The suite keeps one NUnit worker per logical CPU busy with allocation-heavy
decompiles (223 GB allocated per run), so under workstation GC every
gen0/gen1 collection any worker triggers suspends the whole process.
Measured on a 24-thread Windows box (Debug, ILSpy-tests checked out):
27,229 gen0 / 6,919 gen1 collections and 305 s of total GC pause in a
553 s run, at 45% average CPU. With server GC the same run takes 310 s,
1,251 gen0 / 492 gen1, 14 s of pause, 80% CPU, for the same ~46 min of
processor time; the in-suite roundtrip decompiles drop 2-3x
(Random_TestCase_1 353 s -> 133 s, ExplicitConversions 319 s -> 136 s,
NRefactory_CSharp 337 s -> 156 s). Standalone ilspycmd timings are
unaffected, which is what pointed at contention inside the test process
rather than decompiler cost.
Assisted-by: Claude:claude-fable-5:Claude Code
A negative constant operand is usually the two's-complement rendering
of a bit mask or a high unsigned value (an enum member, a sentinel);
the IL view now appends the hexadecimal form as a comment, e.g.
'ldc.i4 -501 // 0xfffffe0b' (#1142). The short forms stay bare: their
operand range is readable as-is. The disassembler round-trip comparer
strips comments, so the new NegativeConstants case pins the rendering
with explicit content asserts.
Assisted-by: Claude:claude-fable-5:Claude Code
An addition or subtraction on an enum whose constant operand's numeric
value does not fit the underlying type (an int constant standing for a
high uint member, e.g. -501 for 0xfffffe0b) failed to resolve as enum
arithmetic and fell back to integer arithmetic with casts, producing
'(uint)((int)value - -501)' and, for the same source expression in an
argument position, '(uint)value - 4294966795u'. Retry the failed
resolution once with constant operands reinterpreted in the enum type;
the reinterpretation is lossless whenever the constant's stack type
matches the enum's underlying stack type, because the IL constant is
the member's bit pattern. Valid non-enum resolutions like 'data - 1'
(enum minus underlying, yielding the enum) are unaffected because the
retry only runs when the plain resolution fails.
Assisted-by: Claude:claude-fable-5:Claude Code
The C# 14 compiler lowers implicit span conversions to calls -
MemoryExtensions.AsSpan(string), ReadOnlySpan<T>.CastUp, and the span
op_Implicit operators - so decompiled code showed the lowered form even
though the conversion and betterness layers already implement the C# 14
rules. CallBuilder now folds those helper calls back into conversions,
riding the existing mechanism: the conversion is built as an explicit
cast, consumption sites make it implicit where the context allows, and
the overload-resolution recheck re-adds a cast when the bare argument
would bind to a different overload (which canonicalizes deliberate
AsSpan disambiguations to the equivalent explicit span cast).
Span conversions compose, so CastCanBeMadeImplicit lets a direct
input-to-target span conversion replace a chained pair; and an rvalue
bound to an in parameter gets the same chance to shed the cast as a
by-value argument, since ChangeDirectionExpressionTo bypasses the
by-value strip.
Part of #829.
Assisted-by: Claude:claude-fable-5:Claude Code
Below C# 7 ref locals are unavailable, so CopyPropagation is allowed to copy
LdFlda/LdElema. When such a copy lands in a StObj target slot whose value is
impure, it violates the invariant checked by StObj.CheckTargetSlot: C# computes
the value to be stored before dereferencing the target, so the exception moves.
ILInlining resolves the same conflict by marking the address as delayed rather
than falling back to a ref local; copy propagation now does the same, which
keeps the generated code unchanged and only repairs the IL.
Unlike inlining, copy propagation has no third arm to fall back to: by the time
DoPropagate runs, the defining store is about to disappear, so every load has to
be replaced and refusing the copy is no longer an option. That decision can only
be made up front, which is what CanPerformCopyPropagation does when ref locals
are requested. The assertion covers the remaining hole, the public Propagate()
entry point, which bypasses that check -- AsyncAwaitDecompiler copies an ldflda
of the builder field through it irrespective of the setting.
