[PATCH] ARM: 2837/2: Re: ARM: Make NWFPE preempt safe

Patch from Richard Purdie

NWFPE used global variables which meant it wasn't safe for use with
preemptive kernels. This patch removes them and communicates the
information between functions in a preempt safe manner. Generation
of some exceptions was broken and this has also been corrected.
Tests with glibc's maths test suite show no change in the results
before/after this patch.

Signed-off-by: Richard Purdie <rpurdie@rpsys.net>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

11 files changed, 304 insertions, 304 deletions

diff --git a/arch/arm/nwfpe/double_cpdo.c b/arch/arm/nwfpe/double_cpdo.c
index 7ffd8cb9bc96..c51d1386a97c 100644
--- a/arch/arm/nwfpe/double_cpdo.c
+++ b/arch/arm/nwfpe/double_cpdo.c
@@ -40,17 +40,17 @@ float64 float64_arccos(float64 rFm);
 float64 float64_pow(float64 rFn, float64 rFm);
 float64 float64_pol(float64 rFn, float64 rFm);
 
-static float64 float64_rsf(float64 rFn, float64 rFm)
+static float64 float64_rsf(struct roundingData *roundData, float64 rFn, float64 rFm)
 {
-	return float64_sub(rFm, rFn);
+	return float64_sub(roundData, rFm, rFn);
 }
 
-static float64 float64_rdv(float64 rFn, float64 rFm)
+static float64 float64_rdv(struct roundingData *roundData, float64 rFn, float64 rFm)
 {
-	return float64_div(rFm, rFn);
+	return float64_div(roundData, rFm, rFn);
 }
 
-static float64 (*const dyadic_double[16])(float64 rFn, float64 rFm) = {
+static float64 (*const dyadic_double[16])(struct roundingData*, float64 rFn, float64 rFm) = {
 	[ADF_CODE >> 20] = float64_add,
 	[MUF_CODE >> 20] = float64_mul,
 	[SUF_CODE >> 20] = float64_sub,
@@ -65,12 +65,12 @@ static float64 (*const dyadic_double[16])(float64 rFn, float64 rFm) = {
 	[FRD_CODE >> 20] = float64_rdv,
 };
 
-static float64 float64_mvf(float64 rFm)
+static float64 float64_mvf(struct roundingData *roundData,float64 rFm)
 {
 	return rFm;
 }
 
-static float64 float64_mnf(float64 rFm)
+static float64 float64_mnf(struct roundingData *roundData,float64 rFm)
 {
 	union float64_components u;
 
@@ -84,7 +84,7 @@ static float64 float64_mnf(float64 rFm)
 	return u.f64;
 }
 
-static float64 float64_abs(float64 rFm)
+static float64 float64_abs(struct roundingData *roundData,float64 rFm)
 {
 	union float64_components u;
 
@@ -98,7 +98,7 @@ static float64 float64_abs(float64 rFm)
 	return u.f64;
 }
 
-static float64 (*const monadic_double[16])(float64 rFm) = {
+static float64 (*const monadic_double[16])(struct roundingData *, float64 rFm) = {
 	[MVF_CODE >> 20] = float64_mvf,
 	[MNF_CODE >> 20] = float64_mnf,
 	[ABS_CODE >> 20] = float64_abs,
@@ -108,7 +108,7 @@ static float64 (*const monadic_double[16])(float64 rFm) = {
 	[NRM_CODE >> 20] = float64_mvf,
 };
 
-unsigned int DoubleCPDO(const unsigned int opcode, FPREG * rFd)
+unsigned int DoubleCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	float64 rFm;
@@ -151,13 +151,13 @@ unsigned int DoubleCPDO(const unsigned int opcode, FPREG * rFd)
 		}
 
 		if (dyadic_double[opc_mask_shift]) {
-			rFd->fDouble = dyadic_double[opc_mask_shift](rFn, rFm);
+			rFd->fDouble = dyadic_double[opc_mask_shift](roundData, rFn, rFm);
 		} else {
 			return 0;
 		}
 	} else {
 		if (monadic_double[opc_mask_shift]) {
-			rFd->fDouble = monadic_double[opc_mask_shift](rFm);
+			rFd->fDouble = monadic_double[opc_mask_shift](roundData, rFm);
 		} else {
 			return 0;
 		}
diff --git a/arch/arm/nwfpe/extended_cpdo.c b/arch/arm/nwfpe/extended_cpdo.c
index c39f68a3449e..65a279ba927f 100644
--- a/arch/arm/nwfpe/extended_cpdo.c
+++ b/arch/arm/nwfpe/extended_cpdo.c
@@ -35,17 +35,17 @@ floatx80 floatx80_arccos(floatx80 rFm);
 floatx80 floatx80_pow(floatx80 rFn, floatx80 rFm);
 floatx80 floatx80_pol(floatx80 rFn, floatx80 rFm);
 
-static floatx80 floatx80_rsf(floatx80 rFn, floatx80 rFm)
+static floatx80 floatx80_rsf(struct roundingData *roundData, floatx80 rFn, floatx80 rFm)
 {
-	return floatx80_sub(rFm, rFn);
+	return floatx80_sub(roundData, rFm, rFn);
 }
 
