xref: /netbsd-src/sys/external/bsd/compiler_rt/dist/lib/builtins/i386/moddi3.S (revision 61f2f2562dcc3e4eb50e7ec5dab0dfa1f093861a)
1156cd587Sjoerg// This file is dual licensed under the MIT and the University of Illinois Open
2156cd587Sjoerg// Source Licenses. See LICENSE.TXT for details.
3156cd587Sjoerg
4156cd587Sjoerg#include "../assembly.h"
5156cd587Sjoerg
6156cd587Sjoerg// di_int __moddi3(di_int a, di_int b);
7156cd587Sjoerg
8156cd587Sjoerg// result = remainder of a / b.
9156cd587Sjoerg// both inputs and the output are 64-bit signed integers.
10156cd587Sjoerg// This will do whatever the underlying hardware is set to do on division by zero.
11156cd587Sjoerg// No other exceptions are generated, as the divide cannot overflow.
12156cd587Sjoerg//
13156cd587Sjoerg// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware
14156cd587Sjoerg// on x86_64.  The performance goal is ~40 cycles per divide, which is faster than
15156cd587Sjoerg// currently possible via simulation of integer divides on the x87 unit.
16156cd587Sjoerg//
17156cd587Sjoerg
18156cd587Sjoerg// Stephen Canon, December 2008
19156cd587Sjoerg
20156cd587Sjoerg#ifdef __i386__
21156cd587Sjoerg
22156cd587Sjoerg.text
23*61f2f256Sjoerg.balign 4
24156cd587SjoergDEFINE_COMPILERRT_FUNCTION(__moddi3)
25156cd587Sjoerg
26156cd587Sjoerg/* This is currently implemented by wrapping the unsigned modulus up in an absolute
27156cd587Sjoerg   value.  This could certainly be improved upon. */
28156cd587Sjoerg
29156cd587Sjoerg	pushl		%esi
30156cd587Sjoerg	movl	 20(%esp),			%edx	// high word of b
31156cd587Sjoerg	movl	 16(%esp),			%eax	// low word of b
32156cd587Sjoerg	movl		%edx,			%ecx
33156cd587Sjoerg	sarl		$31,			%ecx	// (b < 0) ? -1 : 0
34156cd587Sjoerg	xorl		%ecx,			%eax
35156cd587Sjoerg	xorl		%ecx,			%edx	// EDX:EAX = (b < 0) ? not(b) : b
36156cd587Sjoerg	subl		%ecx,			%eax
37156cd587Sjoerg	sbbl		%ecx,			%edx	// EDX:EAX = abs(b)
38156cd587Sjoerg	movl		%edx,		 20(%esp)
39156cd587Sjoerg	movl		%eax,		 16(%esp)	// store abs(b) back to stack
40156cd587Sjoerg
41156cd587Sjoerg	movl	 12(%esp),			%edx	// high word of b
42156cd587Sjoerg	movl	  8(%esp),			%eax	// low word of b
43156cd587Sjoerg	movl		%edx,			%ecx
44156cd587Sjoerg	sarl		$31,			%ecx	// (a < 0) ? -1 : 0
45156cd587Sjoerg	xorl		%ecx,			%eax
46156cd587Sjoerg	xorl		%ecx,			%edx	// EDX:EAX = (a < 0) ? not(a) : a
47156cd587Sjoerg	subl		%ecx,			%eax
48156cd587Sjoerg	sbbl		%ecx,			%edx	// EDX:EAX = abs(a)
49156cd587Sjoerg	movl		%edx,		 12(%esp)
50156cd587Sjoerg	movl		%eax,		  8(%esp)	// store abs(a) back to stack
51156cd587Sjoerg	movl		%ecx,			%esi	// set aside sign of a
52156cd587Sjoerg
53156cd587Sjoerg	pushl		%ebx
54156cd587Sjoerg	movl	 24(%esp),			%ebx	// Find the index i of the leading bit in b.
55156cd587Sjoerg	bsrl		%ebx,			%ecx	// If the high word of b is zero, jump to
56156cd587Sjoerg	jz			9f						// the code to handle that special case [9].
57156cd587Sjoerg
58156cd587Sjoerg	/* High word of b is known to be non-zero on this branch */
59156cd587Sjoerg
60156cd587Sjoerg	movl	 20(%esp),			%eax	// Construct bhi, containing bits [1+i:32+i] of b
61156cd587Sjoerg
62156cd587Sjoerg	shrl		%cl,			%eax	// Practically, this means that bhi is given by:
63156cd587Sjoerg	shrl		%eax					//
64156cd587Sjoerg	notl		%ecx					//		bhi = (high word of b) << (31 - i) |
65156cd587Sjoerg	shll		%cl,			%ebx	//			  (low word of b) >> (1 + i)
66156cd587Sjoerg	orl			%eax,			%ebx	//
67156cd587Sjoerg	movl	 16(%esp),			%edx	// Load the high and low words of a, and jump
68156cd587Sjoerg	movl	 12(%esp),			%eax	// to [2] if the high word is larger than bhi
69156cd587Sjoerg	cmpl		%ebx,			%edx	// to avoid overflowing the upcoming divide.
