|
Packit |
5c3484 |
dnl AMD64 mpn_sqr_basecase optimised for Intel Sandy bridge and Ivy bridge.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjörn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2008, 2009, 2011-2013 Free Software Foundation, Inc.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl This file is part of the GNU MP Library.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify
|
|
Packit |
5c3484 |
dnl it under the terms of either:
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU Lesser General Public License as published by the Free
|
|
Packit |
5c3484 |
dnl Software Foundation; either version 3 of the License, or (at your
|
|
Packit |
5c3484 |
dnl option) any later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU General Public License as published by the Free Software
|
|
Packit |
5c3484 |
dnl Foundation; either version 2 of the License, or (at your option) any
|
|
Packit |
5c3484 |
dnl later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or both in parallel, as here.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but
|
|
Packit |
5c3484 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
|
Packit |
5c3484 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
Packit |
5c3484 |
dnl for more details.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl You should have received copies of the GNU General Public License and the
|
|
Packit |
5c3484 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not,
|
|
Packit |
5c3484 |
dnl see https://www.gnu.org/licenses/.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
include(`../config.m4')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C cycles/limb mul_2 addmul_2 sqr_diag_addlsh1
|
|
Packit |
5c3484 |
C AMD K8,K9 ? ? ?
|
|
Packit |
5c3484 |
C AMD K10 ? ? ?
|
|
Packit |
5c3484 |
C AMD bull ? ? ?
|
|
Packit |
5c3484 |
C AMD pile ? ? ?
|
|
Packit |
5c3484 |
C AMD steam ? ? ?
|
|
Packit |
5c3484 |
C AMD bobcat ? ? ?
|
|
Packit |
5c3484 |
C AMD jaguar ? ? ?
|
|
Packit |
5c3484 |
C Intel P4 ? ? ?
|
|
Packit |
5c3484 |
C Intel core ? ? ?
|
|
Packit |
5c3484 |
C Intel NHM ? ? ?
|
|
Packit |
5c3484 |
C Intel SBR 2.57 2.93 3.0
|
|
Packit |
5c3484 |
C Intel IBR 2.35 2.66 3.0
|
|
Packit |
5c3484 |
C Intel HWL 2.02 2.5 2.5
|
|
Packit |
5c3484 |
C Intel BWL ? ? ?
|
|
Packit |
5c3484 |
C Intel atom ? ? ?
|
|
Packit |
5c3484 |
C VIA nano ? ? ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C The inner loops of this code are the result of running a code generation and
|
|
Packit |
5c3484 |
C optimisation tool suite written by David Harvey and Torbjörn Granlund, except
|
|
Packit |
5c3484 |
C that the sqr_diag_addlsh1 loop was manually written.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Replace current unoptimised sqr_diag_addlsh1 loop, 2.5 c/l should be easy.
|
|
Packit |
5c3484 |
C * Streamline pointer updates.
|
|
Packit |
5c3484 |
C * Perhaps suppress a few more xor insns in feed-in code.
|
|
Packit |
5c3484 |
C * Make sure we write no dead registers in feed-in code.
|
|
Packit |
5c3484 |
C * We might use 32-bit size ops, since n >= 2^32 is non-terminating. Watch
|
|
Packit |
5c3484 |
C out for negative sizes being zero-extended, though.
|
|
Packit |
5c3484 |
C * The straight-line code for n <= 3 comes from the K8 code, and might be
|
|
Packit |
5c3484 |
C quite sub-optimal here. Write specific code, and add code for n = 4.
|
|
Packit |
5c3484 |
C * The mul_2 loop has a 10 insn common sequence in the loop start and the
|
|
Packit |
5c3484 |
C wind-down code. Try re-rolling it.
|
|
Packit |
5c3484 |
C * This file has been the subject to just basic micro-optimisation.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C When playing with pointers, set this to $2 to fall back to conservative
|
|
Packit |
5c3484 |
C indexing in wind-down code.
