/* SCC value numbering for trees
- Copyright (C) 2006-2019 Free Software Foundation, Inc.
+ Copyright (C) 2006-2020 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dan@dberlin.org>
This file is part of GCC.
#include "tree-ssa.h"
#include "dumpfile.h"
#include "cfgloop.h"
-#include "params.h"
#include "tree-ssa-propagate.h"
#include "tree-cfg.h"
#include "domwalk.h"
/* Valueization hook. Valueize NAME if it is an SSA name, otherwise
just return it. */
tree (*vn_valueize) (tree);
+tree vn_valueize_wrapper (tree t, void* context ATTRIBUTE_UNUSED)
+{
+ return vn_valueize (t);
+}
/* This represents the top of the VN lattice, which is the universal
static inline hashval_t hash (const value_type &);
static inline bool equal (const value_type &, const compare_type &);
static inline void mark_deleted (value_type &) {}
+ static const bool empty_zero_p = true;
static inline void mark_empty (value_type &e) { e = NULL; }
static inline bool is_deleted (value_type &) { return false; }
static inline bool is_empty (value_type &e) { return e == NULL; }
break;
case STRING_CST:
case INTEGER_CST:
+ case POLY_INT_CST:
case COMPLEX_CST:
case VECTOR_CST:
case REAL_CST:
{
vn_walk_cb_data (vn_reference_t vr_, tree orig_ref_, tree *last_vuse_ptr_,
vn_lookup_kind vn_walk_kind_, bool tbaa_p_)
- : vr (vr_), last_vuse_ptr (last_vuse_ptr_),
- vn_walk_kind (vn_walk_kind_), tbaa_p (tbaa_p_), known_ranges (NULL)
+ : vr (vr_), last_vuse_ptr (last_vuse_ptr_), last_vuse (NULL_TREE),
+ vn_walk_kind (vn_walk_kind_), tbaa_p (tbaa_p_),
+ saved_operands (vNULL), first_set (-2), known_ranges (NULL)
{
+ if (!last_vuse_ptr)
+ last_vuse_ptr = &last_vuse;
ao_ref_init (&orig_ref, orig_ref_);
}
~vn_walk_cb_data ();
- void *push_partial_def (const pd_data& pd, tree, HOST_WIDE_INT);
+ void *finish (alias_set_type, tree);
+ void *push_partial_def (const pd_data& pd, alias_set_type, HOST_WIDE_INT);
vn_reference_t vr;
ao_ref orig_ref;
tree *last_vuse_ptr;
+ tree last_vuse;
vn_lookup_kind vn_walk_kind;
bool tbaa_p;
+ vec<vn_reference_op_s> saved_operands;
/* The VDEFs of partial defs we come along. */
auto_vec<pd_data, 2> partial_defs;
/* The first defs range to avoid splay tree setup in most cases. */
pd_range first_range;
- tree first_vuse;
+ alias_set_type first_set;
splay_tree known_ranges;
obstack ranges_obstack;
};
splay_tree_delete (known_ranges);
obstack_free (&ranges_obstack, NULL);
}
+ saved_operands.release ();
+}
+
+void *
+vn_walk_cb_data::finish (alias_set_type set, tree val)
+{
+ if (first_set != -2)
+ set = first_set;
+ return vn_reference_lookup_or_insert_for_pieces
+ (last_vuse, set, vr->type,
+ saved_operands.exists () ? saved_operands : vr->operands, val);
}
/* pd_range splay-tree helpers. */
}
/* Push PD to the vector of partial definitions returning a
- value when we are ready to combine things with VUSE and MAXSIZEI,
+ value when we are ready to combine things with VUSE, SET and MAXSIZEI,
NULL when we want to continue looking for partial defs or -1
on failure. */
void *
-vn_walk_cb_data::push_partial_def (const pd_data &pd, tree vuse,
- HOST_WIDE_INT maxsizei)
-{
+vn_walk_cb_data::push_partial_def (const pd_data &pd,
+ alias_set_type set, HOST_WIDE_INT maxsizei)
+{
+ const HOST_WIDE_INT bufsize = 64;
+ /* We're using a fixed buffer for encoding so fail early if the object
+ we want to interpret is bigger. */
+ if (maxsizei > bufsize * BITS_PER_UNIT
+ || CHAR_BIT != 8
+ || BITS_PER_UNIT != 8
+ /* Not prepared to handle PDP endian. */
+ || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN)
+ return (void *)-1;
+
+ bool pd_constant_p = (TREE_CODE (pd.rhs) == CONSTRUCTOR
+ || CONSTANT_CLASS_P (pd.rhs));
if (partial_defs.is_empty ())
{
+ /* If we get a clobber upfront, fail. */
+ if (TREE_CLOBBER_P (pd.rhs))
+ return (void *)-1;
+ if (!pd_constant_p)
+ return (void *)-1;
partial_defs.safe_push (pd);
first_range.offset = pd.offset;
first_range.size = pd.size;
- first_vuse = vuse;
+ first_set = set;
last_vuse_ptr = NULL;
/* Continue looking for partial defs. */
return NULL;
&& ranges_known_overlap_p (r->offset, r->size + 1,
newr.offset, newr.size))
{
- /* Ignore partial defs already covered. */
+ /* Ignore partial defs already covered. Here we also drop shadowed
