Fix stack frame when using runtime async methods by eduardo-vp · Pull Request #127767 · dotnet/runtime

eduardo-vp · 2026-05-04T17:39:43Z

When building the stack frame in Native AOT, inject continuations if they exist and skip frames marked as hidden, matching CoreCLR behavior.

Works on top of #125396.

Fixes#122547.

output native sequence points wip WIP Testing More debug output testing Temp disable il offset lookup on exceptions Walk continuation resume chain and append to stackwalk data structure Use AsyncResumeILStubResolver to get the target method address Rebase fixes and collect native offsets for Env.StackTrace Inject ResumeData frames GetStackFramesData::Elements Truncate stack when async v2 continuation data is present Fix bad merge Addional fixes from previous merge Update to latest Continuation changes in main

Validates that Environment.StackTrace correctly includes async method frames and filters out internal DispatchContinuations frames when a runtime async method resumes via continuation dispatch. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

When executing inside a runtime async (v2) continuation dispatch, the stack trace is augmented by truncating internal dispatch frames and appending continuation DiagnosticIP values from the async continuation chain. This parallels the CoreCLR implementation in debugdebugger.cpp but operates on the NativeAOT IP-address-based stack trace model. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- Fix ExtractContinuationData to load AsyncDispatcherInfo type by name instead of searching the wrong MethodTable (RuntimeAsyncTask<T>) - Fix field name: t_dispatcherInfo -> t_current (matching managed code) - Fix type name: use strstr for RuntimeAsyncTask substring matching instead of exact match against non-existent RuntimeAsyncTaskCore - Add IsDynamicMethod() safety check before AsDynamicMethodDesc() to skip non-stub continuations (Task infrastructure callbacks) - Walk full dispatcher chain via Next pointer (was TODO with break) - Tighten NativeAOT type matching to AsyncHelpers+RuntimeAsyncTask - Update env-stacktrace test to verify OuterMethod appears via continuation injection (not just physical stack presence) - Update reflection tests to expect logical async caller at frame 1 when an async caller exists in the continuation chain Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…action - Register AsyncDispatcherInfo and its t_current field in corelib.h binder instead of using ClassLoader::LoadTypeByNameThrowing and manual field iteration - Use ContinuationObject::GetNext() and GetResumeInfo() accessors instead of hand-parsing fields via ApproxFieldDescIterator with string name matching - Add GetNext() accessor to ContinuationObject in object.h - GC-protect both continuation and pNext as CONTINUATIONREF in the gc struct across calls that may trigger GC - Define ResumeInfoLayout struct with documented field offsets matching the managed ResumeInfo struct layout Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…ameter, tighten boundary check - Use CoreLib binder DEFINE_FIELD entries for AsyncDispatcherInfo.Next and NextContinuation instead of hardcoded DispatcherInfoLayout struct (jkotas) - Remove unused pContinuationMT parameter from ExtractContinuationData (copilot-reviewer) - NativeAOT FindAsyncDispatchBoundary now also checks method name is DispatchContinuations, not just owning type (copilot-reviewer) Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…dation Replace runtime binder field offset lookups with direct struct member access via AsyncDispatcherInfoLayout. Add DEFINE_CLASS_U/DEFINE_FIELD_U entries so debug builds validate the native struct matches the managed layout at startup. Keep DEFINE_CLASS/DEFINE_FIELD for t_current TLS access which is still needed at runtime. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Remove (DWORD) cast on SIZE_T offset to avoid 64-bit truncation. Use already-fetched pFunc instead of redundant pCf null check and triple GetFunction() calls in GetStackFramesCallback. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Fix FindAsyncDispatchBoundary to use the fully qualified NativeAOT metadata type name for the DispatchContinuations boundary frame. NativeAOT metadata uses '.' as the nested type separator (e.g. 'System.Runtime.CompilerServices.AsyncHelpers.RuntimeAsyncTask`1') rather than '+', so the previous substring match never matched. Use exact string match on the fully qualified type name to avoid false matches with user-defined methods. Add AsyncStackTrace test to NativeAOT UnitTests smoke tests that validates OuterMethod -> MiddleMethod -> InnerMethod continuation stitching via Environment.StackTrace. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- Merge async stitching into InitializeForIpAddressArray to eliminate separate scan pass that double-formatted stack metadata (perf fix) - Detect dispatch boundary during frame construction using already-resolved StackFrame metadata (IsAsyncDispatchBoundary) at zero additional cost - Remove NativeAOT smoke test per reviewer guidance (duplicates library test) - Fix comment: use correct type name RuntimeAsyncTask, remove NativeAOT-specific framing Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- Check DiagnosticIP is non-null before appending to continuation list - Skip null pResumeIp entries before constructing EECodeInfo - Add pFunc null guard in else-if branch before dereferencing - Use exact strcmp with fully qualified type name instead of strstr substring Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

