2007-04-01 Jakub Jelinek <jakub@redhat.com>

	* sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c (__nextafterl): Use
	math_opt_barrier and math_force_eval macros.

2007-03-27  Jakub Jelinek  <jakub@redhat.com>

	[BZ #3306]
	* math/math_private.h (math_opt_barrier, math_force_eval): Define.
	* sysdeps/i386/fpu/math_private.h: New file.
	* sysdeps/x86_64/fpu/math_private.h: New file.
	* math/s_nexttowardf.c (__nexttowardf): Use math_opt_barrier and
	math_force_eval macros.  Use "+m" constraint on asm rather than
	"=m" and "m".
	* math/s_nextafter.c (__nextafter): Likewise.
	* sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c (__nexttoward):
	Likewise.
	* sysdeps/ieee754/flt-32/s_nextafterf.c (__nextafterf): Likewise.
	* sysdeps/ieee754/ldbl-128/s_nexttoward.c (__nexttoward): Likewise.
	* sysdeps/ieee754/ldbl-96/s_nexttoward.c (__nexttoward): Likewise.
	* sysdeps/i386/fpu/s_nextafterl.c (__nextafterl): Use
	math_opt_barrier and math_force_eval macros.
	* sysdeps/ieee754/ldbl-128/s_nextafterl.c (__nextafterl): Likewise.
	* sysdeps/ieee754/ldbl-96/s_nextafterl.c (__nextafterl): Likewise.
	* sysdeps/i386/fpu/s_nexttoward.c: Include float.h.
	(__nexttoward): Use math_opt_barrier and
	math_force_eval macros.  Use "+m" constraint on asm rather than
	"=m" and "m".  Only use asm to force double result if
	FLT_EVAL_METHOD is 2.
	* sysdeps/i386/fpu/s_nexttowardf.c: Include float.h.
	(__nexttowardf): Use math_opt_barrier and
	math_force_eval macros.  Use "+m" constraint on asm rather than
	"=m" and "m".  Only use asm to force double result if
	FLT_EVAL_METHOD is not 0.
	* sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c: Include float.h.
	(__nexttowardf): Use math_opt_barrier and
	math_force_eval macros.  If FLT_EVAL_METHOD is not 0, force
	x to float using asm.
	* sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c: Include float.h.
	(__nldbl_nexttowardf): Use math_opt_barrier and
	math_force_eval macros.  If FLT_EVAL_METHOD is not 0, force
	x to float using asm.
	* sysdeps/ieee754/ldbl-96/s_nexttowardf.c: Include float.h.
	(__nexttowardf): Use math_opt_barrier and math_force_eval
	macros.  If FLT_EVAL_METHOD is not 0, force x to float using asm.
	* math/bug-nextafter.c (zero, inf): New variables.
	(main): Add new tests.
	* math/bug-nexttoward.c (zero, inf): New variables.
	(main): Add new tests.

15 files changed, 197 insertions, 139 deletions

diff --git a/sysdeps/i386/fpu/math_private.h b/sysdeps/i386/fpu/math_private.h
new file mode 100644
index 0000000000..a426788ef1
--- /dev/null
+++ b/sysdeps/i386/fpu/math_private.h
@@ -0,0 +1,18 @@
+#ifndef _MATH_PRIVATE_H
+
+#define math_opt_barrier(x) \
+({ __typeof(x) __x;					\
+   __asm ("" : "=t" (__x) : "0" (x));			\
+   __x; })
+#define math_force_eval(x) \
+do							\
+  {							\
+    if (sizeof (x) <= sizeof (double))			\
+      __asm __volatile ("" : : "m" (x));		\
+    else						\
+      __asm __volatile ("" : : "f" (x));		\
+  }							\
+while (0)
+
+#include <math/math_private.h>
+#endif
diff --git a/sysdeps/i386/fpu/s_nextafterl.c b/sysdeps/i386/fpu/s_nextafterl.c
index 5b617cb4e7..aef0a144e5 100644
--- a/sysdeps/i386/fpu/s_nextafterl.c
+++ b/sysdeps/i386/fpu/s_nextafterl.c
@@ -27,7 +27,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 
 #ifdef __STDC__
 	long double __nextafterl(long double x, long double y)
@@ -52,9 +52,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if(x==y) return y;		/* x=y, return y */
 	if((ix|hx|lx)==0) {			/* x == 0 */
+	    long double u;
 	    SET_LDOUBLE_WORDS(x,esy&0x8000,0,1);/* return +-minsubnormal */
-	    y = x*x;
-	    if(y==x) return y; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(esx>=0) {			/* x > 0 */
 	    if(esx>esy||((esx==esy) && (hx>hy||((hx==hy)&&(lx>ly))))) {
@@ -109,12 +112,9 @@ static char rcsid[] = "$NetBSD: $";
 	}
 	esy = esx&0x7fff;
 	if(esy==0x7fff) return x+x;	/* overflow  */
-	if(esy==0) {			/* underflow */
-	    y = x*x;
-	    if(y!=x) {		/* raise underflow flag */
-	        SET_LDOUBLE_WORDS(y,esx,hx,lx);
-		return y;
-	    }
+	if(esy==0) {
+	    long double u = x*x;		/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_LDOUBLE_WORDS(x,esx,hx,lx);
 	return x;
diff --git a/sysdeps/i386/fpu/s_nexttoward.c b/sysdeps/i386/fpu/s_nexttoward.c
index 2bd768e448..9bd86a3724 100644
--- a/sysdeps/i386/fpu/s_nexttoward.c
+++ b/sysdeps/i386/fpu/s_nexttoward.c
@@ -27,7 +27,8 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
+#include <float.h>
 
