Chris@0: /*
Chris@0:  * libmad - MPEG audio decoder library
Chris@0:  * Copyright (C) 2000-2004 Underbit Technologies, Inc.
Chris@0:  *
Chris@0:  * This program is free software; you can redistribute it and/or modify
Chris@0:  * it under the terms of the GNU General Public License as published by
Chris@0:  * the Free Software Foundation; either version 2 of the License, or
Chris@0:  * (at your option) any later version.
Chris@0:  *
Chris@0:  * This program is distributed in the hope that it will be useful,
Chris@0:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@0:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@0:  * GNU General Public License for more details.
Chris@0:  *
Chris@0:  * You should have received a copy of the GNU General Public License
Chris@0:  * along with this program; if not, write to the Free Software
Chris@0:  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
Chris@0:  *
Chris@0:  * $Id: fixed.h,v 1.38 2004/02/17 02:02:03 rob Exp $
Chris@0:  */
Chris@0: 
Chris@0: # ifndef LIBMAD_FIXED_H
Chris@0: # define LIBMAD_FIXED_H
Chris@0: 
Chris@0: # if SIZEOF_INT >= 4
Chris@0: typedef   signed int mad_fixed_t;
Chris@0: 
Chris@0: typedef   signed int mad_fixed64hi_t;
Chris@0: typedef unsigned int mad_fixed64lo_t;
Chris@0: # else
Chris@0: typedef   signed long mad_fixed_t;
Chris@0: 
Chris@0: typedef   signed long mad_fixed64hi_t;
Chris@0: typedef unsigned long mad_fixed64lo_t;
Chris@0: # endif
Chris@0: 
Chris@0: # if defined(_MSC_VER)
Chris@0: #  define mad_fixed64_t  signed __int64
Chris@0: # elif 1 || defined(__GNUC__)
Chris@0: #  define mad_fixed64_t  signed long long
Chris@0: # endif
Chris@0: 
Chris@0: # if defined(FPM_FLOAT)
Chris@0: typedef double mad_sample_t;
Chris@0: # else
Chris@0: typedef mad_fixed_t mad_sample_t;
Chris@0: # endif
Chris@0: 
Chris@0: /*
Chris@0:  * Fixed-point format: 0xABBBBBBB
Chris@0:  * A == whole part      (sign + 3 bits)
Chris@0:  * B == fractional part (28 bits)
Chris@0:  *
Chris@0:  * Values are signed two's complement, so the effective range is:
Chris@0:  * 0x80000000 to 0x7fffffff
Chris@0:  *       -8.0 to +7.9999999962747097015380859375
Chris@0:  *
Chris@0:  * The smallest representable value is:
Chris@0:  * 0x00000001 == 0.0000000037252902984619140625 (i.e. about 3.725e-9)
Chris@0:  *
Chris@0:  * 28 bits of fractional accuracy represent about
Chris@0:  * 8.6 digits of decimal accuracy.
Chris@0:  *
Chris@0:  * Fixed-point numbers can be added or subtracted as normal
Chris@0:  * integers, but multiplication requires shifting the 64-bit result
Chris@0:  * from 56 fractional bits back to 28 (and rounding.)
Chris@0:  *
Chris@0:  * Changing the definition of MAD_F_FRACBITS is only partially
Chris@0:  * supported, and must be done with care.
