[flang][acc] Add ACCUseDeviceCanonicalizer pass (#175228)
This pass canonicalizes the use_device clause on acc.host_data
constructs to enable simpler runtime lowering. For use_device operands
that are box types or references to boxes, the pass:
1. Extracts the host base address for mapping to a device address using
acc.use_device
2. Creates a new boxed descriptor with the device address as the base
address for use inside the host_data region
The pass also removes unused use_device clauses to reduce runtime calls.
This canonicalization hoists load/box_addr patterns out of the host_data
region so they are applied to the host variable before acc.use_device,
ensuring the device pointer is used directly inside the region.
Example transformation for a reference to a box (!fir.ref<!fir.box<>>):
Before:
```
%ptr = acc.use_device varPtr(%ref : !fir.ref<!fir.box<!fir.ptr<i32>>>)
acc.host_data dataOperands(%ptr) {
%box = fir.load %ptr
%addr = fir.box_addr %box
// use %addr
}
```
After:
```
%box = fir.load %ref
%addr = fir.box_addr %box
%dev_ptr = acc.use_device varPtr(%addr : !fir.ptr<i32>)
acc.host_data dataOperands(%dev_ptr) {
%new_box = fir.embox %dev_ptr
// use device pointer through new descriptor
}
```
---------
Co-authored-by: nvptm <pmathew@nvidia.com>
