|
Packit |
5c3484 |
dnl IA-64 mpn_cnd_add_n/mpn_cnd_sub_n.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjörn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 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
|
|
Packit |
5c3484 |
C Itanium: ?
|
|
Packit |
5c3484 |
C Itanium 2: 1.5
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`cnd', `r32')
|
|
Packit |
5c3484 |
define(`rp', `r33')
|
|
Packit |
5c3484 |
define(`up', `r34')
|
|
Packit |
5c3484 |
define(`vp', `r35')
|
|
Packit |
5c3484 |
define(`n', `r36')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_cnd_add_n',`
|
|
Packit |
5c3484 |
define(ADDSUB, add)
|
|
Packit |
5c3484 |
define(CND, ltu)
|
|
Packit |
5c3484 |
define(INCR, 1)
|
|
Packit |
5c3484 |
define(LIM, -1)
|
|
Packit |
5c3484 |
define(func, mpn_cnd_add_n)
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_cnd_sub_n',`
|
|
Packit |
5c3484 |
define(ADDSUB, sub)
|
|
Packit |
5c3484 |
define(CND, gtu)
|
|
Packit |
5c3484 |
define(INCR, -1)
|
|
Packit |
5c3484 |
define(LIM, 0)
|
|
Packit |
5c3484 |
define(func, mpn_cnd_sub_n)
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(PFDIST, 160)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C Some useful aliases for registers we use
|
|
Packit |
5c3484 |
define(`u0',`r14') define(`u1',`r15') define(`u2',`r16') define(`u3',`r17')
|
|
Packit |
5c3484 |
define(`x0',`r20') define(`x1',`r21') define(`x2',`r22') define(`x3',`r23')
|
|
Packit |
5c3484 |
define(`v0',`r24') define(`v1',`r25') define(`v2',`r26') define(`v3',`r27')
|
|
Packit |
5c3484 |
define(`w0',`r28') define(`w1',`r29') define(`w2',`r30') define(`w3',`r31')
|
|
Packit |
5c3484 |
define(`up1',`up') define(`up2',`r8') define(`upadv',`r1')
|
|
Packit |
5c3484 |
define(`vp1',`vp') define(`vp2',`r9') define(`vpadv',`r11')
|
|
Packit |
5c3484 |
define(`rp1',`rp') define(`rp2',`r10')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_cnd_add_n mpn_cnd_sub_n)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
PROLOGUE(func)
|
|
Packit |
5c3484 |
.prologue
|
|
Packit |
5c3484 |
.save ar.lc, r2
|
|
Packit |
5c3484 |
.body
|
|
Packit |
5c3484 |
ifdef(`HAVE_ABI_32',`
|
|
Packit |
5c3484 |
addp4 rp = 0, rp C M I
|
|
Packit |
5c3484 |
addp4 up = 0, up C M I
|
|
Packit |
5c3484 |
nop.i 0
|
|
Packit |
5c3484 |
addp4 vp = 0, vp C M I
|
|
Packit |
5c3484 |
nop.m 0
|
|
Packit |
5c3484 |
zxt4 n = n C I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
{.mmi; and r3 = 3, n C M I
|
|
Packit |
5c3484 |
add n = -1, n C M I
|
|
Packit |
5c3484 |
mov r2 = ar.lc C I0
|
|
Packit |
5c3484 |
}{.mmi; cmp.ne p6, p7 = 0, cnd C M I
|
|
Packit |
5c3484 |
add vp2 = 8, vp C M I
|
|
Packit |
5c3484 |
add up2 = 8, up C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; add upadv = PFDIST, up C M I
|
|
Packit |
5c3484 |
add vpadv = PFDIST, vp C M I
|
|
Packit |
5c3484 |
shr.u n = n, 2 C I0
|
|
Packit |
5c3484 |
.pred.rel "mutex", p6, p7
|
|
Packit |
5c3484 |
}{.mmi; add rp2 = 8, rp C M I
|
|
Packit |
5c3484 |
(p6) mov cnd = -1 C M I
|
|
Packit |
5c3484 |
(p7) mov cnd = 0 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
} cmp.eq p9, p0 = 1, r3 C M I
|
|
Packit |
5c3484 |
cmp.eq p7, p0 = 2, r3 C M I
|
|
Packit |
5c3484 |
cmp.eq p8, p0 = 3, r3 C M I
|
|
Packit |
5c3484 |
(p9) br L(b1) C B
|
|
