the code is through with it, it has to be freed in order to release Haskell
and C resources. Failure to do so result in memory leaks on both the C and
Haskell side.
-
*/
+
#include "PosixSource.h"
#include "Rts.h"
+#include "RtsExternal.h"
#include "RtsUtils.h"
-#include "RtsFlags.h"
#include <stdlib.h>
+#if defined(_WIN32)
+#include <windows.h>
+#endif
+
+#if defined(openbsd_TARGET_OS)
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/mman.h>
+
+/* no C99 header stdint.h on OpenBSD? */
+typedef unsigned long my_uintptr_t;
+#endif
+
/* Heavily arch-specific, I'm afraid.. */
+/*
+ * Allocate len bytes which are readable, writable, and executable.
+ *
+ * ToDo: If this turns out to be a performance bottleneck, one could
+ * e.g. cache the last VirtualProtect/mprotect-ed region and do
+ * nothing in case of a cache hit.
+ */
+static void*
+mallocBytesRWX(int len)
+{
+ void *addr = stgMallocBytes(len, "mallocBytesRWX");
+#if defined(i386_TARGET_ARCH) && defined(_WIN32)
+ /* This could be necessary for processors which distinguish between READ and
+ EXECUTE memory accesses, e.g. Itaniums. */
+ DWORD dwOldProtect = 0;
+ if (VirtualProtect (addr, len, PAGE_EXECUTE_READWRITE, &dwOldProtect) == 0) {
+ barf("mallocBytesRWX: failed to protect 0x%p; error=%lu; old protection: %lu\n",
+ addr, (unsigned long)GetLastError(), (unsigned long)dwOldProtect);
+ }
+#elif defined(openbsd_TARGET_OS)
+ /* malloced memory isn't executable by default on OpenBSD */
+ my_uintptr_t pageSize = sysconf(_SC_PAGESIZE);
+ my_uintptr_t mask = ~(pageSize - 1);
+ my_uintptr_t startOfFirstPage = ((my_uintptr_t)addr ) & mask;
+ my_uintptr_t startOfLastPage = ((my_uintptr_t)addr + len - 1) & mask;
+ my_uintptr_t size = startOfLastPage - startOfFirstPage + pageSize;
+ if (mprotect((void*)startOfFirstPage, (size_t)size, PROT_EXEC | PROT_READ | PROT_WRITE) != 0) {
+ barf("mallocBytesRWX: failed to protect 0x%p\n", addr);
+ }
+#endif
+ return addr;
+}
+
#if defined(i386_TARGET_ARCH)
-/* Now here's something obscure for you:
-
- When generating an adjustor thunk that uses the C calling
- convention, we have to make sure that the thunk kicks off
- the process of jumping into Haskell with a tail jump. Why?
- Because as a result of jumping in into Haskell we may end
- up freeing the very adjustor thunk we came from using
- freeHaskellFunctionPtr(). Hence, we better not return to
- the adjustor code on our way out, since it could by then
- point to junk.
-
- The fix is readily at hand, just include the opcodes
- for the C stack fixup code that we need to perform when
- returning in some static piece of memory and arrange
- to return to it before tail jumping from the adjustor thunk.
-
- For this to work we make the assumption that bytes in .data
- are considered executable.
