kc3-lang/libffi

diff --git a/src/aarch64/ffi.c b/src/aarch64/ffi.c
index f79602b..cf33abc 100644
--- a/src/aarch64/ffi.c
+++ b/src/aarch64/ffi.c
@@ -22,6 +22,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
 #include <stdio.h>
 #include <stdlib.h>
 #include <stdint.h>
+#include <fficonfig.h>
 #include <ffi.h>
 #include <ffi_common.h>
 #include "internal.h"
@@ -70,6 +71,14 @@ ffi_clear_cache (void *start, void *end)
 #endif
 }
 
+#if FFI_EXEC_TRAMPOLINE_TABLE
+
+#ifdef __MACH__
+#include <mach/vm_param.h>
+#endif
+
+#endif
+
 /* 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.  */
@@ -725,240 +734,6 @@ ffi_call_go (ffi_cif *cif, void (*fn) (void), void *rvalue,
 extern void ffi_closure_SYSV (void) FFI_HIDDEN;
 extern void ffi_closure_SYSV_V (void) FFI_HIDDEN;
 
-#if FFI_EXEC_TRAMPOLINE_TABLE
-
-#include <mach/mach.h>
-#include <pthread.h>
-#include <stdio.h>
-#include <stdlib.h>
-
-extern void *ffi_closure_trampoline_table_page;
-
-typedef struct ffi_trampoline_table ffi_trampoline_table;
-typedef struct ffi_trampoline_table_entry ffi_trampoline_table_entry;
-
-struct ffi_trampoline_table
-{
-  /* contiguous writable and executable pages */
-  vm_address_t config_page;
-  vm_address_t trampoline_page;
-
-  /* free list tracking */
-  uint16_t free_count;
-  ffi_trampoline_table_entry *free_list;
-  ffi_trampoline_table_entry *free_list_pool;
-
-  ffi_trampoline_table *prev;
-  ffi_trampoline_table *next;
-};
-
-struct ffi_trampoline_table_entry
-{
-  void *(*trampoline) ();
-  ffi_trampoline_table_entry *next;
-};
-
-/* The trampoline configuration is placed a page prior to the trampoline's entry point */
-#define FFI_TRAMPOLINE_CODELOC_CONFIG(codeloc) ((void **) (((uint8_t *) codeloc) - PAGE_SIZE));
-
-/* Total number of trampolines that fit in one trampoline table */
-#define FFI_TRAMPOLINE_COUNT (PAGE_SIZE / FFI_TRAMPOLINE_SIZE)
-
-static pthread_mutex_t ffi_trampoline_lock = PTHREAD_MUTEX_INITIALIZER;
-static ffi_trampoline_table *ffi_trampoline_tables = NULL;
-
-static ffi_trampoline_table *
-ffi_trampoline_table_alloc ()
-{
-  ffi_trampoline_table *table = NULL;
-
-  /* Loop until we can allocate two contiguous pages */
-  while (table == NULL)
-    {
-      vm_address_t config_page = 0x0;
-      kern_return_t kt;
-
-      /* Try to allocate two pages */
-      kt =
-	vm_allocate (mach_task_self (), &config_page, PAGE_SIZE * 2,
-		     VM_FLAGS_ANYWHERE);
-      if (kt != KERN_SUCCESS)
-	{
-	  fprintf (stderr, "vm_allocate() failure: %d at %s:%d\n", kt,
-		   __FILE__, __LINE__);
-	  break;
-	}
-
-      /* Now drop the second half of the allocation to make room for the trampoline table */
-      vm_address_t trampoline_page = config_page + PAGE_SIZE;
-      kt = vm_deallocate (mach_task_self (), trampoline_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	{
-	  fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		   __FILE__, __LINE__);
-	  break;
-	}
-
-      /* Remap the trampoline table to directly follow the config page */
-      vm_prot_t cur_prot;
-      vm_prot_t max_prot;
-
-      kt =
-	vm_remap (mach_task_self (), &trampoline_page, PAGE_SIZE, 0x0, FALSE,
-		  mach_task_self (),
-		  (vm_address_t) & ffi_closure_trampoline_table_page, FALSE,
-		  &cur_prot, &max_prot, VM_INHERIT_SHARE);
-
-      /* If we lost access to the destination trampoline page, drop our config allocation mapping and retry */
-      if (kt != KERN_SUCCESS)
-	{
-	  /* Log unexpected failures */
-	  if (kt != KERN_NO_SPACE)
-	    {
-	      fprintf (stderr, "vm_remap() failure: %d at %s:%d\n", kt,
-		       __FILE__, __LINE__);
-	    }
-
-	  vm_deallocate (mach_task_self (), config_page, PAGE_SIZE);
-	  continue;
-	}
-
-      /* We have valid trampoline and config pages */
-      table = calloc (1, sizeof (ffi_trampoline_table));
-      table->free_count = FFI_TRAMPOLINE_COUNT;
-      table->config_page = config_page;
-      table->trampoline_page = trampoline_page;
-
-      /* Create and initialize the free list */
-      table->free_list_pool =
-	calloc (FFI_TRAMPOLINE_COUNT, sizeof (ffi_trampoline_table_entry));
-
-      uint16_t i;
-      for (i = 0; i < table->free_count; i++)
-	{
-	  ffi_trampoline_table_entry *entry = &table->free_list_pool[i];
-	  entry->trampoline =
-	    (void *) (table->trampoline_page + (i * FFI_TRAMPOLINE_SIZE));
-
-	  if (i < table->free_count - 1)
-	    entry->next = &table->free_list_pool[i + 1];
-	}
-
-      table->free_list = table->free_list_pool;
-    }
-
-  return table;
-}
-
-void *
-ffi_closure_alloc (size_t size, void **code)
-{
-  /* Create the closure */
-  ffi_closure *closure = malloc (size);
-  if (closure == NULL)
-    return NULL;
-
-  pthread_mutex_lock (&ffi_trampoline_lock);
-
-  /* Check for an active trampoline table with available entries. */
-  ffi_trampoline_table *table = ffi_trampoline_tables;
-  if (table == NULL || table->free_list == NULL)
-    {
-      table = ffi_trampoline_table_alloc ();
-      if (table == NULL)
-	{
-	  free (closure);
-	  return NULL;
-	}
-
-      /* Insert the new table at the top of the list */
-      table->next = ffi_trampoline_tables;
-      if (table->next != NULL)
-	table->next->prev = table;
-
-      ffi_trampoline_tables = table;
-    }
-
-  /* Claim the free entry */
-  ffi_trampoline_table_entry *entry = ffi_trampoline_tables->free_list;
-  ffi_trampoline_tables->free_list = entry->next;
-  ffi_trampoline_tables->free_count--;
-  entry->next = NULL;
-
-  pthread_mutex_unlock (&ffi_trampoline_lock);
-
-  /* Initialize the return values */
-  *code = entry->trampoline;
-  closure->trampoline_table = table;
-  closure->trampoline_table_entry = entry;
-
-  return closure;
-}
-
-void
-ffi_closure_free (void *ptr)
-{
-  ffi_closure *closure = ptr;
-
-  pthread_mutex_lock (&ffi_trampoline_lock);
-
-  /* Fetch the table and entry references */
-  ffi_trampoline_table *table = closure->trampoline_table;
-  ffi_trampoline_table_entry *entry = closure->trampoline_table_entry;
-
-  /* Return the entry to the free list */
-  entry->next = table->free_list;
-  table->free_list = entry;
-  table->free_count++;
-
-  /* If all trampolines within this table are free, and at least one other table exists, deallocate
-   * the table */
-  if (table->free_count == FFI_TRAMPOLINE_COUNT
-      && ffi_trampoline_tables != table)
-    {
-      /* Remove from the list */
-      if (table->prev != NULL)
-	table->prev->next = table->next;
-
-      if (table->next != NULL)
-	table->next->prev = table->prev;
-
-      /* Deallocate pages */
-      kern_return_t kt;
-      kt = vm_deallocate (mach_task_self (), table->config_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		 __FILE__, __LINE__);
-
-      kt =
-	vm_deallocate (mach_task_self (), table->trampoline_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		 __FILE__, __LINE__);
-
-      /* Deallocate free list */
-      free (table->free_list_pool);
-      free (table);
-    }
-  else if (ffi_trampoline_tables != table)
-    {
-      /* Otherwise, bump this table to the top of the list */
-      table->prev = NULL;
-      table->next = ffi_trampoline_tables;
-      if (ffi_trampoline_tables != NULL)
-	ffi_trampoline_tables->prev = table;
-
-      ffi_trampoline_tables = table;
-    }
-
-  pthread_mutex_unlock (&ffi_trampoline_lock);
-
-  /* Free the closure */
-  free (closure);
-}
-
-#endif
-
 ffi_status
 ffi_prep_closure_loc (ffi_closure *closure,
                       ffi_cif* cif,
@@ -977,9 +752,11 @@ ffi_prep_closure_loc (ffi_closure *closure,
     start = ffi_closure_SYSV;
 
