Merge branch 'go-closure' of https://github.com/rth7680/libffi into rth7680-go-closure

Conflicts:
	src/aarch64/ffi.c

Makefile.am

@@ -36,9 +36,9 @@ EXTRA_DIST = LICENSE ChangeLog.v1 ChangeLog.libgcj			\
 	 src/sh/ffi.c src/sh/sysv.S src/sh/ffitarget.h src/sh64/ffi.c	\
 	 src/sh64/sysv.S src/sh64/ffitarget.h src/sparc/v8.S		\
 	 src/sparc/v9.S src/sparc/ffitarget.h src/sparc/ffi.c		\
-	 src/x86/darwin64.S src/x86/ffi.c src/x86/sysv.S		\
+	 src/sparc/ffi64.c src/x86/darwin64.S src/x86/ffi.c src/x86/sysv.S \
-	 src/x86/win32.S src/x86/darwin.S src/x86/win64.S		\
+	 src/x86/darwin.S src/x86/ffiw64.c src/x86/win64.S 		\
-	 src/x86/freebsd.S src/x86/ffi64.c src/x86/unix64.S		\
+	 src/x86/ffi64.c src/x86/unix64.S				\
 	 src/x86/ffitarget.h src/pa/ffitarget.h src/pa/ffi.c		\
 	 src/pa/linux.S src/pa/hpux32.S src/frv/ffi.c src/bfin/ffi.c	\
 	 src/bfin/ffitarget.h src/bfin/sysv.S src/frv/eabi.S		\
@@ -48,9 +48,9 @@ EXTRA_DIST = LICENSE ChangeLog.v1 ChangeLog.libgcj			\
 	 src/xtensa/ffitarget.h src/xtensa/ffi.c src/xtensa/sysv.S	\
 	 ChangeLog.libffi m4/libtool.m4 m4/lt~obsolete.m4		\
 	 m4/ltoptions.m4 m4/ltsugar.m4 m4/ltversion.m4			\
-	 m4/ltversion.m4 src/arm/gentramp.sh src/debug.c msvcc.sh	\
+	 m4/ltversion.m4 src/debug.c msvcc.sh				\
 	 generate-darwin-source-and-headers.py				\
-	 libffi.xcodeproj/project.pbxproj src/arm/trampoline.S		\
+	 libffi.xcodeproj/project.pbxproj				\
 	 libtool-ldflags ChangeLog.libffi-3.1
 
 info_TEXINFOS = doc/libffi.texi
@@ -126,25 +126,22 @@ if BFIN
 nodist_libffi_la_SOURCES += src/bfin/ffi.c src/bfin/sysv.S
 endif
 if X86
-nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/sysv.S src/x86/win32.S
+nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/sysv.S
 endif
 if X86_FREEBSD
-nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/freebsd.S src/x86/win32.S
+nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/sysv.S
 endif
 if X86_WIN32
-nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/win32.S
+nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/sysv.S
 endif
 if X86_WIN64
-nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/win64.S
+nodist_libffi_la_SOURCES += src/x86/ffiw64.c src/x86/win64.S
 endif
 if X86_DARWIN
 nodist_libffi_la_SOURCES += src/x86/ffi.c src/x86/darwin.S src/x86/ffi64.c src/x86/darwin64.S
-if X86_DARWIN32
-nodist_libffi_la_SOURCES += src/x86/win32.S
-endif
 endif
 if SPARC
-nodist_libffi_la_SOURCES += src/sparc/ffi.c src/sparc/v8.S src/sparc/v9.S
+nodist_libffi_la_SOURCES += src/sparc/ffi.c src/sparc/ffi64.c src/sparc/v8.S src/sparc/v9.S
 endif
 if ALPHA
 nodist_libffi_la_SOURCES += src/alpha/ffi.c src/alpha/osf.S
@@ -193,9 +190,6 @@ nodist_libffi_la_SOURCES += src/arc/arcompact.S src/arc/ffi.c
 endif
 if ARM
 nodist_libffi_la_SOURCES += src/arm/sysv.S src/arm/ffi.c
-if FFI_EXEC_TRAMPOLINE_TABLE
-nodist_libffi_la_SOURCES += src/arm/trampoline.S
-endif
 endif
 if AVR32
 nodist_libffi_la_SOURCES += src/avr32/sysv.S src/avr32/ffi.c

include/ffi.h.in

@@ -428,6 +428,22 @@ ffi_prep_java_raw_closure_loc (ffi_java_raw_closure*,
 
 #endif /* FFI_CLOSURES */
 
+#if FFI_GO_CLOSURES
+
+typedef struct {
+  void      *tramp;
+  ffi_cif   *cif;
+  void     (*fun)(ffi_cif*,void*,void**,void*);
+} ffi_go_closure;
+
+ffi_status ffi_prep_go_closure (ffi_go_closure*, ffi_cif *,
+				void (*fun)(ffi_cif*,void*,void**,void*));
+
+void ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
+		  void **avalue, void *closure);
+
+#endif /* FFI_GO_CLOSURES */
+
 /* ---- Public interface definition -------------------------------------- */
 
 ffi_status ffi_prep_cif(ffi_cif *cif,

include/ffi_cfi.h

@@ -0,0 +1,55 @@
+/* -----------------------------------------------------------------------
+   ffi_cfi.h - Copyright (c) 2014  Red Hat, Inc.
+
+   Conditionally assemble cfi directives. Only necessary for building libffi.
+   ----------------------------------------------------------------------- */
+
+#ifndef FFI_CFI_H
+#define FFI_CFI_H
+
+#ifdef HAVE_AS_CFI_PSEUDO_OP
+
+# define cfi_startproc			.cfi_startproc
+# define cfi_endproc			.cfi_endproc
+# define cfi_def_cfa(reg, off)		.cfi_def_cfa reg, off
+# define cfi_def_cfa_register(reg)	.cfi_def_cfa_register reg
+# define cfi_def_cfa_offset(off)	.cfi_def_cfa_offset off
+# define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+# define cfi_offset(reg, off)		.cfi_offset reg, off
+# define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+# define cfi_register(r1, r2)		.cfi_register r1, r2
+# define cfi_return_column(reg)		.cfi_return_column reg
+# define cfi_restore(reg)		.cfi_restore reg
+# define cfi_same_value(reg)		.cfi_same_value reg
+# define cfi_undefined(reg)		.cfi_undefined reg
+# define cfi_remember_state		.cfi_remember_state
+# define cfi_restore_state		.cfi_restore_state
+# define cfi_window_save		.cfi_window_save
+# define cfi_personality(enc, exp)	.cfi_personality enc, exp
+# define cfi_lsda(enc, exp)		.cfi_lsda enc, exp
+# define cfi_escape(...)		.cfi_escape __VA_ARGS__
+
+#else
+
+# define cfi_startproc
+# define cfi_endproc
+# define cfi_def_cfa(reg, off)
+# define cfi_def_cfa_register(reg)
+# define cfi_def_cfa_offset(off)
+# define cfi_adjust_cfa_offset(off)
+# define cfi_offset(reg, off)
+# define cfi_rel_offset(reg, off)
+# define cfi_register(r1, r2)
+# define cfi_return_column(reg)
+# define cfi_restore(reg)
+# define cfi_same_value(reg)
+# define cfi_undefined(reg)
+# define cfi_remember_state
+# define cfi_restore_state
+# define cfi_window_save
+# define cfi_personality(enc, exp)
+# define cfi_lsda(enc, exp)
+# define cfi_escape(...)
+
+#endif /* HAVE_AS_CFI_PSEUDO_OP */
+#endif /* FFI_CFI_H */

src/aarch64/ffitarget.h

@@ -42,22 +42,20 @@ typedef enum ffi_abi
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
-#define FFI_TRAMPOLINE_SIZE 36
+#define FFI_TRAMPOLINE_SIZE 24
 #define FFI_NATIVE_RAW_API 0
 
 /* ---- Internal ---- */
 
 #if defined (__APPLE__)
 #define FFI_TARGET_SPECIFIC_VARIADIC
-#define FFI_EXTRA_CIF_FIELDS unsigned aarch64_flags; unsigned aarch64_nfixedargs
+#define FFI_EXTRA_CIF_FIELDS unsigned aarch64_nfixedargs
 #else
-#define FFI_EXTRA_CIF_FIELDS unsigned aarch64_flags
+/* iOS reserves x18 for the system.  Disable Go closures until
+   a new static chain is chosen.  */
+#define FFI_GO_CLOSURES 1
 #endif
 
-#define AARCH64_FFI_WITH_V_BIT 0
+#define FFI_TARGET_HAS_COMPLEX_TYPE
-
-#define AARCH64_N_XREG 32
-#define AARCH64_N_VREG 32
-#define AARCH64_CALL_CONTEXT_SIZE (AARCH64_N_XREG * 8 + AARCH64_N_VREG * 16)
 
 #endif

src/aarch64/internal.h

@@ -0,0 +1,67 @@
+/* 
+Permission is hereby granted, free of charge, to any person obtaining
+a copy of this software and associated documentation files (the
+``Software''), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+The above copyright notice and this permission notice shall be
+included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+#define AARCH64_RET_VOID	0
+#define AARCH64_RET_INT64	1
+#define AARCH64_RET_INT128	2
+
+#define AARCH64_RET_UNUSED3	3
+#define AARCH64_RET_UNUSED4	4
+#define AARCH64_RET_UNUSED5	5
+#define AARCH64_RET_UNUSED6	6
+#define AARCH64_RET_UNUSED7	7
+
+/* Note that FFI_TYPE_FLOAT == 2, _DOUBLE == 3, _LONGDOUBLE == 4,
+   so _S4 through _Q1 are layed out as (TYPE * 4) + (4 - COUNT).  */
+#define AARCH64_RET_S4		8
+#define AARCH64_RET_S3		9
+#define AARCH64_RET_S2		10
+#define AARCH64_RET_S1		11
+
+#define AARCH64_RET_D4		12
+#define AARCH64_RET_D3		13
+#define AARCH64_RET_D2		14
+#define AARCH64_RET_D1		15
+
+#define AARCH64_RET_Q4		16
+#define AARCH64_RET_Q3		17
+#define AARCH64_RET_Q2		18
+#define AARCH64_RET_Q1		19
+
+/* Note that each of the sub-64-bit integers gets two entries.  */
+#define AARCH64_RET_UINT8	20
+#define AARCH64_RET_UINT16	22
+#define AARCH64_RET_UINT32	24
+
+#define AARCH64_RET_SINT8	26
+#define AARCH64_RET_SINT16	28
+#define AARCH64_RET_SINT32	30
+
+#define AARCH64_RET_MASK	31
+
+#define AARCH64_RET_IN_MEM	(1 << 5)
+#define AARCH64_RET_NEED_COPY	(1 << 6)
+
+#define AARCH64_FLAG_ARG_V_BIT	7
+#define AARCH64_FLAG_ARG_V	(1 << AARCH64_FLAG_ARG_V_BIT)
+
+#define N_X_ARG_REG		8
+#define N_V_ARG_REG		8
+#define CALL_CONTEXT_SIZE	(N_V_ARG_REG * 16 + N_X_ARG_REG * 8)

src/aarch64/sysv.S

@@ -22,6 +22,8 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
 #define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
+#include "internal.h"
 
 #ifdef HAVE_MACHINE_ASM_H
 #include <machine/asm.h>
@@ -37,297 +39,360 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
 #endif
 #endif
 
-#define cfi_adjust_cfa_offset(off)	.cfi_adjust_cfa_offset off
+#ifdef __AARCH64EB__
-#define cfi_rel_offset(reg, off)	.cfi_rel_offset reg, off
+# define BE(X)	X
-#define cfi_restore(reg)		.cfi_restore reg
+#else
-#define cfi_def_cfa_register(reg)	.cfi_def_cfa_register reg
+# define BE(X)	0
+#endif
 
 	.text
-        .globl CNAME(ffi_call_SYSV)
+	.align 4
-#ifdef __ELF__
-        .type CNAME(ffi_call_SYSV), #function
-#endif
-#ifdef __APPLE__
-        .align 2
-#endif
 
-/* ffi_call_SYSV()
+/* ffi_call_SYSV
-
+   extern void ffi_call_SYSV (void *stack, void *frame,
-   Create a stack frame, setup an argument context, call the callee
+			      void (*fn)(void), void *rvalue,
-   and extract the result.
+			      int flags, void *closure);
-
-   The maximum required argument stack size is provided,
-   ffi_call_SYSV() allocates that stack space then calls the
-   prepare_fn to populate register context and stack.  The
-   argument passing registers are loaded from the register
-   context and the callee called, on return the register passing
-   register are saved back to the context.  Our caller will
-   extract the return value from the final state of the saved
-   register context.
-
-   Prototype:
-
-   extern unsigned
-   ffi_call_SYSV (void (*)(struct call_context *context, unsigned char *,
-			   extended_cif *),
-                  struct call_context *context,
-                  extended_cif *,
-                  size_t required_stack_size,
-                  void (*fn)(void));
 
    Therefore on entry we have:
 
