[ARM] Add support code for ARM hardware vector floating point

This cset adds the code to handle the hardware vector floating point
unit found on some ARM926 and later CPUs.  The hardware provides
an implementation for the common cases, and bounces exceptions for
other cases, which have to be handled in software, and signalling
SIGFPE as appropriate.

Documentation/arm/VFP/release-notes.txt

+Release notes for Linux Kernel VFP support code
+-----------------------------------------------
+
+Date: 	20 May 2004
+Author:	Russell King
+
+This is the first release of the Linux Kernel VFP support code.  It
+provides support for the exceptions bounced from VFP hardware found
+on ARM926EJ-S.
+
+This release has been validated against the SoftFloat-2b library by
+John R. Hauser using the TestFloat-2a test suite.  Details of this
+library and test suite can be found at:
+
+   http://www.cs.berkeley.edu/~jhauser/arithmetic/SoftFloat.html
+
+The operations which have been tested with this package are:
+
+ - fdiv
+ - fsub
+ - fadd
+ - fmul
+ - fcmp
+ - fcmpe
+ - fcvtd
+ - fcvts
+ - fsito
+ - ftosi
+ - fsqrt
+
+All the above pass softfloat tests with the following exceptions:
+
+- fadd/fsub shows some differences in the handling of +0 / -0 results
+  when input operands differ in signs.
+- the handling of underflow exceptions is slightly different.  If a
+  result underflows before rounding, but becomes a normalised number
+  after rounding, we do not signal an underflow exception.
+
+Other operations which have been tested by basic assembly-only tests
+are:
+
+ - fcpy
+ - fabs
+ - fneg
+ - ftoui
+ - ftosiz
+ - ftouiz
+
+The combination operations have not been tested:
+
+ - fmac
+ - fnmac
+ - fmsc
+ - fnmsc
+ - fnmul

arch/arm/Kconfig

@@ -421,6 +421,17 @@ config FPE_FASTFPE
 	  If you do not feel you need a faster FP emulation you should better
 	  choose NWFPE.
 
+config VFP
+	bool "VFP-format floating point maths"
+	help
+	  Say Y to include VFP support code in the kernel. This is needed
+	  if your hardware includes a VFP unit.
+
+	  Please see <file:Documentation/arm/VFP/release-notes.txt> for
+	  release notes and additional status information.
+
+	  Say N if your target does not have VFP hardware.
+
 source "fs/Kconfig.binfmt"
 
 source "drivers/base/Kconfig"

arch/arm/Makefile

@@ -119,6 +119,7 @@ core-y				+= arch/arm/mach-$(machine-y)/
 endif
 core-$(CONFIG_FPE_NWFPE)	+= arch/arm/nwfpe/
 core-$(CONFIG_FPE_FASTFPE)	+= $(FASTFPE_OBJ)
+core-$(CONFIG_VFP)		+= arch/arm/vfp/
 
 drivers-$(CONFIG_OPROFILE)      += arch/arm/oprofile/
 drivers-$(CONFIG_ARCH_CLPS7500)	+= drivers/acorn/char/

arch/arm/kernel/entry-armv.S

@@ -19,6 +19,7 @@
 #include <asm/thread_info.h>
 #include <asm/glue.h>
 #include <asm/ptrace.h>
+#include <asm/vfpmacros.h>
 
 #include "entry-header.S"
 
@@ -1198,8 +1199,13 @@ call_fpe:	enable_irq r10				@ Enable interrupts
 		mov	pc, lr				@ CP#7
 		mov	pc, lr				@ CP#8
 		mov	pc, lr				@ CP#9
+#ifdef CONFIG_VFP
+		b	do_vfp				@ CP#10 (VFP)
+		b	do_vfp				@ CP#11 (VFP)
+#else
 		mov	pc, lr				@ CP#10 (VFP)
 		mov	pc, lr				@ CP#11 (VFP)
+#endif
 		mov	pc, lr				@ CP#12
 		mov	pc, lr				@ CP#13
 		mov	pc, lr				@ CP#14 (Debug)
@@ -1260,6 +1266,13 @@ ENTRY(__switch_to)
 		ldr	r3, [r2, #TI_CPU_DOMAIN]!
 		stmia	ip, {r4 - sl, fp, sp, lr}	@ Store most regs on stack
 		mcr	p15, 0, r3, c3, c0, 0		@ Set domain register
+#ifdef CONFIG_VFP
+		@ Always disable VFP so we can lazily save/restore the old
+		@ state. This occurs in the context of the previous thread.
+		VFPFMRX	r4, FPEXC
+		bic	r4, r4, #FPEXC_ENABLE
+		VFPFMXR	FPEXC, r4
+#endif
 		ldmib	r2, {r4 - sl, fp, sp, pc}	@ Load all regs saved previously
 
 		__INIT

arch/arm/kernel/process.c

@@ -314,10 +314,16 @@ void flush_thread(void)
 	memset(thread->used_cp, 0, sizeof(thread->used_cp));
 	memset(&tsk->thread.debug, 0, sizeof(struct debug_info));
 	fp_init(&thread->fpstate);
+#if defined(CONFIG_VFP)
+	vfp_flush_thread(&thread->vfpstate);
+#endif
 }
 
 void release_thread(struct task_struct *dead_task)
 {
+#if defined(CONFIG_VFP)
+	vfp_release_thread(&dead_task->thread_info->vfpstate);
+#endif
 }
 
 asmlinkage void ret_from_fork(void) __asm__("ret_from_fork");

arch/arm/vfp/Makefile

+#
+# linux/arch/arm/vfp/Makefile
+#
+# Copyright (C) 2001 ARM Limited
+#
+
+# EXTRA_CFLAGS := -DDEBUG
+# EXTRA_AFLAGS := -DDEBUG
+
+obj-y			+= vfp.o
+
+vfp-$(CONFIG_VFP)	+= entry.o vfpmodule.o vfphw.o vfpsingle.o vfpdouble.o

