tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
sh: Provide sdivsi3/udivsi3/udivdi3 for sh64, kill off libgcc linking.
This moves in the necessary libgcc bits and kills off the libgcc linking
for sh64 kernels as well.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
+314
-5
+1
-3
arch/sh/Makefile
+1
-3
arch/sh/Makefile
···
177
177
KBUILD_CPPFLAGS += $(cflags-y)
178
178
KBUILD_AFLAGS += $(cflags-y)
179
179
180
-
LIBGCC := $(shell $(CC) $(KBUILD_CFLAGS) -print-libgcc-file-name)
181
-
182
180
libs-$(CONFIG_SUPERH32) := arch/sh/lib/ $(libs-y)
183
-
libs-$(CONFIG_SUPERH64) := arch/sh/lib64/ $(libs-y) $(LIBGCC)
181
+
libs-$(CONFIG_SUPERH64) := arch/sh/lib64/ $(libs-y)
184
182
185
183
PHONY += maketools FORCE
186
184
-2
arch/sh/kernel/sh_ksyms_64.c
-2
arch/sh/kernel/sh_ksyms_64.c
+3
arch/sh/lib64/Makefile
+3
arch/sh/lib64/Makefile
+131
arch/sh/lib64/sdivsi3.S
+131
arch/sh/lib64/sdivsi3.S
···
1
+
.global __sdivsi3
2
+
.section .text..SHmedia32,"ax"
3
+
.align 2
4
+
5
+
/* inputs: r4,r5 */
6
+
/* clobbered: r1,r18,r19,r20,r21,r25,tr0 */
7
+
/* result in r0 */
8
+
__sdivsi3:
9
+
ptb __div_table,tr0
10
+
11
+
nsb r5, r1
12
+
shlld r5, r1, r25 /* normalize; [-2 ..1, 1..2) in s2.62 */
13
+
shari r25, 58, r21 /* extract 5(6) bit index (s2.4 with hole -1..1) */
14
+
/* bubble */
15
+
gettr tr0,r20
16
+
ldx.ub r20, r21, r19 /* u0.8 */
17
+
shari r25, 32, r25 /* normalize to s2.30 */
18
+
shlli r21, 1, r21
19
+
muls.l r25, r19, r19 /* s2.38 */
20
+
ldx.w r20, r21, r21 /* s2.14 */
21
+
ptabs r18, tr0
22
+
shari r19, 24, r19 /* truncate to s2.14 */
23
+
sub r21, r19, r19 /* some 11 bit inverse in s1.14 */
24
+
muls.l r19, r19, r21 /* u0.28 */
25
+
sub r63, r1, r1
26
+
addi r1, 92, r1
27
+
muls.l r25, r21, r18 /* s2.58 */
28
+
shlli r19, 45, r19 /* multiply by two and convert to s2.58 */
29
+
/* bubble */
30
+
sub r19, r18, r18
31
+
shari r18, 28, r18 /* some 22 bit inverse in s1.30 */
32
+
muls.l r18, r25, r0 /* s2.60 */
33
+
muls.l r18, r4, r25 /* s32.30 */
34
+
/* bubble */
35
+
shari r0, 16, r19 /* s-16.44 */
36
+
muls.l r19, r18, r19 /* s-16.74 */
37
+
shari r25, 63, r0
38
+
shari r4, 14, r18 /* s19.-14 */
39
+
shari r19, 30, r19 /* s-16.44 */
40
+
muls.l r19, r18, r19 /* s15.30 */
41
+
xor r21, r0, r21 /* You could also use the constant 1 << 27. */
42
+
add r21, r25, r21
43
+
sub r21, r19, r21
44
+
shard r21, r1, r21
45
+
sub r21, r0, r0
46
+
blink tr0, r63
47
+
48
+
/* This table has been generated by divtab.c .
