On Wed, Mar 16, 2011 at 12:47 PM, Ira Rosen <ira.ro...@linaro.org> wrote:
> On 16 March 2011 12:29, Richard Guenther <richard.guent...@gmail.com> wrote:
>> On Wed, Mar 16, 2011 at 7:49 AM, Ira Rosen <ira.ro...@linaro.org> wrote:
>>> Hi,
>>>
>>> This patch adds a support of conditional store sinking for cases with
>>> multiple data references in then and else basic blocks. The
>>> correctness of the transformation is checked by verifying that there
>>> are no read-after-write and write-after-write dependencies.
>>>
>>> Bootstrapped and tested on powerpc64-suse-linux.
>>> OK for trunk?
>>
>> I will look at the patch later, but can you add a testcase that we don't sink
>>
>>  if (..)
>>   {
>>      *a = i;
>>      *b = j;
>>   }
>>  else
>>   {
>>      *b = j;
>>      *a = i;
>>   }
>>
>> if *a and *b may alias?
>
> Done.

Comments inline

> Thanks,
> Ira
>
> ChangeLog:
>
>    * tree-data-ref.c (dr_equal_offsets_p1): Moved and renamed from
>    tree-vect-data-refs.c vect_equal_offsets.
>    (dr_equal_offsets_p): New function.
>    * tree-data-ref.h (dr_equal_offsets_p): Declare.
>    * tree-vect-data-refs.c (vect_equal_offsets): Move to tree-data-ref.c.
>    (vect_drs_dependent_in_basic_block): Update calls to vect_equal_offsets.
>    (vect_check_interleaving): Likewise.
>    * tree-ssa-phiopt.c: Include cfgloop.h and tree-data-ref.h.
>    (cond_if_else_store_replacement): Rename to...
>    (cond_if_else_store_replacement_1): ... this. Change arguments and
>    documentation.
>    (cond_if_else_store_replacement): New function.
>    * Makefile.in (tree-ssa-phiopt.o): Adjust dependencies.
>
> testsuite/ChangeLog:
>
>    * gcc.dg/vect/vect-cselim-1.c: New test.
>    * gcc.dg/vect/vect-cselim-2.c: New test.
>
>
> Index: tree-data-ref.c
> ===================================================================
> --- tree-data-ref.c     (revision 170712)
> +++ tree-data-ref.c     (working copy)
> @@ -991,6 +991,48 @@ create_data_ref (loop_p nest, loop_p loop, tree me
>   return dr;
>  }
>
> +/* Check if OFFSET1 and OFFSET2 (DR_OFFSETs of some data-refs) are identical
> +   expressions.  */
> +static bool
> +dr_equal_offsets_p1 (tree offset1, tree offset2)
> +{
> +  bool res;
> +
> +  STRIP_NOPS (offset1);
> +  STRIP_NOPS (offset2);
> +
> +  if (offset1 == offset2)
> +    return true;
> +
> +  if (TREE_CODE (offset1) != TREE_CODE (offset2)
> +      || (!BINARY_CLASS_P (offset1) && !UNARY_CLASS_P (offset1)))
> +    return false;
> +
> +  res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 0),
> +                             TREE_OPERAND (offset2, 0));
> +
> +  if (!res || !BINARY_CLASS_P (offset1))
> +    return res;
> +
> +  res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 1),
> +                             TREE_OPERAND (offset2, 1));
> +
> +  return res;
> +}
> +
> +/* Check if DRA and DRB have equal offsets.  */
> +bool
> +dr_equal_offsets_p (struct data_reference *dra,
> +                    struct data_reference *drb)
> +{
> +  tree offset1, offset2;
> +
> +  offset1 = DR_OFFSET (dra);
> +  offset2 = DR_OFFSET (drb);
> +
> +  return dr_equal_offsets_p1 (offset1, offset2);
> +}
> +
>  /* Returns true if FNA == FNB.  */
>
>  static bool
> Index: tree-data-ref.h
> ===================================================================
> --- tree-data-ref.h     (revision 170712)
> +++ tree-data-ref.h     (working copy)
> @@ -430,6 +430,8 @@ extern void compute_all_dependences (VEC (data_ref
>  extern void create_rdg_vertices (struct graph *, VEC (gimple, heap) *);
>  extern bool dr_may_alias_p (const struct data_reference *,
>                            const struct data_reference *);
> +extern bool dr_equal_offsets_p (struct data_reference *,
> +                                struct data_reference *);
>
>
>  /* Return true when the base objects of data references A and B are
> Index: tree-vect-data-refs.c
> ===================================================================
> --- tree-vect-data-refs.c       (revision 170712)
> +++ tree-vect-data-refs.c       (working copy)
> @@ -289,39 +289,6 @@ vect_update_interleaving_chain (struct data_refere
>     }
>  }
>
> -
> -/* Function vect_equal_offsets.
> -
> -   Check if OFFSET1 and OFFSET2 are identical expressions.  */
> -
> -static bool
> -vect_equal_offsets (tree offset1, tree offset2)
> -{
> -  bool res;
> -
> -  STRIP_NOPS (offset1);
> -  STRIP_NOPS (offset2);
> -
> -  if (offset1 == offset2)
> -    return true;
> -
> -  if (TREE_CODE (offset1) != TREE_CODE (offset2)
> -      || (!BINARY_CLASS_P (offset1) && !UNARY_CLASS_P (offset1)))
> -    return false;
> -
> -  res = vect_equal_offsets (TREE_OPERAND (offset1, 0),
> -                           TREE_OPERAND (offset2, 0));
> -
> -  if (!res || !BINARY_CLASS_P (offset1))
> -    return res;
> -
> -  res = vect_equal_offsets (TREE_OPERAND (offset1, 1),
