@Yatekii, @ithinuel: I am more concerned with the correctness than the quality of the code (although I applied the change you suggested and ran cargo clippy):

The new version yields identical results for dump-devices with 1500+ packs.

$ time ./target/release/cmsis-cli dump-devices -b minidom-boards.json -d minidom-devices.json
...
real	0m9.319s
user	0m1.556s
sys	0m7.580s

$ time ./target/release/cmsis-cli dump-devices -b roxmltree-boards.json -d roxmltree-devices.json
...
real	0m9.333s
user	0m1.563s
sys	0m7.571s

$ cmp minidom-boards.json roxmltree-boards.json ; echo $?
0
$ jq -S "." minidom-devices.json > minidom-devices.jq.json
$ jq -S "." roxmltree-devices.json > roxmltree-devices.jq.json
$ cmp minidom-devices.jq.json roxmltree-devices.jq.json ; echo $?
0

Also there seems to be no performance change.

Could you suggest some target-gen tests?

I have tested with target-gen arm with this branch and except for one minor thing it generates identical YAML files.

There are 3 files where some of the memory entries are sometimes listed in a different order. As an example SAMD51.yaml generated from Microchip.SAMD51_DFP.3.8.246.pdsc:

--- minidom/SAMD51.yaml
+++ roxmltree/SAMD51.yaml
@@ -27,7 +27,7 @@
     cores:
     - main
   - !Generic
-    name: IRAM3
+    name: IRAM2
     range:
       start: 0x20000000
       end: 0x20008000
@@ -36,7 +36,7 @@
     access:
       execute: false
   - !Generic
-    name: IRAM2
+    name: IRAM3
     range:
       start: 0x20000000
       end: 0x20008000

Those entries have the same start/end addresses. The information for those entries is passed in HashMap (keyed by the name) which explains why they can have a different order.

Amazing stuff! I went through all the code yesterday and I did not see any deviations in the logic. Considering target-gen chucks away happily, I would merge this :) Thanks a lot for the work!

I squashed the commits and did some more clippy fixes.