http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58453
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
loop distribution distributes this into
q2 = ( g2a(i+1) * g31a(i+1) * v1(i+1,j,k)
1 - g2a(i ) * g31a(i ) * v1(i ,j,k) )
2 * dvl1ai(i)
3 + ( g32a(j+1) * v2(i,j+1,k)
4 - g32a(j ) * v2(i,j ,k) )
5 * g2bi(i) * dvl2ai(j)
6 + ( v3(i,j,k+1) - v3(i,j,k) )
7 * g31bi(i) * g32bi(j) * dvl3ai(k)
q2 = q2 * q1
e(i,j,k) = ( 1.0 - q2 ) / ( 1.0 + q2 ) * e(i,j,k)
and a memcpy for
dlo(i,j,k) = d(i,j,k)
and
eod(i,j,k) = e(i,j,k) / d(i,j,k)
(re-computing e(i,j,k) instead of loading it from the stored value - a known
deficiency)
This doesn't look wrong on the first glance (but it's probably slower).
What the revision in question changed was remove some very odd code
from rdg_flag_uses:
- if (gimple_code (stmt) != GIMPLE_PHI)
- {
- if ((use_p = gimple_vuse_op (stmt)) != NULL_USE_OPERAND_P)
- {
- tree use = USE_FROM_PTR (use_p);
-
- if (TREE_CODE (use) == SSA_NAME
- && !SSA_NAME_IS_DEFAULT_DEF (use))
- {
- gimple def_stmt = SSA_NAME_DEF_STMT (use);
- int v = rdg_vertex_for_stmt (rdg, def_stmt);
-
- if (v >= 0
- && !already_processed_vertex_p (processed, v))
- rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
- processed);
- }
- }
- }
that just doesn't make sense, but it likely made sure everything ended up
in a single partition.
Does the benchmark fail if you build with -ftree-loop-distribution
-fno-tree-loop-distribute-patterns? (it should emit a loop instead of the
memcpy call)
