|
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 SP r12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M r32
|
|
kusano |
2b45e8 |
#define A r34
|
|
kusano |
2b45e8 |
#define LDA r35
|
|
kusano |
2b45e8 |
#define X r36
|
|
kusano |
2b45e8 |
#define INCX r37
|
|
kusano |
2b45e8 |
#define Y r38
|
|
kusano |
2b45e8 |
#define INCY r39
|
|
kusano |
2b45e8 |
#define BUFFER r33
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define I r14
|
|
kusano |
2b45e8 |
#define IS r15
|
|
kusano |
2b45e8 |
#define A1 r16
|
|
kusano |
2b45e8 |
#define A2 r17
|
|
kusano |
2b45e8 |
#define A3 r18
|
|
kusano |
2b45e8 |
#define A4 r19
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define NEW_X r20
|
|
kusano |
2b45e8 |
#define NEW_Y r21
|
|
kusano |
2b45e8 |
#define XX r22
|
|
kusano |
2b45e8 |
#define YY r23
|
|
kusano |
2b45e8 |
#define TEMP r24
|
|
kusano |
2b45e8 |
#define YYS r25
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define PREA1 loc0
|
|
kusano |
2b45e8 |
#define PREA2 loc1
|
|
kusano |
2b45e8 |
#define PREA3 loc2
|
|
kusano |
2b45e8 |
#define PREA4 loc3
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define A11 loc4
|
|
kusano |
2b45e8 |
#define A21 loc5
|
|
kusano |
2b45e8 |
#define A31 loc6
|
|
kusano |
2b45e8 |
#define A41 loc7
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define PREX r8
|
|
kusano |
2b45e8 |
#define PREY r9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ARLC r29
|
|
kusano |
2b45e8 |
#define PR r30
|
|
kusano |
2b45e8 |
#define ARPFS r31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef DOUBLE
|
|
kusano |
2b45e8 |
#define RPREFETCH (16 * 3 + 4)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define RPREFETCH (16 * 3 + 16)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#define PREFETCH lfetch.nt1
|
|
kusano |
2b45e8 |
#define PREFETCHW lfetch.excl.nt1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define alpha f8
|
|
kusano |
2b45e8 |
#define atemp1 f6
|
|
kusano |
2b45e8 |
#define atemp2 f7
|
|
kusano |
2b45e8 |
#define atemp3 f10
|
|
kusano |
2b45e8 |
#define atemp4 f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define xsum1 f12
|
|
kusano |
2b45e8 |
#define xsum2 f13
|
|
kusano |
2b45e8 |
#define xsum3 f14
|
|
kusano |
2b45e8 |
#define xsum4 f15
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
.prologue
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
.save ar.pfs, ARPFS
|
|
kusano |
2b45e8 |
alloc ARPFS = ar.pfs, 8, 16, 8, 0
|
|
kusano |
2b45e8 |
mov ARLC = ar.lc
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov PR = pr
|
|
kusano |
2b45e8 |
adds r14 = 16, SP
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds r8 = -8 * 16, SP
|
|
kusano |
2b45e8 |
adds r9 = -7 * 16, SP
|
|
kusano |
2b45e8 |
adds SP = -8 * 16, SP
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
stf.spill [r8] = f16, 32
|
|
kusano |
2b45e8 |
stf.spill [r9] = f17, 32
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
stf.spill [r8] = f18, 32
|
|
kusano |
2b45e8 |
stf.spill [r9] = f19, 32
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
stf.spill [r8] = f20, 32
|
|
kusano |
2b45e8 |
stf.spill [r9] = f21, 32
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
stf.spill [r8] = f22
|
|
kusano |
2b45e8 |
stf.spill [r9] = f23
|
|
kusano |
2b45e8 |
.body
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ld8 BUFFER = [r14]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd LDA = LDA, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
shladd INCX = INCX, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
shladd INCY = INCY, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.ge p7, p0 = 0, M
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p7) br.cond.dpnt .L999
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov NEW_X = X
|
|
kusano |
2b45e8 |
cmp.eq p10, p0 = SIZE, INCX
|
|
kusano |
2b45e8 |
(p10) br.cond.dptk .L10
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.L10:
|
|
kusano |
2b45e8 |
mov NEW_Y = Y
|
|
kusano |
2b45e8 |
cmp.eq p10, p0 = SIZE, INCY
|
|
kusano |
2b45e8 |
(p10) br.cond.dptk .L20
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L20:
|
|
kusano |
2b45e8 |
mov IS = 0
|
|
kusano |
2b45e8 |
cmp.gt p10, p0 = 4, M
|
|
kusano |
2b45e8 |
(p10) br.cond.dpnt .L30
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.L21:
|
|
kusano |
2b45e8 |
mov A1 = A
|
|
kusano |
2b45e8 |
add A2 = LDA, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd A3 = LDA, 1, A
|
|
kusano |
2b45e8 |
shladd A4 = LDA, 1, A2
|
|
kusano |
2b45e8 |
shladd A = LDA, 2, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds PREX = RPREFETCH * SIZE, NEW_X
|
|
kusano |
2b45e8 |
adds PREY = RPREFETCH * SIZE, NEW_Y
|
|
kusano |
2b45e8 |
adds PREA1 = RPREFETCH * SIZE, A1
|
|
kusano |
2b45e8 |
adds PREA2 = RPREFETCH * SIZE, A2
|
|
kusano |
2b45e8 |
adds PREA3 = RPREFETCH * SIZE, A3
|
|
kusano |
2b45e8 |
adds PREA4 = RPREFETCH * SIZE, A4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd TEMP = IS, BASE_SHIFT, NEW_X
