Skip to content

Commit

Permalink
Disable FP exceptions in H5T init code (HDFGroup#3837)
Browse files Browse the repository at this point in the history 
The H5T floating-point datatype initialization code can raise exceptions when handling signaling NaNs. This change disables FE_INVALID exceptions during initialization.

Also removes the -ieee=full change for NAG Fortran as that shouldn't be necessary anymore.

Fixes HDFGroup#3831
  • Loading branch information
@derobins @jhendersonHDF
derobins authored and jhendersonHDF committed Dec 7, 2023
1 parent ada0656 commit 4722808
Show file tree
Hide file tree
Showing 3 changed files with 53 additions and 24 deletions.
22 changes: 16 additions & 6 deletions release_docs/RELEASE.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,22 @@ Bug Fixes since HDF5-1.14.3 release
===================================
Library
-------
- Suppressed floating-point exceptions in H5T init code

The floating-point datatype initialization code in H5Tinit_float.c
could raise FE_INVALID exceptions while munging bits and performing
comparisons that might involve NaN. This was not a problem when the
initialization code was executed in H5detect at compile time (prior
to 1.14.3), but now that the code is executed at library startup
(1.14.3+), these exceptions can be caught by user code, as is the
default in the NAG Fortran compiler.

Starting in 1.14.4, we now suppress floating-point exceptions while
initializing the floating-point types and clear FE_INVALID before
restoring the original environment.

Fixes GitHub #3831

- Fixed a file handle leak in the core VFD

When opening a file with the core VFD and a file image, if the file
Expand Down Expand Up @@ -339,12 +355,6 @@ Known Problems
-O1 for the time being:
https://forums.developer.nvidia.com/t/hdf5-no-longer-compiles-with-nv-23-9/269045.

IEEE standard arithmetic enables software to raise exceptions such as overflow,
division by zero, and other illegal operations without interrupting or halting
the program flow. The HDF5 C library intentionally performs these exceptions.
Therefore, the "-ieee=full" nagfor switch is necessary when compiling a program
to avoid stopping on an exception.

CMake files do not behave correctly with paths containing spaces.
Do not use spaces in paths because the required escaping for handling spaces
results in very complex and fragile build files.
Expand Down
54 changes: 36 additions & 18 deletions src/H5Tinit_float.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -51,19 +51,23 @@
* Function: DETECT_F
*
* Purpose: This macro takes a floating point type like `double' and
* a base name like `natd' and detects byte order, mantissa
* location, exponent location, sign bit location, presence or
* absence of implicit mantissa bit, and exponent bias and
* initializes a detected_t structure with those properties.
* and detects byte order, mantissa location, exponent location,
* sign bit location, presence or absence of implicit mantissa
* bit, and exponent bias and initializes a detected_t structure
* with those properties.
*
* Note that these operations can raise floating-point
* exceptions and building with some compiler options
* (especially Fortran) can cause problems.
*-------------------------------------------------------------------------
*/
#define DETECT_F(TYPE, VAR, INFO) \
#define DETECT_F(TYPE, INFO) \
do { \
TYPE _v1, _v2, _v3; \
unsigned char _buf1[sizeof(TYPE)], _buf3[sizeof(TYPE)]; \
unsigned char _pad_mask[sizeof(TYPE)]; \
unsigned char _byte_mask; \
int _i, _j, _last = (-1); \
TYPE _v1, _v2, _v3; \
uint8_t _buf1[sizeof(TYPE)], _buf3[sizeof(TYPE)]; \
uint8_t _pad_mask[sizeof(TYPE)]; \
uint8_t _byte_mask; \
int _i, _j, _last = -1; \
\
memset(&INFO, 0, sizeof(INFO)); \
INFO.size = sizeof(TYPE); \
Expand All @@ -81,7 +85,7 @@
_v1 = (TYPE)4.0L; \
H5MM_memcpy(_buf1, (const void *)&_v1, sizeof(TYPE)); \
for (_i = 0; _i < (int)sizeof(TYPE); _i++) \
for (_byte_mask = (unsigned char)1; _byte_mask; _byte_mask = (unsigned char)(_byte_mask << 1)) { \
for (_byte_mask = (uint8_t)1; _byte_mask; _byte_mask = (uint8_t)(_byte_mask << 1)) { \
_buf1[_i] ^= _byte_mask; \
H5MM_memcpy((void *)&_v2, (const void *)_buf1, sizeof(TYPE)); \
H5_GCC_CLANG_DIAG_OFF("float-equal") \
Expand Down Expand Up @@ -118,20 +122,19 @@
_v1 = (TYPE)1.0L; \
_v2 = (TYPE)-1.0L; \
if (H5T__bit_cmp(sizeof(TYPE), INFO.perm, &_v1, &_v2, _pad_mask, &(INFO.sign)) < 0) \
HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "failed to detect byte order"); \
HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "failed to determine sign bit"); \
\
/* Mantissa */ \
INFO.mpos = 0; \
\
_v1 = (TYPE)1.0L; \
_v2 = (TYPE)1.5L; \
if (H5T__bit_cmp(sizeof(TYPE), INFO.perm, &_v1, &_v2, _pad_mask, &(INFO.msize)) < 0) \
HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "failed to detect byte order"); \
HGOTO_ERROR(H5E_DATATYPE, H5E_CANTINIT, FAIL, "failed to determine mantissa"); \
INFO.msize += 1 + (unsigned)(INFO.imp ? 0 : 1) - INFO.mpos; \
\
/* Exponent */ \
INFO.epos = INFO.mpos + INFO.msize; \
\
INFO.epos = INFO.mpos + INFO.msize; \
INFO.esize = INFO.sign - INFO.epos; \
\
_v1 = (TYPE)1.0L; \
Expand Down Expand Up @@ -456,17 +459,24 @@ H5T__set_precision(H5T_fpoint_det_t *d)
herr_t H5_NO_UBSAN
H5T__init_native_float_types(void)
{
fenv_t saved_fenv;
H5T_fpoint_det_t det;
H5T_t *dt = NULL;
herr_t ret_value = SUCCEED;