Chris@0:  */
Chris@0: 
Chris@0: # define MAD_F_FRACBITS		28
Chris@0: 
Chris@0: # if MAD_F_FRACBITS == 28
Chris@0: #  define MAD_F(x)		((mad_fixed_t) (x##L))
Chris@0: # else
Chris@0: #  if MAD_F_FRACBITS < 28
Chris@0: #   warning "MAD_F_FRACBITS < 28"
Chris@0: #   define MAD_F(x)		((mad_fixed_t)  \
Chris@0: 				 (((x##L) +  \
Chris@0: 				   (1L << (28 - MAD_F_FRACBITS - 1))) >>  \
Chris@0: 				  (28 - MAD_F_FRACBITS)))
Chris@0: #  elif MAD_F_FRACBITS > 28
Chris@0: #   error "MAD_F_FRACBITS > 28 not currently supported"
Chris@0: #   define MAD_F(x)		((mad_fixed_t)  \
Chris@0: 				 ((x##L) << (MAD_F_FRACBITS - 28)))
Chris@0: #  endif
Chris@0: # endif
Chris@0: 
Chris@0: # define MAD_F_MIN		((mad_fixed_t) -0x80000000L)
Chris@0: # define MAD_F_MAX		((mad_fixed_t) +0x7fffffffL)
Chris@0: 
Chris@0: # define MAD_F_ONE		MAD_F(0x10000000)
Chris@0: 
Chris@0: # define mad_f_tofixed(x)	((mad_fixed_t)  \
Chris@0: 				 ((x) * (double) (1L << MAD_F_FRACBITS) + 0.5))
Chris@0: # define mad_f_todouble(x)	((double)  \
Chris@0: 				 ((x) / (double) (1L << MAD_F_FRACBITS)))
Chris@0: 
Chris@0: # define mad_f_intpart(x)	((x) >> MAD_F_FRACBITS)
Chris@0: # define mad_f_fracpart(x)	((x) & ((1L << MAD_F_FRACBITS) - 1))
Chris@0: 				/* (x should be positive) */
Chris@0: 
Chris@0: # define mad_f_fromint(x)	((x) << MAD_F_FRACBITS)
Chris@0: 
Chris@0: # define mad_f_add(x, y)	((x) + (y))
Chris@0: # define mad_f_sub(x, y)	((x) - (y))
Chris@0: 
Chris@0: # if defined(FPM_FLOAT)
Chris@0: #  error "FPM_FLOAT not yet supported"
Chris@0: 
Chris@0: #  undef MAD_F
Chris@0: #  define MAD_F(x)		mad_f_todouble(x)
Chris@0: 
Chris@0: #  define mad_f_mul(x, y)	((x) * (y))
Chris@0: #  define mad_f_scale64
Chris@0: 
Chris@0: #  undef ASO_ZEROCHECK
Chris@0: 
Chris@0: # elif defined(FPM_64BIT)
Chris@0: 
Chris@0: /*
Chris@0:  * This version should be the most accurate if 64-bit types are supported by
Chris@0:  * the compiler, although it may not be the most efficient.
Chris@0:  */
Chris@0: #  if defined(OPT_ACCURACY)
Chris@0: #   define mad_f_mul(x, y)  \
Chris@0:     ((mad_fixed_t)  \
Chris@0:      ((((mad_fixed64_t) (x) * (y)) +  \
Chris@0:        (1L << (MAD_F_SCALEBITS - 1))) >> MAD_F_SCALEBITS))
Chris@0: #  else
Chris@0: #   define mad_f_mul(x, y)  \
Chris@0:     ((mad_fixed_t) (((mad_fixed64_t) (x) * (y)) >> MAD_F_SCALEBITS))
Chris@0: #  endif
Chris@0: 
Chris@0: #  define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: 
Chris@0: /* --- Intel --------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_INTEL)
Chris@0: 
Chris@0: #  if defined(_MSC_VER)
Chris@0: #   pragma warning(push)
Chris@0: #   pragma warning(disable: 4035)  /* no return value */
Chris@0: static __forceinline
Chris@0: mad_fixed_t mad_f_mul_inline(mad_fixed_t x, mad_fixed_t y)
Chris@0: {
Chris@0:   enum {
Chris@0:     fracbits = MAD_F_FRACBITS
Chris@0:   };
Chris@0: 
Chris@0:   __asm {
Chris@0:     mov eax, x
Chris@0:     imul y
Chris@0:     shrd eax, edx, fracbits
Chris@0:   }
Chris@0: 
Chris@0:   /* implicit return of eax */
Chris@0: }
Chris@0: #   pragma warning(pop)
Chris@0: 
Chris@0: #   define mad_f_mul		mad_f_mul_inline
Chris@0: #   define mad_f_scale64
Chris@0: #  else
Chris@0: /*
Chris@0:  * This Intel version is fast and accurate; the disposition of the least
Chris@0:  * significant bit depends on OPT_ACCURACY via mad_f_scale64().