|
|
Packit |
5c3484 |
define(`I',`$1')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`rp', `%rdi')
|
|
Packit |
5c3484 |
define(`up', `%rsi')
|
|
Packit |
5c3484 |
define(`un_param',`%rdx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_sqr_basecase)
|
|
Packit |
5c3484 |
FUNC_ENTRY(3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
cmp $2, un_param
|
|
Packit |
5c3484 |
jae L(gt1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov %rax, (rp)
|
|
Packit |
5c3484 |
mov %rdx, 8(rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(gt1): jne L(gt2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax
|
|
Packit |
5c3484 |
mov %rax, %r8
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov 8(up), %r11
|
|
Packit |
5c3484 |
mov %rax, (rp)
|
|
Packit |
5c3484 |
mov %r11, %rax
|
|
Packit |
5c3484 |
mov %rdx, %r9
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mov %r11, %rax
|
|
Packit |
5c3484 |
mov %rdx, %r11
|
|
Packit |
5c3484 |
mul %r8
|
|
Packit |
5c3484 |
xor %r8, %r8
|
|
Packit |
5c3484 |
add %rax, %r9
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
adc %r8, %r11
|
|
Packit |
5c3484 |
add %rax, %r9
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
adc %r8, %r11
|
|
Packit |
5c3484 |
mov %r11, 24(rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(gt2): cmp $4, un_param
|
|
Packit |
5c3484 |
jae L(gt3)
|
|
Packit |
5c3484 |
define(`v0', `%r8')
|
|
Packit |
5c3484 |
define(`v1', `%r9')
|
|
Packit |
5c3484 |
define(`w0', `%r10')
|
|
Packit |
5c3484 |
define(`w2', `%r11')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov 8(up), %r11
|
|
Packit |
5c3484 |
mov %rax, (rp)
|
|
Packit |
5c3484 |
mov %r11, %rax
|
|
Packit |
5c3484 |
mov %rdx, 8(rp)
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov 16(up), %rcx
|
|
Packit |
5c3484 |
mov %rax, 16(rp)
|
|
Packit |
5c3484 |
mov %rcx, %rax
|
|
Packit |
5c3484 |
mov %rdx, 24(rp)
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov %rax, 32(rp)
|
|
Packit |
5c3484 |
mov %rdx, 40(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov %r11, %rax
|
|
Packit |
5c3484 |
mul %r10
|
|
Packit |
5c3484 |
mov %rax, %r8
|
|
Packit |
5c3484 |
mov %rcx, %rax
|
|
Packit |
5c3484 |
mov %rdx, %r9
|
|
Packit |
5c3484 |
mul %r10
|
|
Packit |
5c3484 |
xor %r10, %r10
|
|
Packit |
5c3484 |
add %rax, %r9
|
|
Packit |
5c3484 |
mov %r11, %rax
|
|
Packit |
5c3484 |
mov %r10, %r11
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mul %rcx
|
|
Packit |
5c3484 |
add %rax, %r10
|
|
Packit |
5c3484 |
adc %r11, %rdx
|
|
Packit |
5c3484 |
add %r8, %r8
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
adc %rdx, %rdx
|
|
Packit |
5c3484 |
adc %r11, %r11
|
|
Packit |
5c3484 |
add %r8, 8(rp)
|
|
Packit |
5c3484 |
adc %r9, 16(rp)
|
|
Packit |
5c3484 |
adc %r10, 24(rp)
|
|
Packit |
5c3484 |
adc %rdx, 32(rp)
|
|
Packit |
5c3484 |
adc %r11, 40(rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(gt3):
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`v0', `%r8')
|
|
Packit |
5c3484 |
define(`v1', `%r9')
|
|
Packit |
5c3484 |
define(`w0', `%r10')
|
|
Packit |
5c3484 |
define(`w1', `%r11')
|
|
Packit |
5c3484 |
define(`w2', `%rbx')
|
|
Packit |
5c3484 |
define(`w3', `%rbp')
|
|
Packit |
5c3484 |
define(`un', `%r12')
|
|
Packit |
5c3484 |
define(`n', `%rcx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`X0', `%r13')
|
|
Packit |
5c3484 |
define(`X1', `%r14')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(do_mul_2):
|
|
Packit |
5c3484 |
mov (up), v0
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
lea (rp,un_param,8), rp C point rp at R[un]
|
|
Packit |
5c3484 |
mov 8(up), %rax
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
