Fix invalid store offsets in memset polyfill

[dcodeIO]

Some of the stores in memset are invalid for certain arguments due to enforcing a constant offset where we'd instead want the whole offset to wrap around. Traced this back to #1103 where I assumed that we'd be using lowMemoryUnused anyhow (i.e. never storing to the low 1024 bytes of memory), which actually isn't always the case in memset.

Discovered in #1785.

• I've read the contributing guidelines

[dcodeIO]

The change here essentially reverts to the state before #1103.

[MaxGraey]

Hmm, it seems this cause to significant regression in term of code size and perhaps performance. I prefer solve (fix) this on binaryen side instead just revert all changes

[MaxGraey]

Case store<u8>(-2, 0, 3) possible only if you manually set memory.fill(0, 0, 2). But usually it's not possible in managed runtime when we have memoryBase >= 8. So I guess better add fix to binaryen which normalize store<u8>(-2, 0, 3) to store<u8>(1, 0, 0)

[dcodeIO]

I am not sure that this can be fixed on the Binaryen side, since what we are emitting (store with explicit constant offset) is already invalid for certain, even though uncommon, inputs. I agree that the code size regression here is unfortunate, but at least it's correct.

[MaxGraey]

Alright, now it much better