> -                           TREE_OPERAND (offset2, 1));
> -
> -  return res;
> -}
> -
> -
>  /* Check dependence between DRA and DRB for basic block vectorization.
>    If the accesses share same bases and offsets, we can compare their initial
>    constant offsets to decide whether they differ or not.  In case of a read-
> @@ -352,7 +319,7 @@ vect_drs_dependent_in_basic_block (struct data_ref
>            || TREE_CODE (DR_BASE_ADDRESS (drb)) != ADDR_EXPR
>            || TREE_OPERAND (DR_BASE_ADDRESS (dra), 0)
>            != TREE_OPERAND (DR_BASE_ADDRESS (drb),0)))
> -      || !vect_equal_offsets (DR_OFFSET (dra), DR_OFFSET (drb)))
> +      || !dr_equal_offsets_p (dra, drb))
>     return true;
>
>   /* Check the types.  */
> @@ -402,7 +369,7 @@ vect_check_interleaving (struct data_reference *dr
>           || TREE_CODE (DR_BASE_ADDRESS (drb)) != ADDR_EXPR
>           || TREE_OPERAND (DR_BASE_ADDRESS (dra), 0)
>           != TREE_OPERAND (DR_BASE_ADDRESS (drb),0)))
> -      || !vect_equal_offsets (DR_OFFSET (dra), DR_OFFSET (drb))
> +      || !dr_equal_offsets_p (dra, drb)
>       || !tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb))
>       || DR_IS_READ (dra) != DR_IS_READ (drb))
>     return false;
> Index: tree-ssa-phiopt.c
> ===================================================================
> --- tree-ssa-phiopt.c   (revision 170734)
> +++ tree-ssa-phiopt.c   (working copy)
> @@ -34,6 +34,8 @@ along with GCC; see the file COPYING3.  If not see
>  #include "langhooks.h"
>  #include "pointer-set.h"
>  #include "domwalk.h"
> +#include "cfgloop.h"
> +#include "tree-data-ref.h"
>
>  static unsigned int tree_ssa_phiopt (void);
>  static unsigned int tree_ssa_phiopt_worker (bool);
> @@ -1292,35 +1294,18 @@ cond_store_replacement (basic_block middle_bb, bas
>   return true;
>  }
>
> -/* Do the main work of conditional store replacement.  We already know
> -   that the recognized pattern looks like so:
> +/* Do the main work of conditional store replacement.  */
>
> -   split:
> -     if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
> -   THEN_BB:
> -     X = Y;
> -     goto JOIN_BB;
> -   ELSE_BB:
> -     X = Z;
> -     fallthrough (edge E0)
> -   JOIN_BB:
> -     some more
> -
> -   We check that THEN_BB and ELSE_BB contain only one store
> -   that the stores have a "simple" RHS.  */
> -
>  static bool
> -cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
> -                               basic_block join_bb)
> +cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
> +                                 basic_block join_bb, gimple then_assign,
> +                                 gimple else_assign)
>  {
> -  gimple then_assign = last_and_only_stmt (then_bb);
> -  gimple else_assign = last_and_only_stmt (else_bb);
>   tree lhs_base, lhs, then_rhs, else_rhs;
>   source_location then_locus, else_locus;
>   gimple_stmt_iterator gsi;
>   gimple newphi, new_stmt;
>
> -  /* Check if then_bb and else_bb contain only one store each.  */
>   if (then_assign == NULL
>       || !gimple_assign_single_p (then_assign)
>       || else_assign == NULL
> @@ -1385,6 +1370,151 @@ static bool
>   return true;
>  }
>
> +/* Conditional store replacement.  We already know
> +   that the recognized pattern looks like so:
> +
> +   split:
> +     if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
> +   THEN_BB:
> +     ...
> +     X = Y;
> +     ...
> +     goto JOIN_BB;
> +   ELSE_BB:
> +     ...
> +     X = Z;
> +     ...
> +     fallthrough (edge E0)
> +   JOIN_BB:
> +     some more
> +
> +   We check that it is safe to sink the store to JOIN_BB by verifying that
> +   there are no read-after-write or write-after-write dependencies in
> +   THEN_BB and ELSE_BB.  */
> +
> +static bool
> +cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
> +                                basic_block join_bb)
> +{
> +  gimple then_assign = last_and_only_stmt (then_bb);
> +  gimple else_assign = last_and_only_stmt (else_bb);
> +  VEC (data_reference_p, heap) *then_datarefs, *else_datarefs;
> +  VEC (ddr_p, heap) *then_ddrs, *else_ddrs;
> +  gimple then_store, else_store;
> +  bool found, ok = false, res;
> +  struct data_dependence_relation *ddr;
> +  data_reference_p then_dr, else_dr;
> +  int i, j;
> +  tree then_lhs, else_lhs;
> +
> +  /* Handle the case with single statement in THEN_BB and ELSE_BB.  */
> +  if (then_assign && else_assign)
> +    return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
> +                                             then_assign, else_assign);
> +  then_datarefs = VEC_alloc (data_reference_p, heap, 1);
> +  else_datarefs = VEC_alloc (data_reference_p, heap, 1);
> +  then_ddrs = VEC_alloc (ddr_p, heap, 1);
> +  else_ddrs = VEC_alloc (ddr_p, heap, 1);
> +  if (!compute_data_dependences_for_bb (then_bb, false, &then_datarefs,
> +                                        &then_ddrs)

