Wanted: C source for converting gcc (linux) float to FFP

[alkis]

I've looked through the internets and I only managed to find the reverse (get ieee from ffp)
I need to get ffp from ieee (I think).

Context: I've cross-compiled some ACE Basic programs from linux to amiga and I've noticed that floats are broken. ACE uses FFP math. Now linux doesn't know anything about FFP so it provides the code generator with normal floats. I need to change those with FFP ones.

Here is stuff, I've found.

/// https://marc.info/?l=classiccmp&m=145550087725800&w=2

// val is a FFP float, returns IEEE float format
// basically a direct conversion of the motorola FFPTIEEE function.
// comments are from there.
// you are not expected to understand this horrendous mess.
float ffptieee(const uint32_t val)
{
	uint32_t x = val;
	union {
		float f;
		uint32_t i;
	} _fcast;

	x = x + x;		// delete mantissa high bit
	if (x == 0) {
		// if zero, branch zero as finished
		return (float)x;
	}

	uint8_t k = x & 0xff;

	k ^= 0x80;	// to two's complement exponent
	k >>= 1;	// form 8-bit exponent
	k -= 0x82;	// adjust 64 to 127 and excessize

	x = (x & ~0xff) | k;

	x = (x << 16) | (x >> 16);	// swap for high byte placement
	x <<= 7;	// set sign+exp in high byte

	_fcast.i = x;

	return _fcast.f;
}

/// https://www.lysator.liu.se/~celeborn/sync/atari/68000/FFP.ZIP

/*
         TTL       FFPIEEE CONVERSIONS TO/FROM FFP (FFPTIEEE,FFPFIEEE)
******************************************
*  (C)  COPYRIGHT 1981 BY MOTOROLA INC.  *
******************************************
 
****************************************************
*        FFPTIEEE AND FFPFIEEE SUBROUTINES         *
*                                                  *
*   THIS MODULE CONTAINS SINGLE-PRECISION          *
*   CONVERSION ROUTINES FOR IEEE FORMAT FLOATING   *
*   POINT (DRAFT 8.0) TO AND FROM MOTOROLA FAST    *
*   FLOATING POINT (FFP) VALUES.                   *
*   THESE CAN BE USED WHEN LARGE GROUPS OF NUMBERS *
*   NEED TO BE CONVERTED BETWEEN FORMATS.  SEE     *
*   THE MC68344 USER'S GUIDE FOR A FULLER          *
*   EXPLANATION OF THE VARIOUS METHODS FOR IEEE    *
*   FORMAT SUPPORT.                                *
*                                                  *
*   THE FAST FLOATING POINT (NON-IEEE FORMAT)      *
*   PROVIDES RESULTS AS PRECISE AS THOSE REQUIRED  *
*   BY THE IEEE SPECIFICATION.  HOWEVER, THIS      *
*   FORMAT HAS SOME MINOR DIFFERENCES:             *
*    1) IF THE TRUE RESULT OF AN OPERATION         *
*       IS EXACTLY BETWEEN REPRESENTABLE           *
*       VALUES, THE FFP ROUND-TO-NEAREST           *
*       FUNCTION MAY ROUND TO EITHER EVEN OR ODD.  *
*    2) THE FFP EXPONENT ALLOWS HALF OF THE RANGE  *
*       THAT THE SINGLE-PRECISION IEEE FORMAT      *
*       PROVIDES (APPROX. 10 TO THE +-19 DECIMAL). *
*    3) THE IEEE FORMAT SPECIFIES INFINITY,        *
*       NOT-A-NUMBER, AND DENORMALIZED DATA        *
*       TYPES THAT ARE NOT DIRECTLY SUPPORTED      *
*       BY THE FFP FORMAT.  HOWEVER, THEY MAY BE   *
*       ADDED VIA CUSTOMIZING OR USED VIA THE IEEE *
*       FORMAT EQUIVALENT COMPATIBLE CALLS         *
*       DESCRIBED IN THE MC68344 USER'S GUIDE.     *
*    4) ALL ZEROES ARE CONSIDERED POSITIVE         *
*       IN FFP FORMAT.                             *
*    5) THE SLIGHTLY HIGHER PRECISION MULTIPLY     *
*       ROUTINE "FFPMUL2" SHOULD BE SUBSTITUTED    *
*       FOR THE DEFAULT ROUTINE "FFPMUL" FOR       *
*       COMPLETELY EQUIVALENT PRECISION.           *
*                                                  *
****************************************************
         PAGE
FFPIEEE  IDNT   1,1  FFP CONVERSIONS TO/FROM IEEE FORMAT
 
