[NativeAOT] macOS/iOS: Emit simple compact unwinding information #88724

filipnavara · 2023-07-12T10:19:30Z

Apple platforms can represent certain unwinding sequences more efficiently using a single 32-bit "compact unwinding" code instead of a DWARF EH frame. This format cannot accurately unwind prologs and epilogs. However, we already handle epilogs manually since JIT doesn't generate the DWARF code for it, and prologs could be handled in a similar manner.

This is minimum viable prototype that implements the following:

Emits a single specific ARM64 unwind code for an empty frame prolog. Any other prolog would need changes in JIT to produce compatible prologs (eg. point frame pointer to the top of the frame, not bottom, store registers in correct order, always store registers in pairs).
Emit a prolog length into LSDA block if DWARF information is omitted. This allows the runtime to quickly identify if we are in the prolog or not.

This empty frame prolog accounts for about 30% of the DWARF code / methods in a typical executable, so it has significant impact even without handing the more complex cases.

Contributes to #76371, #88292

ghost · 2023-07-12T10:19:44Z

Tagging subscribers to this area: @agocke, @MichalStrehovsky, @jkotas
See info in area-owners.md if you want to be subscribed.

Issue Details

Apple platforms can represent certain unwinding sequences more efficiently using a single 32-bit "compact unwinding" code instead of a DWARF EH frame. This format cannot accurately unwind prologs and epilogs. However, we already handle epilogs manually since JIT doesn't generate the DWARF code for it, and prologs could be handled in a similar manner.

This is minimum viable prototype that implements the following:

Emits a single specific ARM64 unwind code for an empty frame prolog. Any other prolog would need changes in JIT to produce compatible prologs (eg. point frame pointer to the top of the frame, not bottom, store registers in correct order, always store registers in pairs).
Emit a prolog length into LSDA block if DWARF information is omitted. This allows the runtime to quickly identify if we are in the prolog or not.

This empty frame prolog accounts for about 30% of the DWARF code / methods in a typical executable, so it has significant impact even without handing the more complex cases.

Contributes to #76371, #88292

Author:	filipnavara
Assignees:	-
Labels:	`area-NativeAOT-coreclr`
Milestone:	-

src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/DependencyAnalysis/ObjectWriter.cs

jkotas · 2023-07-18T09:35:30Z

The general idea looks good to me. The size savings are very nice.

The only wrinkle is encoding of the prolog sizes. It looks odd that the prolog sizes are encoded explicitly, and the epilog sizes are inferred from the code. What would it take to infer the prolog size from the compact unwind info or by inspecting the code?

jkotas · 2023-07-18T09:37:16Z

What would it take to infer the prolog size from the compact unwind info or by inspecting the code?

It should be only needed during GetReturnAddressHijackInfo. It should not be needed during regular unwind. We should never call general unwind in prolog or epilog.

filipnavara · 2023-07-18T09:45:27Z

I originally inferred the prolog size from the instructions but it feels a bit fragile once you start doing complex prologs (not included in the PR). The compact unwind code by itself is not sufficient. It describes the stack layout but not which instructions were used to produce it. Due to various ways how ARM64 code can fold pointer adjustments into single instruction it would be fragile for the general case (again, not issue for this single unwind code that's produced in this PR).

If you feel strongly about inferring the prolog size from code I will do it for this simple case and we can extend it later to cover the hard cases.

jkotas · 2023-07-18T12:50:51Z

We took on this fragility in epilogs, so I think it would be fine to take on it in prologs too and save a bit more size.

…registers

filipnavara · 2023-08-02T16:03:22Z

We took on this fragility in epilogs, so I think it would be fine to take on it in prologs too and save a bit more size.

Done. Let's see if the CI passes... (but it seemed to work locally and crossgen2 compiled the framework just fine)

jkotas · 2023-08-02T19:25:32Z

/azp run runtime-extra-platforms

azure-pipelines · 2023-08-02T19:25:37Z

No commit pushedDate could be found for PR 88724 in repo dotnet/runtime

jkotas · 2023-08-02T22:18:14Z

/azp run runtime-extra-platforms

azure-pipelines · 2023-08-02T22:18:35Z

Azure Pipelines successfully started running 1 pipeline(s).

jkotas

LGTM. @VSadov Could you please take a look as well?

