Blame mpn/ia64/popcount.asm

Packit 5c3484
dnl  IA-64 mpn_popcount -- mpn population count.
Packit 5c3484
Packit 5c3484
dnl  Contributed to the GNU project by Torbjorn Granlund.
Packit 5c3484
Packit 5c3484
dnl  Copyright 2000-2005 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:       1.5
Packit 5c3484
C Itanium 2:     1
Packit 5c3484
Packit 5c3484
C INPUT PARAMETERS
Packit 5c3484
define(`up', `r32')
Packit 5c3484
define(`n', `r33')
Packit 5c3484
Packit 5c3484
define(`u0',`r16') define(`u1',`r17') define(`u2',`r18') define(`u3',`r19')
Packit 5c3484
define(`c0',`r28') define(`c1',`r29') define(`c2',`r30') define(`c3',`r31')
Packit 5c3484
define(`s',`r8')
Packit 5c3484
Packit 5c3484
Packit 5c3484
ASM_START()
Packit 5c3484
PROLOGUE(mpn_popcount)
Packit 5c3484
	.prologue
Packit 5c3484
ifdef(`HAVE_ABI_32',
Packit 5c3484
`	addp4		up = 0, up		C			M I
Packit 5c3484
	nop.m		0
Packit 5c3484
	zxt4		n = n			C			I
Packit 5c3484
	;;
Packit 5c3484
')
Packit 5c3484
Packit 5c3484
 {.mmi;	add		r9 = 512, up		C prefetch pointer	M I
Packit 5c3484
	ld8		r10 = [up], 8		C load first limb	M01
Packit 5c3484
	mov.i		r2 = ar.lc		C save ar.lc		I0
Packit 5c3484
}{.mmi;	and		r14 = 3, n		C			M I
Packit 5c3484
	cmp.lt		p15, p14 = 4, n		C small count?		M I
Packit 5c3484
	add		n = -5, n		C			M I
Packit 5c3484
	;;
Packit 5c3484
}{.mmi;	cmp.eq		p6, p0 = 1, r14		C			M I
Packit 5c3484
	cmp.eq		p7, p0 = 2, r14		C			M I
Packit 5c3484
	cmp.eq		p8, p0 = 3, r14		C			M I
Packit 5c3484
}{.bbb
Packit 5c3484
  (p6)	br.dptk		.Lb01			C			B
Packit 5c3484
  (p7)	br.dptk		.Lb10			C			B
Packit 5c3484
  (p8)	br.dptk		.Lb11			C			B
Packit 5c3484
}
Packit 5c3484
Packit 5c3484
Packit 5c3484
.Lb00:	ld8		u1 = [up], 8		C			M01
Packit 5c3484
	shr.u		n = n, 2		C			I0
Packit 5c3484
	mov		s = 0			C			M I
Packit 5c3484
	;;
Packit 5c3484
	ld8		u2 = [up], 8		C			M01
Packit 5c3484
	popcnt		c0 = r10		C			I0
Packit 5c3484
	mov.i		ar.lc = n		C			I0
Packit 5c3484
	;;
Packit 5c3484
	ld8		u3 = [up], 8		C			M01
Packit 5c3484
	popcnt		c1 = u1			C			I0
Packit 5c3484
  (p15)	br.cond.dptk	.grt4			C			B
Packit 5c3484
	;;
Packit 5c3484
	nop.m	0				C			-
Packit 5c3484
	nop.m	0				C			-
Packit 5c3484
	popcnt		c2 = u2			C			I0
Packit 5c3484
	;;
Packit 5c3484
	mov		s = c0			C			M I
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	br		.Lcj4			C			B
Packit 5c3484
Packit 5c3484
.grt4:	ld8		u0 = [up], 8		C			M01
Packit 5c3484
	popcnt		c2 = u2			C			I0
Packit 5c3484
	br		.LL00			C			B
Packit 5c3484
Packit 5c3484
Packit 5c3484
.Lb01:
Packit 5c3484
	popcnt		s = r10			C			I0
Packit 5c3484
  (p14)	br.ret.sptk.many b0			C			B
Packit 5c3484
Packit 5c3484
.grt1:	ld8		u0 = [up], 8		C			M01
Packit 5c3484
	shr.u		n = n, 2		C			I0
Packit 5c3484
	;;
Packit 5c3484
	ld8		u1 = [up], 8		C			M01
Packit 5c3484
	mov.i		ar.lc = n		C			I0
Packit 5c3484
	;;
Packit 5c3484
	ld8		u2 = [up], 8		C			M01
Packit 5c3484
	popcnt		c0 = u0			C			I0
Packit 5c3484
