glibc/sysdeps/m88k/m88100/mul_1.S

128 lines
4.0 KiB
ArmAsm
Raw Normal View History

1995-02-18 01:27:10 +00:00
; mc88100 __mpn_mul_1 -- Multiply a limb vector with a single limb and
; store the product in a second limb vector.
; Copyright (C) 1992, 1994 Free Software Foundation, Inc.
; This file is part of the GNU MP Library.
; The GNU MP Library is free software; you can redistribute it and/or modify
; it under the terms of the GNU General Public License as published by
; the Free Software Foundation; either version 2, or (at your option)
; any later version.
; The GNU MP Library is distributed in the hope that it will be useful,
; but WITHOUT ANY WARRANTY; without even the implied warranty of
; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
; GNU General Public License for more details.
; You should have received a copy of the GNU General Public License
; along with the GNU MP Library; see the file COPYING. If not, write to
; the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
; INPUT PARAMETERS
; res_ptr r2
; s1_ptr r3
; size r4
; s2_limb r5
; Common overhead is about 11 cycles/invocation.
; The speed for S2_LIMB >= 0x10000 is approximately 21 cycles/limb. (The
; pipeline stalls 2 cycles due to WB contention.)
; The speed for S2_LIMB < 0x10000 is approximately 16 cycles/limb. (The
; pipeline stalls 2 cycles due to WB contention and 1 cycle due to latency.)
; To enhance speed:
; 1. Unroll main loop 4-8 times.
; 2. Schedule code to avoid WB contention. It might be tempting to move the
; ld instruction in the loops down to save 2 cycles (less WB contention),
; but that looses because the ultimate value will be read from outside
; the allocated space. But if we handle the ultimate multiplication in
; the tail, we can do this.
; 3. Make the multiplication with less instructions. I think the code for
; (S2_LIMB >= 0x10000) is not minimal.
; With these techniques the (S2_LIMB >= 0x10000) case would run in 17 or
; less cycles/limb; the (S2_LIMB < 0x10000) case would run in 11
; cycles/limb. (Assuming infinite unrolling.)
#include "sysdep.h"
ENTRY (__mpn_mul_1)
; Make S1_PTR and RES_PTR point at the end of their blocks
; and negate SIZE.
lda r3,r3[r4]
lda r6,r2[r4] ; RES_PTR in r6 since r2 is retval
subu r4,r0,r4
addu.co r2,r0,r0 ; r2 = cy = 0
ld r9,r3[r4]
mask r7,r5,0xffff ; r7 = lo(S2_LIMB)
extu r8,r5,16 ; r8 = hi(S2_LIMB)
bcnd.n eq0,r8,Lsmall ; jump if (hi(S2_LIMB) == 0)
subu r6,r6,4
; General code for any value of S2_LIMB.
; Make a stack frame and save r25 and r26
subu r31,r31,16
st.d r25,r31,8
; Enter the loop in the middle
br.n L1
addu r4,r4,1
Loop:
ld r9,r3[r4]
st r26,r6[r4]
; bcnd ne0,r0,0 ; bubble
addu r4,r4,1
L1: mul r26,r9,r5 ; low word of product mul_1 WB ld
mask r12,r9,0xffff ; r12 = lo(s1_limb) mask_1
mul r11,r12,r7 ; r11 = prod_0 mul_2 WB mask_1
mul r10,r12,r8 ; r10 = prod_1a mul_3
extu r13,r9,16 ; r13 = hi(s1_limb) extu_1 WB mul_1
mul r12,r13,r7 ; r12 = prod_1b mul_4 WB extu_1
mul r25,r13,r8 ; r25 = prod_2 mul_5 WB mul_2
extu r11,r11,16 ; r11 = hi(prod_0) extu_2 WB mul_3
addu r10,r10,r11 ; addu_1 WB extu_2
; bcnd ne0,r0,0 ; bubble WB addu_1
addu.co r10,r10,r12 ; WB mul_4
mask.u r10,r10,0xffff ; move the 16 most significant bits...
addu.ci r10,r10,r0 ; ...to the low half of the word...
rot r10,r10,16 ; ...and put carry in pos 16.
addu.co r26,r26,r2 ; add old carry limb
bcnd.n ne0,r4,Loop
addu.ci r2,r25,r10 ; compute new carry limb
st r26,r6[r4]
ld.d r25,r31,8
jmp.n r1
addu r31,r31,16
; Fast code for S2_LIMB < 0x10000
Lsmall:
; Enter the loop in the middle
br.n SL1
addu r4,r4,1
SLoop:
ld r9,r3[r4] ;
st r8,r6[r4] ;
addu r4,r4,1 ;
SL1: mul r8,r9,r5 ; low word of product
mask r12,r9,0xffff ; r12 = lo(s1_limb)
extu r13,r9,16 ; r13 = hi(s1_limb)
mul r11,r12,r7 ; r11 = prod_0
mul r12,r13,r7 ; r12 = prod_1b
addu.cio r8,r8,r2 ; add old carry limb
extu r10,r11,16 ; r11 = hi(prod_0)
addu r10,r10,r12 ;
bcnd.n ne0,r4,SLoop
extu r2,r10,16 ; r2 = new carry limb
jmp.n r1
st r8,r6[r4]