JIT: Add a uniform mechanism to track metadata/metrics about the comp…

…ilation (#98653)

Adds a JIT-EE API to report metadata back to the EE side about the JIT
compilation, and adds support for saving these metrics as part of SPMI runs.
Switches a number of adhoc metrics to use this scheme, and also adds a few new
ones.

The metadata is currently only reported with checked JITs.

Also adds support for fast perfscore diffs, and adds perfscore into the reports
generated by `superpmi.py asmdiffs`.

Fix #52877

src/coreclr/inc/corinfo.h

@@ -2916,6 +2916,13 @@ class ICorStaticInfo
             uint32_t                          numMappings         // [IN] Number of rich mappings
             ) = 0;
 
+    // Report back some metadata about the compilation to the EE -- for
+    // example, metrics about the compilation.
+    virtual void reportMetadata(
+        const char* key,
+        const void* value,
+        size_t length) = 0;
+
     /*-------------------------- Misc ---------------------------------------*/
 
     // Used to allocate memory that needs to handed to the EE.

src/coreclr/inc/icorjitinfoimpl_generated.h

@@ -438,6 +438,11 @@ void reportRichMappings(
           ICorDebugInfo::RichOffsetMapping* mappings,
           uint32_t numMappings) override;
 
+void reportMetadata(
+          const char* key,
+          const void* value,
+          size_t length) override;
+
 void* allocateArray(
           size_t cBytes) override;
 

src/coreclr/inc/jiteeversionguid.h

@@ -43,11 +43,11 @@ typedef const GUID *LPCGUID;
 #define GUID_DEFINED
 #endif // !GUID_DEFINED
 
-constexpr GUID JITEEVersionIdentifier = { /* 0fb71692-0ee6-4914-88a8-6446e45f23e8 */
-    0x0fb71692,
-    0x0ee6,
-    0x4914,
-    {0x88, 0xa8, 0x64, 0x46, 0xe4, 0x5f, 0x23, 0xe8}
+constexpr GUID JITEEVersionIdentifier = { /* 1f30d12b-38f1-4f1e-a08a-831def882aa4 */
+    0x1f30d12b,
+    0x38f1,
+    0x4f1e,
+    {0xa0, 0x8a, 0x83, 0x1d, 0xef, 0x88, 0x2a, 0xa4}
   };
 
 //////////////////////////////////////////////////////////////////////////////////////////////////////////

src/coreclr/jit/CMakeLists.txt

@@ -137,6 +137,7 @@ set( JIT_SOURCES
   jitconfig.cpp
   jiteh.cpp
   jithashtable.cpp
+  jitmetadata.cpp
   lclmorph.cpp
   lclvars.cpp
   likelyclass.cpp
@@ -334,6 +335,8 @@ set( JIT_HEADERS
   jitexpandarray.h
   jitgcinfo.h
   jithashtable.h
+  jitmetadata.h
+  jitmetadatalist.h
   jitpch.h
   jitstd.h
   lir.h

src/coreclr/jit/ICorJitInfo_names_generated.h

@@ -108,6 +108,7 @@ DEF_CLR_API(setBoundaries)
 DEF_CLR_API(getVars)
 DEF_CLR_API(setVars)
 DEF_CLR_API(reportRichMappings)
+DEF_CLR_API(reportMetadata)
 DEF_CLR_API(allocateArray)
 DEF_CLR_API(freeArray)
 DEF_CLR_API(getArgNext)

src/coreclr/jit/ICorJitInfo_wrapper_generated.hpp

@@ -1028,6 +1028,16 @@ void WrapICorJitInfo::reportRichMappings(
     API_LEAVE(reportRichMappings);
 }
 
