Hi Jon
Le 03/21/2017 à 09:11 PM, Jonathan Roelofs a écrit :
On 3/21/17 1:53 PM, Christian Bruel via Phabricator wrote:
chrib added a comment.
In https://reviews.llvm.org/D31140#706411, @jroelofs wrote:
Can you clarify the logic here? It's my understanding that:
`-fno-exceptions` does *not* imply `-fno-unwind-tables`
however:
`-fno-unwind-tables` *does* imply that exceptions cannot be used on
targets that require the tables to do unwinding.
Yes, (bad things might happen or (std::terminate will be called, or
destructors not called.)...
But -f[no]-unwind-tables implies the UWTable attribute, not NoUwind
attribute. To toggle NoUnwind, use -fno-exceptions
And this is getting worse with .canunwind which means DoesNotThrow :)
in my understanding, the logic is as follow:
Since "An exception cannot propagate through a function with a
nounwind table. The exception handling runtime environment terminates
the program if it encounters a nounwind table during exception
processing." (ARM Information Center)
The "nounwind" LLVM attribute, which means "Function does not throw"
translates as the EXIDX_CANTUNWIND value in the exception table index
table which needs to be created for the purpose (for the function)
I think the problem is here, actually. "nounwind" implies "does not
throw", but "does not throw" really should not imply "nounwind". This
is something that ought to be clarified in the langref with the
addition of a "does not throw" attribute. Then the optimizer should be
fixed to deduce "does not throw" instead of "nounwind", and we can let
"nounwind" continue to imply .cantunwind.
Absolutely, renaming the nounwind attribute into a nothrow would make
things lot clearer! That would fix the semantic (which is already a lot).
This has been already discussed here and there. e.g for reference this
thread
http://lists.llvm.org/pipermail/llvm-dev/2014-February/070366.html
without any outcome, because the optimizer cannot deduce, without a
context information from clang, if the environment supports EH or not.
One could imaging a global (LTO) analysis to find any cxa_throw in the
IRs, but that would not pass the pre-compiled object barrier or dynamic
dependencies.
So this context information could be well derived either from the
language (! C) or the fno-exception flag. Unfortunately none of them is
known to the optimizer.
The proposal was to make this information of the form of an option
(like here) or eventually an additional attribute of the form
nocantunwind (or noeh, any suggestion welcome). So
attributes #0 = { nounwind nocantunwind}
means for the arm streamer dont' emit the .cantunwind directive required
by the EHABI if the function cant take exceptions
And of course without exception runtime environment (the test here)
we don't need this table. So I can see 3 cases:
- nounwind set : Generate .cantunwind directive
and unwind table
- nounwind set but not EH Do not generate the .cantunwind directive
and do not emit the unwind table
- uwtable set Need to generate the unwind
table (even without EH)
The disable-arm-cantunwind flag means: without EH support if the
function does not throw, do dot generate the exception tables and the
EXIDX_CANTUNWIND value.
I'm not a big fan of this workaround flag. I'd rather see this fixed
by clarifying/fixing the semantics of the IR.
yes we can clarify the semantic of the IR. But that will not be enough
to fix the problem.
nounwind implies emit .cantunwind which implies exception table
nounwind + NOEH implies don't emit the exception table
and we need this NOEH from clang. Now the question is how?
1) a new arm specific flag
2) a target independant no-exception flag
3) a nocantunwind attribute
But I'm not a big fan of adding a new attribute for a arm specific need...
Best Regards
Christian
Jon
https://reviews.llvm.org/D31140
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
