[pypy-dev] Integer arithmetic

Armin Rigo arigo at tunes.org
Fri Apr 8 17:27:30 CEST 2005


Hi all,

(IRC summary with Samuele and myself)

We propose the following rules for how integer operations should behave in
R-Python code:

* operations on plain integers produce longs in CPython when they overflow,
and wrap-around silently on platforms like C.  In other words, they behave
differently when run on top of Python or after translation.  We cannot easily
change that, but instead the following rule seems to be the sanest idea: in
both cases, the only guarantee is that the last 32 (or 64) bits of the result
are correct.  If needed we must call a function, intmask(x), to ensure that we
mask off the high bits on Python.  intmask() is a no-op after translation to
C.

* unsigned arithmetic: the r_uint class is fine for that.  It provides
wrap-around unsigned arithmetic on top of Python, and for the translator it is
a marker that unsigned ints should be generated.  Warning, mixed operations
r_uint+int produce r_uint results; be careful about unlimited propagation.  
Conversion: r_uint(x) for int-to-uint; intmask(x) for uint-to-int.

* a new space.uint_w() turns a non-negative wrapped int or long into an
r_uint.

* overflow detection: only for signed arithmetic, and only explicitely using a
macro-ish function ovfchecked().  It's only allowed to call it as
ovfchecked(<variable> <operation> <variable>); no subexpressions.  It means
that this single operation is the one that is done with overflow checking.  
When running on top of Python, ovfchecked(x) just returns x unless it is too
large, in which case you get an OverflowError.  After translation, the two
basic block operations "z = op(x,y)" and "simple_call(ovfchecked, z)" can be
detected easily and turned into whatever the back-end does for overflow
checking.  (Additionally, hpk proposed a hack that checks the syntax of the
ovfchecked() call when running on top of Python, to avoid problems with
expressions like ovfchecked(x+y+z) where after translation only the last
operation is checked.)

* Christian's very nice r_int is on the way out...  its main usage, in
intobject.py, will be replaced by ovfchecked().  It's just more explicit and
not really longer.  It's also more in line with Python's evoluation which says
that arithmetic operations don't raise OverflowErrors any more.


Comments welcome!

Armin



More information about the Pypy-dev mailing list