auto import from //depot/cupcake/@135843
diff --git a/libc/arch-x86/string/swab.S b/libc/arch-x86/string/swab.S
new file mode 100644
index 0000000..3055860
--- /dev/null
+++ b/libc/arch-x86/string/swab.S
@@ -0,0 +1,67 @@
+/* $OpenBSD: swab.S,v 1.3 2005/08/07 11:30:38 espie Exp $ */
+/*
+ * Written by J.T. Conklin <jtc@netbsd.org>.
+ * Public domain.
+ */
+
+#include <machine/asm.h>
+
+/*
+ * On the i486, this code is negligibly faster than the code generated
+ * by gcc at about half the size. If my i386 databook is correct, it
+ * should be considerably faster than the gcc code on a i386.
+ */
+
+ENTRY(swab)
+ pushl %esi
+ pushl %edi
+ movl 12(%esp),%esi
+ movl 16(%esp),%edi
+ movl 20(%esp),%ecx
+
+ cld # set direction forward
+
+ shrl $1,%ecx
+ testl $7,%ecx # copy first group of 1 to 7 words
+ jz L2 # while swaping alternate bytes.
+ .align 2,0x90
+L1: lodsw
+ rorw $8,%ax
+ stosw
+ decl %ecx
+ testl $7,%ecx
+ jnz L1
+
+L2: shrl $3,%ecx # copy remainder 8 words at a time
+ jz L4 # while swapping alternate bytes.
+ .align 2,0x90
+L3: lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ lodsw
+ rorw $8,%ax
+ stosw
+ decl %ecx
+ jnz L3
+
+L4: popl %edi
+ popl %esi
+ ret