[stdlib] Refactor memset() to be generic

[martinvuyk]

Refactor memset to be generic over scalars and trivial types to be filled with scalar values of the same bitwidth.

[soraros]

Thanks for the contribution and congrats on the performance improvements! However, I do wonder if memset is right place to add such features. It has the semantics of copying raw memory, and init_pointee_copy is not that.

[martinvuyk]

Hi @soraros the implementation in _memset_impl also just copies a value on a vector and then stores it in the memory address range:

alias simd_width = simdwidthof[Scalar[type]]()
var vector_end = _align_down(count, simd_width)

for i in range(0, vector_end, simd_width):
    ptr.store(i, SIMD[type, simd_width](value))

for i in range(vector_end, count):
    ptr.store(i, value)

And the original signature of memset also made me consider doing this as I interpreted it like it is supposed to be generic:

fn memset[
    type: AnyType, address_space: AddressSpace
](ptr: UnsafePointer[type, address_space], value: UInt8, count: Int):
  ...

the only thing this new version assumes differently than the original is that you are able to copy a value of the same type as the pointer into that address range.

IMO these two branches are doing the same thing, copy a thing count times from the pointer start

if dt is not DType.invalid:
    var p = ptr.bitcast[Scalar[dt]]()
    _memset_impl[dt](p, rebind[Scalar[dt]](value), count)
else:
    for i in range(count):
        (ptr + i).init_pointee_copy(value)

Even memcpy also does copies onto vectors before storing in an address range, so there is no "raw" bit-mapping on an electrical level either:

# Copy in 32-byte chunks.
alias chunk_size = 32
var vector_end = _align_down(n, chunk_size)
for i in range(0, vector_end, chunk_size):
    dest_ptr.store(i, src_ptr.load[width=chunk_size](i))
for i in range(vector_end, n):
    dest_ptr.store(i, src_ptr.load(i))

[JoeLoser]

!sync

[JoeLoser]

FYI we'll want to do some benchmarking here (both in comp and runtime) with this change internally, so that's why it has not been merged internally yet. Thanks for your patience and contribution! 🔥

[martinvuyk]

@jackos to reduce the scope of this PR I reverted the changes to make memset generic over memory only types (I have a better idea to not bloat compile times). I also removed the DType.get_dtype() to its own PR #3810. This PR now only makes memset generic over scalars and trivial types to be filled with scalar values of the same bitwidth.

[JoeLoser]

!sync