Add documentation for hot/cold splitting #73029
Conversation
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsThis PR adds documentation for the current status of hot/cold splitting support in the JIT, as of my internship's conclusion.
|
|
@AndyAyersMS PTAL, thank you! |
|
|
||
| This document describes the current state of hot/cold splitting in the JIT. | ||
|
|
||
| Hot/Cold splitting is a PGO optimization where generated code is split into frequently-executed ("hot") and |
There was a problem hiding this comment.
Hot/cold splitting is not strictly Profile Guided Optimization. It is quite effective even without profile data.
|
|
||
| ## Testing the JIT Without Runtime Support | ||
|
|
||
| Without runtime support for hot/cold splitting in .NET Core yet, testing the JIT's existing hot/cold splitting support |
There was a problem hiding this comment.
| Without runtime support for hot/cold splitting in .NET Core yet, testing the JIT's existing hot/cold splitting support | |
| Without runtime support for hot/cold splitting in .NET as of summer 2022, testing the JIT's existing hot/cold splitting support |
Nit: We have retired .NET Core as the name a few years ago. We are calling the product just .NET now.
|
Thanks for the review @jkotas, I've added your feedback in. |
| Hot/Cold splitting in the JIT was previously implemented for AOT-compiled NGEN images in .NET Framework. With hot/cold | ||
| splitting support in Crossgen2 [in progress](https://github.com/dotnet/runtimelab/tree/feature/hot-cold-splitting) | ||
| (and no existing support for splitting dynamically-generated code), the feature has gone untested in the JIT since | ||
| retiring .NET Framework -- as such, regressions have surely been introduced. Furthermore, this existing implementation | ||
| has several major limitations; functions with certain features, like exception handling or switch tables, cannot be | ||
| split yet. Finally, with ARM64/x64 performance parity being a high priority for CoreCLR, it is key hot/cold splitting | ||
| is implemented for ARM64 code generation, too. |
There was a problem hiding this comment.
I think somebody could get confused by the mixed use of tense here and not be entirely sure what the "current state" actually is.
Consider sectioning this part off as ## Background and describing just the before state in past tense.
| Hot/Cold splitting is an optimization where generated code is split into frequently-executed ("hot") and | ||
| rarely-executed ("cold") parts and placed in separate memory regions. Increased hot code density better leverages |
There was a problem hiding this comment.
Nit -- keep an eye out for passive voice and try to avoid it where you can. Here it seems pretty easy:
| Hot/Cold splitting is an optimization where generated code is split into frequently-executed ("hot") and | |
| rarely-executed ("cold") parts and placed in separate memory regions. Increased hot code density better leverages | |
| Hot/Cold splitting is an optimization that splits code into frequently-executed ("hot") and | |
| rarely-executed ("cold") parts and places the parts in separate memory regions. Increased hot code density better leverages |
|
|
||
| Hot/Cold splitting is an optimization where generated code is split into frequently-executed ("hot") and | ||
| rarely-executed ("cold") parts and placed in separate memory regions. Increased hot code density better leverages | ||
| spatial locality, which can improve application performance via fewer instruction cache misses, less OS paging, and |
There was a problem hiding this comment.
| spatial locality, which can improve application performance via fewer instruction cache misses, less OS paging, and | |
| spatial locality, improving application performance via fewer instruction cache misses, less OS paging, and |
| architecture-specific workarounds. However, if the JIT supports splitting functions multiple times in the future, this | ||
| may be worth revisiting. | ||
|
|
||
| In the absence of PGO data, the JIT assumes exceptions occur rarely to justify moving handlers to the cold section. |
There was a problem hiding this comment.
| In the absence of PGO data, the JIT assumes exceptions occur rarely to justify moving handlers to the cold section. | |
| In the absence of PGO data, the JIT assumes exceptions occur rarely; this justifies moving handlers to the cold section. |
|
@AndyAyersMS thank you for the review -- I've cleaned up some of the phrasing. |
This PR adds documentation for the current status of hot/cold splitting support in the JIT, as of my internship's conclusion.