arch/arm/vfp/entry.S

+/*
+ *  linux/arch/arm/vfp/entry.S
+ *
+ *  Copyright (C) 2004 ARM Limited.
+ *  Written by Deep Blue Solutions Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * Basic entry code, called from the kernel's undefined instruction trap.
+ *  r0  = faulted instruction
+ *  r5  = faulted PC+4
+ *  r9  = successful return
+ *  r10 = thread_info structure
+ *  lr  = failure return
+ */
+#include <linux/linkage.h>
+#include <linux/init.h>
+#include <asm/thread_info.h>
+#include <asm/vfpmacros.h>
+
+	.globl	do_vfp
+do_vfp:
+ 	ldr	r4, .LCvfp
+	add	r10, r10, #TI_VFPSTATE	@ r10 = workspace
+	ldr	pc, [r4]		@ call VFP entry point
+
+.LCvfp:
+	.word	vfp_vector
+
+@ This code is called if the VFP does not exist. It needs to flag the
+@ failure to the VFP initialisation code.
+
+	__INIT
+	.globl	vfp_testing_entry
+vfp_testing_entry:
+	ldr	r0, VFP_arch_address
+	str	r5, [r0]		@ known non-zero value
+	mov	pc, r9			@ we have handled the fault
+
+VFP_arch_address:
+	.word	VFP_arch
+
+	__FINIT