Can we avoid computing dependencies if the other BB would have no
data-refs?  Thus, split collecting datarefs and computing dependences?

> +      || !compute_data_dependences_for_bb (else_bb, false, &else_datarefs,
> +                                           &else_ddrs)
> +      || !VEC_length (data_reference_p, then_datarefs)
> +      || !VEC_length (data_reference_p, else_datarefs))
> +    {
> +      free_data_refs (then_datarefs);
> +      free_data_refs (else_datarefs);
> +      free_dependence_relations (then_ddrs);
> +      free_dependence_relations (else_ddrs);
> +      return false;
> +    }
> +
> +  /* Check that there are no read-after-write or write-after-write 
> dependencies
> +     in THEN_BB.  */
> +  FOR_EACH_VEC_ELT (ddr_p, then_ddrs, i, ddr)
> +    {
> +      struct data_reference *dra = DDR_A (ddr);
> +      struct data_reference *drb = DDR_B (ddr);
> +
> +      if (DDR_ARE_DEPENDENT (ddr) != chrec_known
> +          && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
> +               && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
> +              || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
> +                  && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
> +              || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))

The gimple_uids are not initialized here, you need to make sure to
call renumber_gimple_stmt_uids () before starting.  Note that phiopt
incrementally changes the IL, so I'm not sure those uids will stay
valid as stmts are moved across blocks.

