/* Expand the basic unary and binary arithmetic operations, for GNU compiler.
Copyright (C) 1987, 1988, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
- 1999, 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
+ 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
This file is part of GCC.
case REALIGN_LOAD_EXPR:
return vec_realign_load_optab;
+ case WIDEN_SUM_EXPR:
+ return TYPE_UNSIGNED (type) ? usum_widen_optab : ssum_widen_optab;
+
+ case DOT_PROD_EXPR:
+ return TYPE_UNSIGNED (type) ? udot_prod_optab : sdot_prod_optab;
+
case REDUC_MAX_EXPR:
return TYPE_UNSIGNED (type) ? reduc_umax_optab : reduc_smax_optab;
}
\f
+/* Expand vector widening operations.
+
+ There are two different classes of operations handled here:
+ 1) Operations whose result is wider than all the arguments to the operation.
+ Examples: VEC_UNPACK_HI/LO_EXPR, VEC_WIDEN_MULT_HI/LO_EXPR
+ In this case OP0 and optionally OP1 would be initialized,
+ but WIDE_OP wouldn't (not relevant for this case).
+ 2) Operations whose result is of the same size as the last argument to the
+ operation, but wider than all the other arguments to the operation.
+ Examples: WIDEN_SUM_EXPR, VEC_DOT_PROD_EXPR.
+ In the case WIDE_OP, OP0 and optionally OP1 would be initialized.
+
+ E.g, when called to expand the following operations, this is how
+ the arguments will be initialized:
+ nops OP0 OP1 WIDE_OP
+ widening-sum 2 oprnd0 - oprnd1
+ widening-dot-product 3 oprnd0 oprnd1 oprnd2
+ widening-mult 2 oprnd0 oprnd1 -
+ type-promotion (vec-unpack) 1 oprnd0 - - */
+
+rtx
+expand_widen_pattern_expr (tree exp, rtx op0, rtx op1, rtx wide_op, rtx target,
+ int unsignedp)
+{
+ tree oprnd0, oprnd1, oprnd2;
+ enum machine_mode wmode = 0, tmode0, tmode1 = 0;
+ optab widen_pattern_optab;
+ int icode;
+ enum machine_mode xmode0, xmode1 = 0, wxmode = 0;
+ rtx temp;
+ rtx pat;
+ rtx xop0, xop1, wxop;
+ int nops = TREE_CODE_LENGTH (TREE_CODE (exp));
+
+ oprnd0 = TREE_OPERAND (exp, 0);
+ tmode0 = TYPE_MODE (TREE_TYPE (oprnd0));
+ widen_pattern_optab =
+ optab_for_tree_code (TREE_CODE (exp), TREE_TYPE (oprnd0));
+ icode = (int) widen_pattern_optab->handlers[(int) tmode0].insn_code;
+ gcc_assert (icode != CODE_FOR_nothing);
+ xmode0 = insn_data[icode].operand[1].mode;
+
+ if (nops >= 2)
+ {
+ oprnd1 = TREE_OPERAND (exp, 1);
+ tmode1 = TYPE_MODE (TREE_TYPE (oprnd1));
+ xmode1 = insn_data[icode].operand[2].mode;
+ }
+
+ /* The last operand is of a wider mode than the rest of the operands. */
+ if (nops == 2)
+ {
+ wmode = tmode1;
+ wxmode = xmode1;
+ }
+ else if (nops == 3)
+ {
+ gcc_assert (tmode1 == tmode0);
+ gcc_assert (op1);
+ oprnd2 = TREE_OPERAND (exp, 2);
+ wmode = TYPE_MODE (TREE_TYPE (oprnd2));
+ wxmode = insn_data[icode].operand[3].mode;
+ }
+
+ if (!wide_op)
+ wmode = wxmode = insn_data[icode].operand[0].mode;
+
+ if (!target
+ || ! (*insn_data[icode].operand[0].predicate) (target, wmode))
+ temp = gen_reg_rtx (wmode);
+ else
+ temp = target;
+
+ xop0 = op0;
+ xop1 = op1;
+ wxop = wide_op;
+
+ /* In case the insn wants input operands in modes different from
+ those of the actual operands, convert the operands. It would
+ seem that we don't need to convert CONST_INTs, but we do, so
+ that they're properly zero-extended, sign-extended or truncated
+ for their mode. */
+
+ if (GET_MODE (op0) != xmode0 && xmode0 != VOIDmode)
+ xop0 = convert_modes (xmode0,
+ GET_MODE (op0) != VOIDmode
+ ? GET_MODE (op0)
+ : tmode0,
+ xop0, unsignedp);
+
+ if (op1)
+ if (GET_MODE (op1) != xmode1 && xmode1 != VOIDmode)
+ xop1 = convert_modes (xmode1,
+ GET_MODE (op1) != VOIDmode
+ ? GET_MODE (op1)
+ : tmode1,
+ xop1, unsignedp);
+
+ if (wide_op)
+ if (GET_MODE (wide_op) != wxmode && wxmode != VOIDmode)
+ wxop = convert_modes (wxmode,
