coverage: Discard code regions that might cause fatal errors in llvm-cov

Author: Zalathar

compiler/rustc_mir_transform/src/coverage/mod.rs

@@ -286,7 +286,7 @@ fn make_code_region(
         start_line = source_map.doctest_offset_line(&file.name, start_line);
         end_line = source_map.doctest_offset_line(&file.name, end_line);
     }
-    Some(CodeRegion {
+    check_code_region(CodeRegion {
         file_name,
         start_line: start_line as u32,
         start_col: start_col as u32,
@@ -295,6 +295,37 @@ fn make_code_region(
     })
 }
 
+/// If `llvm-cov` sees a code region that is improperly ordered (end < start),
+/// it will immediately exit with a fatal error. To prevent that from happening,
+/// discard regions that are improperly ordered, or might be interpreted in a
+/// way that makes them improperly ordered.
+fn check_code_region(code_region: CodeRegion) -> Option<CodeRegion> {
+    let CodeRegion { file_name: _, start_line, start_col, end_line, end_col } = code_region;
+
+    // Line/column coordinates are supposed to be 1-based. If we ever emit
+    // coordinates of 0, `llvm-cov` might misinterpret them.
+    let all_nonzero = [start_line, start_col, end_line, end_col].into_iter().all(|x| x != 0);
+    // Coverage mappings use the high bit of `end_col` to indicate that a
+    // region is actually a "gap" region, so make sure it's unset.
+    let end_col_has_high_bit_unset = (end_col & (1 << 31)) == 0;
+    // If a region is improperly ordered (end < start), `llvm-cov` will exit
+    // with a fatal error, which is inconvenient for users and hard to debug.
+    let is_ordered = (start_line, start_col) <= (end_line, end_col);
+
+    if all_nonzero && end_col_has_high_bit_unset && is_ordered {
+        Some(code_region)
+    } else {
+        debug!(
+            ?code_region,
+            ?all_nonzero,
+            ?end_col_has_high_bit_unset,
+            ?is_ordered,
+            "Skipping code region that would be misinterpreted or rejected by LLVM"
+        );
+        None
+    }
+}
+
 fn is_eligible_for_coverage(tcx: TyCtxt<'_>, def_id: LocalDefId) -> bool {
     // Only instrument functions, methods, and closures (not constants since they are evaluated
     // at compile time by Miri).