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD atemp1 = [TEMP], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD atemp2 = [TEMP], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD atemp3 = [TEMP], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD atemp4 = [TEMP], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMPY atemp1 = alpha, atemp1
|
|
kusano |
2b45e8 |
FMPY atemp2 = alpha, atemp2
|
|
kusano |
2b45e8 |
FMPY atemp3 = alpha, atemp3
|
|
kusano |
2b45e8 |
FMPY atemp4 = alpha, atemp4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov xsum1 = f0
|
|
kusano |
2b45e8 |
mov xsum2 = f0
|
|
kusano |
2b45e8 |
mov xsum3 = f0
|
|
kusano |
2b45e8 |
mov xsum4 = f0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov XX = NEW_X
|
|
kusano |
2b45e8 |
mov YY = NEW_Y
|
|
kusano |
2b45e8 |
mov YYS = NEW_Y
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shr I = IS, 2
|
|
kusano |
2b45e8 |
mov pr.rot = 0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.ec = 3
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = 0, I
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L28
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L22:
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFPD f32, f35 = [A1], 2 * SIZE
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f95, 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum1 = f82, f34, xsum1
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f94 = atemp1, f34, f94
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p17) LDFD f90 = [XX], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum2 = f82, f46, xsum2
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f98 = atemp1, f37, f98
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFPD f44, f47 = [A2], 2 * SIZE
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f99, 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum3 = f82, f58, xsum3
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f102 = atemp1, f40, f102
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCHW [PREY], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f92 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum4 = f82, f70, xsum4
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f106 = atemp1, f43, f106
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFPD f56, f59 = [A3], 2 * SIZE
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f103, 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum1 = f85, f37, xsum1
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f94 = atemp2, f46, f94
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFD f96 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum2 = f85, f49, xsum2
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f98 = atemp2, f49, f98
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFPD f68, f71 = [A4], 2 * SIZE
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f107, 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum3 = f85, f61, xsum3
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f102 = atemp2, f52, f102
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFD f100 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum4 = f85, f73, xsum4
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f106 = atemp2, f55, f106
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCH [PREA1], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFPD f38, f41 = [A1], 2 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum1 = f88, f40, xsum1
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f94 = atemp3, f58, f94
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFD f104 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum2 = f88, f52, xsum2
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f98 = atemp3, f61, f98
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCH [PREA2], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFPD f50, f53 = [A2], 2 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum3 = f88, f64, xsum3
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f102 = atemp3, f64, f102
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCH [PREX], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f80 = [XX], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum4 = f88, f76, xsum4
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f106 = atemp3, f67, f106
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCH [PREA3], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFPD f62, f65 = [A3], 2 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum1 = f91, f43, xsum1
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f94 = atemp4, f70, f94
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFD f83 = [XX], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum2 = f91, f55, xsum2
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f98 = atemp4, f73, f98
