On Tue, 9 Jul 2019, Jan Hubicka wrote:
> Hi,
> this is updated patch. I based the logic on gimple_compare_field_offset but
> dropped bits about PLACEHOLDER_EXPR since my understanding is that to get
> comparsion done I would then need to compare operand #2 of COMPONENT_REF which
> I don't.
>
> I also wrote the range checks using polyint since I believe that most code is
> supposed to be updated this way (shall we update gimple_compare_field_offset?
For consistency yes I guess but IIRC they cannot really appear in
FIELD_DECLs.
> It is used in canonical type merging and considering fields different may lead
> to wrong code I would say, but I do not know enough about SVE to construct
> testcase).
>
> I updated documentation of return values which I also find somewhat confusing
> since 1/0 meanings in nonoverlapping_* is reversed compared to
> aliasing_component_refs. Main entry is called refs_may_alias so I am
> considering rename all the functions into *_may_alias and make them return 0
> for disambiguation, 1 for non-disambiguation and use -1's to keep decidion
> whether to work harder.
>
> However for now I am sticking with the nonoverlapping reversed semantics.
>
>
> There are no changes in tramp3d stats (there are no unmerged types)
> Here are stats on cc1plus build:
>
> Alias oracle query stats:
> refs_may_alias_p: 43435216 disambiguations, 51989307 queries
> ref_maybe_used_by_call_p: 60843 disambiguations, 44040532 queries
> call_may_clobber_ref_p: 6051 disambiguations, 9115 queries
> nonoverlapping_component_refs_p: 0 disambiguations, 2535 queries
> nonoverlapping_component_refs_since_match_p: 12297 disambiguations, 40458
> must overlaps, 53243 queries
> aliasing_component_refs_p: 70892 disambiguations, 374789 queries
> TBAA oracle: 12680127 disambiguations 32179405 queries
> 10383370 are in alias set 0
> 5747729 queries asked about the same object
> 148 queries asked about the same alias set
> 0 access volatile
> 2482808 are dependent in the DAG
> 885223 are aritificially in conflict with void *
>
> PTA query stats:
> pt_solution_includes: 562382 disambiguations, 7467931 queries
> pt_solutions_intersect: 412873 disambiguations, 7818955 queries
>
> It seems that nonoverlapping_component_refs_since_match_p is pretty good on
> making decision and there are only few cases where it return -1 which is good.
> So i guess teaching it about ARRAY_REFs is logical next step.
>
> If I would like to benchmark alias oracle compile time, what is
> reasonable way to do that?
>
> Bootstrapped/regtested x86_64-linux, OK?
>
> Honza
>
> * tree-ssa-alias.c (nonoverlapping_component_refs_p_1): Break out
> from ...; work also on duplicated types.
> (nonoverlapping_component_refs_since_match): ... here
> (ncr_type_uid): Break out from ...
> (ncr_compar): ... here; look for TYPE_UID of canonical type if
> available.
> (nonoverlapping_component_refs_p): Use same_type_for_tbaa to match
> the types and nonoverlapping_component_refs_p_1 to disambiguate.
> * g++.dg/lto/alias-3_0.C: New file.
> * g++.dg/lto/alias-3_1.c: New file.
>
> Index: tree-ssa-alias.c
> ===================================================================
> --- tree-ssa-alias.c (revision 273193)
> +++ tree-ssa-alias.c (working copy)
> @@ -1128,6 +1128,90 @@ aliasing_component_refs_p (tree ref1,
> return false;
> }
>
> +/* FIELD1 and FIELD2 are two component refs whose bases either do
> + not overlap at all or their addresses are the same.
> +
> + Return 1 if FIELD1 and FIELD2 are non-overlapping
> +
> + Return 0 if FIELD1 and FIELD2 are posisbly overlapping but in case of
> + overlap their addresses are the same.
> +
> + Return -1 otherwise.
> +
> + Main difference between 0 and -1 is to let
> + nonoverlapping_component_refs_since_match_p discover the semnatically
> + equivalent part of the access path. */
> +
> +static int
> +nonoverlapping_component_refs_p_1 (const_tree field1, const_tree field2)
> +{
> + /* If both fields are of the same type, we could save hard work of
> + comparing offsets.
> + ??? We cannot simply use the type of operand #0 of the refs here
> + as the Fortran compiler smuggles type punning into COMPONENT_REFs
> + for common blocks instead of using unions like everyone else. */
> + tree type1 = DECL_CONTEXT (field1);
> + tree type2 = DECL_CONTEXT (field2);
> +
drop the vertical space
> + if (type1 == type2 && TREE_CODE (type1) == RECORD_TYPE)
> + {
> + if (field1 == field2)
> + return 0;
> + /* A field and its representative need to be considered the
> + same. */
> + if (DECL_BIT_FIELD_REPRESENTATIVE (field1) == field2
> + || DECL_BIT_FIELD_REPRESENTATIVE (field2) == field1)
> + return 0;
> + /* Different fields of the same record type cannot overlap.
> + ??? Bitfields can overlap at RTL level so punt on them. */
> + if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> + return 0;
> + /* Assume that different FIELD_DECLs never overlap in a RECORD_TYPE.
> */
> + return 1;
> + }
> + else
> + {
drop the else {
> + /* Different fields of the same record type cannot overlap.
> + ??? Bitfields can overlap at RTL level so punt on them. */
> + if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> + return 0;
> +
don't you need the DECL_BIT_FIELD_REPRESENTATIVE check here as well?
I'd do
if (DECL_BIT_FIELD_REPRESENTATIVE (field1))
field1 = DECL_BIT_FIELD_REPRESENTATIVE (field1);
if (DECL_BIT_FIELD_REPRESENTATIVE (field2))
field2 = DECL_BIT_FIELD_REPRESENTATIVE (field2);
thus use the representative for the overlap check. It might
be the case that we can improve here and if we do this
can do the DECL_BIT_FIELD check after this (hoping the
representative doesn't have it set).
> + if (tree_int_cst_equal (DECL_FIELD_OFFSET (field1),
> + DECL_FIELD_OFFSET (field2))
> + && tree_int_cst_equal (DECL_FIELD_BIT_OFFSET (field1),
> + DECL_FIELD_BIT_OFFSET (field2)))
> + return 0;
In gimple_compare_field_offset this was fast-pathed for
DECL_OFFSET_ALIGN (f1) == DECL_OFFSET_ALIGN (f2) so I suggest to
do that here as well. Note that DECL_FIELD_OFFSET can be
a non-constant which means you cannot use tree_int_cst_equal
unconditionally here but you have to use operand_equal_p.
> + /* Note that it may be possible to use component_ref_field_offset
> + which would provide offsets as trees. However constructing and folding
> + trees is expensive and does not seem to be worth the compile time
> + cost. */
> +
> + poly_uint64 offset1, offset2;
> + poly_uint64 bit_offset1, bit_offset2;
> + poly_uint64 size1, size2;
I think you need poly_offset_int here since you convert to bits below.
The gimple_compare_field_offset checking way looks cheaper btw, so
I wonder why you don't simply call it but replicate things here?
When do we expect to have partially overlapping field decls? Even
when considering canonical type merging?
> + if (poly_int_tree_p (DECL_FIELD_OFFSET (field1), &offset1)
> + && poly_int_tree_p (DECL_FIELD_OFFSET (field2), &offset2)
> + && poly_int_tree_p (DECL_FIELD_BIT_OFFSET (field1), &bit_offset1)
> + && poly_int_tree_p (DECL_FIELD_BIT_OFFSET (field2), &bit_offset2)
> + && poly_int_tree_p (DECL_SIZE (field1), &size1)
> + && poly_int_tree_p (DECL_SIZE (field2), &size2))
> + {
> + offset1 = (offset1 << LOG2_BITS_PER_UNIT) + bit_offset1;
> + offset2 = (offset2 << LOG2_BITS_PER_UNIT) + bit_offset2;
> +
> + if (known_eq (offset1, offset2))
> + return 0;
> + if (!ranges_maybe_overlap_p (offset1, size1, offset2, size2))
> + return 1;
> + }
> + }
> + /* Resort to slower overlap checking by looking for matching types in
> + the middle of access path. */
> + return -1;
> +}
> +
> /* Try to disambiguate REF1 and REF2 under the assumption that MATCH1 and
> MATCH2 either point to the same address or are disjoint.
> MATCH1 and MATCH2 are assumed to be ref in the access path of REF1 and
> REF2
> @@ -1224,6 +1308,7 @@ nonoverlapping_component_refs_since_matc
> case the return value will precisely be false. */
> while (true)
> {
> + bool seen_noncomponent_ref_p = false;
> do
> {
> if (component_refs1.is_empty ())
> @@ -1233,6 +1318,8 @@ nonoverlapping_component_refs_since_matc
> return 0;
> }
> ref1 = component_refs1.pop ();
> + if (TREE_CODE (ref1) != COMPONENT_REF)
> + seen_noncomponent_ref_p = true;
> }
> while (!RECORD_OR_UNION_TYPE_P (TREE_TYPE (TREE_OPERAND (ref1, 0))));
>
> @@ -1245,17 +1332,15 @@ nonoverlapping_component_refs_since_matc
> return 0;
> }
> ref2 = component_refs2.pop ();
> + if (TREE_CODE (ref2) != COMPONENT_REF)
> + seen_noncomponent_ref_p = true;
> }
> while (!RECORD_OR_UNION_TYPE_P (TREE_TYPE (TREE_OPERAND (ref2, 0))));
>
> - /* Beware of BIT_FIELD_REF. */
> - if (TREE_CODE (ref1) != COMPONENT_REF
> - || TREE_CODE (ref2) != COMPONENT_REF)
> - {
> - ++alias_stats
> - .nonoverlapping_component_refs_since_match_p_may_alias;
> - return -1;
> - }
> + /* BIT_FIELD_REF and VIEW_CONVERT_EXPR are taken off the vectors
> + earlier. */
> + gcc_checking_assert (TREE_CODE (ref1) == COMPONENT_REF
> + && TREE_CODE (ref2) == COMPONENT_REF);
>
> tree field1 = TREE_OPERAND (ref1, 1);
> tree field2 = TREE_OPERAND (ref2, 1);
> @@ -1266,33 +1351,27 @@ nonoverlapping_component_refs_since_matc
> tree type1 = DECL_CONTEXT (field1);
> tree type2 = DECL_CONTEXT (field2);
>
> - /* We cannot disambiguate fields in a union or qualified union. */
> - if (type1 != type2 || TREE_CODE (type1) != RECORD_TYPE)
> + /* If we skipped array refs on type of different sizes, we can
> + no longer be sure that there are not partial overlaps. */
> + if (seen_noncomponent_ref_p
> + && !operand_equal_p (TYPE_SIZE (type1), TYPE_SIZE (type2), 0))
> {
> - ++alias_stats.nonoverlapping_component_refs_since_match_p_may_alias;
> + ++alias_stats
> + .nonoverlapping_component_refs_since_match_p_may_alias;
> return -1;
> }
>
> - if (field1 != field2)
> + int cmp = nonoverlapping_component_refs_p_1 (field1, field2);
> + if (cmp == -1)
> {
> - /* A field and its representative need to be considered the
> - same. */
> - if (DECL_BIT_FIELD_REPRESENTATIVE (field1) == field2
> - || DECL_BIT_FIELD_REPRESENTATIVE (field2) == field1)
> - {
> - ++alias_stats
> - .nonoverlapping_component_refs_since_match_p_must_overlap;
> - return 0;
> - }
> - /* Different fields of the same record type cannot overlap.
> - ??? Bitfields can overlap at RTL level so punt on them. */
> - if (DECL_BIT_FIELD (field1) && DECL_BIT_FIELD (field2))
> - {
> - ++alias_stats
> - .nonoverlapping_component_refs_since_match_p_must_overlap;
> - return 0;
> - }
> - ++alias_stats.nonoverlapping_component_refs_since_match_p_no_alias;
> + ++alias_stats
> + .nonoverlapping_component_refs_since_match_p_may_alias;
> + return -1;
> + }
> + else if (cmp == 1)
> + {
> + ++alias_stats
> + .nonoverlapping_component_refs_since_match_p_no_alias;
> return 1;
> }
> }
> @@ -1301,6 +1380,24 @@ nonoverlapping_component_refs_since_matc
> return 0;
> }
>
> +/* Return TYPE_UID which can be used to match record types we consider
> + same for TBAA purposes. */
> +
> +static inline int
> +ncr_type_uid (const_tree field)
> +{
> + /* ??? We cannot simply use the type of operand #0 of the refs here
> + as the Fortran compiler smuggles type punning into COMPONENT_REFs
> + for common blocks instead of using unions like everyone else. */
> + tree type = DECL_FIELD_CONTEXT (field);
> + /* With LTO types considered same_type_for_tbaa_p
> + from different translation unit may not have same
> + main variant. They however have same TYPE_CANONICAL. */
> + if (TYPE_CANONICAL (type))
> + return TYPE_UID (TYPE_CANONICAL (type));
> + return TYPE_UID (type);
> +}
> +
> /* qsort compare function to sort FIELD_DECLs after their
> DECL_FIELD_CONTEXT TYPE_UID. */
>
> @@ -1309,8 +1406,9 @@ ncr_compar (const void *field1_, const v
> {
> const_tree field1 = *(const_tree *) const_cast <void *>(field1_);
> const_tree field2 = *(const_tree *) const_cast <void *>(field2_);
> - unsigned int uid1 = TYPE_UID (DECL_FIELD_CONTEXT (field1));
> - unsigned int uid2 = TYPE_UID (DECL_FIELD_CONTEXT (field2));
> + unsigned int uid1 = ncr_type_uid (field1);
> + unsigned int uid2 = ncr_type_uid (field2);
> +
> if (uid1 < uid2)
> return -1;
> else if (uid1 > uid2)
> @@ -1377,10 +1475,9 @@ nonoverlapping_component_refs_p (const_t
> if (fieldsx.length () == 1
> && fieldsy.length () == 1)
> {
> - if ((DECL_FIELD_CONTEXT (fieldsx[0])
> - == DECL_FIELD_CONTEXT (fieldsy[0]))
> - && fieldsx[0] != fieldsy[0]
> - && !(DECL_BIT_FIELD (fieldsx[0]) && DECL_BIT_FIELD (fieldsy[0])))
> + if (same_type_for_tbaa (DECL_FIELD_CONTEXT (fieldsx[0]),
> + DECL_FIELD_CONTEXT (fieldsy[0])) == 1
> + && nonoverlapping_component_refs_p_1 (fieldsx[0], fieldsy[0]) == 1)
> {
> ++alias_stats.nonoverlapping_component_refs_p_no_alias;
> return true;
> @@ -1413,31 +1510,18 @@ nonoverlapping_component_refs_p (const_t
> {
> const_tree fieldx = fieldsx[i];
> const_tree fieldy = fieldsy[j];
> - tree typex = DECL_FIELD_CONTEXT (fieldx);
> - tree typey = DECL_FIELD_CONTEXT (fieldy);
> - if (typex == typey)
> - {
> - /* We're left with accessing different fields of a structure,
> - no possible overlap. */
> - if (fieldx != fieldy)
> - {
> - /* A field and its representative need to be considered the
> - same. */
> - if (DECL_BIT_FIELD_REPRESENTATIVE (fieldx) == fieldy
> - || DECL_BIT_FIELD_REPRESENTATIVE (fieldy) == fieldx)
> - ;
> - /* Different fields of the same record type cannot overlap.
> - ??? Bitfields can overlap at RTL level so punt on them. */
> - else if (DECL_BIT_FIELD (fieldx) && DECL_BIT_FIELD (fieldy))
> - ;
> - else
> - {
> - ++alias_stats.nonoverlapping_component_refs_p_no_alias;
> - return true;
> - }
> - }
> +
> + /* We're left with accessing different fields of a structure,
> + no possible overlap. */
> + if (same_type_for_tbaa (DECL_FIELD_CONTEXT (fieldx),
> + DECL_FIELD_CONTEXT (fieldy)) == 1
> + && nonoverlapping_component_refs_p_1 (fieldx, fieldy) == 1)
> + {
> + ++alias_stats.nonoverlapping_component_refs_p_no_alias;
> + return true;
> }
> - if (TYPE_UID (typex) < TYPE_UID (typey))
> +
> + if (ncr_type_uid (fieldx) < ncr_type_uid (fieldy))
> {
> i++;
> if (i == fieldsx.length ())
> Index: testsuite/g++.dg/lto/alias-3_0.C
> ===================================================================
> --- testsuite/g++.dg/lto/alias-3_0.C (nonexistent)
> +++ testsuite/g++.dg/lto/alias-3_0.C (working copy)
> @@ -0,0 +1,27 @@
> +/* { dg-lto-do run } */
> +/* { dg-lto-options { { -O3 -flto -fno-early-inlining } } } */
> +
> +struct a
> +{
> + int foo,bar;
> +};
> +struct b
> +{
> + struct a a[10];
> +};
> +
> +__attribute__ ((used)) struct b b, *bptr=&b, *bptr2=&b;
> +__attribute__ ((used)) int i,j;
> +
> +extern "C" void inline_me_late (void);
> +
> +int
> +main (void)
> +{
> + int jj=j;
> + bptr2->a[jj].bar = 0;
> + inline_me_late ();
> + if (!__builtin_constant_p (bptr2->a[jj].bar == 0))
> + __builtin_abort ();
> + return 0;
> +}
> Index: testsuite/g++.dg/lto/alias-3_1.c
> ===================================================================
> --- testsuite/g++.dg/lto/alias-3_1.c (nonexistent)
> +++ testsuite/g++.dg/lto/alias-3_1.c (working copy)
> @@ -0,0 +1,20 @@
> +/* { dg-lto-do run } */
> +/* { dg-lto-options { { -O3 -flto -fno-early-inlining } } } */
> +struct a
> +{
> + int foo,bar;
> +};
> +struct b
> +{
> + struct a a[10];
> +};
> +
> +extern struct b *bptr;
> +extern int i;
> +
> +void
> +inline_me_late (void)
> +{
> + bptr->a[i].foo=1;
> +}
> +
>
--
Richard Biener <rguent...@suse.de>
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG Nürnberg)
