1734405 - Compare small constant strings using inline assembly

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement)

Description

[André Bargull [:anba]]

The current string comparison code is always using a VM-call to determine if two strings are equal to each other (after pointer, atom, and length checks), cf. MacroAssembler::compareStrings(). We should improve the case when one operand is a compile-time constant string, similar to how Clang and GCC are optimising:

bool test(const char* s) {
    return memcmp(s, "spidermonkey", 12);
}

This optimisation helps to remove a few million VM-calls in benchmarks (Speedometer2, JetStream2, ARES6) and it also applies to real-world code, for example when scrolling down on https://www.twitch.tv/directory, whenever data is reloaded, 20K VM-calls are performed alone for string comparisons.

This bug is different from bug 1326193 insofar that only small constant strings will be optimised.

Comment 1

[Jan de Mooij [:jandem]]

Nice, I've definitely seen this show up when looking at VM call counts (bug 1730863).

Comment 2

[André Bargull [:anba]]

Attachment: Bug 1734405 - Part 1: Fold comparison with empty string. r=jandem!

This optimises str == "" comparison which are now rewritten to str.length == 0.

The change results in less generated code and is also slightly faster.

Comment 3

[André Bargull [:anba]]

Attachment: Bug 1734405 - Part 2: Add MacroAssembler::cmp{8,16,64}Set(). r=jandem!

Add cmp{8,16,64}() in preparation for a later patch in this stack.

Depends on D127665

Comment 4

[André Bargull [:anba]]

Attachment: Bug 1734405 - Part 3: Add MacroAssembler::branch{8,16}(). r=jandem!

Add branch{8,16}() in preparation for a later patch in this stack.

Depends on D127666

Comment 5

[André Bargull [:anba]]

Attachment: Bug 1734405 - Part 4: Compare small constant strings using inline assembly. r=jandem!

Add an optimisation to compare small constant strings using inline assembly code.
For example x == "spidermonkey" is now optimised to:

; rcx: string
; rax: output

; Compare the first eight characters "spidermo".
; 0x73 = s, 0x70 = p, etc.
    cmovzq 0x08(%rcx), %rax
    mov $0x6F6D726564697073, %r11
    cmpq %r11, (%rax)
    jnz <not-equal>

; Compare the next four characters "nkey".
    cmpl $0x79656B6E, 0x08(%rax)
    jnz <not-equal>

; Return true when both strings are equal.
    mov $0x01, %eax
    jmp <done>

; Return false when the strings aren't equal.
not-equal:
    xor %eax, %eax

done:

The current maximum for inline string comparison is 32 Latin-1 characters. This
number was chosen because it's a multiple of eight and covers most constant
strings observed while randomly browsing around. This maximum means we generate
at most four comparison instructions on 64-bit systems.

Depends on D127667

Comment 6

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/a6aebcfcdc67
Part 1: Fold comparison with empty string. r=jandem
https://hg.mozilla.org/integration/autoland/rev/f3a3a2b4de50
Part 2: Add MacroAssembler::cmp{8,16,64}Set(). r=jandem
https://hg.mozilla.org/integration/autoland/rev/fac49dce2402
Part 3: Add MacroAssembler::branch{8,16}(). r=jandem
https://hg.mozilla.org/integration/autoland/rev/14f5245a7859
Part 4: Compare small constant strings using inline assembly. r=jandem

Comment 7

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/a6aebcfcdc67
https://hg.mozilla.org/mozilla-central/rev/f3a3a2b4de50
https://hg.mozilla.org/mozilla-central/rev/fac49dce2402
https://hg.mozilla.org/mozilla-central/rev/14f5245a7859