-*/
-static unsigned char __obscure_ccall_ret_code [] =
- { 0x83, 0xc4, 0x04 /* addl $0x4, %esp */
- , 0xc3 /* ret */
- };
+static unsigned char *obscure_ccall_ret_code;
#endif
#if defined(alpha_TARGET_ARCH)
/* To get the definition of PAL_imb: */
-#include <machine/pal.h>
+# if defined(linux_TARGET_OS)
+# include <asm/pal.h>
+# else
+# include <machine/pal.h>
+# endif
+#endif
+
+#if defined(ia64_TARGET_ARCH)
+#include "Storage.h"
+
+/* Layout of a function descriptor */
+typedef struct _IA64FunDesc {
+ StgWord64 ip;
+ StgWord64 gp;
+} IA64FunDesc;
+
+static void *
+stgAllocStable(size_t size_in_bytes, StgStablePtr *stable)
+{
+ StgArrWords* arr;
+ nat data_size_in_words, total_size_in_words;
+
+ /* round up to a whole number of words */
+ data_size_in_words = (size_in_bytes + sizeof(W_) + 1) / sizeof(W_);
+ total_size_in_words = sizeofW(StgArrWords) + data_size_in_words;
+
+ /* allocate and fill it in */
+ arr = (StgArrWords *)allocate(total_size_in_words);
+ SET_ARR_HDR(arr, &stg_ARR_WORDS_info, CCCS, data_size_in_words);
+
+ /* obtain a stable ptr */
+ *stable = getStablePtr((StgPtr)arr);
+
+ /* and return a ptr to the goods inside the array */
+ return(BYTE_ARR_CTS(arr));
+}
+#endif
+
+#if defined(powerpc64_TARGET_ARCH)
+// We don't need to generate dynamic code on powerpc64-[linux|AIX],
+// but we do need a piece of (static) inline assembly code:
+
+static void
+adjustorCodeWrittenInAsm()
+{
+ __asm__ volatile (
+ "adjustorCode:\n\t"
+ "mr 10,8\n\t"
+ "mr 9,7\n\t"
+ "mr 8,6\n\t"
+ "mr 7,5\n\t"
+ "mr 6,4\n\t"
+ "mr 5,3\n\t"
+ "mr 3,11\n\t"
+ "ld 0,0(2)\n\t"
+ "ld 11,16(2)\n\t"
+ "mtctr 0\n\t"
+ "ld 2,8(2)\n\t"
+ "bctr"
+ : : );
+}
#endif
void*
<c>: ff e0 jmp %eax # and jump to it.
# the callee cleans up the stack
*/
- if ((adjustor = stgMallocBytes(14, "createAdjustor")) != NULL) {
+ adjustor = mallocBytesRWX(14);
+ {
unsigned char *const adj_code = (unsigned char *)adjustor;
adj_code[0x00] = (unsigned char)0x58; /* popl %eax */
<00>: 68 ef be ad de pushl $0xdeadbeef # constant is large enough to
# hold a StgStablePtr
<05>: b8 fa ef ff 00 movl $0x00ffeffa, %eax # load up wptr
- <0a>: 68 ef be ad de pushl $__obscure_ccall_ret_code # push the return address
+ <0a>: 68 ef be ad de pushl $obscure_ccall_ret_code # push the return address
<0f>: ff e0 jmp *%eax # jump to wptr
The ccall'ing version is a tad different, passing in the return
via the stable pointer.) (The auto-generated C stub is in on this
game, don't worry :-)
- See the comment next to __obscure_ccall_ret_code why we need to
+ See the comment next to obscure_ccall_ret_code why we need to
perform a tail jump instead of a call, followed by some C stack
fixup.
That's (thankfully) the case here with the restricted set of
return types that we support.
*/
- if ((adjustor = stgMallocBytes(17, "createAdjustor")) != NULL) {
+ adjustor = mallocBytesRWX(17);
+ {
unsigned char *const adj_code = (unsigned char *)adjustor;
adj_code[0x00] = (unsigned char)0x68; /* pushl hptr (which is a dword immediate ) */
adj_code[0x05] = (unsigned char)0xb8; /* movl $wptr, %eax */
*((StgFunPtr*)(adj_code + 0x06)) = (StgFunPtr)wptr;
- adj_code[0x0a] = (unsigned char)0x68; /* pushl __obscure_ccall_ret_code */
- *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)__obscure_ccall_ret_code;
+ adj_code[0x0a] = (unsigned char)0x68; /* pushl obscure_ccall_ret_code */
+ *((StgFunPtr*)(adj_code + 0x0b)) = (StgFunPtr)obscure_ccall_ret_code;
adj_code[0x0f] = (unsigned char)0xff; /* jmp *%eax */
adj_code[0x10] = (unsigned char)0xe0;
}
#elif defined(sparc_TARGET_ARCH)
- /* Magic constant computed by inspecting the code length of
- the following assembly language snippet
- (offset and machine code prefixed):
-
- <00>: BA 10 00 1B mov %i3, %i5
- <04>: B8 10 00 1A mov %i2, %i4
- <08>: B6 10 00 19 mov %i1, %i3
- <0c>: B4 10 00 18 mov %i0, %i2
- <10>: 13 00 3f fb sethi %hi(0x00ffeffa), %o1 # load up wptr (1 of 2)
- <14>: 11 37 ab 6f sethi %hi(0xdeadbeef), %o0 # load up hptr (1 of 2)
- <18>: 81 c2 63 fa jmp %o1+%lo(0x00ffeffa) # jump to wptr (load 2 of 2)
- <1c>: 90 12 22 ef or %o0, %lo(0xdeadbeef), %o0 # load up hptr (2 of 2)
- # [in delay slot]
- <20>: de ad be ef # Place the value of the StgStablePtr somewhere readable
-
- ccall'ing on a SPARC leaves little to be performed by the caller.