arch/arm/vfp/vfp.h

+/*
+ *  linux/arch/arm/vfp/vfp.h
+ *
+ *  Copyright (C) 2004 ARM Limited.
+ *  Written by Deep Blue Solutions Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+
+static inline u32 vfp_shiftright32jamming(u32 val, unsigned int shift)
+{
+	if (shift) {
+		if (shift < 32)
+			val = val >> shift | ((val << (32 - shift)) != 0);
+		else
+			val = val != 0;
+	}
+	return val;
+}
+
+static inline u64 vfp_shiftright64jamming(u64 val, unsigned int shift)
+{
+	if (shift) {
+		if (shift < 64)
+			val = val >> shift | ((val << (64 - shift)) != 0);
+		else
+			val = val != 0;
+	}
+	return val;
+}
+
+static inline u32 vfp_hi64to32jamming(u64 val)
+{
+	u32 v;
+
+	asm(
+	"cmp	%Q1, #1		@ vfp_hi64to32jamming\n\t"
+	"movcc	%0, %R1\n\t"
+	"orrcs	%0, %R1, #1"
+	: "=r" (v) : "r" (val) : "cc");
+
+	return v;
+}
+
+static inline void add128(u64 *resh, u64 *resl, u64 nh, u64 nl, u64 mh, u64 ml)
+{
+	asm(	"adds	%Q0, %Q2, %Q4\n\t"
+		"adcs	%R0, %R2, %R4\n\t"
+		"adcs	%Q1, %Q3, %Q5\n\t"
+		"adc	%R1, %R3, %R5"
+	    : "=r" (nl), "=r" (nh)
+	    : "0" (nl), "1" (nh), "r" (ml), "r" (mh)
+	    : "cc");
+	*resh = nh;
+	*resl = nl;
+}
+
+static inline void sub128(u64 *resh, u64 *resl, u64 nh, u64 nl, u64 mh, u64 ml)
+{
+	asm(	"subs	%Q0, %Q2, %Q4\n\t"
+		"sbcs	%R0, %R2, %R4\n\t"
+		"sbcs	%Q1, %Q3, %Q5\n\t"
+		"sbc	%R1, %R3, %R5\n\t"
+	    : "=r" (nl), "=r" (nh)
+	    : "0" (nl), "1" (nh), "r" (ml), "r" (mh)
+	    : "cc");
+	*resh = nh;
+	*resl = nl;
+}
+
+static inline void mul64to128(u64 *resh, u64 *resl, u64 n, u64 m)
+{
+	u32 nh, nl, mh, ml;
+	u64 rh, rma, rmb, rl;
+
+	nl = n;
+	ml = m;
+	rl = (u64)nl * ml;
+
+	nh = n >> 32;
+	rma = (u64)nh * ml;
+
+	mh = m >> 32;
+	rmb = (u64)nl * mh;
+	rma += rmb;
+
+	rh = (u64)nh * mh;
+	rh += ((u64)(rma < rmb) << 32) + (rma >> 32);
+
+	rma <<= 32;
+	rl += rma;
+	rh += (rl < rma);
+
+	*resl = rl;
+	*resh = rh;
+}
+
+static inline void shift64left(u64 *resh, u64 *resl, u64 n)
+{
+	*resh = n >> 63;
+	*resl = n << 1;
+}
+
+static inline u64 vfp_hi64multiply64(u64 n, u64 m)
+{
+	u64 rh, rl;
+	mul64to128(&rh, &rl, n, m);
+	return rh | (rl != 0);
+}
+
+static inline u64 vfp_estimate_div128to64(u64 nh, u64 nl, u64 m)
+{
+	u64 mh, ml, remh, reml, termh, terml, z;
+
+	if (nh >= m)
+		return ~0ULL;
+	mh = m >> 32;
+	z = (mh << 32 <= nh) ? 0xffffffff00000000ULL : (nh / mh) << 32;
+	mul64to128(&termh, &terml, m, z);
+	sub128(&remh, &reml, nh, nl, termh, terml);
+	ml = m << 32;
+	while ((s64)remh < 0) {
+		z -= 0x100000000ULL;
+		add128(&remh, &reml, remh, reml, mh, ml);
+	}
+	remh = (remh << 32) | (reml >> 32);
+	z |= (mh << 32 <= remh) ? 0xffffffff : remh / mh;
+	return z;
+}
+
+/*
+ * Operations on unpacked elements
+ */
+#define vfp_sign_negate(sign)	(sign ^ 0x8000)
+
+/*
+ * Single-precision
+ */
+struct vfp_single {
+	s16	exponent;
+	u16	sign;
+	u32	significand;
+};
+
+extern s32 vfp_get_float(unsigned int reg);
+extern void vfp_put_float(unsigned int reg, s32 val);
+
+/*
+ * VFP_SINGLE_MANTISSA_BITS - number of bits in the mantissa
+ * VFP_SINGLE_EXPONENT_BITS - number of bits in the exponent
+ * VFP_SINGLE_LOW_BITS - number of low bits in the unpacked significand
+ *  which are not propagated to the float upon packing.
+ */
+#define VFP_SINGLE_MANTISSA_BITS	(23)
+#define VFP_SINGLE_EXPONENT_BITS	(8)
+#define VFP_SINGLE_LOW_BITS		(32 - VFP_SINGLE_MANTISSA_BITS - 2)
+#define VFP_SINGLE_LOW_BITS_MASK	((1 << VFP_SINGLE_LOW_BITS) - 1)
+
+/*
+ * The bit in an unpacked float which indicates that it is a quiet NaN
+ */
+#define VFP_SINGLE_SIGNIFICAND_QNAN	(1 << (VFP_SINGLE_MANTISSA_BITS - 1 + VFP_SINGLE_LOW_BITS))
+
+/*
+ * Operations on packed single-precision numbers
+ */
+#define vfp_single_packed_sign(v)	((v) & 0x80000000)
+#define vfp_single_packed_negate(v)	((v) ^ 0x80000000)
+#define vfp_single_packed_abs(v)	((v) & ~0x80000000)
+#define vfp_single_packed_exponent(v)	(((v) >> VFP_SINGLE_MANTISSA_BITS) & ((1 << VFP_SINGLE_EXPONENT_BITS) - 1))
+#define vfp_single_packed_mantissa(v)	((v) & ((1 << VFP_SINGLE_MANTISSA_BITS) - 1))
+
+/*
+ * Unpack a single-precision float.  Note that this returns the magnitude
+ * of the single-precision float mantissa with the 1. if necessary,
+ * aligned to bit 30.
+ */
+static inline void vfp_single_unpack(struct vfp_single *s, s32 val)
+{
+	u32 significand;
+
+	s->sign = vfp_single_packed_sign(val) >> 16,
