octo/yuzu-android
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ARM: Merge latest VFP fixes from 3dmoo team.

Browse Source
This commit is contained in:
bunnei
2014-10-29 22:25:54 -04:00
parent 2ca12e7f38
commit bc6989b075
4 changed files with 2161 additions and 1869 deletions
Download Patch File Download Diff File Expand all files Collapse all files

178
src/core/arm/skyeye_common/vfp/vfp.cpp
View File

@@ -28,6 +28,8 @@
#include "core/arm/skyeye_common/armdefs.h"
#include "core/arm/skyeye_common/vfp/vfp.h"
#define DEBUG DBG
//ARMul_State* persistent_state; /* function calls from SoftFloat lib don't have an access to ARMul_state. */
unsigned
@@ -41,11 +43,11 @@ VFPInit (ARMul_State *state)
//persistent_state = state;
/* Reset only specify VFP_FPEXC_EN = '0' */
return No_exp;
return 0;
}
unsigned
VFPMRC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value)
VFPMRC (ARMul_State * state, unsigned type, u32 instr, u32 * value)
{
/* MRC<c> <coproc>,<opc1>,<Rt>,<CRn>,<CRm>{,<opc2>} */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -59,20 +61,19 @@ VFPMRC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value)
/* CRn/opc1 CRm/opc2 */
if (CoProc == 10 || CoProc == 11)
{
#define VFP_MRC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MRC_TRANS
if (CoProc == 10 || CoProc == 11) {
#define VFP_MRC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MRC_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, CRn %x, CRm %x, OPC_2 %x\n",
DEBUG("Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, CRn %x, CRm %x, OPC_2 %x\n",
instr, CoProc, OPC_1, Rt, CRn, CRm, OPC_2);
return ARMul_CANT;
}
unsigned
VFPMCR (ARMul_State * state, unsigned type, ARMword instr, ARMword value)
VFPMCR (ARMul_State * state, unsigned type, u32 instr, u32 value)
{
/* MCR<c> <coproc>,<opc1>,<Rt>,<CRn>,<CRm>{,<opc2>} */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -85,20 +86,19 @@ VFPMCR (ARMul_State * state, unsigned type, ARMword instr, ARMword value)
/* TODO check access permission */
/* CRn/opc1 CRm/opc2 */
if (CoProc == 10 || CoProc == 11)
{
#define VFP_MCR_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MCR_TRANS
if (CoProc == 10 || CoProc == 11) {
#define VFP_MCR_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MCR_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, CRn %x, CRm %x, OPC_2 %x\n",
DEBUG("Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, CRn %x, CRm %x, OPC_2 %x\n",
instr, CoProc, OPC_1, Rt, CRn, CRm, OPC_2);
return ARMul_CANT;
}
unsigned
VFPMRRC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value1, ARMword * value2)
VFPMRRC (ARMul_State * state, unsigned type, u32 instr, u32 * value1, u32 * value2)
{
/* MCRR<c> <coproc>,<opc1>,<Rt>,<Rt2>,<CRm> */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -107,20 +107,19 @@ VFPMRRC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value1, AR
int Rt2 = BITS (16, 19);
int CRm = BITS (0, 3);
if (CoProc == 10 || CoProc == 11)
{
#define VFP_MRRC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MRRC_TRANS
if (CoProc == 10 || CoProc == 11) {
#define VFP_MRRC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MRRC_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, Rt2 %x, CRm %x\n",