+ GET_MODE (wide_op) != VOIDmode
+ ? GET_MODE (wide_op)
+ : wmode,
+ wxop, unsignedp);
+
+ /* Now, if insn's predicates don't allow our operands, put them into
+ pseudo regs. */
+
+ if (! (*insn_data[icode].operand[1].predicate) (xop0, xmode0)
+ && xmode0 != VOIDmode)
+ xop0 = copy_to_mode_reg (xmode0, xop0);
+
+ if (op1)
+ {
+ if (! (*insn_data[icode].operand[2].predicate) (xop1, xmode1)
+ && xmode1 != VOIDmode)
+ xop1 = copy_to_mode_reg (xmode1, xop1);
+
+ if (wide_op)
+ {
+ if (! (*insn_data[icode].operand[3].predicate) (wxop, wxmode)
+ && wxmode != VOIDmode)
+ wxop = copy_to_mode_reg (wxmode, wxop);
+
+ pat = GEN_FCN (icode) (temp, xop0, xop1, wxop);
+ }
+ else
+ pat = GEN_FCN (icode) (temp, xop0, xop1);
+ }
+ else
+ {
+ if (wide_op)
+ {
+ if (! (*insn_data[icode].operand[2].predicate) (wxop, wxmode)
+ && wxmode != VOIDmode)
+ wxop = copy_to_mode_reg (wxmode, wxop);
+
+ pat = GEN_FCN (icode) (temp, xop0, wxop);
+ }
+ else
+ pat = GEN_FCN (icode) (temp, xop0);
+ }
+
+ emit_insn (pat);
+ return temp;
+}
+
/* Generate code to perform an operation specified by TERNARY_OPTAB
on operands OP0, OP1 and OP2, with result having machine-mode MODE.
enum optab_methods methods)
{
if (CONSTANT_P (op0) && CONSTANT_P (op1))
- return simplify_gen_binary (binoptab->code, mode, op0, op1);
- else
- return expand_binop (mode, binoptab, op0, op1, target, unsignedp, methods);
+ {
+ rtx x = simplify_binary_operation (binoptab->code, mode, op0, op1);
+
+ if (x)
+ return x;
+ }
+
+ return expand_binop (mode, binoptab, op0, op1, target, unsignedp, methods);
}
/* Like simplify_expand_binop, but always put the result in TARGET.
subword_label = gen_label_rtx ();
done_label = gen_label_rtx ();
+ NO_DEFER_POP;
do_compare_rtx_and_jump (cmp1, cmp2, cmp_code, false, op1_mode,
0, 0, subword_label);
+ OK_DEFER_POP;
if (!expand_superword_shift (binoptab, outof_input, superword_op1,
outof_target, into_target,
if (expand_doubleword_shift (op1_mode, binoptab,
outof_input, into_input, op1,
outof_target, into_target,
- unsignedp, methods, shift_mask))
+ unsignedp, next_methods, shift_mask))
{
insns = get_insns ();
end_sequence ();
&& ashl_optab->handlers[(int) word_mode].insn_code != CODE_FOR_nothing
&& lshr_optab->handlers[(int) word_mode].insn_code != CODE_FOR_nothing)
{
- rtx insns, equiv_value;
+ rtx insns;
rtx into_target, outof_target;
rtx into_input, outof_input;
rtx inter;
if (inter != 0)
{
- if (binoptab->code != UNKNOWN)
- equiv_value = gen_rtx_fmt_ee (binoptab->code, mode, op0, op1);
- else
- equiv_value = 0;
-
- /* We can't make this a no conflict block if this is a word swap,
- because the word swap case fails if the input and output values
- are in the same register. */
- if (shift_count != BITS_PER_WORD)
- emit_no_conflict_block (insns, target, op0, op1, equiv_value);
- else
- emit_insn (insns);
-
-
+ /* One may be tempted to wrap the insns in a REG_NO_CONFLICT
+ block to help the register allocator a bit. But a multi-word
+ rotate will need all the input bits when setting the output
+ bits, so there clearly is a conflict between the input and
+ output registers. So we can't use a no-conflict block here. */
+ emit_insn (insns);
return target;
}
}
target = gen_reg_rtx (outmode);
emit_libcall_block (insns, target, value,
- gen_rtx_fmt_e (unoptab->code, mode, op0));
+ gen_rtx_fmt_e (unoptab->code, outmode, op0));
return target;
}
emit_move_insn (target, op0);
NO_DEFER_POP;
- /* If this mode is an integer too wide to compare properly,
- compare word by word. Rely on CSE to optimize constant cases. */
- if (GET_MODE_CLASS (mode) == MODE_INT
- && ! can_compare_p (GE, mode, ccp_jump))
- do_jump_by_parts_greater_rtx (mode, 0, target, const0_rtx,
- NULL_RTX, op1);
- else