 #if FFI_EXEC_TRAMPOLINE_TABLE
-  void **config = FFI_TRAMPOLINE_CODELOC_CONFIG (codeloc);
+#ifdef __MACH__
+  void **config = (void **)((uint8_t *)codeloc - PAGE_MAX_SIZE);
   config[0] = closure;
   config[1] = start;
+#endif
 #else
   static const unsigned char trampoline[16] = {
     0x90, 0x00, 0x00, 0x58,	/* ldr	x16, tramp+16	*/
diff --git a/src/aarch64/ffitarget.h b/src/aarch64/ffitarget.h
index 2862ec7..5ded0e4 100644
--- a/src/aarch64/ffitarget.h
+++ b/src/aarch64/ffitarget.h
@@ -48,14 +48,21 @@ typedef enum ffi_abi
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
-#if defined (__APPLE__)
-#define FFI_TRAMPOLINE_SIZE 20
+#define FFI_NATIVE_RAW_API 0
+
+#if defined (FFI_EXEC_TRAMPOLINE_TABLE) && FFI_EXEC_TRAMPOLINE_TABLE
+
+#ifdef __MACH__
+#define FFI_TRAMPOLINE_SIZE 16
 #define FFI_TRAMPOLINE_CLOSURE_OFFSET 16
 #else
+#error "No trampoline table implementation"
+#endif
+
+#else
 #define FFI_TRAMPOLINE_SIZE 24
 #define FFI_TRAMPOLINE_CLOSURE_OFFSET FFI_TRAMPOLINE_SIZE
 #endif
-#define FFI_NATIVE_RAW_API 0
 
 /* ---- Internal ---- */
 
diff --git a/src/aarch64/sysv.S b/src/aarch64/sysv.S
index c1bf9b9..7936c0e 100644
--- a/src/aarch64/sysv.S
+++ b/src/aarch64/sysv.S
@@ -356,16 +356,18 @@ CNAME(ffi_closure_SYSV):
 #endif
 
 #if FFI_EXEC_TRAMPOLINE_TABLE
-    .align 12
+
+#ifdef __MACH__
+#include <mach/vm_param.h>
+    .align PAGE_MAX_SHIFT
 CNAME(ffi_closure_trampoline_table_page):
-    .rept 16384 / FFI_TRAMPOLINE_SIZE
-    adr	x17, -16384
-    adr	x16, -16380
-    ldr x16, [x16]
-    ldr x17, [x17]
-    br	x16
+    .rept PAGE_MAX_SIZE / FFI_TRAMPOLINE_SIZE
+    adr x16, -PAGE_MAX_SIZE
+    ldp x17, x16, [x16]
+    br x16
+	nop		/* each entry in the trampoline config page is 2*sizeof(void*) so the trampoline itself cannot be smaller that 16 bytes */
     .endr
-    
+
     .globl CNAME(ffi_closure_trampoline_table_page)
     #ifdef __ELF__
     	.type	CNAME(ffi_closure_trampoline_table_page), #function
@@ -374,6 +376,8 @@ CNAME(ffi_closure_trampoline_table_page):
     #endif
 #endif
 