+void WrapICorJitInfo::reportMetadata(
+          const char* key,
+          const void* value,
+          size_t length)
+{
+    API_ENTER(reportMetadata);
+    wrapHnd->reportMetadata(key, value, length);
+    API_LEAVE(reportMetadata);
+}
+
 void* WrapICorJitInfo::allocateArray(
           size_t cBytes)
 {

src/coreclr/jit/codegencommon.cpp

@@ -315,10 +315,8 @@ void CodeGen::genPrepForCompiler()
         }
     }
     VarSetOps::AssignNoCopy(compiler, genLastLiveSet, VarSetOps::MakeEmpty(compiler));
-    genLastLiveMask = RBM_NONE;
-#ifdef DEBUG
-    compiler->fgBBcountAtCodegen = compiler->fgBBcount;
-#endif
+    genLastLiveMask                        = RBM_NONE;
+    compiler->Metrics.BasicBlocksAtCodegen = compiler->fgBBcount;
 }
 
 //------------------------------------------------------------------------
@@ -2042,7 +2040,7 @@ void CodeGen::genEmitMachineCode()
 
         printf("; Total bytes of code %d, prolog size %d, PerfScore %.2f, instruction count %d, allocated bytes for "
                "code %d",
-               codeSize, prologSize, compiler->info.compPerfScore, instrCount,
+               codeSize, prologSize, compiler->Metrics.PerfScore, instrCount,
                GetEmitter()->emitTotalHotCodeSize + GetEmitter()->emitTotalColdCodeSize);
 
         if (dspMetrics)

src/coreclr/jit/compiler.cpp

@@ -1793,9 +1793,13 @@ void Compiler::compInit(ArenaAllocator*       pAlloc,
     info.compMethodName = eeGetMethodName(methodHnd);
     info.compClassName  = eeGetClassName(info.compClassHnd);
     info.compFullName   = eeGetMethodFullName(methodHnd);
-    info.compPerfScore  = 0.0;
 
     info.compMethodSuperPMIIndex = g_jitHost->getIntConfigValue(W("SuperPMIMethodContextNumber"), -1);
+
+    if (!compIsForInlining())
+    {
+        JitMetadata::report(this, JitMetadata::MethodFullName, info.compFullName, strlen(info.compFullName));
+    }
 #endif // defined(DEBUG) || defined(LATE_DISASM) || DUMP_FLOWGRAPHS
 
 #if defined(DEBUG)
@@ -1863,9 +1867,7 @@ void Compiler::compInit(ArenaAllocator*       pAlloc,
         //
         // Initialize all the per-method statistics gathering data structures.
         //
-
-        optLoopsCloned = 0;
-
+        CLANG_FORMAT_COMMENT_ANCHOR;
 #if LOOP_HOIST_STATS
         m_loopsConsidered             = 0;
         m_curLoopHasHoistedExpression = false;
@@ -1965,6 +1967,8 @@ void Compiler::compInit(ArenaAllocator*       pAlloc,
     compUsesThrowHelper = false;
 
     m_preferredInitCctor = CORINFO_HELP_UNDEF;
+
+    new (&Metrics, jitstd::placement_t()) JitMetrics();
 }
 
 /*****************************************************************************
@@ -5215,7 +5219,7 @@ void Compiler::compCompile(void** methodCodePtr, uint32_t* methodCodeSize, JitFl
 #ifdef DEBUG
         if (JitConfig.JitMetrics() > 0)
         {
-            sprintf_s(metricPart, 128, ", perfScore=%.2f, numCse=%u", info.compPerfScore, optCSEcount);
+            sprintf_s(metricPart, 128, ", perfScore=%.2f, numCse=%u", Metrics.PerfScore, optCSEcount);
         }
 #endif
 
@@ -5419,7 +5423,7 @@ PhaseStatus Compiler::placeLoopAlignInstructions()
         {
             block->SetFlags(BBF_LOOP_ALIGN);
             BitVecOps::AddElemD(&loopTraits, alignedLoops, loop->GetIndex());
-            INDEBUG(loopAlignCandidates++);
+            Metrics.LoopAlignmentCandidates++;
 
             BasicBlock* prev = block->Prev();
             // shouldAlignLoop should have guaranteed these properties.
@@ -5468,7 +5472,7 @@ PhaseStatus Compiler::placeLoopAlignInstructions()
         }
     }
 
