vfpdouble.c

linux-2.6.15.6

/*
 *  linux/arch/arm/vfp/vfpdouble.c
 *
 * This code is derived in part from John R. Housers softfloat library, which
 * carries the following notice:
 *
 * ===========================================================================
 * This C source file is part of the SoftFloat IEC/IEEE Floating-point
 * Arithmetic Package, Release 2.
 *
 * Written by John R. Hauser.  This work was made possible in part by the
 * International Computer Science Institute, located at Suite 600, 1947 Center
 * Street, Berkeley, California 94704.  Funding was partially provided by the
 * National Science Foundation under grant MIP-9311980.  The original version
 * of this code was written as part of a project to build a fixed-point vector
 * processor in collaboration with the University of California at Berkeley,
 * overseen by Profs. Nelson Morgan and John Wawrzynek.  More information
 * is available through the web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
 * arithmetic/softfloat.html'.
 *
 * THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE.  Although reasonable effort
 * has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
 * TIMES RESULT IN INCORRECT BEHAVIOR.  USE OF THIS SOFTWARE IS RESTRICTED TO
 * PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
 * AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
 *
 * Derivative works are acceptable, even for commercial purposes, so long as
 * (1) they include prominent notice that the work is derivative, and (2) they
 * include prominent notice akin to these three paragraphs for those parts of
 * this code that are retained.
 * ===========================================================================
 */
#include <linux/kernel.h>
#include <linux/bitops.h>

#include <asm/div64.h>
#include <asm/ptrace.h>
#include <asm/vfp.h>

#include "vfpinstr.h"
#include "vfp.h"

static struct vfp_double vfp_double_default_qnan = {
	.exponent	= 2047,
	.sign		= 0,
	.significand	= VFP_DOUBLE_SIGNIFICAND_QNAN,
};

static void vfp_double_dump(const char *str, struct vfp_double *d)
{
	pr_debug("VFP: %s: sign=%d exponent=%d significand=%016llx\n",
		 str, d->sign != 0, d->exponent, d->significand);
}

static void vfp_double_normalise_denormal(struct vfp_double *vd)
{
	int bits = 31 - fls(vd->significand >> 32);
	if (bits == 31)
		bits = 62 - fls(vd->significand);

	vfp_double_dump("normalise_denormal: in", vd);

	if (bits) {
		vd->exponent -= bits - 1;
		vd->significand <<= bits;
	}

	vfp_double_dump("normalise_denormal: out", vd);
}

u32 vfp_double_normaliseround(int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func)
{
	u64 significand, incr;
	int exponent, shift, underflow;
	u32 rmode;

	vfp_double_dump("pack: in", vd);

	/*
	 * Infinities and NaNs are a special case.
	 */
	if (vd->exponent == 2047 && (vd->significand == 0 || exceptions))
		goto pack;

	/*
	 * Special-case zero.
	 */
	if (vd->significand == 0) {
		vd->exponent = 0;
		goto pack;
	}

	exponent = vd->exponent;
	significand = vd->significand;

	shift = 32 - fls(significand >> 32);
	if (shift == 32)
		shift = 64 - fls(significand);
	if (shift) {
		exponent -= shift;
		significand <<= shift;
	}

#ifdef DEBUG
	vd->exponent = exponent;
	vd->significand = significand;
	vfp_double_dump("pack: normalised", vd);
#endif

	/*
	 * Tiny number?
	 */
	underflow = exponent < 0;
	if (underflow) {
		significand = vfp_shiftright64jamming(significand, -exponent);
		exponent = 0;
#ifdef DEBUG
		vd->exponent = exponent;
		vd->significand = significand;
		vfp_double_dump("pack: tiny number", vd);
#endif
		if (!(significand & ((1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1)))
			underflow = 0;
	}

	/*
	 * Select rounding increment.
	 */
	incr = 0;
	rmode = fpscr & FPSCR_RMODE_MASK;

	if (rmode == FPSCR_ROUND_NEAREST) {
		incr = 1ULL << VFP_DOUBLE_LOW_BITS;
		if ((significand & (1ULL << (VFP_DOUBLE_LOW_BITS + 1))) == 0)
			incr -= 1;
	} else if (rmode == FPSCR_ROUND_TOZERO) {
		incr = 0;
	} else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vd->sign != 0))
		incr = (1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1;

	pr_debug("VFP: rounding increment = 0x%08llx\n", incr);

