r310401 - PR19668, PR23034: Fix handling of move constructors and deleted copy

[Diana Picus via cfe-commits]

Ugh, I'm not sure that last email made it through.

Here's the IR with -Xclang -disable-llvm-passes, at r310400 and r310401:
https://goo.gl/4n84SR
https://goo.gl/gxhxp9

The diff between them is small enough so I don't think there's much
point in reducing it further.
Hope that helps.

On 15 August 2017 at 17:51, Diana Picus <diana.picus at linaro.org> wrote:
> Actually, I'm not sure I need to reduce it, the diff between the IR
> with -Xclang -no-llvm-passes is pretty small for the version
> with/without the patch. See attached. I hope this helps, it does look
> like a problem with va_arg handling.
>
> Sorry again that this is going so slowly...
>
> On 15 August 2017 at 16:52, Diana Picus <diana.picus at linaro.org> wrote:
>> On 15 August 2017 at 01:25, Richard Smith <richard at metafoo.co.uk> wrote:
>>> On 14 August 2017 at 03:27, Diana Picus via cfe-commits
>>> <cfe-commits at lists.llvm.org> wrote:
>>>>
>>>> Hi,
>>>>
>>>> Strangely enough, it turns out that if I run
>>>> Asan-armhf-with-calls-Noinst-Test on the command line it fails,
>>>> although it doesn't fail when run with lit.
>>>
>>>
>>> Looks like the crash is within the "use_colors == true" portion of
>>> ColoredPrintf, so this would make sense if running the test within lit turns
>>> off color support (perhaps because the output is not a terminal).
>>>
>>
>> That's a good catch, it seems if I run it with --gtest-color=no it
>> still passes, but prints some garbage at the end:
>> [ PASSED ] 97 tests.
>> YOU HAVE -1094542056 DISABLED F]�뀼��_8�pG����D2
>>
>>
>>>>
>>>> I've attached the stack
>>>> trace from gdb. It looks like some trouble passing down va_arg
>>>> parameters, but I haven't looked into too much details. The segfault
>>>> happens when we try to do a   ldrb   r3, [r0, r1], with r1 set to 0 by
>>>> the current function and r0 passed down from the caller. I'm not sure
>>>> if this is the exact same problem as the other tests, but feel free to
>>>> have a look at that code.
>>>
>>>
>>> Have you tried running Asan-armhf-with-calls-Noinst-Test on the command line
>>> without this patch applied? (It's possible that this is a pre-existing bug
>>> in ARM varargs call lowering, and is unrelated to the bug we're trying to
>>> track down. This code path performs an unrelated varargs call between a
>>> va_start / va_end pair, which seems like quite a rare situation, and I could
>>> easily believe there's something wrong with our lowering that allows some
>>> portion of the outer va_list state to be clobbered in that scenario.)
>>>
>>
>> It seems to work without the patch. I'm currently trying to reduce
>> Asan-armhf-with-calls-Test and I'll get back to you with the IR or
>> assembly files (whichever makes any difference).
>>
>>>> Meanwhile, I've removed some clutter from Asan-armhf-with-calls-Test
>>>> (which is the original failure that we were seeing) and left only one
>>>> failing test that seemed small enough. I'll try to look at the
>>>> disassembly before/after the patch and maybe even run valgrind on it
>>>> (running it on the original binary naturally takes forever).
>>>>
>>>> Let me know if there's anything else I could try. I can also send you
>>>> disassembly or even LLVM IR for the Asan-armhf-with-calls-Noinst-Test
>>>> if you think it helps.
>>>>
>>>> Cheers,
>>>> Diana