-   x0 prepare_fn
+   x0 stack
-   x1 &context
+   x1 frame
-   x2 &ecif
+   x2 fn
-   x3 bytes
+   x3 rvalue
-   x4 fn
+   x4 flags
+   x5 closure
+*/
 
-   This function uses the following stack frame layout:
+	cfi_startproc
-
-   ==
-                saved x30(lr)
-   x29(fp)->    saved x29(fp)
-                saved x24
-                saved x23
-                saved x22
-   sp'    ->    saved x21
-                ...
-   sp     ->    (constructed callee stack arguments)
-   ==
-
-   Voila! */
-
-#define ffi_call_SYSV_FS (8 * 4)
-
-        .cfi_startproc
 CNAME(ffi_call_SYSV):
-        stp     x29, x30, [sp, #-16]!
+	/* Use a stack frame allocated by our caller.  */
-	cfi_adjust_cfa_offset (16)
+	cfi_def_cfa(x1, 32);
+	stp	x29, x30, [x1]
+	mov	x29, x1
+	mov	sp, x0
+	cfi_def_cfa_register(x29)
 	cfi_rel_offset (x29, 0)
 	cfi_rel_offset (x30, 8)
 
-        mov     x29, sp
+	mov	x9, x2			/* save fn */
-	cfi_def_cfa_register (x29)
+	mov	x8, x3			/* install structure return */
-        sub     sp, sp, #ffi_call_SYSV_FS
+#ifdef FFI_GO_CLOSURES
+	mov	x18, x5			/* install static chain */
+#endif
+	stp	x3, x4, [x29, #16]	/* save rvalue and flags */
 
-        stp     x21, x22, [sp, #0]
+	/* Load the vector argument passing registers, if necessary.  */
-        cfi_rel_offset (x21, 0 - ffi_call_SYSV_FS)
+	tbz	w4, #AARCH64_FLAG_ARG_V_BIT, 1f
-        cfi_rel_offset (x22, 8 - ffi_call_SYSV_FS)
+	ldp     q0, q1, [sp, #0]
-
+	ldp     q2, q3, [sp, #32]
-        stp     x23, x24, [sp, #16]
+	ldp     q4, q5, [sp, #64]
-        cfi_rel_offset (x23, 16 - ffi_call_SYSV_FS)
+	ldp     q6, q7, [sp, #96]
-        cfi_rel_offset (x24, 24 - ffi_call_SYSV_FS)
-
-        mov     x21, x1
-        mov     x22, x2
-        mov     x24, x4
-
-        /* Allocate the stack space for the actual arguments, many
-           arguments will be passed in registers, but we assume
-           worst case and allocate sufficient stack for ALL of
-           the arguments.  */
-        sub     sp, sp, x3
-
-        /* unsigned (*prepare_fn) (struct call_context *context,
-				   unsigned char *stack, extended_cif *ecif);
-	 */
-        mov     x23, x0
-        mov     x0, x1
-        mov     x1, sp
-        /* x2 already in place */
-        blr     x23
-
-        /* Preserve the flags returned.  */
-        mov     x23, x0
-
-        /* Figure out if we should touch the vector registers.  */
-        tbz     x23, #AARCH64_FFI_WITH_V_BIT, 1f
-
-        /* Load the vector argument passing registers.  */
-        ldp     q0, q1, [x21, #8*32 +  0]
-        ldp     q2, q3, [x21, #8*32 + 32]
-        ldp     q4, q5, [x21, #8*32 + 64]
-        ldp     q6, q7, [x21, #8*32 + 96]
 1:
-        /* Load the core argument passing registers.  */
+	/* Load the core argument passing registers, including
-        ldp     x0, x1, [x21,  #0]
+	   the structure return pointer.  */
-        ldp     x2, x3, [x21, #16]
+	ldp     x0, x1, [sp, #16*N_V_ARG_REG + 0]
-        ldp     x4, x5, [x21, #32]
+	ldp     x2, x3, [sp, #16*N_V_ARG_REG + 16]
-        ldp     x6, x7, [x21, #48]
+	ldp     x4, x5, [sp, #16*N_V_ARG_REG + 32]
+	ldp     x6, x7, [sp, #16*N_V_ARG_REG + 48]
 
-        /* Don't forget x8 which may be holding the address of a return buffer.
+	/* Deallocate the context, leaving the stacked arguments.  */
-	 */
+	add	sp, sp, #CALL_CONTEXT_SIZE
-        ldr     x8,     [x21, #8*8]
 
-        blr     x24
+	blr     x9			/* call fn */
 
-        /* Save the core argument passing registers.  */
+	ldp	x3, x4, [x29, #16]	/* reload rvalue and flags */
-        stp     x0, x1, [x21,  #0]
-        stp     x2, x3, [x21, #16]
-        stp     x4, x5, [x21, #32]
-        stp     x6, x7, [x21, #48]
-
-        /* Note nothing useful ever comes back in x8!  */
-
-        /* Figure out if we should touch the vector registers.  */
-        tbz     x23, #AARCH64_FFI_WITH_V_BIT, 1f
-
-        /* Save the vector argument passing registers.  */
-        stp     q0, q1, [x21, #8*32 + 0]
-        stp     q2, q3, [x21, #8*32 + 32]
-        stp     q4, q5, [x21, #8*32 + 64]
-        stp     q6, q7, [x21, #8*32 + 96]
-1:
-        /* All done, unwind our stack frame.  */
-        ldp     x21, x22, [x29,  # - ffi_call_SYSV_FS]
-        cfi_restore (x21)
-        cfi_restore (x22)
-
-        ldp     x23, x24, [x29,  # - ffi_call_SYSV_FS + 16]
-        cfi_restore (x23)
-        cfi_restore (x24)
 
+	/* Partially deconstruct the stack frame.  */
 	mov     sp, x29
 	cfi_def_cfa_register (sp)
+	ldp     x29, x30, [x29]
 
-        ldp     x29, x30, [sp], #16
+	/* Save the return value as directed.  */
-	cfi_adjust_cfa_offset (-16)
+	adr	x5, 0f
-        cfi_restore (x29)
+	and	w4, w4, #AARCH64_RET_MASK
-        cfi_restore (x30)
+	add	x5, x5, x4, lsl #3
+	br	x5
 
+	/* Note that each table entry is 2 insns, and thus 8 bytes.
+	   For integer data, note that we're storing into ffi_arg
+	   and therefore we want to extend to 64 bits; these types
+	   have two consecutive entries allocated for them.  */
+	.align	4
+0:	ret				/* VOID */
+	nop
+1:	str	x0, [x3]		/* INT64 */
 	ret
+2:	stp	x0, x1, [x3]		/* INT128 */
+	ret
+3:	brk	#1000			/* UNUSED */
+	ret
+4:	brk	#1000			/* UNUSED */
+	ret
+5:	brk	#1000			/* UNUSED */
+	ret
+6:	brk	#1000			/* UNUSED */
+	ret
+7:	brk	#1000			/* UNUSED */
+	ret
+8:	st4	{ v0.s-v3.s }[0], [x3]	/* S4 */
+	ret
+9:	st3	{ v0.s-v2.s }[0], [x3]	/* S3 */
+	ret
+10:	stp	s0, s1, [x3]		/* S2 */
+	ret
+11:	str	s0, [x3]		/* S1 */
+	ret
+12:	st4	{ v0.d-v3.d }[0], [x3]	/* D4 */
+	ret
+13:	st3	{ v0.d-v2.d }[0], [x3]	/* D3 */
+	ret
+14:	stp	d0, d1, [x3]		/* D2 */
+	ret
+15:	str	d0, [x3]		/* D1 */
+	ret
+16:	str	q3, [x3, #48]		/* Q4 */
+	nop
+17:	str	q2, [x3, #32]		/* Q3 */
+	nop
+18:	stp	q0, q1, [x3]		/* Q2 */
+	ret
+19:	str	q0, [x3]		/* Q1 */
+	ret
+20:	uxtb	w0, w0			/* UINT8 */
+	str	x0, [x3]
+21:	ret				/* reserved */
+	nop
+22:	uxth	w0, w0			/* UINT16 */
+	str	x0, [x3]
+23:	ret				/* reserved */
+	nop
+24:	mov	w0, w0			/* UINT32 */
+	str	x0, [x3]
+25:	ret				/* reserved */
+	nop
+26:	sxtb	x0, w0			/* SINT8 */
+	str	x0, [x3]
+27:	ret				/* reserved */
+	nop
+28:	sxth	x0, w0			/* SINT16 */
+	str	x0, [x3]
+29:	ret				/* reserved */
+	nop
+30:	sxtw	x0, w0			/* SINT32 */
+	str	x0, [x3]
+31:	ret				/* reserved */
+	nop
 
-        .cfi_endproc
+	cfi_endproc
+
+	.globl	CNAME(ffi_call_SYSV)
 #ifdef __ELF__
+	.type	CNAME(ffi_call_SYSV), #function
+	.hidden	CNAME(ffi_call_SYSV)
 	.size CNAME(ffi_call_SYSV), .-CNAME(ffi_call_SYSV)
 #endif
 
-#define ffi_closure_SYSV_FS (8 * 2 + AARCH64_CALL_CONTEXT_SIZE)
-
 /* ffi_closure_SYSV
 
    Closure invocation glue. This is the low level code invoked directly by
    the closure trampoline to setup and call a closure.
 
-   On entry x17 points to a struct trampoline_data, x16 has been clobbered
+   On entry x17 points to a struct ffi_closure, x16 has been clobbered
    all other registers are preserved.
 
    We allocate a call context and save the argument passing registers,
    then invoked the generic C ffi_closure_SYSV_inner() function to do all
    the real work, on return we load the result passing registers back from
    the call context.
+*/
 
-   On entry
+#define ffi_closure_SYSV_FS (8*2 + CALL_CONTEXT_SIZE + 64)
 
-   extern void
+	.align 4
-   ffi_closure_SYSV (struct trampoline_data *);
+CNAME(ffi_closure_SYSV_V):
-
+	cfi_startproc
-   struct trampoline_data
+	stp     x29, x30, [sp, #-ffi_closure_SYSV_FS]!
-   {
+	cfi_adjust_cfa_offset (ffi_closure_SYSV_FS)
-        UINT64 *ffi_closure;
-        UINT64 flags;
-   };
-
-   This function uses the following stack frame layout:
-
-   ==
-                saved x30(lr)
-   x29(fp)->    saved x29(fp)
-                saved x22
-                saved x21
-                ...
-   sp     ->    call_context
-   ==
-
-   Voila!  */
-
-        .text
-        .globl CNAME(ffi_closure_SYSV)
-#ifdef __APPLE__
-        .align 2
-#endif
-        .cfi_startproc
-CNAME(ffi_closure_SYSV):
-        stp     x29, x30, [sp, #-16]!
-	cfi_adjust_cfa_offset (16)
 	cfi_rel_offset (x29, 0)
 	cfi_rel_offset (x30, 8)
 
-        mov     x29, sp
-        cfi_def_cfa_register (x29)
-
-        sub     sp, sp, #ffi_closure_SYSV_FS
-
-        stp     x21, x22, [x29, #-16]
-        cfi_rel_offset (x21, -16)
-        cfi_rel_offset (x22, -8)
-
-        /* Load x21 with &call_context.  */
-        mov     x21, sp
-        /* Preserve our struct trampoline_data *  */
-        mov     x22, x17
-
-        /* Save the rest of the argument passing registers.  */
-        stp     x0, x1, [x21, #0]
-        stp     x2, x3, [x21, #16]
-        stp     x4, x5, [x21, #32]
-        stp     x6, x7, [x21, #48]
-        /* Don't forget we may have been given a result scratch pad address.
-	 */
-        str     x8,     [x21, #64]
-
-        /* Figure out if we should touch the vector registers.  */
-        ldr     x0, [x22, #8]
-        tbz     x0, #AARCH64_FFI_WITH_V_BIT, 1f
-
 	/* Save the argument passing vector registers.  */
-        stp     q0, q1, [x21, #8*32 + 0]
+	stp     q0, q1, [sp, #16 + 0]
-        stp     q2, q3, [x21, #8*32 + 32]
+	stp     q2, q3, [sp, #16 + 32]
-        stp     q4, q5, [x21, #8*32 + 64]
+	stp     q4, q5, [sp, #16 + 64]
-        stp     q6, q7, [x21, #8*32 + 96]
+	stp     q6, q7, [sp, #16 + 96]
-1:
+	b	0f
-        /* Load &ffi_closure..  */
+	cfi_endproc
-        ldr     x0, [x22, #0]
-        mov     x1, x21
-        /* Compute the location of the stack at the point that the
-           trampoline was called.  */
-        add     x2, x29, #16
 