Chris@0:  */
Chris@0: #   define MAD_F_MLX(hi, lo, x, y)  \
Chris@0:     asm ("imull %3"  \
Chris@0: 	 : "=a" (lo), "=d" (hi)  \
Chris@0: 	 : "%a" (x), "rm" (y)  \
Chris@0: 	 : "cc")
Chris@0: 
Chris@0: #   if defined(OPT_ACCURACY)
Chris@0: /*
Chris@0:  * This gives best accuracy but is not very fast.
Chris@0:  */
Chris@0: #    define MAD_F_MLA(hi, lo, x, y)  \
Chris@0:     ({ mad_fixed64hi_t __hi;  \
Chris@0:        mad_fixed64lo_t __lo;  \
Chris@0:        MAD_F_MLX(__hi, __lo, (x), (y));  \
Chris@0:        asm ("addl %2,%0\n\t"  \
Chris@0: 	    "adcl %3,%1"  \
Chris@0: 	    : "=rm" (lo), "=rm" (hi)  \
Chris@0: 	    : "r" (__lo), "r" (__hi), "0" (lo), "1" (hi)  \
Chris@0: 	    : "cc");  \
Chris@0:     })
Chris@0: #   endif  /* OPT_ACCURACY */
Chris@0: 
Chris@0: #   if defined(OPT_ACCURACY)
Chris@0: /*
Chris@0:  * Surprisingly, this is faster than SHRD followed by ADC.
Chris@0:  */
Chris@0: #    define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed64hi_t __hi_;  \
Chris@0:        mad_fixed64lo_t __lo_;  \
Chris@0:        mad_fixed_t __result;  \
Chris@0:        asm ("addl %4,%2\n\t"  \
Chris@0: 	    "adcl %5,%3"  \
Chris@0: 	    : "=rm" (__lo_), "=rm" (__hi_)  \
Chris@0: 	    : "0" (lo), "1" (hi),  \
Chris@0: 	      "ir" (1L << (MAD_F_SCALEBITS - 1)), "ir" (0)  \
Chris@0: 	    : "cc");  \
Chris@0:        asm ("shrdl %3,%2,%1"  \
Chris@0: 	    : "=rm" (__result)  \
Chris@0: 	    : "0" (__lo_), "r" (__hi_), "I" (MAD_F_SCALEBITS)  \
Chris@0: 	    : "cc");  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: #   elif defined(OPT_INTEL)
Chris@0: /*
Chris@0:  * Alternate Intel scaling that may or may not perform better.
Chris@0:  */
Chris@0: #    define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed_t __result;  \
Chris@0:        asm ("shrl %3,%1\n\t"  \
Chris@0: 	    "shll %4,%2\n\t"  \
Chris@0: 	    "orl %2,%1"  \
Chris@0: 	    : "=rm" (__result)  \
Chris@0: 	    : "0" (lo), "r" (hi),  \
Chris@0: 	      "I" (MAD_F_SCALEBITS), "I" (32 - MAD_F_SCALEBITS)  \
Chris@0: 	    : "cc");  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: #   else
Chris@0: #    define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed_t __result;  \
Chris@0:        asm ("shrdl %3,%2,%1"  \
Chris@0: 	    : "=rm" (__result)  \
Chris@0: 	    : "0" (lo), "r" (hi), "I" (MAD_F_SCALEBITS)  \
Chris@0: 	    : "cc");  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: #   endif  /* OPT_ACCURACY */
Chris@0: 
Chris@0: #   define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: #  endif
Chris@0: 
Chris@0: /* --- ARM ----------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_ARM)
Chris@0: 
Chris@0: /* 
Chris@0:  * This ARM V4 version is as accurate as FPM_64BIT but much faster. The
Chris@0:  * least significant bit is properly rounded at no CPU cycle cost!
Chris@0:  */
Chris@0: # if 1
Chris@0: /*
Chris@0:  * This is faster than the default implementation via MAD_F_MLX() and
Chris@0:  * mad_f_scale64().