lea (up,un_param,8), up C point up right after U's end
|
|
Packit |
5c3484 |
mov %rax, v1
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
mov $1, R32(un) C free up rdx
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
sub un_param, un
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push un
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
mov %rax, (rp,un,8)
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
test $1, R8(un)
|
|
Packit |
5c3484 |
jnz L(m2b1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2b0):lea 2(un), n
|
|
Packit |
5c3484 |
xor R32(w1), R32(w1) C FIXME
|
|
Packit |
5c3484 |
xor R32(w2), R32(w2) C FIXME
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
jmp L(m2l0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2b1):lea 1(un), n
|
|
Packit |
5c3484 |
xor R32(w3), R32(w3) C FIXME
|
|
Packit |
5c3484 |
xor R32(w0), R32(w0) C FIXME
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
jmp L(m2l1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(m2tp):
|
|
Packit |
5c3484 |
L(m2l0):mul v0
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w1, w0
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
mov w0, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
L(m2l1):mul v0
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
add w3, w2
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov w2, (rp,n,8)
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
jnc L(m2tp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(m2ed):mul v0
|
|
Packit |
5c3484 |
add %rax, w0
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mov I(-8(up),-8(up,n,8)), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w1, w0
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add %rax, w2
|
|
Packit |
5c3484 |
mov w0, I(-8(rp),-8(rp,n,8))
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
add w3, w2
|
|
Packit |
5c3484 |
mov w2, I((rp),(rp,n,8))
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, I(8(rp),8(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add $2, un C decrease |un|
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(do_addmul_2):
|
|
Packit |
5c3484 |
L(outer):
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
cmp $-2, R32(un) C jump if un C {-1,0} FIXME jump if un C {-2,1}
|
|
Packit |
5c3484 |
jge L(corner) C FIXME: move to before the lea above
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -8(up,un,8), v0
|
|
Packit |
5c3484 |
mov (up,un,8), %rax
|
|
Packit |
5c3484 |
mov %rax, v1
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
test $1, R8(un)
|
|
Packit |
5c3484 |
jnz L(a1x1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a1x0):mov (rp,un,8), X0
|
|
Packit |
5c3484 |
xor w0, w0
|
|
Packit |
5c3484 |
mov 8(rp,un,8), X1
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
xor w2, w2
|
|
Packit |
5c3484 |
mov X0, (rp,un,8)
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jnz L(a110)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a100):lea 2(un), n C un = 4, 8, 12, ...
|
|
Packit |
5c3484 |
jmp L(lo0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a110):lea (un), n C un = 2, 6, 10, ...
|
|
Packit |
5c3484 |
jmp L(lo2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a1x1):mov (rp,un,8), X1
|
|
Packit |
5c3484 |
xor w2, w2
|
|
Packit |
5c3484 |
mov 8(rp,un,8), X0
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
xor w0, w0
|
|
Packit |
5c3484 |
mov 8(up,un,8), %rax
|
|
Packit |
5c3484 |
test $2, R8(un)
|
|
Packit |
5c3484 |
jz L(a111)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a101):lea 3(un), n C un = 1, 5, 9, ...
|
|
Packit |
5c3484 |
jmp L(lo1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(a111):lea 1(un), n C un = 3, 7, 11, ...