+	.globl	CNAME(ffi_closure_SYSV_V)
+#ifdef __ELF__
+	.type	CNAME(ffi_closure_SYSV_V), #function
+	.hidden	CNAME(ffi_closure_SYSV_V)
+	.size	CNAME(ffi_closure_SYSV_V), . - CNAME(ffi_closure_SYSV_V)
+#endif
+
+	.align	4
+	cfi_startproc
+CNAME(ffi_closure_SYSV):
+	stp     x29, x30, [sp, #-ffi_closure_SYSV_FS]!
+	cfi_adjust_cfa_offset (ffi_closure_SYSV_FS)
+	cfi_rel_offset (x29, 0)
+	cfi_rel_offset (x30, 8)
+0:
+	mov     x29, sp
+
+	/* Save the argument passing core registers.  */
+	stp     x0, x1, [sp, #16 + 16*N_V_ARG_REG + 0]
+	stp     x2, x3, [sp, #16 + 16*N_V_ARG_REG + 16]
+	stp     x4, x5, [sp, #16 + 16*N_V_ARG_REG + 32]
+	stp     x6, x7, [sp, #16 + 16*N_V_ARG_REG + 48]
+
+	/* Load ffi_closure_inner arguments.  */
+	ldp	x0, x1, [x17, #FFI_TRAMPOLINE_SIZE]	/* load cif, fn */
+	ldr	x2, [x17, #FFI_TRAMPOLINE_SIZE+16]	/* load user_data */
+.Ldo_closure:
+	add	x3, sp, #16				/* load context */
+	add	x4, sp, #ffi_closure_SYSV_FS		/* load stack */
+	add	x5, sp, #16+CALL_CONTEXT_SIZE		/* load rvalue */
+	mov	x6, x8					/* load struct_rval */
 	bl      CNAME(ffi_closure_SYSV_inner)
 
-        /* Figure out if we should touch the vector registers.  */
+	/* Load the return value as directed.  */
-        ldr     x0, [x22, #8]
+	adr	x1, 0f
-        tbz     x0, #AARCH64_FFI_WITH_V_BIT, 1f
+	and	w0, w0, #AARCH64_RET_MASK
+	add	x1, x1, x0, lsl #3
+	add	x3, sp, #16+CALL_CONTEXT_SIZE
+	br	x1
 
-        /* Load the result passing vector registers.  */
+	/* Note that each table entry is 2 insns, and thus 8 bytes.  */
-        ldp     q0, q1, [x21, #8*32 + 0]
+	.align	4
-        ldp     q2, q3, [x21, #8*32 + 32]
+0:	b	99f			/* VOID */
-        ldp     q4, q5, [x21, #8*32 + 64]
+	nop
-        ldp     q6, q7, [x21, #8*32 + 96]
+1:	ldr	x0, [x3]		/* INT64 */
-1:
+	b	99f
-        /* Load the result passing core registers.  */
+2:	ldp	x0, x1, [x3]		/* INT128 */
-        ldp     x0, x1, [x21,  #0]
+	b	99f
-        ldp     x2, x3, [x21, #16]
+3:	brk	#1000			/* UNUSED */
-        ldp     x4, x5, [x21, #32]
+	nop
-        ldp     x6, x7, [x21, #48]
+4:	brk	#1000			/* UNUSED */
-        /* Note nothing useful is returned in x8.  */
+	nop
-
+5:	brk	#1000			/* UNUSED */
-        /* We are done, unwind our frame.  */
+	nop
-        ldp     x21, x22, [x29,  #-16]
+6:	brk	#1000			/* UNUSED */
-        cfi_restore (x21)
+	nop
-        cfi_restore (x22)
+7:	brk	#1000			/* UNUSED */
-
+	nop
-        mov     sp, x29
+8:	ldr	s3, [x3, #12]		/* S4 */
-        cfi_def_cfa_register (sp)
+	nop
-
+9:	ldr	s2, [x2, #8]		/* S3 */
-        ldp     x29, x30, [sp], #16
+	nop
-	cfi_adjust_cfa_offset (-16)
+10:	ldp	s0, s1, [x3]		/* S2 */
+	b	99f
+11:	ldr	s0, [x3]		/* S1 */
+	b	99f
+12:	ldr	d3, [x3, #24]		/* D4 */
+	nop
+13:	ldr	d2, [x3, #16]		/* D3 */
+	nop
+14:	ldp	d0, d1, [x3]		/* D2 */
+	b	99f
+15:	ldr	d0, [x3]		/* D1 */
+	b	99f
+16:	ldr	q3, [x3, #48]		/* Q4 */
+	nop
+17:	ldr	q2, [x3, #32]		/* Q3 */
+	nop
+18:	ldp	q0, q1, [x3]		/* Q2 */
+	b	99f
+19:	ldr	q0, [x3]		/* Q1 */
+	b	99f
+20:	ldrb	w0, [x3, #BE(7)]	/* UINT8 */
+	b	99f
+21:	brk	#1000			/* reserved */
+	nop
+22:	ldrh	w0, [x3, #BE(6)]	/* UINT16 */
+	b	99f
+23:	brk	#1000			/* reserved */
+	nop
+24:	ldr	w0, [x3, #BE(4)]	/* UINT32 */
+	b	99f
+25:	brk	#1000			/* reserved */
+	nop
+26:	ldrsb	x0, [x3, #BE(7)]	/* SINT8 */
+	b	99f
+27:	brk	#1000			/* reserved */
+	nop
+28:	ldrsh	x0, [x3, #BE(6)]	/* SINT16 */
+	b	99f
+29:	brk	#1000			/* reserved */
+	nop
+30:	ldrsw	x0, [x3, #BE(4)]	/* SINT32 */
+	nop
+31:					/* reserved */
+99:	ldp     x29, x30, [sp], #ffi_closure_SYSV_FS
+	cfi_adjust_cfa_offset (-ffi_closure_SYSV_FS)
 	cfi_restore (x29)
 	cfi_restore (x30)
-
 	ret
-        .cfi_endproc
+	cfi_endproc
+
+	.globl	CNAME(ffi_closure_SYSV)
 #ifdef __ELF__
+	.type	CNAME(ffi_closure_SYSV), #function
+	.hidden	CNAME(ffi_closure_SYSV)
 	.size	CNAME(ffi_closure_SYSV), . - CNAME(ffi_closure_SYSV)
 #endif
+
+#ifdef FFI_GO_CLOSURES
+	.align 4
+CNAME(ffi_go_closure_SYSV_V):
+	cfi_startproc
+	stp     x29, x30, [sp, #-ffi_closure_SYSV_FS]!
+	cfi_adjust_cfa_offset (ffi_closure_SYSV_FS)
+	cfi_rel_offset (x29, 0)
+	cfi_rel_offset (x30, 8)
+
+	/* Save the argument passing vector registers.  */
+	stp     q0, q1, [sp, #16 + 0]
+	stp     q2, q3, [sp, #16 + 32]
+	stp     q4, q5, [sp, #16 + 64]
+	stp     q6, q7, [sp, #16 + 96]
+	b	0f
+	cfi_endproc
+
+	.globl	CNAME(ffi_go_closure_SYSV_V)
+#ifdef __ELF__
+	.type	CNAME(ffi_go_closure_SYSV_V), #function
+	.hidden	CNAME(ffi_go_closure_SYSV_V)
+	.size	CNAME(ffi_go_closure_SYSV_V), . - CNAME(ffi_go_closure_SYSV_V)
+#endif
+
+	.align	4
+	cfi_startproc
+CNAME(ffi_go_closure_SYSV):
+	stp     x29, x30, [sp, #-ffi_closure_SYSV_FS]!
+	cfi_adjust_cfa_offset (ffi_closure_SYSV_FS)
+	cfi_rel_offset (x29, 0)
+	cfi_rel_offset (x30, 8)
+0:
+	mov     x29, sp
+
+	/* Save the argument passing core registers.  */
+	stp     x0, x1, [sp, #16 + 16*N_V_ARG_REG + 0]
+	stp     x2, x3, [sp, #16 + 16*N_V_ARG_REG + 16]
+	stp     x4, x5, [sp, #16 + 16*N_V_ARG_REG + 32]
+	stp     x6, x7, [sp, #16 + 16*N_V_ARG_REG + 48]
+
+	/* Load ffi_closure_inner arguments.  */
+	ldp	x0, x1, [x18, #8]			/* load cif, fn */
+	mov	x2, x18					/* load user_data */
+	b	.Ldo_closure
+	cfi_endproc
+
+	.globl	CNAME(ffi_go_closure_SYSV)
+#ifdef __ELF__
+	.type	CNAME(ffi_go_closure_SYSV), #function
+	.hidden	CNAME(ffi_go_closure_SYSV)
+	.size	CNAME(ffi_go_closure_SYSV), . - CNAME(ffi_go_closure_SYSV)
+#endif
+#endif /* FFI_GO_CLOSURES */

src/alpha/ffi.c

@@ -28,6 +28,7 @@
 #include <ffi.h>
 #include <ffi_common.h>
 #include <stdlib.h>
+#include "internal.h"
 
 /* Force FFI_TYPE_LONGDOUBLE to be different than FFI_TYPE_DOUBLE;
    all further uses in this file will refer to the 128-bit type.  */
@@ -40,135 +41,286 @@
 # define FFI_TYPE_LONGDOUBLE 4
 #endif
 
-extern void ffi_call_osf(void *, unsigned long, unsigned, void *, void (*)(void))
+extern void ffi_call_osf(void *stack, void *frame, unsigned flags,
+			 void *raddr, void (*fn)(void), void *closure)
 	FFI_HIDDEN;
 extern void ffi_closure_osf(void) FFI_HIDDEN;
+extern void ffi_go_closure_osf(void) FFI_HIDDEN;
 
+/* Promote a float value to its in-register double representation.
+   Unlike actually casting to double, this does not trap on NaN.  */
+static inline UINT64 lds(void *ptr)
+{
+  UINT64 ret;
+  asm("lds %0,%1" : "=f"(ret) : "m"(*(UINT32 *)ptr));
+  return ret;
+}
 
-ffi_status
+/* And the reverse.  */
+static inline void sts(void *ptr, UINT64 val)
+{
+  asm("sts %1,%0" : "=m"(*(UINT32 *)ptr) : "f"(val));
+}
+
+ffi_status FFI_HIDDEN
 ffi_prep_cif_machdep(ffi_cif *cif)
 {
-  /* Adjust cif->bytes to represent a minimum 6 words for the temporary
+  size_t bytes = 0;
-     register argument loading area.  */
+  int flags, i, avn;
-  if (cif->bytes < 6*FFI_SIZEOF_ARG)
+  ffi_type *rtype, *itype;
-    cif->bytes = 6*FFI_SIZEOF_ARG;
 
-  /* Set the return type flag */
+  if (cif->abi != FFI_OSF)
-  switch (cif->rtype->type)
+    return FFI_BAD_ABI;
+
+  /* Compute the size of the argument area.  */
+  for (i = 0, avn = cif->nargs; i < avn; i++)
     {
-    case FFI_TYPE_STRUCT:
+      itype = cif->arg_types[i];
+      switch (itype->type)
+	{
+	case FFI_TYPE_INT:
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_POINTER:
 	case FFI_TYPE_FLOAT:
 	case FFI_TYPE_DOUBLE:
-      cif->flags = cif->rtype->type;
+	case FFI_TYPE_LONGDOUBLE:
+	  /* All take one 8 byte slot.  */
+	  bytes += 8;
 	  break;
 
-    case FFI_TYPE_LONGDOUBLE:
+	case FFI_TYPE_VOID:
-      /* 128-bit long double is returned in memory, like a struct.  */
+	case FFI_TYPE_STRUCT:
-      cif->flags = FFI_TYPE_STRUCT;
+	  /* Passed by value in N slots.  */
+	  bytes += ALIGN(itype->size, FFI_SIZEOF_ARG);
+	  break;
+
+	case FFI_TYPE_COMPLEX:
+	  /* _Complex long double passed by reference; others in 2 slots.  */
+	  if (itype->elements[0]->type == FFI_TYPE_LONGDOUBLE)
+	    bytes += 8;
+	  else
+	    bytes += 16;
 	  break;
 
 	default:
-      cif->flags = FFI_TYPE_INT;
+	  abort();
+	}
+    }
+
+  /* Set the return type flag */
+  rtype = cif->rtype;
+  switch (rtype->type)
+    {
+    case FFI_TYPE_VOID:
+      flags = ALPHA_FLAGS(ALPHA_ST_VOID, ALPHA_LD_VOID);
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_SINT32:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_INT32);
+      break;
+    case FFI_TYPE_FLOAT:
+      flags = ALPHA_FLAGS(ALPHA_ST_FLOAT, ALPHA_LD_FLOAT);
+      break;
+    case FFI_TYPE_DOUBLE:
+      flags = ALPHA_FLAGS(ALPHA_ST_DOUBLE, ALPHA_LD_DOUBLE);
+      break;
+    case FFI_TYPE_UINT8:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_UINT8);
+      break;
+    case FFI_TYPE_SINT8:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_SINT8);
+      break;
+    case FFI_TYPE_UINT16:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_UINT16);
+      break;
+    case FFI_TYPE_SINT16:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_SINT16);
+      break;
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+    case FFI_TYPE_POINTER:
+      flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_INT64);
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+    case FFI_TYPE_STRUCT:
+      /* Passed in memory, with a hidden pointer.  */
+      flags = ALPHA_RET_IN_MEM;
+      break;
+    case FFI_TYPE_COMPLEX:
+      itype = rtype->elements[0];
+      switch (itype->type)
+	{
+	case FFI_TYPE_FLOAT:
+	  flags = ALPHA_FLAGS(ALPHA_ST_CPLXF, ALPHA_LD_CPLXF);
+	  break;
+	case FFI_TYPE_DOUBLE:
+	  flags = ALPHA_FLAGS(ALPHA_ST_CPLXD, ALPHA_LD_CPLXD);
+	  break;
+	default:
+	  if (rtype->size <= 8)
+	    flags = ALPHA_FLAGS(ALPHA_ST_INT, ALPHA_LD_INT64);
+	  else
+	    flags = ALPHA_RET_IN_MEM;
 	  break;
 	}
+      break;
+    default:
+      abort();
+    }
+  cif->flags = flags;
+
+  /* Include the hidden structure pointer in args requirement.  */
+  if (flags == ALPHA_RET_IN_MEM)
+    bytes += 8;
+  /* Minimum size is 6 slots, so that ffi_call_osf can pop them.  */
+  if (bytes < 6*8)
+    bytes = 6*8;
+  cif->bytes = bytes;
 
