Fix probable_prime over large shift

In the probable_prime() function we behave slightly different if the number
of bits we are interested in is <= BN_BITS2 (the num of bits in a BN_ULONG).
As part of the calculation we work out a size_limit as follows:

    size_limit = (((BN_ULONG)1) << bits) - BN_get_word(rnd) - 1;

There is a problem though if bits == BN_BITS2. Shifting by that much causes
undefined behaviour. I did some tests. On my system BN_BITS2 == 64. So I
set bits to 64 and calculated the result of:

    (((BN_ULONG)1) << bits)

I was expecting to get the result 0. I actually got 1! Strangely this...

    (((BN_ULONG)0) << BN_BITS2)

...does equal 0! This means that, on my system at least, size_limit will be
off by 1 when bits == BN_BITS2.

This commit fixes the behaviour so that we always get consistent results.
Reviewed-by: NAndy Polyakov <appro@openssl.org>

crypto/bn/bn_prime.c

@@ -518,7 +518,17 @@ static int probable_prime(BIGNUM *rnd, int bits)
      * additionally don't want to exceed that many bits.
      */
     if (is_single_word) {
-        BN_ULONG size_limit = (((BN_ULONG)1) << bits) - BN_get_word(rnd) - 1;
+        BN_ULONG size_limit;
+        
+        if (bits == BN_BITS2) {
+            /*
+             * Shifting by this much has undefined behaviour so we do it a
+             * different way
+             */
+            size_limit = ~((BN_ULONG)0) - BN_get_word(rnd);
+        } else {
+            size_limit = (((BN_ULONG)1) << bits) - BN_get_word(rnd) - 1;
+        }
         if (size_limit < maxdelta)
             maxdelta = size_limit;
     }