hotspot/src/share/vm/opto/phase.cpp
changeset 26913 9ad70cd32368
parent 22872 b6902ee5bc8d
child 28648 102bdbb42723
--- a/hotspot/src/share/vm/opto/phase.cpp	Mon Sep 29 08:40:51 2014 +0200
+++ b/hotspot/src/share/vm/opto/phase.cpp	Thu Sep 25 12:10:57 2014 +0400
@@ -30,56 +30,14 @@
 #include "opto/node.hpp"
 #include "opto/phase.hpp"
 
-#ifndef PRODUCT
 int Phase::_total_bytes_compiled = 0;
 
 elapsedTimer Phase::_t_totalCompilation;
 elapsedTimer Phase::_t_methodCompilation;
 elapsedTimer Phase::_t_stubCompilation;
-#endif
 
-// The next timers used for LogCompilation
-elapsedTimer Phase::_t_parser;
-elapsedTimer Phase::_t_optimizer;
-elapsedTimer   Phase::_t_escapeAnalysis;
-elapsedTimer     Phase::_t_connectionGraph;
-elapsedTimer   Phase::_t_idealLoop;
-elapsedTimer   Phase::_t_ccp;
-elapsedTimer Phase::_t_matcher;
-elapsedTimer Phase::_t_registerAllocation;
-elapsedTimer Phase::_t_output;
-
-#ifndef PRODUCT
-elapsedTimer Phase::_t_graphReshaping;
-elapsedTimer Phase::_t_scheduler;
-elapsedTimer Phase::_t_blockOrdering;
-elapsedTimer Phase::_t_macroEliminate;
-elapsedTimer Phase::_t_macroExpand;
-elapsedTimer Phase::_t_peephole;
-elapsedTimer Phase::_t_postalloc_expand;
-elapsedTimer Phase::_t_codeGeneration;
-elapsedTimer Phase::_t_registerMethod;
-elapsedTimer Phase::_t_temporaryTimer1;
-elapsedTimer Phase::_t_temporaryTimer2;
-elapsedTimer Phase::_t_idealLoopVerify;
-
-// Subtimers for _t_optimizer
-elapsedTimer   Phase::_t_iterGVN;
-elapsedTimer   Phase::_t_iterGVN2;
-elapsedTimer   Phase::_t_incrInline;
-
-// Subtimers for _t_registerAllocation
-elapsedTimer   Phase::_t_ctorChaitin;
-elapsedTimer   Phase::_t_buildIFGphysical;
-elapsedTimer   Phase::_t_computeLive;
-elapsedTimer   Phase::_t_regAllocSplit;
-elapsedTimer   Phase::_t_postAllocCopyRemoval;
-elapsedTimer   Phase::_t_fixupSpills;
-
-// Subtimers for _t_output
-elapsedTimer   Phase::_t_instrSched;
-elapsedTimer   Phase::_t_buildOopMaps;
-#endif
+// The counters to use for LogCompilation
+elapsedTimer Phase::timers[max_phase_timers];
 
 //------------------------------Phase------------------------------------------
 Phase::Phase( PhaseNumber pnum ) : _pnum(pnum), C( pnum == Compiler ? NULL : Compile::current()) {
@@ -89,93 +47,132 @@
   CompileBroker::maybe_block();
 }
 
-#ifndef PRODUCT
-static const double minimum_reported_time             = 0.0001; // seconds
-static const double expected_method_compile_coverage  = 0.97;   // %
-static const double minimum_meaningful_method_compile = 2.00;   // seconds
+void Phase::print_timers() {
+  tty->print_cr ("    C2 Compile Time:      %7.3f s", Phase::_t_totalCompilation.seconds());
+  tty->print_cr ("       Parse:               %7.3f s", timers[_t_parser].seconds());
 
