Retro68/gcc/libgcc/config/arc/ieee-754/arc600-dsp/muldf3.S
2017-10-07 02:16:47 +02:00

232 lines
5.2 KiB
ArmAsm

/* Copyright (C) 2008-2017 Free Software Foundation, Inc.
Contributor: Joern Rennecke <joern.rennecke@embecosm.com>
on behalf of Synopsys Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
Under Section 7 of GPL version 3, you are granted additional
permissions described in the GCC Runtime Library Exception, version
3.1, as published by the Free Software Foundation.
You should have received a copy of the GNU General Public License and
a copy of the GCC Runtime Library Exception along with this program;
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
<http://www.gnu.org/licenses/>. */
#include "../arc-ieee-754.h"
#if 0 /* DEBUG */
.global __muldf3
.balign 4
__muldf3:
push_s blink
push_s r2
push_s r3
push_s r0
bl.d __muldf3_c
push_s r1
ld_s r2,[sp,12]
ld_s r3,[sp,8]
st_s r0,[sp,12]
st_s r1,[sp,8]
pop_s r1
bl.d __muldf3_asm
pop_s r0
pop_s r3
pop_s r2
pop_s blink
cmp r0,r2
cmp.eq r1,r3
jeq_s [blink]
b abort
#define __muldf3 __muldf3_asm
#endif /* DEBUG */
__muldf3_support: /* This label makes debugger output saner. */
.balign 4
FUNC(__muldf3)
.Ldenorm_2:
breq.d DBL1L,0,.Lret0_2 ; 0 input -> 0 output
norm.f r12,DBL1L
mov.mi r12,21
add.pl r12,r12,22
neg r11,r12
asl_s r12,r12,20
lsr.f DBL1H,DBL1L,r11
ror DBL1L,DBL1L,r11
sub_s DBL0H,DBL0H,r12
mov.eq DBL1H,DBL1L
sub_l DBL1L,DBL1L,DBL1H
/* Fall through. */
.global __muldf3
.balign 4
__muldf3:
mululw 0,DBL0L,DBL1L
machulw r4,DBL0L,DBL1L
ld.as r9,[pcl,0x67] ; ((.L7ff00000-.+2)/4)]
bmsk r6,DBL0H,19
bset r6,r6,20
mov r8,acc2
mululw 0,r4,1
and r11,DBL0H,r9
breq.d r11,0,.Ldenorm_dbl0
and r12,DBL1H,r9
breq.d r12,0,.Ldenorm_dbl1
maclw 0,r6,DBL1L
machulw 0,r6,DBL1L
breq.d r11,r9,.Linf_nan
bmsk r10,DBL1H,19
breq.d r12,r9,.Linf_nan
bset r10,r10,20
maclw 0,r10,DBL0L
machulw r5,r10,DBL0L
add_s r12,r12,r11 ; add exponents
mov r4,acc2
mululw 0,r5,1
maclw 0,r6,r10
machulw r7,r6,r10 ; fraction product in r7:acc2:r4:r8
tst r8,r8
bclr r8,r9,30 ; 0x3ff00000
bset.ne r4,r4,0 ; put least significant word into sticky bit
bclr r6,r9,20 ; 0x7fe00000
lsr.f r10,r7,9
rsub.eq r8,r8,r9 ; 0x40000000
sub r12,r12,r8 ; subtract bias + implicit 1
brhs.d r12,r6,.Linf_denorm
rsub r10,r10,12
.Lshift_frac:
neg r8,r10
asl r6,r4,r10
lsr DBL0L,r4,r8
add.f 0,r6,r6
btst.eq DBL0L,0
cmp.eq r4,r4 ; round to nearest / round to even
asl r4,acc2,r10
lsr r5,acc2,r8
adc.f DBL0L,DBL0L,r4
xor.f 0,DBL0H,DBL1H
asl r7,r7,r10
add_s r12,r12,r5
adc DBL0H,r12,r7
j_s.d [blink]
bset.mi DBL0H,DBL0H,31
/* N.B. This is optimized for ARC700.
