Fix signature for 64-bit delegate profiling helper (#74500)

* Fix signature for 64-bit delegate profiling helper

I missed updating the signature of the 64-bit variant here when the
vtable and delegate profiling helpers were originally split up.

Fix #74295

* JIT: Randomly collect 64-bit counts

Add some testing.

* Disable collecting 64 bit counters on 32-bit

This needs some work so disable it for now.

Author: jakobbotsch

src/coreclr/jit/compiler.cpp

@@ -3208,6 +3208,21 @@ void Compiler::compInitOptions(JitFlags* jitFlags)
 #endif
     }
 
+#ifdef TARGET_64BIT
+    opts.compCollect64BitCounts = JitConfig.JitCollect64BitCounts() != 0;
+
+#ifdef DEBUG
+    if (JitConfig.JitRandomlyCollect64BitCounts() != 0)
+    {
+        CLRRandom rng;
+        rng.Init(info.compMethodHash() ^ JitConfig.JitRandomlyCollect64BitCounts() ^ 0x3485e20e);
+        opts.compCollect64BitCounts = rng.Next(2) == 0;
+    }
+#endif
+#else
+    opts.compCollect64BitCounts = false;
+#endif
+
 #ifdef DEBUG
 
     // Now, set compMaxUncheckedOffsetForNullObject for STRESS_NULL_OBJECT_CHECK

src/coreclr/jit/compiler.h

@@ -9335,6 +9335,9 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         // Use early multi-dimensional array operator expansion (expand after loop optimizations; before lowering).
         bool compJitEarlyExpandMDArrays;
 
+        // Collect 64 bit counts for PGO data.
+        bool compCollect64BitCounts;
+
     } opts;
 
     static bool                s_pAltJitExcludeAssembliesListInitialized;

src/coreclr/jit/fgprofile.cpp

@@ -570,7 +570,7 @@ void BlockCountInstrumentor::BuildSchemaElements(BasicBlock* block, Schema& sche
     ICorJitInfo::PgoInstrumentationSchema schemaElem;
     schemaElem.Count               = 1;
     schemaElem.Other               = 0;
-    schemaElem.InstrumentationKind = JitConfig.JitCollect64BitCounts()
+    schemaElem.InstrumentationKind = m_comp->opts.compCollect64BitCounts
                                          ? ICorJitInfo::PgoInstrumentationKind::BasicBlockLongCount
                                          : ICorJitInfo::PgoInstrumentationKind::BasicBlockIntCount;
     schemaElem.ILOffset = offset;
@@ -1314,7 +1314,7 @@ void EfficientEdgeCountInstrumentor::BuildSchemaElements(BasicBlock* block, Sche
         ICorJitInfo::PgoInstrumentationSchema schemaElem;
         schemaElem.Count               = 1;
         schemaElem.Other               = targetKey;
-        schemaElem.InstrumentationKind = JitConfig.JitCollect64BitCounts()
+        schemaElem.InstrumentationKind = m_comp->opts.compCollect64BitCounts
                                              ? ICorJitInfo::PgoInstrumentationKind::EdgeLongCount
                                              : ICorJitInfo::PgoInstrumentationKind::EdgeIntCount;
         schemaElem.ILOffset = sourceKey;
@@ -1503,7 +1503,7 @@ class BuildHandleHistogramProbeSchemaGen
             schemaElem.Other |= ICorJitInfo::HandleHistogram32::DELEGATE_FLAG;
         }
 
-        schemaElem.InstrumentationKind = JitConfig.JitCollect64BitCounts()
+        schemaElem.InstrumentationKind = compiler->opts.compCollect64BitCounts
                                              ? ICorJitInfo::PgoInstrumentationKind::HandleHistogramLongCount
                                              : ICorJitInfo::PgoInstrumentationKind::HandleHistogramIntCount;
         schemaElem.ILOffset = (int32_t)call->gtHandleHistogramProfileCandidateInfo->ilOffset;

