Google Git
Sign in
gnu / gcc / 2c564e813c0626802e5bfb066c094933d5e6a774 / . / gcc / tree-ssa-reassoc.c
blob: db9fb4e1cac8e7428cf6f2e6dcdf3044d9c3f07c [file] [log] [blame]
/* Reassociation for trees.
Copyright (C) 2005-2021 Free Software Foundation, Inc.
Contributed by Daniel Berlin <dan@dberlin.org>
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3, or (at your option)
any later version.
GCC is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "target.h"
#include "rtl.h"
#include "tree.h"
#include "gimple.h"
#include "cfghooks.h"
#include "alloc-pool.h"
#include "tree-pass.h"
#include "memmodel.h"
#include "tm_p.h"
#include "ssa.h"
#include "optabs-tree.h"
#include "gimple-pretty-print.h"
#include "diagnostic-core.h"
#include "fold-const.h"
#include "stor-layout.h"
#include "cfganal.h"
#include "gimple-fold.h"
#include "tree-eh.h"
#include "gimple-iterator.h"
#include "gimplify-me.h"
#include "tree-cfg.h"
#include "tree-ssa-loop.h"
#include "flags.h"
#include "tree-ssa.h"
#include "langhooks.h"
#include "cfgloop.h"
#include "builtins.h"
#include "gimplify.h"
#include "case-cfn-macros.h"
#include "tree-ssa-reassoc.h"
#include "tree-ssa-math-opts.h"
#include "gimple-range.h"
/* This is a simple global reassociation pass. It is, in part, based
on the LLVM pass of the same name (They do some things more/less
than we do, in different orders, etc).
It consists of five steps:
1. Breaking up subtract operations into addition + negate, where
it would promote the reassociation of adds.
2. Left linearization of the expression trees, so that (A+B)+(C+D)
becomes (((A+B)+C)+D), which is easier for us to rewrite later.
During linearization, we place the operands of the binary
expressions into a vector of operand_entry_*
3. Optimization of the operand lists, eliminating things like a +
-a, a & a, etc.
3a. Combine repeated factors with the same occurrence counts
into a __builtin_powi call that will later be optimized into
an optimal number of multiplies.
4. Rewrite the expression trees we linearized and optimized so
they are in proper rank order.
5. Repropagate negates, as nothing else will clean it up ATM.
A bit of theory on #4, since nobody seems to write anything down
about why it makes sense to do it the way they do it:
We could do this much nicer theoretically, but don't (for reasons
explained after how to do it theoretically nice :P).
In order to promote the most redundancy elimination, you want
binary expressions whose operands are the same rank (or
preferably, the same value) exposed to the redundancy eliminator,
for possible elimination.
So the way to do this if we really cared, is to build the new op
tree from the leaves to the roots, merging as you go, and putting the
new op on the end of the worklist, until you are left with one
thing on the worklist.
IE if you have to rewrite the following set of operands (listed with
rank in parentheses), with opcode PLUS_EXPR:
a (1), b (1), c (1), d (2), e (2)
We start with our merge worklist empty, and the ops list with all of
those on it.
You want to first merge all leaves of the same rank, as much as
possible.
So first build a binary op of
mergetmp = a + b, and put "mergetmp" on the merge worklist.
Because there is no three operand form of PLUS_EXPR, c is not going to
be exposed to redundancy elimination as a rank 1 operand.
So you might as well throw it on the merge worklist (you could also
consider it to now be a rank two operand, and merge it with d and e,
but in this case, you then have evicted e from a binary op. So at
least in this situation, you can't win.)
Then build a binary op of d + e
mergetmp2 = d + e
and put mergetmp2 on the merge worklist.
so merge worklist = {mergetmp, c, mergetmp2}
Continue building binary ops of these operations until you have only
one operation left on the worklist.
So we have
build binary op
mergetmp3 = mergetmp + c
worklist = {mergetmp2, mergetmp3}
mergetmp4 = mergetmp2 + mergetmp3
worklist = {mergetmp4}
because we have one operation left, we can now just set the original
statement equal to the result of that operation.
This will at least expose a + b and d + e to redundancy elimination
as binary operations.
For extra points, you can reuse the old statements to build the
mergetmps, since you shouldn't run out.
So why don't we do this?
Because it's expensive, and rarely will help. Most trees we are
reassociating have 3 or less ops. If they have 2 ops, they already
will be written into a nice single binary op. If you have 3 ops, a
single simple check suffices to tell you whether the first two are of the
same rank. If so, you know to order it
mergetmp = op1 + op2
newstmt = mergetmp + op3
instead of
mergetmp = op2 + op3
newstmt = mergetmp + op1
If all three are of the same rank, you can't expose them all in a
single binary operator anyway, so the above is *still* the best you
can do.
Thus, this is what we do. When we have three ops left, we check to see
what order to put them in, and call it a day. As a nod to vector sum
reduction, we check if any of the ops are really a phi node that is a
destructive update for the associating op, and keep the destructive
update together for vector sum reduction recognition. */
/* Enable insertion of __builtin_powi calls during execute_reassoc. See
point 3a in the pass header comment. */
static bool reassoc_insert_powi_p;
/* Enable biasing ranks of loop accumulators. We don't want this before
vectorization, since it interferes with reduction chains. */
static bool reassoc_bias_loop_carried_phi_ranks_p;
/* Statistics */
static struct
{
int linearized;
int constants_eliminated;
int ops_eliminated;
int rewritten;
int pows_encountered;
int pows_created;
} reassociate_stats;
static object_allocator<operand_entry> operand_entry_pool
("operand entry pool");
/* This is used to assign a unique ID to each struct operand_entry
so that qsort results are identical on different hosts. */
static unsigned int next_operand_entry_id;
/* Starting rank number for a given basic block, so that we can rank
operations using unmovable instructions in that BB based on the bb
depth. */
static int64_t *bb_rank;
/* Operand->rank hashtable. */
static hash_map<tree, int64_t> *operand_rank;
/* SSA_NAMEs that are forms of loop accumulators and whose ranks need to be
biased. */
static auto_bitmap biased_names;
/* Vector of SSA_NAMEs on which after reassociate_bb is done with
all basic blocks the CFG should be adjusted - basic blocks
split right after that SSA_NAME's definition statement and before
the only use, which must be a bit ior. */
static vec<tree> reassoc_branch_fixups;
/* Forward decls. */
static int64_t get_rank (tree);
static bool reassoc_stmt_dominates_stmt_p (gimple *, gimple *);
/* Wrapper around gsi_remove, which adjusts gimple_uid of debug stmts
possibly added by gsi_remove. */
static bool
reassoc_remove_stmt (gimple_stmt_iterator *gsi)
{
gimple *stmt = gsi_stmt (*gsi);
if (!MAY_HAVE_DEBUG_BIND_STMTS || gimple_code (stmt) == GIMPLE_PHI)
return gsi_remove (gsi, true);
gimple_stmt_iterator prev = *gsi;
gsi_prev (&prev);
unsigned uid = gimple_uid (stmt);
basic_block bb = gimple_bb (stmt);
bool ret = gsi_remove (gsi, true);
if (!gsi_end_p (prev))
gsi_next (&prev);
else
prev = gsi_start_bb (bb);
gimple *end_stmt = gsi_stmt (*gsi);
while ((stmt = gsi_stmt (prev)) != end_stmt)
{
gcc_assert (stmt && is_gimple_debug (stmt) && gimple_uid (stmt) == 0);
gimple_set_uid (stmt, uid);
gsi_next (&prev);
}
return ret;
}
/* Bias amount for loop-carried phis. We want this to be larger than
the depth of any reassociation tree we can see, but not larger than
the rank difference between two blocks. */
#define PHI_LOOP_BIAS (1 << 15)
/* Return TRUE iff PHI_LOOP_BIAS should be propagated from one of the STMT's
operands to the STMT's left-hand side. The goal is to preserve bias in code
like this:
x_1 = phi(x_0, x_2)
a = x_1 | 1
b = a ^ 2
.MEM = b
c = b + d
x_2 = c + e
That is, we need to preserve bias along single-use chains originating from
loop-carried phis. Only GIMPLE_ASSIGNs to SSA_NAMEs are considered to be
uses, because only they participate in rank propagation. */
static bool
propagate_bias_p (gimple *stmt)
{
use_operand_p use;
imm_use_iterator use_iter;
gimple *single_use_stmt = NULL;
if (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_reference)
return false;
FOR_EACH_IMM_USE_FAST (use, use_iter, gimple_assign_lhs (stmt))
{
gimple *current_use_stmt = USE_STMT (use);
if (is_gimple_assign (current_use_stmt)
&& TREE_CODE (gimple_assign_lhs (current_use_stmt)) == SSA_NAME)
{
if (single_use_stmt != NULL && single_use_stmt != current_use_stmt)
return false;
single_use_stmt = current_use_stmt;
}
}
if (single_use_stmt == NULL)
return false;
if (gimple_bb (stmt)->loop_father
!= gimple_bb (single_use_stmt)->loop_father)
return false;
return true;
}
/* Rank assigned to a phi statement. If STMT is a loop-carried phi of
an innermost loop, and the phi has only a single use which is inside
the loop, then the rank is the block rank of the loop latch plus an
extra bias for the loop-carried dependence. This causes expressions
calculated into an accumulator variable to be independent for each
iteration of the loop. If STMT is some other phi, the rank is the
block rank of its containing block. */
static int64_t
phi_rank (gimple *stmt)
{
basic_block bb = gimple_bb (stmt);
class loop *father = bb->loop_father;
tree res;
unsigned i;
use_operand_p use;
gimple *use_stmt;
if (!reassoc_bias_loop_carried_phi_ranks_p)
return bb_rank[bb->index];
/* We only care about real loops (those with a latch). */
if (!father->latch)
return bb_rank[bb->index];
/* Interesting phis must be in headers of innermost loops. */
if (bb != father->header
|| father->inner)
return bb_rank[bb->index];
/* Ignore virtual SSA_NAMEs. */
res = gimple_phi_result (stmt);
if (virtual_operand_p (res))
return bb_rank[bb->index];
/* The phi definition must have a single use, and that use must be
within the loop. Otherwise this isn't an accumulator pattern. */
if (!single_imm_use (res, &use, &use_stmt)
|| gimple_bb (use_stmt)->loop_father != father)
return bb_rank[bb->index];
/* Look for phi arguments from within the loop. If found, bias this phi. */
for (i = 0; i < gimple_phi_num_args (stmt); i++)
{
tree arg = gimple_phi_arg_def (stmt, i);
if (TREE_CODE (arg) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (arg))
{
gimple *def_stmt = SSA_NAME_DEF_STMT (arg);
if (gimple_bb (def_stmt)->loop_father == father)
return bb_rank[father->latch->index] + PHI_LOOP_BIAS;
}
}
/* Must be an uninteresting phi. */
return bb_rank[bb->index];
}
/* Return the maximum of RANK and the rank that should be propagated
from expression OP. For most operands, this is just the rank of OP.
For loop-carried phis, the value is zero to avoid undoing the bias
in favor of the phi. */
static int64_t
propagate_rank (int64_t rank, tree op, bool *maybe_biased_p)
{
int64_t op_rank;
op_rank = get_rank (op);
/* Check whether op is biased after the get_rank () call, since it might have
updated biased_names. */
if (TREE_CODE (op) == SSA_NAME
&& bitmap_bit_p (biased_names, SSA_NAME_VERSION (op)))
{
if (maybe_biased_p == NULL)
return rank;
*maybe_biased_p = true;
}
return MAX (rank, op_rank);
}
/* Look up the operand rank structure for expression E. */
static inline int64_t
find_operand_rank (tree e)
{
int64_t *slot = operand_rank->get (e);
return slot ? *slot : -1;
}
/* Insert {E,RANK} into the operand rank hashtable. */
static inline void
insert_operand_rank (tree e, int64_t rank)
{
gcc_assert (rank > 0);
gcc_assert (!operand_rank->put (e, rank));
}
/* Given an expression E, return the rank of the expression. */
static int64_t
get_rank (tree e)
{
/* SSA_NAME's have the rank of the expression they are the result
of.