         SECTION   9
 
         XDEF      FFPTIEEE,FFPFIEEE
 
****************************************************
*               FFPTIEEE                           *
*                                                  *
*  FAST FLOATING POINT TO IEEE FORMAT              *
*                                                  *
*   INPUT:   D7 - FAST FLOATING POINT VALUE        *
*                                                  *
*   OUTPUT:  D7 - IEEE FORMAT FLOATING POINT VALUE *
*                                                  *
*   CONDITION CODES:                               *
*            N - SET IF THE RESULT IS NEGATIVE     *
*            Z - SET IF THE RESULT IS ZERO         *
*            V - UNDEFINED                         *
*            C - UNDEFINED                         *
*            X - UNDEFINED                         *
*                                                  *
*   NOTES:                                         *
*     1) NO WORK STORAGE OR REGISTERS REQUIRED.    *
*     2) ALL ZEROES WILL BE CONVERTED POSITIVE.    *
*     3) NO NOT-A-NUMBER, INIFINITY, DENORMALIZED, *
*        OR INDEFINITES GENERATED. (UNLESS         *
*        USER CUSTOMIZED.)                         *
*                                                  *
*   TIMES (ASSUMING IN-LINE CODE):                 *
*           VALUE ZERO      18 CYCLES              *
*           VALUE NOT ZERO  66 CYCLES              *
*                                                  *
****************************************************
 
FFPTIEEE EQU       *
 
         ADD.L     D7,D7     DELETE MANTISSA HIGH BIT
         BEQ.S     DONE1     BRANCH ZERO AS FINISHED
         EOR.B     #$80,D7   TO TWOS COMPLEMENT EXPONENT
         ASR.B     #1,D7     FORM 8-BIT EXPONENT
         SUB.B     #$82,D7   ADJUST 64 TO 127 AND EXCESSIZE
         SWAP.W    D7        SWAP FOR HIGH BYTE PLACEMENT
         ROL.L     #7,D7     SET SIGN+EXP IN HIGH BYTE
DONE1    EQU       *
 
         RTS                 RETURN TO CALLER
         PAGE
************************************************************
*                     FFPFIEEE                             *
*         FAST FLOATING POINT FROM IEEE FORMAT             *
*                                                          *
*   INPUT:   D7 - IEEE FORMAT FLOATING POINT VALUE         *
*   OUTPUT:  D7 - FFP FORMAT FLOATING POINT VALUE          *
*                                                          *
*   CONDITION CODES:                                       *
*            N - UNDEFINED                                 *
*            Z - SET IF THE RESULT IS ZERO                 *
*            V - SET IF RESULT OVERFLOWED FFP FORMAT       *
*            C - UNDEFINED                                 *
*            X - UNDEFINED                                 *
*                                                          *
*   NOTES:                                                 *
*     1) REGISTER D5 IS USED FOR WORK STORAGE AND NOT      *
*        TRANSPARENT.                                      *
*     2) NOT-A-NUMBER, INIFINITY, AND DENORMALIZED         *
*        TYPES AS WELL AS AN EXPONENT OUTSIDE OF FFP RANGE *
*        GENERATE A BRANCH TO A SPECIFIC PART OF THE       *
*        ROUTINE.  CUSTOMIZING MAY EASILY BE DONE THERE.   *
*                                                          *
*        THE DEFAULT ACTIONS FOR THE VARIOUS TYPES ARE:    *
*      LABEL      TYPE         DEFAULT SUBSTITUTION        *
*      -----      ----         --------------------        *
*       NAN    NOT-A-NUMBER    ZERO                        *
*       INF    INFINITY        LARGEST FFP VALUE SAME SIGN *
*                              ("V" SET IN CCR)            *
*       DENOR  DENORMALIZED    ZERO                        *
*       EXPHI  EXP TOO LARGE   LARGEST FFP VALUE SAME SIGN *
*                              ("V" SET IN CCR)            *
*       EXPLO  EXP TOO SMALL   ZERO                        *
*                                                          *
*   TIMES (ASSUMING IN-LINE CODE):                         *
*           VALUE ZERO      78 CYCLES                      *
*           VALUE NOT ZERO  72 CYCLES                      *
*                                                          *
************************************************************
 
