This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Reducing number of alias checks in vectorization.


On Tue, 1 Oct 2013, Cong Hou wrote:

> When alias exists between data refs in a loop, to vectorize it GCC
> does loop versioning and adds runtime alias checks. Basically for each
> pair of data refs with possible data dependence, there will be two
> comparisons generated to make sure there is no aliasing between them
> in each iteration of the vectorized loop. If there are many such data
> refs pairs, the number of comparisons can be very large, which is a
> big overhead.
> 
> However, in some cases it is possible to reduce the number of those
> comparisons. For example, for the following loop, we can detect that
> b[0] and b[1] are two consecutive member accesses so that we can
> combine the alias check between a[0:100]&b[0] and a[0:100]&b[1] into
> checking a[0:100]&b[0:2]:
> 
> void foo(int*a, int* b)
> {
>    for (int i = 0; i < 100; ++i)
>     a[i] = b[0] + b[1];
> }
> 
> Actually, the requirement of consecutive memory accesses is too
> strict. For the following loop, we can still combine the alias checks
> between a[0:100]&b[0] and a[0:100]&b[100]:
> 
> void foo(int*a, int* b)
> {
>    for (int i = 0; i < 100; ++i)
>     a[i] = b[0] + b[100];
> }
> 
> This is because if b[0] is not in a[0:100] and b[100] is not in
> a[0:100] then a[0:100] cannot be between b[0] and b[100]. We only need
> to check a[0:100] and b[0:101] don't overlap.
> 
> More generally, consider two pairs of data refs (a, b1) and (a, b2).
> Suppose addr_b1 and addr_b2 are basic addresses of data ref b1 and b2;
> offset_b1 and offset_b2 (offset_b1 < offset_b2) are offsets of b1 and
> b2, and segment_length_a, segment_length_b1, and segment_length_b2 are
> segment length of a, b1, and b2. Then we can combine the two
> comparisons into one if the following condition is satisfied:
> 
> offset_b2- offset_b1 - segment_length_b1 < segment_length_a
> 
> 
> This patch detects those combination opportunities to reduce the
> number of alias checks. It is tested on an x86-64 machine.

Apart from the other comments you got (to which I agree) the patch
seems to do two things, namely also:

+  /* Extract load and store statements on pointers with zero-stride
+     accesses.  */
+  if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
+    {

which I'd rather see in a separate patch (and done also when
the loop doesn't require versioning for alias).

Also combining the alias checks in vect_create_cond_for_alias_checks
is nice but doesn't properly fix the use of the
vect-max-version-for-alias-checks param which currently inhibits
vectorization of the HIMENO benchmark by default (and make us look bad 
compared to LLVM).

So I believe this merging should be done incrementally when
we collect the DDRs we need to test in vect_mark_for_runtime_alias_test.

Thanks for working on this,
Richard.

