re PR libffi/29152 (64-bit darwin ppc port needed for libffi)

libffi:
	PR libffi/29152
	PR libffi/42378
	* src/powerpc/darwin_closure.S: Provide Darwin64 implementation,
	update comments.
	* src/powerpc/ffitarget.h (POWERPC_DARWIN64): New,
	(FFI_TRAMPOLINE_SIZE): Update for Darwin64.
	* src/powerpc/darwin.S: Provide Darwin64 implementation,
	update comments.
	* src/powerpc/ffi_darwin.c: Likewise.

From-SVN: r168016

libffi/ChangeLog

@@ -1,3 +1,15 @@
+2010-12-18  Iain Sandoe  <iains@gcc.gnu.org>
+
+	PR libffi/29152
+	PR libffi/42378
+	* src/powerpc/darwin_closure.S: Provide Darwin64 implementation,
+	update comments.
+	* src/powerpc/ffitarget.h (POWERPC_DARWIN64): New,
+	(FFI_TRAMPOLINE_SIZE): Update for Darwin64.
+	* src/powerpc/darwin.S: Provide Darwin64 implementation,
+	update comments.
+	* src/powerpc/ffi_darwin.c: Likewise.
+	
 2010-12-06  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>
 
 	* configure.ac (libffi_cv_as_ascii_pseudo_op): Use double

libffi/src/powerpc/darwin.S

@@ -1,6 +1,6 @@
 /* -----------------------------------------------------------------------
    darwin.S - Copyright (c) 2000 John Hornkvist
-	      Copyright (c) 2004 Free Software Foundation, Inc.
+	      Copyright (c) 2004, 2010 Free Software Foundation, Inc.
 
    PowerPC Assembly glue.
 
@@ -24,51 +24,92 @@
    OTHER DEALINGS IN THE SOFTWARE.
    ----------------------------------------------------------------------- */
 
+#define LIBFFI_ASM
 #if defined(__ppc64__)
 #define MODE_CHOICE(x, y) y
 #else
 #define MODE_CHOICE(x, y) x
 #endif
 
-#define g_long  MODE_CHOICE(long, quad)         /* usage is ".g_long" */
+#define machine_choice	MODE_CHOICE(ppc7400,ppc64)
 
-#define LOG2_GPR_BYTES  MODE_CHOICE(2,3)        /* log2(GPR_BYTES) */
+; Define some pseudo-opcodes for size-independent load & store of GPRs ...
+#define lgu		MODE_CHOICE(lwzu, ldu)
+#define lg		MODE_CHOICE(lwz,ld)
+#define sg		MODE_CHOICE(stw,std)
+#define sgu		MODE_CHOICE(stwu,stdu)
+#define sgux		MODE_CHOICE(stwux,stdux)
+
+; ... and the size of GPRs and their storage indicator.
+#define GPR_BYTES	MODE_CHOICE(4,8)
+#define LOG2_GPR_BYTES	MODE_CHOICE(2,3)	/* log2(GPR_BYTES) */
+#define g_long		MODE_CHOICE(long, quad)	/* usage is ".g_long" */
+
+; From the ABI doc: "Mac OS X ABI Function Call Guide" Version 2009-02-04.
+#define LINKAGE_SIZE	MODE_CHOICE(24,48)
+#define PARAM_AREA	MODE_CHOICE(32,64)
+#define SAVED_LR_OFFSET	MODE_CHOICE(8,16)	/* save position for lr */
+
+/* If there is any FP stuff we make space for all of the regs.  */
+#define SAVED_FPR_COUNT 13
+#define FPR_SIZE	8
+#define RESULT_BYTES	16
+
+/* This should be kept in step with the same value in ffi_darwin.c.  */
+#define ASM_NEEDS_REGISTERS 4
+#define SAVE_REGS_SIZE (ASM_NEEDS_REGISTERS * GPR_BYTES)
 
-#define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
+
 #define JUMPTARGET(name) name
 #define L(x) x
-.text
-	.align 2
-.globl _ffi_prep_args
 
-.text
+	.text
 	.align 2
-.globl _ffi_call_DARWIN
+	.globl _ffi_prep_args
-.text
+
 	.align 2
+	.globl _ffi_call_DARWIN
+
+	/* We arrive here with:
+	   r3 = ptr to extended cif.
+	   r4 = -bytes.
+	   r5 = cif flags.
+	   r6 = ptr to return value.
+	   r7 = fn pointer (user func).
+	   r8 = fn pointer (ffi_prep_args).
+	   r9 = ffi_type* for the ret val.  */
+
 _ffi_call_DARWIN:
-LFB0:
+Lstartcode:
 	mr   	r12,r8	/* We only need r12 until the call,
-			   so it doesn't have to be saved.  */
+			   so it does not have to be saved.  */
 LFB1:
 	/* Save the old stack pointer as AP.  */
 	mr	r8,r1
 LCFI0:
+	
+	/* Save the retval type in parents frame.  */
+	sg	r9,(LINKAGE_SIZE+6*GPR_BYTES)(r8)
+
 	/* Allocate the stack space we need.  */
-	stwux	r1,r1,r4
+	sgux	r1,r1,r4
 
 	/* Save registers we use.  */
 	mflr	r9
+	sg	r9,SAVED_LR_OFFSET(r8)
 
-	stw	r28,-16(r8)	
+	sg	r28,-(4 * GPR_BYTES)(r8)	
-	stw	r29,-12(r8)
+	sg	r29,-(3 * GPR_BYTES)(r8)
-	stw	r30,-8(r8)
+	sg	r30,-(2 * GPR_BYTES)(r8)
-	stw	r31,-4(r8)
+	sg	r31,-(    GPR_BYTES)(r8)
+
+#if !defined(POWERPC_DARWIN)
+	/* The TOC slot is reserved in the Darwin ABI and r2 is volatile.  */
+	sg	r2,(5 * GPR_BYTES)(r1)
+#endif
 
-	stw	r9,8(r8)
-	stw	r2,20(r1)
 LCFI1:
 
 	/* Save arguments over call.  */
@@ -77,14 +118,17 @@ LCFI1:
 	mr	r29,r7	/* function address,  */
 	mr	r28,r8	/* our AP.  */
 LCFI2:
-	/* Call ffi_prep_args.  */
+	/* Call ffi_prep_args. r3 = extended cif, r4 = stack ptr copy.  */
 	mr	r4,r1
 	li	r9,0
 
 	mtctr	r12 /* r12 holds address of _ffi_prep_args.  */
 	bctrl
-	lwz     r2,20(r1)
 