-static floatx80 floatx80_rdv(floatx80 rFn, floatx80 rFm)
+static floatx80 floatx80_rdv(struct roundingData *roundData, floatx80 rFn, floatx80 rFm)
 {
-	return floatx80_div(rFm, rFn);
+	return floatx80_div(roundData, rFm, rFn);
 }
 
-static floatx80 (*const dyadic_extended[16])(floatx80 rFn, floatx80 rFm) = {
+static floatx80 (*const dyadic_extended[16])(struct roundingData*, floatx80 rFn, floatx80 rFm) = {
 	[ADF_CODE >> 20] = floatx80_add,
 	[MUF_CODE >> 20] = floatx80_mul,
 	[SUF_CODE >> 20] = floatx80_sub,
@@ -60,24 +60,24 @@ static floatx80 (*const dyadic_extended[16])(floatx80 rFn, floatx80 rFm) = {
 	[FRD_CODE >> 20] = floatx80_rdv,
 };
 
-static floatx80 floatx80_mvf(floatx80 rFm)
+static floatx80 floatx80_mvf(struct roundingData *roundData, floatx80 rFm)
 {
 	return rFm;
 }
 
-static floatx80 floatx80_mnf(floatx80 rFm)
+static floatx80 floatx80_mnf(struct roundingData *roundData, floatx80 rFm)
 {
 	rFm.high ^= 0x8000;
 	return rFm;
 }
 
-static floatx80 floatx80_abs(floatx80 rFm)
+static floatx80 floatx80_abs(struct roundingData *roundData, floatx80 rFm)
 {
 	rFm.high &= 0x7fff;
 	return rFm;
 }
 
-static floatx80 (*const monadic_extended[16])(floatx80 rFm) = {
+static floatx80 (*const monadic_extended[16])(struct roundingData*, floatx80 rFm) = {
 	[MVF_CODE >> 20] = floatx80_mvf,
 	[MNF_CODE >> 20] = floatx80_mnf,
 	[ABS_CODE >> 20] = floatx80_abs,
@@ -87,7 +87,7 @@ static floatx80 (*const monadic_extended[16])(floatx80 rFm) = {
 	[NRM_CODE >> 20] = floatx80_mvf,
 };
 
-unsigned int ExtendedCPDO(const unsigned int opcode, FPREG * rFd)
+unsigned int ExtendedCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	floatx80 rFm;
@@ -138,13 +138,13 @@ unsigned int ExtendedCPDO(const unsigned int opcode, FPREG * rFd)
 		}
 
 		if (dyadic_extended[opc_mask_shift]) {
-			rFd->fExtended = dyadic_extended[opc_mask_shift](rFn, rFm);
+			rFd->fExtended = dyadic_extended[opc_mask_shift](roundData, rFn, rFm);
 		} else {
 			return 0;
 		}
 	} else {
 		if (monadic_extended[opc_mask_shift]) {
-			rFd->fExtended = monadic_extended[opc_mask_shift](rFm);
+			rFd->fExtended = monadic_extended[opc_mask_shift](roundData, rFm);
 		} else {
 			return 0;
 		}
diff --git a/arch/arm/nwfpe/fpa11.c b/arch/arm/nwfpe/fpa11.c
index bf61696865ec..7690f731ee87 100644
--- a/arch/arm/nwfpe/fpa11.c
+++ b/arch/arm/nwfpe/fpa11.c
@@ -51,48 +51,42 @@ static void resetFPA11(void)
 	fpa11->fpsr = FP_EMULATOR | BIT_AC;
 }
 
-void SetRoundingMode(const unsigned int opcode)
+int8 SetRoundingMode(const unsigned int opcode)
 {
 	switch (opcode & MASK_ROUNDING_MODE) {
 	default:
 	case ROUND_TO_NEAREST:
-		float_rounding_mode = float_round_nearest_even;
-		break;
+		return float_round_nearest_even;
 
 	case ROUND_TO_PLUS_INFINITY:
-		float_rounding_mode = float_round_up;
-		break;
+		return float_round_up;
 
 	case ROUND_TO_MINUS_INFINITY:
-		float_rounding_mode = float_round_down;
-		break;
+		return float_round_down;
 
 	case ROUND_TO_ZERO:
-		float_rounding_mode = float_round_to_zero;
-		break;
+		return float_round_to_zero;
 	}
 }
 
