mem: Add documentations for REP string insturctions

Signed-off-by: Joe Richey <[email protected]>

src/mem/x86_64.rs

@@ -1,5 +1,18 @@
 use super::c_int;
 
+// On most modern Intel and AMD processors, "rep movsq" and "rep stosq" have
+// been enhanced to perform better than an simple qword loop, making them ideal
+// for implementing memcpy/memset.
+//
+// On certain recent Intel processors, "rep movsb" and "rep stosb" have been
+// further enhanced to automatically select the best microarchitectural
+// implementation based on length and alignment.
+// See "Intel® 64 and IA-32 Architectures Optimization Reference Manual"
+//  - Enhanced REP MOVSB and STOSB operation (ERMSB) (section 3.7.7)
+//  - Fast Short REP MOV (FSRM)
+// However, to avoid run-time feature detection, we don't use these byte-based
+// instructions for most of the copying, instead using the qword variants.
+
 #[cfg_attr(all(feature = "mem", not(feature = "mangled-names")), no_mangle)]
 pub unsafe extern "C" fn memcpy(dest: *mut u8, src: *const u8, count: usize) -> *mut u8 {
     asm!(