This is an optimisation for middle-end worker-partitioning support (used
to support multiple workers on AMD GCN). At present, barriers may be
emitted in cases where they aren't needed and cannot be optimised away.
This patch stops the extraneous barriers from being emitted in the
first place.
One exception to the above (where the barrier is still needed) is for
predicated blocks of code that perform a write to gang-private shared
memory from one worker. We must execute a barrier before other workers
read that shared memory location.
OK for og10 branch?
Julian
ChangeLog
gcc/
* config/gcn/gcn.c (gimple.h): Include.
(gcn_fork_join): Emit barrier for worker-level joins.
* omp-sese.c (find_local_vars_to_propagate): Add writes_gangprivate
bitmap parameter. Set bit for blocks containing gang-private variable
writes.
(worker_single_simple): Don't emit barrier after predicated block.
(worker_single_copy): Don't emit barrier if we're not broadcasting
anything and the block contains no gang-private writes.
(neuter_worker_single): Don't predicate blocks that only contain NOPs
or internal marker functions. Pass has_gangprivate_write argument to
worker_single_copy.
(oacc_do_neutering): Add writes_gangprivate bitmap handling.
---
gcc/config/gcn/gcn.c | 9 +++-
gcc/omp-sese.c | 115 +++++++++++++++++++++++++++++++++----------
2 files changed, 97 insertions(+), 27 deletions(-)
diff --git a/gcc/config/gcn/gcn.c b/gcc/config/gcn/gcn.c
index bf996b461547..35b2ef5e752b 100644
--- a/gcc/config/gcn/gcn.c
+++ b/gcc/config/gcn/gcn.c
@@ -50,6 +50,7 @@
#include "varasm.h"
#include "intl.h"
#include "rtl-iter.h"
+#include "gimple.h"
/* This file should be included last. */
#include "target-def.h"
@@ -4898,9 +4899,15 @@ gcn_oacc_dim_pos (int dim)
/* Implement TARGET_GOACC_FORK_JOIN. */
static bool
-gcn_fork_join (gcall *ARG_UNUSED (call), const int *ARG_UNUSED (dims),
+gcn_fork_join (gcall *call, const int *ARG_UNUSED (dims),
bool ARG_UNUSED (is_fork))
{
+ tree arg = gimple_call_arg (call, 2);
+ unsigned axis = TREE_INT_CST_LOW (arg);
+
+ if (!is_fork && axis == GOMP_DIM_WORKER && dims[axis] != 1)
+ return true;
+
return false;
}
diff --git a/gcc/omp-sese.c b/gcc/omp-sese.c
index 4dd3417066c6..80697358efec 100644
--- a/gcc/omp-sese.c
+++ b/gcc/omp-sese.c
@@ -768,16 +768,19 @@ static void
find_local_vars_to_propagate (parallel_g *par, unsigned outer_mask,
hash_set<tree> *partitioned_var_uses,
hash_set<tree> *gangprivate_vars,
+ bitmap writes_gangprivate,
vec<propagation_set *> *prop_set)
{
unsigned mask = outer_mask | par->mask;
if (par->inner)
find_local_vars_to_propagate (par->inner, mask, partitioned_var_uses,
- gangprivate_vars, prop_set);
+ gangprivate_vars, writes_gangprivate,
+ prop_set);
if (par->next)
find_local_vars_to_propagate (par->next, outer_mask, partitioned_var_uses,
- gangprivate_vars, prop_set);
+ gangprivate_vars, writes_gangprivate,
+ prop_set);
if (!(mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)))
{
@@ -798,8 +801,7 @@ find_local_vars_to_propagate (parallel_g *par, unsigned
outer_mask,
if (!VAR_P (var)
|| is_global_var (var)
|| AGGREGATE_TYPE_P (TREE_TYPE (var))
- || !partitioned_var_uses->contains (var)
