Add more compiler timeline durations

runtime/vm/compiler.cc

 // Copyright (c) 2012, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/compiler.h"
 
 #include "vm/assembler.h"
 
 #include "vm/ast_printer.h"
 #include "vm/block_scheduler.h"
@@ skipping 396 matching lines @@
                                         bool optimized,
                                         intptr_t osr_id) {
   const Function& function = parsed_function->function();
   if (optimized && !function.IsOptimizable()) {
     return false;
   }
   bool is_compiled = false;
   Thread* const thread = Thread::Current();
   Zone* const zone = thread->zone();
   Isolate* const isolate = thread->isolate();
+  TimelineStream* compiler_timeline = isolate->GetCompilerStream();
   CSTAT_TIMER_SCOPE(thread, codegen_timer);
   HANDLESCOPE(thread);
 
   // Get current generation count so that we can check and ensure that the code
   // was not invalidated while we were compiling in the background.
   uint32_t cha_invalidation_gen_at_start = isolate->cha_invalidation_gen();
   uint32_t field_invalidation_gen_at_start = isolate->field_invalidation_gen();
   uint32_t prefix_invalidation_gen_at_start =
       isolate->prefix_invalidation_gen();
 
@@ skipping 41 matching lines @@
           if (FLAG_print_ic_data_map) {
             for (intptr_t i = 0; i < ic_data_array->length(); i++) {
               if ((*ic_data_array)[i] != NULL) {
                 THR_Print("%" Pd " ", i);
                 FlowGraphPrinter::PrintICData(*(*ic_data_array)[i]);
               }
             }
           }
         }
 
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "BuildFlowGraph");
         flow_graph = pipeline->BuildFlowGraph(zone,
                                               parsed_function,
                                               *ic_data_array,
                                               osr_id);
       }
 
       const bool print_flow_graph =
           (FLAG_print_flow_graph ||
           (optimized && FLAG_print_flow_graph_optimized)) &&
           FlowGraphPrinter::ShouldPrint(function);
 
       if (print_flow_graph) {
         if (osr_id == Compiler::kNoOSRDeoptId) {
           FlowGraphPrinter::PrintGraph("Before Optimizations", flow_graph);
         } else {
           FlowGraphPrinter::PrintGraph("For OSR", flow_graph);
         }
       }
 
       BlockScheduler block_scheduler(flow_graph);
       const bool reorder_blocks =
           FlowGraph::ShouldReorderBlocks(function, optimized);
       if (reorder_blocks) {
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "BlockScheduler::AssignEdgeWeights");
         block_scheduler.AssignEdgeWeights();
       }
 
       if (optimized) {
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "ComputeSSA");
         CSTAT_TIMER_SCOPE(thread, ssa_timer);
         // Transform to SSA (virtual register 0 and no inlining arguments).
         flow_graph->ComputeSSA(0, NULL);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
         if (print_flow_graph) {
           FlowGraphPrinter::PrintGraph("After SSA", flow_graph);
         }
       }
 
       // Maps inline_id_to_function[inline_id] -> function. Top scope
       // function has inline_id 0. The map is populated by the inliner.
       GrowableArray<const Function*> inline_id_to_function;
       // For a given inlining-id(index) specifies the caller's inlining-id.
       GrowableArray<intptr_t> caller_inline_id;
       // Collect all instance fields that are loaded in the graph and
       // have non-generic type feedback attached to them that can
       // potentially affect optimizations.
       if (optimized) {
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "OptimizationPasses");
         inline_id_to_function.Add(&function);
         // Top scope function has no caller (-1).
         caller_inline_id.Add(-1);
         CSTAT_TIMER_SCOPE(thread, graphoptimizer_timer);
 
         FlowGraphOptimizer optimizer(flow_graph,
                                      use_speculative_inlining,
                                      &inlining_black_list);
         if (Compiler::always_optimize()) {
           optimizer.PopulateWithICData();
 
           optimizer.ApplyClassIds();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
           FlowGraphTypePropagator::Propagate(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
         optimizer.ApplyICData();
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
         // Optimize (a << b) & c patterns, merge operations.
         // Run early in order to have more opportunity to optimize left shifts.
         optimizer.TryOptimizePatterns();
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
         FlowGraphInliner::SetInliningId(flow_graph, 0);
 