-void SetRoundingPrecision(const unsigned int opcode)
+int8 SetRoundingPrecision(const unsigned int opcode)
 {
 #ifdef CONFIG_FPE_NWFPE_XP
 	switch (opcode & MASK_ROUNDING_PRECISION) {
 	case ROUND_SINGLE:
-		floatx80_rounding_precision = 32;
-		break;
+		return 32;
 
 	case ROUND_DOUBLE:
-		floatx80_rounding_precision = 64;
-		break;
+		return 64;
 
 	case ROUND_EXTENDED:
-		floatx80_rounding_precision = 80;
-		break;
+		return 80;
 
 	default:
-		floatx80_rounding_precision = 80;
+		return 80;
 	}
 #endif
+	return 80;
 }
 
 void nwfpe_init_fpa(union fp_state *fp)
@@ -103,8 +97,6 @@ void nwfpe_init_fpa(union fp_state *fp)
 #endif
  	memset(fpa11, 0, sizeof(FPA11));
 	resetFPA11();
-	SetRoundingMode(ROUND_TO_NEAREST);
-	SetRoundingPrecision(ROUND_EXTENDED);
 	fpa11->initflag = 1;
 }
 
diff --git a/arch/arm/nwfpe/fpa11.h b/arch/arm/nwfpe/fpa11.h
index e4a61aea534b..93523ae4b7a1 100644
--- a/arch/arm/nwfpe/fpa11.h
+++ b/arch/arm/nwfpe/fpa11.h
@@ -37,6 +37,13 @@
 /* includes */
 #include "fpsr.h"		/* FP control and status register definitions */
 #include "milieu.h"
+
+struct roundingData {
+    int8 mode;
+    int8 precision;
+    signed char exception;
+};
+
 #include "softfloat.h"
 
 #define		typeNone		0x00
@@ -84,8 +91,8 @@ typedef struct tagFPA11 {
 				   initialised. */
 } FPA11;
 
-extern void SetRoundingMode(const unsigned int);
-extern void SetRoundingPrecision(const unsigned int);
+extern int8 SetRoundingMode(const unsigned int);
+extern int8 SetRoundingPrecision(const unsigned int);
 extern void nwfpe_init_fpa(union fp_state *fp);
 
 #endif
diff --git a/arch/arm/nwfpe/fpa11_cpdo.c b/arch/arm/nwfpe/fpa11_cpdo.c
index 1bea67437b6f..4a31dfd94068 100644
--- a/arch/arm/nwfpe/fpa11_cpdo.c
+++ b/arch/arm/nwfpe/fpa11_cpdo.c
@@ -24,15 +24,16 @@
 #include "fpa11.h"
 #include "fpopcode.h"
 
-unsigned int SingleCPDO(const unsigned int opcode, FPREG * rFd);
-unsigned int DoubleCPDO(const unsigned int opcode, FPREG * rFd);
-unsigned int ExtendedCPDO(const unsigned int opcode, FPREG * rFd);
+unsigned int SingleCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd);
+unsigned int DoubleCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd);
+unsigned int ExtendedCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd);
 
 unsigned int EmulateCPDO(const unsigned int opcode)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	FPREG *rFd;
 	unsigned int nType, nDest, nRc;
+	struct roundingData roundData;
 
 	/* Get the destination size.  If not valid let Linux perform
 	   an invalid instruction trap. */
@@ -40,7 +41,9 @@ unsigned int EmulateCPDO(const unsigned int opcode)
 	if (typeNone == nDest)
 		return 0;
 
-	SetRoundingMode(opcode);
+	roundData.mode = SetRoundingMode(opcode);
+	roundData.precision = SetRoundingPrecision(opcode);
+	roundData.exception = 0;
 
 	/* Compare the size of the operands in Fn and Fm.
 	   Choose the largest size and perform operations in that size,
@@ -63,14 +66,14 @@ unsigned int EmulateCPDO(const unsigned int opcode)
 
 	switch (nType) {
 	case typeSingle:
-		nRc = SingleCPDO(opcode, rFd);
+		nRc = SingleCPDO(&roundData, opcode, rFd);
 		break;
 	case typeDouble:
-		nRc = DoubleCPDO(opcode, rFd);
+		nRc = DoubleCPDO(&roundData, opcode, rFd);
 		break;
 #ifdef CONFIG_FPE_NWFPE_XP
 	case typeExtended:
-		nRc = ExtendedCPDO(opcode, rFd);
+		nRc = ExtendedCPDO(&roundData, opcode, rFd);
 		break;
 #endif
 	default:
@@ -93,9 +96,9 @@ unsigned int EmulateCPDO(const unsigned int opcode)
 			case typeSingle:
 				{
 					if (typeDouble == nType)
-						rFd->fSingle = float64_to_float32(rFd->fDouble);
+						rFd->fSingle = float64_to_float32(&roundData, rFd->fDouble);
 					else
-						rFd->fSingle = floatx80_to_float32(rFd->fExtended);
+						rFd->fSingle = floatx80_to_float32(&roundData, rFd->fExtended);
 				}
 				break;
 
@@ -104,7 +107,7 @@ unsigned int EmulateCPDO(const unsigned int opcode)
 					if (typeSingle == nType)
 						rFd->fDouble = float32_to_float64(rFd->fSingle);
 					else
-						rFd->fDouble = floatx80_to_float64(rFd->fExtended);
+						rFd->fDouble = floatx80_to_float64(&roundData, rFd->fExtended);
 				}
 				break;
 