- || gangprivate_vars->contains (var))
+ || !partitioned_var_uses->contains (var))
continue;
if (stmt_may_clobber_ref_p (stmt, var))
@@ -813,6 +815,14 @@ find_local_vars_to_propagate (parallel_g *par, unsigned
outer_mask,
fprintf (dump_file, "\n");
}
+ if (gangprivate_vars->contains (var))
+ {
+ /* If we write a gang-private variable, we want a
+ barrier at the end of the block. */
+ bitmap_set_bit (writes_gangprivate, block->index);
+ continue;
+ }
+
if (!(*prop_set)[block->index])
(*prop_set)[block->index] = new propagation_set;
@@ -924,14 +934,6 @@ worker_single_simple (basic_block from, basic_block to,
}
}
}
-
- gsi = gsi_start_bb (skip_block);
-
- decl = builtin_decl_explicit (BUILT_IN_GOACC_BARRIER);
- gimple *acc_bar = gimple_build_call (decl, 0);
-
- gsi_insert_before (&gsi, acc_bar, GSI_SAME_STMT);
- update_stmt (acc_bar);
}
/* This is a copied and renamed omp-low.c:omp_build_component_ref. */
@@ -1009,7 +1011,7 @@ worker_single_copy (basic_block from, basic_block to,
hash_set<tree> *def_escapes_block,
hash_set<tree> *worker_partitioned_uses,
tree record_type, unsigned HOST_WIDE_INT placement,
- bool isolate_broadcasts)
+ bool isolate_broadcasts, bool has_gangprivate_write)
{
/* If we only have virtual defs, we'll have no record type, but we still want
to emit single_copy_start and (particularly) single_copy_end to act as
@@ -1090,14 +1092,19 @@ worker_single_copy (basic_block from, basic_block to,
edge ef = make_edge (from, barrier_block, EDGE_FALSE_VALUE);
ef->probability = et->probability.invert ();
- decl = builtin_decl_explicit (BUILT_IN_GOACC_BARRIER);
- gimple *acc_bar = gimple_build_call (decl, 0);
-
gimple_stmt_iterator bar_gsi = gsi_start_bb (barrier_block);
- gsi_insert_before (&bar_gsi, acc_bar, GSI_NEW_STMT);
-
cond = gimple_build_cond (NE_EXPR, recv_tmp, zero_ptr, NULL_TREE, NULL_TREE);
- gsi_insert_after (&bar_gsi, cond, GSI_NEW_STMT);
+
+ if (record_type != char_type_node || has_gangprivate_write)
+ {
+ decl = builtin_decl_explicit (BUILT_IN_GOACC_BARRIER);
+ gimple *acc_bar = gimple_build_call (decl, 0);
+
+ gsi_insert_before (&bar_gsi, acc_bar, GSI_NEW_STMT);
+ gsi_insert_after (&bar_gsi, cond, GSI_NEW_STMT);
+ }
+ else
+ gsi_insert_before (&bar_gsi, cond, GSI_NEW_STMT);
edge et2 = split_block (barrier_block, cond);
et2->flags &= ~EDGE_FALLTHRU;
@@ -1259,7 +1266,8 @@ neuter_worker_single (parallel_g *par, unsigned
outer_mask,
bitmap worker_single, bitmap vector_single,
vec<propagation_set *> *prop_set,
hash_set<tree> *partitioned_var_uses,
- blk_offset_map_t *blk_offset_map)
+ blk_offset_map_t *blk_offset_map,
+ bitmap writes_gangprivate)
{
unsigned mask = outer_mask | par->mask;
@@ -1345,19 +1353,69 @@ neuter_worker_single (parallel_g *par, unsigned
outer_mask,
(*prop_set)[block->index] = 0;
}
- tree record_type = (tree) block->aux;
+ bool only_marker_fns = true;
+ bool join_block = false;
+
+ for (gimple_stmt_iterator gsi = gsi_start_bb (block);
+ !gsi_end_p (gsi);
+ gsi_next (&gsi))
+ {
+ gimple *stmt = gsi_stmt (gsi);
+ if (gimple_code (stmt) == GIMPLE_CALL
+ && gimple_call_internal_p (stmt, IFN_UNIQUE))
+ {
+ enum ifn_unique_kind k = ((enum ifn_unique_kind)