Chris@0:  */
Chris@0: #  define mad_f_mul(x, y)  \
Chris@0:     ({ mad_fixed64hi_t __hi;  \
Chris@0:        mad_fixed64lo_t __lo;  \
Chris@0:        mad_fixed_t __result;  \
Chris@0:        asm ("smull	%0, %1, %3, %4\n\t"  \
Chris@0: 	    "movs	%0, %0, lsr %5\n\t"  \
Chris@0: 	    "adc	%2, %0, %1, lsl %6"  \
Chris@0: 	    : "=&r" (__lo), "=&r" (__hi), "=r" (__result)  \
Chris@0: 	    : "%r" (x), "r" (y),  \
Chris@0: 	      "M" (MAD_F_SCALEBITS), "M" (32 - MAD_F_SCALEBITS)  \
Chris@0: 	    : "cc");  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: # endif
Chris@0: 
Chris@0: #  define MAD_F_MLX(hi, lo, x, y)  \
Chris@0:     asm ("smull	%0, %1, %2, %3"  \
Chris@0: 	 : "=&r" (lo), "=&r" (hi)  \
Chris@0: 	 : "%r" (x), "r" (y))
Chris@0: 
Chris@0: #  define MAD_F_MLA(hi, lo, x, y)  \
Chris@0:     asm ("smlal	%0, %1, %2, %3"  \
Chris@0: 	 : "+r" (lo), "+r" (hi)  \
Chris@0: 	 : "%r" (x), "r" (y))
Chris@0: 
Chris@0: #  define MAD_F_MLN(hi, lo)  \
Chris@0:     asm ("rsbs	%0, %2, #0\n\t"  \
Chris@0: 	 "rsc	%1, %3, #0"  \
Chris@0: 	 : "=r" (lo), "=r" (hi)  \
Chris@0: 	 : "0" (lo), "1" (hi)  \
Chris@0: 	 : "cc")
Chris@0: 
Chris@0: #  define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed_t __result;  \
Chris@0:        asm ("movs	%0, %1, lsr %3\n\t"  \
Chris@0: 	    "adc	%0, %0, %2, lsl %4"  \
Chris@0: 	    : "=&r" (__result)  \
Chris@0: 	    : "r" (lo), "r" (hi),  \
Chris@0: 	      "M" (MAD_F_SCALEBITS), "M" (32 - MAD_F_SCALEBITS)  \
Chris@0: 	    : "cc");  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: 
Chris@0: #  define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: 
Chris@0: /* --- MIPS ---------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_MIPS)
Chris@0: 
Chris@0: /*
Chris@0:  * This MIPS version is fast and accurate; the disposition of the least
Chris@0:  * significant bit depends on OPT_ACCURACY via mad_f_scale64().
Chris@0:  */
Chris@0: #  define MAD_F_MLX(hi, lo, x, y)  \
Chris@0:     asm ("mult	%2,%3"  \
Chris@0: 	 : "=l" (lo), "=h" (hi)  \
Chris@0: 	 : "%r" (x), "r" (y))
Chris@0: 
Chris@0: # if defined(HAVE_MADD_ASM)
Chris@0: #  define MAD_F_MLA(hi, lo, x, y)  \
Chris@0:     asm ("madd	%2,%3"  \
Chris@0: 	 : "+l" (lo), "+h" (hi)  \
Chris@0: 	 : "%r" (x), "r" (y))
Chris@0: # elif defined(HAVE_MADD16_ASM)
Chris@0: /*
Chris@0:  * This loses significant accuracy due to the 16-bit integer limit in the
Chris@0:  * multiply/accumulate instruction.
Chris@0:  */
Chris@0: #  define MAD_F_ML0(hi, lo, x, y)  \
Chris@0:     asm ("mult	%2,%3"  \
Chris@0: 	 : "=l" (lo), "=h" (hi)  \
Chris@0: 	 : "%r" ((x) >> 12), "r" ((y) >> 16))
Chris@0: #  define MAD_F_MLA(hi, lo, x, y)  \
Chris@0:     asm ("madd16	%2,%3"  \
Chris@0: 	 : "+l" (lo), "+h" (hi)  \
Chris@0: 	 : "%r" ((x) >> 12), "r" ((y) >> 16))
Chris@0: #  define MAD_F_MLZ(hi, lo)  ((mad_fixed_t) (lo))
Chris@0: # endif
Chris@0: 
Chris@0: # if defined(OPT_SPEED)
Chris@0: #  define mad_f_scale64(hi, lo)  \
Chris@0:     ((mad_fixed_t) ((hi) << (32 - MAD_F_SCALEBITS)))
Chris@0: #  define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: # endif
Chris@0: 
Chris@0: /* --- SPARC --------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_SPARC)
Chris@0: 
Chris@0: /*
Chris@0:  * This SPARC V8 version is fast and accurate; the disposition of the least
Chris@0:  * significant bit depends on OPT_ACCURACY via mad_f_scale64().
