diff options
Diffstat (limited to 'sysdeps/sparc/add_n.S')
| -rw-r--r-- | sysdeps/sparc/add_n.S | 304 |
1 files changed, 194 insertions, 110 deletions
diff --git a/sysdeps/sparc/add_n.S b/sysdeps/sparc/add_n.S index 13704d32d2..80c3b99640 100644 --- a/sysdeps/sparc/add_n.S +++ b/sysdeps/sparc/add_n.S @@ -1,7 +1,7 @@ ! sparc __mpn_add_n -- Add two limb vectors of the same length > 0 and store ! sum in a third limb vector. -! Copyright (C) 1992, 1994, 1995 Free Software Foundation, Inc. +! Copyright (C) 1995 Free Software Foundation, Inc. ! This file is part of the GNU MP Library. @@ -21,10 +21,10 @@ ! INPUT PARAMETERS -! res_ptr %o0 -! s1_ptr %o1 -! s2_ptr %o2 -! size %o3 +#define res_ptr %o0 +#define s1_ptr %o1 +#define s2_ptr %o2 +#define size %o3 #include "sysdep.h" @@ -32,108 +32,192 @@ .align 4 .global C_SYMBOL_NAME(__mpn_add_n) C_SYMBOL_NAME(__mpn_add_n): - ld [%o1+0],%o4 ! read first limb from s1_ptr - srl %o3,4,%g1 - ld [%o2+0],%o5 ! read first limb from s2_ptr - - sub %g0,%o3,%o3 - andcc %o3,(16-1),%o3 - be Lzero - mov %o4,%g2 ! put first s1_limb in g2 too - - sll %o3,2,%o3 ! multiply by 4 - sub %o0,%o3,%o0 ! adjust res_ptr - sub %o1,%o3,%o1 ! adjust s1_ptr - sub %o2,%o3,%o2 ! adjust s2_ptr - -#if PIC - mov %o7,%g4 ! Save return address register - call 1f - add %o7,Lbase-1f,%g3 -1: mov %g4,%o7 ! Restore return address register -#else - sethi %hi(Lbase),%g3 - or %g3,%lo(Lbase),%g3 -#endif - sll %o3,2,%o3 ! multiply by 4 - jmp %g3+%o3 - mov %o5,%g3 ! put first s2_limb in g3 too - -Loop: addxcc %g2,%g3,%o3 - add %o1,64,%o1 - st %o3,[%o0+60] - add %o2,64,%o2 - ld [%o1+0],%o4 - add %o0,64,%o0 - ld [%o2+0],%o5 -Lzero: sub %g1,1,%g1 ! add 0 + 16r limbs (adjust loop counter) -Lbase: ld [%o1+4],%g2 - addxcc %o4,%o5,%o3 - ld [%o2+4],%g3 - st %o3,[%o0+0] - ld [%o1+8],%o4 ! add 15 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+8],%o5 - st %o3,[%o0+4] - ld [%o1+12],%g2 ! add 14 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+12],%g3 - st %o3,[%o0+8] - ld [%o1+16],%o4 ! add 13 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+16],%o5 - st %o3,[%o0+12] - ld [%o1+20],%g2 ! add 12 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+20],%g3 - st %o3,[%o0+16] - ld [%o1+24],%o4 ! add 11 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+24],%o5 - st %o3,[%o0+20] - ld [%o1+28],%g2 ! add 10 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+28],%g3 - st %o3,[%o0+24] - ld [%o1+32],%o4 ! add 9 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+32],%o5 - st %o3,[%o0+28] - ld [%o1+36],%g2 ! add 8 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+36],%g3 - st %o3,[%o0+32] - ld [%o1+40],%o4 ! add 