+ TREE_INT_CST_LOW (gimple_call_arg (stmt, 0)));
+ if (k != IFN_UNIQUE_OACC_PRIVATE
+ && k != IFN_UNIQUE_OACC_JOIN
+ && k != IFN_UNIQUE_OACC_FORK
+ && k != IFN_UNIQUE_OACC_HEAD_MARK
+ && k != IFN_UNIQUE_OACC_TAIL_MARK)
+ only_marker_fns = false;
+ else if (k == IFN_UNIQUE_OACC_JOIN)
+ /* The JOIN marker is special in that it *cannot* be
+ predicated for worker zero, because it may be lowered
+ to a barrier instruction and all workers must typically
+ execute that barrier. We shouldn't be doing any
+ broadcasts from the join block anyway. */
+ join_block = true;
+ }
+ else if (gimple_code (stmt) == GIMPLE_CALL
+ && gimple_call_internal_p (stmt, IFN_GOACC_LOOP))
+ /* Empty. */;
+ else if (gimple_nop_p (stmt))
+ /* Empty. */;
+ else
+ only_marker_fns = false;
+ }
+
+ /* We can skip predicating this block for worker zero if the only
+ thing it contains is marker functions that will be removed in the
+ oaccdevlow pass anyway.
+ Don't do this if the block has (any) phi nodes, because those
+ might define SSA names that need broadcasting.
+ TODO: We might be able to skip transforming blocks that only
+ contain some other trivial statements too. */
+ if (only_marker_fns && !phi_nodes (block))
+ continue;
+
+ gcc_assert (!join_block);
if (has_defs)
{
+ tree record_type = (tree) block->aux;
auto off_rngalloc = blk_offset_map->get (block);
gcc_assert (!record_type || off_rngalloc);
unsigned HOST_WIDE_INT offset
= off_rngalloc ? off_rngalloc->first : 0;
bool range_allocated
= off_rngalloc ? off_rngalloc->second : true;
+ bool has_gangprivate_write
+ = bitmap_bit_p (writes_gangprivate, block->index);
worker_single_copy (block, block, &def_escapes_block,
&worker_partitioned_uses, record_type,
- offset, !range_allocated);
+ offset, !range_allocated,
+ has_gangprivate_write);
}
else
worker_single_simple (block, block, &def_escapes_block);
@@ -1394,10 +1452,12 @@ neuter_worker_single (parallel_g *par, unsigned
outer_mask,
if (par->inner)
neuter_worker_single (par->inner, mask, worker_single, vector_single,
- prop_set, partitioned_var_uses, blk_offset_map);
+ prop_set, partitioned_var_uses, blk_offset_map,
+ writes_gangprivate);
if (par->next)
neuter_worker_single (par->next, outer_mask, worker_single, vector_single,
- prop_set, partitioned_var_uses, blk_offset_map);
+ prop_set, partitioned_var_uses, blk_offset_map,
+ writes_gangprivate);
}
@@ -1595,11 +1655,13 @@ oacc_do_neutering (unsigned HOST_WIDE_INT bounds_lo,
hash_set<tree> partitioned_var_uses;
hash_set<tree> gangprivate_vars;
+ auto_bitmap writes_gangprivate;
find_gangprivate_vars (&gangprivate_vars);
find_partitioned_var_uses (par, mask, &partitioned_var_uses);
find_local_vars_to_propagate (par, mask, &partitioned_var_uses,
- &gangprivate_vars, &prop_set);
+ &gangprivate_vars, writes_gangprivate,
+ &prop_set);
FOR_ALL_BB_FN (bb, cfun)
{
@@ -1747,7 +1809,8 @@ oacc_do_neutering (unsigned HOST_WIDE_INT bounds_lo,
sbitmap_vector_free (reachable);
neuter_worker_single (par, mask, worker_single, vector_single, &prop_set,
- &partitioned_var_uses, &blk_offset_map);
+ &partitioned_var_uses, &blk_offset_map,
+ writes_gangprivate);
prop_set.release ();
--
2.23.0