 #ifdef __STDC__
 	double __nexttoward(double x, long double y)
@@ -52,10 +53,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if((ix|lx)==0) {			/* x == 0 */
-	    double x2;
+	    double u;
 	    INSERT_WORDS(x,(esy&0x8000)<<16,1); /* return +-minsub */
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if (esy>=0x8000||((ix>>20)&0x7ff)>iy-0x3c00
@@ -85,16 +88,14 @@ static char rcsid[] = "$NetBSD: $";
 	hy = hx&0x7ff00000;
 	if(hy>=0x7ff00000) {
 	  x = x+x;	/* overflow  */
-	  /* Force conversion to double.  */
-	  asm ("" : "=m"(x) : "m"(x));
+	  if (FLT_EVAL_METHOD != 0 && FLT_EVAL_METHOD != 1)
+	    /* Force conversion to double.  */
+	    asm ("" : "+m"(x));
 	  return x;
 	}
-	if(hy<0x00100000) {		/* underflow */
-	    double x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        INSERT_WORDS(x2,hx,lx);
-		return x2;
-	    }
+	if(hy<0x00100000) {
+	    double u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	INSERT_WORDS(x,hx,lx);
 	return x;
diff --git a/sysdeps/i386/fpu/s_nexttowardf.c b/sysdeps/i386/fpu/s_nexttowardf.c
index 3fbe53c338..25f70e4f4d 100644
--- a/sysdeps/i386/fpu/s_nexttowardf.c
+++ b/sysdeps/i386/fpu/s_nexttowardf.c
@@ -19,7 +19,8 @@ static char rcsid[] = "$NetBSD: $";
 #endif
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
+#include <float.h>
 
 #ifdef __STDC__
 	float __nexttowardf(float x, long double y)
@@ -44,10 +45,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if(ix==0) {				/* x == 0 */
-	    float x2;
+	    float u;
 	    SET_FLOAT_WORD(x,((esy&0x8000)<<16)|1);/* return +-minsub*/
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if(esy>=0x8000||((ix>>23)&0xff)>iy-0x3f80
@@ -69,16 +72,14 @@ static char rcsid[] = "$NetBSD: $";
 	hy = hx&0x7f800000;
 	if(hy>=0x7f800000) {
 	  x = x+x;	/* overflow  */
-	  /* Force conversion to float.  */
-	  asm ("" : "=m"(x) : "m"(x));
+	  if (FLT_EVAL_METHOD != 0)
+	    /* Force conversion to float.  */
+	    asm ("" : "+m"(x));
 	  return x;
 	}
-	if(hy<0x00800000) {		/* underflow */
-	    float x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        SET_FLOAT_WORD(x2,hx);
-		return x2;
-	    }
+	if(hy<0x00800000) {
+	    float u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_FLOAT_WORD(x,hx);
 	return x;
diff --git a/sysdeps/ieee754/flt-32/s_nextafterf.c b/sysdeps/ieee754/flt-32/s_nextafterf.c
index e1568e24c9..600c14621f 100644
--- a/sysdeps/ieee754/flt-32/s_nextafterf.c
+++ b/sysdeps/ieee754/flt-32/s_nextafterf.c
@@ -18,7 +18,7 @@ static char rcsid[] = "$NetBSD: s_nextafterf.c,v 1.4 1995/05/10 20:48:01 jtc Exp
 #endif
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 #include <float.h>
 