|
|
Packit |
5c3484 |
jmp L(lo3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(top): mul v1
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add w2, X0
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
mov -16(up,n,8), %rax
|
|
Packit |
5c3484 |
L(lo1): mul v0
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
mov -16(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov X1, -24(rp,n,8)
|
|
Packit |
5c3484 |
mov -8(rp,n,8), X1
|
|
Packit |
5c3484 |
add w3, X0
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
mov X0, -16(rp,n,8)
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
add w0, X1
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
L(lo0): mul v0
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
mov (rp,n,8), X0
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mov %rdx, w0
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
L(lo3): mul v0
|
|
Packit |
5c3484 |
add w2, X0
|
|
Packit |
5c3484 |
mov X1, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w0
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
add w3, X0
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
mov 8(rp,n,8), X1
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
mov %rdx, w2
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov X0, (rp,n,8)
|
|
Packit |
5c3484 |
L(lo2): mul v0
|
|
Packit |
5c3484 |
add w0, X1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
mov 16(rp,n,8), X0
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
jnc L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): mul v1
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add w2, %rax
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov X1, I(-8(rp),-24(rp,n,8))
|
|
Packit |
5c3484 |
add w3, %rax
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rax, I((rp),-16(rp,n,8))
|
|
Packit |
5c3484 |
mov %rdx, I(8(rp),-8(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
add $2, un C decrease |un|
|
|
Packit |
5c3484 |
jmp L(outer) C loop until a small corner remains
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(corner):
|
|
Packit |
5c3484 |
pop n
|
|
Packit |
5c3484 |
jg L(small_corner)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
mov -24(up), v0
|
|
Packit |
5c3484 |
mov -16(up), %rax
|
|
Packit |
5c3484 |
mov %rax, v1
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
mov -24(rp), X0
|
|
Packit |
5c3484 |
mov -16(rp), X1
|
|
Packit |
5c3484 |
add %rax, X0
|
|
Packit |
5c3484 |
mov %rdx, w1
|
|
Packit |
5c3484 |
adc $0, w1
|
|
Packit |
5c3484 |
xor w2, w2
|
|
Packit |
5c3484 |
mov X0, -24(rp)
|
|
Packit |
5c3484 |
mov -8(up), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
add $0, X1
|
|
Packit |
5c3484 |
mov %rdx, w3
|
|
Packit |
5c3484 |
adc $0, w2
|
|
Packit |
5c3484 |
add %rax, X1
|
|
Packit |
5c3484 |
mov -8(up), %rax
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
mul v1
|
|
Packit |
5c3484 |
add w1, X1
|
|
Packit |
5c3484 |
adc $0, w3
|
|
Packit |
5c3484 |
add w2, %rax
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov X1, -16(rp)
|
|
Packit |
5c3484 |
jmp L(com)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(small_corner):
|
|
Packit |
5c3484 |
mov -8(rp), w3
|
|
Packit |
5c3484 |
mov -16(up), v0
|
|
Packit |
5c3484 |
mov -8(up), %rax
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
L(com): add w3, %rax
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rax, -8(rp)
|
|
Packit |
5c3484 |
mov %rdx, (rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(sqr_diag_addlsh1):
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
shl n
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
mov %rax, (rp,n,8)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
xor R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov 8(rp,n,8), %r8
|
|
Packit |
5c3484 |
mov 16(rp,n,8), %r9
|
|
Packit |
5c3484 |
jmp L(dm)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(dtop):add %r8, %r10
|
|
Packit |
5c3484 |
adc %r9, %rax
|
|
Packit |
5c3484 |
mov 8(rp,n,8), %r8
|
|
Packit |
5c3484 |
mov 16(rp,n,8), %r9
|
|
Packit |
5c3484 |
mov %r10, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %rax, (rp,n,8)
|
|
Packit |
5c3484 |
L(dm): adc %r8, %r8
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov (up,n,4), %rax
|
|
Packit |
5c3484 |
lea (%rdx,%rbx), %r10
|
|
Packit |
5c3484 |
setc R8(%rbx)
|
|
Packit |
5c3484 |
mul %rax
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
js L(dtop)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(dend):add %r8, %r10
|
|
Packit |
5c3484 |
adc %r9, %rax
|
|
Packit |
5c3484 |
mov %r10, I(-8(rp),-8(rp,n,8))
|
|
Packit |
5c3484 |
mov %rax, I((rp),(rp,n,8))
|
|
Packit |
5c3484 |
adc %rbx, %rdx
|
|
Packit |
5c3484 |
mov %rdx, I(8(rp),8(rp,n,8))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|