Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/power/zdot_cell.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
#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)
kusano 2b45e8
#define RESULT	r3
kusano 2b45e8
#define N	r4
kusano 2b45e8
#define X	r5
kusano 2b45e8
#define INCX	r6	
kusano 2b45e8
#define Y	r7
kusano 2b45e8
#define INCY	r8
kusano 2b45e8
#define PREA	r9
kusano 2b45e8
#else
kusano 2b45e8
#define N	r3
kusano 2b45e8
#define X	r4
kusano 2b45e8
#define INCX	r5	
kusano 2b45e8
#define Y	r6
kusano 2b45e8
#define INCY	r7
kusano 2b45e8
#define PREA	r8
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#define INCXM1	r10
kusano 2b45e8
#define INCYM1	r11
kusano 2b45e8
kusano 2b45e8
#define FZERO	f0
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(N)
kusano 2b45e8
	LDINT	INCX, 0(INCX)
kusano 2b45e8
	LDINT	INCY, 0(INCY)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	slwi	INCX, INCX, ZBASE_SHIFT
kusano 2b45e8
	slwi	INCY, INCY, ZBASE_SHIFT
kusano 2b45e8
kusano 2b45e8
	subi	INCXM1, INCX, SIZE
kusano 2b45e8
	subi	INCYM1, INCY, SIZE
kusano 2b45e8
kusano 2b45e8
	fmr	f1,  FZERO
kusano 2b45e8
	fmr	f2,  FZERO
kusano 2b45e8
	fmr	f3,  FZERO
kusano 2b45e8
	fmr	f4,  FZERO
kusano 2b45e8
	fmr	f5,  FZERO
kusano 2b45e8
	fmr	f6,  FZERO
kusano 2b45e8
	fmr	f7,  FZERO
kusano 2b45e8
kusano 2b45e8
	fmr	f24,  FZERO
kusano 2b45e8
	fmr	f25,  FZERO
kusano 2b45e8
	fmr	f26,  FZERO
kusano 2b45e8
	fmr	f27,  FZERO
kusano 2b45e8
	fmr	f28,  FZERO
kusano 2b45e8
	fmr	f29,  FZERO
kusano 2b45e8
	fmr	f30,  FZERO
kusano 2b45e8
	fmr	f31,  FZERO
kusano 2b45e8
kusano 2b45e8
	li	PREA, 16 * 10 * SIZE
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, N, 0
kusano 2b45e8
	ble-	LL(999)
kusano 2b45e8
kusano 2b45e8
	cmpwi	cr0, INCX, 2 * SIZE
kusano 2b45e8
	bne-	cr0, LL(100)
kusano 2b45e8
	cmpwi	cr0, INCY, 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	f8,     0 * SIZE(X)
kusano 2b45e8
	LFD	f9,     1 * SIZE(X)
kusano 2b45e8
	LFD	f10,    2 * SIZE(X)
kusano 2b45e8
	LFD	f11,    3 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	LFD	f16,    0 * SIZE(Y)
kusano 2b45e8
	LFD	f17,    1 * SIZE(Y)
kusano 2b45e8
	LFD	f18,    2 * SIZE(Y)
kusano 2b45e8
	LFD	f19,    3 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	LFD	f12,    4 * SIZE(X)
kusano 2b45e8
	LFD	f13,    5 * SIZE(X)
kusano 2b45e8
	LFD	f14,    6 * SIZE(X)
kusano 2b45e8
	LFD	f15,    7 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	LFD	f20,    4 * SIZE(Y)
kusano 2b45e8
	LFD	f21,    5 * SIZE(Y)
kusano 2b45e8
	LFD	f23,    7 * SIZE(Y)
kusano 2b45e8
	bdz	LL(20)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(10):
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFD	f22,   6 * SIZE(Y)
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFD	f8,    8 * SIZE(X)
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFD	f17,   9 * SIZE(Y)
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFD	f9,    9 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFD	f16,   8 * SIZE(Y)
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFD	f10,  10 * SIZE(X)
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFD	f19,  11 * SIZE(Y)
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFD	f11,  11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFD	f18,  10 * SIZE(Y)
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFD	f12,  12 * SIZE(X)
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFD	f21,  13 * SIZE(Y)
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFD	f13,  13 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFD	f20,  12 * SIZE(Y)
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFD	f14,  14 * SIZE(X)
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFD	f23,  15 * SIZE(Y)
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFD	f15,  15 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFD	f22,  14 * SIZE(Y)
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFD	f8,   16 * SIZE(X)
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFD	f17,  17 * SIZE(Y)
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFD	f9,   17 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFD	f16,  16 * SIZE(Y)
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFD	f10,  18 * SIZE(X)
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFD	f19,  19 * SIZE(Y)
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFD	f11,  19 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFD	f18,  18 * SIZE(Y)
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFD	f12,   20 * SIZE(X)
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFD	f21,   21 * SIZE(Y)
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFD	f13,   21 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFD	f20,   20 * SIZE(Y)
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFD	f14,   22 * SIZE(X)
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFD	f23,   23 * SIZE(Y)
