vernier 1.10.0 → 1.11.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: a504e7ef885ed393a864128d15ca792af4e6ab60f30cdf90afbc36aa5d577b34
-  data.tar.gz: 87ab42b6cc3ea1effae71f3b95bdc4d90c117abf1aadc3d425bb30b4573dafa9
+  metadata.gz: 6e8d610b954485d3bccefa47408932dddd5806b618caa2958d689f8c0c8fbb3d
+  data.tar.gz: 2423bf5e06dd1168e21f0a5063050e897137c37bd09ce68a540af9d84b0a7546
 SHA512:
-  metadata.gz: f52d1d392009084c1511b7b2cc24c51fdb5da983d0fe1dd60286ed9fe0564229ec409513cd179f2af3ceb96f9a085e2f9f48b556ac5ae6b597aba7c13bd56cf8
-  data.tar.gz: 7b8255043801305c221a967a25974ff1e5896393adf8c820784dd980a2eeecfa49f63509b7eeddf1c0f81a7279c29ea87ffa88622de306d0d5b430c23a40c109
+  metadata.gz: 942b097d843b63457650254b2e9717698939df9ee40c02ff367832d344121e6ffdafffbc2e0824d70229bf5d4187d1cf0eb64491402a4c54c4c5f6dd8fb0b98f
+  data.tar.gz: 2f710c6b9eeb8e4e352290902aed1a4142ea5b48df7039ee49de66dc7c9f7411e25cb77319addba69a15f072f08b3820a51535fe19070a4e77341aa6c000a57b

data/.ruby-version

@@ -1 +1 @@
-3.4.7
+4.0

data/Gemfile

@@ -12,3 +12,5 @@ gem "rake-compiler"
 gem "benchmark-ips"
 
 gem "minitest", "~> 5.0"
+
+gem "zlib", ">= 3.2.1"

data/ext/vernier/heap_tracker.cc

@@ -138,6 +138,9 @@ class HeapTracker {
       if (RTEST(tp_newobj)) {
         rb_tracepoint_disable(tp_newobj);
         tp_newobj = Qnil;
+        if (tombstones * 10 > object_list.size()) {
+          rebuild();
+        }
       }
     }
 

data/lib/vernier/collector.rb

@@ -29,6 +29,7 @@ module Vernier
         result.instance_variable_set(:@threads, {
           0 => {
             tid: 0,
+            is_main: true,
             name: "custom",
             started_at: @started_at,
             samples: @samples,
@@ -80,6 +81,7 @@ module Vernier
         result.instance_variable_set(:@threads, {
           0 => {
             tid: 0,
+            is_main: true,
             name: "retained memory",
             started_at: @started_at,
             samples: samples,

data/lib/vernier/output/cpuprofile.rb

@@ -1,8 +1,12 @@
 require "json"
 
+require_relative "../output_helpers"
+
 module Vernier
   module Output
     class Cpuprofile
+      include OutputHelpers
+
       def initialize(profile)
         @profile = profile
       end
@@ -98,7 +102,9 @@ module Vernier
         line = stack_table.frame_line_no(frame_idx) - 1
 
         func_name = stack_table.func_name(func_idx)
+        func_name = sanitize_string(func_name) if func_name
         filename = stack_table.func_filename(func_idx)
+        filename = sanitize_string(filename) if filename
 
         node_id = nodes.length
         node = {

data/lib/vernier/output/file_listing.rb

@@ -1,23 +1,19 @@
 # frozen_string_literal: true
 
 require_relative "filename_filter"
+require_relative "../output_helpers"
 require "cgi/escape"
 
 module Vernier
   module Output
     class FileListing
+      include OutputHelpers
+
       class SamplesByLocation
         attr_accessor :self, :total
         def initialize
           @self = @total = 0
         end
-
-        def +(other)
-          ret = SamplesByLocation.new
-          ret.self = @self + other.self
-          ret.total = @total + other.total
-          ret
-        end
       end
 
       def initialize(profile)
@@ -25,6 +21,8 @@ module Vernier
       end
 
       def samples_by_file
+        return @samples_by_file if defined?(@samples_by_file)
+
         thread = @profile.main_thread
         if Hash === thread
           # live profile
@@ -38,22 +36,9 @@ module Vernier
         weights = thread[:weights]
         samples = thread[:samples]
 
