Re: Optimization techniques and undefined behavior

On 30/04/2019 13:46, David Brown wrote:
> On 29/04/2019 18:10, Christian Gollwitzer wrote:


>> P5
>> 100 200
>> 255
>> ..jdk hlhdhqkd.. here comes the binary data
>>
>>
>> The 100 and 200 are the width and height of the image data, the 255 is
>> the highest possible value (for 16 bit it would be different).
>> Obviously, you'd read in the width and height, then multiply them to
>> compute the memory needed for the data, and  - oops - how do you make
>> sure that no overflow occurs? In the past, there had been security
>> problems in image libraries with exactly this kind of problem: integer
>> overflow due to unreasonable image sizes.
>
> It is really incredibly simple (at least in this case).  Do the
> multiplications using types that won't overflow.  That might be an
> unsigned type if its range is suitable (not because it has defined
> overflow behaviour, but use it if it has enough range) or a bigger
> signed integer type.


That's just kicking the can further down the road.


If you have two unknown values A and B, and need to multiply, you won't
know if the result will overflow.


Using a type that is double the width might help, unless A and B are
already using the widest type. But if they are int32, and you do the
calculation like this:


          (int64)A * (int64)B


then suppose you had to work out A*B*C; do you use:


          (int128)*(int64)A * (int64)*B) * (int128)C ?


(Real example: imagebytes := width * height * bytes_per_pixel)


It doesn't really scale. And it doesn't help here:


          int32 A, B, C;


          C = (int64)A * (int64)B;


as you will need to check the int64 intermediate result for overflow. It
all gets very messy.


I think a lot of this can be handled with application code dealing with
validation; it doesn't really benefit from UB in the language.


In the example posed, you have the additional problem that the input can
be this:


        P5
        389000000000000000000000000000 9200000000000000000000000000


with both dimensions exceeding int64. You need to get past that first,
and that might indicate some error before you get around to doing any
multiplying.


These are all real, practical problems that are starting to get a long
way from UB in a language.


> There are plenty of cases where it is difficult to write code that is
> efficient even on poorer compilers, and correct code so that it works on
> good compilers.  No one claims that programming is always easy.  And
> sometimes the best solution is code that is not portable or correct by
> the standards, but works well with the implementations you need to use.
> Most code, after all, is only ever compiled on the one compiler.
>
>
>> The simplest thing would be to reject any width or height > 100,000,
>> claiming that noone can handle this images, but what about an image of
>> size 200,000 x 3 ? If C++ would provide an easy way to detect / branch
>> on overflow, for example throw an exception, then this could be handled
>> easily. I can't see how you can claim that your code never overflows,
>> unless you don't handle untrusted user input data.
>>


> All C++ compilers with any self-respect support 64-bit integer types.
> Do you think it's reasonable to reject any image dimension greater than
> 2,000,000,000 ?  I do.


I remember being shown a typesetting machine some decades ago, which
(IIRC) was fed PostScript code and produced an image on film at some
16,000 dpi.


2 billion pixels (or dots) would be only 8 square inches of image. While
A4 scanners working at 9600dpi (1-bit depth) would result in an image of
some 9 billion pixels. So 2 billion pixels is not really that far out of
the ball-park.

Return to the comp.compilers page.
Search the comp.compilers archives again.