sigfpe(3C)
handle_sigfpes - Floating-point exception handler package
As shipped in IRIX 6.5.22. Last changed in IRIX 6.5.19.
NAME handle_sigfpes - Floating-point exception handler package SYNOPSIS #include <sigfpe.h> void handle_sigfpes (int onoff, int en_mask, void (*user_routine)(unsigned[5], int[2]), int abort_action, void (*abort_routine)(unsigned int **)); void (*user_routine)(unsigned[5], int[2]), int abort_action, void (*abort_routine)(unsigned int **)); typedef void (*user_t)(unsigned[5], int[2]); typedef void (*abort_t)(unsigned int **); struct sigfpe_template { int repls; int count; int trace; int abort; int exit; }; extern struct sigfpe_template sigfpe_[_N_EXCEPTION_TYPES+1]; extern int invalidop_results_[_N_INVALIDOP_RESULTS+1]; extern int invalidop_operands_[_N_INVALIDOP_OPERANDS+1]; DESCRIPTION The floating-point accelerator may raise floating-point exceptions, signal SIGFPE, due to five conditions: _OVERFL(overflow), _UNDERFL(underflow), _DIVZERO(divide-by-zero), _INEXACT(inexact result), or _INVALID(invalid operand, e.g., infinity). Usually these conditions are masked, and do not cause a floating-point exception. Instead, a default value is substituted for the result of the operation, and the program continues silently. This event may be intercepted by causing an exception to be raised. When this occurs, the operating system generates a SIGFPE signal. The integer arithmetic instructions add, addi, dadd, daddi, sub, and dsub also generate a SIGFPE signal when the result of the operation overflows (condition _INT_OVERFL). (Currently, SGI compilers generate only unsigned versions of these instructions, which do not generate a signal on overflow. However, it is still possible to generate these instructions via assembly language). For -o32 programs, the compiler generates additional instructions to detect and trap on integer-divide-by-zero. No similar detection code sequences are generated in -n32 or -64 programs. Once an exception is raised, the specific conditions which caused the exception may be determined, and more appropriate action taken. The libfpe.so library provides two methods to unmask and handle these conditions: the subroutine handle_sigfpes, and the environment variable TRAP_FPE. Both methods provide a mechanism for unmasking each of these conditions except _INEXACT, for handling and classifying exceptions arising from them, and for substituting either a default value or a chosen one. They also provide mechanisms to count, trace, exit or abort on enabled exceptions. If you supply your own call to handle_sigfpes, you should leave environment variable TRAP_FPE undefined or set to OFF. TRAP_FPE is supported for Fortran, C, C++ and Pascal. handle_sigfpes is supported for C, C++ and Fortran. Calling the subroutine is the preferred method when preparing software for others to use, since it relieves the user of any need to know about the TRAP_FPE environment variable. The environment variable is preferable if you want to experiment or allow any user to experiment with different trap behaviors with minimum effort. libfpe uses System V signal handling and will not work with programs that use Berkeley signal handling. Note that the preferred method for flushing denormals to zero is to set the FS bit to 1 in the floating point control status register. If the FS bit is set to 1, the user should not handle underflows through the floating point trap handler, because a denormal result flushed to zero will cause an underflow. Note that calls to the floating point trap handler consume many thousands of cpu cycles. The following routine (written in C) can be used to set/clear the FS bit. #include <sys/fpu.h> void flush_to_zero(int on_off) { union fpc_csr n; n.fc_word = get_fpc_csr(); if ( on_off == 0 ) { n.fc_struct.flush = 0; } else { n.fc_struct.flush = 1; } set_fpc_csr(n.fc_word); } Note that if the FS bit in the floating point control status register is set, it remains set after calling handle_sigfpes(_OFF, ... Be sure to read the NOTES FOR R8000 section on this man page, which describes the behavior of libfpe.so on the R8000 processor. HANDLE_SIGFPES SUBROUTINE The values in the global arrays are described in the "WHEN AN EXCEPTION IS ENCOUNTERED" section on this man page. The arguments to handle_sigfpes are as follows: onoff Indicates whether handling is turned on (onoff == _ON) or off (onoff == _OFF). Specifying (onoff == _DEBUG) is another