Recognize setjmp and friends as builtins even if jmp_buf is not declared yet.

[rsmith]

Summary

This happens in glibc's headers. It's important that we recognize these
functions so that we can mark them as returns_twice.

[rjmccall]

LGTM. I like this better as a fix for PR40692 anyway.

[tambre]

LGTM

[mstorsjo]

This broke use of setjmp for mingw on x86_64. IIRC, in MSVC environments, the _setjmp function is considered built-in, which is given an implicit second argument by the compiler. In mingw targets on the other hand, the compiler normally doesn't know of any such extra mechanisms, and it's hooked up e.g. like this:

#define setjmp(BUF) _setjmp((BUF), __builtin_frame_address (0))
int __cdecl __attribute__ ((__nothrow__,__returns_twice__)) _setjmp(jmp_buf _Buf, void *_Ctx);

After this commit, in C++, this errors out with "error: too many arguments to function call, expected 1, have 2", while in C it merely produces a warning "warning: incompatible redeclaration of library function '_setjmp'" (and that warning is produced by the SDK headers, where those warnings normally are silenced).

A full repro case here follows:

typedef __attribute__ ((__aligned__ (16))) struct _SETJMP_FLOAT128 {
  __extension__ unsigned long long Part[2];
} SETJMP_FLOAT128;
#define _JBLEN 16
typedef SETJMP_FLOAT128 _JBTYPE;
typedef _JBTYPE jmp_buf[_JBLEN];

#ifdef __cplusplus
extern "C" {
#endif

#define setjmp(BUF) _setjmp((BUF), __builtin_frame_address (0))
int __cdecl __attribute__ ((__nothrow__,__returns_twice__)) _setjmp(jmp_buf _Buf, void *_Ctx);

#ifdef __cplusplus
}
#endif

void other(void);
jmp_buf buf;
int func(void) {
    if (setjmp(buf)) {
        return 1;
    }
    other();
    return 0;
}

$ clang++ -x c++ -target x86_64-w64-mingw32 -c setjmp.c
setjmp.c:29:9: error: too many arguments to function call, expected 1, have 2
    if (setjmp(buf)) {
        ^~~~~~~~~~~
setjmp.c:12:36: note: expanded from macro 'setjmp'
#define setjmp(BUF) _setjmp((BUF), __builtin_frame_address (0))
                                   ^~~~~~~~~~~~~~~~~~~~~~~~~~~

(The exact functions that is used for setjmp in mingw configuration is a bit of a jungle, with different CRT setups on different architectures: https://github.com/mingw-w64/mingw-w64/blob/master/mingw-w64-headers/crt/setjmp.h#L213-L243)

[rsmith]

In D88518#2308085, @mstorsjo wrote:

This broke use of setjmp for mingw on x86_64.

Thanks, should be fixed in rG8fb2a235b0f22dedba72b8b559ba33171a8dcd09. Now that we properly recognize _setjmp even if it has a weird signature, the MSVC-specific check for exactly one argument was breaking the MinGW case. I've removed that check and instead we only apply the custom MSVC-specific codegen rule if the declaration happens to be of the right form (but we apply the returns_twice attribute regardless).

[mstorsjo]

In D88518#2309538, @rsmith wrote:

In D88518#2308085, @mstorsjo wrote:

This broke use of setjmp for mingw on x86_64.

Thanks, should be fixed in rG8fb2a235b0f22dedba72b8b559ba33171a8dcd09.

Thanks for the quick fix!

clang/include/clang/Basic/Builtins.h

bool isPredefinedRuntimeFunction(unsigned ID) const {
     return strchr(getRecord(ID).Attributes, 'i') != nullptr;
   }
 
   /// Determines whether this builtin has custom typechecking.
   bool hasCustomTypechecking(unsigned ID) const {
     return strchr(getRecord(ID).Attributes, 't') != nullptr;
   }
 
+  /// Determines whether a declaration of this builtin should be recognized
+  /// even if the type doesn't match the specified signature.
+  bool allowTypeMismatch(unsigned ID) const {
+    return strchr(getRecord(ID).Attributes, 'T') != nullptr ||
+           hasCustomTypechecking(ID);
+  }
+
   /// Determines whether this builtin has a result or any arguments which
   /// are pointer types.
   bool hasPtrArgsOrResult(unsigned ID) const {
     return strchr(getRecord(ID).Type, '*') != nullptr;
   }
 
