Change gtGetThisArg not to return nullptr. #44398

sandreenko · 2020-11-09T03:37:05Z

bb0c5dc: Don't wrap string literal const as nop for CoreRT.

It was probably an old workaround for another Jit bug, it is most likely fixed by now.

2ae10db1383: Change gtGetThisArg not to return nullptr.

There was only 1 case where a null return was tolerated - for a tail call marked as virtual in optAssertionGen.
Check this case before we call gtGetThisArg.

Fixes dotnet/runtimelab#298

sandreenko · 2020-11-09T05:57:20Z

src/coreclr/src/jit/assertionprop.cpp

-            if ((tree->gtFlags & GTF_CALL_NULLCHECK) || tree->AsCall()->IsVirtual())
+            // Ignore tail calls because they have 'this` pointer in the regular arg list and an implicit null check.
+            GenTreeCall* const call = tree->AsCall();
+            if (call->NeedsNullCheck() || (call->IsVirtual() && !call->IsTailCall()))


In fgMorphTailCallViaJitHelper we clear NeedsNullCheck, should not we also clear IsVirtual?
Then this condition will be just if (call->NeedsNullCheck() || call->IsVirtual())

@jakobbotsch could you please advise me here?

Sorry I don't know. fgMorphTailCallViaJitHelper is the old (from before #341) tailcall mechanism used only on x86. This mechanism might be used directly for virtual calls (or VSDs), but I don't remember how the address is computed in those cases.

sandreenko · 2020-11-09T08:57:35Z

PTAL @dotnet/jit-contrib

sandreenko · 2020-11-09T20:30:08Z

/azp run runtime-coreclr outerloop

azure-pipelines · 2020-11-09T20:30:30Z

Azure Pipelines successfully started running 1 pipeline(s).

sandreenko · 2020-11-09T21:03:36Z

src/coreclr/src/jit/morph.cpp

@@ -7783,6 +7783,12 @@ GenTree* Compiler::fgMorphTailCallViaHelpers(GenTreeCall* call, CORINFO_TAILCALL
        call->gtFlags &= ~GTF_CALL_VIRT_STUB;


We do it for stub and in my understanding for vtable we should do the same, let's see what ci thinks.

This is fgMorphTailCallViaHelpers which is for the new mechanism. Indeed those turn into normal direct calls always.

it looks like I will need the same fix for fgMorphTailCallViaJitHelper: sandreenko@53cea4e , but I want to see if our ci catches it.

It looks like we remove all virtual kinds below:

runtime/src/coreclr/src/jit/morph.cpp

Lines 7917 to 7921 in dc3101b

// This is now a direct call to the store args stub and not a tailcall.

call->gtCallType = CT_USER_FUNC;

call->gtCallMethHnd = help.hStoreArgs;

call->gtFlags &= ~GTF_CALL_VIRT_KIND_MASK;

call->gtCallMoreFlags &= ~(GTF_CALL_M_TAILCALL | GTF_CALL_M_DELEGATE_INV | GTF_CALL_M_WRAPPER_DELEGATE_INV);

I guess this line is not required but doesn't hurt since this is where we remove the arg for virtual stubs.

I see, thank you. I removed changes in fgMorphTailCallViaHelpers and fixed fgMorphTailCallViaJitHelper in the same way.

It was probably an old workaround for another Jit bug, it is most likely fixed by now.

There was only 1 case where a null return was tolerated - for a tail call in `optAssertionGen` marked as virtual. However, a transformed tail call is never a virtual, fix `fgMorphTailCallViaJitHelper` to unset virtual flag.

sandreenko · 2020-11-10T22:43:09Z

/azp run runtime-coreclr outerloop

azure-pipelines · 2020-11-10T22:43:28Z

Azure Pipelines successfully started running 1 pipeline(s).

sandreenko · 2020-11-11T06:29:52Z

src/coreclr/src/jit/morph.cpp

@@ -8435,6 +8435,10 @@ void Compiler::fgMorphTailCallViaJitHelper(GenTreeCall* call)
            thisPtr = objp;
        }

+        // TODO-Cleanup: we leave it as a virtual stub call to


I made a try to clear GTF_CALL_VIRT_KIND_MASK here but it caused failures because logic in LowerCall depends on this flag to be set in some cases, so I reverted that change.

