JIT: Teach middle-end liveness to DCE dead calls without sideeffects

[jakobbotsch]

Currently only the backend liveness pass will DCE calls without side effects when their result is unused. However, unused runtime lookups can end up expanded to control flow that we will be unable to DCE. This teaches the middle-end liveness pass to DCE dead calls as well.

Fix #106129

[jakobbotsch]

@EgorBot -amd -arm64

using System.Net;
using BenchmarkDotNet.Attributes;

namespace genericlist
{
    public sealed class BinaryTrie
    {
        [System.Runtime.CompilerServices.InlineArray(2)]
        public struct TwoBranches<TNodeLeaf>
        {
            private TNodeLeaf? _ref0;
        }
        public class Node<TNodeLeaf>
        {
            public TwoBranches<Node<TNodeLeaf>> Branches;
            public Node<TNodeLeaf> n0;
            public Node<TNodeLeaf> n1;
            public bool IsLeaf;
            public TNodeLeaf? NodeLeaf;
        }

        public static TLeaf? LookupInlineArray<TLeaf>(Node<TLeaf> root)
        {
            Node<TLeaf> currentNode = root;
            TLeaf? result = default;
            for (int i = 0; i < 8; i++)
            {
                int bit = i % 2;
                Node<TLeaf>? branch = currentNode.Branches[bit];
                if (branch is not null)
                {
                    return result;
                }
            }

            return result;
        }

        public static TLeaf? LookupTernary<TLeaf>(Node<TLeaf> root)
        {
            Node<TLeaf> currentNode = root;
            TLeaf? result = default;
            for (int i = 0; i < 8; i++)
            {
                int bit = i % 2;
                Node<TLeaf>? branch = bit == 0 ? currentNode.n0 : currentNode.n1;
                if (branch is not null)
                {
                    return result;
                }
            }

            return result;
        }
    }

    [DisassemblyDiagnoser]
    public class Benchmarks
    {
        BinaryTrie.Node<bool> _bool;
        BinaryTrie.Node<IPAddress> _ipaddress;

        [GlobalSetup]
        public void Setup()
        {
            _bool = new BinaryTrie.Node<bool>();
            _ipaddress = new BinaryTrie.Node<IPAddress>();
        }

        [Benchmark]
        public bool LookupBoolInlineArray()
        {
            return BinaryTrie.LookupInlineArray(_bool);
        }

        [Benchmark]
        public bool LookupBoolTernary()
        {
            return BinaryTrie.LookupTernary(_bool);
        }

        [Benchmark(Baseline = true)]
        public IPAddress LookupIPAddressInlineArray()
        {
            return BinaryTrie.LookupInlineArray(_ipaddress);
        }

        [Benchmark]
        public IPAddress LookupIPAddressTernary()
        {
            return BinaryTrie.LookupTernary(_ipaddress);
        }
    }
}

[EgorBot]

Benchmark results on Amd

BenchmarkDotNet v0.14.0, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
AMD EPYC 7763, 1 CPU, 8 logical and 4 physical cores
  Job-JTTXUS : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2
  Job-LXMITZ : .NET 9.0.0 (42.42.42.42424), X64 RyuJIT AVX2

Method	Toolchain	Mean	Error	Ratio	Code Size
LookupBoolInlineArray	Main	5.115 ns	0.0874 ns	0.63	10 B
LookupBoolTernary	Main	6.980 ns	0.1289 ns	0.86	10 B
LookupIPAddressInlineArray	Main	8.143 ns	0.0133 ns	1.00	20 B
LookupIPAddressTernary	Main	7.487 ns	0.2102 ns	0.92	20 B
LookupBoolInlineArray	PR	4.967 ns	0.1312 ns	0.61	10 B
LookupBoolTernary	PR	6.749 ns	0.1179 ns	0.83	10 B
LookupIPAddressInlineArray	PR	5.354 ns	0.1748 ns	0.66	20 B
LookupIPAddressTernary	PR	7.517 ns	0.1750 ns	0.92	20 B

BDN_Artifacts.zip

[EgorBot]

Benchmark results on Arm64

BenchmarkDotNet v0.14.0, Ubuntu 22.04.4 LTS (Jammy Jellyfish)
Unknown processor
  Job-QAGPZC : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD
  Job-DXQMVU : .NET 9.0.0 (42.42.42.42424), Arm64 RyuJIT AdvSIMD

Method	Toolchain	Mean	Error	Ratio	Code Size
LookupBoolInlineArray	Main	8.004 ns	0.0023 ns	0.57	36 B
LookupBoolTernary	Main	12.217 ns	0.0022 ns	0.87	36 B
LookupIPAddressInlineArray	Main	14.077 ns	0.0014 ns	1.00	48 B
LookupIPAddressTernary	Main	12.776 ns	0.0046 ns	0.91	48 B
LookupBoolInlineArray	PR	8.040 ns	0.0005 ns	0.57	36 B
LookupBoolTernary	PR	12.220 ns	0.0024 ns	0.87	36 B
LookupIPAddressInlineArray	PR	8.403 ns	0.0005 ns	0.60	48 B
LookupIPAddressTernary	PR	13.106 ns	0.0020 ns	0.93	48 B

BDN_Artifacts.zip

[dotnet-policy-service]

Draft Pull Request was automatically closed for 30 days of inactivity. Please let us know if you'd like to reopen it.

[iSazonov]

Should we keep #106129 opened if the PR is not merged?