Packit |
5c3484 |
(p7) br L(b2) C B
|
|
Packit |
5c3484 |
(p8) br L(b3) C B
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
L(b0):
|
|
Packit |
5c3484 |
{.mmi; ld8 v2 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v3 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
mov ar.lc = n C I0
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
} ld8 u2 = [up1], 16 C M01
|
|
Packit |
5c3484 |
ld8 u3 = [up2], 16 C M01
|
|
Packit |
5c3484 |
and x2 = v2, cnd C M I
|
|
Packit |
5c3484 |
and x3 = v3, cnd C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ADDSUB w2 = u2, x2 C M I
|
|
Packit |
5c3484 |
ADDSUB w3 = u3, x3 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ld8 v0 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v1 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
cmp.CND p8, p0 = w2, u2 C M I
|
|
Packit |
5c3484 |
cmp.CND p9, p0 = w3, u3 C M I
|
|
Packit |
5c3484 |
br L(lo0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): ld8 v1 = [vp1], 8 C M01
|
|
Packit |
5c3484 |
add vp2 = 8, vp2 C M I
|
|
Packit |
5c3484 |
add rp2 = 8, rp2 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ld8 u1 = [up1], 8 C M01
|
|
Packit |
5c3484 |
add up2 = 8, up2 C M I
|
|
Packit |
5c3484 |
and x1 = v1, cnd C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ADDSUB w1 = u1, x1 C M I
|
|
Packit |
5c3484 |
cmp.ne p10, p0 = 0, n
|
|
Packit |
5c3484 |
add n = -1, n
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
cmp.CND p7, p0 = w1, u1 C M I
|
|
Packit |
5c3484 |
st8 [rp1] = w1, 8 C M23
|
|
Packit |
5c3484 |
(p10) br L(b0)
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
mov r8 = 0 C M I
|
|
Packit |
5c3484 |
br L(e1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): ld8 v3 = [vp1], 8 C M01
|
|
Packit |
5c3484 |
add vp2 = 8, vp2 C M I
|
|
Packit |
5c3484 |
add rp2 = 8, rp2 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ld8 u3 = [up1], 8 C M01
|
|
Packit |
5c3484 |
add up2 = 8, up2 C M I
|
|
Packit |
5c3484 |
and x3 = v3, cnd C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ADDSUB w3 = u3, x3 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
cmp.CND p9, p0 = w3, u3 C M I
|
|
Packit |
5c3484 |
st8 [rp1] = w3, 8 C M23
|
|
Packit |
5c3484 |
C fall through
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b2):
|
|
Packit |
5c3484 |
{.mmi; ld8 v0 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v1 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
mov ar.lc = n C I0
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
} ld8 u0 = [up1], 16 C M01
|
|
Packit |
5c3484 |
ld8 u1 = [up2], 16 C M01
|
|
Packit |
5c3484 |
and x0 = v0, cnd C M I
|
|
Packit |
5c3484 |
and x1 = v1, cnd C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
ADDSUB w0 = u0, x0 C M I
|
|
Packit |
5c3484 |
ADDSUB w1 = u1, x1 C M I
|
|
Packit |
5c3484 |
br.cloop.dptk L(gt2) C B
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
cmp.CND p6, p0 = w0, u0 C M I
|
|
Packit |
5c3484 |
br L(e2) C B
|
|
Packit |
5c3484 |
L(gt2):
|
|
Packit |
5c3484 |
ld8 v2 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v3 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
cmp.CND p6, p0 = w0, u0 C M I
|
|
Packit |
5c3484 |
cmp.CND p7, p0 = w1, u1 C M I
|
|
Packit |
5c3484 |
br L(lo2) C B
