* sysdeps/alpha/div.S: Save and restore FPCR around fp operations. * sysdeps/alpha/divl.S, sysdeps/alpha/divq.S, sysdeps/alpha/divqu.S, sysdeps/alpha/ldiv.S, sysdeps/alpha/reml.S, sysdeps/alpha/remq.S, sysdeps/alpha/remqu.S: Likewise. * sysdeps/alpha/div_libc.h (FRAME): Increase to 64.

2004-09-05  Richard Henderson  <rth@redhat.com>

	* sysdeps/alpha/div.S: Save and restore FPCR around fp operations.
	* sysdeps/alpha/divl.S, sysdeps/alpha/divq.S, sysdeps/alpha/divqu.S,
	sysdeps/alpha/ldiv.S, sysdeps/alpha/reml.S, sysdeps/alpha/remq.S,
	sysdeps/alpha/remqu.S: Likewise.
	* sysdeps/alpha/div_libc.h (FRAME): Increase to 64.

9 files changed, 130 insertions, 29 deletions

diff --git a/sysdeps/alpha/div.S b/sysdeps/alpha/div.S
index e0eb7e9796..d1a724d375 100644
--- a/sysdeps/alpha/div.S
+++ b/sysdeps/alpha/div.S
@@ -48,6 +48,8 @@ div:
 #endif
 
 	beq	$18, $divbyzero
+	excb
+	mf_fpcr	$f10
 
 	_ITOFT2	$17, $f0, 0, $18, $f1, 8
 
@@ -55,7 +57,8 @@ div:
 	cvtqt	$f1, $f1
 	divt/c	$f0, $f1, $f0
 	cvttq/c	$f0, $f0
-
+	excb
+	mt_fpcr	$f10
 	_FTOIT	$f0, $0, 0
 
 	mull	$0, $18, $1
diff --git a/sysdeps/alpha/div_libc.h b/sysdeps/alpha/div_libc.h
index 2f06282102..62b4470355 100644
--- a/sysdeps/alpha/div_libc.h
+++ b/sysdeps/alpha/div_libc.h
@@ -71,7 +71,7 @@
 
 /* In order to make the below work, all top-level divide routines must
    use the same frame size.  */
-#define FRAME	48
+#define FRAME	64
 
 /* Code fragment to generate an integer divide-by-zero fault.  When
    building libc.so, we arrange for there to be one copy of this code
diff --git a/sysdeps/alpha/divl.S b/sysdeps/alpha/divl.S
index 90cd6862a7..408d66db00 100644
--- a/sysdeps/alpha/divl.S
+++ b/sysdeps/alpha/divl.S
@@ -22,7 +22,12 @@
    registers are t10 and t11, the result goes in t12.  Only t12 and AT may
    be clobbered.
 
-   The FPU can handle all input values except zero.  Whee!  */
+   The FPU can handle all input values except zero.  Whee!
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 #ifndef EXTEND
 #define EXTEND(S,D)	sextl S, D
@@ -41,25 +46,34 @@ __divl:
 	cfi_def_cfa_offset (FRAME)
 	CALL_MCOUNT
 	stt	$f0, 0(sp)
-	stt	$f1, 8(sp)
+	excb
 	beq	Y, DIVBYZERO
+
+	stt	$f1, 8(sp)
+	stt	$f2, 16(sp)
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f2, 16)
+	mf_fpcr	$f2
 
 	EXTEND	(X, RV)
 	EXTEND	(Y, AT)
-	_ITOFT2	RV, $f0, 16, AT, $f1, 24
+	_ITOFT2	RV, $f0, 24, AT, $f1, 32
 	cvtqt	$f0, $f0
 	cvtqt	$f1, $f1
 	divt/c	$f0, $f1, $f0
 	cvttq/c	$f0, $f0
-	_FTOIT	$f0, RV, 16
+	excb
+	mt_fpcr	$f2
+	_FTOIT	$f0, RV, 24
 
