xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/x86_64/floatundixf.S (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
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// xf_float __floatundixf(du_int a);
8
9#ifdef __x86_64__
10
11CONST_SECTION
12
13	.balign 16
14twop64:
15	.quad 0x43f0000000000000
16
17#define REL_ADDR(_a)	(_a)(%rip)
18
19	.text
20
21	.balign 4
22DEFINE_COMPILERRT_FUNCTION(__floatundixf)
23	movq	%rdi,	 -8(%rsp)
24	fildq	-8(%rsp)
25	test	%rdi,		%rdi
26	js		1f
27	ret
281:	faddl	REL_ADDR(twop64)
29	ret
30END_COMPILERRT_FUNCTION(__floatundixf)
31
32#endif // __x86_64__
33
34
35/* Branch-free implementation is ever so slightly slower, but more beautiful.
36   It is likely superior for inlining, so I kept it around for future reference.
37
38#ifdef __x86_64__
39
40CONST_SECTION
41
42	.balign 4
43twop52:
44	.quad 0x4330000000000000
45twop84_plus_twop52_neg:
46	.quad 0xc530000000100000
47twop84:
48	.quad 0x4530000000000000
49
50#define REL_ADDR(_a)	(_a)(%rip)
51
52.text
53.balign 4
54DEFINE_COMPILERRT_FUNCTION(__floatundixf)
55	movl	%edi,				%esi			// low 32 bits of input
56	shrq	$32,				%rdi			// hi 32 bits of input
57	orq		REL_ADDR(twop84),	%rdi			// 2^84 + hi (as a double)
58	orq		REL_ADDR(twop52),	%rsi			// 2^52 + lo (as a double)
59	movq	%rdi,			 -8(%rsp)
60	movq	%rsi,			-16(%rsp)
61	fldl	REL_ADDR(twop84_plus_twop52_neg)
62	faddl	-8(%rsp)	// hi - 2^52 (as double extended, no rounding occurs)
63	faddl	-16(%rsp)	// hi + lo (as double extended)
64	ret
65END_COMPILERRT_FUNCTION(__floatundixf)
66
67#endif // __x86_64__
68
69*/
70
71NO_EXEC_STACK_DIRECTIVE
72
73