4 files changed, 33 insertions, 23 deletions
diff --git a/sysdeps/i386/fpu/s_ceil.S b/sysdeps/i386/fpu/s_ceil.S
index f32fa26d34..1226bb2f87 100644
--- a/sysdeps/i386/fpu/s_ceil.S
+++ b/sysdeps/i386/fpu/s_ceil.S
@@ -9,10 +9,10 @@ RCSID("$NetBSD: s_ceil.S,v 1.4 1995/05/08 23:52:13 jtc Exp $")
 
 ENTRY(__ceil)
 	fldl	4(%esp)
-	subl	$8,%esp
-	cfi_adjust_cfa_offset (8)
+	subl	$32,%esp
+	cfi_adjust_cfa_offset (32)
 
-	fstcw	4(%esp)			/* store fpu control word */
+	fnstenv	4(%esp)			/* store fpu environment */
 
 	/* We use here %edx although only the low 1 bits are defined.
 	   But none of the operations should care and they are faster
@@ -25,10 +25,10 @@ ENTRY(__ceil)
 
 	frndint				/* round */
 
-	fldcw	4(%esp)			/* restore original control word */
+	fldenv	4(%esp)			/* restore original environment */
 
-	addl	$8,%esp
-	cfi_adjust_cfa_offset (-8)
+	addl	$32,%esp
+	cfi_adjust_cfa_offset (-32)
 	ret
 END (__ceil)
 weak_alias (__ceil, ceil)
diff --git a/sysdeps/i386/fpu/s_ceilf.S b/sysdeps/i386/fpu/s_ceilf.S
index 4fe703b179..d345c0973b 100644
--- a/sysdeps/i386/fpu/s_ceilf.S
+++ b/sysdeps/i386/fpu/s_ceilf.S
@@ -9,10 +9,10 @@ RCSID("$NetBSD: s_ceilf.S,v 1.3 1995/05/08 23:52:44 jtc Exp $")
 
 ENTRY(__ceilf)
 	flds	4(%esp)
-	subl	$8,%esp
-	cfi_adjust_cfa_offset (8)
+	subl	$32,%esp
+	cfi_adjust_cfa_offset (32)
 
-	fstcw	4(%esp)			/* store fpu control word */
+	fnstenv	4(%esp)			/* store fpu environment */
 
 	/* We use here %edx although only the low 1 bits are defined.
 	   But none of the operations should care and they are faster
@@ -25,10 +25,10 @@ ENTRY(__ceilf)
 
 	frndint				/* round */
 
-	fldcw	4(%esp)			/* restore original control word */
+	fldenv	4(%esp)			/* restore original environment */
 
-	addl	$8,%esp
-	cfi_adjust_cfa_offset (-8)
+	addl	$32,%esp
+	cfi_adjust_cfa_offset (-32)
 	ret
 END (__ceilf)
 weak_alias (__ceilf, ceilf)
diff --git a/sysdeps/i386/fpu/s_ceill.S b/sysdeps/i386/fpu/s_ceill.S
index 4b272c522a..7c08f43b24 100644
--- a/sysdeps/i386/fpu/s_ceill.S
+++ b/sysdeps/i386/fpu/s_ceill.S
@@ -10,10 +10,10 @@ RCSID("$NetBSD: $")
 
 ENTRY(__ceill)
 	fldt	4(%esp)
-	subl	$8,%esp
-	cfi_adjust_cfa_offset (8)
+	subl	$32,%esp
+	cfi_adjust_cfa_offset (32)
 
-	fstcw	4(%esp)			/* store fpu control word */
+	fnstenv	4(%esp)			/* store fpu environment */
 
 	/* We use here %edx although only the low 1 bits are defined.
 	   But none of the operations should care and they are faster
@@ -26,10 +26,15 @@ ENTRY(__ceill)
 
 	frndint				/* round */
 
-	fldcw	4(%esp)			/* restore original control word */
+	/* Preserve "invalid" exceptions from sNaN input.  */
+	fnstsw
+	andl	$0x1, %eax
+	orl	%eax, 8(%esp)
 
-	addl	$8,%esp
-	cfi_adjust_cfa_offset (-8)
+	fldenv	4(%esp)			/* restore original environment */
+
+	addl	$32,%esp
+	cfi_adjust_cfa_offset (-32)
 	ret
 END (__ceill)
 weak_alias (__ceill, ceill)
diff --git a/sysdeps/x86_64/fpu/s_ceill.S b/sysdeps/x86_64/fpu/s_ceill.S
index 910c371d58..9d8b79dbee 100644
--- a/sysdeps/x86_64/fpu/s_ceill.S
+++ b/sysdeps/x86_64/fpu/s_ceill.S
@@ -11,20 +11,25 @@
 ENTRY(__ceill)
 	fldt	8(%rsp)
 
-	fstcw	-4(%rsp)		/* store fpu control word */
+	fnstenv	-28(%rsp)		/* store fpu environment */
 
 	/* We use here %edx although only the low 1 bits are defined.
 	   But none of the operations should care and they are faster
 	   than the 16 bit operations.  */
 	movl	$0x0800,%edx		/* round towards +oo */
-	orl	-4(%rsp),%edx
+	orl	-28(%rsp),%edx
 	andl	$0xfbff,%edx
-	movl	%edx,-8(%rsp)
-	fldcw	-8(%rsp)		/* load modified control word */
+	movl	%edx,-32(%rsp)
+	fldcw	-32(%rsp)		/* load modified control word */
 
 	frndint				/* round */
 
-	fldcw	-4(%rsp)		/* restore original control word */
+	/* Preserve "invalid" exceptions from sNaN input.  */
+	fnstsw
+	andl	$0x1, %eax
+	orl	%eax, -24(%rsp)
+
+	fldenv	-28(%rsp)		/* restore original environment */
 
 	ret
 END (__ceill)