float32_t f32_div( float32_t a, float32_t b )
{
union ui32_f32 uA;
uint_fast32_t uiA;
bool signA;
int_fast16_t expA;
uint_fast32_t sigA;
union ui32_f32 uB;
uint_fast32_t uiB;
bool signB;
int_fast16_t expB;
uint_fast32_t sigB;
bool signZ;
struct exp16_sig32 normExpSig;
int_fast16_t expZ;
#ifdef SOFTFLOAT_FAST_DIV64TO32
uint_fast64_t sig64A;
uint_fast32_t sigZ;
#else
uint_fast32_t sigZ;
uint_fast64_t rem;
#endif
uint_fast32_t uiZ;
union ui32_f32 uZ;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
uA.f = a;
uiA = uA.ui;
signA = signF32UI( uiA );
expA = expF32UI( uiA );
sigA = fracF32UI( uiA );
uB.f = b;
uiB = uB.ui;
signB = signF32UI( uiB );
expB = expF32UI( uiB );
sigB = fracF32UI( uiB );
signZ = signA ^ signB;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
if ( expA == 0xFF ) {
if ( sigA ) goto propagateNaN;
if ( expB == 0xFF ) {
if ( sigB ) goto propagateNaN;
goto invalid;
}
goto infinity;
}
if ( expB == 0xFF ) {
if ( sigB ) goto propagateNaN;
goto zero;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
if ( ! expB ) {
if ( ! sigB ) {
if ( ! (expA | sigA) ) goto invalid;
softfloat_raiseFlags( softfloat_flag_infinite );
goto infinity;
}
normExpSig = softfloat_normSubnormalF32Sig( sigB );
expB = normExpSig.exp;
sigB = normExpSig.sig;
}
if ( ! expA ) {
if ( ! sigA ) goto zero;
normExpSig = softfloat_normSubnormalF32Sig( sigA );
expA = normExpSig.exp;
sigA = normExpSig.sig;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
expZ = expA - expB + 0x7E;
sigA |= 0x00800000;
sigB |= 0x00800000;
#ifdef SOFTFLOAT_FAST_DIV64TO32
if ( sigA < sigB ) {
--expZ;
sig64A = (uint_fast64_t) sigA<<31;
} else {
sig64A = (uint_fast64_t) sigA<<30;
}
sigZ = sig64A / sigB;
if ( ! (sigZ & 0x3F) ) sigZ |= ((uint_fast64_t) sigB * sigZ != sig64A);
#else
if ( sigA < sigB ) {
--expZ;
sigA <<= 8;
} else {
sigA <<= 7;
}
sigB <<= 8;
sigZ = ((uint_fast64_t) sigA * softfloat_approxRecip32_1( sigB ))>>32;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
sigZ += 2;
if ( (sigZ & 0x3F) < 2 ) {
sigZ &= ~3;
#ifdef SOFTFLOAT_FAST_INT64
rem = ((uint_fast64_t) sigA<<31) - (uint_fast64_t) sigZ * sigB;
#else
rem = ((uint_fast64_t) sigA<<32) - (uint_fast64_t) (sigZ<<1) * sigB;
#endif
if ( rem & UINT64_C( 0x8000000000000000 ) ) {
sigZ -= 4;
} else {
if ( rem ) sigZ |= 1;
}
}
#endif
return softfloat_roundPackToF32( signZ, expZ, sigZ );
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
propagateNaN:
uiZ = softfloat_propagateNaNF32UI( uiA, uiB );
goto uiZ;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
invalid:
softfloat_raiseFlags( softfloat_flag_invalid );
uiZ = defaultNaNF32UI;
goto uiZ;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
infinity:
uiZ = packToF32UI( signZ, 0xFF, 0 );
goto uiZ;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
zero:
uiZ = packToF32UI( signZ, 0, 0 );
uiZ:
uZ.ui = uiZ;
return uZ.f;
}
float16_t f16_rem( float16_t a, float16_t b )
{
union ui16_f16 uA;
uint_fast16_t uiA;
bool signA;
int_fast8_t expA;
uint_fast16_t sigA;
union ui16_f16 uB;
uint_fast16_t uiB;
int_fast8_t expB;
uint_fast16_t sigB;
struct exp8_sig16 normExpSig;
uint16_t rem;
int_fast8_t expDiff;
uint_fast16_t q;
uint32_t recip32, q32;
uint16_t altRem, meanRem;
bool signRem;
uint_fast16_t uiZ;
union ui16_f16 uZ;
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
uA.f = a;
uiA = uA.ui;
signA = signF16UI( uiA );
expA = expF16UI( uiA );
sigA = fracF16UI( uiA );
uB.f = b;
uiB = uB.ui;
expB = expF16UI( uiB );
sigB = fracF16UI( uiB );
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
if ( expA == 0x1F ) {
if ( sigA || ((expB == 0x1F) && sigB) ) goto propagateNaN;
goto invalid;
}
if ( expB == 0x1F ) {
if ( sigB ) goto propagateNaN;
return a;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
if ( ! expB ) {
if ( ! sigB ) goto invalid;
normExpSig = softfloat_normSubnormalF16Sig( sigB );
expB = normExpSig.exp;
sigB = normExpSig.sig;
}
if ( ! expA ) {
if ( ! sigA ) return a;
normExpSig = softfloat_normSubnormalF16Sig( sigA );
expA = normExpSig.exp;
sigA = normExpSig.sig;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
rem = sigA | 0x0400;
sigB |= 0x0400;
expDiff = expA - expB;
if ( expDiff < 1 ) {
if ( expDiff < -1 ) return a;
sigB <<= 3;
if ( expDiff ) {
rem <<= 2;
q = 0;
} else {
rem <<= 3;
q = (sigB <= rem);
if ( q ) rem -= sigB;
}
} else {
recip32 = softfloat_approxRecip32_1( (uint_fast32_t) sigB<<21 );
/*--------------------------------------------------------------------
| Changing the shift of `rem' here requires also changing the initial
| subtraction from `expDiff'.
*--------------------------------------------------------------------*/
rem <<= 4;
expDiff -= 31;
/*--------------------------------------------------------------------
| The scale of `sigB' affects how many bits are obtained during each
| cycle of the loop. Currently this is 29 bits per loop iteration,
| which is believed to be the maximum possible.
*--------------------------------------------------------------------*/
sigB <<= 3;
for (;;) {
q32 = (rem * (uint_fast64_t) recip32)>>16;
if ( expDiff < 0 ) break;
rem = -((uint_fast16_t) q32 * sigB);
expDiff -= 29;
}
/*--------------------------------------------------------------------
| (`expDiff' cannot be less than -30 here.)
*--------------------------------------------------------------------*/
q32 >>= ~expDiff & 31;
q = q32;
rem = (rem<<(expDiff + 30)) - q * sigB;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
do {
altRem = rem;
++q;
rem -= sigB;
} while ( ! (rem & 0x8000) );
meanRem = rem + altRem;
if ( (meanRem & 0x8000) || (! meanRem && (q & 1)) ) rem = altRem;
signRem = signA;
if ( 0x8000 <= rem ) {
signRem = ! signRem;
rem = -rem;
}
return softfloat_normRoundPackToF16( signRem, expB, rem );
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
propagateNaN:
uiZ = softfloat_propagateNaNF16UI( uiA, uiB );
goto uiZ;
invalid:
softfloat_raiseFlags( softfloat_flag_invalid );
uiZ = defaultNaNF16UI;
uiZ:
uZ.ui = uiZ;
return uZ.f;
}