 #ifdef __STDC__
@@ -40,9 +40,12 @@ static char rcsid[] = "$NetBSD: s_nextafterf.c,v 1.4 1995/05/10 20:48:01 jtc Exp
 	   return x+y;
 	if(x==y) return y;		/* x=y, return y */
 	if(ix==0) {				/* x == 0 */
+	    float u;
 	    SET_FLOAT_WORD(x,(hy&0x80000000)|1);/* return +-minsubnormal */
-	    y = x*x;
-	    if(y==x) return y; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u*u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if(hx>hy) {				/* x > y, x -= ulp */
@@ -61,15 +64,12 @@ static char rcsid[] = "$NetBSD: s_nextafterf.c,v 1.4 1995/05/10 20:48:01 jtc Exp
 	if(hy>=0x7f800000) {
 	  x = x+x;	/* overflow  */
 	  if (FLT_EVAL_METHOD != 0)
-	    asm ("" : "=m"(x) : "m"(x));
+	    asm ("" : "+m"(x));
 	  return x;	/* overflow  */
 	}
-	if(hy<0x00800000) {		/* underflow */
-	    y = x*x;
-	    if(y!=x) {		/* raise underflow flag */
-	        SET_FLOAT_WORD(y,hx);
-		return y;
-	    }
+	if(hy<0x00800000) {
+	    float u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_FLOAT_WORD(x,hx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-128/s_nextafterl.c b/sysdeps/ieee754/ldbl-128/s_nextafterl.c
index d3df668178..7bc869a518 100644
--- a/sysdeps/ieee754/ldbl-128/s_nextafterl.c
+++ b/sysdeps/ieee754/ldbl-128/s_nextafterl.c
@@ -25,7 +25,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 
 #ifdef __STDC__
 	long double __nextafterl(long double x, long double y)
@@ -47,9 +47,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if(x==y) return y;		/* x=y, return y */
 	if((ix|lx)==0) {			/* x == 0 */
+	    long double u;
 	    SET_LDOUBLE_WORDS64(x,hy&0x8000000000000000ULL,1);/* return +-minsubnormal */
-	    y = x*x;
-	    if(y==x) return y; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {			/* x > 0 */
 	    if(hx>hy||((hx==hy)&&(lx>ly))) {	/* x > y, x -= ulp */
@@ -70,12 +73,9 @@ static char rcsid[] = "$NetBSD: $";
 	}
 	hy = hx&0x7fff000000000000LL;
 	if(hy==0x7fff000000000000LL) return x+x;/* overflow  */
-	if(hy==0) {				/* underflow */
-	    y = x*x;
-	    if(y!=x) {		/* raise underflow flag */
-	        SET_LDOUBLE_WORDS64(y,hx,lx);
-		return y;
-	    }
+	if(hy==0) {
+	    long double u = x*x;		/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_LDOUBLE_WORDS64(x,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-128/s_nexttoward.c b/sysdeps/ieee754/ldbl-128/s_nexttoward.c
index 553e401973..178505c58f 100644
--- a/sysdeps/ieee754/ldbl-128/s_nexttoward.c
+++ b/sysdeps/ieee754/ldbl-128/s_nexttoward.c
@@ -26,7 +26,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 #include <float.h>
 
