From f1c0ac72e79ce6e16c383fe05d877f0b4c9e7b5b Mon Sep 17 00:00:00 2001 From: "Christoph M. Wintersteiger" Date: Mon, 24 Jul 2017 20:54:29 +0100 Subject: [PATCH] Fix for fp.fma encoding. Relates to #872. --- src/ast/fpa/fpa2bv_converter.cpp | 117 ++++++++++++++++++------------- src/util/mpf.cpp | 2 +- 2 files changed, 70 insertions(+), 49 deletions(-) diff --git a/src/ast/fpa/fpa2bv_converter.cpp b/src/ast/fpa/fpa2bv_converter.cpp index 99b8c5ac8..5b71b2c2d 100644 --- a/src/ast/fpa/fpa2bv_converter.cpp +++ b/src/ast/fpa/fpa2bv_converter.cpp @@ -1516,7 +1516,8 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, mul_sgn = m_bv_util.mk_bv_xor(2, signs); dbg_decouple("fpa2bv_fma_mul_sgn", mul_sgn); - mul_exp = m_bv_util.mk_bv_add(m_bv_util.mk_bv_sub(a_exp_ext, a_lz_ext), + mul_exp = m_bv_util.mk_bv_add( + m_bv_util.mk_bv_sub(a_exp_ext, a_lz_ext), m_bv_util.mk_bv_sub(b_exp_ext, b_lz_ext)); dbg_decouple("fpa2bv_fma_mul_exp", mul_exp); @@ -1529,11 +1530,14 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, // The product has the form [-1][0].[2*sbits - 2]. // 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))); + expr_ref c_sig_ext(m); + c_sig_ext = 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(c_sig) == 2 * sbits); + SASSERT(m_bv_util.get_bv_size(c_sig_ext) == 2 * sbits); + dbg_decouple("fpa2bv_fma_c_sig_ext", c_sig_ext); + dbg_decouple("fpa2bv_fma_c_exp_ext", c_exp_ext); expr_ref swap_cond(m); swap_cond = m_bv_util.mk_sle(mul_exp, c_exp_ext); @@ -1542,10 +1546,10 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, expr_ref e_sgn(m), e_sig(m), e_exp(m), f_sgn(m), f_sig(m), f_exp(m); m_simp.mk_ite(swap_cond, c_sgn, mul_sgn, e_sgn); - m_simp.mk_ite(swap_cond, c_sig, mul_sig, e_sig); // has 2 * sbits + m_simp.mk_ite(swap_cond, c_sig_ext, mul_sig, e_sig); // has 2 * sbits 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_sig, c_sig, f_sig); // has 2 * sbits + m_simp.mk_ite(swap_cond, mul_sig, c_sig_ext, f_sig); // has 2 * sbits m_simp.mk_ite(swap_cond, mul_exp, c_exp_ext, f_exp); // has ebits + 2 SASSERT(is_well_sorted(m, e_sgn)); @@ -1560,6 +1564,11 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, SASSERT(m_bv_util.get_bv_size(e_exp) == ebits + 2); SASSERT(m_bv_util.get_bv_size(f_exp) == ebits + 2); + dbg_decouple("fpa2bv_fma_e_sig", e_sig); + dbg_decouple("fpa2bv_fma_e_exp", e_exp); + dbg_decouple("fpa2bv_fma_f_sig", f_sig); + dbg_decouple("fpa2bv_fma_f_exp", f_exp); + expr_ref res_sgn(m), res_sig(m), res_exp(m); expr_ref exp_delta(m); @@ -1582,6 +1591,7 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, m_bv_util.mk_zero_extend((3*sbits)-(ebits+2), exp_delta)); 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); + dbg_decouple("fpa2bv_fma_shifted_f_sig", shifted_f_sig); SASSERT(m_bv_util.get_bv_size(shifted_f_sig) == 