@@ -121,12 +124,15 @@ unsigned int EmulateCPDO(const unsigned int opcode)
 #else
 		if (nDest != nType) {
 			if (nDest == typeSingle)
-				rFd->fSingle = float64_to_float32(rFd->fDouble);
+				rFd->fSingle = float64_to_float32(&roundData, rFd->fDouble);
 			else
 				rFd->fDouble = float32_to_float64(rFd->fSingle);
 		}
 #endif
 	}
 
+	if (roundData.exception)
+		float_raise(roundData.exception);
+
 	return nRc;
 }
diff --git a/arch/arm/nwfpe/fpa11_cpdt.c b/arch/arm/nwfpe/fpa11_cpdt.c
index 95fb63fa9d18..b0db5cbcc3b1 100644
--- a/arch/arm/nwfpe/fpa11_cpdt.c
+++ b/arch/arm/nwfpe/fpa11_cpdt.c
@@ -96,7 +96,7 @@ static inline void loadMultiple(const unsigned int Fn, const unsigned int __user
 	}
 }
 
-static inline void storeSingle(const unsigned int Fn, unsigned int __user *pMem)
+static inline void storeSingle(struct roundingData *roundData, const unsigned int Fn, unsigned int __user *pMem)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	union {
@@ -106,12 +106,12 @@ static inline void storeSingle(const unsigned int Fn, unsigned int __user *pMem)
 
 	switch (fpa11->fType[Fn]) {
 	case typeDouble:
-		val.f = float64_to_float32(fpa11->fpreg[Fn].fDouble);
+		val.f = float64_to_float32(roundData, fpa11->fpreg[Fn].fDouble);
 		break;
 
 #ifdef CONFIG_FPE_NWFPE_XP
 	case typeExtended:
-		val.f = floatx80_to_float32(fpa11->fpreg[Fn].fExtended);
+		val.f = floatx80_to_float32(roundData, fpa11->fpreg[Fn].fExtended);
 		break;
 #endif
 
@@ -122,7 +122,7 @@ static inline void storeSingle(const unsigned int Fn, unsigned int __user *pMem)
 	put_user(val.i[0], pMem);
 }
 
-static inline void storeDouble(const unsigned int Fn, unsigned int __user *pMem)
+static inline void storeDouble(struct roundingData *roundData, const unsigned int Fn, unsigned int __user *pMem)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	union {
@@ -137,7 +137,7 @@ static inline void storeDouble(const unsigned int Fn, unsigned int __user *pMem)
 
 #ifdef CONFIG_FPE_NWFPE_XP
 	case typeExtended:
-		val.f = floatx80_to_float64(fpa11->fpreg[Fn].fExtended);
+		val.f = floatx80_to_float64(roundData, fpa11->fpreg[Fn].fExtended);
 		break;
 #endif
 
@@ -259,8 +259,11 @@ unsigned int PerformSTF(const unsigned int opcode)
 {
 	unsigned int __user *pBase, *pAddress, *pFinal;
 	unsigned int nRc = 1, write_back = WRITE_BACK(opcode);
+	struct roundingData roundData;
 
-	SetRoundingMode(ROUND_TO_NEAREST);
+	roundData.mode = SetRoundingMode(opcode);
+	roundData.precision = SetRoundingPrecision(opcode);
+	roundData.exception = 0;
 
 	pBase = (unsigned int __user *) readRegister(getRn(opcode));
 	if (REG_PC == getRn(opcode)) {
@@ -281,10 +284,10 @@ unsigned int PerformSTF(const unsigned int opcode)
 
 	switch (opcode & MASK_TRANSFER_LENGTH) {
 	case TRANSFER_SINGLE:
-		storeSingle(getFd(opcode), pAddress);
+		storeSingle(&roundData, getFd(opcode), pAddress);
 		break;
 	case TRANSFER_DOUBLE:
-		storeDouble(getFd(opcode), pAddress);
+		storeDouble(&roundData, getFd(opcode), pAddress);
 		break;
 #ifdef CONFIG_FPE_NWFPE_XP
 	case TRANSFER_EXTENDED:
@@ -295,6 +298,9 @@ unsigned int PerformSTF(const unsigned int opcode)
 		nRc = 0;
 	}
 
+	if (roundData.exception)
+		float_raise(roundData.exception);
+
 	if (write_back)
 		writeRegister(getRn(opcode), (unsigned long) pFinal);
 	return nRc;
diff --git a/arch/arm/nwfpe/fpa11_cprt.c b/arch/arm/nwfpe/fpa11_cprt.c
index db01fbc97216..adf8d3000540 100644
--- a/arch/arm/nwfpe/fpa11_cprt.c
+++ b/arch/arm/nwfpe/fpa11_cprt.c
@@ -33,8 +33,6 @@ extern flag floatx80_is_nan(floatx80);
 extern flag float64_is_nan(float64);
 extern flag float32_is_nan(float32);
 
-void SetRoundingMode(const unsigned int opcode);
-
 unsigned int PerformFLT(const unsigned int opcode);
 unsigned int PerformFIX(const unsigned int opcode);
 
