1*0a6a1f1dSLionel Sambuc// This file is dual licensed under the MIT and the University of Illinois Open 2*0a6a1f1dSLionel Sambuc// Source Licenses. See LICENSE.TXT for details. 3*0a6a1f1dSLionel Sambuc 4*0a6a1f1dSLionel Sambuc#include "../assembly.h" 5*0a6a1f1dSLionel Sambuc 6*0a6a1f1dSLionel Sambuc// di_int __moddi3(di_int a, di_int b); 7*0a6a1f1dSLionel Sambuc 8*0a6a1f1dSLionel Sambuc// result = remainder of a / b. 9*0a6a1f1dSLionel Sambuc// both inputs and the output are 64-bit signed integers. 10*0a6a1f1dSLionel Sambuc// This will do whatever the underlying hardware is set to do on division by zero. 11*0a6a1f1dSLionel Sambuc// No other exceptions are generated, as the divide cannot overflow. 12*0a6a1f1dSLionel Sambuc// 13*0a6a1f1dSLionel Sambuc// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware 14*0a6a1f1dSLionel Sambuc// on x86_64. The performance goal is ~40 cycles per divide, which is faster than 15*0a6a1f1dSLionel Sambuc// currently possible via simulation of integer divides on the x87 unit. 16*0a6a1f1dSLionel Sambuc// 17*0a6a1f1dSLionel Sambuc 18*0a6a1f1dSLionel Sambuc// Stephen Canon, December 2008 19*0a6a1f1dSLionel Sambuc 20*0a6a1f1dSLionel Sambuc#ifdef __i386__ 21*0a6a1f1dSLionel Sambuc 22*0a6a1f1dSLionel Sambuc.text 23*0a6a1f1dSLionel Sambuc.balign 4 24*0a6a1f1dSLionel SambucDEFINE_COMPILERRT_FUNCTION(__moddi3) 25*0a6a1f1dSLionel Sambuc 26*0a6a1f1dSLionel Sambuc/* This is currently implemented by wrapping the unsigned modulus up in an absolute 27*0a6a1f1dSLionel Sambuc value. This could certainly be improved upon. */ 28*0a6a1f1dSLionel Sambuc 29*0a6a1f1dSLionel Sambuc pushl %esi 30*0a6a1f1dSLionel Sambuc movl 20(%esp), %edx // high word of b 31*0a6a1f1dSLionel Sambuc movl 16(%esp), %eax // low word of b 32*0a6a1f1dSLionel Sambuc movl %edx, %ecx 33*0a6a1f1dSLionel Sambuc sarl $31, %ecx // (b < 0) ? -1 : 0 34*0a6a1f1dSLionel Sambuc xorl %ecx, %eax 35*0a6a1f1dSLionel Sambuc xorl %ecx, %edx // EDX:EAX = (b < 0) ? not(b) : b 36*0a6a1f1dSLionel Sambuc subl %ecx, %eax 37*0a6a1f1dSLionel Sambuc sbbl %ecx, %edx // EDX:EAX = abs(b) 38*0a6a1f1dSLionel Sambuc movl %edx, 20(%esp) 39*0a6a1f1dSLionel Sambuc movl %eax, 16(%esp) // store abs(b) back to stack 40*0a6a1f1dSLionel Sambuc 41*0a6a1f1dSLionel Sambuc movl 12(%esp), %edx // high word of b 42*0a6a1f1dSLionel Sambuc movl 8(%esp), %eax // low word of b 43*0a6a1f1dSLionel Sambuc movl %edx, %ecx 44*0a6a1f1dSLionel Sambuc sarl $31, %ecx // (a < 0) ? -1 : 0 45*0a6a1f1dSLionel Sambuc xorl %ecx, %eax 46*0a6a1f1dSLionel Sambuc xorl %ecx, %edx // EDX:EAX = (a < 0) ? not(a) : a 47*0a6a1f1dSLionel Sambuc subl %ecx, %eax 48*0a6a1f1dSLionel Sambuc sbbl %ecx, %edx // EDX:EAX = abs(a) 49*0a6a1f1dSLionel Sambuc movl %edx, 12(%esp) 50*0a6a1f1dSLionel Sambuc movl %eax, 8(%esp) // store abs(a) back to stack 51*0a6a1f1dSLionel Sambuc movl %ecx, %esi // set aside sign of a 52*0a6a1f1dSLionel Sambuc 53*0a6a1f1dSLionel Sambuc pushl %ebx 54*0a6a1f1dSLionel Sambuc movl 24(%esp), %ebx // Find the index i of the leading bit in b. 55*0a6a1f1dSLionel Sambuc bsrl %ebx, %ecx // If the high word of b is zero, jump to 56*0a6a1f1dSLionel Sambuc jz 9f // the code to handle that special case [9]. 