| 1 |
/* Copyright (C) 2008-2025 Free Software Foundation, Inc. |
| 2 |
Contributor: Joern Rennecke <[email protected]> |
| 3 |
on behalf of Synopsys Inc. |
| 4 |
|
| 5 |
This file is part of GCC. |
| 6 |
|
| 7 |
GCC is free software; you can redistribute it and/or modify it under |
| 8 |
the terms of the GNU General Public License as published by the Free |
| 9 |
Software Foundation; either version 3, or (at your option) any later |
| 10 |
version. |
| 11 |
|
| 12 |
GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| 13 |
WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| 14 |
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 15 |
for more details. |
| 16 |
|
| 17 |
Under Section 7 of GPL version 3, you are granted additional |
| 18 |
permissions described in the GCC Runtime Library Exception, version |
| 19 |
3.1, as published by the Free Software Foundation. |
| 20 |
|
| 21 |
You should have received a copy of the GNU General Public License and |
| 22 |
a copy of the GCC Runtime Library Exception along with this program; |
| 23 |
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 24 |
<http://www.gnu.org/licenses/>. */ |
| 25 |
|
| 26 |
/* XMAC schedule: directly back-to-back multiplies stall; the third |
| 27 |
instruction after a multiply stalls unless it is also a multiply. */ |
| 28 |
#include "arc-ieee-754.h" |
| 29 |
|
| 30 |
#if 0 /* DEBUG */ |
| 31 |
.global __muldf3 |
| 32 |
.balign 4 |
| 33 |
__muldf3: |
| 34 |
push_s blink |
| 35 |
push_s r2 |
| 36 |
push_s r3 |
| 37 |
push_s r0 |
| 38 |
bl.d __muldf3_c |
| 39 |
push_s r1 |
| 40 |
ld_s r2,[sp,12] |
| 41 |
ld_s r3,[sp,8] |
| 42 |
st_s r0,[sp,12] |
| 43 |
st_s r1,[sp,8] |
| 44 |
pop_s r1 |
| 45 |
bl.d __muldf3_asm |
| 46 |
pop_s r0 |
| 47 |
pop_s r3 |
| 48 |
pop_s r2 |
| 49 |
pop_s blink |
| 50 |
cmp r0,r2 |
| 51 |
cmp.eq r1,r3 |
| 52 |
jeq_s [blink] |
| 53 |
b abort |
| 54 |
#define __muldf3 __muldf3_asm |
| 55 |
#endif /* DEBUG */ |
| 56 |
/* N.B. This is optimized for ARC700. |
| 57 |
ARC600 has very different scheduling / instruction selection criteria. */ |
| 58 |
/* For the standard multiplier, instead of mpyu rx,DBL0L,DBL1L; tst rx,rx , |
| 59 |
we can do: |
| 60 |
sub rx,DBL0L,1; bic rx,DBL0L,rx; lsr rx,rx; norm rx,rx; asl.f 0,DBL1L,rx */ |
| 61 |
|
| 62 |
__muldf3_support: /* This label makes debugger output saner. */ |
| 63 |
/* If one number is denormal, subtract some from the exponent of the other |
| 64 |
one (if the other exponent is too small, return 0), and normalize the |
