Do not encode safe points with -1 offset. #104336

VSadov · 2024-07-03T00:41:37Z

When the code position after a call site is a safe point, its GC info is now applicable regardless of how instruction pointer has arrived at this position - by executing a call (and unwinding), returning from a call or jumping around the call,...
That holds in both fully and partially interruptible code.

Therefore leaf/nonleaf/faulted-in-partial/faulted-in-interruptible scenarios all can work the same and we no longer need to resort to various disambiguation schemes.
In particular we do not need to pre-adjust safe point positions in partially interruptible code by -1, into the middle of a previous instruction, and then carefully compensate for that in bunch of places.

An additional benefit is that code offsets in GC info are now all instruction-aligned, thus on RISC architectures fewer bits could be used to encode those.

Fixes: #5677

======== Size impact of this change on arm64:
Using NativeAOT compiled System.Collections.Concurrent.Tests.exe for Win-arm64 as an example

Before the change:
21,920,768 bytes

After the change:
21,779,456 bytes

executable size reduction is roughly 141K ( 0.6% )

NB: the change does not affect the legacy X86 GC info

VSadov · 2024-07-03T19:47:43Z

/azp run runtime-coreclr gcstress0x3-gcstress0xc

azure-pipelines · 2024-07-03T19:47:52Z

Azure Pipelines successfully started running 1 pipeline(s).

VSadov · 2024-07-04T01:06:27Z

These are timing out currently in other PRs as well.

runtime (Build linux-x64 Debug NativeAOT) Cancelled after 120m — Build linux-x64 Debug NativeAOT was canceled
runtime (Build linux-x64 Release NativeAOT) Cancelled after 120m — Build linux-x64 Release NativeAOT was canceled

The rest is green. Including gcstress0x3-gcstress0xc

VSadov · 2024-07-04T01:23:02Z

CC: @dotnet/jit-contrib

VSadov · 2024-07-04T01:23:42Z

CC: @jkotas @janvorli

VSadov · 2024-07-04T01:38:30Z

src/coreclr/jit/codegencommon.cpp

-            {
-                noway_assert(varDsc->lvIsParam);
-
-                gcInfo.gcMarkRegPtrVal(varDsc->GetArgReg(), varDsc->TypeGet());


We should have correct GC info at the label after GS cookie check as cookie check has no effect on GC liveness. (maybe it did in the past?)

This was just noop or was messing things up by forcing argument registers to be GC-alive when they would not have GC references in them. Perhaps it was compensating for some behavior that is long gone.

This was hidden because we did not allow interrupting thread on this instruction in a leaf frame (but we can and should) and in nonleaf this would not show up because this is a tailcall.

VSadov · 2024-07-04T01:45:11Z

src/coreclr/jit/emitinl.h

@@ -594,8 +594,7 @@ bool emitter::emitGenNoGCLst(Callback& cb)
            emitter::instrDesc* id = emitFirstInstrDesc(ig->igData);
            assert(id != nullptr);
            assert(id->idCodeSize() > 0);
-            if (!cb(ig->igFuncIdx, ig->igOffs, ig->igSize, id->idCodeSize(),
-                    ig->igFlags & (IGF_FUNCLET_PROLOG | IGF_FUNCLET_EPILOG | IGF_EPILOG)))


I realized that including epilogs in this set was just compensating for the bug related to GS cookies and for overly conservative asserts in NativeAOT hijacker.

It is fairly common on ARM64 to see:

bl <somewhere> ldp <some regs> // <-- this IP is certainly unwindable, how else can we unwind from the call above? ldp <more regs> ldp fp, lr, [sp], #0x60

The first instruction of an epilog is no different from a first instruction in a regular no-GC region.
We have not executed any instructions yet that destroy unwindability or GC-safeness.

jkotas · 2024-07-04T02:18:06Z

src/coreclr/inc/gcinfotypes.h

@@ -678,8 +678,8 @@ void FASTCALL decodeCallPattern(int         pattern,
 #define NORMALIZE_SIZE_OF_STACK_AREA(x) ((x)>>2)
 #define DENORMALIZE_SIZE_OF_STACK_AREA(x) ((x)<<2)
 #define CODE_OFFSETS_NEED_NORMALIZATION 1
-#define NORMALIZE_CODE_OFFSET(x) (x)   // Instructions are 2/4 bytes long in Thumb/ARM states,


Matching change will need to be made in https://github.com/dotnet/diagnostics/tree/main/src/shared/gcdump so that GC info dumping in SOS works.

