softfloat: Implement fused multiply-add
Implement fused multiply-add as a softfloat primitive. This implements
"a+b*c" as a single step without any intermediate rounding; it is
specified in IEEE 754-2008 and implemented in a number of CPUs.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
diff --git a/fpu/softfloat-specialize.h b/fpu/softfloat-specialize.h
index c165205..c5e2dab 100644
--- a/fpu/softfloat-specialize.h
+++ b/fpu/softfloat-specialize.h
@@ -420,6 +420,82 @@
#endif
/*----------------------------------------------------------------------------
+| Select which NaN to propagate for a three-input operation.
+| For the moment we assume that no CPU needs the 'larger significand'
+| information.
+| Return values : 0 : a; 1 : b; 2 : c; 3 : default-NaN
+*----------------------------------------------------------------------------*/
+#if defined(TARGET_ARM)
+static int pickNaNMulAdd(flag aIsQNaN, flag aIsSNaN, flag bIsQNaN, flag bIsSNaN,
+ flag cIsQNaN, flag cIsSNaN, flag infzero STATUS_PARAM)
+{
+ /* For ARM, the (inf,zero,qnan) case sets InvalidOp and returns
+ * the default NaN
+ */
+ if (infzero && cIsQNaN) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ return 3;
+ }
+
+ /* This looks different from the ARM ARM pseudocode, because the ARM ARM
+ * puts the operands to a fused mac operation (a*b)+c in the order c,a,b.
+ */
+ if (cIsSNaN) {
+ return 2;
+ } else if (aIsSNaN) {
+ return 0;
+ } else if (bIsSNaN) {
+ return 1;
+ } else if (cIsQNaN) {
+ return 2;
+ } else if (aIsQNaN) {
+ return 0;
+ } else {
+ return 1;
+ }
+}
+#elif defined(TARGET_PPC)
+static int pickNaNMulAdd(flag aIsQNaN, flag aIsSNaN, flag bIsQNaN, flag bIsSNaN,
+ flag cIsQNaN, flag cIsSNaN, flag infzero STATUS_PARAM)
+{
+ /* For PPC, the (inf,zero,qnan) case sets InvalidOp, but we prefer
+ * to return an input NaN if we have one (ie c) rather than generating
+ * a default NaN
+ */
+ if (infzero) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ return 2;
+ }
+
+ /* If fRA is a NaN return it; otherwise if fRB is a NaN return it;
+ * otherwise return fRC. Note that muladd on PPC is (fRA * fRC) + frB
+ */
+ if (aIsSNaN || aIsQNaN) {
+ return 0;
+ } else if (cIsSNaN || cIsQNaN) {
+ return 2;
+ } else {
+ return 1;
+ }
+}
+#else
+/* A default implementation: prefer a to b to c.
+ * This is unlikely to actually match any real implementation.
+ */
+static int pickNaNMulAdd(flag aIsQNaN, flag aIsSNaN, flag bIsQNaN, flag bIsSNaN,
+ flag cIsQNaN, flag cIsSNaN, flag infzero STATUS_PARAM)
+{
+ if (aIsSNaN || aIsQNaN) {
+ return 0;
+ } else if (bIsSNaN || bIsQNaN) {
+ return 1;
+ } else {
+ return 2;
+ }
+}
+#endif
+
+/*----------------------------------------------------------------------------
| Takes two single-precision floating-point values `a' and `b', one of which
| is a NaN, and returns the appropriate NaN result. If either `a' or `b' is a
| signaling NaN, the invalid exception is raised.
@@ -460,6 +536,57 @@
}
/*----------------------------------------------------------------------------
+| Takes three single-precision floating-point values `a', `b' and `c', one of
+| which is a NaN, and returns the appropriate NaN result. If any of `a',
+| `b' or `c' is a signaling NaN, the invalid exception is raised.
+| The input infzero indicates whether a*b was 0*inf or inf*0 (in which case
+| obviously c is a NaN, and whether to propagate c or some other NaN is
+| implementation defined).
