Pool alignment information for common bases

[Richard Sandiford <richard dot sandiford at linaro dot org>]

We know that if a vectorised loop is reached, all statements in that
loop execute at least once, so it should be safe to pool the alignment
information for all the statements we're vectorising.  The only catch is
that DR_REFs for masked loads and stores only occur if the mask value is
nonzero.  For example, in:

  struct s __attribute__((aligned(32))) {
    int misaligner;
    int array[N];
  };

  int *ptr;
  for (int i = 0; i < n; ++i)
    ptr[i] = c[i] ? ((struct s *) (ptr - 1))->array[i] : 0;

we can only guarantee that ptr points to a "struct s" if at least
one c[i] is true.

This patch adds a DR_IS_CONDITIONAL_IN_STMT flag to record whether
the DR_REF is guaranteed to occur every time that the statement
executes to completion.  It then pools the alignment information
for references that aren't conditional in this sense.

Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?

Richard


2017-06-28  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* tree-vectorizer.h: Include tree-hash-traits.h.
	(vec_base_alignments): New typedef.
	(vec_info): Add a base_alignments field.
	(vect_record_base_alignments: Declare.
	* tree-data-ref.h (data_reference): Add an is_conditional_in_stmt
	field.
	(DR_IS_CONDITIONAL_IN_STMT): New macro.
	(create_data_ref): Add an is_conditional_in_stmt argument.
	* tree-data-ref.c (create_data_ref): Likewise.  Use it to initialize
	the is_conditional_in_stmt field.
	(data_ref_loc): Add an is_conditional_in_stmt field.
	(get_references_in_stmt): Set the is_conditional_in_stmt field.
	(find_data_references_in_stmt): Update call to create_data_ref.
	(graphite_find_data_references_in_stmt): Likewise.
	* tree-ssa-loop-prefetch.c (determine_loop_nest_reuse): Likewise.
	* tree-vect-data-refs.c (vect_analyze_data_refs): Likewise.
	(vect_record_base_alignment): New function.
	(vect_record_base_alignments): Likewise.
	(vect_compute_data_ref_alignment): Adjust base_addr and aligned_to
	for nested statements even if we fail to compute a misalignment.
	Use pooled base alignments for unconditional references.
	(vect_find_same_alignment_drs): Compare base addresses instead
	of base objects.
	(vect_compute_data_ref_alignment): Call vect_record_base_alignments.
	* tree-vect-slp.c (vect_slp_analyze_bb_1): Likewise.
	(new_bb_vec_info): Initialize base_alignments.
	* tree-vect-loop.c (new_loop_vec_info): Likewise.
	* tree-vectorizer.c (vect_destroy_datarefs): Release base_alignments.

gcc/testsuite/
	* gcc.dg/vect/pr81136.c: Add scan test.

Index: gcc/tree-vectorizer.h
===================================================================
--- gcc/tree-vectorizer.h	2017-06-28 14:26:19.653051245 +0100
+++ gcc/tree-vectorizer.h	2017-06-28 14:26:24.027879710 +0100
@@ -22,6 +22,7 @@ Software Foundation; either version 3, o
 #define GCC_TREE_VECTORIZER_H
 
 #include "tree-data-ref.h"
+#include "tree-hash-traits.h"
 #include "target.h"
 
 /* Used for naming of new temporaries.  */
@@ -84,6 +85,10 @@ struct stmt_info_for_cost {
 
 typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
 
+/* Maps base addresses to the largest alignment that we've been able
+   to calculate for them.  */
+typedef hash_map<tree_operand_hash, unsigned int> vec_base_alignments;
+
 /************************************************************************
   SLP
  ************************************************************************/
@@ -156,6 +161,10 @@ struct vec_info {
   /* All data references.  */
   vec<data_reference_p> datarefs;
 
+  /* Maps the base addresses of all data references in DATAREFS to the
+     largest alignment that we've been able to calculate for them.  */
+  vec_base_alignments base_alignments;
+
   /* All data dependences.  */
   vec<ddr_p> ddrs;
 
