[27/40] openacc: Handle internal function calls in pass_lim

Message ID 20211215155447.19379-28-frederik@codesourcery.com
State New
Headers
Series OpenACC "kernels" Improvements |

Commit Message

Frederik Harwath Dec. 15, 2021, 3:54 p.m. UTC
  The loop invariant motion pass correctly refuses to move statements
out of a loop if any other statement in the loop is unanalyzable.  The
pass does not know how to handle the OpenACC internal function calls
which was not necessary until recently when the OpenACC device
lowering pass was moved to a later position in the pass pipeline.

This commit changes pass_lim to ignore the OpenACC internal function
calls which do not contain any memory references. The hoisting enabled
by this change can be useful for the data-dependence analysis in
Graphite; for instance, in the outlined functions for OpenACC regions,
all invariant accesses to the ".omp_data_i" struct should be hoisted
out of the OpenACC loop.  This is particularly important for variables
that were scalars in the original loop and which have been turned into
accesses to the struct by the outlining process.  Not hoisting those
can prevent scalar evolution analysis which is crucial for Graphite.
Since any hoisting that introduces intermediate names - and hence,
"fake" dependences - inside the analyzed nest can be harmful to
data-dependence analysis, a flag to restrict the hoisting in OpenACC
functions is added to the pass. The pass instance that executes before
Graphite now runs with this flag set to true and the pass instance
after Graphite runs unrestricted.

A more precise way of selecting the statements for which hoisting
should be enabled is left for a future improvement.

gcc/ChangeLog:
        * passes.def: Set restrict_oacc_hoisting to true for the early
        pass_lim instance.
        * tree-ssa-loop-im.c (movement_possibility): Add
        restrict_oacc_hoisting flag to function; restrict movement if set.
        (compute_invariantness): Add restrict_oacc_hoisting flag and pass it on.
        (gather_mem_refs_stmt): Skip IFN_GOACC_LOOP and IFN_UNIQUE
        calls.
        (loop_invariant_motion_in_fun): Add restrict_oacc_hoisting flag and
        pass it on.
        (pass_lim::execute): Pass on new flags.
        * tree-ssa-loop-manip.h (loop_invariant_motion_in_fun): Adjust
        declaration.
        * gimple-loop-interchange.cc (pass_linterchange::execute): Adjust call to
        loop_invariant_motion_in_fun.
---
 gcc/gimple-loop-interchange.cc |  2 +-
 gcc/passes.def                 |  2 +-
 gcc/tree-ssa-loop-im.c         | 57 ++++++++++++++++++++++++++++------
 gcc/tree-ssa-loop-manip.h      |  2 +-
 4 files changed, 51 insertions(+), 12 deletions(-)

--
2.33.0

-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
  

Patch

diff --git a/gcc/gimple-loop-interchange.cc b/gcc/gimple-loop-interchange.cc
index ccd5083145f8..7c9b7b2345fa 100644
--- a/gcc/gimple-loop-interchange.cc
+++ b/gcc/gimple-loop-interchange.cc
@@ -2107,7 +2107,7 @@  pass_linterchange::execute (function *fun)
   if (changed_p)
     {
       unsigned todo = TODO_update_ssa_only_virtuals;
-      todo |= loop_invariant_motion_in_fun (cfun, false);
+      todo |= loop_invariant_motion_in_fun (cfun, false, false);
       scev_reset ();
       return todo;
     }
diff --git a/gcc/passes.def b/gcc/passes.def
index 681392f8f79f..1da9382bac53 100644
--- a/gcc/passes.def
+++ b/gcc/passes.def
@@ -250,7 +250,7 @@  along with GCC; see the file COPYING3.  If not see
       NEXT_PASS (pass_cse_sincos);
       NEXT_PASS (pass_optimize_bswap);
       NEXT_PASS (pass_laddress);
-      NEXT_PASS (pass_lim);
+      NEXT_PASS (pass_lim, true /* restrict_oacc_hoisting */);
       NEXT_PASS (pass_walloca, false);
       NEXT_PASS (pass_pre);
       NEXT_PASS (pass_sink_code);
diff --git a/gcc/tree-ssa-loop-im.c b/gcc/tree-ssa-loop-im.c
index 4b187c2cdafe..466dc494fb52 100644
--- a/gcc/tree-ssa-loop-im.c
+++ b/gcc/tree-ssa-loop-im.c
@@ -47,6 +47,8 @@  along with GCC; see the file COPYING3.  If not see
 #include "builtins.h"
 #include "tree-dfa.h"
 #include "dbgcnt.h"
+#include "graphite-oacc.h"
+#include "internal-fn.h"

 /* TODO:  Support for predicated code motion.  I.e.

@@ -327,11 +329,23 @@  enum move_pos
    Otherwise return MOVE_IMPOSSIBLE.  */

 enum move_pos
-movement_possibility (gimple *stmt)
+movement_possibility (gimple *stmt, bool restrict_oacc_hoisting)
 {
   tree lhs;
   enum move_pos ret = MOVE_POSSIBLE;

+  if (restrict_oacc_hoisting && oacc_get_fn_attrib (cfun->decl)
+      && gimple_code (stmt) == GIMPLE_ASSIGN)
+    {
+      tree rhs = gimple_assign_rhs1 (stmt);
+
+      if (TREE_CODE (rhs) == VIEW_CONVERT_EXPR)
+       rhs = TREE_OPERAND (rhs, 0);
+
+      if (TREE_CODE (rhs) == ARRAY_REF)
+         return MOVE_IMPOSSIBLE;
+    }
+
   if (flag_unswitch_loops
       && gimple_code (stmt) == GIMPLE_COND)
     {
@@ -981,7 +995,7 @@  rewrite_bittest (gimple_stmt_iterator *bsi)
    statements.  */

 static void
-compute_invariantness (basic_block bb)
+compute_invariantness (basic_block bb, bool restrict_oacc_hoisting)
 {
   enum move_pos pos;
   gimple_stmt_iterator bsi;
@@ -1009,7 +1023,7 @@  compute_invariantness (basic_block bb)
       {
        stmt = gsi_stmt (bsi);