-void Phase::print_timers() {
-  tty->print_cr ("Accumulated compiler times:");
-  tty->print_cr ("---------------------------");
-  tty->print_cr ("  Total compilation: %3.3f sec.", Phase::_t_totalCompilation.seconds());
-  tty->print    ("    method compilation   : %3.3f sec", Phase::_t_methodCompilation.seconds());
-  tty->print    ("/%d bytes",_total_bytes_compiled);
-  tty->print_cr (" (%3.0f bytes per sec) ", Phase::_total_bytes_compiled / Phase::_t_methodCompilation.seconds());
-  tty->print_cr ("    stub compilation     : %3.3f sec.", Phase::_t_stubCompilation.seconds());
-  tty->print_cr ("  Phases:");
-  tty->print_cr ("    parse          : %3.3f sec", Phase::_t_parser.seconds());
-  tty->print_cr ("    optimizer      : %3.3f sec", Phase::_t_optimizer.seconds());
-  if( Verbose || WizardMode ) {
+  {
+    tty->print_cr ("       Optimize:            %7.3f s", timers[_t_optimizer].seconds());
     if (DoEscapeAnalysis) {
       // EA is part of Optimizer.
-      tty->print_cr ("      escape analysis: %3.3f sec", Phase::_t_escapeAnalysis.seconds());
-      tty->print_cr ("        connection graph: %3.3f sec", Phase::_t_connectionGraph.seconds());
-      tty->print_cr ("      macroEliminate : %3.3f sec", Phase::_t_macroEliminate.seconds());
+      tty->print_cr ("         Escape Analysis:     %7.3f s", timers[_t_escapeAnalysis].seconds());
+      tty->print_cr ("           Conn Graph:          %7.3f s", timers[_t_connectionGraph].seconds());
+      tty->print_cr ("           Macro Eliminate:     %7.3f s", timers[_t_macroEliminate].seconds());
     }
-    tty->print_cr ("      iterGVN        : %3.3f sec", Phase::_t_iterGVN.seconds());
-    tty->print_cr ("      incrInline     : %3.3f sec", Phase::_t_incrInline.seconds());
-    tty->print_cr ("      idealLoop      : %3.3f sec", Phase::_t_idealLoop.seconds());
-    tty->print_cr ("      idealLoopVerify: %3.3f sec", Phase::_t_idealLoopVerify.seconds());
-    tty->print_cr ("      ccp            : %3.3f sec", Phase::_t_ccp.seconds());
-    tty->print_cr ("      iterGVN2       : %3.3f sec", Phase::_t_iterGVN2.seconds());
-    tty->print_cr ("      macroExpand    : %3.3f sec", Phase::_t_macroExpand.seconds());
-    tty->print_cr ("      graphReshape   : %3.3f sec", Phase::_t_graphReshaping.seconds());
-    double optimizer_subtotal = Phase::_t_iterGVN.seconds() + Phase::_t_iterGVN2.seconds() +
-      Phase::_t_escapeAnalysis.seconds() + Phase::_t_macroEliminate.seconds() +
-      Phase::_t_idealLoop.seconds() + Phase::_t_ccp.seconds() +
-      Phase::_t_macroExpand.seconds() + Phase::_t_graphReshaping.seconds();
-    double percent_of_optimizer = ((optimizer_subtotal == 0.0) ? 0.0 : (optimizer_subtotal / Phase::_t_optimizer.seconds() * 100.0));
-    tty->print_cr ("      subtotal       : %3.3f sec,  %3.2f %%", optimizer_subtotal, percent_of_optimizer);
-  }
-  tty->print_cr ("    matcher        : %3.3f sec", Phase::_t_matcher.seconds());
-  tty->print_cr ("    scheduler      : %3.3f sec", Phase::_t_scheduler.seconds());
-  tty->print_cr ("    regalloc       : %3.3f sec", Phase::_t_registerAllocation.seconds());
-  if( Verbose || WizardMode ) {
-    tty->print_cr ("      ctorChaitin    : %3.3f sec", Phase::_t_ctorChaitin.seconds());
-    tty->print_cr ("      buildIFG       : %3.3f sec", Phase::_t_buildIFGphysical.seconds());
-    tty->print_cr ("      computeLive    : %3.3f sec", Phase::_t_computeLive.seconds());
-    tty->print_cr ("      regAllocSplit  : %3.3f sec", Phase::_t_regAllocSplit.seconds());
-    tty->print_cr ("      postAllocCopyRemoval: %3.3f sec", Phase::_t_postAllocCopyRemoval.seconds());
-    tty->print_cr ("      fixupSpills    : %3.3f sec", Phase::_t_fixupSpills.seconds());
-    double regalloc_subtotal = Phase::_t_ctorChaitin.seconds() +
-      Phase::_t_buildIFGphysical.seconds() + Phase::_t_computeLive.seconds() +
-      Phase::_t_regAllocSplit.seconds()    + Phase::_t_fixupSpills.seconds() +
-      Phase::_t_postAllocCopyRemoval.seconds();
-    double percent_of_regalloc = ((regalloc_subtotal == 0.0) ? 0.0 : (regalloc_subtotal / Phase::_t_registerAllocation.seconds() * 100.0));
-    tty->print_cr ("      subtotal       : %3.3f sec,  %3.2f %%", regalloc_subtotal, percent_of_regalloc);
+    tty->print_cr ("         GVN 1:               %7.3f s", timers[_t_iterGVN].seconds());
+
+    {
+       tty->print_cr ("         Incremental Inline:  %7.3f s", timers[_t_incrInline].seconds());
+       tty->print_cr ("           IdealLoop:           %7.3f s", timers[_t_incrInline_ideal].seconds());
+       tty->print_cr ("           IGVN:                %7.3f s", timers[_t_incrInline_igvn].seconds());
+       tty->print_cr ("           Inline:              %7.3f s", timers[_t_incrInline_inline].seconds());
+       tty->print_cr ("           Prune Useless:       %7.3f s", timers[_t_incrInline_pru].seconds());
+
+       double other = timers[_t_incrInline].seconds() -
+        (timers[_t_incrInline_ideal].seconds() +
+         timers[_t_incrInline_igvn].seconds() +
