ffi 1.0.0 → 1.17.1

data/ext/ffi_c/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
-.text
+	.globl _ffi_prep_args
+
 	.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)	
-	stw	r29,-12(r8)
-	stw	r30,-8(r8)
-	stw	r31,-4(r8)
+	sg	r28,-(4 * GPR_BYTES)(r8)	
+	sg	r29,-(3 * GPR_BYTES)(r8)
+	sg	r30,-(2 * GPR_BYTES)(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)
-	lwz     r4,20+(2*4)(r1)
-	lwz     r5,20+(3*4)(r1)
-	lwz     r6,20+(4*4)(r1)
+	lg     r3, (LINKAGE_SIZE                )(r1)
+	lg     r4, (LINKAGE_SIZE +     GPR_BYTES)(r1)
+	lg     r5, (LINKAGE_SIZE + 2 * GPR_BYTES)(r1)
+	lg     r6, (LINKAGE_SIZE + 3 * GPR_BYTES)(r1)
 	nop
-	lwz     r7,20+(5*4)(r1)
-	lwz     r8,20+(6*4)(r1)
-	lwz     r9,20+(7*4)(r1)
-	lwz     r10,20+(8*4)(r1)
+	lg     r7, (LINKAGE_SIZE + 4 * GPR_BYTES)(r1)
+	lg     r8, (LINKAGE_SIZE + 5 * GPR_BYTES)(r1)
+	lg     r9, (LINKAGE_SIZE + 6 * GPR_BYTES)(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	f2,-16-(12*8)(r28)
-	lfd	f3,-16-(11*8)(r28)
-	lfd	f4,-16-(10*8)(r28)
+	lfd	f1, -SAVE_REGS_SIZE-(13*FPR_SIZE)(r28)
+	lfd	f2, -SAVE_REGS_SIZE-(12*FPR_SIZE)(r28)
+	lfd	f3, -SAVE_REGS_SIZE-(11*FPR_SIZE)(r28)
+	lfd	f4, -SAVE_REGS_SIZE-(10*FPR_SIZE)(r28)
 	nop
-	lfd	f5,-16-(9*8)(r28)
-	lfd	f6,-16-(8*8)(r28)
-	lfd	f7,-16-(7*8)(r28)
-	lfd	f8,-16-(6*8)(r28)
+	lfd	f5, -SAVE_REGS_SIZE-( 9*FPR_SIZE)(r28)
+	lfd	f6, -SAVE_REGS_SIZE-( 8*FPR_SIZE)(r28)
+	lfd	f7, -SAVE_REGS_SIZE-( 7*FPR_SIZE)(r28)
+	lfd	f8, -SAVE_REGS_SIZE-( 6*FPR_SIZE)(r28)
 	nop
-	lfd     f9,-16-(5*8)(r28)
-	lfd     f10,-16-(4*8)(r28)
-	lfd     f11,-16-(3*8)(r28)
-	lfd     f12,-16-(2*8)(r28)
+	lfd     f9, -SAVE_REGS_SIZE-( 5*FPR_SIZE)(r28)
+	lfd     f10,-SAVE_REGS_SIZE-( 4*FPR_SIZE)(r28)
+	lfd     f11,-SAVE_REGS_SIZE-( 3*FPR_SIZE)(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)
-	bt	29,L(fp_return_value)
-	stw	r3,0(r30)
-	bf	28,L(done_return_value)
-	stw	r4,4(r30)
+	/* m64 structure returns can occupy the same set of registers as
+	   would be used to pass such a structure as arg0 - so take care 
+	   not to step on any possibly hot regs.  */
 
-	/* 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):
-	/* Restore the registers we used and return.  */
-	lwz	r9,8(r28)
-	lwz	r31,-4(r28)
-	mtlr	r9
-	lwz	r30,-8(r28)
-	lwz	r29,-12(r28)
-	lwz	r28,-16(r28)
-	lwz	r1,0(r1)
-	blr
+	/* OK, we have to map the return back to a mem struct.
+	   We are about to trample the parents param area, so recover the
+	   return type.  r29 is free, since the call is done.  */
+	lg	r29,(LINKAGE_SIZE + 6 * GPR_BYTES)(r28)
+
+	sg	r3, (LINKAGE_SIZE                )(r28)
+	sg	r4, (LINKAGE_SIZE +     GPR_BYTES)(r28)
+	sg	r5, (LINKAGE_SIZE + 2 * GPR_BYTES)(r28)
+	sg	r6, (LINKAGE_SIZE + 3 * GPR_BYTES)(r28)
+	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 +273,52 @@ 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
-	.align 2
-.globl _ffi_call_AIX
-.text
+	.text
+	.globl _ffi_call_AIX
 	.align 2
 _ffi_call_AIX:
 	blr
 /* END(_ffi_call_AIX)  */
 
-.data
-.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms
+/* EH stuff.  */
+
+#define EH_DATA_ALIGN_FACT MODE_CHOICE(0x7c,0x78)
+
+	.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support
 EH_frame1:
 	.set	L$set$0,LECIE1-LSCIE1
 	.long	L$set$0	; Length of Common Information Entry
@@ -193,28 +327,29 @@ LSCIE1:
 	.byte	0x1	; CIE Version
 	.ascii	"zR\0"	; CIE Augmentation
 	.byte	0x1	; uleb128 0x1; CIE Code Alignment Factor
-	.byte	0x7c	; sleb128 -4; CIE Data Alignment Factor
+	.byte	EH_DATA_ALIGN_FACT ; sleb128 -4; CIE Data Alignment Factor
 	.byte	0x41	; CIE RA Column
 	.byte	0x1	; uleb128 0x1; Augmentation size
-	.byte	0x90	; FDE Encoding (indirect pcrel)
+	.byte	0x10	; FDE Encoding (pcrel)
 	.byte	0xc	; DW_CFA_def_cfa
 	.byte	0x1	; uleb128 0x1
 	.byte	0x0	; uleb128 0x0
 	.align	LOG2_GPR_BYTES
 LECIE1:
-.globl _ffi_call_DARWIN.eh
+
+	.globl _ffi_call_DARWIN.eh
 _ffi_call_DARWIN.eh:
 LSFDE1:
 	.set	L$set$1,LEFDE1-LASFDE1
 	.long	L$set$1	; FDE Length
 LASFDE1:
 	.long	LASFDE1-EH_frame1 ; FDE CIE offset
-	.g_long	LLFB0$non_lazy_ptr-.	; FDE initial location
-	.set	L$set$3,LFE1-LFB0
+	.g_long	Lstartcode-.	; FDE initial location
+	.set	L$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$4
 	.byte	0xd	; DW_CFA_def_cfa_register
 	.byte	0x08	; uleb128 0x08
@@ -239,7 +374,5 @@ LASFDE1:
 	.byte	0x1c	; uleb128 0x1c
 	.align LOG2_GPR_BYTES
 LEFDE1:
-.data
-	.align LOG2_GPR_BYTES
-LLFB0$non_lazy_ptr:
-	.g_long LFB0
+	.align 1
+