Propagating an address also un-inlines its arguments into fresh stack slots, and
those stores are copy-propagation candidates in their own right. They are
inserted before the store being replaced, so the block scan used to step right
past them: a slot loaded more than once could never be inlined back and survived
into the output as a ref local -- for `s.ShortField >>>= 5` the copied ldflda
left behind `stloc C_0(ldloca s)`, printed as `ref CustomStruct reference = ref
s`. Rewinding the scan to the first of those stores lets them propagate too.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Structs whose assembly is missing decompile through the unresolved-type
path, which is not covered anywhere: the fixture pins the constructor
shapes that path has to recognize, and the reference-type cases that must
keep falling through to a plain call.
Assisted-by: Claude:claude-opus-5:Claude Code
Comments used to be child nodes flushed by InsertSpecialsDecorator when the
next node started printing, which put the marker of an init-only setter right
after the keyword. In the slot AST comments are leading/trailing trivia, so the
accessor's trailing trivia moved the marker behind the accessor body. The
placement cannot go back to trivia on the body either: the auto-property
transform drops the body, and the marker with it. The accessor now carries the
init-only fact itself and the printer writes the marker next to the keyword.
Assisted-by: Claude:claude-opus-5:Claude Code
Overload resolution reports no error for an empty candidate set - there is
no best candidate to attach one to - so the null result passed for success
and was dereferenced while checking the call target. Decompiling
FSharp.DataFrame from nuget.org crashes that way: F# compiles its comparison
members to instance methods carrying operator metadata names, and the
operator candidate search looks at the operand types rather than at the
receiver type the member belongs to.
The new fixture pins that such an assembly decompiles at all. It still
renders those instance methods as operators and drops the receiver at the
call sites, which is the misclassification behind the empty candidate set
and is handled separately; this is the guard that keeps an empty candidate
set from being read as a resolved call.
Assisted-by: Claude:claude-opus-5:Claude Code
Backing-field references inside a property's own get/set/init accessors
are emitted as the `field` keyword at IL-to-AST translation time
(ExpressionBuilder.ConvertField), so arbitrary accessor bodies become
expressible and no separate rewrite pass is needed. Compiler-generated
trivial accessors then collapse individually to `get;`/`set;`, which
also handles mixed shapes like `{ get; set { ... field ... } }`; the
backing-field declaration is removed with its remaining attributes
re-hosted as `field:` sections, and constructor stores become property
initializers (or property assignments, for setter-less properties).
Implicit zero-stores that auto-default struct constructors emit for
unassigned backing fields are dropped rather than lifted.
Recognition stays AST/metadata-based rather than mirroring the ILAst
analysis used for automatic events: events must prove compiler-generated
bodies before discarding them, while the field keyword discards nothing,
so name association plus the accessor context is sufficient.
Below C# 14 (or with the new FieldKeyword setting off), the field
declaration survives under its metadata name, so the UI keeps showing
the truth; EscapeInvalidIdentifiers - the transform the compilable-output
flows (project export, VS, tests) already add - now maps
`<P>k__BackingField` to the readable `P__BackingField` instead of the
generic character escape. A genuine field literally named "field" is
qualified as `this.field` inside accessors, and locals are not named
"field" there, since C# 14 rebinds the bare identifier. Bodiless
accessors mixed into multi-line properties get their own line in the
output.
The fixture covering the feature surface lands with the implementation rather
than as a separate xfailed commit. It is excluded from the test-assembly
compilation because its nullable annotations would trip warnings-as-errors
there.
Assisted-by: Claude:claude-fable-5:Claude Code
The parameter-list-less anonymous method form is compatible with any
delegate signature, and C# code must rely on exactly that when a
delegate's parameter types cannot be named at the use site: IL, unlike
C#, permits a delegate signature to reference less accessible types.
