From f702e7d43f2aec71640d0db7ecf1543ba75f37c3 Mon Sep 17 00:00:00 2001 From: Richard Sandiford Date: Fri, 22 Sep 2017 16:44:29 +0000 Subject: [PATCH] Let the target choose a vectorisation alignment The vectoriser aligned vectors to TYPE_ALIGN unconditionally, although there was also a hard-coded assumption that this was equal to the type size. This was inconvenient for SVE for two reasons: - When compiling for a specific power-of-2 SVE vector length, we might want to align to a full vector. However, the TYPE_ALIGN is governed by the ABI alignment, which is 128 bits regardless of size. - For vector-length-agnostic code it doesn't usually make sense to align, since the runtime vector length might not be a power of two. Even for power of two sizes, there's no guarantee that aligning to the previous 16 bytes will be an improveent. This patch therefore adds a target hook to control the preferred vectoriser (as opposed to ABI) alignment. 2017-09-22 Richard Sandiford Alan Hayward David Sherwood gcc/ * target.def (preferred_vector_alignment): New hook. * doc/tm.texi.in (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): New hook. * doc/tm.texi: Regenerate. * targhooks.h (default_preferred_vector_alignment): Declare. * targhooks.c (default_preferred_vector_alignment): New function. * tree-vectorizer.h (dataref_aux): Add a target_alignment field. Expand commentary. (DR_TARGET_ALIGNMENT): New macro. (aligned_access_p): Update commentary. (vect_known_alignment_in_bytes): New function. * tree-vect-data-refs.c (vect_calculate_required_alignment): New function. (vect_compute_data_ref_alignment): Set DR_TARGET_ALIGNMENT. Calculate the misalignment based on the target alignment rather than the vector size. (vect_update_misalignment_for_peel): Use DR_TARGET_ALIGMENT rather than TYPE_ALIGN / BITS_PER_UNIT to update the misalignment. (vect_enhance_data_refs_alignment): Mask the byte misalignment with the target alignment, rather than masking the element misalignment with the number of elements in a vector. Also use the target alignment when calculating the maximum number of peels. (vect_find_same_alignment_drs): Use vect_calculate_required_alignment instead of TYPE_ALIGN_UNIT. (vect_duplicate_ssa_name_ptr_info): Remove stmt_info parameter. Measure DR_MISALIGNMENT relative to DR_TARGET_ALIGNMENT. (vect_create_addr_base_for_vector_ref): Update call accordingly. (vect_create_data_ref_ptr): Likewise. (vect_setup_realignment): Realign by ANDing with -DR_TARGET_MISALIGNMENT. * tree-vect-loop-manip.c (vect_gen_prolog_loop_niters): Calculate the number of peels based on DR_TARGET_ALIGNMENT. * tree-vect-stmts.c (get_group_load_store_type): Compare the gap with the guaranteed alignment boundary when deciding whether overrun is OK. (vectorizable_mask_load_store): Interpret DR_MISALIGNMENT relative to DR_TARGET_ALIGNMENT instead of TYPE_ALIGN_UNIT. (ensure_base_align): Remove stmt_info parameter. Get the target base alignment from DR_TARGET_ALIGNMENT. (vectorizable_store): Update call accordingly. Interpret DR_MISALIGNMENT relative to DR_TARGET_ALIGNMENT instead of TYPE_ALIGN_UNIT. (vectorizable_load): Likewise. gcc/testsuite/ * gcc.dg/vect/vect-outer-3a.c: Adjust dump scan for new wording of alignment message. * gcc.dg/vect/vect-outer-3a-big-array.c: Likewise. Co-Authored-By: Alan Hayward Co-Authored-By: David Sherwood From-SVN: r253101 --- gcc/ChangeLog | 48 ++++++++++ gcc/doc/tm.texi | 12 +++ gcc/doc/tm.texi.in | 2 + gcc/target.def | 14 +++ gcc/targhooks.c | 9 ++ gcc/targhooks.h | 1 + gcc/testsuite/ChangeLog | 8 ++ .../gcc.dg/vect/vect-outer-3a-big-array.c | 2 +- gcc/testsuite/gcc.dg/vect/vect-outer-3a.c | 2 +- gcc/tree-vect-data-refs.c | 92 +++++++++++-------- gcc/tree-vect-loop-manip.c | 51 +++++----- gcc/tree-vect-stmts.c | 83 +++++++++-------- gcc/tree-vectorizer.h | 23 ++++- 13 files changed, 242 insertions(+), 105 deletions(-) diff --git a/gcc/ChangeLog b/gcc/ChangeLog index 43e885b7614..e29256106b2 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,51 @@ +2017-09-22 Richard Sandiford + Alan Hayward + David Sherwood + + * target.def (preferred_vector_alignment): New hook. + * doc/tm.texi.in (TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT): New + hook. + * doc/tm.texi: Regenerate. + * targhooks.h (default_preferred_vector_alignment): Declare. + * targhooks.c (default_preferred_vector_alignment): New function. + * tree-vectorizer.h (dataref_aux): Add a target_alignment field. + Expand commentary. + (DR_TARGET_ALIGNMENT): New macro. + (aligned_access_p): Update commentary. + (vect_known_alignment_in_bytes): New function. + * tree-vect-data-refs.c (vect_calculate_required_alignment): New + function. + (vect_compute_data_ref_alignment): Set DR_TARGET_ALIGNMENT. + Calculate the misalignment based on the target alignment rather than + the vector size. + (vect_update_misalignment_for_peel): Use DR_TARGET_ALIGMENT + rather than TYPE_ALIGN / BITS_PER_UNIT to update the misalignment. + (vect_enhance_data_refs_alignment): Mask the byte misalignment with + the target alignment, rather than masking the element misalignment + with the number of elements in a vector. Also use the target + alignment when calculating the maximum number of peels. + (vect_find_same_alignment_drs): Use vect_calculate_required_alignment + instead of TYPE_ALIGN_UNIT. + (vect_duplicate_ssa_name_ptr_info): Remove stmt_info parameter. + Measure DR_MISALIGNMENT relative to DR_TARGET_ALIGNMENT. + (vect_create_addr_base_for_vector_ref): Update call accordingly. + (vect_create_data_ref_ptr): Likewise. + (vect_setup_realignment): Realign by ANDing with + -DR_TARGET_MISALIGNMENT. + * tree-vect-loop-manip.c (vect_gen_prolog_loop_niters): Calculate + the number of peels based on DR_TARGET_ALIGNMENT. + * tree-vect-stmts.c (get_group_load_store_type): Compare the gap + with the guaranteed alignment boundary when deciding whether + overrun is OK. + (vectorizable_mask_load_store): Interpret DR_MISALIGNMENT + relative to DR_TARGET_ALIGNMENT instead of TYPE_ALIGN_UNIT. + (ensure_base_align): Remove stmt_info parameter. Get the + target base alignment from DR_TARGET_ALIGNMENT. + (vectorizable_store): Update call accordingly. Interpret + DR_MISALIGNMENT relative to DR_TARGET_ALIGNMENT instead of + TYPE_ALIGN_UNIT. + (vectorizable_load): Likewise. + 2017-09-22 Richard Sandiford Alan Hayward David Sherwood diff --git a/gcc/doc/tm.texi b/gcc/doc/tm.texi index 07ae66a2e61..1f8c0f12942 100644 --- a/gcc/doc/tm.texi +++ b/gcc/doc/tm.texi @@ -5758,6 +5758,18 @@ For vector memory operations the cost may depend on type (@var{vectype}) and misalignment value (@var{misalign}). @end deftypefn +@deftypefn {Target Hook} HOST_WIDE_INT TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT (const_tree @var{type}) +This hook returns the preferred alignment in bits for accesses to +vectors of type @var{type} in vectorized code. This might be less than +or greater than the ABI-defined value returned by +@code{TARGET_VECTOR_ALIGNMENT}. It can be equal to the alignment of +a single element, in which case the vectorizer will not try to optimize +for alignment. + +The default hook returns @code{TYPE_ALIGN (@var{type})}, which is +correct for most targets. +@end deftypefn + @deftypefn {Target Hook} bool TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE (const_tree @var{type}, bool @var{is_packed}) Return true if vector alignment is reachable (by peeling N iterations) for the given scalar type @var{type}. @var{is_packed} is false if the scalar access using @var{type} is known to be naturally aligned. @end deftypefn diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in index 6a794371342..fa3ff0164bd 100644 --- a/gcc/doc/tm.texi.in +++ b/gcc/doc/tm.texi.in @@ -4088,6 +4088,8 @@ address; but often a machine-dependent strategy can generate better code. @hook TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST +@hook TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT + @hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE @hook TARGET_VECTORIZE_VEC_PERM_CONST_OK diff --git a/gcc/target.def b/gcc/target.def index ae22d7a6183..9e5ddc00a43 100644 --- a/gcc/target.def +++ b/gcc/target.def @@ -1820,6 +1820,20 @@ misalignment value (@var{misalign}).", int, (enum vect_cost_for_stmt type_of_cost, tree vectype, int misalign), default_builtin_vectorization_cost) +DEFHOOK +(preferred_vector_alignment, + "This hook returns the preferred alignment in bits for accesses to\n\ +vectors of type @var{type} in vectorized code. This might be less than\n\ +or greater than the ABI-defined value returned by\n\ +@code{TARGET_VECTOR_ALIGNMENT}. It can be equal to the alignment of\n\ +a single element, in which case the vectorizer will not try to optimize\n\ +for alignment.\n\ +\n\ +The default hook returns @code{TYPE_ALIGN (@var{type})}, which is\n\ +correct for most targets.", + HOST_WIDE_INT, (const_tree type), + default_preferred_vector_alignment) + /* Return true if vector alignment is reachable (by peeling N iterations) for the given scalar type. */ DEFHOOK diff --git a/gcc/targhooks.c b/gcc/targhooks.c index d87d6c7f430..171583d419e 100644 --- a/gcc/targhooks.c +++ b/gcc/targhooks.c @@ -1175,6 +1175,15 @@ default_vector_alignment (const_tree type) return align; } +/* The default implementation of + TARGET_VECTORIZE_PREFERRED_VECTOR_ALIGNMENT. */ + +HOST_WIDE_INT +default_preferred_vector_alignment (const_tree type) +{ + return TYPE_ALIGN (type); +} + /* By default assume vectors of element TYPE require a multiple of the natural alignment of TYPE. TYPE is naturally aligned if IS_PACKED is false. */ bool diff --git a/gcc/targhooks.h b/gcc/targhooks.h index a70992d034e..e577888654d 100644 --- a/gcc/targhooks.h +++ b/gcc/targhooks.h @@ -95,6 +95,7 @@ extern tree default_builtin_reciprocal (tree); extern HOST_WIDE_INT default_vector_alignment (const_tree); +extern HOST_WIDE_INT default_preferred_vector_alignment (const_tree); extern bool default_builtin_vector_alignment_reachable (const_tree, bool); extern bool default_builtin_support_vector_misalignment (machine_mode mode, diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog index eaae531a84e..39838f595fb 100644 --- a/gcc/testsuite/ChangeLog +++ b/gcc/testsuite/ChangeLog @@ -1,3 +1,11 @@ +2017-09-22 Richard Sandiford + Alan Hayward + David Sherwood + + * gcc.dg/vect/vect-outer-3a.c: Adjust dump scan for new wording + of alignment message. + * gcc.dg/vect/vect-outer-3a-big-array.c: Likewise. + 2017-09-22 Martin Sebor PR c/81854 diff --git a/gcc/testsuite/gcc.dg/vect/vect-outer-3a-big-array.c b/gcc/testsuite/gcc.dg/vect/vect-outer-3a-big-array.c index 1316e80a657..b84f5afa0e2 100644 --- a/gcc/testsuite/gcc.dg/vect/vect-outer-3a-big-array.c +++ b/gcc/testsuite/gcc.dg/vect/vect-outer-3a-big-array.c @@ -49,4 +49,4 @@ int main (void) } /* { dg-final { scan-tree-dump-times "OUTER LOOP VECTORIZED" 1 "vect" { xfail { vect_no_align && { ! vect_hw_misalign } } } } } */ -/* { dg-final { scan-tree-dump-times "step doesn't divide the vector-size" 1 "vect" } } */ +/* { dg-final { scan-tree-dump-times "step doesn't divide the vector alignment" 1 "vect" } } */ diff --git a/gcc/testsuite/gcc.dg/vect/vect-outer-3a.c b/gcc/testsuite/gcc.dg/vect/vect-outer-3a.c index 6a7ab916ec4..d3ba837c95d 100644 --- a/gcc/testsuite/gcc.dg/vect/vect-outer-3a.c +++ b/gcc/testsuite/gcc.dg/vect/vect-outer-3a.c @@ -49,4 +49,4 @@ int main (void) } /* { dg-final { scan-tree-dump-times "OUTER LOOP VECTORIZED" 1 "vect" { xfail { vect_no_align && { ! vect_hw_misalign } } } } } */ -/* { dg-final { scan-tree-dump-times "step doesn't divide the vector-size" 1 "vect" } } */ +/* { dg-final { scan-tree-dump-times "step doesn't divide the vector alignment" 1 "vect" } } */ diff --git a/gcc/tree-vect-data-refs.c b/gcc/tree-vect-data-refs.c index 175052a290b..3c57e5c0be0 100644 --- a/gcc/tree-vect-data-refs.c +++ b/gcc/tree-vect-data-refs.c @@ -775,6 +775,17 @@ vect_record_base_alignments (vec_info *vinfo) } } +/* Return the target alignment for the vectorized form of DR. */ + +static unsigned int +vect_calculate_target_alignment (struct data_reference *dr) +{ + gimple *stmt = DR_STMT (dr); + stmt_vec_info stmt_info = vinfo_for_stmt (stmt); + tree vectype = STMT_VINFO_VECTYPE (stmt_info); + return targetm.vectorize.preferred_vector_alignment (vectype); +} + /* Function vect_compute_data_ref_alignment Compute the misalignment of the data reference DR. @@ -811,6 +822,10 @@ vect_compute_data_ref_alignment (struct data_reference *dr) innermost_loop_behavior *drb = vect_dr_behavior (dr); bool step_preserves_misalignment_p; + unsigned HOST_WIDE_INT vector_alignment + = vect_calculate_target_alignment (dr) / BITS_PER_UNIT; + DR_TARGET_ALIGNMENT (dr) = vector_alignment; + /* No step for BB vectorization. */ if (!loop) { @@ -823,43 +838,41 @@ vect_compute_data_ref_alignment (struct data_reference *dr) relative to the outer-loop (LOOP). This is ok only if the misalignment stays the same throughout the execution of the inner-loop, which is why we have to check that the stride of the dataref in the inner-loop evenly - divides by the vector size. */ + divides by the vector alignment. */ else if (nested_in_vect_loop_p (loop, stmt)) { step_preserves_misalignment_p - = (DR_STEP_ALIGNMENT (dr) - % GET_MODE_SIZE (TYPE_MODE (vectype))) == 0; + = (DR_STEP_ALIGNMENT (dr) % vector_alignment) == 0; if (dump_enabled_p ()) { if (step_preserves_misalignment_p) dump_printf_loc (MSG_NOTE, vect_location, - "inner step divides the vector-size.\n"); + "inner step divides the vector alignment.\n"); else dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "inner step doesn't divide the vector-size.\n"); + "inner step doesn't divide the vector" + " alignment.\n"); } } /* Similarly we can only use base and misalignment information relative to an innermost loop if the misalignment stays the same throughout the execution of the loop. As above, this is the case if the stride of - the dataref evenly divides by the vector size. */ + the dataref evenly divides by the alignment. */ else { unsigned vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); step_preserves_misalignment_p - = ((DR_STEP_ALIGNMENT (dr) * vf) - % GET_MODE_SIZE (TYPE_MODE (vectype))) == 0; + = ((DR_STEP_ALIGNMENT (dr) * vf) % vector_alignment) == 0; if (!step_preserves_misalignment_p && dump_enabled_p ()) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, - "step doesn't divide the vector-size.\n"); + "step doesn't divide the vector alignment.\n"); } unsigned int base_alignment = drb->base_alignment; unsigned int base_misalignment = drb->base_misalignment; - unsigned HOST_WIDE_INT vector_alignment = TYPE_ALIGN_UNIT (vectype); /* Calculate the maximum of the pooled base address alignment and the alignment that we can compute for DR itself. */ @@ -1007,9 +1020,8 @@ vect_update_misalignment_for_peel (struct data_reference *dr, { bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; int misal = DR_MISALIGNMENT (dr); - tree vectype = STMT_VINFO_VECTYPE (stmt_info); misal += negative ? -npeel * dr_size : npeel * dr_size; - misal &= (TYPE_ALIGN (vectype) / BITS_PER_UNIT) - 1; + misal &= DR_TARGET_ALIGNMENT (dr) - 1; SET_DR_MISALIGNMENT (dr, misal); return; } @@ -1657,16 +1669,17 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) { if (known_alignment_for_access_p (dr)) { - unsigned int npeel_tmp = 0; + unsigned int npeel_tmp = 0; bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0; - vectype = STMT_VINFO_VECTYPE (stmt_info); - nelements = TYPE_VECTOR_SUBPARTS (vectype); - mis = DR_MISALIGNMENT (dr) / vect_get_scalar_dr_size (dr); + vectype = STMT_VINFO_VECTYPE (stmt_info); + nelements = TYPE_VECTOR_SUBPARTS (vectype); + unsigned int target_align = DR_TARGET_ALIGNMENT (dr); + unsigned int dr_size = vect_get_scalar_dr_size (dr); + mis = (negative ? DR_MISALIGNMENT (dr) : -DR_MISALIGNMENT (dr)); if (DR_MISALIGNMENT (dr) != 0) - npeel_tmp = (negative ? (mis - nelements) - : (nelements - mis)) & (nelements - 1); + npeel_tmp = (mis & (target_align - 1)) / dr_size; /* For multiple types, it is possible that the bigger type access will have more than one peeling option. E.g., a loop with two @@ -1701,7 +1714,7 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) { vect_peeling_hash_insert (&peeling_htab, loop_vinfo, dr, npeel_tmp); - npeel_tmp += nelements; + npeel_tmp += target_align / dr_size; } one_misalignment_known = true; @@ -1922,7 +1935,6 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) stmt = DR_STMT (dr0); stmt_info = vinfo_for_stmt (stmt); vectype = STMT_VINFO_VECTYPE (stmt_info); - nelements = TYPE_VECTOR_SUBPARTS (vectype); if (known_alignment_for_access_p (dr0)) { @@ -1935,9 +1947,10 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) updating DR_MISALIGNMENT values. The peeling factor is the vectorization factor minus the misalignment as an element count. */ - mis = DR_MISALIGNMENT (dr0) / vect_get_scalar_dr_size (dr0); - npeel = ((negative ? mis - nelements : nelements - mis) - & (nelements - 1)); + mis = negative ? DR_MISALIGNMENT (dr0) : -DR_MISALIGNMENT (dr0); + unsigned int target_align = DR_TARGET_ALIGNMENT (dr0); + npeel = ((mis & (target_align - 1)) + / vect_get_scalar_dr_size (dr0)); } /* For interleaved data access every iteration accesses all the @@ -1976,10 +1989,8 @@ vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) unsigned max_peel = npeel; if (max_peel == 0) { - gimple *dr_stmt = DR_STMT (dr0); - stmt_vec_info vinfo = vinfo_for_stmt (dr_stmt); - tree vtype = STMT_VINFO_VECTYPE (vinfo); - max_peel = TYPE_VECTOR_SUBPARTS (vtype) - 1; + unsigned int target_align = DR_TARGET_ALIGNMENT (dr0); + max_peel = target_align / vect_get_scalar_dr_size (dr0) - 1; } if (max_peel > max_allowed_peel) { @@ -2201,8 +2212,10 @@ vect_find_same_alignment_drs (struct data_dependence_relation *ddr) if (diff != 0) { /* Get the wider of the two alignments. */ - unsigned int align_a = TYPE_ALIGN_UNIT (STMT_VINFO_VECTYPE (stmtinfo_a)); - unsigned int align_b = TYPE_ALIGN_UNIT (STMT_VINFO_VECTYPE (stmtinfo_b)); + unsigned int align_a = (vect_calculate_target_alignment (dra) + / BITS_PER_UNIT); + unsigned int align_b = (vect_calculate_target_alignment (drb) + / BITS_PER_UNIT); unsigned int max_align = MAX (align_a, align_b); /* Require the gap to be a multiple of the larger vector alignment. */ @@ -3995,16 +4008,15 @@ vect_get_new_ssa_name (tree type, enum vect_var_kind var_kind, const char *name) /* Duplicate ptr info and set alignment/misaligment on NAME from DR. */ static void -vect_duplicate_ssa_name_ptr_info (tree name, data_reference *dr, - stmt_vec_info stmt_info) +vect_duplicate_ssa_name_ptr_info (tree name, data_reference *dr) { duplicate_ssa_name_ptr_info (name, DR_PTR_INFO (dr)); - unsigned int align = TYPE_ALIGN_UNIT (STMT_VINFO_VECTYPE (stmt_info)); int misalign = DR_MISALIGNMENT (dr); if (misalign == DR_MISALIGNMENT_UNKNOWN) mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (name)); else - set_ptr_info_alignment (SSA_NAME_PTR_INFO (name), align, misalign); + set_ptr_info_alignment (SSA_NAME_PTR_INFO (name), + DR_TARGET_ALIGNMENT (dr), misalign); } /* Function vect_create_addr_base_for_vector_ref. @@ -4109,7 +4121,7 @@ vect_create_addr_base_for_vector_ref (gimple *stmt, && TREE_CODE (addr_base) == SSA_NAME && !SSA_NAME_PTR_INFO (addr_base)) { - vect_duplicate_ssa_name_ptr_info (addr_base, dr, stmt_info); + vect_duplicate_ssa_name_ptr_info (addr_base, dr); if (offset || byte_offset) mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (addr_base)); } @@ -4368,8 +4380,8 @@ vect_create_data_ref_ptr (gimple *stmt, tree aggr_type, struct loop *at_loop, /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) { - vect_duplicate_ssa_name_ptr_info (indx_before_incr, dr, stmt_info); - vect_duplicate_ssa_name_ptr_info (indx_after_incr, dr, stmt_info); + vect_duplicate_ssa_name_ptr_info (indx_before_incr, dr); + vect_duplicate_ssa_name_ptr_info (indx_after_incr, dr); } if (ptr_incr) *ptr_incr = incr; @@ -4398,8 +4410,8 @@ vect_create_data_ref_ptr (gimple *stmt, tree aggr_type, struct loop *at_loop, /* Copy the points-to information if it exists. */ if (DR_PTR_INFO (dr)) { - vect_duplicate_ssa_name_ptr_info (indx_before_incr, dr, stmt_info); - vect_duplicate_ssa_name_ptr_info (indx_after_incr, dr, stmt_info); + vect_duplicate_ssa_name_ptr_info (indx_before_incr, dr); + vect_duplicate_ssa_name_ptr_info (indx_after_incr, dr); } if (ptr_incr) *ptr_incr = incr; @@ -5003,10 +5015,10 @@ vect_setup_realignment (gimple *stmt, gimple_stmt_iterator *gsi, new_temp = copy_ssa_name (ptr); else new_temp = make_ssa_name (TREE_TYPE (ptr)); + unsigned int align = DR_TARGET_ALIGNMENT (dr); new_stmt = gimple_build_assign (new_temp, BIT_AND_EXPR, ptr, - build_int_cst (TREE_TYPE (ptr), - -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); + build_int_cst (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); gcc_assert (!new_bb); data_ref diff --git a/gcc/tree-vect-loop-manip.c b/gcc/tree-vect-loop-manip.c index f78e4b420c3..5787d53a83a 100644 --- a/gcc/tree-vect-loop-manip.c +++ b/gcc/tree-vect-loop-manip.c @@ -956,8 +956,7 @@ vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo, gimple *dr_stmt = DR_STMT (dr); stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt); tree vectype = STMT_VINFO_VECTYPE (stmt_info); - int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT; - int nelements = TYPE_VECTOR_SUBPARTS (vectype); + unsigned int target_align = DR_TARGET_ALIGNMENT (dr); if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0) { @@ -978,32 +977,36 @@ vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo, tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt, &stmts, offset); tree type = unsigned_type_for (TREE_TYPE (start_addr)); - tree vectype_align_minus_1 = build_int_cst (type, vectype_align - 1); - HOST_WIDE_INT elem_size = - int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); + tree target_align_minus_1 = build_int_cst (type, target_align - 1); + HOST_WIDE_INT elem_size + = int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); tree elem_size_log = build_int_cst (type, exact_log2 (elem_size)); - tree nelements_minus_1 = build_int_cst (type, nelements - 1); - tree nelements_tree = build_int_cst (type, nelements); - tree byte_misalign; - tree elem_misalign; - - /* Create: byte_misalign = addr & (vectype_align - 1) */ - byte_misalign = - fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), - vectype_align_minus_1); - - /* Create: elem_misalign = byte_misalign / element_size */ - elem_misalign = - fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log); - - /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */ + HOST_WIDE_INT align_in_elems = target_align / elem_size; + tree align_in_elems_minus_1 = build_int_cst (type, align_in_elems - 1); + tree align_in_elems_tree = build_int_cst (type, align_in_elems); + tree misalign_in_bytes; + tree misalign_in_elems; + + /* Create: misalign_in_bytes = addr & (target_align - 1). */ + misalign_in_bytes + = fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), + target_align_minus_1); + + /* Create: misalign_in_elems = misalign_in_bytes / element_size. */ + misalign_in_elems + = fold_build2 (RSHIFT_EXPR, type, misalign_in_bytes, elem_size_log); + + /* Create: (niters_type) ((align_in_elems - misalign_in_elems) + & (align_in_elems - 1)). */ if (negative) - iters = fold_build2 (MINUS_EXPR, type, elem_misalign, nelements_tree); + iters = fold_build2 (MINUS_EXPR, type, misalign_in_elems, + align_in_elems_tree); else - iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign); - iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1); + iters = fold_build2 (MINUS_EXPR, type, align_in_elems_tree, + misalign_in_elems); + iters = fold_build2 (BIT_AND_EXPR, type, iters, align_in_elems_minus_1); iters = fold_convert (niters_type, iters); - *bound = nelements - 1; + *bound = align_in_elems - 1; } if (dump_enabled_p ()) diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c index 8f0d3d09096..29b733324d7 100644 --- a/gcc/tree-vect-stmts.c +++ b/gcc/tree-vect-stmts.c @@ -1737,6 +1737,7 @@ get_group_load_store_type (gimple *stmt, tree vectype, bool slp, loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); struct loop *loop = loop_vinfo ? LOOP_VINFO_LOOP (loop_vinfo) : NULL; gimple *first_stmt = GROUP_FIRST_ELEMENT (stmt_info); + data_reference *first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)); unsigned int group_size = GROUP_SIZE (vinfo_for_stmt (first_stmt)); bool single_element_p = (stmt == first_stmt && !GROUP_NEXT_ELEMENT (stmt_info)); @@ -1780,10 +1781,13 @@ get_group_load_store_type (gimple *stmt, tree vectype, bool slp, " non-consecutive accesses\n"); return false; } - /* If the access is aligned an overrun is fine. */ + /* An overrun is fine if the trailing elements are smaller + than the alignment boundary B. Every vector access will + be a multiple of B and so we are guaranteed to access a + non-gap element in the same B-sized block. */ if (overrun_p - && aligned_access_p - (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))) + && gap < (vect_known_alignment_in_bytes (first_dr) + / vect_get_scalar_dr_size (first_dr))) overrun_p = false; if (overrun_p && !can_overrun_p) { @@ -1804,14 +1808,15 @@ get_group_load_store_type (gimple *stmt, tree vectype, bool slp, /* If there is a gap at the end of the group then these optimizations would access excess elements in the last iteration. */ bool would_overrun_p = (gap != 0); - /* If the access is aligned an overrun is fine, but only if the - overrun is not inside an unused vector (if the gap is as large - or larger than a vector). */ + /* An overrun is fine if the trailing elements are smaller than the + alignment boundary B. Every vector access will be a multiple of B + and so we are guaranteed to access a non-gap element in the + same B-sized block. */ if (would_overrun_p - && gap < nunits - && aligned_access_p - (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))) + && gap < (vect_known_alignment_in_bytes (first_dr) + / vect_get_scalar_dr_size (first_dr))) would_overrun_p = false; + if (!STMT_VINFO_STRIDED_P (stmt_info) && (can_overrun_p || !would_overrun_p) && compare_step_with_zero (stmt) > 0) @@ -2351,7 +2356,7 @@ vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, TYPE_SIZE_UNIT (vectype)); } - align = TYPE_ALIGN_UNIT (vectype); + align = DR_TARGET_ALIGNMENT (dr); if (aligned_access_p (dr)) misalign = 0; else if (DR_MISALIGNMENT (dr) == -1) @@ -2404,7 +2409,7 @@ vectorizable_mask_load_store (gimple *stmt, gimple_stmt_iterator *gsi, TYPE_SIZE_UNIT (vectype)); } - align = TYPE_ALIGN_UNIT (vectype); + align = DR_TARGET_ALIGNMENT (dr); if (aligned_access_p (dr)) misalign = 0; else if (DR_MISALIGNMENT (dr) == -1) @@ -5553,25 +5558,25 @@ vectorizable_operation (gimple *stmt, gimple_stmt_iterator *gsi, return true; } -/* A helper function to ensure data reference DR's base alignment - for STMT_INFO. */ +/* A helper function to ensure data reference DR's base alignment. */ static void -ensure_base_align (stmt_vec_info stmt_info, struct data_reference *dr) +ensure_base_align (struct data_reference *dr) { if (!dr->aux) return; if (DR_VECT_AUX (dr)->base_misaligned) { - tree vectype = STMT_VINFO_VECTYPE (stmt_info); tree base_decl = DR_VECT_AUX (dr)->base_decl; + unsigned int align_base_to = DR_TARGET_ALIGNMENT (dr) * BITS_PER_UNIT; + if (decl_in_symtab_p (base_decl)) - symtab_node::get (base_decl)->increase_alignment (TYPE_ALIGN (vectype)); + symtab_node::get (base_decl)->increase_alignment (align_base_to); else { - SET_DECL_ALIGN (base_decl, TYPE_ALIGN (vectype)); + SET_DECL_ALIGN (base_decl, align_base_to); DECL_USER_ALIGN (base_decl) = 1; } DR_VECT_AUX (dr)->base_misaligned = false; @@ -5775,7 +5780,7 @@ vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, /* Transform. */ - ensure_base_align (stmt_info, dr); + ensure_base_align (dr); if (memory_access_type == VMAT_GATHER_SCATTER) { @@ -6417,7 +6422,7 @@ vectorizable_store (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, dataref_offset ? dataref_offset : build_int_cst (ref_type, 0)); - align = TYPE_ALIGN_UNIT (vectype); + align = DR_TARGET_ALIGNMENT (first_dr); if (aligned_access_p (first_dr)) misalign = 0; else if (DR_MISALIGNMENT (first_dr) == -1) @@ -6813,7 +6818,7 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, /* Transform. */ - ensure_base_align (stmt_info, dr); + ensure_base_align (dr); if (memory_access_type == VMAT_GATHER_SCATTER) { @@ -7512,7 +7517,7 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, dataref_offset ? dataref_offset : build_int_cst (ref_type, 0)); - align = TYPE_ALIGN_UNIT (vectype); + align = DR_TARGET_ALIGNMENT (dr); if (alignment_support_scheme == dr_aligned) { gcc_assert (aligned_access_p (first_dr)); @@ -7555,11 +7560,12 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, ptr = copy_ssa_name (dataref_ptr); else ptr = make_ssa_name (TREE_TYPE (dataref_ptr)); + unsigned int align = DR_TARGET_ALIGNMENT (first_dr); new_stmt = gimple_build_assign (ptr, BIT_AND_EXPR, dataref_ptr, build_int_cst (TREE_TYPE (dataref_ptr), - -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); + -(HOST_WIDE_INT) align)); vect_finish_stmt_generation (stmt, new_stmt, gsi); data_ref = build2 (MEM_REF, vectype, ptr, @@ -7581,8 +7587,7 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, new_stmt = gimple_build_assign (NULL_TREE, BIT_AND_EXPR, ptr, build_int_cst - (TREE_TYPE (ptr), - -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); + (TREE_TYPE (ptr), -(HOST_WIDE_INT) align)); ptr = copy_ssa_name (ptr, new_stmt); gimple_assign_set_lhs (new_stmt, ptr); vect_finish_stmt_generation (stmt, new_stmt, gsi); @@ -7592,20 +7597,22 @@ vectorizable_load (gimple *stmt, gimple_stmt_iterator *gsi, gimple **vec_stmt, break; } case dr_explicit_realign_optimized: - if (TREE_CODE (dataref_ptr) == SSA_NAME) - new_temp = copy_ssa_name (dataref_ptr); - else - new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); - new_stmt = gimple_build_assign - (new_temp, BIT_AND_EXPR, dataref_ptr, - build_int_cst - (TREE_TYPE (dataref_ptr), - -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype))); - vect_finish_stmt_generation (stmt, new_stmt, gsi); - data_ref - = build2 (MEM_REF, vectype, new_temp, - build_int_cst (ref_type, 0)); - break; + { + if (TREE_CODE (dataref_ptr) == SSA_NAME) + new_temp = copy_ssa_name (dataref_ptr); + else + new_temp = make_ssa_name (TREE_TYPE (dataref_ptr)); + unsigned int align = DR_TARGET_ALIGNMENT (first_dr); + new_stmt = gimple_build_assign + (new_temp, BIT_AND_EXPR, dataref_ptr, + build_int_cst (TREE_TYPE (dataref_ptr), + -(HOST_WIDE_INT) align)); + vect_finish_stmt_generation (stmt, new_stmt, gsi); + data_ref + = build2 (MEM_REF, vectype, new_temp, + build_int_cst (ref_type, 0)); + break; + } default: gcc_unreachable (); } diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h index 5d273ca1a68..06224f9f94d 100644 --- a/gcc/tree-vectorizer.h +++ b/gcc/tree-vectorizer.h @@ -790,7 +790,11 @@ STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo) #define STMT_SLP_TYPE(S) (S)->slp_type struct dataref_aux { + /* The misalignment in bytes of the reference, or -1 if not known. */ int misalignment; + /* The byte alignment that we'd ideally like the reference to have, + and the value that misalignment is measured against. */ + int target_alignment; /* If true the alignment of base_decl needs to be increased. */ bool base_misaligned; tree base_decl; @@ -1037,7 +1041,11 @@ dr_misalignment (struct data_reference *dr) #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL) #define DR_MISALIGNMENT_UNKNOWN (-1) -/* Return TRUE if the data access is aligned, and FALSE otherwise. */ +/* Only defined once DR_MISALIGNMENT is defined. */ +#define DR_TARGET_ALIGNMENT(DR) DR_VECT_AUX (DR)->target_alignment + +/* Return true if data access DR is aligned to its target alignment + (which may be less than a full vector). */ static inline bool aligned_access_p (struct data_reference *data_ref_info) @@ -1054,6 +1062,19 @@ known_alignment_for_access_p (struct data_reference *data_ref_info) return (DR_MISALIGNMENT (data_ref_info) != DR_MISALIGNMENT_UNKNOWN); } +/* Return the minimum alignment in bytes that the vectorized version + of DR is guaranteed to have. */ + +static inline unsigned int +vect_known_alignment_in_bytes (struct data_reference *dr) +{ + if (DR_MISALIGNMENT (dr) == DR_MISALIGNMENT_UNKNOWN) + return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr))); + if (DR_MISALIGNMENT (dr) == 0) + return DR_TARGET_ALIGNMENT (dr); + return DR_MISALIGNMENT (dr) & -DR_MISALIGNMENT (dr); +} + /* Return the behavior of DR with respect to the vectorization context (which for outer loop vectorization might not be the behavior recorded in DR itself). */ -- 2.30.2