VSadov · 2023-08-02T23:25:52Z

@VSadov Could you please take a look as well?

I will take a look.

src/coreclr/nativeaot/Runtime/unix/UnixNativeCodeManager.cpp

VSadov

LGTM. (just a couple nits). Thanks!!!

VSadov · 2023-08-03T14:50:23Z

Thanks!

jkotas · 2023-08-04T04:58:45Z

This is causing a lot of intermittent crashes on Linux arm64 . For example: https://helixre8s23ayyeko0k025g8.blob.core.windows.net/dotnet-runtime-refs-pull-89969-merge-fdbcd47e91a0464083/System.Collections.Concurrent.Tests/1/console.07e57ce1.log?helixlogtype=result

In checked builds, the crash is typically assertion failure in UnixNativeCodeManager::GetReturnAddressHijackInfo:

* thread #1, name = 'Microsoft.Exten', stop reason = signal SIGABRT
  * frame #0: 0x0000ff5c2f85deac
    frame #1: 0x0000ff5c2f84aaa0
    frame #2: 0x0000aad8f9112abc Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`Assert(expr=<unavailable>, file=<unavailable>, line_num=<unavailable>, message=<unavailable>)
    frame #3: 0x0000aad8f9170dd0 Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`UnixNativeCodeManager::GetReturnAddressHijackInfo(this=0x0000000000000001, pMethodInfo=0x0000ff5c2f99ac00, pRegisterSet=0x0000ff1b91ffb0d0, ppvRetAddrLocation=0x0000ff5c2f99a000, pRetValueKind=0x0000ff1b91ff78bc) at P:810:5
    frame #4: 0x0000aad8f911dfb0 Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`Thread::HijackReturnAddressWorker(this=0x0000ff1b91ffb7f8, frameIterator=0x0000ff1b91ff7918, pfnHijackFunction=(Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`RhpGcProbeHijack))())
    frame #5: 0x0000aad8f911dbdc Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`Thread::HijackCallback(UNIX_CONTEXT*, void*) [inlined] Thread::HijackReturnAddress(this=0x0000ff1b91ffb7f8, pSuspendCtx=0x0000ff1b91ff7ca0)())
    frame #6: 0x0000aad8f911db6c Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`Thread::HijackCallback(pThreadContext=<unavailable>, pThreadToHijack=<unavailable>)
    frame #7: 0x0000aad8f9167810 Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`ActivationHandler(code=34, siginfo=0x0000ff1b91ff7c20, context=0x0000ff1b91ff7ca0)
    frame #8: 0x0000ff5c2fabe5b0 linux-vdso.so.1
    frame #9: 0x0000aad8f9508434 Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`S_P_CoreLib_System_Collections_Generic_List_1<System___Canon>__Sort_2(this=<unavailable>, comparison=<unavailable>) at List.cs:1091
    frame #10: 0x0000aad8f943c130 Microsoft.Extensions.Configuration.EnvironmentVariables.Tests`Microsoft_Extensions_Configuration_Microsoft_Extensions_Configuration_ConfigurationProvider__GetChildKeys(this=<unavailable>, earlierKeys=<unavailable>, parentPath=<unavailable>) at ConfigurationProvider.cs:91

…ion (#88724)" This reverts commit 5bf45ce.

filipnavara · 2023-08-04T06:54:36Z

This is causing a lot of intermittent crashes on Linux arm64 .

