Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/izamax.S

kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
/* Copyright 2009, 2010 The University of Texas at Austin.           */
kusano 2b45e8
/* All rights reserved.                                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* Redistribution and use in source and binary forms, with or        */
kusano 2b45e8
/* without modification, are permitted provided that the following   */
kusano 2b45e8
/* conditions are met:                                               */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   1. Redistributions of source code must retain the above         */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer.                                                  */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   2. Redistributions in binary form must reproduce the above      */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer in the documentation and/or other materials       */
kusano 2b45e8
/*      provided with the distribution.                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
kusano 2b45e8
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
kusano 2b45e8
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
kusano 2b45e8
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
kusano 2b45e8
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
kusano 2b45e8
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
kusano 2b45e8
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
kusano 2b45e8
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
kusano 2b45e8
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
kusano 2b45e8
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
kusano 2b45e8
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
kusano 2b45e8
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
kusano 2b45e8
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
kusano 2b45e8
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* The views and conclusions contained in the software and           */
kusano 2b45e8
/* documentation are those of the authors and should not be          */
kusano 2b45e8
/* interpreted as representing official policies, either expressed   */
kusano 2b45e8
/* or implied, of The University of Texas at Austin.                 */
kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
kusano 2b45e8
#define ASSEMBLER
kusano 2b45e8
#include "common.h"
kusano 2b45e8
	
kusano 2b45e8
#define RET	r3
kusano 2b45e8
#define X	r4
kusano 2b45e8
#define INCX	r5	
kusano 2b45e8
kusano 2b45e8
#define N	r6
kusano 2b45e8
#define NN	r7
kusano 2b45e8
#define XX	r8
kusano 2b45e8
#define PREA	r9
kusano 2b45e8
#define	INCXM1	r10
kusano 2b45e8
kusano 2b45e8
#define FZERO	f1
kusano 2b45e8
kusano 2b45e8
#define STACKSIZE 160
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, -STACKSIZE
kusano 2b45e8
	li	r0,   0
kusano 2b45e8
kusano 2b45e8
	stfd	f14,    0(SP)
kusano 2b45e8
	stfd	f15,    8(SP)
kusano 2b45e8
	stfd	f16,   16(SP)
kusano 2b45e8
	stfd	f17,   24(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f18,   32(SP)
kusano 2b45e8
	stfd	f19,   40(SP)
kusano 2b45e8
	stfd	f20,   48(SP)
kusano 2b45e8
	stfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f22,   64(SP)
