Unsigned ints no wrap

[Shaikh-Ubaid]

towards #2173

[Shaikh-Ubaid]

This is a Work in progress (WIP). The tests might not pass.

[Shaikh-Ubaid]

$ cat examples/expr2.py    
from lpython import u16, bitnot_u16, i32

u: u16 = u16(255)
u = bitnot_u16(u)
print(u)
print(~i32(u))
$ lpython examples/expr2.py
65280
255
$ python examples/expr2.py 
65280
-65281

It seems we are unable to use bitnot() and ~ successively.

[certik]

The bitnot seems correct but ~ doesn't work? We have to debug it.

[Shaikh-Ubaid]

That issue is fixed. The following is a concern:

$ cat examples/expr2.py 
from lpython import i16, u64, bitnot_u64

w: u64 = u64(123457889998923)
x: u64 = bitnot_u64(w)
print(w, i16(w), ~i16(w))
print(x, i16(x), ~i16(x))
$ python examples/expr2.py 
123457889998923 123457889998923 -123457889998924
18446620615819552692 18446620615819552692 -18446620615819552693
$ lpython examples/expr2.py
123457889998923 -26549 26548
18446620615819552692 26548 -26549

It seems u8, u16, u32 and u64 need to be some functions that could convert values appropriately in CPython.

[certik]

the i16 and u16 in LPython must give a runtime (or compile time) error when the value you are converting is out of range. The contract is that if LPython works, then CPython must work. So in CPython we don't need to do any checking and all these casts become a no-op.

[Shaikh-Ubaid]

If the casts in CPython become no-op, the output value of CPython is not matching the output of LPython (Example: #2176 (comment)).

[Shaikh-Ubaid]

For unsigned unary minus, shall we throw error in LPython similar to unsigned bit not?

[certik]

The contract is that if LPython compiles and runs this without errors, then CPython produces the same value. The design of all i8-i64, u8-u64 is such that they are just no-op in CPython and they always produce exactly the same number as in LPython, since LPython only allows such casts that don't change the value. If the value is out of range, we give an error message. So this design should work.

[Shaikh-Ubaid]

only allows such casts that don't change the value. If the value is out of range, we give an error message.

Got it. Thank you so much for the clarification!

[certik]

For unsigned unary minus, shall we throw error in LPython similar to unsigned bit not?

Given that -0 = 0, the unary minus should be allowed for unsigned, but obviously it will only work for 0, for all other values it must give a runtime error.

[Shaikh-Ubaid]

The CI might temporarily pass now. As a next step, we need to support

• unsigned unary minus (based on Unsigned ints no wrap #2176 (comment))
• uncomment the specific tests for unary minus (and possibly update if the values are not in appropriate range)
• uncomment the specific tests for bitshift and update the tests so that it does not overflow.
• add error tests for ~ and -.

[Shaikh-Ubaid]

Further steps could be to support compile time overflow checking and later runtime error checking for which we have a dedicated issue here #2183.

[certik]

Otherwise it looks good. After you fix the above, please do not merge it, I have to manually check and update our internal codes if needed before we can merge it.

[certik]

I think this change is good. I'll mark it as draft, so that we don't accidentally merge it. I need to see if anything broke first.

[certik]

I tested the compiled mode (LPython) and it gives a nice error message (it could be improved to give back exactly the expression to use, but it's good enough for now) and after using bitnot_u64 it works.

I now need to test the emulation mode (lpython.py), which is the more fragile mode that might have some bugs.

[certik]

Everything works for us, so I am merging this PR.

Thanks @Shaikh-Ubaid !

[certik]

I made a new release with this change:

https://github.com/lcompilers/lpython/releases/tag/v0.18.4
conda-forge/lpython-feedstock#14