@@ -77,14 +75,17 @@ unsigned int EmulateCPRT(const unsigned int opcode)
 unsigned int PerformFLT(const unsigned int opcode)
 {
 	FPA11 *fpa11 = GET_FPA11();
-	SetRoundingMode(opcode);
-	SetRoundingPrecision(opcode);
+	struct roundingData roundData;
+
+	roundData.mode = SetRoundingMode(opcode);
+	roundData.precision = SetRoundingPrecision(opcode);
+	roundData.exception = 0;
 
 	switch (opcode & MASK_ROUNDING_PRECISION) {
 	case ROUND_SINGLE:
 		{
 			fpa11->fType[getFn(opcode)] = typeSingle;
-			fpa11->fpreg[getFn(opcode)].fSingle = int32_to_float32(readRegister(getRd(opcode)));
+			fpa11->fpreg[getFn(opcode)].fSingle = int32_to_float32(&roundData, readRegister(getRd(opcode)));
 		}
 		break;
 
@@ -108,6 +109,9 @@ unsigned int PerformFLT(const unsigned int opcode)
 		return 0;
 	}
 
+	if (roundData.exception)
+		float_raise(roundData.exception);
+
 	return 1;
 }
 
@@ -115,26 +119,29 @@ unsigned int PerformFIX(const unsigned int opcode)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	unsigned int Fn = getFm(opcode);
+	struct roundingData roundData;
 
-	SetRoundingMode(opcode);
+	roundData.mode = SetRoundingMode(opcode);
+	roundData.precision = SetRoundingPrecision(opcode);
+	roundData.exception = 0;
 
 	switch (fpa11->fType[Fn]) {
 	case typeSingle:
 		{
-			writeRegister(getRd(opcode), float32_to_int32(fpa11->fpreg[Fn].fSingle));
+			writeRegister(getRd(opcode), float32_to_int32(&roundData, fpa11->fpreg[Fn].fSingle));
 		}
 		break;
 
 	case typeDouble:
 		{
-			writeRegister(getRd(opcode), float64_to_int32(fpa11->fpreg[Fn].fDouble));
+			writeRegister(getRd(opcode), float64_to_int32(&roundData, fpa11->fpreg[Fn].fDouble));
 		}
 		break;
 
 #ifdef CONFIG_FPE_NWFPE_XP
 	case typeExtended:
 		{
-			writeRegister(getRd(opcode), floatx80_to_int32(fpa11->fpreg[Fn].fExtended));
+			writeRegister(getRd(opcode), floatx80_to_int32(&roundData, fpa11->fpreg[Fn].fExtended));
 		}
 		break;
 #endif
@@ -143,6 +150,9 @@ unsigned int PerformFIX(const unsigned int opcode)
 		return 0;
 	}
 
+	if (roundData.exception)
+		float_raise(roundData.exception);
+
 	return 1;
 }
 
diff --git a/arch/arm/nwfpe/fpmodule.c b/arch/arm/nwfpe/fpmodule.c
index 12885f31d347..2dfe1ac42ee8 100644
--- a/arch/arm/nwfpe/fpmodule.c
+++ b/arch/arm/nwfpe/fpmodule.c
@@ -116,8 +116,6 @@ fpmodule.c to integrate with the NetBSD kernel (I hope!).
 code to access data in user space in some other source files at the 
 moment (grep for get_user / put_user calls).  --philb]
 
-float_exception_flags is a global variable in SoftFloat.
-
 This function is called by the SoftFloat routines to raise a floating
 point exception.  We check the trap enable byte in the FPSR, and raise
 a SIGFPE exception if necessary.  If not the relevant bits in the 
@@ -129,15 +127,14 @@ void float_raise(signed char flags)
 	register unsigned int fpsr, cumulativeTraps;
 
 #ifdef CONFIG_DEBUG_USER
-	printk(KERN_DEBUG
-	       "NWFPE: %s[%d] takes exception %08x at %p from %08lx\n",
-	       current->comm, current->pid, flags,
-	       __builtin_return_address(0), GET_USERREG()->ARM_pc);
+ 	/* Ignore inexact errors as there are far too many of them to log */
+ 	if (flags & ~BIT_IXC)
+ 		printk(KERN_DEBUG
+		       "NWFPE: %s[%d] takes exception %08x at %p from %08lx\n",
+		       current->comm, current->pid, flags,
+		       __builtin_return_address(0), GET_USERREG()->ARM_pc);
 #endif
 
-	/* Keep SoftFloat exception flags up to date.  */
-	float_exception_flags |= flags;
-
 	/* Read fpsr and initialize the cumulativeTraps.  */
 	fpsr = readFPSR();
 	cumulativeTraps = 0;
diff --git a/arch/arm/nwfpe/single_cpdo.c b/arch/arm/nwfpe/single_cpdo.c
index 705808e88d9d..c66981d682cf 100644
--- a/arch/arm/nwfpe/single_cpdo.c
+++ b/arch/arm/nwfpe/single_cpdo.c
@@ -36,17 +36,17 @@ float32 float32_arccos(float32 rFm);
 float32 float32_pow(float32 rFn, float32 rFm);
 float32 float32_pol(float32 rFn, float32 rFm);
 
