linux/arch/arm/vfp/vfp.h
Catalin Marinas c642846489 [ARM] 4111/1: Allow VFP to work with thread migration on SMP
The current lazy saving of the VFP registers is no longer possible
with thread migration on SMP. This patch implements a per-CPU
vfp-state pointer and the saving of the VFP registers at every context
switch. The registers restoring is still performed in a lazy way.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
2007-01-25 16:35:29 +00:00

377 lines
9.3 KiB
C

/*
* 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;
if (mh << 32 <= nh) {
z = 0xffffffff00000000ULL;
} else {
z = nh;
do_div(z, mh);
z <<= 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);
if (mh << 32 <= remh) {
z |= 0xffffffff;
} else {
do_div(remh, mh);
z |= remh;
}
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(s32 val, unsigned int reg);
/*
* 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;
};
/*
* VFP_REG_ZERO is a special register number for vfp_get_double
* which returns (double)0.0. This is useful for the compare with
* zero instructions.
*/
#define VFP_REG_ZERO 16
extern u64 vfp_get_double(unsigned int reg);
extern void vfp_put_double(u64 val, unsigned int reg);
#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);
u32 vfp_estimate_sqrt_significand(u32 exponent, u32 significand);
/*
* A special flag to tell the normalisation code not to normalise.
*/
#define VFP_NAN_FLAG 0x100
/*
* A bit pattern used to indicate the initial (unset) value of the
* exception mask, in case nothing handles an instruction. This
* doesn't include the NAN flag, which get masked out before
* we check for an error.
*/
#define VFP_EXCEPTION_ERROR ((u32)-1 & ~VFP_NAN_FLAG)
/*
* A flag to tell vfp instruction type.
* OP_SCALAR - this operation always operates in scalar mode
* OP_SD - the instruction exceptionally writes to a single precision result.
* OP_DD - the instruction exceptionally writes to a double precision result.
*/
#define OP_SCALAR (1 << 0)
#define OP_SD (1 << 1)
#define OP_DD (1 << 1)
struct op {
u32 (* const fn)(int dd, int dn, int dm, u32 fpscr);
u32 flags;
};
#ifdef CONFIG_SMP
extern void vfp_save_state(void *location, u32 fpexc);
#endif