@@ -1123,6 +1132,7 @@ extern bool vect_prune_runtime_alias_tes
 extern bool vect_check_gather_scatter (gimple *, loop_vec_info,
 				       gather_scatter_info *);
 extern bool vect_analyze_data_refs (vec_info *, int *);
+extern void vect_record_base_alignments (vec_info *);
 extern tree vect_create_data_ref_ptr (gimple *, tree, struct loop *, tree,
 				      tree *, gimple_stmt_iterator *,
 				      gimple **, bool, bool *,
Index: gcc/tree-data-ref.h
===================================================================
--- gcc/tree-data-ref.h	2017-06-28 14:26:19.651051322 +0100
+++ gcc/tree-data-ref.h	2017-06-28 14:26:24.025879789 +0100
@@ -119,6 +119,11 @@ struct data_reference
   /* True when the data reference is in RHS of a stmt.  */
   bool is_read;
 
+  /* True when the data reference is conditional within STMT,
+     i.e. if it might not occur even when the statement is executed
+     and runs to completion.  */
+  bool is_conditional_in_stmt;
+
   /* The alignment of INNERMOST.base_address, in bits.  This is logically
      part of INNERMOST, but is kept here to avoid unnecessary padding.  */
   unsigned int base_alignment;
@@ -142,6 +147,7 @@ #define DR_ACCESS_FN(DR, I)        DR_AC
 #define DR_NUM_DIMENSIONS(DR)      DR_ACCESS_FNS (DR).length ()
 #define DR_IS_READ(DR)             (DR)->is_read
 #define DR_IS_WRITE(DR)            (!DR_IS_READ (DR))
+#define DR_IS_CONDITIONAL_IN_STMT(DR) (DR)->is_conditional_in_stmt
 #define DR_BASE_ADDRESS(DR)        (DR)->innermost.base_address
 #define DR_BASE_ALIGNMENT(DR)      (DR)->base_alignment
 #define DR_OFFSET(DR)              (DR)->innermost.offset
@@ -355,7 +361,8 @@ extern bool graphite_find_data_reference
 						   vec<data_reference_p> *);
 tree find_data_references_in_loop (struct loop *, vec<data_reference_p> *);
 bool loop_nest_has_data_refs (loop_p loop);
-struct data_reference *create_data_ref (loop_p, loop_p, tree, gimple *, bool);
+struct data_reference *create_data_ref (loop_p, loop_p, tree, gimple *, bool,
+					bool);
 extern bool find_loop_nest (struct loop *, vec<loop_p> *);
 extern struct data_dependence_relation *initialize_data_dependence_relation
      (struct data_reference *, struct data_reference *, vec<loop_p>);
Index: gcc/tree-data-ref.c
===================================================================
--- gcc/tree-data-ref.c	2017-06-28 14:26:19.651051322 +0100
+++ gcc/tree-data-ref.c	2017-06-28 14:26:24.025879789 +0100
@@ -1065,15 +1065,19 @@ free_data_ref (data_reference_p dr)
   free (dr);
 }
 
-/* Analyzes memory reference MEMREF accessed in STMT.  The reference
-   is read if IS_READ is true, write otherwise.  Returns the
-   data_reference description of MEMREF.  NEST is the outermost loop
-   in which the reference should be instantiated, LOOP is the loop in
-   which the data reference should be analyzed.  */
+/* Analyze memory reference MEMREF, which is accessed in STMT.
+   The reference is a read if IS_READ is true, otherwise it is a write.
+   IS_CONDITIONAL_IN_STMT indicates that the reference is conditional
+   within STMT, i.e. that it might not occur even if STMT is executed
+   and runs to completion.
+
+   Return the data_reference description of MEMREF.  NEST is the outermost
+   loop in which the reference should be instantiated, LOOP is the loop
+   in which the data reference should be analyzed.  */
 
 struct data_reference *
 create_data_ref (loop_p nest, loop_p loop, tree memref, gimple *stmt,
-		 bool is_read)
+		 bool is_read, bool is_conditional_in_stmt)
 {
   struct data_reference *dr;
 
@@ -1088,6 +1092,7 @@ create_data_ref (loop_p nest, loop_p loo
   DR_STMT (dr) = stmt;
   DR_REF (dr) = memref;
   DR_IS_READ (dr) = is_read;
+  DR_IS_CONDITIONAL_IN_STMT (dr) = is_conditional_in_stmt;
 
   dr_analyze_innermost (dr, memref, nest != NULL ? loop : NULL);
   dr_analyze_indices (dr, nest, loop);
@@ -4458,6 +4463,11 @@ struct data_ref_loc
 
