Re: [boost] [MSM] Comparison with ad-hoc FSM implementation

Subject: Re: [boost] [MSM] Comparison with ad-hoc FSM implementation
From: Edouard A. (edouard_at_[hidden])
Date: 2009-12-10 09:08:14

On Thu, 10 Dec 2009 05:50:43 -0800, Steven Watanabe <watanabesj_at_[hidden]>
wrote:

> There are many ways that a compiler can handle a switch statement.
> a) sequential if/else
> b) binary search
> c) jump table

And probably other ways, as the cases number rise I've seen only jump
tables in binaries. if/else are hidden with the rest of the code making it
not trivial to see it's not a switch.

I'm not sure I understand what you're referring to when you talk about
binary search? What kind of implementation is this and what are the
benefits?

> Also, a jump table does not necessarily prevent the processor
> from predicting the branch.

Probably, I guess there is much more intelligence in the branch prediction
algorithms than I can imagine.

However, AFAIK when the branch prediction fails it needs to wait for all
the registers to be loaded and the computation to be done to get the good
address. That's a serious stall. Well perhaps it's not a major issue if you
consider the penalty you get for any branch fail, but I think this penalty
has been greatly reduced since Core 2 (compared to P4).

Main point : switches are slower for the same reason a if/then/else maze is
slow.

-- 
EA