> 
> thanks,
> Cong
> 
> 
> 
> Index: gcc/tree-vect-loop-manip.c
> ===================================================================
> --- gcc/tree-vect-loop-manip.c (revision 202662)
> +++ gcc/tree-vect-loop-manip.c (working copy)
> @@ -19,6 +19,10 @@ You should have received a copy of the G
>  along with GCC; see the file COPYING3.  If not see
>  <http://www.gnu.org/licenses/>.  */
> 
> +#include <vector>
> +#include <utility>
> +#include <algorithm>
> +
>  #include "config.h"
>  #include "system.h"
>  #include "coretypes.h"
> @@ -2248,6 +2252,74 @@ vect_vfa_segment_size (struct data_refer
>    return segment_length;
>  }
> 
> +namespace
> +{
> +
> +/* struct dr_addr_with_seg_len
> +
> +   A struct storing information of a data reference, including the data
> +   ref itself, its basic address, the access offset and the segment length
> +   for aliasing checks.  */
> +
> +struct dr_addr_with_seg_len
> +{
> +  dr_addr_with_seg_len (data_reference* d, tree addr, tree off, tree len)
> +    : dr (d), basic_addr (addr), offset (off), seg_len (len) {}
> +
> +  data_reference* dr;
> +  tree basic_addr;
> +  tree offset;
> +  tree seg_len;
> +};
> +
> +/* Operator == between two dr_addr_with_seg_len objects.
> +
> +   This equality operator is used to make sure two data refs
> +   are the same one so that we will consider to combine the
> +   aliasing checks of those two pairs of data dependent data
> +   refs.  */
> +
> +bool operator == (const dr_addr_with_seg_len& d1,
> +  const dr_addr_with_seg_len& d2)
> +{
> +  return operand_equal_p (d1.basic_addr, d2.basic_addr, 0)
> + && operand_equal_p (d1.offset, d2.offset, 0)
> + && operand_equal_p (d1.seg_len, d2.seg_len, 0);
> +}
> +
> +typedef std::pair <dr_addr_with_seg_len, dr_addr_with_seg_len>
> + dr_addr_with_seg_len_pair_t;
> +
> +
> +/* Operator < between two dr_addr_with_seg_len_pair_t objects.
> +
> +   This operator is used to sort objects of dr_addr_with_seg_len_pair_t
> +   so that we can combine aliasing checks during one scan.  */
> +
> +bool operator < (const dr_addr_with_seg_len_pair_t& p1,
> + const dr_addr_with_seg_len_pair_t& p2)
> +{
> +  const dr_addr_with_seg_len& p11 = p1.first;
> +  const dr_addr_with_seg_len& p12 = p1.second;
> +  const dr_addr_with_seg_len& p21 = p2.first;
> +  const dr_addr_with_seg_len& p22 = p2.second;
> +
> +  if (p11.basic_addr != p21.basic_addr)
> +    return p11.basic_addr < p21.basic_addr;
> +  if (p12.basic_addr != p22.basic_addr)
> +    return p12.basic_addr < p22.basic_addr;
> +  if (TREE_CODE (p11.offset) != INTEGER_CST
> +      || TREE_CODE (p21.offset) != INTEGER_CST)
> +    return p11.offset < p21.offset;
> +  if (int_cst_value (p11.offset) != int_cst_value (p21.offset))
> +    return int_cst_value (p11.offset) < int_cst_value (p21.offset);
> +  if (TREE_CODE (p12.offset) != INTEGER_CST
> +      || TREE_CODE (p22.offset) != INTEGER_CST)
> +    return p12.offset < p22.offset;
> +  return int_cst_value (p12.offset) < int_cst_value (p22.offset);
> +}
> +
> +}
> 
>  /* Function vect_create_cond_for_alias_checks.
> 
> @@ -2292,20 +2364,51 @@ vect_create_cond_for_alias_checks (loop_
>    if (may_alias_ddrs.is_empty ())
>      return;
> 
> +
> +  /* Basically, for each pair of dependent data refs store_ptr_0
> +     and load_ptr_0, we create an expression:
> +
> +     ((store_ptr_0 + store_segment_length_0) <= load_ptr_0)
> +     || (load_ptr_0 + load_segment_length_0) <= store_ptr_0))
> +
> +     for aliasing checks. However, in some cases we can decrease
> +     the number of checks by combining two checks into one. For
> +     example, suppose we have another pair of data refs store_ptr_0
> +     and load_ptr_1, and if the following condition is satisfied:
> +
> +     load_ptr_0 < load_ptr_1  &&
> +     load_ptr_1 - load_ptr_0 - load_segment_length_0 < store_segment_length_0
