blob: 8984627437a614a274329535cbe6f7c1a96c6112 [file] [log] [blame]
Stephen Hines2d1fdb22014-05-28 23:58:16 -07001// This file is dual licensed under the MIT and the University of Illinois Open
2// Source Licenses. See LICENSE.TXT for details.
3
4#include "../assembly.h"
5
6// float __floatundisf(du_int a);
7
8// Note that there is a hardware instruction, fildll, that does most of what
9// this function needs to do. However, because of our ia32 ABI, it will take
10// a write-small read-large stall, so the software implementation here is
11// actually several cycles faster.
12
13// This is a branch-free implementation. A branchy implementation might be
14// faster for the common case if you know something a priori about the input
15// distribution.
16
17/* branch-free x87 implementation - one cycle slower than without x87.
18
19#ifdef __i386__
20
21.const
22.balign 3
23
24 .quad 0x43f0000000000000
25twop64: .quad 0x0000000000000000
26
27#define TWOp64 twop64-0b(%ecx,%eax,8)
28
29.text
30.balign 4
31DEFINE_COMPILERRT_FUNCTION(__floatundisf)
32 movl 8(%esp), %eax
33 movd 8(%esp), %xmm1
34 movd 4(%esp), %xmm0
35 punpckldq %xmm1, %xmm0
36 calll 0f
370: popl %ecx
38 sarl $31, %eax
39 movq %xmm0, 4(%esp)
40 fildll 4(%esp)
41 faddl TWOp64
42 fstps 4(%esp)
43 flds 4(%esp)
44 ret
45END_COMPILERRT_FUNCTION(__floatundisf)
46
47#endif // __i386__
48
49*/
50
51/* branch-free, x87-free implementation - faster at the expense of code size */
52
53#ifdef __i386__
54
55#ifndef __ELF__
56.const
57#endif
58.balign 8
59twop52: .quad 0x4330000000000000
60 .quad 0x0000000000000fff
61sticky: .quad 0x0000000000000000
62 .long 0x00000012
63twelve: .long 0x00000000
64
65#define TWOp52 twop52-0b(%ecx)
66#define STICKY sticky-0b(%ecx,%eax,8)
67
68.text
69.balign 4
70DEFINE_COMPILERRT_FUNCTION(__floatundisf)
71 movl 8(%esp), %eax
72 movd 8(%esp), %xmm1
73 movd 4(%esp), %xmm0
74 punpckldq %xmm1, %xmm0
75
76 calll 0f
770: popl %ecx
78 shrl %eax // high 31 bits of input as sint32
79 addl $0x7ff80000, %eax
80 sarl $31, %eax // (big input) ? -1 : 0
81 movsd STICKY, %xmm1 // (big input) ? 0xfff : 0
82 movl $12, %edx
83 andl %eax, %edx // (big input) ? 12 : 0
84 movd %edx, %xmm3
85 andpd %xmm0, %xmm1 // (big input) ? input & 0xfff : 0
86 movsd TWOp52, %xmm2 // 0x1.0p52
87 psrlq %xmm3, %xmm0 // (big input) ? input >> 12 : input
88 orpd %xmm2, %xmm1 // 0x1.0p52 + ((big input) ? input & 0xfff : input)
89 orpd %xmm1, %xmm0 // 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input)
90 subsd %xmm2, %xmm0 // (double)((big input) ? (input >> 12 | input & 0xfff) : input)
91 cvtsd2ss %xmm0, %xmm0 // (float)((big input) ? (input >> 12 | input & 0xfff) : input)
92 pslld $23, %xmm3
93 paddd %xmm3, %xmm0 // (float)input
94 movd %xmm0, 4(%esp)
95 flds 4(%esp)
96 ret
97END_COMPILERRT_FUNCTION(__floatundisf)
98
99#endif // __i386__