| Message ID | 20241129132032.476978-12-adhemerval.zanella@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | Add remaining CORE-MATH binary32 implementations to libm | expand |
On Fri, 29 Nov 2024, Adhemerval Zanella wrote: > For some correctly rounded inputs where infinity might generate > a number (like atanf), comparing to a pre-defined constant does not > yield the expected result in all roundind modes. > > The most straightforward way to handle it would be to get the expected > result from mpfr, where it handles all the rounding mode. This is missing updates to the comment at the top of gen-auto-libm-tests.c ("Inputs are either finite real numbers or ...") to explain the handling of infinities.
On 29/11/24 13:35, Joseph Myers wrote: > On Fri, 29 Nov 2024, Adhemerval Zanella wrote: > >> For some correctly rounded inputs where infinity might generate >> a number (like atanf), comparing to a pre-defined constant does not >> yield the expected result in all roundind modes. >> >> The most straightforward way to handle it would be to get the expected >> result from mpfr, where it handles all the rounding mode. > > This is missing updates to the comment at the top of gen-auto-libm-tests.c > ("Inputs are either finite real numbers or ...") to explain the handling > of infinities. > Right, I changed to: Other lines are test lines, of the form "function input1 input2 ... [flag1 flag2 ...]". Inputs are either finite real numbers, positive or negative infinite (in the form of "inf" or "-inf"), or integers, depending on the function under test. Real numbers may be in any form acceptable to mpfr_strtofr (base 0), infinite may in any form acceptable to mpfr_set_inf, and integers in any form acceptable to mpz_set_str (base 0). In addition, real numbers may be certain special strings such as "pi", as listed in the special_real_inputs array.
diff --git a/math/gen-auto-libm-tests.c b/math/gen-auto-libm-tests.c index 0b1e307fae..457610ef04 100644 --- a/math/gen-auto-libm-tests.c +++ b/math/gen-auto-libm-tests.c @@ -981,6 +981,26 @@ special_fill_e_minus_1 (mpfr_t res0, mpfr_t res1, fp_format format) return 2; } +/* Set the precision of RES0 based on FORMAT and initialize as an + infinite number. */ +static size_t +special_fill_inf (mpfr_t res0, mpfr_t res1 __attribute__ ((unused)), + fp_format format) +{ + mpfr_init2 (res0, fp_formats[format].mant_dig); + mpfr_set_inf (res0, 0); + return 1; +} + +static size_t +special_fill_minus_inf (mpfr_t res0, mpfr_t res1 __attribute__ ((unused)), + fp_format format) +{ + mpfr_init2 (res0, fp_formats[format].mant_dig); + mpfr_set_inf (res0, -1); + return 1; +} + /* A special string accepted in input arguments. */ typedef struct { @@ -1016,6 +1036,8 @@ static const special_real_input special_real_inputs[] = { "e", special_fill_e }, { "1/e", special_fill_1_e }, { "e-1", special_fill_e_minus_1 }, + { "inf", special_fill_inf }, + { "-inf", special_fill_minus_inf }, }; /* Given a real number R computed in round-to-zero mode, set the @@ -1062,7 +1084,6 @@ round_real (mpfr_t res[rm_num_modes], unsigned int exc_after[rm_num_modes], mpfr_t r, fp_format format) { - assert (mpfr_number_p (r)); for (rounding_mode m = rm_first_mode; m < rm_num_modes; m++) { mpfr_init2 (res[m], fp_formats[format].mant_dig);