57*0a6a1f1dSLionel Sambuc 58*0a6a1f1dSLionel Sambuc /* High word of b is known to be non-zero on this branch */ 59*0a6a1f1dSLionel Sambuc 60*0a6a1f1dSLionel Sambuc movl 20(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b 61*0a6a1f1dSLionel Sambuc 62*0a6a1f1dSLionel Sambuc shrl %cl, %eax // Practically, this means that bhi is given by: 63*0a6a1f1dSLionel Sambuc shrl %eax // 64*0a6a1f1dSLionel Sambuc notl %ecx // bhi = (high word of b) << (31 - i) | 65*0a6a1f1dSLionel Sambuc shll %cl, %ebx // (low word of b) >> (1 + i) 66*0a6a1f1dSLionel Sambuc orl %eax, %ebx // 67*0a6a1f1dSLionel Sambuc movl 16(%esp), %edx // Load the high and low words of a, and jump 68*0a6a1f1dSLionel Sambuc movl 12(%esp), %eax // to [2] if the high word is larger than bhi 69*0a6a1f1dSLionel Sambuc cmpl %ebx, %edx // to avoid overflowing the upcoming divide. 70*0a6a1f1dSLionel Sambuc jae 2f 71*0a6a1f1dSLionel Sambuc 72*0a6a1f1dSLionel Sambuc /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 73*0a6a1f1dSLionel Sambuc 74*0a6a1f1dSLionel Sambuc divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r 75*0a6a1f1dSLionel Sambuc 76*0a6a1f1dSLionel Sambuc pushl %edi 77*0a6a1f1dSLionel Sambuc notl %ecx 78*0a6a1f1dSLionel Sambuc shrl %eax 79*0a6a1f1dSLionel Sambuc shrl %cl, %eax // q = qs >> (1 + i) 80*0a6a1f1dSLionel Sambuc movl %eax, %edi 81*0a6a1f1dSLionel Sambuc mull 24(%esp) // q*blo 82*0a6a1f1dSLionel Sambuc movl 16(%esp), %ebx 83*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // ECX:EBX = a 84*0a6a1f1dSLionel Sambuc subl %eax, %ebx 85*0a6a1f1dSLionel Sambuc sbbl %edx, %ecx // ECX:EBX = a - q*blo 86*0a6a1f1dSLionel Sambuc movl 28(%esp), %eax 87*0a6a1f1dSLionel Sambuc imull %edi, %eax // q*bhi 88*0a6a1f1dSLionel Sambuc subl %eax, %ecx // ECX:EBX = a - q*b 89*0a6a1f1dSLionel Sambuc 90*0a6a1f1dSLionel Sambuc jnc 1f // if positive, this is the result. 91*0a6a1f1dSLionel Sambuc addl 24(%esp), %ebx // otherwise 92*0a6a1f1dSLionel Sambuc adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result 93*0a6a1f1dSLionel Sambuc1: movl %ebx, %eax 94*0a6a1f1dSLionel Sambuc movl %ecx, %edx 95*0a6a1f1dSLionel Sambuc 96*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 97*0a6a1f1dSLionel Sambuc adcl %esi, %edx 98*0a6a1f1dSLionel Sambuc xorl %esi, %eax 99*0a6a1f1dSLionel Sambuc xorl %esi, %edx 100*0a6a1f1dSLionel Sambuc popl %edi // Restore callee-save registers 101*0a6a1f1dSLionel Sambuc popl %ebx 102*0a6a1f1dSLionel Sambuc popl %esi 103*0a6a1f1dSLionel Sambuc retl // Return 104*0a6a1f1dSLionel Sambuc 105*0a6a1f1dSLionel Sambuc2: /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 106*0a6a1f1dSLionel Sambuc 107*0a6a1f1dSLionel Sambuc subl %ebx, %edx // subtract bhi from ahi so that divide will not 108*0a6a1f1dSLionel Sambuc divl %ebx // overflow, and find q and r such that 109*0a6a1f1dSLionel Sambuc // 110*0a6a1f1dSLionel Sambuc // ahi:alo = (1:q)*bhi + r 111*0a6a1f1dSLionel Sambuc // 112*0a6a1f1dSLionel Sambuc // Note that q is a number in (31-i).