/* Lower vector operations to scalar operations.
- Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010
- 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 "tree-flow.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 "gimple.h"
+#include "gimple-iterator.h"
+#include "gimplify-me.h"
+#include "gimple-ssa.h"
+#include "tree-cfg.h"
+#include "stringpool.h"
+#include "tree-ssanames.h"
#include "tree-iterator.h"
#include "tree-pass.h"
#include "flags.h"
-#include "ggc.h"
#include "diagnostic.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"
static tree
build_replicated_const (tree type, tree inner_type, HOST_WIDE_INT value)
{
- int width = tree_low_cst (TYPE_SIZE (inner_type), 1);
- int n = HOST_BITS_PER_WIDE_INT / width;
- unsigned HOST_WIDE_INT low, high, mask;
- tree ret;
+ int width = tree_to_uhwi (TYPE_SIZE (inner_type));
+ int n = (TYPE_PRECISION (type) + HOST_BITS_PER_WIDE_INT - 1)
+ / HOST_BITS_PER_WIDE_INT;
+ unsigned HOST_WIDE_INT low, mask;
+ HOST_WIDE_INT a[WIDE_INT_MAX_ELTS];
+ int i;
- gcc_assert (n);
+ gcc_assert (n && n <= WIDE_INT_MAX_ELTS);
if (width == HOST_BITS_PER_WIDE_INT)
low = value;
low = (unsigned HOST_WIDE_INT) ~0 / mask * (value & mask);
}
- if (TYPE_PRECISION (type) < HOST_BITS_PER_WIDE_INT)
- low &= ((HOST_WIDE_INT)1 << TYPE_PRECISION (type)) - 1, high = 0;
- else if (TYPE_PRECISION (type) == HOST_BITS_PER_WIDE_INT)
- high = 0;
- else if (TYPE_PRECISION (type) == 2 * HOST_BITS_PER_WIDE_INT)
- high = low;
- else
- gcc_unreachable ();
+ for (i = 0; i < n; i++)
+ a[i] = low;
- ret = build_int_cst_wide (type, low, high);
- return ret;
+ gcc_assert (TYPE_PRECISION (type) <= MAX_BITSIZE_MODE_ANY_INT);
+ return wide_int_to_tree
+ (type, wide_int::from_array (a, n, TYPE_PRECISION (type)));
}
static GTY(()) tree vector_inner_type;
do_binop (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b,
tree bitpos, tree bitsize, enum tree_code code)
{
- a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
- b = tree_vec_extract (gsi, inner_type, b, bitsize, bitpos);
+ 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);
return gimplify_build2 (gsi, code, inner_type, a, b);
}
tree type, tree inner_type,
tree a, tree b, enum tree_code code)
{
- VEC(constructor_elt,gc) *v;
+ vec<constructor_elt, va_gc> *v;
tree part_width = TYPE_SIZE (inner_type);
tree index = bitsize_int (0);
int nunits = TYPE_VECTOR_SUBPARTS (type);
- int delta = tree_low_cst (part_width, 1)
- / tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
+ int delta = tree_to_uhwi (part_width)
+ / tree_to_uhwi (TYPE_SIZE (TREE_TYPE (type)));
int i;
location_t loc = gimple_location (gsi_stmt (*gsi));
warning_at (loc, OPT_Wvector_operation_performance,
"vector operation will be expanded in parallel");
- v = VEC_alloc(constructor_elt, gc, (nunits + delta - 1) / delta);
+ vec_alloc (v, (nunits + delta - 1) / delta);
for (i = 0; i < nunits;
i += delta, index = int_const_binop (PLUS_EXPR, index, part_width))
{
tree result = f (gsi, inner_type, a, b, index, part_width, code);
- constructor_elt *ce = VEC_quick_push (constructor_elt, v, NULL);
- ce->index = NULL_TREE;
- ce->value = result;
+ constructor_elt ce = {NULL_TREE, result};
+ v->quick_push (ce);
}
return build_constructor (type, v);
enum tree_code code)
{
tree result, compute_type;
- enum machine_mode mode;
- int n_words = tree_low_cst (TYPE_SIZE_UNIT (type), 1) / UNITS_PER_WORD;
+ machine_mode mode;
+ int n_words = tree_to_uhwi (TYPE_SIZE_UNIT (type)) / UNITS_PER_WORD;
location_t loc = gimple_location (gsi_stmt (*gsi));
/* We have three strategies. If the type is already correct, just do
else
{
/* Use a single scalar operation with a mode no wider than word_mode. */
- mode = mode_for_size (tree_low_cst (TYPE_SIZE (type), 1), MODE_INT, 0);
+ mode = mode_for_size (tree_to_uhwi (TYPE_SIZE (type)), MODE_INT, 0);
compute_type = lang_hooks.types.type_for_mode (mode, 1);
result = f (gsi, compute_type, a, b, NULL_TREE, NULL_TREE, code);
warning_at (loc, OPT_Wvector_operation_performance,
tree type, tree a, tree b, enum tree_code code)
{
int parts_per_word = UNITS_PER_WORD
- / tree_low_cst (TYPE_SIZE_UNIT (TREE_TYPE (type)), 1);
+ / tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (type)));
if (INTEGRAL_TYPE_P (TREE_TYPE (type))
&& parts_per_word >= 4
a, b, code);
}
-/* Check if vector VEC consists of all the equal elements and
- that the number of elements corresponds to the type of VEC.
