overby.me
this repo has no description
1/* double ceil(double )
2 *
3 * Reimplemented by Steve Canon, based on Ian Ollmann's implementations
4 * tuned for increased performance on in-order machines (but faster on
5 * out-of-order machines as well).
6 *
7 * Copyright 2009, Apple Inc.
8 */
9
10#ifdef __i386__
11
12#ifdef __SSE3__
13
14.text
15.align 4
16.globl _ceil
17_ceil:
18 mov 8(%esp), %ecx
19 fldl 4(%esp)
20 cmp $0x43300000, %ecx // if x is negative or x > 0x1.0p53
21 jae 2f // goto 2
22
23 fld %st(0)
24 fistpll 4(%esp)
25 fildll 4(%esp) // rint(x)
26 fucomi %st(1), %st(0)
27 fstp %st(1)
28 jae 1f // if rint(x) >= x, return rint(x)
29 fld1
30 faddp %st(0), %st(1) // otherwise, return rint(x) + 1.0
311: ret
32
332: and $0x7fffffff, %ecx
34 cmp $0x43300000, %ecx // if |x| >= 0x1.0p53
35 jae 3f
36
37 fisttpll 4(%esp)
38 fldz
39 fildll 4(%esp) // ceil(x), up to sign of zero
40 fucomi %st(1), %st(0) // if result == 0
41 fstp %st(1) //
42 jne 3f //
43 fchs // patch up sign bit
443: ret
45
46#else // i386, no SSE3
47
48.text
49.align 4
50.globl _ceil
51_ceil:
52 mov 8(%esp), %ecx
53 fldl 4(%esp)
54 cmp $0x43300000, %ecx // if x is negative or x > 0x1.0p53
55 jae 2f // goto 2
56
57 fld %st(0)
58 fistpll 4(%esp)
59 fildll 4(%esp) // rint(x)
60 fucomi %st(1), %st(0)
61 fstp %st(1)
62 jae 1f // if rint(x) >= x, return rint(x)
63 fld1
64 faddp // otherwise, return rint(x) + 1.0
651: ret
66
672: and $0x7fffffff, %ecx
68 cmp $0x43300000, %ecx // if |x| >= 0x1.0p53
69 fld %st(0)
70 jae 4f
71
72 fistpll 4(%esp)
73 fldz
74 fildll 4(%esp) // rint(x)
75 fucomi %st(2), %st(0)
76 fstp %st(2)
77 jae 3f // if rint(x) < x, add one
78 fld1
79 faddp %st(2), %st(0)
803: fucomi %st(1), %st(0)
81 jne 4f // if ceil(x) == 0, patch up sign bit
82 fchs
83 fstp %st(1)
84 ret
85
864: fstp %st(0)
87 ret
88
89
90#endif // SSE3
91
92#else // x86_64
93
94.const
95.align 4
96one: .quad 0x3ff0000000000000
97absmask:.quad 0x7fffffffffffffff
98thresh: .quad 0x4330000000000000
99
100.text
101.align 4
102.globl _ceil
103_ceil:
104 movd %xmm0, %rcx
105 andq absmask(%rip), %rcx // |x|
106 cmpq thresh(%rip), %rcx // if |x| >= 0x1.0p52 or isnan(x)
107 movsd absmask(%rip), %xmm2 //
108 jae 1f // early out, returning x.
109
110 cvttsd2si %xmm0, %rax
111 andnpd %xmm0, %xmm2 // signbit(x)
112 cvtsi2sd %rax, %xmm1 // trunc(x), except for sign of zero
113 cmplesd %xmm1, %xmm0
114 orpd %xmm2, %xmm1 // trunc(x)
115 andnpd one(%rip), %xmm0 // x <= trunc(x) ? 0.0 : 1.0
116 orpd %xmm2, %xmm0 // x <= trunc(x) ? copysign(0.0, x) : 1.0
117 addsd %xmm1, %xmm0 // ceil(x)
1181: ret
119
120#endif