libffi · tdermendjiev · Feb 28, 2018 · Mar 1, 2018 · Mar 16, 2018 · Mar 16, 2018
diff --git a/configure.ac b/configure.ac
@@ -72,6 +72,7 @@ AM_CONDITIONAL(TESTSUBDIR, test -d $srcdir/testsuite)
 
 TARGETDIR="unknown"
 HAVE_LONG_DOUBLE_VARIANT=0
+HAVE_EXT_VECTOR=0
 
 . ${srcdir}/configure.host
 
@@ -106,6 +107,7 @@ if test -z "$HAVE_LONG_DOUBLE"; then
 fi
 AC_SUBST(HAVE_LONG_DOUBLE)
 AC_SUBST(HAVE_LONG_DOUBLE_VARIANT)
+AC_SUBST(HAVE_EXT_VECTOR)
 
 AC_C_BIGENDIAN
 
@@ -203,6 +205,12 @@ case "$target" in
                    alternative means])
      ;;
 esac
+
+case "$target" in
+     aarch64-apple-darwin*)
+       HAVE_EXT_VECTOR=1
+esac
+
 AM_CONDITIONAL(FFI_EXEC_TRAMPOLINE_TABLE, test x$FFI_EXEC_TRAMPOLINE_TABLE = x1)
 AC_SUBST(FFI_EXEC_TRAMPOLINE_TABLE)
 

diff --git a/include/ffi.h.in b/include/ffi.h.in
@@ -478,9 +478,18 @@ ffi_status ffi_get_struct_offsets (ffi_abi abi, ffi_type *struct_type,
 #define FFI_TYPE_STRUCT     13
 #define FFI_TYPE_POINTER    14
 #define FFI_TYPE_COMPLEX    15
+#if @HAVE_EXT_VECTOR@
+#define FFI_TYPE_EXT_VECTOR 16
+#else
+#define FFI_TYPE_EXT_VECTOR FFI_TYPE_STRUCT
+#endif
 
 /* This should always refer to the last type code (for sanity checks).  */
-#define FFI_TYPE_LAST       FFI_TYPE_COMPLEX
+#if @HAVE_EXT_VECTOR@
+#define FFI_TYPE_LAST  FFI_TYPE_EXT_VECTOR
+#else
+#define FFI_TYPE_LAST  FFI_TYPE_COMPLEX
+#endif  
 
 #ifdef __cplusplus
 }

diff --git a/src/aarch64/ffi.c b/src/aarch64/ffi.c
@@ -81,6 +81,12 @@ ffi_clear_cache (void *start, void *end)
 
 #endif
 
+static int intlog2(int n) {
+    int targetlevel = 0;
+    while (n >>= 1) ++targetlevel;
+    return targetlevel;
+}
+
 /* A subroutine of is_vfp_type.  Given a structure type, return the type code
    of the first non-structure element.  Recurse for structure elements.
    Return -1 if the structure is in fact empty, i.e. no nested elements.  */
@@ -95,7 +101,7 @@ is_hfa0 (const ffi_type *ty)
     for (i = 0; elements[i]; ++i)
       {
         ret = elements[i]->type;
-        if (ret == FFI_TYPE_STRUCT || ret == FFI_TYPE_COMPLEX)
+        if (ret == FFI_TYPE_STRUCT || ret == FFI_TYPE_COMPLEX || ret == FFI_TYPE_EXT_VECTOR)
           {
             ret = is_hfa0 (elements[i]);
             if (ret < 0)
@@ -107,6 +113,30 @@ is_hfa0 (const ffi_type *ty)
   return ret;
 }
 
+static size_t is_simd(const ffi_type *ty) // return 0 if no SIMD elements
+{
+    if (ty->type == FFI_TYPE_EXT_VECTOR || ty->type == FFI_TYPE_COMPLEX) {
+        return ty->size;
+    }
+    ffi_type **elements = ty->elements;
+    int i;
+
+    if (elements != NULL)
+    {
+        for (i = 0; elements[i]; ++i)
+        {
+            int element_type = elements[i]->type;
+            if (element_type == FFI_TYPE_STRUCT || element_type == FFI_TYPE_COMPLEX || element_type == FFI_TYPE_EXT_VECTOR)
+            {
+                return is_simd(elements[i]);
+            }
+        }
+    }
+
+    return 0;
+
+}
+
 /* A subroutine of is_vfp_type.  Given a structure type, return true if all
    of the non-structure elements are the same as CANDIDATE.  */
 