- The function returns first element of the vector
- or NULL_TREE if the vector is not uniform. */
+/* Try to expand vector comparison expression OP0 CODE OP1 by
+ querying optab if the following expression:
+ VEC_COND_EXPR< OP0 CODE OP1, {-1,...}, {0,...}>
+ can be expanded. */
static tree
-uniform_vector_p (tree vec)
+expand_vector_comparison (gimple_stmt_iterator *gsi, tree type, tree op0,
+ tree op1, enum tree_code code)
{
- tree first, t, els;
- unsigned i;
+ tree t;
+ if (! expand_vec_cond_expr_p (type, TREE_TYPE (op0)))
+ t = expand_vector_piecewise (gsi, do_compare, type,
+ TREE_TYPE (TREE_TYPE (op0)), op0, op1, code);
+ else
+ t = NULL_TREE;
- if (vec == NULL_TREE)
- return NULL_TREE;
+ return t;
+}
+
+/* Helper function of expand_vector_divmod. Gimplify a RSHIFT_EXPR in type
+ of OP0 with shift counts in SHIFTCNTS array and return the temporary holding
+ the result if successful, otherwise return NULL_TREE. */
+static tree
+add_rshift (gimple_stmt_iterator *gsi, tree type, tree op0, int *shiftcnts)
+{
+ optab op;
+ unsigned int i, nunits = TYPE_VECTOR_SUBPARTS (type);
+ bool scalar_shift = true;
- if (TREE_CODE (vec) == VECTOR_CST)
+ for (i = 1; i < nunits; i++)
{
- els = TREE_VECTOR_CST_ELTS (vec);
- first = TREE_VALUE (els);
- els = TREE_CHAIN (els);
+ if (shiftcnts[i] != shiftcnts[0])
+ scalar_shift = false;
+ }
- for (t = els; t; t = TREE_CHAIN (t))
- if (!operand_equal_p (first, TREE_VALUE (t), 0))
- return NULL_TREE;
+ if (scalar_shift && shiftcnts[0] == 0)
+ return op0;
- return first;
+ if (scalar_shift)
+ {
+ op = optab_for_tree_code (RSHIFT_EXPR, type, optab_scalar);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing)
+ return gimplify_build2 (gsi, RSHIFT_EXPR, type, op0,
+ build_int_cst (NULL_TREE, shiftcnts[0]));
}
- else if (TREE_CODE (vec) == CONSTRUCTOR)
+ op = optab_for_tree_code (RSHIFT_EXPR, type, optab_vector);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing)
{
- first = error_mark_node;
+ tree *vec = XALLOCAVEC (tree, nunits);
+ for (i = 0; i < nunits; i++)
+ vec[i] = build_int_cst (TREE_TYPE (type), shiftcnts[i]);
+ return gimplify_build2 (gsi, RSHIFT_EXPR, type, op0,
+ build_vector (type, vec));
+ }
- FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (vec), i, t)
- {
- if (i == 0)
- {
- first = t;
- continue;
- }
- if (!operand_equal_p (first, t, 0))
+ return NULL_TREE;
+}
+
+/* Try to expand integer vector division by constant using
+ widening multiply, shifts and additions. */
+static tree
+expand_vector_divmod (gimple_stmt_iterator *gsi, tree type, tree op0,
+ tree op1, enum tree_code code)
+{
+ bool use_pow2 = true;
+ bool has_vector_shift = true;
+ int mode = -1, this_mode;
+ int pre_shift = -1, post_shift;
+ unsigned int nunits = TYPE_VECTOR_SUBPARTS (type);
+ int *shifts = XALLOCAVEC (int, nunits * 4);
+ int *pre_shifts = shifts + nunits;
+ int *post_shifts = pre_shifts + nunits;
+ int *shift_temps = post_shifts + nunits;
+ unsigned HOST_WIDE_INT *mulc = XALLOCAVEC (unsigned HOST_WIDE_INT, nunits);
+ int prec = TYPE_PRECISION (TREE_TYPE (type));
+ int dummy_int;
+ unsigned int i;
+ signop sign_p = TYPE_SIGN (TREE_TYPE (type));
+ unsigned HOST_WIDE_INT mask = GET_MODE_MASK (TYPE_MODE (TREE_TYPE (type)));
+ tree *vec;
+ tree cur_op, mulcst, tem;
+ optab op;
+
+ if (prec > HOST_BITS_PER_WIDE_INT)
+ return NULL_TREE;
+
+ op = optab_for_tree_code (RSHIFT_EXPR, type, optab_vector);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ has_vector_shift = false;
+
+ /* Analysis phase. Determine if all op1 elements are either power
+ of two and it is possible to expand it using shifts (or for remainder
+ using masking). Additionally compute the multiplicative constants
+ and pre and post shifts if the division is to be expanded using
+ widening or high part multiplication plus shifts. */
+ for (i = 0; i < nunits; i++)
+ {
+ tree cst = VECTOR_CST_ELT (op1, i);
+ unsigned HOST_WIDE_INT ml;
+
+ if (TREE_CODE (cst) != INTEGER_CST || integer_zerop (cst))
+ return NULL_TREE;
+ pre_shifts[i] = 0;
+ post_shifts[i] = 0;
+ mulc[i] = 0;
+ if (use_pow2
+ && (!integer_pow2p (cst) || tree_int_cst_sgn (cst) != 1))
+ use_pow2 = false;
+ if (use_pow2)
+ {
+ shifts[i] = tree_log2 (cst);
+ if (shifts[i] != shifts[0]
+ && code == TRUNC_DIV_EXPR
+ && !has_vector_shift)
+ use_pow2 = false;
+ }
+ if (mode == -2)
+ continue;
+ if (sign_p == UNSIGNED)
+ {
+ unsigned HOST_WIDE_INT mh;
+ unsigned HOST_WIDE_INT d = TREE_INT_CST_LOW (cst) & mask;
+
+ if (d >= ((unsigned HOST_WIDE_INT) 1 << (prec - 1)))