Expanding such an anonymous method into a lambda would force the
unnameable type into a parameter list. Keep the delegate form, with its
parameter list dropped, when the parameters are unused and one of their
types is not accessible from the current context.
Assisted-by: Claude:claude-fable-5:Claude Code
VBPretty had no coverage of VB's anonymous types, so nothing caught that their
use sites decompiled to the raw metadata names while their definitions were
hidden from the output. The expected C# is written as it should read once the
generated-name predicates agree with each other; it fails until then.
Roslyn 2.10 targeting .NET Core 2.2 is branched off with #if: there the query
operator calls are not restored to extension-method syntax, so no query
expression is formed and the lowered form survives.
Assisted-by: Claude:claude-fable-5:Claude Code
IsCopyConstructor required the copy constructor to be private on a sealed
record and protected otherwise, but IsGeneratedCopyConstructor in the same
class accepts protected regardless of sealedness. Since the former gates the
latter, a sealed record whose copy constructor stayed protected -- what you get
when a record is sealed after it was compiled -- was not recognized as one at
all: it fell through to the general constructor handling, where its base call
made it count as unchained, and the primary-constructor invariant then failed.
Found by fuzzing nuget.org. Beyond silencing the assertion this improves the
output, as the affected types now decompile to positional records.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
The accessor forwarding-stub matcher bounded the IL body from above but not
from below. Reference assemblies keep the method RVA while stripping the body
to zero bytes, so an explicit interface accessor there sailed past the size
check and the first opcode read ran off the end of the blob, aborting the
whole type with a BadImageFormatException. The sibling matcher in
TransformDisplayClassUsage already guards its lower bound; this one did not.
Found by fuzzing nuget.org: every package resolving to the
Microsoft.NETFramework.ReferenceAssemblies packs was affected.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Roslyn compiles top-level statements into a synthesized Program class
whose entry point is called '<Main>$', a name that cannot be declared
in C#. Decompiled output kept it verbatim (escaped to
_003CMain_003E_0024 when exporting a project), and since C# accepts
only a method called 'Main' as an entry point, the exported executable
did not compile (CS5001).
Give that method the name 'Main'. Per the decision recorded in #829 we
do not reconstruct top-level statements, so this is the level of
support the output needs to compile.
An async top-level program needs the name in a different place: it
compiles to '<Main>$' holding the statements plus a '<Main>' entry
point that only awaits it. That wrapper carries the .entrypoint marker
but is hidden from the output, so the name goes to the method it
awaits instead - unless AsyncAwait is off, when the wrapper is
emitted and keeps the name itself.
Assisted-by: Claude:claude-fable-5:Claude Code
Corrupt or hand-written metadata can contain a property with no
MethodSemantics rows at all. MetadataTypeDefinition.Properties already
skips properties with neither a visible getter nor setter, so neither
the parameterized-property path nor the ordinary one ever sees them;
the accessor-method emission also tolerates a missing accessor by
construction. Pin that with an ILPretty case containing both a
parameterized and an ordinary accessor-less property.
Assisted-by: Claude:claude-fable-5:Claude Code
C# has no syntax for named properties with parameters, so the expected
output declares the accessors as ordinary methods, keeps the
property-level attributes under the inert 'property:' attribute target
(ignored by csc with CS0657), and consumes the properties through
direct accessor calls, which Roslyn permits exactly for properties
whose shape C# cannot bind. Red against the current decompiler, which
emits a parameterless property whose body references the vanished
parameters.
Assisted-by: Claude:claude-fable-5:Claude Code
DebugInfoGenerator asserted that every local variable's decompiled type
is equivalent to the metadata local-signature type at its slot. That
holds at IL-read time but not afterwards: variable splitting gives one
slot several typed variables, pinned-region locals are modeled as
pointers (int* vs int& pinned), and generic-context type-parameter
identity differs. The assertion therefore aborted PDB generation (Debug
builds) for ordinary inputs such as any method with a fixed statement.