DEBUG("Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, Rt2 %x, CRm %x\n",
instr, CoProc, OPC_1, Rt, Rt2, CRm);
return ARMul_CANT;
}
unsigned
VFPMCRR (ARMul_State * state, unsigned type, ARMword instr, ARMword value1, ARMword value2)
VFPMCRR (ARMul_State * state, unsigned type, u32 instr, u32 value1, u32 value2)
{
/* MCRR<c> <coproc>,<opc1>,<Rt>,<Rt2>,<CRm> */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -133,20 +132,19 @@ VFPMCRR (ARMul_State * state, unsigned type, ARMword instr, ARMword value1, ARMw
/* CRn/opc1 CRm/opc2 */
if (CoProc == 11 || CoProc == 10)
{
#define VFP_MCRR_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MCRR_TRANS
if (CoProc == 11 || CoProc == 10) {
#define VFP_MCRR_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_MCRR_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, Rt2 %x, CRm %x\n",
DEBUG("Can't identify %x, CoProc %x, OPC_1 %x, Rt %x, Rt2 %x, CRm %x\n",
instr, CoProc, OPC_1, Rt, Rt2, CRm);
return ARMul_CANT;
}
unsigned
VFPSTC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value)
VFPSTC (ARMul_State * state, unsigned type, u32 instr, u32 * value)
{
/* STC{L}<c> <coproc>,<CRd>,[<Rn>],<option> */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -161,35 +159,34 @@ VFPSTC (ARMul_State * state, unsigned type, ARMword instr, ARMword * value)
/* TODO check access permission */
/* VSTM */
if ( (P|U|D|W) == 0 )
{
DEBUG_LOG(ARM11, "In %s, UNDEFINED\n", __FUNCTION__); exit(-1);
if ( (P|U|D|W) == 0 ) {
DEBUG("In %s, UNDEFINED\n", __FUNCTION__);
exit(-1);
}
if (CoProc == 10 || CoProc == 11)
{
#if 1
if (P == 0 && U == 0 && W == 0)
{
DEBUG_LOG(ARM11, "VSTM Related encodings\n"); exit(-1);
if (CoProc == 10 || CoProc == 11) {
#if 1
if (P == 0 && U == 0 && W == 0) {
DEBUG("VSTM Related encodings\n");
exit(-1);
}
if (P == U && W == 1)
{
DEBUG_LOG(ARM11, "UNDEFINED\n"); exit(-1);
if (P == U && W == 1) {
DEBUG("UNDEFINED\n");
exit(-1);
}
#endif
#endif
#define VFP_STC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_STC_TRANS
#define VFP_STC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_STC_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, CRd %x, Rn %x, imm8 %x, P %x, U %x, D %x, W %x\n",
DEBUG("Can't identify %x, CoProc %x, CRd %x, Rn %x, imm8 %x, P %x, U %x, D %x, W %x\n",
instr, CoProc, CRd, Rn, imm8, P, U, D, W);
return ARMul_CANT;
}
unsigned
VFPLDC (ARMul_State * state, unsigned type, ARMword instr, ARMword value)
VFPLDC (ARMul_State * state, unsigned type, u32 instr, u32 value)
{
/* LDC{L}<c> <coproc>,<CRd>,[<Rn>] */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -203,24 +200,23 @@ VFPLDC (ARMul_State * state, unsigned type, ARMword instr, ARMword value)
/* TODO check access permission */
if ( (P|U|D|W) == 0 )
{
DEBUG_LOG(ARM11, "In %s, UNDEFINED\n", __FUNCTION__); exit(-1);
if ( (P|U|D|W) == 0 ) {
DEBUG("In %s, UNDEFINED\n", __FUNCTION__);
exit(-1);
}
if (CoProc == 10 || CoProc == 11)
{
#define VFP_LDC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_LDC_TRANS
if (CoProc == 10 || CoProc == 11) {
#define VFP_LDC_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_LDC_TRANS
}
DEBUG_LOG(ARM11, "Can't identify %x, CoProc %x, CRd %x, Rn %x, imm8 %x, P %x, U %x, D %x, W %x\n",