+ /* FIXME: Can transform this into op0 >= op1 ? 1 : 0. */
return NULL_TREE;
- }
- if (i != TYPE_VECTOR_SUBPARTS (TREE_TYPE (vec)))
+
+ if (d <= 1)
+ {
+ mode = -2;
+ continue;
+ }
+
+ /* Find a suitable multiplier and right shift count
+ instead of multiplying with D. */
+ mh = choose_multiplier (d, prec, prec, &ml, &post_shift, &dummy_int);
+
+ /* If the suggested multiplier is more than SIZE bits, we can
+ do better for even divisors, using an initial right shift. */
+ if ((mh != 0 && (d & 1) == 0)
+ || (!has_vector_shift && pre_shift != -1))
+ {
+ if (has_vector_shift)
+ pre_shift = floor_log2 (d & -d);
+ else if (pre_shift == -1)
+ {
+ unsigned int j;
+ for (j = 0; j < nunits; j++)
+ {
+ tree cst2 = VECTOR_CST_ELT (op1, j);
+ unsigned HOST_WIDE_INT d2;
+ int this_pre_shift;
+
+ if (!tree_fits_uhwi_p (cst2))
+ return NULL_TREE;
+ d2 = tree_to_uhwi (cst2) & mask;
+ if (d2 == 0)
+ return NULL_TREE;
+ this_pre_shift = floor_log2 (d2 & -d2);
+ if (pre_shift == -1 || this_pre_shift < pre_shift)
+ pre_shift = this_pre_shift;
+ }
+ if (i != 0 && pre_shift != 0)
+ {
+ /* Restart. */
+ i = -1U;
+ mode = -1;
+ continue;
+ }
+ }
+ if (pre_shift != 0)
+ {
+ if ((d >> pre_shift) <= 1)
+ {
+ mode = -2;
+ continue;
+ }
+ mh = choose_multiplier (d >> pre_shift, prec,
+ prec - pre_shift,
+ &ml, &post_shift, &dummy_int);
+ gcc_assert (!mh);
+ pre_shifts[i] = pre_shift;
+ }
+ }
+ if (!mh)
+ this_mode = 0;
+ else
+ this_mode = 1;
+ }
+ else
+ {
+ HOST_WIDE_INT d = TREE_INT_CST_LOW (cst);
+ unsigned HOST_WIDE_INT abs_d;
+
+ if (d == -1)
+ return NULL_TREE;
+
+ /* Since d might be INT_MIN, we have to cast to
+ unsigned HOST_WIDE_INT before negating to avoid
+ undefined signed overflow. */
+ abs_d = (d >= 0
+ ? (unsigned HOST_WIDE_INT) d
+ : - (unsigned HOST_WIDE_INT) d);
+
+ /* n rem d = n rem -d */
+ if (code == TRUNC_MOD_EXPR && d < 0)
+ d = abs_d;
+ else if (abs_d == (unsigned HOST_WIDE_INT) 1 << (prec - 1))
+ {
+ /* This case is not handled correctly below. */
+ mode = -2;
+ continue;
+ }
+ if (abs_d <= 1)
+ {
+ mode = -2;
+ continue;
+ }
+
+ choose_multiplier (abs_d, prec, prec - 1, &ml,
+ &post_shift, &dummy_int);
+ if (ml >= (unsigned HOST_WIDE_INT) 1 << (prec - 1))
+ {
+ this_mode = 4 + (d < 0);
+ ml |= (~(unsigned HOST_WIDE_INT) 0) << (prec - 1);
+ }
+ else
+ this_mode = 2 + (d < 0);
+ }
+ mulc[i] = ml;
+ post_shifts[i] = post_shift;
+ if ((i && !has_vector_shift && post_shifts[0] != post_shift)
+ || post_shift >= prec
+ || pre_shifts[i] >= prec)
+ this_mode = -2;
+
+ if (i == 0)
+ mode = this_mode;
+ else if (mode != this_mode)
+ mode = -2;
+ }
+
+ vec = XALLOCAVEC (tree, nunits);
+
+ if (use_pow2)
+ {
+ tree addend = NULL_TREE;
+ if (sign_p == SIGNED)
+ {
+ tree uns_type;
+
+ /* Both division and remainder sequences need
+ op0 < 0 ? mask : 0 computed. It can be either computed as
+ (type) (((uns_type) (op0 >> (prec - 1))) >> (prec - shifts[i]))
+ if none of the shifts is 0, or as the conditional. */
+ for (i = 0; i < nunits; i++)
+ if (shifts[i] == 0)
+ break;
+ uns_type
+ = build_vector_type (build_nonstandard_integer_type (prec, 1),
+ nunits);
+ if (i == nunits && TYPE_MODE (uns_type) == TYPE_MODE (type))
+ {
+ for (i = 0; i < nunits; i++)
+ shift_temps[i] = prec - 1;
+ cur_op = add_rshift (gsi, type, op0, shift_temps);
+ if (cur_op != NULL_TREE)
+ {
+ cur_op = gimplify_build1 (gsi, VIEW_CONVERT_EXPR,
+ uns_type, cur_op);
+ for (i = 0; i < nunits; i++)
+ shift_temps[i] = prec - shifts[i];
+ cur_op = add_rshift (gsi, uns_type, cur_op, shift_temps);
+ if (cur_op != NULL_TREE)
+ addend = gimplify_build1 (gsi, VIEW_CONVERT_EXPR,
+ type, cur_op);
+ }
+ }
+ if (addend == NULL_TREE
+ && expand_vec_cond_expr_p (type, type))
+ {
+ tree zero, cst, cond;
+ gimple stmt;
+
+ zero = build_zero_cst (type);
+ cond = build2 (LT_EXPR, type, op0, zero);
+ for (i = 0; i < nunits; i++)
+ vec[i] = build_int_cst (TREE_TYPE (type),
+ ((unsigned HOST_WIDE_INT) 1
+ << shifts[i]) - 1);
+ cst = build_vector (type, vec);
+ addend = make_ssa_name (type);
+ stmt = gimple_build_assign (addend, VEC_COND_EXPR, cond,
+ cst, zero);
+ gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
+ }
+ }
+ if (code == TRUNC_DIV_EXPR)
+ {
+ if (sign_p == UNSIGNED)
+ {
+ /* q = op0 >> shift; */
+ cur_op = add_rshift (gsi, type, op0, shifts);
+ if (cur_op != NULL_TREE)
+ return cur_op;
+ }
+ else if (addend != NULL_TREE)
+ {
+ /* t1 = op0 + addend;
+ q = t1 >> shift; */
+ op = optab_for_tree_code (PLUS_EXPR, type, optab_default);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing)
+ {
