ZJIT: Add more *_send_count stats (ruby#14689)

Author: k0kubun

zjit.rb

@@ -162,9 +162,16 @@ def stats_string
     print_counters_with_prefix(prefix: 'compile_error_', prompt: 'compile error reasons', buf:, stats:, limit: 20)
     print_counters_with_prefix(prefix: 'exit_', prompt: 'side exit reasons', buf:, stats:, limit: 20)
 
-    # Show the most important stats ratio_in_zjit at the end
+    # Show no-prefix counters, having the most important stat `ratio_in_zjit` at the end
     print_counters([
+      :send_count,
       :dynamic_send_count,
+      :optimized_send_count,
+      :iseq_optimized_send_count,
+      :inline_cfunc_optimized_send_count,
+      :variadic_cfunc_optimized_send_count,
+    ], buf:, stats:, right_align: true, base: :send_count)
+    print_counters([
       :dynamic_getivar_count,
       :dynamic_setivar_count,
 
@@ -202,12 +209,18 @@ def assert_compiles # :nodoc:
   # :stopdoc:
   private
 
-  def print_counters(keys, buf:, stats:)
-    left_pad = keys.map { |key| key.to_s.sub(/_time_ns\z/, '_time').size }.max + 1
+  def print_counters(keys, buf:, stats:, right_align: false, base: nil)
+    key_pad = keys.map { |key| key.to_s.sub(/_time_ns\z/, '_time').size }.max + 1
+    key_align = '-' unless right_align
+    value_pad = keys.filter_map { |key| stats[key] }.map { |value| number_with_delimiter(value).size }.max
+
     keys.each do |key|
       # Some stats like vm_insn_count and ratio_in_zjit are not supported on the release build
       next unless stats.key?(key)
       value = stats[key]
+      if base && key != base
+        ratio = " (%4.1f%%)" % (100.0 * value / stats[base])
+      end
 
       case key
       when :ratio_in_zjit
@@ -219,7 +232,7 @@ def print_counters(keys, buf:, stats:)
         value = number_with_delimiter(value)
       end
 
-      buf << "#{"%-#{left_pad}s" % "#{key}:"} #{value}\n"
+      buf << "%#{key_align}*s %*s%s\n" % [key_pad, "#{key}:", value_pad, value, ratio]
     end
   end
 

zjit/src/codegen.rs

@@ -661,6 +661,7 @@ fn gen_patch_point(jit: &mut JITState, asm: &mut Assembler, invariant: &Invarian
 /// Lowering for [`Insn::CCall`]. This is a low-level raw call that doesn't know
 /// anything about the callee, so handling for e.g. GC safety is dealt with elsewhere.
 fn gen_ccall(asm: &mut Assembler, cfun: *const u8, args: Vec<Opnd>) -> lir::Opnd {
+    gen_incr_counter(asm, Counter::inline_cfunc_optimized_send_count);
     asm.ccall(cfun, args)
 }
 
@@ -675,6 +676,8 @@ fn gen_ccall_variadic(
     cme: *const rb_callable_method_entry_t,
     state: &FrameState,
 ) -> lir::Opnd {
+    gen_incr_counter(asm, Counter::variadic_cfunc_optimized_send_count);
+
     gen_prepare_non_leaf_call(jit, asm, state);
 
     let stack_growth = state.stack_size();
@@ -1051,6 +1054,8 @@ fn gen_send_without_block_direct(
     args: Vec<Opnd>,
     state: &FrameState,
 ) -> lir::Opnd {
+    gen_incr_counter(asm, Counter::iseq_optimized_send_count);
+
     let local_size = unsafe { get_iseq_body_local_table_size(iseq) }.as_usize();
     let stack_growth = state.stack_size() + local_size + unsafe { get_iseq_body_stack_max(iseq) }.as_usize();
     gen_stack_overflow_check(jit, asm, state, stack_growth);

zjit/src/stats.rs

@@ -21,6 +21,9 @@ macro_rules! make_counters {
         dynamic_send {
             $($dynamic_send_counter_name:ident,)+
         }
+        optimized_send {
+            $($optimized_send_counter_name:ident,)+
+        }
         $($counter_name:ident,)+
     ) => {
         /// Struct containing the counter values
@@ -29,6 +32,7 @@ macro_rules! make_counters {
             $(pub $default_counter_name: u64,)+
             $(pub $exit_counter_name: u64,)+
             $(pub $dynamic_send_counter_name: u64,)+
+            $(pub $optimized_send_counter_name: u64,)+
             $(pub $counter_name: u64,)+
         }
 
