summaryrefslogtreecommitdiff
path: root/lib/i386/moddi3.S
diff options
context:
space:
mode:
Diffstat (limited to 'lib/i386/moddi3.S')
-rw-r--r--lib/i386/moddi3.S166
1 files changed, 166 insertions, 0 deletions
diff --git a/lib/i386/moddi3.S b/lib/i386/moddi3.S
new file mode 100644
index 000000000000..85f5b3811d5c
--- /dev/null
+++ b/lib/i386/moddi3.S
@@ -0,0 +1,166 @@
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+
+#include "../assembly.h"
+
+// di_int __moddi3(di_int a, di_int b);
+
+// result = remainder of a / b.
+// both inputs and the output are 64-bit signed integers.
+// This will do whatever the underlying hardware is set to do on division by zero.
+// No other exceptions are generated, as the divide cannot overflow.
+//
+// This is targeted at 32-bit x86 *only*, as this can be done directly in hardware
+// on x86_64. The performance goal is ~40 cycles per divide, which is faster than
+// currently possible via simulation of integer divides on the x87 unit.
+//
+
+// Stephen Canon, December 2008
+
+#ifdef __i386__
+
+.text
+.align 4
+DEFINE_COMPILERRT_FUNCTION(__moddi3)
+
+/* This is currently implemented by wrapping the unsigned modulus up in an absolute
+ value. This could certainly be improved upon. */
+
+ pushl %esi
+ movl 20(%esp), %edx // high word of b
+ movl 16(%esp), %eax // low word of b
+ movl %edx, %ecx
+ sarl $31, %ecx // (b < 0) ? -1 : 0
+ xorl %ecx, %eax
+ xorl %ecx, %edx // EDX:EAX = (b < 0) ? not(b) : b
+ subl %ecx, %eax
+ sbbl %ecx, %edx // EDX:EAX = abs(b)
+ movl %edx, 20(%esp)
+ movl %eax, 16(%esp) // store abs(b) back to stack
+
+ movl 12(%esp), %edx // high word of b
+ movl 8(%esp), %eax // low word of b
+ movl %edx, %ecx
+ sarl $31, %ecx // (a < 0) ? -1 : 0
+ xorl %ecx, %eax
+ xorl %ecx, %edx // EDX:EAX = (a < 0) ? not(a) : a
+ subl %ecx, %eax
+ sbbl %ecx, %edx // EDX:EAX = abs(a)
+ movl %edx, 12(%esp)
+ movl %eax, 8(%esp) // store abs(a) back to stack
+ movl %ecx, %esi // set aside sign of a
+
+ pushl %ebx
+ movl 24(%esp), %ebx // Find the index i of the leading bit in b.
+ bsrl %ebx, %ecx // If the high word of b is zero, jump to
+ jz 9f // the code to handle that special case [9].
+
+ /* High word of b is known to be non-zero on this branch */
+
+ movl 20(%esp), %eax // Construct bhi, containing bits [1+i:32+i] of b
+
+ shrl %cl, %eax // Practically, this means that bhi is given by:
+ shrl %eax //
+ notl %ecx // bhi = (high word of b) << (31 - i) |
+ shll %cl, %ebx // (low word of b) >> (1 + i)
+ orl %eax, %ebx //
+ movl 16(%esp), %edx // Load the high and low words of a, and jump
+ movl 12(%esp), %eax // to [2] if the high word is larger than bhi
+ cmpl %ebx, %edx // to avoid overflowing the upcoming divide.
+ jae 2f
+
+ /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
+
+ divl %ebx // eax <-- qs, edx <-- r such that ahi:alo = bs*qs + r
+
+ pushl %edi
+ notl %ecx
+ shrl %eax
+ shrl %cl, %eax // q = qs >> (1 + i)
+ movl %eax, %edi
+ mull 24(%esp) // q*blo
+ movl 16(%esp), %ebx
+ movl 20(%esp), %ecx // ECX:EBX = a
+ subl %eax, %ebx
+ sbbl %edx, %ecx // ECX:EBX = a - q*blo
+ movl 28(%esp), %eax
+ imull %edi, %eax // q*bhi
+ subl %eax, %ecx // ECX:EBX = a - q*b
+
+ jnc 1f // if positive, this is the result.
+ addl 24(%esp), %ebx // otherwise
+ adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result
+1: movl %ebx, %eax
+ movl %ecx, %edx
+
+ addl %esi, %eax // Restore correct sign to result
+ adcl %esi, %edx
+ xorl %esi, %eax
+ xorl %esi, %edx
+ popl %edi // Restore callee-save registers
+ popl %ebx
+ popl %esi
+ retl // Return
+
+2: /* High word of a is greater than or equal to (b >> (1 + i)) on this branch */
+
+ subl %ebx, %edx // subtract bhi from ahi so that divide will not
+ divl %ebx // overflow, and find q and r such that
+ //
+ // ahi:alo = (1:q)*bhi + r
+ //
+ // Note that q is a number in (31-i).(1+i)
+ // fix point.
+
+ pushl %edi
+ notl %ecx
+ shrl %eax
+ orl $0x80000000, %eax
+ shrl %cl, %eax // q = (1:qs) >> (1 + i)
+ movl %eax, %edi
+ mull 24(%esp) // q*blo
+ movl 16(%esp), %ebx
+ movl 20(%esp), %ecx // ECX:EBX = a
+ subl %eax, %ebx
+ sbbl %edx, %ecx // ECX:EBX = a - q*blo
+ movl 28(%esp), %eax
+ imull %edi, %eax // q*bhi
+ subl %eax, %ecx // ECX:EBX = a - q*b
+
+ jnc 3f // if positive, this is the result.
+ addl 24(%esp), %ebx // otherwise
+ adcl 28(%esp), %ecx // ECX:EBX = a - (q-1)*b = result
+3: movl %ebx, %eax
+ movl %ecx, %edx
+
+ addl %esi, %eax // Restore correct sign to result
+ adcl %esi, %edx
+ xorl %esi, %eax
+ xorl %esi, %edx
+ popl %edi // Restore callee-save registers
+ popl %ebx
+ popl %esi
+ retl // Return
+
+9: /* High word of b is zero on this branch */
+
+ movl 16(%esp), %eax // Find qhi and rhi such that
+ movl 20(%esp), %ecx //
+ xorl %edx, %edx // ahi = qhi*b + rhi with 0 ≤ rhi < b
+ divl %ecx //
+ movl %eax, %ebx //
+ movl 12(%esp), %eax // Find rlo such that
+ divl %ecx //
+ movl %edx, %eax // rhi:alo = qlo*b + rlo with 0 ≤ rlo < b
+ popl %ebx //
+ xorl %edx, %edx // and return 0:rlo
+
+ addl %esi, %eax // Restore correct sign to result
+ adcl %esi, %edx
+ xorl %esi, %eax
+ xorl %esi, %edx
+ popl %esi
+ retl // Return
+
+
+#endif // __i386__