+#if !defined(POWERPC_DARWIN)
+	/* The TOC slot is reserved in the Darwin ABI and r2 is volatile.  */
+	lg     r2,(5 * GPR_BYTES)(r1)
+#endif
 	/* Now do the call.
 	   Set up cr1 with bits 4-7 of the flags.  */
 	mtcrf	0x40,r31
@@ -92,71 +136,130 @@ LCFI2:
 	mtctr	r29
 	/* Load all those argument registers.
 	   We have set up a nice stack frame, just load it into registers.  */
-	lwz     r3,20+(1*4)(r1)
+	lg     r3, (LINKAGE_SIZE                )(r1)
-	lwz     r4,20+(2*4)(r1)
+	lg     r4, (LINKAGE_SIZE +     GPR_BYTES)(r1)
-	lwz     r5,20+(3*4)(r1)
+	lg     r5, (LINKAGE_SIZE + 2 * GPR_BYTES)(r1)
-	lwz     r6,20+(4*4)(r1)
+	lg     r6, (LINKAGE_SIZE + 3 * GPR_BYTES)(r1)
 	nop
-	lwz     r7,20+(5*4)(r1)
+	lg     r7, (LINKAGE_SIZE + 4 * GPR_BYTES)(r1)
-	lwz     r8,20+(6*4)(r1)
+	lg     r8, (LINKAGE_SIZE + 5 * GPR_BYTES)(r1)
-	lwz     r9,20+(7*4)(r1)
+	lg     r9, (LINKAGE_SIZE + 6 * GPR_BYTES)(r1)
-	lwz     r10,20+(8*4)(r1)
+	lg     r10,(LINKAGE_SIZE + 7 * GPR_BYTES)(r1)
 
 L1:
-	/* Load all the FP registers.  */
+	/* ... Load all the FP registers.  */
 	bf	6,L2	/* No floats to load.  */
-	lfd	f1,-16-(13*8)(r28)
+	lfd	f1, -SAVE_REGS_SIZE-(13*FPR_SIZE)(r28)
-	lfd	f2,-16-(12*8)(r28)
+	lfd	f2, -SAVE_REGS_SIZE-(12*FPR_SIZE)(r28)
-	lfd	f3,-16-(11*8)(r28)
+	lfd	f3, -SAVE_REGS_SIZE-(11*FPR_SIZE)(r28)
-	lfd	f4,-16-(10*8)(r28)
+	lfd	f4, -SAVE_REGS_SIZE-(10*FPR_SIZE)(r28)
 	nop
-	lfd	f5,-16-(9*8)(r28)
+	lfd	f5, -SAVE_REGS_SIZE-( 9*FPR_SIZE)(r28)
-	lfd	f6,-16-(8*8)(r28)
+	lfd	f6, -SAVE_REGS_SIZE-( 8*FPR_SIZE)(r28)
-	lfd	f7,-16-(7*8)(r28)
+	lfd	f7, -SAVE_REGS_SIZE-( 7*FPR_SIZE)(r28)
-	lfd	f8,-16-(6*8)(r28)
+	lfd	f8, -SAVE_REGS_SIZE-( 6*FPR_SIZE)(r28)
 	nop
-	lfd     f9,-16-(5*8)(r28)
+	lfd     f9, -SAVE_REGS_SIZE-( 5*FPR_SIZE)(r28)
-	lfd     f10,-16-(4*8)(r28)
+	lfd     f10,-SAVE_REGS_SIZE-( 4*FPR_SIZE)(r28)
-	lfd     f11,-16-(3*8)(r28)
+	lfd     f11,-SAVE_REGS_SIZE-( 3*FPR_SIZE)(r28)
-	lfd     f12,-16-(2*8)(r28)
+	lfd     f12,-SAVE_REGS_SIZE-( 2*FPR_SIZE)(r28)
 	nop
-	lfd     f13,-16-(1*8)(r28)
+	lfd     f13,-SAVE_REGS_SIZE-( 1*FPR_SIZE)(r28)
 
 L2:
 	mr	r12,r29	/* Put the target address in r12 as specified.  */
 	mtctr  	r12
 	nop
 	nop
+
 	/* Make the call.  */
 	bctrl
 
 	/* Now, deal with the return value.  */
-	mtcrf	0x01,r31
 
-	bt	30,L(done_return_value)
+	/* m64 structure returns can occupy the same set of registers as
-	bt	29,L(fp_return_value)
+	   would be used to pass such a structure as arg0 - so take care 
-	stw	r3,0(r30)
+	   not to step on any possibly hot regs.  */
-	bf	28,L(done_return_value)
-	stw	r4,4(r30)
 
-	/* Fall through.  */
+	/* Get the flags.. */
+	mtcrf	0x03,r31 ; we need c6 & cr7 now.
+	; FLAG_RETURNS_NOTHING also covers struct ret-by-ref.
+	bt	30,L(done_return_value)	  ; FLAG_RETURNS_NOTHING
+	bf	27,L(scalar_return_value) ; not FLAG_RETURNS_STRUCT
+	
+	/* OK, so we have a struct.  */
+#if defined(__ppc64__)
+	bt	31,L(maybe_return_128) ; FLAG_RETURNS_128BITS, special case 
 