-static float32 float32_rsf(float32 rFn, float32 rFm)
+static float32 float32_rsf(struct roundingData *roundData, float32 rFn, float32 rFm)
 {
-	return float32_sub(rFm, rFn);
+	return float32_sub(roundData, rFm, rFn);
 }
 
-static float32 float32_rdv(float32 rFn, float32 rFm)
+static float32 float32_rdv(struct roundingData *roundData, float32 rFn, float32 rFm)
 {
-	return float32_div(rFm, rFn);
+	return float32_div(roundData, rFm, rFn);
 }
 
-static float32 (*const dyadic_single[16])(float32 rFn, float32 rFm) = {
+static float32 (*const dyadic_single[16])(struct roundingData *, float32 rFn, float32 rFm) = {
 	[ADF_CODE >> 20] = float32_add,
 	[MUF_CODE >> 20] = float32_mul,
 	[SUF_CODE >> 20] = float32_sub,
@@ -60,22 +60,22 @@ static float32 (*const dyadic_single[16])(float32 rFn, float32 rFm) = {
 	[FRD_CODE >> 20] = float32_rdv,
 };
 
-static float32 float32_mvf(float32 rFm)
+static float32 float32_mvf(struct roundingData *roundData, float32 rFm)
 {
 	return rFm;
 }
 
-static float32 float32_mnf(float32 rFm)
+static float32 float32_mnf(struct roundingData *roundData, float32 rFm)
 {
 	return rFm ^ 0x80000000;
 }
 
-static float32 float32_abs(float32 rFm)
+static float32 float32_abs(struct roundingData *roundData, float32 rFm)
 {
 	return rFm & 0x7fffffff;
 }
 
-static float32 (*const monadic_single[16])(float32 rFm) = {
+static float32 (*const monadic_single[16])(struct roundingData*, float32 rFm) = {
 	[MVF_CODE >> 20] = float32_mvf,
 	[MNF_CODE >> 20] = float32_mnf,
 	[ABS_CODE >> 20] = float32_abs,
@@ -85,7 +85,7 @@ static float32 (*const monadic_single[16])(float32 rFm) = {
 	[NRM_CODE >> 20] = float32_mvf,
 };
 
-unsigned int SingleCPDO(const unsigned int opcode, FPREG * rFd)
+unsigned int SingleCPDO(struct roundingData *roundData, const unsigned int opcode, FPREG * rFd)
 {
 	FPA11 *fpa11 = GET_FPA11();
 	float32 rFm;
@@ -108,13 +108,13 @@ unsigned int SingleCPDO(const unsigned int opcode, FPREG * rFd)
 		if (fpa11->fType[Fn] == typeSingle &&
 		    dyadic_single[opc_mask_shift]) {
 			rFn = fpa11->fpreg[Fn].fSingle;
-			rFd->fSingle = dyadic_single[opc_mask_shift](rFn, rFm);
+			rFd->fSingle = dyadic_single[opc_mask_shift](roundData, rFn, rFm);
 		} else {
 			return 0;
 		}
 	} else {
 		if (monadic_single[opc_mask_shift]) {
-			rFd->fSingle = monadic_single[opc_mask_shift](rFm);
+			rFd->fSingle = monadic_single[opc_mask_shift](roundData, rFm);
 		} else {
 			return 0;
 		}
diff --git a/arch/arm/nwfpe/softfloat.c b/arch/arm/nwfpe/softfloat.c
index e038dd3be9b3..8b75a6e7cb3a 100644
--- a/arch/arm/nwfpe/softfloat.c
+++ b/arch/arm/nwfpe/softfloat.c
@@ -36,16 +36,6 @@ this code that are retained.
 
 /*
 -------------------------------------------------------------------------------
-Floating-point rounding mode, extended double-precision rounding precision,
-and exception flags.
--------------------------------------------------------------------------------
-*/
-int8 float_rounding_mode = float_round_nearest_even;
-int8 floatx80_rounding_precision = 80;
-int8 float_exception_flags;
-
-/*
--------------------------------------------------------------------------------
 Primitive arithmetic functions, including multi-word arithmetic, and
 division and square root approximations.  (Can be specialized to target if
 desired.)
@@ -77,14 +67,14 @@ input is too large, however, the invalid exception is raised and the largest
 positive or negative integer is returned.
 -------------------------------------------------------------------------------
 */
-static int32 roundAndPackInt32( flag zSign, bits64 absZ )
+static int32 roundAndPackInt32( struct roundingData *roundData, flag zSign, bits64 absZ )
 {
     int8 roundingMode;
     flag roundNearestEven;
     int8 roundIncrement, roundBits;
     int32 z;
 