kusano 2b45e8
	stfd	f23,   72(SP)
kusano 2b45e8
	stfd	f24,   80(SP)
kusano 2b45e8
	stfd	f25,   88(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f26,   96(SP)
kusano 2b45e8
	stfd	f27,  104(SP)
kusano 2b45e8
	stfd	f28,  112(SP)
kusano 2b45e8
	stfd	f29,  120(SP)
kusano 2b45e8
kusano 2b45e8
	stfd	f30,  128(SP)
kusano 2b45e8
	stfd	f31,  136(SP)
kusano 2b45e8
kusano 2b45e8
	stw	r0,   144(SP)
kusano 2b45e8
	lfs	FZERO,144(SP)
kusano 2b45e8
kusano 2b45e8
#ifdef F_INTERFACE
kusano 2b45e8
	LDINT	N,    0(r3)
kusano 2b45e8
	LDINT	INCX, 0(INCX)
kusano 2b45e8
#else
kusano 2b45e8
	mr	N, r3
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	li	RET, 0
kusano 2b45e8
	mr	NN, N
kusano 2b45e8
	mr	XX, X
kusano 2b45e8
kusano 2b45e8
	slwi	INCX, INCX, ZBASE_SHIFT
kusano 2b45e8
	subi	INCXM1, INCX, SIZE
kusano 2b45e8
kusano 2b45e8
	li	PREA, L1_PREFETCHSIZE
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	LL(9999)
kusano 2b45e8
	cmpwi	cr0, INCX, 0
kusano 2b45e8
	ble-	LL(9999)
kusano 2b45e8
kusano 2b45e8
	LFD	f1, 0 * SIZE(X)
kusano 2b45e8
	LFD	f2, 1 * SIZE(X)
kusano 2b45e8
	add	X, X, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f1, f1
kusano 2b45e8
	fabs	f2, f2
kusano 2b45e8
	fadd	f1, f1, f2
kusano 2b45e8
kusano 2b45e8
	fmr	f0, f1
kusano 2b45e8
	fmr	f2, f1
kusano 2b45e8
	fmr	f3, f1
kusano 2b45e8
kusano 2b45e8
	subi	N, N, 1
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCX, 2 * SIZE
kusano 2b45e8
	bne-	cr0, LL(100)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq-	cr0, LL(50)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
	LFD	f24,   0 * SIZE(X)
kusano 2b45e8
	LFD	f25,   1 * SIZE(X)
kusano 2b45e8
	LFD	f26,   2 * SIZE(X)
kusano 2b45e8
	LFD	f27,   3 * SIZE(X)
kusano 2b45e8
	LFD	f28,   4 * SIZE(X)
kusano 2b45e8
	LFD	f29,   5 * SIZE(X)
kusano 2b45e8
	LFD	f30,   6 * SIZE(X)
kusano 2b45e8
	LFD	f31,   7 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f24,   8 * SIZE(X)
kusano 2b45e8
	LFD	f25,   9 * SIZE(X)
kusano 2b45e8
	LFD	f26,  10 * SIZE(X)
kusano 2b45e8
	LFD	f27,  11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  12 * SIZE(X)
kusano 2b45e8
	LFD	f29,  13 * SIZE(X)
kusano 2b45e8
	LFD	f30,  14 * SIZE(X)
kusano 2b45e8
	LFD	f31,  15 * SIZE(X)
kusano 2b45e8
	bdz	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	fadd	f4,  f8,  f9
kusano 2b45e8
	fadd	f5,  f10, f11
kusano 2b45e8
	fadd	f6,  f12, f13
kusano 2b45e8
	fadd	f7,  f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  16 * SIZE(X)
kusano 2b45e8
	LFD	f25,  17 * SIZE(X)
kusano 2b45e8
	LFD	f26,  18 * SIZE(X)
kusano 2b45e8
	LFD	f27,  19 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  20 * SIZE(X)
kusano 2b45e8
	LFD	f29,  21 * SIZE(X)
kusano 2b45e8
	LFD	f30,  22 * SIZE(X)
kusano 2b45e8
	LFD	f31,  23 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f4
kusano 2b45e8
	fsub	f17, f1,  f5
kusano 2b45e8
	fsub	f18, f2,  f6
kusano 2b45e8
	fsub	f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fadd	f20, f8,  f9
kusano 2b45e8
	fadd	f21, f10, f11
kusano 2b45e8
	fadd	f22, f12, f13
kusano 2b45e8
	fadd	f23, f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  24 * SIZE(X)
kusano 2b45e8
	LFD	f25,  25 * SIZE(X)
kusano 2b45e8
	LFD	f26,  26 * SIZE(X)
kusano 2b45e8
	LFD	f27,  27 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f4