+	s->exponent = vfp_single_packed_exponent(val);
+
+	significand = (u32) val;
+	significand = (significand << (32 - VFP_SINGLE_MANTISSA_BITS)) >> 2;
+	if (s->exponent && s->exponent != 255)
+		significand |= 0x40000000;
+	s->significand = significand;
+}
+
+/*
+ * Re-pack a single-precision float.  This assumes that the float is
+ * already normalised such that the MSB is bit 30, _not_ bit 31.
+ */
+static inline s32 vfp_single_pack(struct vfp_single *s)
+{
+	u32 val;
+	val = (s->sign << 16) +
+	      (s->exponent << VFP_SINGLE_MANTISSA_BITS) +
+	      (s->significand >> VFP_SINGLE_LOW_BITS);
+	return (s32)val;
+}
+
+#define VFP_NUMBER		(1<<0)
+#define VFP_ZERO		(1<<1)
+#define VFP_DENORMAL		(1<<2)
+#define VFP_INFINITY		(1<<3)
+#define VFP_NAN			(1<<4)
+#define VFP_NAN_SIGNAL		(1<<5)
+
+#define VFP_QNAN		(VFP_NAN)
+#define VFP_SNAN		(VFP_NAN|VFP_NAN_SIGNAL)
+
+static inline int vfp_single_type(struct vfp_single *s)
+{
+	int type = VFP_NUMBER;
+	if (s->exponent == 255) {
+		if (s->significand == 0)
+			type = VFP_INFINITY;
+		else if (s->significand & VFP_SINGLE_SIGNIFICAND_QNAN)
+			type = VFP_QNAN;
+		else
+			type = VFP_SNAN;
+	} else if (s->exponent == 0) {
+		if (s->significand == 0)
+			type |= VFP_ZERO;
+		else
+			type |= VFP_DENORMAL;
+	}
+	return type;
+}
+
+#ifndef DEBUG
+#define vfp_single_normaliseround(sd,vsd,fpscr,except,func) __vfp_single_normaliseround(sd,vsd,fpscr,except)
+u32 __vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions);
+#else
+u32 vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions, const char *func);
+#endif
+
+/*
+ * Double-precision
+ */
+struct vfp_double {
+	s16	exponent;
+	u16	sign;
+	u64	significand;
+};
+
+extern u64 vfp_get_double(unsigned int reg);
+extern void vfp_put_double(unsigned int reg, u64 val);
+
+#define VFP_DOUBLE_MANTISSA_BITS	(52)
+#define VFP_DOUBLE_EXPONENT_BITS	(11)
+#define VFP_DOUBLE_LOW_BITS		(64 - VFP_DOUBLE_MANTISSA_BITS - 2)
+#define VFP_DOUBLE_LOW_BITS_MASK	((1 << VFP_DOUBLE_LOW_BITS) - 1)
+
+/*
+ * The bit in an unpacked double which indicates that it is a quiet NaN
+ */
+#define VFP_DOUBLE_SIGNIFICAND_QNAN	(1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1 + VFP_DOUBLE_LOW_BITS))
+
+/*
+ * Operations on packed single-precision numbers
+ */
+#define vfp_double_packed_sign(v)	((v) & (1ULL << 63))
+#define vfp_double_packed_negate(v)	((v) ^ (1ULL << 63))
+#define vfp_double_packed_abs(v)	((v) & ~(1ULL << 63))
+#define vfp_double_packed_exponent(v)	(((v) >> VFP_DOUBLE_MANTISSA_BITS) & ((1 << VFP_DOUBLE_EXPONENT_BITS) - 1))
+#define vfp_double_packed_mantissa(v)	((v) & ((1ULL << VFP_DOUBLE_MANTISSA_BITS) - 1))
+
+/*
+ * Unpack a double-precision float.  Note that this returns the magnitude
+ * of the double-precision float mantissa with the 1. if necessary,
+ * aligned to bit 62.
+ */
+static inline void vfp_double_unpack(struct vfp_double *s, s64 val)
+{
+	u64 significand;
+
+	s->sign = vfp_double_packed_sign(val) >> 48;
+	s->exponent = vfp_double_packed_exponent(val);
+
+	significand = (u64) val;
+	significand = (significand << (64 - VFP_DOUBLE_MANTISSA_BITS)) >> 2;
+	if (s->exponent && s->exponent != 2047)
+		significand |= (1ULL << 62);
+	s->significand = significand;
+}
+
+/*
+ * Re-pack a double-precision float.  This assumes that the float is
+ * already normalised such that the MSB is bit 30, _not_ bit 31.
+ */
+static inline s64 vfp_double_pack(struct vfp_double *s)
+{
+	u64 val;
+	val = ((u64)s->sign << 48) +
+	      ((u64)s->exponent << VFP_DOUBLE_MANTISSA_BITS) +
+	      (s->significand >> VFP_DOUBLE_LOW_BITS);
+	return (s64)val;
+}
+
+static inline int vfp_double_type(struct vfp_double *s)
+{
+	int type = VFP_NUMBER;
+	if (s->exponent == 2047) {
+		if (s->significand == 0)
+			type = VFP_INFINITY;
+		else if (s->significand & VFP_DOUBLE_SIGNIFICAND_QNAN)
+			type = VFP_QNAN;
+		else
+			type = VFP_SNAN;
+	} else if (s->exponent == 0) {
+		if (s->significand == 0)
+			type |= VFP_ZERO;
+		else
+			type |= VFP_DENORMAL;
+	}
+	return type;
+}
+
+u32 vfp_double_normaliseround(int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func);
+
+/*
+ * System registers
+ */
+extern u32 vfp_get_sys(unsigned int reg);
+extern void vfp_put_sys(unsigned int reg, u32 val);
+
+u32 vfp_estimate_sqrt_significand(u32 exponent, u32 significand);

