[NativeAOT] Some cleanup of assembly helpers in hijack area. #72542

VSadov · 2022-07-20T19:13:48Z

Mostly removing stuff to simplify porting to unix-arm64

Deleting some dead code
Merged macros with too few uses into callers, where it made the code more clear or similar across platforms.
Making win-x64 to use the same shape of hijack helpers as other working implementations
(re:[NativeAOT] thread suspension on Unix #71187 (comment))
Making RhpGcProbeHijack to be similar between win-x64, unix-x64, win-arm64.
Especially the win-arm64 implementation that had many differences.
Added/tweaked a few comments

There should be no new functionality or any behavior changes here. Just deleting/refactoring code.

VSadov · 2022-07-22T03:51:21Z

src/coreclr/nativeaot/Runtime/unix/unixasmmacrosamd64.inc

@@ -320,7 +320,7 @@ C_FUNC(\Name):
 DEFAULT_FRAME_SAVE_FLAGS = PTFF_SAVE_ALL_PRESERVED + PTFF_SAVE_RSP

 .macro PUSH_COOP_PINVOKE_FRAME trashReg
-    push_nonvol_reg rbp                         // push RBP frame  // TODO: do we need this? not on windows.


I have checked and Windows debuggers do not have issues with unwinding without introducing rbp frame, so I did not change anything.

jkotas · 2022-07-22T03:56:46Z

src/coreclr/nativeaot/Runtime/thread.cpp

@@ -1413,7 +1352,7 @@ COOP_PINVOKE_HELPER(void, RhpReversePInvoke, (ReversePInvokeFrame * pFrame))
    if (pCurThread->InlineTryFastReversePInvoke(pFrame))
        return;

-    RhpReversePInvokeAttachOrTrapThread2(pFrame);
+    pCurThread->ReversePInvokeAttachOrTrapThread(pFrame);


Is the C++ code generated for this going to be as good as before? It looks like this was carefully setup to allow the slow path to be tailcalled.

The slow path would be very rare. It handles a case when a thread has not seen managed code yet.
I thought that the helper was added to call from assembly and now we don't

I will check if we are still tailcalling.

c++ actually inlines RhpReversePInvokeAttachOrTrapThread2 after the change.
The impact on the fast part of the method is not too bad, but keeping the slow path separate still has some minor benefits like pushing fewer registers in prolog.

I will revert this part.

jkotas · 2022-07-22T03:57:32Z

src/coreclr/nativeaot/Runtime/thread.cpp

-    for (size_t i = 0; i < ARRAY_SIZE(NormalHijackTargets); i++)
-    {
-        if (NormalHijackTargets[i] == address)
+        if (&RhpGcProbeHijack == address)


The indentation looks off.

VSadov · 2022-07-22T04:01:15Z

src/coreclr/nativeaot/Runtime/arm64/GcProbe.asm

    ;; Build a map of symbols representing offsets into the transition frame (see PInvokeTransitionFrame in
    ;; rhbinder.h) and keep these two in sync.
    map 0
            field OFFSETOF__PInvokeTransitionFrame__m_PreservedRegs
            field 10 * 8 ; x19..x28
 m_CallersSP field 8      ; SP at routine entry
-            field 19 * 8 ; x0..x18


There is no point in saving/restoring volatile registers as at return site they contain garbage.
lr,fp are saved as a part of stack alloc/dealloc, so no need to save lr again.
NZCV was not actually stored and there is no need.

We do need to save returns though as they may have live data.

jkotas · 2022-07-22T04:25:55Z

src/coreclr/nativeaot/Runtime/amd64/GcProbe.S

+        jnz         LOCAL_LABEL(DoRhpGcProbe)
+        ret
+
+LOCAL_LABEL(DoRhpGcProbe):
        or          ecx, DEFAULT_FRAME_SAVE_FLAGS  + PTFF_SAVE_RAX + PTFF_SAVE_RDX
        jmp         C_FUNC(RhpGcProbe)
 NESTED_END RhpGcProbeHijack, _TEXT

 NESTED_ENTRY RhpGcProbe, _TEXT, NoHandler


RhpGcProbe does not match what the funciton does now that RhpTrapThreads check was removed from it. Should it be called RhpWaitForGC instead?

RhpGcProbe was not the best name.
The actual purpose of the function is to set up the stackwalk-enabling frame around the RhpWaitForGC2, which does the actual waiting. It is hard to capture that in couple words.

RhpWaitForGC seems a bit better. The function does typically result in waiting for GC.

jkotas · 2022-07-22T04:29:02Z

src/coreclr/nativeaot/Runtime/arm/GcProbe.asm

@@ -326,8 +286,8 @@ __PPF_ThreadReg SETS "r2"
    NESTED_ENTRY RhpGcProbeRare


Delete RhpGcProbe and rename RhpGcProbeRare to RhpGcProbe to match other platforms?

I did not put much effort in keeping 32bit platforms in sync. I only did some easy opportunistic changes.

I think we may be iterating on 64bit implementation a bit more and keeping 32bit behind is not a big deal. When/if we get to 32bit, absorbing all the changes should not be hard. It might even be that porting 64bit code, that works and tested, would be a shorter path than adjusting existing 32bit code. Probably it will be a mix of both approaches.