+ clobbers arriving here at the floor. */
if (known_subrange_p (newr.offset, newr.size, r->offset, r->size))
return NULL;
r->size = MAX (r->offset + r->size, newr.offset + newr.size) - r->offset;
rafter->offset + rafter->size) - r->offset;
splay_tree_remove (known_ranges, (splay_tree_key)&rafter->offset);
}
+ /* If we get a clobber, fail. */
+ if (TREE_CLOBBER_P (pd.rhs))
+ return (void *)-1;
+ /* Non-constants are OK as long as they are shadowed by a constant. */
+ if (!pd_constant_p)
+ return (void *)-1;
partial_defs.safe_push (pd);
/* Now we have merged newr into the range tree. When we have covered
[offseti, sizei] then the tree will contain exactly one node which has
the desired properties and it will be 'r'. */
- if (!known_subrange_p (0, maxsizei / BITS_PER_UNIT, r->offset, r->size))
+ if (!known_subrange_p (0, maxsizei, r->offset, r->size))
/* Continue looking for partial defs. */
return NULL;
/* Now simply native encode all partial defs in reverse order. */
unsigned ndefs = partial_defs.length ();
/* We support up to 512-bit values (for V8DFmode). */
- unsigned char buffer[64];
+ unsigned char buffer[bufsize + 1];
+ unsigned char this_buffer[bufsize + 1];
int len;
+ memset (buffer, 0, bufsize + 1);
+ unsigned needed_len = ROUND_UP (maxsizei, BITS_PER_UNIT) / BITS_PER_UNIT;
while (!partial_defs.is_empty ())
{
pd_data pd = partial_defs.pop ();
+ unsigned int amnt;
if (TREE_CODE (pd.rhs) == CONSTRUCTOR)
- /* Empty CONSTRUCTOR. */
- memset (buffer + MAX (0, pd.offset),
- 0, MIN ((HOST_WIDE_INT)sizeof (buffer) - MAX (0, pd.offset),
- pd.size + MIN (0, pd.offset)));
+ {
+ /* Empty CONSTRUCTOR. */
+ if (pd.size >= needed_len * BITS_PER_UNIT)
+ len = needed_len;
+ else
+ len = ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT;
+ memset (this_buffer, 0, len);
+ }
else
{
- unsigned pad = 0;
- if (BYTES_BIG_ENDIAN
- && is_a <scalar_mode> (TYPE_MODE (TREE_TYPE (pd.rhs))))
- {
- /* On big-endian the padding is at the 'front' so just skip
- the initial bytes. */
- fixed_size_mode mode
- = as_a <fixed_size_mode> (TYPE_MODE (TREE_TYPE (pd.rhs)));
- pad = GET_MODE_SIZE (mode) - pd.size;
- }
- len = native_encode_expr (pd.rhs, buffer + MAX (0, pd.offset),
- sizeof (buffer) - MAX (0, pd.offset),
- MAX (0, -pd.offset) + pad);
- if (len <= 0 || len < (pd.size - MAX (0, -pd.offset)))
+ len = native_encode_expr (pd.rhs, this_buffer, bufsize,
+ MAX (0, -pd.offset) / BITS_PER_UNIT);
+ if (len <= 0
+ || len < (ROUND_UP (pd.size, BITS_PER_UNIT) / BITS_PER_UNIT
+ - MAX (0, -pd.offset) / BITS_PER_UNIT))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Failed to encode %u "
return (void *)-1;
}
}
+
+ unsigned char *p = buffer;
+ HOST_WIDE_INT size = pd.size;
+ if (pd.offset < 0)
+ size -= ROUND_DOWN (-pd.offset, BITS_PER_UNIT);
+ this_buffer[len] = 0;
+ if (BYTES_BIG_ENDIAN)
+ {
+ /* LSB of this_buffer[len - 1] byte should be at
+ pd.offset + pd.size - 1 bits in buffer. */
+ amnt = ((unsigned HOST_WIDE_INT) pd.offset
+ + pd.size) % BITS_PER_UNIT;
+ if (amnt)
+ shift_bytes_in_array_right (this_buffer, len + 1, amnt);
+ unsigned char *q = this_buffer;
+ unsigned int off = 0;
+ if (pd.offset >= 0)
+ {
+ unsigned int msk;
+ off = pd.offset / BITS_PER_UNIT;
+ gcc_assert (off < needed_len);
+ p = buffer + off;
+ if (size <= amnt)
+ {
+ msk = ((1 << size) - 1) << (BITS_PER_UNIT - amnt);
+ *p = (*p & ~msk) | (this_buffer[len] & msk);
+ size = 0;
+ }
+ else
+ {
+ if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
+ q = (this_buffer + len
+ - (ROUND_UP (size - amnt, BITS_PER_UNIT)
+ / BITS_PER_UNIT));
+ if (pd.offset % BITS_PER_UNIT)
+ {
+ msk = -1U << (BITS_PER_UNIT
+ - (pd.offset % BITS_PER_UNIT));
+ *p = (*p & msk) | (*q & ~msk);
+ p++;
+ q++;
+ off++;
+ size -= BITS_PER_UNIT - (pd.offset % BITS_PER_UNIT);
+ gcc_assert (size >= 0);
+ }
+ }
+ }
+ else if (TREE_CODE (pd.rhs) != CONSTRUCTOR)
+ {
+ q = (this_buffer + len
+ - (ROUND_UP (size - amnt, BITS_PER_UNIT)
+ / BITS_PER_UNIT));
+ if (pd.offset % BITS_PER_UNIT)
+ {
+ q++;
+ size -= BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) pd.offset
+ % BITS_PER_UNIT);
+ gcc_assert (size >= 0);
+ }
+ }
+ if ((unsigned HOST_WIDE_INT) size / BITS_PER_UNIT + off