-        self_samples_by_frame = Hash.new do |h, k|
-          h[k] = SamplesByLocation.new
-        end
-
-        samples.zip(weights).each do |stack_idx, weight|
-          # self time
-          top_frame_index = stack_table.stack_frame_idx(stack_idx)
-          self_samples_by_frame[top_frame_index].self += weight
+        stack_weights = collapse_stack_weights(samples, weights)
 
-          # total time
-          while stack_idx
-            frame_idx = stack_table.stack_frame_idx(stack_idx)
-            self_samples_by_frame[frame_idx].total += weight
-            stack_idx = stack_table.stack_parent_idx(stack_idx)
-          end
-        end
+        self_by_frame, total_by_frame = frame_weights(stack_table, stack_weights)
 
         samples_by_file = Hash.new do |h, k|
           h[k] = Hash.new do |h2, k2|
@@ -61,19 +46,64 @@ module Vernier
           end
         end
 
-        self_samples_by_frame.each do |frame, samples|
+        self_by_frame.each_with_index do |self_weight, frame|
+          total_weight = total_by_frame[frame]
+          next if self_weight == 0 && total_weight == 0
+
           line = stack_table.frame_line_no(frame)
           func_index = stack_table.frame_func_idx(frame)
           filename = stack_table.func_filename(func_index)
 
-          samples_by_file[filename][line] += samples
+          location = samples_by_file[filename][line]
+          location.self += self_weight
+          location.total += total_weight
         end
 
-        samples_by_file.transform_keys! do |filename|
+        @samples_by_file = samples_by_file.transform_keys! do |filename|
           filename_filter.call(filename)
         end
       end
 
+      # Returns [self_weight, total_weight] arrays indexed by frame, from a
+      # hash of per-stack weights.
+      #
+      # The stack table is a prefix tree whose parent is always interned
+      # before its children (parent idx < child idx), so total weights are
+      # accumulated bottom-up in a single reverse pass.
+      private def frame_weights(stack_table, stack_weights)
+        # Scatter the sparse weights into an array indexed by stack.
+        stack_totals = Array.new(stack_table.stack_count, 0)
+        stack_weights.each do |stack_idx, weight|
+          stack_totals[stack_idx] = weight
+        end
+
+        # Accumulate each stack's total into its parent, children first,
+        # so a single descending pass completes every subtree's total.
+        (stack_totals.length - 1).downto(0) do |stack_idx|
+          parent_idx = stack_table.stack_parent_idx(stack_idx)
+          next unless parent_idx
+
+          if parent_idx >= stack_idx
+            raise "Invalid profile: stack table is not parent-before-child ordered"
+          end
+
+          stack_totals[parent_idx] += stack_totals[stack_idx]
+        end
+
+        self_by_frame = Array.new(stack_table.frame_count, 0)
+        total_by_frame = Array.new(stack_table.frame_count, 0)
+        stack_weights.each do |stack_idx, weight|
+          self_by_frame[stack_table.stack_frame_idx(stack_idx)] += weight
+        end
+        stack_totals.each_with_index do |total, stack_idx|
+          next if total == 0
+
+          total_by_frame[stack_table.stack_frame_idx(stack_idx)] += total
+        end
+
+        [self_by_frame, total_by_frame]
+      end
+
       def output(template: nil)
         output = +""
 
@@ -98,8 +128,7 @@ module Vernier
       end
 
       def total
-        thread = @profile.main_thread
-        thread[:weights].sum
+        @total ||= @profile.main_thread[:weights].sum
       end
 
       def format_file(output, filename, all_samples, total:)

data/lib/vernier/output/firefox.rb

@@ -4,6 +4,7 @@ require "json"
 require "rbconfig"
 
 require_relative "filename_filter"
+require_relative "../output_helpers"
 
 module Vernier
   module Output
@@ -267,6 +268,13 @@ module Vernier
       end
 
       class Thread
+        include OutputHelpers
+
+        SAMPLE_CATEGORY_NAMES = {
+          1 => "Idle",
+          2 => "Stalled"
+        }.freeze
+
         attr_reader :profile, :is_start
 
