167 lines · plain
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-exception4 5#include "../assembly.h"6 7// di_int __divdi3(di_int a, di_int b);8 9// result = a / b.10// both inputs and the output are 64-bit signed 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 hardware15// on x86_64. The performance goal is ~40 cycles per divide, which is faster than16// currently possible via simulation of integer divides on the x87 unit.17//18// Stephen Canon, December 200819 20#ifdef __i386__21 22.text23.balign 424DEFINE_COMPILERRT_FUNCTION(__divdi3)25 26// This is currently implemented by wrapping the unsigned divide up in an absolute27// value, then restoring the correct sign at the end of the computation. This could28// certainly be improved upon.29 30 pushl %esi31 movl 20(%esp), %edx // high word of b32 movl 16(%esp), %eax // low word of b33 movl %edx, %ecx34 sarl $31, %ecx // (b < 0) ? -1 : 035 xorl %ecx, %eax36 xorl %ecx, %edx // EDX:EAX = (b < 0) ? not(b) : b37 subl %ecx, %eax38 sbbl %ecx, %edx // EDX:EAX = abs(b)39 movl %edx, 20(%esp)40 movl %eax, 16(%esp) // store abs(b) back to stack41 movl %ecx, %esi // set aside sign of b42 43 movl 12(%esp), %edx // high word of b44 movl 8(%esp), %eax // low word of b45 movl %edx, %ecx46 sarl $31, %ecx // (a < 0) ? -1 : 047 xorl %ecx, %eax48 xorl %ecx, %edx // EDX:EAX = (a < 0) ? not(a) : a49 subl %ecx, %eax50 sbbl %ecx, %edx // EDX:EAX = abs(a)51 movl %edx, 12(%esp)52 movl %eax, 8(%esp) // store abs(a) back to stack53 xorl %ecx, %esi // sign of result = (sign of a) ^ (sign of b)54 55 pushl %ebx56 movl 24(%esp), %ebx // Find the index i of the leading bit in b.57 bsrl %ebx, %ecx // If the high word of b is zero, jump to58 jz 9f // the code to handle that special case [9].59 60 // High word of b is known to be non-zero on this branch61 62 movl 20(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b63 64 shrl %cl, %eax // Practically, this means that bhi is given by:65 shrl %eax //66 notl %ecx // bhi = (high word of b) << (31 - i) |67 shll %cl, %ebx // (low word of b) >> (1 + i)68 orl %eax, %ebx //69 movl 16(%esp), %edx // Load the high and low words of a, and jump70 movl 12(%esp), %eax // to [1] if the high word is larger than bhi71 cmpl %ebx, %edx // to avoid overflowing the upcoming divide.72 jae 1f73 74 // High word of a is greater than or equal to (b >> (1 + i)) on this branch75 76 divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r77 78 pushl %edi79 notl %ecx80 shrl %eax81 shrl %cl, %eax // q = qs >> (1 + i)82 movl %eax, %edi83 mull 24(%esp) // q*blo84 movl 16(%esp), %ebx85 movl 20(%esp), %ecx // ECX:EBX = a86 subl %eax, %ebx87 sbbl %edx, %ecx // ECX:EBX = a - q*blo88 movl 28(%esp), %eax89 imull %edi, %eax // q*bhi90 subl %eax, %ecx // ECX:EBX = a - q*b91 sbbl $0, %edi // decrement q if remainder is negative92 xorl %edx, %edx93 movl %edi, %eax94 95 addl %esi, %eax // Restore correct sign to result96 adcl %esi, %edx97 xorl %esi, %eax98 xorl %esi, %edx99 popl %edi // Restore callee-save registers100 popl %ebx101 popl %esi102 retl // Return103 104 1051: // High word of a is greater than or equal to (b >> (1 + i)) on this branch106 107 subl %ebx, %edx // subtract bhi from ahi so that divide will not108 divl %ebx // overflow, and find q and r such that109 //110 // ahi:alo = (1:q)*bhi + r111 //112 // Note that q is a number in (31-i).(1+i)113 // fix point.114 115 pushl %edi116 notl %ecx117 shrl %eax118 orl $0x80000000, %eax119 shrl %cl, %eax // q = (1:qs) >> (1 + i)120 movl %eax, %edi121 mull 24(%esp) // q*blo122 movl 16(%esp), %ebx123 movl 20(%esp), %ecx // ECX:EBX = a124 subl %eax, %ebx125 sbbl %edx, %ecx // ECX:EBX = a - q*blo126 movl 28(%esp), %eax127 imull %edi, %eax // q*bhi128 subl %eax, %ecx // ECX:EBX = a - q*b129 sbbl $0, %edi // decrement q if remainder is negative130 xorl %edx, %edx131 movl %edi, %eax132 133 addl %esi, %eax // Restore correct sign to result134 adcl %esi, %edx135 xorl %esi, %eax136 xorl %esi, %edx137 popl %edi // Restore callee-save registers138 popl %ebx139 popl %esi140 retl // Return141 142 1439: // High word of b is zero on this branch144 145 movl 16(%esp), %eax // Find qhi and rhi such that146 movl 20(%esp), %ecx //147 xorl %edx, %edx // ahi = qhi*b + rhi with 0 ≤ rhi < b148 divl %ecx //149 movl %eax, %ebx //150 movl 12(%esp), %eax // Find qlo such that151 divl %ecx //152 movl %ebx, %edx // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b153 154 addl %esi, %eax // Restore correct sign to result155 adcl %esi, %edx156 xorl %esi, %eax157 xorl %esi, %edx158 popl %ebx // Restore callee-save registers159 popl %esi160 retl // Return161END_COMPILERRT_FUNCTION(__divdi3)162 163#endif // __i386__164 165NO_EXEC_STACK_DIRECTIVE166 167