- Fix the division by 0, improve comments.

- Avoid doing the division if the result would be outside the ]0,1<<24[ range:
-- if the numerator is nonpositive, ie <=0, truncate the result to 0,
-- if the numerator is greater or equal than the denominator, the result will be outside the allowed range, hence truncate the result to 1<<24.
-- otherwise, the result will be inside the range. Knowing that the denominator is greater than the numerator, if the numerator has the last 24 bits non zero, the denominator can't be less than 1<<24, hence the denominator won't be truncated to 0.
- Add comment details to help who doesn't know math. Big deal!
This commit is contained in:
Edoardo Prezioso 2014-11-25 16:44:12 +01:00
parent 4b2af7074e
commit cc4e66f976

View file

@ -737,18 +737,50 @@ bool PIT_CheckLine(line_t *ld, const FBoundingBox &box, FCheckPosition &tm)
{ // Find the point on the line closest to the actor's center, and use
// that to calculate openings
// [EP] Use 64 bit integers in order to keep the exact result of the
// multiplication, because in the worst case, which is by the map limit
// (32767 units, which is 2147418112 in fixed_t notation), the result
// would occupy 62 bits (if I consider also the addition with another
// and possible 62 bit value, it's 63 bits).
// This privilege could not be available if the starting data would be
// 64 bit long.
// With this, the division is exact as the 32 bit float counterpart,
// though I don't know why I had to discard the first 24 bits from the
// divisor.
SQWORD r_num = ((SQWORD(tm.x - ld->v1->x)*ld->dx) + (SQWORD(tm.y - ld->v1->y)*ld->dy));
SQWORD r_den = (SQWORD(ld->dx)*ld->dx + SQWORD(ld->dy)*ld->dy) / (1 << 24);
fixed_t r = (fixed_t)(r_num / r_den);
// multiplication, because in the case the vertexes have both the
// distance coordinates equal to the map limit (32767 units, which is
// 2147418112 in fixed_t notation), the product result would occupy
// 62 bits and the sum of two products would occupy 63 bits
// in the worst case. If instead the vertexes are very close (1 in
// fixed_t notation, which is 1.52587890625e-05 in float notation), the
// product and the sum can be 1 in the worst case, which is very tiny.
SQWORD r_num = ((SQWORD(tm.x - ld->v1->x)*ld->dx) +
(SQWORD(tm.y - ld->v1->y)*ld->dy));
// The denominator is always positive. Use this to avoid useless
// calculations.
SQWORD r_den = (SQWORD(ld->dx)*ld->dx + SQWORD(ld->dy)*ld->dy);
fixed_t r = 0;
if (r_num <= 0) {
// [EP] The numerator is less or equal to zero, hence the closest
// point on the line is the first vertex. Truncate the result to 0.
r = 0;
} else if (r_num >= r_den) {
// [EP] The division is greater or equal to 1, hence the closest
// point on the line is the second vertex. Truncate the result to
// 1 << 24.
r = (1 << 24);
} else {
// [EP] Deal with the limited bits. The original formula is:
// r = (r_num << 24) / r_den,
// but r_num might be big enough to make the shift overflow.
// Since the numerator can't be saved in a 128bit integer,
// the denominator must be right shifted. If the denominator is
// less than (1 << 24), there would be a division by zero.
// Thanks to the fact that in this code path the denominator is less
// than the numerator, it's possible to avoid this bad situation by
// just checking the last 24 bits of the numerator.
if ((r_num >> (63-24)) != 0) {
// [EP] In fact, if the numerator is greater than
// (1 << (63-24)), the denominator must be greater than
// (1 << (63-24)), hence the denominator won't be zero after
// the right shift by 24 places.
r = (r_num)/(r_den >> 24);
} else {
// [EP] Having the last 24 bits all zero allows right shifting
// the numerator by 24 bits.
r = (r_num << 24)/r_den;
}
}
/* Printf ("%d:%d: %d (%d %d %d %d) (%d %d %d %d)\n", level.time, ld-lines, r,
ld->frontsector->floorplane.a,
ld->frontsector->floorplane.b,