123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418 |
- //
- // math.s
- // x86 assembly-language math routines.
- #include "asm_i386.h"
- #include "quakeasm.h"
- #if id386
- .data
- .align 4
- Ljmptab: .long Lcase0, Lcase1, Lcase2, Lcase3
- .long Lcase4, Lcase5, Lcase6, Lcase7
- .text
- // TODO: rounding needed?
- // stack parameter offset
- #define val 4
- .globl C(Invert24To16)
- C(Invert24To16):
- movl val(
- movl $0x100,
- cmpl
- jle LOutOfRange
- subl
- divl
- ret
- LOutOfRange:
- movl $0xFFFFFFFF,
- ret
- #define in 4
- #define out 8
- .align 2
- .globl C(TransformVector)
- C(TransformVector):
- movl in(
- movl out(
- flds (
- fmuls C(vright) // in[0]*vright[0]
- flds (
- fmuls C(vup) // in[0]*vup[0] | in[0]*vright[0]
- flds (
- fmuls C(vpn) // in[0]*vpn[0] | in[0]*vup[0] | in[0]*vright[0]
- flds 4(
- fmuls C(vright)+4 // in[1]*vright[1] | ...
- flds 4(
- fmuls C(vup)+4 // in[1]*vup[1] | in[1]*vright[1] | ...
- flds 4(
- fmuls C(vpn)+4 // in[1]*vpn[1] | in[1]*vup[1] | in[1]*vright[1] | ...
- fxch
- faddp
- faddp
- faddp
- flds 8(
- fmuls C(vright)+8 // in[2]*vright[2] | ...
- flds 8(
- fmuls C(vup)+8 // in[2]*vup[2] | in[2]*vright[2] | ...
- flds 8(
- fmuls C(vpn)+8 // in[2]*vpn[2] | in[2]*vup[2] | in[2]*vright[2] | ...
- fxch
- faddp
- faddp
- faddp
- fstps 8(
- fstps 4(
- fstps (
- ret
- #define EMINS 4+4
- #define EMAXS 4+8
- #define P 4+12
- .align 2
- .globl C(BoxOnPlaneSide)
- C(BoxOnPlaneSide):
- pushl
- movl P(
- movl EMINS(
- xorl
- movl EMAXS(
- movb pl_signbits(
- cmpb $8,
- jge Lerror
- flds pl_normal(
- fld
- jmp Ljmptab(,
- //dist1= p->normal[0]*emaxs[0] + p->normal[1]*emaxs[1] + p->normal[2]*emaxs[2];
- //dist2= p->normal[0]*emins[0] + p->normal[1]*emins[1] + p->normal[2]*emins[2];
- Lcase0:
- fmuls (
- flds pl_normal+4(
- // p->normal[0]
- fxch
- // p->normal[1]
- fmuls (
- // p->normal[0]*emaxs[0] | p->normal[1]
- fxch
- // p->normal[0]*emins[0]
- fld
- // p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fmuls 4(
- // p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- flds pl_normal+8(
- // p->normal[1] | p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fxch
- // p->normal[2] | p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fmuls 4(
- // p->normal[1]*emaxs[1] |
- // p->normal[2] | p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fxch
- // p->normal[1]*emins[1] |
- // p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fld
- // p->normal[1]*emaxs[1] |
- // p->normal[1]*emins[1] |
- // p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fmuls 8(
- // p->normal[2] |
- // p->normal[1]*emaxs[1] |
- // p->normal[1]*emins[1] |
- // p->normal[0]*emaxs[0] |
- // p->normal[0]*emins[0]
- fxch
- // p->normal[2] |
- // p->normal[1]*emaxs[1] |
- // p->normal[1]*emins[1] |
- // p->normal[0]*emaxs[0] |
- // p->normal[2]*emaxs[2]
- faddp
- // p->normal[1]*emaxs[1] |
- // p->normal[1]*emins[1]+p->normal[0]*emins[0]|
- // p->normal[0]*emaxs[0] |
- // p->normal[2]*emaxs[2]
- fmuls 8(
- // p->normal[1]*emaxs[1] |
- // p->normal[1]*emins[1]+p->normal[0]*emins[0]|
- // p->normal[0]*emaxs[0] |
- // p->normal[2]*emaxs[2]
- fxch
- // p->normal[2]*emins[2] |
- // p->normal[1]*emins[1]+p->normal[0]*emins[0]|
- // p->normal[0]*emaxs[0] |
- // p->normal[2]*emaxs[2]
- faddp
- // p->normal[1]*emins[1]+p->normal[0]*emins[0]|
- // p->normal[0]*emaxs[0]+p->normal[1]*emaxs[1]|
- // p->normal[2]*emaxs[2]
- fxch
- // p->normal[1]*emins[1]+p->normal[0]*emins[0]|
- // p->normal[0]*emaxs[0]+p->normal[1]*emaxs[1]|
- // p->normal[2]*emins[2]
- faddp
- // dist1 | p->normal[2]*emins[2]
- jmp LSetSides
- //dist1= p->normal[0]*emins[0] + p->normal[1]*emaxs[1] + p->normal[2]*emaxs[2];
- //dist2= p->normal[0]*emaxs[0] + p->normal[1]*emins[1] + p->normal[2]*emins[2];
- Lcase1:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emaxs[0] + p->normal[1]*emins[1] + p->normal[2]*emaxs[2];
- //dist2= p->normal[0]*emins[0] + p->normal[1]*emaxs[1] + p->normal[2]*emins[2];
- Lcase2:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emins[0] + p->normal[1]*emins[1] + p->normal[2]*emaxs[2];
- //dist2= p->normal[0]*emaxs[0] + p->normal[1]*emaxs[1] + p->normal[2]*emins[2];
- Lcase3:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emaxs[0] + p->normal[1]*emaxs[1] + p->normal[2]*emins[2];
- //dist2= p->normal[0]*emins[0] + p->normal[1]*emins[1] + p->normal[2]*emaxs[2];
- Lcase4:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emins[0] + p->normal[1]*emaxs[1] + p->normal[2]*emins[2];
- //dist2= p->normal[0]*emaxs[0] + p->normal[1]*emins[1] + p->normal[2]*emaxs[2];
- Lcase5:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emaxs[0] + p->normal[1]*emins[1] + p->normal[2]*emins[2];
- //dist2= p->normal[0]*emins[0] + p->normal[1]*emaxs[1] + p->normal[2]*emaxs[2];
- Lcase6:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- jmp LSetSides
- //dist1= p->normal[0]*emins[0] + p->normal[1]*emins[1] + p->normal[2]*emins[2];
- //dist2= p->normal[0]*emaxs[0] + p->normal[1]*emaxs[1] + p->normal[2]*emaxs[2];
- Lcase7:
- fmuls (
- flds pl_normal+4(
- fxch
- fmuls (
- fxch
- fld
- fmuls 4(
- flds pl_normal+8(
- fxch
- fmuls 4(
- fxch
- fld
- fmuls 8(
- fxch
- faddp
- fmuls 8(
- fxch
- faddp
- fxch
- faddp
- LSetSides:
- // sides = 0;
- // if (dist1 >= p->dist)
- // sides = 1;
- // if (dist2 < p->dist)
- // sides |= 2;
- faddp
- fcomps pl_dist(
- xorl
- fnstsw
- fcomps pl_dist(
- andb $1,
- xorb $1,
- addb
- fnstsw
- andb $1,
- addb
- addb
- // return sides;
- popl
- movl
- ret
- Lerror:
- call C(BOPS_Error)
- #endif // id386
|