llvm-gcc-4.2/gcc/config/c4x/c4x-modes.def

/* Definitions of target machine for GNU compiler.  TMS320C[34]x
   Copyright (C) 2002, 2004 Free Software Foundation, Inc.

   Contributed by Michael Hayes (m.hayes@elec.canterbury.ac.nz)
              and Herman Ten Brugge (Haj.Ten.Brugge@net.HCC.nl).

   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 2, 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 COPYING.  If not, write to
   the Free Software Foundation, 51 Franklin Street, Fifth Floor,
   Boston, MA 02110-1301, USA.  */

/* C4x wants 1- and 2-word float modes, in its own peculiar format.
   FIXME: Give this port a way to get rid of SFmode, DFmode, and all
   the other modes it doesn't use.  */
FLOAT_MODE (QF, 1, c4x_single_format);
FLOAT_MODE (HF, 2, c4x_extended_format);
RESET_FLOAT_FORMAT (SF, 0);  /* not used */
RESET_FLOAT_FORMAT (DF, 0);  /* not used */

/* Add any extra modes needed to represent the condition code.

   On the C4x, we have a "no-overflow" mode which is used when an ADD,
   SUB, NEG, or MPY insn is used to set the condition code.  This is
   to prevent the combiner from optimizing away a following CMP of the
   result with zero when a signed conditional branch or load insn
   follows.

   The problem is a subtle one and deals with the manner in which the
   negative condition (N) flag is used on the C4x.  This flag does not
   reflect the status of the actual result but of the ideal result had
   no overflow occurred (when considering signed operands).

   For example, 0x7fffffff + 1 => 0x80000000 Z=0 V=1 N=0 C=0.  Here
   the flags reflect the untruncated result, not the actual result.
   While the actual result is less than zero, the N flag is not set
   since the ideal result of the addition without truncation would
   have been positive.
   
   Note that the while the N flag is handled differently to most other
   architectures, the use of it is self consistent and is not the
   cause of the problem.

   Logical operations set the N flag to the MSB of the result so if
   the result is negative, N is 1.  However, integer and floating
   point operations set the N flag to be the MSB of the result
   exclusive ored with the overflow (V) flag.  Thus if an overflow
   occurs and the result does not have the MSB set (i.e., the result
   looks like a positive number), the N flag is set.  Conversely, if
   an overflow occurs and the MSB of the result is set, N is set to 0.
   Thus the N flag represents the sign of the result if it could have
   been stored without overflow but does not represent the apparent
   sign of the result.  Note that most architectures set the N flag to
   be the MSB of the result.

   The C4x approach to setting the N flag simplifies signed
   conditional branches and loads which only have to test the state of
   the N flag, whereas most architectures have to look at both the N
   and V flags.  The disadvantage is that there is no flag giving the
   status of the sign bit of the operation.  However, there are no
   conditional load or branch instructions that make use of this
   feature (e.g., BMI---branch minus) instruction.  Note that BN and
   BLT are identical in the C4x.
   
   To handle the problem where the N flag is set differently whenever
   there is an overflow we use a different CC mode, CC_NOOVmode which
   says that the CC reflects the comparison of the result against zero
   if no overflow occurred.

   For example, 

   [(set (reg:CC_NOOV 21)
         (compare:CC_NOOV (minus:QI (match_operand:QI 1 "src_operand" "")
                                    (match_operand:QI 2 "src_operand" ""))
                          (const_int 0)))
    (set (match_operand:QI 0 "ext_reg_operand" "")
         (minus:QI (match_dup 1)
                   (match_dup 2)))]

   Note that there is no problem for insns that don't return a result
   like CMP, since the CC reflects the effect of operation.

   An example of a potential problem is when GCC
   converts   (LTU (MINUS (0x80000000) (0x7fffffff) (0x80000000)))
   to         (LEU (MINUS (0x80000000) (0x7fffffff) (0x7fffffff)))
   to         (GE  (MINUS (0x80000000) (0x7fffffff) (0x00000000)))

   Now (MINUS (0x80000000) (0x7fffffff)) returns 0x00000001 but the
   C4x sets the N flag since the result without overflow would have
   been 0xffffffff when treating the operands as signed integers.
   Thus (GE (MINUS (0x80000000) (0x7fffffff) (0x00000000))) sets the N
   flag but (GE (0x00000001)) does not set the N flag.

   The upshot is that we cannot use signed branch and conditional
   load instructions after an add, subtract, neg, abs or multiply.
   We must emit a compare insn to check the result against 0.  */

CC_MODE (CC_NOOV);