Shifting operators should not adopt the same implicit conversion rule as other arithmetic operators

[strongoier]

Shifting operators like << should take the left-side type as its result type (i8 << u32 = i8) instead of following rules for other arithmetic operators (i8 + u32 = u32).

[yuanming-hu]

A not so minimal example to reproduce this issue (by @GCChen97, #4613):

@ti.func
def set_bit(switches, i, dt):
    switches  |= (ti.cast(1, dt) << i)
    return switches

@ti.kernel
def test():
    for _ in range(1):
        for i in range(8):
            eight_flags = ti.cast(0, ti.u8)
            eight_flags = set_bit(eight_flags, i, ti.u8)
            print(ti.cast(eight_flags, ti.i32))
test()

You will get a warning because i8 << i32 leads to i32, which is later stored into u8 (lose of precision).

[yuanming-hu]

The place to look at:

taichi/taichi/transforms/type_check.cpp

Line 294 in 6df23cb

void visit(BinaryOpStmt *stmt) override {

For shift operators we should use the left operand type as result type.