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C *** MAIN LOOP START ***
|
|
Packit |
5c3484 |
C ALIGN(32)
|
|
Packit |
5c3484 |
L(top):
|
|
Packit |
5c3484 |
{.mmi; ld8 v2 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v3 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
cmp.CND p6, p0 = w0, u0 C M I
|
|
Packit |
5c3484 |
}{.mmi; st8 [rp1] = w2, 16 C M23
|
|
Packit |
5c3484 |
st8 [rp2] = w3, 16 C M23
|
|
Packit |
5c3484 |
cmp.CND p7, p0 = w1, u1 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
L(lo2):
|
|
Packit |
5c3484 |
{.mmi; ld8 u2 = [up1], 16 C M01
|
|
Packit |
5c3484 |
ld8 u3 = [up2], 16 C M01
|
|
Packit |
5c3484 |
(p9) cmpeqor p6, p0 = LIM, w0 C M I
|
|
Packit |
5c3484 |
}{.mmi; and x2 = v2, cnd C M I
|
|
Packit |
5c3484 |
and x3 = v3, cnd C M I
|
|
Packit |
5c3484 |
(p9) add w0 = INCR, w0 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; ADDSUB w2 = u2, x2 C M I
|
|
Packit |
5c3484 |
(p6) cmpeqor p7, p0 = LIM, w1 C M I
|
|
Packit |
5c3484 |
(p6) add w1 = INCR, w1 C M I
|
|
Packit |
5c3484 |
}{.mmi; ADDSUB w3 = u3, x3 C M I
|
|
Packit |
5c3484 |
lfetch [upadv], 32
|
|
Packit |
5c3484 |
nop 0
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; ld8 v0 = [vp1], 16 C M01
|
|
Packit |
5c3484 |
ld8 v1 = [vp2], 16 C M01
|
|
Packit |
5c3484 |
cmp.CND p8, p0 = w2, u2 C M I
|
|
Packit |
5c3484 |
}{.mmi; st8 [rp1] = w0, 16 C M23
|
|
Packit |
5c3484 |
st8 [rp2] = w1, 16 C M23
|
|
Packit |
5c3484 |
cmp.CND p9, p0 = w3, u3 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
L(lo0):
|
|
Packit |
5c3484 |
{.mmi; ld8 u0 = [up1], 16 C M01
|
|
Packit |
5c3484 |
ld8 u1 = [up2], 16 C M01
|
|
Packit |
5c3484 |
(p7) cmpeqor p8, p0 = LIM, w2 C M I
|
|
Packit |
5c3484 |
}{.mmi; and x0 = v0, cnd C M I
|
|
Packit |
5c3484 |
and x1 = v1, cnd C M I
|
|
Packit |
5c3484 |
(p7) add w2 = INCR, w2 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; ADDSUB w0 = u0, x0 C M I
|
|
Packit |
5c3484 |
(p8) cmpeqor p9, p0 = LIM, w3 C M I
|
|
Packit |
5c3484 |
(p8) add w3 = INCR, w3 C M I
|
|
Packit |
5c3484 |
}{.mmb; ADDSUB w1 = u1, x1 C M I
|
|
Packit |
5c3484 |
lfetch [vpadv], 32
|
|
Packit |
5c3484 |
br.cloop.dptk L(top) C B
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
C *** MAIN LOOP END ***
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end):
|
|
Packit |
5c3484 |
{.mmi; st8 [rp1] = w2, 16 C M23
|
|
Packit |
5c3484 |
st8 [rp2] = w3, 16 C M23
|
|
Packit |
5c3484 |
cmp.CND p6, p0 = w0, u0 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
L(e2):
|
|
Packit |
5c3484 |
{.mmi; cmp.CND p7, p0 = w1, u1 C M I
|
|
Packit |
5c3484 |
(p9) cmpeqor p6, p0 = LIM, w0 C M I
|
|
Packit |
5c3484 |
(p9) add w0 = INCR, w0 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; mov r8 = 0 C M I
|
|
Packit |
5c3484 |
(p6) cmpeqor p7, p0 = LIM, w1 C M I
|
|
Packit |
5c3484 |
(p6) add w1 = INCR, w1 C M I
|
|
Packit |
5c3484 |
;;
|
|
Packit |
5c3484 |
}{.mmi; st8 [rp1] = w0, 16 C M23
|
|
Packit |
5c3484 |
st8 [rp2] = w1, 16 C M23
|
|
Packit |
5c3484 |
mov ar.lc = r2 C I0
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
L(e1):
|
|
Packit |
5c3484 |
{.mmb; nop 0
|
|
Packit |
5c3484 |
(p7) mov r8 = 1 C M I
|
|
Packit |
5c3484 |
br.ret.sptk.many b0 C B
|
|
Packit |
5c3484 |
}
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
ASM_END()
|