- The callee shifts the window on entry and restores it on exit.
- Input paramters and results are passed via registers. (%o0 in the
- code above contains the input paramter to wptr.) The return address
- is stored in %o7/%i7. Since we don't shift the window in this code,
- the return address is preserved and wptr will return to our caller.
-
- JRS, 21 Aug 01: the above para is a fiction. The caller passes
- args in %i0 .. %i5 and then the rest at [%sp+92]. We want to
- tailjump to wptr, passing hptr as the new first arg, and a dummy
- second arg, which would be where the return address is on x86.
- That means we have to shuffle the original caller's args along by
- two.
-
- We do a half-correct solution which works only if the original
- caller passed 4 or fewer arg words. Move %i0 .. %i3 into %i3
- .. %i6, so we can park hptr in %i0 and a bogus arg in %i1. The
- fully correct solution would be to subtract 8 from %sp and then
- place %i4 and %i5 at [%sp+92] and [%sp+96] respectively. This
- machinery should then work in all cases. (Or would it? Perhaps
- it would trash parts of the caller's frame. Dunno).
-
- SUP, 25 Apr 02: We are quite lucky to push a multiple of 8 bytes in
- front of the existing arguments, because %sp must stay double-word
- aligned at all times, see: http://www.sparc.org/standards/psABI3rd.pdf
- Although we extend the *caller's* stack frame, this shouldn't cause
- any problems for a C-like caller: alloca is implemented similarly, and
- local variables should be accessed via %fp, not %sp. In a nutshell:
- This should work. (Famous last words! :-)
+ /* Magic constant computed by inspecting the code length of the following
+ assembly language snippet (offset and machine code prefixed):
+
+ <00>: 9C23A008 sub %sp, 8, %sp ! make room for %o4/%o5 in caller's frame
+ <04>: DA23A060 st %o5, [%sp + 96] ! shift registers by 2 positions
+ <08>: D823A05C st %o4, [%sp + 92]
+ <0C>: 9A10000B mov %o3, %o5
+ <10>: 9810000A mov %o2, %o4
+ <14>: 96100009 mov %o1, %o3
+ <18>: 94100008 mov %o0, %o2
+ <1C>: 13000000 sethi %hi(wptr), %o1 ! load up wptr (1 of 2)
+ <20>: 11000000 sethi %hi(hptr), %o0 ! load up hptr (1 of 2)
+ <24>: 81C26000 jmp %o1 + %lo(wptr) ! jump to wptr (load 2 of 2)
+ <28>: 90122000 or %o0, %lo(hptr), %o0 ! load up hptr (2 of 2, delay slot)
+ <2C> 00000000 ! place for getting hptr back easily
+
+ ccall'ing on SPARC is easy, because we are quite lucky to push a
+ multiple of 8 bytes (1 word hptr + 1 word dummy arg) in front of the
+ existing arguments (note that %sp must stay double-word aligned at
+ all times, see ABI spec at http://www.sparc.org/standards/psABI3rd.pdf).