There is a challenge though: SOS wants to support all runtimes, so the SOS copy will need to work for both old and new GC info encoding - we will need to switch based on the current runtime version somehow.

Also, same problem exists for https://github.com/dotnet/runtime/tree/main/src/coreclr/tools/aot/ILCompiler.Reflection.ReadyToRun/Amd64

Also, same problem exists for https://github.com/dotnet/runtime/tree/main/src/coreclr/tools/aot/ILCompiler.Reflection.ReadyToRun/Amd64

There is no corresponding location for arm[64]. Do we use Amd64 for all cases?
It is the RISC arches that are more interesting, since the encoding size changes.

Figured that:

// Arm, Arm64, LoongArch64 and RISCV64 use the same GcInfo format as Amd64

The arch specific behavior is done via consulting with _readyToRunReader.Machine`

I think there are no changes needed on x64 for the uses of GcInfo that do enumeration. The offsets are different, but that is a matter of FindSafePoint(offset) kind of use, as the offset would need to no longer do -1 adjustment in nonleaf cases. I do not see such use in ReadyToRun thing.

Basically - for the use like foreach(safePoint in ...) {foreach(slot in SlotsAt(safePoint)} , on x64 everything should work as before.

The RISC is more interesting as GcInfo deserializer would need to denormalize code offsets based on GcInfo version. Not just the offsets used in safepoints, there are other places.

I would expect that GcInfo deserializer that wants to support different runtimes already needs to figure the version of GCInfo. This is not the first time it changes.

The diagnostics/SOS GcInfo deserializer looks like just a copy of the one in the runtime. It is version-aware (can handle v1 and v2) and gets the version from the supplied token.

We may just need to copy/paste the new version once this change goes in.

Logged a follow-up issue for that: dotnet/diagnostics#4795

jkotas

Nice simplification!

@dotnet/jit-contrib Could you please review the JIT changes?

AndyAyersMS

JIT changes LGTM

VSadov · 2024-07-17T17:47:30Z

Thanks!!

This reverts commit fa77959.

* Partial re-revert of #104336. Only JIT fixes are included. * fix for stress issues * comments and some cleanup * JIT format * keep existing code for x86

VSadov added the area-VM-coreclr label Jul 3, 2024

dotnet-policy-service bot assigned VSadov Jul 3, 2024

VSadov added 7 commits July 2, 2024 23:10

Do not record safe points with -1 adjustmnt

4bf5e03

fix for arm64

7ecaab7

suppress an assert

91bdf63

comments about emitDisableGC

09e0dfb

JIT format

7d2fd22

NORMALIZE_CODE_OFFSET on RISC

ee1a050

JIT format again

8b84d1f

VSadov force-pushed the noIPadj branch from f4d93d2 to 8b84d1f Compare July 3, 2024 06:10

This was referenced Jul 3, 2024

The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#3008

Open

TimeProviderTests.TestProviderTimer failed in CI #103459

Closed

The job running on agent NetCore-Public ran longer than the maximum time #104044

Closed

do not resurrect random regs after GS check

649fabb

dotnet deleted a comment from azure-pipelines bot Jul 3, 2024

VSadov marked this pull request as ready for review July 4, 2024 01:22

VSadov requested a review from MichalStrehovsky as a code owner July 4, 2024 01:22

VSadov commented Jul 4, 2024

View reviewed changes

jkotas reviewed Jul 4, 2024

View reviewed changes

denormalize code offsets in ILCompiler.Reflection.ReadyToRun.Amd64

698dc61

VSadov mentioned this pull request Jul 16, 2024

Update GCInfo deserializer once https://github.com/dotnet/runtime/pull/104336 merges dotnet/diagnostics#4795

Open

jkotas approved these changes Jul 17, 2024

View reviewed changes

AndyAyersMS approved these changes Jul 17, 2024

View reviewed changes

VSadov merged commit fa77959 into dotnet:main Jul 17, 2024
107 checks passed

VSadov deleted the noIPadj branch July 17, 2024 17:48

jakobbotsch added a commit to jakobbotsch/runtime that referenced this pull request Jul 22, 2024