kusano 2b45e8
	fsel	f1,  f17, f1,  f5
kusano 2b45e8
	fsel	f2,  f18, f2,  f6
kusano 2b45e8
	fsel	f3,  f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  28 * SIZE(X)
kusano 2b45e8
	LFD	f29,  29 * SIZE(X)
kusano 2b45e8
	LFD	f30,  30 * SIZE(X)
kusano 2b45e8
	LFD	f31,  31 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f20
kusano 2b45e8
	fsub	f17, f1,  f21
kusano 2b45e8
	fsub	f18, f2,  f22
kusano 2b45e8
	fsub	f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f20
kusano 2b45e8
	fsel	f1,  f17, f1,  f21
kusano 2b45e8
	fsel	f2,  f18, f2,  f22
kusano 2b45e8
	fsel	f3,  f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
#ifndef POWER6
kusano 2b45e8
	L1_PREFETCH	X, PREA
kusano 2b45e8
#endif
kusano 2b45e8
	addi	X, X, 16 * SIZE
kusano 2b45e8
#ifdef POWER6
kusano 2b45e8
	L1_PREFETCH	X, PREA
kusano 2b45e8
#endif
kusano 2b45e8
	bdnz	LL(10)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	fadd	f4,  f8,  f9
kusano 2b45e8
	fadd	f5,  f10, f11
kusano 2b45e8
	fadd	f6,  f12, f13
kusano 2b45e8
	fadd	f7,  f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f4
kusano 2b45e8
	fsub	f17, f1,  f5
kusano 2b45e8
	fsub	f18, f2,  f6
kusano 2b45e8
	fsub	f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fadd	f20, f8,  f9
kusano 2b45e8
	fadd	f21, f10, f11
kusano 2b45e8
	fadd	f22, f12, f13
kusano 2b45e8
	fadd	f23, f14, f15
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f4
kusano 2b45e8
	fsel	f1,  f17, f1,  f5
kusano 2b45e8
	fsel	f2,  f18, f2,  f6
kusano 2b45e8
	fsel	f3,  f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f20
kusano 2b45e8
	fsub	f17, f1,  f21
kusano 2b45e8
	fsub	f18, f2,  f22
kusano 2b45e8
	fsub	f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f20
kusano 2b45e8
	fsel	f1,  f17, f1,  f21
kusano 2b45e8
	fsel	f2,  f18, f2,  f22
kusano 2b45e8
	fsel	f3,  f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
	addi	X, X, 16 * SIZE
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(50):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(60):
kusano 2b45e8
	LFD	f8,  0 * SIZE(X)
kusano 2b45e8
	LFD	f9,  1 * SIZE(X)
kusano 2b45e8
	addi	X, X,  2 * SIZE
kusano 2b45e8
kusano 2b45e8
	fabs	f8, f8
kusano 2b45e8
	fabs	f9, f9
kusano 2b45e8
	fadd	f8, f8, f9
kusano 2b45e8
	fsub	f16, f1, f8
kusano 2b45e8
	fsel	f1, f16, f1, f8
kusano 2b45e8
	bdnz	LL(60)
kusano 2b45e8
	b	LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(100):
kusano 2b45e8
	sub	X, X, INCXM1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(150)
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   X, INCXM1
kusano 2b45e8
	LFDUX	f25,   X, INCX
kusano 2b45e8
	LFDX	f26,   X, INCXM1
kusano 2b45e8
	LFDUX	f27,   X, INCX
kusano 2b45e8
	LFDX	f28,   X, INCXM1
kusano 2b45e8
	LFDUX	f29,   X, INCX
kusano 2b45e8
	LFDX	f30,   X, INCXM1
kusano 2b45e8
	LFDUX	f31,   X, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   X, INCXM1
kusano 2b45e8
	LFDUX	f25,   X, INCX
kusano 2b45e8
	LFDX	f26,   X, INCXM1
kusano 2b45e8
	LFDUX	f27,   X, INCX
kusano 2b45e8
	LFDX	f28,   X, INCXM1
kusano 2b45e8
	LFDUX	f29,   X, INCX
kusano 2b45e8
	LFDX	f30,   X, INCXM1
kusano 2b45e8
	LFDUX	f31,   X, INCX
kusano 2b45e8
kusano 2b45e8
	bdz	LL(120)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(110):
kusano 2b45e8
	fadd	f4,  f8,  f9
kusano 2b45e8
