Uneccessary deep copy operations on donated inputs

[ricardoV94]

Description

import pytensor
import pytensor.tensor as pt

x = pt.vector("x")
out = x[0]

fn = pytensor.function([pytensor.In(x, borrow=True)], out)
fn.dprint()
# DeepCopyOp [id A] 1
#  └─ Subtensor{i} [id B] 0
#     ├─ x [id C]
#     └─ 0 [id D]

It's the same with mutable=True which also means borrow=True. The DeepCopyOp shouldn't be introduced.

In a more complex example, things work as expected:

import pytensor
import pytensor.tensor as pt

x = pt.vector("x")
out = x[0] * 2

fn = pytensor.function([pytensor.In(x, mutable=True)], out)
fn.dprint(print_view_map=True, print_destroy_map=True)
# Mul [id A] d={0: [1]} 1
#  ├─ 2.0 [id B]
#  └─ Subtensor{i} [id C] v={0: [0]} 0
#     ├─ x [id D]
#     └─ 0 [id E]

We are mutating the first input because we were allowed.

[ricardoV94]

Apparently pytensor.Out also has a borrow kwarg. When that's set the deepcopy is avoided, although now there's a funky View Op, probably for rewrite sake?

import pytensor
import pytensor.tensor as pt

x = pt.vector("x")
out = x[0]

fn = pytensor.function([pytensor.In(x, borrow=True)], pytensor.Out(out, borrow=True))
fn.dprint(print_view_map=True, print_destroy_map=True)
# ViewOp [id A] v={0: [0]} 1
#  └─ Subtensor{i} [id B] v={0: [0]} 0
#     ├─ x [id C]
#     └─ 0 [id D]

[ricardoV94]

So I guess In(borrow=True) is useless without a corresponding Out(borrow=True)