Add core::intrinsics::simd

[calebzulawski]

Intended to close rust-lang/portable-simd#381.

r? ralfjung

[RalfJung]

Cc @rust-lang/opsem

[workingjubilee]

🤨 I forgot about that "fun" detail. Nevermind, I guess.

[workingjubilee]

actually, while we're here, does anyone know if these are actually incompatible signatures, or is that lint just overly strict...?

[digama0]

doesn't that mean the rename should also happen in portable-simd?

I can't imagine that that generic parameter names would matter here, but it still seems like a good idea to ensure the names are the same for clarity.

[RalfJung]

Yeah that looks like an overeager lint to me. So it can be allowed temporarily until the names are back in sync.

[RalfJung]

I have no idea what this means.^^ There's too many different notions of "first" here. Also, there's two cases to consider (output a packed int vs output an array).

What about something like this:
No matter whether the output is an array or an unsigned integer, it is treated as a single contiguous list of bits. The bitmask is always packed on the least-significant side of the output, and padded with 0s in the most-significant bits. The order of the bits depends on endianess:

• On little endian, the least significant bit corresponds to the first vector element.
• On big endian, the least significant bit corresponds to the last vector element.

I think this also needs examples to have any chance of being comprehensible.

As always, please first state the types and then start discussing the details; without knowing what U is the rest of this is even harder to understand.

[calebzulawski]

This is much better. I took your text and added a simple example

[calebzulawski]

I thought those intrinsics were added? Or is this something bootstrap compiler related?

[RalfJung]

Yeah this probably needs #[cfg(not(bootstrap))]

[RalfJung]

Looking good. :) We can always refine this later if more things come up.
I assume the next step is to change portable-simd to import this module rather than directly importing the intrinsics?

@bors r+

[bors]

📌 Commit e61aaf9 has been approved by RalfJung

It is now in the queue for this repository.

[calebzulawski]

Looking good. :) We can always refine this later if more things come up. I assume the next step is to change portable-simd to import this module rather than directly importing the intrinsics?

Yep! And one more feature gate to remove

[bors]

⌛ Testing commit e61aaf9 with merge 558ac1c...

[bors]

☀️ Test successful - checks-actions
Approved by: RalfJung
Pushing 558ac1c to master...

[rust-timer]

Finished benchmarking commit (558ac1c): comparison URL.

Overall result: no relevant changes - no action needed

@rustbot label: -perf-regression

Instruction count

This benchmark run did not return any relevant results for this metric.

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	1.4%	[0.5%, 3.9%]	5
Regressions ❌ (secondary)	3.9%	[2.6%, 4.5%]	3
Improvements ✅ (primary)	-2.0%	[-2.0%, -2.0%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.8%	[-2.0%, 3.9%]	6

Cycles

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-0.5%	[-0.5%, -0.5%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-0.5%	[-0.5%, -0.5%]	1

Binary size

This benchmark run did not return any relevant results for this metric.

Bootstrap: 672.757s -> 673.009s (0.04%)
Artifact size: 312.46 MiB -> 312.42 MiB (-0.01%)

[RalfJung]

In implementing this, I am confused. Isn't this equivalent to simd_gather?

[farnoy]

gather accepts a vector of N pointers, where each element will be loaded from its corresponding pointer.

masked_load accepts a single pointer and all elements of the resulting vector (when unmasked) are loadded from a constant offset from that pointer. i.e the first element will be loaded from ptr, second from ptr.offset(1), and so on.

[RalfJung]

Then the documentation here is wrong. For masked_load it says

/// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`.

[farnoy]

Yeah, looks like we missed this. I'll create a follow up PR soon

[farnoy]

#119203