	fadd	f5,  f10, f11
kusano 2b45e8
	fadd	f6,  f12, f13
kusano 2b45e8
	fadd	f7,  f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   X, INCXM1
kusano 2b45e8
	LFDUX	f25,   X, INCX
kusano 2b45e8
	LFDX	f26,   X, INCXM1
kusano 2b45e8
	LFDUX	f27,   X, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f28,   X, INCXM1
kusano 2b45e8
	LFDUX	f29,   X, INCX
kusano 2b45e8
	LFDX	f30,   X, INCXM1
kusano 2b45e8
	LFDUX	f31,   X, INCX
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f4
kusano 2b45e8
	fsub	f17, f1,  f5
kusano 2b45e8
	fsub	f18, f2,  f6
kusano 2b45e8
	fsub	f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fadd	f20, f8,  f9
kusano 2b45e8
	fadd	f21, f10, f11
kusano 2b45e8
	fadd	f22, f12, f13
kusano 2b45e8
	fadd	f23, f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   X, INCXM1
kusano 2b45e8
	LFDUX	f25,   X, INCX
kusano 2b45e8
	LFDX	f26,   X, INCXM1
kusano 2b45e8
	LFDUX	f27,   X, INCX
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f4
kusano 2b45e8
	fsel	f1,  f17, f1,  f5
kusano 2b45e8
	fsel	f2,  f18, f2,  f6
kusano 2b45e8
	fsel	f3,  f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f28,   X, INCXM1
kusano 2b45e8
	LFDUX	f29,   X, INCX
kusano 2b45e8
	LFDX	f30,   X, INCXM1
kusano 2b45e8
	LFDUX	f31,   X, INCX
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f20
kusano 2b45e8
	fsub	f17, f1,  f21
kusano 2b45e8
	fsub	f18, f2,  f22
kusano 2b45e8
	fsub	f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f20
kusano 2b45e8
	fsel	f1,  f17, f1,  f21
kusano 2b45e8
	fsel	f2,  f18, f2,  f22
kusano 2b45e8
	fsel	f3,  f19, f3,  f23
kusano 2b45e8
	bdnz	LL(110)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(120):
kusano 2b45e8
	fadd	f4,  f8,  f9
kusano 2b45e8
	fadd	f5,  f10, f11
kusano 2b45e8
	fadd	f6,  f12, f13
kusano 2b45e8
	fadd	f7,  f14, f15
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f4
kusano 2b45e8
	fsub	f17, f1,  f5
kusano 2b45e8
	fsub	f18, f2,  f6
kusano 2b45e8
	fsub	f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fadd	f20, f8,  f9
kusano 2b45e8
	fadd	f21, f10, f11
kusano 2b45e8
	fadd	f22, f12, f13
kusano 2b45e8
	fadd	f23, f14, f15
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f4
kusano 2b45e8
	fsel	f1,  f17, f1,  f5
kusano 2b45e8
	fsel	f2,  f18, f2,  f6
kusano 2b45e8
	fsel	f3,  f19, f3,  f7
kusano 2b45e8
kusano 2b45e8
	fsub	f16, f0,  f20
kusano 2b45e8
	fsub	f17, f1,  f21
kusano 2b45e8
	fsub	f18, f2,  f22
kusano 2b45e8
	fsub	f19, f3,  f23
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f16, f0,  f20
kusano 2b45e8
	fsel	f1,  f17, f1,  f21
kusano 2b45e8
	fsel	f2,  f18, f2,  f22
kusano 2b45e8
	fsel	f3,  f19, f3,  f23
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(150):
kusano 2b45e8
	andi.	r0,  N, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq	LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(160):
kusano 2b45e8
	LFDX	f8,    X, INCXM1
kusano 2b45e8
	LFDUX	f9,    X, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f8, f8
kusano 2b45e8
	fabs	f9, f9
kusano 2b45e8
	fadd	f8, f8, f9
kusano 2b45e8
	fsub	f16, f1, f8
kusano 2b45e8
	fsel	f1, f16, f1, f8
kusano 2b45e8
	bdnz	LL(160)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	fsub	f8,  f0,  f1
kusano 2b45e8
	fsub	f9,  f2,  f3
kusano 2b45e8
kusano 2b45e8
	fsel	f0,  f8,  f0,  f1
kusano 2b45e8
