X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=src%2Fmath%2Ffma.c;h=741ccd757ab1544d5ed4d6fd2bd135d7cb0586e3;hb=fd427c4eaeed4d1099acc7b6188e41c6cfa5bca5;hp=89def795943573d1568c5328ee95b2c9efc27b36;hpb=ffd8ac2dd50f99c3c83d7d9d845df9874ec3e7d5;p=musl

diff --git a/src/math/fma.c b/src/math/fma.c
index 89def795..741ccd75 100644
--- a/src/math/fma.c
+++ b/src/math/fma.c
@@ -2,16 +2,6 @@
 #include "libm.h"
 
 #if LDBL_MANT_DIG==64 && LDBL_MAX_EXP==16384
-union ld80 {
-	long double x;
-	struct {
-		uint64_t m;
-		uint16_t e : 15;
-		uint16_t s : 1;
-		uint16_t pad;
-	} bits;
-};
-
 /* exact add, assumes exponent_x >= exponent_y */
 static void add(long double *hi, long double *lo, long double x, long double y)
 {
@@ -45,25 +35,25 @@ return an adjusted hi so that rounding it to double (or less) precision is corre
 */
 static long double adjust(long double hi, long double lo)
 {
-	union ld80 uhi, ulo;
+	union ldshape uhi, ulo;
 
 	if (lo == 0)
 		return hi;
-	uhi.x = hi;
-	if (uhi.bits.m & 0x3ff)
+	uhi.f = hi;
+	if (uhi.i.m & 0x3ff)
 		return hi;
-	ulo.x = lo;
-	if (uhi.bits.s == ulo.bits.s)
-		uhi.bits.m++;
+	ulo.f = lo;
+	if ((uhi.i.se & 0x8000) == (ulo.i.se & 0x8000))
+		uhi.i.m++;
 	else {
-		uhi.bits.m--;
 		/* handle underflow and take care of ld80 implicit msb */
-		if (uhi.bits.m == (uint64_t)-1/2) {
-			uhi.bits.m *= 2;
-			uhi.bits.e--;
+		if (uhi.i.m << 1 == 0) {
+			uhi.i.m = 0;
+			uhi.i.se--;
 		}
+		uhi.i.m--;
 	}
-	return uhi.x;
+	return uhi.f;
 }
 
 /* adjusted add so the result is correct when rounded to double (or less) precision */
@@ -82,9 +72,9 @@ static long double dmul(long double x, long double y)
 
 static int getexp(long double x)
 {
-	union ld80 u;
-	u.x = x;
-	return u.bits.e;
+	union ldshape u;
+	u.f = x;
+	return u.i.se & 0x7fff;
 }
 
 double fma(double x, double y, double z)
@@ -242,16 +232,16 @@ static inline struct dd dd_add(double a, double b)
 static inline double add_adjusted(double a, double b)
 {
 	struct dd sum;
-	uint64_t hibits, lobits;
+	union {double f; uint64_t i;} uhi, ulo;
 
 	sum = dd_add(a, b);
 	if (sum.lo != 0) {
-		EXTRACT_WORD64(hibits, sum.hi);
-		if ((hibits & 1) == 0) {
+		uhi.f = sum.hi;
+		if ((uhi.i & 1) == 0) {
 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
-			EXTRACT_WORD64(lobits, sum.lo);
-			hibits += 1 - ((hibits ^ lobits) >> 62);
-			INSERT_WORD64(sum.hi, hibits);
+			ulo.f = sum.lo;
+			uhi.i += 1 - ((uhi.i ^ ulo.i) >> 62);
+			sum.hi = uhi.f;
 		}
 	}
 	return (sum.hi);
@@ -265,7 +255,7 @@ static inline double add_adjusted(double a, double b)
 static inline double add_and_denormalize(double a, double b, int scale)
 {
 	struct dd sum;
-	uint64_t hibits, lobits;
+	union {double f; uint64_t i;} uhi, ulo;
 	int bits_lost;
 
 	sum = dd_add(a, b);
@@ -281,13 +271,13 @@ static inline double add_and_denormalize(double a, double b, int scale)
 	 * break the ties manually.
 	 */
 	if (sum.lo != 0) {
-		EXTRACT_WORD64(hibits, sum.hi);
-		bits_lost = -((int)(hibits >> 52) & 0x7ff) - scale + 1;
-		if (bits_lost != 1 ^ (int)(hibits & 1)) {
+		uhi.f = sum.hi;
+		bits_lost = -((int)(uhi.i >> 52) & 0x7ff) - scale + 1;
+		if ((bits_lost != 1) ^ (int)(uhi.i & 1)) {
 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
-			EXTRACT_WORD64(lobits, sum.lo);
-			hibits += 1 - (((hibits ^ lobits) >> 62) & 2);
-			INSERT_WORD64(sum.hi, hibits);
+			ulo.f = sum.lo;
+			uhi.i += 1 - (((uhi.i ^ ulo.i) >> 62) & 2);
+			sum.hi = uhi.f;
 		}
 	}
 	return scalbn(sum.hi, scale);
@@ -441,10 +431,24 @@ double fma(double x, double y, double z)
 		/*
 		 * There is no need to worry about double rounding in directed
 		 * rounding modes.
+		 * But underflow may not be raised properly, example in downward rounding:
+		 * fma(0x1.000000001p-1000, 0x1.000000001p-30, -0x1p-1066)
 		 */
+		double ret;
+#if defined(FE_INEXACT) && defined(FE_UNDERFLOW)
+		int e = fetestexcept(FE_INEXACT);
+		feclearexcept(FE_INEXACT);
+#endif
 		fesetround(oround);
 		adj = r.lo + xy.lo;
-		return scalbn(r.hi + adj, spread);
+		ret = scalbn(r.hi + adj, spread);
+#if defined(FE_INEXACT) && defined(FE_UNDERFLOW)
+		if (ilogb(ret) < -1022 && fetestexcept(FE_INEXACT))
+			feraiseexcept(FE_UNDERFLOW);
+		else if (e)
+			feraiseexcept(FE_INEXACT);
+#endif
+		return ret;
 	}
 
 	adj = add_adjusted(r.lo, xy.lo);