On 10/28/25 1:31 PM, Andrew MacLeod wrote:
I added a chunk to print the range being returned. With your path
ranger, R is [irange] BITMAP_WORD [1, +INF] , which I assuem si the
value you are looking for?
Yea, we just need to know the value is nonzero. Essentially the loop
doesn't terminate if presented with a zero value and thus transformation
into a ctz instruction wouldn't be appropriate. Once we prove the value
is known to be nonzero, then we zap the loop entirely and replace it
with a ctz. The loop branch happens to be the single biggest source of
branch mispredicts for 502.gcc input #2 on our design (and probably
others given there really isn't a pattern for a branch predictor to lock
onto).
Jeff
