[BUG] Integer overflow during interval analysis and simplification

[sgrechanik-h]

Consider the following code:

import tvm

dtype = 'int32'
cnst = lambda c: tvm.const(c, dtype)
vs = [tvm.var("x" + str(i), dtype) for i in range(4)]
[x0, x1,x2, x3] = vs
vranges = {v: tvm.Range(tvm.const(-256*256*2560, dtype), tvm.const(256*256*256*20, dtype))
               for v in vs}

expr = x0*cnst(2) + x1*cnst(2) + x2*cnst(2) + cnst(-1) < x3*cnst(-1)

res = tvm.ir_pass.CanonicalSimplify(expr, vranges)

print(vranges)
print(expr)
print(res)

(Note that all the specified ranges are within the limits for int32).
CanonicalSimplify simplifies the expression to 1 which is wrong. The reason is that during interval analysis (EvalSet) an integer overflow happens. Note that although internally values are represented as int64, which should be enough, the function IntImm::make from HalideIR intentionally performs truncation of higher bits.
I don't know what to do with this problem, the truncation seems important if we want to preserve the exact behavior of integer operations, but I think it should be disabled for interval analysis somehow.

[tqchen]

given that the real code overflows, it is hard to argue this as a bug for now. Ideally, we will want to move most of the integer analysis to int64 and do a bound analysis to narrow down necessary ones to int32. This will likely mean a refactor of the current integer analysis. I would suggest we put this in our list to be noticed, and come back to it when we refactor integer analysis. I hope that can happen in the 0.6 release cycle.

[tqchen]

Consolidate this issue to #2588