diff options
author | Joseph Myers <joseph@codesourcery.com> | 2020-06-17 23:20:50 +0000 |
---|---|---|
committer | Paolo Bonzini <pbonzini@redhat.com> | 2020-06-26 09:39:38 -0400 |
commit | 5eebc49d2d0aa5fc7e90eeac97533051bb7b72fa (patch) | |
tree | d35769aeb97cc808b4aec74170d8bf8b3e3fbb62 /target | |
parent | 5ef396e2ba865f34a4766dbd60c739fb4bcb4fcc (diff) | |
download | qemu-5eebc49d2d0aa5fc7e90eeac97533051bb7b72fa.zip qemu-5eebc49d2d0aa5fc7e90eeac97533051bb7b72fa.tar.gz qemu-5eebc49d2d0aa5fc7e90eeac97533051bb7b72fa.tar.bz2 |
target/i386: reimplement fyl2xp1 using floatx80 operations
The x87 fyl2xp1 emulation is currently based around conversion to
double. This is inherently unsuitable for a good emulation of any
floatx80 operation, even before considering that it is a particularly
naive implementation using double (adding 1 then using log rather than
attempting a better emulation using log1p).
Reimplement using the soft-float operations, as was done for f2xm1; as
in that case, m68k has related operations but not exactly this one and
it seemed safest to implement directly rather than reusing the m68k
code to avoid accumulation of errors.
A test is included with many randomly generated inputs. The
assumption of the test is that the result in round-to-nearest mode
should always be one of the two closest floating-point numbers to the
mathematical value of y * log2(x + 1); the implementation aims to do
somewhat better than that (about 70 correct bits before rounding). I
haven't investigated how accurate hardware is.
Intel manuals describe a narrower range of valid arguments to this
instruction than AMD manuals. The implementation accepts the wider
range (it's needed anyway for the core code to be reusable in a
subsequent patch reimplementing fyl2x), but the test only has inputs
in the narrower range so that it's valid on hardware that may reject
or produce poor results for inputs outside that range.
Code in the previous implementation that sets C2 for some out-of-range
arguments is not carried forward to the new implementation; C2 is
undefined for this instruction and I suspect that code was just
cut-and-pasted from the trigonometric instructions (fcos, fptan, fsin,
fsincos) where C2 *is* defined to be set for out-of-range arguments.
Signed-off-by: Joseph Myers <joseph@codesourcery.com>
Message-Id: <alpine.DEB.2.21.2006172320190.20587@digraph.polyomino.org.uk>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Diffstat (limited to 'target')
-rw-r--r-- | target/i386/fpu_helper.c | 211 |
1 files changed, 202 insertions, 9 deletions
diff --git a/target/i386/fpu_helper.c b/target/i386/fpu_helper.c index 8f34ea9..884e84c 100644 --- a/target/i386/fpu_helper.c +++ b/target/i386/fpu_helper.c @@ -1373,19 +1373,212 @@ void helper_fprem(CPUX86State *env) helper_fprem_common(env, true); } +/* 128-bit significand of log2(e). */ +#define log2_e_sig_high 0xb8aa3b295c17f0bbULL +#define log2_e_sig_low 0xbe87fed0691d3e89ULL + +/* + * Polynomial coefficients for an approximation to log2((1+x)/(1-x)), + * with only odd powers of x used, for x in the interval [2*sqrt(2)-3, + * 3-2*sqrt(2)], which corresponds to logarithms of numbers in the + * interval [sqrt(2)/2, sqrt(2)]. + */ +#define fyl2x_coeff_0 make_floatx80(0x4000, 0xb8aa3b295c17f0bcULL) +#define fyl2x_coeff_0_low make_floatx80(0xbfbf, 0x834972fe2d7bab1bULL) +#define fyl2x_coeff_1 make_floatx80(0x3ffe, 0xf6384ee1d01febb8ULL) +#define fyl2x_coeff_2 make_floatx80(0x3ffe, 0x93bb62877cdfa2e3ULL) +#define fyl2x_coeff_3 make_floatx80(0x3ffd, 0xd30bb153d808f269ULL) +#define fyl2x_coeff_4 make_floatx80(0x3ffd, 0xa42589eaf451499eULL) +#define