summaryrefslogtreecommitdiffstats
path: root/test
diff options
context:
space:
mode:
Diffstat (limited to 'test')
-rw-r--r--test/dtypes.c32
1 files changed, 28 insertions, 4 deletions
diff --git a/test/dtypes.c b/test/dtypes.c
index 58a03d4..627d7bf 100644
--- a/test/dtypes.c
+++ b/test/dtypes.c
@@ -5737,7 +5737,7 @@ my_isnan(dtype_t type, void *val)
} else if (FLT_LDOUBLE==type) {
long double x;
HDmemcpy(&x, val, sizeof(long double));
- sprintf(s, "%LLg", x);
+ sprintf(s, "%Lg", x);
#endif
} else {
return 0;
@@ -6617,9 +6617,6 @@ test_conv_int_float(const char *name, hbool_t run_special, hid_t src, hid_t dst)
/* Check the results from the library against hardware */
for (j=0; j<nelmts; j++) {
- if(FLT_FLOAT==src_type || FLT_DOUBLE==src_type || FLT_LDOUBLE==src_type)
- if(my_isnan(src_type, saved+j*src_size))
- continue;
if (FLT_FLOAT==dst_type) {
hw = (unsigned char*)&hw_float;
@@ -7059,6 +7056,33 @@ test_conv_int_float(const char *name, hbool_t run_special, hid_t src, hid_t dst)
}
}
}
+
+ /* Test library's handling of NaN:
+ * Hardware usually doesn't handle NaN too gracefully. The
+ * hardware conversion result is usually garbage so we must handle
+ * those cases differetly when checking results.
+ *
+ * Test user's exception handler when NaN happens:
+ * Try to follow the except_func callback function to check if the
+ * desired value was set.
+ */
+ if (run_special && (FLT_FLOAT==src_type || FLT_DOUBLE==src_type || FLT_LDOUBLE==src_type)) {
+ if (my_isnan(src_type, src_bits)) {
+ if(!except_set) {
+ if(!strcmp(name, "sw")) {
+ if (H5T_bit_find(dst_bits, 0, dst_nbits-1, H5T_BIT_LSB, 1)<0)
+ continue; /*no error. Software conversion set integer to 0.*/
+ } else if(!strcmp(name, "hw")) {
+ continue; /*don't compare because hardware conversion may fill in garbage*/
+ }
+ } else {
+ /* fill_value is small so we know only the 1st byte is set */
+ if (dst_bits[0] == fill_value)
+ continue; /*no error*/
+ }
+ }
+ }
+
/* On some machines (notably the SGI and Solaris 64-bit machines) unsigned long
* values are not converted to float or double values correctly, they are
* consistently off by the lowest bit being rounded oppositely to our