[Bugfix] Fixed: Bitwise ops on floats causing wrong code generation and crashes.

[dpankratz]

Bitwise Ops on Floats

I encountered a few bugs involving using bitwise operators on floating point types. The first bug is that the << and >> operators are essentially no-ops when applied to floats which can be misleading. For example:

shape = (1,1)
c = tvm.compute(shape,lambda i,j: tvm.const(10) << 2.0)
s = tvm.create_schedule([c.op])
f = tvm.build(s,[c])
c_tvm= tvm.nd.array(numpy.zeros(shape,dtype='float32'))
f(c_tvm)
print(c_tvm) #Prints [[10.]] not expected [[40.]]

The other bitwise operators |, &, ^, and ~ throw an error in the LLVM backend when applied to floats. For reference I tested using LLVM-8.

This bug is fixed by adding checks to filter out floats for the |, &, ^, ~, <<, and >> operators.

TypeError Crashes

I also encountered a few more cases where operators failed to promote python types to tvm ExprOps which causes a python TypeError.

a = tvm.var()
10 % a #crashes
10 << a #crashes
10 >> a #crashes

This bug is fixed by adding the missing operators to expr.py

This pull request also includes a regression test for both in test_lang_basic.py

@tqchen @ZihengJiang

[tqchen]

Thanks @dpankratz !