xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/i386/moddi3.S (revision b077aed33b7b6aefca7b17ddb250cf521f938613)
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// di_int __moddi3(di_int a, di_int b);
8
9// result = remainder of a / b.
10// both inputs and the output are 64-bit signed 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(__moddi3)
26
27// This is currently implemented by wrapping the unsigned modulus up in an absolute
28// value.  This could certainly be improved upon.
29
30	pushl		%esi
31	movl	 20(%esp),			%edx	// high word of b
32	movl	 16(%esp),			%eax	// low word of b
33	movl		%edx,			%ecx
34	sarl		$31,			%ecx	// (b < 0) ? -1 : 0
35	xorl		%ecx,			%eax
36	xorl		%ecx,			%edx	// EDX:EAX = (b < 0) ? not(b) : b
37	subl		%ecx,			%eax
38	sbbl		%ecx,			%edx	// EDX:EAX = abs(b)
39	movl		%edx,		 20(%esp)
40	movl		%eax,		 16(%esp)	// store abs(b) back to stack
41
42	movl	 12(%esp),			%edx	// high word of b
43	movl	  8(%esp),			%eax	// low word of b
44	movl		%edx,			%ecx
45	sarl		$31,			%ecx	// (a < 0) ? -1 : 0
46	xorl		%ecx,			%eax
47	xorl		%ecx,			%edx	// EDX:EAX = (a < 0) ? not(a) : a
48	subl		%ecx,			%eax
49	sbbl		%ecx,			%edx	// EDX:EAX = abs(a)
50	movl		%edx,		 12(%esp)
51	movl		%eax,		  8(%esp)	// store abs(a) back to stack
52	movl		%ecx,			%esi	// set aside sign of a
53
54	pushl		%ebx
55	movl	 24(%esp),			%ebx	// Find the index i of the leading bit in b.
56	bsrl		%ebx,			%ecx	// If the high word of b is zero, jump to
57	jz			9f						// the code to handle that special case [9].
58
59	// High word of b is known to be non-zero on this branch
60
61	movl	 20(%esp),			%eax	// Construct bhi, containing bits [1+i:32+i] of b
62
63	shrl		%cl,			%eax	// Practically, this means that bhi is given by:
64	shrl		%eax					//
65	notl		%ecx					//		bhi = (high word of b) << (31 - i) |
66	shll		%cl,			%ebx	//			  (low word of b) >> (1 + i)
67	orl			%eax,			%ebx	//
68	movl	 16(%esp),			%edx	// Load the high and low words of a, and jump
69	movl	 12(%esp),			%eax	// to [2] if the high word is larger than bhi
70	cmpl		%ebx,			%edx	// to avoid overflowing the upcoming divide.
71	jae			2f
72
73	// High word of a is greater than or equal to (b >> (1 + i)) on this branch
74
75	divl		%ebx					// eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
76
77	pushl		%edi
78	notl		%ecx
79	shrl		%eax
80	shrl		%cl,			%eax	// q = qs >> (1 + i)
81	movl		%eax,			%edi
82	mull	 24(%esp)					// q*blo
83	movl	 16(%esp),			%ebx
84	movl	 20(%esp),			%ecx	// ECX:EBX = a
85	subl		%eax,			%ebx
86	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
87	movl	 28(%esp),			%eax
88	imull		%edi,			%eax	// q*bhi
89	subl		%eax,			%ecx	// ECX:EBX = a - q*b
90
91	jnc			1f						// if positive, this is the result.
92	addl	 24(%esp),			%ebx	// otherwise
93	adcl	 28(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
941:	movl		%ebx,			%eax
95	movl		%ecx,			%edx
96
97	addl		%esi,			%eax	// Restore correct sign to result
98	adcl		%esi,			%edx
99	xorl		%esi,			%eax
100	xorl		%esi,			%edx
101	popl		%edi					// Restore callee-save registers
102	popl		%ebx
103	popl		%esi
104	retl								// Return
105
1062:	// High word of a is greater than or equal to (b >> (1 + i)) on this branch
107
108	subl		%ebx,			%edx	// subtract bhi from ahi so that divide will not
109	divl		%ebx					// overflow, and find q and r such that
110										//
111										//		ahi:alo = (1:q)*bhi + r
112										//
113										// Note that q is a number in (31-i).(1+i)
114										// fix point.
115
116	pushl		%edi
117	notl		%ecx
118	shrl		%eax
119	orl			$0x80000000,	%eax
120	shrl		%cl,			%eax	// q = (1:qs) >> (1 + i)
121	movl		%eax,			%edi
122	mull	 24(%esp)					// q*blo
123	movl	 16(%esp),			%ebx
124	movl	 20(%esp),			%ecx	// ECX:EBX = a
125	subl		%eax,			%ebx
126	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
127	movl	 28(%esp),			%eax
128	imull		%edi,			%eax	// q*bhi
129	subl		%eax,			%ecx	// ECX:EBX = a - q*b
130
131	jnc			3f						// if positive, this is the result.
132	addl	 24(%esp),			%ebx	// otherwise
133	adcl	 28(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
1343:	movl		%ebx,			%eax
135	movl		%ecx,			%edx
136
137	addl		%esi,			%eax	// Restore correct sign to result
138	adcl		%esi,			%edx
139	xorl		%esi,			%eax
140	xorl		%esi,			%edx
141	popl		%edi					// Restore callee-save registers
142	popl		%ebx
143	popl		%esi
144	retl								// Return
145
1469:	// High word of b is zero on this branch
147
148	movl	 16(%esp),			%eax	// Find qhi and rhi such that
149	movl	 20(%esp),			%ecx	//
150	xorl		%edx,			%edx	//		ahi = qhi*b + rhi	with	0 ≤ rhi < b
151	divl		%ecx					//
152	movl		%eax,			%ebx	//
153	movl	 12(%esp),			%eax	// Find rlo such that
154	divl		%ecx					//
155	movl		%edx,			%eax	//		rhi:alo = qlo*b + rlo  with 0 ≤ rlo < b
156	popl		%ebx					//
157	xorl		%edx,			%edx	// and return 0:rlo
158
159	addl		%esi,			%eax	// Restore correct sign to result
160	adcl		%esi,			%edx
161	xorl		%esi,			%eax
162	xorl		%esi,			%edx
163	popl		%esi
164	retl								// Return
165END_COMPILERRT_FUNCTION(__moddi3)
166
167#endif // __i386__
168
169NO_EXEC_STACK_DIRECTIVE
170
171