The type is never written to the PDB - only the slot index and name are -
so the mismatch cannot affect the debugging experience. The slot index,
the only emitted value, is correct by construction: it is the IL
ldloc/stloc operand, sourced from the signature slot when the variable is
created and copied verbatim by SplitVariables; it is never reassigned.
Keep only the index-bound check and document why the type is not verified.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Decompiler warnings (ILFunction.Warnings, e.g. the DetectPinnedRegions
block-duplication notice) are surfaced as an EmptyStatement carrying only
a comment. VisitEmptyStatement prints no semicolon for it, and
EmptyStatement derives its StartLocation/EndLocation from its Location
field, which is only set when that semicolon token is written. The
statement was therefore left without a text location, and
SequencePointBuilder then asserted on the empty start location while
generating PDB sequence points - aborting PDB generation for any assembly
whose decompilation emits such a warning (e.g. System.Net.Requests).
Point the empty statement at the comment it carries (already printed by
the time the node ends, so its location is known), falling back to the
collapsed end-of-last-token position. Every printed node then has a
location, so the SequencePointBuilder invariant holds without
special-casing, and the statement lines up with the text it represents.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Extends the breakpoint-map comparison to hidden sequence points, anchoring
each hidden point to the visible point it follows so the descriptor stays
independent of the IL offsets the decompiler reconstructs. Adds PdbGen cases
spanning try/catch/finally, switch, async/await, yield, loops, LINQ, pattern
matching and more, pinning the known residuals where the decompiler folds a
compiler-hidden branch into an adjacent point.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The project listed 201 explicit <Compile Include> and 119 <None Include> items
with EnableDefaultItems=false. Many test-case sources had been marked <None>
only because the C# compiler available when they were written could not build
them; the current Roslyn compiles them fine. The hand-maintained lists had also
drifted: five committed fixtures (Operators.cs, Issue3751.cs, three ILPretty
.cs) were on disk with passing tests but in no item list.
Switch to the SDK default **/*.cs glob and exclude only the sources that still
cannot be compiled into the test assembly, determined empirically by building
with everything included and removing what failed:
- IL-pretty inputs that are not valid standalone C# and *.Expected.cs golden
outputs that reuse type names (compile errors / duplicate definitions).
- MetadataAttributes.cs, which applies assembly/module attributes that break
NUnit test discovery (zero tests found) when compiled into the assembly.
The excluded sources stay <None> for IDE visibility; the harness compiles them
from disk at test time regardless. This compiles 23 more fixtures than before
while keeping every previously-compiled file. Default None globbing stays off so
the non-C# inputs (.il/.vb/.fs) remain the authoritative list.
Operators.cs, now part of the build, is normalized to the repo's tab indentation
by the format hook (a CS110 block used spaces). Verified: the full suite still
reports 2257 passed, 0 failed, 35 skipped.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The generated GetChildNodes materialized a List<AstNode> (plus a boxed
List.Enumerator at every foreach) for each node, so AstNode.Children and
the visitor's per-node child walk allocated three objects per traversal.
Decompiling System.Private.CoreLib that came to ~1.7 GB of extra garbage,
roughly +7% over the linked-list model the slot tree replaced. A yield
iterator removes the List but trades it for an equally costly per-node
state machine, so it is not enough on its own.
Enumerate children through a by-value struct enumerator over the existing
FirstChild/NextSibling primitives, capturing each child's successor before
it is yielded so a transform may still remove or replace the current child
mid-traversal. AstNodeCollection<T> gets the same struct treatment for a
direct foreach. Child enumeration now allocates nothing, bringing total
allocations back to the linked-list baseline at byte-identical output
(full Pretty suite green with CheckInvariant active).