	mov		c3 = 0			C			I0
Packit 5c3484
Packit 5c3484
	;;
Packit 5c3484
	ld8		u3 = [up], 8		C			M01
Packit 5c3484
	popcnt		c1 = u1			C			I0
Packit 5c3484
	br.cloop.dptk	.Loop			C			B
Packit 5c3484
	br		.Lend			C			B
Packit 5c3484
Packit 5c3484
Packit 5c3484
.Lb10:	ld8		u3 = [up], 8		C			M01
Packit 5c3484
	shr.u		n = n, 2		C			I0
Packit 5c3484
  (p15)	br.cond.dptk	.grt2			C			B
Packit 5c3484
Packit 5c3484
	popcnt		s = r10			C			I0
Packit 5c3484
	;;
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	br		.Lcj2			C			B
Packit 5c3484
Packit 5c3484
.grt2:	ld8		u0 = [up], 8		C			M01
Packit 5c3484
	mov.i		ar.lc = n		C			I0
Packit 5c3484
	popcnt		c2 = r10		C			I0
Packit 5c3484
	;;
Packit 5c3484
	ld8		u1 = [up], 8		C			M01
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	mov		s = 0			C			M I
Packit 5c3484
	;;
Packit 5c3484
	ld8		u2 = [up], 8		C			M01
Packit 5c3484
	popcnt		c0 = u0			C			I0
Packit 5c3484
	br		.LL10			C			B
Packit 5c3484
Packit 5c3484
Packit 5c3484
.Lb11:	ld8		u2 = [up], 8		C			M01
Packit 5c3484
	shr.u		n = n, 2		C			I0
Packit 5c3484
	mov		s = 0			C			M I
Packit 5c3484
	;;
Packit 5c3484
	ld8		u3 = [up], 8		C			M01
Packit 5c3484
	popcnt		s = r10			C			I0
Packit 5c3484
  (p15)	br.cond.dptk	.grt3			C			B
Packit 5c3484
Packit 5c3484
	popcnt		c2 = u2			C			I0
Packit 5c3484
	;;
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	br		.Lcj3			C			B
Packit 5c3484
Packit 5c3484
.grt3:	ld8		u0 = [up], 8		C			M01
Packit 5c3484
	popcnt		c2 = u2			C			I0
Packit 5c3484
	mov.i		ar.lc = n		C			I0
Packit 5c3484
	mov		c1 = 0
Packit 5c3484
	;;
Packit 5c3484
	ld8		u1 = [up], 8		C			M01
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	br		.LL11			C			B
Packit 5c3484
Packit 5c3484
Packit 5c3484
.Loop:	ld8		u0 = [up], 8		C			M01
Packit 5c3484
	popcnt		c2 = u2			C			I0
Packit 5c3484
	add		s = s, c3		C			M I
Packit 5c3484
	;;
Packit 5c3484
.LL00:	ld8		u1 = [up], 8		C			M01
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	add		s = s, c0		C			M I
Packit 5c3484
	;;
Packit 5c3484
.LL11:	ld8		u2 = [up], 8		C			M01
Packit 5c3484
	popcnt		c0 = u0			C			I0
Packit 5c3484
	add		s = s, c1		C			M I
Packit 5c3484
	;;
Packit 5c3484
.LL10:	ld8		u3 = [up], 8		C			M01
Packit 5c3484
	popcnt		c1 = u1			C			I0
Packit 5c3484
	add		s = s, c2		C			M I
Packit 5c3484
	lfetch		[r9], 32		C			M01
Packit 5c3484
	nop.m		0			C			-
Packit 5c3484
	br.cloop.dptk	.Loop			C			B
Packit 5c3484
	;;
Packit 5c3484
Packit 5c3484
.Lend:	popcnt		c2 = u2			C			I0
Packit 5c3484
	add		s = s, c3		C			M I
Packit 5c3484
	;;
Packit 5c3484
	popcnt		c3 = u3			C			I0
Packit 5c3484
	add		s = s, c0		C			M I
Packit 5c3484
	;;
Packit 5c3484
.Lcj4:	add		s = s, c1		C			M I
Packit 5c3484
	;;
Packit 5c3484
.Lcj3:	add		s = s, c2		C			M I
Packit 5c3484
	;;
Packit 5c3484
.Lcj2:	add		s = s, c3		C			M I
Packit 5c3484
	mov.i		ar.lc = r2		C			I0
Packit 5c3484
	br.ret.sptk.many b0			C			B
Packit 5c3484
EPILOGUE()
Packit 5c3484
ASM_END()