+ > needed_len)
+ size = (needed_len - off) * BITS_PER_UNIT;
+ memcpy (p, q, size / BITS_PER_UNIT);
+ if (size % BITS_PER_UNIT)
+ {
+ unsigned int msk
+ = -1U << (BITS_PER_UNIT - (size % BITS_PER_UNIT));
+ p += size / BITS_PER_UNIT;
+ q += size / BITS_PER_UNIT;
+ *p = (*q & msk) | (*p & ~msk);
+ }
+ }
+ else
+ {
+ size = MIN (size, (HOST_WIDE_INT) needed_len * BITS_PER_UNIT);
+ if (pd.offset >= 0)
+ {
+ /* LSB of this_buffer[0] byte should be at pd.offset bits
+ in buffer. */
+ unsigned int msk;
+ amnt = pd.offset % BITS_PER_UNIT;
+ if (amnt)
+ shift_bytes_in_array_left (this_buffer, len + 1, amnt);
+ unsigned int off = pd.offset / BITS_PER_UNIT;
+ gcc_assert (off < needed_len);
+ p = buffer + off;
+ if (amnt + size < BITS_PER_UNIT)
+ {
+ /* Low amnt bits come from *p, then size bits
+ from this_buffer[0] and the remaining again from
+ *p. */
+ msk = ((1 << size) - 1) << amnt;
+ *p = (*p & ~msk) | (this_buffer[0] & msk);
+ size = 0;
+ }
+ else if (amnt)
+ {
+ msk = -1U << amnt;
+ *p = (*p & ~msk) | (this_buffer[0] & msk);
+ p++;
+ size -= (BITS_PER_UNIT - amnt);
+ }
+ }
+ else
+ {
+ amnt = (unsigned HOST_WIDE_INT) pd.offset % BITS_PER_UNIT;
+ if (amnt)
+ shift_bytes_in_array_left (this_buffer, len + 1, amnt);
+ }
+ memcpy (p, this_buffer + (amnt != 0), size / BITS_PER_UNIT);
+ p += size / BITS_PER_UNIT;
+ if (size % BITS_PER_UNIT)
+ {
+ unsigned int msk = -1U << (size % BITS_PER_UNIT);
+ *p = (this_buffer[(amnt != 0) + size / BITS_PER_UNIT]
+ & ~msk) | (*p & msk);
+ }
+ }
}
tree type = vr->type;
access size. */
if (INTEGRAL_TYPE_P (vr->type) && maxsizei != TYPE_PRECISION (vr->type))
type = build_nonstandard_integer_type (maxsizei, TYPE_UNSIGNED (type));
- tree val = native_interpret_expr (type, buffer, maxsizei / BITS_PER_UNIT);
+ tree val;
+ if (BYTES_BIG_ENDIAN)
+ {
+ unsigned sz = needed_len;
+ if (maxsizei % BITS_PER_UNIT)
+ shift_bytes_in_array_right (buffer, needed_len,
+ BITS_PER_UNIT
+ - (maxsizei % BITS_PER_UNIT));
+ if (INTEGRAL_TYPE_P (type))
+ sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
+ if (sz > needed_len)
+ {
+ memcpy (this_buffer + (sz - needed_len), buffer, needed_len);
+ val = native_interpret_expr (type, this_buffer, sz);
+ }
+ else
+ val = native_interpret_expr (type, buffer, needed_len);
+ }
+ else
+ val = native_interpret_expr (type, buffer, bufsize);
/* If we chop off bits because the types precision doesn't match the memory
access size this is ok when optimizing reads but not when called from
the DSE code during elimination. */
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"Successfully combined %u partial definitions\n", ndefs);
- return vn_reference_lookup_or_insert_for_pieces
- (first_vuse, vr->set, vr->type, vr->operands, val);
+ /* We are using the alias-set of the first store we encounter which
+ should be appropriate here. */
+ return finish (first_set, val);
}
else
{
return NULL;
if (data->last_vuse_ptr)
- *data->last_vuse_ptr = vuse;
+ {
+ *data->last_vuse_ptr = vuse;
+ data->last_vuse = vuse;
+ }
/* Fixup vuse and hash. */
if (vr->vuse)
hash = vr->hashcode;
slot = valid_info->references->find_slot_with_hash (vr, hash, NO_INSERT);
if (slot)
- return *slot;
+ {
+ if ((*slot)->result && data->saved_operands.exists ())
+ return data->finish (vr->set, (*slot)->result);
+ return *slot;
+ }
return NULL;
}
vn_reference_t vr = data->vr;
gimple *def_stmt = SSA_NAME_DEF_STMT (vuse);
tree base = ao_ref_base (ref);
- HOST_WIDE_INT offseti, maxsizei;
+ HOST_WIDE_INT offseti = 0, maxsizei, sizei = 0;
static vec<vn_reference_op_s> lhs_ops;
ao_ref lhs_ref;
bool lhs_ref_ok = false;
}
}
- /* If we are looking for redundant stores do not create new hashtable
- entries from aliasing defs with made up alias-sets. */
- if (*disambiguate_only > TR_TRANSLATE || !data->tbaa_p)
+ if (*disambiguate_only > TR_TRANSLATE)
return (void *)-1;
/* If we cannot constrain the size of the reference we cannot
poly_int64 offset = ref->offset;
poly_int64 maxsize = ref->max_size;
- /* We can't deduce anything useful from clobbers. */
- if (gimple_clobber_p (def_stmt))
- return (void *)-1;
-
/* def_stmt may-defs *ref. See if we can derive a value for *ref
from that definition.