It would be nice to clear it and fix the failures but it is out of the scope of this bug fix for CoreRT.

sandreenko · 2020-11-14T01:46:26Z

ping @dotnet/jit-contrib

sandreenko · 2020-11-19T00:40:50Z

Thanks @BruceForstall for the review!

sandreenko added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Nov 9, 2020

sandreenko force-pushed the fixCoreRTConstStr branch from f1158f0 to 84b1152 Compare November 9, 2020 05:53

sandreenko commented Nov 9, 2020

View reviewed changes

sandreenko force-pushed the fixCoreRTConstStr branch from 84b1152 to 671fedd Compare November 9, 2020 20:29

sandreenko commented Nov 9, 2020

View reviewed changes

Don't wrap string literal const as nop for CoreRT.

bb0c5dc

It was probably an old workaround for another Jit bug, it is most likely fixed by now.

sandreenko force-pushed the fixCoreRTConstStr branch from 53cea4e to 2ae10db Compare November 10, 2020 00:21

sandreenko changed the title ~~Fix CoreRT const string handling.~~ Change gtGetThisArg not to return nullptr. Nov 10, 2020

Sergey Andreenko added 2 commits November 10, 2020 10:58

Change gtGetThisArg not to return nullptr.

007f31f

There was only 1 case where a null return was tolerated - for a tail call in `optAssertionGen` marked as virtual. However, a transformed tail call is never a virtual, fix `fgMorphTailCallViaJitHelper` to unset virtual flag.

Keep x86 tail call via jit helper as a virtual stub call.

8663da9

sandreenko force-pushed the fixCoreRTConstStr branch from 2ae10db to 8663da9 Compare November 10, 2020 19:09

sandreenko commented Nov 11, 2020

View reviewed changes

runfoapp bot mentioned this pull request Nov 11, 2020

Runtime Tests Linux_musl arm fails with "The "FindDotNetCliPackage" task failed unexpectedly." #44541

Closed

BruceForstall approved these changes Nov 19, 2020

View reviewed changes

sandreenko merged commit 044ee8c into dotnet:master Nov 19, 2020

jkotas mentioned this pull request Nov 21, 2020

Merge from dotnet/runtime dotnet/runtimelab#374

Merged

sandreenko mentioned this pull request Nov 23, 2020

Fix CoreRT frozen strings handling. #45095

Merged

ghost locked as resolved and limited conversation to collaborators Dec 19, 2020

sandreenko deleted the fixCoreRTConstStr branch May 17, 2021 07:50

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Change gtGetThisArg not to return nullptr. #44398

Change gtGetThisArg not to return nullptr. #44398

sandreenko commented Nov 9, 2020 •

edited

Loading

sandreenko Nov 9, 2020

sandreenko Nov 9, 2020

jakobbotsch Nov 9, 2020

sandreenko commented Nov 9, 2020

sandreenko commented Nov 9, 2020

azure-pipelines bot commented Nov 9, 2020

sandreenko Nov 9, 2020

jakobbotsch Nov 9, 2020

sandreenko Nov 9, 2020

jakobbotsch Nov 9, 2020

sandreenko Nov 10, 2020

sandreenko commented Nov 10, 2020

azure-pipelines bot commented Nov 10, 2020

sandreenko Nov 11, 2020

sandreenko commented Nov 14, 2020

sandreenko commented Nov 19, 2020

		@@ -7783,6 +7783,12 @@ GenTree* Compiler::fgMorphTailCallViaHelpers(GenTreeCall* call, CORINFO_TAILCALL
		call->gtFlags &= ~GTF_CALL_VIRT_STUB;

	// This is now a direct call to the store args stub and not a tailcall.
	call->gtCallType = CT_USER_FUNC;
	call->gtCallMethHnd = help.hStoreArgs;
	call->gtFlags &= ~GTF_CALL_VIRT_KIND_MASK;
	call->gtCallMoreFlags &= ~(GTF_CALL_M_TAILCALL \| GTF_CALL_M_DELEGATE_INV \| GTF_CALL_M_WRAPPER_DELEGATE_INV);

Change gtGetThisArg not to return nullptr. #44398

Change gtGetThisArg not to return nullptr. #44398

Conversation

sandreenko commented Nov 9, 2020 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sandreenko commented Nov 9, 2020

sandreenko commented Nov 9, 2020

azure-pipelines bot commented Nov 9, 2020

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

sandreenko commented Nov 10, 2020

azure-pipelines bot commented Nov 10, 2020

Choose a reason for hiding this comment

sandreenko commented Nov 14, 2020

sandreenko commented Nov 19, 2020

sandreenko commented Nov 9, 2020 •

edited

Loading