   return FFI_OK;
 }
 
-
+static unsigned long
-void
+extend_basic_type(void *valp, int type, int argn)
-ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 {
-  unsigned long *stack, *argp;
+  switch (type)
-  long i, avn;
-  ffi_type **arg_types;
-  
-  /* If the return value is a struct and we don't have a return
-     value address then we need to make one.  */
-  if (rvalue == NULL && cif->flags == FFI_TYPE_STRUCT)
-    rvalue = alloca(cif->rtype->size);
-
-  /* Allocate the space for the arguments, plus 4 words of temp
-     space for ffi_call_osf.  */
-  argp = stack = alloca(cif->bytes + 4*FFI_SIZEOF_ARG);
-
-  if (cif->flags == FFI_TYPE_STRUCT)
-    *(void **) argp++ = rvalue;
-
-  i = 0;
-  avn = cif->nargs;
-  arg_types = cif->arg_types;
-
-  while (i < avn)
-    {
-      size_t size = (*arg_types)->size;
-
-      switch ((*arg_types)->type)
     {
     case FFI_TYPE_SINT8:
-	  *(SINT64 *) argp = *(SINT8 *)(* avalue);
+      return *(SINT8 *)valp;
-	  break;
-		  
     case FFI_TYPE_UINT8:
-	  *(SINT64 *) argp = *(UINT8 *)(* avalue);
+      return *(UINT8 *)valp;
-	  break;
-		  
     case FFI_TYPE_SINT16:
-	  *(SINT64 *) argp = *(SINT16 *)(* avalue);
+      return *(SINT16 *)valp;
-	  break;
-		  
     case FFI_TYPE_UINT16:
-	  *(SINT64 *) argp = *(UINT16 *)(* avalue);
+      return *(UINT16 *)valp;
-	  break;
 
+    case FFI_TYPE_FLOAT:
+      if (argn < 6)
+	return lds(valp);
+      /* FALLTHRU */
+
+    case FFI_TYPE_INT:
     case FFI_TYPE_SINT32:
     case FFI_TYPE_UINT32:
       /* Note that unsigned 32-bit quantities are sign extended.  */
-	  *(SINT64 *) argp = *(SINT32 *)(* avalue);
+      return *(SINT32 *)valp;
-	  break;
 
     case FFI_TYPE_SINT64:
     case FFI_TYPE_UINT64:
     case FFI_TYPE_POINTER:
-	  *(UINT64 *) argp = *(UINT64 *)(* avalue);
-	  break;
-
-	case FFI_TYPE_FLOAT:
-	  if (argp - stack < 6)
-	    {
-	      /* Note the conversion -- all the fp regs are loaded as
-		 doubles.  The in-register format is the same.  */
-	      *(double *) argp = *(float *)(* avalue);
-	    }
-	  else
-	    *(float *) argp = *(float *)(* avalue);
-	  break;
-
     case FFI_TYPE_DOUBLE:
-	  *(double *) argp = *(double *)(* avalue);
+      return *(UINT64 *)valp;
+
+    default:
+      abort();
+    }
+}
+
+static void
+ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	      void **avalue, void *closure)
+{
+  unsigned long *argp;
+  long i, avn, argn, flags = cif->flags;
+  ffi_type **arg_types;
+  void *frame;
+
+  /* If the return value is a struct and we don't have a return
+     value address then we need to make one.  */
+  if (rvalue == NULL && flags == ALPHA_RET_IN_MEM)
+    rvalue = alloca(cif->rtype->size);
+
+  /* Allocate the space for the arguments, plus 4 words of temp
+     space for ffi_call_osf.  */
+  argp = frame = alloca(cif->bytes + 4*FFI_SIZEOF_ARG);
+  frame += cif->bytes;
+
+  argn = 0;
+  if (flags == ALPHA_RET_IN_MEM)
+    argp[argn++] = (unsigned long)rvalue;
+
+  avn = cif->nargs;
+  arg_types = cif->arg_types;
+
+  for (i = 0, avn = cif->nargs; i < avn; i++)
+    {
+      ffi_type *ty = arg_types[i];
+      void *valp = avalue[i];
+      int type = ty->type;
+      size_t size;
+
+      switch (type)
+	{
+	case FFI_TYPE_INT:
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_POINTER:
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_DOUBLE:
+	  argp[argn] = extend_basic_type(valp, type, argn);
+	  argn++;
 	  break;
 
 	case FFI_TYPE_LONGDOUBLE:
-	  /* 128-bit long double is passed by reference.  */
+	by_reference:
-	  *(long double **) argp = (long double *)(* avalue);
+	  /* Note that 128-bit long double is passed by reference.  */
-	  size = sizeof (long double *);
+	  argp[argn++] = (unsigned long)valp;
 	  break;
 
+	case FFI_TYPE_VOID:
 	case FFI_TYPE_STRUCT:
-	  memcpy(argp, *avalue, (*arg_types)->size);
+	  size = ty->size;
+	  memcpy(argp + argn, valp, size);
+	  argn += ALIGN(size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
+	  break;
+
+	case FFI_TYPE_COMPLEX:
+	  type = ty->elements[0]->type;
+	  if (type == FFI_TYPE_LONGDOUBLE)
+	    goto by_reference;
+
+	  /* Most complex types passed as two separate arguments.  */
+	  size = ty->elements[0]->size;
+	  argp[argn] = extend_basic_type(valp, type, argn);
+	  argp[argn + 1] = extend_basic_type(valp + size, type, argn + 1);
+	  argn += 2;
 	  break;
 
 	default:
-	  FFI_ASSERT(0);
+	  abort();
+	}
     }
 
-      argp += ALIGN(size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
+  flags = (flags >> ALPHA_ST_SHIFT) & 0xff;
-      i++, arg_types++, avalue++;
+  ffi_call_osf(argp, frame, flags, rvalue, fn, closure);
 }
 
-  ffi_call_osf(stack, cif->bytes, cif->flags, rvalue, fn);
+void
+ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, NULL);
 }
 
+void
+ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	     void **avalue, void *closure)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, closure);
+}
 
 ffi_status
 ffi_prep_closure_loc (ffi_closure* closure,
@@ -204,39 +356,56 @@ ffi_prep_closure_loc (ffi_closure* closure,
   return FFI_OK;
 }
 
+ffi_status
+ffi_prep_go_closure (ffi_go_closure* closure,
+		     ffi_cif* cif,
+		     void (*fun)(ffi_cif*, void*, void**, void*))
+{
+  if (cif->abi != FFI_OSF)
+    return FFI_BAD_ABI;
+
+  closure->tramp = (void *)ffi_go_closure_osf;
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
+}
 
 long FFI_HIDDEN
-ffi_closure_osf_inner(ffi_closure *closure, void *rvalue, unsigned long *argp)
+ffi_closure_osf_inner (ffi_cif *cif,
+		       void (*fun)(ffi_cif*, void*, void**, void*),
+		       void *user_data,
+		       void *rvalue, unsigned long *argp)
 {
-  ffi_cif *cif;
   void **avalue;
   ffi_type **arg_types;
-  long i, avn, argn;
+  long i, avn, argn, flags;
 
-  cif = closure->cif;
   avalue = alloca(cif->nargs * sizeof(void *));
-
+  flags = cif->flags;
   argn = 0;
 
   /* Copy the caller's structure return address to that the closure
      returns the data directly to the caller.  */
-  if (cif->flags == FFI_TYPE_STRUCT)
+  if (flags == ALPHA_RET_IN_MEM)
     {
       rvalue = (void *) argp[0];
       argn = 1;
     }
 
-  i = 0;
-  avn = cif->nargs;
   arg_types = cif->arg_types;
 
   /* Grab the addresses of the arguments from the stack frame.  */
-  while (i < avn)
+  for (i = 0, avn = cif->nargs; i < avn; i++)
     {
-      size_t size = arg_types[i]->size;
+      ffi_type *ty = arg_types[i];
+      int type = ty->type;
+      void *valp = &argp[argn];
+      size_t size;
 
-      switch (arg_types[i]->type)
+      switch (type)
 	{
+	case FFI_TYPE_INT:
 	case FFI_TYPE_SINT8:
 	case FFI_TYPE_UINT8:
 	case FFI_TYPE_SINT16:
@@ -246,43 +415,107 @@ ffi_closure_osf_inner(ffi_closure *closure, void *rvalue, unsigned long *argp)
 	case FFI_TYPE_SINT64:
 	case FFI_TYPE_UINT64:
 	case FFI_TYPE_POINTER:
+	  argn += 1;
+	  break;
+
+	case FFI_TYPE_VOID:
 	case FFI_TYPE_STRUCT:
-	  avalue[i] = &argp[argn];
+	  size = ty->size;
+	  argn += ALIGN(size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
 	  break;
 
 	case FFI_TYPE_FLOAT:
-	  if (argn < 6)
-	    {
 	  /* Floats coming from registers need conversion from double
 	     back to float format.  */
-	      *(float *)&argp[argn - 6] = *(double *)&argp[argn - 6];
+	  if (argn < 6)
-	      avalue[i] = &argp[argn - 6];
+	    {
+	      valp = &argp[argn - 6];
+	      sts(valp, argp[argn - 6]);
 	    }
-	  else
+	  argn += 1;
-	    avalue[i] = &argp[argn];
 	  break;
 
 	case FFI_TYPE_DOUBLE:
-	  avalue[i] = &argp[argn - (argn < 6 ? 6 : 0)];
+	  if (argn < 6)
+	    valp = &argp[argn - 6];
+	  argn += 1;
 	  break;
 
 	case FFI_TYPE_LONGDOUBLE:
+	by_reference:
 	  /* 128-bit long double is passed by reference.  */
-	  avalue[i] = (long double *) argp[argn];
+	  valp = (void *)argp[argn];
-	  size = sizeof (long double *);
+	  argn += 1;
+	  break;
+
+	case FFI_TYPE_COMPLEX:
+	  type = ty->elements[0]->type;
+	  switch (type)
+	    {
+	    case FFI_TYPE_SINT64:
+	    case FFI_TYPE_UINT64:
+	      /* Passed as separate arguments, but they wind up sequential.  */
+	      break;
+
+	    case FFI_TYPE_INT:
+	    case FFI_TYPE_SINT8:
+	    case FFI_TYPE_UINT8:
+	    case FFI_TYPE_SINT16:
+	    case FFI_TYPE_UINT16:
+	    case FFI_TYPE_SINT32:
+	    case FFI_TYPE_UINT32:
+	      /* Passed as separate arguments.  Disjoint, but there's room
+		 enough in one slot to hold the pair.  */
+	      size = ty->elements[0]->size;
+	      memcpy(valp + size, valp + 8, size);
+	      break;
+
+	    case FFI_TYPE_FLOAT:
+	      /* Passed as separate arguments.  Disjoint, and each piece
+		 may need conversion back to float.  */
+	      if (argn < 6)
+		{
+		  valp = &argp[argn - 6];
+		  sts(valp, argp[argn - 6]);
+		}
+	      if (argn + 1 < 6)
+		sts(valp + 4, argp[argn + 1 - 6]);
+	      else
+		*(UINT32 *)(valp + 4) = argp[argn + 1];
+	      break;
+
+	    case FFI_TYPE_DOUBLE:
+	      /* Passed as separate arguments.  Only disjoint if one part
+		 is in fp regs and the other is on the stack.  */
+	      if (argn < 5)
+		valp = &argp[argn - 6];
+	      else if (argn == 5)
+		{
+		  valp = alloca(16);
+		  ((UINT64 *)valp)[0] = argp[5 - 6];
+		  ((UINT64 *)valp)[1] = argp[6];
+		}
+	      break;
+
+	    case FFI_TYPE_LONGDOUBLE:
+	      goto by_reference;
+
+	    default:
+	      abort();
+	    }
+	  argn += 2;
 	  break;
 
 	default:
 	  abort ();
 	}
 
-      argn += ALIGN(size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
+      avalue[i] = valp;
-      i++;
     }
 
   /* Invoke the closure.  */
-  closure->fun (cif, rvalue, avalue, closure->user_data);
+  fun (cif, rvalue, avalue, user_data);
 