1) Memset. */
if (is_gimple_reg_type (vr->type)
- && gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET)
+ && (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET)
+ || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMSET_CHK))
&& (integer_zerop (gimple_call_arg (def_stmt, 1))
|| ((TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST
|| (INTEGRAL_TYPE_P (vr->type) && known_eq (ref->size, 8)))
- && CHAR_BIT == 8 && BITS_PER_UNIT == 8
+ && CHAR_BIT == 8
+ && BITS_PER_UNIT == 8
+ && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
&& offset.is_constant (&offseti)
- && offseti % BITS_PER_UNIT == 0))
- && poly_int_tree_p (gimple_call_arg (def_stmt, 2))
+ && ref->size.is_constant (&sizei)
+ && (offseti % BITS_PER_UNIT == 0
+ || TREE_CODE (gimple_call_arg (def_stmt, 1)) == INTEGER_CST)))
+ && (poly_int_tree_p (gimple_call_arg (def_stmt, 2))
+ || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
+ && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)))))
&& (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
|| TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME))
{
else
return (void *)-1;
tree len = gimple_call_arg (def_stmt, 2);
- HOST_WIDE_INT leni, offset2i, offseti;
+ HOST_WIDE_INT leni, offset2i;
+ if (TREE_CODE (len) == SSA_NAME)
+ len = SSA_VAL (len);
+ /* Sometimes the above trickery is smarter than alias analysis. Take
+ advantage of that. */
+ if (!ranges_maybe_overlap_p (offset, maxsize, offset2,
+ (wi::to_poly_offset (len)
+ << LOG2_BITS_PER_UNIT)))
+ return NULL;
if (data->partial_defs.is_empty ()
&& known_subrange_p (offset, maxsize, offset2,
wi::to_poly_offset (len) << LOG2_BITS_PER_UNIT))
if (integer_zerop (gimple_call_arg (def_stmt, 1)))
val = build_zero_cst (vr->type);
else if (INTEGRAL_TYPE_P (vr->type)
- && known_eq (ref->size, 8))
+ && known_eq (ref->size, 8)
+ && offseti % BITS_PER_UNIT == 0)
{
gimple_match_op res_op (gimple_match_cond::UNCOND, NOP_EXPR,
vr->type, gimple_call_arg (def_stmt, 1));
}
else
{
- unsigned len = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type));
- unsigned char *buf = XALLOCAVEC (unsigned char, len);
+ unsigned buflen = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (vr->type)) + 1;
+ if (INTEGRAL_TYPE_P (vr->type))
+ buflen = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (vr->type)) + 1;
+ unsigned char *buf = XALLOCAVEC (unsigned char, buflen);
memset (buf, TREE_INT_CST_LOW (gimple_call_arg (def_stmt, 1)),
- len);
- val = native_interpret_expr (vr->type, buf, len);
+ buflen);
+ if (BYTES_BIG_ENDIAN)
+ {
+ unsigned int amnt
+ = (((unsigned HOST_WIDE_INT) offseti + sizei)
+ % BITS_PER_UNIT);
+ if (amnt)
+ {
+ shift_bytes_in_array_right (buf, buflen,
+ BITS_PER_UNIT - amnt);
+ buf++;
+ buflen--;
+ }
+ }
+ else if (offseti % BITS_PER_UNIT != 0)
+ {
+ unsigned int amnt
+ = BITS_PER_UNIT - ((unsigned HOST_WIDE_INT) offseti
+ % BITS_PER_UNIT);
+ shift_bytes_in_array_left (buf, buflen, amnt);
+ buf++;
+ buflen--;
+ }
+ val = native_interpret_expr (vr->type, buf, buflen);
if (!val)
return (void *)-1;
}
- return vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
+ return data->finish (0, val);
}
/* For now handle clearing memory with partial defs. */
else if (known_eq (ref->size, maxsize)
&& integer_zerop (gimple_call_arg (def_stmt, 1))
+ && tree_fits_poly_int64_p (len)
&& tree_to_poly_int64 (len).is_constant (&leni)
+ && leni <= INTTYPE_MAXIMUM (HOST_WIDE_INT) / BITS_PER_UNIT
&& offset.is_constant (&offseti)
&& offset2.is_constant (&offset2i)
- && maxsize.is_constant (&maxsizei))
+ && maxsize.is_constant (&maxsizei)
+ && ranges_known_overlap_p (offseti, maxsizei, offset2i,
+ leni << LOG2_BITS_PER_UNIT))
{
pd_data pd;
pd.rhs = build_constructor (NULL_TREE, NULL);
- pd.offset = (offset2i - offseti) / BITS_PER_UNIT;
- pd.size = leni;
- return data->push_partial_def (pd, vuse, maxsizei);
+ pd.offset = offset2i - offseti;
+ pd.size = leni << LOG2_BITS_PER_UNIT;
+ return data->push_partial_def (pd, 0, maxsizei);
}
}
if (data->partial_defs.is_empty ()
&& known_subrange_p (offset, maxsize, offset2, size2))
{
+ /* While technically undefined behavior do not optimize
+ a full read from a clobber. */