Revert "Do not encode safe points with -1 offset. (dotnet#104336)"

af8c618

This reverts commit fa77959.

jakobbotsch mentioned this pull request Jul 22, 2024

Revert "Do not encode safe points with -1 offset. (#104336)" #105234

Merged

jakobbotsch added a commit that referenced this pull request Jul 22, 2024

Revert "Do not encode safe points with -1 offset. (#104336)" (#105234)

c881a7a

This reverts commit fa77959.

janvorli mentioned this pull request Jul 23, 2024

Test failure: baseservices/exceptions/unhandled/unhandledTester/unhandledTester.cmd #105229

Closed

VSadov added a commit to VSadov/runtime that referenced this pull request Jul 27, 2024

Partial re-revert of dotnet#104336. Only JIT fixes are included.

de17d7c

VSadov added a commit to VSadov/runtime that referenced this pull request Jul 31, 2024

Partial re-revert of dotnet#104336. Only JIT fixes are included.

310a7a3

VSadov added a commit to VSadov/runtime that referenced this pull request Jul 31, 2024

Partial re-revert of dotnet#104336. Only JIT fixes are included.

82ab1b9

VSadov added a commit to VSadov/runtime that referenced this pull request Aug 1, 2024

Partial re-revert of dotnet#104336. Only JIT fixes are included.

50adb34

VSadov added a commit to VSadov/runtime that referenced this pull request Aug 1, 2024

Partial re-revert of dotnet#104336. Only JIT fixes are included.

699219a

krwq mentioned this pull request Aug 9, 2024

[9.0.100-preview.7.24375.12] Markdig nuget package thrown "The parser is in an invalid infinite loop while trying to parse inlines for block..." exception when parsering markdown string to html #105792

Closed

1 task

github-actions bot locked and limited conversation to collaborators Aug 17, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Do not encode safe points with -1 offset. #104336

Do not encode safe points with -1 offset. #104336

VSadov commented Jul 3, 2024 •

edited

Loading

VSadov commented Jul 3, 2024

azure-pipelines bot commented Jul 3, 2024

VSadov commented Jul 4, 2024

VSadov commented Jul 4, 2024 •

edited

Loading

VSadov commented Jul 4, 2024

VSadov Jul 4, 2024 •

edited

Loading

VSadov Jul 4, 2024

VSadov Jul 4, 2024 •

edited

Loading

jkotas Jul 4, 2024

jkotas Jul 4, 2024

VSadov Jul 16, 2024 •

edited

Loading

VSadov Jul 16, 2024 •

edited

Loading

VSadov Jul 16, 2024

VSadov Jul 17, 2024

jkotas left a comment

AndyAyersMS left a comment

VSadov commented Jul 17, 2024

Do not encode safe points with -1 offset. #104336

Do not encode safe points with -1 offset. #104336

Conversation

VSadov commented Jul 3, 2024 • edited Loading

VSadov commented Jul 3, 2024

azure-pipelines bot commented Jul 3, 2024

VSadov commented Jul 4, 2024

VSadov commented Jul 4, 2024 • edited Loading

VSadov commented Jul 4, 2024

VSadov Jul 4, 2024 • edited Loading

Choose a reason for hiding this comment

VSadov Jul 4, 2024

Choose a reason for hiding this comment

VSadov Jul 4, 2024 • edited Loading

Choose a reason for hiding this comment

jkotas Jul 4, 2024

Choose a reason for hiding this comment

jkotas Jul 4, 2024

Choose a reason for hiding this comment

VSadov Jul 16, 2024 • edited Loading

Choose a reason for hiding this comment

VSadov Jul 16, 2024 • edited Loading

Choose a reason for hiding this comment

VSadov Jul 16, 2024

Choose a reason for hiding this comment

VSadov Jul 17, 2024

Choose a reason for hiding this comment

jkotas left a comment

Choose a reason for hiding this comment

AndyAyersMS left a comment

Choose a reason for hiding this comment

VSadov commented Jul 17, 2024

VSadov commented Jul 3, 2024 •

edited

Loading

VSadov commented Jul 4, 2024 •

edited

Loading

VSadov Jul 4, 2024 •

edited

Loading

VSadov Jul 4, 2024 •

edited

Loading

VSadov Jul 16, 2024 •

edited

Loading

VSadov Jul 16, 2024 •

edited

Loading