JIT optimization: loop head alignment

[JosephTremoulet]

(I'm creating tracking issues for some optimizations that RyuJit doesn't perform, so we'll have a place to reference/note when we see the lack of them affecting particular benchmarks)

This one actually has an implementation plumbed through, but it's triggered only by the unsupported COMPlus_JitAlignLoops flag. It should probably be enabled by default in at least some configurations/methods.

category:cq
theme:loop-opt
skill-level:expert
cost:medium

[JosephTremoulet]

This came up in #4400

[AndyAyersMS]

Eugene has some evidence that 32 byte alignment matters for hot loops on recent CPUs. More crucial when loop bodies are small.

We'd also, as a prerequisite, need to align method entries. Right now for x64 we always use 8 byte alignment (while somewhat oddly, on x86, we do things differently).

R2R images don't honor alignment requests (fragile NIs do though).

I would like to to collect some distribution data on native code sizes so we could guess at likely costs for different method alignment algorithms. Probably need SPMI working in Core or could try it over in desktop. Benefit would be harder to assess.

So, lots to sort through here.

[adamsitnik]

it's triggered only by the unsupported COMPlus_JitAlignLoops

what do you mean by unsupported? Even if I set COMPlus_JitAlignLoops env var it is not going to work?

[AndyAyersMS]

IIRC setting the env var will do something, but is not guaranteed to do what you'd expect.

[adamsitnik]

@AndyAyersMS thanks!

btw if you would like to test the performance of CoreCLR with some flag enabled/disabled, then you can use following config to run defeault settings vs flag configured:

static void Main(string[] args)
    => BenchmarkRunner.Run<Benchmarks>(
        DefaultConfig.Instance
            .With(DisassemblyDiagnoser.Create(new DisassemblyDiagnoserConfig(printAsm: true, printPrologAndEpilog: true)))
            .With(Job.Default.With(Runtime.Core).AsBaseline())
            .With(Job.Default.With(Runtime.Core).With(new EnvironmentVariable[1] { new EnvironmentVariable("COMPlus_JitAlignLoops", "1") }))
        );

[voinokin]

I wonder whether it will possible some time in the future to allow loop head alignment on per-method basis, eg. with [MethodImpl(MethodImplOptions.AlignLoopHeads)]
In my hi-perf sorting app there are lots of different kinds of loops and I'd like to fine-tune some methods to squeeze as much additional perf as possible out of it without affecting some others.
According to Intel and Agner Fog's manuals the loop alignment affects behavior related to branch predition and uops cache.

[kunalspathak]

This has been enabled for x86/x64 by #44370.