Revert long double checks (#3133)

* Revert "Remove long double conversion work-arounds (#3097)"

This reverts commit 1e1dac1dac58fa18f6b7788346d1ba7d3315b0f9.

* Update comments to reflect newer systems

3 files changed, 403 insertions, 0 deletions

diff --git a/config/cmake/ConfigureChecks.cmake b/config/cmake/ConfigureChecks.cmake
index 32faed4..8c401d9 100644
--- a/config/cmake/ConfigureChecks.cmake
+++ b/config/cmake/ConfigureChecks.cmake
@@ -560,6 +560,19 @@ endif ()
 MARK_AS_ADVANCED (HDF5_STRICT_FORMAT_CHECKS)
 
 # ----------------------------------------------------------------------
+# Decide whether the data accuracy has higher priority during data
+# conversions.  If not, some hard conversions will still be preferred even
+# though the data may be wrong (for example, some compilers don't
+# support denormalized floating values) to maximize speed.
+#-----------------------------------------------------------------------------
+option (HDF5_WANT_DATA_ACCURACY "IF data accuracy is guaranteed during data conversions" ON)
+mark_as_advanced (HDF5_WANT_DATA_ACCURACY)
+if (HDF5_WANT_DATA_ACCURACY)
+  set (${HDF_PREFIX}_WANT_DATA_ACCURACY 1)
+endif ()
+MARK_AS_ADVANCED (HDF5_WANT_DATA_ACCURACY)
+
+# ----------------------------------------------------------------------
 # Decide whether the presence of user's exception handling functions is
 # checked and data conversion exceptions are returned.  This is mainly
 # for the speed optimization of hard conversions.  Soft conversions can
@@ -847,3 +860,83 @@ if (HDF5_BUILD_FORTRAN)
   message (STATUS "maximum decimal precision for C var - ${${HDF_PREFIX}_PAC_C_MAX_REAL_PRECISION}")
 
