llvm dev - Apr 2012 - [LLVMdev] Support for "foreign" Visual Studio 64 bits C++ exceptions?

Hi,

We are currently porting our application to Windows 64 bits (amd64), and we are
getting regressions from 32 bits relatively to LLVM 3.0's ability to get
traversed by "foreign" Visual Studio C++ exceptions. That is, there is
a crash when C++ exceptions which are thrown from Visual Studio code, are
unwinding through LLVM code before being eventually catch by other Visual Studio
code (more details below).

I am not an expert on the subject of exceptions, however I was not able to find
clear answers about this specific issue. Should it be supported, and if not, is
there is any planned effort from the LLVM team to work on this?

Note that we are not using the LLVM3.0 exception API (launchingPad with
personality routine). In fact, this leads to another related question: is there
even a "personality routine" available when compiling with Visual
Studio? From what I read it looks like this would require support for
Windows' SEH, which is not implemented yet. Any news or plans on that front?
SEH was mentionned in the design discussion of LLVM3.0's exception API,
without mentioning any plans to support it:
http://groups.google.com/group/llvm-dev/browse_thread/thread/b7fc0999e9f11b8f/2cc43ddaba3bcff6?lnk=gst&q=SEH+exception#2cc43ddaba3bcff6

More details about the crash:

I was able to repro the issue simply by modifying the Fibonacci example to call
an external C++ function which throws an exception, much like in
ExceptionsDemo.cpp: works on 32 bits but not on 64 bits Visual Studio. The crash
happens within Windows' _cxxThrowException routine. I read that the 64 bits
exception handling in Windows is doing some unwinding in a different fashion
than on 32 bits (as explained here:
http://lackingrhoticity.blogspot.ca/2011/11/stack-unwinding-risks-on-64-bit-windows.html).
In particular, this blog seems to state that exceptions won't unwind
properly through functions that don't provide some unwind info, which is the
case for LLVM I guess.

Thank you,

Jerome Couture-Gagnon

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120412/e06b6ba9/attachment.html>

On 12.04.2012 23:02, Jérôme Couture-Gagnon wrote:
> Hi,
>
> We are currently porting our application to Windows 64 bits (amd64), 
> and we are getting regressions from 32 bits relatively to LLVM 3.0's 
> ability to get traversed by "foreign" Visual Studio C++
exceptions.
> That is, there is a crash when C++ exceptions which are thrown from 
> Visual Studio code, are unwinding through LLVM code before being 
> eventually catch by other Visual Studio code (more details below).
>
> I am not an expert on the subject of exceptions, however I was not 
> able to find clear answers about this specific issue. Should it be 
> supported,
No. Not sure if the current Windows builds use SJLJ or Dwarf2 exception 
handling, but either way it is simply incompatible with SEH stack 
unwinding. Apparently LLVM doesn't generate even the minimal support for 
SEH unwinding through its code to work.
> and if not, is there is any planned effort from the LLVM team to work 
> on this?
Not AFAIK, but I'm pretty sure any and all work on 64-bit SEH support 
would be appreciated. Currently, there is some work in Clang ongoing to 
emulate the Microsoft ABI, and eventually, SEH support will be required 
for completeness.

Sebastian
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120413/cfa0b1c0/attachment.html>