arch/arm/vfp/vfphw.S

+/*
+ *  linux/arch/arm/vfp/vfphw.S
+ *
+ *  Copyright (C) 2004 ARM Limited.
+ *  Written by Deep Blue Solutions Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This code is called from the kernel's undefined instruction trap.
+ * r9 holds the return address for successful handling.
+ * lr holds the return address for unrecognised instructions.
+ * r10 points at the start of the private FP workspace in the thread structure
+ * sp points to a struct pt_regs (as defined in include/asm/proc/ptrace.h)
+ */
+#include <asm/thread_info.h>
+#include <asm/vfpmacros.h>
+#include "../kernel/entry-header.S"
+
+	.macro	DBGSTR, str
+#ifdef DEBUG
+	stmfd	sp!, {r0-r3, ip, lr}
+	add	r0, pc, #4
+	bl	printk
+	b	1f
+	.asciz  "<7>VFP: \str\n"
+	.balign 4
+1:	ldmfd	sp!, {r0-r3, ip, lr}
+#endif
+	.endm
+
+	.macro  DBGSTR1, str, arg
+#ifdef DEBUG
+	stmfd	sp!, {r0-r3, ip, lr}
+	mov	r1, \arg
+	add	r0, pc, #4
+	bl	printk
+	b	1f
+	.asciz  "<7>VFP: \str\n"
+	.balign 4
+1:	ldmfd	sp!, {r0-r3, ip, lr}
+#endif
+	.endm
+
+	.macro  DBGSTR3, str, arg1, arg2, arg3
+#ifdef DEBUG
+	stmfd	sp!, {r0-r3, ip, lr}
+	mov	r3, \arg3
+	mov	r2, \arg2
+	mov	r1, \arg1
+	add	r0, pc, #4
+	bl	printk
+	b	1f
+	.asciz  "<7>VFP: \str\n"
+	.balign 4
+1:	ldmfd	sp!, {r0-r3, ip, lr}
+#endif
+	.endm
+
+
+@ VFP hardware support entry point.
+@
+@  r0  = faulted instruction
+@  r5  = faulted PC+4
+@  r9  = successful return
+@  r10 = vfp_state union
+@  lr  = failure return
+
+	.globl	vfp_support_entry
+vfp_support_entry:
+	DBGSTR3	"instr %08x pc %08x state %p", r0, r5, r10
+
+	VFPFMRX	r1, FPEXC		@ Is the VFP enabled?
+	DBGSTR1	"fpexc %08x", r1
+	tst	r1, #FPEXC_ENABLE
+	bne	look_for_VFP_exceptions	@ VFP is already enabled
+
+	DBGSTR1 "enable %x", r10
+	ldr	r3, last_VFP_context_address
+	orr	r1, r1, #FPEXC_ENABLE	@ user FPEXC has the enable bit set
+	ldr	r4, [r3]		@ last_VFP_context pointer
+	bic	r2, r1, #FPEXC_EXCEPTION @ make sure exceptions are disabled
+	cmp	r4, r10
+	beq	check_for_exception	@ we are returning to the same
+					@ process, so the registers are
+					@ still there.  In this case, we do
+					@ not want to drop a pending exception.
+
+	VFPFMXR	FPEXC, r2		@ enable VFP, disable any pending
+					@ exceptions, so we can get at the
+					@ rest of it
+
+	@ Save out the current registers to the old thread state
+
+	DBGSTR1	"save old state %p", r4
+	cmp	r4, #0
+	beq	no_old_VFP_process
+	VFPFMRX	r2, FPSCR		@ current status
+	VFPFMRX	r6, FPINST		@ FPINST (always there, rev0 onwards)
+	tst	r1, #FPEXC_FPV2		@ is there an FPINST2 to read?
+	VFPFMRX	r8, FPINST2, NE		@ FPINST2 if needed - avoids reading
+					@ nonexistant reg on rev0
+	VFPFSTMIA r4 			@ save the working registers
+	add	r4, r4, #8*16+4
+	stmia	r4, {r1, r2, r6, r8}	@ save FPEXC, FPSCR, FPINST, FPINST2
+					@ and point r4 at the word at the
+					@ start of the register dump
+
+no_old_VFP_process:
+	DBGSTR1	"load state %p", r10
+	str	r10, [r3]		@ update the last_VFP_context pointer
+					@ Load the saved state back into the VFP
+	add	r4, r10, #8*16+4
+	ldmia	r4, {r1, r2, r6, r8}	@ load FPEXC, FPSCR, FPINST, FPINST2
+	VFPFLDMIA r10	 		@ reload the working registers while
+					@ FPEXC is in a safe state
+	tst	r1, #FPEXC_FPV2		@ is there an FPINST2 to write?
+	VFPFMXR	FPINST2, r8, NE		@ FPINST2 if needed - avoids writing
+					@ nonexistant reg on rev0
+	VFPFMXR	FPINST, r6
+	VFPFMXR	FPSCR, r2		@ restore status
+
+check_for_exception:
+	tst	r1, #FPEXC_EXCEPTION
+	bne	process_exception	@ might as well handle the pending
+					@ exception before retrying branch
+					@ out before setting an FPEXC that
+					@ stops us reading stuff
+	VFPFMXR	FPEXC, r1		@ restore FPEXC last
+	sub	r5, r5, #4
+	str	r5, [sp, #S_PC]		@ retry the instruction
+	mov	pc, r9			@ we think we have handled things
+
+
+look_for_VFP_exceptions:
+	tst	r1, #FPEXC_EXCEPTION
+	bne	process_exception
+	VFPFMRX	r2, FPSCR
+	tst	r2, #FPSCR_IXE		@ IXE doesn't set FPEXC_EXCEPTION !
+	bne	process_exception
+
+	@ Fall into hand on to next handler - appropriate coproc instr
+	@ not recognised by VFP
+
+	DBGSTR	"not VFP"
+	mov	pc, lr
+
+process_exception:
+	DBGSTR	"bounce"
+	sub	r5, r5, #4
+	str	r5, [sp, #S_PC]		@ retry the instruction on exit from
+					@ the imprecise exception handling in
+					@ the support code
+	mov	r2, sp			@ nothing stacked - regdump is at TOS
+	mov	lr, r9			@ setup for a return to the user code.
+
+	@ Now call the C code to package up the bounce to the support code
+	@   r0 holds the trigger instruction
+	@   r1 holds the FPEXC value
+	@   r2 pointer to register dump
+	b	VFP9_bounce		@ we have handled this - the support
+					@ code will raise an exception if
+					@ required. If not, the user code will
+					@ retry the faulted instruction
+
+last_VFP_context_address:
+	.word	last_VFP_context
+
+	.globl	vfp_get_float
+vfp_get_float:
+	add	pc, pc, r0, lsl #3
+	mov	r0, r0
+	.irp	dr,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+	mrc	p10, 0, r0, c\dr, c0, 0	@ fmrs	r0, s0
+	mov	pc, lr
+	mrc	p10, 0, r0, c\dr, c0, 4	@ fmrs	r0, s1
+	mov	pc, lr
+	.endr
+
+	.globl	vfp_put_float
+vfp_put_float:
+	add	pc, pc, r0, lsl #3
+	mov	r0, r0
+	.irp	dr,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+	mcr	p10, 0, r1, c\dr, c0, 0	@ fmsr	r0, s0
+	mov	pc, lr
+	mcr	p10, 0, r1, c\dr, c0, 4	@ fmsr	r0, s1
+	mov	pc, lr
+	.endr
+
+	.globl	vfp_get_double
+vfp_get_double:
+	mov	r0, r0, lsr #1
+	add	pc, pc, r0, lsl #3
+	mov	r0, r0
+	.irp	dr,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+	mrrc	p10, 1, r0, r1, c\dr	@ fmrrd	r0, r1, d\dr
+	mov	pc, lr
+	.endr
+
+	.globl	vfp_put_double
+vfp_put_double:
+	mov	r0, r0, lsr #1
+	add	pc, pc, r0, lsl #3
+	mov	r0, r0
+	.irp	dr,0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
+	mcrr	p10, 1, r1, r2, c\dr	@ fmrrd	r1, r2, d\dr
+	mov	pc, lr
+	.endr

