mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2025-04-15 13:28:47 +00:00
Code Activity

FPA: precision bugfixes for FMA

Browse source 
Signed-off-by: Christoph M. Wintersteiger <cwinter@microsoft.com>
This commit is contained in:
Christoph M. Wintersteiger 2013-06-27 16:08:25 +01:00
parent 0d2a7f922c
commit 42b3a81ef6
1 changed files with 32 additions and 23 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

55
src/tactic/fpa/fpa2bv_converter.cpp
View file

@ -1134,6 +1134,9 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
mk_is_neg(z, z_is_neg); mk_is_neg(z, z_is_neg);
mk_is_inf(z, z_is_inf); mk_is_inf(z, z_is_inf);
expr_ref rm_is_to_neg(m);
mk_is_rm(rm, BV_RM_TO_NEGATIVE, rm_is_to_neg);
dbg_decouple("fpa2bv_fma_x_is_nan", x_is_nan); dbg_decouple("fpa2bv_fma_x_is_nan", x_is_nan);
dbg_decouple("fpa2bv_fma_x_is_zero", x_is_zero); dbg_decouple("fpa2bv_fma_x_is_zero", x_is_zero);
dbg_decouple("fpa2bv_fma_x_is_pos", x_is_pos); dbg_decouple("fpa2bv_fma_x_is_pos", x_is_pos);
@ -1187,31 +1190,28 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
m_simp.mk_or(x_is_zero, inf_cond, inf_or); m_simp.mk_or(x_is_zero, inf_cond, inf_or);
mk_ite(inf_or, nan, neg_x_sgn_inf, v5); mk_ite(inf_or, nan, neg_x_sgn_inf, v5);
// z is +-INF -> Z. // z is +-INF -> z.
mk_is_inf(z, c6); mk_is_inf(z, c6);
v6 = z; v6 = z;
// (x is 0) || (y is 0) -> x but with sign = x.sign ^ y.sign // (x is 0) || (y is 0) -> z
m_simp.mk_or(x_is_zero, y_is_zero, c7); m_simp.mk_or(x_is_zero, y_is_zero, c7);
expr_ref sign_xor(m); expr_ref ite_c(m);
m_simp.mk_xor(x_is_pos, y_is_pos, sign_xor); m_simp.mk_and(z_is_zero, m.mk_not(rm_is_to_neg), ite_c);
mk_ite(sign_xor, nzero, pzero, v7); mk_ite(ite_c, pzero, z, v7);
// else comes the fused multiplication. // else comes the fused multiplication.
unsigned ebits = m_util.get_ebits(f->get_range()); unsigned ebits = m_util.get_ebits(f->get_range());
unsigned sbits = m_util.get_sbits(f->get_range()); unsigned sbits = m_util.get_sbits(f->get_range());
SASSERT(ebits <= sbits); SASSERT(ebits <= sbits);
expr_ref rm_is_to_neg(m);
mk_is_rm(rm, BV_RM_TO_NEGATIVE, rm_is_to_neg);
expr_ref a_sgn(m), a_sig(m), a_exp(m), a_lz(m); expr_ref a_sgn(m), a_sig(m), a_exp(m), a_lz(m);
expr_ref b_sgn(m), b_sig(m), b_exp(m), b_lz(m); expr_ref b_sgn(m), b_sig(m), b_exp(m), b_lz(m);
expr_ref c_sgn(m), c_sig(m), c_exp(m), c_lz(m); expr_ref c_sgn(m), c_sig(m), c_exp(m), c_lz(m);
unpack(x, a_sgn, a_sig, a_exp, a_lz, true); unpack(x, a_sgn, a_sig, a_exp, a_lz, true);
unpack(y, b_sgn, b_sig, b_exp, b_lz, true); unpack(y, b_sgn, b_sig, b_exp, b_lz, true);
unpack(z, c_sgn, c_sig, c_exp, c_lz, true); unpack(z, c_sgn, c_sig, c_exp, c_lz, true);
expr_ref a_lz_ext(m), b_lz_ext(m), c_lz_ext(m); expr_ref a_lz_ext(m), b_lz_ext(m), c_lz_ext(m);
a_lz_ext = m_bv_util.mk_zero_extend(2, a_lz); a_lz_ext = m_bv_util.mk_zero_extend(2, a_lz);
@ -1220,12 +1220,12 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
expr_ref a_sig_ext(m), b_sig_ext(m); expr_ref a_sig_ext(m), b_sig_ext(m);
a_sig_ext = m_bv_util.mk_zero_extend(sbits, a_sig); a_sig_ext = m_bv_util.mk_zero_extend(sbits, a_sig);
