Display LSRA stats in method summary (#53901)

* Lsrstats

* jit format

* use blocks iterator

src/coreclr/jit/codegencommon.cpp

@@ -2403,11 +2403,19 @@ void CodeGen::genEmitMachineCode()
     if (compiler->opts.disAsm || verbose)
     {
         printf("\n; Total bytes of code %d, prolog size %d, PerfScore %.2f, instruction count %d, allocated bytes for "
-               "code %d (MethodHash=%08x) for "
-               "method %s\n",
+               "code %d",
                codeSize, prologSize, compiler->info.compPerfScore, instrCount,
-               GetEmitter()->emitTotalHotCodeSize + GetEmitter()->emitTotalColdCodeSize,
-               compiler->info.compMethodHash(), compiler->info.compFullName);
+               GetEmitter()->emitTotalHotCodeSize + GetEmitter()->emitTotalColdCodeSize);
+
+#if TRACK_LSRA_STATS
+        if (JitConfig.DisplayLsraStats() == 3)
+        {
+            compiler->m_pLinearScan->dumpLsraStatsSummary(jitstdout);
+        }
+#endif // TRACK_LSRA_STATS
+
+        printf(" (MethodHash=%08x) for method %s\n", compiler->info.compMethodHash(), compiler->info.compFullName);
+
         printf("; ============================================================\n\n");
         printf(""); // in our logic this causes a flush
     }

src/coreclr/jit/compiler.h

@@ -1131,7 +1131,8 @@ public:
     virtual void recordVarLocationsAtStartOfBB(BasicBlock* bb) = 0;
     virtual bool willEnregisterLocalVars() const               = 0;
 #if TRACK_LSRA_STATS
-    virtual void dumpLsraStatsCsv(FILE* file) = 0;
+    virtual void dumpLsraStatsCsv(FILE* file)     = 0;
+    virtual void dumpLsraStatsSummary(FILE* file) = 0;
 #endif // TRACK_LSRA_STATS
 };
 

src/coreclr/jit/jitconfigvalues.h

@@ -26,10 +26,12 @@ CONFIG_INTEGER(DiffableDasm, W("JitDiffableDasm"), 0)          // Make the disas
 CONFIG_INTEGER(JitDasmWithAddress, W("JitDasmWithAddress"), 0) // Print the process address next to each instruction of
                                                                // the disassembly
 CONFIG_INTEGER(DisplayLoopHoistStats, W("JitLoopHoistStats"), 0) // Display JIT loop hoisting statistics
-CONFIG_INTEGER(DisplayLsraStats, W("JitLsraStats"), 0) // Display JIT Linear Scan Register Allocator statistics
-                                                       // if set to 1. If set to "2", display the stats in csv format.
-                                                       // Recommended to use with JitStdOutFile flag.
-CONFIG_STRING(JitLsraOrdering, W("JitLsraOrdering"))   // LSRA heuristics ordering
+CONFIG_INTEGER(DisplayLsraStats, W("JitLsraStats"), 0)       // Display JIT Linear Scan Register Allocator statistics
+                                                             // If set to "1", display the stats in textual format.
+                                                             // If set to "2", display the stats in csv format.
+                                                             // If set to "3", display the stats in summarize format.
+                                                             // Recommended to use with JitStdOutFile flag.
+CONFIG_STRING(JitLsraOrdering, W("JitLsraOrdering"))         // LSRA heuristics ordering
 CONFIG_INTEGER(DumpJittedMethods, W("DumpJittedMethods"), 0) // Prints all jitted methods to the console
 CONFIG_INTEGER(EnablePCRelAddr, W("JitEnablePCRelAddr"), 1)  // Whether absolute addr be encoded as PC-rel offset by
                                                              // RyuJIT where possible

src/coreclr/jit/lsra.cpp

@@ -8621,7 +8621,8 @@ void LinearScan::dumpLsraStats(FILE* file)
 