| 65 |
denormal. Then re-run the computation. */ |
| 66 |
.balign 4 |
| 67 |
FUNC(__muldf3) |
| 68 |
.Ldenorm_dbl0: |
| 69 |
mov_s r12,DBL0L |
| 70 |
mov_s DBL0L,DBL1L |
| 71 |
mov_s DBL1L,r12 |
| 72 |
mov_s r12,DBL0H |
| 73 |
mov_s DBL0H,DBL1H |
| 74 |
mov_s DBL1H,r12 |
| 75 |
and r11,DBL0H,r9 |
| 76 |
.Ldenorm_dbl1: |
| 77 |
brhs r11,r9,.Linf_nan |
| 78 |
brhs 0x3ca00001,r11,.Lret0 |
| 79 |
sub_s DBL0H,DBL0H,DBL1H |
| 80 |
bmsk_s DBL1H,DBL1H,30 |
| 81 |
add_s DBL0H,DBL0H,DBL1H |
| 82 |
breq_s DBL1H,0,.Ldenorm_2 |
| 83 |
norm r12,DBL1H |
| 84 |
|
| 85 |
sub_s r12,r12,10 |
| 86 |
asl r5,r12,20 |
| 87 |
asl_s DBL1H,DBL1H,r12 |
| 88 |
sub DBL0H,DBL0H,r5 |
| 89 |
neg r5,r12 |
| 90 |
lsr r6,DBL1L,r5 |
| 91 |
asl_s DBL1L,DBL1L,r12 |
| 92 |
b.d __muldf3 |
| 93 |
add_s DBL1H,DBL1H,r6 |
| 94 |
|
| 95 |
.balign 4 |
| 96 |
.Linf_nan: |
| 97 |
bclr r12,DBL1H,31 |
| 98 |
xor_s DBL1H,DBL1H,DBL0H |
| 99 |
bclr_s DBL0H,DBL0H,31 |
| 100 |
max r8,DBL0H,r12 ; either NaN -> NaN ; otherwise inf |
| 101 |
or.f 0,DBL0H,DBL0L |
| 102 |
mov_s DBL0L,0 |
| 103 |
or.ne.f DBL1L,DBL1L,r12 |
| 104 |
not_s DBL0H,DBL0L ; inf * 0 -> NaN |
| 105 |
mov.ne DBL0H,r8 |
| 106 |
tst_s DBL1H,DBL1H |
| 107 |
j_s.d [blink] |
| 108 |
bset.mi DBL0H,DBL0H,31 |
| 109 |
|
| 110 |
.Lret0: xor_s DBL0H,DBL0H,DBL1H |
| 111 |
bclr DBL1H,DBL0H,31 |
| 112 |
xor_s DBL0H,DBL0H,DBL1H |
| 113 |
j_s.d [blink] |
| 114 |
mov_l DBL0L,0 |
| 115 |
|
| 116 |
.balign 4 |
| 117 |
.Ldenorm_2: |
| 118 |
breq_s DBL1L,0,.Lret0 ; 0 input -> 0 output |
| 119 |
norm.f r12,DBL1L |
| 120 |
|
| 121 |
mov.mi r12,21 |
| 122 |
add.pl r12,r12,22 |
| 123 |
neg r11,r12 |
| 124 |
asl_s r12,r12,20 |
| 125 |
lsr.f DBL1H,DBL1L,r11 |
| 126 |
ror DBL1L,DBL1L,r11 |
| 127 |
sub_s DBL0H,DBL0H,r12 |
| 128 |
mov.eq DBL1H,DBL1L |
| 129 |
sub_s DBL1L,DBL1L,DBL1H |
| 130 |
/* Fall through. */ |
| 131 |
.global __muldf3 |
| 132 |
.balign 4 |
| 133 |
__muldf3: |
| 134 |
ld.as r9,[pcl,0x4b] ; ((.L7ff00000-.+2)/4)] |
| 135 |
MPYHU r4,DBL0L,DBL1L |
| 136 |
bmsk r6,DBL0H,19 |
| 137 |
bset r6,r6,20 |
| 138 |
mpyu r7,r6,DBL1L |
| 139 |
and r11,DBL0H,r9 |
| 140 |
breq r11,0,.Ldenorm_dbl0 |
| 141 |
MPYHU r8,r6,DBL1L |
| 142 |
bmsk r10,DBL1H,19 |
| 143 |
bset r10,r10,20 |
| 144 |
MPYHU r5,r10,DBL0L |
| 145 |
add.f r4,r4,r7 |
| 146 |
and r12,DBL1H,r9 |
| 147 |
MPYHU r7,r6,r10 |
| 148 |
breq r12,0,.Ldenorm_dbl1 |
| 149 |
adc.f r5,r5,r8 |
| 150 |
mpyu r8,r10,DBL0L |
| 151 |
breq r11,r9,.Linf_nan |
| 152 |
breq r12,r9,.Linf_nan |
| 153 |
mpyu r6,r6,r10 |
| 154 |
add.cs r7,r7,1 |