 endif()
+
+#-----------------------------------------------------------------------------
+# Macro to determine the various conversion capabilities
+#-----------------------------------------------------------------------------
+macro (H5ConversionTests TEST msg)
+  if (NOT DEFINED ${TEST})
+    TRY_RUN (${TEST}_RUN   ${TEST}_COMPILE
+        ${CMAKE_BINARY_DIR}
+        ${HDF_RESOURCES_DIR}/ConversionTests.c
+        CMAKE_FLAGS -DCOMPILE_DEFINITIONS:STRING=-D${TEST}_TEST
+        OUTPUT_VARIABLE OUTPUT
+    )
+    if (${TEST}_COMPILE)
+      if (${TEST}_RUN EQUAL "0")
+        set (${TEST} 1 CACHE INTERNAL ${msg})
+        message (VERBOSE "${msg}... yes")
+      else ()
+        set (${TEST} "" CACHE INTERNAL ${msg})
+        message (VERBOSE "${msg}... no")
+        file (APPEND ${CMAKE_BINARY_DIR}/CMakeFiles/CMakeError.log
+              "Test ${TEST} Run failed with the following output and exit code:\n ${OUTPUT}\n"
+        )
+      endif ()
+    else ()
+      set (${TEST} "" CACHE INTERNAL ${msg})
+      message (VERBOSE "${msg}... no")
+      file (APPEND ${CMAKE_BINARY_DIR}/CMakeFiles/CMakeError.log
+          "Test ${TEST} Compile failed with the following output:\n ${OUTPUT}\n"
+      )
+    endif ()
+
+  endif ()
+endmacro ()
+
+#-----------------------------------------------------------------------------
+# Check various conversion capabilities
+#-----------------------------------------------------------------------------
+
+# ----------------------------------------------------------------------
+# Set the flag to indicate that the machine is using a special algorithm toconvert
+# 'long double' to '(unsigned) long' values.  (This flag should only be set for
+# the IBM Power Linux.  When the bit sequence of long double is
+# 0x4351ccf385ebc8a0bfcc2a3c3d855620, the converted value of (unsigned)long
+# is 0x004733ce17af227f, not the same as the library's conversion to 0x004733ce17af2282.
+# The machine's conversion gets the correct value.  We define the macro and disable
+# this kind of test until we figure out what algorithm they use.
+#-----------------------------------------------------------------------------
+H5ConversionTests (${HDF_PREFIX}_LDOUBLE_TO_LONG_SPECIAL  "Checking IF your system converts long double to (unsigned) long values with special algorithm")
+# ----------------------------------------------------------------------
+# Set the flag to indicate that the machine is using a special algorithm
+# to convert some values of '(unsigned) long' to 'long double' values.
+# (This flag should be off for all machines, except for IBM Power Linux,
+# when the bit sequences are 003fff..., 007fff..., 00ffff..., 01ffff...,
+# ..., 7fffff..., the compiler uses a unknown algorithm.  We define a
+# macro and skip the test for now until we know about the algorithm.
+#-----------------------------------------------------------------------------
+H5ConversionTests (${HDF_PREFIX}_LONG_TO_LDOUBLE_SPECIAL "Checking IF your system can convert (unsigned) long to long double values with special algorithm")
+# ----------------------------------------------------------------------
+# Set the flag to indicate that the machine can accurately convert
+# 'long double' to '(unsigned) long long' values.  (This flag should be set for
+# all machines, except for Mac OS 10.4 and SGI IRIX64 6.5.  When the bit sequence
+# of long double is 0x4351ccf385ebc8a0bfcc2a3c..., the values of (unsigned)long long
+# start to go wrong on these two machines.  Adjusting it higher to
+# 0x4351ccf385ebc8a0dfcc... or 0x4351ccf385ebc8a0ffcc... will make the converted
+# values wildly wrong.  This test detects this wrong behavior and disable the test.
+#-----------------------------------------------------------------------------
+H5ConversionTests (${HDF_PREFIX}_LDOUBLE_TO_LLONG_ACCURATE "Checking IF correctly converting long double to (unsigned) long long values")
+# ----------------------------------------------------------------------
+# Set the flag to indicate that the machine can accurately convert
+# '(unsigned) long long' to 'long double' values.  (This flag should be set for
+# all machines, except for Mac OS 10.4, when the bit sequences are 003fff...,
+# 007fff..., 00ffff..., 01ffff..., ..., 7fffff..., the converted values are twice
+# as big as they should be.
+#-----------------------------------------------------------------------------
+H5ConversionTests (${HDF_PREFIX}_LLONG_TO_LDOUBLE_CORRECT "Checking IF correctly converting (unsigned) long long to long double values")
+# ----------------------------------------------------------------------
+# Set the flag to indicate that the machine can accurately convert
+# some long double values
+#-----------------------------------------------------------------------------
+H5ConversionTests (${HDF_PREFIX}_DISABLE_SOME_LDOUBLE_CONV "Checking IF the cpu is power9 and cannot correctly converting long double values")