         // Inlining (mutates the flow graph)
         if (FLAG_use_inlining) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "Inlining");
           CSTAT_TIMER_SCOPE(thread, graphinliner_timer);
           // Propagate types to create more inlining opportunities.
           FlowGraphTypePropagator::Propagate(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
           // Use propagated class-ids to create more inlining opportunities.
           optimizer.ApplyClassIds();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
           FlowGraphInliner inliner(flow_graph,
                                    &inline_id_to_function,
                                    &caller_inline_id,
                                    use_speculative_inlining,
                                    &inlining_black_list);
           inliner.Inline();
           // Use lists are maintained and validated by the inliner.
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
 
         // Propagate types and eliminate more type tests.
         FlowGraphTypePropagator::Propagate(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        // Use propagated class-ids to optimize further.
-        optimizer.ApplyClassIds();
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "ApplyClassIds");
+          // Use propagated class-ids to optimize further.
+          optimizer.ApplyClassIds();
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        }
 
         // Propagate types for potentially newly added instructions by
         // ApplyClassIds(). Must occur before canonicalization.
         FlowGraphTypePropagator::Propagate(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        // Do optimizations that depend on the propagated type information.
-        if (optimizer.Canonicalize()) {
-          // Invoke Canonicalize twice in order to fully canonicalize patterns
-          // like "if (a & const == 0) { }".
-          optimizer.Canonicalize();
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "Canonicalize");

[Florian Schneider, 2015/12/15 17:15:35]

Remove scope for canonicalize.

[Cutch, 2015/12/15 17:51:56]

Done.

+          // Do optimizations that depend on the propagated type information.
+          if (optimizer.Canonicalize()) {
+            // Invoke Canonicalize twice in order to fully canonicalize patterns
+            // like "if (a & const == 0) { }".
+            optimizer.Canonicalize();
+          }
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        BranchSimplifier::Simplify(flow_graph);
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "BranchSimplifier");
+          BranchSimplifier::Simplify(flow_graph);
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        IfConverter::Simplify(flow_graph);
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
+          IfConverter::Simplify(flow_graph);
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        }
 
         if (FLAG_constant_propagation) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "ConstantPropagation");
           ConstantPropagator::Optimize(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
           // A canonicalization pass to remove e.g. smi checks on smi constants.
           optimizer.Canonicalize();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
           // Canonicalization introduced more opportunities for constant
           // propagation.
           ConstantPropagator::Optimize(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
 
         // Optimistically convert loop phis that have a single non-smi input
         // coming from the loop pre-header into smi-phis.
         if (FLAG_loop_invariant_code_motion) {
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Remove this scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "LICM");
           LICM licm(flow_graph);
           licm.OptimisticallySpecializeSmiPhis();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
 
         // Propagate types and eliminate even more type tests.
         // Recompute types after constant propagation to infer more precise
         // types for uses that were previously reached by now eliminated phis.
         FlowGraphTypePropagator::Propagate(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        // Where beneficial convert Smi operations into Int32 operations.
-        // Only meanigful for 32bit platforms right now.
-        optimizer.WidenSmiToInt32();
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "WidenSmiToInt32");
+          // Where beneficial convert Smi operations into Int32 operations.
+          // Only meanigful for 32bit platforms right now.
+          optimizer.WidenSmiToInt32();
+        }
 
-        // Unbox doubles. Performed after constant propagation to minimize
-        // interference from phis merging double values and tagged
-        // values coming from dead paths.
-        optimizer.SelectRepresentations();
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        {
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Merge with previous scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "UnboxSelectRepresentations");
+          // Unbox doubles. Performed after constant propagation to minimize
+          // interference from phis merging double values and tagged
+          // values coming from dead paths.
+          optimizer.SelectRepresentations();
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        }
 

[Florian Schneider, 2015/12/15 17:15:35]

ComputeBlockEffects, CSE, LICM in one scope.