…c stitching - Replace expensive GetFullyQualifiedNameForClassNestedAware string formatting in debugdebugger.cpp with a MethodDesc pointer comparison via CoreLibBinder. Register RuntimeAsyncTask\1 and DispatchContinuations in corelib.h binder. - Fix NativeAOT FindAsyncDispatchBoundary to use '.' separator instead of '+' for nested type names in stack trace formatting. - Add shared EnvironmentStackTrace_AsyncContinuationStitching library test that validates async v2 continuation stitching on both CoreCLR and NativeAOT. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

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

…o assertion messages - Replace Task.Yield() with Task.Delay(1) for consistency with library test - Move stack trace output from Console.WriteLine to assertion failure messages to reduce CI log noise while preserving diagnostics on failure Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

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

Hide non-async frames from Environment.StackTrace once runtime async v2 frames are present, making pre-await and post-await traces consistent. DOTNET_HideAsyncDispatchFrames controls behavior: 0 = show all frames (no filtering) 1 = hide all non-async frames below first async frame (default) 2 = keep non-async frames between async frames, truncate trailing Implemented for both CoreCLR (C++ stack walk callback + post-processing) and NativeAOT (managed frame loop + IsAsyncMethod metadata flag). NativeAOT plumbing: IsAsyncMethod flag flows through StackTraceData command byte (0x20) -> ILC StackTraceRecordFlags -> metadata blob -> StackTraceMetadata decoder -> StackFrame.IsAsyncV2Method property. Tests: - env-stacktrace: mode 1 (default) pre/post-await consistency - env-stacktrace-nohide: mode 0 full stack visibility - env-stacktrace-truncate: mode 2 interim kept, trailing truncated - Library StackTraceTests: in-process DefaultOn test Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Update config and code comments to make clear that mode 0 shows all frames including non-async ones, but async continuation stitching (dispatch boundary truncation + continuation splicing) still occurs. This distinguishes it from PhysicalOnly which skips stitching entirely. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Drop the 'V2' implementation detail from the property name to align with reviewer feedback. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Add mode 3 which disables async continuation stitching entirely, returning only the physical call stack. This preserves the pre-async- stitching behavior as an opt-out for the breaking change. CoreCLR: skip async detection and continuation extraction in callback when mode=3. NativeAOT: skip CollectAsyncContinuationIPs when mode=3. Add env-stacktrace-physical test that verifies post-await traces do not contain suspended callers when stitching is disabled. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Extract inline string literals into constants AsyncDispatchMethodName and AsyncDispatchOwningType so there is only one place to update if the method is renamed. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Add comment explaining that MinimumFunctionAlignment = 4 guarantees bits 0 and 1 are clear in method RVAs, making them safe for flag packing. Extract FlagsMask constant for cleaner bit manipulation. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

On ARM32, RhFindMethodStartAddress adds the THUMB bit (bit 0) to method addresses. This bit flows into RVA computations used for BinarySearch lookups in StackTraceMetadata. Packing IsAsyncMethodFlag into bit 0 would collide with the THUMB bit, causing every ARM32 method to appear as an async method. Move IsAsyncMethod to a separate auto-property. IsHidden remains packed in bit 1, which is safe since THUMB only uses bit 0 and MinimumFunctionAlignment = 4 guarantees bit 1 is always zero. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

ExtractContinuationData and CollectAsyncContinuationIPs previously walked the entire AsyncDispatcherInfo.t_current linked list. Only the innermost (first) dispatcher represents the currently executing async scope; outer dispatchers represent already-completed scopes whose continuations should not appear in the stack trace. CoreCLR: add break after processing the first dispatcher with continuations. NativeAOT: replace the for loop with direct access to pInfo->NextContinuation. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Revert IsAsyncMethod from a separate bool field back to bit 0 of the packed RVA, per review feedback that StackTraceData is allocated for all method bodies and the extra field increases array size. Add GetAlignedMethodRva helper that masks off bits 0-1 (& ~0x3) at all three RVA computation sites. MinimumFunctionAlignment = 4 guarantees these bits are always zero for real method RVAs. On ARM32, the THUMB bit (bit 0) is present in addresses from RhFindMethodStartAddress and must be stripped before storage or lookup. Validated on ARM32 hardware (Raspberry Pi 3) — exit code 100. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Update config description to accurately reflect that mode 0 shows all frames above the dispatch boundary with async continuation stitching, not all frames unconditionally. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

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

… isolate tests - Cache CoreLibBinder::GetMethod(METHOD__RUNTIME_ASYNC_TASK__DISPATCH_CONTINUATIONS) in GetStackFramesData to avoid repeated binder lookups during stack walk. - Add explicit #include <sarray.h> in debugdebugger.h. - Pass computed hideMode into InitializeForIpAddressArray (NativeAOT) to eliminate redundant DOTNET_HideAsyncDispatchFrames environment read. - Wrap async stitching and frame hiding library tests in RemoteExecutor with explicit DOTNET_HideAsyncDispatchFrames=1 for test isolation. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