2 * sbits); SASSERT(is_well_sorted(m, shifted_f_sig)); @@ -1594,14 +1604,16 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, expr * or_args[2] = { shifted_f_sig, sticky }; shifted_f_sig = m_bv_util.mk_bv_or(2, or_args); SASSERT(is_well_sorted(m, shifted_f_sig)); + dbg_decouple("fpa2bv_fma_f_sig_or_sticky", shifted_f_sig); // Significant addition. - // Two extra bits for catching the overflow. + // Two extra bits for the sign and for catching overflows. e_sig = m_bv_util.mk_zero_extend(2, e_sig); shifted_f_sig = m_bv_util.mk_zero_extend(2, shifted_f_sig); expr_ref eq_sgn(m); m_simp.mk_eq(e_sgn, f_sgn, eq_sgn); + dbg_decouple("fpa2bv_fma_eq_sgn", eq_sgn); SASSERT(m_bv_util.get_bv_size(e_sig) == 2*sbits + 2); SASSERT(m_bv_util.get_bv_size(shifted_f_sig) == 2*sbits + 2); @@ -1611,27 +1623,24 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, expr_ref sum(m), e_plus_f(m), e_minus_f(m); e_plus_f = m_bv_util.mk_bv_add(e_sig, shifted_f_sig); - e_minus_f = m_bv_util.mk_bv_sub(e_sig, shifted_f_sig), - m_simp.mk_ite(eq_sgn, e_plus_f, e_minus_f, sum); - + e_minus_f = m_bv_util.mk_bv_sub(e_sig, shifted_f_sig); + m_simp.mk_ite(eq_sgn, e_plus_f, e_minus_f, sum); + SASSERT(m_bv_util.get_bv_size(sum) == 2*sbits + 2); SASSERT(is_well_sorted(m, sum)); - dbg_decouple("fpa2bv_fma_add_sum", sum); expr_ref sign_bv(m), n_sum(m); sign_bv = m_bv_util.mk_extract(2*sbits+1, 2*sbits+1, sum); n_sum = m_bv_util.mk_bv_neg(sum); + dbg_decouple("fpa2bv_fma_add_sign_bv", sign_bv); + dbg_decouple("fpa2bv_fma_add_n_sum", n_sum); expr_ref res_sig_eq(m), sig_abs(m), one_1(m); one_1 = m_bv_util.mk_numeral(1, 1); m_simp.mk_eq(sign_bv, one_1, res_sig_eq); m_simp.mk_ite(res_sig_eq, n_sum, sum, sig_abs); - dbg_decouple("fpa2bv_fma_add_sign_bv", sign_bv); - dbg_decouple("fpa2bv_fma_add_n_sum", n_sum); dbg_decouple("fpa2bv_fma_add_sig_abs", sig_abs); - res_exp = e_exp; - family_id bvfid = m_bv_util.get_fid(); expr_ref res_sgn_c1(m), res_sgn_c2(m), res_sgn_c3(m); expr_ref not_e_sgn(m), not_f_sgn(m), not_sign_bv(m); @@ -1643,44 +1652,57 @@ void fpa2bv_converter::mk_fma(func_decl * f, unsigned num, expr * const * args, res_sgn_c3 = m.mk_app(bvfid, OP_BAND, e_sgn, f_sgn); expr * res_sgn_or_args[3] = { res_sgn_c1, res_sgn_c2, res_sgn_c3 }; res_sgn = m_bv_util.mk_bv_or(3, res_sgn_or_args); + dbg_decouple("fpa2bv_fma_res_sgn", res_sgn); + + expr_ref is_sig_neg(m); + is_sig_neg = m.mk_eq(one_1, m_bv_util.mk_extract(2 * sbits + 1, 2 * sbits + 1, sig_abs)); + sig_abs = m.mk_ite(is_sig_neg, m_bv_util.mk_bv_neg(sig_abs), sig_abs); + dbg_decouple("fpa2bv_fma_is_sig_neg", is_sig_neg); // Result could have overflown into 4.xxx. SASSERT(m_bv_util.get_bv_size(sig_abs) == 2 * sbits + 2); - expr_ref ovfl_into_4(m); - ovfl_into_4 = m.mk_eq(m_bv_util.mk_extract(2 * sbits + 1, 2 * sbits, sig_abs), - m_bv_util.mk_numeral(1, 2)); - dbg_decouple("fpa2bv_fma_ovfl_into_4", ovfl_into_4); - if (sbits > 5) { - expr_ref sticky_raw(m), sig_upper(m), sticky_redd(m), res_sig_norm(m); - sticky_raw = m_bv_util.mk_extract(sbits - 4, 0, sig_abs); - sig_upper = m_bv_util.mk_extract(2 * sbits, sbits - 3, sig_abs); - SASSERT(m_bv_util.get_bv_size(sig_upper) == sbits + 4); - sticky_redd = m.mk_app(bvfid, OP_BREDOR, sticky_raw.get()); - sticky = m_bv_util.mk_zero_extend(sbits + 3, sticky_redd); - expr * res_or_args[2] = { sig_upper, sticky }; - res_sig_norm = m_bv_util.mk_bv_or(2, res_or_args); + expr_ref extra(m), extra_is_zero(m); + extra = m_bv_util.mk_extract(2 * sbits + 1, 2 * sbits, sig_abs); + extra_is_zero = m.mk_eq(extra, m_bv_util.mk_numeral(0, 2)); + dbg_decouple("fpa2bv_fma_extra", extra); - expr_ref sticky_raw_ovfl(m), sig_upper_ovfl(m), sticky_redd_ovfl(m), sticky_ovfl(m), res_sig_ovfl(m); - sticky_raw_ovfl = m_bv_util.mk_extract(sbits - 4, 0, sig_abs); - sig_upper_ovfl = m_bv_util.mk_extract(2 * sbits, sbits - 3, sig_abs); - SASSERT(m_bv_util.get_bv_size(sig_upper_ovfl) == sbits + 4); - sticky_redd_ovfl = m.mk_app(bvfid, OP_BREDOR, sticky_raw_ovfl.get()); - sticky_ovfl = m_bv_util.mk_zero_extend(sbits + 3, sticky_redd_ovfl); - expr * res_or_args_ovfl[2] = { sig_upper_ovfl, sticky_ovfl }; - res_sig_ovfl = m_bv_util.mk_bv_or(2, res_or_args_ovfl); - - res_sig = m.mk_ite(ovfl_into_4, res_sig_ovfl, res_sig_norm); - res_exp = m.mk_ite(ovfl_into_4, m_bv_util.mk_bv_add(res_exp, m_bv_util.mk_numeral(1, ebits+2)), - res_exp); - } - else { - unsigned too_short = 6 - sbits; + unsigned too_short = 0; + if (sbits < 5) { + too_short = 6 - sbits + 1; sig_abs = m_bv_util.mk_concat(sig_abs, m_bv_util.mk_numeral(0, too_short)); - res_sig = m_bv_util.mk_extract(sbits + 3, 0, sig_abs); } - dbg_decouple("fpa2bv_fma_add_sum_sticky", sticky); - dbg_decouple("fpa2bv_fma_sig_abs", sig_abs); + + expr_ref sig_abs_h1(m), sticky_h1(m), sticky_h1_red(m), sig_abs_h1_f(m), res_sig_1(m); + sticky_h1 = m_bv_util.mk_extract(sbits - 5 + too_short, 0, sig_abs); + sig_abs_h1 = m_bv_util.mk_extract(2 * sbits + 1 + too_short, sbits - 4 + too_short, sig_abs); + sticky_h1_red = m_bv_util.mk_zero_extend(sbits + 5, m.mk_app(m_bv_util.get_fid(), OP_BREDOR, sticky_h1.get())); + expr * sticky_h1_red_args[2] = { sig_abs_h1, sticky_h1_red }; + sig_abs_h1_f = m_bv_util.mk_bv_or(2, sticky_h1_red_args); + res_sig_1 = m_bv_util.mk_extract(sbits + 3, 0, sig_abs_h1_f); + SASSERT(m_bv_util.get_bv_size(sticky_h1) == sbits - 4 + too_short); + SASSERT(m_bv_util.get_bv_size(sig_abs_h1) == sbits + 6); + SASSERT(m_bv_util.get_bv_size(sticky_h1_red) == sbits + 6); + SASSERT(m_bv_util.get_bv_size(sig_abs_h1_f) == sbits + 6); + SASSERT(m_bv_util.get_bv_size(res_sig_1) == sbits + 4); + + expr_ref sig_abs_h2(m), sticky_h2(m), sticky_h2_red(m), sig_abs_h2_f(m), res_sig_2(m); + sticky_h2 = m_bv_util.mk_extract(sbits - 4 + too_short, 0, sig_abs); + sig_abs_h2 = m_bv_util.mk_extract(2 * sbits + 1 + too_short, sbits - 3 + too_short, sig_abs); + sticky_h2_red = m_bv_util.mk_zero_extend(sbits + 4, m.mk_app(m_bv_util.get_fid(), OP_BREDOR, sticky_h1.get())); + expr * sticky_h2_red_args[2] = { sig_abs_h2, sticky_h2_red }; + sig_abs_h2_f = m_bv_util.mk_zero_extend(1, m_bv_util.mk_bv_or(2, sticky_h2_red_args)); + res_sig_2 = m_bv_util.mk_extract(sbits + 3, 0, sig_abs_h2_f); + SASSERT(m_bv_util.get_bv_size(sticky_h2) == sbits - 3 + too_short); + SASSERT(m_bv_util.get_bv_size(sig_abs_h2) == sbits + 5); + SASSERT(m_bv_util.get_bv_size(sticky_h2_red) == sbits + 5); + SASSERT(m_bv_util.get_bv_size(sig_abs_h2_f) == sbits + 6); + SASSERT(m_bv_util.get_bv_size(res_sig_2) == sbits + 4); + + res_sig = m.mk_ite(extra_is_zero, res_sig_1, res_sig_2); + res_exp = m.mk_ite(extra_is_zero, e_exp, m_bv_util.mk_bv_add(e_exp, m_bv_util.mk_numeral(1, ebits + 2))); + dbg_decouple("fpa2bv_fma_res_sig", res_sig); + dbg_decouple("fpa2bv_fma_res_exp", res_exp); SASSERT(m_bv_util.get_bv_size(res_sig) == sbits + 4); expr_ref is_zero_sig(m), nil_sbits4(m); @@ -3889,8 +3911,8 @@ void fpa2bv_converter::round(sort * s, expr_ref & rm, expr_ref & sgn, expr_ref & mk_min_exp(ebits, e_min); mk_max_exp(ebits, e_max); - TRACE("fpa2bv_dbg", tout << "e_min = " << mk_ismt2_pp(e_min, m) << std::endl << - "e_max = " << mk_ismt2_pp(e_max, m) << std::endl;); + TRACE("fpa2bv_dbg", tout << "e_min = " << mk_ismt2_pp(e_min, m) << std::endl; + tout << "e_max = " << mk_ismt2_pp(e_max, m) << std::endl;); expr_ref OVF1(m), e_top_three(m), sigm1(m), e_eq_emax_and_sigm1(m), e_eq_emax(m); expr_ref e3(m), ne3(m), e2(m), e1(m), e21(m), one_1(m), h_exp(m), sh_exp(m), th_exp(m); @@ -4106,7 +4128,6 @@ void fpa2bv_converter::round(sort * s, expr_ref & rm, expr_ref & sgn, expr_ref & dbg_decouple("fpa2bv_rnd_rm_is_to_neg", rm_is_to_neg); dbg_decouple("fpa2bv_rnd_rm_is_to_pos", rm_is_to_pos); - expr_ref sgn_is_zero(m), zero1(m); zero1 = m_bv_util.mk_numeral(0, 1); m_simp.mk_eq(sgn, zero1, sgn_is_zero); diff --git a/src/util/mpf.cpp b/src/util/mpf.cpp index e9d108cec..7b5c8d62d 100644 --- a/src/util/mpf.cpp +++ b/src/util/mpf.cpp @@ -886,7 +886,7 @@ void mpf_manager::fma(mpf_rounding_mode rm, mpf const & x, mpf const & y, mpf co unsigned extra = 0; // Result could overflow into 4.xxx ... SASSERT(m_mpz_manager.lt(o.significand, m_powers2(2 * x.sbits + 2))); - if(m_mpz_manager.ge(o.significand, m_powers2(2 * x.sbits + 1))) + if (m_mpz_manager.ge(o.significand, m_powers2(2 * x.sbits + 1))) { extra++; o.exponent++;