[Cutch, 2015/12/15 17:51:56]

Done.

         if (FLAG_common_subexpression_elimination ||
             FLAG_loop_invariant_code_motion) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "ComputeBlockEffects");
           flow_graph->ComputeBlockEffects();
         }
 
         if (FLAG_common_subexpression_elimination) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "DominatorBasedCSE");
           if (DominatorBasedCSE::Optimize(flow_graph)) {
             DEBUG_ASSERT(flow_graph->VerifyUseLists());
             optimizer.Canonicalize();
             // Do another round of CSE to take secondary effects into account:
             // e.g. when eliminating dependent loads (a.x[0] + a.x[0])
             // TODO(fschneider): Change to a one-pass optimization pass.
             if (DominatorBasedCSE::Optimize(flow_graph)) {
               optimizer.Canonicalize();
             }
             DEBUG_ASSERT(flow_graph->VerifyUseLists());
           }
         }
 
         // Run loop-invariant code motion right after load elimination since it
         // depends on the numbering of loads from the previous load-elimination.
         if (FLAG_loop_invariant_code_motion) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "LICM");
           LICM licm(flow_graph);
           licm.Optimize();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
         flow_graph->RemoveRedefinitions();
 
         // Optimize (a << b) & c patterns, merge operations.
         // Run after CSE in order to have more opportunity to merge
         // instructions that have same inputs.
         optimizer.TryOptimizePatterns();
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        DeadStoreElimination::Optimize(flow_graph);
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "DeadStoreElimination");
+          DeadStoreElimination::Optimize(flow_graph);
+        }
 
         if (FLAG_range_analysis) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "RangeAnalysis");
           // Propagate types after store-load-forwarding. Some phis may have
           // become smi phis that can be processed by range analysis.
           FlowGraphTypePropagator::Propagate(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
           // We have to perform range analysis after LICM because it
           // optimistically moves CheckSmi through phis into loop preheaders
           // making some phis smi.
           optimizer.InferIntRanges();
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
 
         if (FLAG_constant_propagation) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "ConstantPropagator::OptimizeBranches");
           // Constant propagation can use information from range analysis to
           // find unreachable branch targets and eliminate branches that have
           // the same true- and false-target.
           ConstantPropagator::OptimizeBranches(flow_graph);
           DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
 
         // Recompute types after code movement was done to ensure correct
         // reaching types for hoisted values.
         FlowGraphTypePropagator::Propagate(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        // Optimize try-blocks.
-        TryCatchAnalyzer::Optimize(flow_graph);
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "TryCatchAnalyzer::Optimize");
+          // Optimize try-blocks.
+          TryCatchAnalyzer::Optimize(flow_graph);
+        }
 
-        // Detach environments from the instructions that can't deoptimize.
-        // Do it before we attempt to perform allocation sinking to minimize
-        // amount of materializations it has to perform.
-        optimizer.EliminateEnvironments();
+        {
+          // Detach environments from the instructions that can't deoptimize.
+          // Do it before we attempt to perform allocation sinking to minimize
+          // amount of materializations it has to perform.
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Remove this scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "EliminateEnvironments");
+          optimizer.EliminateEnvironments();
+        }
 
-        DeadCodeElimination::EliminateDeadPhis(flow_graph);
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "EliminateDeadPhis");
+          DeadCodeElimination::EliminateDeadPhis(flow_graph);
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
+        }
 
         if (optimizer.Canonicalize()) {
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Remove this scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "Canonicalize");
           optimizer.Canonicalize();
         }
 
         // Attempt to sink allocations of temporary non-escaping objects to
         // the deoptimization path.
         AllocationSinking* sinking = NULL;
         if (FLAG_allocation_sinking &&
             (flow_graph->graph_entry()->SuccessorCount()  == 1)) {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "AllocationSinking::Optimize");
           // TODO(fschneider): Support allocation sinking with try-catch.
           sinking = new AllocationSinking(flow_graph);
           sinking->Optimize();
         }
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
         DeadCodeElimination::EliminateDeadPhis(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
         FlowGraphTypePropagator::Propagate(flow_graph);
         DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
-        // Ensure that all phis inserted by optimization passes have consistent
-        // representations.
-        optimizer.SelectRepresentations();
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "SelectRepresentations");
+          // Ensure that all phis inserted by optimization passes have
+          // consistent representations.
+          optimizer.SelectRepresentations();
+        }
 
