gcc/0312-Add-late-slp-vectorization-pass-with-additional-chec.patch

From 9df4a0bd76299734ae47f2f4e236b10f6c156994 Mon Sep 17 00:00:00 2001
From: d84370931 <dementiev.daniil@h-partners.com>
Date: Thu, 14 Nov 2024 17:08:40 +0800
Subject: [PATCH 3/8] Add late slp vectorization pass with additional checks.

Add expansion of data reference offset using affine trees to check
if data references may alias.

Add check if a group of interleaving data references is smaller than
max vector register size.

Add operands swap for commutative operations.
Swapping operands is necessary for better vector constructing.
For example for operations
  _1 = a * b;
  _2 = b * c;
Construction vectors (a, c) * (b, b) is more profitable
than (a, b) * (b, c).

Add tests and special param flags for each check:
  --param=vect-addr-expand-for-alias-check={0,1}
  --param=vect-swap-operands={0,1}
  --param=vect-register-size-check={0,1}

Add enabling flag for late slp pass:
  -ftree-slp-late
---
 gcc/common.opt                                |  4 ++
 gcc/params.opt                                | 12 ++++++
 gcc/passes.def                                |  4 ++
 gcc/testsuite/gcc.dg/vect/vect-alias-expand.c | 12 ++++++
 gcc/testsuite/gcc.dg/vect/vect-op-swap.c      | 10 +++++
 gcc/testsuite/gcc.dg/vect/vect-regsize.c      | 18 +++++++++
 gcc/timevar.def                               |  1 +
 gcc/tree-data-ref.cc                          | 12 ++++++
 gcc/tree-pass.h                               |  1 +
 gcc/tree-vect-data-refs.cc                    | 15 +++++++
 gcc/tree-vect-slp.cc                          | 28 +++++++++++++
 gcc/tree-vectorizer.cc                        | 39 +++++++++++++++++++
 12 files changed, 156 insertions(+)
 create mode 100644 gcc/testsuite/gcc.dg/vect/vect-alias-expand.c
 create mode 100644 gcc/testsuite/gcc.dg/vect/vect-op-swap.c
 create mode 100644 gcc/testsuite/gcc.dg/vect/vect-regsize.c

diff --git a/gcc/common.opt b/gcc/common.opt
index 78cfc333a..c3c64ceaf 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -3268,6 +3268,10 @@ ftree-slp-transpose-vectorize
 Common Var(flag_tree_slp_transpose_vectorize) Optimization Init(0)
 Enable basic block vectorization (SLP) for transposed stores and loads on trees.
 
+ftree-slp-late
+Common Var(flag_slp_late) Init(0) Optimization
+Enable additional SLP vectorization pass after reassociation.
+
 fvect-cost-model=
 Common Joined RejectNegative Enum(vect_cost_model) Var(flag_vect_cost_model) Init(VECT_COST_MODEL_DEFAULT) Optimization
 -fvect-cost-model=[unlimited|dynamic|cheap|very-cheap]	Specifies the cost model for vectorization.
diff --git a/gcc/params.opt b/gcc/params.opt
index 3ddfaf5b2..bb4dc1825 100644
--- a/gcc/params.opt
+++ b/gcc/params.opt
@@ -1213,6 +1213,18 @@ The maximum factor which the loop vectorizer applies to the cost of statements i
 Common Joined UInteger Var(param_vect_induction_float) Init(1) IntegerRange(0, 1) Param Optimization
 Enable loop vectorization of floating point inductions.
 