src/coreclr/jit/jitconfigvalues.h

@@ -561,7 +561,9 @@ CONFIG_STRING(JitEnablePgoRange, W("JitEnablePgoRange"))         // Enable pgo d
 CONFIG_INTEGER(JitRandomEdgeCounts, W("JitRandomEdgeCounts"), 0) // Substitute random values for edge counts
 CONFIG_INTEGER(JitCrossCheckDevirtualizationAndPGO, W("JitCrossCheckDevirtualizationAndPGO"), 0)
 CONFIG_INTEGER(JitNoteFailedExactDevirtualization, W("JitNoteFailedExactDevirtualization"), 0)
-#endif // debug
+CONFIG_INTEGER(JitRandomlyCollect64BitCounts, W("JitRandomlyCollect64BitCounts"), 0) // Collect 64-bit counts randomly
+                                                                                     // for some methods.
+#endif                                                                               // debug
 
 // Devirtualize virtual calls with getExactClasses (NativeAOT only for now)
 CONFIG_INTEGER(JitEnableExactDevirtualization, W("JitEnableExactDevirtualization"), 1)

src/coreclr/vm/jithelpers.cpp

@@ -5571,7 +5571,7 @@ HCIMPL2(void, JIT_DelegateProfile32, Object *obj, ICorJitInfo::HandleHistogram32
 HCIMPLEND
 
 // Version of helper above used when the count is 64-bit
-HCIMPL3(void, JIT_DelegateProfile64, Object *obj, CORINFO_METHOD_HANDLE baseMethod, ICorJitInfo::HandleHistogram64* methodProfile)
+HCIMPL2(void, JIT_DelegateProfile64, Object *obj, ICorJitInfo::HandleHistogram64* methodProfile)
 {
     FCALL_CONTRACT;
     FC_GC_POLL_NOT_NEEDED();

src/tests/Common/testenvironment.proj

@@ -66,6 +66,7 @@
       COMPlus_JitRandomGuardedDevirtualization;
       COMPlus_JitRandomEdgeCounts;
       COMPlus_JitRandomOnStackReplacement;
+      COMPlus_JitRandomlyCollect64BitCounts;
       COMPlus_JitForceControlFlowGuard;
       COMPlus_JitCFGUseDispatcher;
       RunningIlasmRoundTrip
@@ -214,10 +215,10 @@
     <TestEnvironment Include="dynamicpgo" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" />
     <TestEnvironment Include="fullpgo" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0"/>
     <TestEnvironment Include="fullpgo_methodprofiling" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitDelegateProfiling="1" JitVTableProfiling="1" />
-    <TestEnvironment Include="fullpgo_random_gdv" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1"/>
-    <TestEnvironment Include="fullpgo_random_gdv_methodprofiling_only" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1" JitClassProfiling="0" JitDelegateProfiling="1" JitVTableProfiling="1" />
-    <TestEnvironment Include="fullpgo_random_edge" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomEdgeCounts="1"/>
-    <TestEnvironment Include="fullpgo_random_gdv_edge" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1" JitRandomEdgeCounts="1"/>
+    <TestEnvironment Include="fullpgo_random_gdv" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1" JitRandomlyCollect64BitCounts="1" />
+    <TestEnvironment Include="fullpgo_random_gdv_methodprofiling_only" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1" JitClassProfiling="0" JitDelegateProfiling="1" JitVTableProfiling="1" JitRandomlyCollect64BitCounts="1" />
+    <TestEnvironment Include="fullpgo_random_edge" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomEdgeCounts="1" JitRandomlyCollect64BitCounts="1" />
+    <TestEnvironment Include="fullpgo_random_gdv_edge" TieredPGO="1" TieredCompilation="1" TC_QuickJitForLoops="1" ReadyToRun="0" JitRandomGuardedDevirtualization="1" JitRandomEdgeCounts="1" JitRandomlyCollect64BitCounts="1" />
     <TestEnvironment Include="gcstandalone" Condition="'$(TargetsWindows)' == 'true'" GCName="clrgc.dll"/>
     <TestEnvironment Include="gcstandalone" Condition="'$(TargetsWindows)' != 'true'" GCName="libclrgc.so"/>
     <TestEnvironment Include="gcstandaloneserver" Condition="'$(TargetsWindows)' == 'true'" gcServer="1" GCName="clrgc.dll"/>