Re: Cost of handling misaligned access

Michael S <already5chosen@yahoo.com> writes:

This resulted in just 12 instructions to handle 32 tests.
 

>
That sounds suboptimal.
By unrolling outer loop by 2 or 3 you can greatly reduce the number of
memory accesses per comparison.

Looking at the inner loop code shown in
<2025Feb6.113049@mips.complang.tuwien.ac.at>, the 12 instructions do
not include the loop overhead and are already unrolled by a factor of
4 (32 for the scalar code).  The loop overhead is 3 instructions, for
a total of 15 instructions per iteration.

The speed up would depend on specific
microarchiture, but I would guess that at least 1.2x speedup is here.

Even if you completely eliminate the loop overhead, the number of
instructions is reduced by at most a factor 1.25, and I expect that
the speedup from further unrolling is a factor of at most 1 on most
CPUs (factor <1 can come from handling the remaining elements slowly,
which does not seem unlikely for code coming out of gcc and clang).

- anton
--
'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'
  Mitch Alsup, <c17fcd89-f024-40e7-a594-88a85ac10d20o@googlegroups.com>