+-param=vect-swap-operands=
+Common Joined UInteger Var(param_vect_swap_operands) Init(0) IntegerRange(0, 1) Param Optimization
+Enable swapping operands for commutative operations in vectorization analysis.
+
+-param=addr-expand-for-alias-check=
+Common Joined UInteger Var(param_addr_expand_for_alias_check) Init(0) IntegerRange(0, 1) Param Optimization
+Enable data reference address expansion for alias check.
+
+-param=vect-register-size-check=
+Common Joined UInteger Var(param_vect_register_size_check) Init(0) IntegerRange(0, 1) Param Optimization
+Enable checking if a group of interleaving data references may not fit in vector register.
+
 -param=vrp1-mode=
 Common Joined Var(param_vrp1_mode) Enum(vrp_mode) Init(VRP_MODE_VRP) Param Optimization
 --param=vrp1-mode=[vrp|ranger] Specifies the mode VRP1 should operate in.
diff --git a/gcc/passes.def b/gcc/passes.def
index e945af96a..529cc5093 100644
--- a/gcc/passes.def
+++ b/gcc/passes.def
@@ -337,6 +337,10 @@ along with GCC; see the file COPYING3.  If not see
       NEXT_PASS (pass_lower_switch);
       NEXT_PASS (pass_cse_reciprocals);
       NEXT_PASS (pass_reassoc, false /* early_p */);
+      NEXT_PASS (pass_slp_vectorize_late);
+      PUSH_INSERT_PASSES_WITHIN (pass_slp_vectorize_late)
+	  NEXT_PASS (pass_slp_vectorize);
+      POP_INSERT_PASSES ()
       NEXT_PASS (pass_strength_reduction);
       NEXT_PASS (pass_split_paths);
       NEXT_PASS (pass_tracer);
diff --git a/gcc/testsuite/gcc.dg/vect/vect-alias-expand.c b/gcc/testsuite/gcc.dg/vect/vect-alias-expand.c
new file mode 100644
index 000000000..a68f4baf8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-alias-expand.c
@@ -0,0 +1,12 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -ftree-vectorize --param=addr-expand-for-alias-check=1 -fdump-tree-slp-details" } */
+
+extern float arr[2][2];
+
+void foo (int i, int j, float a, float b)
+{
+  arr[i][j] *= a;
+  arr[i][j+1] *= b;
+}
+
+/* { dg-final { scan-tree-dump "Basic block will be vectorized using SLP" "slp2" } } */
diff --git a/gcc/testsuite/gcc.dg/vect/vect-op-swap.c b/gcc/testsuite/gcc.dg/vect/vect-op-swap.c
new file mode 100644
index 000000000..4872dc414
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-op-swap.c
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -ftree-vectorize --param=vect-swap-operands=1 -fdump-tree-slp-details" } */
+
+void foo (float *res, float a, float b, float c)
+{
+  res[0] = a * b;
+  res[1] = b * c;
+}
+
+/* { dg-final { scan-tree-dump "Swapped operands for" "slp2" } } */
diff --git a/gcc/testsuite/gcc.dg/vect/vect-regsize.c b/gcc/testsuite/gcc.dg/vect/vect-regsize.c
new file mode 100644
index 000000000..bcd81e6df
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-regsize.c
@@ -0,0 +1,18 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -ftree-vectorize --param=vect-register-size-check=1 -fdump-tree-slp-details" } */
+
+extern float arr[256][256][1024];
+
+void foo (int i, int j, float a, float b)
+{
+  arr[i][j][0] += a;
+  arr[i][j][1] += b;
+  arr[i][j+1][0] += a;
+  arr[i][j+1][1] += b;
+  arr[i+1][j][0] += a;
+  arr[i+1][j][1] += b;
+  arr[i+1][j+1][0] += a;
+  arr[i+1][j+1][1] += b;
+}
+
+/* { dg-final { scan-tree-dump "Basic block will be vectorized using SLP" "slp2" } } */
diff --git a/gcc/timevar.def b/gcc/timevar.def
index fc2b1e1e7..7560e930a 100644
--- a/gcc/timevar.def
+++ b/gcc/timevar.def
@@ -205,6 +205,7 @@ DEFTIMEVAR (TV_SCALAR_CLEANUP        , "scalar cleanup")
 DEFTIMEVAR (TV_TREE_PARALLELIZE_LOOPS, "tree parallelize loops")
 DEFTIMEVAR (TV_TREE_VECTORIZATION    , "tree vectorization")
 DEFTIMEVAR (TV_TREE_SLP_VECTORIZATION, "tree slp vectorization")
+DEFTIMEVAR (TV_TREE_LATE_SLP         , "late slp vectorization")
 DEFTIMEVAR (TV_GRAPHITE              , "Graphite")
 DEFTIMEVAR (TV_GRAPHITE_TRANSFORMS   , "Graphite loop transforms")
 DEFTIMEVAR (TV_GRAPHITE_DATA_DEPS    , "Graphite data dep analysis")
diff --git a/gcc/tree-data-ref.cc b/gcc/tree-data-ref.cc
index a05073c51..5eb4ac102 100644
--- a/gcc/tree-data-ref.cc
+++ b/gcc/tree-data-ref.cc
@@ -3021,6 +3021,18 @@ dr_may_alias_p (const struct data_reference *a, const struct data_reference *b,
       get_inner_reference_aff (DR_REF (b), &off2, &size2);
       aff_combination_scale (&off1, -1);
       aff_combination_add (&off2, &off1);
+
+      if (param_addr_expand_for_alias_check)
+	{
+	  using tree_expand_map_t = hash_map<tree, name_expansion *>;
+	  /* Cache used by aff_combination_expand.  */
+	  tree_expand_map_t *cache = NULL;
+
+	  if (off2.n)
+	    aff_combination_expand (&off2, &cache);
+	  free_affine_expand_cache (&cache);
+	}
+
       if (aff_comb_cannot_overlap_p (&off2, size1, size2))
 	return false;
     }
diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
index 18b0f8022..2ed79f353 100644
--- a/gcc/tree-pass.h
+++ b/gcc/tree-pass.h
@@ -390,6 +390,7 @@ extern gimple_opt_pass *make_pass_slp_vectorize (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_complete_unroll (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_complete_unrolli (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_pre_slp_scalar_cleanup (gcc::context *ctxt);
+extern gimple_opt_pass *make_pass_slp_vectorize_late (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_parallelize_loops (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_loop_prefetch (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_iv_optimize (gcc::context *ctxt);
diff --git a/gcc/tree-vect-data-refs.cc b/gcc/tree-vect-data-refs.cc
index aae7f62f3..ee58c8f6c 100644
--- a/gcc/tree-vect-data-refs.cc
+++ b/gcc/tree-vect-data-refs.cc
@@ -3234,6 +3234,21 @@ vect_analyze_data_ref_accesses (vec_info *vinfo,
 		      != type_size_a))
 		break;
 