fyl2x_coeff_5 make_floatx80(0x3ffd, 0x864d42c0f8f17517ULL) +#define fyl2x_coeff_6 make_floatx80(0x3ffc, 0xe3476578adf26272ULL) +#define fyl2x_coeff_7 make_floatx80(0x3ffc, 0xc506c5f874e6d80fULL) +#define fyl2x_coeff_8 make_floatx80(0x3ffc, 0xac5cf50cc57d6372ULL) +#define fyl2x_coeff_9 make_floatx80(0x3ffc, 0xb1ed0066d971a103ULL) + void helper_fyl2xp1(CPUX86State *env) { - double fptemp = floatx80_to_double(env, ST0); - - if ((fptemp + 1.0) > 0.0) { - fptemp = log(fptemp + 1.0) / log(2.0); /* log2(ST + 1.0) */ - fptemp *= floatx80_to_double(env, ST1); - ST1 = double_to_floatx80(env, fptemp); - fpop(env); + uint8_t old_flags = save_exception_flags(env); + uint64_t arg0_sig = extractFloatx80Frac(ST0); + int32_t arg0_exp = extractFloatx80Exp(ST0); + bool arg0_sign = extractFloatx80Sign(ST0); + uint64_t arg1_sig = extractFloatx80Frac(ST1); + int32_t arg1_exp = extractFloatx80Exp(ST1); + bool arg1_sign = extractFloatx80Sign(ST1); + + if (floatx80_is_signaling_nan(ST0, &env->fp_status)) { + float_raise(float_flag_invalid, &env->fp_status); + ST1 = floatx80_silence_nan(ST0, &env->fp_status); + } else if (floatx80_is_signaling_nan(ST1, &env->fp_status)) { + float_raise(float_flag_invalid, &env->fp_status); + ST1 = floatx80_silence_nan(ST1, &env->fp_status); + } else if (floatx80_invalid_encoding(ST0) || + floatx80_invalid_encoding(ST1)) { + float_raise(float_flag_invalid, &env->fp_status); + ST1 = floatx80_default_nan(&env->fp_status); + } else if (floatx80_is_any_nan(ST0)) { + ST1 = ST0; + } else if (floatx80_is_any_nan(ST1)) { + /* Pass this NaN through. */ + } else if (arg0_exp > 0x3ffd || + (arg0_exp == 0x3ffd && arg0_sig > (arg0_sign ? + 0x95f619980c4336f7ULL : + 0xd413cccfe7799211ULL))) { + /* + * Out of range for the instruction (ST0 must have absolute + * value less than 1 - sqrt(2)/2 = 0.292..., according to + * Intel manuals; AMD manuals allow a range from sqrt(2)/2 - 1 + * to sqrt(2) - 1, which we allow here), treat as invalid. + */ + float_raise(float_flag_invalid, &env->fp_status); + ST1 = floatx80_default_nan(&env->fp_status); + } else if (floatx80_is_zero(ST0) || floatx80_is_zero(ST1) || + arg1_exp == 0x7fff) { + /* + * One argument is zero, or multiplying by infinity; correct + * result is exact and can be obtained by multiplying the + * arguments. + */ + ST1 = floatx80_mul(ST0, ST1, &env->fp_status); + } else if (arg0_exp < 0x3fb0) { + /* + * Multiplying both arguments and an extra-precision version + * of log2(e) is sufficiently precise. + */ + uint64_t sig0, sig1, sig2; + int32_t exp; + if (arg0_exp == 0) { + normalizeFloatx80Subnormal(arg0_sig, &arg0_exp, &arg0_sig); + } + if (arg1_exp == 0) { + normalizeFloatx80Subnormal(arg1_sig, &arg1_exp, &arg1_sig); + } + mul128By64To192(log2_e_sig_high, log2_e_sig_low, arg0_sig, + &sig0, &sig1, &sig2); + exp = arg0_exp + 1; + mul128By64To192(sig0, sig1, arg1_sig, &sig0, &sig1, &sig2); + exp += arg1_exp - 0x3ffe; + /* This result is inexact. */ + sig1 |= 1; + ST1 = normalizeRoundAndPackFloatx80(80, arg0_sign ^ arg1_sign, exp, + sig0, sig1, &env->fp_status); } else { - env->fpus &= ~0x4700; - env->fpus |= 0x400; + bool asign; + uint32_t dexp, texp, aexp; + uint64_t dsig0, dsig1, tsig0, tsig1, rsig0, rsig1, rsig2; + uint64_t msig0, msig1, msig2, t2sig0, t2sig1, t2sig2, t2sig3; + uint64_t asig0, asig1, asig2, asig3, bsig0, bsig1; + floatx80 t2, accum; + FloatRoundMode save_mode = env->fp_status.float_rounding_mode; + signed char save_prec = env->fp_status.floatx80_rounding_precision; + env->fp_status.float_rounding_mode = float_round_nearest_even; + env->fp_status.floatx80_rounding_precision = 80; + + /* + * Compute an approximation of