	fsel	f2,  f9,  f2,  f3
kusano 2b45e8
	fsub	f8,  f0,  f2
kusano 2b45e8
	fsel	f1,  f8,  f0,  f2
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
LL(1000):
kusano 2b45e8
	cmpwi	cr0, INCX, SIZE * 2
kusano 2b45e8
	bne-	cr0, LL(1100)
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, NN, 3
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq-	cr0, LL(1050)
kusano 2b45e8
kusano 2b45e8
	LFD	f24,   0 * SIZE(XX)
kusano 2b45e8
	LFD	f25,   1 * SIZE(XX)
kusano 2b45e8
	LFD	f26,   2 * SIZE(XX)
kusano 2b45e8
	LFD	f27,   3 * SIZE(XX)
kusano 2b45e8
	LFD	f28,   4 * SIZE(XX)
kusano 2b45e8
	LFD	f29,   5 * SIZE(XX)
kusano 2b45e8
	LFD	f30,   6 * SIZE(XX)
kusano 2b45e8
	LFD	f31,   7 * SIZE(XX)
kusano 2b45e8
	bdz	LL(1020)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1010):
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFD	f24,   8 * SIZE(XX)
kusano 2b45e8
	LFD	f25,   9 * SIZE(XX)
kusano 2b45e8
	LFD	f26,  10 * SIZE(XX)
kusano 2b45e8
	LFD	f27,  11 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  12 * SIZE(XX)
kusano 2b45e8
	LFD	f29,  13 * SIZE(XX)
kusano 2b45e8
	LFD	f30,  14 * SIZE(XX)
kusano 2b45e8
	LFD	f31,  15 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFD	f24,  16 * SIZE(XX)
kusano 2b45e8
	LFD	f25,  17 * SIZE(XX)
kusano 2b45e8
	LFD	f26,  18 * SIZE(XX)
kusano 2b45e8
	LFD	f27,  19 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  20 * SIZE(XX)
kusano 2b45e8
	LFD	f29,  21 * SIZE(XX)
kusano 2b45e8
	LFD	f30,  22 * SIZE(XX)
kusano 2b45e8
	LFD	f31,  23 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 16 * SIZE
kusano 2b45e8
	bdnz	LL(1010)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1020):
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFD	f24,   8 * SIZE(XX)
kusano 2b45e8
	LFD	f25,   9 * SIZE(XX)
kusano 2b45e8
	LFD	f26,  10 * SIZE(XX)
kusano 2b45e8
	LFD	f27,  11 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFD	f28,  12 * SIZE(XX)
kusano 2b45e8
	LFD	f29,  13 * SIZE(XX)
kusano 2b45e8
	LFD	f30,  14 * SIZE(XX)
kusano 2b45e8
	LFD	f31,  15 * SIZE(XX)
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	XX, XX, 16 * SIZE
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1050):
kusano 2b45e8
	andi.	r0,  NN, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq	LL(9999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1060):
kusano 2b45e8
	LFD	f8,  0 * SIZE(XX)
kusano 2b45e8
	LFD	f9,  1 * SIZE(XX)
kusano 2b45e8
	addi	XX, XX,  2 * SIZE
kusano 2b45e8
kusano 2b45e8
	fabs	f8, f8
kusano 2b45e8
	fabs	f9, f9
kusano 2b45e8
	fadd	f8, f8, f9
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f8
kusano 2b45e8
	beq	cr0, LL(9999)	
kusano 2b45e8
	bdnz	LL(1060)
kusano 2b45e8
	b	LL(9999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1100):
kusano 2b45e8
	sub	XX, XX, INCXM1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, NN, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(1150)
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   XX, INCXM1
kusano 2b45e8
	LFDUX	f25,   XX, INCX
kusano 2b45e8
	LFDX	f26,   XX, INCXM1
kusano 2b45e8
	LFDUX	f27,   XX, INCX
kusano 2b45e8
	LFDX	f28,   XX, INCXM1
kusano 2b45e8
	LFDUX	f29,   XX, INCX
kusano 2b45e8
	LFDX	f30,   XX, INCXM1
kusano 2b45e8
	LFDUX	f31,   XX, INCX
kusano 2b45e8
	bdz	LL(1120)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1110):