DEBUG("Can't identify %x, CoProc %x, CRd %x, Rn %x, imm8 %x, P %x, U %x, D %x, W %x\n",
instr, CoProc, CRd, Rn, imm8, P, U, D, W);
return ARMul_CANT;
}
unsigned
VFPCDP (ARMul_State * state, unsigned type, ARMword instr)
VFPCDP (ARMul_State * state, unsigned type, u32 instr)
{
/* CDP<c> <coproc>,<opc1>,<CRd>,<CRn>,<CRm>,<opc2> */
int CoProc = BITS (8, 11); /* 10 or 11 */
@@ -230,15 +226,59 @@ VFPCDP (ARMul_State * state, unsigned type, ARMword instr)
int CRm = BITS (0, 3);
int OPC_2 = BITS (5, 7);
//ichfly
/*if ((instr & 0x0FBF0FD0) == 0x0EB70AC0) //vcvt.f64.f32 d8, s16 (s is bit 0-3 and LSB bit 22) (d is bit 12 - 15 MSB is Bit 6)
{
struct vfp_double vdd;
struct vfp_single vsd;
int dn = BITS(12, 15) + (BIT(22) << 4);
int sd = (BITS(0, 3) << 1) + BIT(5);
s32 n = vfp_get_float(state, sd);
vfp_single_unpack(&vsd, n);
if (vsd.exponent & 0x80)
{
vdd.exponent = (vsd.exponent&~0x80) | 0x400;
}
else
{
vdd.exponent = vsd.exponent | 0x380;
}
vdd.sign = vsd.sign;
vdd.significand = (u64)(vsd.significand & ~0xC0000000) << 32; // I have no idea why but the 2 uppern bits are not from the significand
vfp_put_double(state, vfp_double_pack(&vdd), dn);
return ARMul_DONE;
}
if ((instr & 0x0FBF0FD0) == 0x0EB70BC0) //vcvt.f32.f64 s15, d6
{
struct vfp_double vdd;
struct vfp_single vsd;
int sd = BITS(0, 3) + (BIT(5) << 4);
int dn = (BITS(12, 15) << 1) + BIT(22);
vfp_double_unpack(&vdd, vfp_get_double(state, sd));
if (vdd.exponent & 0x400) //todo if the exponent is to low or to high for this convert
{
vsd.exponent = (vdd.exponent) | 0x80;
}
else
{
vsd.exponent = vdd.exponent & ~0x80;
}
vsd.exponent &= 0xFF;
// vsd.exponent = vdd.exponent >> 3;
vsd.sign = vdd.sign;
vsd.significand = ((u64)(vdd.significand ) >> 32)& ~0xC0000000;
vfp_put_float(state, vfp_single_pack(&vsd), dn);
return ARMul_DONE;
}*/
/* TODO check access permission */
/* CRn/opc1 CRm/opc2 */
if (CoProc == 10 || CoProc == 11)
{
#define VFP_CDP_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_CDP_TRANS
if (CoProc == 10 || CoProc == 11) {
#define VFP_CDP_TRANS
#include "core/arm/skyeye_common/vfp/vfpinstr.cpp"
#undef VFP_CDP_TRANS
int exceptions = 0;
if (CoProc == 10)
@@ -250,7 +290,7 @@ VFPCDP (ARMul_State * state, unsigned type, ARMword instr)
return ARMul_DONE;
}
DEBUG_LOG(ARM11, "Can't identify %x\n", instr);
DEBUG("Can't identify %x\n", instr);
return ARMul_CANT;
}
@@ -301,13 +341,13 @@ VFPCDP (ARMul_State * state, unsigned type, ARMword instr)
/* Miscellaneous functions */
int32_t vfp_get_float(arm_core_t* state, unsigned int reg)
{
DBG("VFP get float: s%d=[%08x]\n", reg, state->ExtReg[reg]);
DEBUG("VFP get float: s%d=[%08x]\n", reg, state->ExtReg[reg]);
return state->ExtReg[reg];
}
void vfp_put_float(arm_core_t* state, int32_t val, unsigned int reg)
{
DBG("VFP put float: s%d <= [%08x]\n", reg, val);
DEBUG("VFP put float: s%d <= [%08x]\n", reg, val);
state->ExtReg[reg] = val;
}
@@ -315,13 +355,13 @@ uint64_t vfp_get_double(arm_core_t* state, unsigned int reg)
{
uint64_t result;
result = ((uint64_t) state->ExtReg[reg*2+1])<<32 | state->ExtReg[reg*2];
DBG("VFP get double: s[%d-%d]=[%016llx]\n", reg*2+1, reg*2, result);
DEBUG("VFP get double: s[%d-%d]=[%016llx]\n", reg*2+1, reg*2, result);