VBIT     EQU       $02       CONDITION CODE REGISTER "V" BIT MASK
 
FFPFIEEE EQU       *
 
         SWAP.W    D7        SWAP WORD HALVES
         ROR.L     #7,D7     EXPONENT TO LOW BYTE
         MOVE.L    #-128,D5  LOAD $80 MASK IN WORK REGISTER
         EOR.B     D5,D7     CONVERT FROM EXCESS 127 TO TWO'S-COMPLEMENT
         ADD.B     D7,D7     FROM 8 TO 7 BIT EXPONENT
         BVS.S     FFPOVF    BRANCH WILL NOT FIT
         ADD.B     #2<<1+1,D7 ADJUST EXCESS 127 TO 64 AND SET MANTISSA HIGH BIT
         BVS.S     EXPHI     BRANCH EXPONENT TOO LARGE (OVERFLOW)
         EOR.B     D5,D7     BACK TO EXCESS 64
         ROR.L     #1,D7     TO FAST FLOAT REPRESENTATION ("V" CLEARED)
DONE2    EQU       *
 
         RTS                 RETURN TO CALLER
         PAGE
* OVERFLOW DETECTED - CAUSED BY ONE OF THE FOLLOWING:
*        - FALSE OVERFLOW DUE TO DIFFERENCE BETWEEN EXCESS 127 AND 64 FORMAT
*        - EXPONENT TOO HIGH OR LOW TO FIT IN 7 BITS (EXPONENT OVER/UNDERFLOW)
*        - AN EXPONENT OF $FF REPRESENTING AN INFINITY
*        - AN EXPONENT OF $00 REPRESENTING A ZERO, NAN, OR DENORMALIZED VALUE
FFPOVF   BCC.S     FFPOVLW   BRANCH IF OVERFLOW (EXPONENT $FF OR TOO LARGE)
* OVERFLOW - CHECK FOR POSSIBLE FALSE OVERFLOW DUE TO DIFFERENT EXCESS FORMATS
         CMP.B     #$7C,D7   ? WAS ORIGINAL ARGUMENT REALLY IN RANGE
         BEQ.S     FFPOVFLS  YES, BRANCH FALSE OVERFLOW
         CMP.B     #$7E,D7   ? WAS ORIGINAL ARGUMENT REALLY IN RANGE
         BNE.S     FFPTOVF   NO, BRANCH TRUE OVERFLOW
FFPOVFLS ADD.B     #$80+2<<1+1,D7  EXCESS 64 ADJUSTMENT AND MANTISSA HIGH BIT
         ROR.L     #1,D7     FINALIZE TO FAST FLOATING POINT FORMAT
         TST.B     D7        INSURE NO ILLEGAL ZERO SIGN+EXPONENT BYTE
         BNE.S     DONE2     DONE IF DOES NOT SET S+EXP ALL ZEROES
         BRA.S     EXPLO     TREAT AS UNDERFLOWED EXPONENT OTHERWISE
* EXPONENT LOW - CHECK FOR ZERO, DENORMALIZED VALUE, OR TOO SMALL AN EXPONENT
FFPTOVF  AND.W     #$FEFF,D7 CLEAR SIGN BIT OUT
         TST.L     D7        ? ENTIRE VALUE NOW ZERO
         BEQ.S     DONE2     BRANCH IF VALUE IS ZERO
         TST.B     D7        ? DENORMALIZED NUMBER (SIGNIFICANT#0, EXP=0)
         BEQ.S     DENOR     BRANCH IF DENORMALIZED
 
