ahatanak wrote:
Sorry for reporting this over a year after this PR was merged, but we recently
found a regression it introduced in `-Wunreachable-code`.
The warning fires when there's code following calls to functions annotated with
`analyzer_noreturn`. The warning is a false positive, as those functions,
unlike functions annotated with `noreturn`, do return, so code after the calls
isn't dead code.
For example:
```
extern void Assert(const char *) __attribute__((analyzer_noreturn));
void f(int x) {
if (x) return;
Assert("bad");
use(3); // warning: code will never be executed
}
The change is correct for almost every other analysis and warning. For example,
it fixes false positives in `-Wreturn-type` and `-Wimplicit-fallthrough`.
Gábor and I discussed two ways to fix it:
1. Have `-Wunreachable-code` consume a different CFG that does not respect the
attribute. This may cost compile time, since the CFG could be built more than
once per function.
2. Instead of omitting the edges, add a special "analyzer noreturn" edge or
marker that each analysis can choose to consume or ignore. This is the
preferred option.
Does that direction sound right?
https://github.com/llvm/llvm-project/pull/150952
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
