1582772 - Use branchSub32 to optimize some generated loops

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, task)

Description

[Jan de Mooij [:jandem]]

I was telling Benjamin about branchSub32 and that made me wonder about some loops in Baseline Interpreter and IC code where we count down to zero, like this:

L0:
  test reg, reg
  jz L1
  ...
  sub 1, reg
  jmp L0
L1:

With branchSub32 we could do this a bit more efficiently:

  test reg, reg
  jz L1
L0:
  ...
  sub 1, reg
  jnz L0
L1:

It's what C++ compilers typically do. It executes fewer instructions but it also allows CPUs to do macro-op fusion: Intel CPUs at least can fuse the sub/jnz instructions.

The perf difference is pretty measurable, on my MBP with --no-baseline I get the following for the contrived micro-benchmark below (this tests the blinterp emitInitializeLocals loop):

before: 85 ms
after:  68 ms

var s = "";
for (var i = 0; i < 100; i++) {
    s += "var x" + i + ";";
}
function f() {
    var g = Function(s);
    var t = new Date;
    for (var i = 0; i < 1000000; i++) {
        g();
    }
    print(new Date - t);
}
f();

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Bug 1582772 - Use branchSub32 to micro-optimize some loops in generated code. r?iain!

This is a bit more efficient. Bug 1582772 comment 0 has more data.

Comment 2

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/4d81f578aa35
Use branchSub32 to micro-optimize some loops in generated code. r=iain

Comment 3

[Jan de Mooij [:jandem]]

Hm pushCallArguments can be optimized and simplified a bit more. NI myself to fix next week.

Comment 4

[Daniel Varga [:dvarga]]

https://hg.mozilla.org/mozilla-central/rev/4d81f578aa35

Comment 5

[Jan de Mooij [:jandem]]

(In reply to Jan de Mooij [:jandem] from comment #3)

Hm pushCallArguments can be optimized and simplified a bit more. NI myself to fix next week.

Bug 1583104.