   /// Return true if this builtin has a result or any arguments which are
   /// reference types.

clang/include/clang/Basic/Builtins.def

 // The third value provided to the macro specifies information about attributes
 // of the function.  These must be kept in sync with the predicates in the
 // Builtin::Context class.  Currently we have:
 //  n -> nothrow
 //  r -> noreturn
 //  U -> pure
 //  c -> const
 //  t -> signature is meaningless, use custom typechecking
+//  T -> type is not important to semantic analysis and codegen; recognize as
+//       builtin even if type doesn't match signature, and don't warn if we
+//       can't be sure the type is right
 //  F -> this is a libc/libm function with a '__builtin_' prefix added.
 //  f -> this is a libc/libm function without the '__builtin_' prefix. It can
 //       be followed by ':headername:' to state which header this function
 //       comes from.
 //  h -> this function requires a specific header or an explicit declaration.
 //  i -> this is a runtime library implemented function without the
 //       '__builtin_' prefix. It will be implemented in compiler-rt or libgcc.
 //  p:N: -> this is a printf-like function whose Nth argument is the format

 LANGBUILTIN(_rotr16, "UsUsUc",    "n", ALL_MS_LANGUAGES)
 LANGBUILTIN(_rotr,   "UiUii",     "n", ALL_MS_LANGUAGES)
 LANGBUILTIN(_lrotr,  "ULiULii",   "n", ALL_MS_LANGUAGES)
 LANGBUILTIN(_rotr64, "UWiUWii",   "n", ALL_MS_LANGUAGES)
 LANGBUILTIN(__va_start,       "vc**.", "nt", ALL_MS_LANGUAGES)
 LANGBUILTIN(__fastfail, "vUi",    "nr", ALL_MS_LANGUAGES)
 
 // Microsoft library builtins.
-LIBBUILTIN(_setjmpex, "iJ", "fj",   "setjmpex.h", ALL_MS_LANGUAGES)
+LIBBUILTIN(_setjmpex, "iJ", "fjT", "setjmpex.h", ALL_MS_LANGUAGES)
 
 // C99 library functions
 // C99 stdarg.h
 LIBBUILTIN(va_start, "vA.",       "fn",    "stdarg.h", ALL_LANGUAGES)
 LIBBUILTIN(va_end, "vA",          "fn",    "stdarg.h", ALL_LANGUAGES)
 LIBBUILTIN(va_copy, "vAA",        "fn",    "stdarg.h", ALL_LANGUAGES)
 // C99 stdlib.h
 LIBBUILTIN(abort, "v",            "fr",    "stdlib.h", ALL_LANGUAGES)

 LIBBUILTIN(wmemcmp, "iwC*wC*z", "f", "wchar.h", ALL_LANGUAGES)
 LIBBUILTIN(wmemcpy, "w*w*wC*z", "f", "wchar.h", ALL_LANGUAGES)
 LIBBUILTIN(wmemmove,"w*w*wC*z", "f", "wchar.h", ALL_LANGUAGES)
 
 // C99
 // In some systems setjmp is a macro that expands to _setjmp. We undefine
 // it here to avoid having two identical LIBBUILTIN entries.
 #undef setjmp
-LIBBUILTIN(setjmp, "iJ",          "fj",    "setjmp.h", ALL_LANGUAGES)
-LIBBUILTIN(longjmp, "vJi",        "fr",    "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(setjmp, "iJ",          "fjT",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(longjmp, "vJi",        "frT",   "setjmp.h", ALL_LANGUAGES)
 
 // Non-C library functions, active in GNU mode only.
 // Functions with (returns_twice) attribute (marked as "j") are still active in
 // all languages, because losing this attribute would result in miscompilation
 // when these functions are used in non-GNU mode. PR16138.
 LIBBUILTIN(alloca, "v*z",         "f",     "stdlib.h", ALL_GNU_LANGUAGES)
 // POSIX string.h
 LIBBUILTIN(memccpy, "v*v*vC*iz",  "f",     "string.h", ALL_GNU_LANGUAGES)