 	ldt	$f0, 0(sp)
 	ldt	$f1, 8(sp)
+	ldt	$f2, 16(sp)
 	lda	sp, FRAME(sp)
 	cfi_restore ($f0)
 	cfi_restore ($f1)
+	cfi_restore ($f2)
 	cfi_def_cfa_offset (0)
 	sextl	RV, RV
 	ret	$31, (RA), 1
diff --git a/sysdeps/alpha/divq.S b/sysdeps/alpha/divq.S
index cab6c34ad1..7f245ac056 100644
--- a/sysdeps/alpha/divq.S
+++ b/sysdeps/alpha/divq.S
@@ -33,7 +33,12 @@
    When the dividend is outside the range for which we can compute exact
    results, we use the fp quotent as an estimate from which we begin refining
    an exact integral value.  This reduces the number of iterations in the
-   shift-and-subtract loop significantly.  */
+   shift-and-subtract loop significantly.
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 	.text
 	.align	4
@@ -53,10 +58,15 @@ __divq:
 	   ready -- all the time in the world to figure out how we're
 	   going to use the results.  */
 	stt	$f0, 0(sp)
-	stt	$f1, 8(sp)
+	excb
 	beq	Y, DIVBYZERO
+
+	stt	$f1, 8(sp)
+	stt	$f3, 48(sp)
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f3, 48)
+	mf_fpcr	$f3
 
 	_ITOFT2	X, $f0, 16, Y, $f1, 24
 	cvtqt	$f0, $f0
@@ -73,12 +83,16 @@ __divq:
 	/* If we get here, we're expecting exact results from the division.
 	   Do nothing else besides convert and clean up.  */
 	cvttq/c	$f0, $f0
+	excb
+	mt_fpcr	$f3
 	_FTOIT	$f0, RV, 16
 
 	ldt	$f0, 0(sp)
+	ldt	$f3, 48(sp)
 	cfi_restore ($f1)
 	cfi_remember_state
 	cfi_restore ($f0)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	lda	sp, FRAME(sp)
 	ret	$31, (RA), 1
@@ -121,9 +135,9 @@ $fix_sign_in_ret2:
 	cfi_rel_offset (t3, 0)
 
 	mulq	Q, Y, QY
-	unop
+	excb
 	stq	t4, 8(sp)
-	unop
+	mt_fpcr	$f3
 	cfi_rel_offset (t4, 8)
 
 	subq	QY, X, R
@@ -147,6 +161,7 @@ $fix_sign_out_ret:
 	ldq	t3, 0(sp)
 	ldq	t4, 8(sp)
 	ldq	t5, 40(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_remember_state
 	cfi_restore (t0)
@@ -155,6 +170,7 @@ $fix_sign_out_ret:
 	cfi_restore (t3)
 	cfi_restore (t4)
 	cfi_restore (t5)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	ret	$31, (RA), 1
 
diff --git a/sysdeps/alpha/divqu.S b/sysdeps/alpha/divqu.S
index 63b575faa7..fc00fa133c 100644
--- a/sysdeps/alpha/divqu.S
+++ b/sysdeps/alpha/divqu.S
@@ -33,7 +33,12 @@
    When the dividend is outside the range for which we can compute exact
    results, we use the fp quotent as an estimate from which we begin refining
    an exact integral value.  This reduces the number of iterations in the
-   shift-and-subtract loop significantly.  */
+   shift-and-subtract loop significantly.
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 	.text
 	.align	4
@@ -53,10 +58,15 @@ __divqu:
 	   ready -- all the time in the world to figure out how we're
 	   going to use the results.  */
 	stt	$f0, 0(sp)
-	stt	$f1, 8(sp)
+	excb
 	beq	Y, DIVBYZERO
+
+	stt	$f1, 8(sp)
+	stt	$f3, 48(sp)
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f3, 48)
+	mf_fpcr	$f3
 
 	_ITOFT2	X, $f0, 16, Y, $f1, 24
 	cvtqt	$f0, $f0
@@ -65,10 +75,7 @@ __divqu:
 	divt/c	$f0, $f1, $f0
 
 	/* Check to see if Y was mis-converted as signed value.  */
-	.align	4
 	ldt	$f1, 8(sp)
-	unop
-	nop
 	blt	Y, $y_is_neg
 
 	/* Check to see if X fit in the double as an exact value.  */
@@ -78,11 +85,16 @@ __divqu:
 	/* If we get here, we're expecting exact results from the division.
 	   Do nothing else besides convert and clean up.  */
 	cvttq/c	$f0, $f0
+	excb
+	mt_fpcr	$f3
 	_FTOIT	$f0, RV, 16
+
 	ldt	$f0, 0(sp)
+	ldt	$f3, 48(sp)
 	cfi_remember_state
 	cfi_restore ($f0)
 	cfi_restore ($f1)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	lda	sp, FRAME(sp)
 	ret	$31, (RA), 1
@@ -140,9 +152,9 @@ $x_big:
 
 	.align	4
 	stq	t4, 8(sp)
-	unop
+	excb
 	ldt	$f0, 0(sp)
-	unop
+	mt_fpcr	$f3
 	cfi_rel_offset (t4, 8)
 	cfi_restore ($f0)
 