+*----------------------------------------------------------------------------*/
+
+static float32 propagateFloat32MulAddNaN(float32 a, float32 b,
+ float32 c, flag infzero STATUS_PARAM)
+{
+ flag aIsQuietNaN, aIsSignalingNaN, bIsQuietNaN, bIsSignalingNaN,
+ cIsQuietNaN, cIsSignalingNaN;
+ int which;
+
+ aIsQuietNaN = float32_is_quiet_nan(a);
+ aIsSignalingNaN = float32_is_signaling_nan(a);
+ bIsQuietNaN = float32_is_quiet_nan(b);
+ bIsSignalingNaN = float32_is_signaling_nan(b);
+ cIsQuietNaN = float32_is_quiet_nan(c);
+ cIsSignalingNaN = float32_is_signaling_nan(c);
+
+ if (aIsSignalingNaN | bIsSignalingNaN | cIsSignalingNaN) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ }
+
+ which = pickNaNMulAdd(aIsQuietNaN, aIsSignalingNaN,
+ bIsQuietNaN, bIsSignalingNaN,
+ cIsQuietNaN, cIsSignalingNaN, infzero STATUS_VAR);
+
+ if (STATUS(default_nan_mode)) {
+ /* Note that this check is after pickNaNMulAdd so that function
+ * has an opportunity to set the Invalid flag.
+ */
+ return float32_default_nan;
+ }
+
+ switch (which) {
+ case 0:
+ return float32_maybe_silence_nan(a);
+ case 1:
+ return float32_maybe_silence_nan(b);
+ case 2:
+ return float32_maybe_silence_nan(c);
+ case 3:
+ default:
+ return float32_default_nan;
+ }
+}
+
+/*----------------------------------------------------------------------------
| Returns 1 if the double-precision floating-point value `a' is a quiet
| NaN; otherwise returns 0.
*----------------------------------------------------------------------------*/
@@ -596,6 +723,57 @@
}
/*----------------------------------------------------------------------------
+| Takes three double-precision floating-point values `a', `b' and `c', one of
+| which is a NaN, and returns the appropriate NaN result. If any of `a',
+| `b' or `c' is a signaling NaN, the invalid exception is raised.
+| The input infzero indicates whether a*b was 0*inf or inf*0 (in which case
+| obviously c is a NaN, and whether to propagate c or some other NaN is
+| implementation defined).
+*----------------------------------------------------------------------------*/
+
+static float64 propagateFloat64MulAddNaN(float64 a, float64 b,
+ float64 c, flag infzero STATUS_PARAM)
+{
+ flag aIsQuietNaN, aIsSignalingNaN, bIsQuietNaN, bIsSignalingNaN,
+ cIsQuietNaN, cIsSignalingNaN;
+ int which;
+
+ aIsQuietNaN = float64_is_quiet_nan(a);
+ aIsSignalingNaN = float64_is_signaling_nan(a);
+ bIsQuietNaN = float64_is_quiet_nan(b);
+ bIsSignalingNaN = float64_is_signaling_nan(b);
+ cIsQuietNaN = float64_is_quiet_nan(c);
+ cIsSignalingNaN = float64_is_signaling_nan(c);
+
+ if (aIsSignalingNaN | bIsSignalingNaN | cIsSignalingNaN) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ }
+
+ which = pickNaNMulAdd(aIsQuietNaN, aIsSignalingNaN,
+ bIsQuietNaN, bIsSignalingNaN,
+ cIsQuietNaN, cIsSignalingNaN, infzero STATUS_VAR);
+
+ if (STATUS(default_nan_mode)) {
+ /* Note that this check is after pickNaNMulAdd so that function
+ * has an opportunity to set the Invalid flag.