-    roundingMode = float_rounding_mode;
+    roundingMode = roundData->mode;
     roundNearestEven = ( roundingMode == float_round_nearest_even );
     roundIncrement = 0x40;
     if ( ! roundNearestEven ) {
@@ -107,10 +97,10 @@ static int32 roundAndPackInt32( flag zSign, bits64 absZ )
     z = absZ;
     if ( zSign ) z = - z;
     if ( ( absZ>>32 ) || ( z && ( ( z < 0 ) ^ zSign ) ) ) {
-        float_exception_flags |= float_flag_invalid;
+        roundData->exception |= float_flag_invalid;
         return zSign ? 0x80000000 : 0x7FFFFFFF;
     }
-    if ( roundBits ) float_exception_flags |= float_flag_inexact;
+    if ( roundBits ) roundData->exception |= float_flag_inexact;
     return z;
 
 }
@@ -224,14 +214,14 @@ The handling of underflow and overflow follows the IEC/IEEE Standard for
 Binary Floating-point Arithmetic.
 -------------------------------------------------------------------------------
 */
-static float32 roundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
+static float32 roundAndPackFloat32( struct roundingData *roundData, flag zSign, int16 zExp, bits32 zSig )
 {
     int8 roundingMode;
     flag roundNearestEven;
     int8 roundIncrement, roundBits;
     flag isTiny;
 
-    roundingMode = float_rounding_mode;
+    roundingMode = roundData->mode;
     roundNearestEven = ( roundingMode == float_round_nearest_even );
     roundIncrement = 0x40;
     if ( ! roundNearestEven ) {
@@ -254,7 +244,7 @@ static float32 roundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
              || (    ( zExp == 0xFD )
                   && ( (sbits32) ( zSig + roundIncrement ) < 0 ) )
            ) {
-            float_raise( float_flag_overflow | float_flag_inexact );
+            roundData->exception |= float_flag_overflow | float_flag_inexact;
             return packFloat32( zSign, 0xFF, 0 ) - ( roundIncrement == 0 );
         }
         if ( zExp < 0 ) {
@@ -265,10 +255,10 @@ static float32 roundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
             shift32RightJamming( zSig, - zExp, &zSig );
             zExp = 0;
             roundBits = zSig & 0x7F;
-            if ( isTiny && roundBits ) float_raise( float_flag_underflow );
+            if ( isTiny && roundBits ) roundData->exception |= float_flag_underflow;
         }
     }
-    if ( roundBits ) float_exception_flags |= float_flag_inexact;
+    if ( roundBits ) roundData->exception |= float_flag_inexact;
     zSig = ( zSig + roundIncrement )>>7;
     zSig &= ~ ( ( ( roundBits ^ 0x40 ) == 0 ) & roundNearestEven );
     if ( zSig == 0 ) zExp = 0;
@@ -287,12 +277,12 @@ point exponent.
 -------------------------------------------------------------------------------
 */
 static float32
- normalizeRoundAndPackFloat32( flag zSign, int16 zExp, bits32 zSig )
+ normalizeRoundAndPackFloat32( struct roundingData *roundData, flag zSign, int16 zExp, bits32 zSig )
 {
     int8 shiftCount;
 
     shiftCount = countLeadingZeros32( zSig ) - 1;
-    return roundAndPackFloat32( zSign, zExp - shiftCount, zSig<<shiftCount );
+    return roundAndPackFloat32( roundData, zSign, zExp - shiftCount, zSig<<shiftCount );
 
 }
 
@@ -395,14 +385,14 @@ The handling of underflow and overflow follows the IEC/IEEE Standard for
 Binary Floating-point Arithmetic.
 -------------------------------------------------------------------------------
 */
-static float64 roundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
+static float64 roundAndPackFloat64( struct roundingData *roundData, flag zSign, int16 zExp, bits64 zSig )
 {
     int8 roundingMode;
     flag roundNearestEven;
     int16 roundIncrement, roundBits;
     flag isTiny;
 
-    roundingMode = float_rounding_mode;
+    roundingMode = roundData->mode;
     roundNearestEven = ( roundingMode == float_round_nearest_even );
     roundIncrement = 0x200;
     if ( ! roundNearestEven ) {
@@ -427,7 +417,7 @@ static float64 roundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
            ) {
             //register int lr = __builtin_return_address(0);
             //printk("roundAndPackFloat64 called from 0x%08x\n",lr);
-            float_raise( float_flag_overflow | float_flag_inexact );
+            roundData->exception |= float_flag_overflow | float_flag_inexact;
             return packFloat64( zSign, 0x7FF, 0 ) - ( roundIncrement == 0 );
         }
         if ( zExp < 0 ) {
@@ -438,10 +428,10 @@ static float64 roundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
             shift64RightJamming( zSig, - zExp, &zSig );
             zExp = 0;
             roundBits = zSig & 0x3FF;
-            if ( isTiny && roundBits ) float_raise( float_flag_underflow );
+            if ( isTiny && roundBits ) roundData->exception |= float_flag_underflow;
         }
     }
-    if ( roundBits ) float_exception_flags |= float_flag_inexact;
+    if ( roundBits ) roundData->exception |= float_flag_inexact;
     zSig = ( zSig + roundIncrement )>>10;
     zSig &= ~ ( ( ( roundBits ^ 0x200 ) == 0 ) & roundNearestEven );
     if ( zSig == 0 ) zExp = 0;
@@ -460,12 +450,12 @@ point exponent.
 -------------------------------------------------------------------------------
 */
 static float64
- normalizeRoundAndPackFloat64( flag zSign, int16 zExp, bits64 zSig )
+ normalizeRoundAndPackFloat64( struct roundingData *roundData, flag zSign, int16 zExp, bits64 zSig )
 {
     int8 shiftCount;
 