-L(done_return_value):
+	/* OK, we have to map the return back to a mem struct.
-	/* Restore the registers we used and return.  */
+	   We are about to trample the parents param area, so recover the
-	lwz	r9,8(r28)
+	   return type.  r29 is free, since the call is done.  */
-	lwz	r31,-4(r28)
+	lg	r29,(LINKAGE_SIZE + 6 * GPR_BYTES)(r28)
-	mtlr	r9
+
-	lwz	r30,-8(r28)
+	sg	r3, (LINKAGE_SIZE                )(r28)
-	lwz	r29,-12(r28)
+	sg	r4, (LINKAGE_SIZE +     GPR_BYTES)(r28)
-	lwz	r28,-16(r28)
+	sg	r5, (LINKAGE_SIZE + 2 * GPR_BYTES)(r28)
-	lwz	r1,0(r1)
+	sg	r6, (LINKAGE_SIZE + 3 * GPR_BYTES)(r28)
-	blr
+	nop
+	sg	r7, (LINKAGE_SIZE + 4 * GPR_BYTES)(r28)
+	sg	r8, (LINKAGE_SIZE + 5 * GPR_BYTES)(r28)
+	sg	r9, (LINKAGE_SIZE + 6 * GPR_BYTES)(r28)
+	sg	r10,(LINKAGE_SIZE + 7 * GPR_BYTES)(r28)
+	/* OK, so do the block move - we trust that memcpy will not trample
+	   the fprs...  */
+	mr 	r3,r30 ; dest
+	addi	r4,r28,LINKAGE_SIZE ; source
+	/* The size is a size_t, should be long.  */
+	lg	r5,0(r29)
+	/* Figure out small structs */
+	cmpi	0,r5,4
+	bgt	L3	; 1, 2 and 4 bytes have special rules.
+	cmpi	0,r5,3
+	beq	L3	; not 3
+	addi	r4,r4,8
+	subf	r4,r5,r4
+L3:
+	bl	_memcpy
+	
+	/* ... do we need the FP registers? - recover the flags.. */
+	mtcrf	0x03,r31 ; we need c6 & cr7 now.
+	bf	29,L(done_return_value)	/* No floats in the struct.  */
+	stfd	f1, -SAVE_REGS_SIZE-(13*FPR_SIZE)(r28)
+	stfd	f2, -SAVE_REGS_SIZE-(12*FPR_SIZE)(r28)
+	stfd	f3, -SAVE_REGS_SIZE-(11*FPR_SIZE)(r28)
+	stfd	f4, -SAVE_REGS_SIZE-(10*FPR_SIZE)(r28)
+	nop
+	stfd	f5, -SAVE_REGS_SIZE-( 9*FPR_SIZE)(r28)
+	stfd	f6, -SAVE_REGS_SIZE-( 8*FPR_SIZE)(r28)
+	stfd	f7, -SAVE_REGS_SIZE-( 7*FPR_SIZE)(r28)
+	stfd	f8, -SAVE_REGS_SIZE-( 6*FPR_SIZE)(r28)
+	nop
+	stfd	f9, -SAVE_REGS_SIZE-( 5*FPR_SIZE)(r28)
+	stfd	f10,-SAVE_REGS_SIZE-( 4*FPR_SIZE)(r28)
+	stfd	f11,-SAVE_REGS_SIZE-( 3*FPR_SIZE)(r28)
+	stfd	f12,-SAVE_REGS_SIZE-( 2*FPR_SIZE)(r28)
+	nop
+	stfd	f13,-SAVE_REGS_SIZE-( 1*FPR_SIZE)(r28)
+
+	mr	r3,r29	; ffi_type *
+	mr	r4,r30	; dest
+	addi	r5,r28,-SAVE_REGS_SIZE-(13*FPR_SIZE) ; fprs
+	xor	r6,r6,r6
+	sg	r6,(LINKAGE_SIZE + 7 * GPR_BYTES)(r28)
+	addi	r6,r28,(LINKAGE_SIZE + 7 * GPR_BYTES) ; point to a zeroed counter.
+	bl 	_darwin64_struct_floats_to_mem
+
+	b L(done_return_value)
+#else
+	stw	r3,0(r30) ; m32 the only struct return in reg is 4 bytes.
+#endif
+	b L(done_return_value)
 
 L(fp_return_value):
 	/* Do we have long double to store?  */
-	bf	31,L(fd_return_value)
+	bf	31,L(fd_return_value) ; FLAG_RETURNS_128BITS
 	stfd	f1,0(r30)
-	stfd	f2,8(r30)
+	stfd	f2,FPR_SIZE(r30)
 	b	L(done_return_value)
 
 L(fd_return_value):
@@ -170,21 +273,57 @@ L(float_return_value):
 	stfs	f1,0(r30)
 	b	L(done_return_value)
 
+L(scalar_return_value):
+	bt	29,L(fp_return_value)	; FLAG_RETURNS_FP
+	; ffi_arg is defined as unsigned long. 
+	sg	r3,0(r30)		; Save the reg.
+	bf	28,L(done_return_value) ; not FLAG_RETURNS_64BITS 
+
+#if defined(__ppc64__)
+L(maybe_return_128):
+	std	r3,0(r30)
+	bf	31,L(done_return_value) ; not FLAG_RETURNS_128BITS 
+	std	r4,8(r30)
+#else
+	stw	r4,4(r30)
+#endif
+
+	/* Fall through.  */
+	/* We want this at the end to simplify eh epilog computation.  */
+
+L(done_return_value):
+	/* Restore the registers we used and return.  */
+	lg	r29,SAVED_LR_OFFSET(r28)
+	; epilog
+	lg	r31,-(1 * GPR_BYTES)(r28)
+	mtlr	r29
+	lg	r30,-(2 * GPR_BYTES)(r28)
+	lg	r29,-(3 * GPR_BYTES)(r28)
+	lg	r28,-(4 * GPR_BYTES)(r28)
+	lg	r1,0(r1)
+	blr
 LFE1:
+	.align	1
 /* END(_ffi_call_DARWIN)  */
 
 /* Provide a null definition of _ffi_call_AIX.  */
-.text
+	.text
-	.align 2
+	.globl _ffi_call_AIX
-.globl _ffi_call_AIX
-.text
 	.align 2
 _ffi_call_AIX:
 	blr
 /* END(_ffi_call_AIX)  */
 
