[Analyzer]: Recommend usage of constant in stackalloc expressions

[just-ero]

.NET 8 introduces some performance-related warnings. This made me think of another commonly discussed issue; using constant sizes in stackalloc expressions.

A dynamically-sized stackalloc array can potentially be worse for performance. An analyzer warning for such a situation could be beneficial.

Span<byte> data = len <= 128 ? stackalloc[len] : new byte[len]; // warning: Use constant size in `stackalloc` expression

// fix (have to use `.Slice(0, len)` when using the span)
Span<byte> data = len <= 128 ? stackalloc[128] : new byte[len];

Here are some relevant benchmark results: https://gist.github.com/just-ero/7ef8dd2b49fa5c1cf8d47bbbc6112bd6.

[ghost]

Tagging subscribers to this area: @dotnet/area-meta
See info in area-owners.md if you want to be subscribed.

Issue Details

---

.NET 8 introduces some performance-related warnings. This made me think of another commonly discussed issue; using constant sizes in stackalloc expressions.

A dynamically-sized stackalloc array can potentially be worse for performance. An analyzer warning for such a situation could be beneficial.

Span<byte> data = len <= 128 ? stackalloc[len] : new byte[len]; // warning: Use constant size in `stackalloc` expression

// fix (have to use `.Slice(0, len)` when using the span)
Span<byte> data = len <= 128 ? stackalloc[128] : new byte[len];

Here are some relevant benchmark results: https://gist.github.com/just-ero/7ef8dd2b49fa5c1cf8d47bbbc6112bd6.

Author:	just-ero
Assignees:	-
Labels:	area-Meta
Milestone:	-

[EgorBo]

1. What's the threshold when analyzer should not suggest that, e.g. should we promote len <= 4096 ? stackalloc[len]?
2. In some cases it might increase overall stack usage when the actual len is small. On macOS all secondary threads have 512Kb of stack by default.
3. In some cases non-constant stackalloc might be faster if the actual length is small, I was able to reproduce it:

4. With SkipLocalsInit the performance difference might be similiar for sizes >= 32
5. In theory, we could optimize this in JIT