Refactored working with built-in strings, symbols and small integers. #848

VadimZhestikov · 2025-02-06T21:00:43Z

Proposed changes

This is series of commits which introduce internal atomic representation for
strings, symbols and small numbers as keys for access object properties.

These commits are intended to speed up access to object properties.

Checklist

Before creating a PR, run through this checklist and mark each as complete:

I have read the CONTRIBUTING document
If applicable, I have added tests that prove my fix is effective or that my feature works
If applicable, I have checked that any relevant tests pass after adding my changes

xeioex · 2025-03-27T01:58:03Z

Hi @VadimZhestikov,

I will not comment on exact lines as there are too many changes.

In no particular order:

we do not need two atom tables, njs_atom_map.h can be removed as redundant
njs_flathsh_obj_t is the same as njs_flathsh_t with new special implementation for njs_flathsh_find() and njs_flathsh_insert().
in lexer atom_id can fully replace unique_id
vm->atom_hash_mem_pool and vm->atom_hash_atom_id_cell can be removed as well.

In addition, in order to improve the current patch the following function can be improved with direct uint32_t atom_id argument instead of njs_value_t *key: njs_property_query(), njs_value_property(), njs_value_property_set(), njs_value_property_delete()

See all the proposed improvements here:

"Improved atom_id use." should be applied on top of "Use atom_ids in the code." and eventually merged with it.
Two more improvements on top are: "Introduced NJS_VMCODE_PROPERTY_ATOM_SET instruction.", "Introduced NJS_VMCODE_PROPERTY_ATOM_GET instruction."

Benchmarks on my arm64:

MASTER
Richards: 603
Crypto: 1386
RayTrace: 618
NavierStokes: 2009
----
Score (version 7): 1009

VADIM INLINE
Richards: 971
Crypto: 1514
RayTrace: 674
NavierStokes: 2387
----
Score (version 7): 1240

XEIOEX
Richards: 984
Crypto: 1525
RayTrace: 804
NavierStokes: 2362
----
Score (version 7): 1299

VadimZhestikov · 2025-03-27T02:05:26Z

Good changes, It will take me few days to review them and update all together.

VadimZhestikov · 2025-04-11T17:02:57Z

Just intermediate changes (to be squashed later)

xeioex · 2025-04-22T00:28:32Z

@VadimZhestikov

Some small patches on top

From 7a4e97ae689fdcff5dc779edc8b208e45b0807a4 Mon Sep 17 00:00:00 2001
From: Dmitry Volyntsev <[email protected]>
Date: Mon, 21 Apr 2025 17:06:33 -0700
Subject: [PATCH] Fixed DefineOwnProperty/key-is-not-numeric-index-throws v3

---
 src/njs_object.h         | 8 +++-----
 src/njs_object_prop.c    | 2 +-
 src/test/njs_unit_test.c | 5 +++++
 3 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/src/njs_object.h b/src/njs_object.h
index 7bc03a9e6..4ece7e631 100644
--- a/src/njs_object.h
+++ b/src/njs_object.h
@@ -19,7 +19,7 @@ typedef enum {
     NJS_OBJECT_PROP_CONFIGURABLE = 16,
     NJS_OBJECT_PROP_WRITABLE = 32,
     NJS_OBJECT_PROP_UNSET = 64,
-    NJS_OBJECT_PROP_NOT_STRING = 128,
+    NJS_OBJECT_PROP_IS_STRING = 128,
 #define NJS_OBJECT_PROP_VALUE_ECW (NJS_OBJECT_PROP_VALUE                     \
                                    | NJS_OBJECT_PROP_ENUMERABLE              \
                                    | NJS_OBJECT_PROP_CONFIGURABLE            \
@@ -269,10 +269,8 @@ njs_object_prop_define_val(njs_vm_t *vm, njs_value_t *object, njs_value_t *name,
         }
     }
 
-    if (!njs_is_string(name)) {
-       flags |= NJS_OBJECT_PROP_NOT_STRING;
-    } else {
-       flags &= ~NJS_OBJECT_PROP_NOT_STRING;
+    if (njs_is_string(name)) {
+       flags |= NJS_OBJECT_PROP_IS_STRING;
     }
 
     return njs_object_prop_define(vm, object, name->atom_id, value, flags);
diff --git a/src/njs_object_prop.c b/src/njs_object_prop.c
index b72152727..4b9a1ec88 100644
--- a/src/njs_object_prop.c
+++ b/src/njs_object_prop.c
@@ -256,7 +256,7 @@ njs_object_prop_define(njs_vm_t *vm, njs_value_t *object, unsigned atom_id,
         }
 
         if (njs_slow_path(njs_is_typed_array(object) &&
-           !(flags & NJS_OBJECT_PROP_NOT_STRING)))
+           (flags & NJS_OBJECT_PROP_IS_STRING)))
         {
             /* Integer-Indexed Exotic Objects [[DefineOwnProperty]]. */
 
diff --git a/src/test/njs_unit_test.c b/src/test/njs_unit_test.c
index 5ee222d6b..b8155ac5a 100644
--- a/src/test/njs_unit_test.c
+++ b/src/test/njs_unit_test.c
@@ -5880,6 +5880,11 @@ static njs_unit_test_t  njs_test[] =
               "           return njs.dump(a) === `${v.name} [1,1,1]`})"),
       njs_str("true") },
 
