diff --git a/include/pass_mpi_collective.hpp b/include/pass_mpi_collective.hpp
index 44d96f7bcc1aa996854072ce785c7ecbd89df2d6..8bea3b348ef7818284acc3da387605ba11056e07 100644
--- a/include/pass_mpi_collective.hpp
+++ b/include/pass_mpi_collective.hpp
@@ -29,7 +29,8 @@ struct bb_data
std::set<basic_block> post_dom;
std::set<basic_block> dom_front;
std::set<basic_block> post_dom_front;
- int mark; // for graph parkour
+ int mark1; // for graph parkour
+ int mark2;
int collective_rank;
};
@@ -76,13 +77,14 @@ public:
void reset_bb_mark(function *fun);
void alloc_edge_aux(function *fun);
void free_edge_aux(function *fun);
+
+ // loop detection
void mark_edge(function *fun);
- void mark_edge(basic_block bb);
// rank definition
void rank_collective(function *fun);
- int rank_collective(basic_block bb,
- mpi_collective_code mpi_code, int rank);
+ int __rank_collective(basic_block bb,
+ mpi_collective_code mpi_code, int rank);
private:
// MPI function / collectives detections
diff --git a/src/pass_mpi_collective.cpp b/src/pass_mpi_collective.cpp
index 3e8640e365ed0861fd71663def7c8482f1d62def..e605fda7dee02a444fa6caedea9de9e2472368e0 100644
--- a/src/pass_mpi_collective.cpp
+++ b/src/pass_mpi_collective.cpp
@@ -12,6 +12,8 @@
#include <filesystem>
// set
#include <set>
+// list
+#include <list>
// our pass
#include "pass_mpi_collective.hpp"
@@ -259,12 +261,6 @@ void pass_mpi_collective::label_dom(function *fun)
}
}
- //for (basic_block bb2 : *dom)
- //{
- // printf("%d dom %d\n", bb->index, bb2->index);
- //}
- //((bb_data *) bb->aux)->dom = dominated;
-
post_dom_gccvec = get_all_dominated_blocks(CDI_POST_DOMINATORS, bb);
size_post_dom = post_dom_gccvec.length();
for (size_t i = 0; i < size_post_dom; ++i)
@@ -274,11 +270,6 @@ void pass_mpi_collective::label_dom(function *fun)
data->post_dom.insert(post_dom_gccvec[i]);
}
}
- //for (basic_block bb2 : *post_dom)
- //{
- // printf("%d post dom %d\n", bb->index, bb2->index);
- //}
- //((bb_data *) bb->aux)->post_dom = post_dominated;
}
}
@@ -359,7 +350,8 @@ void pass_mpi_collective::reset_bb_mark(function *fun)
FOR_ALL_BB_FN(bb, fun)
{
- ((bb_data *) bb->aux)->mark = 0;
+ ((bb_data *) bb->aux)->mark1 = 0;
+ ((bb_data *) bb->aux)->mark2 = 0;
}
}
@@ -397,27 +389,101 @@ void pass_mpi_collective::free_edge_aux(function *fun)
}
}
+// naive version of mark_edge (work but Omax(2^n))
+//void pass_mpi_collective::mark_edge(function *fun)
+//{
+// mark_edge(ENTRY_BLOCK_PTR_FOR_FN(fun));
+//}
+//
+//void pass_mpi_collective::mark_edge(basic_block bb)
+//{
+// edge e;
+// edge_iterator ei;
+//
+// ((bb_data *) bb->aux)->mark = 1;
+// FOR_EACH_EDGE(e, ei, bb->succs)
+// {
+// if (((bb_data *) e->dest->aux)->mark)
+// {
+// ((edge_data *) e->aux)->loop = true;
+// } else {
+// mark_edge(e->dest);
+// }
+// }
+// ((bb_data *) bb->aux)->mark = 0;
+//}
+
void pass_mpi_collective::mark_edge(function *fun)
{
- mark_edge(ENTRY_BLOCK_PTR_FOR_FN(fun));
