Re: Any more graceful way to saturate adds?

On Apr 18, 3:57 pm, dave <spamt...@[EMAIL PROTECTED]
> wrote:
> On the real 8088 IBM PC the cycle count is total bytes in the
> instruction and data being referenced times the wait states.  The PC
> had so many waits on memory access that only a few instructions such
> as multiply and divide would take more CPU cycles than the fetch.

....but they still had to execute.  Just because an opcode takes 4 or 8
cycles to fetch doesn't mean its execution time is magically erased.
Since the (tiny) prefetch queue can be full after a slow operation,
optimization still matters.

I originally asked about the topic because a jump clears the prefetch
queue, and was trying to find a way to clamp without jumping.  I have
to ask, though:  The penalty on my CPU for jumping is 1. clearing the
prefetch queue and 2. 17 cycles.  I will have to look at Terje's
suggestions and see if the solutions he provided actually beat that,
which may have been what he was referring to when he wrote "unless
clamping is very common, it is hard to beat the branch
predictors!"