+    { njs_str(NJS_TYPED_ARRAY_LIST
+              ".every(v=>{var a = new v([0]); var desc = Object.getOwnPropertyDescriptor(a, '0');"
+              "           try { Object.defineProperty(a, '1', desc) } catch (e) { return e.name == 'TypeError' }})"),
+      njs_str("true") },
+
     { njs_str(NJS_TYPED_ARRAY_LIST
               ".every(v=>{try {var a = new v([1,1]); Object.defineProperty(a, '1', {configurable:true})} "
               "           catch (e) { return e.message == 'Cannot redefine property: \"1\"'}})"),

From 923881004ea2ebd6566055c1a370ea3568132775 Mon Sep 17 00:00:00 2001
From: Dmitry Volyntsev <[email protected]>
Date: Mon, 21 Apr 2025 17:15:09 -0700
Subject: [PATCH] Added test for built-ins/Array/S15.4_A1.1_T10

---
 src/test/njs_unit_test.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/test/njs_unit_test.c b/src/test/njs_unit_test.c
index b8155ac5a..13c9e00c6 100644
--- a/src/test/njs_unit_test.c
+++ b/src/test/njs_unit_test.c
@@ -4172,6 +4172,11 @@ static njs_unit_test_t  njs_test[] =
     { njs_str("var a = [ 1, 2, 3 ]; a[4294967296] = 4; a + a[4294967296]"),
       njs_str("1,2,34") },
 
+    { njs_str("var x = []; var k = 1;"
+              "for (var i = 0; i < 32; i++) { k = k * 2; x[k - 2] = k; };"
+              "k = 1; for (i = 0; i < 32; i++) { k = k * 2; if (x[k - 2] != k) { throw 'error'; } }"),
+      njs_str("undefined") },
+
     { njs_str("delete[]['4e9']"),
       njs_str("true") },

I suggest the following resulting patches

Main patch:

Improved atom_id use.

Use atom_ids in the code.

 Fixed built-ins/TypedArrayConstructors/internals/DefineOwnProperty/key-is-not-numeric-index-throws v.2

Fixed built-ins/TypedArrayConstructors/internals/DefineOwnProperty/key-is-not-numeric-index-throws.

Fixed built-ins/Array/S15.4_A1.1_T10.

Separated namespaces for predefined STRING and SYMBOL names.

Fixed DefineOwnProperty/key-is-not-numeric-index-throws v3

Added test for built-ins/Array/S15.4_A1.1_T10

ATOM_GET patch:

Introduced NJS_VMCODE_PROPERTY_ATOM_GET instruction.

Put ATOM_GET changes from "Used more NJS_VMCODE_PROPERTY_ATOM_[SET|GET] instructions."

ATOM_SET patch:

Introduced NJS_VMCODE_PROPERTY_ATOM_SET instruction.

Put ATOM_SET changes from "Used more NJS_VMCODE_PROPERTY_ATOM_[SET|GET] instructions."

After the squash we will do the last cleanup/polish round with the main patch.

VadimZhestikov · 2025-04-22T03:30:52Z

It was just a test before big squash.

xeioex · 2025-04-25T06:05:27Z

Hi @VadimZhestikov,

See minor patches for the "Use atom_ids in the code." here:

- remove not used njs_flathsh_alloc_copy() - not used in the code.

- remove static from declarators. not relevant to atomics  - this change while useful is not needed for the patch. Feel free to prepare as a separate patch.

- making explicit token_type masks - trying to make njs_atom_defs magic number more readable.

- back njs_string_eq() - restored the original code, it seems performance is not impacted by this.

- style and diff minimization - style changes, and which makes the main patch shorter.

VadimZhestikov · 2025-04-29T18:10:07Z

Merged latest changes from @xeioex, both code and commit message.

xeioex

Otherwise looks good.

src/njs_value.c

- Implemented atom IDs for strings, symbols and small numbers, enabling equality checks via ID comparison - Optimized string operations for faster property lookups and comparisons - Removed short string inlining from njs_value_t structure Performance improvements (arewefastyet/benchmarks/v8-v7 benchmark): - Richards: +57% (631 → 989) - Crypto: +7% (1445 → 1551) - RayTrace: +37% (562 → 772) - NavierStokes: +20% (2062 → 2465) - Overall score: +29% (1014 → 1307) In collaboration with Dmitry Volyntsev.

fixed

xeioex

Looks good.