return result;
}
void vfp_put_double(arm_core_t* state, uint64_t val, unsigned int reg)
{
DBG("VFP put double: s[%d-%d] <= [%08x-%08x]\n", reg*2+1, reg*2, (uint32_t) (val>>32), (uint32_t) (val & 0xffffffff));
DEBUG("VFP put double: s[%d-%d] <= [%08x-%08x]\n", reg*2+1, reg*2, (uint32_t) (val>>32), (uint32_t) (val & 0xffffffff));
state->ExtReg[reg*2] = (uint32_t) (val & 0xffffffff);
state->ExtReg[reg*2+1] = (uint32_t) (val>>32);
}
@@ -338,7 +378,7 @@ void vfp_raise_exceptions(ARMul_State* state, u32 exceptions, u32 inst, u32 fpsc
vfpdebug("VFP: raising exceptions %08x\n", exceptions);
if (exceptions == VFP_EXCEPTION_ERROR) {
DEBUG_LOG(ARM11, "unhandled bounce %x\n", inst);
DEBUG("unhandled bounce %x\n", inst);
exit(-1);
return;
}

9
src/core/arm/skyeye_common/vfp/vfp_helper.h
View File

@@ -44,7 +44,7 @@
#define pr_info //printf
#define pr_debug //printf
static u32 vfp_fls(int x);
static u32 fls(int x);
#define do_div(n, base) {n/=base;}
/* From vfpinstr.h */
@@ -502,7 +502,7 @@ struct op {
u32 flags;
};
static u32 vfp_fls(int x)
static u32 fls(int x)
{
int r = 32;
@@ -532,4 +532,9 @@ static u32 vfp_fls(int x)
}
u32 vfp_double_normaliseroundintern(ARMul_State* state, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func);
u32 vfp_double_multiply(struct vfp_double *vdd, struct vfp_double *vdn, struct vfp_double *vdm, u32 fpscr);
u32 vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn, struct vfp_double *vdm, u32 fpscr);
u32 vfp_double_fcvtsinterncutting(ARMul_State* state, int sd, struct vfp_double* dm, u32 fpscr);
#endif

229
src/core/arm/skyeye_common/vfp/vfpdouble.cpp
View File

@@ -56,9 +56,9 @@
#include "core/arm/skyeye_common/vfp/asm_vfp.h"
static struct vfp_double vfp_double_default_qnan = {
//.exponent = 2047,
//.sign = 0,
//.significand = VFP_DOUBLE_SIGNIFICAND_QNAN,
2047,
0,
VFP_DOUBLE_SIGNIFICAND_QNAN,
};
static void vfp_double_dump(const char *str, struct vfp_double *d)
@@ -69,9 +69,9 @@ static void vfp_double_dump(const char *str, struct vfp_double *d)
static void vfp_double_normalise_denormal(struct vfp_double *vd)
{
int bits = 31 - vfp_fls(vd->significand >> 32);
int bits = 31 - fls((ARMword)(vd->significand >> 32));
if (bits == 31)
bits = 63 - vfp_fls(vd->significand);
bits = 63 - fls((ARMword)vd->significand);
vfp_double_dump("normalise_denormal: in", vd);
@@ -83,6 +83,134 @@ static void vfp_double_normalise_denormal(struct vfp_double *vd)
vfp_double_dump("normalise_denormal: out", vd);
}
u32 vfp_double_normaliseroundintern(ARMul_State* state, 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((ARMword)(significand >> 32));
if (shift == 32)
shift = 64 - fls((ARMword)significand);
if (shift) {
exponent -= shift;
significand <<= shift;
}
#if 1
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;
#if 1
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;
#if 1
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:
return 0;
}
u32 vfp_double_normaliseround(ARMul_State* state, int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func)
{
u64 significand, incr;
@@ -108,9 +236,9 @@ u32 vfp_double_normaliseround(ARMul_State* state, int dd, struct vfp_double *vd,
exponent = vd->exponent;
