sysdeps/m88k/m88100/mul_1.s

   1 ; mc88100 __mpn_mul_1 -- Multiply a limb vector with a single limb and
   2 ; store the product in a second limb vector.
   3
   4 ; Copyright (C) 1992, 1994 Free Software Foundation, Inc.
   5
   6 ; This file is part of the GNU MP Library.
   7
   8 ; The GNU MP Library is free software; you can redistribute it and/or modify
   9 ; it under the terms of the GNU Lesser General Public License as published by
  10 ; the Free Software Foundation; either version 2.1 of the License, or (at your
  11 ; option) any later version.
  12
  13 ; The GNU MP Library is distributed in the hope that it will be useful, but
  14 ; WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15 ; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  16 ; License for more details.
  17
  18 ; You should have received a copy of the GNU Lesser General Public License
  19 ; along with the GNU MP Library; see the file COPYING.LIB.  If not, write to
  20 ; the Free Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21
  22
  23 ; INPUT PARAMETERS
  24 ; res_ptr       r2
  25 ; s1_ptr        r3
  26 ; size          r4
  27 ; s2_limb       r5
  28
  29 ; Common overhead is about 11 cycles/invocation.
  30
  31 ; The speed for S2_LIMB >= 0x10000 is approximately 21 cycles/limb.  (The
  32 ; pipeline stalls 2 cycles due to WB contention.)
  33
  34 ; The speed for S2_LIMB < 0x10000 is approximately 16 cycles/limb.  (The
  35 ; pipeline stalls 2 cycles due to WB contention and 1 cycle due to latency.)
  36
  37 ; To enhance speed:
  38 ; 1. Unroll main loop 4-8 times.
  39 ; 2. Schedule code to avoid WB contention.  It might be tempting to move the
  40 ;    ld instruction in the loops down to save 2 cycles (less WB contention),
  41 ;    but that looses because the ultimate value will be read from outside
  42 ;    the allocated space.  But if we handle the ultimate multiplication in
  43 ;    the tail, we can do this.
  44 ; 3. Make the multiplication with less instructions.  I think the code for
  45 ;    (S2_LIMB >= 0x10000) is not minimal.
  46 ; With these techniques the (S2_LIMB >= 0x10000) case would run in 17 or
  47 ; less cycles/limb; the (S2_LIMB < 0x10000) case would run in 11
  48 ; cycles/limb.  (Assuming infinite unrolling.)
  49
  50         text
  51         align    16
  52         global   ___mpn_mul_1
  53 ___mpn_mul_1:
  54
  55         ; Make S1_PTR and RES_PTR point at the end of their blocks
  56         ; and negate SIZE.
  57         lda      r3,r3[r4]
  58         lda      r6,r2[r4]              ; RES_PTR in r6 since r2 is retval
  59         subu     r4,r0,r4
  60
  61         addu.co  r2,r0,r0               ; r2 = cy = 0
  62         ld       r9,r3[r4]
  63         mask     r7,r5,0xffff           ; r7 = lo(S2_LIMB)
  64         extu     r8,r5,16               ; r8 = hi(S2_LIMB)
  65         bcnd.n   eq0,r8,Lsmall          ; jump if (hi(S2_LIMB) == 0)
  66          subu    r6,r6,4
  67
  68 ; General code for any value of S2_LIMB.
  69
  70         ; Make a stack frame and save r25 and r26
  71         subu     r31,r31,16
  72         st.d     r25,r31,8
  73
  74         ; Enter the loop in the middle
  75         br.n    L1
  76         addu     r4,r4,1
  77
  78 Loop:
  79         ld       r9,r3[r4]
  80         st       r26,r6[r4]
  81 ; bcnd  ne0,r0,0                        ; bubble
  82         addu     r4,r4,1
  83 L1:     mul      r26,r9,r5              ; low word of product   mul_1   WB ld
  84         mask     r12,r9,0xffff          ; r12 = lo(s1_limb)     mask_1
  85         mul      r11,r12,r7             ; r11 =  prod_0         mul_2   WB mask_1
  86         mul      r10,r12,r8             ; r10 = prod_1a         mul_3
  87         extu     r13,r9,16              ; r13 = hi(s1_limb)     extu_1  WB mul_1
  88         mul      r12,r13,r7             ; r12 = prod_1b         mul_4   WB extu_1
  89         mul      r25,r13,r8             ; r25  = prod_2         mul_5   WB mul_2
  90         extu     r11,r11,16             ; r11 = hi(prod_0)      extu_2  WB mul_3
  91         addu     r10,r10,r11            ;                       addu_1  WB extu_2
  92 ; bcnd  ne0,r0,0                        ; bubble                        WB addu_1
  93         addu.co  r10,r10,r12            ;                               WB mul_4
  94         mask.u   r10,r10,0xffff         ; move the 16 most significant bits...
  95         addu.ci  r10,r10,r0             ; ...to the low half of the word...
  96         rot      r10,r10,16             ; ...and put carry in pos 16.
  97         addu.co  r26,r26,r2             ; add old carry limb
  98         bcnd.n   ne0,r4,Loop
  99          addu.ci r2,r25,r10             ; compute new carry limb
 100
 101         st       r26,r6[r4]
 102         ld.d     r25,r31,8
 103         jmp.n    r1
 104          addu    r31,r31,16
 105
 106 ; Fast code for S2_LIMB < 0x10000
 107 Lsmall:
 108         ; Enter the loop in the middle
 109         br.n    SL1
 110         addu     r4,r4,1
 111
 112 SLoop:
 113         ld       r9,r3[r4]              ;
 114         st       r8,r6[r4]              ;
 115         addu     r4,r4,1                ;
 116 SL1:    mul      r8,r9,r5               ; low word of product
 117         mask     r12,r9,0xffff          ; r12 = lo(s1_limb)
 118         extu     r13,r9,16              ; r13 = hi(s1_limb)
 119         mul      r11,r12,r7             ; r11 =  prod_0
 120         mul      r12,r13,r7             ; r12 = prod_1b
 121         addu.cio r8,r8,r2               ; add old carry limb
 122         extu     r10,r11,16             ; r11 = hi(prod_0)
 123         addu     r10,r10,r12            ;
 124         bcnd.n   ne0,r4,SLoop
 125         extu     r2,r10,16              ; r2 = new carry limb
 126
 127         jmp.n    r1
 128         st       r8,r6[r4]