+ cur_op = gimplify_build2 (gsi, PLUS_EXPR, type, op0, addend);
+ cur_op = add_rshift (gsi, type, cur_op, shifts);
+ if (cur_op != NULL_TREE)
+ return cur_op;
+ }
+ }
+ }
+ else
+ {
+ tree mask;
+ for (i = 0; i < nunits; i++)
+ vec[i] = build_int_cst (TREE_TYPE (type),
+ ((unsigned HOST_WIDE_INT) 1
+ << shifts[i]) - 1);
+ mask = build_vector (type, vec);
+ op = optab_for_tree_code (BIT_AND_EXPR, type, optab_default);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type)) != CODE_FOR_nothing)
+ {
+ if (sign_p == UNSIGNED)
+ /* r = op0 & mask; */
+ return gimplify_build2 (gsi, BIT_AND_EXPR, type, op0, mask);
+ else if (addend != NULL_TREE)
+ {
+ /* t1 = op0 + addend;
+ t2 = t1 & mask;
+ r = t2 - addend; */
+ op = optab_for_tree_code (PLUS_EXPR, type, optab_default);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type))
+ != CODE_FOR_nothing)
+ {
+ cur_op = gimplify_build2 (gsi, PLUS_EXPR, type, op0,
+ addend);
+ cur_op = gimplify_build2 (gsi, BIT_AND_EXPR, type,
+ cur_op, mask);
+ op = optab_for_tree_code (MINUS_EXPR, type,
+ optab_default);
+ if (op != unknown_optab
+ && optab_handler (op, TYPE_MODE (type))
+ != CODE_FOR_nothing)
+ return gimplify_build2 (gsi, MINUS_EXPR, type,
+ cur_op, addend);
+ }
+ }
+ }
+ }
+ }
+
+ if (mode == -2 || BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN)
+ return NULL_TREE;
+
+ if (!can_mult_highpart_p (TYPE_MODE (type), TYPE_UNSIGNED (type)))
+ return NULL_TREE;
+
+ cur_op = op0;
+
+ switch (mode)
+ {
+ case 0:
+ gcc_assert (sign_p == UNSIGNED);
+ /* t1 = oprnd0 >> pre_shift;
+ t2 = t1 h* ml;
+ q = t2 >> post_shift; */
+ cur_op = add_rshift (gsi, type, cur_op, pre_shifts);
+ if (cur_op == NULL_TREE)
return NULL_TREE;
+ break;
+ case 1:
+ gcc_assert (sign_p == UNSIGNED);
+ for (i = 0; i < nunits; i++)
+ {
+ shift_temps[i] = 1;
+ post_shifts[i]--;
+ }
+ break;
+ case 2:
+ case 3:
+ case 4:
+ case 5:
+ gcc_assert (sign_p == SIGNED);
+ for (i = 0; i < nunits; i++)
+ shift_temps[i] = prec - 1;
+ break;
+ default:
+ return NULL_TREE;
+ }
- return first;
+ for (i = 0; i < nunits; i++)
+ vec[i] = build_int_cst (TREE_TYPE (type), mulc[i]);
+ mulcst = build_vector (type, vec);
+
+ cur_op = gimplify_build2 (gsi, MULT_HIGHPART_EXPR, type, cur_op, mulcst);
+
+ switch (mode)
+ {
+ case 0:
+ /* t1 = oprnd0 >> pre_shift;
+ t2 = t1 h* ml;
+ q = t2 >> post_shift; */
+ cur_op = add_rshift (gsi, type, cur_op, post_shifts);
+ break;
+ case 1:
+ /* t1 = oprnd0 h* ml;
+ t2 = oprnd0 - t1;
+ t3 = t2 >> 1;
+ t4 = t1 + t3;
+ q = t4 >> (post_shift - 1); */
+ op = optab_for_tree_code (MINUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ tem = gimplify_build2 (gsi, MINUS_EXPR, type, op0, cur_op);
+ tem = add_rshift (gsi, type, tem, shift_temps);
+ op = optab_for_tree_code (PLUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ tem = gimplify_build2 (gsi, PLUS_EXPR, type, cur_op, tem);
+ cur_op = add_rshift (gsi, type, tem, post_shifts);
+ if (cur_op == NULL_TREE)
+ return NULL_TREE;
+ break;
+ case 2:
+ case 3:
+ case 4:
+ case 5:
+ /* t1 = oprnd0 h* ml;
+ t2 = t1; [ iff (mode & 2) != 0 ]
+ t2 = t1 + oprnd0; [ iff (mode & 2) == 0 ]
+ t3 = t2 >> post_shift;
+ t4 = oprnd0 >> (prec - 1);
+ q = t3 - t4; [ iff (mode & 1) == 0 ]
+ q = t4 - t3; [ iff (mode & 1) != 0 ] */
+ if ((mode & 2) == 0)
+ {
+ op = optab_for_tree_code (PLUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ cur_op = gimplify_build2 (gsi, PLUS_EXPR, type, cur_op, op0);
+ }
+ cur_op = add_rshift (gsi, type, cur_op, post_shifts);
+ if (cur_op == NULL_TREE)
+ return NULL_TREE;
+ tem = add_rshift (gsi, type, op0, shift_temps);
+ if (tem == NULL_TREE)
+ return NULL_TREE;
+ op = optab_for_tree_code (MINUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ if ((mode & 1) == 0)
+ cur_op = gimplify_build2 (gsi, MINUS_EXPR, type, cur_op, tem);
+ else
+ cur_op = gimplify_build2 (gsi, MINUS_EXPR, type, tem, cur_op);
+ break;
+ default:
+ gcc_unreachable ();
}
- return NULL_TREE;
+ if (code == TRUNC_DIV_EXPR)
+ return cur_op;
+
+ /* We divided. Now finish by:
+ t1 = q * oprnd1;
+ r = oprnd0 - t1; */
+ op = optab_for_tree_code (MULT_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ tem = gimplify_build2 (gsi, MULT_EXPR, type, cur_op, op1);
+ op = optab_for_tree_code (MINUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return NULL_TREE;
+ return gimplify_build2 (gsi, MINUS_EXPR, type, op0, tem);
}
-/* Try to expand vector comparison expression OP0 CODE OP1 by
- querying optab if the following expression:
- VEC_COND_EXPR< OP0 CODE OP1, {-1,...}, {0,...}>
- can be expanded. */
-static tree
