move ADD_PC around to optimize PC manipluiations

This commit introduces new attribute handles_flame and if that is
_not_ the case, places ADD_PC right after INC_SP.  This improves
locality of PC manipulations to prevents unnecessary register spill-
outs. As a result, it reduces the size of vm_exec_core from 32,688
bytes to 32,384 bytes on my machine.

Speedup is very faint, but certain.

-----------------------------------------------------------
benchmark results:
minimum results in each 3 measurements.
Execution time (sec)
name    before  after
so_ackermann     0.476  0.464
so_array         0.742  0.728
so_binary_trees  5.493  5.466
so_concatenate   3.619  3.395
so_count_words   0.190  0.184
so_exception     0.249  0.239
so_fannkuch      0.994  0.953
so_fasta         1.369  1.374
so_k_nucleotide  1.111  1.111
so_lists         0.470  0.481
so_mandelbrot    2.059  2.050
so_matrix        0.466  0.465
so_meteor_contest        2.712  2.781
so_nbody         1.154  1.204
so_nested_loop   0.852  0.846
so_nsieve        1.636  1.623
so_nsieve_bits   2.073  2.039
so_object        0.616  0.584
so_partial_sums  1.464  1.481
so_pidigits      1.075  1.082
so_random        0.321  0.317
so_reverse_complement    0.555  0.558
so_sieve         0.495  0.490
so_spectralnorm  1.634  1.627

Speedup ratio: compare with the result of `before' (greater is better)
name    after
so_ackermann    1.025
so_array        1.019
so_binary_trees 1.005
so_concatenate  1.066
so_count_words  1.030
so_exception    1.040
so_fannkuch     1.043
so_fasta        0.996
so_k_nucleotide 1.000
so_lists        0.978
so_mandelbrot   1.004
so_matrix       1.001
so_meteor_contest       0.975
so_nbody        0.959
so_nested_loop  1.007
so_nsieve       1.008
so_nsieve_bits  1.017
so_object       1.056
so_partial_sums 0.989
so_pidigits     0.994
so_random       1.014
so_reverse_complement   0.996
so_sieve        1.010
so_spectralnorm 1.004

git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@62051 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

insns.def

@@ -680,6 +680,7 @@ defineclass
 (ID id, ISEQ class_iseq, rb_num_t flags)
 (VALUE cbase, VALUE super)
 (VALUE val)
+// attr bool handles_frame = true;
 {
     VALUE klass = vm_find_or_create_class_by_id(id, flags, cbase, super);
 
@@ -829,6 +830,7 @@ leave
 ()
 (VALUE val)
 (VALUE val)
+// attr bool handles_frame = true;
 {
     if (OPT_CHECKED_RUN) {
 	const VALUE *const bp = vm_base_ptr(reg_cfp);
@@ -1380,6 +1382,7 @@ opt_call_c_function
 (rb_insn_func_t funcptr)
 ()
 ()
+// attr bool handles_frame = true;
 {
     reg_cfp = (funcptr)(ec, reg_cfp);
 

tool/ruby_vm/models/bare_instructions.rb

@@ -33,6 +33,7 @@ class RubyVM::BareInstructions
       h[a.key] = a
     }
     @attrs_orig = @attrs.dup
+    predefine_attributes
   end
 
   def pretty_name
@@ -109,8 +110,8 @@ class RubyVM::BareInstructions
     }.join
   end
 
-  def pushs_frame?
+  def handles_frame?
-    opes.any? {|o| /CALL_INFO/ =~ o[:type] }
+    /\b(false|0)\b/ !~ @attrs['handles_frame'].expr.expr
   end
 
   def inspect
@@ -126,7 +127,13 @@ class RubyVM::BareInstructions
       type: t, \
       location: [], \
       expr: v.to_s + ';'
-    return @attrs[k] = attr
+    return @attrs[k] ||= attr
+  end
+
+  def predefine_attributes
+    generate_attribute 'sp_inc', 'rb_snum_t', rets.size - pops.size
+    generate_attribute 'handles_frame', 'bool', \
+      opes.any? {|o| /CALL_INFO/ =~ o[:type] }
   end
 
   def typesplit a

tool/ruby_vm/views/_insn_entry.erb

@@ -28,8 +28,9 @@ INSN_ENTRY(<%= insn.name %>)
     <%= pop[:name] %> = <%= insn.cast_from_VALUE pop, "TOPN(#{i})"%>;
 % end
     DEBUG_ENTER_INSN(<%=cstr insn.name %>);
+% if insn.handles_frame?
     ADD_PC(<%= insn.width %>);
-    PREFETCH(GET_PC());
+% end
 % unless insn.pops.empty?
     POPN(<%= insn.pops.size %>);
 % end
@@ -44,7 +45,10 @@ INSN_ENTRY(<%= insn.name %>)
     PUSH(<%= insn.cast_to_VALUE ret %>);
 %   end
 % end
-%
+% unless insn.handles_frame?
+    ADD_PC(<%= insn.width %>);
+    PREFETCH(GET_PC());
+% end
     END_INSN(<%= insn.name %>);
 }
 #undef NAME_OF_CURRENT_INSN