For globals and uninitialized values, the rank is 0.
For function arguments, use the pre-setup rank.
For PHI nodes, stores, asm statements, etc, we use the rank of
the BB.
For simple operations, the rank is the maximum rank of any of
its operands, or the bb_rank, whichever is less.
I make no claims that this is optimal, however, it gives good
results. */
/* We make an exception to the normal ranking system to break
dependences of accumulator variables in loops. Suppose we
have a simple one-block loop containing:
x_1 = phi(x_0, x_2)
b = a + x_1
c = b + d
x_2 = c + e
As shown, each iteration of the calculation into x is fully
dependent upon the iteration before it. We would prefer to
see this in the form:
x_1 = phi(x_0, x_2)
b = a + d
c = b + e
x_2 = c + x_1
If the loop is unrolled, the calculations of b and c from
different iterations can be interleaved.
To obtain this result during reassociation, we bias the rank
of the phi definition x_1 upward, when it is recognized as an
accumulator pattern. The artificial rank causes it to be
added last, providing the desired independence. */
if (TREE_CODE (e) == SSA_NAME)
{
ssa_op_iter iter;
gimple *stmt;
int64_t rank;
tree op;
/* If we already have a rank for this expression, use that. */
rank = find_operand_rank (e);
if (rank != -1)
return rank;
stmt = SSA_NAME_DEF_STMT (e);
if (gimple_code (stmt) == GIMPLE_PHI)
{
rank = phi_rank (stmt);
if (rank != bb_rank[gimple_bb (stmt)->index])
bitmap_set_bit (biased_names, SSA_NAME_VERSION (e));
}
else if (!is_gimple_assign (stmt))
rank = bb_rank[gimple_bb (stmt)->index];
else
{
bool biased_p = false;
bool *maybe_biased_p = propagate_bias_p (stmt) ? &biased_p : NULL;
/* Otherwise, find the maximum rank for the operands. As an
exception, remove the bias from loop-carried phis when propagating
the rank so that dependent operations are not also biased. */
/* Simply walk over all SSA uses - this takes advatage of the
fact that non-SSA operands are is_gimple_min_invariant and
thus have rank 0. */
rank = 0;
FOR_EACH_SSA_TREE_OPERAND (op, stmt, iter, SSA_OP_USE)
rank = propagate_rank (rank, op, maybe_biased_p);
rank += 1;
if (biased_p)
bitmap_set_bit (biased_names, SSA_NAME_VERSION (e));
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Rank for ");
print_generic_expr (dump_file, e);
fprintf (dump_file, " is %" PRId64 "\n", rank);
}
/* Note the rank in the hashtable so we don't recompute it. */
insert_operand_rank (e, rank);
return rank;
}
/* Constants, globals, etc., are rank 0 */
return 0;
}
/* We want integer ones to end up last no matter what, since they are
the ones we can do the most with. */
#define INTEGER_CONST_TYPE 1 << 4
#define FLOAT_ONE_CONST_TYPE 1 << 3
#define FLOAT_CONST_TYPE 1 << 2
#define OTHER_CONST_TYPE 1 << 1
/* Classify an invariant tree into integer, float, or other, so that
we can sort them to be near other constants of the same type. */
static inline int
constant_type (tree t)
{
if (INTEGRAL_TYPE_P (TREE_TYPE (t)))
return INTEGER_CONST_TYPE;
else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (t)))
{
/* Sort -1.0 and 1.0 constants last, while in some cases
const_binop can't optimize some inexact operations, multiplication
by -1.0 or 1.0 can be always merged with others. */
if (real_onep (t) || real_minus_onep (t))
return FLOAT_ONE_CONST_TYPE;
return FLOAT_CONST_TYPE;
}
else
return OTHER_CONST_TYPE;
}
/* qsort comparison function to sort operand entries PA and PB by rank
so that the sorted array is ordered by rank in decreasing order. */
static int
sort_by_operand_rank (const void *pa, const void *pb)
{
const operand_entry *oea = *(const operand_entry *const *)pa;
const operand_entry *oeb = *(const operand_entry *const *)pb;
if (oeb->rank != oea->rank)
return oeb->rank > oea->rank ? 1 : -1;
/* It's nicer for optimize_expression if constants that are likely
to fold when added/multiplied/whatever are put next to each
other. Since all constants have rank 0, order them by type. */
if (oea->rank == 0)
{
if (constant_type (oeb->op) != constant_type (oea->op))
return constant_type (oea->op) - constant_type (oeb->op);
else
/* To make sorting result stable, we use unique IDs to determine
order. */
return oeb->id > oea->id ? 1 : -1;
}
if (TREE_CODE (oea->op) != SSA_NAME)
{
if (TREE_CODE (oeb->op) != SSA_NAME)
return oeb->id > oea->id ? 1 : -1;
else
return 1;
}
else if (TREE_CODE (oeb->op) != SSA_NAME)
return -1;
/* Lastly, make sure the versions that are the same go next to each
other. */
if (SSA_NAME_VERSION (oeb->op) != SSA_NAME_VERSION (oea->op))
{
/* As SSA_NAME_VERSION is assigned pretty randomly, because we reuse
versions of removed SSA_NAMEs, so if possible, prefer to sort
based on basic block and gimple_uid of the SSA_NAME_DEF_STMT.
See PR60418. */
gimple *stmta = SSA_NAME_DEF_STMT (oea->op);
gimple *stmtb = SSA_NAME_DEF_STMT (oeb->op);
basic_block bba = gimple_bb (stmta);
basic_block bbb = gimple_bb (stmtb);
if (bbb != bba)
{
/* One of the SSA_NAMEs can be defined in oeN->stmt_to_insert
but the other might not. */
if (!bba)
return 1;
if (!bbb)
return -1;
/* If neither is, compare bb_rank. */
if (bb_rank[bbb->index] != bb_rank[bba->index])
return (bb_rank[bbb->index] >> 16) - (bb_rank[bba->index] >> 16);
}
bool da = reassoc_stmt_dominates_stmt_p (stmta, stmtb);
bool db = reassoc_stmt_dominates_stmt_p (stmtb, stmta);
if (da != db)
return da ? 1 : -1;
return SSA_NAME_VERSION (oeb->op) > SSA_NAME_VERSION (oea->op) ? 1 : -1;
}
return oeb->id > oea->id ? 1 : -1;
}
/* Add an operand entry to *OPS for the tree operand OP. */
static void
add_to_ops_vec (vec<operand_entry *> *ops, tree op, gimple *stmt_to_insert = NULL)
{
operand_entry *oe = operand_entry_pool.allocate ();
oe->op = op;
oe->rank = get_rank (op);
oe->id = next_operand_entry_id++;
oe->count = 1;
oe->stmt_to_insert = stmt_to_insert;
ops->safe_push (oe);
}
/* Add an operand entry to *OPS for the tree operand OP with repeat
count REPEAT. */
static void
add_repeat_to_ops_vec (vec<operand_entry *> *ops, tree op,
HOST_WIDE_INT repeat)
{
operand_entry *oe = operand_entry_pool.allocate ();
oe->op = op;
oe->rank = get_rank (op);
oe->id = next_operand_entry_id++;
oe->count = repeat;
oe->stmt_to_insert = NULL;
ops->safe_push (oe);
reassociate_stats.pows_encountered++;
}
/* Returns true if we can associate the SSA def OP. */
static bool
can_reassociate_op_p (tree op)
{
if (TREE_CODE (op) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op))
return false;
/* Make sure asm goto outputs do not participate in reassociation since
we have no way to find an insertion place after asm goto. */
if (TREE_CODE (op) == SSA_NAME
&& gimple_code (SSA_NAME_DEF_STMT (op)) == GIMPLE_ASM
&& gimple_asm_nlabels (as_a <gasm *> (SSA_NAME_DEF_STMT (op))) != 0)
return false;
return true;
}
/* Returns true if we can reassociate operations of TYPE.
That is for integral or non-saturating fixed-point types, and for
floating point type when associative-math is enabled. */
static bool
can_reassociate_type_p (tree type)
{
if ((ANY_INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_WRAPS (type))
|| NON_SAT_FIXED_POINT_TYPE_P (type)
|| (flag_associative_math && FLOAT_TYPE_P (type)))
return true;
return false;
}
/* Return true if STMT is reassociable operation containing a binary
operation with tree code CODE, and is inside LOOP. */
static bool
is_reassociable_op (gimple *stmt, enum tree_code code, class loop *loop)
{
basic_block bb = gimple_bb (stmt);
if (gimple_bb (stmt) == NULL)
return false;
if (!flow_bb_inside_loop_p (loop, bb))
return false;
if (is_gimple_assign (stmt)
&& gimple_assign_rhs_code (stmt) == code
&& has_single_use (gimple_assign_lhs (stmt)))
{
tree rhs1 = gimple_assign_rhs1 (stmt);
tree rhs2 = gimple_assign_rhs2 (stmt);
if (!can_reassociate_op_p (rhs1)
|| (rhs2 && !can_reassociate_op_p (rhs2)))
return false;
return true;
}
return false;
}
/* Return true if STMT is a nop-conversion. */
static bool
gimple_nop_conversion_p (gimple *stmt)
{
if (gassign *ass = dyn_cast <gassign *> (stmt))
{
if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (ass))
&& tree_nop_conversion_p (TREE_TYPE (gimple_assign_lhs (ass)),
TREE_TYPE (gimple_assign_rhs1 (ass))))
return true;
}
return false;
}
/* Given NAME, if NAME is defined by a unary operation OPCODE, return the
operand of the negate operation. Otherwise, return NULL. */
static tree
get_unary_op (tree name, enum tree_code opcode)
{
gimple *stmt = SSA_NAME_DEF_STMT (name);
/* Look through nop conversions (sign changes). */
if (gimple_nop_conversion_p (stmt)
&& TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME)
stmt = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
if (!is_gimple_assign (stmt))
return NULL_TREE;
if (gimple_assign_rhs_code (stmt) == opcode)
return gimple_assign_rhs1 (stmt);
return NULL_TREE;
}
/* Return true if OP1 and OP2 have the same value if casted to either type. */
static bool
ops_equal_values_p (tree op1, tree op2)
{
if (op1 == op2)
return true;
tree orig_op1 = op1;
if (TREE_CODE (op1) == SSA_NAME)
{
gimple *stmt = SSA_NAME_DEF_STMT (op1);
if (gimple_nop_conversion_p (stmt))
{
op1 = gimple_assign_rhs1 (stmt);
if (op1 == op2)
return true;
}
}
if (TREE_CODE (op2) == SSA_NAME)
{
gimple *stmt = SSA_NAME_DEF_STMT (op2);
if (gimple_nop_conversion_p (stmt))
{
op2 = gimple_assign_rhs1 (stmt);
if (op1 == op2
|| orig_op1 == op2)
return true;
}
}
return false;
}
/* If CURR and LAST are a pair of ops that OPCODE allows us to
eliminate through equivalences, do so, remove them from OPS, and
return true. Otherwise, return false. */
static bool
eliminate_duplicate_pair (enum tree_code opcode,
vec<operand_entry *> *ops,
bool *all_done,
unsigned int i,
operand_entry *curr,
operand_entry *last)
{
/* If we have two of the same op, and the opcode is & |, min, or max,
we can eliminate one of them.
