1 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
2 | // See https://llvm.org/LICENSE.txt for license information. |
3 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
4 | |
5 | #include "../assembly.h" |
6 | |
7 | // float __floatundisf(du_int a); |
8 | |
9 | // Note that there is a hardware instruction, fildll, that does most of what |
10 | // this function needs to do. However, because of our ia32 ABI, it will take |
11 | // a write-small read-large stall, so the software implementation here is |
12 | // actually several cycles faster. |
13 | |
14 | // This is a branch-free implementation. A branchy implementation might be |
15 | // faster for the common case if you know something a priori about the input |
16 | // distribution. |
17 | |
18 | /* branch-free x87 implementation - one cycle slower than without x87. |
19 | |
20 | #ifdef __i386__ |
21 | |
22 | CONST_SECTION |
23 | .balign 3 |
24 | |
25 | .quad 0x43f0000000000000 |
26 | twop64: .quad 0x0000000000000000 |
27 | |
28 | #define TWOp64 twop64-0b(%ecx,%eax,8) |
29 | |
30 | .text |
31 | .balign 4 |
32 | DEFINE_COMPILERRT_FUNCTION(__floatundisf) |
33 | movl 8(%esp), %eax |
34 | movd 8(%esp), %xmm1 |
35 | movd 4(%esp), %xmm0 |
36 | punpckldq %xmm1, %xmm0 |
37 | calll 0f |
38 | 0: popl %ecx |
39 | sarl $31, %eax |
40 | movq %xmm0, 4(%esp) |
41 | fildll 4(%esp) |
42 | faddl TWOp64 |
43 | fstps 4(%esp) |
44 | flds 4(%esp) |
45 | ret |
46 | END_COMPILERRT_FUNCTION(__floatundisf) |
47 | |
48 | #endif // __i386__ |
49 | |
50 | */ |
51 | |
52 | // branch-free, x87-free implementation - faster at the expense of code size |
53 | |
54 | #ifdef __i386__ |
55 | |
56 | CONST_SECTION |
57 | |
58 | .balign 16 |
59 | twop52: |
60 | .quad 0x4330000000000000 |
61 | .quad 0x0000000000000fff |
62 | |
63 | .balign 16 |
64 | sticky: |
65 | .quad 0x0000000000000000 |
66 | .long 0x00000012 |
67 | |
68 | .balign 16 |
69 | twelve: |
70 | .long 0x00000000 |
71 | |
72 | #define TWOp52 twop52-0b(%ecx) |
73 | #define STICKY sticky-0b(%ecx,%eax,8) |
74 | |
75 | .text |
76 | .balign 4 |
77 | DEFINE_COMPILERRT_FUNCTION(__floatundisf) |
78 | movl 8(%esp), %eax |
79 | movd 8(%esp), %xmm1 |
80 | movd 4(%esp), %xmm0 |
81 | punpckldq %xmm1, %xmm0 |
82 | |
83 | calll 0f |
84 | 0: popl %ecx |
85 | shrl %eax // high 31 bits of input as sint32 |
86 | addl $0x7ff80000, %eax |
87 | sarl $31, %eax // (big input) ? -1 : 0 |
88 | movsd STICKY, %xmm1 // (big input) ? 0xfff : 0 |
89 | movl $12, %edx |
90 | andl %eax, %edx // (big input) ? 12 : 0 |
91 | movd %edx, %xmm3 |
92 | andpd %xmm0, %xmm1 // (big input) ? input & 0xfff : 0 |
93 | movsd TWOp52, %xmm2 // 0x1.0p52 |
94 | psrlq %xmm3, %xmm0 // (big input) ? input >> 12 : input |
95 | orpd %xmm2, %xmm1 // 0x1.0p52 + ((big input) ? input & 0xfff : input) |
96 | orpd %xmm1, %xmm0 // 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input) |
97 | subsd %xmm2, %xmm0 // (double)((big input) ? (input >> 12 | input & 0xfff) : input) |
98 | cvtsd2ss %xmm0, %xmm0 // (float)((big input) ? (input >> 12 | input & 0xfff) : input) |
99 | pslld $23, %xmm3 |
100 | paddd %xmm3, %xmm0 // (float)input |
101 | movd %xmm0, 4(%esp) |
102 | flds 4(%esp) |
103 | ret |
104 | END_COMPILERRT_FUNCTION(__floatundisf) |
105 | |
106 | #endif // __i386__ |
107 | |
108 | NO_EXEC_STACK_DIRECTIVE |
109 | |
110 | |