[LLVMdev] How to prevent insertion of memcpy()

[Chandler Carruth]

On Tue, May 29, 2012 at 10:56 AM, Dmitry Vyukov <dvyukov at google.com> wrote:

> On Tue, May 29, 2012 at 9:50 PM, Chandler Carruth <chandlerc at google.com>wrote:
>
>>   > How do I disable that feature? I've tried -fno-builtin and/or
>>>>>>>> -ffreestanding
>>>>>>>> > with no success.
>>>>>>>> clang (as well as gcc) requires that freestanding environment
>>>>>>>> provides
>>>>>>>> memcpy, memmove, memset and memcmp.
>>>>>>>>
>>>>>>>> PS: Consider emailing cfedev, not llvmdev.
>>>>>>>>
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Thanks. I've emailed cfe-dev.
>>>>>>> We absolutely need clang/llvm to not insert the calls into our code.
>>>>>>>
>>>>>>
>>>>>> This really isn't possible.
>>>>>>
>>>>>> The C++ standard essentially requires the compiler to insert calls to
>>>>>> memcpy for certain code patterns.
>>>>>>
>>>>>> What do you really need here? Clearly you have some way of handling
>>>>>> when the user writes memcpy; what is different about Clang or LLVM
>>>>>> inserting memcpy?
>>>>>>
>>>>>
>>>>> I need it for ThreadSanitizer runtime. In particular
>>>>>
>>>>> http://llvm.org/viewvc/llvm-project/compiler-rt/trunk/lib/tsan/rtl/tsan_interceptors.cc?view=annotate
>>>>> line 1238. But I had similar problems in other places.
>>>>> Both memory access processing and signal handling are quite tricky, we
>>>>> can't allow recursion.
>>>>>
>>>>
>>>> The first thing to think about is that you *do* need to use
>>>> -fno-builtin / -ffreestanding when compiling the runtime because it
>>>> provides its own implementations of memcpy.
>>>>
>>>
>>> We used both at some points in time, but the problem is that they do not
>>> help to solve the problem. I think we use -fno-builtin now, I am not sure
>>> about -ffreestanding.
>>>
>>> The second is that there is no way to write fully generic C++ code w/o
>>>> inserting calls to memcpy. =/ If you are writing your memcpy
>>>> implementation, you'll have to go to great lengths to use C constructs that
>>>> are guaranteed to not cause this behavior, or to manually call an
>>>> un-instrumented memcpy implementation. I don't know of any easy ways around
>>>> this.
>>>>
>>>
>>> What are these magic constructs. I had problems with both struct copies
>>> and for loops.
>>>
>>
>> Don't copy things by value ever. =/ It is really, *really* hard to do
>> this.
>>
>
> Do you mean 'don't do struct copies'? Are there other problems aside from
> implicit memcpy calls?
>

Don't do copies outside of a restricted set of primitive types (sizeof(T)
<= sizeof(T*) would be my rule of thumb, but there is no hard-and-fast rule
here to avoid these problems).


>
>
>>  If at all possible, I would build your runtime against an
>> un-instrumented memcpy (perhaps defined within the runtime), and then use
>> aliases or other techniques to wrap the instrumented functions in the
>> exported names necessary for use when intercepting memcpy calls from the
>> instrumented program.
>>
>
> I am not sure I understand it.
> We can't afford function calls scattered at random places. It will cost
> 30% of performance of so.
>

These won't end up actually being function calls... Clang lowers them to
'memcpy', and LLVM will try to lower them to actual loads and stores where
possible.

We should discuss these issues separately though:

1) Get the runtime working w/o worrying about memcpy being inserted or not
by having a clear barrier between instrumented functions and
non-instrumented functions, and making the non-instrumented ones available
when compiling and linking the runtime, but not when compiling / linking
the instrumented program.

2) Deal with any performance fallout of the thusly built runtime. We can
fix the LLVM optimizers until they generate the optimal code. =]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20120529/502198a1/attachment.html>