From e337595266ee9be3d0b805c522d1abeead8bc7ff Mon Sep 17 00:00:00 2001 From: Yuao Ma Date: Sun, 1 Jun 2025 15:46:59 +0800 Subject: [PATCH] gcc: middle-end opt for trigonometric pi-based functions builtins This patch partially handled PR118592. This patch builds upon r16-710-g591d3d02664c7b and r16-711-g89935d56f768b4. It introduces middle-end optimizations, such as constant folding, for our trigonometric pi-based function built-ins. For MPFR versions older than 4.2.0, we've included our own folding functions. I've also added tests to ensure everything works as expected. gcc/ChangeLog: * fold-const-call.cc (mpfr_acospi, mpfr_asinpi, mpfr_atanpi, mpfr_atan2pi, mpfr_cospi, mpfr_sinpi, mpfr_tanpi): Workaround for mpfr < 4.2.0. (fold_const_call_ss): Constant fold for single arg. (fold_const_call_sss): Constant fold for double args. * fold-const.cc (negate_mathfn_p): asinpi/atanpi is odd func. (tree_call_nonnegative_warnv_p): acospi always non-neg, asinpi/atanpi non-neg iff arg non-neg. * tree-call-cdce.cc (can_test_argument_range): Add acospi/asinpi. (edom_only_function): Add acospi/asinpi/cospi/sinpi. Remove atan because atan NEVER produce EDOM. (get_no_error_domain): Add acospi/asinpi. gcc/testsuite/ChangeLog: * gcc.dg/torture/builtin-math-1.c: Add tests for builtin math fold. * gcc.dg/torture/floatn-builtin.h: Add tests for float-nx. Signed-off-by: Yuao Ma --- gcc/fold-const-call.cc | 173 ++++++++++++++++++ gcc/fold-const.cc | 14 ++ gcc/testsuite/gcc.dg/torture/builtin-math-1.c | 35 ++++ gcc/testsuite/gcc.dg/torture/floatn-builtin.h | 24 +++ gcc/tree-call-cdce.cc | 18 +- 5 files changed, 262 insertions(+), 2 deletions(-) diff --git a/gcc/fold-const-call.cc b/gcc/fold-const-call.cc index 8230b83b929..b9bb6350e14 100644 --- a/gcc/fold-const-call.cc +++ b/gcc/fold-const-call.cc @@ -171,6 +171,145 @@ do_mpfr_sincos (real_value *result_sin, real_value *result_cos, return ok; } +#if MPFR_VERSION < MPFR_VERSION_NUM(4, 2, 0) +static int +mpfr_acospi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t pi, tmp; + mpfr_inits2 (2 * mpfr_get_prec (op), pi, tmp, NULL); + mpfr_const_pi (pi, rnd); + mpfr_acos (tmp, op, rnd); + mpfr_div (rop, tmp, pi, rnd); + mpfr_clears (pi, tmp, NULL); + return 0; +} + +static int +mpfr_asinpi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t pi, tmp; + mpfr_inits2 (2 * mpfr_get_prec (op), pi, tmp, NULL); + mpfr_const_pi (pi, rnd); + mpfr_asin (tmp, op, rnd); + mpfr_div (rop, tmp, pi, rnd); + mpfr_clears (pi, tmp, NULL); + return 0; +} + +static int +mpfr_atanpi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t pi, tmp; + mpfr_inits2 (2 * mpfr_get_prec (op), pi, tmp, NULL); + mpfr_const_pi (pi, rnd); + mpfr_atan (tmp, op, rnd); + mpfr_div (rop, tmp, pi, rnd); + mpfr_clears (pi, tmp, NULL); + return 0; +} + +static int +mpfr_atan2pi (mpfr_ptr rop, mpfr_srcptr y, mpfr_srcptr x, mpfr_rnd_t rnd) +{ + mpfr_t pi, tmp; + mpfr_inits2 (2 * mpfr_get_prec (x), pi, tmp, NULL); + mpfr_const_pi (pi, rnd); + mpfr_atan2 (tmp, y, x, rnd); + mpfr_div (rop, tmp, pi, rnd); + mpfr_clears (pi, tmp, NULL); + return 0; +} + +static int +mpfr_cospi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t cs, n, r, two; + int s; + + mpfr_inits2 (2 * mpfr_get_prec (op), cs, n, r, two, NULL); + + mpfr_abs (r, op, rnd); + mpfr_modf (n, r, r, rnd); + + if (mpfr_cmp_d (r, 0.5) == 0) + { + mpfr_set_ui (rop, 0, rnd); + return 0; + } + + mpfr_set_ui (two, 2, rnd); + mpfr_fmod (cs, n, two, rnd); + s = mpfr_cmp_ui (cs, 0) == 0 ? 