There's actually a couple of things going on here.
First of all, this is actually a great example of why we always tell people
that we can't "just" optimize functions ahead of time. You may notice that
"add" is in the --trace-opt twice; it's optimized exactly when you request
it, but then as soon as that optimized code is executed, it immediately
deopts due to insufficient type feedback for the add operation (this is
called a "soft" deopt). You'll see the deopt/bailout if you run with
--trace-deopt. Then, later, it's optimised again from being run lots of
times, which is a concurrent compilation that may or may not complete
before your loop ends. To get "add" properly optimized using
OptimizeFunctionOnNextCall, you have to
1. Call %PrepareFunctionForOptimization(add) -- this makes sure that the
add function can collect type feedback and that its bytecode is not flushed
away by GC
2. Call add(1,2) -- this is a call with the correct types, two small
integers, which teach it the correct type feedback for the sum
3. Call %OptimizeFunctionOnNextCall(add)
Secondly, this is a bit of a printing failure -- the "code" you're seeing
is on the JSFunction as you'd expect, but the "no optimized code" is
actually printing as part of printing the "feedback vector" of the
JSFunction. The reason is that there is only one closure (JSFunction) in
one context that implements "add", which means that we can do some extra
optimisations called "function context specialization"; this is the common
case, but in some cases you may have JSFunctions implementing the same
underlying code (e.g. function factory() { return function add(a+b) {
return a+b; } } always returns a new JSFunction but one that implements the
same code). For those cases, we skip the "function context specialization",
and share code between these different JSFunctions by putting it on the
feedback vector.
Putting this together, you'll see what you originally expected by running:
function factory() {
return function add(a, b) { return a + b; }
}
// Create two closures for 'add', to turn off function context
specialization.
add = factory();
some_other_add = factory();
%PrepareFunctionForOptimization(add);
// Warm up the type feedback.
add(1, 2);
%OptimizeFunctionOnNextCall(add);
// Do the optimization.
add(1, 2);
// Run the loop.
let sum = 0;
for (let i = 0; i < 100000; i++) {
sum += add(i, i);
}
%DebugPrint(add);
- Leszek
On Tuesday, December 7, 2021 at 5:45:51 AM UTC+1 zin...@gmail.com wrote:
> Hi all,
>
> I'm trying to mannually optimize my function [1] by
> calling %OptimizeFunctionOnNextCall.
>
> I run `./d8 --allow-natives-syntax --trace-opt --print-opt-code foo.js`,
> it seems that the function is compiled:
>
> [manually marking 0x1e6e082935b5 <JSFunction add (sfi = 0x1e6e08293485)>
> for non-concurrent optimization]
> [compiling method 0x1e6e082935b5 <JSFunction add (sfi = 0x1e6e08293485)>
> (target TURBOFAN) using TurboFan]
> --- Optimized code ---
> optimization_id = 0
> source_position = 12
> kind = TURBOFAN
> name = add
> stack_slots = 6
> compiler = turbofan
> address = 0x1e6e00044001
> ...
>
> But according to `%DebugPrint(add)` in the last line, it says:
>
> - code: 0x1e6e00045841 <Code TURBOFAN>
> ...
> - shared function info: 0x1e6e08293485 <SharedFunctionInfo add>
> - no optimized code
> - optimization marker: OptimizationMarker::kNone
> - optimization tier: OptimizationTier::kNone
>
> I'm curious why it has `<Code TURBOFAN>` but `no optimized code` and
> `OptimizationMarker::kNone`? Shouldn't the function be compiled with
> `OptimizationMarker::kCompileOptimized`? Is there something I'm missing?
> Thanks in advance!
>
> Regards,
> Jiading
>
> [1] add.js:
> function add(a, b) { return a + b; }
> %OptimizeFunctionOnNextCall(add);
> add(1, 2);
> let sum = 0;
> for (let i = 0; i < 100000; i++) {
> sum += add(i, i);
> }
> %DebugPrint(add);
>
--
--
v8-users mailing list
v8-users@googlegroups.com
http://groups.google.com/group/v8-users
---
You received this message because you are subscribed to the Google Groups
"v8-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to v8-users+unsubscr...@googlegroups.com.
To view this discussion on the web visit
https://groups.google.com/d/msgid/v8-users/e23b2218-3880-46f4-810c-6348d3d562acn%40googlegroups.com.