On Thu, Apr 12, 2012 at 2:02 PM, Jérôme Couture-Gagnon
<Jerome.Couture-Gagnon at fabricengine.com> wrote:
> Hi,
>
> We are currently porting our application to Windows 64 bits (amd64), and we
> are getting regressions from 32 bits relatively to LLVM 3.0's ability
to get
> traversed by "foreign" Visual Studio C++ exceptions. That is,
there is a
> crash when C++ exceptions which are thrown from Visual Studio code, are
> unwinding through LLVM code before being eventually catch by other Visual
> Studio code (more details below).
>
> I am not an expert on the subject of exceptions, however I was not able to
> find clear answers about this specific issue. Should it be supported, and
if
> not, is there is any planned effort from the LLVM team to work on this?
>
> Note that we are not using the LLVM3.0 exception API (launchingPad with
> personality routine). In fact, this leads to another related question: is
> there even a "personality routine" available when compiling with
Visual
> Studio? From what I read it looks like this would require support for
> Windows' SEH, which is not implemented yet. Any news or plans on that
front?
> SEH was mentionned in the design discussion of LLVM3.0's exception API,
> without mentioning any plans to support it:
>
http://groups.google.com/group/llvm-dev/browse_thread/thread/b7fc0999e9f11b8f/2cc43ddaba3bcff6?lnk=gst&q=SEH+exception#2cc43ddaba3bcff6
>
> More details about the crash:
>
> I was able to repro the issue simply by modifying the Fibonacci example to
> call an external C++ function which throws an exception, much like in
> ExceptionsDemo.cpp: works on 32 bits but not on 64 bits Visual Studio. The
> crash happens within Windows' _cxxThrowException routine. I read that
the 64
> bits exception handling in Windows is doing some unwinding in a different
> fashion than on 32 bits (as explained
>
here: http://lackingrhoticity.blogspot.ca/2011/11/stack-unwinding-risks-on-64-bit-windows.html).
> In particular, this blog seems to state that exceptions won't unwind
> properly through functions that don't provide some unwind info, which
is the
> case for LLVM I guess.
>
> Thank you,
>
> Jerome Couture-Gagnon
Windows 64 does not have SEH. LLVM has code to generate proper unwind
information for Windows 64, but it may have bugs, or not generate the
proper info for this case.  Clang also does no generate proper MSVC
C++ exceptions. This is as much as I know.

- Michael Spencer

Sorry for mixing up terms, reading through forums about Windows 64 bits
exceptions gives quite a fuzzy picture (some talk about SEH for Windows 64 but
probably just refer to the Windows specific mechanism).

If the scenario I described below should work properly, then I will submit a bug
with the small example that I built. I am not sure however if that crash is
specific to Visual Studio's amd64 (which is not officially supported by
LLVM).

Thank you,

Jerome Couture-Gagnon

________________________________________
From: Michael Spencer [bigcheesegs at gmail.com]
Sent: Friday, April 13, 2012 11:19 AM
To: Jérôme Couture-Gagnon
Cc: llvmdev at cs.uiuc.edu
Subject: Re: [LLVMdev] Support for "foreign" Visual Studio 64 bits C++
exceptions?

On Thu, Apr 12, 2012 at 2:02 PM, Jérôme Couture-Gagnon
<Jerome.Couture-Gagnon at fabricengine.com> wrote:
> Hi,
>
> We are currently porting our application to Windows 64 bits (amd64), and we
> are getting regressions from 32 bits relatively to LLVM 3.0's ability
to get
> traversed by "foreign" Visual Studio C++ exceptions. That is,
there is a
> crash when C++ exceptions which are thrown from Visual Studio code, are
> unwinding through LLVM code before being eventually catch by other Visual
> Studio code (more details below).
>
> I am not an expert on the subject of exceptions, however I was not able to
> find clear answers about this specific issue. Should it be supported, and
if
> not, is there is any planned effort from the LLVM team to work on this?
>
> Note that we are not using the LLVM3.0 exception API (launchingPad with
> personality routine). In fact, this leads to another related question: is
> there even a "personality routine" available when compiling with
Visual
> Studio? From what I read it looks like this would require support for
> Windows' SEH, which is not implemented yet. Any news or plans on that
front?
> SEH was mentionned in the design discussion of LLVM3.0's exception API,
> without mentioning any plans to support it:
>
http://groups.google.com/group/llvm-dev/browse_thread/thread/b7fc0999e9f11b8f/2cc43ddaba3bcff6?lnk=gst&q=SEH+exception#2cc43ddaba3bcff6
>
> More details about the crash:
>
> I was able to repro the issue simply by modifying the Fibonacci example to
> call an external C++ function which throws an exception, much like in
> ExceptionsDemo.cpp: works on 32 bits but not on 64 bits Visual Studio. The
> crash happens within Windows' _cxxThrowException routine. I read that
the 64
> bits exception handling in Windows is doing some unwinding in a different
> fashion than on 32 bits (as explained
> here:
http://lackingrhoticity.blogspot.ca/2011/11/stack-unwinding-risks-on-64-bit-windows.html).
> In particular, this blog seems to state that exceptions won't unwind
> properly through functions that don't provide some unwind info, which
is the
> case for LLVM I guess.
>
> Thank you,
>
> Jerome Couture-Gagnon
Windows 64 does not have SEH. LLVM has code to generate proper unwind
information for Windows 64, but it may have bugs, or not generate the
proper info for this case.  Clang also does no generate proper MSVC
C++ exceptions. This is as much as I know.

- Michael Spencer