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFD	f15,   23 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	dcbt	X, PREA
kusano 2b45e8
	addi	X, X, 16 * SIZE
kusano 2b45e8
	dcbt	Y, PREA
kusano 2b45e8
	addi	Y, Y, 16 * SIZE
kusano 2b45e8
	bdnz	LL(10)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(20):
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFD	f22,    6 * SIZE(Y)
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFD	f8,    8 * SIZE(X)
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFD	f17,   9 * SIZE(Y)
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFD	f9,    9 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFD	f16,   8 * SIZE(Y)
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFD	f10,  10 * SIZE(X)
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFD	f19,  11 * SIZE(Y)
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFD	f11,  11 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFD	f18,  10 * SIZE(Y)
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFD	f12,  12 * SIZE(X)
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFD	f21,  13 * SIZE(Y)
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFD	f13,  13 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFD	f20,  12 * SIZE(Y)
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFD	f14,  14 * SIZE(X)
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFD	f23,  15 * SIZE(Y)
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFD	f15,  15 * SIZE(X)
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFD	f22,  14 * SIZE(Y)
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	addi	X, X, 16 * SIZE
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	addi	Y, Y, 16 * SIZE
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	nop
kusano 2b45e8
	
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
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
	LFD	f16, 0 * SIZE(Y)
kusano 2b45e8
	LFD	f17, 1 * SIZE(Y)
kusano 2b45e8
kusano 2b45e8
	addi	X, X,  2 * SIZE
kusano 2b45e8
	addi	Y, Y,  2 * SIZE
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(60)
kusano 2b45e8
	b	LL(999)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(100):
kusano 2b45e8
#ifdef F_INTERFACE
kusano 2b45e8
	cmpwi	cr0, INCX, 0
kusano 2b45e8
	bge+	LL(102)
kusano 2b45e8
kusano 2b45e8
	subi	r0, N, 1
kusano 2b45e8
	mullw	r0, r0, INCX
kusano 2b45e8
	sub	X, X, r0
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(102):
kusano 2b45e8
	cmpwi	cr0, INCY, 0
kusano 2b45e8
	bge+	LL(104)
kusano 2b45e8
kusano 2b45e8
	subi	r0, N, 1
kusano 2b45e8
	mullw	r0, r0, INCY
kusano 2b45e8
	sub	Y, Y, r0
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(104):
kusano 2b45e8
#endif
kusano 2b45e8
	sub	X, X, INCXM1
kusano 2b45e8
	sub	Y, Y, INCYM1
kusano 2b45e8
kusano 2b45e8
	srawi.	r0, N, 3
kusano 2b45e8
	mtspr	CTR,  r0
kusano 2b45e8
	beq-	LL(150)
kusano 2b45e8
kusano 2b45e8
	LFDX	f8,   X, INCXM1
kusano 2b45e8
	LFDX	f16,  Y, INCYM1
kusano 2b45e8
	LFDUX	f9,   X, INCX
kusano 2b45e8
	LFDUX	f17,  Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFDX	f10,  X, INCXM1
kusano 2b45e8
	LFDX	f18,  Y, INCYM1
kusano 2b45e8
	LFDUX	f11,  X, INCX
kusano 2b45e8
	LFDUX	f19,  Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFDX	f12,   X, INCXM1
kusano 2b45e8
	LFDX	f20,   Y, INCYM1
kusano 2b45e8
	LFDUX	f13,   X, INCX
kusano 2b45e8
	LFDUX	f21,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	LFDX	f14,   X, INCXM1
kusano 2b45e8
	LFDUX	f15,   X, INCX
kusano 2b45e8
	bdz	LL(120)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(110):
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFDX	f22,   Y, INCYM1
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFDX	f8,   X, INCXM1
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFDUX	f23,   Y, INCY
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFDUX	f9,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFDX	f16,  Y, INCYM1
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFDX	f10,  X, INCXM1
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFDUX	f17,  Y, INCY
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFDUX	f11,  X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFDX	f18,  Y, INCYM1
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFDX	f12,   X, INCXM1
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFDUX	f19,  Y, INCY
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFDUX	f13,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFDX	f20,   Y, INCYM1
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFDX	f14,   X, INCXM1
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFDUX	f21,   Y, INCY
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFDUX	f15,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFDX	f22,   Y, INCYM1
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFDX	f8,    X, INCXM1
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFDUX	f23,   Y, INCY
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFDUX	f9,    X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFDX	f16,  Y, INCYM1