   /* True if the memory reference is read.  */
   bool is_read;
+
+  /* True if the data reference is conditional within the containing
+     statement, i.e. if it might not occur even when the statement
+     is executed and runs to completion.  */
+  bool is_conditional_in_stmt;
 };
 
 
@@ -4524,6 +4534,7 @@ get_references_in_stmt (gimple *stmt, ve
 	{
 	  ref.ref = op1;
 	  ref.is_read = true;
+	  ref.is_conditional_in_stmt = false;
 	  references->safe_push (ref);
 	}
     }
@@ -4551,6 +4562,7 @@ get_references_in_stmt (gimple *stmt, ve
 	      type = TREE_TYPE (gimple_call_arg (stmt, 3));
 	    if (TYPE_ALIGN (type) != align)
 	      type = build_aligned_type (type, align);
+	    ref.is_conditional_in_stmt = true;
 	    ref.ref = fold_build2 (MEM_REF, type, gimple_call_arg (stmt, 0),
 				   ptr);
 	    references->safe_push (ref);
@@ -4570,6 +4582,7 @@ get_references_in_stmt (gimple *stmt, ve
 	    {
 	      ref.ref = op1;
 	      ref.is_read = true;
+	      ref.is_conditional_in_stmt = false;
 	      references->safe_push (ref);
 	    }
 	}
@@ -4583,6 +4596,7 @@ get_references_in_stmt (gimple *stmt, ve
     {
       ref.ref = op0;
       ref.is_read = false;
+      ref.is_conditional_in_stmt = false;
       references->safe_push (ref);
     }
   return clobbers_memory;
@@ -4647,8 +4661,8 @@ find_data_references_in_stmt (struct loo
 
   FOR_EACH_VEC_ELT (references, i, ref)
     {
-      dr = create_data_ref (nest, loop_containing_stmt (stmt),
-			    ref->ref, stmt, ref->is_read);
+      dr = create_data_ref (nest, loop_containing_stmt (stmt), ref->ref,
+			    stmt, ref->is_read, ref->is_conditional_in_stmt);
       gcc_assert (dr != NULL);
       datarefs->safe_push (dr);
     }
@@ -4677,7 +4691,8 @@ graphite_find_data_references_in_stmt (l
 
   FOR_EACH_VEC_ELT (references, i, ref)
     {
-      dr = create_data_ref (nest, loop, ref->ref, stmt, ref->is_read);
+      dr = create_data_ref (nest, loop, ref->ref, stmt, ref->is_read,
+			    ref->is_conditional_in_stmt);
       gcc_assert (dr != NULL);
       datarefs->safe_push (dr);
     }
Index: gcc/tree-ssa-loop-prefetch.c
===================================================================
--- gcc/tree-ssa-loop-prefetch.c	2017-06-26 19:41:19.549571836 +0100
+++ gcc/tree-ssa-loop-prefetch.c	2017-06-28 14:26:24.025879789 +0100
@@ -1633,7 +1633,7 @@ determine_loop_nest_reuse (struct loop *
     for (ref = gr->refs; ref; ref = ref->next)
       {
 	dr = create_data_ref (nest, loop_containing_stmt (ref->stmt),
-			      ref->mem, ref->stmt, !ref->write_p);
+			      ref->mem, ref->stmt, !ref->write_p, false);
 
 	if (dr)
 	  {
Index: gcc/tree-vect-data-refs.c
===================================================================
--- gcc/tree-vect-data-refs.c	2017-06-28 14:26:19.652051284 +0100
+++ gcc/tree-vect-data-refs.c	2017-06-28 14:26:24.026879750 +0100
@@ -647,6 +647,67 @@ vect_slp_analyze_instance_dependence (sl
   return res;
 }
 