   /* Tell ffi_closure_osf how to perform return type promotions.  */
-  return cif->rtype->type;
+  return (flags >> ALPHA_LD_SHIFT) & 0xff;
 }

src/alpha/ffitarget.h

@@ -44,9 +44,13 @@ typedef enum ffi_abi {
 } ffi_abi;
 #endif
 
+#define FFI_TARGET_SPECIFIC_STACK_SPACE_ALLOCATION
+#define FFI_TARGET_HAS_COMPLEX_TYPE
+
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
+#define FFI_GO_CLOSURES 1
 #define FFI_TRAMPOLINE_SIZE 24
 #define FFI_NATIVE_RAW_API 0
 

src/alpha/internal.h

@@ -0,0 +1,23 @@
+#define ALPHA_ST_VOID	0
+#define ALPHA_ST_INT	1
+#define ALPHA_ST_FLOAT	2
+#define ALPHA_ST_DOUBLE	3
+#define ALPHA_ST_CPLXF	4
+#define ALPHA_ST_CPLXD	5
+
+#define ALPHA_LD_VOID	0
+#define ALPHA_LD_INT64	1
+#define ALPHA_LD_INT32	2
+#define ALPHA_LD_UINT16	3
+#define ALPHA_LD_SINT16	4
+#define ALPHA_LD_UINT8	5
+#define ALPHA_LD_SINT8	6
+#define ALPHA_LD_FLOAT	7
+#define ALPHA_LD_DOUBLE	8
+#define ALPHA_LD_CPLXF	9
+#define ALPHA_LD_CPLXD	10
+
+#define ALPHA_ST_SHIFT		0
+#define ALPHA_LD_SHIFT		8
+#define ALPHA_RET_IN_MEM	0x10000
+#define ALPHA_FLAGS(S, L)	(((L) << ALPHA_LD_SHIFT) | (S))

src/alpha/osf.S

@@ -1,5 +1,5 @@
 /* -----------------------------------------------------------------------
-   osf.S - Copyright (c) 1998, 2001, 2007, 2008, 2011 Red Hat
+   osf.S - Copyright (c) 1998, 2001, 2007, 2008, 2011, 2014 Red Hat
 
    Alpha/OSF Foreign Function Interface
 
@@ -27,37 +27,46 @@
 #define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
+#include "internal.h"
 
 	.arch ev6
 	.text
 
-/* ffi_call_osf (void *args, unsigned long bytes, unsigned flags,
+/* Aid in building a direct addressed jump table, 4 insns per entry.  */
-		 void *raddr, void (*fnaddr)(void));
+.macro E index
+	.align	4
+	.org	99b + \index * 16
+.endm
 
-   Bit o trickiness here -- ARGS+BYTES is the base of the stack frame
+/* ffi_call_osf (void *stack, void *frame, unsigned flags,
+		 void *raddr, void (*fnaddr)(void), void *closure)
+
+   Bit o trickiness here -- FRAME is the base of the stack frame
    for this function.  This has been allocated by ffi_call.  We also
    deallocate some of the stack that has been alloca'd.  */
 
-	.align	3
+	.align	4
 	.globl	ffi_call_osf
 	.ent	ffi_call_osf
 	FFI_HIDDEN(ffi_call_osf)
 
 ffi_call_osf:
-	.frame	$15, 32, $26, 0
+	cfi_startproc
-	.mask   0x4008000, -32
+	cfi_def_cfa($17, 32)
-$LFB1:
-	addq	$16,$17,$1
 	mov	$16, $30
-	stq	$26, 0($1)
+	stq	$26, 0($17)
-	stq	$15, 8($1)
+	stq	$15, 8($17)
-	stq	$18, 16($1)
+	mov	$17, $15
-	mov	$1, $15
-$LCFI1:
 	.prologue 0
+	cfi_def_cfa_register($15)
+	cfi_rel_offset($26, 0)
+	cfi_rel_offset($15, 8)
 
-	stq	$19, 24($1)
+	stq	$18, 16($17)		# save flags into frame
-	mov	$20, $27
+	stq	$19, 24($17)		# save rvalue into frame
+	mov	$20, $27		# fn into place for call
+	mov	$21, $1			# closure into static chain
 
 	# Load up all of the (potential) argument registers.
 	ldq	$16, 0($30)
@@ -77,310 +86,196 @@ $LCFI1:
 	lda	$30, 48($30)
 
 	jsr	$26, ($27), 0
-	ldgp	$29, 0($26)
+0:
-
+	ldah	$29, 0($26)		!gpdisp!1
-	# If the return value pointer is NULL, assume no return value.
+	ldq	$2, 24($15)		# reload rvalue
-	ldq	$19, 24($15)
+	lda	$29, 0($29)		!gpdisp!1
-	ldq	$18, 16($15)
+	ldq	$3, 16($15)		# reload flags
+	lda	$1, 99f-0b($26)
 	ldq	$26, 0($15)
-$LCFI2:
-	beq	$19, $noretval
-
-	# Store the return value out in the proper type.
-	cmpeq	$18, FFI_TYPE_INT, $1
-	bne	$1, $retint
-	cmpeq	$18, FFI_TYPE_FLOAT, $2
-	bne	$2, $retfloat
-	cmpeq	$18, FFI_TYPE_DOUBLE, $3
-	bne	$3, $retdouble
-
-	.align	3
-$noretval:
 	ldq	$15, 8($15)
-	ret
+	cfi_restore($26)
+	cfi_restore($15)
+	cfi_def_cfa($sp, 0)
+	cmoveq	$2, ALPHA_ST_VOID, $3	# mash null rvalue to void
+	addq	$3, $3, $3
+	s8addq	$3, $1, $1		# 99f + stcode * 16
+	jmp	$31, ($1), $st_int
 
 	.align	4
-$retint:
+99:
-	stq	$0, 0($19)
+E ALPHA_ST_VOID
-	nop
+	ret
-	ldq	$15, 8($15)
+E ALPHA_ST_INT
+$st_int:
+	stq	$0, 0($2)
+	ret
+E ALPHA_ST_FLOAT
+	sts	$f0, 0($2)
+	ret
+E ALPHA_ST_DOUBLE
+	stt	$f0, 0($2)
+	ret
+E ALPHA_ST_CPLXF
+	sts	$f0, 0($2)
+	sts	$f1, 4($2)
+	ret
+E ALPHA_ST_CPLXD
+	stt	$f0, 0($2)
+	stt	$f1, 8($2)
 	ret
 
-	.align	4
+	cfi_endproc
-$retfloat:
-	sts	$f0, 0($19)
-	nop
-	ldq	$15, 8($15)
-	ret
-
-	.align	4
-$retdouble:
-	stt	$f0, 0($19)
-	nop
-	ldq	$15, 8($15)
-	ret
-$LFE1:
-
 	.end	ffi_call_osf
 
 /* ffi_closure_osf(...)
 
    Receives the closure argument in $1.   */
 
-	.align	3
+#define CLOSURE_FS	(16*8)
+
+	.align	4
+	.globl	ffi_go_closure_osf
+	.ent	ffi_go_closure_osf
+	FFI_HIDDEN(ffi_go_closure_osf)
+
+ffi_go_closure_osf:
+	cfi_startproc
+	ldgp	$29, 0($27)
+	subq	$30, CLOSURE_FS, $30
+	cfi_adjust_cfa_offset(CLOSURE_FS)
+	stq	$26, 0($30)
+	.prologue 1
+	cfi_rel_offset($26, 0)
+
+	stq	$16, 10*8($30)
+	stq	$17, 11*8($30)
+	stq	$18, 12*8($30)
+
+	ldq	$16, 8($1)			# load cif
+	ldq	$17, 16($1)			# load fun
+	mov	$1, $18				# closure is user_data
+	br	$do_closure
+
+	cfi_endproc
+	.end	ffi_go_closure_osf
+
+	.align	4
 	.globl	ffi_closure_osf
 	.ent	ffi_closure_osf
 	FFI_HIDDEN(ffi_closure_osf)
 
 ffi_closure_osf:
-	.frame	$30, 16*8, $26, 0
+	cfi_startproc
-	.mask	0x4000000, -16*8
-$LFB2:
 	ldgp	$29, 0($27)
-	subq	$30, 16*8, $30
+	subq	$30, CLOSURE_FS, $30
-$LCFI5:
+	cfi_adjust_cfa_offset(CLOSURE_FS)
 	stq	$26, 0($30)
-$LCFI6:
 	.prologue 1
+	cfi_rel_offset($26, 0)
 
 	# Store all of the potential argument registers in va_list format.
+	stq	$16, 10*8($30)
+	stq	$17, 11*8($30)
+	stq	$18, 12*8($30)
+
+	ldq	$16, 24($1)			# load cif
+	ldq	$17, 32($1)			# load fun
+	ldq	$18, 40($1)			# load user_data
+
+$do_closure:
+	stq	$19, 13*8($30)
+	stq	$20, 14*8($30)
+	stq	$21, 15*8($30)
 	stt	$f16, 4*8($30)
 	stt	$f17, 5*8($30)
 	stt	$f18, 6*8($30)
 	stt	$f19, 7*8($30)
 	stt	$f20, 8*8($30)
 	stt	$f21, 9*8($30)
-	stq	$16, 10*8($30)
-	stq	$17, 11*8($30)
-	stq	$18, 12*8($30)
-	stq	$19, 13*8($30)
-	stq	$20, 14*8($30)
-	stq	$21, 15*8($30)
 
 	# Call ffi_closure_osf_inner to do the bulk of the work.
-	mov	$1, $16
+	lda	$19, 2*8($30)
-	lda	$17, 2*8($30)
+	lda	$20, 10*8($30)
-	lda	$18, 10*8($30)
 	jsr	$26, ffi_closure_osf_inner
-	ldgp	$29, 0($26)
+0:
+	ldah	$29, 0($26)			!gpdisp!2
+	lda	$2, 99f-0b($26)
+	s4addq	$0, 0, $1			# ldcode * 4
+	ldq	$0, 16($30)			# preload return value
+	s4addq	$1, $2, $1			# 99f + ldcode * 16
+	lda	$29, 0($29)			!gpdisp!2
 	ldq	$26, 0($30)
-
+	cfi_restore($26)
-	# Load up the return value in the proper type.
-	lda	$1, $load_table
-	s4addq	$0, $1, $1
-	ldl	$1, 0($1)
-	addq	$1, $29, $1
 	jmp	$31, ($1), $load_32
 
-	.align 4
+.macro epilogue
-$load_none:
+	addq	$30, CLOSURE_FS, $30
-	addq	$30, 16*8, $30
+	cfi_adjust_cfa_offset(-CLOSURE_FS)
 	ret
+	.align	4
+	cfi_adjust_cfa_offset(CLOSURE_FS)
+.endm
 
 	.align 4
-$load_float:
+99:
-	lds	$f0, 16($30)
+E ALPHA_LD_VOID
-	nop
+	epilogue
-	addq	$30, 16*8, $30
-	ret
 
-	.align 4
+E ALPHA_LD_INT64
-$load_double:
+	epilogue
-	ldt	$f0, 16($30)
-	nop
-	addq	$30, 16*8, $30
-	ret
 
-	.align 4
+E ALPHA_LD_INT32
-$load_u8:
+$load_32:
-#ifdef __alpha_bwx__
+	sextl	$0, $0
-	ldbu	$0, 16($30)
+	epilogue
-	nop
-#else
-	ldq	$0, 16($30)
-	and	$0, 255, $0
-#endif
-	addq	$30, 16*8, $30
-	ret
 
-	.align 4
+E ALPHA_LD_UINT16
-$load_s8:
-#ifdef __alpha_bwx__
-	ldbu	$0, 16($30)
-	sextb	$0, $0
-#else
-	ldq	$0, 16($30)
-	sll	$0, 56, $0
-	sra	$0, 56, $0
-#endif
-	addq	$30, 16*8, $30
-	ret
-
-	.align 4
-$load_u16:
-#ifdef __alpha_bwx__
-	ldwu	$0, 16($30)
-	nop
-#else
-	ldq	$0, 16($30)
 	zapnot	$0, 3, $0
-#endif
+	epilogue
-	addq	$30, 16*8, $30
-	ret
 
-	.align 4
+E ALPHA_LD_SINT16
-$load_s16:
 #ifdef __alpha_bwx__
-	ldwu	$0, 16($30)
 	sextw	$0, $0
 #else
-	ldq	$0, 16($30)
 	sll	$0, 48, $0
 	sra	$0, 48, $0
 #endif
-	addq	$30, 16*8, $30
+	epilogue
-	ret
 
-	.align 4
+E ALPHA_LD_UINT8
-$load_32:
+	and	$0, 0xff, $0
-	ldl	$0, 16($30)
+	epilogue
-	nop
-	addq	$30, 16*8, $30
-	ret
 
