Changes

a subtraction approach to determine sliding piece attacks with bitboards, proposed by [[Michael Hoffmann]] <ref>[http://www.talkchess.com/forum/viewtopic.php?t=29087 DIRECT BITBOARD MOVEGENERATION] by [[Michael Hoffmann]], [[CCC]], July 23, 2009</ref>. The idea is to get disjoint [[On an empty Board#PositiveRays|positive]] and [[On an empty Board#NegativeRays|negative ray masks]] of the [[Occupancy|occupancy]], and to subtract the [[General Setwise Operations#TheMostSignificantOneBitMS1B|most significant one bit]] of the negative ray from the doubled [[General Setwise Operations#TheLeastSignificantOneBitLS1B|least siginificat significant one bit]] (LS1B) of the positive ray. The difference is finally intersected with the linemask, excluding the [[Sliding Pieces|sliding piece]] itself from the attack set. However, in 2022, [[Daniel Infuehr]] found a faster solution to calculate the obstruction difference: subtract the most significant one bit of the negative ray from the the positive ray obstruction, followed by the [[General Setwise Operations#ExclusiveOr|symmetric difference]] to flip all positive ray obstruction bits for the proper result <ref>[https://www.talkchess.com/forum3/viewtopic.php?f=7&t=79701&start=2 Re: Sliding Piece Algorithm by Abstract Syntax Tree Sifting <nowiki>[RELEASE]</nowiki>] by [[Daniel Infuehr]], [[CCC]], April 17, 2022</ref>.

* @author Michael Hoffmann, Gerd IsenbergDaniel Infuehr

U64 lower, upper, mMS1B, ls1bms1B, odiff;

mMS1B ms1B = -C64(10x8000000000000000u) << bsr64>> lzcnt(lower | 1); // ms1b of lower (at least bit zero) // ls1b = upper & -upper; // odiff = 2 * ls1b + mMS1B- ms1B; odiff = upper ^ (upper - ms1B); // x86 lea option due to add -ms1bDaniel Infuehr's improvement

Since with two's complement i << x == - (-i << x), 2*high-low is done by add for the option to use [http://www.cs.virginia.edu/%7Eevans/cs216/guides/x86.html x86 lea-instruction]. The routine looks competitive for CPUs with fast [[x86-64#gpinstructions|bsr]] or [[x86-64#gpinstructions|lzcnt]] but has problems on processors with slow bitScan or 32-bit processors. Register usage is moderate and should suffice all x88-64 scratch registers for two lines, even for [https://en.wikipedia.org/wiki/Visual_C%2B%2B Visual C++].

* @author Michael Hoffmann, Gerd IsenbergDaniel Infuehr

long mMs1b = 0x8000000000000000 >>> Long.numberOfLeadingZeros (lower | 1); long ls1b odiff = upper & ^ (upper -upper; long odiff = 2*ls1b + mMs1b);