+ */
+ return float64_default_nan;
+ }
+
+ switch (which) {
+ case 0:
+ return float64_maybe_silence_nan(a);
+ case 1:
+ return float64_maybe_silence_nan(b);
+ case 2:
+ return float64_maybe_silence_nan(c);
+ case 3:
+ default:
+ return float64_default_nan;
+ }
+}
+
+/*----------------------------------------------------------------------------
| Returns 1 if the extended double-precision floating-point value `a' is a
| quiet NaN; otherwise returns 0. This slightly differs from the same
| function for other types as floatx80 has an explicit bit.
diff --git a/fpu/softfloat.c b/fpu/softfloat.c
index 3aafa81..81a7d1a 100644
--- a/fpu/softfloat.c
+++ b/fpu/softfloat.c
@@ -2118,6 +2118,213 @@
}
/*----------------------------------------------------------------------------
+| Returns the result of multiplying the single-precision floating-point values
+| `a' and `b' then adding 'c', with no intermediate rounding step after the
+| multiplication. The operation is performed according to the IEC/IEEE
+| Standard for Binary Floating-Point Arithmetic 754-2008.
+| The flags argument allows the caller to select negation of the
+| addend, the intermediate product, or the final result. (The difference
+| between this and having the caller do a separate negation is that negating
+| externally will flip the sign bit on NaNs.)
+*----------------------------------------------------------------------------*/
+
+float32 float32_muladd(float32 a, float32 b, float32 c, int flags STATUS_PARAM)
+{
+ flag aSign, bSign, cSign, zSign;
+ int aExp, bExp, cExp, pExp, zExp, expDiff;
+ uint32_t aSig, bSig, cSig;
+ flag pInf, pZero, pSign;
+ uint64_t pSig64, cSig64, zSig64;
+ uint32_t pSig;
+ int shiftcount;
+ flag signflip, infzero;
+
+ a = float32_squash_input_denormal(a STATUS_VAR);
+ b = float32_squash_input_denormal(b STATUS_VAR);
+ c = float32_squash_input_denormal(c STATUS_VAR);
+ aSig = extractFloat32Frac(a);
+ aExp = extractFloat32Exp(a);
+ aSign = extractFloat32Sign(a);
+ bSig = extractFloat32Frac(b);
+ bExp = extractFloat32Exp(b);
+ bSign = extractFloat32Sign(b);
+ cSig = extractFloat32Frac(c);
+ cExp = extractFloat32Exp(c);
+ cSign = extractFloat32Sign(c);
+
+ infzero = ((aExp == 0 && aSig == 0 && bExp == 0xff && bSig == 0) ||
+ (aExp == 0xff && aSig == 0 && bExp == 0 && bSig == 0));
+
+ /* It is implementation-defined whether the cases of (0,inf,qnan)
+ * and (inf,0,qnan) raise InvalidOperation or not (and what QNaN
+ * they return if they do), so we have to hand this information
+ * off to the target-specific pick-a-NaN routine.
+ */
+ if (((aExp == 0xff) && aSig) ||
+ ((bExp == 0xff) && bSig) ||
+ ((cExp == 0xff) && cSig)) {
+ return propagateFloat32MulAddNaN(a, b, c, infzero STATUS_VAR);
+ }
+
+ if (infzero) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ return float32_default_nan;
+ }
+
+ if (flags & float_muladd_negate_c) {
+ cSign ^= 1;
+ }
+
+ signflip = (flags & float_muladd_negate_result) ? 1 : 0;
+
+ /* Work out the sign and type of the product */
+ pSign = aSign ^ bSign;
+ if (flags & float_muladd_negate_product) {
+ pSign ^= 1;
+ }
+ pInf = (aExp == 0xff) || (bExp == 0xff);
+ pZero = ((aExp | aSig) == 0) || ((bExp | bSig) == 0);
+
+ if (cExp == 0xff) {
+ if (pInf && (pSign ^ cSign)) {
+ /* addition of opposite-signed infinities => InvalidOperation */
+ float_raise(float_flag_invalid STATUS_VAR);
+ return float32_default_nan;
+ }
+ /* Otherwise generate an infinity of the same sign */
+ return packFloat32(cSign ^ signflip, 0xff, 0);
+ }
+
+ if (pInf) {
+ return packFloat32(pSign ^ signflip, 0xff, 0);
+ }
+
+ if (pZero) {
+ if (cExp == 0) {
+ if (cSig == 0) {
+ /* Adding two exact zeroes */
+ if (pSign == cSign) {
+ zSign = pSign;
+ } else if (STATUS(float_rounding_mode) == float_round_down) {
+ zSign = 1;
+ } else {
+ zSign = 0;
+ }
+ return packFloat32(zSign ^ signflip, 0, 0);
+ }
+ /* Exact zero plus a denorm */
+ if (STATUS(flush_to_zero)) {
+ float_raise(float_flag_output_denormal STATUS_VAR);
+ return packFloat32(cSign ^ signflip, 0, 0);
+ }
+ }
+ /* Zero plus something non-zero : just return the something */
+ return c ^ (signflip << 31);
+ }
+
+ if (aExp == 0) {
+ normalizeFloat32Subnormal(aSig, &aExp, &aSig);
+ }
+ if (bExp == 0) {
+ normalizeFloat32Subnormal(bSig, &bExp, &bSig);
+ }
+
+ /* Calculate the actual result a * b + c */
+
+ /* Multiply first; this is easy. */
+ /* NB: we subtract 0x7e where float32_mul() subtracts 0x7f
+ * because we want the true exponent, not the "one-less-than"
+ * flavour that roundAndPackFloat32() takes.