	/*
	 * Is our rounding going to overflow?
	 */
	if ((significand + incr) < significand) {
		exponent += 1;
		significand = (significand >> 1) | (significand & 1);
		incr >>= 1;
#ifdef DEBUG
		vd->exponent = exponent;
		vd->significand = significand;
		vfp_double_dump("pack: overflow", vd);
#endif
	}

	/*
	 * If any of the low bits (which will be shifted out of the
	 * number) are non-zero, the result is inexact.
	 */
	if (significand & ((1 << (VFP_DOUBLE_LOW_BITS + 1)) - 1))
		exceptions |= FPSCR_IXC;

	/*
	 * Do our rounding.
	 */
	significand += incr;

	/*
	 * Infinity?
	 */
	if (exponent >= 2046) {
		exceptions |= FPSCR_OFC | FPSCR_IXC;
		if (incr == 0) {
			vd->exponent = 2045;
			vd->significand = 0x7fffffffffffffffULL;
		} else {
			vd->exponent = 2047;		/* infinity */
			vd->significand = 0;
		}
	} else {
		if (significand >> (VFP_DOUBLE_LOW_BITS + 1) == 0)
			exponent = 0;
		if (exponent || significand > 0x8000000000000000ULL)
			underflow = 0;
		if (underflow)
			exceptions |= FPSCR_UFC;
		vd->exponent = exponent;
		vd->significand = significand >> 1;
	}

 pack:
	vfp_double_dump("pack: final", vd);
	{
		s64 d = vfp_double_pack(vd);
		pr_debug("VFP: %s: d(d%d)=%016llx exceptions=%08x\n", func,
			 dd, d, exceptions);
		vfp_put_double(dd, d);
	}
	return exceptions & ~VFP_NAN_FLAG;
}

/*
 * Propagate the NaN, setting exceptions if it is signalling.
 * 'n' is always a NaN.  'm' may be a number, NaN or infinity.
 */
static u32
vfp_propagate_nan(struct vfp_double *vdd, struct vfp_double *vdn,
		  struct vfp_double *vdm, u32 fpscr)
{
	struct vfp_double *nan;
	int tn, tm = 0;

	tn = vfp_double_type(vdn);

	if (vdm)
		tm = vfp_double_type(vdm);

	if (fpscr & FPSCR_DEFAULT_NAN)
		/*
		 * Default NaN mode - always returns a quiet NaN
		 */
		nan = &vfp_double_default_qnan;
	else {
		/*
		 * Contemporary mode - select the first signalling
		 * NAN, or if neither are signalling, the first
		 * quiet NAN.
		 */
		if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
			nan = vdn;
		else
			nan = vdm;
		/*
		 * Make the NaN quiet.
		 */
		nan->significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
	}

	*vdd = *nan;

	/*
	 * If one was a signalling NAN, raise invalid operation.
	 */
	return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
}

/*
 * Extended operations
 */
static u32 vfp_double_fabs(int dd, int unused, int dm, u32 fpscr)
{
	vfp_put_double(dd, vfp_double_packed_abs(vfp_get_double(dm)));
	return 0;
}

static u32 vfp_double_fcpy(int dd, int unused, int dm, u32 fpscr)
{
	vfp_put_double(dd, vfp_get_double(dm));
	return 0;
}

static u32 vfp_double_fneg(int dd, int unused, int dm, u32 fpscr)
{
	vfp_put_double(dd, vfp_double_packed_negate(vfp_get_double(dm)));
	return 0;
}

static u32 vfp_double_fsqrt(int dd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm, vdd;
	int ret, tm;

	vfp_double_unpack(&vdm, vfp_get_double(dm));
	tm = vfp_double_type(&vdm);
	if (tm & (VFP_NAN|VFP_INFINITY)) {
		struct vfp_double *vdp = &vdd;

		if (tm & VFP_NAN)
			ret = vfp_propagate_nan(vdp, &vdm, NULL, fpscr);
		else if (vdm.sign == 0) {
 sqrt_copy:
			vdp = &vdm;
			ret = 0;
		} else {
 sqrt_invalid:
			vdp = &vfp_double_default_qnan;
			ret = FPSCR_IOC;
		}
		vfp_put_double(dd, vfp_double_pack(vdp));
		return ret;
	}

	/*
	 * sqrt(+/- 0) == +/- 0
	 */
	if (tm & VFP_ZERO)
		goto sqrt_copy;