Assisted-by: Claude:claude-opus-4-8:Claude Code
The pattern matcher walked collections through INode.Role/FirstChild/
NextSibling, skipping siblings of a different role. Now that each
AstNodeCollection<T> is already the per-role child list, the engine matches two
collections by list index, and INode sheds Role/FirstChild/NextSibling
entirely. A collection exposes its IReadOnlyList<INode> view through a cached
adapter rather than implementing the interface directly, so a typed collection
does not become ambiguous for LINQ. Characterization tests pin the matcher's
behavior first.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Source locations were virtual, computed by recursing to the first and last
child, whose leftmost and rightmost leaves are token nodes; sequence-point
coordinates likewise came from reconstructed token nodes. Store locations as
fields assigned while printing, and derive sequence-point coordinates from the
surrounding real nodes plus the decompiler's fixed formatting, so neither
depends on token children. The using/foreach await modifier becomes a plain
bool field. Characterization gates lock the emitted locations and PDB
coordinates, which are unchanged.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The InsertMissingTokensDecorator path (TokenWriter.CreateWriterThatSetsLocationsInAST)
reconstructs token nodes and assigns source locations onto the AST, feeding PDB
sequence points and GUI navigation. The Pretty suite never drives it, so it had no
coverage at all. Before reworking the token model, lock its observable consequences:
the located path emits the same text as the plain path, real nodes receive ordered
locations, location-based navigation resolves into the method body, and sequence
points are produced for a method body.
Assisted-by: Claude:claude-opus-4-8:Claude Code
The "IL with C#" view decompiles each method body as a bare handle, so a
static constructor is decompiled without its type's field declarations in
the syntax tree. MoveFieldInitializersToDeclarations then could not find a
declaration to move the static-field-initializer statement onto, asserted
(kind was Static, not Primary) and dropped the statement -- crashing Debug
builds and silently losing the assignment in Release.
Dropping the statement is only correct for the primary-constructor case,
where the assignment's backing member is synthesized and has no separate
declaration. For static/instance initializers a missing declaration just
means the member is not part of this partial syntax tree, so the
assignment must remain in the constructor body.
Assisted-by: Claude:claude-opus-4-8:Claude Code
On FIPS-mode systems the platform crypto provider refuses to create
SHA-1 instances (OpenSSL: error:03000098 invalid digest), so merely
displaying a strong-named assembly's identity failed. The public-key
token is a non-secret identity hash whose algorithm is fixed by
ECMA-335, so the two token sites now use dotnet/runtime's managed
Sha1ForNonSecretPurposes, vendored with its license header intact and
shielded from the repo formatter via generated_code in .editorconfig
so future upstream syncs diff cleanly. IncrementalHash was considered
and rejected: like SHA1.Create(), it resolves the digest through the
host crypto policy, and Roslyn's equivalent token code also relies on
the platform SHA-1, so it offers no precedent for FIPS safety.
Assisted-by: Claude:claude-fable-5:Claude Code
Windows maps CON, PRN, AUX, NUL, COM1-9 and LPT1-9 to devices -- on many
builds even with an extension appended, so a type named Con made both
whole-project export and the save dialog fail with IOException '\\.\Con'.
CleanUpName only checked for reserved names after re-appending the file
extension, where they never match, and the save-dialog default-name
helpers did not check them at all. The escape appends the underscore to
the base name (con_.txt, not con.txt_) because device-name parsing
ignores everything after the first dot, and is applied per path segment
so reserved directory names produced by namespaces are covered too. The
ILSpy.Tests.Windows fixture verifies on a real Windows filesystem that
the escaped names are creatable.
Assisted-by: Claude:claude-fable-5:Claude Code
The decompiled view only offered navigation on member names; finding
what an override actually overrides required opening the analyzer.
Attaching a reference to the modifier token gives the same
go-to-definition affordance Visual Studio has on 'override'. The
reference resolves via InheritanceHelper.GetBaseMember, so it targets
the nearest overridden member and degrades to plain text when the base
member cannot be resolved.