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) PREFETCH [PREA4], 4 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFPD f74, f77 = [A4], 2 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum3 = f91, f67, xsum3
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p18) FMA f102 = atemp4, f76, f102
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmf
|
|
kusano |
2b45e8 |
(p16) LDFD f86 = [XX], 1 * SIZE
|
|
kusano |
2b45e8 |
(p18) FMA xsum4 = f91, f79, xsum4
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mfb
|
|
kusano |
2b45e8 |
(p18) FMA f106 = atemp4, f79, f106
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L22
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f95, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f99, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f103, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p19) STFD [YYS] = f107, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L28:
|
|
kusano |
2b45e8 |
FMPY xsum1 = alpha, xsum1
|
|
kusano |
2b45e8 |
FMPY xsum2 = alpha, xsum2
|
|
kusano |
2b45e8 |
FMPY xsum3 = alpha, xsum3
|
|
kusano |
2b45e8 |
FMPY xsum4 = alpha, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f64 = [A1], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f65 = [A2], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f66 = [A3], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f67 = [A4], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f68 = [A1], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f69 = [A2], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f70 = [A3], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f71 = [A4], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f72 = [A1], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f73 = [A2], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f74 = [A3], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f75 = [A4], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f76 = [A1], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f77 = [A2], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f78 = [A3], 1 * SIZE
|
|
kusano |
2b45e8 |
LDFD f79 = [A4], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA xsum1 = atemp1, f64, xsum1
|
|
kusano |
2b45e8 |
FMA xsum2 = atemp1, f65, xsum2
|
|
kusano |
2b45e8 |
FMA xsum3 = atemp1, f66, xsum3
|
|
kusano |
2b45e8 |
FMA xsum4 = atemp1, f67, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA xsum1 = atemp2, f65, xsum1
|
|
kusano |
2b45e8 |
FMA xsum2 = atemp2, f69, xsum2
|
|
kusano |
2b45e8 |
FMA xsum3 = atemp2, f70, xsum3
|
|
kusano |
2b45e8 |
FMA xsum4 = atemp2, f71, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA xsum1 = atemp3, f66, xsum1
|
|
kusano |
2b45e8 |
FMA xsum2 = atemp3, f70, xsum2
|
|
kusano |
2b45e8 |
FMA xsum3 = atemp3, f74, xsum3
|
|
kusano |
2b45e8 |
FMA xsum4 = atemp3, f75, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA xsum1 = atemp4, f67, xsum1
|
|
kusano |
2b45e8 |
FMA xsum2 = atemp4, f71, xsum2
|
|
kusano |
2b45e8 |
FMA xsum3 = atemp4, f75, xsum3
|
|
kusano |
2b45e8 |
FMA xsum4 = atemp4, f79, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f36 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f37 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f38 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f39 = [YY], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FADD f36 = f36, xsum1
|
|
kusano |
2b45e8 |
FADD f37 = f37, xsum2
|
|
kusano |
2b45e8 |
FADD f38 = f38, xsum3
|
|
kusano |
2b45e8 |
FADD f39 = f39, xsum4
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YYS] = f36, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YYS] = f37, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YYS] = f38, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YYS] = f39, 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds IS = 4, IS
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds TEMP = 4, IS
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.le p6, p0 = TEMP, M
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L21
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.L30:
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L990:
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
mov r8 = r0
|
|
kusano |
2b45e8 |
adds r9 = 1 * 16, SP
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ldf.fill f16 = [SP], 32
|
|
kusano |
2b45e8 |
ldf.fill f17 = [r9], 32
|
|
kusano |
2b45e8 |
mov ar.lc = ARLC
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ldf.fill f18 = [SP], 32
|
|
kusano |
2b45e8 |
ldf.fill f19 = [r9], 32
|
|
kusano |
2b45e8 |
mov pr = PR, -1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ldf.fill f20 = [SP], 32
|
|
kusano |
2b45e8 |
ldf.fill f21 = [r9], 32
|
|
kusano |
2b45e8 |
mov ar.pfs = ARPFS
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ldf.fill f22 = [SP], 32
|
|
kusano |
2b45e8 |
ldf.fill f23 = [r9]
|
|
kusano |
2b45e8 |
br.ret.sptk.many b0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
EPILOGUE
|