+ if (gimple_clobber_p (def_stmt))
+ return (void *)-1;
tree val = build_zero_cst (vr->type);
- return vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
+ return data->finish (get_alias_set (lhs), val);
}
else if (known_eq (ref->size, maxsize)
&& maxsize.is_constant (&maxsizei)
- && maxsizei % BITS_PER_UNIT == 0
&& offset.is_constant (&offseti)
- && offseti % BITS_PER_UNIT == 0
&& offset2.is_constant (&offset2i)
- && offset2i % BITS_PER_UNIT == 0
&& size2.is_constant (&size2i)
- && size2i % BITS_PER_UNIT == 0)
+ && ranges_known_overlap_p (offseti, maxsizei,
+ offset2i, size2i))
{
+ /* Let clobbers be consumed by the partial-def tracker
+ which can choose to ignore them if they are shadowed
+ by a later def. */
pd_data pd;
pd.rhs = gimple_assign_rhs1 (def_stmt);
- pd.offset = (offset2i - offseti) / BITS_PER_UNIT;
- pd.size = size2i / BITS_PER_UNIT;
- return data->push_partial_def (pd, vuse, maxsizei);
+ pd.offset = offset2i - offseti;
+ pd.size = size2i;
+ return data->push_partial_def (pd, get_alias_set (lhs), maxsizei);
}
}
}
&& is_gimple_reg_type (vr->type)
&& !contains_storage_order_barrier_p (vr->operands)
&& gimple_assign_single_p (def_stmt)
- && CHAR_BIT == 8 && BITS_PER_UNIT == 8
+ && CHAR_BIT == 8
+ && BITS_PER_UNIT == 8
+ && BYTES_BIG_ENDIAN == WORDS_BIG_ENDIAN
/* native_encode and native_decode operate on arrays of bytes
and so fundamentally need a compile-time size and offset. */
&& maxsize.is_constant (&maxsizei)
- && maxsizei % BITS_PER_UNIT == 0
&& offset.is_constant (&offseti)
- && offseti % BITS_PER_UNIT == 0
&& (is_gimple_min_invariant (gimple_assign_rhs1 (def_stmt))
|| (TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
&& is_gimple_min_invariant (SSA_VAL (gimple_assign_rhs1 (def_stmt))))))
&offset2, &size2, &maxsize2, &reverse);
if (base2
&& !reverse
+ && !storage_order_barrier_p (lhs)
&& known_eq (maxsize2, size2)
- && multiple_p (size2, BITS_PER_UNIT)
- && multiple_p (offset2, BITS_PER_UNIT)
&& adjust_offsets_for_equal_base_address (base, &offset,
base2, &offset2)
&& offset.is_constant (&offseti)
&& known_subrange_p (offseti, maxsizei, offset2, size2))
{
/* We support up to 512-bit values (for V8DFmode). */
- unsigned char buffer[64];
+ unsigned char buffer[65];
int len;
tree rhs = gimple_assign_rhs1 (def_stmt);
if (TREE_CODE (rhs) == SSA_NAME)
rhs = SSA_VAL (rhs);
- unsigned pad = 0;
- if (BYTES_BIG_ENDIAN
- && is_a <scalar_mode> (TYPE_MODE (TREE_TYPE (rhs))))
- {
- /* On big-endian the padding is at the 'front' so
- just skip the initial bytes. */
- fixed_size_mode mode
- = as_a <fixed_size_mode> (TYPE_MODE (TREE_TYPE (rhs)));
- pad = GET_MODE_SIZE (mode) - size2i / BITS_PER_UNIT;
- }
len = native_encode_expr (rhs,
- buffer, sizeof (buffer),
- ((offseti - offset2i) / BITS_PER_UNIT
- + pad));
+ buffer, sizeof (buffer) - 1,
+ (offseti - offset2i) / BITS_PER_UNIT);
if (len > 0 && len * BITS_PER_UNIT >= maxsizei)
{
tree type = vr->type;
+ unsigned char *buf = buffer;
+ unsigned int amnt = 0;
/* Make sure to interpret in a type that has a range
covering the whole access size. */
if (INTEGRAL_TYPE_P (vr->type)
&& maxsizei != TYPE_PRECISION (vr->type))
type = build_nonstandard_integer_type (maxsizei,
TYPE_UNSIGNED (type));
- tree val = native_interpret_expr (type, buffer,
- maxsizei / BITS_PER_UNIT);
+ if (BYTES_BIG_ENDIAN)
+ {
+ /* For big-endian native_encode_expr stored the rhs
+ such that the LSB of it is the LSB of buffer[len - 1].
+ That bit is stored into memory at position
+ offset2 + size2 - 1, i.e. in byte
+ base + (offset2 + size2 - 1) / BITS_PER_UNIT.
+ E.g. for offset2 1 and size2 14, rhs -1 and memory
+ previously cleared that is:
+ 0 1
+ 01111111|11111110
+ Now, if we want to extract offset 2 and size 12 from
+ it using native_interpret_expr (which actually works
+ for integral bitfield types in terms of byte size of
+ the mode), the native_encode_expr stored the value
+ into buffer as
+ XX111111|11111111
+ and returned len 2 (the X bits are outside of
+ precision).
+ Let sz be maxsize / BITS_PER_UNIT if not extracting
+ a bitfield, and GET_MODE_SIZE otherwise.
+ We need to align the LSB of the value we want to
+ extract as the LSB of buf[sz - 1].