If we have two of the same op, and the opcode is ^, we can
eliminate both of them. */
if (last && last->op == curr->op)
{
switch (opcode)
{
case MAX_EXPR:
case MIN_EXPR:
case BIT_IOR_EXPR:
case BIT_AND_EXPR:
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, curr->op);
fprintf (dump_file, " [&|minmax] ");
print_generic_expr (dump_file, last->op);
fprintf (dump_file, " -> ");
print_generic_stmt (dump_file, last->op);
}
ops->ordered_remove (i);
reassociate_stats.ops_eliminated ++;
return true;
case BIT_XOR_EXPR:
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, curr->op);
fprintf (dump_file, " ^ ");
print_generic_expr (dump_file, last->op);
fprintf (dump_file, " -> nothing\n");
}
reassociate_stats.ops_eliminated += 2;
if (ops->length () == 2)
{
ops->truncate (0);
add_to_ops_vec (ops, build_zero_cst (TREE_TYPE (last->op)));
*all_done = true;
}
else
{
ops->ordered_remove (i-1);
ops->ordered_remove (i-1);
}
return true;
default:
break;
}
}
return false;
}
static vec<tree> plus_negates;
/* If OPCODE is PLUS_EXPR, CURR->OP is a negate expression or a bitwise not
expression, look in OPS for a corresponding positive operation to cancel
it out. If we find one, remove the other from OPS, replace
OPS[CURRINDEX] with 0 or -1, respectively, and return true. Otherwise,
return false. */
static bool
eliminate_plus_minus_pair (enum tree_code opcode,
vec<operand_entry *> *ops,
unsigned int currindex,
operand_entry *curr)
{
tree negateop;
tree notop;
unsigned int i;
operand_entry *oe;
if (opcode != PLUS_EXPR || TREE_CODE (curr->op) != SSA_NAME)
return false;
negateop = get_unary_op (curr->op, NEGATE_EXPR);
notop = get_unary_op (curr->op, BIT_NOT_EXPR);
if (negateop == NULL_TREE && notop == NULL_TREE)
return false;
/* Any non-negated version will have a rank that is one less than
the current rank. So once we hit those ranks, if we don't find
one, we can stop. */
for (i = currindex + 1;
ops->iterate (i, &oe)
&& oe->rank >= curr->rank - 1 ;
i++)
{
if (negateop
&& ops_equal_values_p (oe->op, negateop))
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, negateop);
fprintf (dump_file, " + -");
print_generic_expr (dump_file, oe->op);
fprintf (dump_file, " -> 0\n");
}
ops->ordered_remove (i);
add_to_ops_vec (ops, build_zero_cst (TREE_TYPE (oe->op)));
ops->ordered_remove (currindex);
reassociate_stats.ops_eliminated ++;
return true;
}
else if (notop
&& ops_equal_values_p (oe->op, notop))
{
tree op_type = TREE_TYPE (oe->op);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, notop);
fprintf (dump_file, " + ~");
print_generic_expr (dump_file, oe->op);
fprintf (dump_file, " -> -1\n");
}
ops->ordered_remove (i);
add_to_ops_vec (ops, build_all_ones_cst (op_type));
ops->ordered_remove (currindex);
reassociate_stats.ops_eliminated ++;
return true;
}
}
/* If CURR->OP is a negate expr without nop conversion in a plus expr:
save it for later inspection in repropagate_negates(). */
if (negateop != NULL_TREE
&& gimple_assign_rhs_code (SSA_NAME_DEF_STMT (curr->op)) == NEGATE_EXPR)
plus_negates.safe_push (curr->op);
return false;
}
/* If OPCODE is BIT_IOR_EXPR, BIT_AND_EXPR, and, CURR->OP is really a
bitwise not expression, look in OPS for a corresponding operand to
cancel it out. If we find one, remove the other from OPS, replace
OPS[CURRINDEX] with 0, and return true. Otherwise, return
false. */
static bool
eliminate_not_pairs (enum tree_code opcode,
vec<operand_entry *> *ops,
unsigned int currindex,
operand_entry *curr)
{
tree notop;
unsigned int i;
operand_entry *oe;
if ((opcode != BIT_IOR_EXPR && opcode != BIT_AND_EXPR)
|| TREE_CODE (curr->op) != SSA_NAME)
return false;
notop = get_unary_op (curr->op, BIT_NOT_EXPR);
if (notop == NULL_TREE)
return false;
/* Any non-not version will have a rank that is one less than
the current rank. So once we hit those ranks, if we don't find
one, we can stop. */
for (i = currindex + 1;
ops->iterate (i, &oe)
&& oe->rank >= curr->rank - 1;
i++)
{
if (oe->op == notop)
{
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, notop);
if (opcode == BIT_AND_EXPR)
fprintf (dump_file, " & ~");
else if (opcode == BIT_IOR_EXPR)
fprintf (dump_file, " | ~");
print_generic_expr (dump_file, oe->op);
if (opcode == BIT_AND_EXPR)
fprintf (dump_file, " -> 0\n");
else if (opcode == BIT_IOR_EXPR)
fprintf (dump_file, " -> -1\n");
}
if (opcode == BIT_AND_EXPR)
oe->op = build_zero_cst (TREE_TYPE (oe->op));
else if (opcode == BIT_IOR_EXPR)
oe->op = build_all_ones_cst (TREE_TYPE (oe->op));
reassociate_stats.ops_eliminated += ops->length () - 1;
ops->truncate (0);
ops->quick_push (oe);
return true;
}
}
return false;
}
/* Use constant value that may be present in OPS to try to eliminate
operands. Note that this function is only really used when we've
eliminated ops for other reasons, or merged constants. Across
single statements, fold already does all of this, plus more. There
is little point in duplicating logic, so I've only included the
identities that I could ever construct testcases to trigger. */
static void
eliminate_using_constants (enum tree_code opcode,
vec<operand_entry *> *ops)
{
operand_entry *oelast = ops->last ();
tree type = TREE_TYPE (oelast->op);
if (oelast->rank == 0
&& (ANY_INTEGRAL_TYPE_P (type) || FLOAT_TYPE_P (type)))
{
switch (opcode)
{
case BIT_AND_EXPR:
if (integer_zerop (oelast->op))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found & 0, removing all other ops\n");
reassociate_stats.ops_eliminated += ops->length () - 1;
ops->truncate (0);
ops->quick_push (oelast);
return;
}
}
else if (integer_all_onesp (oelast->op))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found & -1, removing\n");
ops->pop ();
reassociate_stats.ops_eliminated++;
}
}
break;
case BIT_IOR_EXPR:
if (integer_all_onesp (oelast->op))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found | -1, removing all other ops\n");
reassociate_stats.ops_eliminated += ops->length () - 1;
ops->truncate (0);
ops->quick_push (oelast);
return;
}
}
else if (integer_zerop (oelast->op))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found | 0, removing\n");
ops->pop ();
reassociate_stats.ops_eliminated++;
}
}
break;
case MULT_EXPR:
if (integer_zerop (oelast->op)
|| (FLOAT_TYPE_P (type)
&& !HONOR_NANS (type)
&& !HONOR_SIGNED_ZEROS (type)
&& real_zerop (oelast->op)))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found * 0, removing all other ops\n");
reassociate_stats.ops_eliminated += ops->length () - 1;
ops->truncate (0);
ops->quick_push (oelast);
return;
}
}
else if (integer_onep (oelast->op)
|| (FLOAT_TYPE_P (type)
&& !HONOR_SNANS (type)
&& real_onep (oelast->op)))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found * 1, removing\n");
ops->pop ();
reassociate_stats.ops_eliminated++;
return;
}
}
break;
case BIT_XOR_EXPR:
case PLUS_EXPR:
case MINUS_EXPR:
if (integer_zerop (oelast->op)
|| (FLOAT_TYPE_P (type)
&& (opcode == PLUS_EXPR || opcode == MINUS_EXPR)
&& fold_real_zero_addition_p (type, 0, oelast->op,
opcode == MINUS_EXPR)))
{
if (ops->length () != 1)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Found [|^+] 0, removing\n");
ops->pop ();
reassociate_stats.ops_eliminated++;
return;
}
}
break;
default:
break;
}
}
}
static void linearize_expr_tree (vec<operand_entry *> *, gimple *,
bool, bool);
/* Structure for tracking and counting operands. */
struct oecount {
unsigned int cnt;
unsigned int id;
enum tree_code oecode;
tree op;
};
/* The heap for the oecount hashtable and the sorted list of operands. */
static vec<oecount> cvec;
/* Oecount hashtable helpers. */
struct oecount_hasher : int_hash <int, 0, 1>
{
static inline hashval_t hash (int);
static inline bool equal (int, int);
};
/* Hash function for oecount. */
inline hashval_t
oecount_hasher::hash (int p)
{
const oecount *c = &cvec[p - 42];
return htab_hash_pointer (c->op) ^ (hashval_t)c->oecode;
}
/* Comparison function for oecount. */
inline bool
oecount_hasher::equal (int p1, int p2)
{
const oecount *c1 = &cvec[p1 - 42];
const oecount *c2 = &cvec[p2 - 42];
return c1->oecode == c2->oecode && c1->op == c2->op;
}
/* Comparison function for qsort sorting oecount elements by count. */
static int
oecount_cmp (const void *p1, const void *p2)
{
const oecount *c1 = (const oecount *)p1;
const oecount *c2 = (const oecount *)p2;
if (c1->cnt != c2->cnt)
return c1->cnt > c2->cnt ? 1 : -1;
else
/* If counts are identical, use unique IDs to stabilize qsort. */
return c1->id > c2->id ? 1 : -1;
}
/* Return TRUE iff STMT represents a builtin call that raises OP
to some exponent. */
static bool
stmt_is_power_of_op (gimple *stmt, tree op)
{
if (!is_gimple_call (stmt))
return false;
switch (gimple_call_combined_fn (stmt))
{
CASE_CFN_POW:
CASE_CFN_POWI:
return (operand_equal_p (gimple_call_arg (stmt, 0), op, 0));
default:
return false;
}
}
/* Given STMT which is a __builtin_pow* call, decrement its exponent
in place and return the result. Assumes that stmt_is_power_of_op
was previously called for STMT and returned TRUE. */
static HOST_WIDE_INT
decrement_power (gimple *stmt)
{
REAL_VALUE_TYPE c, cint;
HOST_WIDE_INT power;
tree arg1;
switch (gimple_call_combined_fn (stmt))
{
CASE_CFN_POW:
arg1 = gimple_call_arg (stmt, 1);
c = TREE_REAL_CST (arg1);
power = real_to_integer (&c) - 1;
real_from_integer (&cint, VOIDmode, power, SIGNED);
gimple_call_set_arg (stmt, 1, build_real (TREE_TYPE (arg1), cint));
return power;
CASE_CFN_POWI:
arg1 = gimple_call_arg (stmt, 1);
power = TREE_INT_CST_LOW (arg1) - 1;
gimple_call_set_arg (stmt, 1, build_int_cst (TREE_TYPE (arg1), power));
return power;
default:
gcc_unreachable ();
}
}
/* Replace SSA defined by STMT and replace all its uses with new
SSA. Also return the new SSA. */
static tree
make_new_ssa_for_def (gimple *stmt, enum tree_code opcode, tree op)
{
gimple *use_stmt;
use_operand_p use;
imm_use_iterator iter;
tree new_lhs, new_debug_lhs = NULL_TREE;
tree lhs = gimple_get_lhs (stmt);
new_lhs = make_ssa_name (TREE_TYPE (lhs));