- Rename DOTNET_HideAsyncDispatchFrames to DOTNET_StackTraceAsyncBehavior - Update config description to cover both Environment.StackTrace and System.Diagnostics.StackTrace - Revert pDispatchContinuationsMD caching (CoreLibBinder already caches) - Remove unnecessary null checks: CoreLibBinder::GetField throws on failure, GetThread() is never null from managed code - Remove left-over mode description comment in NativeAOT - Pin DOTNET_StackTraceAsyncBehavior in all env-stacktrace test projects with RequiresProcessIsolation - Cache env var read in NativeAOT GetHideAsyncDispatchMode Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

… comparison - Cache DispatchContinuations method start address in StackTraceMetadata during first metadata resolution instead of matching names at runtime - Add IsAsyncDispatchBoundaryMethod virtual callback for address comparison - Add fallback in NativeAOT: if boundary not found (e.g. inlined), append continuation frames at end of stack trace - Use fully-qualified const strings for boundary method identification Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Exception stack traces should not be modified by DOTNET_StackTraceAsyncBehavior to maintain consistency with CoreCLR, where exception traces go through a separate path that does not apply async frame hiding. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Simplify StackTraceAsyncBehavior config to 3 modes: 0 = show all frames with async continuation stitching 1 = hide non-async frames below first async frame (default) 2 = physical only (no async stitching) The old mode 2 (keep interim non-async frames, truncate trailing) added complexity without clear value. Old mode 3 is renumbered to mode 2. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Copilot

Copilot encountered an error: Your billing is not configured or you have Copilot licenses from multiple standalone organizations or enterprises. To use premium requests, select a billing entity via the GitHub site, under Settings > Copilot > Features.

dotnet-policy-service · 2026-05-04T17:41:04Z

Tagging subscribers to this area: @agocke, @dotnet/ilc-contrib
See info in area-owners.md if you want to be subscribed.

Copilot

Pull request overview

Copilot reviewed 20 out of 20 changed files in this pull request and generated 4 comments.

Comments suppressed due to low confidence (1)

src/coreclr/nativeaot/System.Private.StackTraceMetadata/src/Internal/StackTraceMetadata/StackTraceMetadata.cs:120

Methods that are present in the reflection invoke table are intentionally omitted from the stack-trace mapping blob, so this reflection fallback is the only metadata source for them. Leaving isAsyncMethod hard-coded to false here means reflectable runtime-async methods will never be recognized as async in NativeAOT, which breaks the new hiding/stitching logic for public or otherwise reflectable async methods.

 isStackTraceHidden = false;
isAsyncMethod = false;
// We haven't found information in the stack trace metadata tables, but maybe reflection will have this
if (ReflectionExecution.TryGetMethodMetadataFromStartAddress(methodStartAddress,
out MetadataReader reader,
out TypeDefinitionHandle typeHandle,
out MethodHandle methodHandle))

+ pDispatcherInfo = (AsyncDispatcherInfoLayout*)pDispatcherInfo->Next;
+ continue;


+ // Identify and cache the async dispatch boundary method on first encounter.
+ if (_asyncDispatchBoundaryMethodAddress == IntPtr.Zero &&
+ isStackTraceHidden &&
+ methodName is AsyncDispatchBoundaryMethodName &&
+ owningTypeName is AsyncDispatchBoundaryOwningType)
+ {
+ _asyncDispatchBoundaryMethodAddress = methodStartAddress;


+ IntPtr[]? continuationIPs = StackTrace.CollectAsyncContinuationIPs();
+
+ int visibleIndex = 0;
+ for (int i = 0; i < totalFrameCount; i++)
+ {
+ IntPtr ip = frameArray[i];
+
+ if (stackTraceCallbacks != null)
+ {
+ IntPtr methodStart = RuntimeImports.RhFindMethodStartAddress(ip);
+ int offset = (int)((nint)ip - (nint)methodStart);
+ stackTraceCallbacks.TryGetMethodStackFrameInfo(methodStart, offset, false,
+ out _, out _, out _, out bool isHidden, out _, out _, out _);
+
+ bool isBoundary = stackTraceCallbacks.IsAsyncDispatchBoundaryMethod(methodStart);
+
+ // If this frame is the async dispatch boundary and we have continuations,
+ // inject continuations in place of remaining physical frames.
+ if (isBoundary && continuationIPs is not null)
+ {
+ for (int j = 0; j < continuationIPs.Length; j++)
+ {
+ if (visibleIndex == frameIndex)
+ {
+ ipAddress = continuationIPs[j];
+ break;
+ }
+ visibleIndex++;
+ }
+ InitializeForIpAddress(ipAddress, needFileInfo);


 AppDomain *pDomain;
 BOOL fDoWeHaveAnyFramesFromForeignStackTrace;
+ BOOL fAsyncFramesPresent; // True if async frames were present in the stack
+ DWORD hideAsyncDispatchMode; // 0 = show all above dispatch boundary (with stitching), 1 = hide non-async, 2 = truncate trailing, 3 = physical only (no stitching)


eduardo-vp · 2026-05-04T17:39:43Z