cast<t>(immediate) should be make_const(t, immediate)

[abadams]

Doing a Halide cast of a c++ int constructs an immediate Expr (e.g. IntImm) and then eagerly folds it to a different type of immediate in the Cast constructor. It's better to just construct the immediate using the desired type to begin with.

[mcourteaux]

This would be a good linting rule somehow, if we could define our own.

[abadams]

I found all the cases by adding things like this in IROperator.h:

template<typename T>
Expr cast(int64_t) {
  static_assert(false, "no!");
  return Expr{};
}

but I don't want to leave that because I don't want to break existing code outside the repo over a very very minor inefficiency.

[alexreinking]

Failure is:

Seed: 6251522420976499283
can't prove upper bound: (1 <= 0)
a = 115
b = 10
c = 123
d = 85
e = 85
int32x2(x2(uint1(int32(uint1(c) && (uint1)1))) && uint1x2(ramp(b, 22, 2) % int32x2(ramp((uint32)75, (uint32)4294967192, 2))))
[0, 0]
In vector lane 1:
int32(uint1(int32(uint1(c) && (uint1)1)) && uint1((b + 22) % int32((uint32)4294967267))) -> 1
scope {
	a : [109, 115]
	b : [7, 117]
	c : [120, 123]
	d : [77, 109]
	e : [33, 119]
}

[mcourteaux]

I found all the cases by adding things like this in IROperator.h:

template<typename T>
Expr cast(int64_t) {
  static_assert(false, "no!");
  return Expr{};
}

but I don't want to leave that because I don't want to break existing code outside the repo over a very very minor inefficiency.

Super clever! I wonder if there is a preprocessor define that we can use to check if we're building on CI, or if we're building just libHalide.so, @alexreinking?