1 1.1 joerg // This file is dual licensed under the MIT and the University of Illinois Open 2 1.1 joerg // Source Licenses. See LICENSE.TXT for details. 3 1.1 joerg 4 1.1 joerg #include "../assembly.h" 5 1.1 joerg 6 1.1 joerg // du_int __udivdi3(du_int a, du_int b); 7 1.1 joerg 8 1.1 joerg // result = a / b. 9 1.1 joerg // both inputs and the output are 64-bit unsigned integers. 10 1.1 joerg // This will do whatever the underlying hardware is set to do on division by zero. 11 1.1 joerg // No other exceptions are generated, as the divide cannot overflow. 12 1.1 joerg // 13 1.1 joerg // This is targeted at 32-bit x86 *only*, as this can be done directly in hardware 14 1.1 joerg // on x86_64. The performance goal is ~40 cycles per divide, which is faster than 15 1.1 joerg // currently possible via simulation of integer divides on the x87 unit. 16 1.1 joerg // 17 1.1 joerg // Stephen Canon, December 2008 18 1.1 joerg 19 1.1 joerg #ifdef __i386__ 20 1.1 joerg 21 1.1 joerg .text 22 1.1.1.2 joerg .balign 4 23 1.1 joerg DEFINE_COMPILERRT_FUNCTION(__udivdi3) 24 1.1 joerg 25 1.1 joerg pushl %ebx 26 1.1 joerg movl 20(%esp), %ebx // Find the index i of the leading bit in b. 27 1.1 joerg bsrl %ebx, %ecx // If the high word of b is zero, jump to 28 1.1 joerg jz 9f // the code to handle that special case [9]. 29 1.1 joerg 30 1.1 joerg /* High word of b is known to be non-zero on this branch */ 31 1.1 joerg 32 1.1 joerg movl 16(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b 33 1.1 joerg 34 1.1 joerg shrl %cl, %eax // Practically, this means that bhi is given by: 35 1.1 joerg shrl %eax // 36 1.1 joerg notl %ecx // bhi = (high word of b) << (31 - i) | 37 1.1 joerg shll %cl, %ebx // (low word of b) >> (1 + i) 38 1.1 joerg orl %eax, %ebx // 39 1.1 joerg movl 12(%esp), %edx // Load the high and low words of a, and jump 40 1.1 joerg movl 8(%esp), %eax // to [1] if the high word is larger than bhi 41 1.1 joerg cmpl %ebx, %edx // to avoid overflowing the upcoming divide. 42 1.1 joerg jae 1f 43 1.1 joerg 44 1.1 joerg /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 45 1.1 joerg 46 1.1 joerg divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r 47 1.1 joerg 48 1.1 joerg pushl %edi 49 1.1 joerg notl %ecx 50 1.1 joerg shrl %eax 51 1.1 joerg shrl %cl, %eax // q = qs >> (1 + i) 52 1.1 joerg movl %eax, %edi 53 1.1 joerg mull 20(%esp) // q*blo 54 1.1 joerg movl 12(%esp), %ebx 55 1.1 joerg movl 16(%esp), %ecx // ECX:EBX = a 56 1.1 joerg subl %eax, %ebx 57 1.1 joerg sbbl %edx, %ecx // ECX:EBX = a - q*blo 58 1.1 joerg movl 24(%esp), %eax 59 1.1 joerg imull %edi, %eax // q*bhi 60 1.1 joerg subl %eax, %ecx // ECX:EBX = a - q*b 61 1.1 joerg sbbl $0, %edi // decrement q if remainder is negative 62 1.1 joerg xorl %edx, %edx 63 1.1 joerg movl %edi, %eax 64 1.1 joerg popl %edi 65 1.1 joerg popl %ebx 66 1.1 joerg retl 67 1.1 joerg 68 1.1 joerg 69 1.1 joerg 1: /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 70 1.1 joerg 71 1.1 joerg subl %ebx, %edx // subtract bhi from ahi so that divide will not 72 1.1 joerg divl %ebx // overflow, and find q and r such that 73 1.1 joerg // 74 1.1 joerg // ahi:alo = (1:q)*bhi + r 75 1.1 joerg // 76 1.1 joerg // Note that q is a number in (31-i).(1+i) 77 1.1 joerg // fix point. 78 1.1 joerg 79 1.1 joerg pushl %edi 80 1.1 joerg notl %ecx 81 1.1 joerg shrl %eax 82 1.1 joerg orl $0x80000000, %eax 83 1.1 joerg shrl %cl, %eax // q = (1:qs) >> (1 + i) 84 1.1 joerg movl %eax, %edi 85 1.1 joerg mull 20(%esp) // q*blo 86 1.1 joerg movl 12(%esp), %ebx 87 1.1 joerg movl 16(%esp), %ecx // ECX:EBX = a 88 1.1 joerg subl %eax, %ebx 89 1.1 joerg sbbl %edx, %ecx // ECX:EBX = a - q*blo 90 1.1 joerg movl 24(%esp), %eax 91 1.1 joerg imull %edi, %eax // q*bhi 92 1.1 joerg subl %eax, %ecx // ECX:EBX = a - q*b 93 1.1 joerg sbbl $0, %edi // decrement q if remainder is negative 94 1.1 joerg xorl %edx, %edx 95 1.1 joerg movl %edi, %eax 96 1.1 joerg popl %edi 97 1.1 joerg popl %ebx 98 1.1 joerg retl 99 1.1 joerg 100 1.1 joerg 101 1.1 joerg 9: /* High word of b is zero on this branch */ 102 1.1 joerg 103 1.1 joerg movl 12(%esp), %eax // Find qhi and rhi such that 104 1.1 joerg movl 16(%esp), %ecx // 105 1.1 joerg xorl %edx, %edx // ahi = qhi*b + rhi with 0 rhi < b 106 1.1 joerg divl %ecx // 107 1.1 joerg movl %eax, %ebx // 108 1.1 joerg movl 8(%esp), %eax // Find qlo such that 109 1.1 joerg divl %ecx // 110 1.1 joerg movl %ebx, %edx // rhi:alo = qlo*b + rlo with 0 rlo < b 111 1.1 joerg popl %ebx // 112 1.1 joerg retl // and return qhi:qlo 113 1.1 joerg END_COMPILERRT_FUNCTION(__udivdi3) 114 1.1 joerg 115 1.1 joerg #endif // __i386__ 116
Indexes created Mon Sep 22 13:09:51 GMT 2025