This way, we don't need the MapToMergedBounds logic to split the merged list back into lower/upper.
Also, this commit avoids the quadratic merge-everything-with-everything else -- instead we use a dictionary to compare only types that are equivalent to begin with.
This is the same approach as Roslyn MethodTypeInference.Fix/AddAllCandidates.
ICSharpCode.Decompiler.Generators pins a stable Microsoft.CodeAnalysis.CSharp
(a source generator must not reference a Roslyn newer than the host compiler)
and ILSpy.AddIn.VS2022 pins a stable 4.0.1. Stable Roslyn lives only on
nuget.org, but the repo-root NuGet.config mapped Microsoft.CodeAnalysis.*
exclusively to the dotnet-tools feed, which carries only prerelease builds.
Both projects worked around that with a per-project NuGet.config. dotnet
restore honors it, since settings are computed per project directory, but
Visual Studio starts NuGet.config discovery at the solution directory and
never sees it: the first restore in VS on a machine with a cold NuGet cache
fails with NU1103 exactly as reported. Once any successful restore has put
the stable package into the global packages folder, source mapping is
satisfied from the cache and the problem never reappears on that machine,
which is why #3835 was closed as unreproducible.
NuGet consults every source that declares the longest pattern matching a
package id, so mapping Microsoft.CodeAnalysis.* to nuget.org as well as to
dotnet-tools makes both feeds available for the whole family: the stable
versions resolve from nuget.org and the prerelease $(RoslynVersion) from
dotnet-tools. Both per-project configs are then redundant and removed.
Verified with an empty NUGET_PACKAGES and the root config forced as the only
config (simulating VS's discovery): the old config reproduces the reported
NU1103; with the new config the generator, the decompiler tests (prerelease
Roslyn) and the VS add-in (stable 4.0.1 with its Workspaces dependencies)
all restore; restore.ps1 over ILSpy.sln leaves the lock files unchanged.
Assisted-by: Claude:claude-fable-5:Claude Code
The headless UI tests synchronized with the application by pumping a fixed
number of frames (39 loops of RunJobs/Delay across 19 files) and by pressing
at a point computed once from a control's bounds. Both encode how fast the
machine that wrote the test was: on the loaded Windows Debug CI agent the
frame count comes up short and the point goes stale, which is the recurring
timeout in the tree context-menu tests and the reason each such failure was
repaired one test at a time.
Waiters.WaitForIdleAsync replaces the frame loops. It observes the actual
precondition - no dispatcher job queued at Background priority or above, no
assembly still loading in the background sweep, a frame rendered - and
requires it on two consecutive polls so a thread-pool continuation about to
post back is caught as well.
Window.ClickAsync replaces element-targeted MouseDown/MouseUp pairs. It
re-resolves the target on every poll and presses only once the window's hit
test at the click point answers with that target, reporting the point and
what was hit instead on timeout. That diagnostic exposed one vacuous test:
User_Click_On_Visible_Row_Does_Not_Recentre_Viewport clicked the centre of a
row wider than the tree viewport, which lies under the decompiler text view,
so its assertion held without the row ever being clicked. It now clamps the
point to the viewport like the other tree-row clicks.
Clicks at text positions and press-only gutter clicks stay raw; they do not
target an element.
Assisted-by: Claude:claude-fable-5:Claude Code
Right_Clicking_A_Second_Row_Moves_The_Context_Highlight_To_It still timed
out on the Windows Debug CI job, now in the hit-test wait added for the
second right-click: for the full 60s no hit at the precomputed point matched
the captured row container. A light-dismiss overlay that survives one frame
cannot explain that many rendered frames; a container that is no longer the
one on screen can. The test only waits for three assemblies, so the rest of
the list keeps loading on the slow agent while the test runs, and every
insertion reshuffles the rows - re-realising containers and moving them -
after the row and point were captured.
The wait now resolves the row container and the click point on every poll
and matches the hit by node instead of by container identity, and a timeout
reports the point and what was hit instead so a further failure is
diagnosable from the log.
