hdf5/PPC64LE-Fix-long-double-handling.patch
Dirk Stoecker 5daaa9f016 Accepting request 637096 from home:StefanBruens:branches:science
Main reason to update is the addition/enhancement of 
H5Pset_libver_bounds():
"Added an enumerated value to H5F_libver_t for H5Pset_libver_bounds()." - https://support.hdfgroup.org/ftp/HDF5/releases/hdf5-1.10/hdf5-1.10.2/src/hdf5-1.10.2-RELEASE.txt

This allows to write backwards compatible files with newer hdf5 versions, although this is "opt-in" by the library users. See e.g. https://github.com/Unidata/netcdf-c/issues/250, https://github.com/Unidata/netcdf-c/issues/951


- Update to 1.10.3:
  https://support.hdfgroup.org/ftp/HDF5/releases/hdf5-1.10/hdf5-1.10.3/src/hdf5-1.10.3-RELEASE.txt
- Update to 1.10.2:
  https://support.hdfgroup.org/ftp/HDF5/releases/hdf5-1.10/hdf5-1.10.2/src/hdf5-1.10.2-RELEASE.txt
- Bump soversions (101 -> 103) for libhdf5 and libhdf5_cpp
- Drop Fix-warnings-for-missing-returns.patch, fixed upstream
- Add 0001-Fix-return-value-in-test_libver_bounds_open.patch
- Rebase patches (tab / space conversion):
  * PPC64LE-Fix-long-double-handling.patch
  * hdf5-buildcompare.patch
  * hdf5-mpi.patch
- Rebase Disable-phdf5-tests.patch

OBS-URL: https://build.opensuse.org/request/show/637096
OBS-URL: https://build.opensuse.org/package/show/science/hdf5?expand=0&rev=82
2018-09-21 19:26:10 +00:00

92 lines
4.0 KiB
Diff

From: Egbert Eich <eich@suse.com>
Date: Tue Nov 7 14:16:53 2017 +0100
Subject: [PATCH]PPC64LE: Fix long double handling
Git-commit: ad6559a71b7ba3cacb4b56d4747db63f28a12f55
References:
Signed-off-by: Egbert Eich <eich@suse.com>
Signed-off-by: Egbert Eich <eich@suse.com>
---
hdf5-1.10.1/config/cmake/ConversionTests.c | 16 ++++++++++++++++
hdf5-1.10.1/test/dt_arith.c | 26 ++++++++++++++++++++++++--
2 files changed, 40 insertions(+), 2 deletions(-)
diff --git a/hdf5-1.10.1/config/cmake/ConversionTests.c b/hdf5-1.10.1/config/cmake/ConversionTests.c
index 082dbd4..9618a14 100644
--- a/config/cmake/ConversionTests.c
+++ b/config/cmake/ConversionTests.c
@@ -34,6 +34,14 @@ int main(void)
unsigned char s2[8];
int ret = 1;
+#if defined __powerpc64__ && defined _LITTLE_ENDIAN
+ /* Don't bother checking on ppc64le, we know it'll work, and
+ that what hdf5 calls 'special algorithm' simply is
+ IBM ldouble 128 (i.e. two seperately scaled doubles).
+ The check below assumes big endian. */
+ ret = 0;
+#endif
+
if(sizeof(long double) == 16 && sizeof(long) == 8) {
/*make sure the long double type has 16 bytes in size and
* 11 bits of exponent. If it is,
@@ -157,6 +165,14 @@ int main(void)
unsigned char s[16];
int ret = 0;
+#if defined __powerpc64__ && defined _LITTLE_ENDIAN
+ /* Don't bother checking on ppc64le, we know it'll work, and
+ that what hdf5 calls 'special algorithm' simply is
+ IBM ldouble 128 (i.e. two seperately scaled doubles).
+ The check below assumes big endian. */
+ ret = 0;
+#endif
+
if(sizeof(long double) == 16) {
/*make sure the long double type is the same as the failing type
*which has 16 bytes in size and 11 bits of exponent. If it is,
diff --git a/hdf5-1.10.1/test/dt_arith.c b/hdf5-1.10.1/test/dt_arith.c
index 8f3e6e4..084b47c 100644
--- a/test/dt_arith.c
+++ b/test/dt_arith.c
@@ -3051,7 +3051,18 @@ test_conv_flt_1 (const char *name, int run_test, hid_t src, hid_t dst)
buf, saved, nelmts);
#if H5_SIZEOF_LONG_DOUBLE!=H5_SIZEOF_DOUBLE && H5_SIZEOF_LONG_DOUBLE!=0
} else if(src_type == FLT_LDOUBLE) {
- INIT_FP_SPECIAL(src_size, src_nbits, sendian, LDBL_MANT_DIG, dst_size,
+ size_t mant_dig = LDBL_MANT_DIG;
+ if (mant_dig >= src_nbits) {
+ /* This happens for IBM long double in little endian.
+ The macro LDBL_MANT_DIG says 106 mantissa bits, but the
+ HDF5 detection code actually represents it as a normal 64bit
+ double (52 bit mantissa) with the upper double being
+ unspec bits (which is sort of okay as the testsuite
+ wouldn't deal with that format correctly anyway). So
+ override the mantissa size. */
+ mant_dig = 52;
+ }
+ INIT_FP_SPECIAL(src_size, src_nbits, sendian, mant_dig, dst_size,
buf, saved, nelmts);
#endif
} else
@@ -3712,7 +3723,18 @@ test_conv_int_fp(const char *name, int run_test, hid_t src, hid_t dst)
INIT_FP_DENORM(long double, LDBL_MANT_DIG, src_size, src_nbits, sendian, dst_size,
buf, saved, nelmts);
} else {
- INIT_FP_SPECIAL(src_size, src_nbits, sendian, LDBL_MANT_DIG, dst_size, buf, saved, nelmts);
+ size_t mant_dig = LDBL_MANT_DIG;
+ if (mant_dig >= src_nbits) {
+ /* This happens for IBM long double in little endian.
+ The macro LDBL_MANT_DIG says 106 mantissa bits, but the
+ HDF5 detection code actually represents it as a normal 64bit
+ double (52 bit mantissa) with the upper double being
+ unspec bits (which is sort of okay as the testsuite
+ wouldn't deal with that format correctly anyway). So
+ override the mantissa size. */
+ mant_dig = 52;
+ }
+ INIT_FP_SPECIAL(src_size, src_nbits, sendian, mant_dig, dst_size, buf, saved, nelmts);
}
#endif
} else