Feature #9171: [patch] use fstrings for symbol table

• Status: Closed
• Priority: Normal
• Assignee: nobu (Nobuyoshi Nakada)
• Target version: -

[tmm1 (Aman Karmani)]

Here is a simple patch to use fstrings for the table backing symbols.

Unfortunately it causes a segfault in test/rdoc/test_rdoc_parser_ruby.rb. Maybe someone wants to investigate.

diff --git a/parse.y b/parse.y
index 8207ad7..b1f3112 100644
--- a/parse.y
+++ b/parse.y
@@ -10333,7 +10333,7 @@ register_symid(ID id, const char *name, long len, rb_encoding *enc)
 static ID
 register_symid_str(ID id, VALUE str)
 {
-    OBJ_FREEZE(str);
+    str = rb_fstring(str);
 
     if (RUBY_DTRACE_SYMBOL_CREATE_ENABLED()) {
 	RUBY_DTRACE_SYMBOL_CREATE(RSTRING_PTR(str), rb_sourcefile(), rb_sourceline());
diff --git a/string.c b/string.c
index 231bb2f..6ae33e3 100644
--- a/string.c
+++ b/string.c
@@ -138,6 +138,9 @@ rb_fstring(VALUE str)
     st_data_t fstr;
     Check_Type(str, T_STRING);
 
+    if (!frozen_strings)
+	frozen_strings = st_init_table(&fstring_hash_type);
+
     if (FL_TEST(str, RSTRING_FSTR))
 	return str;
 
@@ -8707,8 +8710,6 @@ Init_String(void)
 #undef rb_intern
 #define rb_intern(str) rb_intern_const(str)
 
-    frozen_strings = st_init_table(&fstring_hash_type);
-
     rb_cString  = rb_define_class("String", rb_cObject);
     rb_include_module(rb_cString, rb_mComparable);
     rb_define_alloc_func(rb_cString, empty_str_alloc);

[hsbt (Hiroshi SHIBATA)]

• Assignee set to ko1 (Koichi Sasada)

ko1: Could you review this?

[ko1 (Koichi Sasada)]

• Category set to core
• Assignee changed from ko1 (Koichi Sasada) to nobu (Nobuyoshi Nakada)

Nobu, could you check it?

[avit (Andrew Vit)]

Would this allow the symbol table to be GC'd or is that a separate issue?

[nobu (Nobuyoshi Nakada)]

different.

[nobu (Nobuyoshi Nakada)]

Seems my reply hasn't caught by the ITS...

(13/11/28 16:55), tmm1 (Aman Gupta) wrote:

diff --git a/parse.y b/parse.y
index 8207ad7..b1f3112 100644
--- a/parse.y
+++ b/parse.y
@@ -10333,7 +10333,7 @@ register_symid(ID id, const char *name, long len, rb_encoding *enc)
 static ID
 register_symid_str(ID id, VALUE str)
 {
-    OBJ_FREEZE(str);
+    str = rb_fstring(str);

At this point, rb_cString is not created yet.
And unfrozen string will be duplicated in rb_fstring().

diff --git i/parse.y w/parse.y
index 8207ad7..9f580e6 100644
--- i/parse.y
+++ w/parse.y
@@ -10334,6 +10334,7 @@ static ID
 register_symid_str(ID id, VALUE str)
 {
     OBJ_FREEZE(str);
+    str = rb_fstring(str);
 
     if (RUBY_DTRACE_SYMBOL_CREATE_ENABLED()) {
 	RUBY_DTRACE_SYMBOL_CREATE(RSTRING_PTR(str), rb_sourcefile(), rb_sourceline());
diff --git i/string.c w/string.c
index 151170c..ec43af7 100644
--- i/string.c
+++ w/string.c
@@ -165,6 +165,9 @@ rb_fstring(VALUE str)
     VALUE fstr = Qnil;
     Check_Type(str, T_STRING);
 
+    if (!frozen_strings)
+	frozen_strings = st_init_table(&fstring_hash_type);
+
     if (FL_TEST(str, RSTRING_FSTR))
 	return str;
 
@@ -173,6 +176,13 @@ rb_fstring(VALUE str)
 }
 
 static int
+fstring_set_class_i(st_data_t key, st_data_t val, st_data_t arg)
+{
+    RBASIC_SET_CLASS((VALUE)key, (VALUE)arg);
+    return ST_CONTINUE;
+}
+
+static int
 fstring_cmp(VALUE a, VALUE b)
 {
     int cmp = rb_str_hash_cmp(a, b);
@@ -8718,8 +8728,6 @@ Init_String(void)
 #undef rb_intern
 #define rb_intern(str) rb_intern_const(str)
 
-    frozen_strings = st_init_table(&fstring_hash_type);
-
     rb_cString  = rb_define_class("String", rb_cObject);
     rb_include_module(rb_cString, rb_mComparable);
     rb_define_alloc_func(rb_cString, empty_str_alloc);
@@ -8891,4 +8899,6 @@ Init_String(void)
     rb_define_method(rb_cSymbol, "swapcase", sym_swapcase, 0);
 
     rb_define_method(rb_cSymbol, "encoding", sym_encoding, 0);
+
+    st_foreach(frozen_strings, fstring_set_class_i, rb_cString);
 }

[tmm1 (Aman Karmani)]

Ah great! Thanks for finding my bug.

With your patch, long-lived strings on our rails app heap are reduced by 6000. I think we should merge it.

[tmm1 (Aman Karmani)]

• Status changed from Open to Closed
• % Done changed from 0 to 100

This issue was solved with changeset r44057.
Aman, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.

---

parse.y: use rb_fstring() for strings stored in the symbol table

• parse.y (register_symid_str): use fstrings in symbol table
  [Bug #9171] [ruby-core:58656]
• parse.y (rb_id2str): ditto
• string.c (rb_fstring): create frozen_strings on first usage. this
  allows rb_fstring() calls from the parser (before cString is created)
• string.c (fstring_set_class_i): set klass on fstrings generated
  before cString was defined
• string.c (Init_String): convert frozen_strings table to String
  objects after boot
• ext/-test-/symbol/type.c (bug_sym_id2str): expose rb_id2str()
• test/-ext-/symbol/test_type.rb (module Test_Symbol): verify symbol
  table entries are fstrings

[nobu (Nobuyoshi Nakada)]

• Tracker changed from Bug to Feature
• Description updated (diff)