+         timers[_t_incrInline_inline].seconds() +
+         timers[_t_incrInline_pru].seconds());
+       if (other > 0) {
+         tty->print_cr("           Other:               %7.3f s", other);
+       }
+    }
+    tty->print_cr ("         IdealLoop:           %7.3f s", timers[_t_idealLoop].seconds());
+    tty->print_cr ("         IdealLoop Verify:    %7.3f s", timers[_t_idealLoopVerify].seconds());
+    tty->print_cr ("         Cond Const Prop:     %7.3f s", timers[_t_ccp].seconds());
+    tty->print_cr ("         GVN 2:               %7.3f s", timers[_t_iterGVN2].seconds());
+    tty->print_cr ("         Macro Expand:        %7.3f s", timers[_t_macroExpand].seconds());
+    tty->print_cr ("         Graph Reshape:       %7.3f s", timers[_t_graphReshaping].seconds());
+
+    double other = timers[_t_optimizer].seconds() -
+      (timers[_t_escapeAnalysis].seconds() +
+       timers[_t_iterGVN].seconds() +
+       timers[_t_incrInline].seconds() +
+       timers[_t_idealLoop].seconds() +
+       timers[_t_idealLoopVerify].seconds() +
+       timers[_t_ccp].seconds() +
+       timers[_t_iterGVN2].seconds() +
+       timers[_t_macroExpand].seconds() +
+       timers[_t_graphReshaping].seconds());
+    if (other > 0) {
+      tty->print_cr("         Other:               %7.3f s", other);
+    }
   }
-  tty->print_cr ("    blockOrdering  : %3.3f sec", Phase::_t_blockOrdering.seconds());
-  tty->print_cr ("    peephole       : %3.3f sec", Phase::_t_peephole.seconds());
-  if (Matcher::require_postalloc_expand) {
-    tty->print_cr ("    postalloc_expand: %3.3f sec", Phase::_t_postalloc_expand.seconds());
-  }
-  tty->print_cr ("    codeGen        : %3.3f sec", Phase::_t_codeGeneration.seconds());
-  tty->print_cr ("    install_code   : %3.3f sec", Phase::_t_registerMethod.seconds());
-  tty->print_cr ("    -------------- : ----------");
-  double phase_subtotal = Phase::_t_parser.seconds() +
-    Phase::_t_optimizer.seconds() + Phase::_t_graphReshaping.seconds() +
-    Phase::_t_matcher.seconds() + Phase::_t_scheduler.seconds() +
-    Phase::_t_registerAllocation.seconds() + Phase::_t_blockOrdering.seconds() +
-    Phase::_t_codeGeneration.seconds() + Phase::_t_registerMethod.seconds();
-  double percent_of_method_compile = ((phase_subtotal == 0.0) ? 0.0 : phase_subtotal / Phase::_t_methodCompilation.seconds()) * 100.0;
-  // counters inside Compile::CodeGen include time for adapters and stubs
-  // so phase-total can be greater than 100%
-  tty->print_cr ("    total          : %3.3f sec,  %3.2f %%", phase_subtotal, percent_of_method_compile);
+
+  tty->print_cr ("       Matcher:             %7.3f s", timers[_t_matcher].seconds());
+  tty->print_cr ("       Scheduler:           %7.3f s", timers[_t_scheduler].seconds());
+
+  {
+    tty->print_cr ("       Regalloc:            %7.3f s", timers[_t_registerAllocation].seconds());
+    tty->print_cr ("         Ctor Chaitin:        %7.3f s", timers[_t_ctorChaitin].seconds());
+    tty->print_cr ("         Build IFG (virt):    %7.3f s", timers[_t_buildIFGvirtual].seconds());
+    tty->print_cr ("         Build IFG (phys):    %7.3f s", timers[_t_buildIFGphysical].seconds());
+    tty->print_cr ("         Compute Liveness:    %7.3f s", timers[_t_computeLive].seconds());
+    tty->print_cr ("         Regalloc Split:      %7.3f s", timers[_t_regAllocSplit].seconds());
+    tty->print_cr ("         Postalloc Copy Rem:  %7.3f s", timers[_t_postAllocCopyRemoval].seconds());
+    tty->print_cr ("         Fixup Spills:        %7.3f s", timers[_t_fixupSpills].seconds());
+    tty->print_cr ("         Compact:             %7.3f s", timers[_t_chaitinCompact].seconds());
+    tty->print_cr ("         Coalesce 1:          %7.3f s", timers[_t_chaitinCoalesce1].seconds());
+    tty->print_cr ("         Coalesce 2:          %7.3f s", timers[_t_chaitinCoalesce2].seconds());
+    tty->print_cr ("         Coalesce 3:          %7.3f s", timers[_t_chaitinCoalesce3].seconds());
+    tty->print_cr ("         Cache LRG:           %7.3f s", timers[_t_chaitinCacheLRG].seconds());
+    tty->print_cr ("         Simplify:            %7.3f s", timers[_t_chaitinSimplify].seconds());
+    tty->print_cr ("         Select:              %7.3f s", timers[_t_chaitinSelect].seconds());
+
+    double other = timers[_t_registerAllocation].seconds() -
+      (timers[_t_ctorChaitin].seconds() +
+       timers[_t_buildIFGvirtual].seconds() +
+       timers[_t_buildIFGphysical].seconds() +
+       timers[_t_computeLive].seconds() +
+       timers[_t_regAllocSplit].seconds() +
+       timers[_t_postAllocCopyRemoval].seconds() +
+       timers[_t_fixupSpills].seconds() +
+       timers[_t_chaitinCompact].seconds() +
+       timers[_t_chaitinCoalesce1].seconds() +
+       timers[_t_chaitinCoalesce2].seconds() +
+       timers[_t_chaitinCoalesce3].seconds() +
+       timers[_t_chaitinCacheLRG].seconds() +
+       timers[_t_chaitinSimplify].seconds() +
+       timers[_t_chaitinSelect].seconds());
 