***************
*****EXPLO - EXPONENT TO SMALL FOR FFP FORMAT
***************
*  THE SIGN BIT WILL BE BIT 8.
EXPLO    MOVE.L    #0,D7     DEFAULT ZERO FOR THIS CASE ("V" CLEARED)
         BRA.S     DONE2     RETURN TO MAINLINE
 
***************
*****DENOR - DENORMALIZED NUMBER
***************
DENOR    MOVE.L    #0,D7     DEFAULT IS TO RETURN A ZERO ("V" CLEARED)
         BRA.S     DONE2     RETURN TO MAINLINE
 
* EXPONENT HIGH - CHECK FOR EXPONENT TOO HIGH, INFINITY, OR NAN
FFPOVLW  CMP.B     #$FE,D7   ? WAS ORIGINAL EXPONENT $FF
         BNE.S     EXPHI     NO, BRANCH EXPONENT TOO LARGE
         LSR.L     #8,D7     SHIFT OUT EXPONENT
         LSR.L     #1,D7     SHIFT OUT SIGN
         BNE.S     NAN       BRANCH NOT-A-NUMBER
 
***************
*****INF - INFINITY
***************
*  THE CARRY AND X BIT REPRESENT THE SIGN
INF      MOVE.L    #-1,D7    SETUP MAXIMUM FFP VALUE
         ROXR.B    #1,D7     SHIFT IN SIGN
         OR.B      #VBIT,SR  SHOW OVERFLOW OCCURED
         BRA.S     DONE2     RETURN WITH MAXIMUM SAME SIGN TO MAINLINE
 
***************
*****EXPHI - EXPONENT TO LARGE FOR FFP FORMAT
***************
*  THE SIGN BIT WILL BE BIT 8.
EXPHI    LSL.W     #8,D7     SET X BIT TO SIGN
         MOVE.L    #-1,D7    SETUP MAXIMUM NUMBER
         ROXR.B    #1,D7     SHIFT IN SIGN
         OR.B      #VBIT,SR  SHOW OVERFLOW OCURRED
         BRA.S     DONE2     RETURN MAXIMUM SAME SIGN TO MAINLINE
 
***************
*****NAN - NOT-A-NUMBER
***************
* BITS 0 THRU 22 CONTAIN THE NAN DATA FIELD
NAN      MOVE.L    #0,D7     DEFAULT TO A ZERO ("V" BIT CLEARED)
         BRA.S     DONE2     RETURN TO MAINLINE
 
         END


 */

[Jobbo]

The 68000 code you posted has the same ieee to ffp conversion code I've used, which I can confirm works well.

Can you port that to C code yourself?

[alkis]

Originally Posted by Jobbo

The 68000 code you posted has the same ieee to ffp conversion code I've used, which I can confirm works well.

Can you port that to C code yourself?

If there is no any already used/tested function lying around in some folk's hard drive, I could give it a shot.

[Docent]

You can use SPFieee function from mathtrans.library - it will convert a number from ieee to ffp format.

[alkis]

Originally Posted by Docent

You can use SPFieee function from mathtrans.library - it will convert a number from ieee to ffp format.

Thanks, but the code has to run on linux.

[alkis]

Ok, this is a working draft
Very very lightly tested, seems to work. Also fixed a bug or two on the initial C code that I found.