-	.align 4
+E ALPHA_LD_SINT8
-$load_64:
+#ifdef __alpha_bwx__
-	ldq	$0, 16($30)
+	sextb	$0, $0
-	nop
-	addq	$30, 16*8, $30
-	ret
-$LFE2:
-
-	.end	ffi_closure_osf
-
-#ifdef __ELF__
-.section .rodata
 #else
-.rdata
+	sll	$0, 56, $0
+	sra	$0, 56, $0
 #endif
-$load_table:
+	epilogue
-	.gprel32 $load_none	# FFI_TYPE_VOID
-	.gprel32 $load_32	# FFI_TYPE_INT
-	.gprel32 $load_float	# FFI_TYPE_FLOAT
-	.gprel32 $load_double	# FFI_TYPE_DOUBLE
-	.gprel32 $load_none	# FFI_TYPE_LONGDOUBLE
-	.gprel32 $load_u8	# FFI_TYPE_UINT8
-	.gprel32 $load_s8	# FFI_TYPE_SINT8
-	.gprel32 $load_u16	# FFI_TYPE_UINT16
-	.gprel32 $load_s16	# FFI_TYPE_SINT16
-	.gprel32 $load_32	# FFI_TYPE_UINT32
-	.gprel32 $load_32	# FFI_TYPE_SINT32
-	.gprel32 $load_64	# FFI_TYPE_UINT64
-	.gprel32 $load_64	# FFI_TYPE_SINT64
-	.gprel32 $load_none	# FFI_TYPE_STRUCT
-	.gprel32 $load_64	# FFI_TYPE_POINTER
 
-/* Assert that the table above is in sync with ffi.h.  */
+E ALPHA_LD_FLOAT
+	lds	$f0, 16($sp)
+	epilogue
 
-#if	   FFI_TYPE_FLOAT != 2		\
+E ALPHA_LD_DOUBLE
-	|| FFI_TYPE_DOUBLE != 3		\
+	ldt	$f0, 16($sp)
-	|| FFI_TYPE_UINT8 != 5		\
+	epilogue
-	|| FFI_TYPE_SINT8 != 6		\
-	|| FFI_TYPE_UINT16 != 7		\
-	|| FFI_TYPE_SINT16 != 8		\
-	|| FFI_TYPE_UINT32 != 9		\
-	|| FFI_TYPE_SINT32 != 10	\
-	|| FFI_TYPE_UINT64 != 11	\
-	|| FFI_TYPE_SINT64 != 12	\
-	|| FFI_TYPE_STRUCT != 13	\
-	|| FFI_TYPE_POINTER != 14	\
-	|| FFI_TYPE_LAST != 14
-#error "osf.S out of sync with ffi.h"
-#endif
 
-#ifdef __ELF__
+E ALPHA_LD_CPLXF
-# define UA_SI		.4byte
+	lds	$f0, 16($sp)
-# define FDE_ENCODING	0x1b	/* pcrel sdata4 */
+	lds	$f1, 20($sp)
-# define FDE_ENCODE(X)	.4byte X-.
+	epilogue
-# define FDE_ARANGE(X)	.4byte X
-#elif defined __osf__
-# define UA_SI		.align 0; .long
-# define FDE_ENCODING	0x50	/* aligned absolute */
-# define FDE_ENCODE(X)	.align 3; .quad X
-# define FDE_ARANGE(X)	.align 0; .quad X
-#endif
 
-#ifdef __ELF__
+E ALPHA_LD_CPLXD
-	.section	.eh_frame,EH_FRAME_FLAGS,@progbits
+	ldt	$f0, 16($sp)
-#elif defined __osf__
+	ldt	$f1, 24($sp)
-	.data
+	epilogue
-	.align 3
-	.globl _GLOBAL__F_ffi_call_osf
-_GLOBAL__F_ffi_call_osf:
-#endif
-__FRAME_BEGIN__:
-	UA_SI	$LECIE1-$LSCIE1	# Length of Common Information Entry
-$LSCIE1:
-	UA_SI	0x0		# CIE Identifier Tag
-	.byte	0x1		# CIE Version
-	.ascii "zR\0"		# CIE Augmentation
-	.byte	0x1		# uleb128 0x1; CIE Code Alignment Factor
-	.byte	0x78		# sleb128 -8; CIE Data Alignment Factor
-	.byte	26		# CIE RA Column
-	.byte	0x1		# uleb128 0x1; Augmentation size
-	.byte	FDE_ENCODING	# FDE Encoding
-	.byte	0xc		# DW_CFA_def_cfa
-	.byte	30		# uleb128 column 30
-	.byte	0		# uleb128 offset 0
-	.align 3
-$LECIE1:
-$LSFDE1:
-	UA_SI	$LEFDE1-$LASFDE1		# FDE Length
-$LASFDE1:
-	UA_SI	$LASFDE1-__FRAME_BEGIN__	# FDE CIE offset
-	FDE_ENCODE($LFB1)			# FDE initial location
-	FDE_ARANGE($LFE1-$LFB1)			# FDE address range
-	.byte	0x0		# uleb128 0x0; Augmentation size
 
-	.byte	0x4		# DW_CFA_advance_loc4
+	cfi_endproc
-	UA_SI	$LCFI1-$LFB1
+	.end	ffi_closure_osf
-	.byte	0x9a		# DW_CFA_offset, column 26
-	.byte	4		# uleb128 4*-8
-	.byte	0x8f		# DW_CFA_offset, column 15
-	.byte	0x3		# uleb128 3*-8
-	.byte	0xc		# DW_CFA_def_cfa
-	.byte	15		# uleb128 column 15
-	.byte	32		# uleb128 offset 32
-
-	.byte	0x4		# DW_CFA_advance_loc4
-	UA_SI	$LCFI2-$LCFI1
-	.byte	0xda		# DW_CFA_restore, column 26
-	.align 3
-$LEFDE1:
-
-$LSFDE3:
-	UA_SI	$LEFDE3-$LASFDE3		# FDE Length
-$LASFDE3:
-	UA_SI	$LASFDE3-__FRAME_BEGIN__	# FDE CIE offset
-	FDE_ENCODE($LFB2)			# FDE initial location
-	FDE_ARANGE($LFE2-$LFB2)			# FDE address range
-	.byte	0x0		# uleb128 0x0; Augmentation size
-
-	.byte	0x4		# DW_CFA_advance_loc4
-	UA_SI	$LCFI5-$LFB2
-	.byte	0xe		# DW_CFA_def_cfa_offset
-	.byte	0x80,0x1	# uleb128 128
-
-	.byte	0x4		# DW_CFA_advance_loc4
-	UA_SI	$LCFI6-$LCFI5
-	.byte	0x9a		# DW_CFA_offset, column 26
-	.byte	16		# uleb128 offset 16*-8
-	.align 3
-$LEFDE3:
-#if defined __osf__
-	.align 0
-	.long	0		# End of Table
-#endif
 
 #if defined __ELF__ && defined __linux__
 	.section	.note.GNU-stack,"",@progbits

src/arm/ffitarget.h

@@ -53,19 +53,17 @@ typedef enum ffi_abi {
 
 #define FFI_EXTRA_CIF_FIELDS			\
   int vfp_used;					\
-  short vfp_reg_free, vfp_nargs;		\
+  unsigned short vfp_reg_free, vfp_nargs;	\
   signed char vfp_args[16]			\
 
-/* Internally used. */
-#define FFI_TYPE_STRUCT_VFP_FLOAT  (FFI_TYPE_LAST + 1)
-#define FFI_TYPE_STRUCT_VFP_DOUBLE (FFI_TYPE_LAST + 2)
-
 #define FFI_TARGET_SPECIFIC_VARIADIC
+#define FFI_TARGET_HAS_COMPLEX_TYPE
 
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
-#define FFI_TRAMPOLINE_SIZE 20
+#define FFI_GO_CLOSURES 1
+#define FFI_TRAMPOLINE_SIZE 12
 #define FFI_NATIVE_RAW_API 0
 
 #endif

src/arm/gentramp.sh

@@ -1,118 +0,0 @@
-#!/bin/sh
-
-# -----------------------------------------------------------------------
-#  gentramp.sh - Copyright (c) 2010, Plausible Labs Cooperative, Inc.
-#  
-#  ARM Trampoline Page Generator
-#
-#  Permission is hereby granted, free of charge, to any person obtaining
-#  a copy of this software and associated documentation files (the
-#  ``Software''), to deal in the Software without restriction, including
-#  without limitation the rights to use, copy, modify, merge, publish,
-#  distribute, sublicense, and/or sell copies of the Software, and to
-#  permit persons to whom the Software is furnished to do so, subject to
-#  the following conditions:
-#
-#  The above copyright notice and this permission notice shall be included
-#  in all copies or substantial portions of the Software.
-#
-#  THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
-#  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-#  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
-#  NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
-#  HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
-#  WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-#  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
-#  DEALINGS IN THE SOFTWARE.
-#  -----------------------------------------------------------------------
-
-PROGNAME=$0
-
-# Each trampoline is exactly 3 instructions, or 12 bytes. If any of these values change,
-# the entire arm trampoline implementation must be updated to match, too.
-
-# Size of an individual trampoline, in bytes
-TRAMPOLINE_SIZE=12
-
-# Page size, in bytes
-PAGE_SIZE=4096
-
-# Compute the size of the reachable config page; The first 16 bytes of the config page
-# are unreachable due to our maximum pc-relative ldr offset.
-PAGE_AVAIL=`expr $PAGE_SIZE - 16`
-
-# Compute the number of of available trampolines. 
-TRAMPOLINE_COUNT=`expr $PAGE_AVAIL / $TRAMPOLINE_SIZE`
-
-header () {
-    echo "# GENERATED CODE - DO NOT EDIT"
-    echo "# This file was generated by $PROGNAME"
-    echo ""
-
-    # Write out the license header
-cat << EOF
-#  Copyright (c) 2010, Plausible Labs Cooperative, Inc.
-#  
-#  Permission is hereby granted, free of charge, to any person obtaining
-#  a copy of this software and associated documentation files (the
-#  ``Software''), to deal in the Software without restriction, including
-#  without limitation the rights to use, copy, modify, merge, publish,
-#  distribute, sublicense, and/or sell copies of the Software, and to
-#  permit persons to whom the Software is furnished to do so, subject to
-#  the following conditions:
-#
-#  The above copyright notice and this permission notice shall be included
-#  in all copies or substantial portions of the Software.
-#
-#  THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
-#  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-#  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
-#  NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
-#  HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
-#  WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-#  OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
-#  DEALINGS IN THE SOFTWARE.
-#  -----------------------------------------------------------------------
-
-EOF
-
-    # Write out the trampoline table, aligned to the page boundary
-    echo ".text"
-    echo ".align 12"
-    echo ".globl _ffi_closure_trampoline_table_page"
-    echo "_ffi_closure_trampoline_table_page:"
-}
-
-
-# WARNING - Don't modify the trampoline code size without also updating the relevant libffi code
-trampoline () {
-    cat << END
-
-    // trampoline
-    // Save to stack
-    stmfd sp!, {r0-r3}
-
-    // Load the context argument from the config page.
-    // This places the first usable config value at _ffi_closure_trampoline_table-4080
-    // This accounts for the above 4-byte stmfd instruction, plus 8 bytes constant when loading from pc.
-    ldr r0, [pc, #-4092]
-
-    // Load the jump address from the config page.
-    ldr pc, [pc, #-4092]
-
-END
-}
-
-main () {
-    # Write out the header
-    header
-
-    # Write out the trampolines
-    local i=0
-    while [ $i -lt ${TRAMPOLINE_COUNT} ]; do
-        trampoline
-        local i=`expr $i + 1`
-    done
-}
-
-main

src/arm/internal.h

@@ -0,0 +1,7 @@
+#define ARM_TYPE_VFP_S	0
+#define ARM_TYPE_VFP_D	1
+#define ARM_TYPE_VFP_N	2
+#define ARM_TYPE_INT64	3
+#define ARM_TYPE_INT	4
+#define ARM_TYPE_VOID	5
+#define ARM_TYPE_STRUCT	6

src/arm/sysv.S

@@ -28,463 +28,307 @@
 #define LIBFFI_ASM	
 #include <fficonfig.h>
 #include <ffi.h>
-#ifdef HAVE_MACHINE_ASM_H
+#include <ffi_cfi.h>
-#include <machine/asm.h>
+#include "internal.h"
-#else
-#ifdef __USER_LABEL_PREFIX__
-#define CONCAT1(a, b) CONCAT2(a, b)
-#define CONCAT2(a, b) a ## b
-
-/* Use the right prefix for global labels.  */
-#define CNAME(x) CONCAT1 (__USER_LABEL_PREFIX__, x)
-#else
-#define CNAME(x) x
-#endif
-#ifdef __APPLE__
-#define ENTRY(x) .globl _##x; _##x:
-#else
-#define ENTRY(x) .globl CNAME(x); .type CNAME(x),%function; CNAME(x):
-#endif /* __APPLE__ */
-#endif
-
-#ifdef __ELF__
-#define LSYM(x) .x
-#else
-#define LSYM(x) x
-#endif
-
-/* Use the SOFTFP return value ABI on Mac OS X, as per the iOS ABI
-  Function Call Guide */
-#ifdef __APPLE__
-#define __SOFTFP__
-#endif
-
-/* We need a better way of testing for this, but for now, this is all 
-   we can do.  */
-@ This selects the minimum architecture level required.
-#define __ARM_ARCH__ 3
-
-#if defined(__ARM_ARCH_4__) || defined(__ARM_ARCH_4T__)
-# undef __ARM_ARCH__
-# define __ARM_ARCH__ 4
-#endif
-        
-#if defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_5T__) \
-	|| defined(__ARM_ARCH_5E__) || defined(__ARM_ARCH_5TE__) \
-	|| defined(__ARM_ARCH_5TEJ__)
-# undef __ARM_ARCH__
-# define __ARM_ARCH__ 5
-#endif
-
-#if defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \
-        || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \
-        || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) \
-	|| defined(__ARM_ARCH_6M__)
-# undef __ARM_ARCH__
-# define __ARM_ARCH__ 6
-#endif
 