 // In some systems str[n]casejmp is a macro that expands to _str[n]icmp.
 // We undefine then here to avoid wrong name.
 #undef strcasecmp
 #undef strncasecmp
 LIBBUILTIN(strcasecmp, "icC*cC*", "f",     "strings.h", ALL_GNU_LANGUAGES)
 LIBBUILTIN(strncasecmp, "icC*cC*z", "f",   "strings.h", ALL_GNU_LANGUAGES)
 // POSIX unistd.h
 LIBBUILTIN(_exit, "vi",           "fr",    "unistd.h", ALL_GNU_LANGUAGES)
-LIBBUILTIN(vfork, "p",            "fj",    "unistd.h", ALL_LANGUAGES)
+LIBBUILTIN(vfork, "p",            "fjT",   "unistd.h", ALL_LANGUAGES)
 // POSIX pthread.h
 // FIXME: Should specify argument types.
 LIBBUILTIN(pthread_create, "",  "fC<2,3>", "pthread.h", ALL_GNU_LANGUAGES)
 
 // POSIX setjmp.h
 
-LIBBUILTIN(_setjmp, "iJ",         "fj",   "setjmp.h", ALL_LANGUAGES)
-LIBBUILTIN(__sigsetjmp, "iSJi",   "fj",   "setjmp.h", ALL_LANGUAGES)
-LIBBUILTIN(sigsetjmp, "iSJi",     "fj",   "setjmp.h", ALL_LANGUAGES)
-LIBBUILTIN(savectx, "iJ",         "fj",   "setjmp.h", ALL_LANGUAGES)
-LIBBUILTIN(getcontext, "iK*",     "fj",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(_setjmp, "iJ",         "fjT",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(__sigsetjmp, "iSJi",   "fjT",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(sigsetjmp, "iSJi",     "fjT",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(savectx, "iJ",         "fjT",   "setjmp.h", ALL_LANGUAGES)
+LIBBUILTIN(getcontext, "iK*",     "fjT",   "setjmp.h", ALL_LANGUAGES)
 
-LIBBUILTIN(_longjmp, "vJi",       "fr",    "setjmp.h", ALL_GNU_LANGUAGES)
-LIBBUILTIN(siglongjmp, "vSJi",    "fr",    "setjmp.h", ALL_GNU_LANGUAGES)
+LIBBUILTIN(_longjmp, "vJi",       "frT",   "setjmp.h", ALL_GNU_LANGUAGES)
+LIBBUILTIN(siglongjmp, "vSJi",    "frT",   "setjmp.h", ALL_GNU_LANGUAGES)
 // non-standard but very common
 LIBBUILTIN(strlcpy, "zc*cC*z",    "f",     "string.h", ALL_GNU_LANGUAGES)
 LIBBUILTIN(strlcat, "zc*cC*z",    "f",     "string.h", ALL_GNU_LANGUAGES)
 //   id objc_msgSend(id, SEL, ...)
 LIBBUILTIN(objc_msgSend, "GGH.",   "f",     "objc/message.h", OBJC_LANG)
 // long double objc_msgSend_fpret(id self, SEL op, ...)
 LIBBUILTIN(objc_msgSend_fpret, "LdGH.", "f", "objc/message.h", OBJC_LANG)
 // _Complex long double objc_msgSend_fp2ret(id self, SEL op, ...)

clang/lib/Sema/SemaDecl.cpp

NamedDecl *Sema::LazilyCreateBuiltin(IdentifierInfo *II, unsigned ID,
   ASTContext::GetBuiltinTypeError Error;
   QualType R = Context.GetBuiltinType(ID, Error);
   if (Error) {
     if (!ForRedeclaration)
       return nullptr;
 
     // If we have a builtin without an associated type we should not emit a
     // warning when we were not able to find a type for it.
-    if (Error == ASTContext::GE_Missing_type)
+    if (Error == ASTContext::GE_Missing_type ||
+        Context.BuiltinInfo.allowTypeMismatch(ID))
       return nullptr;
 
     // If we could not find a type for setjmp it is because the jmp_buf type was
     // not defined prior to the setjmp declaration.
     if (Error == ASTContext::GE_Missing_setjmp) {
       Diag(Loc, diag::warn_implicit_decl_no_jmp_buf)
           << Context.BuiltinInfo.getName(ID);
       return nullptr;
     }
 