@@ -39,6 +43,7 @@ macro_rules! make_counters {
             $($default_counter_name,)+
             $($exit_counter_name,)+
             $($dynamic_send_counter_name,)+
+            $($optimized_send_counter_name,)+
             $($counter_name,)+
         }
 
@@ -48,6 +53,7 @@ macro_rules! make_counters {
                     $( Counter::$default_counter_name => stringify!($default_counter_name).to_string(), )+
                     $( Counter::$exit_counter_name => stringify!($exit_counter_name).to_string(), )+
                     $( Counter::$dynamic_send_counter_name => stringify!($dynamic_send_counter_name).to_string(), )+
+                    $( Counter::$optimized_send_counter_name => stringify!($optimized_send_counter_name).to_string(), )+
                     $( Counter::$counter_name => stringify!($counter_name).to_string(), )+
                 }
             }
@@ -60,6 +66,7 @@ macro_rules! make_counters {
                 $( Counter::$default_counter_name => std::ptr::addr_of_mut!(counters.$default_counter_name), )+
                 $( Counter::$exit_counter_name => std::ptr::addr_of_mut!(counters.$exit_counter_name), )+
                 $( Counter::$dynamic_send_counter_name => std::ptr::addr_of_mut!(counters.$dynamic_send_counter_name), )+
+                $( Counter::$optimized_send_counter_name => std::ptr::addr_of_mut!(counters.$optimized_send_counter_name), )+
                 $( Counter::$counter_name => std::ptr::addr_of_mut!(counters.$counter_name), )+
             }
         }
@@ -80,6 +87,11 @@ macro_rules! make_counters {
             $( Counter::$dynamic_send_counter_name, )+
         ];
 
+        /// List of other counters that are summed as optimized_send_count.
+        pub const OPTIMIZED_SEND_COUNTERS: &'static [Counter] = &[
+            $( Counter::$optimized_send_counter_name, )+
+        ];
+
         /// List of other counters that are available only for --zjit-stats.
         pub const OTHER_COUNTERS: &'static [Counter] = &[
             $( Counter::$counter_name, )+
@@ -140,6 +152,13 @@ make_counters! {
         send_fallback_not_optimized_instruction,
     }
 
+    // Optimized send counters that are summed as optimized_send_count
+    optimized_send {
+        iseq_optimized_send_count,
+        inline_cfunc_optimized_send_count,
+        variadic_cfunc_optimized_send_count,
+    }
+
     // compile_error_: Compile error reasons
     compile_error_iseq_stack_too_large,
     compile_error_exception_handler,
@@ -421,6 +440,16 @@ pub extern "C" fn rb_zjit_stats(_ec: EcPtr, _self: VALUE, target_key: VALUE) ->
     }
     set_stat_usize!(hash, "dynamic_send_count", dynamic_send_count);
 
+    // Set optimized send counters
+    let mut optimized_send_count = 0;
+    for &counter in OPTIMIZED_SEND_COUNTERS {
+        let count = unsafe { *counter_ptr(counter) };
+        optimized_send_count += count;
+        set_stat_usize!(hash, &counter.name(), count);
+    }
+    set_stat_usize!(hash, "optimized_send_count", optimized_send_count);
+    set_stat_usize!(hash, "send_count", dynamic_send_count + optimized_send_count);
+
     // Set send fallback counters for NotOptimizedInstruction
     let send_fallback_counters = ZJITState::get_send_fallback_counters();
     for (op_idx, count) in send_fallback_counters.iter().enumerate().take(VM_INSTRUCTION_SIZE as usize) {