Chris@0:  */
Chris@0: #  define MAD_F_MLX(hi, lo, x, y)  \
Chris@0:     asm ("smul %2, %3, %0\n\t"  \
Chris@0: 	 "rd %%y, %1"  \
Chris@0: 	 : "=r" (lo), "=r" (hi)  \
Chris@0: 	 : "%r" (x), "rI" (y))
Chris@0: 
Chris@0: /* --- PowerPC ------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_PPC)
Chris@0: 
Chris@0: /*
Chris@0:  * This PowerPC version is fast and accurate; the disposition of the least
Chris@0:  * significant bit depends on OPT_ACCURACY via mad_f_scale64().
Chris@0:  */
Chris@0: #  define MAD_F_MLX(hi, lo, x, y)  \
Chris@0:     do {  \
Chris@0:       asm ("mullw %0,%1,%2"  \
Chris@0: 	   : "=r" (lo)  \
Chris@0: 	   : "%r" (x), "r" (y));  \
Chris@0:       asm ("mulhw %0,%1,%2"  \
Chris@0: 	   : "=r" (hi)  \
Chris@0: 	   : "%r" (x), "r" (y));  \
Chris@0:     }  \
Chris@0:     while (0)
Chris@0: 
Chris@0: #  if defined(OPT_ACCURACY)
Chris@0: /*
Chris@0:  * This gives best accuracy but is not very fast.
Chris@0:  */
Chris@0: #   define MAD_F_MLA(hi, lo, x, y)  \
Chris@0:     ({ mad_fixed64hi_t __hi;  \
Chris@0:        mad_fixed64lo_t __lo;  \
Chris@0:        MAD_F_MLX(__hi, __lo, (x), (y));  \
Chris@0:        asm ("addc %0,%2,%3\n\t"  \
Chris@0: 	    "adde %1,%4,%5"  \
Chris@0: 	    : "=r" (lo), "=r" (hi)  \
Chris@0: 	    : "%r" (lo), "r" (__lo),  \
Chris@0: 	      "%r" (hi), "r" (__hi)  \
Chris@0: 	    : "xer");  \
Chris@0:     })
Chris@0: #  endif
Chris@0: 
Chris@0: #  if defined(OPT_ACCURACY)
Chris@0: /*
Chris@0:  * This is slower than the truncating version below it.
Chris@0:  */
Chris@0: #   define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed_t __result, __round;  \
Chris@0:        asm ("rotrwi %0,%1,%2"  \
Chris@0: 	    : "=r" (__result)  \
Chris@0: 	    : "r" (lo), "i" (MAD_F_SCALEBITS));  \
Chris@0:        asm ("extrwi %0,%1,1,0"  \
Chris@0: 	    : "=r" (__round)  \
Chris@0: 	    : "r" (__result));  \
Chris@0:        asm ("insrwi %0,%1,%2,0"  \
Chris@0: 	    : "+r" (__result)  \
Chris@0: 	    : "r" (hi), "i" (MAD_F_SCALEBITS));  \
Chris@0:        asm ("add %0,%1,%2"  \
Chris@0: 	    : "=r" (__result)  \
Chris@0: 	    : "%r" (__result), "r" (__round));  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: #  else
Chris@0: #   define mad_f_scale64(hi, lo)  \
Chris@0:     ({ mad_fixed_t __result;  \
Chris@0:        asm ("rotrwi %0,%1,%2"  \
Chris@0: 	    : "=r" (__result)  \
Chris@0: 	    : "r" (lo), "i" (MAD_F_SCALEBITS));  \
Chris@0:        asm ("insrwi %0,%1,%2,0"  \
Chris@0: 	    : "+r" (__result)  \
Chris@0: 	    : "r" (hi), "i" (MAD_F_SCALEBITS));  \
Chris@0:        __result;  \
Chris@0:     })
Chris@0: #  endif
Chris@0: 
Chris@0: #  define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: 
Chris@0: /* --- Default ------------------------------------------------------------- */
Chris@0: 
Chris@0: # elif defined(FPM_DEFAULT)
Chris@0: 
Chris@0: /*
Chris@0:  * This version is the most portable but it loses significant accuracy.
