copysign updates, part 2
Richard Henderson
rth@redhat.com
Mon Jan 31 09:08:00 GMT 2005
The following implements a version of copysign that uses existing
floating-point absolute value and negation patterns. Not only does
this result in (for particular targets) much less shuffling of data
between fp and integer register files, but it'll be the basis for
doing copysign on the ibm double-double format.
Tested on i686-linux. Which without sse enabled, uses this
extensively.
r~
Index: gcc/optabs.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/optabs.c,v
retrieving revision 1.256
diff -u -p -d -r1.256 optabs.c
--- gcc/optabs.c 31 Jan 2005 04:00:55 -0000 1.256
+++ gcc/optabs.c 31 Jan 2005 08:57:36 -0000
@@ -2657,44 +2657,99 @@ expand_abs (enum machine_mode mode, rtx
return target;
}
-/* Expand the C99 copysign operation. OP0 and OP1 must be the same
- scalar floating point mode. Return NULL if we do not know how to
- expand the operation inline. */
+/* A subroutine of expand_copysign, perform the copysign operation using the
+ abs and neg primitives advertised to exist on the target. The assumption
+ is that we have a split register file, and leaving op0 in fp registers,
+ and not playing with subregs so much, will help the register allocator. */
-rtx
-expand_copysign (rtx op0, rtx op1, rtx target)
+static rtx
+expand_copysign_absneg (enum machine_mode mode, rtx op0, rtx op1, rtx target,
+ int bitpos, bool op0_is_abs)
{
- enum machine_mode mode = GET_MODE (op0);
- const struct real_format *fmt;
enum machine_mode imode;
- int bitpos, word, nwords, i, have_abs;
HOST_WIDE_INT hi, lo;
- rtx temp, insns;
-
- gcc_assert (SCALAR_FLOAT_MODE_P (mode));
- gcc_assert (GET_MODE (op1) == mode);
+ int word;
+ rtx label;
- /* First try to do it with a special instruction. */
- temp = expand_binop (mode, copysign_optab, op0, op1,
- target, 0, OPTAB_DIRECT);
- if (temp)
- return temp;
+ if (target == op1)
+ target = NULL_RTX;
- fmt = REAL_MODE_FORMAT (mode);
- if (fmt == NULL || !fmt->has_signed_zero)
- return NULL_RTX;
+ if (!op0_is_abs)
+ {
+ op0 = expand_unop (mode, abs_optab, op0, target, 0);
+ if (op0 == NULL)
+ return NULL_RTX;
+ target = op0;
+ }
+ else
+ {
+ if (target == NULL_RTX)
+ target = copy_to_reg (op0);
+ else
+ emit_move_insn (target, op0);
+ }
- bitpos = fmt->signbit;
- if (bitpos < 0)
- return NULL_RTX;
+ if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
+ {
+ imode = int_mode_for_mode (mode);
+ if (imode == BLKmode)
+ return NULL_RTX;
+ op1 = gen_lowpart (imode, op1);
+ }
+ else
+ {
+ imode = word_mode;
+ if (FLOAT_WORDS_BIG_ENDIAN)
+ word = (GET_MODE_BITSIZE (mode) - bitpos) / BITS_PER_WORD;
+ else
+ word = bitpos / BITS_PER_WORD;
+ bitpos = bitpos % BITS_PER_WORD;
+ op1 = operand_subword_force (op1, word, mode);
+ }
- have_abs = false;
- if (GET_CODE (op0) == CONST_DOUBLE)
+ if (bitpos < HOST_BITS_PER_WIDE_INT)
{
- if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0)))
- op0 = simplify_unary_operation (ABS, mode, op0, mode);
- have_abs = true;
+ hi = 0;
+ lo = (HOST_WIDE_INT) 1 << bitpos;
}
+ else
+ {
+ hi = (HOST_WIDE_INT) 1 << (bitpos - HOST_BITS_PER_WIDE_INT);