-expand_vector_comparison (gimple_stmt_iterator *gsi, tree type, tree op0,
- tree op1, enum tree_code code)
+/* Expand a vector condition to scalars, by using many conditions
+ on the vector's elements. */
+static void
+expand_vector_condition (gimple_stmt_iterator *gsi)
{
- tree t;
- if (! expand_vec_cond_expr_p (type, TREE_TYPE (op0)))
- t = expand_vector_piecewise (gsi, do_compare, type,
- TREE_TYPE (TREE_TYPE (op0)), op0, op1, code);
- else
- t = NULL_TREE;
+ gassign *stmt = as_a <gassign *> (gsi_stmt (*gsi));
+ tree type = gimple_expr_type (stmt);
+ tree a = gimple_assign_rhs1 (stmt);
+ tree a1 = a;
+ tree a2;
+ bool a_is_comparison = false;
+ tree b = gimple_assign_rhs2 (stmt);
+ tree c = gimple_assign_rhs3 (stmt);
+ vec<constructor_elt, va_gc> *v;
+ tree constr;
+ tree inner_type = TREE_TYPE (type);
+ tree cond_type = TREE_TYPE (TREE_TYPE (a));
+ tree comp_inner_type = cond_type;
+ tree width = TYPE_SIZE (inner_type);
+ tree index = bitsize_int (0);
+ int nunits = TYPE_VECTOR_SUBPARTS (type);
+ int i;
+ location_t loc = gimple_location (gsi_stmt (*gsi));
- return t;
+ if (!is_gimple_val (a))
+ {
+ gcc_assert (COMPARISON_CLASS_P (a));
+ a_is_comparison = true;
+ a1 = TREE_OPERAND (a, 0);
+ a2 = TREE_OPERAND (a, 1);
+ comp_inner_type = TREE_TYPE (TREE_TYPE (a1));
+ }
+
+ if (expand_vec_cond_expr_p (type, TREE_TYPE (a1)))
+ return;
+
+ /* TODO: try and find a smaller vector type. */
+
+ warning_at (loc, OPT_Wvector_operation_performance,
+ "vector condition will be expanded piecewise");
+
+ vec_alloc (v, nunits);
+ for (i = 0; i < nunits;
+ i++, index = int_const_binop (PLUS_EXPR, index, width))
+ {
+ tree aa, result;
+ tree bb = tree_vec_extract (gsi, inner_type, b, width, index);
+ tree cc = tree_vec_extract (gsi, inner_type, c, width, index);
+ if (a_is_comparison)
+ {
+ tree aa1 = tree_vec_extract (gsi, comp_inner_type, a1, width, index);
+ tree aa2 = tree_vec_extract (gsi, comp_inner_type, a2, width, index);
+ aa = build2 (TREE_CODE (a), cond_type, aa1, aa2);
+ }
+ else
+ aa = tree_vec_extract (gsi, cond_type, a, width, index);
+ result = gimplify_build3 (gsi, COND_EXPR, inner_type, aa, bb, cc);
+ constructor_elt ce = {NULL_TREE, result};
+ v->quick_push (ce);
+ }
+
+ constr = build_constructor (type, v);
+ gimple_assign_set_rhs_from_tree (gsi, constr);
+ update_stmt (gsi_stmt (*gsi));
}
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);
return expand_vector_comparison (gsi, type, rhs1, rhs2, code);
}
+
+ case TRUNC_DIV_EXPR:
+ case TRUNC_MOD_EXPR:
+ {
+ tree rhs1 = gimple_assign_rhs1 (assign);
+ tree rhs2 = gimple_assign_rhs2 (assign);
+ tree ret;
+
+ if (!optimize
+ || !VECTOR_INTEGER_TYPE_P (type)
+ || TREE_CODE (rhs2) != VECTOR_CST
+ || !VECTOR_MODE_P (TYPE_MODE (type)))
+ break;
+
+ ret = expand_vector_divmod (gsi, type, rhs1, rhs2, code);
+ if (ret != NULL_TREE)
+ return ret;
+ break;
+ }
+
default:
break;
}
gimple_assign_rhs1 (assign),
gimple_assign_rhs2 (assign), code);
}
+
+/* Try to optimize
+ a_5 = { b_7, b_7 + 3, b_7 + 6, b_7 + 9 };
+ style stmts into:
+ _9 = { b_7, b_7, b_7, b_7 };
+ a_5 = _9 + { 0, 3, 6, 9 };
+ because vector splat operation is usually more efficient
+ than piecewise initialization of the vector. */
+
+static void
+optimize_vector_constructor (gimple_stmt_iterator *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);
+ unsigned int i, j, nelts = TYPE_VECTOR_SUBPARTS (type);
+ bool all_same = true;
+ constructor_elt *elt;
+ tree *cst;
+ gimple g;
+ tree base = NULL_TREE;
+ optab op;
+
+ if (nelts <= 2 || CONSTRUCTOR_NELTS (rhs) != nelts)
+ return;
+ op = optab_for_tree_code (PLUS_EXPR, type, optab_default);
+ if (op == unknown_optab
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ return;
+ FOR_EACH_VEC_SAFE_ELT (CONSTRUCTOR_ELTS (rhs), i, elt)
+ if (TREE_CODE (elt->value) != SSA_NAME
+ || TREE_CODE (TREE_TYPE (elt->value)) == VECTOR_TYPE)
+ return;
+ else
+ {
+ tree this_base = elt->value;
+ if (this_base != CONSTRUCTOR_ELT (rhs, 0)->value)
+ all_same = false;
+ for (j = 0; j < nelts + 1; j++)
+ {
+ g = SSA_NAME_DEF_STMT (this_base);
+ if (is_gimple_assign (g)
+ && gimple_assign_rhs_code (g) == PLUS_EXPR
+ && TREE_CODE (gimple_assign_rhs2 (g)) == INTEGER_CST
+ && TREE_CODE (gimple_assign_rhs1 (g)) == SSA_NAME
+ && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (gimple_assign_rhs1 (g)))
+ this_base = gimple_assign_rhs1 (g);
+ else
+ break;
+ }
+ if (i == 0)
+ base = this_base;
+ else if (this_base != base)
+ return;
+ }
+ if (all_same)
+ return;
+ cst = XALLOCAVEC (tree, nelts);