-.data
+/* EH stuff.  */
-.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms
+
+#define EH_DATA_ALIGN_FACT MODE_CHOICE(0x7c,0x78)
+
+	.static_data
+	.align LOG2_GPR_BYTES
+LLFB0$non_lazy_ptr:
+	.g_long Lstartcode
+
+	.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support
 EH_frame1:
 	.set	L$set$0,LECIE1-LSCIE1
 	.long	L$set$0	; Length of Common Information Entry
@@ -193,7 +332,7 @@ LSCIE1:
 	.byte	0x1	; CIE Version
 	.ascii	"zR\0"	; CIE Augmentation
 	.byte	0x1	; uleb128 0x1; CIE Code Alignment Factor
-	.byte	0x7c	; sleb128 -4; CIE Data Alignment Factor
+	.byte	EH_DATA_ALIGN_FACT ; sleb128 -4; CIE Data Alignment Factor
 	.byte	0x41	; CIE RA Column
 	.byte	0x1	; uleb128 0x1; Augmentation size
 	.byte	0x90	; FDE Encoding (indirect pcrel)
@@ -202,7 +341,8 @@ LSCIE1:
 	.byte	0x0	; uleb128 0x0
 	.align	LOG2_GPR_BYTES
 LECIE1:
-.globl _ffi_call_DARWIN.eh
+
+	.globl _ffi_call_DARWIN.eh
 _ffi_call_DARWIN.eh:
 LSFDE1:
 	.set	L$set$1,LEFDE1-LASFDE1
@@ -210,11 +350,11 @@ LSFDE1:
 LASFDE1:
 	.long	LASFDE1-EH_frame1 ; FDE CIE offset
 	.g_long	LLFB0$non_lazy_ptr-.	; FDE initial location
-	.set	L$set$3,LFE1-LFB0
+	.set	L$set$3,LFE1-Lstartcode
 	.g_long	L$set$3	; FDE address range
 	.byte   0x0     ; uleb128 0x0; Augmentation size
 	.byte	0x4	; DW_CFA_advance_loc4
-	.set	L$set$4,LCFI0-LFB1
+	.set	L$set$4,LCFI0-Lstartcode
 	.long	L$set$4
 	.byte	0xd	; DW_CFA_def_cfa_register
 	.byte	0x08	; uleb128 0x08
@@ -239,7 +379,5 @@ LASFDE1:
 	.byte	0x1c	; uleb128 0x1c
 	.align LOG2_GPR_BYTES
 LEFDE1:
-.data
+	.align 1
-	.align LOG2_GPR_BYTES
+
-LLFB0$non_lazy_ptr:
-	.g_long LFB0

libffi/src/powerpc/darwin_closure.S

@@ -1,6 +1,7 @@
 /* -----------------------------------------------------------------------
-   darwin_closure.S - Copyright (c) 2002, 2003, 2004, Free Software Foundation,
+   darwin_closure.S - Copyright (c) 2002, 2003, 2004, 2010, 
-   Inc. based on ppc_closure.S
+   Free Software Foundation, Inc. 
+   based on ppc_closure.S
 
    PowerPC Assembly glue.
 
@@ -33,91 +34,177 @@
 #define MODE_CHOICE(x, y) x
 #endif
 
-#define lgu     MODE_CHOICE(lwzu, ldu)
+#define machine_choice	MODE_CHOICE(ppc7400,ppc64)
 
-#define g_long  MODE_CHOICE(long, quad)         /* usage is ".g_long" */
+; Define some pseudo-opcodes for size-independent load & store of GPRs ...
+#define lgu		MODE_CHOICE(lwzu, ldu)
+#define lg		MODE_CHOICE(lwz,ld)
+#define sg		MODE_CHOICE(stw,std)
+#define sgu		MODE_CHOICE(stwu,stdu)
 
-#define LOG2_GPR_BYTES  MODE_CHOICE(2,3)        /* log2(GPR_BYTES) */
+; ... and the size of GPRs and their storage indicator.
+#define GPR_BYTES	MODE_CHOICE(4,8)
+#define LOG2_GPR_BYTES	MODE_CHOICE(2,3)	/* log2(GPR_BYTES) */
+#define g_long		MODE_CHOICE(long, quad)	/* usage is ".g_long" */
+
+; From the ABI doc: "Mac OS X ABI Function Call Guide" Version 2009-02-04.
+#define LINKAGE_SIZE	MODE_CHOICE(24,48)
+#define PARAM_AREA	MODE_CHOICE(32,64)
+
+#define SAVED_CR_OFFSET	MODE_CHOICE(4,8)	/* save position for CR */
+#define SAVED_LR_OFFSET	MODE_CHOICE(8,16)	/* save position for lr */
+
+/* WARNING: if ffi_type is changed... here be monsters.  
+   Offsets of items within the result type.  */
+#define FFI_TYPE_TYPE	MODE_CHOICE(6,10)
+#define FFI_TYPE_ELEM	MODE_CHOICE(8,16)
+
+#define SAVED_FPR_COUNT 13
+#define FPR_SIZE	8
+/* biggest m64 struct ret is 8GPRS + 13FPRS = 168 bytes - rounded to 16bytes = 176. */
+#define RESULT_BYTES	MODE_CHOICE(16,176)
+
+; The whole stack frame **MUST** be 16byte-aligned.
+#define SAVE_SIZE (((LINKAGE_SIZE+PARAM_AREA+SAVED_FPR_COUNT*FPR_SIZE+RESULT_BYTES)+15) & -16LL)
+#define PAD_SIZE (SAVE_SIZE-(LINKAGE_SIZE+PARAM_AREA+SAVED_FPR_COUNT*FPR_SIZE+RESULT_BYTES))
+
+#define PARENT_PARM_BASE (SAVE_SIZE+LINKAGE_SIZE)
+#define FP_SAVE_BASE (LINKAGE_SIZE+PARAM_AREA)
+
+#if defined(__ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__) && __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__ >= 1050
+; We no longer need the pic symbol stub for Darwin >= 9.
+#define BLCLS_HELP _ffi_closure_helper_DARWIN
+#define STRUCT_RETVALUE_P _darwin64_struct_ret_by_value_p
+#define PASS_STR_FLOATS _darwin64_pass_struct_floats
+#undef WANT_STUB
+#else
+#define BLCLS_HELP L_ffi_closure_helper_DARWIN$stub
+#define STRUCT_RETVALUE_P L_darwin64_struct_ret_by_value_p$stub
+#define PASS_STR_FLOATS L_darwin64_pass_struct_floats$stub
+#define WANT_STUB
+#endif
+
+/* m32/m64
+
+   The stack layout looks like this:
+
+   |   Additional params...			| |     Higher address
+   ~						~ ~
+   |   Parameters      (at least 8*4/8=32/64)	| | NUM_GPR_ARG_REGISTERS
+   |--------------------------------------------| |
+   |   TOC=R2 (AIX) Reserved (Darwin)   4/8	| |
+   |--------------------------------------------| |
+   |   Reserved                       2*4/8	| |
+   |--------------------------------------------| |
+   |   Space for callee`s LR		4/8	| |
+   |--------------------------------------------| |
+   |   Saved CR [low word for m64]      4/8	| |
+   |--------------------------------------------| |
+   |   Current backchain pointer	4/8	|-/ Parent`s frame.
+   |--------------------------------------------| <+ <<< on entry to
+   |   Result Bytes		       16/176	| |
+   |--------------------------------------------| |
+   ~   padding to 16-byte alignment		~ ~
+   |--------------------------------------------| |
+   |   NUM_FPR_ARG_REGISTERS slots		| |
+   |   here fp13 .. fp1		       13*8	| |
+   |--------------------------------------------| |
+   |   R3..R10			  8*4/8=32/64	| | NUM_GPR_ARG_REGISTERS
+   |--------------------------------------------| |
+   |   TOC=R2 (AIX) Reserved (Darwin)   4/8	| |
+   |--------------------------------------------| |	stack	|
+   |   Reserved [compiler,binder]     2*4/8	| |	grows	|
+   |--------------------------------------------| |	down	V
+   |   Space for callees LR		4/8	| |
+   |--------------------------------------------| |	lower addresses
+   |   Saved CR [low word for m64]      4/8	| |
+   |--------------------------------------------| |     stack pointer here
+   |   Current backchain pointer	4/8	|-/	during
+   |--------------------------------------------|   <<<	call.
+
+*/
 