         def initialize(ruby_thread_id, profile, categorizer, name:, tid:, samples:, weights:, timestamps: nil, sample_categories: nil, markers:, started_at:, stopped_at: nil, allocations: nil, is_main: nil, is_start: nil)
@@ -355,6 +363,31 @@ module Vernier
           @frame_subcategories = @stack_table_hash[:frame_table].fetch(:func).map do |func_idx|
             func_subcategories[func_idx]
           end
+
+          @sample_category_idx = SAMPLE_CATEGORY_NAMES.transform_values do |name|
+            @categorizer.get_category(name).idx
+          end
+
+          @samples.zip(@sample_categories).each do |sample, raw_category|
+            next if raw_category == 0
+
+            @categorized_stacks[[sample, raw_category]]
+          end
+
+          base_stack_frames = @stack_table_hash[:stack_table].fetch(:frame)
+          base_frame_count = @stack_table_hash[:frame_table].fetch(:func).size
+          @extra_frames = []
+          @categorized_frame_map = {}
+
+          @categorized_stacks.each_key do |(stack, raw_category)|
+            original_frame_idx = base_stack_frames[stack]
+            key = [original_frame_idx, raw_category]
+            next if @categorized_frame_map.key?(key)
+
+            new_frame_idx = base_frame_count + @extra_frames.size
+            @categorized_frame_map[key] = new_frame_idx
+            @extra_frames << [original_frame_idx, @sample_category_idx[raw_category]]
+          end
         end
 
         def categorize_filename(filename)
@@ -377,7 +410,7 @@ module Vernier
         def find_category_and_subcategory(filename, categories)
           categories.each do |category_name|
             category = @categorizer.get_category(category_name)
-            subcategory = category.subcategories.detect {|c| c.matches?(filename) }&.idx
+            subcategory = category.subcategories.detect { |c| c.matches?(filename) }&.idx
             return category, subcategory if subcategory
           end
           [nil, nil]
@@ -430,7 +463,7 @@ module Vernier
           categories = []
           data = []
 
-          @markers.each_with_index do |(_, name, start, finish, phase, datum), i|
+          @markers.each do |(_, name, start, finish, phase, datum)|
             string_indexes << @strings[name]
             start_times << (start / 1_000_000.0)
 
@@ -463,12 +496,14 @@ module Vernier
         end
 
         def allocations_table
-          return nil if !@allocations
+          return nil unless @allocations
+
           samples, weights, timestamps = @allocations.values_at(:samples, :weights, :timestamps)
-          return nil if samples.size == 0
+          return nil if samples.empty?
+
           size = samples.size
           timestamps = timestamps.map { _1 / 1_000_000.0 }
-          ret = {
+          {
             "time": timestamps,
             "className": ["Object"]*size,
             "typeName": ["JSObject"]*size,
@@ -478,7 +513,6 @@ module Vernier
             "stack": samples,
             "length": size
           }
-          ret
         end
 
         def samples_table
@@ -518,21 +552,22 @@ module Vernier
         end
 
         def stack_table
-          frames =   @stack_table_hash[:stack_table].fetch(:frame).dup
+          base_frames = @stack_table_hash[:stack_table].fetch(:frame)
+          frames = base_frames.dup
           prefixes = @stack_table_hash[:stack_table].fetch(:parent).dup
-          categories  = frames.map{|idx| @frame_categories[idx].idx }
-          subcategories  = frames.map{|idx| @frame_subcategories[idx] }
+          categories = frames.map { |idx| @frame_categories[idx].idx }
+          subcategories = frames.map { |idx| @frame_subcategories[idx] }
 
-          @categorized_stacks.each_key do |(stack, category)|
-            frames << frames[stack]
+          @categorized_stacks.each_key do |(stack, raw_category)|
+            original_frame_idx = base_frames[stack]
+            frames << @categorized_frame_map[[original_frame_idx, raw_category]]
             prefixes << prefixes[stack]
-            categories << category
+            categories << @sample_category_idx[raw_category]
             subcategories << 0
           end
 
-          size = frames.length
-          raise unless frames.size == size
-          raise unless prefixes.size == size
+          raise unless prefixes.size == frames.size
+
           {
             frame: frames,
             category: categories,
@@ -543,18 +578,27 @@ module Vernier
         end
 
         def frame_table
-          funcs = @stack_table_hash[:frame_table].fetch(:func)
-          lines = @stack_table_hash[:frame_table].fetch(:line)
+          funcs = @stack_table_hash[:frame_table].fetch(:func).dup
+          lines = @stack_table_hash[:frame_table].fetch(:line).dup
           raise unless lines.size == funcs.size
 