+ */
+ pExp = aExp + bExp - 0x7e;
+ aSig = (aSig | 0x00800000) << 7;
+ bSig = (bSig | 0x00800000) << 8;
+ pSig64 = (uint64_t)aSig * bSig;
+ if ((int64_t)(pSig64 << 1) >= 0) {
+ pSig64 <<= 1;
+ pExp--;
+ }
+
+ zSign = pSign ^ signflip;
+
+ /* Now pSig64 is the significand of the multiply, with the explicit bit in
+ * position 62.
+ */
+ if (cExp == 0) {
+ if (!cSig) {
+ /* Throw out the special case of c being an exact zero now */
+ shift64RightJamming(pSig64, 32, &pSig64);
+ pSig = pSig64;
+ return roundAndPackFloat32(zSign, pExp - 1,
+ pSig STATUS_VAR);
+ }
+ normalizeFloat32Subnormal(cSig, &cExp, &cSig);
+ }
+
+ cSig64 = (uint64_t)cSig << (62 - 23);
+ cSig64 |= LIT64(0x4000000000000000);
+ expDiff = pExp - cExp;
+
+ if (pSign == cSign) {
+ /* Addition */
+ if (expDiff > 0) {
+ /* scale c to match p */
+ shift64RightJamming(cSig64, expDiff, &cSig64);
+ zExp = pExp;
+ } else if (expDiff < 0) {
+ /* scale p to match c */
+ shift64RightJamming(pSig64, -expDiff, &pSig64);
+ zExp = cExp;
+ } else {
+ /* no scaling needed */
+ zExp = cExp;
+ }
+ /* Add significands and make sure explicit bit ends up in posn 62 */
+ zSig64 = pSig64 + cSig64;
+ if ((int64_t)zSig64 < 0) {
+ shift64RightJamming(zSig64, 1, &zSig64);
+ } else {
+ zExp--;
+ }
+ } else {
+ /* Subtraction */
+ if (expDiff > 0) {
+ shift64RightJamming(cSig64, expDiff, &cSig64);
+ zSig64 = pSig64 - cSig64;
+ zExp = pExp;
+ } else if (expDiff < 0) {
+ shift64RightJamming(pSig64, -expDiff, &pSig64);
+ zSig64 = cSig64 - pSig64;
+ zExp = cExp;
+ zSign ^= 1;
+ } else {
+ zExp = pExp;
+ if (cSig64 < pSig64) {
+ zSig64 = pSig64 - cSig64;
+ } else if (pSig64 < cSig64) {
+ zSig64 = cSig64 - pSig64;
+ zSign ^= 1;
+ } else {
+ /* Exact zero */
+ zSign = signflip;
+ if (STATUS(float_rounding_mode) == float_round_down) {
+ zSign ^= 1;
+ }
+ return packFloat32(zSign, 0, 0);
+ }
+ }
+ --zExp;
+ /* Normalize to put the explicit bit back into bit 62. */
+ shiftcount = countLeadingZeros64(zSig64) - 1;
+ zSig64 <<= shiftcount;
+ zExp -= shiftcount;
+ }
+ shift64RightJamming(zSig64, 32, &zSig64);
+ return roundAndPackFloat32(zSign, zExp, zSig64 STATUS_VAR);
+}
+
+
+/*----------------------------------------------------------------------------
| Returns the square root of the single-precision floating-point value `a'.
| The operation is performed according to the IEC/IEEE Standard for Binary
| Floating-Point Arithmetic.
