xref: /minix3/sys/external/bsd/compiler_rt/dist/lib/builtins/i386/floatundisf.S (revision 0a6a1f1d05b60e214de2f05a7310ddd1f0e590e7)
1*0a6a1f1dSLionel Sambuc// This file is dual licensed under the MIT and the University of Illinois Open
2*0a6a1f1dSLionel Sambuc// Source Licenses. See LICENSE.TXT for details.
3*0a6a1f1dSLionel Sambuc
4*0a6a1f1dSLionel Sambuc#include "../assembly.h"
5*0a6a1f1dSLionel Sambuc
6*0a6a1f1dSLionel Sambuc// float __floatundisf(du_int a);
7*0a6a1f1dSLionel Sambuc
8*0a6a1f1dSLionel Sambuc// Note that there is a hardware instruction, fildll, that does most of what
9*0a6a1f1dSLionel Sambuc// this function needs to do.  However, because of our ia32 ABI, it will take
10*0a6a1f1dSLionel Sambuc// a write-small read-large stall, so the software implementation here is
11*0a6a1f1dSLionel Sambuc// actually several cycles faster.
12*0a6a1f1dSLionel Sambuc
13*0a6a1f1dSLionel Sambuc// This is a branch-free implementation.  A branchy implementation might be
14*0a6a1f1dSLionel Sambuc// faster for the common case if you know something a priori about the input
15*0a6a1f1dSLionel Sambuc// distribution.
16*0a6a1f1dSLionel Sambuc
17*0a6a1f1dSLionel Sambuc/* branch-free x87 implementation - one cycle slower than without x87.
18*0a6a1f1dSLionel Sambuc
19*0a6a1f1dSLionel Sambuc#ifdef __i386__
20*0a6a1f1dSLionel Sambuc
21*0a6a1f1dSLionel Sambuc.const
22*0a6a1f1dSLionel Sambuc.balign 3
23*0a6a1f1dSLionel Sambuc
24*0a6a1f1dSLionel Sambuc		.quad	0x43f0000000000000
25*0a6a1f1dSLionel Sambuctwop64:	.quad	0x0000000000000000
26*0a6a1f1dSLionel Sambuc
27*0a6a1f1dSLionel Sambuc#define			TWOp64			twop64-0b(%ecx,%eax,8)
28*0a6a1f1dSLionel Sambuc
29*0a6a1f1dSLionel Sambuc.text
30*0a6a1f1dSLionel Sambuc.balign 4
31*0a6a1f1dSLionel SambucDEFINE_COMPILERRT_FUNCTION(__floatundisf)
32*0a6a1f1dSLionel Sambuc	movl		8(%esp),		%eax
33*0a6a1f1dSLionel Sambuc	movd		8(%esp),		%xmm1
34*0a6a1f1dSLionel Sambuc	movd		4(%esp),		%xmm0
35*0a6a1f1dSLionel Sambuc	punpckldq	%xmm1,			%xmm0
36*0a6a1f1dSLionel Sambuc	calll		0f
37*0a6a1f1dSLionel Sambuc0:	popl		%ecx
38*0a6a1f1dSLionel Sambuc	sarl		$31,			%eax
39*0a6a1f1dSLionel Sambuc	movq		%xmm0,			4(%esp)
40*0a6a1f1dSLionel Sambuc	fildll		4(%esp)
41*0a6a1f1dSLionel Sambuc	faddl		TWOp64
42*0a6a1f1dSLionel Sambuc	fstps		4(%esp)
43*0a6a1f1dSLionel Sambuc	flds		4(%esp)
44*0a6a1f1dSLionel Sambuc	ret
45*0a6a1f1dSLionel SambucEND_COMPILERRT_FUNCTION(__floatundisf)
46*0a6a1f1dSLionel Sambuc
47*0a6a1f1dSLionel Sambuc#endif // __i386__
48*0a6a1f1dSLionel Sambuc
49*0a6a1f1dSLionel Sambuc*/
50*0a6a1f1dSLionel Sambuc
51*0a6a1f1dSLionel Sambuc/* branch-free, x87-free implementation - faster at the expense of code size */
52*0a6a1f1dSLionel Sambuc
53*0a6a1f1dSLionel Sambuc#ifdef __i386__
54*0a6a1f1dSLionel Sambuc
55*0a6a1f1dSLionel Sambuc#if defined(__APPLE__)
56*0a6a1f1dSLionel Sambuc	.const