diff --git a/config/cmake/ConversionTests.c b/config/cmake/ConversionTests.c
new file mode 100644
index 0000000..725f049
--- /dev/null
+++ b/config/cmake/ConversionTests.c
@@ -0,0 +1,287 @@
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
+ * Copyright by The HDF Group.                                               *
+ * All rights reserved.                                                      *
+ *                                                                           *
+ * This file is part of HDF5.  The full HDF5 copyright notice, including     *
+ * terms governing use, modification, and redistribution, is contained in    *
+ * the COPYING file, which can be found at the root of the source code       *
+ * distribution tree, or in https://www.hdfgroup.org/licenses.               *
+ * If you do not have access to either file, you may request a copy from     *
+ * help@hdfgroup.org.                                                        *
+ * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+
+#if defined(__has_attribute)
+#   if __has_attribute(no_sanitize)
+#       define HDF_NO_UBSAN __attribute__((no_sanitize("undefined")))
+#   else
+#       define HDF_NO_UBSAN
+#   endif
+#else
+#   define HDF_NO_UBSAN
+#endif
+
+#ifdef H5_LDOUBLE_TO_LONG_SPECIAL_TEST
+
+#include <string.h>
+#include <stdlib.h>
+
+int HDF_NO_UBSAN main(void)
+{
+    long double         ld = 20041683600089727.779961L;
+    long                ll;
+    unsigned long       ull;
+    unsigned char       s[16];
+    unsigned char       s2[8];
+    int                 ret = 1;
+
+    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, the bit sequence should be
+         * like below.  It's not a decent way to check but this info
+         * isn't available.
+         */
+        memcpy(s, &ld, 16);
+        if (s[0]==0x43 && s[1]==0x51 && s[2]==0xcc && s[3]==0xf3 &&
+            s[4]==0x85 && s[5]==0xeb && s[6]==0xc8 && s[7]==0xa0 &&
+            s[8]==0xbf && s[9]==0xcc && s[10]==0x2a && s[11]==0x3c) {
+
+            /* Assign the hexadecimal value of long double type. */
+            s[0]=0x43; s[1]=0x51; s[2]=0xcc; s[3]=0xf3;
+            s[4]=0x85; s[5]=0xeb; s[6]=0xc8; s[7]=0xa0;
+            s[8]=0xbf; s[9]=0xcc; s[10]=0x2a; s[11]=0x3c;
+            s[12]=0x3d; s[13]=0x85; s[14]=0x56; s[15]=0x20;
+
+            memcpy(&ld, s, 16);
+
+            ll = (long)ld;
+            memcpy(s2, &ll, 8);
+
+            /* The library's algorithm converts it to 0x00 47 33 ce 17 af 22 82
+             * and gets wrong value 20041683600089730 on Linux on IBM Power
+             * architecture.
+             *
+             * But Linux on IBM Power converts it to 0x00 47 33 ce 17 af 22 7f
+             * and gets the correct value 20041683600089727.  It uses some special
+             * algorithm.  We're going to define the macro and skip the test until
+             * we can figure out how they do it.
+             */
+            if (s2[0]==0x00 && s2[1]==0x47 && s2[2]==0x33 && s2[3]==0xce &&
+                s2[4]==0x17 && s2[5]==0xaf && s2[6]==0x22 && s2[7]==0x7f)
+
+                ret = 0;
+
+            ull = (unsigned long)ld;
+            memcpy(s2, &ull, 8);
+
+            /* The unsigned long is the same as signed long */
+            if(s2[0]==0x00 && s2[1]==0x47 && s2[2]==0x33 && s2[3]==0xce &&
+               s2[4]==0x17 && s2[5]==0xaf && s2[6]==0x22 && s2[7]==0x7f)
+
+                ret = 0;
+        }
+    }
+
+done:
+    exit(ret);
+}
+
+#endif
+
+#ifdef H5_LONG_TO_LDOUBLE_SPECIAL_TEST
+
+#include <string.h>
+#include <stdlib.h>
+
+int HDF_NO_UBSAN main(void)
+{
+    long double         ld;
+    long                ll;
+    unsigned long       ull;
+    unsigned char       s[16];
+    int                 flag=0, ret=1;
+
+    /* Determine if long double has 16 byte in size, 11 bit exponent, and
+     * the bias is 0x3ff
+     */
+    if (sizeof(long double) == 16) {
+        ld = 1.0L;
+        memcpy(s, &ld, 16);
+
+        if (s[0]==0x3f && s[1]==0xf0 && s[2]==0x00 && s[3]==0x00 &&
+            s[4]==0x00 && s[5]==0x00 && s[6]==0x00 && s[7]==0x00)
+
+            flag = 1;
+    }
+
+    if (flag==1 && sizeof(long)==8) {
+        ll = 0x003fffffffffffffL;
+        ld = (long double)ll;
+        memcpy(s, &ld, 16);
+
+        /* The library converts the value to 0x434fffffffffffff8000000000000000.
+         * In decimal it is 18014398509481982.000000, one value short of the original.
+         *
+         * Linux on IBM Power architecture converts it to
+         * 0x4350000000000000bff0000000000000. The value is correct in decimal.
+         * It uses some special algorithm.  We're going to define the macro and
+         * skip the test until we can figure out how they do it.
+         */
+        if (s[0]==0x43 && s[1]==0x50 && s[2]==0x00 && s[3]==0x00 &&