     shiftCount = countLeadingZeros64( zSig ) - 1;
-    return roundAndPackFloat64( zSign, zExp - shiftCount, zSig<<shiftCount );
+    return roundAndPackFloat64( roundData, zSign, zExp - shiftCount, zSig<<shiftCount );
 
 }
 
@@ -572,14 +562,15 @@ Floating-point Arithmetic.
 */
 static floatx80
  roundAndPackFloatx80(
-     int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
+     struct roundingData *roundData, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
  )
 {
-    int8 roundingMode;
+    int8 roundingMode, roundingPrecision;
     flag roundNearestEven, increment, isTiny;
     int64 roundIncrement, roundMask, roundBits;
 
-    roundingMode = float_rounding_mode;
+    roundingMode = roundData->mode;
+    roundingPrecision = roundData->precision;
     roundNearestEven = ( roundingMode == float_round_nearest_even );
     if ( roundingPrecision == 80 ) goto precision80;
     if ( roundingPrecision == 64 ) {
@@ -623,8 +614,8 @@ static floatx80
             shift64RightJamming( zSig0, 1 - zExp, &zSig0 );
             zExp = 0;
             roundBits = zSig0 & roundMask;
-            if ( isTiny && roundBits ) float_raise( float_flag_underflow );
-            if ( roundBits ) float_exception_flags |= float_flag_inexact;
+            if ( isTiny && roundBits ) roundData->exception |= float_flag_underflow;
+            if ( roundBits ) roundData->exception |= float_flag_inexact;
             zSig0 += roundIncrement;
             if ( (sbits64) zSig0 < 0 ) zExp = 1;
             roundIncrement = roundMask + 1;
@@ -635,7 +626,7 @@ static floatx80
             return packFloatx80( zSign, zExp, zSig0 );
         }
     }
-    if ( roundBits ) float_exception_flags |= float_flag_inexact;
+    if ( roundBits ) roundData->exception |= float_flag_inexact;
     zSig0 += roundIncrement;
     if ( zSig0 < roundIncrement ) {
         ++zExp;
@@ -672,7 +663,7 @@ static floatx80
            ) {
             roundMask = 0;
  overflow:
-            float_raise( float_flag_overflow | float_flag_inexact );
+            roundData->exception |= float_flag_overflow | float_flag_inexact;
             if (    ( roundingMode == float_round_to_zero )
                  || ( zSign && ( roundingMode == float_round_up ) )
                  || ( ! zSign && ( roundingMode == float_round_down ) )
@@ -689,8 +680,8 @@ static floatx80
                 || ( zSig0 < LIT64( 0xFFFFFFFFFFFFFFFF ) );
             shift64ExtraRightJamming( zSig0, zSig1, 1 - zExp, &zSig0, &zSig1 );
             zExp = 0;
-            if ( isTiny && zSig1 ) float_raise( float_flag_underflow );
-            if ( zSig1 ) float_exception_flags |= float_flag_inexact;
+            if ( isTiny && zSig1 ) roundData->exception |= float_flag_underflow;
+            if ( zSig1 ) roundData->exception |= float_flag_inexact;
             if ( roundNearestEven ) {
                 increment = ( (sbits64) zSig1 < 0 );
             }
@@ -710,7 +701,7 @@ static floatx80
             return packFloatx80( zSign, zExp, zSig0 );
         }
     }
-    if ( zSig1 ) float_exception_flags |= float_flag_inexact;
+    if ( zSig1 ) roundData->exception |= float_flag_inexact;
     if ( increment ) {
         ++zSig0;
         if ( zSig0 == 0 ) {
@@ -740,7 +731,7 @@ normalized.
 */
 static floatx80
  normalizeRoundAndPackFloatx80(
-     int8 roundingPrecision, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
+     struct roundingData *roundData, flag zSign, int32 zExp, bits64 zSig0, bits64 zSig1
  )
 {
     int8 shiftCount;
@@ -754,7 +745,7 @@ static floatx80
     shortShift128Left( zSig0, zSig1, shiftCount, &zSig0, &zSig1 );
     zExp -= shiftCount;
     return
-        roundAndPackFloatx80( roundingPrecision, zSign, zExp, zSig0, zSig1 );
+        roundAndPackFloatx80( roundData, zSign, zExp, zSig0, zSig1 );
 
 }
 
@@ -767,14 +758,14 @@ the single-precision floating-point format.  The conversion is performed
 according to the IEC/IEEE Standard for Binary Floating-point Arithmetic.
 -------------------------------------------------------------------------------
 */