b_sig_ext = m_bv_util.mk_zero_extend(sbits, b_sig); b_sig_ext = m_bv_util.mk_zero_extend(sbits, b_sig);
expr_ref a_exp_ext(m), b_exp_ext(m), c_exp_ext(m); expr_ref a_exp_ext(m), b_exp_ext(m), c_exp_ext(m);
a_exp_ext = m_bv_util.mk_sign_extend(2, a_exp); a_exp_ext = m_bv_util.mk_sign_extend(2, a_exp);
b_exp_ext = m_bv_util.mk_sign_extend(2, b_exp); b_exp_ext = m_bv_util.mk_sign_extend(2, b_exp);
c_exp_ext = m_bv_util.mk_sign_extend(2, c_exp); c_exp_ext = m_bv_util.mk_sign_extend(2, c_exp);
dbg_decouple("fpa2bv_fma_a_sig", a_sig_ext); dbg_decouple("fpa2bv_fma_a_sig", a_sig_ext);
dbg_decouple("fpa2bv_fma_b_sig", b_sig_ext); dbg_decouple("fpa2bv_fma_b_sig", b_sig_ext);
@ -1233,6 +1233,9 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
dbg_decouple("fpa2bv_fma_a_exp", a_exp_ext); dbg_decouple("fpa2bv_fma_a_exp", a_exp_ext);
dbg_decouple("fpa2bv_fma_b_exp", b_exp_ext); dbg_decouple("fpa2bv_fma_b_exp", b_exp_ext);
dbg_decouple("fpa2bv_fma_c_exp", c_exp_ext); dbg_decouple("fpa2bv_fma_c_exp", c_exp_ext);
dbg_decouple("fpa2bv_fma_a_lz", a_lz_ext);
dbg_decouple("fpa2bv_fma_b_lz", b_lz_ext);
dbg_decouple("fpa2bv_fma_c_lz", c_lz_ext);
expr_ref mul_sgn(m), mul_sig(m), mul_exp(m); expr_ref mul_sgn(m), mul_sig(m), mul_exp(m);
expr * signs[2] = { a_sgn, b_sgn }; expr * signs[2] = { a_sgn, b_sgn };
@ -1254,6 +1257,7 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
// Extend c // Extend c
c_sig = m_bv_util.mk_zero_extend(1, m_bv_util.mk_concat(c_sig, m_bv_util.mk_numeral(0, sbits-1))); c_sig = m_bv_util.mk_zero_extend(1, m_bv_util.mk_concat(c_sig, m_bv_util.mk_numeral(0, sbits-1)));
c_exp_ext = m_bv_util.mk_bv_sub(c_exp_ext, c_lz_ext);
SASSERT(m_bv_util.get_bv_size(mul_sig) == 2 * sbits); SASSERT(m_bv_util.get_bv_size(mul_sig) == 2 * sbits);
SASSERT(m_bv_util.get_bv_size(c_sig) == 2 * sbits); SASSERT(m_bv_util.get_bv_size(c_sig) == 2 * sbits);
@ -1269,7 +1273,7 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
m_simp.mk_ite(swap_cond, c_exp_ext, mul_exp, e_exp); // has ebits + 2 m_simp.mk_ite(swap_cond, c_exp_ext, mul_exp, e_exp); // has ebits + 2
m_simp.mk_ite(swap_cond, mul_sgn, c_sgn, f_sgn); m_simp.mk_ite(swap_cond, mul_sgn, c_sgn, f_sgn);
m_simp.mk_ite(swap_cond, mul_sig, c_sig, f_sig); // has 2 * sbits m_simp.mk_ite(swap_cond, mul_sig, c_sig, f_sig); // has 2 * sbits
m_simp.mk_ite(swap_cond, mul_exp, c_exp_ext, f_exp); // has ebits + 2 m_simp.mk_ite(swap_cond, mul_exp, c_exp_ext, f_exp); // has ebits + 2
SASSERT(is_well_sorted(m, e_sgn)); SASSERT(is_well_sorted(m, e_sgn));
SASSERT(is_well_sorted(m, e_sig)); SASSERT(is_well_sorted(m, e_sig));
@ -1303,7 +1307,7 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
shifted_big = m_bv_util.mk_bv_lshr( shifted_big = m_bv_util.mk_bv_lshr(
m_bv_util.mk_concat(f_sig, m_bv_util.mk_numeral(0, sbits)), m_bv_util.mk_concat(f_sig, m_bv_util.mk_numeral(0, sbits)),
m_bv_util.mk_zero_extend((3*sbits)-(ebits+2), exp_delta)); m_bv_util.mk_zero_extend((3*sbits)-(ebits+2), exp_delta));
shifted_f_sig = m_bv_util.mk_zero_extend(sbits, m_bv_util.mk_extract(3*sbits-1, 2*sbits, shifted_big)); shifted_f_sig = m_bv_util.mk_extract(3*sbits-1, sbits, shifted_big);
