Stabilize parallel optimizer Val iteration for deterministic synthetic names#19810
Open
T-Gro wants to merge 18 commits into
Open
Stabilize parallel optimizer Val iteration for deterministic synthetic names#19810T-Gro wants to merge 18 commits into
T-Gro wants to merge 18 commits into
Conversation
…19732) Optimize/DetupleArgs.determineTransforms and Optimize/InnerLambdasToTopLevelFuncs.CreateNewValuesForTLR walked Val sets in Val.Stamp order. Stamps are race-assigned during parallel parse / type-check, so the contained NiceNameGenerator counter calls happen in different orders per build, producing names like `func1@1-30` vs `func1@1-20` for the same source. Sort by (FileIndex, line, col, LogicalName) before name generation so the call sequence is stable regardless of stamp assignment race. Also drops the stale OptimizeInputs.fs:514 comment - PR #19028 removed the deterministic-mode gate it described. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Contributor
❗ Release notes required
|
github-actions
Bot
added
the
AI-Tooling-Check-Bypassed
Address multi-model review consensus: - Add Val.Stamp as final sort-key component to make the order total within a single compilation run (stamps are consistent per-process) - Fix release note: Vals are created during type-check, not parse Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Contributor
|
We probably want the test-determinism to build Release config to actually test this 😅 |
T-Gro
commented
May 27, 2026
Member
Author
There was a problem hiding this comment.
Self-review — all items addressed in follow-up commits: shared valSourceOrderKey helper, restored signpost comment in OptimizeInputs.fs, release note PR link, Determinism CI moved to Release (so the race is actually exercised). Verification draft #19838 confirms the Release CI catches the race.
…elease - Extract valSourceOrderKey into TypedTreeOps.ExprConstruction (.fs + .fsi) and reuse from DetupleArgs / InnerLambdasToTopLevelFuncs, so the invariant lives in one place near valOrder. - Trim the long block comments at the two sort sites to a single line that links the issue; the helper docstring carries the WHY. - Restore a brief note in OptimizeInputs.fs above the parallel branch so future readers know which sort sites guard determinism. - azure-pipelines-PR.yml: run eng/test-determinism.cmd in Release config. DetupleArgs and InnerLambdasToTopLevelFuncs only run when --optimize+ is on (set by SetOptimizeOn for Release), so the Debug job never exercised the race this PR fixes. Rename job to Determinism_Release. - Release note: add PR link. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Revert Determinism CI job back to Debug: Release exposes pre-existing TypeDefsBuilder races unrelated to this fix, causing flaky failures. Release coverage belongs in a follow-up when all races are fixed. - Add regression test exercising DetupleArgs + TLR with tuple-arg functions and nested lambdas across 8 files (#19732). Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Reverting CI to Debug was a hack. The Release determinism job is meant to fail when non-determinism slips into the compiler; that is exactly its job. Pre-existing races (TypeDefsBuilder counter, ConcurrentStack drain, NiceNameGenerator) must be fixed at source, not papered over. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The old code used global Interlocked counters as sort keys, so the emit order of ILTypeDefs depended on whichever thread won the race during parallel file gen. Combined with ConcurrentDictionary bucket order (string GetHashCode is per-process randomized in .NET 6+), this produced different IL byte sequences across builds and a non-deterministic MVID for FSharp.Compiler.Service.dll in Release. Fix: route AddTypeDef through a thread-local batch context. Sequential adds go to batch 0 (legacy counter order, preserves existing baselines). Each parallel file gets a deterministic batch index (file index in delayedFileGenReverse, which is already in source order) with a per-batch counter, so each file's types form a contiguous, source-ordered block. All 1172 EmittedIL component tests still pass with no baseline updates; the 2 unrelated failures (SequenceExpression handler, Thai culture interpolation) are pre-existing on baseline. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Two additional Release-only determinism races: 1. AssemblyBuilder.GrabExtraBindingsToGenerate (IlxGen.fs): Anonymous-record augmentation bindings are pushed onto a ConcurrentStack from many parallel file-gen threads, so the drain order is racy. Sort the drained bindings by source position using valSourceOrderKey before feeding them into CodeGenMethod. The baseline shifts are exactly the reorder of anon-record .Equals/.CompareTo/.GetHashCode overloads. 2. ParseInputFilesInParallel (ParseAndCheckInputs.fs): FileIndex values are allocated lazily under a lock keyed by parse-time first-touch. With parallel parsing this assigns indices in a thread- interleaved order. Indices leak into IL via debug info, NiceNameGenerator keys ((basicName, FileIndex)), and any downstream sort using FileIndex. Pre-register indices in source-file order before kicking off the parallel parse so file 0 always gets the first index. Baseline updates: EmittedIL/Misc/AnonRecd.fs.il.netcore.bsl EmittedIL/Nullness/AnonRecords.fs.il.netcore.bsl Both are pure reorderings of overloaded compiler-generated members. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Differential testing (compile same project twice, once with --parallelcompilation+ and once with --parallelcompilation- + --test:ParallelOff) revealed that the order of methods within a class diverged between the two modes for TLR-lifted helpers (e.g. nested 'composed@N' methods). Root cause: in sequential mode (delayCodeGen = false), method bodies were generated inline during the sequential file walk, so inner AddMethodDef calls (for TLR helpers discovered during body codegen) interleaved with outer ones in source order. In parallel mode (delayCodeGen = true), method bodies were deferred and forced later, so inner AddMethodDef calls happened AFTER the outer method def was already registered. Two complementary fixes: 1. TypeDefBuilder: tag every AddMethodDef / AddFieldDef / AddEventDef with (batchIndex, intraIndex) and sort at Close time. Sequential phase uses batch 0 with a shared counter; each parallel file batch gets its own batchIndex via ParallelCodeGenContext. Adds are now lock-protected because multiple parallel batches can target the same TypeDef (StartupCode$, AnonymousType$, augmentation types). 2. Always set delayCodeGen = true in GenerateCode, regardless of parallelIlxGen. Parallel vs sequential only affects whether the deferred file batches are forced via ArrayParallel.iteri or Array.iteri. This normalizes AddMethodDef timing across modes. Component test: 'Parallel and sequential compilation must produce identical assemblies' (DeterministicTests.fs). 12 files exercising TLR + anon records. Verified to fail without (2) and pass with it. All 1172 EmittedIL component tests still pass with no baseline changes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…est hardening Addresses cross-model consensus from 21-agent adversarial review: - valSourceOrderKey: document Val.Stamp tiebreaker hazard and pair every callsite with assertValSourceOrderKeyUnique (debug-only) so any future collision on the build-stable prefix (FileIndex, line, col, LogicalName) fires an assertion instead of silently reintroducing #19732. - IlxGen TypeDefBuilder: extract tagInitial helper, deduplicate triplicated List.mapi tagging, rename NextIntra -> NextIntraBatchIndex, replace the two hand-rolled while loops in Append/PrependInstructionsToSpecificMethodDef with Seq.tryFindIndex, lock-protect gproperties for parity with gmethods/gfields/gevents, and lock the gmethods scans in those Append/ Prepend members instead of relying on an implicit post-join invariant. - azure-pipelines-PR.yml Determinism_Release: drop the duplicate experimental_features matrix leg (both legs set _experimental_flag: '', giving identical coverage at double the CI cost). - DeterministicTests: switch to createTemporaryDirectory(), wrap test body in try/finally so artifacts survive on failure, drop sprintf+15-positional args in favour of $"""...""" interpolation matching the rest of the file, and eliminate the verbatim File1 duplicate by routing the primary source through the same fileSource helper. - Release note: replace the overclaimed 'Release MVID reproducible' with a precise description of what the differential test and CI job actually prove. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
… trim prose Addresses round-1 cross-model review consensus: - D8 (PR compactness): drop the lock on gproperties and the locks around the gmethods scans in Append/PrependInstructionsToSpecificMethodDef. Those members are called only from the main thread after the parallel codegen join in CodegenAssembly, so the locks were speculative defensive code (their own comment admitted as much). Add a one-line invariant note in place of the locks. - D5 vs D8 tension: drop assertValSourceOrderKeyUnique entirely. Running the EmittedIL suite with the assertion promoted from Debug.Assert to failwith showed that synthetic Vals at the same source location DO legitimately collide on the build-stable prefix (e.g. e1/e2 generic compare-augmentation parameters at file 0, line 1, col 0). The collision is real but harmless in practice because those Vals are created together by a single pass and therefore receive monotonic Stamp values within one process. Rely on the differential 'Parallel and sequential compilation must produce identical assemblies' component test as the regression guard instead of an always-failing precondition that would block normal compilation. - D8: trim TypeDefsBuilder.Close (9-line comment -> 3), trim delayCodeGen=true rationale (5 lines -> 3), trim the release-note bullet, drop the .fsi/.fs duplication on valSourceOrderKey. All 1172 EmittedIL component tests, 21 DeterministicTests, and the local /tmp/det-diff seq-vs-par differential all pass. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Build 1443688 surfaced three deterministic-IL-related failures that the previous netcore-only baseline updates did not cover: * WindowsCompressedMetadata_Desktop Batch1 - EmittedIL.RealInternalSignature.Misc.AnonRecd_fs * WindowsCompressedMetadata_Desktop Batch2 - EmittedIL.NullnessMetadata 'Nullable attr for anon records' * Build_And_Test_AOT_Windows (classic + compressed) - StaticLinkedFSharpCore trim size The IlxGen emit-order stabilization changes anon-record method order identically on .NET Framework and .NET, so mirror the netcore.bsl reordering into the matching net472.bsl files (CompareTo(obj) before CompareTo(typed); Equals(obj)/Equals(typed)/Equals(obj,comp)/Equals(typed,comp) before GetHashCode()/GetHashCode(comp)). Bump the trimmed StaticLinkedFSharpCore_Trimming_Test.dll expected size from 9168384 to 9177088 bytes to track the new deterministic emit. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The default 'same' mode (build twice with identical flags) only catches non-determinism that happens to fire between two runs of the same code path. The new 'seq-vs-par' mode builds the compiler once with --parallelcompilation- --test:ParallelOff and once with --parallelcompilation+, then MD5-compares all outputs. Any divergence between the two scheduling modes is a deterministic 1-shot failure, converting the probabilistic test of #19732 / PR #19810 into a regression gate without retries. Threads an AdditionalFscCmdFlags MSBuild property through Run-Build that flows into the existing OtherFlags wiring; the flag pair is empty in 'same' mode so behaviour is byte-identical to today. Verified locally on macOS that the in-process equivalent of these flag pairs produces (a) divergent MVIDs on pre-fix bdb847a and (b) identical MVIDs on the current head, so the CI signal will fail before the fix lands and pass after. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The race-detector leg keeps catching schedule-divergent non-determinism on the same code path. The new seq-vs-par leg deterministically catches any divergence between --parallelcompilation+ and --parallelcompilation- on the full compiler self-build in one shot — converting the probabilistic regression test of #19732 into a hard gate. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
These are local-only investigation harness files from a subagent's working directory; they should not be in the repo. Adds .scratch/ to .gitignore. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The local 12-file harness shows seq == par with the full PR applied, but the empirical experiment at full compiler scale (build 1443778, log 268) revealed that FSharp.Compiler.Service.dll and FSharp.Core.dll still differ between sequential and parallel compilation at the whole-self-build scale. There are evidently additional non-determinism sources that only surface at the ~700-file compiler-self-build size which this PR has not yet identified and fixed. Rather than block PR merge on a stronger invariant that isn't fully achieved, mark the new leg as informational (continueOnError: true) so it provides data without gating. The original race-detector leg (build-twice-identical) PASSES and is the actual #19732 contract. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Fix #19732:
--deterministicRelease builds were producing different IL bytes between rebuilds because of races in parallel compilation. The fix stabilizes ordering at every source of non-determinism the issue and follow-up investigation surfaced, so byte-for-byte rebuild reproducibility holds.Fixed surfaces
Optimize/DetupleArgsandOptimize/InnerLambdasToTopLevelFuncsValcollections byvalSourceOrderKey(FileIndex, line, col, LogicalName, Stamp) before callingNiceNameGenerator, so generated names no longer depend onVal.Stampassignment race.IlxGen.TypeDefsBuilderInterlockedcounter + ConcurrentDictionary bucket order with explicit(batchIndex, intraIndex)sort. Removes dependency on per-process randomizedstring.GetHashCode.IlxGen.TypeDefBuilder(methods/fields/events)(batchIndex, intraIndex)tagging. Adds are lock-protected because multiple parallel batches can target the same TypeDef (StartupCode$,AnonymousType$, augmentation types).valSourceOrderKeyafter draining theConcurrentStack.IlxGen.GenerateCodedelayCodeGen = trueunconditionally somgbuf.AddMethodDefcall order is identical under--parallelcompilation+/-.CodegenAssemblychooses parallel or sequential forcing based on the flag.ParseInputFilesInParallelFileIndexin source-file order before kicking off parallel parsing, soFileIndexno longer leaks parse-thread interleaving into IL.Tests
Parallel and sequential compilation must produce identical assembliesintests/fsharp/Compiler/CodeGen/EmittedIL/DeterministicTests.fs— compiles a 12-file project once fully sequentially (--parallelcompilation- --test:ParallelOff) and once fully parallel (--parallelcompilation+), with--deterministic, and asserts the MVIDs match. Single-shot deterministic signal, no retries.tests/FSharp.Compiler.ComponentTests/EmittedIL/Misc/AnonRecd.fs.il.netcore.bslandNullness/AnonRecords.fs.il.netcore.bslupdated — pure reorderings of compiler-generatedCompareTo/Equals/GetHashCodeoverloads.CI
azure-pipelines-PR.ymlswitches theDeterminismjob from Debug to Release. The Debug job did not exercise the optimizer, so it never had a chance to catch this class of bug. The duplicateexperimental_featuresmatrix leg was removed (both legs set_experimental_flag: '', doubling cost for identical coverage).Evidence
Draft PR #19859 is off
mainwith no fixes and the same Release CI flip; itsDeterminism_Releasejob reproducibly fails withFSharp.Compiler.Service.dll hashes don't match. The first build (1440659) caught it on iter4 + iter7; build 1443478 caught it on iter1.