/* Lower vector operations to scalar operations.
- Copyright (C) 2004-2014 Free Software Foundation, Inc.
+ Copyright (C) 2004-2015 Free Software Foundation, Inc.
This file is part of GCC.
#include "config.h"
#include "system.h"
#include "coretypes.h"
+#include "alias.h"
+#include "symtab.h"
+#include "options.h"
#include "tree.h"
+#include "fold-const.h"
#include "stor-layout.h"
#include "tm.h"
#include "langhooks.h"
+#include "predict.h"
+#include "hard-reg-set.h"
+#include "function.h"
+#include "dominance.h"
+#include "cfg.h"
#include "basic-block.h"
#include "tree-ssa-alias.h"
#include "internal-fn.h"
#include "tree-eh.h"
#include "gimple-expr.h"
-#include "is-a.h"
#include "gimple.h"
#include "gimple-iterator.h"
#include "gimplify-me.h"
#include "target.h"
/* Need to include rtl.h, expr.h, etc. for optabs. */
+#include "rtl.h"
+#include "insn-config.h"
+#include "expmed.h"
+#include "dojump.h"
+#include "explow.h"
+#include "calls.h"
+#include "emit-rtl.h"
+#include "varasm.h"
+#include "stmt.h"
#include "expr.h"
+#include "insn-codes.h"
#include "optabs.h"
enum tree_code code)
{
tree result, compute_type;
- enum machine_mode mode;
+ machine_mode mode;
int n_words = tree_to_uhwi (TYPE_SIZE_UNIT (type)) / UNITS_PER_WORD;
location_t loc = gimple_location (gsi_stmt (*gsi));
((unsigned HOST_WIDE_INT) 1
<< shifts[i]) - 1);
cst = build_vector (type, vec);
- addend = make_ssa_name (type, NULL);
- stmt = gimple_build_assign_with_ops (VEC_COND_EXPR, addend,
- cond, cst, zero);
+ addend = make_ssa_name (type);
+ stmt = gimple_build_assign (addend, VEC_COND_EXPR, cond,
+ cst, zero);
gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
}
}
static void
expand_vector_condition (gimple_stmt_iterator *gsi)
{
- gimple stmt = gsi_stmt (*gsi);
+ gassign *stmt = as_a <gassign *> (gsi_stmt (*gsi));
tree type = gimple_expr_type (stmt);
tree a = gimple_assign_rhs1 (stmt);
tree a1 = a;
static tree
expand_vector_operation (gimple_stmt_iterator *gsi, tree type, tree compute_type,
- gimple assign, enum tree_code code)
+ gassign *assign, enum tree_code code)
{
- enum machine_mode compute_mode = TYPE_MODE (compute_type);
+ machine_mode compute_mode = TYPE_MODE (compute_type);
/* If the compute mode is not a vector mode (hence we are not decomposing
a BLKmode vector to smaller, hardware-supported vectors), we may want
{
case PLUS_EXPR:
case MINUS_EXPR:
- if (!TYPE_OVERFLOW_TRAPS (type))
+ if (ANY_INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type))
return expand_vector_addition (gsi, do_binop, do_plus_minus, type,
gimple_assign_rhs1 (assign),
gimple_assign_rhs2 (assign), code);
break;
case NEGATE_EXPR:
- if (!TYPE_OVERFLOW_TRAPS (type))
+ if (ANY_INTEGRAL_TYPE_P (type) && !TYPE_OVERFLOW_TRAPS (type))
return expand_vector_addition (gsi, do_unop, do_negate, type,
gimple_assign_rhs1 (assign),
NULL_TREE, code);
static void
optimize_vector_constructor (gimple_stmt_iterator *gsi)
{
- gimple stmt = gsi_stmt (*gsi);
+ gassign *stmt = as_a <gassign *> (gsi_stmt (*gsi));
tree lhs = gimple_assign_lhs (stmt);
tree rhs = gimple_assign_rhs1 (stmt);
tree type = TREE_TYPE (rhs);
}
for (i = 0; i < nelts; i++)
CONSTRUCTOR_ELT (rhs, i)->value = base;
- g = gimple_build_assign (make_ssa_name (type, NULL), rhs);
+ g = gimple_build_assign (make_ssa_name (type), rhs);
gsi_insert_before (gsi, g, GSI_SAME_STMT);
- g = gimple_build_assign_with_ops (PLUS_EXPR, lhs, gimple_assign_lhs (g),
- build_vector (type, cst));
+ g = gimple_build_assign (lhs, PLUS_EXPR, gimple_assign_lhs (g),
+ build_vector (type, cst));
