[llvm-commits] [PATCH] Fix tail calls (indirect or large-code-model): PR5729

[Jeffrey Yasskin]

On Sat, Jan 9, 2010 at 7:39 AM, Jon Harrop <jon at ffconsultancy.com> wrote:
> On Saturday 09 January 2010 01:53:17 Jeffrey Yasskin wrote:
>> Bug: http://llvm.org/PR5729
>> Patch: http://llvm.org/bugs/attachment.cgi?id=4017
>> Review at: http://codereview.appspot.com/185044
>>
>> Dan's probably the right person to double-check my tests.
>>
>> Jon and Albert, this probably won't unbreak things for you guys yet
>> since it doesn't change any behavior in -Asserts mode. I'll look for
>> your runtime problems next.
>
> Thanks for trying! I'm still seeing lots of bugs that seem to be TCO related
> and some (on x86) that seem to appear only with TCO and JIT at the same time
> (but not standalone compilation).

The JIT codegens more closely to `llc -code-model=large` than to the
default llc. Which have you tested with?

> I just noticed an LLVM bug in TCO on x86-64 recently that was doing this:
>
> $ ./hlvm
> 8x 8-queens
> Queens
> Queens
> Queens
> Queens
> hlvm: X86ISelLowering.cpp:2006: virtual llvm::SDValue
> llvm::X86TargetLowering::LowerCall(llvm::SDValue, llvm::SDValue, unsigned
> int, bool, bool, const llvm::SmallVectorImpl<llvm::ISD::OutputArg>&, const
> llvm::SmallVectorImpl<llvm::ISD::InputArg>&, llvm::DebugLoc,
> llvm::SelectionDAG&, llvm::SmallVectorImpl<llvm::SDValue>&): Assertion
> `((Callee.getOpcode() == ISD::Register &&
> (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
> cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) || Callee.getOpcode() ==
> ISD::TargetExternalSymbol || Callee.getOpcode() == ISD::TargetGlobalAddress)
> && "Expecting an global address, external symbol, or register"' failed.
> Stack dump:
> 0.      Segmentation fault
>
> I just upgraded to the latest LLVM SVN that I assume contains your fix but now
> I get this:

This thread is about a _patch_ to fix that assertion. It hasn't gone in yet.

> I do think the situation is getting worse, probably because there are no TCO
> tests in the LLVM test suite so people don't know when they're breaking it.
> Can I give you the LLVM IR from some tests to put into LLVM's test suite?

That'd be awesome. Please do check that they still crash or behave
badly with this patch applied, or that they're uncovered variations on
the test/CodeGen/X86/tailcall*.ll tests
(http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/).