> +        {
> +          free_data_refs (then_datarefs);
> +          free_data_refs (else_datarefs);
> +          free_dependence_relations (then_ddrs);
> +          free_dependence_relations (else_ddrs);
> +          return false;
> +        }
> +    }
> +
> +  /* Check that there are no read-after-write or write-after-write 
> dependencies
> +     in ELSE_BB.  */
> +  FOR_EACH_VEC_ELT (ddr_p, else_ddrs, i, ddr)
> +    {
> +      struct data_reference *dra = DDR_A (ddr);
> +      struct data_reference *drb = DDR_B (ddr);
> +
> +      if (DDR_ARE_DEPENDENT (ddr) != chrec_known
> +          && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
> +               && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
> +              || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
> +                  && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
> +              || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
> +        {
> +          free_data_refs (then_datarefs);
> +          free_data_refs (else_datarefs);
> +          free_dependence_relations (then_ddrs);
> +          free_dependence_relations (else_ddrs);
> +          return false;
> +        }
> +    }
> +
> +  /* Found pairs of stores with equal LHS.  */
> +  FOR_EACH_VEC_ELT (data_reference_p, then_datarefs, i, then_dr)
> +    {
> +      if (DR_IS_READ (then_dr))
> +        continue;
> +
> +      then_store = DR_STMT (then_dr);
> +      then_lhs = gimple_assign_lhs (then_store);
> +      found = false;
> +
> +      FOR_EACH_VEC_ELT (data_reference_p, else_datarefs, j, else_dr)
> +        {
> +          if (DR_IS_READ (else_dr))
> +            continue;
> +
> +          else_store = DR_STMT (else_dr);
> +          else_lhs = gimple_assign_lhs (else_store);
> +
> +          if (operand_equal_p (then_lhs, else_lhs, 0))
> +            {
> +              found = true;
> +              break;
> +            }
> +        }
> +
> +      if (!found)
> +        continue;
> +
> +      res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
> +                                              then_store, else_store);

So you are executing if-else store replacement for common data reference
pairs only.  I think it's cheaper to collect those pairs before computing
dependences and only if there is at least one pair perform the optimization.

You basically perform store sinking, creating a PHI node for each store
you sink (that's then probably if-converted by if-conversion later, eventually
redundant with -ftree-loop-if-convert-stores?)

I am concerned that having no bound on the number of stores sunk will
increase register pressure and does not allow scheduling of the stores
in an optimal way.  Consider two BBs similar to

  t = a + b;
  *p = t;
  t = c + d;
  *q = t;

where the transformation undoes a good schedule and makes fixing it
impossible if the remaining statements are not if-convertible.

Thus, I'd rather make this transformation only if in the end the conditional
can be completely if-converted.

I realize that we already do unbound and very aggressive if-conversion
in tree-ifcvt.c regardless of whether the loop will be vectorized or not
(including leaking the if-converted loops to the various loop versions
we create during vectorization, causing only code-size bloat).  But it's
not a good reason to continue down this road ;)

I suppose a simple maximum on the number of stores to sink
controllable by a param should do, eventually disabling this
extended transformation when vectorization is not enabled?

Otherwise the implementation looks good.

Thanks,
Richard.