+/* GCC 4.8 provides __ARM_ARCH; construct it otherwise.  */
+#ifndef __ARM_ARCH
 # if defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) \
      || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) \
      || defined(__ARM_ARCH_7EM__)
-# undef __ARM_ARCH__
+#  define __ARM_ARCH 7
-# define __ARM_ARCH__ 7
+# elif defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) \
-#endif
+        || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6Z__) \
-
+        || defined(__ARM_ARCH_6ZK__) || defined(__ARM_ARCH_6T2__) \
-#if __ARM_ARCH__ >= 5
+	|| defined(__ARM_ARCH_6M__)
-# define call_reg(x)	blx	x
+#  define __ARM_ARCH 6
-#elif defined (__ARM_ARCH_4T__)
+# elif defined(__ARM_ARCH_5__) || defined(__ARM_ARCH_5T__) \
-# define call_reg(x)	mov	lr, pc ; bx	x
+	|| defined(__ARM_ARCH_5E__) || defined(__ARM_ARCH_5TE__) \
-# if defined(__thumb__) || defined(__THUMB_INTERWORK__)
+	|| defined(__ARM_ARCH_5TEJ__)
-#  define __INTERWORKING__
+#  define __ARM_ARCH 5
-# endif
 # else
-# define call_reg(x)	mov	lr, pc ; mov	pc, x
+#  define __ARM_ARCH 4
+# endif
 #endif
 
 /* Conditionally compile unwinder directives.  */
+.macro UNWIND text:vararg
 #ifdef __ARM_EABI__
-#define UNWIND
+	\text
-#else
-#define UNWIND @
-#endif	
-
-.syntax unified
-
-#if defined(__thumb__) && !defined(__THUMB_INTERWORK__)
-#define ARM_FUNC_START(name) \
-	.text; \
-	.align 2; \
-	.thumb; \
-	.thumb_func; \
-	ENTRY(name); \
-	bx pc; \
-	nop; \
-	.arm; \
-	UNWIND .fnstart; \
-_L__##name:
-#else
-#define ARM_FUNC_START(name) \
-	.text; \
-	.align 2; \
-	.arm; \
-	ENTRY(name); \
-	UNWIND .fnstart
-#endif
-
-.macro	RETLDM	regs=, cond=, dirn=ia
-#if defined (__INTERWORKING__)
-	.ifc "\regs",""
-	ldr\cond	lr, [sp], #4
-	.else
-	ldm\cond\dirn	sp!, {\regs, lr}
-	.endif
-	bx\cond	lr
-#else
-	.ifc "\regs",""
-	ldr\cond	pc, [sp], #4
-	.else
-	ldm\cond\dirn	sp!, {\regs, pc}
-	.endif
 #endif	
 .endm
+#if defined(HAVE_AS_CFI_PSEUDO_OP) && defined(__ARM_EABI__)
+	.cfi_sections	.debug_frame
+#endif
 
-	@ r0:   ffi_prep_args
+#define CONCAT(a, b)	CONCAT2(a, b)
-	@ r1:   &ecif
+#define CONCAT2(a, b)	a ## b
-	@ r2:   cif->bytes
-	@ r3:   fig->flags
-	@ sp+0: ecif.rvalue
 
-	@ This assumes we are using gas.
+#ifdef __USER_LABEL_PREFIX__
-ARM_FUNC_START(ffi_call_SYSV)
+# define CNAME(X)	CONCAT (__USER_LABEL_PREFIX__, X)
-	@ Save registers
+#else
-        stmfd	sp!, {r0-r3, fp, lr}
+# define CNAME(X)	X
-	UNWIND .save	{r0-r3, fp, lr}
+#endif
-	mov	fp, sp
+#ifdef __ELF__
+# define SIZE(X)	.size CNAME(X), . - CNAME(X)
+# define TYPE(X, Y)	.type CNAME(X), Y
+#else
+# define SIZE(X)
+# 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); \
+	CNAME(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"
+	.endif
+.endm
+
+	.text
+	.syntax unified
+	.arm
+
+	/* We require interworking on LDM, which implies ARMv5T,
+	   which implies the existance of BLX.  */
+	.arch	armv5t
+
+	/* Note that we use STC and LDC to encode VFP instructions,
+	   so that we do not need ".fpu vfp", nor get that added to
+	   the object file attributes.  These will not be executed
+	   unless the FFI_VFP abi is used.  */
+
+	@ r0:   stack
+	@ r1:   frame
+	@ r2:   fn
+	@ r3:	vfp_used
+
+ARM_FUNC_START(ffi_call_VFP, 1)
+	UNWIND	.fnstart
+	cfi_startproc
+
+	cmp	r3, #3			@ load only d0 if possible
+	ldcle	p11, cr0, [r0]		@ vldrle d0, [sp]
+	ldcgt	p11, cr0, [r0], {16}	@ vldmgt sp, {d0-d7}
+	add	r0, r0, #64		@ discard the vfp register args
+	/* FALLTHRU */
+ARM_FUNC_END(ffi_call_VFP)
+
+ARM_FUNC_START(ffi_call_SYSV, 1)
+	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
+	cfi_def_cfa(fp, 8)
+	cfi_rel_offset(fp, 0)
+	cfi_rel_offset(lr, 4)
 
-	@ Make room for all of the new args.
+	mov	sp, r0		@ install the stack pointer
-	sub	sp, fp, r2
+	mov	lr, r2		@ move the fn pointer out of the way
-
+	ldr	ip, [fp, #16]	@ install the static chain
-	@ Place all of the ffi_prep_args in position
+	ldmia	sp!, {r0-r3}	@ move first 4 parameters in registers.
-	mov	r0, sp
+	blx	lr		@ call fn
-	@     r1 already set
-
-	@ Call ffi_prep_args(stack, &ecif)
-	bl	CNAME(ffi_prep_args_SYSV)
-
-	@ move first 4 parameters in registers
-	ldmia	sp, {r0-r3}
-
-	@ and adjust stack
-	sub	lr, fp, sp	@ cif->bytes == fp - sp
-	ldr	ip, [fp]	@ load fn() in advance
-	cmp	lr, #16
-	movhs	lr, #16
-	add	sp, sp, lr
-
-	@ call (fn) (...)
-	call_reg(ip)
-	
-	@ Remove the space we pushed for the args
-	mov	sp, fp
 
 	@ Load r2 with the pointer to storage for the return value
-	ldr	r2, [sp, #24]
-
 	@ Load r3 with the return type code
-	ldr	r3, [sp, #12]
+	ldr	r2, [fp, #8]
+	ldr	r3, [fp, #12]
 
-	@ If the return value pointer is NULL, assume no return value.
+	@ Deallocate the stack with the arguments.
-	cmp	r2, #0
+	mov	sp, fp
-	beq	LSYM(Lepilogue)
+	cfi_def_cfa_register(sp)
 
-@ return INT
+	@ Store values stored in registers.
-	cmp	r3, #FFI_TYPE_INT
+	.align	3
-#if defined(__SOFTFP__) || defined(__ARM_EABI__)
+	add	pc, pc, r3, lsl #3
-	cmpne	r3, #FFI_TYPE_FLOAT
+	nop
-#endif
+0:
-	streq	r0, [r2]
+E ARM_TYPE_VFP_S
-	beq	LSYM(Lepilogue)
+	stc	p10, cr0, [r2]		@ vstr s0, [r2]
+	pop	{fp,pc}
+E ARM_TYPE_VFP_D
+	stc	p11, cr0, [r2]		@ vstr d0, [r2]
+	pop	{fp,pc}
+E ARM_TYPE_VFP_N
+	stc	p11, cr0, [r2], {8}	@ vstm r2, {d0-d3}
+	pop	{fp,pc}
+E ARM_TYPE_INT64
+	str	r1, [r2, #4]
+	nop
+E ARM_TYPE_INT
+	str	r0, [r2]
+	pop	{fp,pc}
+E ARM_TYPE_VOID
+	pop	{fp,pc}
+	nop
+E ARM_TYPE_STRUCT
+	pop	{fp,pc}
 
-	@ return INT64
+	cfi_endproc
-	cmp	r3, #FFI_TYPE_SINT64
-#if defined(__SOFTFP__) || defined(__ARM_EABI__)
-	cmpne	r3, #FFI_TYPE_DOUBLE
-#endif
-	stmiaeq	r2, {r0, r1}
-
-#if !defined(__SOFTFP__) && !defined(__ARM_EABI__)
-	beq	LSYM(Lepilogue)
-
-@ return FLOAT
-	cmp	r3, #FFI_TYPE_FLOAT
-	stfeqs	f0, [r2]
-	beq	LSYM(Lepilogue)
-
-@ return DOUBLE or LONGDOUBLE
-	cmp	r3, #FFI_TYPE_DOUBLE
-	stfeqd	f0, [r2]
-#endif
-
-LSYM(Lepilogue):
-#if defined (__INTERWORKING__)
-	ldmia   sp!, {r0-r3,fp, lr}
-	bx	lr
-#else
-	ldmia   sp!, {r0-r3,fp, pc}
-#endif
-
-.ffi_call_SYSV_end:
 	UNWIND	.fnend
-#ifdef __ELF__
+ARM_FUNC_END(ffi_call_SYSV)
-        .size    CNAME(ffi_call_SYSV),.ffi_call_SYSV_end-CNAME(ffi_call_SYSV)
-#endif
 
 
 /*
-	unsigned int FFI_HIDDEN
+	int ffi_closure_inner_* (cif, fun, user_data, frame)
-	ffi_closure_inner (closure, respp, args)
-	     ffi_closure *closure;
-	     void **respp;
-  	     void *args;
 */
 
-ARM_FUNC_START(ffi_closure_SYSV)
+ARM_FUNC_START(ffi_go_closure_SYSV, 1)
-	UNWIND .pad #16
+	cfi_startproc
+	stmdb	sp!, {r0-r3}			@ save argument regs
+	cfi_adjust_cfa_offset(16)
+	ldr	r0, [ip, #4]			@ load cif
+	ldr	r1, [ip, #8]			@ load fun
+	mov	r2, ip				@ load user_data
+	b	0f
+	cfi_endproc
+ARM_FUNC_END(ffi_go_closure_SYSV)
+
+ARM_FUNC_START(ffi_closure_SYSV, 1)
+	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
+0:
+	add	ip, sp, #16			@ compute entry sp
+	sub	sp, sp, #64+32			@ allocate frame
+	cfi_adjust_cfa_offset(64+32)
+	stmdb	sp!, {ip,lr}
+
+	/* 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}
+	cfi_adjust_cfa_offset(8)
+	cfi_rel_offset(lr, 4)
+
+	add	r3, sp, #8			@ load frame
+	bl	CNAME(ffi_closure_inner_SYSV)
+
+	@ Load values returned in registers.
+	add	r2, sp, #8+64			@ load result
+	adr	r3, CNAME(ffi_closure_ret)
+	add	pc, r3, r0, lsl #3
+	cfi_endproc
+	UNWIND	.fnend
+ARM_FUNC_END(ffi_closure_SYSV)
+
+ARM_FUNC_START(ffi_go_closure_VFP, 1)
+	cfi_startproc
+	stmdb	sp!, {r0-r3}			@ save argument regs
+	cfi_adjust_cfa_offset(16)
+	ldr	r0, [ip, #4]			@ load cif
+	ldr	r1, [ip, #8]			@ load fun
+	mov	r2, ip				@ load user_data
+	b	0f
+	cfi_endproc
+ARM_FUNC_END(ffi_go_closure_VFP)
+
+ARM_FUNC_START(ffi_closure_VFP, 1)
+	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
+0:
 	add	ip, sp, #16
-	stmfd	sp!, {ip, lr}
+	sub	sp, sp, #64+32			@ allocate frame
-	UNWIND .save	{r0, lr}
+	cfi_adjust_cfa_offset(64+32)
-	add	r2, sp, #8
+	stc	p11, cr0, [sp], {16}		@ vstm sp, {d0-d7}
-	UNWIND .pad #16
+	stmdb	sp!, {ip,lr}
-	sub	sp, sp, #16
-	str	sp, [sp, #8]
-	add	r1, sp, #8
-	bl	CNAME(ffi_closure_inner)
-	cmp	r0, #FFI_TYPE_INT
-	beq	.Lretint
 