arch/arm/vfp/vfpinstr.h

+/*
+ *  linux/arch/arm/vfp/vfpinstr.h
+ *
+ *  Copyright (C) 2004 ARM Limited.
+ *  Written by Deep Blue Solutions Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * VFP instruction masks.
+ */
+#define INST_CPRTDO(inst)	(((inst) & 0x0f000000) == 0x0e000000)
+#define INST_CPRT(inst)		((inst) & (1 << 4))
+#define INST_CPRT_L(inst)	((inst) & (1 << 20))
+#define INST_CPRT_Rd(inst)	(((inst) & (15 << 12)) >> 12)
+#define INST_CPRT_OP(inst)	(((inst) >> 21) & 7)
+#define INST_CPNUM(inst)	((inst) & 0xf00)
+#define CPNUM(cp)		((cp) << 8)
+
+#define FOP_MASK	(0x00b00040)
+#define FOP_FMAC	(0x00000000)
+#define FOP_FNMAC	(0x00000040)
+#define FOP_FMSC	(0x00100000)
+#define FOP_FNMSC	(0x00100040)
+#define FOP_FMUL	(0x00200000)
+#define FOP_FNMUL	(0x00200040)
+#define FOP_FADD	(0x00300000)
+#define FOP_FSUB	(0x00300040)
+#define FOP_FDIV	(0x00800000)
+#define FOP_EXT		(0x00b00040)
+
+#define FOP_TO_IDX(inst)	((inst & 0x00b00000) >> 20 | (inst & (1 << 6)) >> 4)
+
+#define FEXT_MASK	(0x000f0080)
+#define FEXT_FCPY	(0x00000000)
+#define FEXT_FABS	(0x00000080)
+#define FEXT_FNEG	(0x00010000)
+#define FEXT_FSQRT	(0x00010080)
+#define FEXT_FCMP	(0x00040000)
+#define FEXT_FCMPE	(0x00040080)
+#define FEXT_FCMPZ	(0x00050000)
+#define FEXT_FCMPEZ	(0x00050080)
+#define FEXT_FCVT	(0x00070080)
+#define FEXT_FUITO	(0x00080000)
+#define FEXT_FSITO	(0x00080080)
+#define FEXT_FTOUI	(0x000c0000)
+#define FEXT_FTOUIZ	(0x000c0080)
+#define FEXT_FTOSI	(0x000d0000)
+#define FEXT_FTOSIZ	(0x000d0080)
+
+#define FEXT_TO_IDX(inst)	((inst & 0x000f0000) >> 15 | (inst & (1 << 7)) >> 7)
+
+#define vfp_get_sd(inst)	((inst & 0x0000f000) >> 11 | (inst & (1 << 22)) >> 22)
+#define vfp_get_dd(inst)	((inst & 0x0000f000) >> 12)
+#define vfp_get_sm(inst)	((inst & 0x0000000f) << 1 | (inst & (1 << 5)) >> 5)
+#define vfp_get_dm(inst)	((inst & 0x0000000f))
+#define vfp_get_sn(inst)	((inst & 0x000f0000) >> 15 | (inst & (1 << 7)) >> 7)
+#define vfp_get_dn(inst)	((inst & 0x000f0000) >> 16)
+
+#define vfp_single(inst)	(((inst) & 0x0000f00) == 0xa00)
+
+#define FPSCR_N	(1 << 31)
+#define FPSCR_Z	(1 << 30)
+#define FPSCR_C (1 << 29)
+#define FPSCR_V	(1 << 28)
+
+/*
+ * Since we aren't building with -mfpu=vfp, we need to code
+ * these instructions using their MRC/MCR equivalents.
+ */
+#define vfpreg(_vfp_) #_vfp_
+
+#define fmrx(_vfp_) ({			\
+	u32 __v;			\
+	asm("mrc%? p10, 7, %0, " vfpreg(_vfp_) ", cr0, 0 @ fmrx	%0, " #_vfp_	\
+	    : "=r" (__v));		\
+	__v;				\
+ })
+
+#define fmxr(_vfp_,_var_)		\
+	asm("mcr%? p10, 7, %0, " vfpreg(_vfp_) ", cr0, 0 @ fmxr	" #_vfp_ ", %0"	\
+	   : : "r" (_var_))
+
+u32 vfp_single_cpdo(u32 inst, u32 fpscr);
+u32 vfp_single_cprt(u32 inst, u32 fpscr, struct pt_regs *regs);
+
+u32 vfp_double_cpdo(u32 inst, u32 fpscr);

