xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/i386/udivdi3.S (revision 43a5ec4eb41567cc92586503212743d89686d78f)
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 __udivdi3(du_int a, du_int b);
8
9// result = 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// Stephen Canon, December 2008
19
20#ifdef __i386__
21
22.text
23.balign 4
24DEFINE_COMPILERRT_FUNCTION(__udivdi3)
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 [1] if the high word is larger than bhi
42	cmpl		%ebx,			%edx	// to avoid overflowing the upcoming divide.
43	jae			1f
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	sbbl		$0,				%edi	// decrement q if remainder is negative
63	xorl		%edx,			%edx
64	movl		%edi,			%eax
65	popl		%edi
66	popl		%ebx
67	retl
68
69
701:	// High word of a is greater than or equal to (b >> (1 + i)) on this branch
71
72	subl		%ebx,			%edx	// subtract bhi from ahi so that divide will not
73	divl		%ebx					// overflow, and find q and r such that
74										//
75										//		ahi:alo = (1:q)*bhi + r
76										//
77										// Note that q is a number in (31-i).(1+i)
78										// fix point.
79
80	pushl		%edi
81	notl		%ecx
82	shrl		%eax
83	orl			$0x80000000,	%eax
84	shrl		%cl,			%eax	// q = (1:qs) >> (1 + i)
85	movl		%eax,			%edi
86	mull	 20(%esp)					// q*blo
87	movl	 12(%esp),			%ebx
88	movl	 16(%esp),			%ecx	// ECX:EBX = a
89	subl		%eax,			%ebx
90	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
91	movl	 24(%esp),			%eax
92	imull		%edi,			%eax	// q*bhi
93	subl		%eax,			%ecx	// ECX:EBX = a - q*b
94	sbbl		$0,				%edi	// decrement q if remainder is negative
95	xorl		%edx,			%edx
96	movl		%edi,			%eax
97	popl		%edi
98	popl		%ebx
99	retl
100
101
1029:	// High word of b is zero on this branch
103
104	movl	 12(%esp),			%eax	// Find qhi and rhi such that
105	movl	 16(%esp),			%ecx	//
106	xorl		%edx,			%edx	//		ahi = qhi*b + rhi	with	0 ≤ rhi < b
107	divl		%ecx					//
108	movl		%eax,			%ebx	//
109	movl	  8(%esp),			%eax	// Find qlo such that
110	divl		%ecx					//
111	movl		%ebx,			%edx	//		rhi:alo = qlo*b + rlo  with 0 ≤ rlo < b
112	popl		%ebx					//
113	retl								// and return qhi:qlo
114END_COMPILERRT_FUNCTION(__udivdi3)
115
116#endif // __i386__
117
118NO_EXEC_STACK_DIRECTIVE
119
120