(1+i) 113*0a6a1f1dSLionel Sambuc // fix point. 114*0a6a1f1dSLionel Sambuc 115*0a6a1f1dSLionel Sambuc pushl %edi 116*0a6a1f1dSLionel Sambuc notl %ecx 117*0a6a1f1dSLionel Sambuc shrl %eax 118*0a6a1f1dSLionel Sambuc orl $0x80000000, %eax 119*0a6a1f1dSLionel Sambuc shrl %cl, %eax // q = (1:qs) >> (1 + i) 120*0a6a1f1dSLionel Sambuc movl %eax, %edi 121*0a6a1f1dSLionel Sambuc mull 24(%esp) // q*blo 122*0a6a1f1dSLionel Sambuc movl 16(%esp), %ebx 123*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // ECX:EBX = a 124*0a6a1f1dSLionel Sambuc subl %eax, %ebx 125*0a6a1f1dSLionel Sambuc sbbl %edx, %ecx // ECX:EBX = a - q*blo 126*0a6a1f1dSLionel Sambuc movl 28(%esp), %eax 127*0a6a1f1dSLionel Sambuc imull %edi, %eax // q*bhi 128*0a6a1f1dSLionel Sambuc subl %eax, %ecx // ECX:EBX = a - q*b 129*0a6a1f1dSLionel Sambuc 130*0a6a1f1dSLionel Sambuc jnc 3f // if positive, this is the result. 131*0a6a1f1dSLionel Sambuc addl 24(%esp), %ebx // otherwise 132*0a6a1f1dSLionel Sambuc adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result 133*0a6a1f1dSLionel Sambuc3: movl %ebx, %eax 134*0a6a1f1dSLionel Sambuc movl %ecx, %edx 135*0a6a1f1dSLionel Sambuc 136*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 137*0a6a1f1dSLionel Sambuc adcl %esi, %edx 138*0a6a1f1dSLionel Sambuc xorl %esi, %eax 139*0a6a1f1dSLionel Sambuc xorl %esi, %edx 140*0a6a1f1dSLionel Sambuc popl %edi // Restore callee-save registers 141*0a6a1f1dSLionel Sambuc popl %ebx 142*0a6a1f1dSLionel Sambuc popl %esi 143*0a6a1f1dSLionel Sambuc retl // Return 144*0a6a1f1dSLionel Sambuc 145*0a6a1f1dSLionel Sambuc9: /* High word of b is zero on this branch */ 146*0a6a1f1dSLionel Sambuc 147*0a6a1f1dSLionel Sambuc movl 16(%esp), %eax // Find qhi and rhi such that 148*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // 149*0a6a1f1dSLionel Sambuc xorl %edx, %edx // ahi = qhi*b + rhi with 0 ≤ rhi < b 150*0a6a1f1dSLionel Sambuc divl %ecx // 151*0a6a1f1dSLionel Sambuc movl %eax, %ebx // 152*0a6a1f1dSLionel Sambuc movl 12(%esp), %eax // Find rlo such that 153*0a6a1f1dSLionel Sambuc divl %ecx // 154*0a6a1f1dSLionel Sambuc movl %edx, %eax // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b 155*0a6a1f1dSLionel Sambuc popl %ebx // 156*0a6a1f1dSLionel Sambuc xorl %edx, %edx // and return 0:rlo 157*0a6a1f1dSLionel Sambuc 158*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 159*0a6a1f1dSLionel Sambuc adcl %esi, %edx 160*0a6a1f1dSLionel Sambuc xorl %esi, %eax 161*0a6a1f1dSLionel Sambuc xorl %esi, %edx 162*0a6a1f1dSLionel Sambuc popl %esi 163*0a6a1f1dSLionel Sambuc retl // Return 164*0a6a1f1dSLionel SambucEND_COMPILERRT_FUNCTION(__moddi3) 165*0a6a1f1dSLionel Sambuc 166*0a6a1f1dSLionel Sambuc#endif // __i386__ 167