Float32x4 GC mitigation needed!

[gmpassos]

Float32x4 and Int32x4 brings SIMD acceleration, replacing use of normal numbers (double & int) to a lane of 4 numbers.

One side effect of Float32x4 and Int32x4 is the increase of objects to be handled by GC.

Some strategy in the VM implementation of Float32x4 and Int32x4 is needed to reduce GC work. It seems that other
types, likeString, have a better performance.

[mraleph]

@gmpassos could you provide some runnable code and instructions for us to run/benchmark/profile ourselves?

In general most local allocations of SIMD values should be eliminated, but SIMD values still need to be boxed when passed around. It is possible to give SIMD values similar treatment as (non-nullable) double type - which we unbox in AOT mode in most places. However it would require language team decision on #43255 - to allow VM erase identity of SIMD values.

[gmpassos]

I have worked to show that in the last days:

https://github.com/eneural-net/eneural_net_dart/blob/master/example/eneural_net_benchmark_activation_functions.dart

This benchmark is only testing the activation function.

See how the Hertz decreases as memory is being filled.

You can teste with only "ActivationFunctionSigmoid" in the list (line 10).

[gmpassos]

@mraleph FYI: #47488