Local array creation about diospyros HOT 3 OPEN

Yep, to clarify, SIZE was 2 for this example, and that does help clarify why the i64 is used! I will go ahead and translate the memset and memcpy instructions.

from diospyros.

Ah, nice work boiling this down to a small example! This is incredibly helpful to look at!

It looks like the optimizations run ahead of Diospyros are already smart enough to convert that for loops that does initialization into a memset, huh? That's pretty cool.

To confirm: is SIZE equal to 2 here? That would explain this:

%1 = alloca i64, align 8
%tmpcast = bitcast i64* %1 to [2 x float]*

That is, LLVM is using a single i64 to store two floats for x.

Anyway, treating memset & memcpy calls like stores seem like exactly the right thing to do!

from diospyros.

The original problem was mostly fixed, but another similar problem relating in the load/store movement pass is being fixed, where loads and stores cannot be moved around memsets.

from diospyros.