 #ifdef __STDC__
@@ -53,10 +53,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if((ix|lx)==0) {			/* x == 0 */
-	    double x2;
+	    double u;
 	    INSERT_WORDS(x,(u_int32_t)((hy>>32)&0x80000000),1);/* return +-minsub */
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if (hy<0||(ix>>20)>(iy>>48)-0x3c00
@@ -87,16 +89,13 @@ static char rcsid[] = "$NetBSD: $";
 	if(hy>=0x7ff00000) {
 	  x = x+x;	/* overflow  */
 	  if (FLT_EVAL_METHOD != 0 && FLT_EVAL_METHOD != 1)
-	    /* Force conversion to float.  */
-	    asm ("" : "=m"(x) : "m"(x));
+	    /* Force conversion to double.  */
+	    asm ("" : "+m"(x));
 	  return x;
 	}
-	if(hy<0x00100000) {		/* underflow */
-	    double x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        INSERT_WORDS(x2,hx,lx);
-		return x2;
-	    }
+	if(hy<0x00100000) {
+	    double u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	INSERT_WORDS(x,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c b/sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c
index e35ce50829..4f8aa7dec1 100644
--- a/sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c
+++ b/sysdeps/ieee754/ldbl-128ibm/s_nextafterl.c
@@ -24,8 +24,8 @@ static char rcsid[] = "$NetBSD: $";
  *   Special cases:
  */
 
-#include "math.h"
-#include "math_private.h"
+#include <math.h>
+#include <math_private.h>
 #include <math_ldbl_opt.h>
 
 #ifdef __STDC__
@@ -53,9 +53,12 @@ static char rcsid[] = "$NetBSD: $";
 	if(x==y)
 	    return y;		/* x=y, return y */
 	if(ihx == 0 && ilx == 0) {			/* x == 0 */
+	    long double u;
 	    SET_LDOUBLE_WORDS64(x,hy&0x8000000000000000ULL,1);/* return +-minsubnormal */
-	    y = x*x;
-	    if(y==x) return y; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (u);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(ihx>=0) {			/* x > 0 */
 	    if(ihx>ihy||((ihx==ihy)&&(ilx>ily))) {	/* x > y, x -= ulp */
@@ -93,12 +96,9 @@ static char rcsid[] = "$NetBSD: $";
 	}
 	hy = hx&0x7ff0000000000000LL;
 	if(hy==0x7ff0000000000000LL) return x+x;/* overflow  */
-	if(hy==0) {				/* underflow */
-	    y = x*x;
-	    if(y!=x) {		/* raise underflow flag */
-	        SET_LDOUBLE_WORDS64(y,hx,lx);
-		return y;
-	    }
+	if(hy==0) {
+	    long double u = x * x;		/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_LDOUBLE_WORDS64(x,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c b/sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c
index 3335100592..e2f6521f57 100644
--- a/sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c
+++ b/sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c
@@ -26,7 +26,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 #include <math_ldbl_opt.h>
 #include <float.h>
 
@@ -55,10 +55,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if((ix|lx)==0) {			/* x == 0 */
-	    double x2;
+	    double u;
 	    INSERT_WORDS(x,(u_int32_t)((hy>>32)&0x80000000),1);/* return +-minsub */
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if (hy<0||(ix>>20)>(iy>>52)
@@ -89,16 +91,13 @@ static char rcsid[] = "$NetBSD: $";
 	if(hy>=0x7ff00000) {
 	  x = x+x;	/* overflow  */
 	  if (FLT_EVAL_METHOD != 0 && FLT_EVAL_METHOD != 1)
-	    /* Force conversion to float.  */
-	    asm ("" : "=m"(x) : "m"(x));
+	    /* Force conversion to double.  */
+	    asm ("" : "+m"(x));
 	  return x;
 	}
-	if(hy<0x00100000) {		/* underflow */
-	    double x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        INSERT_WORDS(x2,hx,lx);
-		return x2;
-	    }
+	if(hy<0x00100000) {
+	    double u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	INSERT_WORDS(x,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c b/sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c
index a9373ff822..cf655fad16 100644
--- a/sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c
+++ b/sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c
@@ -19,8 +19,9 @@ static char rcsid[] = "$NetBSD: $";
 #endif
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 #include <math_ldbl_opt.h>
+#include <float.h>
 
