[stdlib] Span slicing with negative step

[msaelices]

Fixes #3649

  1> s = String("01234")  
(String) s = "01234"
  2> for i in s.as_bytes_span()[::-1]: 
  3.   print(chr(int(i[]))) 
  4.    
4
3
2
1
0
  5>  

[msaelices]

What's the best way to ensure the buffer allocated memory is destroyed when the span is deleted?

[martinvuyk]

Hmmm, not sure since I don't really understand lifetimes yet. But I think building a var items = List[T](unsafe_pointer=dest_ptr, size=new_len, capacity=new_len) and returning Span[T, lifetime](items^) will work?

Either way I think this should probably return an iterator in the future (once we have them), it doesn't make much sense to allocate for this since it is a read only view on the data.

[soraros]

• Span suppose to be a non-owning view, I don't think this is the right fix.
• Actually, I'm not even sure a span with only a length is capable of representing slicing with negative step at all.

[msaelices]

• Span suppose to be a non-owning view, I don't think this is the right fix.
• Actually, I'm not even sure a span with only a length is capable of representing slicing with negative step at all.

Yes, it's a little bit weird. However, in Python you can slice a slice, like this:

>>> l = list(range(10))[2:5][::-1]
>>> l
[4, 3, 2]
>>> 

How Mojo could support this kind of operations without the support of negative steps?

[bgreni]

• Actually, I'm not even sure a span with only a length is capable of representing slicing with negative step at all.

I made a change previously to kinda hack it by giving Span a Slice member here #2950, but apparently it caused issues and was reverted internally? It feels like something we should be able to do though.

[msaelices]

• Actually, I'm not even sure a span with only a length is capable of representing slicing with negative step at all.

I made a change previously to kinda hack it by giving Span a Slice member here #2950, but apparently it caused issues and was reverted internally? It feels like something we should be able to do though.

That is amazing! I wonder why this was reverted. To me it's a must to have slicing matching up the Python ones.

[soraros]

#2950 essentially demonstrates the points I made above: Span itself is indeed incapable of representing stride. I believe it's still desirable for Span to remain a simple pointer-length pair. Perhaps we could introduce a StridedSpan?

[martinvuyk]

What do you think of this proposal #3653 ?

[bgreni]

#2950 essentially demonstrates the points I made above: Span itself is indeed incapable of representing stride. I believe it's still desirable for Span to remain a simple pointer-length pair. Perhaps we could introduce a StridedSpan?

That's reasonable, I wish we had some context for why that change didn't work internally though.