From: Richard Biener <richard.guenther@gmail.com>
To: GCC Patches <gcc-patches@gcc.gnu.org>,
Richard Sandiford <richard.sandiford@linaro.org>
Subject: Re: Use base inequality for some vector alias checks
Date: Wed, 07 Jun 2017 10:10:00 -0000 [thread overview]
Message-ID: <CAFiYyc0eCV1OmtxkNsH4Li8Tydiwmnt-Odq7oan7kaVX9BBNJQ@mail.gmail.com> (raw)
In-Reply-To: <871sr5o6oc.fsf@linaro.org>
On Wed, May 31, 2017 at 8:56 AM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> Ping
>
> Richard Sandiford <richard.sandiford@linaro.org> writes:
>> This patch checks whether two data references x and y cannot
>> partially overlap and so are independent whenever &x != &y.
>> We can then use this in the vectoriser to optimise alias checks.
>>
>> Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
Looks good to me. Probably needs refactoring now if Bin was faster
with factoring out the machinery to elsewhere.
Thanks,
Richard.
>> Thanks,
>> Richard
>>
>>
>> gcc/
>> 2016-05-03 Richard Sandiford <richard.sandiford@linaro.org>
>>
>> * hash-traits.h (pair_hash): New struct.
>> * tree-data-ref.h (data_dependence_relation): Add object_a and
>> object_b fields.
>> (DDR_OBJECT_A, DDR_OBJECT_B): New macros.
>> * tree-data-ref.c (initialize_data_dependence_relation): Initialize
>> DDR_OBJECT_A and DDR_OBJECT_B.
>> * tree-vectorizer.h (vec_object_pair): New type.
>> (_loop_vec_info): Add a check_unequal_addrs field.
>> (LOOP_VINFO_CHECK_UNEQUAL_ADDRS): New macro.
>> (LOOP_REQUIRES_VERSIONING_FOR_ALIAS): Return true if there is an
>> entry in check_unequal_addrs. Check comp_alias_ddrs instead of
>> may_alias_ddrs.
>> * tree-vect-loop.c (destroy_loop_vec_info): Release
>> LOOP_VINFO_CHECK_UNEQUAL_ADDRS.
>> (vect_analyze_loop_2): Likewise, when restarting.
>> (vect_estimate_min_profitable_iters): Estimate the cost of
>> LOOP_VINFO_CHECK_UNEQUAL_ADDRS.
>> * tree-vect-data-refs.c: Include tree-hash-traits.h.
>> (vect_prune_runtime_alias_test_list): Try to handle conflicts
>> using LOOP_VINFO_CHECK_UNEQUAL_ADDRS, if the data dependence allows.
>> Count such tests in the final summary.
>> * tree-vect-loop-manip.c (chain_cond_expr): New function.
>> (vect_create_cond_for_align_checks): Use it.
>> (vect_create_cond_for_alias_checks): Likewise.
>> (vect_create_cond_for_unequal_addrs): New function.
>> (vect_loop_versioning): Call it.
>>
>> gcc/testsuite/
>> * gcc.dg/vect/vect-alias-check-6.c: New test.
>>
>> Index: gcc/hash-traits.h
>> ===================================================================
>> --- gcc/hash-traits.h 2017-02-23 19:54:15.000000000 +0000
>> +++ gcc/hash-traits.h 2017-05-03 08:48:53.312035228 +0100
>> @@ -301,6 +301,76 @@ struct ggc_cache_ptr_hash : pointer_hash
>>
>> struct nofree_string_hash : string_hash, typed_noop_remove <const char *> {};
>>
>> +/* Traits for pairs of values, using the first to record empty and
>> + deleted slots. */
>> +
>> +template <typename T1, typename T2>
>> +struct pair_hash
>> +{
>> + typedef std::pair <typename T1::value_type,
>> + typename T2::value_type> value_type;
>> + typedef std::pair <typename T1::compare_type,
>> + typename T2::compare_type> compare_type;
>> +
>> + static inline hashval_t hash (const value_type &);
>> + static inline bool equal (const value_type &, const compare_type &);
>> + static inline void remove (value_type &);
>> + static inline void mark_deleted (value_type &);
>> + static inline void mark_empty (value_type &);
>> + static inline bool is_deleted (const value_type &);
>> + static inline bool is_empty (const value_type &);
>> +};
>> +
>> +template <typename T1, typename T2>
>> +inline hashval_t
>> +pair_hash <T1, T2>::hash (const value_type &x)
>> +{
>> + return iterative_hash_hashval_t (T1::hash (x.first), T2::hash (x.second));
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline bool
>> +pair_hash <T1, T2>::equal (const value_type &x, const compare_type &y)
>> +{
>> + return T1::equal (x.first, y.first) && T2::equal (x.second, y.second);
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline void
>> +pair_hash <T1, T2>::remove (value_type &x)
>> +{
>> + T1::remove (x.first);
>> + T2::remove (x.second);
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline void
>> +pair_hash <T1, T2>::mark_deleted (value_type &x)
>> +{
>> + T1::mark_deleted (x.first);
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline void
>> +pair_hash <T1, T2>::mark_empty (value_type &x)
>> +{
>> + T1::mark_empty (x.first);
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline bool
>> +pair_hash <T1, T2>::is_deleted (const value_type &x)
>> +{
>> + return T1::is_deleted (x.first);
>> +}
>> +
>> +template <typename T1, typename T2>
>> +inline bool
>> +pair_hash <T1, T2>::is_empty (const value_type &x)
>> +{
>> + return T1::is_empty (x.first);
>> +}
>> +
>> template <typename T> struct default_hash_traits : T {};
>>
>> template <typename T>
>> Index: gcc/tree-data-ref.h
>> ===================================================================
>> --- gcc/tree-data-ref.h 2017-05-03 08:48:48.737038502 +0100
>> +++ gcc/tree-data-ref.h 2017-05-03 08:48:53.313041828 +0100
>> @@ -240,6 +240,13 @@ struct data_dependence_relation
>> but the analyzer cannot be more specific. */
>> tree are_dependent;
>>
>> + /* If nonnull, COULD_BE_INDEPENDENT_P is true and the accesses are
>> + independent when the runtime addresses of OBJECT_A and OBJECT_B
>> + are different. The addresses of both objects are invariant in the
>> + loop nest. */
>> + tree object_a;
>> + tree object_b;
>> +
>> /* For each subscript in the dependence test, there is an element in
>> this array. This is the attribute that labels the edge A->B of
>> the data_dependence_relation. */
>> @@ -303,6 +310,8 @@ #define DDR_A(DDR) (DDR)->a
>> #define DDR_B(DDR) (DDR)->b
>> #define DDR_AFFINE_P(DDR) (DDR)->affine_p
>> #define DDR_ARE_DEPENDENT(DDR) (DDR)->are_dependent
>> +#define DDR_OBJECT_A(DDR) (DDR)->object_a
>> +#define DDR_OBJECT_B(DDR) (DDR)->object_b
>> #define DDR_SUBSCRIPTS(DDR) (DDR)->subscripts
>> #define DDR_SUBSCRIPT(DDR, I) DDR_SUBSCRIPTS (DDR)[I]
>> #define DDR_NUM_SUBSCRIPTS(DDR) DDR_SUBSCRIPTS (DDR).length ()
>> Index: gcc/tree-data-ref.c
>> ===================================================================
>> --- gcc/tree-data-ref.c 2017-05-03 08:48:48.737038502 +0100
>> +++ gcc/tree-data-ref.c 2017-05-03 08:48:53.313041828 +0100
>> @@ -1715,6 +1715,15 @@ initialize_data_dependence_relation (str
>> }
>>
>> DDR_COULD_BE_INDEPENDENT_P (res) = true;
>> + if (!loop_nest.exists ()
>> + || (object_address_invariant_in_loop_p (loop_nest[0],
>> + struct_ref_a)
>> + && object_address_invariant_in_loop_p (loop_nest[0],
>> + struct_ref_b)))
>> + {
>> + DDR_OBJECT_A (res) = struct_ref_a;
>> + DDR_OBJECT_B (res) = struct_ref_b;
>> + }
>> }
>>
>> DDR_AFFINE_P (res) = true;
>> Index: gcc/tree-vectorizer.h
>> ===================================================================
>> --- gcc/tree-vectorizer.h 2017-05-03 08:48:48.738045102 +0100
>> +++ gcc/tree-vectorizer.h 2017-05-03 08:48:53.315055028 +0100
>> @@ -174,6 +174,10 @@ struct dr_with_seg_len_pair_t
>>
>>
>>
>> +/* Describes two objects whose addresses must be unequal for the vectorized
>> + loop to be valid. */
>> +typedef std::pair<tree, tree> vec_object_pair;
>> +
>> /* Vectorizer state common between loop and basic-block vectorization. */
>> struct vec_info {
>> enum { bb, loop } kind;
>> @@ -270,6 +274,9 @@ typedef struct _loop_vec_info : public v
>> lengths from which the run-time aliasing check is built. */
>> vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
>>
>> + /* Check that the addresses of each pair of objects is unequal. */
>> + vec<vec_object_pair> check_unequal_addrs;
>> +
>> /* Statements in the loop that have data references that are candidates for a
>> runtime (loop versioning) misalignment check. */
>> vec<gimple *> may_misalign_stmts;
>> @@ -364,6 +371,7 @@ #define LOOP_VINFO_UNALIGNED_DR(L)
>> #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
>> #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
>> #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
>> +#define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
>> #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
>> #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
>> #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
>> @@ -383,7 +391,8 @@ #define LOOP_VINFO_ORIG_LOOP_INFO(L)
>> #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
>> ((L)->may_misalign_stmts.length () > 0)
>> #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
>> - ((L)->comp_alias_ddrs.length () > 0)
>> + ((L)->comp_alias_ddrs.length () > 0 \
>> + || (L)->check_unequal_addrs.length () > 0)
>> #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
>> (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
>> #define LOOP_REQUIRES_VERSIONING(L) \
>> Index: gcc/tree-vect-loop.c
>> ===================================================================
>> --- gcc/tree-vect-loop.c 2017-04-18 19:52:35.059158859 +0100
>> +++ gcc/tree-vect-loop.c 2017-05-03 08:48:53.315055028 +0100
>> @@ -1275,6 +1275,8 @@ destroy_loop_vec_info (loop_vec_info loo
>> destroy_cost_data (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo));
>> loop_vinfo->scalar_cost_vec.release ();
>>
>> + LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).release ();
>> +
>> free (loop_vinfo);
>> loop->aux = NULL;
>> }
>> @@ -2299,6 +2301,7 @@ vect_analyze_loop_2 (loop_vec_info loop_
>> }
>> /* Free optimized alias test DDRS. */
>> LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo).release ();
>> + LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).release ();
>> /* Reset target cost data. */
>> destroy_cost_data (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo));
>> LOOP_VINFO_TARGET_COST_DATA (loop_vinfo)
>> @@ -3261,6 +3264,11 @@ vect_estimate_min_profitable_iters (loop
>> unsigned len = LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo).length ();
>> (void) add_stmt_cost (target_cost_data, len, vector_stmt, NULL, 0,
>> vect_prologue);
>> + len = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).length ();
>> + if (len)
>> + /* Count LEN - 1 ANDs and LEN comparisons. */
>> + (void) add_stmt_cost (target_cost_data, len * 2 - 1, scalar_stmt,
>> + NULL, 0, vect_prologue);
>> dump_printf (MSG_NOTE,
>> "cost model: Adding cost of checks for loop "
>> "versioning aliasing.\n");
>> Index: gcc/tree-vect-data-refs.c
>> ===================================================================
>> --- gcc/tree-vect-data-refs.c 2017-05-03 08:48:48.738045102 +0100
>> +++ gcc/tree-vect-data-refs.c 2017-05-03 08:48:53.314048428 +0100
>> @@ -50,6 +50,7 @@ Software Foundation; either version 3, o
>> #include "expr.h"
>> #include "builtins.h"
>> #include "params.h"
>> +#include "tree-hash-traits.h"
>>
>> /* Return true if load- or store-lanes optab OPTAB is implemented for
>> COUNT vectors of type VECTYPE. NAME is the name of OPTAB. */
>> @@ -3085,10 +3086,14 @@ dependence_distance_ge_vf (data_dependen
>> bool
>> vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo)
>> {
>> - vec<ddr_p> may_alias_ddrs =
>> - LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
>> - vec<dr_with_seg_len_pair_t>& comp_alias_ddrs =
>> - LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo);
>> + typedef pair_hash <tree_operand_hash, tree_operand_hash> tree_pair_hash;
>> + hash_set <tree_pair_hash> compared_objects;
>> +
>> + vec<ddr_p> may_alias_ddrs = LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
>> + vec<dr_with_seg_len_pair_t> &comp_alias_ddrs
>> + = LOOP_VINFO_COMP_ALIAS_DDRS (loop_vinfo);
>> + vec<vec_object_pair> &check_unequal_addrs
>> + = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo);
>> int vect_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
>> tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo);
>>
>> @@ -3151,6 +3156,24 @@ vect_prune_runtime_alias_test_list (loop
>> if (dependence_distance_ge_vf (ddr, loop_depth, vect_factor))
>> continue;
>>
>> + if (DDR_OBJECT_A (ddr))
>> + {
>> + vec_object_pair new_pair (DDR_OBJECT_A (ddr), DDR_OBJECT_B (ddr));
>> + if (!compared_objects.add (new_pair))
>> + {
>> + if (dump_enabled_p ())
>> + {
>> + dump_printf_loc (MSG_NOTE, vect_location, "checking that ");
>> + dump_generic_expr (MSG_NOTE, TDF_SLIM, new_pair.first);
>> + dump_printf (MSG_NOTE, " and ");
>> + dump_generic_expr (MSG_NOTE, TDF_SLIM, new_pair.second);
>> + dump_printf (MSG_NOTE, " have different addresses\n");
>> + }
>> + LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo).safe_push (new_pair);
>> + }
>> + continue;
>> + }
>> +
>> dr_a = DDR_A (ddr);
>> stmt_a = DR_STMT (DDR_A (ddr));
>> dr_group_first_a = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_a));
>> @@ -3346,11 +3369,12 @@ vect_prune_runtime_alias_test_list (loop
>> }
>> }
>>
>> + unsigned int count = (comp_alias_ddrs.length ()
>> + + check_unequal_addrs.length ());
>> dump_printf_loc (MSG_NOTE, vect_location,
>> "improved number of alias checks from %d to %d\n",
>> - may_alias_ddrs.length (), comp_alias_ddrs.length ());
>> - if ((int) comp_alias_ddrs.length () >
>> - PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS))
>> + may_alias_ddrs.length (), count);
>> + if ((int) count > PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS))
>> {
>> if (dump_enabled_p ())
>> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> Index: gcc/tree-vect-loop-manip.c
>> ===================================================================
>> --- gcc/tree-vect-loop-manip.c 2017-04-18 19:52:34.027608884 +0100
>> +++ gcc/tree-vect-loop-manip.c 2017-05-03 08:48:53.314048428 +0100
>> @@ -1926,6 +1926,20 @@ vect_create_cond_for_niters_checks (loop
>> *cond_expr = part_cond_expr;
>> }
>>
>> +/* Set *COND_EXPR to a tree that is true when both the original *COND_EXPR
>> + and PART_COND_EXPR are true. Treat a null *COND_EXPR as "true". */
>> +
>> +static void
>> +chain_cond_expr (tree *cond_expr, tree part_cond_expr)
>> +{
>> + if (*cond_expr)
>> + *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
>> + *cond_expr, part_cond_expr);
>> + else
>> + *cond_expr = part_cond_expr;
>> +}
>> +
>> +
>> /* Function vect_create_cond_for_align_checks.
>>
>> Create a conditional expression that represents the alignment checks for
>> @@ -2037,13 +2051,32 @@ vect_create_cond_for_align_checks (loop_
>> ptrsize_zero = build_int_cst (int_ptrsize_type, 0);
>> part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node,
>> and_tmp_name, ptrsize_zero);
>> - if (*cond_expr)
>> - *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
>> - *cond_expr, part_cond_expr);
>> - else
>> - *cond_expr = part_cond_expr;
>> + chain_cond_expr (cond_expr, part_cond_expr);
>> }
>>
>> +
>> +/* If LOOP_VINFO_CHECK_UNEQUAL_ADDRS contains <A1, B1>, ..., <An, Bn>,
>> + create a tree representation of: (&A1 != &B1) && ... && (&An != &Bn).
>> + Set *COND_EXPR to a tree that is true when both the original *COND_EXPR
>> + and this new condition are true. Treat a null *COND_EXPR as "true". */
>> +
>> +static void
>> +vect_create_cond_for_unequal_addrs (loop_vec_info loop_vinfo, tree *cond_expr)
>> +{
>> + vec<vec_object_pair> pairs = LOOP_VINFO_CHECK_UNEQUAL_ADDRS (loop_vinfo);
>> + unsigned int i;
>> + vec_object_pair *pair;
>> + FOR_EACH_VEC_ELT (pairs, i, pair)
>> + {
>> + tree addr1 = build_fold_addr_expr (pair->first);
>> + tree addr2 = build_fold_addr_expr (pair->second);
>> + tree part_cond_expr = fold_build2 (NE_EXPR, boolean_type_node,
>> + addr1, addr2);
>> + chain_cond_expr (cond_expr, part_cond_expr);
>> + }
>> +}
>> +
>> +
>> /* Given two data references and segment lengths described by DR_A and DR_B,
>> create expression checking if the two addresses ranges intersect with
>> each other based on index of the two addresses. This can only be done
>> @@ -2280,11 +2313,7 @@ vect_create_cond_for_alias_checks (loop_
>>
>> /* Create condition expression for each pair data references. */
>> create_intersect_range_checks (loop_vinfo, &part_cond_expr, dr_a, dr_b);
>> - if (*cond_expr)
>> - *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
>> - *cond_expr, part_cond_expr);
>> - else
>> - *cond_expr = part_cond_expr;
>> + chain_cond_expr (cond_expr, part_cond_expr);
>> }
>>
>> if (dump_enabled_p ())
>> @@ -2353,7 +2382,10 @@ vect_loop_versioning (loop_vec_info loop
>> &cond_expr_stmt_list);
>>
>> if (version_alias)
>> - vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr);
>> + {
>> + vect_create_cond_for_unequal_addrs (loop_vinfo, &cond_expr);
>> + vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr);
>> + }
>>
>> cond_expr = force_gimple_operand_1 (cond_expr, &gimplify_stmt_list,
>> is_gimple_condexpr, NULL_TREE);
>> Index: gcc/testsuite/gcc.dg/vect/vect-alias-check-6.c
>> ===================================================================
>> --- /dev/null 2017-05-03 08:16:26.972699664 +0100
>> +++ gcc/testsuite/gcc.dg/vect/vect-alias-check-6.c 2017-05-03 08:48:53.312035228 +0100
>> @@ -0,0 +1,23 @@
>> +/* { dg-do compile } */
>> +/* { dg-require-effective-target vect_int } */
>> +
>> +#define N 16
>> +
>> +struct s { int x[N]; };
>> +
>> +void
>> +f1 (struct s *a, struct s *b)
>> +{
>> + for (int i = 0; i < N - 1; ++i)
>> + a->x[i + 1] += b->x[i];
>> +}
>> +
>> +void
>> +f2 (struct s *a, struct s *b)
>> +{
>> + for (int i = 0; i < N; ++i)
>> + a->x[i] += b->x[N - i - 1];
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times {checking that [^\n]* and [^\n]* have different addresses} 2 "vect" } } */
>> +/* { dg-final { scan-tree-dump-times "LOOP VECTORIZED" 2 "vect" } } */
prev parent reply other threads:[~2017-06-07 10:10 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-05-03 8:04 Richard Sandiford
2017-05-31 6:58 ` Richard Sandiford
2017-06-07 10:10 ` Richard Biener [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CAFiYyc0eCV1OmtxkNsH4Li8Tydiwmnt-Odq7oan7kaVX9BBNJQ@mail.gmail.com \
--to=richard.guenther@gmail.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=richard.sandiford@linaro.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).