+ To do this, we extend the *caller's* stack frame by 2 words and shift
+ the output registers used for argument passing (%o0 - %o5, we are a *leaf*
+ procedure because of the tail-jump) by 2 positions. This makes room in
+ %o0 and %o1 for the additinal arguments, namely hptr and a dummy (used
+ for destination addr of jump on SPARC, return address on x86, ...). This
+ shouldn't cause any problems for a C-like caller: alloca is implemented
+ similarly, and local variables should be accessed via %fp, not %sp. In a
+ nutshell: This should work! (Famous last words! :-)
*/
- if ((adjustor = stgMallocBytes(4*(8+1), "createAdjustor")) != NULL) {
- unsigned long *const adj_code = (unsigned long *)adjustor;
-
- /* mov %o3, %o5 */
- adj_code[0] = (unsigned long)0x9A10000B;
- /* mov %o2, %o4 */
- adj_code[1] = (unsigned long)0x9810000A;
- /* mov %o1, %o3 */
- adj_code[2] = (unsigned long)0x96100009;
- /* mov %o0, %o2 */
- adj_code[3] = (unsigned long)0x94100008;
-
- /* sethi %hi(wptr), %o1 */
- adj_code[4] = (unsigned long)0x13000000;
- adj_code[4] |= ((unsigned long)wptr) >> 10;
-
- /* sethi %hi(hptr), %o0 */
- adj_code[5] = (unsigned long)0x11000000;
- adj_code[5] |= ((unsigned long)hptr) >> 10;
-
- /* jmp %o1+%lo(wptr) */
- adj_code[6] = (unsigned long)0x81c26000;
- adj_code[6] |= ((unsigned long)wptr) & 0x000003ff;
-
- /* or %o0, %lo(hptr), %o0 */
- adj_code[7] = (unsigned long)0x90122000;
- adj_code[7] |= ((unsigned long)hptr) & 0x000003ff;
-
- adj_code[8] = (StgStablePtr)hptr;
-
- /* flush cache */
- asm("flush %0" : : "r" (adj_code ));
- asm("flush %0" : : "r" (adj_code + 2));
- asm("flush %0" : : "r" (adj_code + 4));
- asm("flush %0" : : "r" (adj_code + 6));
-
- /* max. 5 instructions latency, and we need at >= 1 for returning */
- asm("nop");
- asm("nop");
- asm("nop");
- asm("nop");
+ adjustor = mallocBytesRWX(4*(11+1));
+ {
+ unsigned long *const adj_code = (unsigned long *)adjustor;
+
+ adj_code[ 0] = 0x9C23A008UL; /* sub %sp, 8, %sp */
+ adj_code[ 1] = 0xDA23A060UL; /* st %o5, [%sp + 96] */
+ adj_code[ 2] = 0xD823A05CUL; /* st %o4, [%sp + 92] */
+ adj_code[ 3] = 0x9A10000BUL; /* mov %o3, %o5 */
+ adj_code[ 4] = 0x9810000AUL; /* mov %o2, %o4 */
+ adj_code[ 5] = 0x96100009UL; /* mov %o1, %o3 */
+ adj_code[ 6] = 0x94100008UL; /* mov %o0, %o2 */
+ adj_code[ 7] = 0x13000000UL; /* sethi %hi(wptr), %o1 */
+ adj_code[ 7] |= ((unsigned long)wptr) >> 10;
+ adj_code[ 8] = 0x11000000UL; /* sethi %hi(hptr), %o0 */
+ adj_code[ 8] |= ((unsigned long)hptr) >> 10;
+ adj_code[ 9] = 0x81C26000UL; /* jmp %o1 + %lo(wptr) */
+ adj_code[ 9] |= ((unsigned long)wptr) & 0x000003FFUL;
+ adj_code[10] = 0x90122000UL; /* or %o0, %lo(hptr), %o0 */
+ adj_code[10] |= ((unsigned long)hptr) & 0x000003FFUL;
+
+ adj_code[11] = (unsigned long)hptr;
+
+ /* flush cache */
+ asm("flush %0" : : "r" (adj_code ));
+ asm("flush %0" : : "r" (adj_code + 2));
+ asm("flush %0" : : "r" (adj_code + 4));
+ asm("flush %0" : : "r" (adj_code + 6));
+ asm("flush %0" : : "r" (adj_code + 10));
+
+ /* max. 5 instructions latency, and we need at >= 1 for returning */
+ asm("nop");
+ asm("nop");
+ asm("nop");
+ asm("nop");
}
#elif defined(alpha_TARGET_ARCH)
/* Magic constant computed by inspecting the code length of
4 bytes (getting rid of the nop), hence saving memory. [ccshan]
*/
ASSERT(((StgWord64)wptr & 3) == 0);
- if ((adjustor = stgMallocBytes(48, "createAdjustor")) != NULL) {
+ adjustor = mallocBytesRWX(48);
+ {
StgWord64 *const code = (StgWord64 *)adjustor;
code[0] = 0x4610041246520414L;
this code, it only works for up to 6 arguments (when floating point arguments
are involved, this may be more or less, depending on the exact situation).