+#endif /* FFI_EXEC_TRAMPOLINE_TABLE */
+
 #ifdef FFI_GO_CLOSURES
 	.align 4
 CNAME(ffi_go_closure_SYSV_V):
diff --git a/src/arm/ffi.c b/src/arm/ffi.c
index 9c8732d..c24085d 100644
--- a/src/arm/ffi.c
+++ b/src/arm/ffi.c
@@ -28,11 +28,22 @@
    DEALINGS IN THE SOFTWARE.
    ----------------------------------------------------------------------- */
 
+#include <fficonfig.h>
 #include <ffi.h>
 #include <ffi_common.h>
 #include <stdlib.h>
 #include "internal.h"
 
+#if FFI_EXEC_TRAMPOLINE_TABLE
+
+#ifdef __MACH__
+#include <mach/vm_param.h>
+#endif
+
+#else
+extern unsigned int ffi_arm_trampoline[2] FFI_HIDDEN;
+#endif
+
 /* Forward declares. */
 static int vfp_type_p (const ffi_type *);
 static void layout_vfp_args (ffi_cif *);
@@ -530,252 +541,6 @@ void ffi_closure_VFP (void) FFI_HIDDEN;
 void ffi_go_closure_SYSV (void) FFI_HIDDEN;
 void ffi_go_closure_VFP (void) FFI_HIDDEN;
 