1 : -1; + + mpfr_const_pi (cs, rnd); + mpfr_mul (cs, cs, r, rnd); + mpfr_cos (cs, cs, rnd); + mpfr_mul_si (rop, cs, s, rnd); + + mpfr_clears (cs, n, r, two, NULL); + return 0; +} + +static int +mpfr_sinpi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t sn, n, r, two; + int s; + + mpfr_inits2 (2 * mpfr_get_prec (op), sn, n, r, two, NULL); + + mpfr_abs (r, op, rnd); + mpfr_modf (n, r, r, rnd); + + if (mpfr_cmp_d (r, 0.0) == 0) + { + mpfr_set_ui (rop, 0, rnd); + return 0; + } + + mpfr_set_ui (two, 2, rnd); + mpfr_fmod (sn, n, two, rnd); + s = mpfr_cmp_si (op, 0) < 0 ? -1 : 1; + s *= mpfr_cmp_ui (sn, 0) == 0 ? 1 : -1; + + mpfr_const_pi (sn, rnd); + mpfr_mul (sn, sn, r, rnd); + mpfr_sin (sn, sn, rnd); + mpfr_mul_si (rop, sn, s, rnd); + + mpfr_clears (sn, n, r, two, NULL); + return 0; +} + +static int +mpfr_tanpi (mpfr_ptr rop, mpfr_srcptr op, mpfr_rnd_t rnd) +{ + mpfr_t tn, n, r; + int s; + + mpfr_inits2 (2 * mpfr_get_prec (op), tn, n, r, NULL); + + mpfr_abs (r, op, rnd); + mpfr_modf (n, r, r, rnd); + + if (mpfr_cmp_d (r, 0.0) == 0) + { + mpfr_set_ui (rop, 0, rnd); + return 0; + } + + s = mpfr_cmp_si (op, 0) < 0 ? -1 : 1; + + mpfr_const_pi (tn, rnd); + mpfr_mul (tn, tn, r, rnd); + mpfr_tan (tn, tn, rnd); + mpfr_mul_si (rop, tn, s, rnd); + + mpfr_clears (tn, n, r, NULL); + return 0; +} +#endif + /* Try to evaluate: *RESULT = f (*ARG0, *ARG1) @@ -788,6 +927,34 @@ fold_const_call_ss (real_value *result, combined_fn fn, CASE_CFN_TANH_FN: return do_mpfr_arg1 (result, mpfr_tanh, arg, format); + CASE_CFN_ACOSPI: + CASE_CFN_ACOSPI_FN: + return (real_compare (GE_EXPR, arg, &dconstm1) + && real_compare (LE_EXPR, arg, &dconst1) + && do_mpfr_arg1 (result, mpfr_acospi, arg, format)); + + CASE_CFN_ASINPI: + CASE_CFN_ASINPI_FN: + return (real_compare (GE_EXPR, arg, &dconstm1) + && real_compare (LE_EXPR, arg, &dconst1) + && do_mpfr_arg1 (result, mpfr_asinpi, arg, format)); + + CASE_CFN_ATANPI: + CASE_CFN_ATANPI_FN: + return do_mpfr_arg1 (result, mpfr_atanpi, arg, format); + + CASE_CFN_COSPI: + CASE_CFN_COSPI_FN: + return do_mpfr_arg1 (result, mpfr_cospi, arg, format); + + CASE_CFN_SINPI: + CASE_CFN_SINPI_FN: + return do_mpfr_arg1 (result, mpfr_sinpi, arg, format); + + CASE_CFN_TANPI: + CASE_CFN_TANPI_FN: + return do_mpfr_arg1 (result, mpfr_tanpi, arg, format); + CASE_CFN_ERF: CASE_CFN_ERF_FN: return do_mpfr_arg1 (result, mpfr_erf, arg, format); @@ -1432,6 +1599,12 @@ fold_const_call_sss (real_value *result, combined_fn fn, CASE_CFN_ATAN2_FN: return do_mpfr_arg2 (result, mpfr_atan2, arg0, arg1, format); + CASE_CFN_ATAN2PI: + CASE_CFN_ATAN2PI_FN: + return (!(real_compare (EQ_EXPR, arg0, &dconst0) + && real_compare (EQ_EXPR, arg1, &dconst0)) + && do_mpfr_arg2 (result, mpfr_atan2pi, arg0, arg1, format)); + CASE_CFN_FDIM: CASE_CFN_FDIM_FN: return do_mpfr_arg2 (result, mpfr_dim, arg0, arg1, format); diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc index 014f4218793..47492575d00 100644 --- a/gcc/fold-const.cc +++ b/gcc/fold-const.cc @@ -395,10 +395,14 @@ negate_mathfn_p (combined_fn fn) CASE_CFN_ASIN_FN: CASE_CFN_ASINH: CASE_CFN_ASINH_FN: + CASE_CFN_ASINPI: + CASE_CFN_ASINPI_FN: CASE_CFN_ATAN: CASE_CFN_ATAN_FN: CASE_CFN_ATANH: CASE_CFN_ATANH_FN: + CASE_CFN_ATANPI: + CASE_CFN_ATANPI_FN: CASE_CFN_CASIN: CASE_CFN_CASIN_FN: CASE_CFN_CASINH: @@ -432,10 +436,14 @@ negate_mathfn_p (combined_fn fn) CASE_CFN_SIN_FN: CASE_CFN_SINH: CASE_CFN_SINH_FN: + CASE_CFN_SINPI: + CASE_CFN_SINPI_FN: CASE_CFN_TAN: CASE_CFN_TAN_FN: CASE_CFN_TANH: CASE_CFN_TANH_FN: + CASE_CFN_TANPI: + CASE_CFN_TANPI_FN: CASE_CFN_TRUNC: CASE_CFN_TRUNC_FN: return true; @@ -14962,6 +14970,8 @@ tree_call_nonnegative_warnv_p (tree type, combined_fn fn, tree arg0, tree arg1, CASE_CFN_ACOS_FN: CASE_CFN_ACOSH: CASE_CFN_ACOSH_FN: + CASE_CFN_ACOSPI: + CASE_CFN_ACOSPI_FN: CASE_CFN_CABS: CASE_CFN_CABS_FN: CASE_CFN_COSH: @@ -15006,10 +15016,14 @@ tree_call_nonnegative_warnv_p (tree type, combined_fn fn, tree arg0, tree arg1, CASE_CFN_ASINH: CASE_CFN_ASINH_FN: + CASE_CFN_ASINPI: + CASE_CFN_ASINPI_FN: CASE_CFN_ATAN: CASE_CFN_ATAN_FN: CASE_CFN_ATANH: CASE_CFN_ATANH_FN: + CASE_CFN_ATANPI: + CASE_CFN_ATANPI_FN: CASE_CFN_CBRT: CASE_CFN_CBRT_FN: CASE_CFN_CEIL: diff --git a/gcc/testsuite/gcc.dg/torture/builtin-math-1.c b/gcc/testsuite/gcc.dg/torture/builtin-math-1.c index 4ecc98d5f5c..94ae69e7531 100644 --- a/gcc/testsuite/gcc.dg/torture/builtin-math-1.c +++ b/gcc/testsuite/gcc.dg/torture/builtin-math-1.c @@ -36,6 +36,14 @@ extern double tan (double); extern float tanf (float); extern long double tanl (long double); +extern double acospi (double); +extern double asinpi (double); +extern double atanpi (double); +extern double atan2pi (double, double); +extern double cospi (double); +extern double sinpi (double); +extern double tanpi (double); + /* All references to link_error should go away at compile-time. */ extern void link_error(void); @@ -177,6 +185,33 @@ void test (float f, double d, long double ld) if (powl (1.0L, ld) != 1.0L) link_error (); + + if (acospi (0.5) < 0.3333 || acospi (0.5) > 0.3334 || acospi (-0.5) < 0.6666 + || acospi (-0.5) > 0.6667) + link_error (); + + if (asinpi (0.5) < 0.1666 || asinpi (0.5) > 0.1667 || asinpi (-0.5) < -0.1667 + || asinpi (-0.5) > -0.1666) + link_error (); + + if (atanpi (1.0) != 0.25 || atanpi (-1.0) > -0.25) + link_error (); + + if (atan2pi (1.0, 1.0) > 0.2501 || atan2pi (1.0, 1.0) < 0.2499 + || atan2pi (1.0, -1.0) < 0.7499 || atan2pi (1.0, -1.0) > 0.7501) + link_error (); + + if (cospi (1.0 / 3) > 0.5001 || cospi (-1.0 / 3) < 0.4999 + || cospi (4.0 / 3) > -0.4999 || cospi (4.0 / 3) < -0.5001) + link_error (); + + if (sinpi (1.0 / 6) > 0.5001 || sinpi (-1.0 / 6) > -0.4999 + || sinpi (5.0 / 6) < 0.4999 || sinpi (-7.0 / 6) > 0.5001) + link_error (); + + if (tanpi (0.25) != 1.0000 || tanpi (-0.25) != -1.0000 + || tanpi (1.25) != 1.0000 || tanpi (-1.25) != -1.0000) + link_error (); } int main() diff --git a/gcc/testsuite/gcc.dg/torture/floatn-builtin.h b/gcc/testsuite/gcc.dg/torture/floatn-builtin.h index 304a459ee0b..bd72e7cdfb2 100644 --- a/gcc/testsuite/gcc.dg/torture/floatn-builtin.h +++ b/gcc/testsuite/gcc.dg/torture/floatn-builtin.h @@ -73,6 +73,12 @@ extern __typeof (FN (__builtin_remainder) (0, 0)) test_type; extern __typeof (FN (__builtin_remquo) (0, 0, &test_i)) test_type; extern __typeof (FN (__builtin_scalbln) (0, 0L)) test_type; extern __typeof (FN (__builtin_scalbn) (0, 0)) test_type; +extern __typeof (FN (__builtin_acospi) (0)) test_type; +extern __typeof (FN (__builtin_asinpi) (0)) test_type; +extern __typeof (FN (__builtin_atanpi) (0)) test_type; +extern __typeof (FN (__builtin_cospi) (0)) test_type; +extern __typeof (FN (__builtin_sinpi) (0)) test_type; +extern __typeof (FN (__builtin_tanpi) (0)) test_type; volatile TYPE inf_cst = FN (__builtin_inf) (); volatile TYPE huge_val_cst = FN (__builtin_huge_val) (); @@ -115,6 +121,12 @@ volatile TYPE t33 = FN (__builtin_pow) (CST (1.0), CST (2.0)); volatile TYPE t34 = FN (__builtin_remainder) (CST (7.0), CST (4.0)); volatile TYPE t35 = FN (__builtin_scalbln) (CST (1.0), 1L); volatile TYPE t36 = FN (__builtin_scalbn) (CST (1.0), 1); +volatile TYPE t37 = FN (__builtin_acospi) (CST (1.0)); +volatile TYPE t38 = FN (__builtin_asinpi) (CST (0.0)); +volatile TYPE t39 = FN (__builtin_atanpi) (CST (0.0)); +volatile TYPE t40 = FN (__builtin_cospi) (CST (0.0)); +volatile TYPE t41 = FN (__builtin_sinpi) (CST (0.0)); +volatile TYPE t42 = FN (__builtin_tanpi) (CST (0.0)); int main (void) @@ -214,5 +226,17 @@ main (void) abort (); if (t36 != CST (2.0)) abort (); + if (t37 != CST (0.0)) + abort (); + if (t38 != CST (0.0)) + abort (); + if (t39 != CST (0.0)) + abort (); + if (t40 != CST (1.0)) + abort (); + if (t41 != CST (0.0)) + abort (); + if (t42 != CST (0.0)) + abort (); exit (0); } diff --git a/gcc/tree-call-cdce.cc b/gcc/tree-call-cdce.cc index 9ca5fda5126..649c1e2b9f9 100644 --- a/gcc/tree-call-cdce.cc +++ b/gcc/tree-call-cdce.cc @@ -296,8 +296,12 @@ can_test_argument_range (gcall *call) /* Trig functions. */ CASE_FLT_FN (BUILT_IN_ACOS): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOS): + CASE_FLT_FN (BUILT_IN_ACOSPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOSPI): CASE_FLT_FN (BUILT_IN_ASIN): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASIN): + CASE_FLT_FN (BUILT_IN_ASINPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASINPI): /* Hyperbolic functions. */ CASE_FLT_FN (BUILT_IN_ACOSH): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOSH): @@ -351,15 +355,21 @@ edom_only_function (gcall *call) { CASE_FLT_FN (BUILT_IN_ACOS): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOS): + CASE_FLT_FN (BUILT_IN_ACOSPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOSPI): CASE_FLT_FN (BUILT_IN_ASIN): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASIN): - CASE_FLT_FN (BUILT_IN_ATAN): - CASE_FLT_FN_FLOATN_NX (BUILT_IN_ATAN): + CASE_FLT_FN (BUILT_IN_ASINPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASINPI): CASE_FLT_FN (BUILT_IN_COS): CASE_FLT_FN_FLOATN_NX (BUILT_IN_COS): + CASE_FLT_FN (BUILT_IN_COSPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_COSPI): CASE_FLT_FN (BUILT_IN_SIGNIFICAND): CASE_FLT_FN (BUILT_IN_SIN): CASE_FLT_FN_FLOATN_NX (BUILT_IN_SIN): + CASE_FLT_FN (BUILT_IN_SINPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_SINPI): CASE_FLT_FN (BUILT_IN_SQRT): CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT): CASE_FLT_FN (BUILT_IN_FMOD): @@ -694,8 +704,12 @@ get_no_error_domain (enum built_in_function fnc) /* Trig functions: return [-1, +1] */ CASE_FLT_FN (BUILT_IN_ACOS): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOS): + CASE_FLT_FN (BUILT_IN_ACOSPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ACOSPI): CASE_FLT_FN (BUILT_IN_ASIN): CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASIN): + CASE_FLT_FN (BUILT_IN_ASINPI): + CASE_FLT_FN_FLOATN_NX (BUILT_IN_ASINPI): return get_domain (-1, true, true, 1, true, true); /* Hyperbolic functions. */ -- 2.43.0