significand = vd->significand;
shift = 32 - vfp_fls(significand >> 32);
shift = 32 - fls((ARMword)(significand >> 32));
if (shift == 32)
shift = 64 - vfp_fls(significand);
shift = 64 - fls((ARMword)significand);
if (shift) {
exponent -= shift;
significand <<= shift;
@@ -204,7 +332,7 @@ u32 vfp_double_normaliseround(ARMul_State* state, int dd, struct vfp_double *vd,
vd->significand = significand >> 1;
}
pack:
pack:
vfp_double_dump("pack: final", vd);
{
s64 d = vfp_double_pack(vd);
@@ -287,7 +415,7 @@ static u32 vfp_double_fneg(ARMul_State* state, int dd, int unused, int dm, u32 f
static u32 vfp_double_fsqrt(ARMul_State* state, int dd, int unused, int dm, u32 fpscr)
{
pr_debug("In %s\n", __FUNCTION__);
struct vfp_double vdm, vdd, *vdp;
vfp_double vdm, vdd, *vdp;
int ret, tm;
vfp_double_unpack(&vdm, vfp_get_double(state, dm));
@@ -298,11 +426,11 @@ static u32 vfp_double_fsqrt(ARMul_State* state, int dd, int unused, int dm, u32
if (tm & VFP_NAN)
ret = vfp_propagate_nan(vdp, &vdm, NULL, fpscr);
else if (vdm.sign == 0) {
sqrt_copy:
sqrt_copy:
vdp = &vdm;
ret = 0;
} else {
sqrt_invalid:
sqrt_invalid:
vdp = &vfp_double_default_qnan;
ret = FPSCR_IOC;
}
@@ -464,6 +592,49 @@ static u32 vfp_double_fcmpez(ARMul_State* state, int dd, int unused, int dm, u32
return vfp_compare(state, dd, 1, VFP_REG_ZERO, fpscr);
}
u32 vfp_double_fcvtsinterncutting(ARMul_State* state, int sd, struct vfp_double* dm, u32 fpscr) //ichfly for internal use only
{
struct vfp_single vsd;
int tm;
u32 exceptions = 0;
pr_debug("In %s\n", __FUNCTION__);
tm = vfp_double_type(dm);
/*
* If we have a signalling NaN, signal invalid operation.
*/
if (tm == VFP_SNAN)
exceptions = FPSCR_IOC;
if (tm & VFP_DENORMAL)
vfp_double_normalise_denormal(dm);
vsd.sign = dm->sign;
vsd.significand = vfp_hi64to32jamming(dm->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_QNAN)
vsd.significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
goto pack_nan;
}
else if (tm & VFP_ZERO)
vsd.exponent = 0;
else
vsd.exponent = dm->exponent - (1023 - 127);
return vfp_single_normaliseround(state, sd, &vsd, fpscr, exceptions, "fcvts");
pack_nan:
vfp_put_float(state, vfp_single_pack(&vsd), sd);
return exceptions;
}
static u32 vfp_double_fcvts(ARMul_State* state, int sd, int unused, int dm, u32 fpscr)
{
struct vfp_double vdm;
@@ -503,7 +674,7 @@ static u32 vfp_double_fcvts(ARMul_State* state, int sd, int unused, int dm, u32
return vfp_single_normaliseround(state, sd, &vsd, fpscr, exceptions, "fcvts");
pack_nan:
pack_nan:
vfp_put_float(state, vfp_single_pack(&vsd), sd);
return exceptions;
}
@@ -564,7 +735,7 @@ static u32 vfp_double_ftoui(ARMul_State* state, int sd, int unused, int dm, u32
/*
* 2^0 <= m < 2^32-2^8
*/
d = (vdm.significand << 1) >> shift;
d = (ARMword)((vdm.significand << 1) >> shift);
rem = vdm.significand << (65 - shift);
if (rmode == FPSCR_ROUND_NEAREST) {
@@ -645,7 +816,7 @@ static u32 vfp_double_ftosi(ARMul_State* state, int sd, int unused, int dm, u32
int shift = 1023 + 63 - vdm.exponent; /* 58 */
u64 rem, incr = 0;
d = (vdm.significand << 1) >> shift;
d = (ARMword)((vdm.significand << 1) >> shift);
rem = vdm.significand << (65 - shift);
if (rmode == FPSCR_ROUND_NEAREST) {