 #ifdef __STDC__
 	float __nexttowardf(float x, long double y)
@@ -46,10 +47,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if(ix==0) {				/* x == 0 */
-	    float x2;
+	    float u;
 	    SET_FLOAT_WORD(x,(u_int32_t)((hy>>32)&0x80000000)|1);/* return +-minsub*/
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if(hy<0||(ix>>23)>(iy>>52)-0x380
@@ -69,13 +72,16 @@ static char rcsid[] = "$NetBSD: $";
 	    }
 	}
 	hy = hx&0x7f800000;
-	if(hy>=0x7f800000) return x+x;	/* overflow  */
+	if(hy>=0x7f800000) {
+	  x = x+x;	/* overflow  */
+	  if (FLT_EVAL_METHOD != 0)
+	    /* Force conversion to float.  */
+	    asm ("" : "+m"(x));
+	  return x;
+	}
 	if(hy<0x00800000) {		/* underflow */
-	    float x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        SET_FLOAT_WORD(x2,hx);
-		return x2;
-	    }
+	    float u = x*x;
+	    math_force_eval (u);	/* raise underflow flag */
 	}
 	SET_FLOAT_WORD(x,hx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-96/s_nextafterl.c b/sysdeps/ieee754/ldbl-96/s_nextafterl.c
index 1798261b28..ed0877d6dd 100644
--- a/sysdeps/ieee754/ldbl-96/s_nextafterl.c
+++ b/sysdeps/ieee754/ldbl-96/s_nextafterl.c
@@ -26,7 +26,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 
 #ifdef __STDC__
 	long double __nextafterl(long double x, long double y)
@@ -48,9 +48,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if(x==y) return y;		/* x=y, return y */
 	if((ix|hx|lx)==0) {			/* x == 0 */
+	    long double u;
 	    SET_LDOUBLE_WORDS(x,esy&0x8000,0,1);/* return +-minsubnormal */
-	    y = x*x;
-	    if(y==x) return y; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(esx<0x8000) {			/* x > 0 */
 	    if(ix>iy||((ix==iy) && (hx>hy||((hx==hy)&&(lx>ly))))) {
@@ -85,13 +88,10 @@ static char rcsid[] = "$NetBSD: $";
 	    }
 	}
 	esy = esx&0x7fff;
-	if(esy==0x7fff) return x+x;	/* overflow  */
-	if(esy==0) {			/* underflow */
-	    y = x*x;
-	    if(y!=x) {		/* raise underflow flag */
-	        SET_LDOUBLE_WORDS(y,esx,hx,lx);
-		return y;
-	    }
+	if(esy==0x7fff) return x+x;		/* overflow  */
+	if(esy==0) {
+	    long double u = x*x;		/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_LDOUBLE_WORDS(x,esx,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-96/s_nexttoward.c b/sysdeps/ieee754/ldbl-96/s_nexttoward.c
index 7945cb5cb1..e30438482c 100644
--- a/sysdeps/ieee754/ldbl-96/s_nexttoward.c
+++ b/sysdeps/ieee754/ldbl-96/s_nexttoward.c
@@ -26,7 +26,7 @@ static char rcsid[] = "$NetBSD: $";
  */
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
 #include <float.h>
 
 #ifdef __STDC__
@@ -50,10 +50,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if((ix|lx)==0) {			/* x == 0 */
-	    double x2;
+	    double u;
 	    INSERT_WORDS(x,(esy&0x8000)<<16,1); /* return +-minsub */
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if (esy>=0x8000||((ix>>20)&0x7ff)>iy-0x3c00
@@ -84,16 +86,13 @@ static char rcsid[] = "$NetBSD: $";
 	if(hy>=0x7ff00000) {
 	  x = x+x;	/* overflow  */
 	  if (FLT_EVAL_METHOD != 0 && FLT_EVAL_METHOD != 1)
-	    /* Force conversion to float.  */
-	    asm ("" : "=m"(x) : "m"(x));
+	    /* Force conversion to double.  */
+	    asm ("" : "+m"(x));
 	  return x;
 	}
-	if(hy<0x00100000) {		/* underflow */
-	    double x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        INSERT_WORDS(x2,hx,lx);
-		return x2;
-	    }
+	if(hy<0x00100000) {
+	    double u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	INSERT_WORDS(x,hx,lx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-96/s_nexttowardf.c b/sysdeps/ieee754/ldbl-96/s_nexttowardf.c
index a1c38b5d4c..6357975ffc 100644
--- a/sysdeps/ieee754/ldbl-96/s_nexttowardf.c
+++ b/sysdeps/ieee754/ldbl-96/s_nexttowardf.c
@@ -18,7 +18,8 @@ static char rcsid[] = "$NetBSD: $";
 #endif
 
 #include "math.h"
-#include "math_private.h"
+#include <math_private.h>
+#include <float.h>
 
