Re: [boost] [multiprecision review] Possible use of uninitialisedvalue

Subject: Re: [boost] [multiprecision review] Possible use of uninitialisedvalue
From: John Maddock (boost.regex_at_[hidden])
Date: 2012-06-14 11:43:27

> resize() doesn't zero-initialise the new storage.  In this particular 
> case, a == result because the expression being evaluated is something like 
> result = result + other.  So resizing result is also resizing a.  Hence 
> the compare_unsigned looks at the limbs of a that have just been added by 
> the resize, which are undefined.
>
> It appears that this could be fixed by zeroing the new limbs in resize(), 
> but that is unnecessary in other cases.  Is it even correct that a == 
> result is possible in this code?

Yes, inplace subtraction is OK.

Here's the fix I'm testing:

Index: cpp_int.hpp
===================================================================
--- cpp_int.hpp (revision 78943)
+++ cpp_int.hpp (working copy)
@@ -964,12 +964,16 @@
       result.sign(s);
       return;
    }
+   // This isn't used till later, but comparison has to occur before we 
resize the result,
+   // as that may also resize a or b if this is an inplace operation:
+   int c = a.compare_unsigned(b);
+   // Set up the result vector:
    result.resize(x);
+   // Now that a, b, and result are stable, get pointers to their limbs:
    typename cpp_int_backend<MinBits, Signed, Allocator>::const_limb_pointer 
pa = a.limbs();
    typename cpp_int_backend<MinBits, Signed, Allocator>::const_limb_pointer 
pb = b.limbs();
    typename cpp_int_backend<MinBits, Signed, Allocator>::limb_pointer pr = 
result.limbs();
    bool swapped = false;
-   int c = a.compare_unsigned(b);
    if(c < 0)
    {
       swap(pa, pb);

John.