Congratulation!

VadimZhestikov changed the title ~~Atomic strings new~~ Atomic keys introduction for strings, symbols, small numbers. Feb 6, 2025

VadimZhestikov force-pushed the atomic_strings_new branch 13 times, most recently from 269870d to d0dc432 Compare February 14, 2025 22:46

VadimZhestikov force-pushed the atomic_strings_new branch from cc67468 to dd33c04 Compare February 20, 2025 02:06

VadimZhestikov force-pushed the atomic_strings_new branch from dd33c04 to 1430356 Compare February 28, 2025 04:22

VadimZhestikov requested a review from xeioex March 3, 2025 17:51

xeioex mentioned this pull request Mar 19, 2025

Build failure with GCC 15 #857

Closed

VadimZhestikov force-pushed the atomic_strings_new branch from 1430356 to b04a54c Compare April 11, 2025 17:02

xeioex mentioned this pull request Apr 12, 2025

Tag some char arrays with __attribute__((__nonstring__)) #871

Closed

VadimZhestikov force-pushed the atomic_strings_new branch from b04a54c to ad70a7f Compare April 14, 2025 15:35

VadimZhestikov force-pushed the atomic_strings_new branch from ad70a7f to cf46a01 Compare April 22, 2025 03:22

VadimZhestikov force-pushed the atomic_strings_new branch from cf46a01 to d658657 Compare April 22, 2025 04:20

VadimZhestikov force-pushed the atomic_strings_new branch from d658657 to 2cdc940 Compare April 29, 2025 18:07

VadimZhestikov force-pushed the atomic_strings_new branch from 2cdc940 to f2a9f5c Compare April 29, 2025 21:46

xeioex previously requested changes Apr 30, 2025

View reviewed changes

src/njs_value.c Outdated Show resolved Hide resolved

VadimZhestikov and others added 3 commits April 29, 2025 18:01

Introduced NJS_VMCODE_PROPERTY_ATOM_GET instruction.

cb4e515

Introduced NJS_VMCODE_PROPERTY_ATOM_SET instruction.

a5cf037

VadimZhestikov force-pushed the atomic_strings_new branch from f2a9f5c to a5cf037 Compare April 30, 2025 01:02

VadimZhestikov requested a review from xeioex April 30, 2025 01:02

VadimZhestikov changed the title ~~Atomic keys introduction for strings, symbols, small numbers.~~ Refactored working with built-in strings, symbols and small integers. Apr 30, 2025

xeioex approved these changes Apr 30, 2025

View reviewed changes

VadimZhestikov merged commit 2d4788c into nginx:master Apr 30, 2025
1 check passed

VadimZhestikov mentioned this pull request Apr 30, 2025

Fixed regexp undefined value of captured group. #900

Merged

xeioex added this to the 0.9.0 milestone May 6, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Refactored working with built-in strings, symbols and small integers. #848

Refactored working with built-in strings, symbols and small integers. #848

Uh oh!

VadimZhestikov commented Feb 6, 2025 •

edited

Loading

Uh oh!

xeioex commented Mar 27, 2025 •

edited

Loading

Uh oh!

VadimZhestikov commented Mar 27, 2025

Uh oh!

VadimZhestikov commented Apr 11, 2025 •

edited

Loading

Uh oh!

xeioex commented Apr 22, 2025

Uh oh!

VadimZhestikov commented Apr 22, 2025

Uh oh!

xeioex commented Apr 25, 2025

Uh oh!

VadimZhestikov commented Apr 29, 2025

Uh oh!

xeioex left a comment

Uh oh!

Uh oh!

xeioex left a comment

Uh oh!

Uh oh!

Uh oh!

Refactored working with built-in strings, symbols and small integers. #848

Refactored working with built-in strings, symbols and small integers. #848

Uh oh!

Conversation

VadimZhestikov commented Feb 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Proposed changes

Checklist

Uh oh!

xeioex commented Mar 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

VadimZhestikov commented Mar 27, 2025

Uh oh!

VadimZhestikov commented Apr 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xeioex commented Apr 22, 2025

Uh oh!

VadimZhestikov commented Apr 22, 2025

Uh oh!

xeioex commented Apr 25, 2025

Uh oh!

VadimZhestikov commented Apr 29, 2025

Uh oh!

xeioex left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

xeioex left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

VadimZhestikov commented Feb 6, 2025 •

edited

Loading

xeioex commented Mar 27, 2025 •

edited

Loading

VadimZhestikov commented Apr 11, 2025 •

edited

Loading