 #ifdef __STDC__
 	float __nexttowardf(float x, long double y)
@@ -41,10 +42,12 @@ static char rcsid[] = "$NetBSD: $";
 	   return x+y;
 	if((long double) x==y) return y;	/* x=y, return y */
 	if(ix==0) {				/* x == 0 */
-	    float x2;
+	    float u;
 	    SET_FLOAT_WORD(x,((esy&0x8000)<<16)|1);/* return +-minsub*/
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x;	/* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if(esy>=0x8000||((ix>>23)&0xff)>iy-0x3f80
@@ -64,13 +67,16 @@ static char rcsid[] = "$NetBSD: $";
 	    }
 	}
 	hy = hx&0x7f800000;
-	if(hy>=0x7f800000) return x+x;	/* overflow  */
-	if(hy<0x00800000) {		/* underflow */
-	    float x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-	        SET_FLOAT_WORD(x2,hx);
-		return x2;
-	    }
+	if(hy>=0x7f800000) {
+	  x = x+x;	/* overflow  */
+	  if (FLT_EVAL_METHOD != 0)
+	    /* Force conversion to float.  */
+	    asm ("" : "+m"(x));
+	  return x;
+	}
+	if(hy<0x00800000) {
+	    float u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_FLOAT_WORD(x,hx);
 	return x;
diff --git a/sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c b/sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c
index d52526f719..68027f26fa 100644
--- a/sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c
+++ b/sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c
@@ -20,7 +20,10 @@
  *   Special cases:
  */
 
+#include <math.h>
+#include <math_private.h>
 #include <math_ldbl_opt.h>
+#include <float.h>
 
 float __nldbl_nexttowardf(float x, double y);
 
@@ -39,10 +42,12 @@ float __nldbl_nexttowardf(float x, double y)
 	   return x+y;
 	if((double) x==y) return y;		/* x=y, return y */
 	if(ix==0) {				/* x == 0 */
-	    float x2;
+	    float u;
 	    SET_FLOAT_WORD(x,(u_int32_t)(hy&0x80000000)|1);/* return +-minsub*/
-	    x2 = x*x;
-	    if(x2==x) return x2; else return x; /* raise underflow flag */
+	    u = math_opt_barrier (x);
+	    u = u * u;
+	    math_force_eval (u);		/* raise underflow flag */
+	    return x;
 	}
 	if(hx>=0) {				/* x > 0 */
 	    if(hy<0||(ix>>23)>(iy>>20)-0x380
@@ -60,13 +65,16 @@ float __nldbl_nexttowardf(float x, double y)
 		hx += 1;
 	}
 	hy = hx&0x7f800000;
-	if(hy>=0x7f800000) return x+x;	/* overflow  */
-	if(hy<0x00800000) {		/* underflow */
-	    float x2 = x*x;
-	    if(x2!=x) {		/* raise underflow flag */
-		SET_FLOAT_WORD(x2,hx);
-		return x2;
-	    }
+	if(hy>=0x7f800000) {
+	  x = x+x;	/* overflow  */
+	  if (FLT_EVAL_METHOD != 0)
+	    /* Force conversion to float.  */
+	    asm ("" : "+m"(x));
+	  return x;
+	}
+	if(hy<0x00800000) {
+	    float u = x*x;			/* underflow */
+	    math_force_eval (u);		/* raise underflow flag */
 	}
 	SET_FLOAT_WORD(x,hx);
 	return x;
diff --git a/sysdeps/x86_64/fpu/math_private.h b/sysdeps/x86_64/fpu/math_private.h
new file mode 100644
index 0000000000..4febcbb5ec
--- /dev/null
+++ b/sysdeps/x86_64/fpu/math_private.h
@@ -0,0 +1,21 @@
+#ifndef _MATH_PRIVATE_H
+
+#define math_opt_barrier(x) \
+({ __typeof(x) __x;					\
+   if (sizeof (x) <= sizeof (double))			\
+     __asm ("" : "=x" (__x) : "0" (x));			\
+   else							\
+     __asm ("" : "=t" (__x) : "0" (x));			\
+   __x; })
+#define math_force_eval(x) \
+do							\
+  {							\
+    if (sizeof (x) <= sizeof (double))			\
+      __asm __volatile ("" : : "x" (x));		\
+    else						\
+      __asm __volatile ("" : : "f" (x));		\
+  }							\
+while (0)
+
+#include <math/math_private.h>
+#endif