57*0a6a1f1dSLionel Sambuc#elif defined(__ELF__)
58*0a6a1f1dSLionel Sambuc	.section .rodata
59*0a6a1f1dSLionel Sambuc#else
60*0a6a1f1dSLionel Sambuc	.section .rdata,"rd"
61*0a6a1f1dSLionel Sambuc#endif
62*0a6a1f1dSLionel Sambuc
63*0a6a1f1dSLionel Sambuc	.balign 16
64*0a6a1f1dSLionel Sambuctwop52:
65*0a6a1f1dSLionel Sambuc	.quad 0x4330000000000000
66*0a6a1f1dSLionel Sambuc	.quad 0x0000000000000fff
67*0a6a1f1dSLionel Sambuc
68*0a6a1f1dSLionel Sambuc	.balign 16
69*0a6a1f1dSLionel Sambucsticky:
70*0a6a1f1dSLionel Sambuc	.quad 0x0000000000000000
71*0a6a1f1dSLionel Sambuc	.long 0x00000012
72*0a6a1f1dSLionel Sambuc
73*0a6a1f1dSLionel Sambuc	.balign 16
74*0a6a1f1dSLionel Sambuctwelve:
75*0a6a1f1dSLionel Sambuc	.long 0x00000000
76*0a6a1f1dSLionel Sambuc
77*0a6a1f1dSLionel Sambuc#define			TWOp52			twop52-0b(%ecx)
78*0a6a1f1dSLionel Sambuc#define			STICKY			sticky-0b(%ecx,%eax,8)
79*0a6a1f1dSLionel Sambuc
80*0a6a1f1dSLionel Sambuc.text
81*0a6a1f1dSLionel Sambuc.balign 4
82*0a6a1f1dSLionel SambucDEFINE_COMPILERRT_FUNCTION(__floatundisf)
83*0a6a1f1dSLionel Sambuc	movl		8(%esp),		%eax
84*0a6a1f1dSLionel Sambuc	movd		8(%esp),		%xmm1
85*0a6a1f1dSLionel Sambuc	movd		4(%esp),		%xmm0
86*0a6a1f1dSLionel Sambuc	punpckldq	%xmm1,			%xmm0
87*0a6a1f1dSLionel Sambuc
88*0a6a1f1dSLionel Sambuc	calll		0f
89*0a6a1f1dSLionel Sambuc0:	popl		%ecx
90*0a6a1f1dSLionel Sambuc	shrl		%eax					// high 31 bits of input as sint32
91*0a6a1f1dSLionel Sambuc	addl		$0x7ff80000,	%eax
92*0a6a1f1dSLionel Sambuc	sarl		$31,			%eax	// (big input) ? -1 : 0
93*0a6a1f1dSLionel Sambuc	movsd		STICKY,			%xmm1	// (big input) ? 0xfff : 0
94*0a6a1f1dSLionel Sambuc	movl		$12,			%edx
95*0a6a1f1dSLionel Sambuc	andl		%eax,			%edx	// (big input) ? 12 : 0
96*0a6a1f1dSLionel Sambuc	movd		%edx,			%xmm3
97*0a6a1f1dSLionel Sambuc	andpd		%xmm0,			%xmm1	// (big input) ? input & 0xfff : 0
98*0a6a1f1dSLionel Sambuc	movsd		TWOp52,			%xmm2	// 0x1.0p52
99*0a6a1f1dSLionel Sambuc	psrlq		%xmm3,			%xmm0	// (big input) ? input >> 12 : input
100*0a6a1f1dSLionel Sambuc	orpd		%xmm2,			%xmm1	// 0x1.0p52 + ((big input) ? input & 0xfff : input)
101*0a6a1f1dSLionel Sambuc	orpd		%xmm1,			%xmm0	// 0x1.0p52 + ((big input) ? (input >> 12 | input & 0xfff) : input)
102*0a6a1f1dSLionel Sambuc	subsd		%xmm2,			%xmm0	// (double)((big input) ? (input >> 12 | input & 0xfff) : input)
103*0a6a1f1dSLionel Sambuc	cvtsd2ss	%xmm0,			%xmm0	// (float)((big input) ? (input >> 12 | input & 0xfff) : input)
104*0a6a1f1dSLionel Sambuc	pslld		$23,			%xmm3
105*0a6a1f1dSLionel Sambuc	paddd		%xmm3,			%xmm0	// (float)input
106*0a6a1f1dSLionel Sambuc	movd		%xmm0,			4(%esp)
107*0a6a1f1dSLionel Sambuc	flds		4(%esp)
108*0a6a1f1dSLionel Sambuc	ret
109*0a6a1f1dSLionel SambucEND_COMPILERRT_FUNCTION(__floatundisf)
110*0a6a1f1dSLionel Sambuc
111*0a6a1f1dSLionel Sambuc#endif // __i386__
112