171 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 __moddi3(di_int a, di_int b);8 9// result = remainder of 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 19// Stephen Canon, December 200820 21#ifdef __i386__22 23.text24.balign 425DEFINE_COMPILERRT_FUNCTION(__moddi3)26 27// This is currently implemented by wrapping the unsigned modulus up in an absolute28// value. This could 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 42 movl 12(%esp), %edx // high word of b43 movl 8(%esp), %eax // low word of b44 movl %edx, %ecx45 sarl $31, %ecx // (a < 0) ? -1 : 046 xorl %ecx, %eax47 xorl %ecx, %edx // EDX:EAX = (a < 0) ? not(a) : a48 subl %ecx, %eax49 sbbl %ecx, %edx // EDX:EAX = abs(a)50 movl %edx, 12(%esp)51 movl %eax, 8(%esp) // store abs(a) back to stack52 movl %ecx, %esi // set aside sign of a53 54 pushl %ebx55 movl 24(%esp), %ebx // Find the index i of the leading bit in b.56 bsrl %ebx, %ecx // If the high word of b is zero, jump to57 jz 9f // the code to handle that special case [9].58 59 // High word of b is known to be non-zero on this branch60 61 movl 20(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b62 63 shrl %cl, %eax // Practically, this means that bhi is given by:64 shrl %eax //65 notl %ecx // bhi = (high word of b) << (31 - i) |66 shll %cl, %ebx // (low word of b) >> (1 + i)67 orl %eax, %ebx //68 movl 16(%esp), %edx // Load the high and low words of a, and jump69 movl 12(%esp), %eax // to [2] if the high word is larger than bhi70 cmpl %ebx, %edx // to avoid overflowing the upcoming divide.71 jae 2f72 73 // High word of a is greater than or equal to (b >> (1 + i)) on this branch74 75 divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r76 77 pushl %edi78 notl %ecx79 shrl %eax80 shrl %cl, %eax // q = qs >> (1 + i)81 movl %eax, %edi82 mull 24(%esp) // q*blo83 movl 16(%esp), %ebx84 movl 20(%esp), %ecx // ECX:EBX = a85 subl %eax, %ebx86 sbbl %edx, %ecx // ECX:EBX = a - q*blo87 movl 28(%esp), %eax88 imull %edi, %eax // q*bhi89 subl %eax, %ecx // ECX:EBX = a - q*b90 91 jnc 1f // if positive, this is the result.92 addl 24(%esp), %ebx // otherwise93 adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result941: movl %ebx, %eax95 movl %ecx, %edx96 97 addl %esi, %eax // Restore correct sign to result98 adcl %esi, %edx99 xorl %esi, %eax100 xorl %esi, %edx101 popl %edi // Restore callee-save registers102 popl %ebx103 popl %esi104 retl // Return105 1062: // High word of a is greater than or equal to (b >> (1 + i)) on this branch107 108 subl %ebx, %edx // subtract bhi from ahi so that divide will not109 divl %ebx // overflow, and find q and r such that110 //111 // ahi:alo = (1:q)*bhi + r112 //113 // Note that q is a number in (31-i).(1+i)114 // fix point.115 116 pushl %edi117 notl %ecx118 shrl %eax119 orl $0x80000000, %eax120 shrl %cl, %eax // q = (1:qs) >> (1 + i)121 movl %eax, %edi122 mull 24(%esp) // q*blo123 movl 16(%esp), %ebx124 movl 20(%esp), %ecx // ECX:EBX = a125 subl %eax, %ebx126 sbbl %edx, %ecx // ECX:EBX = a - q*blo127 movl 28(%esp), %eax128 imull %edi, %eax // q*bhi129 subl %eax, %ecx // ECX:EBX = a - q*b130 131 jnc 3f // if positive, this is the result.132 addl 24(%esp), %ebx // otherwise133 adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result1343: movl %ebx, %eax135 movl %ecx, %edx136 137 addl %esi, %eax // Restore correct sign to result138 adcl %esi, %edx139 xorl %esi, %eax140 xorl %esi, %edx141 popl %edi // Restore callee-save registers142 popl %ebx143 popl %esi144 retl // Return145 1469: // High word of b is zero on this branch147 148 movl 16(%esp), %eax // Find qhi and rhi such that149 movl 20(%esp), %ecx //150 xorl %edx, %edx // ahi = qhi*b + rhi with 0 ≤ rhi < b151 divl %ecx //152 movl %eax, %ebx //153 movl 12(%esp), %eax // Find rlo such that154 divl %ecx //155 movl %edx, %eax // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b156 popl %ebx //157 xorl %edx, %edx // and return 0:rlo158 159 addl %esi, %eax // Restore correct sign to result160 adcl %esi, %edx161 xorl %esi, %eax162 xorl %esi, %edx163 popl %esi164 retl // Return165END_COMPILERRT_FUNCTION(__moddi3)166 167#endif // __i386__168 169NO_EXEC_STACK_DIRECTIVE170 171