+          categories = @frame_categories.map(&:idx)
+          subcategories = @frame_subcategories.dup
+          implementations = @frame_implementations.dup
+
+          @extra_frames.each do |(frame_idx, category_idx)|
+            funcs << funcs[frame_idx]
+            lines << lines[frame_idx]
+            categories << category_idx
+            subcategories << 0
+            implementations << implementations[frame_idx]
+          end
+
           size = funcs.size
           none = [nil] * size
           default = [0] * size
           unidentified = [-1] * size
 
-          categories = @frame_categories.map(&:idx)
-          subcategories = @frame_subcategories
-
           {
             address: unidentified,
             inlineDepth: default,
@@ -563,7 +607,7 @@ module Vernier
             func: funcs,
             nativeSymbol: none,
             innerWindowID: none,
-            implementation: @frame_implementations,
+            implementation: implementations,
             line: lines,
             column: none,
             length: size
@@ -597,21 +641,7 @@ module Vernier
 
         def string_table
           @strings.keys.map do |string|
-            if string.ascii_only?
-              string
-            elsif string.encoding == Encoding::UTF_8
-              if string.valid_encoding?
-                string
-              else
-                string.scrub
-              end
-            elsif string.encoding == Encoding::BINARY
-              # TODO: We might want to guess UTF-8 and escape the binary more explicitly
-              string.dup.force_encoding("UTF-8").scrub
-            else
-              # TODO: ideally we should attempt to properly re-encode here, but right now I think this is dead code
-              string.dup.force_encoding("UTF-8").scrub
-            end
+            sanitize_string(string)
           end
         end
 

data/lib/vernier/output/top.rb

@@ -1,8 +1,12 @@
 # frozen_string_literal: true
 
+require_relative "../output_helpers"
+
 module Vernier
   module Output
     class Top
+      include OutputHelpers
+
       def initialize(profile, row_limit)
         @profile = profile
         @row_limit = row_limit
@@ -57,10 +61,10 @@ module Vernier
             @profile._stack_table
           end
 
-        stack_weights = Hash.new(0)
-        thread[:samples].zip(thread[:weights]) do |stack_idx, weight|
-          stack_weights[stack_idx] += weight
-        end
+        samples = thread[:samples]
+        weights = thread[:weights]
+
+        stack_weights = collapse_stack_weights(samples, weights)
 
         total = stack_weights.values.sum
 

data/lib/vernier/output_helpers.rb

@@ -0,0 +1,33 @@
+# frozen_string_literal: true
+
+module Vernier
+  module OutputHelpers
+    # Collapses parallel samples/weights arrays into a hash of total weight
+    # per stack index. Retained-mode profiles have one sample per object and
+    # many samples share the same stack, so collapsing first keeps later
+    # aggregation proportional to the number of unique stacks.
+    def collapse_stack_weights(samples, weights)
+      stack_weights = Hash.new(0)
+      samples.each_with_index do |stack_idx, idx|
+        stack_weights[stack_idx] += weights[idx]
+      end
+      stack_weights
+    end
+
+    # Returns a string safe to embed in JSON output: valid UTF-8 with any
+    # invalid bytes replaced. Method names and paths can carry arbitrary
+    # encodings (e.g. BINARY or Shift_JIS source), which would otherwise
+    # raise JSON::GeneratorError.
+    def sanitize_string(string)
+      if string.ascii_only?
+        string
+      elsif string.encoding == Encoding::UTF_8
+        string.valid_encoding? ? string : string.scrub
+      else
+        # We don't know the real encoding; interpret the bytes as UTF-8
+        # and replace anything invalid.
+        string.dup.force_encoding(Encoding::UTF_8).scrub
+      end
+    end
+  end
+end

data/lib/vernier/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Vernier
-  VERSION = "1.10.0"
+  VERSION = "1.11.0"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: vernier
 version: !ruby/object:Gem::Version
-  version: 1.10.0
+  version: 1.11.0
 platform: ruby
 authors:
 - John Hawthorn
@@ -106,6 +106,7 @@ files:
 - lib/vernier/output/firefox.rb
 - lib/vernier/output/markdown.rb
 - lib/vernier/output/top.rb
+- lib/vernier/output_helpers.rb
 - lib/vernier/parsed_profile.rb
 - lib/vernier/result.rb
 - lib/vernier/stack_table.rb
@@ -134,7 +135,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 4.0.3
+rubygems_version: 4.0.16
 specification_version: 4
 summary: A next generation CRuby profiler
 test_files: []