> +
> +     (this condition means, in each iteration of vectorized loop,
> +     the accessed memory of store_ptr_0 cannot be between the memory
> +     of load_ptr_0 and load_ptr_1.)
> +
> +     we then can use only the following expression to finish the
> +     alising checks between store_ptr_0 & load_ptr_0 and
> +     store_ptr_0 & load_ptr_1:
> +
> +     ((store_ptr_0 + store_segment_length_0) <= load_ptr_0)
> +     || (load_ptr_1 + load_segment_length_1 <= store_ptr_0))
> +
> +     Note that we only consider that load_ptr_0 and load_ptr_1 have the
> +     same basic address.  */
> +
> +  std::vector<dr_addr_with_seg_len_pair_t> ddrs_with_seg_len;
> +
> +  /* First, we collect all data ref pairs for aliasing checks.  */
> +
>    FOR_EACH_VEC_ELT (may_alias_ddrs, i, ddr)
>      {
>        struct data_reference *dr_a, *dr_b;
>        gimple dr_group_first_a, dr_group_first_b;
> -      tree addr_base_a, addr_base_b;
>        tree segment_length_a, segment_length_b;
>        gimple stmt_a, stmt_b;
> -      tree seg_a_min, seg_a_max, seg_b_min, seg_b_max;
> 
>        dr_a = DDR_A (ddr);
>        stmt_a = DR_STMT (DDR_A (ddr));
>        dr_group_first_a = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_a));
>        if (dr_group_first_a)
> -        {
> + {
>    stmt_a = dr_group_first_a;
>    dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a));
>   }
> @@ -2314,20 +2417,11 @@ vect_create_cond_for_alias_checks (loop_
>        stmt_b = DR_STMT (DDR_B (ddr));
>        dr_group_first_b = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_b));
>        if (dr_group_first_b)
> -        {
> + {
>    stmt_b = dr_group_first_b;
>    dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b));
>   }
> 
> -      addr_base_a
> - = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_a),
> -   size_binop (PLUS_EXPR, DR_OFFSET (dr_a),
> -       DR_INIT (dr_a)));
> -      addr_base_b
> - = fold_build_pointer_plus (DR_BASE_ADDRESS (dr_b),
> -   size_binop (PLUS_EXPR, DR_OFFSET (dr_b),
> -       DR_INIT (dr_b)));
> -
>        if (!operand_equal_p (DR_STEP (dr_a), DR_STEP (dr_b), 0))
>   length_factor = scalar_loop_iters;
>        else
> @@ -2335,24 +2429,149 @@ vect_create_cond_for_alias_checks (loop_
>        segment_length_a = vect_vfa_segment_size (dr_a, length_factor);
>        segment_length_b = vect_vfa_segment_size (dr_b, length_factor);
> 
> +      dr_addr_with_seg_len_pair_t dr_with_seg_len_pair
> +  (dr_addr_with_seg_len
> +       (dr_a, DR_BASE_ADDRESS (dr_a),
> + size_binop (PLUS_EXPR, DR_OFFSET (dr_a), DR_INIT (dr_a)),
> + segment_length_a),
> +   dr_addr_with_seg_len
> +       (dr_b, DR_BASE_ADDRESS (dr_b),
> + size_binop (PLUS_EXPR, DR_OFFSET (dr_b), DR_INIT (dr_b)),
> + segment_length_b));
> +
> +      if (dr_with_seg_len_pair.first.basic_addr >
> +  dr_with_seg_len_pair.second.basic_addr)
> + std::swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second);
> +
> +      ddrs_with_seg_len.push_back (dr_with_seg_len_pair);
> +    }
> +
> +  /* Second, we sort the collected data ref pairs so that we can scan
> +     them once to combine all possible aliasing checks.  */
> +
> +  std::sort (ddrs_with_seg_len.begin(), ddrs_with_seg_len.end());
> +
> +  /* Remove duplicate data ref pairs.  */
> +  ddrs_with_seg_len.erase (std::unique (ddrs_with_seg_len.begin(),
> + ddrs_with_seg_len.end()),
> +   ddrs_with_seg_len.end());
> +
> +  /* We then scan the sorted dr pairs and check if we can combine
> +     alias checks of two neighbouring dr pairs.  */
> +
> +  for (size_t i = 1; i < ddrs_with_seg_len.size (); ++i)
> +    {
> +      dr_addr_with_seg_len& dr_a1 = ddrs_with_seg_len[i-1].first;
> +      dr_addr_with_seg_len& dr_b1 = ddrs_with_seg_len[i-1].second;
> +      dr_addr_with_seg_len& dr_a2 = ddrs_with_seg_len[i].first;