Assisted-by: Claude:claude-fable-5:Claude Code
The Pdb2Xml command (ILSpy) and the PDB round-trip tests (Decompiler.Tests)
reference Microsoft.DiaSymReader*, previously gated on the build host being
Windows. That made dotnet restore resolve a different graph on Windows than
on Linux -- the packages (and their transitive tail: DiaSymReader.PortablePdb,
the legacy NETCore.Platforms/Win32 packages) appear only on Windows, and
DiaSymReader.Native flips between Direct and CentralTransitive. So a checkout
could not be developed across OSes without the committed packages.lock.json
churning on every Windows restore.
Drop the OS gate (keep Debug-only) so the restored graph, and the committed
lock, are identical on every OS. The consuming code is still gated by DEBUG
and WINDOWS, so on non-Windows the packages are restored but never compiled
in; the native asset only resolves for win-* RIDs.
The "Verify package contents" step (which checks the committed *.filelist
snapshots still match the built VSIX/MSI contents) had been excluding
packages.lock.json from its git diff to tolerate that per-OS churn. With the
locks now identical across OSes the carve-out is unnecessary, so it goes back
to a plain git diff --exit-code.
Assisted-by: Claude:claude-opus-4-8:Claude Code
ICSharpCode.Decompiler.Tests builds for the host RID (so the native
Microsoft.DiaSymReader.Native assets the Windows PDB tests need are copied),
which made its packages.lock.json host-specific (linux-x64 vs win-x64) and
drift between Linux and Windows restores. Add an explicit <RuntimeIdentifiers>
list next to the single <RuntimeIdentifier>: the lock then records every RID
host-independently while the build still targets the host RID. Apply the same
list to the TestRunner it pulls in.
TestRunner's lock is now fully portable. The test project's lock still carries
the OS-conditional DiaSymReader.Native difference in its base graph, which the
*.filelist-scoped "Verify package contents" step already tolerates.
Assisted-by: Claude:claude-opus-4-8:Claude Code
Decompiler.Tests targeted net11.0-windows only because
Microsoft.DiaSymReader.Native is Win-only and Tester.cs hardcoded ilasm.exe
/ ildasm.exe filenames. Both are narrow problems: gate the Native package
to Windows, replace the ".exe" suffix with an OS-aware helper, and drop
the -windows TFM. The Microsoft.NETCore.IL{,D}Asm metapackages' RID graph
already ships ilasm/ildasm for linux + osx via per-RID runtime sub-packages
when RuntimeIdentifier resolves dynamically, so the tool just lands in the
test output dir with the right suffix on every host.
Windows-only test fixtures (RoundtripAssembly + the RunWithTestRunner,
SignAssembly, FindMSBuild call sites) gate via Assert.Ignore +
[Platform("Win")] so they skip on non-Windows instead of failing. CA1416
joins NoWarn -- the analyzer can't see the runtime guards, and the test
inputs (Console.CapsLock, Registry access) are decompilation targets, not
infrastructure.
BuildTools/pre-commit learns to find dotnet-format on non-Windows (XDG
basedir + no .exe suffix) so the hook can run wherever the build now does.
Assisted-by: Claude:claude-opus-4-7:Claude Code
* .NET 11 RC2 minimal changes
* Heuristic for transport feed Roslyn selection
* Microsoft.CodeAnalysis.NetAnalyzers from main NuGet feed
* Use the VS2026 image
* Switch all test projects to net11
* Extract constants
* Include vsix with plain nuget.config files
Various improvements regarding primary constructor decompilation, including:
- introduce `HasPrimaryConstructor` property in the AST, as there is a difference between no primary constructor and a parameterless primary constructor
- improved support for inherited records and forwarded ctor calls
- exclude non-public fields and properties in IsPrintedMember
- introduce an option to always make the decompiler emit primary constructors, when possible
Christoph Wille
Siegfried Pammer
DoctorKrolic
ds5678