-       pos = movement_possibility (stmt);
+       pos = movement_possibility (stmt, restrict_oacc_hoisting);
        if (pos == MOVE_IMPOSSIBLE)
          continue;

@@ -1040,7 +1054,7 @@  compute_invariantness (basic_block bb)
     {
       stmt = gsi_stmt (bsi);

-      pos = movement_possibility (stmt);
+      pos = movement_possibility (stmt, restrict_oacc_hoisting);
       if (pos == MOVE_IMPOSSIBLE)
        {
          if (nonpure_call_p (stmt))
@@ -1465,6 +1479,13 @@  gather_mem_refs_stmt (class loop *loop, gimple *stmt)
   if (!gimple_vuse (stmt))
     return;

+  /* The expansion of those OpenACC internal function calls which occurs in a
+   * later pass does not introduce any memory references. Hence it is safe to
+   * ignore them. */
+  if (gimple_call_internal_p (stmt, IFN_GOACC_LOOP)
+      || gimple_call_internal_p (stmt, IFN_UNIQUE))
+    return;
+
   mem = simple_mem_ref_in_stmt (stmt, &is_stored);
   if (!mem && is_gimple_assign (stmt))
     {
@@ -1506,7 +1527,7 @@  gather_mem_refs_stmt (class loop *loop, gimple *stmt)
       ao_ref_alias_set (&aor);
       HOST_WIDE_INT offset, size, max_size;
       poly_int64 saved_maxsize = aor.max_size, mem_off;
-      tree mem_base;
+      tree mem_base = NULL;
       bool ref_decomposed;
       if (aor.max_size_known_p ()
          && aor.offset.is_constant (&offset)
@@ -3244,7 +3265,8 @@  tree_ssa_lim_finalize (void)
    Only perform store motion if STORE_MOTION is true.  */

 unsigned int
-loop_invariant_motion_in_fun (function *fun, bool store_motion)
+loop_invariant_motion_in_fun (function *fun, bool store_motion,
+                             bool restrict_oacc_hoisting)
 {
   unsigned int todo = 0;

@@ -3262,7 +3284,7 @@  loop_invariant_motion_in_fun (function *fun, bool store_motion)
   /* For each statement determine the outermost loop in that it is
      invariant and cost for computing the invariant.  */
   for (int i = 0; i < n; ++i)
-    compute_invariantness (BASIC_BLOCK_FOR_FN (fun, rpo[i]));
+    compute_invariantness (BASIC_BLOCK_FOR_FN (fun, rpo[i]), restrict_oacc_hoisting);

   /* Execute store motion.  Force the necessary invariants to be moved
      out of the loops as well.  */
@@ -3309,13 +3331,21 @@  class pass_lim : public gimple_opt_pass
 {
 public:
   pass_lim (gcc::context *ctxt)
-    : gimple_opt_pass (pass_data_lim, ctxt)
+    : gimple_opt_pass (pass_data_lim, ctxt), restrict_oacc_hoisting (false)
   {}

+  void set_pass_param (unsigned int n, bool param)
+    {
+      gcc_assert (n == 0);
+      restrict_oacc_hoisting = param;
+    }
+
   /* opt_pass methods: */
   opt_pass * clone () { return new pass_lim (m_ctxt); }
   virtual bool gate (function *) { return flag_tree_loop_im != 0; }
   virtual unsigned int execute (function *);
+private:
+  bool restrict_oacc_hoisting;

 }; // class pass_lim

@@ -3328,7 +3358,16 @@  pass_lim::execute (function *fun)

   if (number_of_loops (fun) <= 1)
     return 0;
-  unsigned int todo = loop_invariant_motion_in_fun (fun, flag_move_loop_stores);
+
+  bool store_motion = flag_move_loop_stores;
+  /* TODO Enabling store motion in OpenACC kernel functions requires further
+     handling of the OpenACC internal function calls.  It can also be harmful
+     to data-dependence analysis. Keep it disabled for now. */
+  if (oacc_function_p (cfun) && graphite_analyze_oacc_target_region_type_p (cfun))
+    store_motion = false;
+
+  unsigned int todo = loop_invariant_motion_in_fun (fun, store_motion,
+                                                   restrict_oacc_hoisting);

   if (!in_loop_pipeline)
     loop_optimizer_finalize ();
diff --git a/gcc/tree-ssa-loop-manip.h b/gcc/tree-ssa-loop-manip.h
index 4f604e1bd24a..864fb9f1d355 100644
--- a/gcc/tree-ssa-loop-manip.h
+++ b/gcc/tree-ssa-loop-manip.h
@@ -53,7 +53,7 @@  extern void tree_transform_and_unroll_loop (class loop *, unsigned,
                                            transform_callback, void *);
 extern void tree_unroll_loop (class loop *, unsigned, tree_niter_desc *);
 extern tree canonicalize_loop_ivs (class loop *, tree *, bool);
-extern unsigned int loop_invariant_motion_in_fun (function *, bool);
+extern unsigned int loop_invariant_motion_in_fun (function *, bool, bool);


 #endif /* GCC_TREE_SSA_LOOP_MANIP_H */