aboutsummaryrefslogtreecommitdiff
path: root/sysdeps/m88k/m88100
diff options
context:
space:
mode:
authorRoland McGrath <roland@gnu.org>1995-02-18 01:27:10 +0000
committerRoland McGrath <roland@gnu.org>1995-02-18 01:27:10 +0000
commit28f540f45bbacd939bfd07f213bcad2bf730b1bf (patch)
tree15f07c4c43d635959c6afee96bde71fb1b3614ee /sysdeps/m88k/m88100
downloadglibc-28f540f45bbacd939bfd07f213bcad2bf730b1bf.tar
glibc-28f540f45bbacd939bfd07f213bcad2bf730b1bf.tar.gz
glibc-28f540f45bbacd939bfd07f213bcad2bf730b1bf.tar.bz2
glibc-28f540f45bbacd939bfd07f213bcad2bf730b1bf.zip
initial import
Diffstat (limited to 'sysdeps/m88k/m88100')
-rw-r--r--sysdeps/m88k/m88100/add_n.S133
-rw-r--r--sysdeps/m88k/m88100/mul_1.S127
-rw-r--r--sysdeps/m88k/m88100/sub_n.S134
3 files changed, 394 insertions, 0 deletions
diff --git a/sysdeps/m88k/m88100/add_n.S b/sysdeps/m88k/m88100/add_n.S
new file mode 100644
index 0000000000..2107eb55cb
--- /dev/null
+++ b/sysdeps/m88k/m88100/add_n.S
@@ -0,0 +1,133 @@
+; mc88100 __mpn_add -- Add two limb vectors of the same length > 0 and store
+; sum in a third 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
+; s2_ptr r4
+; size r5
+
+; This code has been optimized to run one instruction per clock, avoiding
+; load stalls and writeback contention. As a result, the instruction
+; order is not always natural.
+
+; The speed is approximately 4.3 clocks/limb + 18 clocks/limb-vector.
+
+#include "sysdep.h"
+
+ENTRY (__mpn_add_n)
+ ld r6,r3,0 ; read first limb from s1_ptr
+ extu r10,r5,4
+ ld r7,r4,0 ; read first limb from s2_ptr
+
+ subu.co r5,r0,r5 ; (clear carry as side effect)
+ mak r5,r5,4<4>
+ bcnd eq0,r5,Lzero
+
+ or r12,r0,lo16(Lbase)
+ or.u r12,r12,hi16(Lbase)
+ addu r12,r12,r5 ; r12 is address for entering in loop
+
+ extu r5,r5,2 ; divide by 4
+ subu r2,r2,r5 ; adjust res_ptr
+ subu r3,r3,r5 ; adjust s1_ptr
+ subu r4,r4,r5 ; adjust s2_ptr
+
+ or r8,r6,r0
+
+ jmp.n r12
+ or r9,r7,r0
+
+Loop: addu r3,r3,64
+ st r8,r2,60
+ addu r4,r4,64
+ ld r6,r3,0
+ addu r2,r2,64
+ ld r7,r4,0
+Lzero: subu r10,r10,1 ; add 0 + 16r limbs (adjust loop counter)
+Lbase: ld r8,r3,4
+ addu.cio r6,r6,r7
+ ld r9,r4,4
+ st r6,r2,0
+ ld r6,r3,8 ; add 15 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,8
+ st r8,r2,4
+ ld r8,r3,12 ; add 14 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,12
+ st r6,r2,8
+ ld r6,r3,16 ; add 13 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,16
+ st r8,r2,12
+ ld r8,r3,20 ; add 12 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,20
+ st r6,r2,16
+ ld r6,r3,24 ; add 11 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,24
+ st r8,r2,20
+ ld r8,r3,28 ; add 10 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,28
+ st r6,r2,24
+ ld r6,r3,32 ; add 9 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,32
+ st r8,r2,28
+ ld r8,r3,36 ; add 8 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,36
+ st r6,r2,32
+ ld r6,r3,40 ; add 7 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,40
+ st r8,r2,36
+ ld r8,r3,44 ; add 6 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,44
+ st r6,r2,40
+ ld r6,r3,48 ; add 5 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,48
+ st r8,r2,44
+ ld r8,r3,52 ; add 4 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,52
+ st r6,r2,48
+ ld r6,r3,56 ; add 3 + 16r limbs
+ addu.cio r8,r8,r9
+ ld r7,r4,56
+ st r8,r2,52
+ ld r8,r3,60 ; add 2 + 16r limbs
+ addu.cio r6,r6,r7
+ ld r9,r4,60
+ st r6,r2,56
+ bcnd.n ne0,r10,Loop ; add 1 + 16r limbs
+ addu.cio r8,r8,r9
+
+ st r8,r2,60 ; store most significant limb
+
+ jmp.n r1
+ addu.ci r2,r0,r0 ; return carry-out from most sign. limb
diff --git a/sysdeps/m88k/m88100/mul_1.S b/sysdeps/m88k/m88100/mul_1.S
new file mode 100644
index 0000000000..503897b298
--- /dev/null
+++ b/sysdeps/m88k/m88100/mul_1.S
@@ -0,0 +1,127 @@
+; 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]
diff --git a/sysdeps/m88k/m88100/sub_n.S b/sysdeps/m88k/m88100/sub_n.S
new file mode 100644
index 0000000000..927ece4f15
--- /dev/null
+++ b/sysdeps/m88k/m88100/sub_n.S
@@ -0,0 +1,134 @@
+; mc88100 __mpn_sub -- Subtract two limb vectors of the same length > 0 and
+; store difference in a third 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
+; s2_ptr r4
+; size r5
+
+; This code has been optimized to run one instruction per clock, avoiding
+; load stalls and writeback contention. As a result, the instruction
+; order is not always natural.