-#if FFI_EXEC_TRAMPOLINE_TABLE
-
-#include <mach/mach.h>
-#include <pthread.h>
-#include <stdio.h>
-#include <stdlib.h>
-
-extern void *ffi_closure_trampoline_table_page;
-
-typedef struct ffi_trampoline_table ffi_trampoline_table;
-typedef struct ffi_trampoline_table_entry ffi_trampoline_table_entry;
-
-struct ffi_trampoline_table
-{
-  /* contiguous writable and executable pages */
-  vm_address_t config_page;
-  vm_address_t trampoline_page;
-
-  /* free list tracking */
-  uint16_t free_count;
-  ffi_trampoline_table_entry *free_list;
-  ffi_trampoline_table_entry *free_list_pool;
-
-  ffi_trampoline_table *prev;
-  ffi_trampoline_table *next;
-};
-
-struct ffi_trampoline_table_entry
-{
-  void *(*trampoline) ();
-  ffi_trampoline_table_entry *next;
-};
-
-/* Override the standard architecture trampoline size */
-// XXX TODO - Fix
-#undef FFI_TRAMPOLINE_SIZE
-#define FFI_TRAMPOLINE_SIZE 12
-
-/* The trampoline configuration is placed at 4080 bytes prior to the trampoline's entry point */
-#define FFI_TRAMPOLINE_CODELOC_CONFIG(codeloc) ((void **) (((uint8_t *) codeloc) - 4080));
-
-/* The first 16 bytes of the config page are unused, as they are unaddressable from the trampoline page. */
-#define FFI_TRAMPOLINE_CONFIG_PAGE_OFFSET 16
-
-/* Total number of trampolines that fit in one trampoline table */
-#define FFI_TRAMPOLINE_COUNT ((PAGE_SIZE - FFI_TRAMPOLINE_CONFIG_PAGE_OFFSET) / FFI_TRAMPOLINE_SIZE)
-
-static pthread_mutex_t ffi_trampoline_lock = PTHREAD_MUTEX_INITIALIZER;
-static ffi_trampoline_table *ffi_trampoline_tables = NULL;
-
-static ffi_trampoline_table *
-ffi_trampoline_table_alloc ()
-{
-  ffi_trampoline_table *table = NULL;
-
-  /* Loop until we can allocate two contiguous pages */
-  while (table == NULL)
-    {
-      vm_address_t config_page = 0x0;
-      kern_return_t kt;
-
-      /* Try to allocate two pages */
-      kt =
-	vm_allocate (mach_task_self (), &config_page, PAGE_SIZE * 2,
-		     VM_FLAGS_ANYWHERE);
-      if (kt != KERN_SUCCESS)
-	{
-	  fprintf (stderr, "vm_allocate() failure: %d at %s:%d\n", kt,
-		   __FILE__, __LINE__);
-	  break;
-	}
-
-      /* Now drop the second half of the allocation to make room for the trampoline table */
-      vm_address_t trampoline_page = config_page + PAGE_SIZE;
-      kt = vm_deallocate (mach_task_self (), trampoline_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	{
-	  fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		   __FILE__, __LINE__);
-	  break;
-	}
-
-      /* Remap the trampoline table to directly follow the config page */
-      vm_prot_t cur_prot;
-      vm_prot_t max_prot;
-
-      kt =
-	vm_remap (mach_task_self (), &trampoline_page, PAGE_SIZE, 0x0, FALSE,
-		  mach_task_self (),
-		  (vm_address_t) & ffi_closure_trampoline_table_page, FALSE,
-		  &cur_prot, &max_prot, VM_INHERIT_SHARE);
-
-      /* If we lost access to the destination trampoline page, drop our config allocation mapping and retry */
-      if (kt != KERN_SUCCESS)
-	{
-	  /* Log unexpected failures */
-	  if (kt != KERN_NO_SPACE)
-	    {
-	      fprintf (stderr, "vm_remap() failure: %d at %s:%d\n", kt,
-		       __FILE__, __LINE__);
-	    }
-
-	  vm_deallocate (mach_task_self (), config_page, PAGE_SIZE);
-	  continue;
-	}
-
-      /* We have valid trampoline and config pages */
-      table = calloc (1, sizeof (ffi_trampoline_table));
-      table->free_count = FFI_TRAMPOLINE_COUNT;
-      table->config_page = config_page;
-      table->trampoline_page = trampoline_page;
-
-      /* Create and initialize the free list */
-      table->free_list_pool =
-	calloc (FFI_TRAMPOLINE_COUNT, sizeof (ffi_trampoline_table_entry));
-
-      uint16_t i;
-      for (i = 0; i < table->free_count; i++)
-	{
-	  ffi_trampoline_table_entry *entry = &table->free_list_pool[i];
-	  entry->trampoline =
-	    (void *) (table->trampoline_page + (i * FFI_TRAMPOLINE_SIZE));
-
-	  if (i < table->free_count - 1)
-	    entry->next = &table->free_list_pool[i + 1];
-	}
-
-      table->free_list = table->free_list_pool;
-    }
-
-  return table;
-}
-
-void *
-ffi_closure_alloc (size_t size, void **code)
-{
-  /* Create the closure */
-  ffi_closure *closure = malloc (size);
-  if (closure == NULL)
-    return NULL;
-
-  pthread_mutex_lock (&ffi_trampoline_lock);
-
-  /* Check for an active trampoline table with available entries. */
-  ffi_trampoline_table *table = ffi_trampoline_tables;
-  if (table == NULL || table->free_list == NULL)
-    {
-      table = ffi_trampoline_table_alloc ();
-      if (table == NULL)
-	{
-	  free (closure);
-	  return NULL;
-	}
-
-      /* Insert the new table at the top of the list */
-      table->next = ffi_trampoline_tables;
-      if (table->next != NULL)
-	table->next->prev = table;
-
-      ffi_trampoline_tables = table;
-    }
-
-  /* Claim the free entry */
-  ffi_trampoline_table_entry *entry = ffi_trampoline_tables->free_list;
-  ffi_trampoline_tables->free_list = entry->next;
-  ffi_trampoline_tables->free_count--;
-  entry->next = NULL;
-
-  pthread_mutex_unlock (&ffi_trampoline_lock);
-
-  /* Initialize the return values */
-  *code = entry->trampoline;
-  closure->trampoline_table = table;
-  closure->trampoline_table_entry = entry;
-
-  return closure;
-}
-
-void
-ffi_closure_free (void *ptr)
-{
-  ffi_closure *closure = ptr;
-
-  pthread_mutex_lock (&ffi_trampoline_lock);
-
-  /* Fetch the table and entry references */
-  ffi_trampoline_table *table = closure->trampoline_table;
-  ffi_trampoline_table_entry *entry = closure->trampoline_table_entry;
-
-  /* Return the entry to the free list */
-  entry->next = table->free_list;
-  table->free_list = entry;
-  table->free_count++;
-
-  /* If all trampolines within this table are free, and at least one other table exists, deallocate
-   * the table */
-  if (table->free_count == FFI_TRAMPOLINE_COUNT
-      && ffi_trampoline_tables != table)
-    {
-      /* Remove from the list */
-      if (table->prev != NULL)
-	table->prev->next = table->next;
-
-      if (table->next != NULL)
-	table->next->prev = table->prev;
-
-      /* Deallocate pages */
-      kern_return_t kt;
-      kt = vm_deallocate (mach_task_self (), table->config_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		 __FILE__, __LINE__);
-
-      kt =
-	vm_deallocate (mach_task_self (), table->trampoline_page, PAGE_SIZE);
-      if (kt != KERN_SUCCESS)
-	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
-		 __FILE__, __LINE__);
-
-      /* Deallocate free list */
-      free (table->free_list_pool);
-      free (table);
-    }
-  else if (ffi_trampoline_tables != table)
-    {
-      /* Otherwise, bump this table to the top of the list */
-      table->prev = NULL;
-      table->next = ffi_trampoline_tables;
-      if (ffi_trampoline_tables != NULL)
-	ffi_trampoline_tables->prev = table;
-
-      ffi_trampoline_tables = table;
-    }
-
-  pthread_mutex_unlock (&ffi_trampoline_lock);
-
-  /* Free the closure */
-  free (closure);
-}
-
-#else
-
-extern unsigned int ffi_arm_trampoline[2] FFI_HIDDEN;
-
-#endif
-
 /* the cif must already be prep'ed */
 
 ffi_status
@@ -796,7 +561,7 @@ ffi_prep_closure_loc (ffi_closure * closure,
     return FFI_BAD_ABI;
 
 #if FFI_EXEC_TRAMPOLINE_TABLE
-  void **config = FFI_TRAMPOLINE_CODELOC_CONFIG (codeloc);
+  void **config = (void **)((uint8_t *)codeloc - PAGE_MAX_SIZE);
   config[0] = closure;
   config[1] = closure_func;
 #else
diff --git a/src/arm/ffitarget.h b/src/arm/ffitarget.h
index 4f473f9..1cf1036 100644
--- a/src/arm/ffitarget.h
+++ b/src/arm/ffitarget.h
@@ -63,7 +63,20 @@ typedef enum ffi_abi {
 
 #define FFI_CLOSURES 1
 #define FFI_GO_CLOSURES 1
-#define FFI_TRAMPOLINE_SIZE 12
 #define FFI_NATIVE_RAW_API 0
 
+#if defined (FFI_EXEC_TRAMPOLINE_TABLE) && FFI_EXEC_TRAMPOLINE_TABLE
+
+#ifdef __MACH__
+#define FFI_TRAMPOLINE_SIZE 12
+#define FFI_TRAMPOLINE_CLOSURE_OFFSET 8
+#else
+#error "No trampoline table implementation"
+#endif
+
+#else
+#define FFI_TRAMPOLINE_SIZE 12
+#define FFI_TRAMPOLINE_CLOSURE_OFFSET FFI_TRAMPOLINE_SIZE
+#endif
+
 #endif
diff --git a/src/arm/sysv.S b/src/arm/sysv.S
index fd16589..379dcdf 100644
--- a/src/arm/sysv.S
+++ b/src/arm/sysv.S
@@ -25,7 +25,7 @@
    DEALINGS IN THE SOFTWARE.
    ----------------------------------------------------------------------- */
 
