Pointer-bump allocations are fast in isolation, but if your code is doing one it means the compiler has gotten itself confused, and every time it gets confused it makes it harder for later code to be optimized. Compilers only really reason reasonably about value types (and even then you should never assume they're smart[1]).

That said, I don't actually know allocation is the problem; it's a suspect, but unlike a native script I can't just run `perf` to get performance counters and assembly listings. All I really know is that the non-allocating order is a lot faster.

[1]: If you can read assembly, this one's hilarious: https://godbolt.org/g/81qWoq

>[1]: If you can read assembly, this one's hilarious: https://godbolt.org/g/81qWoq

I only barely read assembly, so I was just trying to figure this one out.

So, the compiler has figured out that it just needs to test v against 0 (since the array is immutable), but it does this 2 times, and with a couple of apparently redundant 'je' instructions. Is that what's weird about the output?

In case it changed, this is the output I saw:

  example::exists_in_table:
    push rbp
    mov rbp, rsp
    mov al, 1
    test edi, edi
    je .LBB0_5
    je .LBB0_5
    test edi, edi
    je .LBB0_5
    je .LBB0_5
    xor eax, eax
  .LBB0_5:
    pop rbp
    ret

Basically, yes. The first test and je show the compiler has managed the hard part, but the next je obviously can't be hit, the test after that literally doesn't do anything, and the two jumps after that also can't be hit. It's funny because this is all so trivial.

If I write it as some C

    bool should_jump = edi == 0;
    if (should_jump) goto LBB0_5;
    if (should_jump) goto LBB0_5;
    bool should_jump = edi == 0;
    if (should_jump) goto LBB0_5;
    if (should_jump) goto LBB0_5;

I can guarantee the compiler is going to produce good code from it.

This looks like yet-another-phase-ordering-issue in LLVM. I hit so many of these it's not even funny. Arguably one could find C code that shows the exact problem under clang. Example of optimized LLVM IR for that sample (press LLVM IR on https://play.rust-lang.org/?gist=b60ec85886eff967b21c7abc899...):

    br i1 false, label %bb9, label %bb6

SimplifyCFG would instantly kill that. But it doesn't get to run again after whatever simplified the branch condition. I've heard rumors that LLVM is working on a new pass manager, maybe they will eventually fix this pervasive issue.

Apologies for prodding a sore spot! The phase ordering problem is a well known demon for compilers; makes me wonder why I've not heard anything about unphased compilation.

No worries, I'm just surprised you managed to find a simple testcase that exhibits it, so, really, thank you!

I've opened a (a bit vague for now) issue so at least I don't forget about it: https://github.com/rust-lang/rust/issues/44041

As for phase ordering - someone did show me recently VSDG, which claims to solve at least part of the problem (specifically, https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-705.pdf and http://sro.sussex.ac.uk/7576/1/Stanier%2C_James.pdf).

Awesome references, seems I have some fun reading :). (Trivia: Alan Mycroft actually supervised me for my undergraduate dissertation.)

You can at least use the browser profiler tools.

Browser profiling tools are great but their granularity stops at the function level - they can't tell you anything about what happens within a given function, just how that function compares to others.

I haven't used them much, specially since I tend to choose native projects if possible.

Thanks for the heads up.

I do! It's missing a lot of detail, though; it's optimized much more for overall performance analysis.

> Allocation in modern generational GCs is in a completely different performance category than C-style malloc is.

Yes and no. If enough allocation happens that a GC pass occurs while you're still running your JS hot loop you're pretty much screwed and all your performance vanishes. If you want to see this in the extreme go open an Android systrace or Chrome chrome://tracing trace result and do a JS profile while you do so. The single largest consumer of JS CPU time is the GC at nearly 13%.

The problem with modern generation GC langauges is they tend to be designed assuming that allocation is always cheap/free and it just isn't. That's a fairytale. It's just not true.

Also just because it only takes 6-13ns for an allocation to be given to you doesn't mean anything when the first access of that allocation is a cache line miss compared to if you did the same thing in C and it was on the stack, which is likely sitting in hot L1 primed and ready to go. Reusing the same memory address range is critical to getting maximum performance, and modern generational GCs are miserably bad at that.

> Yes and no. If enough allocation happens that a GC pass occurs while you're still running your JS hot loop you're pretty much screwed and all your performance vanishes.

GCs are expensive relative to not doing anything, but minor GCs are very fast.

> The problem with modern generation GC langauges is they tend to be designed assuming that allocation is always cheap/free and it just isn't. That's a fairytale. It's just not true.

I'm not making an argument about language design here. Obviously value types are a good thing.

> Reusing the same memory address range is critical to getting maximum performance, and modern generational GCs are miserably bad at that.

Not at all. The nursery is usually in cache, just as the stack is. Nurseries are usually implemented as two space copying collectors with small spaces, which are excellent for cache locality (arguably even better than stacks, because stacks can grow deep).

> Not at all. The nursery is usually in cache, just as the stack is. Nurseries are usually implemented as two space copying collectors with small spaces, which are excellent for cache locality (arguably even better than stacks, because stacks can grow deep).

I think you're still working under the assumption that the GC is keeping up with the allocation rate, which if you have a small allocation in a hot loop will largely not be true. Typically GC'd languages rely on escape analysis to handle this and not the generational GC at all, but if that fails then you're SOL because the generational GC is unable to keep up and unable to keep things in the fast path.

> Typically GC'd languages rely on escape analysis to handle this and not the generational GC at all

No, they don't. Most GC'd languages rely on generational GC, because escape analysis on its own doesn't directly provide a lot of performance benefits once you have a generational GC.

> if that fails then you're SOL because the generational GC is unable to keep up and unable to keep things in the fast path.

No, you aren't. Cleaning up dead objects (for example, temporaries created in a hot loop) in a nursery is extremely fast. The entire nursery semispace is typically in L1.

    The entire nursery semispace is typically in L1.

L1 is only going to be about 64k per core (eg, Kaby Lake). That'll have your stack, your working set, and a tiny bit of your nursery in it... as long as you don't get associativity problems.

If you read a bunch of data before making an object, I could see you easily evicting your nursery from L1 into L2, but then you only have about 4x as much space.

Not GP, but as someone who has worked with image manipulation and typed arrays in JS (which seems similar to what the GP was doing) I can say that this often boils down to linearly looping through a typed array and doing very simple operations. You know, one of those "best case" scenarios for cpu utilisation where memory access tends to be the bottleneck. If the allocation happens inside a tight loop like that I can imagine it ruining performance