way to turn on handling. Information from the sigfpe structure will be printed if (onoff == _DEBUG). (The names used in this document are defined in sigfpe.h.) en_mask Indicates which of the five conditions should be unmasked, enabling them to raise floating-point exceptions/or an integer overflow exception. en_mask is valid only if onoff == _ON or onoff == _DEBUG, and is the bitwise or of one or more of the constants _EN_UNDERFL, _EN_OVERFL,_EN_DIVZERO, _EN_INVALID, and _EN_INT_OVERFL (defined in sigfpe.h). user_routine: handle_sigfpes provides a mechanism for setting the result of the operation to any one of a set of well-known values. If full control over the value of selected operations is desired for one or more exception conditions, a function user_routine must be provided. For these selected exception conditions, user_routine will be called to set the value resulting from the operation. Pass a 0 (plain 0 is adequate) if you do not want to provide a user_routine. Typedef user_t is defined in sigfpe.h for convenience in programming. abort_action: If the handler encounters an unexpected condition or an inconsistency, the flag abort_action indicates what action should be taken. Another option is for you to specify that you are supplying your own floating point exception handler as the default handler. The following legal values can be specified: _TURN_OFF_HANDLER_ON_ERROR Instruct the floating-point-accelerator to cease causing exceptions and continue (i.e., disable handling). _ABORT_ON_ERROR Kill the process after giving an error message and calling a user-supplied cleanup routine if one is provided via the abort_routine parameter. _REPLACE_HANDLER_ON_ERROR Install the indicated user routine as the handler when such an error is encountered. Future floating-point exceptions will branch to the user-routine (see signal(2)). _USER_HANDLER Install the indicated user routine as the handler immediately. Future floating-point exceptions will branch to the user-routine (see signal(2)). abort_routine: When a fatal error (i.e., one described under abort_action) is encountered, abort_routine is used as the address of a user routine. If abort_action is _ABORT_ON_ERROR, and abort_routine is valid, it is called before aborting, and passed a pointer to the address of the instruction causing the exception as its single argument. In this case, the user's abort_routine should be defined as follows: void abort_routine( ptr_to_pc ) unsigned int **ptr_to_pc; If abort_action is _REPLACE_HANDLER_ON_ERROR, and abort_routine is valid, it will be installed as the new handler. In this case, abort_routine will be called immediately to handle the current exception. (see signal(2)) Pass a 0 (plain 0 is adequate) if you do not want to provide an abort_action routine. If abort_action is _USER_HANDLER, and abort_routine is valid, it will be installed immediately as the default floating point exception handler. If abort_routine is to be invoked as a floating point exception handler, the following prototype should be used (see <sigfpe.h>, <sys/signal.h>, and signal(2)): int user_handler( sig, code, sc ) int sig, code; struct sigcontext *sc; Typedef abort_t is defined in sigfpe.h for convenience in programming. Routine user_handler should return 0 to continue processing of the user's code and a non-zero value to disconnect user_handler as the floating point exception handler. User_handler should not issue a call to signal(), nor should it update the program counter in the sigcontext area, because these actions are done in the routine which calls it. A user handler can determine which type of exception has occurred by calling routine __fpe_trap_type(). This routine returns one of _UNDERFL, _OVERFL, etc. as appropriate. Users can supply separate handlers for each exception type by making multiple calls to handle_sigfpes. See example 3 that follows. Similarly, trap handling for one or more exception types can be turned off by or-ing the appropriate combination of masks in the second parameter to handle_sigfpes. handle_sigfpes(OFF, 0, ... disables handling of all SIGFPE signals. EXAMPLES OF CALLS TO HANDLE_SIGFPES The following example sets up traps of underflow which change the resulting value to zero: #include <sigfpe.h> /* call this during program startup to set underflowing values to zero */ void my_underflow_to_zero(void) { /* underflow to zero */ sigfpe_[_UNDERFL].repls = _ZERO; /* only trap on underflow */ handle_sigfpes(_ON, _EN_UNDERFL , 0, _ABORT_ON_ERROR, 0); } The following example counts all traps, traces the first five exceptions of each kind, aborts on the first divide by zero, or the 100th overflow. It replaces zero for underflows, max float/double for overflows, max integer for integer overflows, and the default values for divide by zero, invalid operands, and integer overflows. The environment variable example below does the same thing. #include <limits.h> /* to get INT_MAX */ #include <sigfpe.h> main() { /* underflow to zero */ sigfpe_[_UNDERFL].repls = _ZERO; /* substitute max float/double on overflow */ sigfpe_[_OVERFL].repls=_MAX; /* trace first 5 exceptions of each kind */ sigfpe_[_UNDERFL].trace=5; sigfpe_[_OVERFL].trace =5; sigfpe_[_DIVZERO].trace=5; sigfpe_[_INVALID].trace=5; sigfpe_[_INT_OVERFL].trace=5; /* counts at end */ sigfpe_[_UNDERFL].count=INT_MAX; sigfpe_[_OVERFL].count =INT_MAX; sigfpe_[_DIVZERO].count=INT_MAX; sigfpe_[_INVALID].count=INT_MAX; sigfpe_[_INT_OVERFL].count=INT_MAX; /* abort after 100 */ sigfpe_[_UNDERFL].abort=100; sigfpe_[_OVERFL].abort =100; sigfpe_[_INVALID].abort=100; sigfpe_[_INT_OVERFL].abort=100; /* abort on first divide by zero */ sigfpe_[_DIVZERO].abort=1; handle_sigfpes(_ON, _EN_UNDERFL| _EN_OVERFL|_EN_DIVZERO | _EN_INVALID | _EN_INT_OVERFL, 0, _ABORT_ON_ERROR, 0); /* do the real application work here */ } The last example shows how to enable different handlers for various exception types: #include <limits.h> /* to get INT_MAX */ #include <sigfpe.h> extern user_t my_invalid_handler; main() { /* counts at end */ sigfpe_[_UNDERFL].count=INT_MAX; sigfpe_[_OVERFL].count =INT_MAX; sigfpe_[_DIVZERO].count=INT_MAX; sigfpe_[_INVALID].count=INT_MAX; sigfpe_[_INT_OVERFL].count=INT_MAX; /* enable trapping on overflow, using libfpe's trap handler */ handle_sigfpes(_ON, _EN_OVERFL, 0, 0, 0); /* enable trapping on invalid, using user's own trap handler */ handle_sigfpes(_ON, _EN_INVALID, my_invalid_handler, 0, 0); /* do the real application work here */ .... /* turn off trapping of overflows */ handle_sigfpes(_OFF, _EN_OVERFL, 0, 0, 0); .... /* turn off all handling of SIGFPE signals */ handle_sigfpes(_OFF, 0, 0, 0, 0); .... The TRAP_FPE Environment Variable If the code has been linked with libfpe.so, the runtime startup routine will check for the environment variable TRAP_FPE. The string read as the value of TRAP_FPE will be interpreted and handle_sigfpes will be called with the resulting values. TRAP_FPE is read in uppercase letters only. The string assigned to TRAP_FPE may be in uppercase or lowercase. TRAP_FPE can take one of two forms: either a global value, or a list of individual items. Global values: "" or OFF Execute the program with no trap handling enabled. Same as TRAP_FPE undefined. Same as linking without libfpe.so ON Same as TRAP_FPE="ALL=DEFAULT". Alternately, replacement values and actions may be specified for each of the possible trap types individually. This is accomplished by setting the environment variable as follows: setenv TRAP_FPE "item;item;item...." An item can be one of the following: traptype=statuslist Where traptype defines the specific floating point exception to enable, and statuslist defines the list of actions upon encountering the trap. DEBUG Confirm the parsing of the environment variable and the trap actions. Traptype can be one of the following literal strings: UNDERFL underflow OVERFL overflow DIVZERO divide by zero INVALID invalid operand INT_OVERFL integer overflow ALL all of the above Statuslist is a list separated by commas. It contains an optional symbolic replacement value, and an optional list of actions. Symbolic replacement values: DEFAULT Do not override the predefined default values. IEEE Maps to integer code _APPROPRIATE. APPROPRIATE Maps to integer code _APPROPRIATE. ZERO Maps to integer code _ZERO. FLUSH_ZERO Maps to integer code _FLUSH_ZERO (R4000 and later processors). FLUSH_ZERO Maps to integer code _ZERO (other processors). MIN Maps to integer code _MIN. MAX Maps to integer code _MAX. INF Maps to integer code _INF. NAN Maps to integer code _NAN. All actions take an optional integer in parentheses: Note: for any traps that have an action and no specified replacement value, the DEFAULT replacement value will be used. COUNT(n) A count of the trap type will be printed to stderr every nth trap, and at the end of the program. Default