sticky_raw = m_bv_util.mk_extract(sbits-1, 0, shifted_big); sticky_raw = m_bv_util.mk_extract(sbits-1, 0, shifted_big);
SASSERT(m_bv_util.get_bv_size(shifted_f_sig) == 2 * sbits); SASSERT(m_bv_util.get_bv_size(shifted_f_sig) == 2 * sbits);
SASSERT(is_well_sorted(m, shifted_f_sig)); SASSERT(is_well_sorted(m, shifted_f_sig));
@ -1334,7 +1338,7 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
expr_ref sum(m); expr_ref sum(m);
m_simp.mk_ite(eq_sgn, m_simp.mk_ite(eq_sgn,
m_bv_util.mk_bv_add(e_sig, shifted_f_sig), // ADD LZ m_bv_util.mk_bv_add(e_sig, shifted_f_sig),
m_bv_util.mk_bv_sub(e_sig, shifted_f_sig), m_bv_util.mk_bv_sub(e_sig, shifted_f_sig),
sum); sum);
@ -1354,7 +1358,7 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
dbg_decouple("fpa2bv_fma_add_n_sum", n_sum); dbg_decouple("fpa2bv_fma_add_n_sum", n_sum);
dbg_decouple("fpa2bv_fma_add_sig_abs", sig_abs); dbg_decouple("fpa2bv_fma_add_sig_abs", sig_abs);
res_exp = m_bv_util.mk_bv_sub(e_exp, c_lz_ext); res_exp = e_exp;
// Result could overflow into 4.xxx ... // Result could overflow into 4.xxx ...
@ -1371,8 +1375,11 @@ void fpa2bv_converter::mk_fusedma(func_decl * f, unsigned num, expr * const * ar
res_sgn = m_bv_util.mk_bv_or(3, res_sgn_or_args); res_sgn = m_bv_util.mk_bv_or(3, res_sgn_or_args);
sticky_raw = m_bv_util.mk_extract(sbits-5, 0, sig_abs); sticky_raw = m_bv_util.mk_extract(sbits-5, 0, sig_abs);
sticky = m_bv_util.mk_zero_extend(sbits+7, m.mk_app(bvfid, OP_BREDOR, sticky_raw.get())); sticky = m_bv_util.mk_zero_extend(sbits+3, m.mk_app(bvfid, OP_BREDOR, sticky_raw.get()));
res_sig = m_bv_util.mk_extract(2*sbits-1, sbits-4, sig_abs); dbg_decouple("fpa2bv_fma_add_sum_sticky", sticky);
expr * res_or_args[2] = { m_bv_util.mk_extract(2*sbits-1, sbits-4, sig_abs), sticky };
res_sig = m_bv_util.mk_bv_or(2, res_or_args);
SASSERT(m_bv_util.get_bv_size(res_sig) == sbits+4); SASSERT(m_bv_util.get_bv_size(res_sig) == sbits+4);
expr_ref is_zero_sig(m), nil_sbits4(m); expr_ref is_zero_sig(m), nil_sbits4(m);
@ -2264,14 +2271,16 @@ void fpa2bv_converter::unpack(expr * e, expr_ref & sgn, expr_ref & sig, expr_ref
zero_e = m_bv_util.mk_numeral(0, ebits); zero_e = m_bv_util.mk_numeral(0, ebits);
if (normalize) { if (normalize) {
expr_ref is_sig_zero(m), zero_s(m);
zero_s = m_bv_util.mk_numeral(0, sbits);
m_simp.mk_eq(zero_s, denormal_sig, is_sig_zero);
expr_ref lz_d(m); expr_ref lz_d(m);
mk_leading_zeros(denormal_sig, ebits, lz_d); mk_leading_zeros(denormal_sig, ebits, lz_d);
m_simp.mk_ite(is_normal, zero_e, lz_d, lz); m_simp.mk_ite(m.mk_or(is_normal, is_sig_zero), zero_e, lz_d, lz);
dbg_decouple("fpa2bv_unpack_lz", lz); dbg_decouple("fpa2bv_unpack_lz", lz);
expr_ref is_sig_zero(m), shift(m), zero_s(m); expr_ref shift(m);
zero_s = m_bv_util.mk_numeral(0, sbits);
m_simp.mk_eq(zero_s, denormal_sig, is_sig_zero);
m_simp.mk_ite(is_sig_zero, zero_e, lz, shift); m_simp.mk_ite(is_sig_zero, zero_e, lz, shift);
dbg_decouple("fpa2bv_unpack_shift", shift); dbg_decouple("fpa2bv_unpack_shift", shift);
SASSERT(is_well_sorted(m, is_sig_zero)); SASSERT(is_well_sorted(m, is_sig_zero));