arch/arm/vfp/vfpmodule.c

+/*
+ *  linux/arch/arm/vfp/vfpmodule.c
+ *
+ *  Copyright (C) 2004 ARM Limited.
+ *  Written by Deep Blue Solutions Limited.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+#include <linux/module.h>
+#include <linux/config.h>
+#include <linux/types.h>
+#include <linux/kernel.h>
+#include <linux/signal.h>
+#include <linux/sched.h>
+#include <linux/init.h>
+#include <asm/vfp.h>
+
+#include "vfpinstr.h"
+#include "vfp.h"
+
+/*
+ * Our undef handlers (in entry.S)
+ */
+void vfp_testing_entry(void);
+void vfp_support_entry(void);
+
+void (*vfp_vector)(void) = vfp_testing_entry;
+union vfp_state *last_VFP_context;
+
+/*
+ * Dual-use variable.
+ * Used in startup: set to non-zero if VFP checks fail
+ * After startup, holds VFP architecture
+ */
+unsigned int VFP_arch;
+
+/*
+ * Per-thread VFP initialisation.
+ */
+void vfp_flush_thread(union vfp_state *vfp)
+{
+	memset(vfp, 0, sizeof(union vfp_state));
+
+	vfp->hard.fpexc = FPEXC_ENABLE;
+	vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
+
+	/*
+	 * Disable VFP to ensure we initialise it first.
+	 */
+	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_ENABLE);
+
+	/*
+	 * Ensure we don't try to overwrite our newly initialised
+	 * state information on the first fault.
+	 */
+	if (last_VFP_context == vfp)
+		last_VFP_context = NULL;
+}
+
+/*
+ * Per-thread VFP cleanup.
+ */
+void vfp_release_thread(union vfp_state *vfp)
+{
+	if (last_VFP_context == vfp)
+		last_VFP_context = NULL;
+}
+
+/*
+ * Raise a SIGFPE for the current process.
+ * sicode describes the signal being raised.
+ */
+void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
+{
+	siginfo_t info;
+
+	memset(&info, 0, sizeof(info));
+
+	info.si_signo = SIGFPE;
+	info.si_code = sicode;
+	info.si_addr = (void *)(instruction_pointer(regs) - 4);
+
+	/*
+	 * This is the same as NWFPE, because it's not clear what
+	 * this is used for
+	 */
+	current->thread.error_code = 0;
+	current->thread.trap_no = 6;
+
+	force_sig_info(SIGFPE, &info, current);
+}
+
+static void vfp_panic(char *reason)
+{
+	int i;
+
+	printk(KERN_ERR "VFP: Error: %s\n", reason);
+	printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
+		fmrx(FPEXC), fmrx(FPSCR), fmrx(FPINST));
+	for (i = 0; i < 32; i += 2)
+		printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
+		       i, vfp_get_float(i), i+1, vfp_get_float(i+1));
+}
+
+/*
+ * Process bitmask of exception conditions.
+ */
+static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
+{
+	int si_code = 0;
+
+	pr_debug("VFP: raising exceptions %08x\n", exceptions);
+
+	if (exceptions == (u32)-1) {
+		vfp_panic("unhandled bounce");
+		vfp_raise_sigfpe(0, regs);
+		return;
+	}
+
+	/*
+	 * If any of the status flags are set, update the FPSCR.
+	 * Comparison instructions always return at least one of
+	 * these flags set.
+	 */
+	if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
+		fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
+
+	fpscr |= exceptions;
+
+	fmxr(FPSCR, fpscr);
+
+#define RAISE(stat,en,sig)				\
+	if (exceptions & stat && fpscr & en)		\
+		si_code = sig;
+
+	/*
+	 * These are arranged in priority order, least to highest.
+	 */
+	RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
+	RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
+	RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
+	RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
+
+	if (si_code)
+		vfp_raise_sigfpe(si_code, regs);
+}
+
+/*
+ * Emulate a VFP instruction.
+ */
+static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
+{
+	u32 exceptions = (u32)-1;
+
+	pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
+
+	if (INST_CPRTDO(inst)) {
+		if (!INST_CPRT(inst)) {
+			/*
+			 * CPDO
+			 */
+			if (vfp_single(inst)) {
+				exceptions = vfp_single_cpdo(inst, fpscr);
+			} else {
+				exceptions = vfp_double_cpdo(inst, fpscr);
+			}
+		} else {
+			/*
+			 * A CPRT instruction can not appear in FPINST2, nor
+			 * can it cause an exception.  Therefore, we do not
+			 * have to emulate it.
+			 */
+		}
+	} else {
+		/*
+		 * A CPDT instruction can not appear in FPINST2, nor can
+		 * it cause an exception.  Therefore, we do not have to
+		 * emulate it.
+		 */
+	}
+	return exceptions;
+}
+
+/*
+ * Package up a bounce condition.
+ */
+void VFP9_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
+{
+	u32 fpscr, orig_fpscr, exceptions, inst;
+
+	pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
+
+	/*
+	 * Enable access to the VFP so we can handle the bounce.
+	 */
+	fmxr(FPEXC, fpexc & ~(FPEXC_EXCEPTION|FPEXC_INV|FPEXC_UFC|FPEXC_IOC));
+
+	orig_fpscr = fpscr = fmrx(FPSCR);
+
+	/*
+	 * If we are running with inexact exceptions enabled, we need to
+	 * emulate the trigger instruction.  Note that as we're emulating
+	 * the trigger instruction, we need to increment PC.
+	 */
+	if (fpscr & FPSCR_IXE) {
+		regs->ARM_pc += 4;
+		goto emulate;
+	}
+
+	barrier();
+
+	/*
+	 * Modify fpscr to indicate the number of iterations remaining
+	 */
+	if (fpexc & FPEXC_EXCEPTION) {
+		u32 len;
+
+		len = fpexc + (1 << FPEXC_LENGTH_BIT);
+
+		fpscr &= ~FPSCR_LENGTH_MASK;
+		fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
+	}
+
+	/*
+	 * Handle the first FP instruction.  We used to take note of the
+	 * FPEXC bounce reason, but this appears to be unreliable.
+	 * Emulate the bounced instruction instead.
+	 */
+	inst = fmrx(FPINST);
+	exceptions = vfp_emulate_instruction(inst, fpscr, regs);
+	if (exceptions)
+		vfp_raise_exceptions(exceptions, inst, orig_fpscr, regs);
+
+	/*
+	 * If there isn't a second FP instruction, exit now.
+	 */
+	if (!(fpexc & FPEXC_FPV2))
+		return;
+
+	/*
+	 * The barrier() here prevents fpinst2 being read
+	 * before the condition above.
+	 */
+	barrier();
+	trigger = fmrx(FPINST2);
+	fpscr = fmrx(FPSCR);
+
+ emulate:
+	exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
+	if (exceptions)
+		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
+}
+ 
+/*
+ * VFP support code initialisation.
+ */
+static int __init vfp_init(void)
+{
+	unsigned int vfpsid;
+
+	/*
+	 * First check that there is a VFP that we can use.
+	 * The handler is already setup to just log calls, so
+	 * we just need to read the VFPSID register.
+	 */
+	vfpsid = fmrx(FPSID);
+
+	printk(KERN_INFO "VFP support v0.3: ");
+	if (VFP_arch) {
+		printk("not present\n");
+	} else if (vfpsid & FPSID_NODOUBLE) {
+		printk("no double precision support\n");
+	} else {
+		VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;  /* Extract the architecture version */
+		printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
+			(vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
+			(vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
+			(vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
+			(vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
+			(vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
+		vfp_vector = vfp_support_entry;
+	}
+	return 0;
+}
+
+late_initcall(vfp_init);

include/asm-arm/fpstate.h

@@ -13,6 +13,36 @@
 
 #define FP_SIZE 35
 
+#ifndef __ASSEMBLY__
+
+/*
+ * VFP storage area has:
+ *  - FPEXC, FPSCR, FPINST and FPINST2.
+ *  - 16 double precision data registers
+ *  - an implementation-dependant word of state for FLDMX/FSTMX
+ * 
+ *  FPEXC will always be non-zero once the VFP has been used in this process.
+ */
+
+struct vfp_hard_struct {
+	__u64 fpregs[16];
+	__u32 fpmx_state;
+	__u32 fpexc;
+	__u32 fpscr;
+	/*
+	 * VFP implementation specific state
+	 */
+	__u32 fpinst;
+	__u32 fpinst2;
+};
+
+union vfp_state {
+	struct vfp_hard_struct	hard;
+};
+
+extern void vfp_flush_thread(union vfp_state *);
+extern void vfp_release_thread(union vfp_state *);
+
 struct fp_hard_struct {
 	unsigned int save[FP_SIZE];		/* as yet undefined */
 };
@@ -27,3 +57,5 @@ union fp_state {
 };
 
 #endif
+
+#endif

include/asm-arm/thread_info.h

@@ -12,12 +12,13 @@
 
 #ifdef __KERNEL__
 
+#include <asm/fpstate.h>
+
 #ifndef __ASSEMBLY__
 
 struct task_struct;
 struct exec_domain;
 