ST0/(2+ST0), with extra + * precision, as the argument to a polynomial approximation. + * The extra precision is only needed for the first term of + * the approximation, with subsequent terms being + * significantly smaller; the approximation only uses odd + * exponents, and the square of ST0/(2+ST0) is at most + * 17-12*sqrt(2) = 0.029.... + */ + if (arg0_sign) { + dexp = 0x3fff; + shift128RightJamming(arg0_sig, 0, dexp - arg0_exp, &dsig0, &dsig1); + sub128(0, 0, dsig0, dsig1, &dsig0, &dsig1); + } else { + dexp = 0x4000; + shift128RightJamming(arg0_sig, 0, dexp - arg0_exp, &dsig0, &dsig1); + dsig0 |= 0x8000000000000000ULL; + } + texp = arg0_exp - dexp + 0x3ffe; + rsig0 = arg0_sig; + rsig1 = 0; + rsig2 = 0; + if (dsig0 <= rsig0) { + shift128Right(rsig0, rsig1, 1, &rsig0, &rsig1); + ++texp; + } + tsig0 = estimateDiv128To64(rsig0, rsig1, dsig0); + mul128By64To192(dsig0, dsig1, tsig0, &msig0, &msig1, &msig2); + sub192(rsig0, rsig1, rsig2, msig0, msig1, msig2, + &rsig0, &rsig1, &rsig2); + while ((int64_t) rsig0 < 0) { + --tsig0; + add192(rsig0, rsig1, rsig2, 0, dsig0, dsig1, + &rsig0, &rsig1, &rsig2); + } + tsig1 = estimateDiv128To64(rsig1, rsig2, dsig0); + /* + * No need to correct any estimation error in tsig1; even with + * such error, it is accurate enough. Now compute the square + * of that approximation. + */ + mul128To256(tsig0, tsig1, tsig0, tsig1, + &t2sig0, &t2sig1, &t2sig2, &t2sig3); + t2 = normalizeRoundAndPackFloatx80(80, false, texp + texp - 0x3ffe, + t2sig0, t2sig1, &env->fp_status); + + /* Compute the lower parts of the polynomial expansion. */ + accum = floatx80_mul(fyl2x_coeff_9, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_8, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_7, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_6, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_5, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_4, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_3, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_2, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_1, accum, &env->fp_status); + accum = floatx80_mul(accum, t2, &env->fp_status); + accum = floatx80_add(fyl2x_coeff_0_low, accum, &env->fp_status); + + /* + * The full polynomial expansion is fyl2x_coeff_0 + accum + * (where accum has much lower magnitude, and so, in + * particular, carry out of the addition is not possible), + * multiplied by t. (This expansion is only accurate to about + * 70 bits, not 128 bits.) + */ + aexp = extractFloatx80Exp(fyl2x_coeff_0); + asign = extractFloatx80Sign(fyl2x_coeff_0); + shift128RightJamming(extractFloatx80Frac(accum), 0, + aexp - extractFloatx80Exp(accum), + &asig0, &asig1); + bsig0 = extractFloatx80Frac(fyl2x_coeff_0); + bsig1 = 0; + if (asign == extractFloatx80Sign(accum)) { + add128(bsig0, bsig1, asig0, asig1, &asig0, &asig1); + } else { + sub128(bsig0, bsig1, asig0, asig1, &asig0, &asig1); + } + /* + * Multiply by t and by the second argument to compute the + * required result. + */ + mul128To256(asig0, asig1, tsig0, tsig1, + &asig0, &asig1, &asig2, &asig3); + aexp += texp - 0x3ffe; + if (arg1_exp == 0) { + normalizeFloatx80Subnormal(arg1_sig, &arg1_exp, &arg1_sig); + } + mul128By64To192(asig0, asig1, arg1_sig, &asig0, &asig1, &asig2); + aexp += arg1_exp - 0x3ffe; + /* This result is inexact. */ + asig1 |= 1; + env->fp_status.float_rounding_mode = save_mode; + ST1 = normalizeRoundAndPackFloatx80(80, arg0_sign ^ arg1_sign, aexp, + asig0, asig1, &env->fp_status); + env->fp_status.floatx80_rounding_precision = save_prec; } + fpop(env); + merge_exception_flags(env, old_flags); } void helper_fsqrt(CPUX86State *env) |