+            s[4]==0x00 && s[5]==0x00 && s[6]==0x00 && s[7]==0x00 &&
+            s[8]==0xbf && s[9]==0xf0 && s[10]==0x00 && s[11]==0x00 &&
+            s[12]==0x00 && s[13]==0x00 && s[14]==0x00 && s[15]==0x00)
+
+            ret = 0;
+    }
+
+    if (flag==1 && sizeof(unsigned long)==8) {
+        ull = 0xffffffffffffffffUL;
+        ld = (long double)ull;
+        memcpy(s, &ld, 16);
+
+        /* Use a different value from signed long to test. The problem is the
+         * same for both long and unsigned long. The value is 18446744073709551615.
+         * The library converts the value to 0x43effffffffffffffe000000000000000.
+         * In decimal it's 18446744073709548544.000000, very different from the
+         * original. Linux on IBM Power architecture converts it to
+         * 0x43f0000000000000bff0000000000000. The value is correct in decimal.
+         * It uses some special algorithm.  We're going to define the macro and
+         * skip the test until we can figure out how they do it.
+         */
+        if (s[0]==0x43 && s[1]==0xf0 && s[2]==0x00 && s[3]==0x00 &&
+            s[4]==0x00 && s[5]==0x00 && s[6]==0x00 && s[7]==0x00 &&
+            s[8]==0xbf && s[9]==0xf0 && s[10]==0x00 && s[11]==0x00 &&
+            s[12]==0x00 && s[13]==0x00 && s[14]==0x00 && s[15]==0x00)
+
+            ret = 0;
+        }
+done:
+    exit(ret);
+}
+
+#endif
+
+#ifdef H5_LDOUBLE_TO_LLONG_ACCURATE_TEST
+
+#include <string.h>
+#include <stdlib.h>
+
+int HDF_NO_UBSAN main(void)
+{
+    long double         ld = 20041683600089727.779961L;
+    long long           ll;
+    unsigned long long  ull;
+    unsigned char       s[16];
+    int                 ret = 0;
+
+    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,
+         * the bit sequence should be like below.  It's not
+         * a decent way to check but this info isn't available.
+         */
+        memcpy(s, &ld, 16);
+
+        if (s[0]==0x43 && s[1]==0x51 && s[2]==0xcc && s[3]==0xf3 &&
+            s[4]==0x85 && s[5]==0xeb && s[6]==0xc8 && s[7]==0xa0 &&
+            s[8]==0xbf && s[9]==0xcc && s[10]==0x2a && s[11]==0x3c) {
+
+            /* Slightly adjust the bit sequence (s[8]=0xdf).  The converted
+             * values will go wild on Mac OS 10.4 and IRIX64 6.5.
+             */
+            s[0]=0x43; s[1]=0x51; s[2]=0xcc; s[3]=0xf3;
+            s[4]=0x85; s[5]=0xeb; s[6]=0xc8; s[7]=0xa0;
+            s[8]=0xdf; s[9]=0xcc; s[10]=0x2a; s[11]=0x3c;
+            s[12]=0x3d; s[13]=0x85; s[14]=0x56; s[15]=0x20;
+
+            memcpy(&ld, s, 16);
+            ll = (long long)ld;
+            ull = (unsigned long long)ld;
+
+            if (ll != 20041683600089728 || ull != 20041683600089728)
+                ret = 1;
+        }
+    }
+done:
+    exit(ret);
+}
+#endif
+
+#ifdef H5_LLONG_TO_LDOUBLE_CORRECT_TEST
+
+#include <string.h>
+#include <stdlib.h>
+
+int HDF_NO_UBSAN main(void)
+{
+    long double         ld;
+    long long           ll;
+    unsigned long long  ull;
+    unsigned char       s[16];
+    int                 flag=0, ret=0;
+
+    /* Determine if long double has 16 byte in size, 11 bit exponent, and
+     * the bias is 0x3ff
+     */
+    if (sizeof(long double) == 16) {
+        ld = 1.0L;
+        memcpy(s, &ld, 16);
+        if (s[0]==0x3f && s[1]==0xf0 && s[2]==0x00 && s[3]==0x00 &&
+            s[4]==0x00 && s[5]==0x00 && s[6]==0x00 && s[7]==0x00)
+
+            flag = 1;
+    }
+
+    if (flag==1 && sizeof(long long)==8) {
+        ll = 0x01ffffffffffffffLL;
+        ld = (long double)ll;
+        memcpy(s, &ld, 16);
+
+        /* Check if the bit sequence is as expected*/
+        if (s[0]!=0x43 || s[1]!=0x7f || s[2]!=0xff || s[3]!=0xff ||
+            s[4]!=0xff || s[5]!=0xff || s[6]!=0xff || s[7]!=0xff ||
+            s[8]!=0xf0 || s[9]!=0x00 || s[10]!=0x00 || s[11]!=0x00)
+
+            ret = 1;
+    }
+    if (flag==1 && sizeof(unsigned long long)==8) {
+        ull = 0x01ffffffffffffffULL;
+        ld = (long double)ull;
+        memcpy(s, &ld, 16);
+
+        if (s[0]!=0x43 || s[1]!=0x7f || s[2]!=0xff || s[3]!=0xff ||
+            s[4]!=0xff || s[5]!=0xff || s[6]!=0xff || s[7]!=0xff ||
+            s[8]!=0xf0 || s[9]!=0x00 || s[10]!=0x00 || s[11]!=0x00)
+
+            ret = 1;
+    }
+
+done:
+    exit(ret);
+}
+#endif
+
+#ifdef H5_DISABLE_SOME_LDOUBLE_CONV_TEST
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+int HDF_NO_UBSAN main(void)
+{
+    FILE *fp;
+    char cpu[64];
+
+    fp = popen("uname -m", "r");
+
+    fgets(cpu, sizeof(cpu)-1, fp);
+
+    pclose(fp);
+
+    if (strncmp(cpu, "ppc64le", 7) == 0)
+        return 0;
+
+    return 1;
+}
+
+#endif
diff --git a/config/cmake/H5pubconf.h.in b/config/cmake/H5pubconf.h.in
index 444a0e4..d4b0d36 100644
--- a/config/cmake/H5pubconf.h.in
+++ b/config/cmake/H5pubconf.h.in
@@ -32,6 +32,10 @@
 /* Define if dev_t is a scalar */
 #cmakedefine H5_DEV_T_IS_SCALAR @H5_DEV_T_IS_SCALAR@
 