> +      dr_addr_with_seg_len& dr_b2 = ddrs_with_seg_len[i].second;
> +
> +      if (dr_a1 == dr_a2)
> + {
> +  if (dr_b1.basic_addr != dr_b2.basic_addr
> +      || TREE_CODE (dr_b1.offset) != INTEGER_CST
> +      || TREE_CODE (dr_b2.offset) != INTEGER_CST)
> +    continue;
> +
> +  int diff = int_cst_value (dr_b2.offset) -
> +     int_cst_value (dr_b1.offset);
> +
> +  gcc_assert (diff > 0);
> +
> +  if (diff <= vect_factor
> +      || (TREE_CODE (dr_b1.seg_len) == INTEGER_CST
> +  && diff - int_cst_value (dr_b1.seg_len) < vect_factor)
> +      || (TREE_CODE (dr_b1.seg_len) == INTEGER_CST
> +  && TREE_CODE (dr_a1.seg_len) == INTEGER_CST
> +  && diff - int_cst_value (dr_b1.seg_len) <
> +     int_cst_value (dr_a1.seg_len)))
> +    {
> +      if (dump_enabled_p ())
> + {
> +  dump_printf_loc
> +      (MSG_NOTE, vect_location,
> +       "combining two runtime checks for data references ");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b1.dr));
> +  dump_printf (MSG_NOTE, " and ");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b2.dr));
> +  dump_printf (MSG_NOTE, "\n");
> + }
> +
> +      dr_b1.seg_len = size_binop (PLUS_EXPR,
> +  dr_b2.seg_len, size_int (diff));
> +      ddrs_with_seg_len.erase (ddrs_with_seg_len.begin () + i);
> +      --i;
> +    }
> + }
> +      else if (dr_b1 == dr_b2)
> + {
> +  if (dr_a1.basic_addr != dr_a2.basic_addr
> +      || TREE_CODE (dr_a1.offset) != INTEGER_CST
> +      || TREE_CODE (dr_a2.offset) != INTEGER_CST)
> +    continue;
> +
> +  int diff = int_cst_value (dr_a2.offset) -
> +     int_cst_value (dr_a1.offset);
> +
> +  gcc_assert (diff > 0);
> +
> +  if (diff <= vect_factor
> +      || (TREE_CODE (dr_a1.seg_len) == INTEGER_CST
> +  && diff - int_cst_value (dr_a1.seg_len) < vect_factor)
> +      || (TREE_CODE (dr_a1.seg_len) == INTEGER_CST
> +  && TREE_CODE (dr_b1.seg_len) == INTEGER_CST
> +  && diff - int_cst_value (dr_a1.seg_len) <
> +     int_cst_value (dr_b1.seg_len)))
> +    {
> +      if (dump_enabled_p ())
> + {
> +  dump_printf_loc
> +      (MSG_NOTE, vect_location,
> +       "combining two runtime checks for data references ");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a1.dr));
> +  dump_printf (MSG_NOTE, " and ");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a2.dr));
> +  dump_printf (MSG_NOTE, "\n");
> + }
> +
> +      dr_a1.seg_len = size_binop (PLUS_EXPR,
> +  dr_a2.seg_len, size_int (diff));
> +      ddrs_with_seg_len.erase (ddrs_with_seg_len.begin () + i);
> +      --i;
> +    }
> + }
> +    }
> +
> +  for (size_t i = 0, s = ddrs_with_seg_len.size (); i < s; ++i)
> +    {
> +      const dr_addr_with_seg_len& dr_a = ddrs_with_seg_len[i].first;
> +      const dr_addr_with_seg_len& dr_b = ddrs_with_seg_len[i].second;
> +      tree segment_length_a = dr_a.seg_len;
> +      tree segment_length_b = dr_b.seg_len;
> +
> +      tree addr_base_a
> + = fold_build_pointer_plus (dr_a.basic_addr, dr_a.offset);
> +      tree addr_base_b
> + = fold_build_pointer_plus (dr_b.basic_addr, dr_b.offset);
> +
>        if (dump_enabled_p ())
>   {
>    dump_printf_loc (MSG_NOTE, vect_location,
> -                           "create runtime check for data references ");
> -  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a));
> +   "create runtime check for data references ");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_a.dr));
>    dump_printf (MSG_NOTE, " and ");
> -  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b));
> -          dump_printf (MSG_NOTE, "\n");
> +  dump_generic_expr (MSG_NOTE, TDF_SLIM, DR_REF (dr_b.dr));
> +  dump_printf (MSG_NOTE, "\n");
>   }
> 
> -      seg_a_min = addr_base_a;
> -      seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a);
> -      if (tree_int_cst_compare (DR_STEP (dr_a), size_zero_node) < 0)