+
+; The speed is approximately 4.3 clocks/limb + 18 clocks/limb-vector.
+
+#include "sysdep.h"
+
+ENTRY (__mpn_sub_n)
+ ld r6,r3,0 ; read first limb from s1_ptr
+ extu r10,r5,4
+ ld r7,r4,0 ; read first limb from s2_ptr
+
+ subu.co r5,r0,r5 ; (clear carry as side effect)
+ mak r5,r5,4<4>
+ bcnd eq0,r5,Lzero
+
+ or r12,r0,lo16(Lbase)
+ or.u r12,r12,hi16(Lbase)
+ addu r12,r12,r5 ; r12 is address for entering in loop
+
+ extu r5,r5,2 ; divide by 4
+ subu r2,r2,r5 ; adjust res_ptr
+ subu r3,r3,r5 ; adjust s1_ptr
+ subu r4,r4,r5 ; adjust s2_ptr
+
+ or r8,r6,r0
+
+ jmp.n r12
+ or r9,r7,r0
+
+Loop: addu r3,r3,64
+ st r8,r2,60
+ addu r4,r4,64
+ ld r6,r3,0
+ addu r2,r2,64
+ ld r7,r4,0
+Lzero: subu r10,r10,1 ; subtract 0 + 16r limbs (adjust loop counter)
+Lbase: ld r8,r3,4
+ subu.cio r6,r6,r7
+ ld r9,r4,4
+ st r6,r2,0
+ ld r6,r3,8 ; subtract 15 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,8
+ st r8,r2,4
+ ld r8,r3,12 ; subtract 14 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,12
+ st r6,r2,8
+ ld r6,r3,16 ; subtract 13 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,16
+ st r8,r2,12
+ ld r8,r3,20 ; subtract 12 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,20
+ st r6,r2,16
+ ld r6,r3,24 ; subtract 11 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,24
+ st r8,r2,20
+ ld r8,r3,28 ; subtract 10 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,28
+ st r6,r2,24
+ ld r6,r3,32 ; subtract 9 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,32
+ st r8,r2,28
+ ld r8,r3,36 ; subtract 8 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,36
+ st r6,r2,32
+ ld r6,r3,40 ; subtract 7 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,40
+ st r8,r2,36
+ ld r8,r3,44 ; subtract 6 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,44
+ st r6,r2,40
+ ld r6,r3,48 ; subtract 5 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,48
+ st r8,r2,44
+ ld r8,r3,52 ; subtract 4 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,52
+ st r6,r2,48
+ ld r6,r3,56 ; subtract 3 + 16r limbs
+ subu.cio r8,r8,r9
+ ld r7,r4,56
+ st r8,r2,52
+ ld r8,r3,60 ; subtract 2 + 16r limbs
+ subu.cio r6,r6,r7
+ ld r9,r4,60
+ st r6,r2,56
+ bcnd.n ne0,r10,Loop ; subtract 1 + 16r limbs
+ subu.cio r8,r8,r9
+
+ st r8,r2,60 ; store most significant limb
+
+ addu.ci r2,r0,r0 ; return carry-out from most sign. limb
+ jmp.n r1
+ xor r2,r2,1