- do_compare_rtx_and_jump (target, CONST0_RTX (mode), GE, 0, mode,
- NULL_RTX, NULL_RTX, op1);
+ do_compare_rtx_and_jump (target, CONST0_RTX (mode), GE, 0, mode,
+ NULL_RTX, NULL_RTX, op1);
op0 = expand_unop (mode, result_unsignedp ? neg_optab : negv_optab,
target, target, 0);
p->must_stay = true;
}
+/* Encapsulate the block starting at FIRST and ending with LAST, which is
+ logically equivalent to EQUIV, so it gets manipulated as a unit if it
+ is possible to do so. */
+
+static void
+maybe_encapsulate_block (rtx first, rtx last, rtx equiv)
+{
+ if (!flag_non_call_exceptions || !may_trap_p (equiv))
+ {
+ /* We can't attach the REG_LIBCALL and REG_RETVAL notes when the
+ encapsulated region would not be in one basic block, i.e. when
+ there is a control_flow_insn_p insn between FIRST and LAST. */
+ bool attach_libcall_retval_notes = true;
+ rtx insn, next = NEXT_INSN (last);
+
+ for (insn = first; insn != next; insn = NEXT_INSN (insn))
+ if (control_flow_insn_p (insn))
+ {
+ attach_libcall_retval_notes = false;
+ break;
+ }
+
+ if (attach_libcall_retval_notes)
+ {
+ REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last,
+ REG_NOTES (first));
+ REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first,
+ REG_NOTES (last));
+ }
+ }
+}
+
/* Emit code to perform a series of operations on a multi-word quantity, one
word at a time.
else
first = NEXT_INSN (prev);
- /* Encapsulate the block so it gets manipulated as a unit. */
- REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last,
- REG_NOTES (first));
- REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first, REG_NOTES (last));
+ maybe_encapsulate_block (first, last, equiv);
return last;
}
else
first = NEXT_INSN (prev);
- /* Encapsulate the block so it gets manipulated as a unit. */
- if (!flag_non_call_exceptions || !may_trap_p (equiv))
- {
- /* We can't attach the REG_LIBCALL and REG_RETVAL notes
- when the encapsulated region would not be in one basic block,
- i.e. when there is a control_flow_insn_p insn between FIRST and LAST.
- */
- bool attach_libcall_retval_notes = true;
- next = NEXT_INSN (last);
- for (insn = first; insn != next; insn = NEXT_INSN (insn))
- if (control_flow_insn_p (insn))
- {
- attach_libcall_retval_notes = false;
- break;
- }
-
- if (attach_libcall_retval_notes)
- {
- REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last,
- REG_NOTES (first));
- REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first,
- REG_NOTES (last));
- }
- }
+ maybe_encapsulate_block (first, last, equiv);
}
\f
/* Nonzero if we can perform a comparison of mode MODE straightforwardly.
result = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST_MAKE_BLOCK,
word_mode, 2, x, mode, y, mode);
+ /* There are two kinds of comparison routines. Biased routines
+ return 0/1/2, and unbiased routines return -1/0/1. Other parts
+ of gcc expect that the comparison operation is equivalent
+ to the modified comparison. For signed comparisons compare the
+ result against 1 in the biased case, and zero in the unbiased
+ case. For unsigned comparisons always compare against 1 after
+ biasing the unbiased result by adding 1. This gives us a way to
+ represent LTU. */
*px = result;
*pmode = word_mode;
- if (TARGET_LIB_INT_CMP_BIASED)
- /* Integer comparison returns a result that must be compared
- against 1, so that even if we do an unsigned compare
- afterward, there is still a value that can represent the
- result "less than". */
- *py = const1_rtx;
- else
+ *py = const1_rtx;
+
+ if (!TARGET_LIB_INT_CMP_BIASED)
{
- *py = const0_rtx;
- *punsignedp = 1;
+ if (*punsignedp)
+ *px = plus_constant (result, 1);
+ else
+ *py = const0_rtx;
}
return;
}
}
}
- /* Unsigned integer, and no way to convert directly.
