Propagate constant load/store offsets more efficiently

[MaxGraey]

For example:

const index = 4;
store<i32>(ptr + <usize>index * sizeof<i32>(), value);
// emit as: (i32.store (i32.add (get_local $0) (i32.const 16)) (get_local $2))

Good optimize to this in compile time:

store<i32>(ptr, value, 4 * sizeof<i32>());
// emit as: (i32.store offset=16 (get_local $0) (get_local $2))

So may be this good doing inside load / store buildins.

That do you think?

[dcodeIO]

The compiler itself avoids doing optimizations that diverge from the source, and leaves these to Binaryen. Actually, I had a similar idea, but added the constantOffset parameter instead because of this comment.

[MaxGraey]

As mentioned kripken "but you need to know more about the pointer $1 here". So it seems it hard to do on binaryen side but you have much more information about managed pointers.

[dcodeIO]

Still the compiler would have to know for sure that ptr doesn't overflow, which doesn't seem trivial to me considering that the value is passed in by the user. What the compiler already does is, when you access a field on a class, it uses its actual offset as the load/store's offset. It should probably also do this for array accesses eventually.

[MaxGraey]

It should probably also do this for array accesses eventually
It will be perfect! Actually this motivate me open this proposal

[MaxGraey]

Closed via #1103