+/* Record in VINFO that base address BASE_ADDR has alignment
+   BASE_ALIGNMENT.  This information came from DR (which is useful
+   for recording in the dump file).  */
+
+void
+vect_record_base_alignment (vec_info *vinfo, data_reference *dr,
+			    tree base_addr, unsigned int base_alignment)
+{
+  bool existed;
+  unsigned int &entry
+    = vinfo->base_alignments.get_or_insert (base_addr, &existed);
+  if (!existed || entry < base_alignment)
+    {
+      entry = base_alignment;
+      if (dump_enabled_p ())
+	{
+	  dump_printf_loc (MSG_NOTE, vect_location,
+			   "setting base alignment for ");
+	  dump_generic_expr (MSG_NOTE, TDF_SLIM, base_addr);
+	  dump_printf (MSG_NOTE, " to %d, based on ", base_alignment);
+	  dump_gimple_stmt (MSG_NOTE, TDF_SLIM, DR_STMT (dr), 0);
+	}
+    }
+}
+
+/* If the region we're going to vectorize is reached, all unconditional
+   data references occur at least once.  We can therefore pool the base
+   alignment guarantees from each unconditional reference.  Do this by
+   going through all the data references in VINFO and checking whether
+   the containing statement makes the reference unconditionally.  If so,
+   record the alignment of the base address in VINFO so that it can be
+   used for all other references with the same base.  */
+
+void
+vect_record_base_alignments (vec_info *vinfo)
+{
+  loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
+  struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL;
+  data_reference *dr;
+  unsigned int i;
+  FOR_EACH_VEC_ELT (vinfo->datarefs, i, dr)
+    if (!DR_IS_CONDITIONAL_IN_STMT (dr))
+      {
+	vect_record_base_alignment (vinfo, dr,
+				    DR_BASE_ADDRESS (dr),
+				    DR_BASE_ALIGNMENT (dr));
+
+	/* If DR is nested in the loop that is being vectorized, we can also
+	   record the alignment of the base wrt the outer loop.  */
+	gimple *stmt = DR_STMT (dr);
+	if (loop && nested_in_vect_loop_p (loop, DR_STMT (dr)))
+	  {
+	    stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+	    vect_record_base_alignment
+	      (vinfo, dr,
+	       STMT_VINFO_DR_BASE_ADDRESS (stmt_info),
+	       STMT_VINFO_DR_BASE_ALIGNMENT (stmt_info));
+	  }
+      }
+}
+
 /* Function vect_compute_data_ref_alignment
 
    Compute the misalignment of the data reference DR.
@@ -664,6 +725,7 @@ vect_compute_data_ref_alignment (struct
 {
   gimple *stmt = DR_STMT (dr);
   stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
+  vec_base_alignments *base_alignments = &stmt_info->vinfo->base_alignments;
   loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
   struct loop *loop = NULL;
   tree ref = DR_REF (dr);
@@ -702,6 +764,9 @@ vect_compute_data_ref_alignment (struct
     {
       tree step = DR_STEP (dr);
 
+      aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info);
+      base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info);
+      base_alignment = STMT_VINFO_DR_BASE_ALIGNMENT (stmt_info);
       if (tree_fits_shwi_p (step)
 	  && tree_to_shwi (step) % GET_MODE_SIZE (TYPE_MODE (vectype)) == 0)
         {
@@ -709,9 +774,6 @@ vect_compute_data_ref_alignment (struct
             dump_printf_loc (MSG_NOTE, vect_location,
                              "inner step divides the vector-size.\n");
 	  misalign = STMT_VINFO_DR_INIT (stmt_info);
-	  aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info);
-	  base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info);
-	  base_alignment = STMT_VINFO_DR_BASE_ALIGNMENT (stmt_info);
         }
       else
 	{
@@ -742,6 +804,11 @@ vect_compute_data_ref_alignment (struct
 	}
     }
 
+  /* Calculate the maximum of the pooled base address alignment and the
+     alignment that we can compute for DR itself.  */
+  if (unsigned int *entry = base_alignments->get (base_addr))
+    base_alignment = MAX (base_alignment, *entry);
+
   if (base_alignment >= TYPE_ALIGN (TREE_TYPE (vectype)))
     DR_VECT_AUX (dr)->base_element_aligned = true;
 
@@ -2089,8 +2156,7 @@ vect_find_same_alignment_drs (struct dat
   if (dra == drb)
     return;
 