| 155 |
add.f r4,r4,r8 |
| 156 |
mpyu r10,DBL1L,DBL0L |
| 157 |
bclr r8,r9,30 ; 0x3ff00000 |
| 158 |
adc.f r5,r5,r6 |
| 159 |
; XMAC write-back stall / std. mult stall is one cycle later |
| 160 |
bclr r6,r9,20 ; 0x7fe00000 |
| 161 |
add.cs r7,r7,1 ; fraction product in r7:r5:r4 |
| 162 |
tst r10,r10 |
| 163 |
bset.ne r4,r4,0 ; put least significant word into sticky bit |
| 164 |
lsr.f r10,r7,9 |
| 165 |
add_l r12,r12,r11 ; add exponents |
| 166 |
rsub.eq r8,r8,r9 ; 0x40000000 |
| 167 |
sub r12,r12,r8 ; subtract bias + implicit 1 |
| 168 |
brhs.d r12,r6,.Linf_denorm |
| 169 |
rsub r10,r10,12 |
| 170 |
.Lshift_frac: |
| 171 |
neg r8,r10 |
| 172 |
asl r6,r4,r10 |
| 173 |
lsr DBL0L,r4,r8 |
| 174 |
add.f 0,r6,r6 |
| 175 |
btst.eq DBL0L,0 |
| 176 |
cmp.eq r4,r4 ; round to nearest / round to even |
| 177 |
asl r4,r5,r10 |
| 178 |
lsr r5,r5,r8 |
| 179 |
adc.f DBL0L,DBL0L,r4 |
| 180 |
xor.f 0,DBL0H,DBL1H |
| 181 |
asl r7,r7,r10 |
| 182 |
add_s r12,r12,r5 |
| 183 |
adc DBL0H,r12,r7 |
| 184 |
j_s.d [blink] |
| 185 |
bset.mi DBL0H,DBL0H,31 |
| 186 |
|
| 187 |
/* We have checked for infinity / NaN input before, and transformed |
| 188 |
denormalized inputs into normalized inputs. Thus, the worst case |
| 189 |
exponent overflows are: |
| 190 |
1 + 1 - 0x400 == 0xc02 : maximum underflow |
| 191 |
0x7fe + 0x7fe - 0x3ff == 0xbfd ; maximum overflow |
| 192 |
N.B. 0x7e and 0x7f are also values for overflow. |
| 193 |
|
| 194 |
If (r12 <= -54), we have an underflow to zero. */ |
| 195 |
.balign 4 |
| 196 |
.Linf_denorm: |
| 197 |
brlo r12,0xc0000000,.Linf |
| 198 |
asr r6,r12,20 |
| 199 |
mov_s r12,0 |
| 200 |
add.f r10,r10,r6 |
| 201 |
brgt r10,0,.Lshift_frac |
| 202 |
beq_s .Lround_frac |
| 203 |
add.f r10,r10,32 |
| 204 |
.Lshift32_frac: |
| 205 |
tst r4,r4 |
| 206 |
mov r4,r5 |
| 207 |
bset.ne r4,r4,1 |
| 208 |
mov r5,r7 |
| 209 |
mov r7,0 |
| 210 |
brge r10,1,.Lshift_frac |
| 211 |
breq r10,0,.Lround_frac |
| 212 |
add.f r10,r10,32 |
| 213 |
brgt r10,21,.Lshift32_frac |
| 214 |
b_s .Lret0 |
| 215 |
|
| 216 |
.Lround_frac: |
| 217 |
add.f 0,r4,r4 |
| 218 |
btst.eq r5,0 |
| 219 |
mov_s DBL0L,r5 |
| 220 |
mov_s DBL0H,r7 |
| 221 |
adc.eq.f DBL0L,DBL0L,0 |
| 222 |
j_s.d [blink] |
| 223 |
|
| 224 |
adc.eq DBL0H,DBL0H,0 |
| 225 |
|
| 226 |
.Linf: xor.f DBL1H,DBL1H,DBL0H |
| 227 |
mov_s DBL0L,0 |
| 228 |
mov_s DBL0H,r9 |
| 229 |
j_s.d [blink] |
| 230 |
bset.mi DBL0H,DBL0H,31 |
| 231 |
ENDFUNC(__muldf3) |
| 232 |
|
| 233 |
.balign 4 |
| 234 |
.L7ff00000: |
| 235 |
.long 0x7ff00000 |