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// du_int __umoddi3(du_int a, du_int b);
7
8// result = remainder of a / b.
9// both inputs and the output are 64-bit unsigned integers.
10// This will do whatever the underlying hardware is set to do on division by zero.
11// No other exceptions are generated, as the divide cannot overflow.
12//
13// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware
14// on x86_64.  The performance goal is ~40 cycles per divide, which is faster than
15// currently possible via simulation of integer divides on the x87 unit.
16//
17
18// Stephen Canon, December 2008
19
20#ifdef __i386__
21
22.text
23.balign 4
24DEFINE_COMPILERRT_FUNCTION(__umoddi3)
25
26	pushl		%ebx
27	movl	 20(%esp),			%ebx	// Find the index i of the leading bit in b.
28	bsrl		%ebx,			%ecx	// If the high word of b is zero, jump to
29	jz			9f						// the code to handle that special case [9].
30
31	/* High word of b is known to be non-zero on this branch */
32
33	movl	 16(%esp),			%eax	// Construct bhi, containing bits [1+i:32+i] of b
34
35	shrl		%cl,			%eax	// Practically, this means that bhi is given by:
36	shrl		%eax					//
37	notl		%ecx					//		bhi = (high word of b) << (31 - i) |
38	shll		%cl,			%ebx	//			  (low word of b) >> (1 + i)
39	orl			%eax,			%ebx	//
40	movl	 12(%esp),			%edx	// Load the high and low words of a, and jump
41	movl	  8(%esp),			%eax	// to [2] if the high word is larger than bhi
42	cmpl		%ebx,			%edx	// to avoid overflowing the upcoming divide.
43	jae			2f
44
45	/* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
46
47	divl		%ebx					// eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
48
49	pushl		%edi
50	notl		%ecx
51	shrl		%eax
52	shrl		%cl,			%eax	// q = qs >> (1 + i)
53	movl		%eax,			%edi
54	mull	 20(%esp)					// q*blo
55	movl	 12(%esp),			%ebx
56	movl	 16(%esp),			%ecx	// ECX:EBX = a
57	subl		%eax,			%ebx
58	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
59	movl	 24(%esp),			%eax
60	imull		%edi,			%eax	// q*bhi
61	subl		%eax,			%ecx	// ECX:EBX = a - q*b
62
63	jnc			1f						// if positive, this is the result.
64	addl	 20(%esp),			%ebx	// otherwise
65	adcl	 24(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
661:	movl		%ebx,			%eax
67	movl		%ecx,			%edx
68
69	popl		%edi
70	popl		%ebx
71	retl
72
73
742:	/* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
75
76	subl		%ebx,			%edx	// subtract bhi from ahi so that divide will not
77	divl		%ebx					// overflow, and find q and r such that
78										//
79										//		ahi:alo = (1:q)*bhi + r
80										//
81										// Note that q is a number in (31-i).(1+i)
82										// fix point.
83
84	pushl		%edi
85	notl		%ecx
86	shrl		%eax
87	orl			$0x80000000,	%eax
88	shrl		%cl,			%eax	// q = (1:qs) >> (1 + i)
89	movl		%eax,			%edi
90	mull	 20(%esp)					// q*blo
91	movl	 12(%esp),			%ebx
92	movl	 16(%esp),			%ecx	// ECX:EBX = a
93	subl		%eax,			%ebx
94	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
95	movl	 24(%esp),			%eax
96	imull		%edi,			%eax	// q*bhi
97	subl		%eax,			%ecx	// ECX:EBX = a - q*b
98
99	jnc			3f						// if positive, this is the result.
100	addl	 20(%esp),			%ebx	// otherwise
101	adcl	 24(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
1023:	movl		%ebx,			%eax
103	movl		%ecx,			%edx
104
105	popl		%edi
106	popl		%ebx
107	retl
108
109
110
1119:	/* High word of b is zero on this branch */
112
113	movl	 12(%esp),			%eax	// Find qhi and rhi such that
114	movl	 16(%esp),			%ecx	//
115	xorl		%edx,			%edx	//		ahi = qhi*b + rhi	with	0 ≤ rhi < b
116	divl		%ecx					//
117	movl		%eax,			%ebx	//
118	movl	  8(%esp),			%eax	// Find rlo such that
119	divl		%ecx					//
120	movl		%edx,			%eax	//		rhi:alo = qlo*b + rlo  with 0 ≤ rlo < b
121	popl		%ebx					//
122	xorl		%edx,			%edx	// and return 0:rlo
123	retl								//
124END_COMPILERRT_FUNCTION(__umoddi3)
125
126#endif // __i386__
127