+ for (i = 0; i < nelts; i++)
+ {
+ tree this_base = CONSTRUCTOR_ELT (rhs, i)->value;;
+ cst[i] = build_zero_cst (TREE_TYPE (base));
+ while (this_base != base)
+ {
+ g = SSA_NAME_DEF_STMT (this_base);
+ cst[i] = fold_binary (PLUS_EXPR, TREE_TYPE (base),
+ cst[i], gimple_assign_rhs2 (g));
+ if (cst[i] == NULL_TREE
+ || TREE_CODE (cst[i]) != INTEGER_CST
+ || TREE_OVERFLOW (cst[i]))
+ return;
+ this_base = gimple_assign_rhs1 (g);
+ }
+ }
+ for (i = 0; i < nelts; i++)
+ CONSTRUCTOR_ELT (rhs, i)->value = base;
+ g = gimple_build_assign (make_ssa_name (type), rhs);
+ gsi_insert_before (gsi, g, GSI_SAME_STMT);
+ g = gimple_build_assign (lhs, PLUS_EXPR, gimple_assign_lhs (g),
+ build_vector (type, cst));
+ gsi_replace (gsi, g, false);
+}
\f
-/* Return a type for the widest vector mode whose components are of mode
- INNER_MODE, or NULL_TREE if none is found.
- SATP is true for saturating fixed-point types. */
+/* Return a type for the widest vector mode whose components are of type
+ TYPE, or NULL_TREE if none is found. */
static tree
-type_for_widest_vector_mode (enum machine_mode inner_mode, optab op, int satp)
+type_for_widest_vector_mode (tree type, optab op)
{
- 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))
if (best_mode == VOIDmode)
return NULL_TREE;
else
- {
- /* For fixed-point modes, we need to pass satp as the 2nd parameter. */
- if (ALL_FIXED_POINT_MODE_P (best_mode))
- return lang_hooks.types.type_for_mode (best_mode, satp);
-
- return lang_hooks.types.type_for_mode (best_mode, 1);
- }
+ return build_vector_type_for_mode (type, best_mode);
}
returns either the element itself, either BIT_FIELD_REF, or an
ARRAY_REF expression.
- GSI is requred to insert temporary variables while building a
+ GSI is required to insert temporary variables while building a
refernece to the element of the vector VECT.
PTMPVEC is a pointer to the temporary variable for caching
/* Given that we're about to compute a binary modulus,
we don't care about the high bits of the value. */
index = TREE_INT_CST_LOW (idx);
- if (!host_integerp (idx, 1) || index >= elements)
+ if (!tree_fits_uhwi_p (idx) || index >= elements)
{
index &= elements - 1;
idx = build_int_cst (TREE_TYPE (idx), index);
}
if (TREE_CODE (vect) == VECTOR_CST)
+ return VECTOR_CST_ELT (vect, index);
+ else if (TREE_CODE (vect) == CONSTRUCTOR
+ && (CONSTRUCTOR_NELTS (vect) == 0
+ || TREE_CODE (TREE_TYPE (CONSTRUCTOR_ELT (vect, 0)->value))
+ != VECTOR_TYPE))
{
- unsigned i;
- tree vals = TREE_VECTOR_CST_ELTS (vect);
- for (i = 0; vals; vals = TREE_CHAIN (vals), ++i)
- if (i == index)
- return TREE_VALUE (vals);
- return build_zero_cst (vect_elt_type);
- }
- else if (TREE_CODE (vect) == CONSTRUCTOR)
- {
- unsigned i;
- tree elt_i, elt_v;
-
- FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (vect), i, elt_i, elt_v)
- if (operand_equal_p (elt_i, idx, 0))
- return elt_v;
+ if (index < CONSTRUCTOR_NELTS (vect))
+ return CONSTRUCTOR_ELT (vect, index)->value;
return build_zero_cst (vect_elt_type);
}
else
{
tree size = TYPE_SIZE (vect_elt_type);
- tree pos = fold_build2 (MULT_EXPR, TREE_TYPE (idx), idx, size);
- return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos);
+ tree pos = fold_build2 (MULT_EXPR, bitsizetype, bitsize_int (index),
+ size);
+ return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos);
}
}
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);
tree vect_elt_type = TREE_TYPE (vect_type);
tree mask_elt_type = TREE_TYPE (mask_type);
unsigned int elements = TYPE_VECTOR_SUBPARTS (vect_type);
- VEC(constructor_elt,gc) *v;
+ vec<constructor_elt, va_gc> *v;
tree constr, t, si, i_val;
tree vec0tmp = NULL_TREE, vec1tmp = NULL_TREE, masktmp = NULL_TREE;
bool two_operand_p = !operand_equal_p (vec0, vec1, 0);
location_t loc = gimple_location (gsi_stmt (*gsi));
unsigned i;
- if (expand_vec_perm_expr_p (TYPE_MODE (vect_type), vec0, vec1, mask))
+ if (TREE_CODE (mask) == SSA_NAME)
+ {
+ gimple def_stmt = SSA_NAME_DEF_STMT (mask);
+ if (is_gimple_assign (def_stmt)
+ && gimple_assign_rhs_code (def_stmt) == VECTOR_CST)
+ mask = gimple_assign_rhs1 (def_stmt);
+ }
+
+ if (TREE_CODE (mask) == VECTOR_CST)
+ {
+ unsigned char *sel_int = XALLOCAVEC (unsigned char, elements);
+
+ for (i = 0; i < elements; ++i)
+ sel_int[i] = (TREE_INT_CST_LOW (VECTOR_CST_ELT (mask, i))
+ & (2 * elements - 1));
+
+ if (can_vec_perm_p (TYPE_MODE (vect_type), false, sel_int))
+ {
+ gimple_assign_set_rhs3 (stmt, mask);
+ update_stmt (stmt);
+ return;
+ }
+ }
+ else if (can_vec_perm_p (TYPE_MODE (vect_type), true, NULL))