49
+
Defects for bias -330:
50
+
Max defect: 6.081536e-07 at -1.000000e+00
51
+
Min defect: 2.849516e-08 at 1.030651e+00
52
+
Max 2nd step defect: 9.606539e-12 at -1.000000e+00
53
+
Min 2nd step defect: 0.000000e+00 at 0.000000e+00
54
+
Defect at 1: 1.238659e-07
55
+
Defect at -2: 1.061708e-07 */
56
+
57
+
.balign 2
58
+
.type __div_table,@object
59
+
.size __div_table,128
60
+
/* negative division constants */
61
+
.word -16638
62
+
.word -17135
63
+
.word -17737
64
+
.word -18433
65
+
.word -19103
66
+
.word -19751
67
+
.word -20583
68
+
.word -21383
69
+
.word -22343
70
+
.word -23353
71
+
.word -24407
72
+
.word -25582
73
+
.word -26863
74
+
.word -28382
75
+
.word -29965
76
+
.word -31800
77
+
/* negative division factors */
78
+
.byte 66
79
+
.byte 70
80
+
.byte 75
81
+
.byte 81
82
+
.byte 87
83
+
.byte 93
84
+
.byte 101
85
+
.byte 109
86
+
.byte 119
87
+
.byte 130
88
+
.byte 142
89
+
.byte 156
90
+
.byte 172
91
+
.byte 192
92
+
.byte 214
93
+
.byte 241
94
+
.skip 16
95
+
.global __div_table
96
+
__div_table:
97
+
.skip 16
98
+
/* positive division factors */
99
+
.byte 241
100
+
.byte 214
101
+
.byte 192
102
+
.byte 172
103
+
.byte 156
104
+
.byte 142
105
+
.byte 130
106
+
.byte 119
107
+
.byte 109
108
+
.byte 101
109
+
.byte 93
110
+
.byte 87
111
+
.byte 81
112
+
.byte 75
113
+
.byte 70
114
+
.byte 66
115
+
/* positive division constants */
116
+
.word 31801
117
+
.word 29966
118
+
.word 28383
119
+
.word 26864
120
+
.word 25583
121
+
.word 24408
122
+
.word 23354
123
+
.word 22344
124
+
.word 21384
125
+
.word 20584
126
+
.word 19752
127
+
.word 19104
128
+
.word 18434
129
+
.word 17738
130
+
.word 17136
131
+
.word 16639
+120
arch/sh/lib64/udivdi3.S
+120
arch/sh/lib64/udivdi3.S
···
1
+
.section .text..SHmedia32,"ax"
2
+
.align 2
3
+
.global __udivdi3
4
+
__udivdi3:
5
+
shlri r3,1,r4
6
+
nsb r4,r22
7
+
shlld r3,r22,r6
8
+
shlri r6,49,r5
9
+
movi 0xffffffffffffbaf1,r21 /* .l shift count 17. */
10
+
sub r21,r5,r1
11
+
mmulfx.w r1,r1,r4
12
+
mshflo.w r1,r63,r1
13
+
sub r63,r22,r20 // r63 == 64 % 64
14
+
mmulfx.w r5,r4,r4
15
+
pta large_divisor,tr0
16
+
addi r20,32,r9
17
+
msub.w r1,r4,r1
18
+
madd.w r1,r1,r1
19
+
mmulfx.w r1,r1,r4
20
+
shlri r6,32,r7
21
+
bgt/u r9,r63,tr0 // large_divisor
22
+
mmulfx.w r5,r4,r4
23
+
shlri r2,32+14,r19
24
+
addi r22,-31,r0
25
+
msub.w r1,r4,r1
26
+
27
+
mulu.l r1,r7,r4
28
+
addi r1,-3,r5
29
+
mulu.l r5,r19,r5
30
+
sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
31
+
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
32
+
the case may be, %0000000000000000 000.11111111111, still */
33
+
muls.l r1,r4,r4 /* leaving at least one sign bit. */
34
+
mulu.l r5,r3,r8
35
+
mshalds.l r1,r21,r1
36
+
shari r4,26,r4
37
+
shlld r8,r0,r8
38
+
add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
39
+
sub r2,r8,r2
40
+
/* Can do second step of 64 : 32 div now, using r1 and the rest in r2. */
41
+
42
+
shlri r2,22,r21
43
+
mulu.l r21,r1,r21
44
+
shlld r5,r0,r8
45
+
addi r20,30-22,r0
46
+
shlrd r21,r0,r21
47
+
mulu.l r21,r3,r5
48
+
add r8,r21,r8
49
+
mcmpgt.l r21,r63,r21 // See Note 1
50
+
addi r20,30,r0
51
+
mshfhi.l r63,r21,r21
52
+
sub r2,r5,r2
53
+
andc r2,r21,r2
54
+
55
+
/* small divisor: need a third divide step */
56
+
mulu.l r2,r1,r7
57
+
ptabs r18,tr0
58
+
addi r2,1,r2
59
+
shlrd r7,r0,r7
60
+
mulu.l r7,r3,r5
61
+
add r8,r7,r8
62
+
sub r2,r3,r2
63
+