@@ -660,8 +831,8 @@ static u32 vfp_double_ftosi(ARMul_State* state, int sd, int unused, int dm, u32
if ((rem + incr) < rem && d < 0xffffffff)
d += 1;
if (d > 0x7fffffff + (vdm.sign != 0)) {
d = 0x7fffffff + (vdm.sign != 0);
if (d > (0x7fffffff + (vdm.sign != 0))) {
d = (0x7fffffff + (vdm.sign != 0));
exceptions |= FPSCR_IOC;
} else if (rem)
exceptions |= FPSCR_IXC;
@@ -768,16 +939,14 @@ vfp_double_fadd_nonnumber(struct vfp_double *vdd, struct vfp_double *vdn,
return exceptions;
}
static u32
vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,
struct vfp_double *vdm, u32 fpscr)
u32 vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,struct vfp_double *vdm, u32 fpscr)
{
u32 exp_diff;
u64 m_sig;
if (vdn->significand & (1ULL << 63) ||
vdm->significand & (1ULL << 63)) {
pr_info("VFP: bad FP values\n");
pr_info("VFP: bad FP values in %s\n", __func__);
vfp_double_dump("VDN", vdn);
vfp_double_dump("VDM", vdm);
}
@@ -833,7 +1002,7 @@ vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,
return 0;
}
static u32
u32
vfp_double_multiply(struct vfp_double *vdd, struct vfp_double *vdn,
struct vfp_double *vdm, u32 fpscr)
{
@@ -895,7 +1064,7 @@ vfp_double_multiply(struct vfp_double *vdd, struct vfp_double *vdn,
#define NEG_SUBTRACT (1 << 1)
static u32
vfp_double_multiply_accumulate(ARMul_State* state, int dd, int dn, int dm, u32 fpscr, u32 negate, const char *func)
vfp_double_multiply_accumulate(ARMul_State* state, int dd, int dn, int dm, u32 fpscr, u32 negate, char *func)
{
struct vfp_double vdd, vdp, vdn, vdm;
u32 exceptions;
@@ -1136,29 +1305,29 @@ static u32 vfp_double_fdiv(ARMul_State* state, int dd, int dn, int dm, u32 fpscr
}
return vfp_double_normaliseround(state, dd, &vdd, fpscr, 0, "fdiv");
vdn_nan:
vdn_nan:
exceptions = vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr);
pack:
pack:
vfp_put_double(state, vfp_double_pack(&vdd), dd);
return exceptions;
vdm_nan:
vdm_nan:
exceptions = vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr);
goto pack;
zero:
zero:
vdd.exponent = 0;
vdd.significand = 0;
goto pack;
divzero:
divzero:
exceptions = FPSCR_DZC;
infinity:
infinity:
vdd.exponent = 2047;
vdd.significand = 0;
goto pack;
invalid:
invalid:
vfp_put_double(state, vfp_double_pack(&vfp_double_default_qnan), dd);
return FPSCR_IOC;
}
@@ -1258,6 +1427,6 @@ u32 vfp_double_cpdo(ARMul_State* state, u32 inst, u32 fpscr)
}
return exceptions;
invalid:
invalid:
return ~0;
}

128
src/core/arm/skyeye_common/vfp/vfpsingle.cpp
View File

@@ -56,9 +56,9 @@
#include "core/arm/skyeye_common/vfp/vfp.h"
static struct vfp_single vfp_single_default_qnan = {
//.exponent = 255,
//.sign = 0,
//.significand = VFP_SINGLE_SIGNIFICAND_QNAN,
255,
0,
VFP_SINGLE_SIGNIFICAND_QNAN,
};
static void vfp_single_dump(const char *str, struct vfp_single *s)
@@ -69,7 +69,7 @@ static void vfp_single_dump(const char *str, struct vfp_single *s)
static void vfp_single_normalise_denormal(struct vfp_single *vs)
{
int bits = 31 - vfp_fls(vs->significand);
int bits = 31 - fls(vs->significand);
vfp_single_dump("normalise_denormal: in", vs);
@@ -111,7 +111,7 @@ u32 vfp_single_normaliseround(ARMul_State* state, int sd, struct vfp_single *vs,
* bit 31, so we have VFP_SINGLE_LOW_BITS + 1 below the least
* significant bit.