@@ -3465,6 +3672,226 @@
}
/*----------------------------------------------------------------------------
+| Returns the result of multiplying the double-precision floating-point values
+| `a' and `b' then adding 'c', with no intermediate rounding step after the
+| multiplication. The operation is performed according to the IEC/IEEE
+| Standard for Binary Floating-Point Arithmetic 754-2008.
+| The flags argument allows the caller to select negation of the
+| addend, the intermediate product, or the final result. (The difference
+| between this and having the caller do a separate negation is that negating
+| externally will flip the sign bit on NaNs.)
+*----------------------------------------------------------------------------*/
+
+float64 float64_muladd(float64 a, float64 b, float64 c, int flags STATUS_PARAM)
+{
+ flag aSign, bSign, cSign, zSign;
+ int aExp, bExp, cExp, pExp, zExp, expDiff;
+ uint64_t aSig, bSig, cSig;
+ flag pInf, pZero, pSign;
+ uint64_t pSig0, pSig1, cSig0, cSig1, zSig0, zSig1;
+ int shiftcount;
+ flag signflip, infzero;
+
+ a = float64_squash_input_denormal(a STATUS_VAR);
+ b = float64_squash_input_denormal(b STATUS_VAR);
+ c = float64_squash_input_denormal(c STATUS_VAR);
+ aSig = extractFloat64Frac(a);
+ aExp = extractFloat64Exp(a);
+ aSign = extractFloat64Sign(a);
+ bSig = extractFloat64Frac(b);
+ bExp = extractFloat64Exp(b);
+ bSign = extractFloat64Sign(b);
+ cSig = extractFloat64Frac(c);
+ cExp = extractFloat64Exp(c);
+ cSign = extractFloat64Sign(c);
+
+ infzero = ((aExp == 0 && aSig == 0 && bExp == 0x7ff && bSig == 0) ||
+ (aExp == 0x7ff && aSig == 0 && bExp == 0 && bSig == 0));
+
+ /* It is implementation-defined whether the cases of (0,inf,qnan)
+ * and (inf,0,qnan) raise InvalidOperation or not (and what QNaN
+ * they return if they do), so we have to hand this information
+ * off to the target-specific pick-a-NaN routine.
+ */
+ if (((aExp == 0x7ff) && aSig) ||
+ ((bExp == 0x7ff) && bSig) ||
+ ((cExp == 0x7ff) && cSig)) {
+ return propagateFloat64MulAddNaN(a, b, c, infzero STATUS_VAR);
+ }
+
+ if (infzero) {
+ float_raise(float_flag_invalid STATUS_VAR);
+ return float64_default_nan;
+ }
+
+ if (flags & float_muladd_negate_c) {
+ cSign ^= 1;
+ }
+
+ signflip = (flags & float_muladd_negate_result) ? 1 : 0;
+
+ /* Work out the sign and type of the product */
+ pSign = aSign ^ bSign;
+ if (flags & float_muladd_negate_product) {
+ pSign ^= 1;
+ }
+ pInf = (aExp == 0x7ff) || (bExp == 0x7ff);
+ pZero = ((aExp | aSig) == 0) || ((bExp | bSig) == 0);
+
+ if (cExp == 0x7ff) {
+ if (pInf && (pSign ^ cSign)) {
+ /* addition of opposite-signed infinities => InvalidOperation */
+ float_raise(float_flag_invalid STATUS_VAR);
+ return float64_default_nan;
+ }
+ /* Otherwise generate an infinity of the same sign */
+ return packFloat64(cSign ^ signflip, 0x7ff, 0);
+ }
+
+ if (pInf) {
+ return packFloat64(pSign ^ signflip, 0x7ff, 0);
+ }
+
+ if (pZero) {
+ if (cExp == 0) {
+ if (cSig == 0) {
