Some patches to Lee Busby's fpectl mods that accidentally didn't make it

into 1.5a4.
This commit is contained in:
Guido van Rossum 1997-10-20 23:50:01 +00:00
parent 9476a78c1e
commit e85da651dd
3 changed files with 40 additions and 30 deletions

View file

@ -66,9 +66,10 @@ extern "C" {
* 1) Add the *_PROTECT macros to your C code as required to protect
* dangerous floating point sections.
*
* 2) Turn on the inclusion of the code by #defining WANT_SIGFPE_HANDLER in
* config.h.in before you configure, compile, and install Python, and the
* fpectl module, and any other modules which may have conditional code.
* 2) Turn on the inclusion of the code by adding the ``--with-fpectl''
* flag at the time you run configure. If the fpectl or other modules
* which use the *_PROTECT macros are to be dynamically loaded, be
* sure they are compiled with WANT_SIGFPE_HANDLER defined.
*
* 3) When python is built and running, import fpectl, and execute
* fpectl.turnon_sigfpe(). This sets up the signal handler and enables
@ -76,11 +77,11 @@ extern "C" {
* on, any properly trapped SIGFPE should result in the Python
* FloatingPointError exception.
*
* Step 1 has been done already for the Python kernel code, and will be
* done soon for Hugunin's NumPy array package and my Gist graphics module.
* Step 2 is usually done once at python install time. Python's behavior
* with respect to SIGFPE is not changed unless you also do step 3. Thus
* you can control this new facility at compile time, or run time, or both.
* Step 1 has been done already for the Python kernel code, and should be
* done soon for the NumPy array package. Step 2 is usually done once at
* python install time. Python's behavior with respect to SIGFPE is not
* changed unless you also do step 3. Thus you can control this new
* facility at compile time, or run time, or both.
*
********************************
* Using the macros in your code:
@ -89,17 +90,16 @@ extern "C" {
* {
* ....
* PyFPE_START_PROTECT("Error in foobar", return 0)
* dangerous_op(somearg1, somearg2, ...);
* PyFPE_END_PROTECT
* result = dangerous_op(somearg1, somearg2, ...);
* PyFPE_END_PROTECT(result)
* ....
* }
*
* If a floating point error occurs in dangerous_op, foobar returns 0
* (NULL), after setting the associated value of the FloatingPointError
* exception to "Error in foobar". ``Dangerous_op'' can be a single
* operation, or a block, or function calls, or any combination, so long as
* no alternate return is possible before the PyFPE_END_PROTECT macro is
* reached.
* If a floating point error occurs in dangerous_op, foobar returns 0 (NULL),
* after setting the associated value of the FloatingPointError exception to
* "Error in foobar". ``Dangerous_op'' can be a single operation, or a block
* of code, function calls, or any combination, so long as no alternate
* return is possible before the PyFPE_END_PROTECT macro is reached.
*
* The macros can only be used in a function context where an error return
* can be recognized as signaling a Python exception. (Generally, most
@ -121,7 +121,7 @@ extern "C" {
* I therefore decided on a more limited form of nesting, using a counter
* variable (PyFPE_counter) to keep track of any recursion. If an exception
* occurs in an ``inner'' pair of macros, the return will apparently
* come from the top level.
* come from the outermost level.
*
*/
@ -131,23 +131,35 @@ extern "C" {
#include <math.h>
extern jmp_buf PyFPE_jbuf;
extern int PyFPE_counter;
extern double PyFPE_dummy();
extern double PyFPE_dummy(void *);
#define PyFPE_START_PROTECT(err_string, leave_stmt) \
if (!PyFPE_counter++ && setjmp(PyFPE_jbuf)) { \
PyFPE_counter = 0; \
PyErr_SetString(PyExc_FloatingPointError, err_string); \
PyFPE_counter = 0; \
leave_stmt; \
}
/*
* This (following) is a heck of a way to decrement a counter. However,
* code optimizers will sometimes move this statement so that it gets
* executed *before* the unsafe expression which we're trying to protect.
* This pretty well messes things up, of course. So the best I've been able
* to do is to put a (hopefully fast) function call into the expression
* which counts down PyFPE_counter, and thereby monkey wrench the overeager
* optimizer. Better solutions are welcomed....
* unless the macro argument is provided, code optimizers will sometimes move
* this statement so that it gets executed *before* the unsafe expression
* which we're trying to protect. That pretty well messes things up,
* of course.
*
* If the expression(s) you're trying to protect don't happen to return a
* value, you will need to manufacture a dummy result just to preserve the
* correct ordering of statements. Note that the macro passes the address
* of its argument (so you need to give it something which is addressable).
* If your expression returns multiple results, pass the last such result
* to PyFPE_END_PROTECT.
*
* Note that PyFPE_dummy returns a double, which is cast to int.
* This seeming insanity is to tickle the Floating Point Unit (FPU).
* If an exception has occurred in a preceding floating point operation,
* some architectures (notably Intel 80x86) will not deliver the interrupt
* until the *next* floating point operation. This is painful if you've
* already decremented PyFPE_counter.
*/
#define PyFPE_END_PROTECT(v) PyFPE_counter -= (int)PyFPE_dummy(&(v));

View file

@ -180,9 +180,8 @@ static void fpe_reset(Sigfunc *handler)
/*-- Linux ----------------------------------------------------------------*/
#elif defined(linux)
/* Linux delivers SIGFPE by default,
except for log(0), atanh(-1), 0.^0.
*/
#include <i386/fpu_control.h>
__setfpucw(0x1372);
signal(SIGFPE, handler);
/*-- NeXT -----------------------------------------------------------------*/

View file

@ -1,6 +1,5 @@
#include "config.h"
#include "pyfpe.h"
/*
* The signal handler for SIGFPE is actually declared in an external
* module fpectl, or as preferred by the user. These variable
@ -12,5 +11,5 @@
#ifdef WANT_SIGFPE_HANDLER
jmp_buf PyFPE_jbuf;
int PyFPE_counter = 0;
double PyFPE_dummy(){return(1.0);}
double PyFPE_dummy(void *dummy){ return 1.0; }
#endif