gsi_replace (gsi, g, false);
}
\f
static tree
type_for_widest_vector_mode (tree type, optab op)
{
- enum machine_mode inner_mode = TYPE_MODE (type);
- enum machine_mode best_mode = VOIDmode, mode;
+ machine_mode inner_mode = TYPE_MODE (type);
+ machine_mode best_mode = VOIDmode, mode;
int best_nunits = 0;
if (SCALAR_FLOAT_MODE_P (inner_mode))
static void
lower_vec_perm (gimple_stmt_iterator *gsi)
{
- gimple stmt = gsi_stmt (*gsi);
+ gassign *stmt = as_a <gassign *> (gsi_stmt (*gsi));
tree mask = gimple_assign_rhs3 (stmt);
tree vec0 = gimple_assign_rhs1 (stmt);
tree vec1 = gimple_assign_rhs2 (stmt);
update_stmt (gsi_stmt (*gsi));
}
+/* Return type in which CODE operation with optab OP can be
+ computed. */
+
+static tree
+get_compute_type (enum tree_code code, optab op, tree type)
+{
+ /* For very wide vectors, try using a smaller vector mode. */
+ tree compute_type = type;
+ if (op
+ && (!VECTOR_MODE_P (TYPE_MODE (type))
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing))
+ {
+ tree vector_compute_type
+ = type_for_widest_vector_mode (TREE_TYPE (type), op);
+ if (vector_compute_type != NULL_TREE
+ && (TYPE_VECTOR_SUBPARTS (vector_compute_type)
+ < TYPE_VECTOR_SUBPARTS (compute_type))
+ && (optab_handler (op, TYPE_MODE (vector_compute_type))
+ != CODE_FOR_nothing))
+ compute_type = vector_compute_type;
+ }
+
+ /* If we are breaking a BLKmode vector into smaller pieces,
+ type_for_widest_vector_mode has already looked into the optab,
+ so skip these checks. */
+ if (compute_type == type)
+ {
+ machine_mode compute_mode = TYPE_MODE (compute_type);
+ if (VECTOR_MODE_P (compute_mode))
+ {
+ if (op && optab_handler (op, compute_mode) != CODE_FOR_nothing)
+ return compute_type;
+ if (code == MULT_HIGHPART_EXPR
+ && can_mult_highpart_p (compute_mode,
+ TYPE_UNSIGNED (compute_type)))
+ return compute_type;
+ }
+ /* There is no operation in hardware, so fall back to scalars. */
+ compute_type = TREE_TYPE (type);
+ }
+
+ return compute_type;
+}
+
+/* Helper function of expand_vector_operations_1. Return number of
+ vector elements for vector types or 1 for other types. */
+
+static inline int
+count_type_subparts (tree type)
+{
+ return VECTOR_TYPE_P (type) ? TYPE_VECTOR_SUBPARTS (type) : 1;
+}
+
+static tree
+do_cond (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b,
+ tree bitpos, tree bitsize, enum tree_code code)
+{
+ if (TREE_CODE (TREE_TYPE (a)) == VECTOR_TYPE)
+ a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
+ if (TREE_CODE (TREE_TYPE (b)) == VECTOR_TYPE)
+ b = tree_vec_extract (gsi, inner_type, b, bitsize, bitpos);
+ tree cond = gimple_assign_rhs1 (gsi_stmt (*gsi));
+ return gimplify_build3 (gsi, code, inner_type, cond, a, b);
+}
+
+/* Expand a vector COND_EXPR to scalars, piecewise. */
+static void
+expand_vector_scalar_condition (gimple_stmt_iterator *gsi)
+{
+ gassign *stmt = as_a <gassign *> (gsi_stmt (*gsi));
+ tree type = gimple_expr_type (stmt);
+ tree compute_type = get_compute_type (COND_EXPR, mov_optab, type);
+ machine_mode compute_mode = TYPE_MODE (compute_type);
+ gcc_assert (compute_mode != BLKmode);
+ tree lhs = gimple_assign_lhs (stmt);
+ tree rhs2 = gimple_assign_rhs2 (stmt);
+ tree rhs3 = gimple_assign_rhs3 (stmt);
+ tree new_rhs;
+
+ /* If the compute mode is not a vector mode (hence we are not decomposing
+ a BLKmode vector to smaller, hardware-supported vectors), we may want
+ to expand the operations in parallel. */
+ if (GET_MODE_CLASS (compute_mode) != MODE_VECTOR_INT
+ && GET_MODE_CLASS (compute_mode) != MODE_VECTOR_FLOAT