	/*
	 * Normalise a denormalised number
	 */
	if (tm & VFP_DENORMAL)
		vfp_double_normalise_denormal(&vdm);

	/*
	 * sqrt(<0) = invalid
	 */
	if (vdm.sign)
		goto sqrt_invalid;

	vfp_double_dump("sqrt", &vdm);

	/*
	 * Estimate the square root.
	 */
	vdd.sign = 0;
	vdd.exponent = ((vdm.exponent - 1023) >> 1) + 1023;
	vdd.significand = (u64)vfp_estimate_sqrt_significand(vdm.exponent, vdm.significand >> 32) << 31;

	vfp_double_dump("sqrt estimate1", &vdd);

	vdm.significand >>= 1 + (vdm.exponent & 1);
	vdd.significand += 2 + vfp_estimate_div128to64(vdm.significand, 0, vdd.significand);

	vfp_double_dump("sqrt estimate2", &vdd);

	/*
	 * And now adjust.
	 */
	if ((vdd.significand & VFP_DOUBLE_LOW_BITS_MASK) <= 5) {
		if (vdd.significand < 2) {
			vdd.significand = ~0ULL;
		} else {
			u64 termh, terml, remh, reml;
			vdm.significand <<= 2;
			mul64to128(&termh, &terml, vdd.significand, vdd.significand);
			sub128(&remh, &reml, vdm.significand, 0, termh, terml);
			while ((s64)remh < 0) {
				vdd.significand -= 1;
				shift64left(&termh, &terml, vdd.significand);
				terml |= 1;
				add128(&remh, &reml, remh, reml, termh, terml);
			}
			vdd.significand |= (remh | reml) != 0;
		}
	}
	vdd.significand = vfp_shiftright64jamming(vdd.significand, 1);

	return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fsqrt");
}

/*
 * Equal	:= ZC
 * Less than	:= N
 * Greater than	:= C
 * Unordered	:= CV
 */
static u32 vfp_compare(int dd, int signal_on_qnan, int dm, u32 fpscr)
{
	s64 d, m;
	u32 ret = 0;

	m = vfp_get_double(dm);
	if (vfp_double_packed_exponent(m) == 2047 && vfp_double_packed_mantissa(m)) {
		ret |= FPSCR_C | FPSCR_V;
		if (signal_on_qnan || !(vfp_double_packed_mantissa(m) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
			/*
			 * Signalling NaN, or signalling on quiet NaN
			 */
			ret |= FPSCR_IOC;
	}

	d = vfp_get_double(dd);
	if (vfp_double_packed_exponent(d) == 2047 && vfp_double_packed_mantissa(d)) {
		ret |= FPSCR_C | FPSCR_V;
		if (signal_on_qnan || !(vfp_double_packed_mantissa(d) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
			/*
			 * Signalling NaN, or signalling on quiet NaN
			 */
			ret |= FPSCR_IOC;
	}

	if (ret == 0) {
		if (d == m || vfp_double_packed_abs(d | m) == 0) {
			/*
			 * equal
			 */
			ret |= FPSCR_Z | FPSCR_C;
		} else if (vfp_double_packed_sign(d ^ m)) {
			/*
			 * different signs
			 */
			if (vfp_double_packed_sign(d))
				/*
				 * d is negative, so d < m
				 */
				ret |= FPSCR_N;
			else
				/*
				 * d is positive, so d > m
				 */
				ret |= FPSCR_C;
		} else if ((vfp_double_packed_sign(d) != 0) ^ (d < m)) {
			/*
			 * d < m
			 */
			ret |= FPSCR_N;
		} else if ((vfp_double_packed_sign(d) != 0) ^ (d > m)) {
			/*
			 * d > m
			 */
			ret |= FPSCR_C;
		}
	}

	return ret;
}

static u32 vfp_double_fcmp(int dd, int unused, int dm, u32 fpscr)
{
	return vfp_compare(dd, 0, dm, fpscr);
}

static u32 vfp_double_fcmpe(int dd, int unused, int dm, u32 fpscr)
{
	return vfp_compare(dd, 1, dm, fpscr);
}

static u32 vfp_double_fcmpz(int dd, int unused, int dm, u32 fpscr)
{
	return vfp_compare(dd, 0, VFP_REG_ZERO, fpscr);
}

static u32 vfp_double_fcmpez(int dd, int unused, int dm, u32 fpscr)
{
	return vfp_compare(dd, 1, VFP_REG_ZERO, fpscr);
}

static u32 vfp_double_fcvts(int sd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm;
	struct vfp_single vsd;
	int tm;
	u32 exceptions = 0;

	vfp_double_unpack(&vdm, vfp_get_double(dm));

	tm = vfp_double_type(&vdm);

