xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/i386/umoddi3.S (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
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// du_int __umoddi3(du_int a, du_int b);
8
9// result = remainder of a / b.
10// both inputs and the output are 64-bit unsigned integers.
11// This will do whatever the underlying hardware is set to do on division by zero.
12// No other exceptions are generated, as the divide cannot overflow.
13//
14// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware
15// on x86_64.  The performance goal is ~40 cycles per divide, which is faster than
16// currently possible via simulation of integer divides on the x87 unit.
17//
18
19// Stephen Canon, December 2008
20
21#ifdef __i386__
22
23.text
24.balign 4
25DEFINE_COMPILERRT_FUNCTION(__umoddi3)
26
27	pushl		%ebx
28	movl	 20(%esp),			%ebx	// Find the index i of the leading bit in b.
29	bsrl		%ebx,			%ecx	// If the high word of b is zero, jump to
30	jz			9f						// the code to handle that special case [9].
31
32	// High word of b is known to be non-zero on this branch
33
34	movl	 16(%esp),			%eax	// Construct bhi, containing bits [1+i:32+i] of b
35
36	shrl		%cl,			%eax	// Practically, this means that bhi is given by:
37	shrl		%eax					//
38	notl		%ecx					//		bhi = (high word of b) << (31 - i) |
39	shll		%cl,			%ebx	//			  (low word of b) >> (1 + i)
40	orl			%eax,			%ebx	//
41	movl	 12(%esp),			%edx	// Load the high and low words of a, and jump
42	movl	  8(%esp),			%eax	// to [2] if the high word is larger than bhi
43	cmpl		%ebx,			%edx	// to avoid overflowing the upcoming divide.
44	jae			2f
45
46	// High word of a is greater than or equal to (b >> (1 + i)) on this branch
47
48	divl		%ebx					// eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
49
50	pushl		%edi
51	notl		%ecx
52	shrl		%eax
53	shrl		%cl,			%eax	// q = qs >> (1 + i)
54	movl		%eax,			%edi
55	mull	 20(%esp)					// q*blo
56	movl	 12(%esp),			%ebx
57	movl	 16(%esp),			%ecx	// ECX:EBX = a
58	subl		%eax,			%ebx
59	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
60	movl	 24(%esp),			%eax
61	imull		%edi,			%eax	// q*bhi
62	subl		%eax,			%ecx	// ECX:EBX = a - q*b
63
64	jnc			1f						// if positive, this is the result.
65	addl	 20(%esp),			%ebx	// otherwise
66	adcl	 24(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
671:	movl		%ebx,			%eax
68	movl		%ecx,			%edx
69
70	popl		%edi
71	popl		%ebx
72	retl
73
74
752:	// High word of a is greater than or equal to (b >> (1 + i)) on this branch
76
77	subl		%ebx,			%edx	// subtract bhi from ahi so that divide will not
78	divl		%ebx					// overflow, and find q and r such that
79										//
80										//		ahi:alo = (1:q)*bhi + r
81										//
82										// Note that q is a number in (31-i).(1+i)
83										// fix point.
84
85	pushl		%edi
86	notl		%ecx
87	shrl		%eax
88	orl			$0x80000000,	%eax
89	shrl		%cl,			%eax	// q = (1:qs) >> (1 + i)
90	movl		%eax,			%edi
91	mull	 20(%esp)					// q*blo
92	movl	 12(%esp),			%ebx
93	movl	 16(%esp),			%ecx	// ECX:EBX = a
94	subl		%eax,			%ebx
95	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
96	movl	 24(%esp),			%eax
97	imull		%edi,			%eax	// q*bhi
98	subl		%eax,			%ecx	// ECX:EBX = a - q*b
99
100	jnc			3f						// if positive, this is the result.
101	addl	 20(%esp),			%ebx	// otherwise
102	adcl	 24(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
1033:	movl		%ebx,			%eax
104	movl		%ecx,			%edx
105
106	popl		%edi
107	popl		%ebx
108	retl
109
110
111
1129:	// High word of b is zero on this branch
113
114	movl	 12(%esp),			%eax	// Find qhi and rhi such that
115	movl	 16(%esp),			%ecx	//
116	xorl		%edx,			%edx	//		ahi = qhi*b + rhi	with	0 ≤ rhi < b
117	divl		%ecx					//
118	movl		%eax,			%ebx	//
119	movl	  8(%esp),			%eax	// Find rlo such that
120	divl		%ecx					//
121	movl		%edx,			%eax	//		rhi:alo = qlo*b + rlo  with 0 ≤ rlo < b
122	popl		%ebx					//
123	xorl		%edx,			%edx	// and return 0:rlo
124	retl								//
125END_COMPILERRT_FUNCTION(__umoddi3)
126
127#endif // __i386__
128
129NO_EXEC_STACK_DIRECTIVE
130
131