JIT: enable instrumentation for inlinees #119658
Conversation
If we are doing an optimized+instrumented jit pass (like we do for R2R methods) allow the inlinees to be instrumented. This gives us a chance to collect profile data for methods that are always inlined. All inlinees currently share the same profile data segment with each other and with a root compilation of the method (if any). So this profile is "context-free". If there is a schema mismatch (say there is a stale pre-existing R2R schema) then subsequent instrumentation will fail. This is something we need to keep an eye on. Right now we can't distinguish this failure from other kinds of schema allocation failures. Closes dotnet#44372
|
This should address the long-standing regression in #91938. Local runs show it possibly helps but doesn't completely fix the regression. Still need to verify it is actually getting instrumentation data for the key inlinees. Generally speaking, this should help reduce the gap between running with R2R enabled (where we can lose profile data, or optimize based on stale data) and R2R disabled. |
|
I have a bunch more fixes on top of this, will PR them soon. One issue I'm seeing is IL mismatches. Some of it I've tracked down to IL differences between the PGO schema embedded into corelib and the current corelib IL; that's somewhat understandable as the PGO data likely goes stale, and should be (generally) a non-product issue. Would still be good to flag this case and allow the dynamic PGO to take over. I also suspect inlinee schemas may diverge from root schemas even for the same IL, but need to do more work to track this down. |
|
Still some failures to sort out. Also my local perf runs aren't as promising as I'd hoped. |
If we are doing an optimized+instrumented jit pass (like we do for R2R methods) allow the inlinees to be instrumented. This gives us a chance to collect profile data for methods that are always inlined.
All inlinees currently share the same profile data segment with each other and with a root compilation of the method (if any). So this profile is "context-free".
If there is a schema mismatch (say there is a stale pre-existing R2R schema) then subsequent instrumentation will fail. This is something we need to keep an eye on. Right now we can't distinguish this failure from other kinds of schema allocation failures.
Closes #44372