Thanks for alert, I will check. I must have messed up some condition because it should have had no effect on Linux.

filipnavara · 2023-08-04T07:17:27Z

I see the mistake now, will submit a fix shortly...

dotnet-issue-labeler bot added the area-NativeAOT-coreclr label Jul 12, 2023

ghost added the community-contribution Indicates that the PR has been added by a community member label Jul 12, 2023

filipnavara requested a review from VSadov July 12, 2023 10:19

filipnavara requested a review from ivanpovazan July 12, 2023 10:19

filipnavara commented Jul 12, 2023

View reviewed changes

src/coreclr/tools/aot/ILCompiler.Compiler/Compiler/DependencyAnalysis/ObjectWriter.cs Show resolved Hide resolved

build-analysis bot mentioned this pull request Jul 12, 2023

Timeout in Microsoft.Gen.OptionsValidation.Unit.Test.EmitterTest #88784

Closed

filipnavara marked this pull request as ready for review July 14, 2023 06:26

filipnavara requested a review from MichalStrehovsky as a code owner July 14, 2023 06:26

filipnavara requested a review from jkotas July 17, 2023 19:36

filipnavara and others added 3 commits August 2, 2023 17:00

Generate compact unwind sequence for methods with frame and no saved …

e9b1340

…registers

Store prolog length in LSDA data when compact unwinding is used

803f511

Remove prolog length, detect prologs from instruction stream

53117e3

filipnavara force-pushed the compact-unwind branch from 312cadd to 53117e3 Compare August 2, 2023 16:00

dotnet deleted a comment from azure-pipelines bot Aug 2, 2023

jkotas approved these changes Aug 2, 2023

View reviewed changes

VSadov reviewed Aug 3, 2023

View reviewed changes

src/coreclr/nativeaot/Runtime/unix/UnixNativeCodeManager.cpp Outdated Show resolved Hide resolved

VSadov approved these changes Aug 3, 2023

View reviewed changes

build-analysis bot mentioned this pull request Aug 3, 2023

Tracking issue for CI build timeouts #76454

Closed

Update comments

5446ab0

VSadov merged commit 5bf45ce into dotnet:main Aug 3, 2023
106 checks passed

filipnavara deleted the compact-unwind branch August 3, 2023 15:01

jkotas added a commit that referenced this pull request Aug 4, 2023

Revert "[NativeAOT] macOS/iOS: Emit simple compact unwinding informat…

549af5b

…ion (#88724)" This reverts commit 5bf45ce.

jkotas mentioned this pull request Aug 4, 2023

Revert "[NativeAOT] macOS/iOS: Emit simple compact unwinding information" #89994

Closed

filipnavara mentioned this pull request Aug 4, 2023

[NativeAOT] Fix Linux ARM64 checked build asserts #89998

Merged

ghost locked as resolved and limited conversation to collaborators Sep 3, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[NativeAOT] macOS/iOS: Emit simple compact unwinding information #88724

[NativeAOT] macOS/iOS: Emit simple compact unwinding information #88724

filipnavara commented Jul 12, 2023

ghost commented Jul 12, 2023

jkotas commented Jul 18, 2023

jkotas commented Jul 18, 2023

filipnavara commented Jul 18, 2023

jkotas commented Jul 18, 2023

filipnavara commented Aug 2, 2023

jkotas commented Aug 2, 2023

azure-pipelines bot commented Aug 2, 2023

jkotas commented Aug 2, 2023

azure-pipelines bot commented Aug 2, 2023

jkotas left a comment

VSadov commented Aug 2, 2023

VSadov left a comment

VSadov commented Aug 3, 2023

jkotas commented Aug 4, 2023

filipnavara commented Aug 4, 2023

filipnavara commented Aug 4, 2023

[NativeAOT] macOS/iOS: Emit simple compact unwinding information #88724

[NativeAOT] macOS/iOS: Emit simple compact unwinding information #88724

Conversation

filipnavara commented Jul 12, 2023

ghost commented Jul 12, 2023

jkotas commented Jul 18, 2023

jkotas commented Jul 18, 2023

filipnavara commented Jul 18, 2023

jkotas commented Jul 18, 2023

filipnavara commented Aug 2, 2023

jkotas commented Aug 2, 2023

azure-pipelines bot commented Aug 2, 2023

jkotas commented Aug 2, 2023

azure-pipelines bot commented Aug 2, 2023

jkotas left a comment

Choose a reason for hiding this comment

VSadov commented Aug 2, 2023

VSadov left a comment

Choose a reason for hiding this comment

VSadov commented Aug 3, 2023

jkotas commented Aug 4, 2023

filipnavara commented Aug 4, 2023

filipnavara commented Aug 4, 2023