Attribute of typedef dropped, leads to failure of extraction.

[5c4lar]

Minimal example to check this:

typedef unsigned short int sa_family_t;
typedef unsigned int __socklen_t;
typedef __socklen_t socklen_t;
typedef int ares_socket_t;
typedef socklen_t ares_socklen_t;

struct sockaddr {
  sa_family_t sa_family;
  char sa_data[14];
};

typedef union {
  struct sockaddr *__restrict __sockaddr__;
} __SOCKADDR_ARG __attribute__((__transparent_union__));

extern int getsockname(int __fd, __SOCKADDR_ARG __addr,
                       socklen_t *__restrict __len)
    __attribute__((__nothrow__));

int main() {
  struct sockaddr *src_addr;
  ares_socket_t sock;
  ares_socklen_t len;
  getsockname(sock, src_addr, &len);
}

extract with the following command

clang-extract exp.c  -DCE_EXTRACT_FUNCTIONS=main -DCE_DUMP_PASSES                                        05:51:01
exp.c:34:21: error: passing 'struct sockaddr *' to parameter of incompatible type '__SOCKADDR_ARG'
   34 |   getsockname(sock, src_addr, &len);
      |                     ^~~~~~~~
exp.c:25:49: note: passing argument to parameter '__addr' here
   25 | extern int getsockname(int __fd, __SOCKADDR_ARG __addr,
      |                                                 ^

Error on pass: ClosurePass

The original source should have been compiled with clang, examine the dumpped pass, the __attribute__((__transparent_union__)) is dropped:

/** clang-extract: from exp.c:1:1  */
typedef unsigned short int sa_family_t;

/** clang-extract: from exp.c:2:1  */
typedef unsigned int __socklen_t;

/** clang-extract: from exp.c:3:1  */
typedef __socklen_t socklen_t;

/** clang-extract: from exp.c:4:1  */
typedef int ares_socket_t;

/** clang-extract: from exp.c:5:1  */
typedef socklen_t ares_socklen_t;

/** clang-extract: from exp.c:10:1  */
struct sockaddr;

/** clang-extract: from exp.c:12:1  */
typedef union {
  struct sockaddr *__restrict __sockaddr__;
} __SOCKADDR_ARG;

/** clang-extract: from exp.c:16:1  */
extern int getsockname(int __fd, __SOCKADDR_ARG __addr,
                       socklen_t *__restrict __len)
    __attribute__((__nothrow__));

/** clang-extract: from exp.c:20:1  */
int main() {
  struct sockaddr *src_addr;
  ares_socket_t sock;
  ares_socklen_t len;
  getsockname(sock, src_addr, &len);
}

[giulianobelinassi]

Confirmed. We may need to patch PrettyPrint::Get_Expanded_Loc to also look for the attributes applied to the structs or unions that the typedef bounds to.

[giulianobelinassi]

@5c4lar can you check if PR #69 fixes your problem?

[5c4lar]

@5c4lar can you check if PR #69 fixes your problem?

Yes, it fixes the problem! Thanks for your timely response!

[5c4lar]

@giulianobelinassi Hi! I found a new case that will make the extraction fail, can you have a look at it?

typedef long long __m64 __attribute__((__vector_size__(8), __aligned__(8)));

typedef int __v2si __attribute__((__vector_size__(8)));

static __inline__ __m64
    __attribute__((__always_inline__, __nodebug__, __target__("mmx,no-evex512"),
                   __min_vector_width__(64)))
    _mm_cmpgt_pi32(__m64 __m1, __m64 __m2) {
  return (__m64)__builtin_ia32_pcmpgtd((__v2si)__m1, (__v2si)__m2);
}

clang-extract -Wno-everything test.c -DCE_EXTRACT_FUNCTIONS=_mm_cmpgt_pi32 -DCE_OUTPUT_FILE=-
test.c:12:40: error: invalid conversion between vector type '__m64' (vector of 1 'long long' value) and integer type '__v2si' (aka 'int') of different size
   12 |   return (__m64)__builtin_ia32_pcmpgtd((__v2si)__m1, (__v2si)__m2);
      |                                        ^~~~~~~~~~~~
test.c:12:54: error: invalid conversion between vector type '__m64' (vector of 1 'long long' value) and integer type '__v2si' (aka 'int') of different size
   12 |   return (__m64)__builtin_ia32_pcmpgtd((__v2si)__m1, (__v2si)__m2);
      |                                                      ^~~~~~~~~~~~

Error on pass: ClosurePass

It seems that this is also because of the dropped attribute of typedef. But why is that of _m64 kept, but that of __v2si dropped?

cat test.dump.3.ClosurePass.c
/** clang-extract: from test.c:224:1  */
typedef long long __m64 __attribute__((__vector_size__(8), __aligned__(8)));

/** clang-extract: from test.c:226:1  */
typedef int __v2si;

/** clang-extract: from test.c:564:1  */
static __inline__ __m64
    __attribute__((__always_inline__, __nodebug__, __target__("mmx,no-evex512"),
                   __min_vector_width__(64)))
    _mm_cmpgt_pi32(__m64 __m1, __m64 __m2) {
  return (__m64)__builtin_ia32_pcmpgtd((__v2si)__m1, (__v2si)__m2);
}

[giulianobelinassi]

Reopening this. I will take a look at it tomorrow.

