1222656Sed// This file is dual licensed under the MIT and the University of Illinois Open 2222656Sed// Source Licenses. See LICENSE.TXT for details. 3214152Sed 4214152Sed#include "../assembly.h" 5214152Sed 6214152Sed// du_int __udivdi3(du_int a, du_int b); 7214152Sed 8214152Sed// result = a / b. 9214152Sed// both inputs and the output are 64-bit unsigned integers. 10214152Sed// This will do whatever the underlying hardware is set to do on division by zero. 11214152Sed// No other exceptions are generated, as the divide cannot overflow. 12214152Sed// 13214152Sed// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware 14214152Sed// on x86_64. The performance goal is ~40 cycles per divide, which is faster than 15214152Sed// currently possible via simulation of integer divides on the x87 unit. 16214152Sed// 17214152Sed// Stephen Canon, December 2008 18214152Sed 19214152Sed#ifdef __i386__ 20214152Sed 21214152Sed.text 22214152Sed.align 4 23214152SedDEFINE_COMPILERRT_FUNCTION(__udivdi3) 24214152Sed 25214152Sed pushl %ebx 26214152Sed movl 20(%esp), %ebx // Find the index i of the leading bit in b. 27214152Sed bsrl %ebx, %ecx // If the high word of b is zero, jump to 28214152Sed jz 9f // the code to handle that special case [9]. 29214152Sed 30214152Sed /* High word of b is known to be non-zero on this branch */ 31214152Sed 32214152Sed movl 16(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b 33214152Sed 34214152Sed shrl %cl, %eax // Practically, this means that bhi is given by: 35214152Sed shrl %eax // 36214152Sed notl %ecx // bhi = (high word of b) << (31 - i) | 37214152Sed shll %cl, %ebx // (low word of b) >> (1 + i) 38214152Sed orl %eax, %ebx // 39214152Sed movl 12(%esp), %edx // Load the high and low words of a, and jump 40214152Sed movl 8(%esp), %eax // to [1] if the high word is larger than bhi 41214152Sed cmpl %ebx, %edx // to avoid overflowing the upcoming divide. 42214152Sed jae 1f 43214152Sed 44214152Sed /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 45214152Sed 46214152Sed divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r 47214152Sed 48214152Sed pushl %edi 49214152Sed notl %ecx 50214152Sed shrl %eax 51214152Sed shrl %cl, %eax // q = qs >> (1 + i) 52214152Sed movl %eax, %edi 53214152Sed mull 20(%esp) // q*blo 54214152Sed movl 12(%esp), %ebx 55214152Sed movl 16(%esp), %ecx // ECX:EBX = a 56214152Sed subl %eax, %ebx 57214152Sed sbbl %edx, %ecx // ECX:EBX = a - q*blo 58214152Sed movl 24(%esp), %eax 59214152Sed imull %edi, %eax // q*bhi 60214152Sed subl %eax, %ecx // ECX:EBX = a - q*b 61214152Sed sbbl $0, %edi // decrement q if remainder is negative 62214152Sed xorl %edx, %edx 63214152Sed movl %edi, %eax 64214152Sed popl %edi 65214152Sed popl %ebx 66214152Sed retl 67214152Sed 68214152Sed 69214152Sed1: /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 70214152Sed 71214152Sed subl %ebx, %edx // subtract bhi from ahi so that divide will not 72214152Sed divl %ebx // overflow, and find q and r such that 73214152Sed // 74214152Sed // ahi:alo = (1:q)*bhi + r 75214152Sed // 76214152Sed // Note that q is a number in (31-i).(1+i) 77214152Sed // fix point. 78214152Sed 79214152Sed pushl %edi 80214152Sed notl %ecx 81214152Sed shrl %eax 82214152Sed orl $0x80000000, %eax 83214152Sed shrl %cl, %eax // q = (1:qs) >> (1 + i) 84214152Sed movl %eax, %edi 85214152Sed mull 20(%esp) // q*blo 86214152Sed movl 12(%esp), %ebx 87214152Sed movl 16(%esp), %ecx // ECX:EBX = a 88214152Sed subl %eax, %ebx 89214152Sed sbbl %edx, %ecx // ECX:EBX = a - q*blo 90214152Sed movl 24(%esp), %eax 91214152Sed imull %edi, %eax // q*bhi 92214152Sed subl %eax, %ecx // ECX:EBX = a - q*b 93214152Sed sbbl $0, %edi // decrement q if remainder is negative 94214152Sed xorl %edx, %edx 95214152Sed movl %edi, %eax 96214152Sed popl %edi 97214152Sed popl %ebx 98214152Sed retl 99214152Sed 100214152Sed 101214152Sed9: /* High word of b is zero on this branch */ 102214152Sed 103214152Sed movl 12(%esp), %eax // Find qhi and rhi such that 104214152Sed movl 16(%esp), %ecx // 105214152Sed xorl %edx, %edx // ahi = qhi*b + rhi with 0 ��� rhi < b 106214152Sed divl %ecx // 107214152Sed movl %eax, %ebx // 108214152Sed movl 8(%esp), %eax // Find qlo such that 109214152Sed divl %ecx // 110214152Sed movl %ebx, %edx // rhi:alo = qlo*b + rlo with 0 ��� rlo < b 111214152Sed popl %ebx // 112214152Sed retl // and return qhi:qlo 113214152Sed 114214152Sed#endif // __i386__ 115