Chris@0:  * Furthermore, accuracy is biased against the second argument, so care
Chris@0:  * should be taken when ordering operands.
Chris@0:  *
Chris@0:  * The scale factors are constant as this is not used with SSO.
Chris@0:  *
Chris@0:  * Pre-rounding is required to stay within the limits of compliance.
Chris@0:  */
Chris@0: #  if defined(OPT_SPEED)
Chris@0: #   define mad_f_mul(x, y)	(((x) >> 12) * ((y) >> 16))
Chris@0: #  else
Chris@0: #   define mad_f_mul(x, y)	((((x) + (1L << 11)) >> 12) *  \
Chris@0: 				 (((y) + (1L << 15)) >> 16))
Chris@0: #  endif
Chris@0: 
Chris@0: /* ------------------------------------------------------------------------- */
Chris@0: 
Chris@0: # else
Chris@0: #  error "no FPM selected"
Chris@0: # endif
Chris@0: 
Chris@0: /* default implementations */
Chris@0: 
Chris@0: # if !defined(mad_f_mul)
Chris@0: #  define mad_f_mul(x, y)  \
Chris@0:     ({ register mad_fixed64hi_t __hi;  \
Chris@0:        register mad_fixed64lo_t __lo;  \
Chris@0:        MAD_F_MLX(__hi, __lo, (x), (y));  \
Chris@0:        mad_f_scale64(__hi, __lo);  \
Chris@0:     })
Chris@0: # endif
Chris@0: 
Chris@0: # if !defined(MAD_F_MLA)
Chris@0: #  define MAD_F_ML0(hi, lo, x, y)	((lo)  = mad_f_mul((x), (y)))
Chris@0: #  define MAD_F_MLA(hi, lo, x, y)	((lo) += mad_f_mul((x), (y)))
Chris@0: #  define MAD_F_MLN(hi, lo)		((lo)  = -(lo))
Chris@0: #  define MAD_F_MLZ(hi, lo)		((void) (hi), (mad_fixed_t) (lo))
Chris@0: # endif
Chris@0: 
Chris@0: # if !defined(MAD_F_ML0)
Chris@0: #  define MAD_F_ML0(hi, lo, x, y)	MAD_F_MLX((hi), (lo), (x), (y))
Chris@0: # endif
Chris@0: 
Chris@0: # if !defined(MAD_F_MLN)
Chris@0: #  define MAD_F_MLN(hi, lo)		((hi) = ((lo) = -(lo)) ? ~(hi) : -(hi))
Chris@0: # endif
Chris@0: 
Chris@0: # if !defined(MAD_F_MLZ)
Chris@0: #  define MAD_F_MLZ(hi, lo)		mad_f_scale64((hi), (lo))
Chris@0: # endif
Chris@0: 
Chris@0: # if !defined(mad_f_scale64)
Chris@0: #  if defined(OPT_ACCURACY)
Chris@0: #   define mad_f_scale64(hi, lo)  \
Chris@0:     ((((mad_fixed_t)  \
Chris@0:        (((hi) << (32 - (MAD_F_SCALEBITS - 1))) |  \
Chris@0: 	((lo) >> (MAD_F_SCALEBITS - 1)))) + 1) >> 1)
Chris@0: #  else
Chris@0: #   define mad_f_scale64(hi, lo)  \
Chris@0:     ((mad_fixed_t)  \
Chris@0:      (((hi) << (32 - MAD_F_SCALEBITS)) |  \
Chris@0:       ((lo) >> MAD_F_SCALEBITS)))
Chris@0: #  endif
Chris@0: #  define MAD_F_SCALEBITS  MAD_F_FRACBITS
Chris@0: # endif
Chris@0: 
Chris@0: /* C routines */
Chris@0: 
Chris@0: mad_fixed_t mad_f_abs(mad_fixed_t);
Chris@0: mad_fixed_t mad_f_div(mad_fixed_t, mad_fixed_t);
Chris@0: 
Chris@0: # endif