I will check if these changes are easy to do.

I have fixed the entry points so that all platforms look the same - one RhpGcProbeHijack that checks for the trap flag, and if set, sets bits for saved registers and calls RhpWaitForGC. It might even work when we get to 32bit.

jkotas · 2022-07-22T04:30:21Z

src/coreclr/nativeaot/Runtime/i386/GcProbe.asm

@@ -198,7 +162,8 @@ RhpGcProbe  proc
 SynchronousRendezVous:


Delete RhpTrapThreads check to match other platforms?

jkotas

LGTM modulo comments. Thank you!

(as much as can be done without trying to compile and test)

VSadov · 2022-07-22T07:45:21Z

/azp run runtime-extra-platforms

azure-pipelines · 2022-07-22T07:45:43Z

Azure Pipelines successfully started running 1 pipeline(s).

VSadov · 2022-07-22T14:44:12Z

Thanks!!

VSadov added 5 commits July 19, 2022 08:07

remove RhpReversePInvokeAttachOrTrapThread (dead code)

dd9382e

remove RhpWaitForGC

c60f01f

removing some dead code

45dfbfc

remove unused extraStack parameter to PUSH_PROBE_FRAME

eaf6eff

make all working variants of RhpGcProbeHijack to have the same shape

1f29db1

dotnet-issue-labeler bot added the area-NativeAOT-coreclr label Jul 20, 2022

ghost assigned VSadov Jul 20, 2022

VSadov added 5 commits July 20, 2022 13:00

fix Unix build, tweak some comments

7d66dd2

check the trap flag in RhpGcProbeHijack

23fdd0a

add a stub for RhpGcStressHijack on Unix

8c23ae6

save flags for windows x64 should not have rdx

8faffe2

not saving scratch registers on win-arm64

75a01b9

dotnet deleted a comment from azure-pipelines bot Jul 21, 2022

PUSH_PROBE_FRAME on arm64

c321a96

VSadov force-pushed the cleanup1 branch from 201506a to c321a96 Compare July 21, 2022 07:01

dotnet deleted a comment from azure-pipelines bot Jul 21, 2022

VSadov added 2 commits July 21, 2022 11:56

couple comments

e8eea2b

make RhpGcProbeHijack responsible for setting PTFF_SAVE_ bits

d305471

dotnet deleted a comment from azure-pipelines bot Jul 21, 2022

VSadov marked this pull request as ready for review July 21, 2022 20:11

VSadov requested a review from MichalStrehovsky as a code owner July 21, 2022 20:11

VSadov requested review from janvorli and jkotas July 22, 2022 03:49

VSadov commented Jul 22, 2022

View reviewed changes

jkotas reviewed Jul 22, 2022

View reviewed changes

VSadov commented Jul 22, 2022

View reviewed changes

jkotas reviewed Jul 22, 2022

View reviewed changes

jkotas approved these changes Jul 22, 2022

View reviewed changes

VSadov added 4 commits July 21, 2022 23:30

revert RhpReversePInvokeAttachOrTrapThread2 change

bc03016

fix indentation

7b260ac

made 32bit RhpGcStressHijack similar to 64bit counterparts

38e9c17

(as much as can be done without trying to compile and test)

Renamed RhpGcProbe --> RhpWaitForGC

f16bb29

dotnet deleted a comment from azure-pipelines bot Jul 22, 2022

VSadov merged commit 41def7c into dotnet:main Jul 22, 2022

VSadov deleted the cleanup1 branch July 22, 2022 14:44

ghost locked as resolved and limited conversation to collaborators Aug 21, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[NativeAOT] Some cleanup of assembly helpers in hijack area. #72542

[NativeAOT] Some cleanup of assembly helpers in hijack area. #72542

VSadov commented Jul 20, 2022 •

edited

Loading

VSadov Jul 22, 2022 •

edited

Loading

jkotas Jul 22, 2022

VSadov Jul 22, 2022

VSadov Jul 22, 2022

jkotas Jul 22, 2022

VSadov Jul 22, 2022 •

edited

Loading

VSadov Jul 22, 2022

jkotas Jul 22, 2022

VSadov Jul 22, 2022

jkotas Jul 22, 2022

VSadov Jul 22, 2022

VSadov Jul 22, 2022

jkotas Jul 22, 2022

jkotas left a comment

VSadov commented Jul 22, 2022

azure-pipelines bot commented Jul 22, 2022

VSadov commented Jul 22, 2022

		@@ -326,8 +286,8 @@ __PPF_ThreadReg SETS "r2"
		NESTED_ENTRY RhpGcProbeRare

[NativeAOT] Some cleanup of assembly helpers in hijack area. #72542

[NativeAOT] Some cleanup of assembly helpers in hijack area. #72542

Conversation

VSadov commented Jul 20, 2022 • edited Loading

VSadov Jul 22, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

VSadov Jul 22, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jkotas left a comment

Choose a reason for hiding this comment

VSadov commented Jul 22, 2022

azure-pipelines bot commented Jul 22, 2022

VSadov commented Jul 22, 2022

VSadov commented Jul 20, 2022 •

edited

Loading

VSadov Jul 22, 2022 •

edited

Loading

VSadov Jul 22, 2022 •

edited

Loading