 	.file	"darwin_closure.S"
-.text
-	.align LOG2_GPR_BYTES
-.globl _ffi_closure_ASM
 
-.text
+	.machine machine_choice
+
+	.text
+	.globl _ffi_closure_ASM
 	.align LOG2_GPR_BYTES
 _ffi_closure_ASM:
 LFB1:
-	mflr	r0		/* extract return address  */
+Lstartcode:
-	stw	r0,8(r1)	/* save the return address  */
+	mflr	r0			/* extract return address  */
+	sg	r0,SAVED_LR_OFFSET(r1)	/* save the return address  */
 LCFI0:
-	/* 24 Bytes (Linkage Area)
+	sgu	r1,-SAVE_SIZE(r1)	/* skip over caller save area
-	   32 Bytes (outgoing parameter area, always reserved)
+					keep stack aligned to 16.  */
-	   104 Bytes (13*8 from FPR)
-	   16 Bytes (result)
-	   176 Bytes  */
-
-	stwu	r1,-176(r1)	/* skip over caller save area
-				keep stack aligned to 16.  */
 LCFI1:
 	/* We want to build up an area for the parameters passed
 	   in registers. (both floating point and integer)  */
 
-	/* We store gpr 3 to gpr 10 (aligned to 4)
+	/* Put gpr 3 to gpr 10 in the parents outgoing area...
-	   in the parents outgoing area.  */
+	   ... the remainder of any params that overflowed the regs will
-	stw   r3,200(r1)
+	   follow here.  */
-	stw   r4,204(r1)
+	sg	r3, (PARENT_PARM_BASE                )(r1)
-	stw   r5,208(r1)
+	sg	r4, (PARENT_PARM_BASE + GPR_BYTES    )(r1)
-	stw   r6,212(r1)
+	sg	r5, (PARENT_PARM_BASE + GPR_BYTES * 2)(r1)
-	stw   r7,216(r1)
+	sg	r6, (PARENT_PARM_BASE + GPR_BYTES * 3)(r1)
-	stw   r8,220(r1)
+	sg	r7, (PARENT_PARM_BASE + GPR_BYTES * 4)(r1)
-	stw   r9,224(r1)
+	sg	r8, (PARENT_PARM_BASE + GPR_BYTES * 5)(r1)
-	stw   r10,228(r1)
+	sg	r9, (PARENT_PARM_BASE + GPR_BYTES * 6)(r1)
+	sg	r10,(PARENT_PARM_BASE + GPR_BYTES * 7)(r1)
 
-	/* We save fpr 1 to fpr 13. (aligned to 8)  */
+	/* We save fpr 1 to fpr 14 in our own save frame.  */
-	stfd  f1,56(r1)
+	stfd	f1, (FP_SAVE_BASE                 )(r1)
-	stfd  f2,64(r1)
+	stfd	f2, (FP_SAVE_BASE +  FPR_SIZE     )(r1)
-	stfd  f3,72(r1)
+	stfd	f3, (FP_SAVE_BASE +  FPR_SIZE * 2 )(r1)
-	stfd  f4,80(r1)
+	stfd	f4, (FP_SAVE_BASE +  FPR_SIZE * 3 )(r1)
-	stfd  f5,88(r1)
+	stfd	f5, (FP_SAVE_BASE +  FPR_SIZE * 4 )(r1)
-	stfd  f6,96(r1)
+	stfd	f6, (FP_SAVE_BASE +  FPR_SIZE * 5 )(r1)
-	stfd  f7,104(r1)
+	stfd	f7, (FP_SAVE_BASE +  FPR_SIZE * 6 )(r1)
-	stfd  f8,112(r1)
+	stfd	f8, (FP_SAVE_BASE +  FPR_SIZE * 7 )(r1)
-	stfd  f9,120(r1)
+	stfd	f9, (FP_SAVE_BASE +  FPR_SIZE * 8 )(r1)
-	stfd  f10,128(r1)
+	stfd	f10,(FP_SAVE_BASE +  FPR_SIZE * 9 )(r1)
-	stfd  f11,136(r1)
+	stfd	f11,(FP_SAVE_BASE +  FPR_SIZE * 10)(r1)
-	stfd  f12,144(r1)
+	stfd	f12,(FP_SAVE_BASE +  FPR_SIZE * 11)(r1)
-	stfd  f13,152(r1)
+	stfd	f13,(FP_SAVE_BASE +  FPR_SIZE * 12)(r1)
 
 	/* Set up registers for the routine that actually does the work
 	   get the context pointer from the trampoline.  */
-	mr r3,r11
+	mr	r3,r11
 
 	/* Now load up the pointer to the result storage.  */
-	addi r4,r1,160
+	addi	r4,r1,(SAVE_SIZE-RESULT_BYTES)
 
 	/* Now load up the pointer to the saved gpr registers.  */
-	addi r5,r1,200
+	addi	r5,r1,PARENT_PARM_BASE
 
 	/* Now load up the pointer to the saved fpr registers.  */
-	addi r6,r1,56
+	addi	r6,r1,FP_SAVE_BASE
 