@@ -164,6 +176,7 @@ $q_low_ret:
 	ldq	t2, 32(sp)
 
 	ldq	t3, 40(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_remember_state
 	cfi_restore (t0)
@@ -171,6 +184,7 @@ $q_low_ret:
 	cfi_restore (t2)
 	cfi_restore (t3)
 	cfi_restore (t4)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	ret	$31, (RA), 1
 
@@ -227,9 +241,13 @@ $y_is_neg:
 	   from the divide will be completely wrong.  Fortunately, the
 	   quotient must be either 0 or 1, so just compute it directly.  */
 	cmpult	Y, X, RV
+	excb
+	mt_fpcr	$f3
 	ldt	$f0, 0(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_restore ($f0)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	ret	$31, (RA), 1
 
diff --git a/sysdeps/alpha/ldiv.S b/sysdeps/alpha/ldiv.S
index c90edfb784..3909672782 100644
--- a/sysdeps/alpha/ldiv.S
+++ b/sysdeps/alpha/ldiv.S
@@ -53,6 +53,8 @@ ldiv:
 #endif
 
 	beq	Y, $divbyzero
+	excb
+	mf_fpcr	$f10
 
 	_ITOFT2	X, $f0, 0, Y, $f1, 8
 
@@ -71,6 +73,8 @@ ldiv:
 	/* If we get here, we're expecting exact results from the division.
 	   Do nothing else besides convert and clean up.  */
 	cvttq/c	$f0, $f0
+	excb
+	mt_fpcr	$f10
 	_FTOIT	$f0, $0, 0
 
 $egress:
@@ -107,9 +111,10 @@ $fix_sign_in_ret1:
 	cvttq/c	$f0, $f0
 
 	_FTOIT	$f0, Q, 8
-	.align	3
 $fix_sign_in_ret2:
 	mulq	Q, Y, QY
+	excb
+	mt_fpcr	$f10
 
 	.align	4
 	subq	QY, X, R
diff --git a/sysdeps/alpha/reml.S b/sysdeps/alpha/reml.S
index 1bbb978f66..bfc3be5c3f 100644
--- a/sysdeps/alpha/reml.S
+++ b/sysdeps/alpha/reml.S
@@ -24,7 +24,12 @@
    be clobbered.
 
    The FPU can handle the division for all input values except zero.
-   All we have to do is compute the remainder via multiply-and-subtract.  */
+   All we have to do is compute the remainder via multiply-and-subtract.
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 #ifndef EXTEND
 #define EXTEND(S,D)	sextl S, D
@@ -43,26 +48,35 @@ __reml:
 	cfi_def_cfa_offset (FRAME)
 	CALL_MCOUNT
 	stt	$f0, 0(sp)
-	stt	$f1, 8(sp)
+	excb
 	beq	Y, DIVBYZERO
+
+	stt	$f1, 8(sp)
+	stt	$f2, 16(sp)
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f2, 16)
+	mf_fpcr	$f2
 
 	EXTEND	(X, RV)
 	EXTEND	(Y, AT)
-	_ITOFT2	RV, $f0, 16, AT, $f1, 24
+	_ITOFT2	RV, $f0, 24, AT, $f1, 32
 	cvtqt	$f0, $f0
 	cvtqt	$f1, $f1
 	divt/c	$f0, $f1, $f0
 	cvttq/c	$f0, $f0
-	_FTOIT	$f0, RV, 16
+	excb
+	mt_fpcr	$f2
+	_FTOIT	$f0, RV, 24
 
 	ldt	$f0, 0(sp)
 	mull	RV, Y, RV
 	ldt	$f1, 8(sp)
+	ldt	$f2, 16(sp)
 	lda	sp, FRAME(sp)
 	cfi_restore ($f0)
 	cfi_restore ($f1)
+	cfi_restore ($f2)
 	cfi_def_cfa_offset (0)
 	subl	X, RV, RV
 	ret	$31, (RA), 1
diff --git a/sysdeps/alpha/remq.S b/sysdeps/alpha/remq.S
index 40c68d7f10..645a834453 100644
--- a/sysdeps/alpha/remq.S
+++ b/sysdeps/alpha/remq.S
@@ -33,7 +33,12 @@
    When the dividend is outside the range for which we can compute exact
    results, we use the fp quotent as an estimate from which we begin refining
    an exact integral value.  This reduces the number of iterations in the
-   shift-and-subtract loop significantly.  */
+   shift-and-subtract loop significantly.
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 	.text
 	.align	4
@@ -53,10 +58,15 @@ __remq:
 	   ready -- all the time in the world to figure out how we're
 	   going to use the results.  */
 	stt	$f0, 0(sp)
-	stt	$f1, 8(sp)
+	excb
 	beq	Y, DIVBYZERO
+
+	stt	$f1, 8(sp)
+	stt	$f3, 48(sp)
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f3, 48)
+	mf_fpcr	$f3
 