return;
warning_at (loc, OPT_Wvector_operation_performance,
"vector shuffling operation will be expanded piecewise");
-
- v = VEC_alloc (constructor_elt, gc, elements);
+ vec_alloc (v, elements);
for (i = 0; i < elements; i++)
{
si = size_int (i);
unsigned HOST_WIDE_INT index;
index = TREE_INT_CST_LOW (i_val);
- if (!host_integerp (i_val, 1) || index >= elements)
+ if (!tree_fits_uhwi_p (i_val) || index >= elements)
i_val = build_int_cst (mask_elt_type, index & (elements - 1));
if (two_operand_p && (index & elements) != 0)
t = v0_val;
}
- CONSTRUCTOR_APPEND_ELT (v, si, t);
+ CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, t);
}
constr = build_constructor (vect_type, v);
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 = NULL;
+ 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 == VEC_COND_EXPR)
+ {
+ expand_vector_condition (gsi);
+ 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)))
+ && !gimple_clobber_p (stmt)
+ && optimize)
+ {
+ optimize_vector_constructor (gsi);
+ return;
+ }
+
if (rhs_class != GIMPLE_UNARY_RHS && rhs_class != GIMPLE_BINARY_RHS)
return;
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)
type = TREE_TYPE (rhs1);
+ /* For widening/narrowing vector operations, the relevant type is of the
+ arguments, not the widened result. VEC_UNPACK_FLOAT_*_EXPR is
+ calculated in the same way above. */
+ if (code == WIDEN_SUM_EXPR
+ || code == VEC_WIDEN_MULT_HI_EXPR
+ || code == VEC_WIDEN_MULT_LO_EXPR
+ || code == VEC_WIDEN_MULT_EVEN_EXPR
+ || code == VEC_WIDEN_MULT_ODD_EXPR
+ || code == VEC_UNPACK_HI_EXPR
+ || code == VEC_UNPACK_LO_EXPR
+ || code == VEC_PACK_TRUNC_EXPR
+ || code == VEC_PACK_SAT_EXPR
+ || code == VEC_PACK_FIX_TRUNC_EXPR
+ || code == VEC_WIDEN_LSHIFT_HI_EXPR
+ || code == VEC_WIDEN_LSHIFT_LO_EXPR)
+ type = TREE_TYPE (rhs1);
+
/* Choose between vector shift/rotate by vector and vector shift/rotate by
scalar */
if (code == LSHIFT_EXPR
|| code == LROTATE_EXPR
|| code == RROTATE_EXPR)
{
- bool vector_scalar_shift;
- op = optab_for_tree_code (code, type, optab_scalar);
-
- /* Vector/Scalar shift is supported. */
- vector_scalar_shift = (op && (optab_handler (op, TYPE_MODE (type))
- != CODE_FOR_nothing));
+ optab opv;
- /* If the 2nd argument is vector, we need a vector/vector shift.
- Except all the elements in the second vector are the same. */
- if (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (rhs2))))
+ /* Check whether we have vector <op> {x,x,x,x} where x
+ could be a scalar variable or a constant. Transform
+ vector <op> {x,x,x,x} ==> vector <op> scalar. */
+ if (VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2)))
{
tree first;
gimple def_stmt;
- /* Check whether we have vector <op> {x,x,x,x} where x
- could be a scalar variable or a constant. Transform
- vector <op> {x,x,x,x} ==> vector <op> scalar. */
- if (vector_scalar_shift
- && ((TREE_CODE (rhs2) == VECTOR_CST
- && (first = uniform_vector_p (rhs2)) != NULL_TREE)
- || (TREE_CODE (rhs2) == SSA_NAME
- && (def_stmt = SSA_NAME_DEF_STMT (rhs2))
- && gimple_assign_single_p (def_stmt)
- && (first = uniform_vector_p
- (gimple_assign_rhs1 (def_stmt))) != NULL_TREE)))
+ if ((TREE_CODE (rhs2) == VECTOR_CST
+ && (first = uniform_vector_p (rhs2)) != NULL_TREE)
+ || (TREE_CODE (rhs2) == SSA_NAME
+ && (def_stmt = SSA_NAME_DEF_STMT (rhs2))
+ && gimple_assign_single_p (def_stmt)
+ && (first = uniform_vector_p
+ (gimple_assign_rhs1 (def_stmt))) != NULL_TREE))
{
gimple_assign_set_rhs2 (stmt, first);
update_stmt (stmt);
rhs2 = first;
}
- else
- op = optab_for_tree_code (code, type, optab_vector);
}
- /* Try for a vector/scalar shift, and if we don't have one, see if we
- have a vector/vector shift */
- else if (!vector_scalar_shift)
+ opv = optab_for_tree_code (code, type, optab_vector);
+ if (VECTOR_INTEGER_TYPE_P (TREE_TYPE (rhs2)))
+ op = opv;
+ else
{
- op = optab_for_tree_code (code, type, optab_vector);
+ 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. 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 (op && (optab_handler (op, TYPE_MODE (type))
- != CODE_FOR_nothing))
+ 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)))
{
- /* Transform vector <op> scalar => vector <op> {x,x,x,x}. */
- int n_parts = TYPE_VECTOR_SUBPARTS (type);
- int part_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
- tree part_type = lang_hooks.types.type_for_size (part_size, 1);
- tree vect_type = build_vector_type (part_type, n_parts);
-
- rhs2 = fold_convert (part_type, rhs2);
- rhs2 = build_vector_from_val (vect_type, rhs2);