 	/* Make the call.  */
-	bl	Lffi_closure_helper_DARWIN$stub
+	bl	BLCLS_HELP
 
-	/* Now r3 contains the return type
+	/* r3 contains the rtype pointer... save it since we will need
-	   so use it to look up in a table
+	   it later.  */
+	sg	r3,LINKAGE_SIZE(r1)	; ffi_type * result_type
+	lg	r0,0(r3)		; size => r0
+	lhz	r3,FFI_TYPE_TYPE(r3)	; type => r3
+
+	/* The helper will have intercepted struture returns and inserted
+	   the caller`s destination address for structs returned by ref.  */
+
+	/* r3 contains the return type  so use it to look up in a table
 	   so we know how to deal with each type.  */
 
-	/* Look up the proper starting point in table
+	addi	r5,r1,(SAVE_SIZE-RESULT_BYTES) /* Otherwise, our return is here.  */
-	   by using return type as offset.  */
+	bl	Lget_ret_type0_addr	/* Get pointer to Lret_type0 into LR.  */
-	addi  r5,r1,160		  /* Get pointer to results area.  */
+	mflr	r4			/* Move to r4.  */
-	bl    Lget_ret_type0_addr /* Get pointer to Lret_type0 into LR.  */
+	slwi	r3,r3,4			/* Now multiply return type by 16.  */
-	mflr  r4		  /* Move to r4.  */
+	add	r3,r3,r4		/* Add contents of table to table address.  */
-	slwi  r3,r3,4		  /* Now multiply return type by 16.  */
+	mtctr	r3
-	add   r3,r3,r4		  /* Add contents of table to table address.  */
+	bctr			 	 /* Jump to it.  */
-	mtctr r3
-	bctr			  /* Jump to it.  */
 LFE1:
 /* Each of the ret_typeX code fragments has to be exactly 16 bytes long
    (4 instructions). For cache effectiveness we align to a 16 byte boundary
@@ -140,7 +227,7 @@ Lret_type0:
 
 /* case FFI_TYPE_INT  */
 Lret_type1:
-	lwz	r3,0(r5)
+	lg	r3,0(r5)
 	b	Lfinish
 	nop
 	nop
@@ -168,85 +255,224 @@ Lret_type4:
 
 /* case FFI_TYPE_UINT8  */
 Lret_type5:
+#if defined(__ppc64__)
+	lbz	r3,7(r5)
+#else
 	lbz	r3,3(r5)
+#endif
 	b	Lfinish
 	nop
 	nop
 
 /* case FFI_TYPE_SINT8  */
 Lret_type6:
+#if defined(__ppc64__)
+	lbz	r3,7(r5)
+#else
 	lbz	r3,3(r5)
+#endif
 	extsb	r3,r3
 	b	Lfinish
 	nop
 
 /* case FFI_TYPE_UINT16  */
 Lret_type7:
+#if defined(__ppc64__)
+	lhz	r3,6(r5)
+#else
 	lhz	r3,2(r5)
+#endif
 	b	Lfinish
 	nop
 	nop
 
 /* case FFI_TYPE_SINT16  */
 Lret_type8:
+#if defined(__ppc64__)
+	lha	r3,6(r5)
+#else
 	lha	r3,2(r5)
+#endif
 	b	Lfinish
 	nop
 	nop
 
 /* case FFI_TYPE_UINT32  */
 Lret_type9:
+#if defined(__ppc64__)
+	lwz	r3,4(r5)
+#else
 	lwz	r3,0(r5)
+#endif
 	b	Lfinish
 	nop
 	nop
 
 /* case FFI_TYPE_SINT32  */
 Lret_type10:
+#if defined(__ppc64__)
+	lwz	r3,4(r5)
+#else
 	lwz	r3,0(r5)
+#endif
 	b	Lfinish
 	nop
 	nop
 
 /* case FFI_TYPE_UINT64  */
 Lret_type11:
+#if defined(__ppc64__)
+	lg	r3,0(r5)
+	b	Lfinish
+	nop
+#else
 	lwz	r3,0(r5)
 	lwz	r4,4(r5)
 	b	Lfinish
+#endif
 	nop
 
 /* case FFI_TYPE_SINT64  */
 Lret_type12:
+#if defined(__ppc64__)
+	lg	r3,0(r5)
+	b	Lfinish
+	nop
+#else
 	lwz	r3,0(r5)
 	lwz	r4,4(r5)
 	b	Lfinish
+#endif
 	nop
 
 /* case FFI_TYPE_STRUCT  */
 Lret_type13:
+#if defined(__ppc64__)
+	lg	r3,0(r5)		; we need at least this...
+	cmpi	0,r0,4
+	bgt	Lstructend		; not a special small case
+	b	Lsmallstruct		; see if we need more.
+#else
+	cmpi	0,r0,4
+	bgt	Lfinish		; not by value
+	lg	r3,0(r5)
+	b	Lfinish
+#endif
+/* case FFI_TYPE_POINTER  */
+Lret_type14:
+	lg	r3,0(r5)
 	b	Lfinish
-	nop
 	nop
 	nop
 