	/*
	 * If we have a signalling NaN, signal invalid operation.
	 */
	if (tm == VFP_SNAN)
		exceptions = FPSCR_IOC;

	if (tm & VFP_DENORMAL)
		vfp_double_normalise_denormal(&vdm);

	vsd.sign = vdm.sign;
	vsd.significand = vfp_hi64to32jamming(vdm.significand);

	/*
	 * If we have an infinity or a NaN, the exponent must be 255
	 */
	if (tm & (VFP_INFINITY|VFP_NAN)) {
		vsd.exponent = 255;
		if (tm & VFP_NAN)
			vsd.significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
		goto pack_nan;
	} else if (tm & VFP_ZERO)
		vsd.exponent = 0;
	else
		vsd.exponent = vdm.exponent - (1023 - 127);

	return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fcvts");

 pack_nan:
	vfp_put_float(sd, vfp_single_pack(&vsd));
	return exceptions;
}

static u32 vfp_double_fuito(int dd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm;
	u32 m = vfp_get_float(dm);

	vdm.sign = 0;
	vdm.exponent = 1023 + 63 - 1;
	vdm.significand = (u64)m;

	return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fuito");
}

static u32 vfp_double_fsito(int dd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm;
	u32 m = vfp_get_float(dm);

	vdm.sign = (m & 0x80000000) >> 16;
	vdm.exponent = 1023 + 63 - 1;
	vdm.significand = vdm.sign ? -m : m;

	return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fsito");
}

static u32 vfp_double_ftoui(int sd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm;
	u32 d, exceptions = 0;
	int rmode = fpscr & FPSCR_RMODE_MASK;
	int tm;

	vfp_double_unpack(&vdm, vfp_get_double(dm));

	/*
	 * Do we have a denormalised number?
	 */
	tm = vfp_double_type(&vdm);
	if (tm & VFP_DENORMAL)
		exceptions |= FPSCR_IDC;

	if (tm & VFP_NAN)
		vdm.sign = 0;

	if (vdm.exponent >= 1023 + 32) {
		d = vdm.sign ? 0 : 0xffffffff;
		exceptions = FPSCR_IOC;
	} else if (vdm.exponent >= 1023 - 1) {
		int shift = 1023 + 63 - vdm.exponent;
		u64 rem, incr = 0;

		/*
		 * 2^0 <= m < 2^32-2^8
		 */
		d = (vdm.significand << 1) >> shift;
		rem = vdm.significand << (65 - shift);

		if (rmode == FPSCR_ROUND_NEAREST) {
			incr = 0x8000000000000000ULL;
			if ((d & 1) == 0)
				incr -= 1;
		} else if (rmode == FPSCR_ROUND_TOZERO) {
			incr = 0;
		} else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vdm.sign != 0)) {
			incr = ~0ULL;
		}

		if ((rem + incr) < rem) {
			if (d < 0xffffffff)
				d += 1;
			else
				exceptions |= FPSCR_IOC;
		}

		if (d && vdm.sign) {
			d = 0;
			exceptions |= FPSCR_IOC;
		} else if (rem)
			exceptions |= FPSCR_IXC;
	} else {
		d = 0;
		if (vdm.exponent | vdm.significand) {
			exceptions |= FPSCR_IXC;
			if (rmode == FPSCR_ROUND_PLUSINF && vdm.sign == 0)
				d = 1;
			else if (rmode == FPSCR_ROUND_MINUSINF && vdm.sign) {
				d = 0;
				exceptions |= FPSCR_IOC;
			}
		}
	}

	pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);

	vfp_put_float(sd, d);

	return exceptions;
}

static u32 vfp_double_ftouiz(int sd, int unused, int dm, u32 fpscr)
{
	return vfp_double_ftoui(sd, unused, dm, FPSCR_ROUND_TOZERO);
}

static u32 vfp_double_ftosi(int sd, int unused, int dm, u32 fpscr)
{
	struct vfp_double vdm;
	u32 d, exceptions = 0;
	int rmode = fpscr & FPSCR_RMODE_MASK;

	vfp_double_unpack(&vdm, vfp_get_double(dm));
	vfp_double_dump("VDM", &vdm);

	/*
	 * Do we have denormalised number?