libpayload: cbgfx: Fix add_fractions() overflow reduction

log2(1) is 0 and log2(0) is -1. If we have the int64_t 0xffffffff then
log2(0xffffffff >> 31) = log2(0x1) = 0, so the current reduction code
would not shift. That's a bad idea, though, since 0xffffffff when
interpreted as an int32_t would become a negative number.

We need to always shift one more than the current code does to get a
safe reduction. This also means we can get rid of another compare/branch
since -1 is the smallest result log2() can return, so the shift can no
longer go negative now.

Signed-off-by: Julius Werner <jwerner@chromium.org>
Change-Id: Ib1eb6364c35c26924804261c02171139cdbd1034
Reviewed-on: https://review.coreboot.org/c/coreboot/+/42845
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Yu-Ping Wu <yupingso@google.com>
Reviewed-by: Joel Kitching <kitching@google.com>
Reviewed-by: Paul Menzel <paulepanter@users.sourceforge.net>
This commit is contained in:
Julius Werner 2020-06-26 23:27:26 -07:00 committed by Patrick Georgi
parent a94136f082
commit 63358eaff6
1 changed files with 3 additions and 7 deletions

View File

@ -85,13 +85,9 @@ static void add_fractions(struct fraction *out,
n = (int64_t)f1->n * f2->d + (int64_t)f2->n * f1->d;
d = (int64_t)f1->d * f2->d;
/* Simplest way to reduce the fraction until fitting in int32_t */
shift = log2(MAX(ABS(n), ABS(d)) >> 31);
if (shift > 0) {
n >>= shift;
d >>= shift;
}
out->n = n;
out->d = d;
shift = log2(MAX(ABS(n), ABS(d)) >> 31) + 1;
out->n = n >> shift;
out->d = d >> shift;
}
static void add_scales(struct scale *out,