-/* case FFI_TYPE_POINTER  */
+#if defined(__ppc64__)
-Lret_type14:
+Lsmallstruct:
-	lwz	r3,0(r5)
+	beq	Lfour			; continuation of Lret13.
-	b	Lfinish
+	cmpi	0,r0,3
-	nop
+	beq	Lfinish			; don`t adjust this - can`t be any floats here...
-	nop
+	srdi	r3,r3,48
+	cmpi	0,r0,2
+	beq	Lfinish			; .. or here ..
+	srdi	r3,r3,8
+	b 	Lfinish			; .. or here.
+
+Lfour:
+	lg	r6,LINKAGE_SIZE(r1)	; get the result type
+	lg	r6,FFI_TYPE_ELEM(r6)	; elements array pointer
+	lg	r6,0(r6)		; first element
+	lhz	r0,FFI_TYPE_TYPE(r6)	; OK go the type
+	cmpi	0,r0,2			; FFI_TYPE_FLOAT
+	bne	Lfourint
+	lfs	f1,0(r5)		; just one float in the struct.
+	b 	Lfinish
+
+Lfourint:
+	srdi	r3,r3,32		; four bytes.
+	b 	Lfinish
+
+Lstructend:
+	lg	r3,LINKAGE_SIZE(r1)	; get the result type
+	bl	STRUCT_RETVALUE_P
+	cmpi	0,r3,0
+	beq	Lfinish			; nope.
+	/* Recover a pointer to the results.  */
+	addi	r11,r1,(SAVE_SIZE-RESULT_BYTES)
+	lg	r3,0(r11)		; we need at least this...
+	lg	r4,8(r11)
+	cmpi	0,r0,16
+	beq	Lfinish		; special case 16 bytes we don't consider floats.
+
+	/* OK, frustratingly, the process of saving the struct to mem might have
+	   messed with the FPRs, so we have to re-load them :(.
+	   We`ll use our FPRs space again - calling: 
+	   void darwin64_pass_struct_floats (ffi_type *s, char *src, 
+					     unsigned *nfpr, double **fprs) 
+	   We`ll temporarily pinch the first two slots of the param area for local
+	   vars used by the routine.  */
+	xor	r6,r6,r6
+	addi	r5,r1,PARENT_PARM_BASE		; some space
+	sg	r6,0(r5)			; *nfpr zeroed.
+	addi	r6,r5,8				; **fprs
+	addi	r3,r1,FP_SAVE_BASE		; pointer to FPRs space
+	sg	r3,0(r6)
+	mr	r4,r11				; the struct is here...
+	lg	r3,LINKAGE_SIZE(r1)		; ffi_type * result_type.
+	bl	PASS_STR_FLOATS			; get struct floats into FPR save space.
+	/* See if we used any floats  */
+	lwz	r0,(SAVE_SIZE-RESULT_BYTES)(r1)	
+	cmpi	0,r0,0
+	beq	Lstructints			; nope.
+	/* OK load `em up... */
+	lfd	f1, (FP_SAVE_BASE                 )(r1)
+	lfd	f2, (FP_SAVE_BASE +  FPR_SIZE     )(r1)
+	lfd	f3, (FP_SAVE_BASE +  FPR_SIZE * 2 )(r1)
+	lfd	f4, (FP_SAVE_BASE +  FPR_SIZE * 3 )(r1)
+	lfd	f5, (FP_SAVE_BASE +  FPR_SIZE * 4 )(r1)
+	lfd	f6, (FP_SAVE_BASE +  FPR_SIZE * 5 )(r1)
+	lfd	f7, (FP_SAVE_BASE +  FPR_SIZE * 6 )(r1)
+	lfd	f8, (FP_SAVE_BASE +  FPR_SIZE * 7 )(r1)
+	lfd	f9, (FP_SAVE_BASE +  FPR_SIZE * 8 )(r1)
+	lfd	f10,(FP_SAVE_BASE +  FPR_SIZE * 9 )(r1)
+	lfd	f11,(FP_SAVE_BASE +  FPR_SIZE * 10)(r1)
+	lfd	f12,(FP_SAVE_BASE +  FPR_SIZE * 11)(r1)
+	lfd	f13,(FP_SAVE_BASE +  FPR_SIZE * 12)(r1)
+
+	/* point back at our saved struct.  */
+Lstructints:
+	addi	r11,r1,(SAVE_SIZE-RESULT_BYTES)
+	lg	r3,0(r11)			; we end up picking the
+	lg	r4,8(r11)			; first two again.
+	lg	r5,16(r11)
+	lg	r6,24(r11)
+	lg	r7,32(r11)
+	lg	r8,40(r11)
+	lg	r9,48(r11)
+	lg	r10,56(r11)
+#endif
 
 /* case done  */
 Lfinish:
-	addi	r1,r1,176	/* Restore stack pointer.  */
+	addi	r1,r1,SAVE_SIZE		/* Restore stack pointer.  */
-	lwz	r0,8(r1)	/* Get return address.  */
+	lg	r0,SAVED_LR_OFFSET(r1)	/* Get return address.  */
-	mtlr	r0		/* Reset link register.  */
+	mtlr	r0			/* Reset link register.  */
 	blr
-
+Lendcode:
+	.align 1
+	
 /* END(ffi_closure_ASM)  */
 