gimple_set_lhs (stmt, new_lhs);
/* Also need to update GIMPLE_DEBUGs. */
FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
{
tree repl = new_lhs;
if (is_gimple_debug (use_stmt))
{
if (new_debug_lhs == NULL_TREE)
{
new_debug_lhs = make_node (DEBUG_EXPR_DECL);
gdebug *def_temp
= gimple_build_debug_bind (new_debug_lhs,
build2 (opcode, TREE_TYPE (lhs),
new_lhs, op),
stmt);
DECL_ARTIFICIAL (new_debug_lhs) = 1;
TREE_TYPE (new_debug_lhs) = TREE_TYPE (lhs);
SET_DECL_MODE (new_debug_lhs, TYPE_MODE (TREE_TYPE (lhs)));
gimple_set_uid (def_temp, gimple_uid (stmt));
gimple_stmt_iterator gsi = gsi_for_stmt (stmt);
gsi_insert_after (&gsi, def_temp, GSI_SAME_STMT);
}
repl = new_debug_lhs;
}
FOR_EACH_IMM_USE_ON_STMT (use, iter)
SET_USE (use, repl);
update_stmt (use_stmt);
}
return new_lhs;
}
/* Replace all SSAs defined in STMTS_TO_FIX and replace its
uses with new SSAs. Also do this for the stmt that defines DEF
if *DEF is not OP. */
static void
make_new_ssa_for_all_defs (tree *def, enum tree_code opcode, tree op,
vec<gimple *> &stmts_to_fix)
{
unsigned i;
gimple *stmt;
if (*def != op
&& TREE_CODE (*def) == SSA_NAME
&& (stmt = SSA_NAME_DEF_STMT (*def))
&& gimple_code (stmt) != GIMPLE_NOP)
*def = make_new_ssa_for_def (stmt, opcode, op);
FOR_EACH_VEC_ELT (stmts_to_fix, i, stmt)
make_new_ssa_for_def (stmt, opcode, op);
}
/* Find the single immediate use of STMT's LHS, and replace it
with OP. Remove STMT. If STMT's LHS is the same as *DEF,
replace *DEF with OP as well. */
static void
propagate_op_to_single_use (tree op, gimple *stmt, tree *def)
{
tree lhs;
gimple *use_stmt;
use_operand_p use;
gimple_stmt_iterator gsi;
if (is_gimple_call (stmt))
lhs = gimple_call_lhs (stmt);
else
lhs = gimple_assign_lhs (stmt);
gcc_assert (has_single_use (lhs));
single_imm_use (lhs, &use, &use_stmt);
if (lhs == *def)
*def = op;
SET_USE (use, op);
if (TREE_CODE (op) != SSA_NAME)
update_stmt (use_stmt);
gsi = gsi_for_stmt (stmt);
unlink_stmt_vdef (stmt);
reassoc_remove_stmt (&gsi);
release_defs (stmt);
}
/* Walks the linear chain with result *DEF searching for an operation
with operand OP and code OPCODE removing that from the chain. *DEF
is updated if there is only one operand but no operation left. */
static void
zero_one_operation (tree *def, enum tree_code opcode, tree op)
{
tree orig_def = *def;
gimple *stmt = SSA_NAME_DEF_STMT (*def);
/* PR72835 - Record the stmt chain that has to be updated such that
we dont use the same LHS when the values computed are different. */
auto_vec<gimple *, 64> stmts_to_fix;
do
{
tree name;
if (opcode == MULT_EXPR)
{
if (stmt_is_power_of_op (stmt, op))
{
if (decrement_power (stmt) == 1)
{
if (stmts_to_fix.length () > 0)
stmts_to_fix.pop ();
propagate_op_to_single_use (op, stmt, def);
}
break;
}
else if (gimple_assign_rhs_code (stmt) == NEGATE_EXPR)
{
if (gimple_assign_rhs1 (stmt) == op)
{
tree cst = build_minus_one_cst (TREE_TYPE (op));
if (stmts_to_fix.length () > 0)
stmts_to_fix.pop ();
propagate_op_to_single_use (cst, stmt, def);
break;
}
else if (integer_minus_onep (op)
|| real_minus_onep (op))
{
gimple_assign_set_rhs_code
(stmt, TREE_CODE (gimple_assign_rhs1 (stmt)));
break;
}
}
}
name = gimple_assign_rhs1 (stmt);
/* If this is the operation we look for and one of the operands
is ours simply propagate the other operand into the stmts
single use. */
if (gimple_assign_rhs_code (stmt) == opcode
&& (name == op
|| gimple_assign_rhs2 (stmt) == op))
{
if (name == op)
name = gimple_assign_rhs2 (stmt);
if (stmts_to_fix.length () > 0)
stmts_to_fix.pop ();
propagate_op_to_single_use (name, stmt, def);
break;
}
/* We might have a multiply of two __builtin_pow* calls, and
the operand might be hiding in the rightmost one. Likewise
this can happen for a negate. */
if (opcode == MULT_EXPR
&& gimple_assign_rhs_code (stmt) == opcode
&& TREE_CODE (gimple_assign_rhs2 (stmt)) == SSA_NAME
&& has_single_use (gimple_assign_rhs2 (stmt)))
{
gimple *stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt));
if (stmt_is_power_of_op (stmt2, op))
{
if (decrement_power (stmt2) == 1)
propagate_op_to_single_use (op, stmt2, def);
else
stmts_to_fix.safe_push (stmt2);
break;
}
else if (is_gimple_assign (stmt2)
&& gimple_assign_rhs_code (stmt2) == NEGATE_EXPR)
{
if (gimple_assign_rhs1 (stmt2) == op)
{
tree cst = build_minus_one_cst (TREE_TYPE (op));
propagate_op_to_single_use (cst, stmt2, def);
break;
}
else if (integer_minus_onep (op)
|| real_minus_onep (op))
{
stmts_to_fix.safe_push (stmt2);
gimple_assign_set_rhs_code
(stmt2, TREE_CODE (gimple_assign_rhs1 (stmt2)));
break;
}
}
}
/* Continue walking the chain. */
gcc_assert (name != op
&& TREE_CODE (name) == SSA_NAME);
stmt = SSA_NAME_DEF_STMT (name);
stmts_to_fix.safe_push (stmt);
}
while (1);
if (stmts_to_fix.length () > 0 || *def == orig_def)
make_new_ssa_for_all_defs (def, opcode, op, stmts_to_fix);
}
/* Returns true if statement S1 dominates statement S2. Like
stmt_dominates_stmt_p, but uses stmt UIDs to optimize. */
static bool
reassoc_stmt_dominates_stmt_p (gimple *s1, gimple *s2)
{
basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2);
/* If bb1 is NULL, it should be a GIMPLE_NOP def stmt of an (D)
SSA_NAME. Assume it lives at the beginning of function and
thus dominates everything. */
if (!bb1 || s1 == s2)
return true;
/* If bb2 is NULL, it doesn't dominate any stmt with a bb. */
if (!bb2)
return false;
if (bb1 == bb2)
{
/* PHIs in the same basic block are assumed to be
executed all in parallel, if only one stmt is a PHI,
it dominates the other stmt in the same basic block. */
if (gimple_code (s1) == GIMPLE_PHI)
return true;
if (gimple_code (s2) == GIMPLE_PHI)
return false;
gcc_assert (gimple_uid (s1) && gimple_uid (s2));
if (gimple_uid (s1) < gimple_uid (s2))
return true;
if (gimple_uid (s1) > gimple_uid (s2))
return false;
gimple_stmt_iterator gsi = gsi_for_stmt (s1);
unsigned int uid = gimple_uid (s1);
for (gsi_next (&gsi); !gsi_end_p (gsi); gsi_next (&gsi))
{
gimple *s = gsi_stmt (gsi);
if (gimple_uid (s) != uid)
break;
if (s == s2)
return true;
}
return false;
}
return dominated_by_p (CDI_DOMINATORS, bb2, bb1);
}
/* Insert STMT after INSERT_POINT. */
static void
insert_stmt_after (gimple *stmt, gimple *insert_point)
{
gimple_stmt_iterator gsi;
basic_block bb;
if (gimple_code (insert_point) == GIMPLE_PHI)
bb = gimple_bb (insert_point);
else if (!stmt_ends_bb_p (insert_point))
{
gsi = gsi_for_stmt (insert_point);
gimple_set_uid (stmt, gimple_uid (insert_point));
gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
return;
}
else if (gimple_code (insert_point) == GIMPLE_ASM)
/* We have no idea where to insert - it depends on where the
uses will be placed. */
gcc_unreachable ();
else
/* We assume INSERT_POINT is a SSA_NAME_DEF_STMT of some SSA_NAME,
thus if it must end a basic block, it should be a call that can
throw, or some assignment that can throw. If it throws, the LHS
of it will not be initialized though, so only valid places using
the SSA_NAME should be dominated by the fallthru edge. */
bb = find_fallthru_edge (gimple_bb (insert_point)->succs)->dest;
gsi = gsi_after_labels (bb);
if (gsi_end_p (gsi))
{
gimple_stmt_iterator gsi2 = gsi_last_bb (bb);
gimple_set_uid (stmt,
gsi_end_p (gsi2) ? 1 : gimple_uid (gsi_stmt (gsi2)));
}
else
gimple_set_uid (stmt, gimple_uid (gsi_stmt (gsi)));
gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
}
/* Builds one statement performing OP1 OPCODE OP2 using TMPVAR for
the result. Places the statement after the definition of either
OP1 or OP2. Returns the new statement. */
static gimple *
build_and_add_sum (tree type, tree op1, tree op2, enum tree_code opcode)
{
gimple *op1def = NULL, *op2def = NULL;
gimple_stmt_iterator gsi;
tree op;
gassign *sum;
/* Create the addition statement. */
op = make_ssa_name (type);
sum = gimple_build_assign (op, opcode, op1, op2);
/* Find an insertion place and insert. */
if (TREE_CODE (op1) == SSA_NAME)
op1def = SSA_NAME_DEF_STMT (op1);
if (TREE_CODE (op2) == SSA_NAME)
op2def = SSA_NAME_DEF_STMT (op2);
if ((!op1def || gimple_nop_p (op1def))
&& (!op2def || gimple_nop_p (op2def)))
{
gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
if (gsi_end_p (gsi))
{
gimple_stmt_iterator gsi2
= gsi_last_bb (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
gimple_set_uid (sum,
gsi_end_p (gsi2) ? 1 : gimple_uid (gsi_stmt (gsi2)));
}
else
gimple_set_uid (sum, gimple_uid (gsi_stmt (gsi)));
gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
}
else
{
gimple *insert_point;
if ((!op1def || gimple_nop_p (op1def))
|| (op2def && !gimple_nop_p (op2def)
&& reassoc_stmt_dominates_stmt_p (op1def, op2def)))
insert_point = op2def;
else
insert_point = op1def;
insert_stmt_after (sum, insert_point);
}
update_stmt (sum);
return sum;
}
/* Perform un-distribution of divisions and multiplications.
A * X + B * X is transformed into (A + B) * X and A / X + B / X
to (A + B) / X for real X.
The algorithm is organized as follows.
- First we walk the addition chain *OPS looking for summands that
are defined by a multiplication or a real division. This results
in the candidates bitmap with relevant indices into *OPS.
- Second we build the chains of multiplications or divisions for
these candidates, counting the number of occurrences of (operand, code)
pairs in all of the candidates chains.
- Third we sort the (operand, code) pairs by number of occurrence and
process them starting with the pair with the most uses.
* For each such pair we walk the candidates again to build a
second candidate bitmap noting all multiplication/division chains
that have at least one occurrence of (operand, code).
* We build an alternate addition chain only covering these
candidates with one (operand, code) operation removed from their
multiplication/division chain.
* The first candidate gets replaced by the alternate addition chain
multiplied/divided by the operand.
* All candidate chains get disabled for further processing and
processing of (operand, code) pairs continues.