7 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+40],%o5 - st %o3,[%o0+36] - ld [%o1+44],%g2 ! add 6 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+44],%g3 - st %o3,[%o0+40] - ld [%o1+48],%o4 ! add 5 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+48],%o5 - st %o3,[%o0+44] - ld [%o1+52],%g2 ! add 4 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+52],%g3 - st %o3,[%o0+48] - ld [%o1+56],%o4 ! add 3 + 16r limbs - addxcc %g2,%g3,%o3 - ld [%o2+56],%o5 - st %o3,[%o0+52] - ld [%o1+60],%g2 ! add 2 + 16r limbs - addxcc %o4,%o5,%o3 - ld [%o2+60],%g3 - st %o3,[%o0+56] - addx %g0,%g0,%o4 - tst %g1 - bne Loop - subcc %g0,%o4,%g0 ! restore cy (delay slot) - - addxcc %g2,%g3,%o3 - st %o3,[%o0+60] ! store most significant limb - - retl - addx %g0,%g0,%o0 ! return carry-out from most sign. limb + cmp size,8 + mov 0,%o4 ! clear cy-save register + blt,a Ltriv + addcc size,-2,size + xor s2_ptr,res_ptr,%g1 + andcc %g1,4,%g0 + bne L1 ! branch if alignment differs + nop +L0: andcc res_ptr,4,%g0 ! res_ptr unaligned? Side effect: cy=0 + beq L_v1 ! if no, branch + nop +! ** V1a ** +/* Add least significant limb separately to align res_ptr and s2_ptr */ + ld [s1_ptr],%g4 + add s1_ptr,4,s1_ptr + ld [s2_ptr],%g2 + add s2_ptr,4,s2_ptr + add size,-1,size + addcc %g4,%g2,%o4 + st %o4,[res_ptr] + add res_ptr,4,res_ptr + +L_v1: ld [s1_ptr+0],%g4 + ld [s1_ptr+4],%g1 + ldd [s2_ptr+0],%g2 + addx %g0,%g0,%o4 ! save cy in register + addcc size,-10,size + blt Lfin1 + subcc %g0,%o4,%g0 ! restore cy +/* Add blocks of 8 limbs until less than 8 limbs remain */ +Loop1: addxcc %g4,%g2,%o4 + ld [s1_ptr+8],%g4 + addxcc %g1,%g3,%o5 + ld [s1_ptr+12],%g1 + ldd [s2_ptr+8],%g2 + std %o4,[res_ptr+0] + addxcc %g4,%g2,%o4 + ld [s1_ptr+16],%g4 + addxcc %g1,%g3,%o5 + ld [s1_ptr+20],%g1 + ldd [s2_ptr+16],%g2 + std %o4,[res_ptr+8] + addxcc %g4,%g2,%o4 + ld [s1_ptr+24],%g4 + addxcc %g1,%g3,%o5 + ld [s1_ptr+28],%g1 + ldd [s2_ptr+24],%g2 + std %o4,[res_ptr+16] + addxcc %g4,%g2,%o4 + ld [s1_ptr+32],%g4 + addxcc %g1,%g3,%o5 + ld [s1_ptr+36],%g1 + ldd [s2_ptr+32],%g2 + std %o4,[res_ptr+24] + addx %g0,%g0,%o4 ! save cy in register + addcc size,-8,size + add s1_ptr,32,s1_ptr + add s2_ptr,32,s2_ptr + add res_ptr,32,res_ptr + bge Loop1 + subcc %g0,%o4,%g0 ! restore cy + +Lfin1: addcc size,8-2,size + blt Lend1 + subcc %g0,%o4,%g0 ! restore cy +/* Add blocks of 2 limbs until less than 2 limbs remain */ +Loop1b: addxcc %g4,%g2,%o4 + ld [s1_ptr+8],%g4 + addxcc %g1,%g3,%o5 + ld [s1_ptr+12],%g1 + ldd [s2_ptr+8],%g2 + std %o4,[res_ptr+0] + addx %g0,%g0,%o4 ! save cy in register + addcc size,-2,size + add s1_ptr,8,s1_ptr + add s2_ptr,8,s2_ptr + add