-    JITDUMP("Found %u candidates for loop alignment\n", loopAlignCandidates);
+    JITDUMP("Found %d candidates for loop alignment\n", Metrics.LoopAlignmentCandidates);
 
     return madeChanges ? PhaseStatus::MODIFIED_EVERYTHING : PhaseStatus::MODIFIED_NOTHING;
 }
@@ -6441,6 +6445,8 @@ void Compiler::compCompileFinish()
         compArenaAllocator->finishMemStats();
         memAllocHist.record((unsigned)((compArenaAllocator->getTotalBytesAllocated() + 1023) / 1024));
         memUsedHist.record((unsigned)((compArenaAllocator->getTotalBytesUsed() + 1023) / 1024));
+
+        Metrics.BytesAllocated = (int64_t)compArenaAllocator->getTotalBytesUsed();
     }
 
 #ifdef DEBUG
@@ -6624,7 +6630,7 @@ void Compiler::compCompileFinish()
 
         printf(" %3d |", optCallCount);
         printf(" %3d |", optIndirectCallCount);
-        printf(" %3d |", fgBBcountAtCodegen);
+        printf(" %3d |", Metrics.BasicBlocksAtCodegen);
         printf(" %3d |", lvaCount);
 
         if (opts.MinOpts())
@@ -6637,13 +6643,13 @@ void Compiler::compCompileFinish()
             printf(" %3d |", optCSEcount);
         }
 
-        if (info.compPerfScore < 9999.995)
+        if (Metrics.PerfScore < 9999.995)
         {
-            printf(" %7.2f |", info.compPerfScore);
+            printf(" %7.2f |", Metrics.PerfScore);
         }
         else
         {
-            printf(" %7.0f |", info.compPerfScore);
+            printf(" %7.0f |", Metrics.PerfScore);
         }
 
         printf(" %4d |", info.compMethodInfo->ILCodeSize);
@@ -6654,9 +6660,13 @@ void Compiler::compCompileFinish()
         printf(""); // in our logic this causes a flush
     }
 
+    JITDUMP("Final metrics:\n");
+    Metrics.report(this);
+    DBEXEC(verbose, Metrics.dump());
+
     if (verbose)
     {
-        printf("****** DONE compiling %s\n", info.compFullName);
+        printf("\n****** DONE compiling %s\n", info.compFullName);
         printf(""); // in our logic this causes a flush
     }
 
@@ -7149,6 +7159,13 @@ int Compiler::compCompileHelper(CORINFO_MODULE_HANDLE classPtr,
         opts.disAsm = false;
     }
 
+#ifdef DEBUG
+    {
+        const char* tieringName = compGetTieringName(true);
+        JitMetadata::report(this, JitMetadata::TieringName, tieringName, strlen(tieringName));
+    }
+#endif
+
 #if COUNT_BASIC_BLOCKS
     bbCntTable.record(fgBBcount);
 
@@ -9077,12 +9094,12 @@ void JitTimer::PrintCsvMethodStats(Compiler* comp)
     fprintf(s_csvFile, "%u,", comp->info.compILCodeSize);
     fprintf(s_csvFile, "%u,", comp->fgBBcount);
     fprintf(s_csvFile, "%u,", comp->opts.MinOpts());
-    fprintf(s_csvFile, "%u,", comp->optNumNaturalLoopsFound);
-    fprintf(s_csvFile, "%u,", comp->optLoopsCloned);
+    fprintf(s_csvFile, "%d,", comp->Metrics.LoopsFoundDuringOpts);
+    fprintf(s_csvFile, "%d,", comp->Metrics.LoopsCloned);
 #if FEATURE_LOOP_ALIGN
 #ifdef DEBUG
-    fprintf(s_csvFile, "%u,", comp->loopAlignCandidates);
-    fprintf(s_csvFile, "%u,", comp->loopsAligned);
+    fprintf(s_csvFile, "%d,", comp->Metrics.LoopAlignmentCandidates);
+    fprintf(s_csvFile, "%d,", comp->Metrics.LoopsAligned);
 #endif // DEBUG
 #endif // FEATURE_LOOP_ALIGN
     unsigned __int64 totCycles = 0;

src/coreclr/jit/compiler.h

@@ -65,6 +65,8 @@ inline var_types genActualType(T value);
 #include "simd.h"
 #include "simdashwintrinsic.h"
 