+ && GET_MODE_CLASS (compute_mode) != MODE_VECTOR_FRACT
+ && GET_MODE_CLASS (compute_mode) != MODE_VECTOR_UFRACT
+ && GET_MODE_CLASS (compute_mode) != MODE_VECTOR_ACCUM
+ && GET_MODE_CLASS (compute_mode) != MODE_VECTOR_UACCUM)
+ new_rhs = expand_vector_parallel (gsi, do_cond, type, rhs2, rhs3,
+ COND_EXPR);
+ else
+ new_rhs = expand_vector_piecewise (gsi, do_cond, type, compute_type,
+ rhs2, rhs3, COND_EXPR);
+ if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (new_rhs)))
+ new_rhs = gimplify_build1 (gsi, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
+ new_rhs);
+
+ /* NOTE: We should avoid using gimple_assign_set_rhs_from_tree. One
+ way to do it is change expand_vector_operation and its callees to
+ return a tree_code, RHS1 and RHS2 instead of a tree. */
+ gimple_assign_set_rhs_from_tree (gsi, new_rhs);
+ update_stmt (gsi_stmt (*gsi));
+}
+
/* Process one statement. If we identify a vector operation, expand it. */
static void
expand_vector_operations_1 (gimple_stmt_iterator *gsi)
{
- gimple stmt = gsi_stmt (*gsi);
- tree lhs, rhs1, rhs2 = NULL, type, compute_type;
+ tree lhs, rhs1, rhs2 = NULL, type, compute_type = NULL_TREE;
enum tree_code code;
- enum machine_mode compute_mode;
optab op = unknown_optab;
enum gimple_rhs_class rhs_class;
tree new_rhs;
- if (gimple_code (stmt) != GIMPLE_ASSIGN)
+ /* Only consider code == GIMPLE_ASSIGN. */
+ gassign *stmt = dyn_cast <gassign *> (gsi_stmt (*gsi));
+ if (!stmt)
return;
code = gimple_assign_rhs_code (stmt);
return;
}
+ if (code == COND_EXPR
+ && TREE_CODE (TREE_TYPE (gimple_assign_lhs (stmt))) == VECTOR_TYPE
+ && TYPE_MODE (TREE_TYPE (gimple_assign_lhs (stmt))) == BLKmode)
+ {
+ expand_vector_scalar_condition (gsi);
+ return;
+ }
+
if (code == CONSTRUCTOR
&& TREE_CODE (lhs) == SSA_NAME
&& VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (lhs)))
if (TREE_CODE (type) != VECTOR_TYPE)
return;
- if (code == NOP_EXPR
+ if (CONVERT_EXPR_CODE_P (code)
|| code == FLOAT_EXPR
|| code == FIX_TRUNC_EXPR
|| code == VIEW_CONVERT_EXPR)
return;
- gcc_assert (code != CONVERT_EXPR);
-
/* The signedness is determined from input argument. */
if (code == VEC_UNPACK_FLOAT_HI_EXPR
|| code == VEC_UNPACK_FLOAT_LO_EXPR)
{
op = optab_for_tree_code (code, type, optab_scalar);
+ compute_type = get_compute_type (code, op, type);
+ if (compute_type == type)
+ return;
/* The rtl expander will expand vector/scalar as vector/vector
- if necessary. Don't bother converting the stmt here. */
- if (optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing
- && optab_handler (opv, TYPE_MODE (type)) != CODE_FOR_nothing)
+ if necessary. Pick one with wider vector type. */
+ tree compute_vtype = get_compute_type (code, opv, type);
+ if (count_type_subparts (compute_vtype)
+ > count_type_subparts (compute_type))
+ {
+ compute_type = compute_vtype;
+ op = opv;
+ }
+ }
+
+ if (code == LROTATE_EXPR || code == RROTATE_EXPR)
+ {
+ if (compute_type == NULL_TREE)
+ compute_type = get_compute_type (code, op, type);
+ if (compute_type == type)
return;
+ /* Before splitting vector rotates into scalar rotates,
+ see if we can't use vector shifts and BIT_IOR_EXPR
+ instead. For vector by vector rotates we'd also
+ need to check BIT_AND_EXPR and NEGATE_EXPR, punt there
+ for now, fold doesn't seem to create such rotates anyway. */
+ if (compute_type == TREE_TYPE (type)
+ && !VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2)))
+ {
+ optab oplv = vashl_optab, opl = ashl_optab;
+ optab oprv = vlshr_optab, opr = lshr_optab, opo = ior_optab;
