What is the motivation for pre/main/post loops?

[John Rose]

As Roland says, a pre-loop can perform a slow iteration over a part of the iteration space that is not provably safe to do RCE on (= range check elimination).  That is why one technical name of our RCE is “iteration range splitting”:  You find a middle part of the iteration range (range of the index variable) that can be proven safe using loop-invariant computations.  Then you execute that “comfortable” range in the mid-loop.  The pre- and post-loops handle the uncomfortable parts of the iteration range; hopefully they are small.

ALSO, the VERY FIRST execution of the pre-loop is very useful to “shake out” one-time-only checks, such as a null checks of a loop-invariant value.  You might think such checks could be hoisted (predicated) but it is more robust to run the loop just once with all checks enabled, and then run the rest of the iterations with redundant checks suppressed.  Put another way, the pre-loop contains (one or more) peeled iterations of the loop, just to get some redundant checks covered by the dominating early peeled iteration.


On 8 Feb 2024, at 4:45, Roland Westrelin wrote:

>> I would like to hear if there are any other reasons for a pre-loop.
>
> PhaseIdealLoop::do_range_check() does need a pre loop.
>
> Roland.