[TieredCompilation] Virtual calls may be slower with tiering

[kouvel]

With tiering, there is an extra precode that calls would have to go through in some cases. If the processor does not compensate for the extra indirection, virtual calls are slower (normal calls are too, but that overhead is more easily compensated for with inlining).

See https://github.com/dotnet/coreclr/issues/18361 for an example where virtual call overhead is significant in the measurement.

Considerations:

• For virtual calls, I had prototyped using the virtual slot with an indirect call instead of the precode, and it eliminated the regression in the test case above. It looks like this would also close most of the steady-state perf gap in JitBench benchmarks between tiering, and full-opt JIT alone. The idea seems promising, it should be completed.
• For normal calls, I had prototyped using an indirect call instead of jumping over a trampoline. There were marginal improvements during startup, when with tier 0 JIT there would be more calls than at steady-state, but nothing significant since R2R'ed code also does some inlining. Without R2R it may be more interesting.

[kouvel]

A PerfView profile with the -ShowUnknownAddresses switch shows unnamed functions in callees:

Name                                                                                              	Inc %	     Inc	Exc %	   Exc
 fstest!Program+PerformanceTests+trivialTest@67-1.Invoke(class Microsoft.FSharp.Core.FSharpFunc`2)	 95.1	   6,390	 14.6	   983
+ fstest!Program+PerformanceTests+trivialTest@66.Invoke(int32)                                    	 21.6	   1,454	 21.6	 1,452
+ fstest!Program+PerformanceTests+trivialTest@67-2.Invoke(int64)                                  	 14.4	     971	 14.4	   970
+ fstest!Program+PerformanceTests+trivialTest@68-4.Invoke(int64)                                  	 14.1	     949	 14.1	   948
+ fstest!Program+PerformanceTests.arg@1(int32,class Microsoft.FSharp.Core.FSharpFunc`2)           	 10.4	     702	  0.0	     0
+ fstest!Program+PerformanceTests+trivialTest@69-5.Invoke(int64)                                  	  5.6	     378	  5.6	   377
+ ?!0x7ffd8f2f7508                                                                                	  5.2	     352	  5.2	   352
+ ?!0x7ffd8f2f74f0                                                                                	  3.6	     239	  3.6	   239
+ ?!0x7ffd8f2f6638                                                                                	  2.6	     172	  2.5	   171
+ ?!0x7ffd8f2f6650                                                                                	  2.3	     155	  2.3	   155

The unnamed functions are the extra precodes, they don't show up in the caller stacks of virtual functions because the precode jumps to the target. Names should be added for the precodes, this is covered by https://github.com/dotnet/coreclr/issues/18428.

[kouvel]

Fixed by dotnet/coreclr#21292