+ lo = 0;
+ }
+
+ op1 = expand_binop (imode, and_optab, op1,
+ immed_double_const (lo, hi, imode),
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+
+ label = gen_label_rtx ();
+ emit_cmp_and_jump_insns (op1, const0_rtx, EQ, NULL_RTX, imode, 1, label);
+
+ if (GET_CODE (op0) == CONST_DOUBLE)
+ op0 = simplify_unary_operation (NEG, mode, op0, mode);
+ else
+ op0 = expand_unop (mode, neg_optab, op0, target, 0);
+ if (op0 != target)
+ emit_move_insn (target, op0);
+
+ emit_label (label);
+
+ return target;
+}
+
+
+/* A subroutine of expand_copysign, perform the entire copysign operation
+ with integer bitmasks. BITPOS is the position of the sign bit; OP0_IS_ABS
+ is true if op0 is known to have its sign bit clear. */
+
+static rtx
+expand_copysign_bit (enum machine_mode mode, rtx op0, rtx op1, rtx target,
+ int bitpos, bool op0_is_abs)
+{
+ enum machine_mode imode;
+ HOST_WIDE_INT hi, lo;
+ int word, nwords, i;
+ rtx temp, insns;
if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
{
@@ -2741,7 +2796,7 @@ expand_copysign (rtx op0, rtx op1, rtx t
if (i == word)
{
- if (!have_abs)
+ if (!op0_is_abs)
op0_piece = expand_binop (imode, and_optab, op0_piece,
immed_double_const (~lo, ~hi, imode),
NULL_RTX, 1, OPTAB_LIB_WIDEN);
@@ -2772,7 +2827,7 @@ expand_copysign (rtx op0, rtx op1, rtx t
NULL_RTX, 1, OPTAB_LIB_WIDEN);
op0 = gen_lowpart (imode, op0);
- if (!have_abs)
+ if (!op0_is_abs)
op0 = expand_binop (imode, and_optab, op0,
immed_double_const (~lo, ~hi, imode),
NULL_RTX, 1, OPTAB_LIB_WIDEN);
@@ -2784,6 +2839,57 @@ expand_copysign (rtx op0, rtx op1, rtx t
return target;
}
+
+/* Expand the C99 copysign operation. OP0 and OP1 must be the same
+ scalar floating point mode. Return NULL if we do not know how to
+ expand the operation inline. */
+
+rtx
+expand_copysign (rtx op0, rtx op1, rtx target)
+{
+ enum machine_mode mode = GET_MODE (op0);
+ const struct real_format *fmt;
+ int bitpos;
+ bool op0_is_abs;
+ rtx temp;
+
+ gcc_assert (SCALAR_FLOAT_MODE_P (mode));
+ gcc_assert (GET_MODE (op1) == mode);
+
+ /* First try to do it with a special instruction. */
+ temp = expand_binop (mode, copysign_optab, op0, op1,
+ target, 0, OPTAB_DIRECT);
+ if (temp)
+ return temp;
+
+ fmt = REAL_MODE_FORMAT (mode);
+ if (fmt == NULL || !fmt->has_signed_zero)
+ return NULL_RTX;
+
+ bitpos = fmt->signbit;
+ if (bitpos < 0)
+ return NULL_RTX;
+
+ op0_is_abs = false;
+ if (GET_CODE (op0) == CONST_DOUBLE)
+ {
+ if (real_isneg (CONST_DOUBLE_REAL_VALUE (op0)))
+ op0 = simplify_unary_operation (ABS, mode, op0, mode);
+ op0_is_abs = true;
+ }
+
+ if (GET_CODE (op0) == CONST_DOUBLE
+ || (neg_optab->handlers[mode].insn_code != CODE_FOR_nothing
+ && abs_optab->handlers[mode].insn_code != CODE_FOR_nothing))
+ {
+ temp = expand_copysign_absneg (mode, op0, op1, target,
+ bitpos, op0_is_abs);
+ if (temp)
+ return temp;
+ }
+
+ return expand_copysign_bit (mode, op0, op1, target, bitpos, op0_is_abs);
+}
/* Generate an instruction whose insn-code is INSN_CODE,
with two operands: an output TARGET and an input OP0.
More information about the Gcc-patches
mailing list