I see that gcc.dg/tree-ssa/ssa-pre-21.c fails for avr.
The computation of k + 1L is done as unsigned int (16 bit),
but later ++k is performed as unsigned long (32 bit), and thus
array[k+1] is not re-used.

Is this to be considered a big in the optimizers for not
optimizing this, or in the testsuite for insisting that it
should be?

FWIW, equivalent behaviour can be seen natively on
i686-pc-linux-gnu by replacing "long" with "long long",
as shown in the attachment.