num_variables=1) T{Context}
 30:     StoreField(v19 . i = v5 T{_Smi}, NoStoreBarrier)
 32:     StoreField(v19 . Context.parent = v30 T{Context?}, NoStoreBarrier)
 34:     v22 <- BinarySmiOp:32(+ [-o], v5 T{_Smi}, v21) [-4611686018427387903, 1000000] T{_Smi}
 36:     StoreField(v19 . i = v22, NoStoreBarrier)
 38:     ParallelMove rcx <- rax goto:48 B5

though I might be reading it wrong.

/cc @mraleph

(this was on this modified code btw:

        for (int i = 0; i < iterations; i++) {
          (){
            total += i;
          }();
        }

)