+#include "jitmetadata.h"
+
 /*****************************************************************************
  *                  Forward declarations
  */
@@ -4968,7 +4970,6 @@ class Compiler
     unsigned        fgEdgeCount;    // # of control flow edges between the BBs
     unsigned        fgBBcount;      // # of BBs in the method (in the linked list that starts with fgFirstBB)
 #ifdef DEBUG
-    unsigned                     fgBBcountAtCodegen; // # of BBs in the method at the start of codegen
     jitstd::vector<BasicBlock*>* fgBBOrder;          // ordered vector of BBs
 #endif
     // Used as a quick check for whether loop alignment should look for natural loops.
@@ -4997,7 +4998,6 @@ class Compiler
     // 2. All loop exits where bbIsHandlerBeg(exit) is false have only loop predecessors.
     //
     bool optLoopsCanonical;
-    unsigned optNumNaturalLoopsFound; // Number of natural loops found in the loop finding phase
 
     bool fgBBVarSetsInited;
 
@@ -6841,16 +6841,11 @@ class Compiler
 
 public:
     bool fgHasLoops;
-#ifdef DEBUG
-    unsigned loopAlignCandidates; // number of candidates identified by placeLoopAlignInstructions
-    unsigned loopsAligned;        // number of loops actually aligned
-#endif                          // DEBUG
 
 protected:
     unsigned optCallCount;         // number of calls made in the method
     unsigned optIndirectCallCount; // number of virtual, interface and indirect calls made in the method
     unsigned optNativeCallCount;   // number of Pinvoke/Native calls made in the method
-    unsigned optLoopsCloned;       // number of loops cloned in the current method.
 
 #ifdef DEBUG
     void optCheckPreds();
@@ -10155,7 +10150,6 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         const char* compMethodName;
         const char* compClassName;
         const char* compFullName;
-        double      compPerfScore;
         int         compMethodSuperPMIIndex; // useful when debugging under SuperPMI
 
 #endif // defined(DEBUG) || defined(LATE_DISASM) || DUMP_FLOWGRAPHS
@@ -10616,6 +10610,8 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
     static EnregisterStats s_enregisterStats;
 #endif // TRACK_ENREG_STATS
 
+    JitMetrics Metrics;
+
     bool compIsForInlining() const;
     bool compDonotInline();
 

src/coreclr/jit/emit.cpp

@@ -4393,7 +4393,7 @@ size_t emitter::emitIssue1Instr(insGroup* ig, instrDesc* id, BYTE** dp)
     float insExeCost = insEvaluateExecutionCost(id);
     // All compPerfScore calculations must be performed using doubles
     double insPerfScore = (double)(ig->igWeight / (double)BB_UNITY_WEIGHT) * insExeCost;
-    emitComp->info.compPerfScore += insPerfScore;
+    emitComp->Metrics.PerfScore += insPerfScore;
     ig->igPerfScore += insPerfScore;
 #endif // defined(DEBUG) || defined(LATE_DISASM)
 

src/coreclr/jit/emitxarch.cpp

@@ -12475,10 +12475,10 @@ BYTE* emitter::emitOutputAlign(insGroup* ig, instrDesc* id, BYTE* dst)
             assert(paddingToAdd == paddingNeeded);
         }
     }
-
-    emitComp->loopsAligned++;
 #endif
 
+    emitComp->Metrics.LoopsAligned++;
+
 #ifdef DEBUG
     // Under STRESS_EMITTER, if this is the 'align' before the 'jmp' instruction,
     // then add "int3" instruction. Since int3 takes 1 byte, we would only add

src/coreclr/jit/fgbasic.cpp

@@ -49,8 +49,7 @@ void Compiler::fgInit()
     fgBBcount = 0;
 
 #ifdef DEBUG
-    fgBBcountAtCodegen = 0;
-    fgBBOrder          = nullptr;
+    fgBBOrder = nullptr;
 #endif // DEBUG
 
     fgMightHaveNaturalLoops = false;