- Convert as signed, then conditionally adjust the result. */
- if (unsignedp && can_do_signed)
+ /* Unsigned integer, and no way to convert directly. For binary
+ floating point modes, convert as signed, then conditionally adjust
+ the result. */
+ if (unsignedp && can_do_signed && !DECIMAL_FLOAT_MODE_P (GET_MODE (to)))
{
rtx label = gen_label_rtx ();
rtx temp;
This is not needed. Consider, for instance conversion from SFmode
into DImode.
- The hot path trought the code is dealing with inputs smaller than 2^63
+ The hot path through the code is dealing with inputs smaller than 2^63
and doing just the conversion, so there is no bits to lose.
In the other path we know the value is positive in the range 2^63..2^64-1
init_floating_libfuncs (optab optable, const char *opname, int suffix)
{
init_libfuncs (optable, MIN_MODE_FLOAT, MAX_MODE_FLOAT, opname, suffix);
+ init_libfuncs (optable, MIN_MODE_DECIMAL_FLOAT, MAX_MODE_DECIMAL_FLOAT,
+ opname, suffix);
}
/* Initialize the libfunc fields of an entire group of entries of an
/* Zap the nonsensical SYMBOL_REF_DECL for this. What we're left with
are the flags assigned by targetm.encode_section_info. */
- SYMBOL_REF_DECL (symbol) = 0;
+ SET_SYMBOL_REF_DECL (symbol, 0);
return symbol;
}
reduc_splus_optab = init_optab (UNKNOWN);
reduc_uplus_optab = init_optab (UNKNOWN);
+ ssum_widen_optab = init_optab (UNKNOWN);
+ usum_widen_optab = init_optab (UNKNOWN);
+ sdot_prod_optab = init_optab (UNKNOWN);
+ udot_prod_optab = init_optab (UNKNOWN);
+
vec_extract_optab = init_optab (UNKNOWN);
vec_set_optab = init_optab (UNKNOWN);
vec_init_optab = init_optab (UNKNOWN);
/* Conversions. */
init_interclass_conv_libfuncs (sfloat_optab, "float",
MODE_INT, MODE_FLOAT);
+ init_interclass_conv_libfuncs (sfloat_optab, "float",
+ MODE_INT, MODE_DECIMAL_FLOAT);
init_interclass_conv_libfuncs (ufloat_optab, "floatun",
MODE_INT, MODE_FLOAT);
+ init_interclass_conv_libfuncs (ufloat_optab, "floatun",
+ MODE_INT, MODE_DECIMAL_FLOAT);
init_interclass_conv_libfuncs (sfix_optab, "fix",
MODE_FLOAT, MODE_INT);
+ init_interclass_conv_libfuncs (sfix_optab, "fix",
+ MODE_DECIMAL_FLOAT, MODE_INT);
init_interclass_conv_libfuncs (ufix_optab, "fixuns",
MODE_FLOAT, MODE_INT);
+ init_interclass_conv_libfuncs (ufix_optab, "fixuns",
+ MODE_DECIMAL_FLOAT, MODE_INT);
+ init_interclass_conv_libfuncs (ufloat_optab, "floatuns",
+ MODE_INT, MODE_DECIMAL_FLOAT);
/* sext_optab is also used for FLOAT_EXTEND. */
init_intraclass_conv_libfuncs (sext_optab, "extend", MODE_FLOAT, true);
+ init_intraclass_conv_libfuncs (sext_optab, "extend", MODE_DECIMAL_FLOAT, true);
+ init_interclass_conv_libfuncs (sext_optab, "extend", MODE_FLOAT, MODE_DECIMAL_FLOAT);
+ init_interclass_conv_libfuncs (sext_optab, "extend", MODE_DECIMAL_FLOAT, MODE_FLOAT);
init_intraclass_conv_libfuncs (trunc_optab, "trunc", MODE_FLOAT, false);
+ init_intraclass_conv_libfuncs (trunc_optab, "trunc", MODE_DECIMAL_FLOAT, false);
+ init_interclass_conv_libfuncs (trunc_optab, "trunc", MODE_FLOAT, MODE_DECIMAL_FLOAT);
+ init_interclass_conv_libfuncs (trunc_optab, "trunc", MODE_DECIMAL_FLOAT, MODE_FLOAT);
/* Use cabs for double complex abs, since systems generally have cabs.
Don't define any libcall for float complex, so that cabs will be used. */