Andy Ayers AndyAyersMS

JIT CSE Heuristic Improvement via Oracle-Guided Imitation Learning

A methodology writeup for improving the .NET JIT's Common Subexpression Elimination (CSE) heuristic using RL and oracle-derived labels, with an emphasis on the evaluation loop. Written up so somebody else can leverage the pattern for a similar heuristic-improvement task.

Results delivered by this line of work:

v10, v11, v12b — successive imitation-learning models baked into the runtime as CSE_HeuristicImitation (~155KB weights, transformer over

Diffs are based on 2,219,596 contexts (807,447 MinOpts, 1,412,149 FullOpts).

MISSED contexts: base: 4,468 (0.20%), diff: 589 (0.03%)

Base JIT options: JitObjectStackAllocation=0

Overall (-18,212 bytes)

Diffs are based on 1,872,949 contexts (616,222 MinOpts, 1,256,727 FullOpts).

MISSED contexts: base: 3,555 (0.19%), diff: 0 (0.00%)

Base JIT options: JitObjectStackAllocation=0

Overall (-16,169 bytes)

Diffs are based on 1,514,429 contexts (459,672 MinOpts, 1,054,757 FullOpts).

MISSED contexts: base: 12,403 (0.82%), diff: 12,404 (0.82%)

Overall (+117,686 bytes)

|Collection|Base size (bytes)|Diff size (bytes)|

	<Project Sdk="Microsoft.NET.Sdk">
	<PropertyGroup>
	<OutputType>Exe</OutputType>
	<TargetFramework>net10.0</TargetFramework>
	<ImplicitUsings>enable</ImplicitUsings>
	<Nullable>enable</Nullable>
	<Optimize>true</Optimize>
	</PropertyGroup>
	</Project>

	From aad563bbc3784792493afc917d6347da60465a3d Mon Sep 17 00:00:00 2001
	From: Andy Ayers <andya@microsoft.com>
	Date: Wed, 19 Aug 2026 17:51:14 -0700
	Subject: [PATCH 1/2] Use acquire reads for cast cache entries

	Replace the ARM64 load barrier in CastCache.TryGet with acquire reads of the source and target/result fields.

	Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

	Copilot-Session: 6bd43e03-841f-47ee-80f4-4da985b32a67

	using System.Diagnostics;
	using System.Reflection;
	using System.Runtime.CompilerServices;
	using System.Runtime.InteropServices;

	interface ITarget
	{
	int Invoke(int value);
	}

	<Project Sdk="Microsoft.NET.Sdk">
	<PropertyGroup>
	<OutputType>Exe</OutputType>
	<TargetFramework>net10.0</TargetFramework>
	<AllowUnsafeBlocks>true</AllowUnsafeBlocks>
	<ImplicitUsings>enable</ImplicitUsings>
	<Nullable>enable</Nullable>
	<Optimize>true</Optimize>
	</PropertyGroup>
	</Project>

	****** START compiling System.Text.Json.JsonHelpers:TraverseGraphWithTopologicalSort[System.__Canon](System.__Canon,System.Func`2[System.__Canon,System.__Canon],System.Collections.Generic.IEqualityComparer`1[System.__Canon]):System.__Canon[] (MethodHash=ddfd2475)
	Generating code for Unix x64
	OPTIONS: Tier-1 compilation
	OPTIONS: compCodeOpt = FAST_CODE
	OPTIONS: compDbgCode = false
	OPTIONS: compDbgInfo = true
	OPTIONS: compDbgEnC = false
	OPTIONS: compProcedureSplitting = false
	OPTIONS: compProcedureSplittingEH = false
	OPTIONS: optimizer should use profile data

	115838 methods, 34802 methods with cses; 193807 cse candidates, 56973 cses
	Count: Cses Candidates
	1 [n: 10259 o: 0.53]: 0: 4836; 1: 5423
	2 [n: 4913 o: 0.44]: 0: 1760; 1: 2028; 2: 1125
	3 [n: 4058 o: 0.36]: 0: 1351; 1: 1329; 2: 1031; 3: 347
	4 [n: 3454 o: 0.37]: 0: 788; 1: 1100; 2: 860; 3: 570; 4: 136
	5 [n: 2576 o: 0.35]: 0: 586; 1: 497; 2: 792; 3: 470; 4: 175; 5: 56
	6 [n: 1627 o: 0.31]: 0: 264; 1: 452; 2: 474; 3: 182; 4: 227; 5: 21; 6: 7
	7 [n: 999 o: 0.33]: 0: 168; 1: 184; 2: 261; 3: 173; 4: 78; 5: 70; 6: 63; 7: 2
	8 [n: 979 o: 0.28]: 0: 194; 1: 198; 2: 243; 3: 128; 4: 100; 5: 42; 6: 36; 7: 15; 8: 23