-#include <asm/fpstate.h>
 #include <asm/ptrace.h>
 #include <asm/types.h>
 #include <asm/domain.h>
@@ -53,6 +54,7 @@ struct thread_info {
 	struct cpu_context_save	cpu_context;	/* cpu context */
 	__u8			used_cp[16];	/* thread used copro */
 	union fp_state		fpstate;
+	union vfp_state		vfpstate;
 	struct restart_block	restart_block;
 };
 
@@ -110,6 +112,7 @@ extern void free_thread_info(struct thread_info *);
 #define TI_CPU_SAVE	28
 #define TI_USED_CP	76
 #define TI_FPSTATE	(TI_USED_CP+16)
+#define TI_VFPSTATE	(TI_FPSTATE+FP_SIZE*4)
 
 #endif
 

include/asm-arm/vfp.h

+/*
+ * linux/include/asm-arm/vfp.h
+ *
+ * VFP register definitions.
+ * First, the standard VFP set.
+ */
+
+#define FPSID			cr0
+#define FPSCR			cr1
+#define FPEXC			cr8
+
+/* FPSID bits */
+#define FPSID_IMPLEMENTER_BIT	(24)
+#define FPSID_IMPLEMENTER_MASK	(0xff << FPSID_IMPLEMENTER_BIT)
+#define FPSID_SOFTWARE		(1<<23)
+#define FPSID_FORMAT_BIT	(21)
+#define FPSID_FORMAT_MASK	(0x3  << FPSID_FORMAT_BIT)
+#define FPSID_NODOUBLE		(1<<20)
+#define FPSID_ARCH_BIT		(16)
+#define FPSID_ARCH_MASK		(0xF  << FPSID_ARCH_BIT)
+#define FPSID_PART_BIT		(8)
+#define FPSID_PART_MASK		(0xFF << FPSID_PART_BIT)
+#define FPSID_VARIANT_BIT	(4)
+#define FPSID_VARIANT_MASK	(0xF  << FPSID_VARIANT_BIT)
+#define FPSID_REV_BIT		(0)
+#define FPSID_REV_MASK		(0xF  << FPSID_REV_BIT)
+
+/* FPEXC bits */
+#define FPEXC_EXCEPTION		(1<<31)
+#define FPEXC_ENABLE		(1<<30)
+
+/* FPSCR bits */
+#define FPSCR_DEFAULT_NAN	(1<<25)
+#define FPSCR_FLUSHTOZERO	(1<<24)
+#define FPSCR_ROUND_NEAREST	(0<<22)
+#define FPSCR_ROUND_PLUSINF	(1<<22)
+#define FPSCR_ROUND_MINUSINF	(2<<22)
+#define FPSCR_ROUND_TOZERO	(3<<22)
+#define FPSCR_RMODE_BIT		(22)
+#define FPSCR_RMODE_MASK	(3 << FPSCR_RMODE_BIT)
+#define FPSCR_STRIDE_BIT	(20)
+#define FPSCR_STRIDE_MASK	(3 << FPSCR_STRIDE_BIT)
+#define FPSCR_LENGTH_BIT	(16)
+#define FPSCR_LENGTH_MASK	(7 << FPSCR_LENGTH_BIT)
+#define FPSCR_IOE		(1<<8)
+#define FPSCR_DZE		(1<<9)
+#define FPSCR_OFE		(1<<10)
+#define FPSCR_UFE		(1<<11)
+#define FPSCR_IXE		(1<<12)
+#define FPSCR_IDE		(1<<15)
+#define FPSCR_IOC		(1<<0)
+#define FPSCR_DZC		(1<<1)
+#define FPSCR_OFC		(1<<2)
+#define FPSCR_UFC		(1<<3)
+#define FPSCR_IXC		(1<<4)
+#define FPSCR_IDC		(1<<7)
+
+/*
+ * VFP9-S specific.
+ */
+#define FPINST			cr9
+#define FPINST2			cr10
+
+/* FPEXC bits */
+#define FPEXC_FPV2		(1<<28)
+#define FPEXC_LENGTH_BIT	(8)
+#define FPEXC_LENGTH_MASK	(7 << FPEXC_LENGTH_BIT)
+#define FPEXC_INV		(1 << 7)
+#define FPEXC_UFC		(1 << 3)
+#define FPEXC_OFC		(1 << 2)
+#define FPEXC_IOC		(1 << 0)
+
+/* Bit patterns for decoding the packaged operation descriptors */
+#define VFPOPDESC_LENGTH_BIT	(9)
+#define VFPOPDESC_LENGTH_MASK	(0x07 << VFPOPDESC_LENGTH_BIT)
+#define VFPOPDESC_UNUSED_BIT	(24)
+#define VFPOPDESC_UNUSED_MASK	(0xFF << VFPOPDESC_UNUSED_BIT)
+#define VFPOPDESC_OPDESC_MASK	(~(VFPOPDESC_LENGTH_MASK | VFPOPDESC_UNUSED_MASK))

include/asm-arm/vfpmacros.h

+/*
+ * linux/include/asm-arm/vfpmacros.h
+ *
+ * Assembler-only file containing VFP macros and register definitions.
+ */
+#include "vfp.h"
+
+@ Macros to allow building with old toolkits (with no VFP support)
+	.macro	VFPFMRX, rd, sysreg, cond
+	MRC\cond	p10, 7, \rd, \sysreg, cr0, 0	@ FMRX	\rd, \sysreg
+	.endm
+
+	.macro	VFPFMXR, sysreg, rd, cond
+	MCR\cond	p10, 7, \rd, \sysreg, cr0, 0	@ FMXR	\sysreg, \rd
+	.endm
+
+	@ read all the working registers back into the VFP
+	.macro	VFPFLDMIA, base
+	LDC	p11, cr0, [\base],#33*4		    @ FLDMIAX \base!, {d0-d15}
+	.endm
+
+	@ write all the working registers out of the VFP
+	.macro	VFPFSTMIA, base
+	STC	p11, cr0, [\base],#33*4		    @ FSTMIAX \base!, {d0-d15}
+	.endm