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   XX, INCXM1
kusano 2b45e8
	LFDUX	f25,   XX, INCX
kusano 2b45e8
	LFDX	f26,   XX, INCXM1
kusano 2b45e8
	LFDUX	f27,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f28,   XX, INCXM1
kusano 2b45e8
	LFDUX	f29,   XX, INCX
kusano 2b45e8
	LFDX	f30,   XX, INCXM1
kusano 2b45e8
	LFDUX	f31,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   XX, INCXM1
kusano 2b45e8
	LFDUX	f25,   XX, INCX
kusano 2b45e8
	LFDX	f26,   XX, INCXM1
kusano 2b45e8
	LFDUX	f27,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f28,   XX, INCXM1
kusano 2b45e8
	LFDUX	f29,   XX, INCX
kusano 2b45e8
	LFDX	f30,   XX, INCXM1
kusano 2b45e8
	LFDUX	f31,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(1110)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1120):
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	LFDX	f24,   XX, INCXM1
kusano 2b45e8
	LFDUX	f25,   XX, INCX
kusano 2b45e8
	LFDX	f26,   XX, INCXM1
kusano 2b45e8
	LFDUX	f27,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	LFDX	f28,   XX, INCXM1
kusano 2b45e8
	LFDUX	f29,   XX, INCX
kusano 2b45e8
	LFDX	f30,   XX, INCXM1
kusano 2b45e8
	LFDUX	f31,   XX, INCX
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	fabs	f8,  f24
kusano 2b45e8
	fabs	f9,  f25
kusano 2b45e8
	fabs	f10, f26
kusano 2b45e8
	fabs	f11, f27
kusano 2b45e8
kusano 2b45e8
	fabs	f12, f28
kusano 2b45e8
	fabs	f13, f29
kusano 2b45e8
	fabs	f14, f30
kusano 2b45e8
	fabs	f15, f31
kusano 2b45e8
kusano 2b45e8
	fadd	f4, f8,  f9
kusano 2b45e8
	fadd	f5, f10, f11
kusano 2b45e8
	fadd	f6, f12, f13
kusano 2b45e8
	fadd	f7, f14, f15
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f4
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f5
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f6
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f7
kusano 2b45e8
	beq	cr0, LL(9999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1150):
kusano 2b45e8
	andi.	r0,  NN, 7
kusano 2b45e8
	mtspr	CTR, r0
kusano 2b45e8
	beq	LL(9999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(1160):
kusano 2b45e8
	LFDX	f8,    XX, INCXM1
kusano 2b45e8
	LFDUX	f9,    XX, INCX
kusano 2b45e8
kusano 2b45e8
	fabs	f8, f8
kusano 2b45e8
	fabs	f9, f9
kusano 2b45e8
	fadd	f8, f8, f9
kusano 2b45e8
kusano 2b45e8
	addi	RET, RET, 1
kusano 2b45e8
	fcmpu	cr0, f1, f8
kusano 2b45e8
	beq	cr0, LL(9999)	
kusano 2b45e8
	bdnz	LL(1160)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(9999):
kusano 2b45e8
	lfd	f14,    0(SP)
kusano 2b45e8
	lfd	f15,    8(SP)
kusano 2b45e8
	lfd	f16,   16(SP)
kusano 2b45e8
	lfd	f17,   24(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f18,   32(SP)
kusano 2b45e8
	lfd	f19,   40(SP)
kusano 2b45e8
	lfd	f20,   48(SP)
kusano 2b45e8
	lfd	f21,   56(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f22,   64(SP)
kusano 2b45e8
	lfd	f23,   72(SP)
kusano 2b45e8
	lfd	f24,   80(SP)
kusano 2b45e8
	lfd	f25,   88(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f26,   96(SP)
kusano 2b45e8
	lfd	f27,  104(SP)
kusano 2b45e8
	lfd	f28,  112(SP)
kusano 2b45e8
	lfd	f29,  120(SP)
kusano 2b45e8
kusano 2b45e8
	lfd	f30,  128(SP)
kusano 2b45e8
	lfd	f31,  136(SP)
kusano 2b45e8
kusano 2b45e8
	addi	SP, SP, STACKSIZE
kusano 2b45e8
	blr
kusano 2b45e8
kusano 2b45e8
	EPILOGUE