-        if (optimizer.Canonicalize()) {
-          // To fully remove redundant boxing (e.g. BoxDouble used only in
-          // environments and UnboxDouble instructions) instruction we
-          // first need to replace all their uses and then fold them away.
-          // For now we just repeat Canonicalize twice to do that.
-          // TODO(vegorov): implement a separate representation folding pass.
-          optimizer.Canonicalize();
+        {
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Remove this scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "Canonicalize");
+          if (optimizer.Canonicalize()) {
+            // To fully remove redundant boxing (e.g. BoxDouble used only in
+            // environments and UnboxDouble instructions) instruction we
+            // first need to replace all their uses and then fold them away.
+            // For now we just repeat Canonicalize twice to do that.
+            // TODO(vegorov): implement a separate representation folding pass.
+            optimizer.Canonicalize();
+          }
+          DEBUG_ASSERT(flow_graph->VerifyUseLists());
         }
-        DEBUG_ASSERT(flow_graph->VerifyUseLists());
 
         if (sinking != NULL) {
+          TimelineDurationScope tds2(
+              thread,
+              compiler_timeline,
+              "AllocationSinking::DetachMaterializations");
           // Remove all MaterializeObject instructions inserted by allocation
           // sinking from the flow graph and let them float on the side
           // referenced only from environments. Register allocator will consider
           // them as part of a deoptimization environment.
           sinking->DetachMaterializations();
         }
 
-        // Compute and store graph informations (call & instruction counts)
-        // to be later used by the inliner.
-        FlowGraphInliner::CollectGraphInfo(flow_graph, true);
+        {
+          TimelineDurationScope tds2(thread,

[Florian Schneider, 2015/12/15 17:15:35]

Remove this scope.

[Cutch, 2015/12/15 17:51:56]

Done.

+                                     compiler_timeline,
+                                     "CollectGraphInfo");
+          // Compute and store graph informations (call & instruction counts)
+          // to be later used by the inliner.
+          FlowGraphInliner::CollectGraphInfo(flow_graph, true);
+        }
 
-        // Perform register allocation on the SSA graph.
-        FlowGraphAllocator allocator(*flow_graph);
-        allocator.AllocateRegisters();
-        if (reorder_blocks) block_scheduler.ReorderBlocks();
+        {
+          TimelineDurationScope tds2(thread,
+                                     compiler_timeline,
+                                     "AllocateRegisters");
+          // Perform register allocation on the SSA graph.
+          FlowGraphAllocator allocator(*flow_graph);
+          allocator.AllocateRegisters();
+        }
+
+        if (reorder_blocks) {
+          TimelineDurationScope tds(thread,
+                                    compiler_timeline,
+                                    "BlockScheduler::ReorderBlocks");
+          block_scheduler.ReorderBlocks();
+        }
 
         if (print_flow_graph) {
           FlowGraphPrinter::PrintGraph("After Optimizations", flow_graph);
         }
       }
 
       ASSERT(inline_id_to_function.length() == caller_inline_id.length());
       Assembler assembler(use_far_branches);
       FlowGraphCompiler graph_compiler(&assembler, flow_graph,
                                        *parsed_function, optimized,
                                        inline_id_to_function,
                                        caller_inline_id);
       {
         CSTAT_TIMER_SCOPE(thread, graphcompiler_timer);
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "CompileGraph");
         graph_compiler.CompileGraph();
         pipeline->FinalizeCompilation();
       }
       {
+        TimelineDurationScope tds(thread,
+                                  compiler_timeline,
+                                  "FinalizeCompilation");
         // This part of compilation must be at a safepoint.
         if (!Thread::Current()->IsMutatorThread()) {
           // Stop mutator thread before creating the instruction object and
           // installing code.
           // Mutator thread may not run code while we are creating the
           // instruction object, since the creation of instruction object
           // changes code page access permissions (makes them temporary not
           // executable).
           isolate->thread_registry()->SafepointThreads();
         }
@@ skipping 1102 matching lines @@
 }
 
 
 void BackgroundCompiler::EnsureInit(Thread* thread) {
   UNREACHABLE();
 }
 
 #endif  // DART_PRECOMPILED
 
 }  // namespace dart