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 __divdi3(di_int a, di_int b); 7*0a6a1f1dSLionel Sambuc 8*0a6a1f1dSLionel Sambuc// result = 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// Stephen Canon, December 2008 18*0a6a1f1dSLionel Sambuc 19*0a6a1f1dSLionel Sambuc#ifdef __i386__ 20*0a6a1f1dSLionel Sambuc 21*0a6a1f1dSLionel Sambuc.text 22*0a6a1f1dSLionel Sambuc.balign 4 23*0a6a1f1dSLionel SambucDEFINE_COMPILERRT_FUNCTION(__divdi3) 24*0a6a1f1dSLionel Sambuc 25*0a6a1f1dSLionel Sambuc/* This is currently implemented by wrapping the unsigned divide up in an absolute 26*0a6a1f1dSLionel Sambuc value, then restoring the correct sign at the end of the computation. This could 27*0a6a1f1dSLionel Sambuc 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 movl %ecx, %esi // set aside sign of b 41*0a6a1f1dSLionel Sambuc 42*0a6a1f1dSLionel Sambuc movl 12(%esp), %edx // high word of b 43*0a6a1f1dSLionel Sambuc movl 8(%esp), %eax // low word of b 44*0a6a1f1dSLionel Sambuc movl %edx, %ecx 45*0a6a1f1dSLionel Sambuc sarl $31, %ecx // (a < 0) ? -1 : 0 46*0a6a1f1dSLionel Sambuc xorl %ecx, %eax 47*0a6a1f1dSLionel Sambuc xorl %ecx, %edx // EDX:EAX = (a < 0) ? not(a) : a 48*0a6a1f1dSLionel Sambuc subl %ecx, %eax 49*0a6a1f1dSLionel Sambuc sbbl %ecx, %edx // EDX:EAX = abs(a) 50*0a6a1f1dSLionel Sambuc movl %edx, 12(%esp) 51*0a6a1f1dSLionel Sambuc movl %eax, 8(%esp) // store abs(a) back to stack 52*0a6a1f1dSLionel Sambuc xorl %ecx, %esi // sign of result = (sign of a) ^ (sign of b) 53*0a6a1f1dSLionel Sambuc 54*0a6a1f1dSLionel Sambuc pushl %ebx 55*0a6a1f1dSLionel Sambuc movl 24(%esp), %ebx // Find the index i of the leading bit in b. 56*0a6a1f1dSLionel Sambuc bsrl %ebx, %ecx // If the high word of b is zero, jump to 57*0a6a1f1dSLionel Sambuc jz 9f // the code to handle that special case [9]. 58*0a6a1f1dSLionel Sambuc 59*0a6a1f1dSLionel Sambuc /* High word of b is known to be non-zero on this branch */ 60*0a6a1f1dSLionel Sambuc 61*0a6a1f1dSLionel Sambuc movl 20(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b 62*0a6a1f1dSLionel Sambuc 63*0a6a1f1dSLionel Sambuc shrl %cl, %eax // Practically, this means that bhi is given by: 64*0a6a1f1dSLionel Sambuc shrl %eax // 65*0a6a1f1dSLionel Sambuc notl %ecx // bhi = (high word of b) << (31 - i) | 66*0a6a1f1dSLionel Sambuc shll %cl, %ebx // (low word of b) >> (1 + i) 67*0a6a1f1dSLionel Sambuc orl %eax, %ebx // 68*0a6a1f1dSLionel Sambuc movl 16(%esp), %edx // Load the high and low words of a, and jump 69*0a6a1f1dSLionel Sambuc movl 12(%esp), %eax // to [1] if the high word is larger than bhi 70*0a6a1f1dSLionel Sambuc cmpl %ebx, %edx // to avoid overflowing the upcoming divide. 71*0a6a1f1dSLionel Sambuc jae 1f 72*0a6a1f1dSLionel Sambuc 73*0a6a1f1dSLionel Sambuc /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 74*0a6a1f1dSLionel Sambuc 75*0a6a1f1dSLionel Sambuc divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r 76*0a6a1f1dSLionel Sambuc 77*0a6a1f1dSLionel Sambuc pushl %edi 78*0a6a1f1dSLionel Sambuc notl %ecx 79*0a6a1f1dSLionel Sambuc shrl %eax 80*0a6a1f1dSLionel Sambuc shrl %cl, %eax // q = qs >> (1 + i) 81*0a6a1f1dSLionel Sambuc movl %eax, %edi 82*0a6a1f1dSLionel Sambuc mull 24(%esp) // q*blo 83*0a6a1f1dSLionel Sambuc movl 16(%esp), %ebx 84*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // ECX:EBX = a 85*0a6a1f1dSLionel Sambuc subl %eax, %ebx 86*0a6a1f1dSLionel Sambuc sbbl %edx, %ecx // ECX:EBX = a - q*blo 87*0a6a1f1dSLionel Sambuc movl 28(%esp), %eax 88*0a6a1f1dSLionel Sambuc imull %edi, %eax // q*bhi 89*0a6a1f1dSLionel Sambuc subl %eax, %ecx // ECX:EBX = a - q*b 90*0a6a1f1dSLionel