The alternate addition chains built are re-processed by the main
reassociation algorithm which allows optimizing a * x * y + b * y * x
to (a + b ) * x * y in one invocation of the reassociation pass. */
static bool
undistribute_ops_list (enum tree_code opcode,
vec<operand_entry *> *ops, class loop *loop)
{
unsigned int length = ops->length ();
operand_entry *oe1;
unsigned i, j;
unsigned nr_candidates, nr_candidates2;
sbitmap_iterator sbi0;
vec<operand_entry *> *subops;
bool changed = false;
unsigned int next_oecount_id = 0;
if (length <= 1
|| opcode != PLUS_EXPR)
return false;
/* Build a list of candidates to process. */
auto_sbitmap candidates (length);
bitmap_clear (candidates);
nr_candidates = 0;
FOR_EACH_VEC_ELT (*ops, i, oe1)
{
enum tree_code dcode;
gimple *oe1def;
if (TREE_CODE (oe1->op) != SSA_NAME)
continue;
oe1def = SSA_NAME_DEF_STMT (oe1->op);
if (!is_gimple_assign (oe1def))
continue;
dcode = gimple_assign_rhs_code (oe1def);
if ((dcode != MULT_EXPR
&& dcode != RDIV_EXPR)
|| !is_reassociable_op (oe1def, dcode, loop))
continue;
bitmap_set_bit (candidates, i);
nr_candidates++;
}
if (nr_candidates < 2)
return false;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "searching for un-distribute opportunities ");
print_generic_expr (dump_file,
(*ops)[bitmap_first_set_bit (candidates)]->op, TDF_NONE);
fprintf (dump_file, " %d\n", nr_candidates);
}
/* Build linearized sub-operand lists and the counting table. */
cvec.create (0);
hash_table<oecount_hasher> ctable (15);
/* ??? Macro arguments cannot have multi-argument template types in
them. This typedef is needed to workaround that limitation. */
typedef vec<operand_entry *> vec_operand_entry_t_heap;
subops = XCNEWVEC (vec_operand_entry_t_heap, ops->length ());
EXECUTE_IF_SET_IN_BITMAP (candidates, 0, i, sbi0)
{
gimple *oedef;
enum tree_code oecode;
unsigned j;
oedef = SSA_NAME_DEF_STMT ((*ops)[i]->op);
oecode = gimple_assign_rhs_code (oedef);
linearize_expr_tree (&subops[i], oedef,
associative_tree_code (oecode), false);
FOR_EACH_VEC_ELT (subops[i], j, oe1)
{
oecount c;
int *slot;
int idx;
c.oecode = oecode;
c.cnt = 1;
c.id = next_oecount_id++;
c.op = oe1->op;
cvec.safe_push (c);
idx = cvec.length () + 41;
slot = ctable.find_slot (idx, INSERT);
if (!*slot)
{
*slot = idx;
}
else
{
cvec.pop ();
cvec[*slot - 42].cnt++;
}
}
}
/* Sort the counting table. */
cvec.qsort (oecount_cmp);
if (dump_file && (dump_flags & TDF_DETAILS))
{
oecount *c;
fprintf (dump_file, "Candidates:\n");
FOR_EACH_VEC_ELT (cvec, j, c)
{
fprintf (dump_file, " %u %s: ", c->cnt,
c->oecode == MULT_EXPR
? "*" : c->oecode == RDIV_EXPR ? "/" : "?");
print_generic_expr (dump_file, c->op);
fprintf (dump_file, "\n");
}
}
/* Process the (operand, code) pairs in order of most occurrence. */
auto_sbitmap candidates2 (length);
while (!cvec.is_empty ())
{
oecount *c = &cvec.last ();
if (c->cnt < 2)
break;
/* Now collect the operands in the outer chain that contain
the common operand in their inner chain. */
bitmap_clear (candidates2);
nr_candidates2 = 0;
EXECUTE_IF_SET_IN_BITMAP (candidates, 0, i, sbi0)
{
gimple *oedef;
enum tree_code oecode;
unsigned j;
tree op = (*ops)[i]->op;
/* If we undistributed in this chain already this may be
a constant. */
if (TREE_CODE (op) != SSA_NAME)
continue;
oedef = SSA_NAME_DEF_STMT (op);
oecode = gimple_assign_rhs_code (oedef);
if (oecode != c->oecode)
continue;
FOR_EACH_VEC_ELT (subops[i], j, oe1)
{
if (oe1->op == c->op)
{
bitmap_set_bit (candidates2, i);
++nr_candidates2;
break;
}
}
}
if (nr_candidates2 >= 2)
{
operand_entry *oe1, *oe2;
gimple *prod;
int first = bitmap_first_set_bit (candidates2);
/* Build the new addition chain. */
oe1 = (*ops)[first];
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Building (");
print_generic_expr (dump_file, oe1->op);
}
zero_one_operation (&oe1->op, c->oecode, c->op);
EXECUTE_IF_SET_IN_BITMAP (candidates2, first+1, i, sbi0)
{
gimple *sum;
oe2 = (*ops)[i];
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " + ");
print_generic_expr (dump_file, oe2->op);
}
zero_one_operation (&oe2->op, c->oecode, c->op);
sum = build_and_add_sum (TREE_TYPE (oe1->op),
oe1->op, oe2->op, opcode);
oe2->op = build_zero_cst (TREE_TYPE (oe2->op));
oe2->rank = 0;
oe1->op = gimple_get_lhs (sum);
}
/* Apply the multiplication/division. */
prod = build_and_add_sum (TREE_TYPE (oe1->op),
oe1->op, c->op, c->oecode);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, ") %s ", c->oecode == MULT_EXPR ? "*" : "/");
print_generic_expr (dump_file, c->op);
fprintf (dump_file, "\n");
}
/* Record it in the addition chain and disable further
undistribution with this op. */
oe1->op = gimple_assign_lhs (prod);
oe1->rank = get_rank (oe1->op);
subops[first].release ();
changed = true;
}
cvec.pop ();
}
for (i = 0; i < ops->length (); ++i)
subops[i].release ();
free (subops);
cvec.release ();
return changed;
}
/* Pair to hold the information of one specific VECTOR_TYPE SSA_NAME:
first: element index for each relevant BIT_FIELD_REF.
second: the index of vec ops* for each relevant BIT_FIELD_REF. */
typedef std::pair<unsigned, unsigned> v_info_elem;
struct v_info {
tree vec_type;
auto_vec<v_info_elem, 32> vec;
};
typedef v_info *v_info_ptr;
/* Comparison function for qsort on VECTOR SSA_NAME trees by machine mode. */
static int
sort_by_mach_mode (const void *p_i, const void *p_j)
{
const tree tr1 = *((const tree *) p_i);
const tree tr2 = *((const tree *) p_j);
unsigned int mode1 = TYPE_MODE (TREE_TYPE (tr1));
unsigned int mode2 = TYPE_MODE (TREE_TYPE (tr2));
if (mode1 > mode2)
return 1;
else if (mode1 < mode2)
return -1;
if (SSA_NAME_VERSION (tr1) < SSA_NAME_VERSION (tr2))
return -1;
else if (SSA_NAME_VERSION (tr1) > SSA_NAME_VERSION (tr2))
return 1;
return 0;
}
/* Cleanup hash map for VECTOR information. */
static void
cleanup_vinfo_map (hash_map<tree, v_info_ptr> &info_map)
{
for (hash_map<tree, v_info_ptr>::iterator it = info_map.begin ();
it != info_map.end (); ++it)
{
v_info_ptr info = (*it).second;
delete info;
(*it).second = NULL;
}
}
/* Perform un-distribution of BIT_FIELD_REF on VECTOR_TYPE.
V1[0] + V1[1] + ... + V1[k] + V2[0] + V2[1] + ... + V2[k] + ... Vn[k]
is transformed to
Vs = (V1 + V2 + ... + Vn)
Vs[0] + Vs[1] + ... + Vs[k]
The basic steps are listed below:
1) Check the addition chain *OPS by looking those summands coming from
VECTOR bit_field_ref on VECTOR type. Put the information into
v_info_map for each satisfied summand, using VECTOR SSA_NAME as key.
2) For each key (VECTOR SSA_NAME), validate all its BIT_FIELD_REFs are
continuous, they can cover the whole VECTOR perfectly without any holes.
Obtain one VECTOR list which contain candidates to be transformed.
3) Sort the VECTOR list by machine mode of VECTOR type, for each group of
candidates with same mode, build the addition statements for them and
generate BIT_FIELD_REFs accordingly.
TODO:
The current implementation requires the whole VECTORs should be fully
covered, but it can be extended to support partial, checking adjacent
but not fill the whole, it may need some cost model to define the
boundary to do or not.
*/
static bool
undistribute_bitref_for_vector (enum tree_code opcode,
vec<operand_entry *> *ops, struct loop *loop)
{
if (ops->length () <= 1)
return false;
if (opcode != PLUS_EXPR
&& opcode != MULT_EXPR
&& opcode != BIT_XOR_EXPR
&& opcode != BIT_IOR_EXPR
&& opcode != BIT_AND_EXPR)
return false;
hash_map<tree, v_info_ptr> v_info_map;
operand_entry *oe1;
unsigned i;
/* Find those summands from VECTOR BIT_FIELD_REF in addition chain, put the
information into map. */
FOR_EACH_VEC_ELT (*ops, i, oe1)
{
enum tree_code dcode;
gimple *oe1def;
if (TREE_CODE (oe1->op) != SSA_NAME)
continue;
oe1def = SSA_NAME_DEF_STMT (oe1->op);
if (!is_gimple_assign (oe1def))
continue;
dcode = gimple_assign_rhs_code (oe1def);
if (dcode != BIT_FIELD_REF || !is_reassociable_op (oe1def, dcode, loop))
continue;
tree rhs = gimple_assign_rhs1 (oe1def);
tree vec = TREE_OPERAND (rhs, 0);
tree vec_type = TREE_TYPE (vec);
if (TREE_CODE (vec) != SSA_NAME || !VECTOR_TYPE_P (vec_type))
continue;
/* Ignore it if target machine can't support this VECTOR type. */
if (!VECTOR_MODE_P (TYPE_MODE (vec_type)))
continue;
/* Check const vector type, constrain BIT_FIELD_REF offset and size. */
if (!TYPE_VECTOR_SUBPARTS (vec_type).is_constant ())
continue;
if (VECTOR_TYPE_P (TREE_TYPE (rhs))
|| !is_a <scalar_mode> (TYPE_MODE (TREE_TYPE (rhs))))
continue;
/* The type of BIT_FIELD_REF might not be equal to the element type of
the vector. We want to use a vector type with element type the
same as the BIT_FIELD_REF and size the same as TREE_TYPE (vec). */
if (!useless_type_conversion_p (TREE_TYPE (rhs), TREE_TYPE (vec_type)))
{
machine_mode simd_mode;
unsigned HOST_WIDE_INT size, nunits;
unsigned HOST_WIDE_INT elem_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (rhs)));
if (!GET_MODE_BITSIZE (TYPE_MODE (vec_type)).is_constant (&size))
continue;
if (size <= elem_size || (size % elem_size) != 0)
continue;
nunits = size / elem_size;
if (!mode_for_vector (SCALAR_TYPE_MODE (TREE_TYPE (rhs)),
nunits).exists (&simd_mode))
continue;
vec_type = build_vector_type_for_mode (TREE_TYPE (rhs), simd_mode);
/* Ignore it if target machine can't support this VECTOR type. */
if (!VECTOR_MODE_P (TYPE_MODE (vec_type)))
continue;
/* Check const vector type, constrain BIT_FIELD_REF offset and
size. */
if (!TYPE_VECTOR_SUBPARTS (vec_type).is_constant ())
continue;
if (maybe_ne (GET_MODE_SIZE (TYPE_MODE (vec_type)),
GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (vec)))))
continue;
}
tree elem_type = TREE_TYPE (vec_type);
unsigned HOST_WIDE_INT elem_size = tree_to_uhwi (TYPE_SIZE (elem_type));
if (maybe_ne (bit_field_size (rhs), elem_size))
continue;
unsigned idx;
if (!constant_multiple_p (bit_field_offset (rhs), elem_size, &idx))
continue;
/* Ignore it if target machine can't support this type of VECTOR
operation. */
optab op_tab = optab_for_tree_code (opcode, vec_type, optab_vector);
if (optab_handler (op_tab, TYPE_MODE (vec_type)) == CODE_FOR_nothing)
continue;
bool existed;
v_info_ptr &info = v_info_map.get_or_insert (vec, &existed);
if (!existed)
{
info = new v_info;
info->vec_type = vec_type;
}
else if (!types_compatible_p (vec_type, info->vec_type))
continue;
info->vec.safe_push (std::make_pair (idx, i));
}
/* At least two VECTOR to combine. */
if (v_info_map.elements () <= 1)
{
cleanup_vinfo_map (v_info_map);
return false;
}
/* Verify all VECTOR candidates by checking two conditions:
1) sorted offsets are adjacent, no holes.
2) can fill the whole VECTOR perfectly.
