Question on Taming the Optimizer

Question on Taming the Optimizer

[Jeffrey Walton]

Hi All,

What is the preferred way to tame the optimizer when two class
functions are producing incorrect results?

There's not much to the class - its a vector with copy semantics. So
it mirrors all vector's public functions, and forwards calls through
the internal vector pointer. The pointer is a boost shared_ptr.

-O2 is giving problems. -O0 is OK, Intel ICC is OK, and Visual Studio is OK.

Jeff

Re: Question on Taming the Optimizer

[Jonathan Wakely]

On 28 September 2011 00:41, Jeffrey Walton <noloader@gmail.com> wrote:
> Hi All,
>
> What is the preferred way to tame the optimizer when two class
> functions are producing incorrect results?

Fix your code ;-)

> There's not much to the class - its a vector with copy semantics. So
> it mirrors all vector's public functions, and forwards calls through
> the internal vector pointer. The pointer is a boost shared_ptr.
>
> -O2 is giving problems. -O0 is OK, Intel ICC is OK, and Visual Studio is OK.

As suggested at http://gcc.gnu.org/bugs/ you can try
-fno-strict-aliasing and -fwrapv, if they help your code is broken.

You can also turn off other individual optimisations with -fno-foobar,
see the manual for the various optimizations that are active at -O1
and -O2


> Jeff
>

Re: Question on Taming the Optimizer

[Jeffrey Walton]

On Tue, Sep 27, 2011 at 8:03 PM, Jonathan Wakely <jwakely.gcc@gmail.com> wrote:
> On 28 September 2011 00:41, Jeffrey Walton <noloader@gmail.com> wrote:
>> Hi All,
>>
>> What is the preferred way to tame the optimizer when two class
>> functions are producing incorrect results?
>
> Fix your code ;-)
That's what I thought, too. I'm kind of surprised that I haven't
gotten a warning out of any of the tools (I run with maximum on all
platforms).

Below is a brief version of what the declarations and definitions look
like. Obviously, the ASSERTs are removed in the 'release'
configuration. The best I can tell, the call to "m_vector->insert(pos,
first, last);" is not being made, but I could be wrong.

We are using LeBlanc's SafeInt, which itself uses unsigned math (so
there should be no undefined behavior). But I don't believe the
SafeInts are being used on the code path leading to the unexpected
behavior.

Oh, and trying to use GDB to debug this is really fun. Sometimes I
think GDB is guessing where it might be after a 'si'.

>> There's not much to the class - its a vector with copy semantics. So
>> it mirrors all vector's public functions, and forwards calls through
>> the internal vector pointer. The pointer is a boost shared_ptr.
>>
>> -O2 is giving problems. -O0 is OK, Intel ICC is OK, and Visual Studio is OK.
>
> As suggested at http://gcc.gnu.org/bugs/ you can try
> -fno-strict-aliasing and -fwrapv, if they help your code is broken.
I actually decided against -fwrapv because I did not want to affect
all compilation units.

> You can also turn off other individual optimisations with -fno-foobar,
> see the manual for the various optimizations that are active at -O1
> and -O2
I was thinking __attribute__((optimize(0))) for the two functions in
question. Is it available in modern GCC versions (ie, 4.0 and above)?

Jeff

template <typename T>
class SecureArray
{
 public:
    typedef typename std::vector< T, zallocator<T> > SecureVector;
    typedef typename zallocator<T>::size_type size_type;
    typedef typename SecureVector::value_type value_type;
    ...

    // One of the problem functions
    void insert(iterator pos, InputIterator first, InputIterator last);

  private:
    boost::shared_ptr<SecureVector> m_vector;
}

template <typename T>
template <class InputIterator>
void SecureArray<T>::insert(iterator pos, InputIterator first,
InputIterator last)
{
  ASSERT2(first, "Bad first input iterator");
  ASSERT2(first >= last, "Input iterators are not valid");
  if(!(first >= last))
    throw std::length_error("Input iterators are not valid");

  // Not sure what the conatiner does here....
  ASSERT2(first % sizeof(T) == 0, "InputIterator first slices elements");
  ASSERT2(last % sizeof(T) == 0, "InputIterator last slices elements");

  ASSERT1(m_vector.get());
  m_vector->insert(pos, first, last);
}

Re: Question on Taming the Optimizer

[Jonathan Wakely]

On 28 September 2011 01:40, Jeffrey Walton wrote:
> I was thinking __attribute__((optimize(0))) for the two functions in
> question. Is it available in modern GCC versions (ie, 4.0 and above)?

No, it's much more recent than that:
http://gcc.gnu.org/gcc-4.4/changes.html

Re: Question on Taming the Optimizer

[Andrew Haley]

On 09/28/2011 12:41 AM, Jeffrey Walton wrote:

> 
> What is the preferred way to tame the optimizer when two class
> functions are producing incorrect results?

I wouldn't.  I'd try to isolate the problem into a test case that
demonstrates the problem.

Andrew.