cmpgt r2,r5,r5
64
+
add r8,r5,r2
65
+
/* could test r3 here to check for divide by zero. */
66
+
blink tr0,r63
67
+
68
+
large_divisor:
69
+
mmulfx.w r5,r4,r4
70
+
shlrd r2,r9,r25
71
+
shlri r25,32,r8
72
+
msub.w r1,r4,r1
73
+
74
+
mulu.l r1,r7,r4
75
+
addi r1,-3,r5
76
+
mulu.l r5,r8,r5
77
+
sub r63,r4,r4 // Negate to make sure r1 ends up <= 1/r2
78
+
shlri r4,2,r4 /* chop off leading %0000000000000000 001.00000000000 - or, as
79
+
the case may be, %0000000000000000 000.11111111111, still */
80
+
muls.l r1,r4,r4 /* leaving at least one sign bit. */
81
+
shlri r5,14-1,r8
82
+
mulu.l r8,r7,r5
83
+
mshalds.l r1,r21,r1
84
+
shari r4,26,r4
85
+
add r1,r4,r1 // 31 bit unsigned reciprocal now in r1 (msb equiv. 0.5)
86
+
sub r25,r5,r25
87
+
/* Can do second step of 64 : 32 div now, using r1 and the rest in r25. */
88
+
89
+
shlri r25,22,r21
90
+
mulu.l r21,r1,r21
91
+
pta no_lo_adj,tr0
92
+
addi r22,32,r0
93
+
shlri r21,40,r21
94
+
mulu.l r21,r7,r5
95
+
add r8,r21,r8
96
+
shlld r2,r0,r2
97
+
sub r25,r5,r25
98
+
bgtu/u r7,r25,tr0 // no_lo_adj
99
+
addi r8,1,r8
100
+
sub r25,r7,r25
101
+
no_lo_adj:
102
+
mextr4 r2,r25,r2
103
+
104
+
/* large_divisor: only needs a few adjustments. */
105
+
mulu.l r8,r6,r5
106
+
ptabs r18,tr0
107
+
/* bubble */
108
+
cmpgtu r5,r2,r5
109
+
sub r8,r5,r2
110
+
blink tr0,r63
111
+
112
+
/* Note 1: To shift the result of the second divide stage so that the result
113
+
always fits into 32 bits, yet we still reduce the rest sufficiently
114
+
would require a lot of instructions to do the shifts just right. Using
115
+
the full 64 bit shift result to multiply with the divisor would require
116
+
four extra instructions for the upper 32 bits (shift / mulu / shift / sub).
117
+
Fortunately, if the upper 32 bits of the shift result are nonzero, we
118
+
know that the rest after taking this partial result into account will
119
+
fit into 32 bits. So we just clear the upper 32 bits of the rest if the
120
+
upper 32 bits of the partial result are nonzero. */
+59
arch/sh/lib64/udivsi3.S
+59
arch/sh/lib64/udivsi3.S
···
1
+
.global __udivsi3
2
+
.section .text..SHmedia32,"ax"
3
+
.align 2
4
+
5
+
/*
6
+
inputs: r4,r5
7
+
clobbered: r18,r19,r20,r21,r22,r25,tr0
8
+
result in r0.
9
+
*/
10
+
__udivsi3:
11
+
addz.l r5,r63,r22
12
+
nsb r22,r0
13
+
shlld r22,r0,r25
14
+
shlri r25,48,r25
15
+
movi 0xffffffffffffbb0c,r20 /* shift count eqiv 76 */
16
+
sub r20,r25,r21
17
+
mmulfx.w r21,r21,r19
18
+
mshflo.w r21,r63,r21
19
+
ptabs r18,tr0
20
+
mmulfx.w r25,r19,r19
21
+
sub r20,r0,r0
22
+
/* bubble */
23
+
msub.w r21,r19,r19
24
+
25
+
/*
26
+
* It would be nice for scheduling to do this add to r21 before
27
+
* the msub.w, but we need a different value for r19 to keep
28
+
* errors under control.
29
+
*/
30
+
addi r19,-2,r21
31
+
mulu.l r4,r21,r18
32
+
mmulfx.w r19,r19,r19
33
+
shlli r21,15,r21
34
+
shlrd r18,r0,r18
35
+
mulu.l r18,r22,r20
36
+
mmacnfx.wl r25,r19,r21
37
+
/* bubble */
38
+
sub r4,r20,r25
39
+
40
+
mulu.l r25,r21,r19
41
+
addi r0,14,r0
42
+
/* bubble */
43
+
shlrd r19,r0,r19
44
+
mulu.l r19,r22,r20
45
+
add r18,r19,r18
46
+
/* bubble */
47
+
sub.l r25,r20,r25
48
+
49
+
mulu.l r25,r21,r19
50
+
addz.l r25,r63,r25
51
+
sub r25,r22,r25
52
+
shlrd r19,r0,r19
53
+
mulu.l r19,r22,r20
54
+
addi r25,1,r25
55
+
add r18,r19,r18
56
+
57
+
cmpgt r25,r20,r25
58
+
add.l r18,r25,r0
59
+
blink tr0,r63