+/* Define if your system is IBM ppc64le and cannot convert some long double
+   values correctly. */
+#cmakedefine H5_DISABLE_SOME_LDOUBLE_CONV @H5_DISABLE_SOME_LDOUBLE_CONV@
+
 /* Define to dummy `main' function (if any) required to link to the Fortran
    libraries. */
 #cmakedefine H5_FC_DUMMY_MAIN @H5_FC_DUMMY_MAIN@
@@ -371,6 +375,22 @@
 /* Define if new-style references should be used with dimension scales */
 #cmakedefine H5_DIMENSION_SCALES_WITH_NEW_REF @H5_DIMENSION_SCALES_WITH_NEW_REF@
 
+/* Define if your system can convert long double to (unsigned) long long
+   values correctly. */
+#cmakedefine H5_LDOUBLE_TO_LLONG_ACCURATE @H5_LDOUBLE_TO_LLONG_ACCURATE@
+
+/* Define if your system converts long double to (unsigned) long values with
+   special algorithm. */
+#cmakedefine H5_LDOUBLE_TO_LONG_SPECIAL @H5_LDOUBLE_TO_LONG_SPECIAL@
+
+/* Define if your system can convert (unsigned) long long to long double
+   values correctly. */
+#cmakedefine H5_LLONG_TO_LDOUBLE_CORRECT @H5_LLONG_TO_LDOUBLE_CORRECT@
+
+/* Define if your system can convert (unsigned) long to long double values
+   with special algorithm. */
+#cmakedefine H5_LONG_TO_LDOUBLE_SPECIAL @H5_LONG_TO_LDOUBLE_SPECIAL@
+
 /* Define to the sub-directory where libtool stores uninstalled libraries. */
 #cmakedefine H5_LT_OBJDIR @H5_LT_OBJDIR@
 
@@ -599,6 +619,9 @@
 /* Version number of package */
 #define H5_VERSION "@HDF5_PACKAGE_VERSION_STRING@"
 
+/* Data accuracy is preferred to speed during data conversions */
+#cmakedefine H5_WANT_DATA_ACCURACY @H5_WANT_DATA_ACCURACY@
+
 /* Check exception handling functions during data conversions */
 #cmakedefine H5_WANT_DCONV_EXCEPTION @H5_WANT_DCONV_EXCEPTION@