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFDX	f10,  X, INCXM1
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFDUX	f17,  Y, INCY
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFDUX	f11,  X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFDX	f18,  Y, INCYM1
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFDX	f12,   X, INCXM1
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFDUX	f19,  Y, INCY
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFDUX	f13,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFDX	f20,   Y, INCYM1
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFDX	f14,   X, INCXM1
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFDUX	f21,   Y, INCY
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFDUX	f15,   X, INCX
kusano 2b45e8
kusano 2b45e8
	bdnz	LL(110)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(120):
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFDX	f22,   Y, INCYM1
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFDX	f8,   X, INCXM1
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	LFDUX	f23,   Y, INCY
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	LFDUX	f9,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	LFDX	f16,  Y, INCYM1
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	LFDX	f10,  X, INCXM1
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	LFDUX	f17,  Y, INCY
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
	LFDUX	f11,  X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	LFDX	f18,  Y, INCYM1
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	LFDX	f12,   X, INCXM1
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	LFDUX	f19,  Y, INCY
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
	LFDUX	f13,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	LFDX	f20,   Y, INCYM1
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	LFDX	f14,   X, INCXM1
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	LFDUX	f21,   Y, INCY
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
kusano 2b45e8
	LFDUX	f15,   X, INCX
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	LFDX	f22,   Y, INCYM1
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	LFDUX	f23,   Y, INCY
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
kusano 2b45e8
	FMADD	f4, f10, f18,  f4
kusano 2b45e8
	FMADD	f7, f10, f19,  f7
kusano 2b45e8
	FMADD	f5, f11, f19,  f5
kusano 2b45e8
	FMADD	f6, f11, f18,  f6
kusano 2b45e8
kusano 2b45e8
	FMADD	f24, f12, f20,  f24
kusano 2b45e8
	FMADD	f27, f12, f21,  f27
kusano 2b45e8
	FMADD	f25, f13, f21,  f25
kusano 2b45e8
	FMADD	f26, f13, f20,  f26
kusano 2b45e8
kusano 2b45e8
	FMADD	f28, f14, f22,  f28
kusano 2b45e8
	FMADD	f31, f14, f23,  f31
kusano 2b45e8
	FMADD	f29, f15, f23,  f29
kusano 2b45e8
	FMADD	f30, f15, f22,  f30
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
	LFDX	f16,   Y, INCYM1
kusano 2b45e8
	LFDUX	f17,   Y, INCY
kusano 2b45e8
kusano 2b45e8
	FMADD	f0, f8,  f16,  f0
kusano 2b45e8
	FMADD	f3, f8,  f17,  f3
kusano 2b45e8
	FMADD	f1, f9,  f17,  f1
kusano 2b45e8
	FMADD	f2, f9,  f16,  f2
kusano 2b45e8
	bdnz	LL(160)
kusano 2b45e8
	.align 4
kusano 2b45e8
kusano 2b45e8
LL(999):
kusano 2b45e8
	FADD	f0, f0, f4
kusano 2b45e8
	FADD	f1, f1, f5
kusano 2b45e8
	FADD	f2, f2, f6
kusano 2b45e8
	FADD	f3, f3, f7
kusano 2b45e8
kusano 2b45e8
	FADD	f24, f28, f24
kusano 2b45e8
	FADD	f25, f29, f25
kusano 2b45e8
	FADD	f26, f30, f26
kusano 2b45e8
	FADD	f27, f31, f27
kusano 2b45e8
kusano 2b45e8
	FADD	f0, f0, f24
kusano 2b45e8
	FADD	f1, f1, f25
kusano 2b45e8
	FADD	f2, f2, f26
kusano 2b45e8
	FADD	f3, f3, f27
kusano 2b45e8
kusano 2b45e8
#ifndef CONJ
kusano 2b45e8
	FSUB	f1,  f0,  f1
kusano 2b45e8
	FADD	f2,  f2,  f3
kusano 2b45e8
#else
kusano 2b45e8
	FADD	f1,  f0,  f1
kusano 2b45e8
	FSUB	f2,  f3,  f2
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)
kusano 2b45e8
	STFD	f1,    0 * SIZE(RESULT)
kusano 2b45e8
	STFD	f2,    1 * SIZE(RESULT)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
#if defined(F_INTERFACE) && defined(F_INTERFACE_GFORT)
kusano 2b45e8
#ifndef __64BIT__
kusano 2b45e8
#ifndef DOUBLE
kusano 2b45e8
	stfs	f1,   144(SP)
kusano 2b45e8
	stfs	f2,   148(SP)
kusano 2b45e8
	lwz	r3,   144(SP)
kusano 2b45e8
	lwz	r4,   148(SP)
kusano 2b45e8
#else
kusano 2b45e8
	stfd	f1,   144(SP)
kusano 2b45e8
	stfd	f2,   152(SP)
kusano 2b45e8
	lwz	r3,   144(SP)
kusano 2b45e8
	lwz	r4,   148(SP)
kusano 2b45e8
	lwz	r5,   152(SP)
kusano 2b45e8
	lwz	r6,   156(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#else
kusano 2b45e8
#ifndef DOUBLE
kusano 2b45e8
	stfs	f1,   144(SP)
kusano 2b45e8
	stfs	f2,   148(SP)
kusano 2b45e8
	ld	r3,   144(SP)
kusano 2b45e8
#else
kusano 2b45e8
	stfd	f1,   144(SP)
kusano 2b45e8
	stfd	f2,   152(SP)
kusano 2b45e8
	ld	r3,   144(SP)
kusano 2b45e8
	ld	r4,   152(SP)
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
#endif
kusano 2b45e8
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