And add the valid candidates to a vector for further handling. */
auto_vec<tree> valid_vecs (v_info_map.elements ());
for (hash_map<tree, v_info_ptr>::iterator it = v_info_map.begin ();
it != v_info_map.end (); ++it)
{
tree cand_vec = (*it).first;
v_info_ptr cand_info = (*it).second;
unsigned int num_elems
= TYPE_VECTOR_SUBPARTS (cand_info->vec_type).to_constant ();
if (cand_info->vec.length () != num_elems)
continue;
sbitmap holes = sbitmap_alloc (num_elems);
bitmap_ones (holes);
bool valid = true;
v_info_elem *curr;
FOR_EACH_VEC_ELT (cand_info->vec, i, curr)
{
if (!bitmap_bit_p (holes, curr->first))
{
valid = false;
break;
}
else
bitmap_clear_bit (holes, curr->first);
}
if (valid && bitmap_empty_p (holes))
valid_vecs.quick_push (cand_vec);
sbitmap_free (holes);
}
/* At least two VECTOR to combine. */
if (valid_vecs.length () <= 1)
{
cleanup_vinfo_map (v_info_map);
return false;
}
valid_vecs.qsort (sort_by_mach_mode);
/* Go through all candidates by machine mode order, query the mode_to_total
to get the total number for each mode and skip the single one. */
for (unsigned i = 0; i < valid_vecs.length () - 1; ++i)
{
tree tvec = valid_vecs[i];
enum machine_mode mode = TYPE_MODE (TREE_TYPE (tvec));
/* Skip modes with only a single candidate. */
if (TYPE_MODE (TREE_TYPE (valid_vecs[i + 1])) != mode)
continue;
unsigned int idx, j;
gimple *sum = NULL;
tree sum_vec = tvec;
v_info_ptr info_ptr = *(v_info_map.get (tvec));
v_info_elem *elem;
tree vec_type = info_ptr->vec_type;
/* Build the sum for all candidates with same mode. */
do
{
sum = build_and_add_sum (vec_type, sum_vec,
valid_vecs[i + 1], opcode);
if (!useless_type_conversion_p (vec_type,
TREE_TYPE (valid_vecs[i + 1])))
{
/* Update the operands only after build_and_add_sum,
so that we don't have to repeat the placement algorithm
of build_and_add_sum. */
gimple_stmt_iterator gsi = gsi_for_stmt (sum);
tree vce = build1 (VIEW_CONVERT_EXPR, vec_type,
valid_vecs[i + 1]);
tree lhs = make_ssa_name (vec_type);
gimple *g = gimple_build_assign (lhs, VIEW_CONVERT_EXPR, vce);
gimple_set_uid (g, gimple_uid (sum));
gsi_insert_before (&gsi, g, GSI_NEW_STMT);
gimple_assign_set_rhs2 (sum, lhs);
if (sum_vec == tvec)
{
vce = build1 (VIEW_CONVERT_EXPR, vec_type, sum_vec);
lhs = make_ssa_name (vec_type);
g = gimple_build_assign (lhs, VIEW_CONVERT_EXPR, vce);
gimple_set_uid (g, gimple_uid (sum));
gsi_insert_before (&gsi, g, GSI_NEW_STMT);
gimple_assign_set_rhs1 (sum, lhs);
}
update_stmt (sum);
}
sum_vec = gimple_get_lhs (sum);
info_ptr = *(v_info_map.get (valid_vecs[i + 1]));
gcc_assert (types_compatible_p (vec_type, info_ptr->vec_type));
/* Update those related ops of current candidate VECTOR. */
FOR_EACH_VEC_ELT (info_ptr->vec, j, elem)
{
idx = elem->second;
gimple *def = SSA_NAME_DEF_STMT ((*ops)[idx]->op);
/* Set this then op definition will get DCEd later. */
gimple_set_visited (def, true);
if (opcode == PLUS_EXPR
|| opcode == BIT_XOR_EXPR
|| opcode == BIT_IOR_EXPR)
(*ops)[idx]->op = build_zero_cst (TREE_TYPE ((*ops)[idx]->op));
else if (opcode == MULT_EXPR)
(*ops)[idx]->op = build_one_cst (TREE_TYPE ((*ops)[idx]->op));
else
{
gcc_assert (opcode == BIT_AND_EXPR);
(*ops)[idx]->op
= build_all_ones_cst (TREE_TYPE ((*ops)[idx]->op));
}
(*ops)[idx]->rank = 0;
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Generating addition -> ");
print_gimple_stmt (dump_file, sum, 0);
}
i++;
}
while ((i < valid_vecs.length () - 1)
&& TYPE_MODE (TREE_TYPE (valid_vecs[i + 1])) == mode);
/* Referring to first valid VECTOR with this mode, generate the
BIT_FIELD_REF statements accordingly. */
info_ptr = *(v_info_map.get (tvec));
gcc_assert (sum);
tree elem_type = TREE_TYPE (vec_type);
FOR_EACH_VEC_ELT (info_ptr->vec, j, elem)
{
idx = elem->second;
tree dst = make_ssa_name (elem_type);
tree pos = bitsize_int (elem->first
* tree_to_uhwi (TYPE_SIZE (elem_type)));
tree bfr = build3 (BIT_FIELD_REF, elem_type, sum_vec,
TYPE_SIZE (elem_type), pos);
gimple *gs = gimple_build_assign (dst, BIT_FIELD_REF, bfr);
insert_stmt_after (gs, sum);
gimple *def = SSA_NAME_DEF_STMT ((*ops)[idx]->op);
/* Set this then op definition will get DCEd later. */
gimple_set_visited (def, true);
(*ops)[idx]->op = gimple_assign_lhs (gs);
(*ops)[idx]->rank = get_rank ((*ops)[idx]->op);
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Generating bit_field_ref -> ");
print_gimple_stmt (dump_file, gs, 0);
}
}
}
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "undistributiong bit_field_ref for vector done.\n");
cleanup_vinfo_map (v_info_map);
return true;
}
/* If OPCODE is BIT_IOR_EXPR or BIT_AND_EXPR and CURR is a comparison
expression, examine the other OPS to see if any of them are comparisons
of the same values, which we may be able to combine or eliminate.
For example, we can rewrite (a < b) | (a == b) as (a <= b). */
static bool
eliminate_redundant_comparison (enum tree_code opcode,
vec<operand_entry *> *ops,
unsigned int currindex,
operand_entry *curr)
{
tree op1, op2;
enum tree_code lcode, rcode;
gimple *def1, *def2;
int i;
operand_entry *oe;
if (opcode != BIT_IOR_EXPR && opcode != BIT_AND_EXPR)
return false;
/* Check that CURR is a comparison. */
if (TREE_CODE (curr->op) != SSA_NAME)
return false;
def1 = SSA_NAME_DEF_STMT (curr->op);
if (!is_gimple_assign (def1))
return false;
lcode = gimple_assign_rhs_code (def1);
if (TREE_CODE_CLASS (lcode) != tcc_comparison)
return false;
op1 = gimple_assign_rhs1 (def1);
op2 = gimple_assign_rhs2 (def1);
/* Now look for a similar comparison in the remaining OPS. */
for (i = currindex + 1; ops->iterate (i, &oe); i++)
{
tree t;
if (TREE_CODE (oe->op) != SSA_NAME)
continue;
def2 = SSA_NAME_DEF_STMT (oe->op);
if (!is_gimple_assign (def2))
continue;
rcode = gimple_assign_rhs_code (def2);
if (TREE_CODE_CLASS (rcode) != tcc_comparison)
continue;
/* If we got here, we have a match. See if we can combine the
two comparisons. */
tree type = TREE_TYPE (gimple_assign_lhs (def1));
if (opcode == BIT_IOR_EXPR)
t = maybe_fold_or_comparisons (type,
lcode, op1, op2,
rcode, gimple_assign_rhs1 (def2),
gimple_assign_rhs2 (def2));
else
t = maybe_fold_and_comparisons (type,
lcode, op1, op2,
rcode, gimple_assign_rhs1 (def2),
gimple_assign_rhs2 (def2));
if (!t)
continue;
/* maybe_fold_and_comparisons and maybe_fold_or_comparisons
always give us a boolean_type_node value back. If the original
BIT_AND_EXPR or BIT_IOR_EXPR was of a wider integer type,
we need to convert. */
if (!useless_type_conversion_p (TREE_TYPE (curr->op), TREE_TYPE (t)))
t = fold_convert (TREE_TYPE (curr->op), t);
if (TREE_CODE (t) != INTEGER_CST
&& !operand_equal_p (t, curr->op, 0))
{
enum tree_code subcode;
tree newop1, newop2;
if (!COMPARISON_CLASS_P (t))
continue;
extract_ops_from_tree (t, &subcode, &newop1, &newop2);
STRIP_USELESS_TYPE_CONVERSION (newop1);
STRIP_USELESS_TYPE_CONVERSION (newop2);
if (!is_gimple_val (newop1) || !is_gimple_val (newop2))
continue;
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Equivalence: ");
print_generic_expr (dump_file, curr->op);
fprintf (dump_file, " %s ", op_symbol_code (opcode));
print_generic_expr (dump_file, oe->op);
fprintf (dump_file, " -> ");
print_generic_expr (dump_file, t);
fprintf (dump_file, "\n");
}
/* Now we can delete oe, as it has been subsumed by the new combined
expression t. */
ops->ordered_remove (i);
reassociate_stats.ops_eliminated ++;
/* If t is the same as curr->op, we're done. Otherwise we must
replace curr->op with t. Special case is if we got a constant
back, in which case we add it to the end instead of in place of
the current entry. */
if (TREE_CODE (t) == INTEGER_CST)
{
ops->ordered_remove (currindex);
add_to_ops_vec (ops, t);
}
else if (!operand_equal_p (t, curr->op, 0))
{
gimple *sum;
enum tree_code subcode;
tree newop1;
tree newop2;
gcc_assert (COMPARISON_CLASS_P (t));
extract_ops_from_tree (t, &subcode, &newop1, &newop2);
STRIP_USELESS_TYPE_CONVERSION (newop1);
STRIP_USELESS_TYPE_CONVERSION (newop2);
gcc_checking_assert (is_gimple_val (newop1)
&& is_gimple_val (newop2));
sum = build_and_add_sum (TREE_TYPE (t), newop1, newop2, subcode);
curr->op = gimple_get_lhs (sum);
}
return true;
}
return false;
}
/* Transform repeated addition of same values into multiply with
constant. */
static bool
transform_add_to_multiply (vec<operand_entry *> *ops)
{
operand_entry *oe;
tree op = NULL_TREE;
int j;
int i, start = -1, end = 0, count = 0;
auto_vec<std::pair <int, int> > indxs;
bool changed = false;
if (!INTEGRAL_TYPE_P (TREE_TYPE ((*ops)[0]->op))
&& (!SCALAR_FLOAT_TYPE_P (TREE_TYPE ((*ops)[0]->op))
|| !flag_unsafe_math_optimizations))
return false;
/* Look for repeated operands. */
FOR_EACH_VEC_ELT (*ops, i, oe)
{
if (start == -1)
{
count = 1;
op = oe->op;
start = i;
}
else if (operand_equal_p (oe->op, op, 0))
{
count++;
end = i;
}
else
{
if (count > 1)
indxs.safe_push (std::make_pair (start, end));
count = 1;
op = oe->op;
start = i;
}
}
if (count > 1)
indxs.safe_push (std::make_pair (start, end));
for (j = indxs.length () - 1; j >= 0; --j)
{
/* Convert repeated operand addition to multiplication. */
start = indxs[j].first;
end = indxs[j].second;
op = (*ops)[start]->op;
count = end - start + 1;
for (i = end; i >= start; --i)
ops->unordered_remove (i);
tree tmp = make_ssa_name (TREE_TYPE (op));
tree cst = build_int_cst (integer_type_node, count);
gassign *mul_stmt
= gimple_build_assign (tmp, MULT_EXPR,
op, fold_convert (TREE_TYPE (op), cst));
gimple_set_visited (mul_stmt, true);
add_to_ops_vec (ops, tmp, mul_stmt);
changed = true;
}
return changed;
}
/* Perform various identities and other optimizations on the list of
operand entries, stored in OPS. The tree code for the binary
operation between all the operands is OPCODE. */
static void
optimize_ops_list (enum tree_code opcode,
vec<operand_entry *> *ops)
{
unsigned int length = ops->length ();
unsigned int i;
operand_entry *oe;
operand_entry *oelast = NULL;
bool iterate = false;
if (length == 1)
return;
oelast = ops->last ();
/* If the last two are constants, pop the constants off, merge them
and try the next two. */
if (oelast->rank == 0 && is_gimple_min_invariant (oelast->op))
{
operand_entry *oelm1 = (*ops)[length - 2];
if (oelm1->rank == 0
&& is_gimple_min_invariant (oelm1->op)
&& useless_type_conversion_p (TREE_TYPE (oelm1->op),
TREE_TYPE (oelast->op)))
{
tree folded = fold_binary (opcode, TREE_TYPE (oelm1->op),
oelm1->op, oelast->op);
if (folded && is_gimple_min_invariant (folded))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Merging constants\n");
ops->pop ();
ops->pop ();
add_to_ops_vec (ops, folded);
reassociate_stats.constants_eliminated++;
optimize_ops_list (opcode, ops);
return;
}
}
}
eliminate_using_constants (opcode, ops);
oelast = NULL;
for (i = 0; ops->iterate (i, &oe);)
{
bool done = false;
if (eliminate_not_pairs (opcode, ops, i, oe))
return;
if (eliminate_duplicate_pair (opcode, ops, &done, i, oe, oelast)
|| (!done && eliminate_plus_minus_pair (opcode, ops, i, oe))
|| (!done && eliminate_redundant_comparison (opcode, ops, i, oe)))
{
if (done)
return;
iterate = true;
oelast = NULL;
continue;
}
oelast = oe;
i++;
}
if (iterate)
optimize_ops_list (opcode, ops);
}
/* The following functions are subroutines to optimize_range_tests and allow
it to try to change a logical combination of comparisons into a range
test.