-.data
+/* EH frame stuff.  */
-.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support
+#define EH_DATA_ALIGN_FACT MODE_CHOICE(0x7c,0x78)
+/* 176, 400 */
+#define EH_FRAME_OFFSETA MODE_CHOICE(176,0x90)
+#define EH_FRAME_OFFSETB MODE_CHOICE(1,3)
+
+	.static_data
+	.align LOG2_GPR_BYTES
+LLFB1$non_lazy_ptr:
+	.g_long Lstartcode
+
+	.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support
 EH_frame1:
 	.set	L$set$0,LECIE1-LSCIE1
 	.long	L$set$0	; Length of Common Information Entry
@@ -255,7 +481,7 @@ LSCIE1:
 	.byte	0x1	; CIE Version
 	.ascii	"zR\0"	; CIE Augmentation
 	.byte	0x1	; uleb128 0x1; CIE Code Alignment Factor
-	.byte	0x7c	; sleb128 -4; CIE Data Alignment Factor
+	.byte	EH_DATA_ALIGN_FACT ; sleb128 -4; CIE Data Alignment Factor
 	.byte	0x41	; CIE RA Column
 	.byte	0x1	; uleb128 0x1; Augmentation size
 	.byte	0x90	; FDE Encoding (indirect pcrel)
@@ -264,7 +490,7 @@ LSCIE1:
 	.byte	0x0	; uleb128 0x0
 	.align	LOG2_GPR_BYTES
 LECIE1:
-.globl _ffi_closure_ASM.eh
+	.globl _ffi_closure_ASM.eh
 _ffi_closure_ASM.eh:
 LSFDE1:
 	.set	L$set$1,LEFDE1-LASFDE1
@@ -273,45 +499,77 @@ LSFDE1:
 LASFDE1:
 	.long	LASFDE1-EH_frame1	; FDE CIE offset
 	.g_long	LLFB1$non_lazy_ptr-.	; FDE initial location
-	.set	L$set$3,LFE1-LFB1
+	.set	L$set$3,LFE1-Lstartcode
 	.g_long	L$set$3	; FDE address range
 	.byte   0x0     ; uleb128 0x0; Augmentation size
 	.byte	0x4	; DW_CFA_advance_loc4
 	.set	L$set$3,LCFI1-LCFI0
 	.long	L$set$3
 	.byte	0xe	; DW_CFA_def_cfa_offset
-	.byte	176,1	; uleb128 176
+	.byte	EH_FRAME_OFFSETA,EH_FRAME_OFFSETB	; uleb128 176,1/190,3
 	.byte	0x4	; DW_CFA_advance_loc4
-	.set	L$set$4,LCFI0-LFB1
+	.set	L$set$4,LCFI0-Lstartcode
 	.long	L$set$4
 	.byte   0x11    ; DW_CFA_offset_extended_sf
 	.byte	0x41	; uleb128 0x41
 	.byte   0x7e    ; sleb128 -2
 	.align	LOG2_GPR_BYTES
 LEFDE1:
-.data
+	.align 	1
-	.align	LOG2_GPR_BYTES
+
-LDFCM0:
+#ifdef WANT_STUB
-.section __TEXT,__picsymbolstub1,symbol_stubs,pure_instructions,32
+	.section __TEXT,__picsymbolstub1,symbol_stubs,pure_instructions,32
-	.align	LOG2_GPR_BYTES
+	.align 5
-Lffi_closure_helper_DARWIN$stub:
+L_ffi_closure_helper_DARWIN$stub:
-#if 1
 	.indirect_symbol _ffi_closure_helper_DARWIN
-	mflr	r0
+	mflr r0
-	bcl	20,31,LO$ffi_closure_helper_DARWIN
+	bcl 20,31,"L00000000001$spb"
-LO$ffi_closure_helper_DARWIN:
+"L00000000001$spb":
-	mflr	r11
+	mflr r11
-	addis	r11,r11,ha16(L_ffi_closure_helper_DARWIN$lazy_ptr - LO$ffi_closure_helper_DARWIN)
+	addis r11,r11,ha16(L_ffi_closure_helper_DARWIN$lazy_ptr-"L00000000001$spb")
-	mtlr	r0
+	mtlr r0
-	lgu	r12,lo16(L_ffi_closure_helper_DARWIN$lazy_ptr - LO$ffi_closure_helper_DARWIN)(r11)
+	lwzu r12,lo16(L_ffi_closure_helper_DARWIN$lazy_ptr-"L00000000001$spb")(r11)
-	mtctr	r12
+	mtctr r12
 	bctr
-.lazy_symbol_pointer
+	.lazy_symbol_pointer
 L_ffi_closure_helper_DARWIN$lazy_ptr:
 	.indirect_symbol _ffi_closure_helper_DARWIN
-	.g_long dyld_stub_binding_helper
+	.long	dyld_stub_binding_helper
+
+#if defined(__ppc64__)
+	.section __TEXT,__picsymbolstub1,symbol_stubs,pure_instructions,32
+	.align 5
+L_darwin64_struct_ret_by_value_p$stub:
+	.indirect_symbol _darwin64_struct_ret_by_value_p
+	mflr r0
+	bcl 20,31,"L00000000002$spb"
+"L00000000002$spb":
+	mflr r11
+	addis r11,r11,ha16(L_darwin64_struct_ret_by_value_p$lazy_ptr-"L00000000002$spb")
+	mtlr r0
+	lwzu r12,lo16(L_darwin64_struct_ret_by_value_p$lazy_ptr-"L00000000002$spb")(r11)
+	mtctr r12
+	bctr
+	.lazy_symbol_pointer
+L_darwin64_struct_ret_by_value_p$lazy_ptr:
+	.indirect_symbol _darwin64_struct_ret_by_value_p
+	.long	dyld_stub_binding_helper
+	.section __TEXT,__picsymbolstub1,symbol_stubs,pure_instructions,32
+	.align 5
+L_darwin64_pass_struct_floats$stub:
+	.indirect_symbol _darwin64_pass_struct_floats
+	mflr r0
+	bcl 20,31,"L00000000003$spb"
+"L00000000003$spb":
+	mflr r11
+	addis r11,r11,ha16(L_darwin64_pass_struct_floats$lazy_ptr-"L00000000003$spb")
+	mtlr r0
+	lwzu r12,lo16(L_darwin64_pass_struct_floats$lazy_ptr-"L00000000003$spb")(r11)
+	mtctr r12
+	bctr
+	.lazy_symbol_pointer
+L_darwin64_pass_struct_floats$lazy_ptr:
+	.indirect_symbol _darwin64_pass_struct_floats
+	.long	dyld_stub_binding_helper
+#  endif
 #endif
-.data
-	.align LOG2_GPR_BYTES
-LLFB1$non_lazy_ptr:
-	.g_long LFB1

libffi/src/powerpc/ffitarget.h

@@ -1,6 +1,6 @@
 /* -----------------------------------------------------------------*-C-*-
    ffitarget.h - Copyright (c) 1996-2003  Red Hat, Inc.
-   Copyright (C) 2007, 2008 Free Software Foundation, Inc
+   Copyright (C) 2007, 2008, 2010 Free Software Foundation, Inc
    Target configuration macros for PowerPC.
 
    Permission is hereby granted, free of charge, to any person obtaining
@@ -34,10 +34,13 @@
 #ifndef POWERPC64
 #define POWERPC64
 #endif
-#elif defined (POWERPC_DARWIN) && defined (__ppc64__)	/* Darwin */
+#elif defined (POWERPC_DARWIN) && defined (__ppc64__)	/* Darwin64 */
 #ifndef POWERPC64
 #define POWERPC64
 #endif
+#ifndef POWERPC_DARWIN64
+#define POWERPC_DARWIN64
+#endif
 #elif defined (POWERPC_AIX) && defined (__64BIT__)	/* AIX64 */
 #ifndef POWERPC64
 #define POWERPC64
@@ -114,9 +117,13 @@ typedef enum ffi_abi {
 #define FFI_SYSV_TYPE_SMALL_STRUCT (FFI_TYPE_LAST + 2)
 
 #if defined(POWERPC64) || defined(POWERPC_AIX)
-#define FFI_TRAMPOLINE_SIZE 24
+#  if defined(POWERPC_DARWIN64)
+#    define FFI_TRAMPOLINE_SIZE 48
+#  else
+#    define FFI_TRAMPOLINE_SIZE 24
+#  endif
 #else /* POWERPC || POWERPC_AIX */
-#define FFI_TRAMPOLINE_SIZE 40
+#  define FFI_TRAMPOLINE_SIZE 40
 #endif
 
 #ifndef LIBFFI_ASM