+ /* Adding two exact zeroes */
+ if (pSign == cSign) {
+ zSign = pSign;
+ } else if (STATUS(float_rounding_mode) == float_round_down) {
+ zSign = 1;
+ } else {
+ zSign = 0;
+ }
+ return packFloat64(zSign ^ signflip, 0, 0);
+ }
+ /* Exact zero plus a denorm */
+ if (STATUS(flush_to_zero)) {
+ float_raise(float_flag_output_denormal STATUS_VAR);
+ return packFloat64(cSign ^ signflip, 0, 0);
+ }
+ }
+ /* Zero plus something non-zero : just return the something */
+ return c ^ ((uint64_t)signflip << 63);
+ }
+
+ if (aExp == 0) {
+ normalizeFloat64Subnormal(aSig, &aExp, &aSig);
+ }
+ if (bExp == 0) {
+ normalizeFloat64Subnormal(bSig, &bExp, &bSig);
+ }
+
+ /* Calculate the actual result a * b + c */
+
+ /* Multiply first; this is easy. */
+ /* NB: we subtract 0x3fe where float64_mul() subtracts 0x3ff
+ * because we want the true exponent, not the "one-less-than"
+ * flavour that roundAndPackFloat64() takes.
+ */
+ pExp = aExp + bExp - 0x3fe;
+ aSig = (aSig | LIT64(0x0010000000000000))<<10;
+ bSig = (bSig | LIT64(0x0010000000000000))<<11;
+ mul64To128(aSig, bSig, &pSig0, &pSig1);
+ if ((int64_t)(pSig0 << 1) >= 0) {
+ shortShift128Left(pSig0, pSig1, 1, &pSig0, &pSig1);
+ pExp--;
+ }
+
+ zSign = pSign ^ signflip;
+
+ /* Now [pSig0:pSig1] is the significand of the multiply, with the explicit
+ * bit in position 126.
+ */
+ if (cExp == 0) {
+ if (!cSig) {
+ /* Throw out the special case of c being an exact zero now */
+ shift128RightJamming(pSig0, pSig1, 64, &pSig0, &pSig1);
+ return roundAndPackFloat64(zSign, pExp - 1,
+ pSig1 STATUS_VAR);
+ }
+ normalizeFloat64Subnormal(cSig, &cExp, &cSig);
+ }
+
+ /* Shift cSig and add the explicit bit so [cSig0:cSig1] is the
+ * significand of the addend, with the explicit bit in position 126.
+ */
+ cSig0 = cSig << (126 - 64 - 52);
+ cSig1 = 0;
+ cSig0 |= LIT64(0x4000000000000000);
+ expDiff = pExp - cExp;
+
+ if (pSign == cSign) {
+ /* Addition */
+ if (expDiff > 0) {
+ /* scale c to match p */
+ shift128RightJamming(cSig0, cSig1, expDiff, &cSig0, &cSig1);
+ zExp = pExp;
+ } else if (expDiff < 0) {
+ /* scale p to match c */
+ shift128RightJamming(pSig0, pSig1, -expDiff, &pSig0, &pSig1);
+ zExp = cExp;
+ } else {
+ /* no scaling needed */
+ zExp = cExp;
+ }
+ /* Add significands and make sure explicit bit ends up in posn 126 */
+ add128(pSig0, pSig1, cSig0, cSig1, &zSig0, &zSig1);
+ if ((int64_t)zSig0 < 0) {
+ shift128RightJamming(zSig0, zSig1, 1, &zSig0, &zSig1);
+ } else {
+ zExp--;
+ }
+ shift128RightJamming(zSig0, zSig1, 64, &zSig0, &zSig1);
+ return roundAndPackFloat64(zSign, zExp, zSig1 STATUS_VAR);
+ } else {
+ /* Subtraction */
+ if (expDiff > 0) {
+ shift128RightJamming(cSig0, cSig1, expDiff, &cSig0, &cSig1);
+ sub128(pSig0, pSig1, cSig0, cSig1, &zSig0, &zSig1);
+ zExp = pExp;
+ } else if (expDiff < 0) {
+ shift128RightJamming(pSig0, pSig1, -expDiff, &pSig0, &pSig1);
+ sub128(cSig0, cSig1, pSig0, pSig1, &zSig0, &zSig1);
+ zExp = cExp;
+ zSign ^= 1;
