Messages in this thread | | | Date | Thu, 19 Jul 2012 16:07:15 -0700 | From | "H. Peter Anvin" <> | Subject | Re: [RFC][PATCH 2/4 v4] ftrace/x86: Add save_regs for i386 function calls |
| |
On 07/19/2012 04:04 PM, Steven Rostedt wrote: > On Thu, 2012-07-19 at 15:53 -0700, H. Peter Anvin wrote: > >> lea is not typically faster than add, but in the case of Atom, it is >> done in an earlier pipeline stage (AGU instead of ALU) which means lea >> is faster if its inputs are already available as address expressions and >> is consumed by address expressions; the goal is to avoid the ALU->AGU >> forwarding latency. > > Well, the question is, which is faster: > > lea 8(%esp), %esp > addl $8, %esp > > Basically, all we want to do is add 8 to the stack pointer. And this is > for the x86_32 version of whatever hardware is in use. >
What I'm telling you is that it depends on the context.
An address expression needs to be ready in the AGU; a piece of data comes from the ALU. Whenever something moves from the ALU to the AGU, there is a penalty. There is no penalty to move from the AGU to the ALU, since the ALU is in a later stage.
I *believe* the stack adjustments push/pop are done in the AGU, but I have to double-check.
-hpa
| |