For example, both
X == 2 || X == 5 || X == 3 || X == 4
and
X >= 2 && X <= 5
are converted to
(unsigned) (X - 2) <= 3
For more information see comments above fold_test_range in fold-const.c,
this implementation is for GIMPLE. */
/* Dump the range entry R to FILE, skipping its expression if SKIP_EXP. */
void
dump_range_entry (FILE *file, struct range_entry *r, bool skip_exp)
{
if (!skip_exp)
print_generic_expr (file, r->exp);
fprintf (file, " %c[", r->in_p ? '+' : '-');
print_generic_expr (file, r->low);
fputs (", ", file);
print_generic_expr (file, r->high);
fputc (']', file);
}
/* Dump the range entry R to STDERR. */
DEBUG_FUNCTION void
debug_range_entry (struct range_entry *r)
{
dump_range_entry (stderr, r, false);
fputc ('\n', stderr);
}
/* This is similar to make_range in fold-const.c, but on top of
GIMPLE instead of trees. If EXP is non-NULL, it should be
an SSA_NAME and STMT argument is ignored, otherwise STMT
argument should be a GIMPLE_COND. */
void
init_range_entry (struct range_entry *r, tree exp, gimple *stmt)
{
int in_p;
tree low, high;
bool is_bool, strict_overflow_p;
r->exp = NULL_TREE;
r->in_p = false;
r->strict_overflow_p = false;
r->low = NULL_TREE;
r->high = NULL_TREE;
if (exp != NULL_TREE
&& (TREE_CODE (exp) != SSA_NAME || !INTEGRAL_TYPE_P (TREE_TYPE (exp))))
return;
/* Start with simply saying "EXP != 0" and then look at the code of EXP
and see if we can refine the range. Some of the cases below may not
happen, but it doesn't seem worth worrying about this. We "continue"
the outer loop when we've changed something; otherwise we "break"
the switch, which will "break" the while. */
low = exp ? build_int_cst (TREE_TYPE (exp), 0) : boolean_false_node;
high = low;
in_p = 0;
strict_overflow_p = false;
is_bool = false;
if (exp == NULL_TREE)
is_bool = true;
else if (TYPE_PRECISION (TREE_TYPE (exp)) == 1)
{
if (TYPE_UNSIGNED (TREE_TYPE (exp)))
is_bool = true;
else
return;
}
else if (TREE_CODE (TREE_TYPE (exp)) == BOOLEAN_TYPE)
is_bool = true;
while (1)
{
enum tree_code code;
tree arg0, arg1, exp_type;
tree nexp;
location_t loc;
if (exp != NULL_TREE)
{
if (TREE_CODE (exp) != SSA_NAME
|| SSA_NAME_OCCURS_IN_ABNORMAL_PHI (exp))
break;
stmt = SSA_NAME_DEF_STMT (exp);
if (!is_gimple_assign (stmt))
break;
code = gimple_assign_rhs_code (stmt);
arg0 = gimple_assign_rhs1 (stmt);
arg1 = gimple_assign_rhs2 (stmt);
exp_type = TREE_TYPE (exp);
}
else
{
code = gimple_cond_code (stmt);
arg0 = gimple_cond_lhs (stmt);
arg1 = gimple_cond_rhs (stmt);
exp_type = boolean_type_node;
}
if (TREE_CODE (arg0) != SSA_NAME
|| SSA_NAME_OCCURS_IN_ABNORMAL_PHI (arg0))
break;
loc = gimple_location (stmt);
switch (code)
{
case BIT_NOT_EXPR:
if (TREE_CODE (TREE_TYPE (exp)) == BOOLEAN_TYPE
/* Ensure the range is either +[-,0], +[0,0],
-[-,0], -[0,0] or +[1,-], +[1,1], -[1,-] or
-[1,1]. If it is e.g. +[-,-] or -[-,-]
or similar expression of unconditional true or
false, it should not be negated. */
&& ((high && integer_zerop (high))
|| (low && integer_onep (low))))
{
in_p = !in_p;
exp = arg0;
continue;
}
break;
case SSA_NAME:
exp = arg0;
continue;
CASE_CONVERT:
if (is_bool)
{
if ((TYPE_PRECISION (exp_type) == 1
|| TREE_CODE (exp_type) == BOOLEAN_TYPE)
&& TYPE_PRECISION (TREE_TYPE (arg0)) > 1)
return;
}
else if (TYPE_PRECISION (TREE_TYPE (arg0)) == 1)
{
if (TYPE_UNSIGNED (TREE_TYPE (arg0)))
is_bool = true;
else
return;
}
else if (TREE_CODE (TREE_TYPE (arg0)) == BOOLEAN_TYPE)
is_bool = true;
goto do_default;
case EQ_EXPR:
case NE_EXPR:
case LT_EXPR:
case LE_EXPR:
case GE_EXPR:
case GT_EXPR:
is_bool = true;
/* FALLTHRU */
default:
if (!is_bool)
return;
do_default:
nexp = make_range_step (loc, code, arg0, arg1, exp_type,
&low, &high, &in_p,
&strict_overflow_p);
if (nexp != NULL_TREE)
{
exp = nexp;
gcc_assert (TREE_CODE (exp) == SSA_NAME);
continue;
}
break;
}
break;
}
if (is_bool)
{
r->exp = exp;
r->in_p = in_p;
r->low = low;
r->high = high;
r->strict_overflow_p = strict_overflow_p;
}
}
/* Comparison function for qsort. Sort entries
without SSA_NAME exp first, then with SSA_NAMEs sorted
by increasing SSA_NAME_VERSION, and for the same SSA_NAMEs
by increasing ->low and if ->low is the same, by increasing
->high. ->low == NULL_TREE means minimum, ->high == NULL_TREE
maximum. */
static int
range_entry_cmp (const void *a, const void *b)
{
const struct range_entry *p = (const struct range_entry *) a;
const struct range_entry *q = (const struct range_entry *) b;
if (p->exp != NULL_TREE && TREE_CODE (p->exp) == SSA_NAME)
{
if (q->exp != NULL_TREE && TREE_CODE (q->exp) == SSA_NAME)
{
/* Group range_entries for the same SSA_NAME together. */
if (SSA_NAME_VERSION (p->exp) < SSA_NAME_VERSION (q->exp))
return -1;
else if (SSA_NAME_VERSION (p->exp) > SSA_NAME_VERSION (q->exp))
return 1;
/* If ->low is different, NULL low goes first, then by
ascending low. */
if (p->low != NULL_TREE)
{
if (q->low != NULL_TREE)
{
tree tem = fold_binary (LT_EXPR, boolean_type_node,
p->low, q->low);
if (tem && integer_onep (tem))
return -1;
tem = fold_binary (GT_EXPR, boolean_type_node,
p->low, q->low);
if (tem && integer_onep (tem))
return 1;
}
else
return 1;
}
else if (q->low != NULL_TREE)
return -1;
/* If ->high is different, NULL high goes last, before that by
ascending high. */
if (p->high != NULL_TREE)
{
if (q->high != NULL_TREE)
{
tree tem = fold_binary (LT_EXPR, boolean_type_node,
p->high, q->high);
if (tem && integer_onep (tem))
return -1;
tem = fold_binary (GT_EXPR, boolean_type_node,
p->high, q->high);
if (tem && integer_onep (tem))
return 1;
}
else
return -1;
}
else if (q->high != NULL_TREE)
return 1;
/* If both ranges are the same, sort below by ascending idx. */
}
else
return 1;
}
else if (q->exp != NULL_TREE && TREE_CODE (q->exp) == SSA_NAME)
return -1;
if (p->idx < q->idx)
return -1;
else
{
gcc_checking_assert (p->idx > q->idx);
return 1;
}
}
/* Helper function for update_range_test. Force EXPR into an SSA_NAME,
insert needed statements BEFORE or after GSI. */
static tree
force_into_ssa_name (gimple_stmt_iterator *gsi, tree expr, bool before)
{
enum gsi_iterator_update m = before ? GSI_SAME_STMT : GSI_CONTINUE_LINKING;
tree ret = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE, before, m);
if (TREE_CODE (ret) != SSA_NAME)
{
gimple *g = gimple_build_assign (make_ssa_name (TREE_TYPE (ret)), ret);
if (before)
gsi_insert_before (gsi, g, GSI_SAME_STMT);
else
gsi_insert_after (gsi, g, GSI_CONTINUE_LINKING);
ret = gimple_assign_lhs (g);
}
return ret;
}
/* Helper routine of optimize_range_test.
[EXP, IN_P, LOW, HIGH, STRICT_OVERFLOW_P] is a merged range for
RANGE and OTHERRANGE through OTHERRANGE + COUNT - 1 ranges,
OPCODE and OPS are arguments of optimize_range_tests. If OTHERRANGE
is NULL, OTHERRANGEP should not be and then OTHERRANGEP points to
an array of COUNT pointers to other ranges. Return
true if the range merge has been successful.
If OPCODE is ERROR_MARK, this is called from within
maybe_optimize_range_tests and is performing inter-bb range optimization.