     fprintf(file, "----------\n");
 #ifdef DEBUG
-    fprintf(file, "Register selection order: %S\n", JitConfig.JitLsraOrdering());
+    fprintf(file, "Register selection order: %S\n",
+            JitConfig.JitLsraOrdering() == nullptr ? W("ABCDEFGHIJKLMNOPQ") : JitConfig.JitLsraOrdering());
 #endif
     fprintf(file, "Total Tracked Vars:  %d\n", compiler->lvaTrackedCount);
     fprintf(file, "Total Reg Cand Vars: %d\n", regCandidateVarCount);
@@ -8640,6 +8641,7 @@ void LinearScan::dumpLsraStats(FILE* file)
     bool addedBlockHeader = false;
     bool anyNonZeroStat   = false;
 
+    // Iterate for block 0
     for (int statIndex = 0; statIndex < LsraStat::COUNT; statIndex++)
     {
         unsigned lsraStat = blockInfo[0].stats[statIndex];
@@ -8667,6 +8669,7 @@ void LinearScan::dumpLsraStats(FILE* file)
         fprintf(file, "\n");
     }
 
+    // Iterate for remaining blocks
     for (BasicBlock* const block : compiler->Blocks())
     {
         if (block->bbNum > bbNumMaxBeforeResolution)
@@ -8712,7 +8715,6 @@ void LinearScan::dumpLsraStats(FILE* file)
         {
             fprintf(file, "..........\n");
         }
-        // TODO-review: I don't see a point of displaying Stats (SpillCount, etc.) if they are zero. Thoughts?
         if ((regSelectI < firstRegSelStat) || (sumStats[regSelectI] != 0))
         {
             // Print register selection stats
@@ -8780,6 +8782,48 @@ void LinearScan::dumpLsraStatsCsv(FILE* file)
     }
     fprintf(file, ",%.2f\n", compiler->info.compPerfScore);
 }
+
+// -----------------------------------------------------------
+// dumpLsraStatsSummary - dumps Lsra stats summary to given file
+//
+// Arguments:
+//    file    -  file to which stats are to be written.
+//
+void LinearScan::dumpLsraStatsSummary(FILE* file)
+{
+    unsigned             sumStats[LsraStat::STAT_FREE] = {0};
+    BasicBlock::weight_t wtdStats[LsraStat::STAT_FREE] = {0.0};
+
+    // Iterate for block 0
+    for (int statIndex = 0; statIndex < LsraStat::STAT_FREE; statIndex++)
+    {
+        unsigned lsraStat = blockInfo[0].stats[statIndex];
+        sumStats[statIndex] += lsraStat;
+        wtdStats[statIndex] += (lsraStat * blockInfo[0].weight);
+    }
+
+    // Iterate for remaining blocks
+    for (BasicBlock* const block : compiler->Blocks())
+    {
+        if (block->bbNum > bbNumMaxBeforeResolution)
+        {
+            continue;
+        }
+
+        for (int statIndex = 0; statIndex < LsraStat::STAT_FREE; statIndex++)
+        {
+            unsigned lsraStat = blockInfo[block->bbNum].stats[statIndex];
+            sumStats[statIndex] += lsraStat;
+            wtdStats[statIndex] += (lsraStat * block->bbWeight);
+        }
+    }
+
+    for (int regSelectI = 0; regSelectI < LsraStat::STAT_FREE; regSelectI++)
+    {
+        fprintf(file, ", %s %u %sWt %f", getStatName(regSelectI), sumStats[regSelectI], getStatName(regSelectI),
+                wtdStats[regSelectI]);
+    }
+}
 #endif // TRACK_LSRA_STATS
 
 #ifdef DEBUG

src/coreclr/jit/lsra.h

@@ -1455,6 +1455,7 @@ private:
 
 public:
     virtual void dumpLsraStatsCsv(FILE* file);
+    virtual void dumpLsraStatsSummary(FILE* file);
     static const char* getStatName(unsigned stat);
 
 #define INTRACK_STATS(x) x