[giulianobelinassi]

It looks like some attributes are not stored in the Decl itself but rather the QualType of the Decl. From clang's DeclPrinter.cpp:

void DeclPrinter::VisitTypedefDecl(TypedefDecl *D) {
  if (!Policy.SuppressSpecifiers) {
    Out << "typedef ";

    if (D->isModulePrivate())
      Out << "__module_private__ ";
  }
  QualType Ty = D->getTypeSourceInfo()->getType();
  Ty.print(Out, Policy, D->getName(), Indentation);
  prettyPrintAttributes(D);
}

Hence we also need to analyze the QualTypes for the attributes.

[giulianobelinassi]

@5c4lar Please check if #76 fixes this issue.

[5c4lar]

@giulianobelinassi The previous testcase is fixed. But when I test it on other functions in the same project, there seems to be lots of other cases.

typedef short __v8hi __attribute__((__vector_size__(16)));

typedef long long __m128i __attribute__((__vector_size__(16), __aligned__(16)));

typedef signed char __v16qs __attribute__((__vector_size__(16)));

static __inline__ __m128i __attribute__((__always_inline__, __nodebug__, __target__("sse4.1,no-evex512"), __min_vector_width__(128))) _mm_cvtepi8_epi16(__m128i __V) {
  return (__m128i) __builtin_convertvector(
      __builtin_shufflevector((__v16qs)__V, (__v16qs)__V, 0, 1, 2, 3, 4, 5, 6,
                              7),
      __v8hi);
}

/** clang-extract: from test.c:3:1 */
typedef long long __m128i __attribute__((__vector_size__(16), __aligned__(16)));

/** clang-extract: from test.c:5:1 */
typedef signed char __v16qs __attribute__((__vector_size__(16)));

/** clang-extract: from test.c:7:1 */
static __inline__ __m128i __attribute__((__always_inline__, __nodebug__, __target__("sse4.1,no-evex512"), __min_vector_width__(128))) _mm_cvtepi8_epi16(__m128i __V) {
  return (__m128i) __builtin_convertvector(
      __builtin_shufflevector((__v16qs)__V, (__v16qs)__V, 0, 1, 2, 3, 4, 5, 6,
                              7),
      __v8hi);
}

For example, the typedef of __v8hi is dropped completely here.

[5c4lar]

And here is another case:

typedef float __v8sf __attribute__ ((__vector_size__ (32)));

typedef unsigned int __v8su __attribute__ ((__vector_size__ (32)));

typedef float __m256 __attribute__ ((__vector_size__ (32), __aligned__(32)));

typedef long long __m256i __attribute__((__vector_size__(32), __aligned__(32)));

static __inline__ __m256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,no-evex512"), __min_vector_width__(256)))
_mm256_cvtepu32_ps (__m256i __A) {
  return (__m256)__builtin_convertvector((__v8su)__A, __v8sf);
}

where the definition of __v8sf will be dropped

[giulianobelinassi]

So it turns out that __builtin_convertvector is not a function but rather a specific kind of Expr in clang AST:

FunctionDecl 0x557b7ebd4e90 <line:7:1, line:12:1> line:7:135 _mm_cvtepi8_epi16 '__m128i (__m128i)' static inline
| |-ParmVarDecl 0x557b7ebd4d80 <col:153, col:161> col:161 used __V '__m128i':'__attribute__((__vector_size__(2 * sizeof(long long)))) long long'
| |-CompoundStmt 0x557b7ebd56b0 <col:166, line:12:1>                           
| | `-ReturnStmt 0x557b7ebd56a0 <line:8:3, line:11:13>                                                                                                                                        | |   `-CStyleCastExpr 0x557b7ebd5678 <line:8:10, line:11:13> '__m128i':'__attribute__((__vector_size__(2 * sizeof(long long)))) long long' <BitCast>                                         
| |     `-ConvertVectorExpr 0x557b7ebd5640 <line:8:20, line:11:13> '__v8hi':'__attribute__((__vector_size__(8 * sizeof(short)))) short'
| |       `-ShuffleVectorExpr 0x557b7ebd5550 <line:9:7, line:10:32> '__attribute__((__vector_size__(8 * sizeof(signed char)))) signed char'
| |         |-CStyleCastExpr 0x557b7ebd52c8 <line:9:31, col:40> '__v16qs':'__attribute__((__vector_size__(16 * sizeof(signed char)))) signed char' <BitCast>
| |         | `-ImplicitCastExpr 0x557b7ebd52b0 <col:40> '__m128i':'__attribute__((__vector_size__(2 * sizeof(long long)))) long long' <LValueToRValue> part_of_explicit_cast
| |         |   `-DeclRefExpr 0x557b7ebd5280 <col:40> '__m128i':'__attribute__((__vector_size__(2 * sizeof(long long)))) long long' lvalue ParmVar 0x557b7ebd4d80 '__V' '__m128i':'__attribute
__((__vector_size__(2 * sizeof(long long)))) long long' 

And this escapes the Closure computation because of that. I will submit a patch to it soon.

[giulianobelinassi]

@5c4lar I updated #76 with a fix for this. Please check if it works :)

[5c4lar]

@giulianobelinassi Thanks! The solution works well for the scenarios we've discussed previously. However, I've encountered a new situation where clang successfully compiles the code, but clang-extract encounters an error. This issue seems to be unrelated to our previous discussions, so I'll create a separate issue to address it specifically.