-}
-void pass_mpi_collective::mark_edge(basic_block bb)
-{
- edge e;
- edge_iterator ei;
+ edge e1;
+ edge_iterator ei1;
+ std::list<basic_block> nodes1;
+ basic_block bb;
+
+ edge e2;
+ edge_iterator ei2;
+ std::list<basic_block> nodes2;
+ basic_block cursor;
+
+ int index;
- ((bb_data *) bb->aux)->mark = 1;
- FOR_EACH_EDGE(e, ei, bb->succs)
+ index = 0;
+
+ nodes1.push_back(ENTRY_BLOCK_PTR_FOR_FN(fun));
+ ((bb_data *) ENTRY_BLOCK_PTR_FOR_FN(fun)->aux)->mark1 = 1;
+
+ while (!nodes1.empty())
{
- if (((bb_data *) e->dest->aux)->mark)
+ bb = nodes1.front();
+ nodes1.pop_front();
+
+ FOR_EACH_EDGE(e1, ei1, bb->succs)
{
- ((edge_data *) e->aux)->loop = true;
- } else {
- mark_edge(e->dest);
+ if (((bb_data *) e1->dest->aux)->mark1)
+ {
+ continue;
+ }
+ ((bb_data *) bb->aux)->mark1 = 1;
+ nodes1.push_back(e1->dest);
+ }
+
+ if (EDGE_COUNT(bb->preds) < 2)
+ {
+ continue;
+ }
+
+ ++index;
+ nodes2.clear();
+ nodes2.push_back(bb);
+ ((bb_data *) bb->aux)->mark2 = index;
+
+ while (!nodes2.empty())
+ {
+ cursor = nodes2.front();
+ nodes2.pop_front();
+
+ FOR_EACH_EDGE(e2, ei2, cursor->succs)
+ {
+ if (e2->dest == bb)
+ {
+ ((edge_data *) e2->aux)->loop = true;
+ }
+ if (((edge_data *) e2->aux)->loop)
+ {
+ continue;
+ }
+
+ if (((bb_data *) e2->dest->aux)->mark2 >= index)
+ {
+ continue;
+ }
+ ((bb_data *) cursor->aux)->mark2 = index;
+ nodes2.push_back(e2->dest);
+ }
}
}
- ((bb_data *) bb->aux)->mark = 0;
+ reset_bb_mark(fun);
}
void pass_mpi_collective::rank_collective(function *fun)
@@ -427,13 +493,13 @@ void pass_mpi_collective::rank_collective(function *fun)
for (i = 0; i < LAST_AND_UNUSED_MPI_COLLECTIVE_CODE; ++i)
{
- next_rank = rank_collective(EXIT_BLOCK_PTR_FOR_FN(fun),
- (enum mpi_collective_code) i, next_rank);
+ next_rank = __rank_collective(EXIT_BLOCK_PTR_FOR_FN(fun),
+ (enum mpi_collective_code) i, next_rank);
reset_bb_mark(fun);
}
}
-int pass_mpi_collective::rank_collective(basic_block bb,
+int pass_mpi_collective::__rank_collective(basic_block bb,
mpi_collective_code mpi_code, int rank)
{
edge e;
@@ -447,25 +513,27 @@ int pass_mpi_collective::rank_collective(basic_block bb,
{
continue;
}
- if (((bb_data *) e->src->aux)->mark)
+
+ // cache or travell
+ if (((bb_data *) e->src->aux)->mark1)
{
- //printf("caching from %d to %d\n", e->src->index, bb->index);
next_rank = std::max(next_rank,
- ((bb_data *) e->src->aux)->mark);
+ ((bb_data *) e->src->aux)->mark1);
}
else
{
- //printf("going from %d to %d\n", e->src->index, bb->index);
next_rank = std::max(next_rank,
- rank_collective(e->src, mpi_code, rank));
+ __rank_collective(e->src, mpi_code, rank));
}
}
+ // set mpi code
if (((bb_data *) bb->aux)->mpi_code == mpi_code)
{
((bb_data *) bb->aux)->collective_rank = next_rank;
next_rank++;
}
- ((bb_data *) bb->aux)->mark = next_rank;
+ // set mark before recusion
+ ((bb_data *) bb->aux)->mark1 = next_rank;
return next_rank;
}