is INT_MAX. ABORT(n) Core dump and abort the program upon encountering the nth trap. Default is 1. EXIT(n) Exit program upon encountering the nth trap. Default is 1. TRACE(n) If a trap is encountered, print a stack trace to stderr up to n times. Default is 10. EXAMPLES OF TRAP_FPE The following example counts all traps, traces the first five overflows, aborts on the first divide by zero, or the 100th overflow. It replaces zero for underflows, the "appropriate" value for overflows, and the default values for divide by zero, invalid operands, and integer overflows. setenv TRAP_FPE "ALL=COUNT; UNDERFL=ZERO; OVERFL=IEEE,TRACE(5), ABORT(100); DIVZERO=ABORT" WHEN AN EXCEPTION IS ENCOUNTERED When an exception is encountered, the handler examines the instruction causing the exception, the state of the floating-point accelerator and the sigfpe structure to determine the correct action to take, and the program is continued. In the cases of _UNDERFL, _OVERFL, _DIVZERO, _INT_OVERFL, and some instances of _INVALID, an appropriate value is substituted for the result of the operation, and the instruction which caused the exception is skipped. For most exceptions arising due to an invalid operand ( exceptions), more meaningful behavior may be obtained by replacing an erroneous operand. For these conditions, the operand is replaced, and the instruction re-issued. sigfpe: For each enabled exception, the sigfpe structure contains the following fields: repls, count, trace, exit and abort. For each enabled exception <p>, and each non-zero entry <n> in the sigfpe structure, the trap handler will take the following actions: count: A count of all enabled traps will be printed to stderr at the end of execution of the program, and every at <n>th exception <p>. trace: A dbx stack trace will be printed to stderr every exception <p>, up to <n> times. You must have dbx installed on your system to use this option. abort: Core dump and abort program upon encountering the <n>th exception <p>. The abort option takes precedence over the exit option. exit: Exit program upon encountering the <n>th exception <p>. repls: Each of the exceptions _UNDERFL, _OVERFL, _DIVZERO, and _INT_OVERFL has an associated default value which is used as the result of the operation causing the exception. These default values may be overridden by initializing this integer value. This value is interpreted as an integer code used to select one of a set of replacement values, or to indicate that the user_routine routine is responsible for setting the value. These integer codes are listed below: _ZERO Use zero as the replacement value _FLUSH_ZERO Set the flush_zero bit in the Control Status register. This causes a flush to zero without invoking the trap handler. Works only for underflow traps on the R4000 and later processors. Works like _ZERO for the R3000. _MIN Use the appropriately-typed minimum value as the replacement (i.e., the smallest number which is representable in that format without denormalizing). _MAX Use the appropriately-typed maximum value as the replacement. _INF Use the appropriately-typed value for infinity as the replacement. _NAN Use the appropriately-typed value for not-a-number as the replacement. (A quiet not-a-number is used.) _APPROPRIATE Use IEEE standard results as the return result for _UNDERFL, _OVERFL, _DIVZERO, and _INVALID exceptions. _USER_DETERMINED Invoke the routine user_routine (see note) to set the value of the operation. If this is the code used for exceptions, all such exceptions will defer to user_routine to set their value. In this case, invalidop_results_ and invalidop_operands_ will be ignored. _NEG Use the negative of the argument as the replacement operand. This code is valid only for the cases _SQRT_NEG_X and _RSQRT_NEG_X (see below). The default values used as the results of floating-point exceptions are: ---------------------------------------------------------------------- # Element mnemonic Exception condition Default value ---------------------------------------------------------------------- 0 (none) (ignored) 1 _UNDERFL underflow _APPROPRIATE 2 _OVERFL overflow _APPROPRIATE 3 _DIVZERO divide-by-zero _APPROPRIATE 4 _INVALID invalid operand (use tables) 5 _INT_OVERFL integer overflow _MAX ---------------------------------------------------------------------- The default values for _UNDERFL, _OVERFL, _DIVZERO, and _INVALID exceptions will produce the same results