*/
- if ((adjustor = stgMallocBytes(4*13, "createAdjustor")) != NULL) {
+ adjustor = mallocBytesRWX(4*13);
+ {
unsigned long *const adj_code = (unsigned long *)adjustor;
// make room for extra arguments
while(n--)
{
__asm__ volatile ("dcbf 0,%0\n\tsync\n\ticbi 0,%0"
- : : "g" (p));
+ : : "r" (p));
p++;
}
__asm__ volatile ("sync\n\tisync");
}
}
+#elif defined(powerpc64_TARGET_ARCH)
+ // This is for powerpc64 linux and powerpc64 AIX.
+ // It probably won't apply to powerpc64-darwin.
+
+ {
+ typedef struct {
+ StgFunPtr code;
+ void* toc;
+ void* env;
+ } FunDesc;
+
+ FunDesc *desc = malloc(sizeof(FunDesc));
+ extern void *adjustorCode;
+
+ desc->code = (void*) &adjustorCode;
+ desc->toc = (void*) wptr;
+ desc->env = (void*) hptr;
+
+ adjustor = (void*) desc;
+ }
+ break;
+
+#elif defined(ia64_TARGET_ARCH)
+/*
+ Up to 8 inputs are passed in registers. We flush the last two inputs to
+ the stack, initially into the 16-byte scratch region left by the caller.
+ We then shuffle the others along by 4 (taking 2 registers for ourselves
+ to save return address and previous function state - we need to come back
+ here on the way out to restore the stack, so this is a real function
+ rather than just a trampoline).
+
+ The function descriptor we create contains the gp of the target function
+ so gp is already loaded correctly.
+
+ [MLX] alloc r16=ar.pfs,10,2,0
+ movl r17=wptr
+ [MII] st8.spill [r12]=r38,8 // spill in6 (out4)
+ mov r41=r37 // out7 = in5 (out3)
+ mov r40=r36;; // out6 = in4 (out2)
+ [MII] st8.spill [r12]=r39 // spill in7 (out5)
+ mov.sptk b6=r17,50
+ mov r38=r34;; // out4 = in2 (out0)
+ [MII] mov r39=r35 // out5 = in3 (out1)
+ mov r37=r33 // out3 = in1 (loc1)
+ mov r36=r32 // out2 = in0 (loc0)
+ [MLX] adds r12=-24,r12 // update sp
+ movl r34=hptr;; // out0 = hptr
+ [MIB] mov r33=r16 // loc1 = ar.pfs
+ mov r32=b0 // loc0 = retaddr
+ br.call.sptk.many b0=b6;;
+
+ [MII] adds r12=-16,r12
+ mov b0=r32
+ mov.i ar.pfs=r33
+ [MFB] nop.m 0x0
+ nop.f 0x0
+ br.ret.sptk.many b0;;
+*/
+
+/* These macros distribute a long constant into the two words of an MLX bundle */
+#define BITS(val,start,count) (((val) >> (start)) & ((1 << (count))-1))
+#define MOVL_LOWORD(val) (BITS(val,22,18) << 46)
+#define MOVL_HIWORD(val) (BITS(val,40,23) | (BITS(val,0,7) << 36) | (BITS(val,7,9) << 50) \
+ | (BITS(val,16,5) << 55) | (BITS(val,21,1) << 44) | BITS(val,63,1) << 59)
+
+ {
+ StgStablePtr stable;
+ IA64FunDesc *wdesc = (IA64FunDesc *)wptr;
+ StgWord64 wcode = wdesc->ip;
+ IA64FunDesc *fdesc;
+ StgWord64 *code;
+
+ /* we allocate on the Haskell heap since malloc'd memory isn't executable - argh */
+ adjustor = stgAllocStable(sizeof(IA64FunDesc)+18*8, &stable);
+
+ fdesc = (IA64FunDesc *)adjustor;
+ code = (StgWord64 *)(fdesc + 1);
+ fdesc->ip = (StgWord64)code;
+ fdesc->gp = wdesc->gp;
+
+ code[0] = 0x0000058004288004 | MOVL_LOWORD(wcode);
+ code[1] = 0x6000000220000000 | MOVL_HIWORD(wcode);
+ code[2] = 0x029015d818984001;
+ code[3] = 0x8401200500420094;
+ code[4] = 0x886011d8189c0001;
+ code[5] = 0x84011004c00380c0;
+ code[6] = 0x0250210046013800;
+ code[7] = 0x8401000480420084;
+ code[8] = 0x0000233f19a06005 | MOVL_LOWORD((StgWord64)hptr);
