bc: fix for mul overflow in scale calculation in a^b

function                                             old     new   delta
zbc_num_p                                            516     518      +2

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This commit is contained in:
Denys Vlasenko 2021-06-12 12:19:20 +02:00
parent 4d983dcdde
commit e5958f7dda

View File

@ -2152,6 +2152,7 @@ static FAST_FUNC BC_STATUS zbc_num_p(BcNum *a, BcNum *b, BcNum *restrict c, size
BcNum copy;
unsigned long pow;
size_t i, powrdx, resrdx;
size_t a_rdx;
bool neg;
// GNU bc does not allow 2^2.0 - we do
@ -2182,16 +2183,30 @@ static FAST_FUNC BC_STATUS zbc_num_p(BcNum *a, BcNum *b, BcNum *restrict c, size
bc_num_init(&copy, a->len);
bc_num_copy(&copy, a);
a_rdx = a->rdx; // pull it into a CPU register (hopefully)
// a is not used beyond this point
if (!neg) {
if (a->rdx > scale)
scale = a->rdx;
if (a->rdx * pow < scale)
scale = a->rdx * pow;
unsigned long new_scale;
if (a_rdx > scale)
scale = a_rdx;
new_scale = a_rdx * pow;
// Don't fall for multiplication overflow. Example:
// 0.01^2147483648 a_rdx:2 pow:0x80000000, 32bit mul is 0.
//not that it matters with current algorithm, it would OOM on such large powers,
//but it can be improved to detect zero results etc. Example: with scale=0,
//result of 0.01^N for any N>1 is 0: 0.01^2 = 0.0001 ~= 0.00 (trunc to scale)
//then this would matter:
// if (new_scale >= pow) is false, we had overflow, correct
// "new_scale" value is larger than ULONG_MAX, thus larger
// than any possible current value of "scale",
// thus "scale = new_scale" should not be done:
if (new_scale >= pow)
if (new_scale < scale)
scale = new_scale;
}
for (powrdx = a->rdx; !(pow & 1); pow >>= 1) {
for (powrdx = a_rdx; !(pow & 1); pow >>= 1) {
powrdx <<= 1;
s = zbc_num_mul(&copy, &copy, &copy, powrdx);
if (s) goto err;