+ } else {
+ zExp = pExp;
+ if (lt128(cSig0, cSig1, pSig0, pSig1)) {
+ sub128(pSig0, pSig1, cSig0, cSig1, &zSig0, &zSig1);
+ } else if (lt128(pSig0, pSig1, cSig0, cSig1)) {
+ sub128(cSig0, cSig1, pSig0, pSig1, &zSig0, &zSig1);
+ zSign ^= 1;
+ } else {
+ /* Exact zero */
+ zSign = signflip;
+ if (STATUS(float_rounding_mode) == float_round_down) {
+ zSign ^= 1;
+ }
+ return packFloat64(zSign, 0, 0);
+ }
+ }
+ --zExp;
+ /* Do the equivalent of normalizeRoundAndPackFloat64() but
+ * starting with the significand in a pair of uint64_t.
+ */
+ if (zSig0) {
+ shiftcount = countLeadingZeros64(zSig0) - 1;
+ shortShift128Left(zSig0, zSig1, shiftcount, &zSig0, &zSig1);
+ if (zSig1) {
+ zSig0 |= 1;
+ }
+ zExp -= shiftcount;
+ } else {
+ shiftcount = countLeadingZeros64(zSig1) - 1;
+ zSig0 = zSig1 << shiftcount;
+ zExp -= (shiftcount + 64);
+ }
+ return roundAndPackFloat64(zSign, zExp, zSig0 STATUS_VAR);
+ }
+}
+
+/*----------------------------------------------------------------------------
| Returns the square root of the double-precision floating-point value `a'.
| The operation is performed according to the IEC/IEEE Standard for Binary
| Floating-Point Arithmetic.
diff --git a/fpu/softfloat.h b/fpu/softfloat.h
index 618ddee..07c2929 100644
--- a/fpu/softfloat.h
+++ b/fpu/softfloat.h
@@ -212,6 +212,18 @@
void float_raise( int8 flags STATUS_PARAM);
/*----------------------------------------------------------------------------
+| Options to indicate which negations to perform in float*_muladd()
+| Using these differs from negating an input or output before calling
+| the muladd function in that this means that a NaN doesn't have its
+| sign bit inverted before it is propagated.
+*----------------------------------------------------------------------------*/
+enum {
+ float_muladd_negate_c = 1,
+ float_muladd_negate_product = 2,
+ float_muladd_negate_result = 3,
+};
+
+/*----------------------------------------------------------------------------
| Software IEC/IEEE integer-to-floating-point conversion routines.
*----------------------------------------------------------------------------*/
float32 int32_to_float32( int32 STATUS_PARAM );
@@ -269,6 +281,7 @@
float32 float32_mul( float32, float32 STATUS_PARAM );
float32 float32_div( float32, float32 STATUS_PARAM );
float32 float32_rem( float32, float32 STATUS_PARAM );
+float32 float32_muladd(float32, float32, float32, int STATUS_PARAM);
float32 float32_sqrt( float32 STATUS_PARAM );
float32 float32_exp2( float32 STATUS_PARAM );
float32 float32_log2( float32 STATUS_PARAM );
@@ -375,6 +388,7 @@
float64 float64_mul( float64, float64 STATUS_PARAM );
float64 float64_div( float64, float64 STATUS_PARAM );
float64 float64_rem( float64, float64 STATUS_PARAM );
+float64 float64_muladd(float64, float64, float64, int STATUS_PARAM);
float64 float64_sqrt( float64 STATUS_PARAM );
float64 float64_log2( float64 STATUS_PARAM );
int float64_eq( float64, float64 STATUS_PARAM );