+ code[9] = 0x6000000440000000 | MOVL_HIWORD((StgWord64)hptr);
+ code[10] = 0x0200210020010811;
+ code[11] = 0x1080006800006200;
+ code[12] = 0x0000210018406000;
+ code[13] = 0x00aa021000038005;
+ code[14] = 0x000000010000001d;
+ code[15] = 0x0084000880000200;
+
+ /* save stable pointers in convenient form */
+ code[16] = (StgWord64)hptr;
+ code[17] = (StgWord64)stable;
+ }
#else
barf("adjustor creation not supported on this platform");
#endif
#if defined(i386_TARGET_ARCH)
if ( *(unsigned char*)ptr != 0x68 &&
*(unsigned char*)ptr != 0x58 ) {
- fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
return;
}
freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x02)));
}
#elif defined(sparc_TARGET_ARCH)
- if ( *(unsigned long*)ptr != 0x9A10000B ) {
- fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ if ( *(unsigned long*)ptr != 0x9C23A008UL ) {
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
return;
}
/* Free the stable pointer first..*/
- freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 8)));
+ freeStablePtr(*((StgStablePtr*)((unsigned long*)ptr + 11)));
#elif defined(alpha_TARGET_ARCH)
if ( *(StgWord64*)ptr != 0xa77b0018a61b0010L ) {
- fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
return;
}
freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 0x10)));
#elif defined(powerpc_TARGET_ARCH)
if ( *(StgWord*)ptr != 0x7d0a4378 ) {
- fprintf(stderr, "freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
return;
}
freeStablePtr(*((StgStablePtr*)((unsigned char*)ptr + 4*12)));
+#elif defined(ia64_TARGET_ARCH)
+ IA64FunDesc *fdesc = (IA64FunDesc *)ptr;
+ StgWord64 *code = (StgWord64 *)(fdesc+1);
+
+ if (fdesc->ip != (StgWord64)code) {
+ errorBelch("freeHaskellFunctionPtr: not for me, guv! %p\n", ptr);
+ return;
+ }
+ freeStablePtr((StgStablePtr)code[16]);
+ freeStablePtr((StgStablePtr)code[17]);
+ return;
#else
ASSERT(0);
#endif
*((unsigned char*)ptr) = '\0';
- free(ptr);
+ stgFree(ptr);
}
+
+/*
+ * Function: initAdjustor()
+ *
+ * Perform initialisation of adjustor thunk layer (if needed.)
+ */
+void
+initAdjustor(void)
+{
+#if defined(i386_TARGET_ARCH)
+ /* Now here's something obscure for you:
+
+ When generating an adjustor thunk that uses the C calling
+ convention, we have to make sure that the thunk kicks off
+ the process of jumping into Haskell with a tail jump. Why?
+ Because as a result of jumping in into Haskell we may end
+ up freeing the very adjustor thunk we came from using
+ freeHaskellFunctionPtr(). Hence, we better not return to
+ the adjustor code on our way out, since it could by then
+ point to junk.
+
+ The fix is readily at hand, just include the opcodes
+ for the C stack fixup code that we need to perform when
+ returning in some static piece of memory and arrange
+ to return to it before tail jumping from the adjustor thunk.
+ */
+
+ obscure_ccall_ret_code = mallocBytesRWX(4);
+
+ obscure_ccall_ret_code[0x00] = (unsigned char)0x83; /* addl $0x4, %esp */
+ obscure_ccall_ret_code[0x01] = (unsigned char)0xc4;
+ obscure_ccall_ret_code[0x02] = (unsigned char)0x04;
+
+ obscure_ccall_ret_code[0x03] = (unsigned char)0xc3; /* ret */
+#endif
+}