ARC600 has very different scheduling / instruction selection criteria. */
/* If one number is denormal, subtract some from the exponent of the other
one (if the other exponent is too small, return 0), and normalize the
denormal. Then re-run the computation. */
.Lret0_2:
lsr_s DBL0H,DBL0H,31
asl_s DBL0H,DBL0H,31
j_s.d [blink]
mov_s DBL0L,0
.balign 4
.Ldenorm_dbl0:
mov_s r12,DBL0L
mov_s DBL0L,DBL1L
mov_s DBL1L,r12
mov_s r12,DBL0H
mov_s DBL0H,DBL1H
mov_s DBL1H,r12
and r11,DBL0H,r9
.Ldenorm_dbl1:
brhs r11,r9,.Linf_nan
brhs 0x3ca00001,r11,.Lret0
sub_s DBL0H,DBL0H,DBL1H
bmsk.f DBL1H,DBL1H,30
add_s DBL0H,DBL0H,DBL1H
beq.d .Ldenorm_2
norm r12,DBL1H
sub_s r12,r12,10
asl r5,r12,20
asl_s DBL1H,DBL1H,r12
sub DBL0H,DBL0H,r5
neg r5,r12
lsr r6,DBL1L,r5
asl_s DBL1L,DBL1L,r12
b.d __muldf3
add_s DBL1H,DBL1H,r6
.Lret0: xor_s DBL0H,DBL0H,DBL1H
bclr DBL1H,DBL0H,31
xor_s DBL0H,DBL0H,DBL1H
j_s.d [blink]
mov_s DBL0L,0
.balign 4
.Linf_nan:
bclr r12,DBL1H,31
xor_s DBL1H,DBL1H,DBL0H
bclr_s DBL0H,DBL0H,31
max r8,DBL0H,r12 ; either NaN -> NaN ; otherwise inf
or.f 0,DBL0H,DBL0L
mov_s DBL0L,0
or.ne.f DBL1L,DBL1L,r12
not_s DBL0H,DBL0L ; inf * 0 -> NaN
mov.ne DBL0H,r8
tst_s DBL1H,DBL1H
j_s.d [blink]
bset.mi DBL0H,DBL0H,31
/* We have checked for infinity / NaN input before, and transformed
denormalized inputs into normalized inputs. Thus, the worst case
exponent overflows are:
1 + 1 - 0x400 == 0xc02 : maximum underflow
0x7fe + 0x7fe - 0x3ff == 0xbfd ; maximum overflow
N.B. 0x7e and 0x7f are also values for overflow.
If (r12 <= -54), we have an underflow to zero. */
.balign 4
.Linf_denorm:
lsr r6,r12,28
brlo.d r6,0xc,.Linf
asr r6,r12,20
add.f r10,r10,r6
brgt.d r10,0,.Lshift_frac
mov_s r12,0
beq.d .Lround_frac
add r10,r10,32
.Lshift32_frac:
tst r4,r4
mov r4,acc2
bset.ne r4,r4,1
mululw 0,r7,1
brge.d r10,1,.Lshift_frac
mov r7,0
breq.d r10,0,.Lround_frac
add r10,r10,32
brgt r10,21,.Lshift32_frac
b_s .Lret0
.Lround_frac:
add.f 0,r4,r4
btst.eq acc2,0
mov_s DBL0L,acc2
mov_s DBL0H,r7
adc.eq.f DBL0L,DBL0L,0
j_s.d [blink]
adc.eq DBL0H,DBL0H,0
.Linf: mov_s DBL0L,0
xor.f DBL1H,DBL1H,DBL0H
mov_s DBL0H,r9
j_s.d [blink]
bset.mi DBL0H,DBL0H,31
ENDFUNC(__muldf3)
.balign 4
.L7ff00000:
.long 0x7ff00000