[runtime] Fix off-by-one error in minimum integers #916
Conversation
|
From a practical perspective (how many ones-complement machines are there?), this seems OK. It's not clear to me without more digging, though, if this is legal Fortran. The definition of model integers in the standard (F2018 subclause 16.4) defines model integers only through the equivalent of -HUGE(the integer kind under consideration). |
|
@gklimowicz I read that part of the language standard and I was wondering about the same thing. However, neither flang nor gfortran complains about the use of |
|
@bryanpkc Yeah. It's hard to argue against that. I will stop. :-) |
runtime/flang/const.c defines min_int{1,2,4,8} to be -max_int{1,2,4,8}, which
is wrong for the two's complement representation. As a result, the MAXLOC
runtime function will compare each element in the input array with an incorrect
minimum value. If the array contains -huge(integer)-1, MAXLOC will return an
unexpected result.
bryanpkc
force-pushed
the
fix-maxloc-int_min
branch
from
05bf6d8
to
b9bde8d
Compare
|
@bryanpkc please merge if no further changes required. You have approval from everyone i guess. |
|
Thanks @kiranchandramohan. |
runtime/flang/const.c defines min_int{1,2,4,8} to be -max_int{1,2,4,8}, which is wrong for the two's complement representation. As a result, the MAXLOC runtime function will compare each element in the input array with an incorrect minimum value. If the array contains -huge(integer)-1, MAXLOC will return an unexpected result.