70156cd587Sjoerg	jae			2f
71156cd587Sjoerg
72156cd587Sjoerg	/* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
73156cd587Sjoerg
74156cd587Sjoerg	divl		%ebx					// eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
75156cd587Sjoerg
76156cd587Sjoerg	pushl		%edi
77156cd587Sjoerg	notl		%ecx
78156cd587Sjoerg	shrl		%eax
79156cd587Sjoerg	shrl		%cl,			%eax	// q = qs >> (1 + i)
80156cd587Sjoerg	movl		%eax,			%edi
81156cd587Sjoerg	mull	 24(%esp)					// q*blo
82156cd587Sjoerg	movl	 16(%esp),			%ebx
83156cd587Sjoerg	movl	 20(%esp),			%ecx	// ECX:EBX = a
84156cd587Sjoerg	subl		%eax,			%ebx
85156cd587Sjoerg	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
86156cd587Sjoerg	movl	 28(%esp),			%eax
87156cd587Sjoerg	imull		%edi,			%eax	// q*bhi
88156cd587Sjoerg	subl		%eax,			%ecx	// ECX:EBX = a - q*b
89156cd587Sjoerg
90156cd587Sjoerg	jnc			1f						// if positive, this is the result.
91156cd587Sjoerg	addl	 24(%esp),			%ebx	// otherwise
92156cd587Sjoerg	adcl	 28(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
93156cd587Sjoerg1:	movl		%ebx,			%eax
94156cd587Sjoerg	movl		%ecx,			%edx
95156cd587Sjoerg
96156cd587Sjoerg	addl		%esi,			%eax	// Restore correct sign to result
97156cd587Sjoerg	adcl		%esi,			%edx
98156cd587Sjoerg	xorl		%esi,			%eax
99156cd587Sjoerg	xorl		%esi,			%edx
100156cd587Sjoerg	popl		%edi					// Restore callee-save registers
101156cd587Sjoerg	popl		%ebx
102156cd587Sjoerg	popl		%esi
103156cd587Sjoerg	retl								// Return
104156cd587Sjoerg
105156cd587Sjoerg2:	/* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
106156cd587Sjoerg
107156cd587Sjoerg	subl		%ebx,			%edx	// subtract bhi from ahi so that divide will not
108156cd587Sjoerg	divl		%ebx					// overflow, and find q and r such that
109156cd587Sjoerg										//
110156cd587Sjoerg										//		ahi:alo = (1:q)*bhi + r
111156cd587Sjoerg										//
112156cd587Sjoerg										// Note that q is a number in (31-i).(1+i)
113156cd587Sjoerg										// fix point.
114156cd587Sjoerg
115156cd587Sjoerg	pushl		%edi
116156cd587Sjoerg	notl		%ecx
117156cd587Sjoerg	shrl		%eax
118156cd587Sjoerg	orl			$0x80000000,	%eax
119156cd587Sjoerg	shrl		%cl,			%eax	// q = (1:qs) >> (1 + i)
120156cd587Sjoerg	movl		%eax,			%edi
121156cd587Sjoerg	mull	 24(%esp)					// q*blo
122156cd587Sjoerg	movl	 16(%esp),			%ebx
123156cd587Sjoerg	movl	 20(%esp),			%ecx	// ECX:EBX = a
124156cd587Sjoerg	subl		%eax,			%ebx
125156cd587Sjoerg	sbbl		%edx,			%ecx	// ECX:EBX = a - q*blo
126156cd587Sjoerg	movl	 28(%esp),			%eax
127156cd587Sjoerg	imull		%edi,			%eax	// q*bhi
128156cd587Sjoerg	subl		%eax,			%ecx	// ECX:EBX = a - q*b
129156cd587Sjoerg
130156cd587Sjoerg	jnc			3f						// if positive, this is the result.
131156cd587Sjoerg	addl	 24(%esp),			%ebx	// otherwise
132156cd587Sjoerg	adcl	 28(%esp),			%ecx	// ECX:EBX = a - (q-1)*b = result
133156cd587Sjoerg3:	movl		%ebx,			%eax
134156cd587Sjoerg	movl		%ecx,			%edx
135156cd587Sjoerg
136156cd587Sjoerg	addl		%esi,			%eax	// Restore correct sign to result
137156cd587Sjoerg	adcl		%esi,			%edx
138156cd587Sjoerg	xorl		%esi,			%eax
139156cd587Sjoerg	xorl		%esi,			%edx
140156cd587Sjoerg	popl		%edi					// Restore callee-save registers
141156cd587Sjoerg	popl		%ebx
142156cd587Sjoerg	popl		%esi
143156cd587Sjoerg	retl								// Return
144156cd587Sjoerg
145156cd587Sjoerg9:	/* High word of b is zero on this branch */
146156cd587Sjoerg
147156cd587Sjoerg	movl	 16(%esp),			%eax	// Find qhi and rhi such that
148156cd587Sjoerg	movl	 20(%esp),			%ecx	//
149156cd587Sjoerg	xorl		%edx,			%edx	//		ahi = qhi*b + rhi	with	0 ≤ rhi < b
150156cd587Sjoerg	divl		%ecx					//
151156cd587Sjoerg	movl		%eax,			%ebx	//
152156cd587Sjoerg	movl	 12(%esp),			%eax	// Find rlo such that
153156cd587Sjoerg	divl		%ecx					//
154156cd587Sjoerg	movl		%edx,			%eax	//		rhi:alo = qlo*b + rlo  with 0 ≤ rlo < b
155156cd587Sjoerg	popl		%ebx					//
156156cd587Sjoerg	xorl		%edx,			%edx	// and return 0:rlo
157156cd587Sjoerg
158156cd587Sjoerg	addl		%esi,			%eax	// Restore correct sign to result
159156cd587Sjoerg	adcl		%esi,			%edx
160156cd587Sjoerg	xorl		%esi,			%eax
161156cd587Sjoerg	xorl		%esi,			%edx
162156cd587Sjoerg	popl		%esi
163156cd587Sjoerg	retl								// Return
164156cd587SjoergEND_COMPILERRT_FUNCTION(__moddi3)
165156cd587Sjoerg
166156cd587Sjoerg#endif // __i386__
167