     // Generally, we emit a warning that the declaration requires the
     // appropriate header.
     Diag(Loc, diag::warn_implicit_decl_requires_sysheader)
         << getHeaderName(Context.BuiltinInfo, ID, Error)
         << Context.BuiltinInfo.getName(ID);
     return nullptr;
   }
 
   if (!ForRedeclaration &&
       (Context.BuiltinInfo.isPredefinedLibFunction(ID) ||
        Context.BuiltinInfo.isHeaderDependentFunction(ID))) {
     Diag(Loc, diag::ext_implicit_lib_function_decl)
         << Context.BuiltinInfo.getName(ID) << R;
-    if (Context.BuiltinInfo.getHeaderName(ID) &&
-        !Diags.isIgnored(diag::ext_implicit_lib_function_decl, Loc))
+    if (const char *Header = Context.BuiltinInfo.getHeaderName(ID))
       Diag(Loc, diag::note_include_header_or_declare)
-          << Context.BuiltinInfo.getHeaderName(ID)
-          << Context.BuiltinInfo.getName(ID);
+          << Header << Context.BuiltinInfo.getName(ID);
   }
 
   if (R.isNull())
     return nullptr;
 
   FunctionDecl *New = CreateBuiltin(II, R, ID, Loc);
   RegisterLocallyScopedExternCDecl(New, S);
 
if (NewFD->isInvalidDecl()) {
       // C++ [temp.expl.spec]p2. We also allow these declarations as an
       // extension for compatibility with old SWIG code which likes to
       // generate them.
       Diag(NewFD->getLocation(), diag::ext_out_of_line_declaration)
         << D.getCXXScopeSpec().getRange();
     }
   }
 
-  // In C builtins get merged with implicitly lazily created declarations.
-  // In C++ we need to check if it's a builtin and add the BuiltinAttr here.
-  if (getLangOpts().CPlusPlus &&
+  // If this is the first declaration of a library builtin function, add
+  // attributes as appropriate.
+  if (!D.isRedeclaration() &&
       NewFD->getDeclContext()->getRedeclContext()->isFileContext()) {
     if (IdentifierInfo *II = Previous.getLookupName().getAsIdentifierInfo()) {
       if (unsigned BuiltinID = II->getBuiltinID()) {
         if (NewFD->getLanguageLinkage() == CLanguageLinkage) {
-          // Declarations for builtins with custom typechecking by definition
-          // don't make sense. Don't attempt typechecking and simply add the
-          // attribute.
-          if (Context.BuiltinInfo.hasCustomTypechecking(BuiltinID)) {
+          // Validate the type matches unless this builtin is specified as
+          // matching regardless of its declared type.
+          if (Context.BuiltinInfo.allowTypeMismatch(BuiltinID)) {
             NewFD->addAttr(BuiltinAttr::CreateImplicit(Context, BuiltinID));
           } else {
             ASTContext::GetBuiltinTypeError Error;
             LookupNecessaryTypesForBuiltin(S, BuiltinID);
             QualType BuiltinType = Context.GetBuiltinType(BuiltinID, Error);
 
             if (!Error && !BuiltinType.isNull() &&
                 Context.hasSameFunctionTypeIgnoringExceptionSpec(

clang/test/CodeGen/setjmp.c

+// RUN: %clang_cc1 -x c %s -triple x86_64-linux-gnu -emit-llvm -o - | FileCheck %s
+// RUN: %clang_cc1 -x c++ %s -triple x86_64-linux-gnu -emit-llvm -o - | FileCheck %s
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct __jmp_buf_tag { int n; };
+int setjmp(struct __jmp_buf_tag*);
+int sigsetjmp(struct __jmp_buf_tag*, int);
+int _setjmp(struct __jmp_buf_tag*);
+int __sigsetjmp(struct __jmp_buf_tag*, int);
+
+typedef struct __jmp_buf_tag jmp_buf[1];
+typedef struct __jmp_buf_tag sigjmp_buf[1];
+
+#ifdef __cplusplus
+}
+#endif
+
+void f() {
+  jmp_buf jb;
+  // CHECK: call {{.*}}@setjmp(
+  setjmp(jb);
+  // CHECK: call {{.*}}@sigsetjmp(
+  sigsetjmp(jb, 0);
+  // CHECK: call {{.*}}@_setjmp(
+  _setjmp(jb);
+  // CHECK: call {{.*}}@__sigsetjmp(
+  __sigsetjmp(jb, 0);
+}
+
+// CHECK: ; Function Attrs: returns_twice
+// CHECK-NEXT: declare {{.*}} @setjmp(
+
+// CHECK: ; Function Attrs: returns_twice
+// CHECK-NEXT: declare {{.*}} @sigsetjmp(
+
+// CHECK: ; Function Attrs: returns_twice
+// CHECK-NEXT: declare {{.*}} @_setjmp(
+
+// CHECK: ; Function Attrs: returns_twice
+// CHECK-NEXT: declare {{.*}} @__sigsetjmp(
+