@@ -120,7 +150,7 @@ is_hfa1 (const ffi_type *ty, int candidate)
     for (i = 0; elements[i]; ++i)
       {
         int t = elements[i]->type;
-        if (t == FFI_TYPE_STRUCT || t == FFI_TYPE_COMPLEX)
+        if (t == FFI_TYPE_STRUCT || t == FFI_TYPE_COMPLEX || t == FFI_TYPE_EXT_VECTOR)
           {
             if (!is_hfa1 (elements[i], candidate))
               return 0;
@@ -145,7 +175,7 @@ is_vfp_type (const ffi_type *ty)
 {
   ffi_type **elements;
   int candidate, i;
-  size_t size, ele_count;
+  size_t size, ele_count, simd_size;
 
   /* Quickest tests first.  */
   candidate = ty->type;
@@ -170,6 +200,7 @@ is_vfp_type (const ffi_type *ty)
 	}
       return 0;
     case FFI_TYPE_STRUCT:
+    case FFI_TYPE_EXT_VECTOR:
       break;
     }
 
@@ -181,7 +212,9 @@ is_vfp_type (const ffi_type *ty)
   /* Find the type of the first non-structure member.  */
   elements = ty->elements;
   candidate = elements[0]->type;
-  if (candidate == FFI_TYPE_STRUCT || candidate == FFI_TYPE_COMPLEX)
+  /* Find the size of the SIMD member, if any. */
+  simd_size = is_simd(ty);
+  if (candidate == FFI_TYPE_STRUCT || candidate == FFI_TYPE_COMPLEX || candidate == FFI_TYPE_EXT_VECTOR)
     {
       for (i = 0; ; ++i)
         {
@@ -213,14 +246,14 @@ is_vfp_type (const ffi_type *ty)
     default:
       return 0;
     }
-  if (ele_count > 4)
+  if ((ele_count > 4 && !simd_size) || (ele_count > 16 && simd_size))
     return 0;
 
   /* Finally, make sure that all scalar elements are the same type.  */
   for (i = 0; elements[i]; ++i)
     {
       int t = elements[i]->type;
-      if (t == FFI_TYPE_STRUCT || t == FFI_TYPE_COMPLEX)
+      if (t == FFI_TYPE_STRUCT || t == FFI_TYPE_COMPLEX || t == FFI_TYPE_EXT_VECTOR)
         {
           if (!is_hfa1 (elements[i], candidate))
             return 0;
@@ -230,8 +263,16 @@ is_vfp_type (const ffi_type *ty)
     }
 
   /* All tests succeeded.  Encode the result.  */
- done:
-  return candidate * 4 + (4 - (int)ele_count);
+    if (simd_size) {
+        size_t regSize = simd_size;
+
+        int num_registers = (int) size / regSize + (size % regSize ? 1 : 0);
+        int first_level_element_type = FFI_TYPE_FLOAT + intlog2((int)regSize) - 2;
+        return (int) (first_level_element_type * 4 + (4 - num_registers));
+    }
+
+  done:
+    return candidate * 4 + (4 - (int)ele_count);
 }
 
 /* Representation of the procedure call argument marshalling
@@ -497,6 +538,7 @@ ffi_prep_cif_machdep (ffi_cif *cif)
     case FFI_TYPE_DOUBLE:
     case FFI_TYPE_LONGDOUBLE:
     case FFI_TYPE_STRUCT:
+    case FFI_TYPE_EXT_VECTOR:
     case FFI_TYPE_COMPLEX:
       flags = is_vfp_type (rtype);
       if (flags == 0)
@@ -643,6 +685,7 @@ ffi_call_int (ffi_cif *cif, void (*fn)(void), void *orig_rvalue,
 	case FFI_TYPE_DOUBLE:
 	case FFI_TYPE_LONGDOUBLE:
 	case FFI_TYPE_STRUCT:
+    case FFI_TYPE_EXT_VECTOR:
 	case FFI_TYPE_COMPLEX:
 	  {
 	    void *dest;
@@ -866,6 +909,7 @@ ffi_closure_SYSV_inner (ffi_cif *cif,
 	case FFI_TYPE_DOUBLE:
 	case FFI_TYPE_LONGDOUBLE:
 	case FFI_TYPE_STRUCT:
+    case FFI_TYPE_EXT_VECTOR:
 	case FFI_TYPE_COMPLEX:
 	  h = is_vfp_type (ty);
 	  if (h)

diff --git a/testsuite/Makefile.am b/testsuite/Makefile.am
@@ -114,4 +114,6 @@ libffi.go/aa-direct.c libffi.go/ffitest.h libffi.go/go.exp		\
 libffi.go/static-chain.h libffi.bhaible/bhaible.exp			\
 libffi.bhaible/test-call.c libffi.bhaible/alignof.h			\
 libffi.bhaible/testcases.c libffi.bhaible/test-callback.c		\
-libffi.bhaible/Makefile libffi.bhaible/README config/default.exp  
+libffi.bhaible/Makefile libffi.bhaible/README config/default.exp \
+libffi.call/simd_float_matrices.c \
+libffi.call/simd_double_matrices.c