fix: don't mark a method's throws as visited before its blocks are loaded - #3028

Description

MethodThrowsVisitor.visit sets the visited flag and then reads the method's blocks . when it is reached through a call from another class that is processed first , the callee's class isn't loaded yet , its blocks are null , nothing is collected , and the flag stays set . when the callee's own class is processed later the visit is skipped , so the method never gets its throws declaration even though it throws a checked exception . the output then doesn't compile .

this happens with a dependency cycle : A depends on B , B is processed first as a dependency of A , and its call to A.check() visits A.check before A is loaded . the flag is now only set once the blocks exist , so A.check is analyzed when A is processed .

found by decompiling the apk from #2892 . single threaded the output was stable but android.ext.qt.a(String, int, String) came out without throws InterruptedException , although it calls FutureTask.get(long, TimeUnit) and only catches ExecutionException and TimeoutException . luaj.lib.OsLib.a(String) throws new IOException(...) and had no throws either . with this change 10 methods in 3 classes get their throws and nothing else in the 813 classes changes . with -j 8 the throws clauses used to differ between two runs of the same build , with the change two runs are identical .

the new smali test is the cycle above and fails without the change . jadx-core tests : 1034 passed .

one thing this does not fix : a caller that is analyzed before its callee still doesn't see the callee's throws , B.use and A.run in the test stay without it . that needs the analysis to wait for the callee , which is a bigger change , happy to look at it if you want it .