Sambuc sbbl $0, %edi // decrement q if remainder is negative 91*0a6a1f1dSLionel Sambuc xorl %edx, %edx 92*0a6a1f1dSLionel Sambuc movl %edi, %eax 93*0a6a1f1dSLionel Sambuc 94*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 95*0a6a1f1dSLionel Sambuc adcl %esi, %edx 96*0a6a1f1dSLionel Sambuc xorl %esi, %eax 97*0a6a1f1dSLionel Sambuc xorl %esi, %edx 98*0a6a1f1dSLionel Sambuc popl %edi // Restore callee-save registers 99*0a6a1f1dSLionel Sambuc popl %ebx 100*0a6a1f1dSLionel Sambuc popl %esi 101*0a6a1f1dSLionel Sambuc retl // Return 102*0a6a1f1dSLionel Sambuc 103*0a6a1f1dSLionel Sambuc 104*0a6a1f1dSLionel Sambuc1: /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */ 105*0a6a1f1dSLionel Sambuc 106*0a6a1f1dSLionel Sambuc subl %ebx, %edx // subtract bhi from ahi so that divide will not 107*0a6a1f1dSLionel Sambuc divl %ebx // overflow, and find q and r such that 108*0a6a1f1dSLionel Sambuc // 109*0a6a1f1dSLionel Sambuc // ahi:alo = (1:q)*bhi + r 110*0a6a1f1dSLionel Sambuc // 111*0a6a1f1dSLionel Sambuc // Note that q is a number in (31-i).(1+i) 112*0a6a1f1dSLionel Sambuc // fix point. 113*0a6a1f1dSLionel Sambuc 114*0a6a1f1dSLionel Sambuc pushl %edi 115*0a6a1f1dSLionel Sambuc notl %ecx 116*0a6a1f1dSLionel Sambuc shrl %eax 117*0a6a1f1dSLionel Sambuc orl $0x80000000, %eax 118*0a6a1f1dSLionel Sambuc shrl %cl, %eax // q = (1:qs) >> (1 + i) 119*0a6a1f1dSLionel Sambuc movl %eax, %edi 120*0a6a1f1dSLionel Sambuc mull 24(%esp) // q*blo 121*0a6a1f1dSLionel Sambuc movl 16(%esp), %ebx 122*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // ECX:EBX = a 123*0a6a1f1dSLionel Sambuc subl %eax, %ebx 124*0a6a1f1dSLionel Sambuc sbbl %edx, %ecx // ECX:EBX = a - q*blo 125*0a6a1f1dSLionel Sambuc movl 28(%esp), %eax 126*0a6a1f1dSLionel Sambuc imull %edi, %eax // q*bhi 127*0a6a1f1dSLionel Sambuc subl %eax, %ecx // ECX:EBX = a - q*b 128*0a6a1f1dSLionel Sambuc sbbl $0, %edi // decrement q if remainder is negative 129*0a6a1f1dSLionel Sambuc xorl %edx, %edx 130*0a6a1f1dSLionel Sambuc movl %edi, %eax 131*0a6a1f1dSLionel Sambuc 132*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 133*0a6a1f1dSLionel Sambuc adcl %esi, %edx 134*0a6a1f1dSLionel Sambuc xorl %esi, %eax 135*0a6a1f1dSLionel Sambuc xorl %esi, %edx 136*0a6a1f1dSLionel Sambuc popl %edi // Restore callee-save registers 137*0a6a1f1dSLionel Sambuc popl %ebx 138*0a6a1f1dSLionel Sambuc popl %esi 139*0a6a1f1dSLionel Sambuc retl // Return 140*0a6a1f1dSLionel Sambuc 141*0a6a1f1dSLionel Sambuc 142*0a6a1f1dSLionel Sambuc9: /* High word of b is zero on this branch */ 143*0a6a1f1dSLionel Sambuc 144*0a6a1f1dSLionel Sambuc movl 16(%esp), %eax // Find qhi and rhi such that 145*0a6a1f1dSLionel Sambuc movl 20(%esp), %ecx // 146*0a6a1f1dSLionel Sambuc xorl %edx, %edx // ahi = qhi*b + rhi with 0 ≤ rhi < b 147*0a6a1f1dSLionel Sambuc divl %ecx // 148*0a6a1f1dSLionel Sambuc movl %eax, %ebx // 149*0a6a1f1dSLionel Sambuc movl 12(%esp), %eax // Find qlo such that 150*0a6a1f1dSLionel Sambuc divl %ecx // 151*0a6a1f1dSLionel Sambuc movl %ebx, %edx // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b 152*0a6a1f1dSLionel Sambuc 153*0a6a1f1dSLionel Sambuc addl %esi, %eax // Restore correct sign to result 154*0a6a1f1dSLionel Sambuc adcl %esi, %edx 155*0a6a1f1dSLionel Sambuc xorl %esi, %eax 156*0a6a1f1dSLionel Sambuc xorl %esi, %edx 157*0a6a1f1dSLionel Sambuc popl %ebx // Restore callee-save registers 158*0a6a1f1dSLionel Sambuc popl %esi 159*0a6a1f1dSLionel Sambuc retl // Return 160*0a6a1f1dSLionel SambucEND_COMPILERRT_FUNCTION(__divdi3) 161*0a6a1f1dSLionel Sambuc 162*0a6a1f1dSLionel Sambuc#endif // __i386__ 163