[swift-dev] DebugScope Verification Error

[Raj Barik]

Thanks a lot for the clarification Adrain and Vedant.

What confused me is the cycle -- A SILFunction has a DeubgScope  and the
DebugScope points back to the SILFunction. During moving instructions
across functions, one can easily mess up the pointers. As you all pointed
out, one needs to also use ScopeCloner to get around this problem.

Does it make sense not to have the cycle, i.e., DebugScopes are made
independent of SILFunction so that multiple SILFunction(s) can share the
same DebugScope?


On Tue, Jan 16, 2018 at 1:26 PM, Adrian Prantl <aprantl at apple.com> wrote:

>
>
> On Jan 16, 2018, at 1:20 PM, Vedant Kumar <vsk at apple.com> wrote:
>
> + Adrian and Davide, since they work in this area.
>
> Hi Raj,
>
> On Jan 15, 2018, at 4:48 PM, Raj Barik via swift-dev <swift-dev at swift.org>
> wrote:
>
> Thanks Mark. You are right, the following assertion fails
>
> require(!DS || DS->getParentFunction() == I->getFunction(),
>             "debug scope of instruction belongs to a different function");
>
> I am not sure if we need such a check.
>
>
> This check exists to ensure that the debug info generated for a program is
> helpful. For example, a function with the wrong debug scope attached might
> make for confusing backtraces.
>
>
> It's more than just confusing backtraces. If these invariants aren't met
> the LLVM IR generated from this SIL will not pass the LLVM IR verifier or
> in older versions even crash LLVM.
>
> The constraint being checked in this assertion is that each instruction
> must be in a lexical scope whose top-level ancestor is the function itself.
> I.e., it is not legal to move an instruction from one function to another
> with updating that instruction's debug information by either reparenting it
> into the new function or by creating inline information for it.
> SILClonerWithScope implements this re-parenting of instructions into a new
> function for you.
>
> -- adrian
>
>
> vedant
>
>
>
> On Mon, Jan 15, 2018 at 4:16 PM, Mark Lacey <mark.lacey at apple.com> wrote:
>
>>
>>
>> On Jan 15, 2018, at 12:54 PM, Raj Barik <rkbarik at gmail.com> wrote:
>>
>> Mark,
>>
>> Thanks a lot for quick reply.
>>
>> Is there any reason this interface in SILInstruction should be private?
>> void setDebugScope(SILBuilder &B, const SILDebugScope *DS);
>>
>> In my case, I am splicing the old F into the new function NF. While
>> splicing retains the debug scope, the new instruction (InitRef) that I am
>> adding to NF is being created in some random scope (not F) even though I
>> explicitly make Builder's debug scope point to F's.
>>
>>
>> The assert you originally hit appears to indicate that the parentFunction
>> of the debug scope for an instruction should be the same as the function
>> the instruction is in. I am not very familiar with the requirements of our
>> debug information and debug scopes, but this seems like a reasonable
>> expectation.
>>
>> I suggested looking at SILCloner because it is the primary way by which
>> we clone functions, and clearly has to deal with debug scopes when it does
>> so.
>>
>> I’m not sure how much more help I can be on this matter, but perhaps
>> someone more knowledgable can chime in.
>>
>> Mark
>>
>>
>>
>>
>>
>> Best,
>> Raj
>>
>>
>>
>> On Mon, Jan 15, 2018 at 9:52 AM, Mark Lacey <mark.lacey at apple.com> wrote:
>>
>>> I’d suggest looking at SILCloner.h as well as ScopeCloner in
>>> SILBasicBlock.cpp to see how they are dealing with debug scopes.
>>>
>>> Mark
>>>
>>> > On Jan 15, 2018, at 9:24 AM, Raj Barik via swift-dev <
>>> swift-dev at swift.org> wrote:
>>> >
>>> > Hi,
>>> >
>>> > I am running into a debug scope SIL Verifier error when creating a new
>>> function (NF) from an existing one (F). Can someone point me where I am
>>> going wrong?
>>> >
>>> > NF = M.createFunction(...., F->getDebugScope());
>>> > SILBasicBlock *NFBody = NF->createBasicBlock();
>>> > SILBuilder NFBuilder(NFBody);
>>> > SILOpenedArchetypesTracker OpenedArchetypesTrackerNF(NF);
>>> > NFBuilder.setOpenedArchetypesTracker(&OpenedArchetypesTrackerNF);
>>> > NFBuilder.setCurrentDebugScope(NFBody->getParent()->getDebugScope());
>>> > ...
>>> > for (auto &param : params) { /* Assume all are generic types */
>>> >   auto GenericsSILType = ....
>>> >   auto NewArg = NFBody->createFunctionArgument(GenericSILType);
>>> >   auto Conformances = Mod->lookupConformance(...);
>>> >   auto *InitRef = NFBuilder.createInitExistentialRef( Loc,
>>> ArgDesc.Arg->getType(), NewArg->getType().getSwiftRValueType()->getCanonicalType(),
>>> NewArg, Conformances);
>>> >   ...
>>> > }
>>> >
>>> > The InitRef instruction created above runs into SIL verifier error:
>>> >
>>> > SIL verification failed: debug scope of instruction belongs to a
>>> different function: !DS || DS->getParentFunction() == I->getFunction()
>>> > Verifying instruction:
>>> >    %0 = argument of bb0 : $τ_0_0                 // user: %1
>>> > ->   %1 = init_existential_ref %0 : $τ_0_0 : $τ_0_0, $SomeProtocol //
>>> user: %2
>>> >
>>> > The SIL looks correct to me though.
>>> >
>>> > --Raj
>>> > _______________________________________________
>>> > swift-dev mailing list
>>> > swift-dev at swift.org
>>> > https://lists.swift.org/mailman/listinfo/swift-dev
>>>
>>>
>>
>>
> _______________________________________________
> swift-dev mailing list
> swift-dev at swift.org
> https://lists.swift.org/mailman/listinfo/swift-dev
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-dev/attachments/20180116/6dd2b392/attachment.html>