-#define LIBFFI_ASM	
+#define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
 #include <ffi_cfi.h>
@@ -52,11 +52,12 @@
 #endif
 
 /* Conditionally compile unwinder directives.  */
-.macro UNWIND text:vararg
 #ifdef __ARM_EABI__
-	\text
-#endif	
-.endm
+# define UNWIND(...)	__VA_ARGS__
+#else
+# define UNWIND(...)
+#endif
+
 #if defined(HAVE_AS_CFI_PSEUDO_OP) && defined(__ARM_EABI__)
 	.cfi_sections	.debug_frame
 #endif
@@ -77,29 +78,39 @@
 # define TYPE(X, Y)
 #endif
 
-#define ARM_FUNC_START(name, gl) \
-	.align	3; \
-	.ifne gl; .globl CNAME(name); FFI_HIDDEN(CNAME(name)); .endif; \
-	TYPE(name, %function); \
+#define ARM_FUNC_START_LOCAL(name)	\
+	.align	3;			\
+	TYPE(CNAME(name), %function);	\
 	CNAME(name):
 
+#define ARM_FUNC_START(name)		\
+	.globl CNAME(name);		\
+	FFI_HIDDEN(CNAME(name));	\
+	ARM_FUNC_START_LOCAL(name)
+
 #define ARM_FUNC_END(name) \
 	SIZE(name)
 
 /* Aid in defining a jump table with 8 bytes between entries.  */
-.macro E index
-	.if . - 0b - 8*\index
-	.error "type table out of sync"
+/* ??? The clang assembler doesn't handle .if with symbolic expressions.  */
+#ifdef __clang__
+# define E(index)
+#else
+# define E(index)				\
+	.if . - 0b - 8*index;			\
+	.error "type table out of sync";	\
 	.endif
-.endm
+#endif
 
 	.text
 	.syntax unified
 	.arm
 
+#ifndef __clang__
 	/* We require interworking on LDM, which implies ARMv5T,
 	   which implies the existance of BLX.  */
-	.arch	armv5t
+ 	.arch	armv5t
+#endif
 
 	/* Note that we use STC and LDC to encode VFP instructions,
 	   so that we do not need ".fpu vfp", nor get that added to
@@ -111,25 +122,30 @@
 	@ r2:   fn
 	@ r3:	vfp_used
 
-ARM_FUNC_START(ffi_call_VFP, 1)
-	UNWIND	.fnstart
+ARM_FUNC_START(ffi_call_VFP)
+	UNWIND(.fnstart)
 	cfi_startproc
 
 	cmp	r3, #3			@ load only d0 if possible
+#ifdef __clang__
+	vldrle d0, [sp]
+	vldmgt sp, {d0-d7}
+#else
 	ldcle	p11, cr0, [r0]		@ vldrle d0, [sp]
 	ldcgt	p11, cr0, [r0], {16}	@ vldmgt sp, {d0-d7}
+#endif
 	add	r0, r0, #64		@ discard the vfp register args
 	/* FALLTHRU */
 ARM_FUNC_END(ffi_call_VFP)
 
-ARM_FUNC_START(ffi_call_SYSV, 1)
+ARM_FUNC_START(ffi_call_SYSV)
 	stm	r1, {fp, lr}
 	mov	fp, r1
 
 	@ This is a bit of a lie wrt the origin of the unwind info, but
 	@ now we've got the usual frame pointer and two saved registers.
-	UNWIND	.save {fp,lr}
-	UNWIND	.setfp fp, sp
+	UNWIND(.save {fp,lr})
+	UNWIND(.setfp fp, sp)
 	cfi_def_cfa(fp, 8)
 	cfi_rel_offset(fp, 0)
 	cfi_rel_offset(lr, 4)
@@ -154,29 +170,41 @@ ARM_FUNC_START(ffi_call_SYSV, 1)
 	add	pc, pc, r3, lsl #3
 	nop
 0:
-E ARM_TYPE_VFP_S
+E(ARM_TYPE_VFP_S)
+#ifdef __clang__
+	vstr s0, [r2]
+#else
 	stc	p10, cr0, [r2]		@ vstr s0, [r2]
+#endif
 	pop	{fp,pc}
-E ARM_TYPE_VFP_D
+E(ARM_TYPE_VFP_D)
+#ifdef __clang__
+	vstr d0, [r2]
+#else
 	stc	p11, cr0, [r2]		@ vstr d0, [r2]
+#endif
 	pop	{fp,pc}
-E ARM_TYPE_VFP_N
+E(ARM_TYPE_VFP_N)
+#ifdef __clang__
+	vstm r2, {d0-d3}
+#else
 	stc	p11, cr0, [r2], {8}	@ vstm r2, {d0-d3}
+#endif
 	pop	{fp,pc}
-E ARM_TYPE_INT64
+E(ARM_TYPE_INT64)
 	str	r1, [r2, #4]
 	nop
-E ARM_TYPE_INT
+E(ARM_TYPE_INT)
 	str	r0, [r2]
 	pop	{fp,pc}
-E ARM_TYPE_VOID
+E(ARM_TYPE_VOID)
 	pop	{fp,pc}
 	nop
-E ARM_TYPE_STRUCT
+E(ARM_TYPE_STRUCT)
 	pop	{fp,pc}
 
 	cfi_endproc
-	UNWIND	.fnend
+	UNWIND(.fnend)
 ARM_FUNC_END(ffi_call_SYSV)
 
 
@@ -184,7 +212,7 @@ ARM_FUNC_END(ffi_call_SYSV)
 	int ffi_closure_inner_* (cif, fun, user_data, frame)
 */
 
