Port closure optimizations

[saghul]

Ref: bellard/quickjs@ae7219b

[bnoordhuis]

Nice work.

[bnoordhuis]

This field creates a 6 byte padding hole on 64 bits architectures because of the alignment requirement of the next field. Can maybe be eliminated with some reshuffling.

[saghul]

I'll take a look!

[saghul]

I did some AI assisted suffling of fields :-) Can you PTAL?

[saghul]

FWIW, got this from struct_layout.py:

struct ::JSFunctionBytecode [136 Bytes]
     : [JSGCObjectHeader : 24] header
    0:   [int : 4] ref_count                              -- {cache-line 0}
   --- 1 Bytes padding ---
    5:   [unsigned char : 1] dummy1                       -- {cache-line 0}
    6:   [unsigned short : 2] dummy2                      -- {cache-line 0}
     :   [list_head : 16] link
    8:     [list_head* : 8] prev                          -- {cache-line 0}
   16:     [list_head* : 8] next                          -- {cache-line 0}
   --- 8 Bytes padding ---
   32: [unsigned char* : 8] byte_code_buf                 -- {cache-line 0}
   40: [int : 4] byte_code_len                            -- {cache-line 0}
   44: [unsigned int : 4] func_name                       -- {cache-line 0}
   48: [JSVarDef* : 8] vardefs                            -- {cache-line 0}
   56: [JSClosureVar* : 8] closure_var                    -- {cache-line 0}
   64: [unsigned short : 2] arg_count                     -- {cache-line 1}
   66: [unsigned short : 2] var_count                     -- {cache-line 1}
   68: [unsigned short : 2] defined_arg_count             -- {cache-line 1}
   70: [unsigned short : 2] stack_size                    -- {cache-line 1}
   72: [unsigned short : 2] var_ref_count                 -- {cache-line 1}
   74: [unsigned short : 2] closure_var_count             -- {cache-line 1}
   76: [int : 4] cpool_count                              -- {cache-line 1}
   80: [JSContext* : 8] realm                             -- {cache-line 1}
   88: [JSValue* : 8] cpool                               -- {cache-line 1}
   96: [unsigned int : 4] filename                        -- {cache-line 1}
  100: [int : 4] line_num                                 -- {cache-line 1}
  104: [int : 4] col_num                                  -- {cache-line 1}
  108: [int : 4] source_len                               -- {cache-line 1}
  112: [int : 4] pc2line_len                              -- {cache-line 1}
   --- 4 Bytes padding ---
  120: [unsigned char* : 8] pc2line_buf                   -- {cache-line 1}
  128: [char* : 8] source                                 -- {cache-line 2}

Maybe it can be further improved to avoid the 8 byte hole?

[bnoordhuis]

Is that on macos? JSGCObjectHeader size and alignment on my x86_64 Linux system are both 24, not 24 and 32 as seems to be the case on your system. Maybe __attribute__((packed)) helps but that's not entirely without peril.

[saghul]

Yep, that is macOS, aarch64.

[gengjiawen]

seems score very near

Benchmark (Higher scores are better)	QuickJS-ng (local build)	QuickJS Bellard (local build)	V8 --jitless
Richards	840	832	1157
DeltaBlue	823	832	1085
Crypto	746	715	1302
RayTrace	958	1231	4423
EarleyBoyer	1670	1726	6115
RegExp	213	340	4209
Splay	2505	2501	8345
NavierStokes	1157	1264	1995
Score	919	1015	2731