+ tree compute_lvtype = get_compute_type (LSHIFT_EXPR, oplv, type);
+ tree compute_rvtype = get_compute_type (RSHIFT_EXPR, oprv, type);
+ tree compute_otype = get_compute_type (BIT_IOR_EXPR, opo, type);
+ tree compute_ltype = get_compute_type (LSHIFT_EXPR, opl, type);
+ tree compute_rtype = get_compute_type (RSHIFT_EXPR, opr, type);
+ /* The rtl expander will expand vector/scalar as vector/vector
+ if necessary. Pick one with wider vector type. */
+ if (count_type_subparts (compute_lvtype)
+ > count_type_subparts (compute_ltype))
+ {
+ compute_ltype = compute_lvtype;
+ opl = oplv;
+ }
+ if (count_type_subparts (compute_rvtype)
+ > count_type_subparts (compute_rtype))
+ {
+ compute_rtype = compute_rvtype;
+ opr = oprv;
+ }
+ /* Pick the narrowest type from LSHIFT_EXPR, RSHIFT_EXPR and
+ BIT_IOR_EXPR. */
+ compute_type = compute_ltype;
+ if (count_type_subparts (compute_type)
+ > count_type_subparts (compute_rtype))
+ compute_type = compute_rtype;
+ if (count_type_subparts (compute_type)
+ > count_type_subparts (compute_otype))
+ compute_type = compute_otype;
+ /* Verify all 3 operations can be performed in that type. */
+ if (compute_type != TREE_TYPE (type))
+ {
+ if (optab_handler (opl, TYPE_MODE (compute_type))
+ == CODE_FOR_nothing
+ || optab_handler (opr, TYPE_MODE (compute_type))
+ == CODE_FOR_nothing
+ || optab_handler (opo, TYPE_MODE (compute_type))
+ == CODE_FOR_nothing)
+ compute_type = TREE_TYPE (type);
+ }
+ }
}
}
else
&& INTEGRAL_TYPE_P (TREE_TYPE (type)))
op = optab_for_tree_code (MINUS_EXPR, type, optab_default);
- /* For very wide vectors, try using a smaller vector mode. */
- compute_type = type;
- if (!VECTOR_MODE_P (TYPE_MODE (type)) && op)
- {
- tree vector_compute_type
- = type_for_widest_vector_mode (TREE_TYPE (type), op);
- if (vector_compute_type != NULL_TREE
- && (TYPE_VECTOR_SUBPARTS (vector_compute_type)
- < TYPE_VECTOR_SUBPARTS (compute_type))
- && (optab_handler (op, TYPE_MODE (vector_compute_type))
- != CODE_FOR_nothing))
- compute_type = vector_compute_type;
- }
-
- /* If we are breaking a BLKmode vector into smaller pieces,
- type_for_widest_vector_mode has already looked into the optab,
- so skip these checks. */
+ if (compute_type == NULL_TREE)
+ compute_type = get_compute_type (code, op, type);
if (compute_type == type)
- {
- compute_mode = TYPE_MODE (compute_type);
- if (VECTOR_MODE_P (compute_mode))
- {
- if (op && optab_handler (op, compute_mode) != CODE_FOR_nothing)
- return;
- if (code == MULT_HIGHPART_EXPR
- && can_mult_highpart_p (compute_mode,
- TYPE_UNSIGNED (compute_type)))
- return;
- }
- /* There is no operation in hardware, so fall back to scalars. */
- compute_type = TREE_TYPE (type);
- }
+ return;
- gcc_assert (code != VEC_LSHIFT_EXPR && code != VEC_RSHIFT_EXPR);
new_rhs = expand_vector_operation (gsi, type, compute_type, stmt, code);
/* Leave expression untouched for later expansion. */
GIMPLE_PASS, /* type */
"veclower", /* name */
OPTGROUP_VEC, /* optinfo_flags */
- true, /* has_execute */
TV_NONE, /* tv_id */
PROP_cfg, /* properties_required */
PROP_gimple_lvec, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */
- ( TODO_update_ssa
- | TODO_cleanup_cfg ), /* todo_flags_finish */
+ TODO_update_ssa, /* todo_flags_finish */
};
class pass_lower_vector : public gimple_opt_pass
GIMPLE_PASS, /* type */
"veclower2", /* name */
OPTGROUP_VEC, /* optinfo_flags */
- true, /* has_execute */
TV_NONE, /* tv_id */
PROP_cfg, /* properties_required */
PROP_gimple_lvec, /* properties_provided */
projects / gcc.git / blobdiff