1// 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
21CONST_SECTION
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
55CONST_SECTION
56
57	.balign 16
58twop52:
59	.quad 0x4330000000000000
60	.quad 0x0000000000000fff
61
62	.balign 16
63sticky:
64	.quad 0x0000000000000000
65	.long 0x00000012
66
67	.balign 16
68twelve:
69	.long 0x00000000
70
71#define			TWOp52			twop52-0b(%ecx)
72#define			STICKY			sticky-0b(%ecx,%eax,8)
73
74.text
75.balign 4
76DEFINE_COMPILERRT_FUNCTION(__floatundisf)
77	movl		8(%esp),		%eax
78	movd		8(%esp),		%xmm1
79	movd		4(%esp),		%xmm0
80	punpckldq	%xmm1,			%xmm0
81
82	calll		0f
830:	popl		%ecx
84	shrl		%eax					// high 31 bits of input as sint32
85	addl		$0x7ff80000,	%eax
86	sarl		$31,			%eax	// (big input) ? -1 : 0
87	movsd		STICKY,			%xmm1	// (big input) ? 0xfff : 0
88	movl		$12,			%edx
89	andl		%eax,			%edx	// (big input) ? 12 : 0
90	movd		%edx,			%xmm3
91	andpd		%xmm0,			%xmm1	// (big input) ? input & 0xfff : 0
92	movsd		TWOp52,			%xmm2	// 0x1.0p52
93	psrlq		%xmm3,			%xmm0	// (big input) ? input >> 12 : input
94	orpd		%xmm2,			%xmm1	// 0x1.0p52 + ((big input) ? input & 0xfff : input)
95	orpd		%xmm1,			%xmm0	// 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input)
96	subsd		%xmm2,			%xmm0	// (double)((big input) ? (input >> 12 | input & 0xfff) : input)
97	cvtsd2ss	%xmm0,			%xmm0	// (float)((big input) ? (input >> 12 | input & 0xfff) : input)
98	pslld		$23,			%xmm3
99	paddd		%xmm3,			%xmm0	// (float)input
100	movd		%xmm0,			4(%esp)
101	flds		4(%esp)
102	ret
103END_COMPILERRT_FUNCTION(__floatundisf)
104
105#endif // __i386__
106
107NO_EXEC_STACK_DIRECTIVE
108
109