Loop-im and PRE can hoist loads out of loops, creating artificial dependencies
that inhibit graphite's analysis.
do k = 1,4096
do j = 1,4096
do i = 1,4096
c(i,j)= c(i, j) + a(k,j) * b(i, k)
enddo
enddo
enddo
In the preceding loop body, the a(k,j) load can be hoisted out of the
inner-most loop, which is a valuable optimization, however also one that
creates a cross-iteration dependency, inhibiting polynomial transformation of
the nested-loop. An attempt to tile will fail as graphite will assume a
dependency between loop iterations and fail.
By inhibiting hoists until after graphite has run, we preserve the loop
structure, while allowing hoists to be performed by later passes of loop-im
and/or PRE. If graphite is not enabled on the command line, hoists are
performed early as normal.
This change gives ~8% improvement on spec2017/fotonik3d on my system. It should
also help graphite become applicable to further applications that are blocked
by data dependency.