-ARM_FUNC_START(ffi_go_closure_SYSV, 1)
+ARM_FUNC_START(ffi_go_closure_SYSV)
 	cfi_startproc
 	stmdb	sp!, {r0-r3}			@ save argument regs
 	cfi_adjust_cfa_offset(16)
@@ -195,14 +223,18 @@ ARM_FUNC_START(ffi_go_closure_SYSV, 1)
 	cfi_endproc
 ARM_FUNC_END(ffi_go_closure_SYSV)
 
-ARM_FUNC_START(ffi_closure_SYSV, 1)
-	UNWIND	.fnstart
+ARM_FUNC_START(ffi_closure_SYSV)
+	UNWIND(.fnstart)
 	cfi_startproc
 	stmdb	sp!, {r0-r3}			@ save argument regs
 	cfi_adjust_cfa_offset(16)
-	ldr	r0, [ip, #FFI_TRAMPOLINE_SIZE]	  @ load cif
-	ldr	r1, [ip, #FFI_TRAMPOLINE_SIZE+4]  @ load fun
-	ldr	r2, [ip, #FFI_TRAMPOLINE_SIZE+8]  @ load user_data
+
+#if FFI_EXEC_TRAMPOLINE_TABLE
+	ldr ip, [ip]				@ ip points to the config page, dereference to get the ffi_closure*
+#endif
+	ldr	r0, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET]	@ load cif
+	ldr	r1, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET+4]  @ load fun
+	ldr	r2, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET+8]  @ load user_data
 0:
 	add	ip, sp, #16			@ compute entry sp
 	sub	sp, sp, #64+32			@ allocate frame
@@ -212,7 +244,7 @@ ARM_FUNC_START(ffi_closure_SYSV, 1)
 	/* Remember that EABI unwind info only applies at call sites.
 	   We need do nothing except note the save of the stack pointer
 	   and the link registers.  */
-	UNWIND	.save {sp,lr}
+	UNWIND(.save {sp,lr})
 	cfi_adjust_cfa_offset(8)
 	cfi_rel_offset(lr, 4)
 
@@ -224,10 +256,10 @@ ARM_FUNC_START(ffi_closure_SYSV, 1)
 	adr	r3, CNAME(ffi_closure_ret)
 	add	pc, r3, r0, lsl #3
 	cfi_endproc
-	UNWIND	.fnend
+	UNWIND(.fnend)
 ARM_FUNC_END(ffi_closure_SYSV)
 
-ARM_FUNC_START(ffi_go_closure_VFP, 1)
+ARM_FUNC_START(ffi_go_closure_VFP)
 	cfi_startproc
 	stmdb	sp!, {r0-r3}			@ save argument regs
 	cfi_adjust_cfa_offset(16)
@@ -238,23 +270,31 @@ ARM_FUNC_START(ffi_go_closure_VFP, 1)
 	cfi_endproc
 ARM_FUNC_END(ffi_go_closure_VFP)
 
-ARM_FUNC_START(ffi_closure_VFP, 1)
-	UNWIND	.fnstart
+ARM_FUNC_START(ffi_closure_VFP)
+	UNWIND(.fnstart)
 	cfi_startproc
 	stmdb	sp!, {r0-r3}			@ save argument regs
 	cfi_adjust_cfa_offset(16)
-	ldr	r0, [ip, #FFI_TRAMPOLINE_SIZE]	  @ load cif
-	ldr	r1, [ip, #FFI_TRAMPOLINE_SIZE+4]  @ load fun
-	ldr	r2, [ip, #FFI_TRAMPOLINE_SIZE+8]  @ load user_data
+
+#if FFI_EXEC_TRAMPOLINE_TABLE
+	ldr ip, [ip]				@ ip points to the config page, dereference to get the ffi_closure*
+#endif
+	ldr	r0, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET]	@ load cif
+	ldr	r1, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET+4]  @ load fun
+	ldr	r2, [ip, #FFI_TRAMPOLINE_CLOSURE_OFFSET+8]  @ load user_data
 0:
 	add	ip, sp, #16
 	sub	sp, sp, #64+32			@ allocate frame
 	cfi_adjust_cfa_offset(64+32)
+#ifdef __clang__
+	vstm sp, {d0-d7}
+#else
 	stc	p11, cr0, [sp], {16}		@ vstm sp, {d0-d7}
+#endif
 	stmdb	sp!, {ip,lr}
 
 	/* See above.  */
-	UNWIND	.save {sp,lr}
+	UNWIND(.save {sp,lr})
 	cfi_adjust_cfa_offset(8)
 	cfi_rel_offset(lr, 4)
 
@@ -266,63 +306,71 @@ ARM_FUNC_START(ffi_closure_VFP, 1)
 	adr	r3, CNAME(ffi_closure_ret)
 	add	pc, r3, r0, lsl #3
 	cfi_endproc
-	UNWIND	.fnend
+	UNWIND(.fnend)
 ARM_FUNC_END(ffi_closure_VFP)
 
 /* Load values returned in registers for both closure entry points.
    Note that we use LDM with SP in the register set.  This is deprecated
    by ARM, but not yet unpredictable.  */
 