Assisted-by: Claude:claude-fable-5:Claude Code
HandleConditionalOperator collapses `if (c) a = x; else a = y;` into a
conditional operator, innermost first, and keeps going for as long as the
chain does. A source else-if ladder therefore comes back as one expression,
however long it was: the sample in #2027 decompiles to a single
2095-character statement, and one NLog method to 2230 characters nested 29
brackets deep.
ExpandNestedConditionals undoes that past one level, so a statement keeps at
most a single conditional operator.
It runs at the end of the pipeline rather than inside ExpressionTransforms,
because every transform that needs its input to be a single expression has
to see the collapsed form first: object and collection initializers, `with`,
switch expressions, interpolated string handlers, and the query lambdas the
C# stage later rewrites into clauses. Cutting the chain earlier leaves an
if-else between the statements they pattern-match on and they silently stop
matching - an object initializer assigning an init-only member then does not
even compile. The same reasoning ReduceNestingTransform gives for walking
back ConditionDetection's aggressive else-inlining once the structure is
settled.
A chain already stored to a variable is expanded into that variable, so
nothing has to be decided: the variable carries its own type. A chain in any
other position - an argument, a return value, a field store - has nothing to
expand into, and ILExtraction can give it one. The temporary ILExtraction
creates is typed from the stack type though, where `I4` is `int`, `bool`,
`char` and every enum at once, so extracting on that basis turned a bool
into `int num` with `if (num == 0)` and an enum into
`dbType = (IsFixedLength ? 22 : 0)`.
InferExpectedType is the counterpart to InferType that answers this: where
InferType asks what a value is, it asks what the position the value flows
into says it should be - a parameter, a return type, a field, all of which
carry their type in metadata. Extraction is done only where that question
has an answer, and the temporary is typed from it. The receiver of a call
then reads `XPathNavigator xPathNavigator`, not `object obj` with a cast
back, and a field store keeps its enum's member names.
The Pretty fixture covers what must NOT change: an array initializer, a
query lambda, a ref local, a switch expression, an object initializer with
an init-only member, a `with` expression, a catch-when filter and both
constructor-initializer forms. The positive case is an ILPretty test,
because a Pretty fixture is its own input and expected output, and a chain
that round-trips through collapse and expansion has no fixed point there.
The PdbGen test records the cost in breakpoints: the compiler's single
sequence point for the collapsed statement becomes one per expanded
statement, which is inherent to splitting a statement in two.
#2027
Assisted-by: Claude:claude-opus-5:Claude Code
A dynamic call site that names a type passes typeof(T) as its target. When
a later argument contains control flow, the compiler stores that typeof in
a temporary. ILInlining does not undo this: it inlines a store only into
the immediately following instruction, so an intervening statement leaves
the typeof behind, and the expression builder printed the temporary instead
of the type. Substituting the typeof back into the target slot is not an
option either, because a call there has side effects and blocks inlining of
the remaining arguments.
The type is therefore stored on the instruction. Object creation already
did this, but kept the field private, so the expression builder re-derived
the type from the target argument and failed on the spilled form. Member
invocation gets the same field. Both drop the target from Arguments, so the
dead-argument handling removes the store; ArgumentInfo keeps its entry for
the target, because the invocation symbol is built from it.
The type of an object creation is not nullable: the transform returns
before constructing the instruction when it cannot match the typeof, and
substitutes the unknown type otherwise. An unresolvable type therefore
reaches the expression builder as the unresolved type it is, and prints as
`?` like any other, rather than being indistinguishable from a missing one.
Those two are also the only binder method kinds that ever see
CSharpArgumentInfoFlags.IsStaticType. Every other way of naming a type as
the receiver binds statically and leaves only a conversion of the dynamic
operand behind.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Handing the corpus over as `$(cat list)` is a bourne shell construct, so the
one invocation the tools print and the README documents did not work on
Windows. nugetfuzz already reads its package list from an @file for the same
reason; decompdiff now takes its corpus entries the same way, leaving the
shell out of it entirely.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Both tools need a corpus of real assemblies and neither had a way to get
one: nugetfuzz-all.ps1 walks the catalog in publish order, which is fine for
a crash sweep but makes a poor readability corpus, and the alternative was
picking package ids by hand.