#include <stdio.h>
#include <stdint.h>


uint32_t rol(uint32_t x, uint32_t n) {
  //return (x >> n % 32) | (x << (32-n) % 32);
  return (x << n ) | (x >> (32-n));
}

uint32_t ror(uint32_t x, uint32_t n) {
  //return (x >> n % 32) | (x << (32-n) % 32);
  return (x >> n ) | (x << (32-n));
}


float ffptieee(const uint32_t val)
{
	uint32_t x = val;
	union {
		float f;
		uint32_t i;
	} _fcast;

	x = x + x;		// delete mantissa high bit
	if (x == 0) {
		// if zero, branch zero as finished
          _fcast.i = 0;
          return _fcast.f;
	}

	uint8_t k = x & 0xff;

	k ^= 0x80;	// to two's complement exponent
        k >>= 1;	// form 8-bit exponent
        if (k & 0x40)   // do an actual ASR here
          k |= 0x80;    // if previous leftmost bit was 1, do the now leftmost bit 1 as well
        
        
	k -= 0x82;	// adjust 64 to 127 and excessize

	x = (x & ~0xff) | k;

	x = (x << 16) | (x >> 16);	// swap for high byte placement
	x = rol(x, 7);	// set sign+exp in high byte
	_fcast.i = x;

	return _fcast.f;
}


uint32_t ffpfieee(float val) {
  union {
    float f;
    uint32_t i;
  } _fcast;

  _fcast.f = val;

  uint32_t x = _fcast.i;

  x = (x << 16) | (x >> 16);  // swap
  x = ror(x, 7);

  uint8_t k = x & 0xff;

  k ^= 0x80;

  uint8_t orig_k=k;
  uint8_t b = k & 0x80;
  k+=k;
  uint8_t a = k & 0x80;
  if (a != b) {
    //printf("Overflow? (a=%d b=%d)\n", a, b);
    if (k < orig_k) {
      //printf("Carry must be set\n");
      if ((k != 0x7C) && (k != 0x7E)) {
        //printf("Need to implement FFPTOVF (k=%02X)\n", k);
        //x &= 0xFFFFFEFF;  // clear sign bit out
        //if (!x)... no need for code, all branches lead to 0
        return 0;
      } else {
        k+= 0x80+(2<<1)+1;
        x = ror(x, 1);
        if (k)
          return x;
        return 0;
      }
    } else {
      // FFPOVLW, EXPONENT HIGH - CHECK FOR EXPONENT TOO HIGH, INFINITY, OR NAN
      if (k == 0xFE) {
        x >>= 9;
        if (x) {
          // NaN
          return 0;
        } else {
          // INF
          if (x & 0x100) 
            x = 0xFFFFFFFF;
          else
            x = 0xFFFFFF7F;
          return x;
        }
      } else {
        // EXPHI
        if (x & 0x100) 
          x = 0xFFFFFFFF;
        else
          x = 0xFFFFFF7F;
        return x;
      }
    }
  }
  // BVS.S     FFPOVF    BRANCH WILL NOT FIT

  k+=(2<<1)+1;
  b = k & 0x80;
  if (a == 0 && b) {
    // EXPHI
    if (x & 0x100) 
      x = 0xFFFFFFFF;
    else
      x = 0xFFFFFF7F;
    return x;
  }
  // BVS.S     EXPHI     BRANCH EXPONENT TOO LARGE (OVERFLOW)

  k ^= 0x80;

  x = (x & 0xFFFFFF00) | k;
  x = ror(x, 1);
  return x;
}

int main(int argc, char *argv[]) {
  union {
    float f;
    uint32_t i;
  } _fcast, g;

  for (int i = -5; i <= 5; ++i) {
    float f = i*1;
    _fcast.f = f;
    uint32_t ffp = ffpfieee(f);
    g.f = ffptieee(ffp);
    printf("%f %f FFP: %08X %08X (%08X) ", f, ffptieee(ffp), ffp, _fcast.i, g.i);
    if (_fcast.i != g.i)
      printf(" Mismatch\n");
    else
      printf("\n");
  }
  /* for (int i = 0; i < 10; ++i) { */
  /*   float f = i/10.f; */
  /*   _fcast.f = f; */
  /*   printf("$%08X, ", _fcast.i); */
  /* } */
  /* printf("\n"); */
  
  return 0;
}