> +      tree seg_a_min = addr_base_a;
> +      tree seg_a_max = fold_build_pointer_plus (addr_base_a, segment_length_a);
> +      if (tree_int_cst_compare (DR_STEP (dr_a.dr), size_zero_node) < 0)
>   seg_a_min = seg_a_max, seg_a_max = addr_base_a;
> 
> -      seg_b_min = addr_base_b;
> -      seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b);
> -      if (tree_int_cst_compare (DR_STEP (dr_b), size_zero_node) < 0)
> +      tree seg_b_min = addr_base_b;
> +      tree seg_b_max = fold_build_pointer_plus (addr_base_b, segment_length_b);
> +      if (tree_int_cst_compare (DR_STEP (dr_b.dr), size_zero_node) < 0)
>   seg_b_min = seg_b_max, seg_b_max = addr_base_b;
> 
>        part_cond_expr =
> @@ -2477,6 +2696,81 @@ vect_loop_versioning (loop_vec_info loop
>        adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi));
>      }
> 
> +  /* Extract load and store statements on pointers with zero-stride
> +     accesses.  */
> +  if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
> +    {
> +
> +      /* In the loop body, we iterate each statement to check if it is a load
> + or store. Then we check the DR_STEP of the data reference.  If
> + DR_STEP is zero, then we will hoist the load statement to the loop
> + preheader, and move the store statement to the loop exit.  */
> +
> +      for (gimple_stmt_iterator si = gsi_start_bb (loop->header);
> +   !gsi_end_p (si); )
> + {
> +  gimple stmt = gsi_stmt (si);
> +  stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
> +  struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
> +
> +
> +  if (dr && integer_zerop (DR_STEP (dr)))
> +    {
> +      if (DR_IS_READ (dr))
> + {
> +  if (dump_file)
> +    {
> +      fprintf (dump_file,
> +       "Hoist the load to outside of the loop:\n");
> +      print_gimple_stmt (dump_file, stmt, 0,
> + TDF_VOPS|TDF_MEMSYMS);
> +    }
> +
> +  basic_block preheader = loop_preheader_edge (loop)->src;
> +  gimple_stmt_iterator si_dst = gsi_last_bb (preheader);
> +  gsi_move_after (&si, &si_dst);
> + }
> +      else
> + {
> +  gimple_stmt_iterator si_dst =
> +      gsi_last_bb (single_exit (loop)->dest);
> +  gsi_move_after (&si, &si_dst);
> + }
> +              continue;
> +    }
> +  else if (!dr)
> +          {
> +            bool hoist = true;
> +            for (size_t i = 0; i < gimple_num_ops (stmt); i++)
> +            {
> +              tree op = gimple_op (stmt, i);
> +              if (TREE_CODE (op) == INTEGER_CST
> +                  || TREE_CODE (op) == REAL_CST)
> +                continue;
> +              if (TREE_CODE (op) == SSA_NAME)
> +              {
> +                gimple def = SSA_NAME_DEF_STMT (op);
> +                if (def == stmt
> +                    || gimple_nop_p (def)
> +                    || !flow_bb_inside_loop_p (loop, gimple_bb (def)))
> +                  continue;
> +              }
> +              hoist = false;
> +              break;
> +            }
> +
> +            if (hoist)
> +            {
> +              basic_block preheader = loop_preheader_edge (loop)->src;
> +              gimple_stmt_iterator si_dst = gsi_last_bb (preheader);
> +              gsi_move_after (&si, &si_dst);
> +              continue;
> +            }
> +          }
> +          gsi_next (&si);
> + }
> +    }
> +
>    /* End loop-exit-fixes after versioning.  */
> 
>    if (cond_expr_stmt_list)
> Index: gcc/ChangeLog
> ===================================================================
> --- gcc/ChangeLog (revision 202663)
> +++ gcc/ChangeLog (working copy)
> @@ -1,3 +1,8 @@
> +2013-10-01  Cong Hou  <congh@google.com>
> +
> + * tree-vect-loop-manip.c (vect_create_cond_for_alias_checks): Combine
> + alias checks if it is possible to amortize the runtime overhead.
> +
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]