The ids come from an empty search query, which orders by download count.
Downloading them reuses nugetfuzz, which already resolves versions, matches
target frameworks and walks the dependency closure into the same cache
decompdiff reads; --download-only stops it before it decompiles, since the
sweep is the expensive part and a corpus only needs the files.
The result is a list of lib directories rather than a single root, because a
package already restored on this machine is used from the machine-wide NuGet
cache instead of being copied into ours, and a corpus that silently omitted
those would misrepresent what was tested.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Comparing two commits meant preparing a worktree for each by hand before the
tool could be called, which is most of the work of running it and easy to get
wrong: a checkout carrying a stale Release build is reused silently, and the
timestamp in the header line was the only thing that said so.
--old and --new now also take anything git can resolve to a commit, checked
out into a worktree under ~/.cache/decompdiff keyed by that commit. The
worktrees are kept because the Release build inside one is what a rerun would
otherwise repeat: a second run of the same pair drops from minutes to seconds.
Paths keep priority over refs, so an existing directory never changes meaning.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Dock 12.1.0.6 removed the public JsonConverterFactoryList/JsonConverterList<T>
converters from Dock.Serializer.SystemTextJson that ILSpyDockJson used to
deserialize IList<T> into ObservableCollection<T>. Dock now does the same
substitution with an internal JsonTypeInfo modifier that swaps CreateObject
for IList<T> enumerables. ILSpyDockJson mirrors that technique in its own
modifier chain, which also removes the per-element JsonSerializer.Serialize
side effect the old converter had.
Also bumps Xaml.Behaviors.Avalonia, ProDataGrid, AwesomeAssertions, CliWrap,
NUnit3TestAdapter, and the decompiler minor version to 11.1.
Assisted-by: Claude:claude-fable-5:Claude Code
decompdiff indexed the Microsoft.NETFramework.ReferenceAssemblies packs into
one flat name map shared by the whole corpus, and never indexed
Microsoft.NETCore.App.Ref at all. Ordering the packs by name let net45 claim
mscorlib and System.Runtime, so a net9.0 assembly resolved its BCL against
.NET Framework 4.5: Task, ValueTask and the async method builders came back
as UnknownType, AsyncAwaitDecompiler could not match a state machine, and
every async method decompiled as a raw MoveNext.
Both sides of a diff degraded identically, so comparisons stayed valid, but
the corpus stopped representing modern code. Over 28 nuget assemblies the
same run reports 338 //IL_ warnings instead of 18202, 237 leaked <> names
instead of 15654, and 116k fewer lines.
The pack is now chosen per assembly from its TargetFrameworkAttribute and
searched before anything else, matching how nugetfuzz already resolves them.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Right_Clicking_A_Second_Row_Moves_The_Context_Highlight_To_It timed out on
the Windows CI agent waiting for the second context menu to open. A closed
popup's light-dismiss overlay keeps answering hit tests until the scene is
rendered again, and a press that lands on it raises no ContextRequested at
all, so the menu never opens. The test pumped a fixed four frames after
dismissing the first menu to get past that, which is a guess about how long
the overlay survives: enough on a fast machine, not on a loaded agent.
Hit testing the point is the same question the context-request handler asks,
so waiting for it to reach the row is the actual precondition for the click,
whatever number of frames that takes. The failure could not be reproduced
locally - removing the frames entirely still passes here - so this fixes the
documented mechanism rather than a reproduction, and a timeout now reports
which of the two conditions was not met.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
FindRefStructParameters dropped generic instantiations, so a parameter
typed 'ref <>c__DisplayClass0_0<T>' never reached RefStructTypes. Both
consumers therefore missed local functions whose declaring type or method
is generic: the signature test for an obfuscated local function, and
LocalFunctionNeedsAccessibilityChange, which left such a function internal
while its closure struct stayed private - the recompiled output then fails
with CS0051.