 	_ITOFT2	X, $f0, 16, Y, $f1, 24
 	cvtqt	$f0, $f0
@@ -73,12 +83,16 @@ __remq:
 	/* If we get here, we're expecting exact results from the division.
 	   Do nothing else besides convert, compute remainder, clean up.  */
 	cvttq/c	$f0, $f0
+	excb
+	mt_fpcr	$f3
 	_FTOIT	$f0, AT, 16
 	mulq	AT, Y, AT
 	ldt	$f0, 0(sp)
+	ldt	$f3, 48(sp)
 	cfi_restore ($f1)
 	cfi_remember_state
 	cfi_restore ($f0)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	lda	sp, FRAME(sp)
 	subq	X, AT, RV
@@ -122,9 +136,9 @@ $fix_sign_in_ret2:
 	cfi_rel_offset (t3, 0)
 
 	mulq	Q, Y, QY
-	unop
+	excb
 	stq	t4, 8(sp)
-	unop
+	mt_fpcr	$f3
 	cfi_rel_offset (t4, 8)
 
 	subq	QY, X, R
@@ -148,6 +162,7 @@ $fix_sign_out_ret:
 	ldq	t3, 0(sp)
 	ldq	t4, 8(sp)
 	ldq	t5, 40(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_remember_state
 	cfi_restore (t0)
@@ -156,6 +171,7 @@ $fix_sign_out_ret:
 	cfi_restore (t3)
 	cfi_restore (t4)
 	cfi_restore (t5)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	ret	$31, (RA), 1
 
diff --git a/sysdeps/alpha/remqu.S b/sysdeps/alpha/remqu.S
index f8deebbbc1..bfa78dff57 100644
--- a/sysdeps/alpha/remqu.S
+++ b/sysdeps/alpha/remqu.S
@@ -33,7 +33,12 @@
    When the dividend is outside the range for which we can compute exact
    results, we use the fp quotent as an estimate from which we begin refining
    an exact integral value.  This reduces the number of iterations in the
-   shift-and-subtract loop significantly.  */
+   shift-and-subtract loop significantly.
+
+   The FPCR save/restore is due to the fact that the EV6 _will_ set FPCR_INE
+   for cvttq/c even without /sui being set.  It will not, however, properly
+   raise the exception, so we don't have to worry about FPCR_INED being clear
+   and so dying by SIGFPE.  */
 
 	.text
 	.align	4
@@ -57,11 +62,15 @@ __remqu:
 	and	Y, AT, AT
 
 	stt	$f1, 8(sp)
+	excb
+	stt	$f3, 48(sp)
 	beq	AT, $powerof2
 	cfi_rel_offset ($f0, 0)
 	cfi_rel_offset ($f1, 8)
+	cfi_rel_offset ($f3, 48)
 
 	_ITOFT2	X, $f0, 16, Y, $f1, 24
+	mf_fpcr	$f3
 	cvtqt	$f0, $f0
 	cvtqt	$f1, $f1
 
@@ -79,14 +88,18 @@ __remqu:
 	/* If we get here, we're expecting exact results from the division.
 	   Do nothing else besides convert, compute remainder, clean up.  */
 	cvttq/c	$f0, $f0
+	excb
+	mt_fpcr	$f3
 	_FTOIT	$f0, AT, 16
 
 	mulq	AT, Y, AT
 	ldt	$f0, 0(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_remember_state
 	cfi_restore ($f0)
 	cfi_restore ($f1)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 
 	.align	4
@@ -144,9 +157,9 @@ $x_big:
 
 	.align	4
 	stq	t4, 8(sp)
-	unop
+	excb
 	ldt	$f0, 0(sp)
-	unop
+	mt_fpcr	$f3
 	cfi_rel_offset (t4, 8)
 	cfi_restore ($f0)
 
@@ -168,6 +181,7 @@ $q_low_ret:
 	ldq	t2, 32(sp)
 
 	ldq	t3, 40(sp)
+	ldt	$f3, 48(sp)
 	lda	sp, FRAME(sp)
 	cfi_remember_state
 	cfi_restore (t0)
@@ -175,6 +189,7 @@ $q_low_ret:
 	cfi_restore (t2)
 	cfi_restore (t3)
 	cfi_restore (t4)
+	cfi_restore ($f3)
 	cfi_def_cfa_offset (0)
 	ret	$31, (RA), 1