Return to addmul_1.asm CVS log

Up to [local] / OpenXM_contrib / gmp / mpn / sparc64

Annotation of OpenXM_contrib/gmp/mpn/sparc64/addmul_1.asm, Revision 1.1.1.1

1.1       maekawa     1: dnl  SPARC 64-bit mpn_addmul_1 -- Multiply a limb vector with a limb and
                      2: dnl  add the result to a second limb vector.
                      3:
                      4: dnl  Copyright (C) 1998, 2000 Free Software Foundation, Inc.
                      5:
                      6: dnl  This file is part of the GNU MP Library.
                      7:
                      8: dnl  The GNU MP Library is free software; you can redistribute it and/or modify
                      9: dnl  it under the terms of the GNU Lesser General Public License as published
                     10: dnl  by the Free Software Foundation; either version 2.1 of the License, or (at
                     11: dnl  your option) any later version.
                     12:
                     13: dnl  The GNU MP Library is distributed in the hope that it will be useful, but
                     14: dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
                     15: dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
                     16: dnl  License for more details.
                     17:
                     18: dnl  You should have received a copy of the GNU Lesser General Public License
                     19: dnl  along with the GNU MP Library; see the file COPYING.LIB.  If not, write to
                     20: dnl  the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
                     21: dnl  MA 02111-1307, USA.
                     22:
                     23: include(`../config.m4')
                     24:
                     25: C INPUT PARAMETERS
                     26: C res_ptr      i0
                     27: C s1_ptr       i1
                     28: C size         i2
                     29: C s2_limb      i3
                     30:
                     31: ASM_START()
                     32:        .register       %g2,#scratch
                     33:        .register       %g3,#scratch
                     34:
                     35: PROLOGUE(mpn_addmul_1)
                     36:        save    %sp,-256,%sp
                     37:
                     38: C We store 0.0 in f10 and keep it invariant accross thw two
                     39: C function calls below.  Note that this is not ABI conformant,
                     40: C but since the functions are local, that's acceptable.
                     41: ifdef(`PIC',
                     42: `L(pc):        rd      %pc,%o7
                     43:        ld      [%o7+L(noll)-L(pc)],%f10',
                     44: `      sethi   %hh(L(noll)),%g2
                     45:        sethi   %lm(L(noll)),%g1
                     46:        or      %g2,%hm(L(noll)),%g2
                     47:        or      %g1,%lo(L(noll)),%g1
                     48:        sllx    %g2,32,%g2
                     49:        ld      [%g1+%g2],%f10')
                     50:
                     51:        sub     %i1,%i0,%g1
                     52:        srlx    %g1,3,%g1
                     53:        cmp     %g1,%i2
                     54:        bcc,pt  %xcc,L(nooverlap)
                     55:        nop
                     56:
                     57:        sllx    %i2,3,%g2               C compute stack allocation byte count
                     58:        add     %g2,15,%o0
                     59:        and     %o0,-16,%o0
                     60:        sub     %sp,%o0,%sp
                     61:        add     %sp,2223,%o0
                     62:
                     63:        mov     %i1,%o1                 C copy s1_ptr to mpn_copyi's srcp
                     64:        call    mpn_copyi
                     65:        mov     %i2,%o2                 C copy n to mpn_copyi's count parameter
                     66:
                     67:        add     %sp,2223,%i1
                     68:
                     69: L(nooverlap):
                     70: C First multiply-add with low 32 bits of s2_limb
                     71:        mov     %i0,%o0
                     72:        mov     %i1,%o1
                     73:        add     %i2,%i2,%o2
                     74:        call    addmull
                     75:        srl     %i3,0,%o3
                     76:
                     77:        mov     %o0,%l0                 C keep carry-out from accmull
                     78:
                     79: C Now multiply-add with high 32 bits of s2_limb, unless it is zero.
                     80:        srlx    %i3,32,%o3
                     81:        brz,a,pn        %o3,L(small)
                     82:         mov    %o0,%i0
                     83:        mov     %i1,%o1
                     84:        add     %i2,%i2,%o2
                     85:        call    addmulu
                     86:        add     %i0,4,%o0
                     87:
                     88:        add     %l0,%o0,%i0
                     89: L(small):
                     90:        ret
                     91:        restore %g0,%g0,%g0
                     92: EPILOGUE(mpn_addmul_1)
                     93:
                     94: C Put a zero in the text segment to allow us to t the address
                     95: C quickly when compiling for PIC
                     96:        TEXT
                     97:        ALIGN(4)
                     98: L(noll):
                     99:        .word   0
                    100:
                    101: define(`LO',`(+4)')
                    102: define(`HI',`(-4)')
                    103:
                    104: define(`DLO',`(+4)')
                    105: define(`DHI',`(-4)')
                    106: define(`LOWPART')
                    107: define(`E',`L(l.$1)')
                    108: include_mpn(`sparc64/addmul1h.asm')
                    109:
                    110: define(`DLO',`(-4)')
                    111: define(`DHI',`(+4)')
                    112: undefine(`LOWPART')
                    113: define(`E',`L(u.$1)')
                    114: include_mpn(`sparc64/addmul1h.asm')