+ The LSB from memory we need to read is at position
+ offset + maxsize - 1. */
+ HOST_WIDE_INT sz = maxsizei / BITS_PER_UNIT;
+ if (INTEGRAL_TYPE_P (type))
+ sz = GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (type));
+ amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
+ - offseti - maxsizei) % BITS_PER_UNIT;
+ if (amnt)
+ shift_bytes_in_array_right (buffer, len, amnt);
+ amnt = ((unsigned HOST_WIDE_INT) offset2i + size2i
+ - offseti - maxsizei - amnt) / BITS_PER_UNIT;
+ if ((unsigned HOST_WIDE_INT) sz + amnt > (unsigned) len)
+ len = 0;
+ else
+ {
+ buf = buffer + len - sz - amnt;
+ len -= (buf - buffer);
+ }
+ }
+ else
+ {
+ amnt = ((unsigned HOST_WIDE_INT) offset2i
+ - offseti) % BITS_PER_UNIT;
+ if (amnt)
+ {
+ buffer[len] = 0;
+ shift_bytes_in_array_left (buffer, len + 1, amnt);
+ buf = buffer + 1;
+ }
+ }
+ tree val = native_interpret_expr (type, buf, len);
/* If we chop off bits because the types precision doesn't
match the memory access size this is ok when optimizing
reads but not when called from the DSE code during
}
if (val)
- return vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
+ return data->finish (get_alias_set (lhs), val);
}
}
- else if (ranges_known_overlap_p (offseti, maxsizei, offset2i, size2i))
+ else if (ranges_known_overlap_p (offseti, maxsizei, offset2i,
+ size2i))
{
pd_data pd;
tree rhs = gimple_assign_rhs1 (def_stmt);
if (TREE_CODE (rhs) == SSA_NAME)
rhs = SSA_VAL (rhs);
pd.rhs = rhs;
- pd.offset = (offset2i - offseti) / BITS_PER_UNIT;
- pd.size = size2i / BITS_PER_UNIT;
- return data->push_partial_def (pd, vuse, maxsizei);
+ pd.offset = offset2i - offseti;
+ pd.size = size2i;
+ return data->push_partial_def (pd, get_alias_set (lhs), maxsizei);
}
}
}
/* 4) Assignment from an SSA name which definition we may be able
- to access pieces from. */
+ to access pieces from or we can combine to a larger entity. */
else if (known_eq (ref->size, maxsize)
&& is_gimple_reg_type (vr->type)
&& !contains_storage_order_barrier_p (vr->operands)
&& gimple_assign_single_p (def_stmt)
- && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME
- /* A subset of partial defs from non-constants can be handled
- by for example inserting a CONSTRUCTOR, a COMPLEX_EXPR or
- even a (series of) BIT_INSERT_EXPR hoping for simplifications
- downstream, not so much for actually doing the insertion. */
- && data->partial_defs.is_empty ())
+ && TREE_CODE (gimple_assign_rhs1 (def_stmt)) == SSA_NAME)
{
tree lhs = gimple_assign_lhs (def_stmt);
tree base2;
poly_int64 offset2, size2, maxsize2;
+ HOST_WIDE_INT offset2i, size2i, offseti;
bool reverse;
if (lhs_ref_ok)
{
&offset2, &size2, &maxsize2, &reverse);
tree def_rhs = gimple_assign_rhs1 (def_stmt);
if (!reverse
+ && !storage_order_barrier_p (lhs)
&& known_size_p (maxsize2)
&& known_eq (maxsize2, size2)
&& adjust_offsets_for_equal_base_address (base, &offset,
- base2, &offset2)
- && known_subrange_p (offset, maxsize, offset2, size2)
- /* ??? We can't handle bitfield precision extracts without
- either using an alternate type for the BIT_FIELD_REF and
- then doing a conversion or possibly adjusting the offset
- according to endianness. */
- && (! INTEGRAL_TYPE_P (vr->type)
- || known_eq (ref->size, TYPE_PRECISION (vr->type)))
- && multiple_p (ref->size, BITS_PER_UNIT)
- && (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs))
- || type_has_mode_precision_p (TREE_TYPE (def_rhs))))
+ base2, &offset2))
{
- gimple_match_op op (gimple_match_cond::UNCOND,
- BIT_FIELD_REF, vr->type,
- vn_valueize (def_rhs),
- bitsize_int (ref->size),
- bitsize_int (offset - offset2));
- tree val = vn_nary_build_or_lookup (&op);
- if (val
- && (TREE_CODE (val) != SSA_NAME
- || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
+ if (data->partial_defs.is_empty ()
+ && known_subrange_p (offset, maxsize, offset2, size2)
+ /* ??? We can't handle bitfield precision extracts without
+ either using an alternate type for the BIT_FIELD_REF and
+ then doing a conversion or possibly adjusting the offset
+ according to endianness. */
+ && (! INTEGRAL_TYPE_P (vr->type)
+ || known_eq (ref->size, TYPE_PRECISION (vr->type)))
+ && multiple_p (ref->size, BITS_PER_UNIT))
{
- vn_reference_t res = vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
- return res;
+ tree val = NULL_TREE;
+ if (! INTEGRAL_TYPE_P (TREE_TYPE (def_rhs))
+ || type_has_mode_precision_p (TREE_TYPE (def_rhs)))
+ {
+ gimple_match_op op (gimple_match_cond::UNCOND,
+ BIT_FIELD_REF, vr->type,
+ SSA_VAL (def_rhs),
+ bitsize_int (ref->size),
+ bitsize_int (offset - offset2));
+ val = vn_nary_build_or_lookup (&op);
+ }
+ else if (known_eq (ref->size, size2))
+ {
+ gimple_match_op op (gimple_match_cond::UNCOND,
+ VIEW_CONVERT_EXPR, vr->type,
+ SSA_VAL (def_rhs));
+ val = vn_nary_build_or_lookup (&op);
+ }
+ if (val
+ && (TREE_CODE (val) != SSA_NAME
+ || ! SSA_NAME_OCCURS_IN_ABNORMAL_PHI (val)))
+ return data->finish (get_alias_set (lhs), val);
+ }
+ else if (maxsize.is_constant (&maxsizei)
+ && offset.is_constant (&offseti)
+ && offset2.is_constant (&offset2i)
+ && size2.is_constant (&size2i)
+ && ranges_known_overlap_p (offset, maxsize, offset2, size2))
+ {
+ pd_data pd;
+ pd.rhs = SSA_VAL (def_rhs);
+ pd.offset = offset2i - offseti;
+ pd.size = size2i;
+ return data->push_partial_def (pd, get_alias_set (lhs), maxsizei);
}
}