*/
shift = 32 - vfp_fls(significand);
shift = 32 - fls(significand);
if (shift < 32 && shift) {
exponent -= shift;
significand <<= shift;
@@ -205,7 +205,7 @@ u32 vfp_single_normaliseround(ARMul_State* state, int sd, struct vfp_single *vs,
vs->significand = significand >> 1;
}
pack:
pack:
vfp_single_dump("pack: final", vs);
{
s32 d = vfp_single_pack(vs);
@@ -321,7 +321,7 @@ u32 vfp_estimate_sqrt_significand(u32 exponent, u32 significand)
{
u64 v = (u64)a << 31;
do_div(v, z);
return v + (z >> 1);
return (u32)(v + (z >> 1));
}
}
@@ -338,11 +338,11 @@ static u32 vfp_single_fsqrt(ARMul_State* state, int sd, int unused, s32 m, u32 f
if (tm & VFP_NAN)
ret = vfp_propagate_nan(vsp, &vsm, NULL, fpscr);
else if (vsm.sign == 0) {
sqrt_copy:
sqrt_copy:
vsp = &vsm;
ret = 0;
} else {
sqrt_invalid:
sqrt_invalid:
vsp = &vfp_single_default_qnan;
ret = FPSCR_IOC;
}
@@ -491,6 +491,47 @@ static u32 vfp_single_fcmpez(ARMul_State* state, int sd, int unused, s32 m, u32
return vfp_compare(state, sd, 1, 0, fpscr);
}
static s64 vfp_single_to_doubleintern(ARMul_State* state, s32 m, u32 fpscr) //ichfly for internal use only
{
struct vfp_single vsm;
struct vfp_double vdd;
int tm;
u32 exceptions = 0;
vfp_single_unpack(&vsm, m);
tm = vfp_single_type(&vsm);
/*
* If we have a signalling NaN, signal invalid operation.
*/
if (tm == VFP_SNAN)
exceptions = FPSCR_IOC;
if (tm & VFP_DENORMAL)
vfp_single_normalise_denormal(&vsm);
vdd.sign = vsm.sign;
vdd.significand = (u64)vsm.significand << 32;
/*
* If we have an infinity or NaN, the exponent must be 2047.
*/
if (tm & (VFP_INFINITY | VFP_NAN)) {
vdd.exponent = 2047;
if (tm == VFP_QNAN)
vdd.significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
goto pack_nan;
}
else if (tm & VFP_ZERO)
vdd.exponent = 0;
else
vdd.exponent = vsm.exponent + (1023 - 127);
pack_nan:
vfp_double_normaliseroundintern(state, &vdd, fpscr, exceptions, "fcvtd");
return vfp_double_pack(&vdd);
}
static u32 vfp_single_fcvtd(ARMul_State* state, int dd, int unused, s32 m, u32 fpscr)
{
struct vfp_single vsm;
@@ -529,7 +570,7 @@ static u32 vfp_single_fcvtd(ARMul_State* state, int dd, int unused, s32 m, u32 f
return vfp_double_normaliseround(state, dd, &vdd, fpscr, exceptions, "fcvtd");
pack_nan:
pack_nan:
vfp_put_double(state, vfp_double_pack(&vdd), dd);
return exceptions;
}
@@ -684,14 +725,14 @@ static u32 vfp_single_ftosi(ARMul_State* state, int sd, int unused, s32 m, u32 f
if ((rem + incr) < rem && d < 0xffffffff)
d += 1;
if (d > 0x7fffffff + (vsm.sign != 0)) {
d = 0x7fffffff + (vsm.sign != 0);
if (d > (0x7fffffffu + (vsm.sign != 0))) {
d = (0x7fffffffu + (vsm.sign != 0));
exceptions |= FPSCR_IOC;
} else if (rem)
exceptions |= FPSCR_IXC;
if (vsm.sign)
d = -d;
d = 0-d;
} else {
d = 0;
if (vsm.exponent | vsm.significand) {
@@ -800,7 +841,7 @@ vfp_single_add(struct vfp_single *vsd, struct vfp_single *vsn,
if (vsn->significand & 0x80000000 ||
vsm->significand & 0x80000000) {
pr_info("VFP: bad FP values\n");
pr_info("VFP: bad FP values in %s\n", __func__);
vfp_single_dump("VSN", vsn);
vfp_single_dump("VSM", vsm);
}
@@ -843,7 +884,7 @@ vfp_single_add(struct vfp_single *vsd, struct vfp_single *vsn,