- gimple_assign_set_rhs2 (stmt, rhs2);
- update_stmt (stmt);
+ 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
op = optab_for_tree_code (code, type, optab_default);
- /* For widening/narrowing vector operations, the relevant type is of the
- arguments, not the widened result. VEC_UNPACK_FLOAT_*_EXPR is
- calculated in the same way above. */
- if (code == WIDEN_SUM_EXPR
- || code == VEC_WIDEN_MULT_HI_EXPR
- || code == VEC_WIDEN_MULT_LO_EXPR
- || code == VEC_UNPACK_HI_EXPR
- || code == VEC_UNPACK_LO_EXPR
- || code == VEC_PACK_TRUNC_EXPR
- || code == VEC_PACK_SAT_EXPR
- || code == VEC_PACK_FIX_TRUNC_EXPR)
- type = TREE_TYPE (rhs1);
-
/* Optabs will try converting a negation into a subtraction, so
look for it as well. TODO: negation of floating-point vectors
might be turned into an exclusive OR toggling the sign bit. */
- if (op == NULL
+ if (op == unknown_optab
&& code == NEGATE_EXPR
&& 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 (TYPE_MODE (type) == BLKmode && op)
- {
- tree vector_compute_type
- = type_for_widest_vector_mode (TYPE_MODE (TREE_TYPE (type)), op,
- TYPE_SATURATING (TREE_TYPE (type)));
- if (vector_compute_type != NULL_TREE
- && (TYPE_VECTOR_SUBPARTS (vector_compute_type)
- < TYPE_VECTOR_SUBPARTS (compute_type)))
- 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 ((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)
- && op != NULL
- && optab_handler (op, compute_mode) != CODE_FOR_nothing)
- return;
- else
- /* 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. */
/* Use this to lower vector operations introduced by the vectorizer,
if it may need the bit-twiddling tricks implemented in this file. */
-static bool
-gate_expand_vector_operations_ssa (void)
-{
- return optimize == 0;
-}
-
static unsigned int
expand_vector_operations (void)
{
basic_block bb;
bool cfg_changed = false;
- FOR_EACH_BB (bb)
+ FOR_EACH_BB_FN (bb, cfun)
{
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
{
return cfg_changed ? TODO_cleanup_cfg : 0;
}
-struct gimple_opt_pass pass_lower_vector =
-{
- {
- GIMPLE_PASS,
- "veclower", /* name */
- gate_expand_vector_operations_ssa, /* gate */
- expand_vector_operations, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_NONE, /* tv_id */
- PROP_cfg, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- 0, /* todo_flags_start */
- TODO_update_ssa /* todo_flags_finish */
- | TODO_verify_ssa
- | TODO_verify_stmts | TODO_verify_flow
- | TODO_cleanup_cfg
- }
+namespace {
+
+const pass_data pass_data_lower_vector =
+{
+ GIMPLE_PASS, /* type */
+ "veclower", /* name */
+ OPTGROUP_VEC, /* optinfo_flags */
+ TV_NONE, /* tv_id */
+ PROP_cfg, /* properties_required */
+ PROP_gimple_lvec, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_update_ssa, /* todo_flags_finish */
};
-struct gimple_opt_pass pass_lower_vector_ssa =
-{
- {
- GIMPLE_PASS,
- "veclower2", /* name */
- 0, /* gate */
- expand_vector_operations, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_NONE, /* tv_id */
- PROP_cfg, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- 0, /* todo_flags_start */
- TODO_update_ssa /* todo_flags_finish */
- | TODO_verify_ssa
- | TODO_verify_stmts | TODO_verify_flow
- | TODO_cleanup_cfg
- }
+class pass_lower_vector : public gimple_opt_pass
+{
+public:
+ pass_lower_vector (gcc::context *ctxt)
+ : gimple_opt_pass (pass_data_lower_vector, ctxt)
+ {}
+
+ /* opt_pass methods: */
+ virtual bool gate (function *fun)
+ {
+ return !(fun->curr_properties & PROP_gimple_lvec);
+ }
+
+ virtual unsigned int execute (function *)
+ {
+ return expand_vector_operations ();
+ }
+
+}; // class pass_lower_vector
+
+} // anon namespace
+
+gimple_opt_pass *
+make_pass_lower_vector (gcc::context *ctxt)
+{
+ return new pass_lower_vector (ctxt);
+}
+
+namespace {
+
+const pass_data pass_data_lower_vector_ssa =
+{
+ GIMPLE_PASS, /* type */
+ "veclower2", /* name */
+ OPTGROUP_VEC, /* optinfo_flags */
+ 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 */
};
+class pass_lower_vector_ssa : public gimple_opt_pass
+{
+public:
+ pass_lower_vector_ssa (gcc::context *ctxt)
+ : gimple_opt_pass (pass_data_lower_vector_ssa, ctxt)
+ {}
+
+ /* opt_pass methods: */
+ opt_pass * clone () { return new pass_lower_vector_ssa (m_ctxt); }
+ virtual unsigned int execute (function *)
+ {
+ return expand_vector_operations ();
+ }
+
+}; // class pass_lower_vector_ssa
+
+} // anon namespace
+
+gimple_opt_pass *
+make_pass_lower_vector_ssa (gcc::context *ctxt)
+{
+ return new pass_lower_vector_ssa (ctxt);
+}
+
#include "gt-tree-vect-generic.h"