as if the instruction were re-issued with the original operand(s) and floating-point traps disabled. Valid values for sigfpe_[_INT_OVERFL].repls are: _MAX, APPROPRIATE, _ZERO, and _USER_DETERMINED, the default being _MAX. For exceptions, the correct action may be either to set the result and skip the instruction, or to replace an operand and retry the instruction. There are four cases in which the result is set. The array named invalidop_results_ is consulted for replacement codes for these cases: ---------------------------------------------------------------------- # Element mnemonic Exception condition Default value ---------------------------------------------------------------------- 0 (none) (ignored) 1 _MAGNITUDE_INF_SUBTRACTION oo - oo _NAN 2 _ZERO_TIMES_INF 0 * oo _NAN 3 _ZERO_DIV_ZERO 0/0 _NAN 4 _INF_DIV_INF oo / oo _NAN ---------------------------------------------------------------------- There are ten cases in which an offending operand is replaced. An array named invalidop_operands_ is consulted for user-initialized codes for these cases. Cases 8 through 11 are valid only for the mips3 and later architectures. Array invalidop_operands_ has only 8 entries (0-7) for the earlier processors. Each element governs the following cases: ---------------------------------------------------------------------- # Element mnemonic Exception condition Default value ---------------------------------------------------------------------- 0 (none) (ignored) 1 _SQRT_NEG_X sqrt(-x) reissue 2 (unused) (ignored) with original 3 _CVTW_OVERFL conversion to integer caused target to overflow operands and 4 _CVTW_NA conversion of NaN to int floating point 5 _CVTW_INF conversion of oo to int traps disabled 6 _UNORDERED_CMP comparison to NaN 7 _SNAN_OP operand was Signaling NaN 8 _CVTL_OVERFL conversion to long long caused target to overflow 9 _CVTL_NAN conversion of NaN to long long 10 _CVTL_INF conversion of oo to long long 11 _RSQRT_NEG_X reciprocal sqrt(-x) ---------------------------------------------------------------------- WARNINGS handle_sigfpes is not reentrant. In particular, it must not be called from within an exception handler. NOTES "Use of user_routine to set values" If the integer code defining the replacement value for a particular exception condition is _USER_DETERMINED, the user- supplied routine user_routine is called: (*user_routine)(exception_parameters, value); value An array of two ints into which user_routine should store the replacement value. If an operand is being replaced, value has a copy of the current operand. exception_parameters An array of five unsigned ints which describe the following exception condition: -------------------------------------------------------------------- # Element mnemonic Description -------------------------------------------------------------------- 0 _EXCEPTION_TYPE The exception type (_DIVZERO, etc). 1 _INVALID_ACTION value = _SET_RESULT if result is being set. value = _REPL_OPERAND if an operand is being replaced. This element is meaningful only if the exception type is . 2 _INVALID_TYPE This element is meaningful only if the exception type is . It is the index corresponding to the particular conditions giving rise to the exception. In conjunction with element 1, this value uniquely determines the exception condition. (e.g., if _INVALID_ACTION is _SET_RESULT and _INVALID_TYPE is 2, the exception is due to _ZERO_TIMES_INF.) 3 _VALUE_TYPE The type of the replacement value - either _SINGLE, _DOUBLE, _WORD, or _LONGWORD 4 _VALUE_SIGN T{ The suggested sign user_routine should use for the replacement value - either _POSITIVE or _NEGATIVE. -------------------------------------------------------------------- NOTES FOR R8000 Due to the nature of parallel operations on the R8000 processor, it is not possible to determine the true value of the program counter when a floating point exception occurs. Therefore, on that processor, libfpe.so will not update either operands or results when floating point exceptions occur. Another anomaly on this processor is that counts of floating point exceptions for a particular program may vary from run to run. When the R8000 processor executes in precise exception mode, libfpe behaves as it does on other processors, i.e. operands and results can be updated when floating point exceptions occur. See fpmode(1). SEE ALSO fpc(3c), fsigfpe(3f), get_fpc_csr(3c), set_fpc_csr(3c), signal(3c)