diff options
| author | Joseph Myers <joseph@codesourcery.com> | 2017-01-09 22:40:59 +0000 |
|---|---|---|
| committer | Joseph Myers <joseph@codesourcery.com> | 2017-01-09 22:40:59 +0000 |
| commit | 08f7b95dbdf7ddc25899fe5b521a36db03fe6e2a (patch) | |
| tree | 9cd4770bac4f35d3c84eb24c641901f46bbb9b71 /math/gen-auto-libm-tests.c | |
| parent | c6be839efded78768ad3c70d7f2e913a9fb8603c (diff) | |
Improve libm-test XFAILing for ibm128-libgcc.
This patch further improves XFAILing for ibm128-libgcc of tests in
auto-libm-test-*.
The bulk of the cases needing XFAILing are
xfail-rounding:ibm128-libgcc used for inputs where (possibly after
rounding the inputs to another floating-point type) the result
overflows (and the result in non-default rounding modes may be wildly
wrong with unpatched libgcc) or underflows near 0 (and the result in
non-default rounding modes may end up having the wrong sign). This
patch makes gen-auto-libm-tests detect such cases and apply
xfail-rounding:ibm128-libgcc automatically to them, so most of the
manual XFAILs in auto-libm-test-in are no longer needed (some are
still needed if e.g. the result is very close to overflow, resulting
in an internal overflow in libgcc in some rounding modes). A few
manual XFAILs are added for cases not covered by this
gen-auto-libm-tests change, and a few existing such XFAILs are left
in.
Tested for powerpc.
* math/gen-auto-libm-tests.c (output_for_one_input_case): Apply
xfail-rounding:ibm128-libgcc automatically to tests overflowing
and those that can underflow to zero.
* math/auto-libm-test-in: Remove most XFAILs for ibm128-libgcc and
add others.
* math/auto-libm-test-out: Regenerated.
Diffstat (limited to 'math/gen-auto-libm-tests.c')
| -rw-r--r-- | math/gen-auto-libm-tests.c | 15 |
1 files changed, 15 insertions, 0 deletions
diff --git a/math/gen-auto-libm-tests.c b/math/gen-auto-libm-tests.c index 22dd73f916..9f497e019f 100644 --- a/math/gen-auto-libm-tests.c +++ b/math/gen-auto-libm-tests.c @@ -1907,6 +1907,7 @@ output_for_one_input_case (FILE *fp, const char *filename, test_function *tf, error (EXIT_FAILURE, errno, "write to '%s'", filename); /* Print outputs. */ bool must_erange = false; + bool some_underflow_zero = false; for (size_t i = 0; i < tf->num_ret; i++) { generic_value g; @@ -1924,6 +1925,10 @@ output_for_one_input_case (FILE *fp, const char *filename, test_function *tf, && (all_exc_before[i][m] & (1U << exc_underflow)) != 0) must_erange = true; + if (mpfr_zero_p (all_res[i][rm_towardzero]) + && (all_exc_before[i][m] + & (1U << exc_underflow)) != 0) + some_underflow_zero = true; mpfr_init2 (g.value.f, fp_formats[f].mant_dig); assert_exact (mpfr_set (g.value.f, all_res[i][m], MPFR_RNDN)); @@ -1971,6 +1976,16 @@ output_for_one_input_case (FILE *fp, const char *filename, test_function *tf, default: break; } + /* For the ibm128 format, expect incorrect overflowing + results in rounding modes other than to nearest; + likewise incorrect results where the result may + underflow to 0. */ + if (f == fp_ldbl_128ibm + && m != rm_tonearest + && (some_underflow_zero + || (merged_exc_before[m] & (1U << exc_overflow)) != 0)) + if (fputs (" xfail:ibm128-libgcc", fp) < 0) + error (EXIT_FAILURE, errno, "write to '%s'", filename); /* Print exception flags and compute errno expectations where not already computed. */ bool may_edom = false; |