================
@@ -1165,7 +1165,112 @@ void
Sema::checkFortifiedBuiltinMemoryFunction(FunctionDecl *FD,
unsigned BuiltinID = UseDecl->getBuiltinID(/*ConsiderWrappers=*/true);
- if (!BuiltinID)
+ // Some libc I/O functions are intentionally not Clang builtins:
+ // * "read", "write", "readlink", "readlinkat", and "getcwd" are common
+ // identifiers (or names that appear in asm-label wrappers); declaring
+ // them as builtins triggers -Wincompatible-library-redeclaration on
+ // harmless local declarations (an error under -Werror).
+ // * pread/pwrite prototypes use off_t, whose width is platform- and
+ // macro-dependent (notably _FILE_OFFSET_BITS); a fixed builtin signature
+ // would clash with the system header on some targets.
+ // Recognize them by name, but only after checking the full POSIX prototype
+ // so that an unrelated function happening to share the name is not
+ // diagnosed as if it were the libc function.
+ enum class LibCDispatch {
+ None,
+ Read,
+ Write,
+ PRead,
+ PWrite,
+ ReadLink,
+ ReadLinkAt,
+ GetCWD,
+ };
+ LibCDispatch LibC = LibCDispatch::None;
+ StringRef LibCName;
+ if (!BuiltinID && FD->isExternC() && FD->getIdentifier() &&
----------------
nuclearcat wrote:
Thanks, I agree, table + one equivalence loop would read much better. Sorry for
the back and forth on this, I just want to make sure we get the libc-divergence
edge cases right.
The issue is that three of the types we're matching aren't fixed QualTypes:
ssize_t is long on glibc but int on musl/Bionic/ILP32, so
hasSameUnqualifiedType against any single expected type rejects one libc or the
other. The current code matches structurally (signed int, width == size_t
width).
off_t width depends on _FILE_OFFSET_BITS;
pread vs pread64 can't be told apart by type equivalence at all, both are just
typedefs to a signed int in the AST, so the disambiguator has to be name +
width.
We could look those up as typedefs from the AST, but that's the same
triple-hardcoding pattern you (rightly) pushed back on in the poll PR.
What I think works: i can keep a table of expected signatures, but let each
parameter slot be either a QualType or a small predicate (SSizeTLike, OffTLike,
Off64TLike). Generic equivalence loop dispatches on the variant. Most slots
collapse to QualType; only the irregular ones use a predicate. That readability
win without breaking the ibc-divergence edge cases.
If you like that, I'll proceed with the implementation. If you have a better
idea, I'm all ears!
https://github.com/llvm/llvm-project/pull/196499
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