m_sig = vsn->significand - m_sig;
if ((s32)m_sig < 0) {
vsd->sign = vfp_sign_negate(vsd->sign);
m_sig = -m_sig;
m_sig = 0-m_sig;
} else if (m_sig == 0) {
vsd->sign = (fpscr & FPSCR_RMODE_MASK) ==
FPSCR_ROUND_MINUSINF ? 0x8000 : 0;
@@ -917,12 +958,16 @@ vfp_single_multiply(struct vfp_single *vsd, struct vfp_single *vsn, struct vfp_s
#define NEG_SUBTRACT (1 << 1)
static u32
vfp_single_multiply_accumulate(ARMul_State* state, int sd, int sn, s32 m, u32 fpscr, u32 negate, const char *func)
vfp_single_multiply_accumulate(ARMul_State* state, int sd, int sn, s32 m, u32 fpscr, u32 negate, char *func)
{
{
struct vfp_single vsd, vsp, vsn, vsm;
u32 exceptions;
s32 v;
v = vfp_get_float(state, sn);
pr_debug("VFP: s%u = %08x\n", sn, v);
vfp_single_unpack(&vsn, v);
@@ -934,6 +979,7 @@ vfp_single_multiply_accumulate(ARMul_State* state, int sd, int sn, s32 m, u32 fp
vfp_single_normalise_denormal(&vsm);
exceptions = vfp_single_multiply(&vsp, &vsn, &vsm, fpscr);
if (negate & NEG_MULTIPLY)
vsp.sign = vfp_sign_negate(vsp.sign);
@@ -946,6 +992,38 @@ vfp_single_multiply_accumulate(ARMul_State* state, int sd, int sn, s32 m, u32 fp
exceptions |= vfp_single_add(&vsd, &vsn, &vsp, fpscr);
return vfp_single_normaliseround(state, sd, &vsd, fpscr, exceptions, func);
}
struct vfp_double vsd, vsp, vsn, vsm;
u32 exceptions;
s32 v;
s64 vd;
s64 md;
v = vfp_get_float(state, sn);
vd = vfp_single_to_doubleintern(state, v, fpscr);
vfp_double_unpack(&vsn, vd);
md = vfp_single_to_doubleintern(state, m, fpscr);
vfp_double_unpack(&vsm, md);
exceptions = vfp_double_multiply(&vsp, &vsn, &vsm, fpscr);
if (negate & NEG_MULTIPLY)
vsp.sign = vfp_sign_negate(vsp.sign);
v = vfp_get_float(state, sd);
vd = vfp_single_to_doubleintern(state, v, fpscr);
vfp_double_unpack(&vsn, vd);
if (negate & NEG_SUBTRACT)
vsn.sign = vfp_sign_negate(vsn.sign);
exceptions |= vfp_double_add(&vsd, &vsn, &vsp, fpscr);
s64 debug = vfp_double_pack(&vsd);
return vfp_double_fcvtsinterncutting(state, sd, &vsd, fpscr);
}
/*
@@ -1148,36 +1226,36 @@ static u32 vfp_single_fdiv(ARMul_State* state, int sd, int sn, s32 m, u32 fpscr)
{
u64 significand = (u64)vsn.significand << 32;
do_div(significand, vsm.significand);
vsd.significand = significand;
vsd.significand = (u32)significand;
}
if ((vsd.significand & 0x3f) == 0)
vsd.significand |= ((u64)vsm.significand * vsd.significand != (u64)vsn.significand << 32);
return vfp_single_normaliseround(state, sd, &vsd, fpscr, 0, "fdiv");
vsn_nan:
vsn_nan:
exceptions = vfp_propagate_nan(&vsd, &vsn, &vsm, fpscr);
pack:
pack:
vfp_put_float(state, vfp_single_pack(&vsd), sd);
return exceptions;
vsm_nan:
vsm_nan:
exceptions = vfp_propagate_nan(&vsd, &vsm, &vsn, fpscr);
goto pack;
zero:
zero:
vsd.exponent = 0;
vsd.significand = 0;
goto pack;
divzero:
divzero:
exceptions = FPSCR_DZC;
infinity:
infinity:
vsd.exponent = 255;
vsd.significand = 0;
goto pack;
invalid:
invalid:
vfp_put_float(state, vfp_single_pack(&vfp_single_default_qnan), sd);
return FPSCR_IOC;
}
@@ -1273,6 +1351,6 @@ u32 vfp_single_cpdo(ARMul_State* state, u32 inst, u32 fpscr)
}
return exceptions;
invalid:
invalid:
return (u32)-1;
}