-  assert( percent_of_method_compile > expected_method_compile_coverage ||
-          phase_subtotal < minimum_meaningful_method_compile,
-          "Must account for method compilation");
+    if (other > 0) {
+      tty->print_cr("         Other:               %7.3f s", other);
+    }
+  }
+  tty->print_cr ("       Block Ordering:      %7.3f s", timers[_t_blockOrdering].seconds());
+  tty->print_cr ("       Peephole:            %7.3f s", timers[_t_peephole].seconds());
+  if (Matcher::require_postalloc_expand) {
+    tty->print_cr ("       Postalloc Expand:    %7.3f s", timers[_t_postalloc_expand].seconds());
+  }
+  tty->print_cr ("       Code Emission:         %7.3f s", timers[_t_output].seconds());
+  tty->print_cr ("         Insn Scheduling:     %7.3f s", timers[_t_instrSched].seconds());
+  tty->print_cr ("         Build OOP maps:      %7.3f s", timers[_t_buildOopMaps].seconds());
+  tty->print_cr ("       Code Installation:   %7.3f s", timers[_t_registerMethod].seconds());
 
-  if( Phase::_t_temporaryTimer1.seconds() > minimum_reported_time ) {
+  if( timers[_t_temporaryTimer1].seconds() > 0 ) {
     tty->cr();
-    tty->print_cr ("    temporaryTimer1: %3.3f sec", Phase::_t_temporaryTimer1.seconds());
+    tty->print_cr ("       Temp Timer 1:        %7.3f s", timers[_t_temporaryTimer1].seconds());
   }
-  if( Phase::_t_temporaryTimer2.seconds() > minimum_reported_time ) {
+  if( timers[_t_temporaryTimer2].seconds() > 0 ) {
     tty->cr();
-    tty->print_cr ("    temporaryTimer2: %3.3f sec", Phase::_t_temporaryTimer2.seconds());
+    tty->print_cr ("       Temp Timer 2:        %7.3f s", timers[_t_temporaryTimer2].seconds());
   }
-  tty->print_cr ("    output         : %3.3f sec", Phase::_t_output.seconds());
-  tty->print_cr ("      isched         : %3.3f sec", Phase::_t_instrSched.seconds());
-  tty->print_cr ("      bldOopMaps     : %3.3f sec", Phase::_t_buildOopMaps.seconds());
+
+   double other = Phase::_t_totalCompilation.seconds() -
+      (timers[_t_parser].seconds() +
+       timers[_t_optimizer].seconds() +
+       timers[_t_matcher].seconds() +
+       timers[_t_scheduler].seconds() +
+       timers[_t_registerAllocation].seconds() +
+       timers[_t_blockOrdering].seconds() +
+       timers[_t_peephole].seconds() +
+       timers[_t_postalloc_expand].seconds() +
+       timers[_t_output].seconds() +
+       timers[_t_registerMethod].seconds() +
+       timers[_t_temporaryTimer1].seconds() +
+       timers[_t_temporaryTimer2].seconds());
+    if (other > 0) {
+      tty->print_cr("       Other:               %7.3f s", other);
+    }
+
 }
-#endif