clang/test/Sema/builtin-setjmp.c

-// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DNO_JMP_BUF %s
-// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DNO_JMP_BUF %s -ast-dump | FileCheck %s
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DWRONG_JMP_BUF %s -ast-dump | FileCheck %s
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DRIGHT_JMP_BUF %s -ast-dump | FileCheck %s
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DONLY_JMP_BUF %s -ast-dump | FileCheck %s
+// RUN: %clang_cc1 -triple x86_64-unknown-unknown -fsyntax-only -verify -DNO_SETJMP %s -ast-dump 2>&1 | FileCheck %s
 
 #ifdef NO_JMP_BUF
-extern long setjmp(long *);   // expected-warning {{declaration of built-in function 'setjmp' requires the declaration of the 'jmp_buf' type, commonly provided in the header <setjmp.h>.}}
-#else
+// This happens in some versions of glibc: the declaration of __sigsetjmp
+// precedes the declaration of sigjmp_buf.
+extern long setjmp(long *); // Can't check, so we trust that this is the right type
+// FIXME: We could still diagnose the missing `jmp_buf` at the point of the call.
+// expected-no-diagnostics
+#elif WRONG_JMP_BUF
 typedef long jmp_buf;
-extern int setjmp(char);      // expected-warning@8 {{incompatible redeclaration of library function 'setjmp'}}
-                              // expected-note@8    {{'setjmp' is a builtin with type 'int (jmp_buf)' (aka 'int (long)')}}
+extern int setjmp(char); // expected-warning {{incompatible redeclaration of library function 'setjmp'}}
+                         // expected-note@-1 {{'setjmp' is a builtin with type 'int (jmp_buf)' (aka 'int (long)')}}
+#elif RIGHT_JMP_BUF
+typedef long jmp_buf;
+extern int setjmp(long); // OK, right type.
+// expected-no-diagnostics
+#elif ONLY_JMP_BUF
+typedef int *jmp_buf;
+#endif
+
+void use() {
+  setjmp(0);
+  #ifdef NO_SETJMP
+  // expected-warning@-2 {{implicit declaration of function 'setjmp' is invalid in C99}}
+  #elif ONLY_JMP_BUF
+  // expected-warning@-4 {{implicitly declaring library function 'setjmp' with type 'int (jmp_buf)' (aka 'int (int *)')}}
+  // expected-note@-5 {{include the header <setjmp.h> or explicitly provide a declaration for 'setjmp'}}
   #endif
+
+  #ifdef NO_SETJMP
+  // In this case, the regular AST dump doesn't dump the implicit declaration of 'setjmp'.
+  #pragma clang __debug dump setjmp
+  #endif
+}
+
+// CHECK: FunctionDecl {{.*}} used setjmp
+// CHECK: BuiltinAttr {{.*}} Implicit
+// CHECK: ReturnsTwiceAttr {{.*}} Implicit

clang/test/Sema/implicit-builtin-decl.c

 int
 main(int argc, char *argv[])
 {
   snprintf();
 }
 
 void snprintf() { }
 
-// PR8316 & PR40692
-void longjmp(); // expected-warning{{declaration of built-in function 'longjmp' requires the declaration of the 'jmp_buf' type, commonly provided in the header <setjmp.h>.}}
+void longjmp();
 
 extern float fmaxf(float, float);
 
 struct __jmp_buf_tag {};
-void sigsetjmp(struct __jmp_buf_tag[1], int); // expected-warning{{declaration of built-in function 'sigsetjmp' requires the declaration of the 'jmp_buf' type, commonly provided in the header <setjmp.h>.}}
+void sigsetjmp(struct __jmp_buf_tag[1], int);
 
 // PR40692
 void pthread_create(); // no warning expected