Cross-module signatures still drop out, because the generic type part of an
instantiation goes through GetTypeFromReference, which returns nil.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Whether an unresolvable type is a reference type is not a property of the
type but of the metadata that mentioned it: a signature spelling it
`valuetype T` yields false, a bare TypeRef yields null. UnknownType.Equals
compares the flag, so the two spellings of one missing type compared
unequal and EquivalentTypes reported false - the decompiler then emitted a
cast between a type and itself.
Erasing the flag in NormalizeTypeVisitor keeps the relaxation inside the
comparisons that ask for erasure, next to the nullability, modopt and tuple
erasure that are use-site spellings of the same kind. Dropping the term
from UnknownType.Equals instead was measured and rejected: Equals also keys
CSharpConversions' implicit-conversion cache, where merging the two
spellings lets whichever conversion is computed first answer for both,
adding 398 boxing casts across two real-world assemblies.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
This was due to StackType.O doing double-duty as `object` and `other`.
While ExpressionBuilder would often improve the type of such locals, the `object` nevertheless ended up used in a couple of places, e.g. via the `typeHint`. This could result in value types being boxed even though the original IL didn't contain any `box` instruction.
This is an attempt to use better types for stack slot variables created by ILReader. The idea is: there aren't many IL instructions that produce "other" value types, and `InferType()` already handles pretty much all of them, so we can use that to assign types to our stack slots.
It's a bit more tricky if the stack is pushed to on multiple branches that join together before the value is used: here the variable type must be suitable for both assignments. In this case, we go back to the previously-used stacktype.
ListBox.ScrollIntoView realises a container for its target, arranges it at its
own desired width, parks it aside for a few layout passes and then drops the
reference. A container those passes do not adopt back into the realized range
stays a visible child of the panel that nothing arranges again: it keeps
painting its old item, at its old position and its own narrow width, over
whatever row now occupies that spot - the ghost row drawn across another.
Rows here are a uniform height, so the offset a row's index implies reaches the
same place without ever entering that path. ScrollIntoView is hidden on
SharpTreeView so the obvious call lands on the safe one; hiding is not
enforcement, since a call through an ItemsControl-typed reference still reaches
the base method, but no call site has such a reference.
Assisted-by: Claude:claude-opus-5:Claude Code
An ID that resolves to no member left the tree on an empty selection with no
indication of what happened, because supplying --navigateto also suppresses the
single-assembly selection that opening a file otherwise makes. Only a target
that actually resolved should claim the selection; "none" still counts as
handled, since the VS add-in uses it to deliberately leave the tree empty.
The target arrives from a command line, so it goes through the omission-tolerant
search rather than exact resolution, and that can name several members. All of
them are selected. Landing on one would hide that there was a choice, and
falling back to the declaring type would bury the group in a large type's
decompilation - Enumerable.Where would decompile some two hundred members to
show four. The tree already multi-selects, so the overloads appear together at
the level the ID was pointing at.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Typing "M:System.Linq.Enumerable.Where" at a command line is a reasonable thing
to do, and it found nothing: resolution compares the whole id string, so a form
without the parameter list only ever matched a member that genuinely takes none.
Spelling the signature out is no answer, because it means knowing the overload
count before asking. The same goes for a generic arity - and the exact spelling,
Dictionary`2, does not even survive an unquoted bash prompt, where a backtick
starts command substitution.
None of that makes the short form legal. Measured against Roslyn: its own
DocumentationCommentId resolver accepts no abbreviation at all, and the compiler
never emits one - a cref is a different grammar, which the compiler binds and
rewrites into a full id, warning CS0419 and picking one member when the cref is
ambiguous. A prefixed cref is copied through unvalidated, so an id in a
documentation file can be anything a human typed.
