opensourcemirror
/
postgresql
mirror of git://git.postgresql.org/git/postgresql.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Fix power() for large inputs yet more. 

Buildfarm results for commit e532b1d57 reveal the error in my thinking
about the unexpected-EDOM case.  I'd supposed this was no longer really
a live issue, but it seems the fix for glibc's bug #3866 is not all that
old, and we still have at least one buildfarm animal (lapwing) with the
bug.  Hence, resurrect essentially the previous logic (but, I hope, less
opaquely presented), and explain what it is we're really doing here.

Also, blindly try to fix fossa's failure by tweaking the logic that
figures out whether y is an odd integer when x is -inf.  This smells
a whole lot like a compiler bug, but I lack access to icc to try to
pin it down.  Maybe doing division instead of multiplication will
dodge the issue.

Discussion: https://postgr.es/m/E1jkU7H-00024V-NZ@gemulon.postgresql.org
This commit is contained in:
Tom Lane 2020-06-15 19:10:30 -04:00
parent 2961c9711c
commit 5674eb9876
1 changed files with 23 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
src/backend/utils/adt/float.c
View File

@ -1583,7 +1583,7 @@ dpow(PG_FUNCTION_ARGS)
if (arg2 == floor(arg2))
{
/* y is integral; it's odd if y/2 is not integral */
double halfy = arg2 * 0.5; /* should be computed exactly */
double halfy = arg2 / 2; /* should be computed exactly */
if (halfy != floor(halfy))
yisoddinteger = true;
@ -1608,17 +1608,29 @@ dpow(PG_FUNCTION_ARGS)
if (errno == EDOM || isnan(result))
{
/*
* We eliminated all the possible domain errors above, or should
* have; but if pow() has a more restrictive test for "is y an
* integer?" than we do, we could get here anyway. Historical
* evidence suggests that some platforms once implemented the test
* as "y == (long) y", which of course misbehaves beyond LONG_MAX.
* There's not a lot of choice except to accept the platform's
* conclusion that we have a domain error.
* We handled all possible domain errors above, so this should be
* impossible. However, old glibc versions on x86 have a bug that
* causes them to fail this way for abs(y) greater than 2^63:
*
* https://sourceware.org/bugzilla/show_bug.cgi?id=3866
*
* Hence, if we get here, assume y is finite but large (large
* enough to be certainly even). The result should be 0 if x == 0,
* 1.0 if abs(x) == 1.0, otherwise an overflow or underflow error.
*/
ereport(ERROR,
(errcode(ERRCODE_INVALID_ARGUMENT_FOR_POWER_FUNCTION),
errmsg("a negative number raised to a non-integer power yields a complex result")));
if (arg1 == 0.0)
result = 0.0; /* we already verified y is positive */
else
{
double absx = fabs(arg1);
if (absx == 1.0)
result = 1.0;
else if (arg2 >= 0.0 ? (absx > 1.0) : (absx < 1.0))
float_overflow_error();
else
float_underflow_error();
}
}
else if (errno == ERANGE)
{