-ARM_FUNC_START(ffi_closure_ret, 0)
+ARM_FUNC_START_LOCAL(ffi_closure_ret)
 	cfi_startproc
 	cfi_rel_offset(sp, 0)
 	cfi_rel_offset(lr, 4)
 0:
-E ARM_TYPE_VFP_S
+E(ARM_TYPE_VFP_S)
+#ifdef __clang__
+	vldr s0, [r2]
+#else
 	ldc	p10, cr0, [r2]			@ vldr s0, [r2]
+#endif
 	ldm	sp, {sp,pc}
-E ARM_TYPE_VFP_D
+E(ARM_TYPE_VFP_D)
+#ifdef __clang__
+	vldr d0, [r2]
+#else
 	ldc	p11, cr0, [r2]			@ vldr d0, [r2]
+#endif
 	ldm	sp, {sp,pc}
-E ARM_TYPE_VFP_N
+E(ARM_TYPE_VFP_N)
+#ifdef __clang__
+	vldm r2, {d0-d3}
+#else
 	ldc	p11, cr0, [r2], {8}		@ vldm r2, {d0-d3}
+#endif
 	ldm	sp, {sp,pc}
-E ARM_TYPE_INT64
+E(ARM_TYPE_INT64)
 	ldr	r1, [r2, #4]
 	nop
-E ARM_TYPE_INT
+E(ARM_TYPE_INT)
 	ldr	r0, [r2]
 	ldm	sp, {sp,pc}
-E ARM_TYPE_VOID
+E(ARM_TYPE_VOID)
 	ldm	sp, {sp,pc}
 	nop
-E ARM_TYPE_STRUCT
+E(ARM_TYPE_STRUCT)
 	ldm	sp, {sp,pc}
 	cfi_endproc
 ARM_FUNC_END(ffi_closure_ret)
 
 #if FFI_EXEC_TRAMPOLINE_TABLE
 
-/* ??? The iOS support should be updated.  The first insn used to
-   be STMFD, but that's been moved into ffi_closure_SYSV.  If the
-   writable page is put after this one we can make use of the
-   pc+8 feature of the architecture.  We can also reduce the size
-   of the thunk to 8 and pack more of these into the page.
-
-   In the meantime, simply replace the STMFD with a NOP so as to
-   keep all the magic numbers the same within ffi.c.  */
+#ifdef __MACH__
+#include <mach/vm_param.h>
 
-	.align	12
+.align	PAGE_MAX_SHIFT
 ARM_FUNC_START(ffi_closure_trampoline_table_page)
-.rept	4096 / 12
-	nop
-	ldr	ip, [pc, #-4092]
-	ldr	pc, [pc, #-4092]
+.rept	PAGE_MAX_SIZE / FFI_TRAMPOLINE_SIZE
+	adr ip, #-PAGE_MAX_SIZE   @ the config page is PAGE_MAX_SIZE behind the trampoline page
+	sub ip, #8				  @ account for pc bias
+	ldr	pc, [ip, #4]		  @ jump to ffi_closure_SYSV or ffi_closure_VFP
 .endr
+ARM_FUNC_END(ffi_closure_trampoline_table_page)
+#endif
 
 #else
 
-ARM_FUNC_START(ffi_arm_trampoline, 1)
+ARM_FUNC_START(ffi_arm_trampoline)
 0:	adr	ip, 0b
 	ldr	pc, 1f
 1:	.long	0
diff --git a/src/closures.c b/src/closures.c
index 721ff00..3dec0e3 100644
--- a/src/closures.c
+++ b/src/closures.c
@@ -30,6 +30,7 @@
 #define _GNU_SOURCE 1
 #endif
 