}
|| TREE_CODE (gimple_assign_rhs1 (def_stmt)) == MEM_REF
|| handled_component_p (gimple_assign_rhs1 (def_stmt))))
{
+ tree lhs = gimple_assign_lhs (def_stmt);
tree base2;
int i, j, k;
auto_vec<vn_reference_op_s> rhs;
}
/* Now re-write REF to be based on the rhs of the assignment. */
- copy_reference_ops_from_ref (gimple_assign_rhs1 (def_stmt), &rhs);
+ tree rhs1 = gimple_assign_rhs1 (def_stmt);
+ copy_reference_ops_from_ref (rhs1, &rhs);
/* Apply an extra offset to the inner MEM_REF of the RHS. */
if (maybe_ne (extra_off, 0))
extra_off));
}
+ /* Save the operands since we need to use the original ones for
+ the hash entry we use. */
+ if (!data->saved_operands.exists ())
+ data->saved_operands = vr->operands.copy ();
+
/* We need to pre-pend vr->operands[0..i] to rhs. */
vec<vn_reference_op_s> old = vr->operands;
if (i + 1 + rhs.length () > vr->operands.length ())
if (val)
{
if (data->partial_defs.is_empty ())
- return vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
+ return data->finish (get_alias_set (lhs), val);
/* This is the only interesting case for partial-def handling
coming from targets that like to gimplify init-ctors as
aggregate copies from constant data like aarch64 for
PR83518. */
- if (maxsize.is_constant (&maxsizei)
- && known_eq (ref->size, maxsize))
+ if (maxsize.is_constant (&maxsizei) && known_eq (ref->size, maxsize))
{
pd_data pd;
pd.rhs = val;
pd.offset = 0;
- pd.size = maxsizei / BITS_PER_UNIT;
- return data->push_partial_def (pd, vuse, maxsizei);
+ pd.size = maxsizei;
+ return data->push_partial_def (pd, get_alias_set (lhs),
+ maxsizei);
}
}
return (void *)-1;
/* Adjust *ref from the new operands. */
- if (!ao_ref_init_from_vn_reference (&r, vr->set, vr->type, vr->operands))
+ if (!ao_ref_init_from_vn_reference (&r, get_alias_set (rhs1),
+ vr->type, vr->operands))
return (void *)-1;
/* This can happen with bitfields. */
if (maybe_ne (ref->size, r.size))
/* Invalidate the original access path since it now contains
the wrong base. */
data->orig_ref.ref = NULL_TREE;
+ /* Use the alias-set of this LHS for recording an eventual result. */
+ if (data->first_set == -2)
+ data->first_set = get_alias_set (lhs);
/* Keep looking for the adjusted *REF / VR pair. */
return NULL;
&& is_gimple_reg_type (vr->type)
/* ??? Handle BCOPY as well. */
&& (gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY)
+ || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMCPY_CHK)
|| gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY)
- || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE))
+ || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMPCPY_CHK)
+ || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE)
+ || gimple_call_builtin_p (def_stmt, BUILT_IN_MEMMOVE_CHK))
&& (TREE_CODE (gimple_call_arg (def_stmt, 0)) == ADDR_EXPR
|| TREE_CODE (gimple_call_arg (def_stmt, 0)) == SSA_NAME)
&& (TREE_CODE (gimple_call_arg (def_stmt, 1)) == ADDR_EXPR
|| TREE_CODE (gimple_call_arg (def_stmt, 1)) == SSA_NAME)
- && poly_int_tree_p (gimple_call_arg (def_stmt, 2), ©_size)
+ && (poly_int_tree_p (gimple_call_arg (def_stmt, 2), ©_size)
+ || (TREE_CODE (gimple_call_arg (def_stmt, 2)) == SSA_NAME
+ && poly_int_tree_p (SSA_VAL (gimple_call_arg (def_stmt, 2)),
+ ©_size)))
/* Handling this is more complicated, give up for now. */
&& data->partial_defs.is_empty ())
{
else
return (void *)-1;
}
- if (TREE_CODE (rhs) != SSA_NAME
- && TREE_CODE (rhs) != ADDR_EXPR)
+ if (TREE_CODE (rhs) == SSA_NAME)
+ rhs = SSA_VAL (rhs);
+ else if (TREE_CODE (rhs) != ADDR_EXPR)
return (void *)-1;
/* The bases of the destination and the references have to agree. */
if (!known_subrange_p (at, byte_maxsize, lhs_offset, copy_size))
return (void *)-1;
+ /* Save the operands since we need to use the original ones for
+ the hash entry we use. */
+ if (!data->saved_operands.exists ())
+ data->saved_operands = vr->operands.copy ();
+
/* Make room for 2 operands in the new reference. */
if (vr->operands.length () < 2)
{
/* Try folding the new reference to a constant. */
tree val = fully_constant_vn_reference_p (vr);
if (val)
- return vn_reference_lookup_or_insert_for_pieces
- (vuse, vr->set, vr->type, vr->operands, val);
+ return data->finish (0, val);
/* Adjust *ref from the new operands. */
- if (!ao_ref_init_from_vn_reference (&r, vr->set, vr->type, vr->operands))
+ if (!ao_ref_init_from_vn_reference (&r, 0, vr->type, vr->operands))
return (void *)-1;
/* This can happen with bitfields. */
if (maybe_ne (ref->size, r.size))
/* Invalidate the original access path since it now contains
the wrong base. */
data->orig_ref.ref = NULL_TREE;
+ /* Use the alias-set of this stmt for recording an eventual result. */
+ if (data->first_set == -2)
+ data->first_set = 0;
/* Keep looking for the adjusted *REF / VR pair. */
return NULL;
&& vr1.vuse)
{
ao_ref r;
- unsigned limit = PARAM_VALUE (PARAM_SCCVN_MAX_ALIAS_QUERIES_PER_ACCESS);
+ unsigned limit = param_sccvn_max_alias_queries_per_access;
vn_walk_cb_data data (&vr1, NULL_TREE, NULL, kind, true);
if (ao_ref_init_from_vn_reference (&r, set, type, vr1.operands))
*vnresult =
{
vn_reference_t wvnresult;
ao_ref r;
- unsigned limit = PARAM_VALUE (PARAM_SCCVN_MAX_ALIAS_QUERIES_PER_ACCESS);
+ unsigned limit = param_sccvn_max_alias_queries_per_access;
/* Make sure to use a valueized reference if we valueized anything.