> +      ok = ok || res;
> +    }
> +
> +  free_data_refs (then_datarefs);
> +  free_data_refs (else_datarefs);
> +  free_dependence_relations (then_ddrs);
> +  free_dependence_relations (else_ddrs);
> +
> +  return ok;
> +}
> +
>  /* Always do these optimizations if we have SSA
>    trees to work on.  */
>  static bool
> Index: Makefile.in
> ===================================================================
> --- Makefile.in (revision 170712)
> +++ Makefile.in (working copy)
> @@ -2422,7 +2422,8 @@ tree-ssa-ifcombine.o : tree-ssa-ifcombine.c $(CONF
>  tree-ssa-phiopt.o : tree-ssa-phiopt.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
>    $(TM_H) $(GGC_H) $(TREE_H) $(TM_P_H) $(BASIC_BLOCK_H) \
>    $(TREE_FLOW_H) $(TREE_PASS_H) $(TREE_DUMP_H) langhooks.h $(FLAGS_H) \
> -   $(DIAGNOSTIC_H) $(TIMEVAR_H) pointer-set.h domwalk.h
> +   $(DIAGNOSTIC_H) $(TIMEVAR_H) pointer-set.h domwalk.h $(CFGLOOP_H) \
> +   $(TREE_DATA_REF_H)
>  tree-nrv.o : tree-nrv.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \
>    $(TM_H) $(TREE_H) $(FUNCTION_H) $(BASIC_BLOCK_H) $(FLAGS_H) \
>    $(DIAGNOSTIC_H) $(TREE_FLOW_H) $(TIMEVAR_H) $(TREE_DUMP_H) $(TREE_PASS_H) \
> Index: testsuite/gcc.dg/vect/vect-cselim-1.c
> ===================================================================
> --- testsuite/gcc.dg/vect/vect-cselim-1.c       (revision 0)
> +++ testsuite/gcc.dg/vect/vect-cselim-1.c       (revision 0)
> @@ -0,0 +1,86 @@
> +/* { dg-require-effective-target vect_int } */
> +
> +#include <stdarg.h>
> +#include "tree-vect.h"
> +
> +#define N 50
> +
> +typedef struct {
> +  short a;
> +  short b;
> +} data;
> +
> +data in1[N], in2[N], out[N];
> +short result[N*2] =
> {7,-7,9,-6,11,-5,13,-4,15,-3,17,-2,19,-1,21,0,23,1,25,2,27,3,29,4,31,5,33,6,35,7,37,8,39,9,41,10,43,11,45,12,47,13,49,14,51,15,53,16,55,17,57,18,59,19,61,20,63,21,65,22,67,23,69,24,71,25,73,26,75,27,77,28,79,29,81,30,83,31,85,32,87,33,89,34,91,35,93,36,95,37,97,38,99,39,101,40,103,41,105,42};
> +short out1[N], out2[N];
> +
> +__attribute__ ((noinline)) void
> +foo ()
> +{
> +  int i;
> +  short c, d;
> +
> +  /* Vectorizable with conditional store sinking.  */
> +  for (i = 0; i < N; i++)
> +    {
> +      c = in1[i].b;
> +      d = in2[i].b;
> +
> +      if (c >= d)
> +        {
> +          out[i].b = c;
> +          out[i].a = d + 5;
> +        }
> +      else
> +        {
> +          out[i].b = d - 12;
> +          out[i].a = c + d;
> +        }
> +    }
> +
> +  /* Not vectorizable.  */
> +  for (i = 0; i < N; i++)
> +    {
> +      c = in1[i].b;
> +      d = in2[i].b;
> +
> +      if (c >= d)
> +        {
> +          out1[i] = c;
> +        }
> +      else
> +        {
> +          out2[i] = c + d;
> +        }
> +    }
> +}
> +
> +int
> +main (void)
> +{
> +  int i;
> +
> +  check_vect ();
> +
> +  for (i = 0; i < N; i++)
> +    {
> +      in1[i].a = i;
> +      in1[i].b = i + 2;
> +      in2[i].a = 5;
> +      in2[i].b = i + 5;
> +      __asm__ volatile ("");
> +    }
> +
> +  foo ();
> +
> +  for (i = 0; i < N; i++)
> +    {
> +      if (out[i].a != result[2*i] || out[i].b != result[2*i+1])
> +        abort ();
> +    }
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect"  } } */
> +/* { dg-final { cleanup-tree-dump "vect" } } */
> Index: testsuite/gcc.dg/vect/vect-cselim-2.c
> ===================================================================
> --- testsuite/gcc.dg/vect/vect-cselim-2.c       (revision 0)
> +++ testsuite/gcc.dg/vect/vect-cselim-2.c       (revision 0)
> @@ -0,0 +1,65 @@
> +/* { dg-require-effective-target vect_int } */
> +
> +#include <stdarg.h>
> +#include "tree-vect.h"
> +
> +#define N 50
> +
> +int a[N], b[N], in1[N], in2[N];
> +int result[2*N] =
> {5,-7,7,-6,9,-5,11,-4,13,-3,15,-2,17,-1,19,0,21,1,23,2,25,3,27,4,29,5,31,6,33,7,35,8,37,9,39,10,41,11,43,12,45,13,47,14,49,15,51,16,53,17,55,18,57,19,59,20,61,21,63,22,65,23,67,24,69,25,71,26,73,27,75,28,77,29,79,30,81,31,83,32,85,33,87,34,89,35,91,36,93,37,95,38,97,39,99,40,101,41,103,42};
> +
> +__attribute__ ((noinline)) void
> +foo (int *pa, int *pb)
> +{
> +  int i;
> +  int c, d;
> +
> +  /* Store sinking should not work here since the pointers may alias.  */
> +  for (i = 0; i < N; i++)
> +    {
> +      c = in1[i];
> +      d = in2[i];
> +
> +      if (c >= d)
> +        {
> +          *pa = c;
> +          *pb = d + 5;
> +        }
> +      else
> +        {
> +          *pb = d - 12;
> +          *pa = c + d;
> +        }
> +
> +      pa++;
> +      pb++;
> +    }
> +}
> +
> +int
> +main (void)
> +{
> +  int i;
> +
> +  check_vect ();
> +
> +  for (i = 0; i < N; i++)
> +    {
> +      in1[i] = i;
> +      in2[i] = i + 5;
> +      __asm__ volatile ("");
> +    }
> +
> +  foo (a, b);
> +
> +  for (i = 0; i < N; i++)
> +    {
> +      if (a[i] != result[2*i] || b[i] != result[2*i+1])
> +        abort ();
> +    }
> +
> +  return 0;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 0 "vect"  } } */
> +/* { dg-final { cleanup-tree-dump "vect" } } */
>

Reply via email to