+	      if (param_vect_register_size_check)
+		{
+		  tree scalar_type = TREE_TYPE (DR_REF (dra));
+		  tree vec_type = get_related_vectype_for_scalar_type (
+		      vinfo->vector_mode, scalar_type);
+		  poly_uint64 vec_size = TYPE_VECTOR_SUBPARTS (vec_type);
+
+		  /* If we have a large interleaving group (especially a group
+		     of loads with gaps) that does not fit in vector register,
+		     we should split this group to chunks we support.  */
+		  if (maybe_ge (((unsigned HOST_WIDE_INT)init_b - init_prev)
+				/ type_size_a, vec_size))
+		    break;
+		}
+
 	      /* If the step (if not zero or non-constant) is smaller than the
 		 difference between data-refs' inits this splits groups into
 		 suitable sizes.  */
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index fbd638333..79026fb5b 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -687,6 +687,34 @@ vect_get_and_check_slp_defs (vec_info *vinfo, unsigned char swap,
   if (first)
     return 0;
 
+  /* If different statements in the group of commutative operations
+     have the same arguments but in different places, swap them to
+     group the same operands in one vector.
+
+     Check if swapping is enabled, operation is commutative and has
+     two operands of the same type.
+     If one of the operands in current statement match the operand
+     on another place of the first statement in the group we
+     swap operands in current statement.  */
+  if (param_vect_swap_operands && commutative_op == 0 && !first
+      && is_a <bb_vec_info> (vinfo) && number_of_oprnds == 2
+      && vect_def_types_match (dts[0], dts[1]))
+    {
+      slp_oprnd_info oprnd_info0 = (*oprnds_info)[0];
+      slp_oprnd_info oprnd_info1 = (*oprnds_info)[1];
+      if (oprnd_info1->ops[stmt_num] == oprnd_info0->ops[0]
+	  || oprnd_info0->ops[stmt_num] == oprnd_info1->ops[0])
+      {
+	std::swap (oprnd_info0->def_stmts[stmt_num],
+		   oprnd_info1->def_stmts[stmt_num]);
+	std::swap (oprnd_info0->ops[stmt_num],
+		   oprnd_info1->ops[stmt_num]);
+	if (dump_enabled_p ())
+	  dump_printf_loc (MSG_NOTE, vect_location,
+			   "Swapped operands for %G", stmt_info->stmt);
+      }
+    }
+
   /* Now match the operand definition types to that of the first stmt.  */
   for (i = 0; i < number_of_oprnds;)
     {
diff --git a/gcc/tree-vectorizer.cc b/gcc/tree-vectorizer.cc
index a63fa3912..c363ce490 100644
--- a/gcc/tree-vectorizer.cc
+++ b/gcc/tree-vectorizer.cc
@@ -1524,6 +1524,45 @@ make_pass_slp_vectorize (gcc::context *ctxt)
   return new pass_slp_vectorize (ctxt);
 }
 
+/*  The late SLP vectorization pass.  */
+
+namespace {
+
+const pass_data pass_data_slp_vectorize_late =
+{
+  GIMPLE_PASS, /* type.  */
+  "slp_late", /* name.  */
+  OPTGROUP_NONE, /* optinfo_flags.  */
+  TV_TREE_LATE_SLP, /* tv_id.  */
+  PROP_cfg, /* properties_required.  */
+  0, /* properties_provided.  */
+  0, /* properties_destroyed.  */
+  0, /* todo_flags_start.  */
+  0, /* todo_flags_finish.  */
+};
+
+class pass_slp_vectorize_late : public gimple_opt_pass
+{
+public:
+  pass_slp_vectorize_late (gcc::context *ctxt)
+    : gimple_opt_pass (pass_data_slp_vectorize_late, ctxt)
+  {}
+
+  /* opt_pass methods: */
+  virtual bool gate (function *)
+  {
+    return flag_slp_late != 0;
+  }
+
+}; // class pass_slp_vectorize_late
+
+} // anon namespace
+
+gimple_opt_pass *
+make_pass_slp_vectorize_late (gcc::context *ctxt)
+{
+  return new pass_slp_vectorize_late (ctxt);
+}
 
 /* Increase alignment of global arrays to improve vectorization potential.
    TODO:
-- 
2.33.0