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
5 #include "../assembly.h"
7 // du_int __udivdi3(du_int a, du_int 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.
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.
18 // Stephen Canon, December 2008
24 DEFINE_COMPILERRT_FUNCTION(__udivdi3)
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].
31 // High word of b is known to be non-zero on this branch
33 movl 16(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b
35 shrl %cl, %eax // Practically, this means that bhi is given by:
37 notl %ecx // bhi = (high word of b) << (31 - i) |
38 shll %cl, %ebx // (low word of b) >> (1 + i)
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.
45 // High word of a is greater than or equal to (b >> (1 + i)) on this branch
47 divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
52 shrl %cl, %eax // q = qs >> (1 + i)
54 mull 20(%esp) // q*blo
56 movl 16(%esp), %ecx // ECX:EBX = a
58 sbbl %edx, %ecx // ECX:EBX = a - q*blo
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
70 1: // High word of a is greater than or equal to (b >> (1 + i)) on this branch
72 subl %ebx, %edx // subtract bhi from ahi so that divide will not
73 divl %ebx // overflow, and find q and r such that
75 // ahi:alo = (1:q)*bhi + r
77 // Note that q is a number in (31-i).(1+i)
84 shrl %cl, %eax // q = (1:qs) >> (1 + i)
86 mull 20(%esp) // q*blo
88 movl 16(%esp), %ecx // ECX:EBX = a
90 sbbl %edx, %ecx // ECX:EBX = a - q*blo
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
102 9: // High word of b is zero on this branch
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
109 movl 8(%esp), %eax // Find qlo such that
111 movl %ebx, %edx // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b
113 retl // and return qhi:qlo
114 END_COMPILERRT_FUNCTION(__udivdi3)
118 NO_EXEC_STACK_DIRECTIVE