+#include <fficonfig.h>
 #include <ffi.h>
 #include <ffi_common.h>
 
@@ -64,11 +65,247 @@
 
 #if FFI_CLOSURES
 
-# if FFI_EXEC_TRAMPOLINE_TABLE
+#if FFI_EXEC_TRAMPOLINE_TABLE
+
+#ifdef __MACH__
+
+#include <mach/mach.h>
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+extern void *ffi_closure_trampoline_table_page;
+
+typedef struct ffi_trampoline_table ffi_trampoline_table;
+typedef struct ffi_trampoline_table_entry ffi_trampoline_table_entry;
+
+struct ffi_trampoline_table
+{
+  /* contiguous writable and executable pages */
+  vm_address_t config_page;
+  vm_address_t trampoline_page;
+
+  /* free list tracking */
+  uint16_t free_count;
+  ffi_trampoline_table_entry *free_list;
+  ffi_trampoline_table_entry *free_list_pool;
+
+  ffi_trampoline_table *prev;
+  ffi_trampoline_table *next;
+};
+
+struct ffi_trampoline_table_entry
+{
+  void *(*trampoline) ();
+  ffi_trampoline_table_entry *next;
+};
+
+/* Total number of trampolines that fit in one trampoline table */
+#define FFI_TRAMPOLINE_COUNT (PAGE_MAX_SIZE / FFI_TRAMPOLINE_SIZE)
+
+static pthread_mutex_t ffi_trampoline_lock = PTHREAD_MUTEX_INITIALIZER;
+static ffi_trampoline_table *ffi_trampoline_tables = NULL;
+
+static ffi_trampoline_table *
+ffi_trampoline_table_alloc ()
+{
+  ffi_trampoline_table *table = NULL;
+
+  /* Loop until we can allocate two contiguous pages */
+  while (table == NULL)
+    {
+      vm_address_t config_page = 0x0;
+      kern_return_t kt;
+
+      /* Try to allocate two pages */
+      kt =
+	vm_allocate (mach_task_self (), &config_page, PAGE_MAX_SIZE * 2,
+		     VM_FLAGS_ANYWHERE);
+      if (kt != KERN_SUCCESS)
+	{
+	  fprintf (stderr, "vm_allocate() failure: %d at %s:%d\n", kt,
+		   __FILE__, __LINE__);
+	  break;
+	}
+
+      /* Now drop the second half of the allocation to make room for the trampoline table */
+      vm_address_t trampoline_page = config_page + PAGE_MAX_SIZE;
+      kt = vm_deallocate (mach_task_self (), trampoline_page, PAGE_MAX_SIZE);
+      if (kt != KERN_SUCCESS)
+	{
+	  fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
+		   __FILE__, __LINE__);
+	  break;
+	}
+
+      /* Remap the trampoline table to directly follow the config page */
+      vm_prot_t cur_prot;
+      vm_prot_t max_prot;
+
+	  vm_address_t trampoline_page_template = (vm_address_t)&ffi_closure_trampoline_table_page;
+#ifdef __arm__
+	  /* ffi_closure_trampoline_table_page can be thumb-biased on some ARM archs */
+	  trampoline_page_template &= ~1UL;
+#endif
+
+      kt =
+	vm_remap (mach_task_self (), &trampoline_page, PAGE_MAX_SIZE, 0x0, FALSE,
+		  mach_task_self (), trampoline_page_template, FALSE,
+		  &cur_prot, &max_prot, VM_INHERIT_SHARE);
+
+      /* If we lost access to the destination trampoline page, drop our config allocation mapping and retry */
+      if (kt != KERN_SUCCESS)
+	{
+	  /* Log unexpected failures */
+	  if (kt != KERN_NO_SPACE)
+	    {
+	      fprintf (stderr, "vm_remap() failure: %d at %s:%d\n", kt,
+		       __FILE__, __LINE__);
+	    }
+
+	  vm_deallocate (mach_task_self (), config_page, PAGE_SIZE);
+	  continue;
+	}
+
+      /* We have valid trampoline and config pages */
+      table = calloc (1, sizeof (ffi_trampoline_table));
+      table->free_count = FFI_TRAMPOLINE_COUNT;
+      table->config_page = config_page;
+      table->trampoline_page = trampoline_page;
+
+      /* Create and initialize the free list */
+      table->free_list_pool =
+	calloc (FFI_TRAMPOLINE_COUNT, sizeof (ffi_trampoline_table_entry));
+
+      uint16_t i;
+      for (i = 0; i < table->free_count; i++)
+	{
+	  ffi_trampoline_table_entry *entry = &table->free_list_pool[i];
+	  entry->trampoline =
+	    (void *) (table->trampoline_page + (i * FFI_TRAMPOLINE_SIZE));
+
+	  if (i < table->free_count - 1)
+	    entry->next = &table->free_list_pool[i + 1];
+	}
+
+      table->free_list = table->free_list_pool;
+    }
+
+  return table;
+}
+
+void *
+ffi_closure_alloc (size_t size, void **code)
+{
+  /* Create the closure */
+  ffi_closure *closure = malloc (size);
+  if (closure == NULL)
+    return NULL;
+
+  pthread_mutex_lock (&ffi_trampoline_lock);
+
+  /* Check for an active trampoline table with available entries. */
+  ffi_trampoline_table *table = ffi_trampoline_tables;
+  if (table == NULL || table->free_list == NULL)
+    {
+      table = ffi_trampoline_table_alloc ();
+      if (table == NULL)
+	{
+	  free (closure);
+	  return NULL;
+	}
+
+      /* Insert the new table at the top of the list */
+      table->next = ffi_trampoline_tables;
+      if (table->next != NULL)
+	table->next->prev = table;
+
+      ffi_trampoline_tables = table;
+    }
+
+  /* Claim the free entry */
+  ffi_trampoline_table_entry *entry = ffi_trampoline_tables->free_list;
+  ffi_trampoline_tables->free_list = entry->next;
+  ffi_trampoline_tables->free_count--;
+  entry->next = NULL;
+
+  pthread_mutex_unlock (&ffi_trampoline_lock);
+
+  /* Initialize the return values */
+  *code = entry->trampoline;
+  closure->trampoline_table = table;
+  closure->trampoline_table_entry = entry;
+
+  return closure;
+}
+
+void
+ffi_closure_free (void *ptr)
+{
+  ffi_closure *closure = ptr;
+
+  pthread_mutex_lock (&ffi_trampoline_lock);
+
+  /* Fetch the table and entry references */
+  ffi_trampoline_table *table = closure->trampoline_table;
+  ffi_trampoline_table_entry *entry = closure->trampoline_table_entry;
+
+  /* Return the entry to the free list */
+  entry->next = table->free_list;
+  table->free_list = entry;
+  table->free_count++;
+
+  /* If all trampolines within this table are free, and at least one other table exists, deallocate
+   * the table */
+  if (table->free_count == FFI_TRAMPOLINE_COUNT
+      && ffi_trampoline_tables != table)
+    {
+      /* Remove from the list */
+      if (table->prev != NULL)
+	table->prev->next = table->next;
+
+      if (table->next != NULL)
+	table->next->prev = table->prev;
+
+      /* Deallocate pages */
+      kern_return_t kt;
+      kt = vm_deallocate (mach_task_self (), table->config_page, PAGE_SIZE);
+      if (kt != KERN_SUCCESS)
+	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
+		 __FILE__, __LINE__);
+
+      kt =
+	vm_deallocate (mach_task_self (), table->trampoline_page, PAGE_SIZE);
+      if (kt != KERN_SUCCESS)
+	fprintf (stderr, "vm_deallocate() failure: %d at %s:%d\n", kt,
+		 __FILE__, __LINE__);
+
+      /* Deallocate free list */
+      free (table->free_list_pool);
+      free (table);
+    }
+  else if (ffi_trampoline_tables != table)
+    {
+      /* Otherwise, bump this table to the top of the list */
+      table->prev = NULL;
+      table->next = ffi_trampoline_tables;
+      if (ffi_trampoline_tables != NULL)
+	ffi_trampoline_tables->prev = table;
+
+      ffi_trampoline_tables = table;
+    }
+
+  pthread_mutex_unlock (&ffi_trampoline_lock);
+
+  /* Free the closure */
+  free (closure);
+}
+
+#endif
 
 // Per-target implementation; It's unclear what can reasonable be shared between two OS/architecture implementations.
 
-# elif FFI_MMAP_EXEC_WRIT /* !FFI_EXEC_TRAMPOLINE_TABLE */
+#elif FFI_MMAP_EXEC_WRIT /* !FFI_EXEC_TRAMPOLINE_TABLE */
 
 #define USE_LOCKS 1
 #define USE_DL_PREFIX 1