res_ptr,8,res_ptr + bge Loop1b + subcc %g0,%o4,%g0 ! restore cy +Lend1: addxcc %g4,%g2,%o4 + addxcc %g1,%g3,%o5 + std %o4,[res_ptr+0] + addx %g0,%g0,%o4 ! save cy in register + + andcc size,1,%g0 + be Lret1 + subcc %g0,%o4,%g0 ! restore cy +/* Add last limb */ + ld [s1_ptr+8],%g4 + ld [s2_ptr+8],%g2 + addxcc %g4,%g2,%o4 + st %o4,[res_ptr+8] + +Lret1: retl + addx %g0,%g0,%o0 ! return carry-out from most sign. limb + +L1: xor s1_ptr,res_ptr,%g1 + andcc %g1,4,%g0 + bne L2 + nop +! ** V1b ** + mov s2_ptr,%g1 + mov s1_ptr,s2_ptr + b L0 + mov %g1,s1_ptr + +! ** V2 ** +/* If we come here, the alignment of s1_ptr and res_ptr as well as the + alignment of s2_ptr and res_ptr differ. Since there are only two ways + things can be aligned (that we care about) we now know that the alignment + of s1_ptr and s2_ptr are the same. */ + +L2: andcc s1_ptr,4,%g0 ! s1_ptr unaligned? Side effect: cy=0 + beq L_v2 ! if no, branch + nop +/* Add least significant limb separately to align res_ptr and s2_ptr */ + ld [s1_ptr],%g4 + add s1_ptr,4,s1_ptr + ld [s2_ptr],%g2 + add s2_ptr,4,s2_ptr + add size,-1,size + addcc %g4,%g2,%o4 + st %o4,[res_ptr] + add res_ptr,4,res_ptr + +L_v2: addx %g0,%g0,%o4 ! save cy in register + addcc size,-8,size + blt Lfin2 + subcc %g0,%o4,%g0 ! restore cy +/* Add blocks of 8 limbs until less than 8 limbs remain */ +Loop2: ldd [s1_ptr+0],%g2 + ldd [s2_ptr+0],%o4 + addxcc %g2,%o4,%g2 + st %g2,[res_ptr+0] + addxcc %g3,%o5,%g3 + st %g3,[res_ptr+4] + ldd [s1_ptr+8],%g2 + ldd [s2_ptr+8],%o4 + addxcc %g2,%o4,%g2 + st %g2,[res_ptr+8] + addxcc %g3,%o5,%g3 + st %g3,[res_ptr+12] + ldd [s1_ptr+16],%g2 + ldd [s2_ptr+16],%o4 + addxcc %g2,%o4,%g2 + st %g2,[res_ptr+16] + addxcc %g3,%o5,%g3 + st %g3,[res_ptr+20] + ldd [s1_ptr+24],%g2 + ldd [s2_ptr+24],%o4 + addxcc %g2,%o4,%g2 + st %g2,[res_ptr+24] + addxcc %g3,%o5,%g3 + st %g3,[res_ptr+28] + addx %g0,%g0,%o4 ! save cy in register + addcc size,-8,size + add s1_ptr,32,s1_ptr + add s2_ptr,32,s2_ptr + add res_ptr,32,res_ptr + bge Loop2 + subcc %g0,%o4,%g0 ! restore cy + +Lfin2: addcc size,8-2,size +Ltriv: blt Lend2 + subcc %g0,%o4,%g0 ! restore cy +Loop2b: ldd [s1_ptr+0],%g2 + ldd [s2_ptr+0],%o4 + addxcc %g2,%o4,%g2 + st %g2,[res_ptr+0] + addxcc %g3,%o5,%g3 + st %g3,[res_ptr+4] + addx %g0,%g0,%o4 ! save cy in register + addcc size,-2,size + add s1_ptr,8,s1_ptr + add s2_ptr,8,s2_ptr + add res_ptr,8,res_ptr + bge Loop2b + subcc %g0,%o4,%g0 ! restore cy +Lend2: andcc size,1,%g0 + be Lret2 + subcc %g0,%o4,%g0 ! restore cy +/* Add last limb */ + ld [s1_ptr],%g4 + ld [s2_ptr],%g2 + addxcc %g4,%g2,%o4 + st %o4,[res_ptr] + +Lret2: retl + addx %g0,%g0,%o0 ! return carry-out from most sign. limb |