[clr-interp] Implement cached virtual/interface dispatch

[BrzVlad]

We create a simple hashtable (InterpDispatchCache) that maps DispatchToken + target MT to target method to be called. The cache is similar to the DispatchCache used by VSD. It holds a single mapping per index, when a collision happens the entry will be replaced with the new one. Replaced entries are freed during GC. The expectation is that there will be few collisions given only a subset of methods are being interpreted.

This makes a microbenchmark from the suite 4x faster.

[dotnet-policy-service]

Tagging subscribers to this area: @BrzVlad, @janvorli, @kg
See info in area-owners.md if you want to be subscribed.

[Copilot]

Pull request overview

This pull request implements cached virtual and interface method dispatch for the CoreCLR interpreter to improve performance of virtual method calls. The implementation adds a simple hash table (InterpDispatchCache) that caches the mapping from (DispatchToken, MethodTable*) to the resolved MethodDesc*, similar to the DispatchCache used by Virtual Stub Dispatch (VSD).

Changes:

• Implements a new InterpDispatchCache structure that holds 4096 entries (12-bit cache) with collision replacement strategy
• Refactors DispatchToken::GetHash() to be shared between VSD and interpreter caches
• Integrates cache cleanup into GC sync points
• Adds dispatch token caching slots to INTOP_CALLVIRT and INTOP_CALLVIRT_TAIL instructions

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/coreclr/vm/interpexec.cpp	Adds InterpDispatchCache implementation with lookup/insert/reclaim logic, CreateDispatchTokenForMethod helper, and integrates cache into virtual call execution path
src/coreclr/vm/virtualcallstub.cpp	Refactors hash function by moving token hashing logic into DispatchToken::GetHash() method to enable sharing with interpreter cache
src/coreclr/vm/contractimpl.h	Adds GetHash() method declaration to DispatchToken struct for computing 12-bit cache hash
src/coreclr/vm/syncclean.cpp	Adds call to InterpDispatchCache_ReclaimAll() during GC sync point to clean up dead cache entries
src/coreclr/interpreter/inc/intops.def	Changes instruction size from 4 to 5 bytes for INTOP_CALLVIRT and INTOP_CALLVIRT_TAIL to accommodate dispatch token cache slot
src/coreclr/interpreter/compiler.h	Adds GetNewDataItemIndex() helper method to allocate non-shared data item slots for runtime caching
src/coreclr/interpreter/compiler.cpp	Updates EmitCall to allocate dispatch token cache slots for virtual call instructions

[BrzVlad]

@jkotas @davidwrighton Simplistic implementation from the discussion from the other day. Let me know if this approach seems ok.

[davidwrighton]

I want to look at this more on Monday, but this is about what I thought we should build here.

[davidwrighton]

Also this needs to handle collectibility correctly. There is a callback when an assembly is in the process of collecting where we clear the vsd cache. We should do the same here.

[BrzVlad]

@davidwrighton Added callback for when LoaderAllocator is destroyed.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 7 comments.

[BrzVlad]

@davidwrighton Tested 2 benchmarks in full interpreted mode (json and blazor):

default cache size 4k => ~40 collisions on json, 740 on blazor
cache size 8k => ~20k collisions on json, ~90 on blazor
cache size 16k => ~30 collisions on json, 850k on blazor
cache size 32k => 0 collisions on json, ~10 on blazor
cache size 64k => 0 collisions on json, ~10 on blazor

So it seems like heavily increasing the cache size reduces collisions but it is not so deterministic and incremental. I believe we should keep the default 4k until we have more widespread performance capability.

[davidwrighton]

Hmm, that looks like there might be real value to increasing the bucket size a bit or tweaking the hash function. Or maybe making the hash size a prime number. In any case, let's do that after merging the general concept in.

[Copilot]

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 3 comments.