Otherwise preserve the full reference for advanced TBAA. */
if (!valuezied_anything
return NULL_TREE;
}
+ if (last_vuse_ptr)
+ *last_vuse_ptr = vr1.vuse;
return vn_reference_lookup_1 (&vr1, vnresult);
}
honor_nans = flag_trapping_math && !flag_finite_math_only;
honor_snans = flag_signaling_nans != 0;
}
- else if (INTEGRAL_TYPE_P (type)
- && TYPE_OVERFLOW_TRAPS (type))
+ else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type))
honor_trapv = true;
}
if (nary->length >= 2)
rhs2 = nary->op[1];
ret = operation_could_trap_helper_p (nary->opcode, fp_operation,
- honor_trapv,
- honor_nans, honor_snans, rhs2,
- &handled);
- if (handled
- && ret)
+ honor_trapv, honor_nans, honor_snans,
+ rhs2, &handled);
+ if (handled && ret)
return true;
for (i = 0; i < nary->length; ++i)
/* If this is an assignment from our leader (which
happens in the case the value-number is a constant)
- then there is nothing to do. */
- if (gimple_assign_single_p (stmt)
+ then there is nothing to do. Likewise if we run into
+ inserted code that needed a conversion because of
+ our type-agnostic value-numbering of loads. */
+ if ((gimple_assign_single_p (stmt)
+ || (is_gimple_assign (stmt)
+ && (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt))
+ || gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR)))
&& sprime == gimple_assign_rhs1 (stmt))
return;
&& (TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME
|| is_gimple_min_invariant (gimple_assign_rhs1 (stmt))))
{
- tree val;
tree rhs = gimple_assign_rhs1 (stmt);
vn_reference_t vnresult;
- val = vn_reference_lookup (lhs, gimple_vuse (stmt), VN_WALKREWRITE,
- &vnresult, false);
+ /* ??? gcc.dg/torture/pr91445.c shows that we lookup a boolean
+ typed load of a byte known to be 0x11 as 1 so a store of
+ a boolean 1 is detected as redundant. Because of this we
+ have to make sure to lookup with a ref where its size
+ matches the precision. */
+ tree lookup_lhs = lhs;
+ if (INTEGRAL_TYPE_P (TREE_TYPE (lhs))
+ && (TREE_CODE (lhs) != COMPONENT_REF
+ || !DECL_BIT_FIELD_TYPE (TREE_OPERAND (lhs, 1)))
+ && !type_has_mode_precision_p (TREE_TYPE (lhs)))
+ {
+ if (TREE_CODE (lhs) == COMPONENT_REF
+ || TREE_CODE (lhs) == MEM_REF)
+ {
+ tree ltype = build_nonstandard_integer_type
+ (TREE_INT_CST_LOW (TYPE_SIZE (TREE_TYPE (lhs))),
+ TYPE_UNSIGNED (TREE_TYPE (lhs)));
+ if (TREE_CODE (lhs) == COMPONENT_REF)
+ {
+ tree foff = component_ref_field_offset (lhs);
+ tree f = TREE_OPERAND (lhs, 1);
+ if (!poly_int_tree_p (foff))
+ lookup_lhs = NULL_TREE;
+ else
+ lookup_lhs = build3 (BIT_FIELD_REF, ltype,
+ TREE_OPERAND (lhs, 0),
+ TYPE_SIZE (TREE_TYPE (lhs)),
+ bit_from_pos
+ (foff, DECL_FIELD_BIT_OFFSET (f)));
+ }
+ else
+ lookup_lhs = build2 (MEM_REF, ltype,
+ TREE_OPERAND (lhs, 0),
+ TREE_OPERAND (lhs, 1));
+ }
+ else
+ lookup_lhs = NULL_TREE;
+ }
+ tree val = NULL_TREE;
+ if (lookup_lhs)
+ val = vn_reference_lookup (lookup_lhs, gimple_vuse (stmt),
+ VN_WALKREWRITE, &vnresult, false);
if (TREE_CODE (rhs) == SSA_NAME)
rhs = VN_INFO (rhs)->valnum;
if (val
- && operand_equal_p (val, rhs, 0))
+ && (operand_equal_p (val, rhs, 0)
+ /* Due to the bitfield lookups above we can get bit
+ interpretations of the same RHS as values here. Those
+ are redundant as well. */
+ || (TREE_CODE (val) == SSA_NAME
+ && gimple_assign_single_p (SSA_NAME_DEF_STMT (val))
+ && (val = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (val)))
+ && TREE_CODE (val) == VIEW_CONVERT_EXPR
+ && TREE_OPERAND (val, 0) == rhs)))
{
/* We can only remove the later store if the former aliases
at least all accesses the later one does or if the store
if (bb->loop_father->nb_iterations)
bb->loop_father->nb_iterations
= simplify_replace_tree (bb->loop_father->nb_iterations,
- NULL_TREE, NULL_TREE, vn_valueize);
+ NULL_TREE, NULL_TREE, &vn_valueize_wrapper);
}
/* Value-number all defs in the basic-block. */
if (iterate)
{
loop_p loop;
- unsigned max_depth = PARAM_VALUE (PARAM_RPO_VN_MAX_LOOP_DEPTH);
+ unsigned max_depth = param_rpo_vn_max_loop_depth;
FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)
if (loop_depth (loop) > max_depth)
for (unsigned i = 2;