One thing that comes to mind is that GCC's version could have a bit more overhead due to ESP-folding [0]. Basically, reading ESP/RSP directly can incur some overhead, because the register renamer is playing tricks to avoid actually adjusting the stack pointer on every push/pop until you actually read the stack pointer's value. It's unclear here why GCC chose RSP over some other register.

[0] the only reference I could find at the moment, but I've seen it documented elsewhere too: http://homepage.ntlworld.com/jonathan.deboynepollard/FGA/fun...