[PATCH] D108435: [CSSPGO] split context string II - reader/writer changes

[Wenlei He via Phabricator via llvm-commits]

wenlei added inline comments.


================
Comment at: llvm/lib/ProfileData/SampleProfReader.cpp:326-327
+        // otherwise it's treated as context-less function name only.
+        auto Context = createContext(FName);
+        FContext.setContext(Context);
         ++CSProfileCount;
----------------
hoy wrote:
> wenlei wrote:
> > hoy wrote:
> > > wenlei wrote:
> > > > hoy wrote:
> > > > > wenlei wrote:
> > > > > > Even though we only need to deal with string context in profile reader/writer for text profile, it's probably still cleaner to keep all string context related parsing into SampleContext. 
> > > > > > 
> > > > > > `createContext` is more like a ctor. I'd prefer keep string decoding, createContext in its original place in SampleContext. That way, we can construct a context from string, SampleContext::setContext remain a private helper too, and the logic here can be simpler, just like before.
> > > > > > 
> > > > > > SampleContext does still have getContextString and toString, so it's not really isolated from string representation, might as well keep all string stuff together there for consistency. 
> > > > > Makes sense. Moved back to SampleContext.
> > > > Thanks. Can you move it back in D107299, so we don't see the change back and forth, just for review?
> > > > 
> > > > Can we also fold `FName.startswith("[")` in to SampleContext? 
> > > > 
> > > > Additionally, why we do need a `createContextFromString` instead of using overload ctors? This is also inconsistent between how SampleContext is created from text profile (createContextFromString) vs from binary profile (ctor). What I was thinking is have SampleContext decide how to create the object, so there's no changed needed here, just like before. 
> > > Constructing a CS context will require additional parameter than non-CS profile, especially the underlying context vector. That is causing the inconsistency with the context split work and I'm separating the construction of CS and non-CS contexts. Currently CS context is only constructible from a `SampleContextFrameVector`. 
> > > 
> > > Another reason is that I was hoping to construct `SampleContext` in a quick manner from `StringRef` to favor non-CS profile.
> > > 
> > > 
> > > Can you move it back in D107299, so we don't see the change back and forth, just for review?
> > 
> > Sorry my bad, I meant to say D108433. 
> > 
> > I think the key blocker for everything to be taken care from within ctor is that you need reader to own the context created from the string. How about having a ctor with StringRef and CSNameTable as parameter - it puts the context into CSNameTable (owned by reader) for CS profile, and the CSNameTable would be ignored for non-CS profile. 
> > 
> > The current implementation works, but reader has to be aware of the actual string representation of context. I thought it'd be cleaner if such representation is all dealt with from within SampleContext. Currently, it's indeed mostly handled by SampleContext, except it's bleeding into reader here. 
> > 
> > 
> Yeah, that's why I originally put everything related to context string parsing in the reader. I thought they were related to profile-specific representation that `SampleContext` shouldn't care. 
> 
> I guess moving the parsing code back is better for extension and code sharing, when we have new module doing the same thing in the future.  Adding a new constructor with StringRef and CSNameTable as parameter can work but feel like it's easier and more clear for the reader to know what it constructs. Hiding that from the reader is fine, but asking the reader to provide an underlying CS name table which may not be needed for non-CS sounds a bit confusing. That also kinds of exposes reader-specific implementation to `SampleContext`.
> 
> I think having the reader be aware of the specific representation might be reasonable, since context representation is a part of the profile format. What do you think?
> 
> 
> 
> 
> asking the reader to provide an underlying CS name table which may not be needed for non-CS sounds a bit confusing. That also kinds of exposes reader-specific implementation to SampleContext. 

I don't see this as exposing reader-specific stuff to sample context. It'd just be the expectation of the sample context API that requires a buffer to hold newly created context. 

This is in essence no different from how string buffer is passed in getRepInFormat (and only used for MD5), and I don't see it as coupling between reader and context.  

> I think having the reader be aware of the specific representation might be reasonable, since context representation is a part of the profile format. 

I think the actual string presentation of context is not tied to the format, but rather the format is using the string representation directly. Also if string representation is considered part of format, then all of the string related stuff should be part of reader (like your earlier change). The possible reusing of string functions you mentioned also shows that the string representation is not format-specific.


================
Comment at: llvm/lib/ProfileData/SampleProfReader.cpp:638
 
-  auto FName(readStringFromTable());
-  if (std::error_code EC = FName.getError())
+  auto FContext(readNameFromTable(IsContextName));
+  if (std::error_code EC = FContext.getError())
----------------
hoy wrote:
> wenlei wrote:
> > wenlei wrote:
> > > hoy wrote:
> > > > wenlei wrote:
> > > > > nit: this can be confusing, readNameFromTable can mislead people to think FContext is string (or vector of strings) type. the auto also isn't helping. Either spell out the type name, or rename `readNameFromTable` to something like `readSampleContextFromTable`
> > > > > 
> > > > > 
> > > > Fixed by using explicit type.
> > > Ok, but on 2nd thought, why do we call it readName while it's actually returning a context? Especially given that we've renamed addName to AddContext, writeNameIdx to writeContextIdx.
> > I think it'd be good to establish a convention as to when we call things `Name` vs `Context`. My thought is it goes with the type, what do you think? 
> Yeah, I've been using `Name` as an identifier of of the context and function name string, and using `Context` for CS profile and `String` for function names. But sometimes `Name` and `String` as mixed. 
> 
> We are using `SampleContext` for both CS and non-CS. And we are also using the word context specifically for CS. Sounds like we need a more general name (instead of `Name`) in the reader for both of them. How about `readIdFromTable`?
>  I've been using Name as an identifier of of the context and function name string, and using Context for CS profile and String for function names. 

The problem is you can't do that cleanly because there're cases that covers both CS and non-CS. readNameFromTable is one example.  

I think that name it according to the type makes things clearer. For non-CS profile we're still using SampleContext as key in the profile map anyways. Adding a new notion of `Id` in addition to name and context seem unnecessary (we have Idx already which can confuse people if we add Id). So I think readSampleContextFromTable/readContextFromTable is better. 


================
Comment at: llvm/lib/ProfileData/SampleProfReader.cpp:1117
 
-    SampleContext FContext(*FName);
-    bool ProfileInMap = Profiles.count(FContext);
+    bool ProfileInMap = Profiles.count(FContext.get());
 
----------------
hoy wrote:
> wenlei wrote:
> > hoy wrote:
> > > wenlei wrote:
> > > > nit: peal/hoist the `get()` so we don't have to call it for every use.
> > > You mean save `FContext.get()` in a temp and use it in the loop? Thought the compiler would do it.
> > Yeah, it's less about optimization but for readability to make the code less verbose. 
> Changed to using `*FContext` for readability, like other places. Look good?
> 
> Alternatively, we can have two variables such as `FContextError` and `FContext`. 
> 
> 
either works, thanks.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D108435/new/

https://reviews.llvm.org/D108435