-
Notifications
You must be signed in to change notification settings - Fork 5.3k
Closed
Labels
area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMICLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI
Milestone
Description
runtime/src/coreclr/jit/codegencommon.cpp
Lines 2005 to 2011 in 64fe78e
| #if defined(DEBUG) || defined(LATE_DISASM) | |
| // Add code size information into the Perf Score | |
| // All compPerfScore calculations must be performed using doubles | |
| compiler->info.compPerfScore += ((double)compiler->info.compTotalHotCodeSize * (double)PERFSCORE_CODESIZE_COST_HOT); | |
| compiler->info.compPerfScore += | |
| ((double)compiler->info.compTotalColdCodeSize * (double)PERFSCORE_CODESIZE_COST_COLD); | |
| #endif // DEBUG || LATE_DISASM |
This was part of the original design dotnet/coreclr#26456 but I don't think it makes sense, either from a "units" standpoint (speed is not the same as size, and existence of a widely valid conversion ratio is not obvious) or generally.
And I think it may also be complicating attempts to leverage perf score for learning ,eg #92915 (comment)
We have a separate size metric, so in cases where we care about both, we can just look at that as well.
Metadata
Metadata
Assignees
Labels
area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMICLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI