/* Vectorizer Specific Loop Manipulations
- Copyright (C) 2003-2017 Free Software Foundation, Inc.
+ Copyright (C) 2003-2019 Free Software Foundation, Inc.
Contributed by Dorit Naishlos <dorit@il.ibm.com>
and Ira Rosen <irar@il.ibm.com>
#include "tree-scalar-evolution.h"
#include "tree-vectorizer.h"
#include "tree-ssa-loop-ivopts.h"
+#include "gimple-fold.h"
+#include "tree-ssa-loop-niter.h"
+#include "internal-fn.h"
+#include "stor-layout.h"
+#include "optabs-query.h"
+#include "vec-perm-indices.h"
/*************************************************************************
Simple Loop Peeling Utilities
ssa_op_iter iter;
edge e;
edge_iterator ei;
- struct loop *loop = bb->loop_father;
- struct loop *outer_loop = NULL;
+ class loop *loop = bb->loop_father;
+ class loop *outer_loop = NULL;
if (rename_from_outer_loop)
{
gimple_bb (update_phi));
}
-/* Make the LOOP iterate NITERS times. This is done by adding a new IV
- that starts at zero, increases by one and its limit is NITERS.
+/* Define one loop mask MASK from loop LOOP. INIT_MASK is the value that
+ the mask should have during the first iteration and NEXT_MASK is the
+ value that it should have on subsequent iterations. */
- Assumption: the exit-condition of LOOP is the last stmt in the loop. */
+static void
+vect_set_loop_mask (class loop *loop, tree mask, tree init_mask,
+ tree next_mask)
+{
+ gphi *phi = create_phi_node (mask, loop->header);
+ add_phi_arg (phi, init_mask, loop_preheader_edge (loop), UNKNOWN_LOCATION);
+ add_phi_arg (phi, next_mask, loop_latch_edge (loop), UNKNOWN_LOCATION);
+}
-void
-slpeel_make_loop_iterate_ntimes (struct loop *loop, tree niters)
+/* Add SEQ to the end of LOOP's preheader block. */
+
+static void
+add_preheader_seq (class loop *loop, gimple_seq seq)
+{
+ if (seq)
+ {
+ edge pe = loop_preheader_edge (loop);
+ basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
+ gcc_assert (!new_bb);
+ }
+}
+
+/* Add SEQ to the beginning of LOOP's header block. */
+
+static void
+add_header_seq (class loop *loop, gimple_seq seq)
+{
+ if (seq)
+ {
+ gimple_stmt_iterator gsi = gsi_after_labels (loop->header);
+ gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
+ }
+}
+
+/* Return true if the target can interleave elements of two vectors.
+ OFFSET is 0 if the first half of the vectors should be interleaved
+ or 1 if the second half should. When returning true, store the
+ associated permutation in INDICES. */
+
+static bool
+interleave_supported_p (vec_perm_indices *indices, tree vectype,
+ unsigned int offset)
+{
+ poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (vectype);
+ poly_uint64 base = exact_div (nelts, 2) * offset;
+ vec_perm_builder sel (nelts, 2, 3);
+ for (unsigned int i = 0; i < 3; ++i)
+ {
+ sel.quick_push (base + i);
+ sel.quick_push (base + i + nelts);
+ }
+ indices->new_vector (sel, 2, nelts);
+ return can_vec_perm_const_p (TYPE_MODE (vectype), *indices);
+}
+
+/* Try to use permutes to define the masks in DEST_RGM using the masks
+ in SRC_RGM, given that the former has twice as many masks as the
+ latter. Return true on success, adding any new statements to SEQ. */
+
+static bool
+vect_maybe_permute_loop_masks (gimple_seq *seq, rgroup_masks *dest_rgm,
+ rgroup_masks *src_rgm)
+{
+ tree src_masktype = src_rgm->mask_type;
+ tree dest_masktype = dest_rgm->mask_type;
+ machine_mode src_mode = TYPE_MODE (src_masktype);
+ if (dest_rgm->max_nscalars_per_iter <= src_rgm->max_nscalars_per_iter
+ && optab_handler (vec_unpacku_hi_optab, src_mode) != CODE_FOR_nothing
+ && optab_handler (vec_unpacku_lo_optab, src_mode) != CODE_FOR_nothing)
+ {
+ /* Unpacking the source masks gives at least as many mask bits as
+ we need. We can then VIEW_CONVERT any excess bits away. */
+ tree unpack_masktype = vect_halve_mask_nunits (src_masktype);
+ for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i)
+ {
+ tree src = src_rgm->masks[i / 2];
+ tree dest = dest_rgm->masks[i];
+ tree_code code = ((i & 1) == (BYTES_BIG_ENDIAN ? 0 : 1)
+ ? VEC_UNPACK_HI_EXPR
+ : VEC_UNPACK_LO_EXPR);
+ gassign *stmt;
+ if (dest_masktype == unpack_masktype)
+ stmt = gimple_build_assign (dest, code, src);
+ else
+ {
+ tree temp = make_ssa_name (unpack_masktype);
+ stmt = gimple_build_assign (temp, code, src);
+ gimple_seq_add_stmt (seq, stmt);
+ stmt = gimple_build_assign (dest, VIEW_CONVERT_EXPR,
+ build1 (VIEW_CONVERT_EXPR,
+ dest_masktype, temp));
+ }
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ return true;
+ }
+ vec_perm_indices indices[2];
+ if (dest_masktype == src_masktype
+ && interleave_supported_p (&indices[0], src_masktype, 0)
+ && interleave_supported_p (&indices[1], src_masktype, 1))
+ {
+ /* The destination requires twice as many mask bits as the source, so
+ we can use interleaving permutes to double up the number of bits. */
+ tree masks[2];
+ for (unsigned int i = 0; i < 2; ++i)
+ masks[i] = vect_gen_perm_mask_checked (src_masktype, indices[i]);
+ for (unsigned int i = 0; i < dest_rgm->masks.length (); ++i)
+ {
+ tree src = src_rgm->masks[i / 2];
+ tree dest = dest_rgm->masks[i];
+ gimple *stmt = gimple_build_assign (dest, VEC_PERM_EXPR,
+ src, src, masks[i & 1]);
+ gimple_seq_add_stmt (seq, stmt);
+ }
+ return true;
+ }
+ return false;
+}
+
+/* Helper for vect_set_loop_condition_masked. Generate definitions for
+ all the masks in RGM and return a mask that is nonzero when the loop
+ needs to iterate. Add any new preheader statements to PREHEADER_SEQ.
+ Use LOOP_COND_GSI to insert code before the exit gcond.
+
+ RGM belongs to loop LOOP. The loop originally iterated NITERS
+ times and has been vectorized according to LOOP_VINFO.
+
+ If NITERS_SKIP is nonnull, the first iteration of the vectorized loop
+ starts with NITERS_SKIP dummy iterations of the scalar loop before
+ the real work starts. The mask elements for these dummy iterations
+ must be 0, to ensure that the extra iterations do not have an effect.
+
+ It is known that:
+
+ NITERS * RGM->max_nscalars_per_iter
+
+ does not overflow. However, MIGHT_WRAP_P says whether an induction
+ variable that starts at 0 and has step:
+
+ VF * RGM->max_nscalars_per_iter
+
+ might overflow before hitting a value above:
+
+ (NITERS + NITERS_SKIP) * RGM->max_nscalars_per_iter
+
+ This means that we cannot guarantee that such an induction variable
+ would ever hit a value that produces a set of all-false masks for RGM. */
+
+static tree
+vect_set_loop_masks_directly (class loop *loop, loop_vec_info loop_vinfo,
+ gimple_seq *preheader_seq,
+ gimple_stmt_iterator loop_cond_gsi,
+ rgroup_masks *rgm, tree niters, tree niters_skip,
+ bool might_wrap_p)
+{
+ tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo);
+ tree iv_type = LOOP_VINFO_MASK_IV_TYPE (loop_vinfo);
+ tree mask_type = rgm->mask_type;
+ unsigned int nscalars_per_iter = rgm->max_nscalars_per_iter;
+ poly_uint64 nscalars_per_mask = TYPE_VECTOR_SUBPARTS (mask_type);
+ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+
+ /* Calculate the maximum number of scalar values that the rgroup
+ handles in total, the number that it handles for each iteration
+ of the vector loop, and the number that it should skip during the
+ first iteration of the vector loop. */
+ tree nscalars_total = niters;
+ tree nscalars_step = build_int_cst (iv_type, vf);
+ tree nscalars_skip = niters_skip;
+ if (nscalars_per_iter != 1)
+ {
+ /* We checked before choosing to use a fully-masked loop that these
+ multiplications don't overflow. */
+ tree compare_factor = build_int_cst (compare_type, nscalars_per_iter);
+ tree iv_factor = build_int_cst (iv_type, nscalars_per_iter);
+ nscalars_total = gimple_build (preheader_seq, MULT_EXPR, compare_type,
+ nscalars_total, compare_factor);
+ nscalars_step = gimple_build (preheader_seq, MULT_EXPR, iv_type,
+ nscalars_step, iv_factor);
+ if (nscalars_skip)
+ nscalars_skip = gimple_build (preheader_seq, MULT_EXPR, compare_type,
+ nscalars_skip, compare_factor);
+ }
+
+ /* Create an induction variable that counts the number of scalars
+ processed. */
+ tree index_before_incr, index_after_incr;
+ gimple_stmt_iterator incr_gsi;
+ bool insert_after;
+ standard_iv_increment_position (loop, &incr_gsi, &insert_after);
+ create_iv (build_int_cst (iv_type, 0), nscalars_step, NULL_TREE, loop,
+ &incr_gsi, insert_after, &index_before_incr, &index_after_incr);
+
+ tree zero_index = build_int_cst (compare_type, 0);
+ tree test_index, test_limit, first_limit;
+ gimple_stmt_iterator *test_gsi;
+ if (might_wrap_p)
+ {
+ /* In principle the loop should stop iterating once the incremented
+ IV reaches a value greater than or equal to:
+
+ NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP
+
+ However, there's no guarantee that this addition doesn't overflow
+ the comparison type, or that the IV hits a value above it before
+ wrapping around. We therefore adjust the limit down by one
+ IV step:
+
+ (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP)
+ -[infinite-prec] NSCALARS_STEP
+
+ and compare the IV against this limit _before_ incrementing it.
+ Since the comparison type is unsigned, we actually want the
+ subtraction to saturate at zero:
+
+ (NSCALARS_TOTAL +[infinite-prec] NSCALARS_SKIP)
+ -[sat] NSCALARS_STEP
+
+ And since NSCALARS_SKIP < NSCALARS_STEP, we can reassociate this as:
+
+ NSCALARS_TOTAL -[sat] (NSCALARS_STEP - NSCALARS_SKIP)
+
+ where the rightmost subtraction can be done directly in
+ COMPARE_TYPE. */
+ test_index = index_before_incr;
+ tree adjust = gimple_convert (preheader_seq, compare_type,
+ nscalars_step);
+ if (nscalars_skip)
+ adjust = gimple_build (preheader_seq, MINUS_EXPR, compare_type,
+ adjust, nscalars_skip);
+ test_limit = gimple_build (preheader_seq, MAX_EXPR, compare_type,
+ nscalars_total, adjust);
+ test_limit = gimple_build (preheader_seq, MINUS_EXPR, compare_type,
+ test_limit, adjust);
+ test_gsi = &incr_gsi;
+
+ /* Get a safe limit for the first iteration. */
+ if (nscalars_skip)
+ {
+ /* The first vector iteration can handle at most NSCALARS_STEP
+ scalars. NSCALARS_STEP <= CONST_LIMIT, and adding
+ NSCALARS_SKIP to that cannot overflow. */
+ tree const_limit = build_int_cst (compare_type,
+ LOOP_VINFO_VECT_FACTOR (loop_vinfo)
+ * nscalars_per_iter);
+ first_limit = gimple_build (preheader_seq, MIN_EXPR, compare_type,
+ nscalars_total, const_limit);
+ first_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type,
+ first_limit, nscalars_skip);
+ }
+ else
+ /* For the first iteration it doesn't matter whether the IV hits
+ a value above NSCALARS_TOTAL. That only matters for the latch
+ condition. */
+ first_limit = nscalars_total;
+ }
+ else
+ {
+ /* Test the incremented IV, which will always hit a value above
+ the bound before wrapping. */
+ test_index = index_after_incr;
+ test_limit = nscalars_total;
+ if (nscalars_skip)
+ test_limit = gimple_build (preheader_seq, PLUS_EXPR, compare_type,
+ test_limit, nscalars_skip);
+ test_gsi = &loop_cond_gsi;
+
+ first_limit = test_limit;
+ }
+
+ /* Convert the IV value to the comparison type (either a no-op or
+ a demotion). */
+ gimple_seq test_seq = NULL;
+ test_index = gimple_convert (&test_seq, compare_type, test_index);
+ gsi_insert_seq_before (test_gsi, test_seq, GSI_SAME_STMT);
+
+ /* Provide a definition of each mask in the group. */
+ tree next_mask = NULL_TREE;
+ tree mask;
+ unsigned int i;
+ FOR_EACH_VEC_ELT_REVERSE (rgm->masks, i, mask)
+ {
+ /* Previous masks will cover BIAS scalars. This mask covers the
+ next batch. */
+ poly_uint64 bias = nscalars_per_mask * i;
+ tree bias_tree = build_int_cst (compare_type, bias);
+ gimple *tmp_stmt;
+
+ /* See whether the first iteration of the vector loop is known
+ to have a full mask. */
+ poly_uint64 const_limit;
+ bool first_iteration_full
+ = (poly_int_tree_p (first_limit, &const_limit)
+ && known_ge (const_limit, (i + 1) * nscalars_per_mask));
+
+ /* Rather than have a new IV that starts at BIAS and goes up to
+ TEST_LIMIT, prefer to use the same 0-based IV for each mask
+ and adjust the bound down by BIAS. */
+ tree this_test_limit = test_limit;
+ if (i != 0)
+ {
+ this_test_limit = gimple_build (preheader_seq, MAX_EXPR,
+ compare_type, this_test_limit,
+ bias_tree);
+ this_test_limit = gimple_build (preheader_seq, MINUS_EXPR,
+ compare_type, this_test_limit,
+ bias_tree);
+ }
+
+ /* Create the initial mask. First include all scalars that
+ are within the loop limit. */
+ tree init_mask = NULL_TREE;
+ if (!first_iteration_full)
+ {
+ tree start, end;
+ if (first_limit == test_limit)
+ {
+ /* Use a natural test between zero (the initial IV value)
+ and the loop limit. The "else" block would be valid too,
+ but this choice can avoid the need to load BIAS_TREE into
+ a register. */
+ start = zero_index;
+ end = this_test_limit;
+ }
+ else
+ {
+ /* FIRST_LIMIT is the maximum number of scalars handled by the
+ first iteration of the vector loop. Test the portion
+ associated with this mask. */
+ start = bias_tree;
+ end = first_limit;
+ }
+
+ init_mask = make_temp_ssa_name (mask_type, NULL, "max_mask");
+ tmp_stmt = vect_gen_while (init_mask, start, end);
+ gimple_seq_add_stmt (preheader_seq, tmp_stmt);
+ }
+
+ /* Now AND out the bits that are within the number of skipped
+ scalars. */
+ poly_uint64 const_skip;
+ if (nscalars_skip
+ && !(poly_int_tree_p (nscalars_skip, &const_skip)
+ && known_le (const_skip, bias)))
+ {
+ tree unskipped_mask = vect_gen_while_not (preheader_seq, mask_type,
+ bias_tree, nscalars_skip);
+ if (init_mask)
+ init_mask = gimple_build (preheader_seq, BIT_AND_EXPR, mask_type,
+ init_mask, unskipped_mask);
+ else
+ init_mask = unskipped_mask;
+ }
+
+ if (!init_mask)
+ /* First iteration is full. */
+ init_mask = build_minus_one_cst (mask_type);
+
+ /* Get the mask value for the next iteration of the loop. */
+ next_mask = make_temp_ssa_name (mask_type, NULL, "next_mask");
+ gcall *call = vect_gen_while (next_mask, test_index, this_test_limit);
+ gsi_insert_before (test_gsi, call, GSI_SAME_STMT);
+
+ vect_set_loop_mask (loop, mask, init_mask, next_mask);
+ }
+ return next_mask;
+}
+
+/* Make LOOP iterate NITERS times using masking and WHILE_ULT calls.
+ LOOP_VINFO describes the vectorization of LOOP. NITERS is the
+ number of iterations of the original scalar loop that should be
+ handled by the vector loop. NITERS_MAYBE_ZERO and FINAL_IV are
+ as for vect_set_loop_condition.
+
+ Insert the branch-back condition before LOOP_COND_GSI and return the
+ final gcond. */
+
+static gcond *
+vect_set_loop_condition_masked (class loop *loop, loop_vec_info loop_vinfo,
+ tree niters, tree final_iv,
+ bool niters_maybe_zero,
+ gimple_stmt_iterator loop_cond_gsi)
+{
+ gimple_seq preheader_seq = NULL;
+ gimple_seq header_seq = NULL;
+
+ tree compare_type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo);
+ unsigned int compare_precision = TYPE_PRECISION (compare_type);
+ tree orig_niters = niters;
+
+ /* Type of the initial value of NITERS. */
+ tree ni_actual_type = TREE_TYPE (niters);
+ unsigned int ni_actual_precision = TYPE_PRECISION (ni_actual_type);
+ tree niters_skip = LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo);
+
+ /* Convert NITERS to the same size as the compare. */
+ if (compare_precision > ni_actual_precision
+ && niters_maybe_zero)
+ {
+ /* We know that there is always at least one iteration, so if the
+ count is zero then it must have wrapped. Cope with this by
+ subtracting 1 before the conversion and adding 1 to the result. */
+ gcc_assert (TYPE_UNSIGNED (ni_actual_type));
+ niters = gimple_build (&preheader_seq, PLUS_EXPR, ni_actual_type,
+ niters, build_minus_one_cst (ni_actual_type));
+ niters = gimple_convert (&preheader_seq, compare_type, niters);
+ niters = gimple_build (&preheader_seq, PLUS_EXPR, compare_type,
+ niters, build_one_cst (compare_type));
+ }
+ else
+ niters = gimple_convert (&preheader_seq, compare_type, niters);
+
+ widest_int iv_limit = vect_iv_limit_for_full_masking (loop_vinfo);
+
+ /* Iterate over all the rgroups and fill in their masks. We could use
+ the first mask from any rgroup for the loop condition; here we
+ arbitrarily pick the last. */
+ tree test_mask = NULL_TREE;
+ rgroup_masks *rgm;
+ unsigned int i;
+ vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
+ FOR_EACH_VEC_ELT (*masks, i, rgm)
+ if (!rgm->masks.is_empty ())
+ {
+ /* First try using permutes. This adds a single vector
+ instruction to the loop for each mask, but needs no extra
+ loop invariants or IVs. */
+ unsigned int nmasks = i + 1;
+ if ((nmasks & 1) == 0)
+ {
+ rgroup_masks *half_rgm = &(*masks)[nmasks / 2 - 1];
+ if (!half_rgm->masks.is_empty ()
+ && vect_maybe_permute_loop_masks (&header_seq, rgm, half_rgm))
+ continue;
+ }
+
+ /* See whether zero-based IV would ever generate all-false masks
+ before wrapping around. */
+ bool might_wrap_p
+ = (iv_limit == -1
+ || (wi::min_precision (iv_limit * rgm->max_nscalars_per_iter,
+ UNSIGNED)
+ > compare_precision));
+
+ /* Set up all masks for this group. */
+ test_mask = vect_set_loop_masks_directly (loop, loop_vinfo,
+ &preheader_seq,
+ loop_cond_gsi, rgm,
+ niters, niters_skip,
+ might_wrap_p);
+ }
+
+ /* Emit all accumulated statements. */
+ add_preheader_seq (loop, preheader_seq);
+ add_header_seq (loop, header_seq);
+
+ /* Get a boolean result that tells us whether to iterate. */
+ edge exit_edge = single_exit (loop);
+ tree_code code = (exit_edge->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR;
+ tree zero_mask = build_zero_cst (TREE_TYPE (test_mask));
+ gcond *cond_stmt = gimple_build_cond (code, test_mask, zero_mask,
+ NULL_TREE, NULL_TREE);
+ gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT);
+
+ /* The loop iterates (NITERS - 1) / VF + 1 times.
+ Subtract one from this to get the latch count. */
+ tree step = build_int_cst (compare_type,
+ LOOP_VINFO_VECT_FACTOR (loop_vinfo));
+ tree niters_minus_one = fold_build2 (PLUS_EXPR, compare_type, niters,
+ build_minus_one_cst (compare_type));
+ loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, compare_type,
+ niters_minus_one, step);
+
+ if (final_iv)
+ {
+ gassign *assign = gimple_build_assign (final_iv, orig_niters);
+ gsi_insert_on_edge_immediate (single_exit (loop), assign);
+ }
+
+ return cond_stmt;
+}
+
+/* Like vect_set_loop_condition, but handle the case in which there
+ are no loop masks. */
+
+static gcond *
+vect_set_loop_condition_unmasked (class loop *loop, tree niters,
+ tree step, tree final_iv,
+ bool niters_maybe_zero,
+ gimple_stmt_iterator loop_cond_gsi)
{
tree indx_before_incr, indx_after_incr;
gcond *cond_stmt;
gcond *orig_cond;
+ edge pe = loop_preheader_edge (loop);
edge exit_edge = single_exit (loop);
- gimple_stmt_iterator loop_cond_gsi;
gimple_stmt_iterator incr_gsi;
bool insert_after;
- tree init = build_int_cst (TREE_TYPE (niters), 0);
- tree step = build_int_cst (TREE_TYPE (niters), 1);
- source_location loop_loc;
enum tree_code code;
+ tree niters_type = TREE_TYPE (niters);
orig_cond = get_loop_exit_condition (loop);
gcc_assert (orig_cond);
loop_cond_gsi = gsi_for_stmt (orig_cond);
+ tree init, limit;
+ if (!niters_maybe_zero && integer_onep (step))
+ {
+ /* In this case we can use a simple 0-based IV:
+
+ A:
+ x = 0;
+ do
+ {
+ ...
+ x += 1;
+ }
+ while (x < NITERS); */
+ code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR;
+ init = build_zero_cst (niters_type);
+ limit = niters;
+ }
+ else
+ {
+ /* The following works for all values of NITERS except 0:
+
+ B:
+ x = 0;
+ do
+ {
+ ...
+ x += STEP;
+ }
+ while (x <= NITERS - STEP);
+
+ so that the loop continues to iterate if x + STEP - 1 < NITERS
+ but stops if x + STEP - 1 >= NITERS.
+
+ However, if NITERS is zero, x never hits a value above NITERS - STEP
+ before wrapping around. There are two obvious ways of dealing with
+ this:
+
+ - start at STEP - 1 and compare x before incrementing it
+ - start at -1 and compare x after incrementing it
+
+ The latter is simpler and is what we use. The loop in this case
+ looks like:
+
+ C:
+ x = -1;
+ do
+ {
+ ...
+ x += STEP;
+ }
+ while (x < NITERS - STEP);
+
+ In both cases the loop limit is NITERS - STEP. */
+ gimple_seq seq = NULL;
+ limit = force_gimple_operand (niters, &seq, true, NULL_TREE);
+ limit = gimple_build (&seq, MINUS_EXPR, TREE_TYPE (limit), limit, step);
+ if (seq)
+ {
+ basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
+ gcc_assert (!new_bb);
+ }
+ if (niters_maybe_zero)
+ {
+ /* Case C. */
+ code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR;
+ init = build_all_ones_cst (niters_type);
+ }
+ else
+ {
+ /* Case B. */
+ code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GT_EXPR : LE_EXPR;
+ init = build_zero_cst (niters_type);
+ }
+ }
+
standard_iv_increment_position (loop, &incr_gsi, &insert_after);
create_iv (init, step, NULL_TREE, loop,
&incr_gsi, insert_after, &indx_before_incr, &indx_after_incr);
-
indx_after_incr = force_gimple_operand_gsi (&loop_cond_gsi, indx_after_incr,
true, NULL_TREE, true,
GSI_SAME_STMT);
- niters = force_gimple_operand_gsi (&loop_cond_gsi, niters, true, NULL_TREE,
+ limit = force_gimple_operand_gsi (&loop_cond_gsi, limit, true, NULL_TREE,
true, GSI_SAME_STMT);
- code = (exit_edge->flags & EDGE_TRUE_VALUE) ? GE_EXPR : LT_EXPR;
- cond_stmt = gimple_build_cond (code, indx_after_incr, niters, NULL_TREE,
+ cond_stmt = gimple_build_cond (code, indx_after_incr, limit, NULL_TREE,
NULL_TREE);
gsi_insert_before (&loop_cond_gsi, cond_stmt, GSI_SAME_STMT);
- /* Remove old loop exit test: */
- gsi_remove (&loop_cond_gsi, true);
- free_stmt_vec_info (orig_cond);
+ /* Record the number of latch iterations. */
+ if (limit == niters)
+ /* Case A: the loop iterates NITERS times. Subtract one to get the
+ latch count. */
+ loop->nb_iterations = fold_build2 (MINUS_EXPR, niters_type, niters,
+ build_int_cst (niters_type, 1));
+ else
+ /* Case B or C: the loop iterates (NITERS - STEP) / STEP + 1 times.
+ Subtract one from this to get the latch count. */
+ loop->nb_iterations = fold_build2 (TRUNC_DIV_EXPR, niters_type,
+ limit, step);
- loop_loc = find_loop_location (loop);
- if (dump_enabled_p ())
+ if (final_iv)
{
- if (LOCATION_LOCUS (loop_loc) != UNKNOWN_LOCATION)
- dump_printf (MSG_NOTE, "\nloop at %s:%d: ", LOCATION_FILE (loop_loc),
- LOCATION_LINE (loop_loc));
- dump_gimple_stmt (MSG_NOTE, TDF_SLIM, cond_stmt, 0);
+ gassign *assign = gimple_build_assign (final_iv, MINUS_EXPR,
+ indx_after_incr, init);
+ gsi_insert_on_edge_immediate (single_exit (loop), assign);
}
- /* Record the number of latch iterations. */
- loop->nb_iterations = fold_build2 (MINUS_EXPR, TREE_TYPE (niters), niters,
- build_int_cst (TREE_TYPE (niters), 1));
+ return cond_stmt;
+}
+
+/* If we're using fully-masked loops, make LOOP iterate:
+
+ N == (NITERS - 1) / STEP + 1
+
+ times. When NITERS is zero, this is equivalent to making the loop
+ execute (1 << M) / STEP times, where M is the precision of NITERS.
+ NITERS_MAYBE_ZERO is true if this last case might occur.
+
+ If we're not using fully-masked loops, make LOOP iterate:
+
+ N == (NITERS - STEP) / STEP + 1
+
+ times, where NITERS is known to be outside the range [1, STEP - 1].
+ This is equivalent to making the loop execute NITERS / STEP times
+ when NITERS is nonzero and (1 << M) / STEP times otherwise.
+ NITERS_MAYBE_ZERO again indicates whether this last case might occur.
+
+ If FINAL_IV is nonnull, it is an SSA name that should be set to
+ N * STEP on exit from the loop.
+
+ Assumption: the exit-condition of LOOP is the last stmt in the loop. */
+
+void
+vect_set_loop_condition (class loop *loop, loop_vec_info loop_vinfo,
+ tree niters, tree step, tree final_iv,
+ bool niters_maybe_zero)
+{
+ gcond *cond_stmt;
+ gcond *orig_cond = get_loop_exit_condition (loop);
+ gimple_stmt_iterator loop_cond_gsi = gsi_for_stmt (orig_cond);
+
+ if (loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
+ cond_stmt = vect_set_loop_condition_masked (loop, loop_vinfo, niters,
+ final_iv, niters_maybe_zero,
+ loop_cond_gsi);
+ else
+ cond_stmt = vect_set_loop_condition_unmasked (loop, niters, step,
+ final_iv, niters_maybe_zero,
+ loop_cond_gsi);
+
+ /* Remove old loop exit test. */
+ stmt_vec_info orig_cond_info;
+ if (loop_vinfo
+ && (orig_cond_info = loop_vinfo->lookup_stmt (orig_cond)))
+ loop_vinfo->remove_stmt (orig_cond_info);
+ else
+ gsi_remove (&loop_cond_gsi, true);
+
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location, "New loop exit condition: %G",
+ cond_stmt);
}
/* Helper routine of slpeel_tree_duplicate_loop_to_edge_cfg.
}
if (TREE_CODE (from_arg) != SSA_NAME)
gcc_assert (operand_equal_p (from_arg, to_arg, 0));
- else
+ else if (TREE_CODE (to_arg) == SSA_NAME
+ && from_arg != to_arg)
{
if (get_current_def (to_arg) == NULL_TREE)
- set_current_def (to_arg, get_current_def (from_arg));
+ {
+ gcc_assert (types_compatible_p (TREE_TYPE (to_arg),
+ TREE_TYPE (get_current_def
+ (from_arg))));
+ set_current_def (to_arg, get_current_def (from_arg));
+ }
}
gsi_next (&gsi_from);
gsi_next (&gsi_to);
basic blocks from SCALAR_LOOP instead of LOOP, but to either the
entry or exit of LOOP. */
-struct loop *
-slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *loop,
- struct loop *scalar_loop, edge e)
+class loop *
+slpeel_tree_duplicate_loop_to_edge_cfg (class loop *loop,
+ class loop *scalar_loop, edge e)
{
- struct loop *new_loop;
+ class loop *new_loop;
basic_block *new_bbs, *bbs, *pbbs;
bool at_exit;
bool was_imm_dom;
*/
bool
-slpeel_can_duplicate_loop_p (const struct loop *loop, const_edge e)
+slpeel_can_duplicate_loop_p (const class loop *loop, const_edge e)
{
edge exit_e = single_exit (loop);
edge entry_e = loop_preheader_edge (loop);
uses should be renamed. */
static void
-create_lcssa_for_virtual_phi (struct loop *loop)
+create_lcssa_for_virtual_phi (class loop *loop)
{
gphi_iterator gsi;
edge exit_e = single_exit (loop);
location is calculated.
Return the loop location if succeed and NULL if not. */
-source_location
-find_loop_location (struct loop *loop)
+dump_user_location_t
+find_loop_location (class loop *loop)
{
gimple *stmt = NULL;
basic_block bb;
gimple_stmt_iterator si;
if (!loop)
- return UNKNOWN_LOCATION;
+ return dump_user_location_t ();
stmt = get_loop_exit_condition (loop);
if (stmt
&& LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION)
- return gimple_location (stmt);
+ return stmt;
/* If we got here the loop is probably not "well formed",
try to estimate the loop location */
if (!loop->header)
- return UNKNOWN_LOCATION;
+ return dump_user_location_t ();
bb = loop->header;
{
stmt = gsi_stmt (si);
if (LOCATION_LOCUS (gimple_location (stmt)) > BUILTINS_LOCATION)
- return gimple_location (stmt);
+ return stmt;
}
- return UNKNOWN_LOCATION;
+ return dump_user_location_t ();
}
-/* Return true if PHI defines an IV of the loop to be vectorized. */
+/* Return true if the phi described by STMT_INFO defines an IV of the
+ loop to be vectorized. */
static bool
-iv_phi_p (gphi *phi)
+iv_phi_p (stmt_vec_info stmt_info)
{
+ gphi *phi = as_a <gphi *> (stmt_info->stmt);
if (virtual_operand_p (PHI_RESULT (phi)))
return false;
- stmt_vec_info stmt_info = vinfo_for_stmt (phi);
- gcc_assert (stmt_info != NULL);
if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def
|| STMT_VINFO_DEF_TYPE (stmt_info) == vect_double_reduction_def)
return false;
bool
vect_can_advance_ivs_p (loop_vec_info loop_vinfo)
{
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block bb = loop->header;
gphi_iterator gsi;
tree evolution_part;
gphi *phi = gsi.phi ();
+ stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi);
if (dump_enabled_p ())
- {
- dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: ");
- dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0);
- }
+ dump_printf_loc (MSG_NOTE, vect_location, "Analyze phi: %G",
+ phi_info->stmt);
/* Skip virtual phi's. The data dependences that are associated with
virtual defs/uses (i.e., memory accesses) are analyzed elsewhere.
Skip reduction phis. */
- if (!iv_phi_p (phi))
+ if (!iv_phi_p (phi_info))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
/* Analyze the evolution function. */
- evolution_part
- = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (vinfo_for_stmt (phi));
+ evolution_part = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info);
if (evolution_part == NULL_TREE)
{
if (dump_enabled_p ())
tree niters, edge update_e)
{
gphi_iterator gsi, gsi1;
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
basic_block update_bb = update_e->dest;
basic_block exit_bb = single_exit (loop)->dest;
gphi *phi = gsi.phi ();
gphi *phi1 = gsi1.phi ();
+ stmt_vec_info phi_info = loop_vinfo->lookup_stmt (phi);
if (dump_enabled_p ())
- {
- dump_printf_loc (MSG_NOTE, vect_location,
- "vect_update_ivs_after_vectorizer: phi: ");
- dump_gimple_stmt (MSG_NOTE, TDF_SLIM, phi, 0);
- }
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "vect_update_ivs_after_vectorizer: phi: %G", phi);
/* Skip reduction and virtual phis. */
- if (!iv_phi_p (phi))
+ if (!iv_phi_p (phi_info))
{
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
}
type = TREE_TYPE (gimple_phi_result (phi));
- step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (vinfo_for_stmt (phi));
+ step_expr = STMT_VINFO_LOOP_PHI_EVOLUTION_PART (phi_info);
step_expr = unshare_expr (step_expr);
/* FORNOW: We do not support IVs whose evolution function is a polynomial
}
}
+/* Return a gimple value containing the misalignment (measured in vector
+ elements) for the loop described by LOOP_VINFO, i.e. how many elements
+ it is away from a perfectly aligned address. Add any new statements
+ to SEQ. */
+
+static tree
+get_misalign_in_elems (gimple **seq, loop_vec_info loop_vinfo)
+{
+ dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+ stmt_vec_info stmt_info = dr_info->stmt;
+ tree vectype = STMT_VINFO_VECTYPE (stmt_info);
+
+ poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
+ unsigned HOST_WIDE_INT target_align_c;
+ tree target_align_minus_1;
+
+ bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
+ size_zero_node) < 0;
+ tree offset = (negative
+ ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1)
+ : size_zero_node);
+ tree start_addr = vect_create_addr_base_for_vector_ref (stmt_info, seq,
+ offset);
+ tree type = unsigned_type_for (TREE_TYPE (start_addr));
+ if (target_align.is_constant (&target_align_c))
+ target_align_minus_1 = build_int_cst (type, target_align_c - 1);
+ else
+ {
+ tree vla = build_int_cst (type, target_align);
+ tree vla_align = fold_build2 (BIT_AND_EXPR, type, vla,
+ fold_build2 (MINUS_EXPR, type,
+ build_int_cst (type, 0), vla));
+ target_align_minus_1 = fold_build2 (MINUS_EXPR, type, vla_align,
+ build_int_cst (type, 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));
+
+ /* Create: misalign_in_bytes = addr & (target_align - 1). */
+ tree int_start_addr = fold_convert (type, start_addr);
+ tree misalign_in_bytes = fold_build2 (BIT_AND_EXPR, type, int_start_addr,
+ target_align_minus_1);
+
+ /* Create: misalign_in_elems = misalign_in_bytes / element_size. */
+ tree misalign_in_elems = fold_build2 (RSHIFT_EXPR, type, misalign_in_bytes,
+ elem_size_log);
+
+ return misalign_in_elems;
+}
+
/* Function vect_gen_prolog_loop_niters
Generate the number of iterations which should be peeled as prolog for the
If the misalignment of DR is known at compile time:
addr_mis = int mis = DR_MISALIGNMENT (dr);
Else, compute address misalignment in bytes:
- addr_mis = addr & (vectype_align - 1)
+ addr_mis = addr & (target_align - 1)
prolog_niters = ((VF - addr_mis/elem_size)&(VF-1))/step
vect_gen_prolog_loop_niters (loop_vec_info loop_vinfo,
basic_block bb, int *bound)
{
- struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+ dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
tree var;
tree niters_type = TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo));
gimple_seq stmts = NULL, new_stmts = NULL;
tree iters, iters_name;
- gimple *dr_stmt = DR_STMT (dr);
- stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt);
+ stmt_vec_info stmt_info = dr_info->stmt;
tree vectype = STMT_VINFO_VECTYPE (stmt_info);
- unsigned int target_align = DR_TARGET_ALIGNMENT (dr);
+ poly_uint64 target_align = DR_TARGET_ALIGNMENT (dr_info);
if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
{
}
else
{
- bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
- tree offset = negative
- ? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node;
- tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt,
- &stmts, offset);
- tree type = unsigned_type_for (TREE_TYPE (start_addr));
- tree target_align_minus_1 = build_int_cst (type, target_align - 1);
+ tree misalign_in_elems = get_misalign_in_elems (&stmts, loop_vinfo);
+ tree type = TREE_TYPE (misalign_in_elems);
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));
- HOST_WIDE_INT align_in_elems = target_align / elem_size;
- tree align_in_elems_minus_1 = build_int_cst (type, align_in_elems - 1);
+ /* We only do prolog peeling if the target alignment is known at compile
+ time. */
+ poly_uint64 align_in_elems =
+ exact_div (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)). */
+ bool negative = tree_int_cst_compare (DR_STEP (dr_info->dr),
+ size_zero_node) < 0;
if (negative)
iters = fold_build2 (MINUS_EXPR, type, misalign_in_elems,
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 = align_in_elems - 1;
+ unsigned HOST_WIDE_INT align_in_elems_c;
+ if (align_in_elems.is_constant (&align_in_elems_c))
+ *bound = align_in_elems_c - 1;
+ else
+ *bound = -1;
}
if (dump_enabled_p ())
- {
- dump_printf_loc (MSG_NOTE, vect_location,
- "niters for prolog loop: ");
- dump_generic_expr (MSG_NOTE, TDF_SLIM, iters);
- dump_printf (MSG_NOTE, "\n");
- }
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "niters for prolog loop: %T\n", iters);
var = create_tmp_var (niters_type, "prolog_loop_niters");
iters_name = force_gimple_operand (iters, &new_stmts, false, var);
/* Function vect_update_init_of_dr
- NITERS iterations were peeled from LOOP. DR represents a data reference
- in LOOP. This function updates the information recorded in DR to
- account for the fact that the first NITERS iterations had already been
- executed. Specifically, it updates the OFFSET field of DR. */
+ If CODE is PLUS, the vector loop starts NITERS iterations after the
+ scalar one, otherwise CODE is MINUS and the vector loop starts NITERS
+ iterations before the scalar one (using masking to skip inactive
+ elements). This function updates the information recorded in DR to
+ account for the difference. Specifically, it updates the OFFSET
+ field of DR. */
static void
-vect_update_init_of_dr (struct data_reference *dr, tree niters)
+vect_update_init_of_dr (struct data_reference *dr, tree niters, tree_code code)
{
tree offset = DR_OFFSET (dr);
niters = fold_build2 (MULT_EXPR, sizetype,
fold_convert (sizetype, niters),
fold_convert (sizetype, DR_STEP (dr)));
- offset = fold_build2 (PLUS_EXPR, sizetype,
+ offset = fold_build2 (code, sizetype,
fold_convert (sizetype, offset), niters);
DR_OFFSET (dr) = offset;
}
/* Function vect_update_inits_of_drs
- NITERS iterations were peeled from the loop represented by LOOP_VINFO.
- This function updates the information recorded for the data references in
- the loop to account for the fact that the first NITERS iterations had
- already been executed. Specifically, it updates the initial_condition of
- the access_function of all the data_references in the loop. */
+ Apply vect_update_inits_of_dr to all accesses in LOOP_VINFO.
+ CODE and NITERS are as for vect_update_inits_of_dr. */
static void
-vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters)
+vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters,
+ tree_code code)
{
unsigned int i;
vec<data_reference_p> datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
struct data_reference *dr;
- if (dump_enabled_p ())
- dump_printf_loc (MSG_NOTE, vect_location,
- "=== vect_update_inits_of_dr ===\n");
+ DUMP_VECT_SCOPE ("vect_update_inits_of_dr");
/* Adjust niters to sizetype and insert stmts on loop preheader edge. */
if (!types_compatible_p (sizetype, TREE_TYPE (niters)))
}
FOR_EACH_VEC_ELT (datarefs, i, dr)
- vect_update_init_of_dr (dr, niters);
+ {
+ dr_vec_info *dr_info = loop_vinfo->lookup_dr (dr);
+ if (!STMT_VINFO_GATHER_SCATTER_P (dr_info->stmt))
+ vect_update_init_of_dr (dr, niters, code);
+ }
}
+/* For the information recorded in LOOP_VINFO prepare the loop for peeling
+ by masking. This involves calculating the number of iterations to
+ be peeled and then aligning all memory references appropriately. */
+
+void
+vect_prepare_for_masked_peels (loop_vec_info loop_vinfo)
+{
+ tree misalign_in_elems;
+ tree type = LOOP_VINFO_MASK_COMPARE_TYPE (loop_vinfo);
+
+ gcc_assert (vect_use_loop_mask_for_alignment_p (loop_vinfo));
+
+ /* From the information recorded in LOOP_VINFO get the number of iterations
+ that need to be skipped via masking. */
+ if (LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
+ {
+ poly_int64 misalign = (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
+ - LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo));
+ misalign_in_elems = build_int_cst (type, misalign);
+ }
+ else
+ {
+ gimple_seq seq1 = NULL, seq2 = NULL;
+ misalign_in_elems = get_misalign_in_elems (&seq1, loop_vinfo);
+ misalign_in_elems = fold_convert (type, misalign_in_elems);
+ misalign_in_elems = force_gimple_operand (misalign_in_elems,
+ &seq2, true, NULL_TREE);
+ gimple_seq_add_seq (&seq1, seq2);
+ if (seq1)
+ {
+ edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo));
+ basic_block new_bb = gsi_insert_seq_on_edge_immediate (pe, seq1);
+ gcc_assert (!new_bb);
+ }
+ }
+
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "misalignment for fully-masked loop: %T\n",
+ misalign_in_elems);
+
+ LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo) = misalign_in_elems;
+
+ vect_update_inits_of_drs (loop_vinfo, misalign_in_elems, MINUS_EXPR);
+}
/* This function builds ni_name = number of iterations. Statements
are emitted on the loop preheader edge. If NEW_VAR_P is not NULL, set
/* Calculate the number of iterations above which vectorized loop will be
preferred than scalar loop. NITERS_PROLOG is the number of iterations
of prolog loop. If it's integer const, the integer number is also passed
- in INT_NITERS_PROLOG. BOUND_PROLOG is the upper bound (included) of
- number of iterations of prolog loop. VFM1 is vector factor minus one.
- If CHECK_PROFITABILITY is true, TH is the threshold below which scalar
- (rather than vectorized) loop will be executed. This function stores
- upper bound (included) of the result in BOUND_SCALAR. */
+ in INT_NITERS_PROLOG. BOUND_PROLOG is the upper bound (inclusive) of the
+ number of iterations of the prolog loop. BOUND_EPILOG is the corresponding
+ value for the epilog loop. If CHECK_PROFITABILITY is true, TH is the
+ threshold below which the scalar (rather than vectorized) loop will be
+ executed. This function stores the upper bound (inclusive) of the result
+ in BOUND_SCALAR. */
static tree
vect_gen_scalar_loop_niters (tree niters_prolog, int int_niters_prolog,
- int bound_prolog, int vfm1, int th,
- int *bound_scalar, bool check_profitability)
+ int bound_prolog, poly_int64 bound_epilog, int th,
+ poly_uint64 *bound_scalar,
+ bool check_profitability)
{
tree type = TREE_TYPE (niters_prolog);
tree niters = fold_build2 (PLUS_EXPR, type, niters_prolog,
- build_int_cst (type, vfm1));
+ build_int_cst (type, bound_epilog));
- *bound_scalar = vfm1 + bound_prolog;
+ *bound_scalar = bound_prolog + bound_epilog;
if (check_profitability)
{
/* TH indicates the minimum niters of vectorized loop, while we
/* Peeling for constant times. */
if (int_niters_prolog >= 0)
{
- *bound_scalar = (int_niters_prolog + vfm1 < th
- ? th
- : vfm1 + int_niters_prolog);
+ *bound_scalar = upper_bound (int_niters_prolog + bound_epilog, th);
return build_int_cst (type, *bound_scalar);
}
- /* Peeling for unknown times. Note BOUND_PROLOG is the upper
- bound (inlcuded) of niters of prolog loop. */
- if (th >= vfm1 + bound_prolog)
+ /* Peeling an unknown number of times. Note that both BOUND_PROLOG
+ and BOUND_EPILOG are inclusive upper bounds. */
+ if (known_ge (th, bound_prolog + bound_epilog))
{
*bound_scalar = th;
return build_int_cst (type, th);
}
- /* Need to do runtime comparison, but BOUND_SCALAR remains the same. */
- else if (th > vfm1)
- return fold_build2 (MAX_EXPR, type, build_int_cst (type, th), niters);
+ /* Need to do runtime comparison. */
+ else if (maybe_gt (th, bound_epilog))
+ {
+ *bound_scalar = upper_bound (*bound_scalar, th);
+ return fold_build2 (MAX_EXPR, type,
+ build_int_cst (type, th), niters);
+ }
}
return niters;
}
-/* This function generates the following statements:
+/* NITERS is the number of times that the original scalar loop executes
+ after peeling. Work out the maximum number of iterations N that can
+ be handled by the vectorized form of the loop and then either:
+
+ a) set *STEP_VECTOR_PTR to the vectorization factor and generate:
+
+ niters_vector = N
- niters = number of iterations loop executes (after peeling)
- niters_vector = niters / vf
+ b) set *STEP_VECTOR_PTR to one and generate:
- and places them on the loop preheader edge. NITERS_NO_OVERFLOW is
- true if NITERS doesn't overflow. */
+ niters_vector = N / vf
+
+ In both cases, store niters_vector in *NITERS_VECTOR_PTR and add
+ any new statements on the loop preheader edge. NITERS_NO_OVERFLOW
+ is true if NITERS doesn't overflow (i.e. if NITERS is always nonzero). */
void
vect_gen_vector_loop_niters (loop_vec_info loop_vinfo, tree niters,
- tree *niters_vector_ptr, bool niters_no_overflow)
+ tree *niters_vector_ptr, tree *step_vector_ptr,
+ bool niters_no_overflow)
{
tree ni_minus_gap, var;
- tree niters_vector, type = TREE_TYPE (niters);
- int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ tree niters_vector, step_vector, type = TREE_TYPE (niters);
+ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
edge pe = loop_preheader_edge (LOOP_VINFO_LOOP (loop_vinfo));
- tree log_vf = build_int_cst (type, exact_log2 (vf));
+ tree log_vf = NULL_TREE;
/* If epilogue loop is required because of data accesses with gaps, we
subtract one iteration from the total number of iterations here for
else
ni_minus_gap = niters;
- /* Create: niters >> log2(vf) */
- /* If it's known that niters == number of latch executions + 1 doesn't
- overflow, we can generate niters >> log2(vf); otherwise we generate
- (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio
- will be at least one. */
- if (niters_no_overflow)
- niters_vector = fold_build2 (RSHIFT_EXPR, type, ni_minus_gap, log_vf);
+ unsigned HOST_WIDE_INT const_vf;
+ if (vf.is_constant (&const_vf)
+ && !LOOP_VINFO_FULLY_MASKED_P (loop_vinfo))
+ {
+ /* Create: niters >> log2(vf) */
+ /* If it's known that niters == number of latch executions + 1 doesn't
+ overflow, we can generate niters >> log2(vf); otherwise we generate
+ (niters - vf) >> log2(vf) + 1 by using the fact that we know ratio
+ will be at least one. */
+ log_vf = build_int_cst (type, exact_log2 (const_vf));
+ if (niters_no_overflow)
+ niters_vector = fold_build2 (RSHIFT_EXPR, type, ni_minus_gap, log_vf);
+ else
+ niters_vector
+ = fold_build2 (PLUS_EXPR, type,
+ fold_build2 (RSHIFT_EXPR, type,
+ fold_build2 (MINUS_EXPR, type,
+ ni_minus_gap,
+ build_int_cst (type, vf)),
+ log_vf),
+ build_int_cst (type, 1));
+ step_vector = build_one_cst (type);
+ }
else
- niters_vector
- = fold_build2 (PLUS_EXPR, type,
- fold_build2 (RSHIFT_EXPR, type,
- fold_build2 (MINUS_EXPR, type, ni_minus_gap,
- build_int_cst (type, vf)),
- log_vf),
- build_int_cst (type, 1));
+ {
+ niters_vector = ni_minus_gap;
+ step_vector = build_int_cst (type, vf);
+ }
if (!is_gimple_val (niters_vector))
{
gsi_insert_seq_on_edge_immediate (pe, stmts);
/* Peeling algorithm guarantees that vector loop bound is at least ONE,
we set range information to make niters analyzer's life easier. */
- if (stmts != NULL)
+ if (stmts != NULL && log_vf)
set_range_info (niters_vector, VR_RANGE,
wi::to_wide (build_int_cst (type, 1)),
wi::to_wide (fold_build2 (RSHIFT_EXPR, type,
log_vf)));
}
*niters_vector_ptr = niters_vector;
+ *step_vector_ptr = step_vector;
return;
}
tree niters_vector,
tree *niters_vector_mult_vf_ptr)
{
- int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ /* We should be using a step_vector of VF if VF is variable. */
+ int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo).to_constant ();
+ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
tree type = TREE_TYPE (niters_vector);
tree log_vf = build_int_cst (type, exact_log2 (vf));
basic_block exit_bb = single_exit (loop)->dest;
static void
slpeel_update_phi_nodes_for_loops (loop_vec_info loop_vinfo,
- struct loop *first, struct loop *second,
+ class loop *first, class loop *second,
bool create_lcssa_for_iv_phis)
{
gphi_iterator gsi_update, gsi_orig;
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
edge first_latch_e = EDGE_SUCC (first->latch, 0);
edge second_preheader_e = loop_preheader_edge (second);
tree arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, first_latch_e);
/* Generate lcssa PHI node for the first loop. */
gphi *vect_phi = (loop == first) ? orig_phi : update_phi;
- if (create_lcssa_for_iv_phis || !iv_phi_p (vect_phi))
+ stmt_vec_info vect_phi_info = loop_vinfo->lookup_stmt (vect_phi);
+ if (create_lcssa_for_iv_phis || !iv_phi_p (vect_phi_info))
{
tree new_res = copy_ssa_name (PHI_RESULT (orig_phi));
gphi *lcssa_phi = create_phi_node (new_res, between_bb);
in the update_loop's PHI node with the result of new PHI result. */
static void
-slpeel_update_phi_nodes_for_guard1 (struct loop *skip_loop,
- struct loop *update_loop,
+slpeel_update_phi_nodes_for_guard1 (class loop *skip_loop,
+ class loop *update_loop,
edge guard_edge, edge merge_edge)
{
- source_location merge_loc, guard_loc;
+ location_t merge_loc, guard_loc;
edge orig_e = loop_preheader_edge (skip_loop);
edge update_e = loop_preheader_edge (update_loop);
gphi_iterator gsi_orig, gsi_update;
NULL. */
static tree
-find_guard_arg (struct loop *loop, struct loop *epilog ATTRIBUTE_UNUSED,
+find_guard_arg (class loop *loop, class loop *epilog ATTRIBUTE_UNUSED,
gphi *lcssa_phi)
{
gphi_iterator gsi;
in exit_bb will also be updated. */
static void
-slpeel_update_phi_nodes_for_guard2 (struct loop *loop, struct loop *epilog,
+slpeel_update_phi_nodes_for_guard2 (class loop *loop, class loop *epilog,
edge guard_edge, edge merge_edge)
{
gphi_iterator gsi;
the arg of its loop closed ssa PHI needs to be updated. */
static void
-slpeel_update_phi_nodes_for_lcssa (struct loop *epilog)
+slpeel_update_phi_nodes_for_lcssa (class loop *epilog)
{
gphi_iterator gsi;
basic_block exit_bb = single_exit (epilog)->dest;
- TH, CHECK_PROFITABILITY: Threshold of niters to vectorize loop if
CHECK_PROFITABILITY is true.
Output:
- - NITERS_VECTOR: The number of iterations of loop after vectorization.
+ - *NITERS_VECTOR and *STEP_VECTOR describe how the main loop should
+ iterate after vectorization; see vect_set_loop_condition for details.
+ - *NITERS_VECTOR_MULT_VF_VAR is either null or an SSA name that
+ should be set to the number of scalar iterations handled by the
+ vector loop. The SSA name is only used on exit from the loop.
This function peels prolog and epilog from the loop, adds guards skipping
PROLOG and EPILOG for various conditions. As a result, the changed CFG
versioning conditions if loop versioning is needed. */
-struct loop *
+class loop *
vect_do_peeling (loop_vec_info loop_vinfo, tree niters, tree nitersm1,
- tree *niters_vector, int th, bool check_profitability,
- bool niters_no_overflow)
+ tree *niters_vector, tree *step_vector,
+ tree *niters_vector_mult_vf_var, int th,
+ bool check_profitability, bool niters_no_overflow)
{
edge e, guard_e;
tree type = TREE_TYPE (niters), guard_cond;
basic_block guard_bb, guard_to;
profile_probability prob_prolog, prob_vector, prob_epilog;
- int bound_prolog = 0, bound_scalar = 0, bound = 0;
- int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
- int prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
- bool epilog_peeling = (LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo)
- || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo));
+ int estimated_vf;
+ int prolog_peeling = 0;
+ /* We currently do not support prolog peeling if the target alignment is not
+ known at compile time. 'vect_gen_prolog_loop_niters' depends on the
+ target alignment being constant. */
+ dr_vec_info *dr_info = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
+ if (dr_info && !DR_TARGET_ALIGNMENT (dr_info).is_constant ())
+ return NULL;
+
+ if (!vect_use_loop_mask_for_alignment_p (loop_vinfo))
+ prolog_peeling = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
+
+ poly_uint64 vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ poly_uint64 bound_epilog = 0;
+ if (!LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
+ && LOOP_VINFO_PEELING_FOR_NITER (loop_vinfo))
+ bound_epilog += vf - 1;
+ if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
+ bound_epilog += 1;
+ bool epilog_peeling = maybe_ne (bound_epilog, 0U);
+ poly_uint64 bound_scalar = bound_epilog;
if (!prolog_peeling && !epilog_peeling)
return NULL;
prob_vector = profile_probability::guessed_always ().apply_scale (9, 10);
- if ((vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo)) == 2)
- vf = 3;
+ estimated_vf = vect_vf_for_cost (loop_vinfo);
+ if (estimated_vf == 2)
+ estimated_vf = 3;
prob_prolog = prob_epilog = profile_probability::guessed_always ()
- .apply_scale (vf - 1, vf);
- vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
+ .apply_scale (estimated_vf - 1, estimated_vf);
- struct loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo);
- struct loop *first_loop = loop;
+ class loop *prolog, *epilog = NULL, *loop = LOOP_VINFO_LOOP (loop_vinfo);
+ class loop *first_loop = loop;
bool irred_flag = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
create_lcssa_for_virtual_phi (loop);
update_ssa (TODO_update_ssa_only_virtuals);
}
initialize_original_copy_tables ();
+ /* Record the anchor bb at which the guard should be placed if the scalar
+ loop might be preferred. */
+ basic_block anchor = loop_preheader_edge (loop)->src;
+
+ /* Generate the number of iterations for the prolog loop. We do this here
+ so that we can also get the upper bound on the number of iterations. */
+ tree niters_prolog;
+ int bound_prolog = 0;
+ if (prolog_peeling)
+ niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor,
+ &bound_prolog);
+ else
+ niters_prolog = build_int_cst (type, 0);
+
/* Prolog loop may be skipped. */
bool skip_prolog = (prolog_peeling != 0);
/* Skip to epilog if scalar loop may be preferred. It's only needed
when we peel for epilog loop and when it hasn't been checked with
loop versioning. */
- bool skip_vector = (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
- && !LOOP_REQUIRES_VERSIONING (loop_vinfo));
+ bool skip_vector = (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+ ? maybe_lt (LOOP_VINFO_INT_NITERS (loop_vinfo),
+ bound_prolog + bound_epilog)
+ : !LOOP_REQUIRES_VERSIONING (loop_vinfo));
/* Epilog loop must be executed if the number of iterations for epilog
loop is known at compile time, otherwise we need to add a check at
the end of vector loop and skip to the end of epilog loop. */
bool skip_epilog = (prolog_peeling < 0
- || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo));
+ || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
+ || !vf.is_constant ());
/* PEELING_FOR_GAPS is special because epilog loop must be executed. */
if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo))
skip_epilog = false;
- /* Record the anchor bb at which guard should be placed if scalar loop
- may be preferred. */
- basic_block anchor = loop_preheader_edge (loop)->src;
if (skip_vector)
{
split_edge (loop_preheader_edge (loop));
needs to be scaled back later. */
basic_block bb_before_loop = loop_preheader_edge (loop)->src;
if (prob_vector.initialized_p ())
- scale_bbs_frequencies (&bb_before_loop, 1, prob_vector);
- scale_loop_profile (loop, prob_vector, bound);
+ {
+ scale_bbs_frequencies (&bb_before_loop, 1, prob_vector);
+ scale_loop_profile (loop, prob_vector, 0);
+ }
}
- tree niters_prolog = build_int_cst (type, 0);
- source_location loop_loc = find_loop_location (loop);
- struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
+ dump_user_location_t loop_loc = find_loop_location (loop);
+ class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
if (prolog_peeling)
{
e = loop_preheader_edge (loop);
"slpeel_tree_duplicate_loop_to_edge_cfg failed.\n");
gcc_unreachable ();
}
+ prolog->force_vectorize = false;
slpeel_update_phi_nodes_for_loops (loop_vinfo, prolog, loop, true);
first_loop = prolog;
reset_original_copy_tables ();
- /* Generate and update the number of iterations for prolog loop. */
- niters_prolog = vect_gen_prolog_loop_niters (loop_vinfo, anchor,
- &bound_prolog);
- slpeel_make_loop_iterate_ntimes (prolog, niters_prolog);
+ /* Update the number of iterations for prolog loop. */
+ tree step_prolog = build_one_cst (TREE_TYPE (niters_prolog));
+ vect_set_loop_condition (prolog, NULL, niters_prolog,
+ step_prolog, NULL_TREE, false);
/* Skip the prolog loop. */
if (skip_prolog)
scale_loop_profile (prolog, prob_prolog, bound_prolog);
}
/* Update init address of DRs. */
- vect_update_inits_of_drs (loop_vinfo, niters_prolog);
+ vect_update_inits_of_drs (loop_vinfo, niters_prolog, PLUS_EXPR);
/* Update niters for vector loop. */
LOOP_VINFO_NITERS (loop_vinfo)
= fold_build2 (MINUS_EXPR, type, niters, niters_prolog);
"slpeel_tree_duplicate_loop_to_edge_cfg failed.\n");
gcc_unreachable ();
}
+ epilog->force_vectorize = false;
slpeel_update_phi_nodes_for_loops (loop_vinfo, loop, epilog, false);
/* Scalar version loop may be preferred. In this case, add guard
if (skip_vector)
{
/* Additional epilogue iteration is peeled if gap exists. */
- bool peel_for_gaps = LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo);
tree t = vect_gen_scalar_loop_niters (niters_prolog, prolog_peeling,
- bound_prolog,
- peel_for_gaps ? vf : vf - 1,
+ bound_prolog, bound_epilog,
th, &bound_scalar,
check_profitability);
/* Build guard against NITERSM1 since NITERS may overflow. */
overflows. */
niters_no_overflow |= (prolog_peeling > 0);
vect_gen_vector_loop_niters (loop_vinfo, niters,
- niters_vector, niters_no_overflow);
- vect_gen_vector_loop_niters_mult_vf (loop_vinfo, *niters_vector,
- &niters_vector_mult_vf);
+ niters_vector, step_vector,
+ niters_no_overflow);
+ if (!integer_onep (*step_vector))
+ {
+ /* On exit from the loop we will have an easy way of calcalating
+ NITERS_VECTOR / STEP * STEP. Install a dummy definition
+ until then. */
+ niters_vector_mult_vf = make_ssa_name (TREE_TYPE (*niters_vector));
+ SSA_NAME_DEF_STMT (niters_vector_mult_vf) = gimple_build_nop ();
+ *niters_vector_mult_vf_var = niters_vector_mult_vf;
+ }
+ else
+ vect_gen_vector_loop_niters_mult_vf (loop_vinfo, *niters_vector,
+ &niters_vector_mult_vf);
/* Update IVs of original loop as if they were advanced by
niters_vector_mult_vf steps. */
gcc_checking_assert (vect_can_advance_ivs_p (loop_vinfo));
scale_bbs_frequencies (&bb_before_epilog, 1, prob_epilog);
}
- scale_loop_profile (epilog, prob_epilog, bound);
+ scale_loop_profile (epilog, prob_epilog, 0);
}
else
slpeel_update_phi_nodes_for_lcssa (epilog);
- bound = LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) ? vf - 1 : vf - 2;
- /* We share epilog loop with scalar version loop. */
- bound = MAX (bound, bound_scalar - 1);
- record_niter_bound (epilog, bound, false, true);
+ unsigned HOST_WIDE_INT bound;
+ if (bound_scalar.is_constant (&bound))
+ {
+ gcc_assert (bound != 0);
+ /* -1 to convert loop iterations to latch iterations. */
+ record_niter_bound (epilog, bound - 1, false, true);
+ }
delete_update_ssa ();
adjust_vec_debug_stmts ();
/* Function vect_create_cond_for_niters_checks.
Create a conditional expression that represents the run-time checks for
- loop's niter. The loop is guaranteed to to terminate if the run-time
+ loop's niter. The loop is guaranteed to terminate if the run-time
checks hold.
Input:
tree *cond_expr,
gimple_seq *cond_expr_stmt_list)
{
- vec<gimple *> may_misalign_stmts
+ vec<stmt_vec_info> may_misalign_stmts
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
- gimple *ref_stmt;
+ stmt_vec_info stmt_info;
int mask = LOOP_VINFO_PTR_MASK (loop_vinfo);
tree mask_cst;
unsigned int i;
/* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
of the first vector of the i'th data reference. */
- FOR_EACH_VEC_ELT (may_misalign_stmts, i, ref_stmt)
+ FOR_EACH_VEC_ELT (may_misalign_stmts, i, stmt_info)
{
gimple_seq new_stmt_list = NULL;
tree addr_base;
tree addr_tmp_name;
tree new_or_tmp_name;
gimple *addr_stmt, *or_stmt;
- stmt_vec_info stmt_vinfo = vinfo_for_stmt (ref_stmt);
- tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
+ tree vectype = STMT_VINFO_VECTYPE (stmt_info);
bool negative = tree_int_cst_compare
- (DR_STEP (STMT_VINFO_DATA_REF (stmt_vinfo)), size_zero_node) < 0;
+ (DR_STEP (STMT_VINFO_DATA_REF (stmt_info)), size_zero_node) < 0;
tree offset = negative
? size_int (-TYPE_VECTOR_SUBPARTS (vectype) + 1) : size_zero_node;
/* create: addr_tmp = (int)(address_of_first_vector) */
addr_base =
- vect_create_addr_base_for_vector_ref (ref_stmt, &new_stmt_list,
+ vect_create_addr_base_for_vector_ref (stmt_info, &new_stmt_list,
offset);
if (new_stmt_list != NULL)
gimple_seq_add_seq (cond_expr_stmt_list, new_stmt_list);
}
}
+/* Create an expression that is true when all lower-bound conditions for
+ the vectorized loop are met. Chain this condition with *COND_EXPR. */
+
+static void
+vect_create_cond_for_lower_bounds (loop_vec_info loop_vinfo, tree *cond_expr)
+{
+ vec<vec_lower_bound> lower_bounds = LOOP_VINFO_LOWER_BOUNDS (loop_vinfo);
+ for (unsigned int i = 0; i < lower_bounds.length (); ++i)
+ {
+ tree expr = lower_bounds[i].expr;
+ tree type = unsigned_type_for (TREE_TYPE (expr));
+ expr = fold_convert (type, expr);
+ poly_uint64 bound = lower_bounds[i].min_value;
+ if (!lower_bounds[i].unsigned_p)
+ {
+ expr = fold_build2 (PLUS_EXPR, type, expr,
+ build_int_cstu (type, bound - 1));
+ bound += bound - 1;
+ }
+ tree part_cond_expr = fold_build2 (GE_EXPR, boolean_type_node, expr,
+ build_int_cstu (type, bound));
+ chain_cond_expr (cond_expr, part_cond_expr);
+ }
+}
+
/* Function vect_create_cond_for_alias_checks.
Create a conditional expression that represents the run-time checks for
The versioning precondition(s) are placed in *COND_EXPR and
*COND_EXPR_STMT_LIST. */
-void
+class loop *
vect_loop_versioning (loop_vec_info loop_vinfo,
unsigned int th, bool check_profitability,
poly_uint64 versioning_threshold)
{
- struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop;
- struct loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
+ class loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *nloop;
+ class loop *scalar_loop = LOOP_VINFO_SCALAR_LOOP (loop_vinfo);
basic_block condition_bb;
gphi_iterator gsi;
gimple_stmt_iterator cond_exp_gsi;
bool version_align = LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo);
bool version_alias = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo);
bool version_niter = LOOP_REQUIRES_VERSIONING_FOR_NITERS (loop_vinfo);
+ tree version_simd_if_cond
+ = LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (loop_vinfo);
if (check_profitability)
cond_expr = fold_build2 (GE_EXPR, boolean_type_node, scalar_loop_iters,
vect_create_cond_for_niters_checks (loop_vinfo, &cond_expr);
if (cond_expr)
- cond_expr = force_gimple_operand_1 (cond_expr, &cond_expr_stmt_list,
+ cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr),
+ &cond_expr_stmt_list,
is_gimple_condexpr, NULL_TREE);
if (version_align)
if (version_alias)
{
vect_create_cond_for_unequal_addrs (loop_vinfo, &cond_expr);
+ vect_create_cond_for_lower_bounds (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,
+ if (version_simd_if_cond)
+ {
+ gcc_assert (dom_info_available_p (CDI_DOMINATORS));
+ if (flag_checking)
+ if (basic_block bb
+ = gimple_bb (SSA_NAME_DEF_STMT (version_simd_if_cond)))
+ gcc_assert (bb != loop->header
+ && dominated_by_p (CDI_DOMINATORS, loop->header, bb)
+ && (scalar_loop == NULL
+ || (bb != scalar_loop->header
+ && dominated_by_p (CDI_DOMINATORS,
+ scalar_loop->header, bb))));
+ tree zero = build_zero_cst (TREE_TYPE (version_simd_if_cond));
+ tree c = fold_build2 (NE_EXPR, boolean_type_node,
+ version_simd_if_cond, zero);
+ if (cond_expr)
+ cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
+ c, cond_expr);
+ else
+ cond_expr = c;
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "created versioning for simd if condition check.\n");
+ }
+
+ cond_expr = force_gimple_operand_1 (unshare_expr (cond_expr),
+ &gimplify_stmt_list,
is_gimple_condexpr, NULL_TREE);
gimple_seq_add_seq (&cond_expr_stmt_list, gimplify_stmt_list);
- initialize_original_copy_tables ();
- if (scalar_loop)
+ /* Compute the outermost loop cond_expr and cond_expr_stmt_list are
+ invariant in. */
+ class loop *outermost = outermost_invariant_loop_for_expr (loop, cond_expr);
+ for (gimple_stmt_iterator gsi = gsi_start (cond_expr_stmt_list);
+ !gsi_end_p (gsi); gsi_next (&gsi))
{
- edge scalar_e;
- basic_block preheader, scalar_preheader;
+ gimple *stmt = gsi_stmt (gsi);
+ update_stmt (stmt);
+ ssa_op_iter iter;
+ use_operand_p use_p;
+ basic_block def_bb;
+ FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
+ if ((def_bb = gimple_bb (SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p))))
+ && flow_bb_inside_loop_p (outermost, def_bb))
+ outermost = superloop_at_depth (loop, bb_loop_depth (def_bb) + 1);
+ }
- /* We don't want to scale SCALAR_LOOP's frequencies, we need to
- scale LOOP's frequencies instead. */
- nloop = loop_version (scalar_loop, cond_expr, &condition_bb,
- prob, prob.invert (), prob, prob.invert (), true);
- scale_loop_frequencies (loop, prob);
- /* CONDITION_BB was created above SCALAR_LOOP's preheader,
- while we need to move it above LOOP's preheader. */
- e = loop_preheader_edge (loop);
- scalar_e = loop_preheader_edge (scalar_loop);
- gcc_assert (empty_block_p (e->src)
- && single_pred_p (e->src));
- gcc_assert (empty_block_p (scalar_e->src)
- && single_pred_p (scalar_e->src));
- gcc_assert (single_pred_p (condition_bb));
- preheader = e->src;
- scalar_preheader = scalar_e->src;
- scalar_e = find_edge (condition_bb, scalar_preheader);
- e = single_pred_edge (preheader);
- redirect_edge_and_branch_force (single_pred_edge (condition_bb),
- scalar_preheader);
- redirect_edge_and_branch_force (scalar_e, preheader);
- redirect_edge_and_branch_force (e, condition_bb);
- set_immediate_dominator (CDI_DOMINATORS, condition_bb,
- single_pred (condition_bb));
- set_immediate_dominator (CDI_DOMINATORS, scalar_preheader,
- single_pred (scalar_preheader));
- set_immediate_dominator (CDI_DOMINATORS, preheader,
- condition_bb);
+ /* Search for the outermost loop we can version. Avoid versioning of
+ non-perfect nests but allow if-conversion versioned loops inside. */
+ class loop *loop_to_version = loop;
+ if (flow_loop_nested_p (outermost, loop))
+ {
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "trying to apply versioning to outer loop %d\n",
+ outermost->num);
+ if (outermost->num == 0)
+ outermost = superloop_at_depth (loop, 1);
+ /* And avoid applying versioning on non-perfect nests. */
+ while (loop_to_version != outermost
+ && single_exit (loop_outer (loop_to_version))
+ && (!loop_outer (loop_to_version)->inner->next
+ || vect_loop_vectorized_call (loop_to_version))
+ && (!loop_outer (loop_to_version)->inner->next
+ || !loop_outer (loop_to_version)->inner->next->next))
+ loop_to_version = loop_outer (loop_to_version);
+ }
+
+ /* Apply versioning. If there is already a scalar version created by
+ if-conversion re-use that. Note we cannot re-use the copy of
+ an if-converted outer-loop when vectorizing the inner loop only. */
+ gcond *cond;
+ gimple *call;
+ if ((!loop_to_version->inner || loop == loop_to_version)
+ && (call = vect_loop_vectorized_call (loop_to_version, &cond)))
+ {
+ gcc_assert (scalar_loop);
+ condition_bb = gimple_bb (cond);
+ gimple_cond_set_condition_from_tree (cond, cond_expr);
+ update_stmt (cond);
+
+ if (cond_expr_stmt_list)
+ {
+ cond_exp_gsi = gsi_for_stmt (call);
+ gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
+ GSI_SAME_STMT);
+ }
+
+ /* if-conversion uses profile_probability::always () for both paths,
+ reset the paths probabilities appropriately. */
+ edge te, fe;
+ extract_true_false_edges_from_block (condition_bb, &te, &fe);
+ te->probability = prob;
+ fe->probability = prob.invert ();
+ /* We can scale loops counts immediately but have to postpone
+ scaling the scalar loop because we re-use it during peeling. */
+ scale_loop_frequencies (loop_to_version, te->probability);
+ LOOP_VINFO_SCALAR_LOOP_SCALING (loop_vinfo) = fe->probability;
+
+ nloop = scalar_loop;
+ if (dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "reusing %sloop version created by if conversion\n",
+ loop_to_version != loop ? "outer " : "");
}
else
- nloop = loop_version (loop, cond_expr, &condition_bb,
- prob, prob.invert (), prob, prob.invert (), true);
+ {
+ if (loop_to_version != loop
+ && dump_enabled_p ())
+ dump_printf_loc (MSG_NOTE, vect_location,
+ "applying loop versioning to outer loop %d\n",
+ loop_to_version->num);
+
+ initialize_original_copy_tables ();
+ nloop = loop_version (loop_to_version, cond_expr, &condition_bb,
+ prob, prob.invert (), prob, prob.invert (), true);
+ gcc_assert (nloop);
+ nloop = get_loop_copy (loop);
+
+ /* Kill off IFN_LOOP_VECTORIZED_CALL in the copy, nobody will
+ reap those otherwise; they also refer to the original
+ loops. */
+ class loop *l = loop;
+ while (gimple *call = vect_loop_vectorized_call (l))
+ {
+ call = SSA_NAME_DEF_STMT (get_current_def (gimple_call_lhs (call)));
+ fold_loop_internal_call (call, boolean_false_node);
+ l = loop_outer (l);
+ }
+ free_original_copy_tables ();
+
+ if (cond_expr_stmt_list)
+ {
+ cond_exp_gsi = gsi_last_bb (condition_bb);
+ gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
+ GSI_SAME_STMT);
+ }
+
+ /* Loop versioning violates an assumption we try to maintain during
+ vectorization - that the loop exit block has a single predecessor.
+ After versioning, the exit block of both loop versions is the same
+ basic block (i.e. it has two predecessors). Just in order to simplify
+ following transformations in the vectorizer, we fix this situation
+ here by adding a new (empty) block on the exit-edge of the loop,
+ with the proper loop-exit phis to maintain loop-closed-form.
+ If loop versioning wasn't done from loop, but scalar_loop instead,
+ merge_bb will have already just a single successor. */
+
+ merge_bb = single_exit (loop_to_version)->dest;
+ if (EDGE_COUNT (merge_bb->preds) >= 2)
+ {
+ gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2);
+ new_exit_bb = split_edge (single_exit (loop_to_version));
+ new_exit_e = single_exit (loop_to_version);
+ e = EDGE_SUCC (new_exit_bb, 0);
+
+ for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi);
+ gsi_next (&gsi))
+ {
+ tree new_res;
+ orig_phi = gsi.phi ();
+ new_res = copy_ssa_name (PHI_RESULT (orig_phi));
+ new_phi = create_phi_node (new_res, new_exit_bb);
+ arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
+ add_phi_arg (new_phi, arg, new_exit_e,
+ gimple_phi_arg_location_from_edge (orig_phi, e));
+ adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi));
+ }
+ }
+
+ update_ssa (TODO_update_ssa);
+ }
if (version_niter)
{
loop_constraint_set (loop, LOOP_C_INFINITE);
}
- if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION
+ if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION
&& dump_enabled_p ())
{
if (version_alias)
- dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
+ dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING,
+ vect_location,
"loop versioned for vectorization because of "
"possible aliasing\n");
if (version_align)
- dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
+ dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | MSG_PRIORITY_USER_FACING,
+ vect_location,
"loop versioned for vectorization to enhance "
"alignment\n");
}
- free_original_copy_tables ();
-
- /* Loop versioning violates an assumption we try to maintain during
- vectorization - that the loop exit block has a single predecessor.
- After versioning, the exit block of both loop versions is the same
- basic block (i.e. it has two predecessors). Just in order to simplify
- following transformations in the vectorizer, we fix this situation
- here by adding a new (empty) block on the exit-edge of the loop,
- with the proper loop-exit phis to maintain loop-closed-form.
- If loop versioning wasn't done from loop, but scalar_loop instead,
- merge_bb will have already just a single successor. */
-
- merge_bb = single_exit (loop)->dest;
- if (scalar_loop == NULL || EDGE_COUNT (merge_bb->preds) >= 2)
- {
- gcc_assert (EDGE_COUNT (merge_bb->preds) >= 2);
- new_exit_bb = split_edge (single_exit (loop));
- new_exit_e = single_exit (loop);
- e = EDGE_SUCC (new_exit_bb, 0);
-
- for (gsi = gsi_start_phis (merge_bb); !gsi_end_p (gsi); gsi_next (&gsi))
- {
- tree new_res;
- orig_phi = gsi.phi ();
- new_res = copy_ssa_name (PHI_RESULT (orig_phi));
- new_phi = create_phi_node (new_res, new_exit_bb);
- arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
- add_phi_arg (new_phi, arg, new_exit_e,
- gimple_phi_arg_location_from_edge (orig_phi, e));
- adjust_phi_and_debug_stmts (orig_phi, e, PHI_RESULT (new_phi));
- }
- }
-
- /* End loop-exit-fixes after versioning. */
- if (cond_expr_stmt_list)
- {
- cond_exp_gsi = gsi_last_bb (condition_bb);
- gsi_insert_seq_before (&cond_exp_gsi, cond_expr_stmt_list,
- GSI_SAME_STMT);
- }
- update_ssa (TODO_update_ssa);
+ return nloop;
}