[proxy/json]: optimisations for string escaping

[conradludgate]

I spent some time optimising the string routines. Much of the time in json encoding is spent on strings. We can skip a bunch of that with keys that are statically known to not need escaping. I've also spent some effort optimising the assembly of the string escaping. It's very marginal gains but it's gains nonetheless.

Before:

small                   time:   [6.3211 ns 6.3785 ns 6.4460 ns]

large_fmt               time:   [206.45 ns 208.20 ns 209.96 ns]

After:

small                   time:   [5.3861 ns 5.4242 ns 5.4664 ns]
                        change: [-14.858% -14.163% -13.512%] (p = 0.00 < 0.05)

small_static            time:   [2.9270 ns 2.9369 ns 2.9481 ns]

large_fmt               time:   [194.09 ns 196.21 ns 198.54 ns]
                        change: [-6.8999% -6.1206% -5.2425%] (p = 0.00 < 0.05)

Frankly, these micro-optimisations are unnecessary but they make me happy 😅

some statistics graphs

Small string

Large string with fmt

Small strings, const escaped vs dynamic escaped.

comparing the assembly for the hot loops

Before

https://godbolt.org/z/vsYqTzv4r

.LBB2_5:
        xor     r13d, r13d
.LBB2_6:
        lea     rax, [r14 + r13]
        cmp     rax, rbp
        je      .LBB2_11
        movzx   ebx, byte ptr [rax]
        movzx   r15d, byte ptr [rbx + rdx]
        inc     r13
        test    r15d, r15d
        je      .LBB2_6 ; MAIN LOOP, 8 INSTRUCTIONS
        lea     rsi, [rcx + r13]
        dec     rsi
        cmp     rsi, rcx
        jb      .LBB2_13
        cmp     rsi, r9
        ja      .LBB2_10
        lea     rdx, [r13 - 1]
        mov     rax, qword ptr [rdi]
        sub     rax, r12
        cmp     rdx, rax
        mov     qword ptr [rsp + 24], r8
        ja      .LBB2_15
.LBB2_16:
        ; 88 lines of inlined write_char_escape ommited for brevity
        jmp     .LBB2_5

After

https://godbolt.org/z/dsbz7M6WP

.LBB2_5:
        xor     edx, edx
.LBB2_6:
        cmp     r14, rdx
        je      .LBB2_7
        movzx   eax, byte ptr [rsi + rdx]
        inc     rdx
        cmp     byte ptr [rax + r15], 0
        je      .LBB2_6 ; MAIN LOOP, 6 INSTRUCTIONS
        lea     r12, [rsi + rdx]
        sub     r14, rdx
        mov     rdi, rbx
        call    example::write_char_escape::hc44aa4c48ff73fd1
        mov     rsi, r12
        jmp     .LBB2_5

I'm gonna attempt to push some of these changes back into serde_json as well.

[github-actions]

8855 tests run: 8192 passed, 0 failed, 663 skipped (full report)

---

Flaky tests (2)

Postgres 17

• test_ps_unavailable_after_delete[DeletionAPIKind.FORCE]: release-x86-64-with-lfc

Postgres 14

• test_ps_unavailable_after_delete[DeletionAPIKind.FORCE]: release-x86-64-with-lfc

Code coverage* (full report)

• functions: 35.1% (8837 of 25151 functions)
• lines: 46.4% (71122 of 153313 lines)

* collected from Rust tests only

---

_{The comment gets automatically updated with the latest test results
ce6bbca at 2025-07-18T10:35:03.965Z :recycle:}