In that case, whether an op is BIT_AND_EXPR or BIT_IOR_EXPR is found in
oe->rank. */
static bool
update_range_test (struct range_entry *range, struct range_entry *otherrange,
struct range_entry **otherrangep,
unsigned int count, enum tree_code opcode,
vec<operand_entry *> *ops, tree exp, gimple_seq seq,
bool in_p, tree low, tree high, bool strict_overflow_p)
{
operand_entry *oe = (*ops)[range->idx];
tree op = oe->op;
gimple *stmt = op ? SSA_NAME_DEF_STMT (op)
: last_stmt (BASIC_BLOCK_FOR_FN (cfun, oe->id));
location_t loc = gimple_location (stmt);
tree optype = op ? TREE_TYPE (op) : boolean_type_node;
tree tem = build_range_check (loc, optype, unshare_expr (exp),
in_p, low, high);
enum warn_strict_overflow_code wc = WARN_STRICT_OVERFLOW_COMPARISON;
gimple_stmt_iterator gsi;
unsigned int i, uid;
if (tem == NULL_TREE)
return false;
/* If op is default def SSA_NAME, there is no place to insert the
new comparison. Give up, unless we can use OP itself as the
range test. */
if (op && SSA_NAME_IS_DEFAULT_DEF (op))
{
if (op == range->exp
&& ((TYPE_PRECISION (optype) == 1 && TYPE_UNSIGNED (optype))
|| TREE_CODE (optype) == BOOLEAN_TYPE)
&& (op == tem
|| (TREE_CODE (tem) == EQ_EXPR
&& TREE_OPERAND (tem, 0) == op
&& integer_onep (TREE_OPERAND (tem, 1))))
&& opcode != BIT_IOR_EXPR
&& (opcode != ERROR_MARK || oe->rank != BIT_IOR_EXPR))
{
stmt = NULL;
tem = op;
}
else
return false;
}
if (strict_overflow_p && issue_strict_overflow_warning (wc))
warning_at (loc, OPT_Wstrict_overflow,
"assuming signed overflow does not occur "
"when simplifying range test");
if (dump_file && (dump_flags & TDF_DETAILS))
{
struct range_entry *r;
fprintf (dump_file, "Optimizing range tests ");
dump_range_entry (dump_file, range, false);
for (i = 0; i < count; i++)
{
if (otherrange)
r = otherrange + i;
else
r = otherrangep[i];
if (r->exp
&& r->exp != range->exp
&& TREE_CODE (r->exp) == SSA_NAME)
{
fprintf (dump_file, " and ");
dump_range_entry (dump_file, r, false);
}
else
{
fprintf (dump_file, " and");
dump_range_entry (dump_file, r, true);
}
}
fprintf (dump_file, "\n into ");
print_generic_expr (dump_file, tem);
fprintf (dump_file, "\n");
}
if (opcode == BIT_IOR_EXPR
|| (opcode == ERROR_MARK && oe->rank == BIT_IOR_EXPR))
tem = invert_truthvalue_loc (loc, tem);
tem = fold_convert_loc (loc, optype, tem);
if (stmt)
{
gsi = gsi_for_stmt (stmt);
uid = gimple_uid (stmt);
}
else
{
gsi = gsi_none ();
uid = 0;
}
if (stmt == NULL)
gcc_checking_assert (tem == op);
/* In rare cases range->exp can be equal to lhs of stmt.
In that case we have to insert after the stmt rather then before
it. If stmt is a PHI, insert it at the start of the basic block. */
else if (op != range->exp)
{
gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
tem = force_into_ssa_name (&gsi, tem, true);
gsi_prev (&gsi);
}
else if (gimple_code (stmt) != GIMPLE_PHI)
{
gsi_insert_seq_after (&gsi, seq, GSI_CONTINUE_LINKING);
tem = force_into_ssa_name (&gsi, tem, false);
}
else
{
gsi = gsi_after_labels (gimple_bb (stmt));
if (!gsi_end_p (gsi))
uid = gimple_uid (gsi_stmt (gsi));
else
{
gsi = gsi_start_bb (gimple_bb (stmt));
uid = 1;
while (!gsi_end_p (gsi))
{
uid = gimple_uid (gsi_stmt (gsi));
gsi_next (&gsi);
}
}
gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT);
tem = force_into_ssa_name (&gsi, tem, true);
if (gsi_end_p (gsi))
gsi = gsi_last_bb (gimple_bb (stmt));
else
gsi_prev (&gsi);
}
for (; !gsi_end_p (gsi); gsi_prev (&gsi))
if (gimple_uid (gsi_stmt (gsi)))
break;
else
gimple_set_uid (gsi_stmt (gsi), uid);
oe->op = tem;
range->exp = exp;
range->low = low;
range->high = high;
range->in_p = in_p;
range->strict_overflow_p = false;
for (i = 0; i < count; i++)
{
if (otherrange)
range = otherrange + i;
else
range = otherrangep[i];
oe = (*ops)[range->idx];
/* Now change all the other range test immediate uses, so that
those tests will be optimized away. */
if (opcode == ERROR_MARK)
{
if (oe->op)
oe->op = build_int_cst (TREE_TYPE (oe->op),
oe->rank == BIT_IOR_EXPR ? 0 : 1);
else
oe->op = (oe->rank == BIT_IOR_EXPR
? boolean_false_node : boolean_true_node);
}
else
oe->op = error_mark_node;
range->exp = NULL_TREE;
range->low = NULL_TREE;
range->high = NULL_TREE;
}
return true;
}
/* Optimize X == CST1 || X == CST2
if popcount (CST1 ^ CST2) == 1 into
(X & ~(CST1 ^ CST2)) == (CST1 & ~(CST1 ^ CST2)).
Similarly for ranges. E.g.
X != 2 && X != 3 && X != 10 && X != 11
will be transformed by the previous optimization into
!((X - 2U) <= 1U || (X - 10U) <= 1U)
and this loop can transform that into
!(((X & ~8) - 2U) <= 1U). */
static bool
optimize_range_tests_xor (enum tree_code opcode, tree type,
tree lowi, tree lowj, tree highi, tree highj,
vec<operand_entry *> *ops,
struct range_entry *rangei,
struct range_entry *rangej)
{
tree lowxor, highxor, tem, exp;
/* Check lowi ^ lowj == highi ^ highj and
popcount (lowi ^ lowj) == 1. */
lowxor = fold_binary (BIT_XOR_EXPR, type, lowi, lowj);
if (lowxor == NULL_TREE || TREE_CODE (lowxor) != INTEGER_CST)
return false;
if (!integer_pow2p (lowxor))
return false;
highxor = fold_binary (BIT_XOR_EXPR, type, highi, highj);
if (!tree_int_cst_equal (lowxor, highxor))
return false;
exp = rangei->exp;
scalar_int_mode mode = as_a <scalar_int_mode> (TYPE_MODE (type));
int prec = GET_MODE_PRECISION (mode);
if (TYPE_PRECISION (type) < prec
|| (wi::to_wide (TYPE_MIN_VALUE (type))
!= wi::min_value (prec, TYPE_SIGN (type)))
|| (wi::to_wide (TYPE_MAX_VALUE (type))
!= wi::max_value (prec, TYPE_SIGN (type))))
{
type = build_nonstandard_integer_type (prec, TYPE_UNSIGNED (type));
exp = fold_convert (type, exp);
lowxor = fold_convert (type, lowxor);
lowi = fold_convert (type, lowi);
highi = fold_convert (type, highi);
}
tem = fold_build1 (BIT_NOT_EXPR, type, lowxor);
exp = fold_build2 (BIT_AND_EXPR, type, exp, tem);
lowj = fold_build2 (BIT_AND_EXPR, type, lowi, tem);
highj = fold_build2 (BIT_AND_EXPR, type, highi, tem);
if (update_range_test (rangei, rangej, NULL, 1, opcode, ops, exp,
NULL, rangei->in_p, lowj, highj,
rangei->strict_overflow_p
|| rangej->strict_overflow_p))
return true;
return false;
}
/* Optimize X == CST1 || X == CST2
if popcount (CST2 - CST1) == 1 into
((X - CST1) & ~(CST2 - CST1)) == 0.
Similarly for ranges. E.g.
X == 43 || X == 76 || X == 44 || X == 78 || X == 77 || X == 46
|| X == 75 || X == 45
will be transformed by the previous optimization into
(X - 43U) <= 3U || (X - 75U) <= 3U
and this loop can transform that into
((X - 43U) & ~(75U - 43U)) <= 3U. */
static bool
optimize_range_tests_diff (enum tree_code opcode, tree type,
tree lowi, tree lowj, tree highi, tree highj,
vec<operand_entry *> *ops,
struct range_entry *rangei,
struct range_entry *rangej)
{
tree tem1, tem2, mask;
/* Check highi - lowi == highj - lowj. */
tem1 = fold_binary (MINUS_EXPR, type, highi, lowi);
if (tem1 == NULL_TREE || TREE_CODE (tem1) != INTEGER_CST)
return false;
tem2 = fold_binary (MINUS_EXPR, type, highj, lowj);
if (!tree_int_cst_equal (tem1, tem2))
return false;
/* Check popcount (lowj - lowi) == 1. */
tem1 = fold_binary (MINUS_EXPR, type, lowj, lowi);
if (tem1 == NULL_TREE || TREE_CODE (tem1) != INTEGER_CST)
return false;
if (!integer_pow2p (tem1))
return false;
scalar_int_mode mode = as_a <scalar_int_mode> (TYPE_MODE (type));
int prec = GET_MODE_PRECISION (mode);
if (TYPE_PRECISION (type) < prec
|| (wi::to_wide (TYPE_MIN_VALUE (type))
!= wi::min_value (prec, TYPE_SIGN (type)))
|| (wi::to_wide (TYPE_MAX_VALUE (type))
!= wi::max_value (prec, TYPE_SIGN (type))))
type = build_nonstandard_integer_type (prec, 1);
else
type = unsigned_type_for (type);
tem1 = fold_convert (type, tem1);
tem2 = fold_convert (type, tem2);
lowi = fold_convert (type, lowi);
mask = fold_build1 (BIT_NOT_EXPR, type, tem1);
tem1 = fold_build2 (MINUS_EXPR, type,
fold_convert (type, rangei->exp), lowi);
tem1 = fold_build2 (BIT_AND_EXPR, type, tem1, mask);
lowj = build_int_cst (type, 0);
if (update_range_test (rangei, rangej, NULL, 1, opcode, ops, tem1,
NULL, rangei->in_p, lowj, tem2,
rangei->strict_overflow_p
|| rangej->strict_overflow_p))
return true;
return false;
}
/* It does some common checks for function optimize_range_tests_xor and
optimize_range_tests_diff.
If OPTIMIZE_XOR is TRUE, it calls optimize_range_tests_xor.
Else it calls optimize_range_tests_diff. */
static bool
optimize_range_tests_1 (enum tree_code opcode, int first, int length,
bool optimize_xor, vec<operand_entry *> *ops,
struct range_entry *ranges)
{
int i, j;
bool any_changes = false;
for (i = first; i < length; i++)
{
tree lowi, highi, lowj, highj, type, tem;
if (ranges[i].exp == NULL_TREE || ranges[i].in_p)
continue;
type = TREE_TYPE (ranges[i].exp);
if (!INTEGRAL_TYPE_P (type))
continue;
lowi = ranges[i].low;
if (lowi == NULL_TREE)
lowi = TYPE_MIN_VALUE (type);
highi = ranges[i].high;
if (highi == NULL_TREE)
continue;
for (j = i + 1; j < length && j < i + 64; j++)
{
bool changes;
if (ranges[i].exp != ranges[j].exp || ranges[j].in_p)
continue;
lowj = ranges[j].low;
if (lowj == NULL_TREE)
continue;
highj = ranges[j].high;
if (highj == NULL_TREE)
highj = TYPE_MAX_VALUE (type);
/* Check lowj > highi. */
tem = fold_binary (GT_EXPR, boolean_type_node,
lowj, highi);
if (tem == NULL_TREE || !integer_onep (tem))
continue;
if (optimize_xor)
changes = optimize_range_tests_xor (opcode, type, lowi, lowj,
highi, highj, ops,
ranges + i, ranges + j);
else
changes = optimize_range_tests_diff (opcode, type, lowi, lowj,
highi, highj, ops,
ranges + i, ranges + j);
if (changes)
{
any_changes = true;
break;
}
}
}
return any_changes;
}
/* Helper function of optimize_range_tests_to_bit_test. Handle a single
range, EXP, LOW, HIGH, compute bit mask of bits to test and return
EXP on success, NULL otherwise. */
static tree
extract_bit_test_mask (tree exp, int prec, tree totallow, tree low, tree high,
wide_int *mask, tree *totallowp)
{
tree tem = int_const_binop (MINUS_EXPR, high, low);
if (tem == NULL_TREE
|| TREE_CODE (tem) != INTEGER_CST
|| TREE_OVERFLOW (tem)
|| tree_int_cst_sgn (tem) == -1
|| compare_tree_int (tem, prec) != -1)