-  if (!operand_equal_p (DR_BASE_OBJECT (dra), DR_BASE_OBJECT (drb),
-			OEP_ADDRESS_OF)
+  if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)
       || !operand_equal_p (DR_OFFSET (dra), DR_OFFSET (drb), 0)
       || !operand_equal_p (DR_STEP (dra), DR_STEP (drb), 0))
     return;
@@ -2148,6 +2214,7 @@ vect_analyze_data_refs_alignment (loop_v
   vec<data_reference_p> datarefs = vinfo->datarefs;
   struct data_reference *dr;
 
+  vect_record_base_alignments (vinfo);
   FOR_EACH_VEC_ELT (datarefs, i, dr)
     {
       stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr));
@@ -3346,7 +3413,8 @@ vect_analyze_data_refs (vec_info *vinfo,
 	    {
 	      struct data_reference *newdr
 		= create_data_ref (NULL, loop_containing_stmt (stmt),
-				   DR_REF (dr), stmt, maybe_scatter ? false : true);
+				   DR_REF (dr), stmt, !maybe_scatter,
+				   DR_IS_CONDITIONAL_IN_STMT (dr));
 	      gcc_assert (newdr != NULL && DR_REF (newdr));
 	      if (DR_BASE_ADDRESS (newdr)
 		  && DR_OFFSET (newdr)
Index: gcc/tree-vect-slp.c
===================================================================
--- gcc/tree-vect-slp.c	2017-06-26 19:41:19.549571836 +0100
+++ gcc/tree-vect-slp.c	2017-06-28 14:26:24.027879710 +0100
@@ -2367,6 +2367,7 @@ new_bb_vec_info (gimple_stmt_iterator re
   gimple_stmt_iterator gsi;
 
   res = (bb_vec_info) xcalloc (1, sizeof (struct _bb_vec_info));
+  new (&res->base_alignments) vec_base_alignments ();
   res->kind = vec_info::bb;
   BB_VINFO_BB (res) = bb;
   res->region_begin = region_begin;
@@ -2741,6 +2742,8 @@ vect_slp_analyze_bb_1 (gimple_stmt_itera
       return NULL;
     }
 
+  vect_record_base_alignments (bb_vinfo);
+
   /* Analyze and verify the alignment of data references and the
      dependence in the SLP instances.  */
   for (i = 0; BB_VINFO_SLP_INSTANCES (bb_vinfo).iterate (i, &instance); )
Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c	2017-06-26 19:41:19.549571836 +0100
+++ gcc/tree-vect-loop.c	2017-06-28 14:26:24.026879750 +0100
@@ -1157,6 +1157,7 @@ new_loop_vec_info (struct loop *loop)
   LOOP_VINFO_VECT_FACTOR (res) = 0;
   LOOP_VINFO_LOOP_NEST (res) = vNULL;
   LOOP_VINFO_DATAREFS (res) = vNULL;
+  new (&res->base_alignments) vec_base_alignments ();
   LOOP_VINFO_DDRS (res) = vNULL;
   LOOP_VINFO_UNALIGNED_DR (res) = NULL;
   LOOP_VINFO_MAY_MISALIGN_STMTS (res) = vNULL;
Index: gcc/tree-vectorizer.c
===================================================================
--- gcc/tree-vectorizer.c	2017-06-26 19:41:19.549571836 +0100
+++ gcc/tree-vectorizer.c	2017-06-28 14:26:24.027879710 +0100
@@ -370,6 +370,8 @@ vect_destroy_datarefs (vec_info *vinfo)
       }
 
   free_data_refs (vinfo->datarefs);
+
+  vinfo->base_alignments.~vec_base_alignments ();
 }
 
 /* A helper function to free scev and LOOP niter information, as well as
Index: gcc/testsuite/gcc.dg/vect/pr81136.c
===================================================================
--- gcc/testsuite/gcc.dg/vect/pr81136.c	2017-06-28 14:25:58.810888422 +0100
+++ gcc/testsuite/gcc.dg/vect/pr81136.c	2017-06-28 14:26:24.024879829 +0100
@@ -14,3 +14,5 @@ fn1 (int n)
   for (int i = 0; i < n; i++)
     a->bar[i] = b[i];
 }
+
+/* { dg-final { scan-tree-dump-not "Unknown misalignment" "vect" } } */