So the id grammar stays exact and IdStringProvider stays with it, which is what
lets cref-following trust its answer. The tolerance belongs to the callers that
serve people typing, and lives in DocumentationIdSearch as a ladder that loosens
one thing at a time: the exact id, then the id without its parameter list, then
without generic arities. Stating a detail wrongly still finds nothing; only
leaving one out asks for any. A rung may match several members and all of them
are returned, because which to present is the caller's decision and hiding the
rest would hide that the id was ambiguous.
ilspycmd shows every member of the group, headed by a comment naming the
ambiguity, and accepts the shapes people actually type: no prefix, a shortened
namespace, and arity written the cref or C# way.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
This also matters in the `1 => DateTime.Now, 2 => null` case -- BestCommonType infers `DateTime` here, but we need `DateTime?` instead. But both had `StackType.O` so this went wrong prior to this commit.
* merge object/dynamic distinctions like we do with tuple element names. This fixes BestCommonType(object, dynamic).
* add a test that `new[] { 1, null }` has the "best common type" = `int`. The conversion error from `null` to `int` only happens later, it's not related to the best common type computation.
The previous wording described what an AI-assisted commit looks like
without saying the trailer is mandatory, which left it readable as a
convention for substantial changes only. The marker is what lets a
reader tell which work came from an agent, so it has to hold across all
of it -- an exemption for trivial commits makes the absence of the
trailer meaningless. The same reasoning closed a matching gap in how
agents comment on PRs, where a single disclaimer on a review body left
eleven inline comments unmarked.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
An agent review of PR #4065 posted twelve comments and eight had to be
withdrawn. Most were output of a decompdiff corpus sweep run against the
PR branch, which measures the whole decompiler rather than the diff, and
one called a cast a regression when it was the correct emission -- the
old form compiled only because C# target-types the switch expression.
The guide lives in .github/ rather than inline because most sessions
never review anything and should not carry it, and because it is as
useful to a human reviewer as to an agent. CLAUDE.md keeps only the two
rules that have to hold before you have decided you are reviewing at
all, plus the pointer.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
A Roslyn local function name is "<caller>g__name|x_y", where the trailing digits
are a synthetic disambiguator that SplitName has to strip before the scope-local
renumbering can run. An obfuscated name has no such suffix, so running it through
the same path renames "smethod_1" to "smethod_" -- a needless second mangling on
top of what the obfuscator already did.
#3202
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Obfuscators strip the CompilerGeneratedAttribute and rewrite the
"<caller>g__name|x_y" name, which is all IsLocalFunctionMethod had to go on.
The method then stays an ordinary static method, its display struct escapes by
ref into a plain call, and TransformDisplayClassUsage correctly refuses to
scalar-replace it -- so the closure fields leak into the output as
"<>c__DisplayClass29_0_.iid" (issue #3202).
The one marker an obfuscator cannot remove is the signature: Roslyn emits
struct closures exclusively for local functions, and no hand-written C# can
name a "<>c__DisplayClass" type, so a by-ref parameter of one identifies the
method regardless of what it is called.
#3202
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
The ilspycmd dump and the contributor notes described themselves by reference to
the UI's ILAst language; the pane is what walks that pipeline now.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
The file this branch adds takes the contributor's name rather than
AlphaSierraPapa, and three comments it added drop their en-GB spelling.
EndOpenGroups now requires its target depth: zero is the one value that closes
groups the caller does not own, which is the misattribution the depth argument
was added to prevent.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Three copies of the same pre-order walk across two test files become
TreeTraversal.PreOrder, and the stepper fixture builds its decompiler through
the file-name constructor instead of assembling the type system by hand.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
The ref parameter was only there because a partial method could not return one.
An extended partial can, so the caller reads the answer where it uses it. The
implementation stays behind DEBUG, which now needs a Release counterpart: an
extended partial must have one in every configuration, and there the step limit
is never set, so it never writes.
Assisted-by: Claude:claude-opus-5[1m]:Claude Code
Daniel Grunwald
Siegfried Pammer
Siegfried Pammer
Christoph Wille