FUNC_ENTER_PACKAGE

/* Turn off floating-point exceptions while initializing to avoid
* tripping over signaling NaNs while looking at "don't care" bits.
*/
if (feholdexcept(&saved_fenv) != 0)
HSYS_GOTO_ERROR(H5E_DATATYPE, H5E_CANTSET, FAIL, "can't save floating-point environment");

/* H5T_NATIVE_FLOAT */

/* Get the type's characteristics */
memset(&det, 0, sizeof(H5T_fpoint_det_t));
DETECT_F(float, FLOAT, det);
DETECT_F(float, det);

/* Allocate and fill type structure */
if (NULL == (dt = H5T__alloc()))
Expand Down Expand Up @@ -497,7 +507,7 @@ H5T__init_native_float_types(void)

/* Get the type's characteristics */
memset(&det, 0, sizeof(H5T_fpoint_det_t));
DETECT_F(double, DOUBLE, det);
DETECT_F(double, det);

/* Allocate and fill type structure */
if (NULL == (dt = H5T__alloc()))
Expand Down Expand Up @@ -528,7 +538,7 @@ H5T__init_native_float_types(void)

/* Get the type's characteristics */
memset(&det, 0, sizeof(H5T_fpoint_det_t));
DETECT_F(long double, LDOUBLE, det);
DETECT_F(long double, det);

/* Allocate and fill type structure */
if (NULL == (dt = H5T__alloc()))
Expand Down Expand Up @@ -561,6 +571,14 @@ H5T__init_native_float_types(void)
H5T_native_order_g = det.order;

done:
/* Clear any FE_INVALID exceptions from NaN handling */
if (feclearexcept(FE_INVALID) != 0)
HSYS_GOTO_ERROR(H5E_DATATYPE, H5E_CANTSET, FAIL, "can't clear floating-point exceptions");

/* Restore the original environment */
if (feupdateenv(&saved_fenv) != 0)
HSYS_GOTO_ERROR(H5E_DATATYPE, H5E_CANTSET, FAIL, "can't restore floating-point environment");

if (ret_value < 0) {
if (dt != NULL) {
dt->shared = H5FL_FREE(H5T_shared_t, dt->shared);
Expand Down
1 change: 1 addition & 0 deletions src/H5private.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@
#include <ctype.h>
#include <errno.h>
#include <fcntl.h>
#include <fenv.h>
#include <float.h>
#include <math.h>
#include <setjmp.h>
Expand Down

0 comments on commit 4722808

Please sign in to comment.