-	cmp	r0, #FFI_TYPE_FLOAT
+	/* See above.  */
-#if defined(__SOFTFP__) || defined(__ARM_EABI__)
+	UNWIND	.save {sp,lr}
-	beq	.Lretint
+	cfi_adjust_cfa_offset(8)
-#else
+	cfi_rel_offset(lr, 4)
-	beq	.Lretfloat
-#endif
 
-	cmp	r0, #FFI_TYPE_DOUBLE
+	add	r3, sp, #8			@ load frame
-#if defined(__SOFTFP__) || defined(__ARM_EABI__)
+	bl	CNAME(ffi_closure_inner_VFP)
-	beq	.Lretlonglong
-#else
-	beq	.Lretdouble
-#endif
 
-	cmp	r0, #FFI_TYPE_LONGDOUBLE
+	@ Load values returned in registers.
-#if defined(__SOFTFP__) || defined(__ARM_EABI__)
+	add	r2, sp, #8+64			@ load result
-	beq	.Lretlonglong
+	adr	r3, CNAME(ffi_closure_ret)
-#else
+	add	pc, r3, r0, lsl #3
-	beq	.Lretlongdouble
+	cfi_endproc
-#endif
-
-	cmp	r0, #FFI_TYPE_SINT64
-	beq	.Lretlonglong
-.Lclosure_epilogue:
-	add	sp, sp, #16
-	ldmfd	sp, {sp, pc}
-.Lretint:
-	ldr	r0, [sp]
-	b	.Lclosure_epilogue
-.Lretlonglong:
-	ldr	r0, [sp]
-	ldr	r1, [sp, #4]
-	b	.Lclosure_epilogue
-
-#if !defined(__SOFTFP__) && !defined(__ARM_EABI__)
-.Lretfloat:
-	ldfs	f0, [sp]
-	b	.Lclosure_epilogue
-.Lretdouble:
-	ldfd	f0, [sp]
-	b	.Lclosure_epilogue
-.Lretlongdouble:
-	ldfd	f0, [sp]
-	b	.Lclosure_epilogue
-#endif
-
-.ffi_closure_SYSV_end:
 	UNWIND	.fnend
-#ifdef __ELF__
+ARM_FUNC_END(ffi_closure_VFP)
-        .size    CNAME(ffi_closure_SYSV),.ffi_closure_SYSV_end-CNAME(ffi_closure_SYSV)
-#endif
 
+/* 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.  */
 
-/* Below are VFP hard-float ABI call and closure implementations.
+ARM_FUNC_START(ffi_closure_ret, 0)
-   Add VFP FPU directive here. This is only compiled into the library
+	cfi_startproc
-   under EABI.  */
+	cfi_rel_offset(sp, 0)
-#ifdef __ARM_EABI__
+	cfi_rel_offset(lr, 4)
-	.fpu	vfp
+0:
+E ARM_TYPE_VFP_S
+	ldc	p10, cr0, [r2]			@ vldr s0, [r2]
+	ldm	sp, {sp,pc}
+E ARM_TYPE_VFP_D
+	ldc	p11, cr0, [r2]			@ vldr d0, [r2]
+	ldm	sp, {sp,pc}
+E ARM_TYPE_VFP_N
+	ldc	p11, cr0, [r2], {8}		@ vldm r2, {d0-d3}
+	ldm	sp, {sp,pc}
+E ARM_TYPE_INT64
+	ldr	r1, [r2, #4]
+	nop
+E ARM_TYPE_INT
+	ldr	r0, [r2]
+	ldm	sp, {sp,pc}
+E ARM_TYPE_VOID
+	ldm	sp, {sp,pc}
+	nop
+E ARM_TYPE_STRUCT
+	ldm	sp, {sp,pc}
+	cfi_endproc
+ARM_FUNC_END(ffi_closure_ret)
 
-	@ r0:   fn
+#if FFI_EXEC_TRAMPOLINE_TABLE
-	@ r1:   &ecif
-	@ r2:   cif->bytes
-	@ r3:   fig->flags
-	@ sp+0: ecif.rvalue
 
-ARM_FUNC_START(ffi_call_VFP)
+/* ??? The iOS support should be updated.  The first insn used to
-	@ Save registers
+   be STMFD, but that's been moved into ffi_closure_SYSV.  If the
-        stmfd	sp!, {r0-r3, fp, lr}
+   writable page is put after this one we can make use of the
-	UNWIND .save	{r0-r3, fp, lr}
+   pc+8 feature of the architecture.  We can also reduce the size
-	mov	fp, sp
+   of the thunk to 8 and pack more of these into the page.
-	UNWIND .setfp	fp, sp
 
-	@ Make room for all of the new args.
+   In the meantime, simply replace the STMFD with a NOP so as to
-	sub	sp, sp, r2
+   keep all the magic numbers the same within ffi.c.  */
 
-	@ Make room for loading VFP args
+	.align	12
-	sub	sp, sp, #64
+ARM_FUNC_START(ffi_closure_trampoline_table_page)
+.rept	4096 / 12
+	nop
+	ldr	ip, [pc, #-4092]
+	ldr	pc, [pc, #-4092]
+.endr
 
-	@ Place all of the ffi_prep_args in position
+#else
-	mov	r0, sp
-	@     r1 already set
-	sub	r2, fp, #64   @ VFP scratch space
 
-	@ Call ffi_prep_args(stack, &ecif, vfp_space)
+ARM_FUNC_START(ffi_arm_trampoline, 1)
-	bl	CNAME(ffi_prep_args_VFP)
+0:	adr	ip, 0b
+	ldr	pc, 1f
+1:	.long	0
+ARM_FUNC_END(ffi_arm_trampoline)
 
-	@ Load VFP register args if needed
+#endif /* FFI_EXEC_TRAMPOLINE_TABLE */
-	cmp	r0, #0
-	mov	ip, fp
-	beq	LSYM(Lbase_args)
-
-	@ Load only d0 if possible
-	cmp	r0, #3
-	sub	ip, fp, #64
-	flddle	d0, [ip]
-	fldmiadgt	ip, {d0-d7}
-
-LSYM(Lbase_args):
-	@ move first 4 parameters in registers
-	ldmia	sp, {r0-r3}
-
-	@ and adjust stack
-	sub	lr, ip, sp	@ cif->bytes == (fp - 64) - sp
-	ldr	ip, [fp]	@ load fn() in advance
-        cmp	lr, #16
-	movhs	lr, #16
-        add	sp, sp, lr
-
-	@ call (fn) (...)
-	call_reg(ip)
-
-	@ Remove the space we pushed for the args
-	mov	sp, fp
-
-	@ Load r2 with the pointer to storage for
-	@ the return value
-	ldr	r2, [sp, #24]
-
-	@ Load r3 with the return type code 
-	ldr	r3, [sp, #12]
-
-	@ If the return value pointer is NULL,
-	@ assume no return value.
-	cmp	r2, #0
-	beq	LSYM(Lepilogue_vfp)
-
-	cmp	r3, #FFI_TYPE_INT
-	streq	r0, [r2]
-	beq	LSYM(Lepilogue_vfp)
-
-	cmp	r3, #FFI_TYPE_SINT64
-	stmeqia	r2, {r0, r1}
-	beq	LSYM(Lepilogue_vfp)
-
-	cmp	r3, #FFI_TYPE_FLOAT
-	fstseq	s0, [r2]
-	beq	LSYM(Lepilogue_vfp)
-	
-	cmp	r3, #FFI_TYPE_DOUBLE
-	fstdeq	d0, [r2]
-	beq	LSYM(Lepilogue_vfp)
-
-	cmp	r3, #FFI_TYPE_STRUCT_VFP_FLOAT
-	cmpne	r3, #FFI_TYPE_STRUCT_VFP_DOUBLE
-	fstmiadeq	r2, {d0-d3}
-
-LSYM(Lepilogue_vfp):
-	RETLDM	"r0-r3,fp"
-
-.ffi_call_VFP_end:
-	UNWIND .fnend
-        .size    CNAME(ffi_call_VFP),.ffi_call_VFP_end-CNAME(ffi_call_VFP)
-
-
-ARM_FUNC_START(ffi_closure_VFP)
-	fstmfdd	sp!, {d0-d7}
-	@ r0-r3, then d0-d7
-	UNWIND .pad #80
-	add	ip, sp, #80
-	stmfd	sp!, {ip, lr}
-	UNWIND .save	{r0, lr}
-	add	r2, sp, #72
-	add	r3, sp, #8
-	UNWIND .pad #72
-	sub	sp, sp, #72
-	str	sp, [sp, #64]
-	add	r1, sp, #64
-	bl	CNAME(ffi_closure_inner)
-
-	cmp	r0, #FFI_TYPE_INT
-	beq	.Lretint_vfp
-
-	cmp	r0, #FFI_TYPE_FLOAT
-	beq	.Lretfloat_vfp
-
-	cmp	r0, #FFI_TYPE_DOUBLE
-	cmpne	r0, #FFI_TYPE_LONGDOUBLE
-	beq	.Lretdouble_vfp
-
-	cmp	r0, #FFI_TYPE_SINT64
-	beq	.Lretlonglong_vfp
-
-	cmp	r0, #FFI_TYPE_STRUCT_VFP_FLOAT
-	beq	.Lretfloat_struct_vfp
-
-	cmp	r0, #FFI_TYPE_STRUCT_VFP_DOUBLE
-	beq	.Lretdouble_struct_vfp
-	
-.Lclosure_epilogue_vfp:
-	add	sp, sp, #72
-	ldmfd	sp, {sp, pc}
-
-.Lretfloat_vfp:
-	flds	s0, [sp]
-	b	.Lclosure_epilogue_vfp
-.Lretdouble_vfp:
-	fldd	d0, [sp]
-	b	.Lclosure_epilogue_vfp
-.Lretint_vfp:
-	ldr	r0, [sp]
-	b	.Lclosure_epilogue_vfp
-.Lretlonglong_vfp:
-	ldmia	sp, {r0, r1}
-	b	.Lclosure_epilogue_vfp
-.Lretfloat_struct_vfp:
-	fldmiad	sp, {d0-d1}
-	b	.Lclosure_epilogue_vfp
-.Lretdouble_struct_vfp:
-	fldmiad	sp, {d0-d3}
-	b	.Lclosure_epilogue_vfp
-
-.ffi_closure_VFP_end:
-	UNWIND .fnend
-        .size    CNAME(ffi_closure_VFP),.ffi_closure_VFP_end-CNAME(ffi_closure_VFP)
-#endif
-
-ENTRY(ffi_arm_trampoline)
-	stmfd sp!, {r0-r3}
-	ldr r0, [pc]
-	ldr pc, [pc]
 
 #if defined __ELF__ && defined __linux__
 	.section	.note.GNU-stack,"",%progbits

src/prep_cif.c

@@ -147,9 +147,6 @@ ffi_status FFI_HIDDEN ffi_prep_cif_core(ffi_cif *cif, ffi_abi abi,
 #if !defined FFI_TARGET_SPECIFIC_STACK_SPACE_ALLOCATION
   /* Make space for the return structure pointer */
   if (cif->rtype->type == FFI_TYPE_STRUCT
-#ifdef SPARC
-      && (cif->abi != FFI_V9 || cif->rtype->size > 32)
-#endif
 #ifdef TILE
       && (cif->rtype->size > 10 * FFI_SIZEOF_ARG)
 #endif
@@ -179,14 +176,6 @@ ffi_status FFI_HIDDEN ffi_prep_cif_core(ffi_cif *cif, ffi_abi abi,
       FFI_ASSERT_VALID_TYPE(*ptr);
 
 #if !defined FFI_TARGET_SPECIFIC_STACK_SPACE_ALLOCATION
-#ifdef SPARC
-      if (((*ptr)->type == FFI_TYPE_STRUCT
-	   && ((*ptr)->size > 16 || cif->abi != FFI_V9))
-	  || ((*ptr)->type == FFI_TYPE_LONGDOUBLE
-	      && cif->abi != FFI_V9))
-	bytes += sizeof(void*);
-      else
-#endif
 	{
 	  /* Add any padding if necessary */
 	  if (((*ptr)->alignment - 1) & bytes)

src/sparc/ffi.c

@@ -27,463 +27,304 @@
 
 #include <ffi.h>
 #include <ffi_common.h>
-
 #include <stdlib.h>
+#include "internal.h"
 
+#ifndef SPARC64
 
-/* ffi_prep_args is called by the assembly routine once stack space
+/* Force FFI_TYPE_LONGDOUBLE to be different than FFI_TYPE_DOUBLE;
-   has been allocated for the function's arguments */
+   all further uses in this file will refer to the 128-bit type.  */
-
-void ffi_prep_args_v8(char *stack, extended_cif *ecif)
-{
-  int i;
-  void **p_argv;
-  char *argp;
-  ffi_type **p_arg;
-
-  /* Skip 16 words for the window save area */
-  argp = stack + 16*sizeof(int);
-
-  /* This should only really be done when we are returning a structure,
-     however, it's faster just to do it all the time...
-
-  if ( ecif->cif->rtype->type == FFI_TYPE_STRUCT ) */
-  *(int *) argp = (long)ecif->rvalue;
-
-  /* And 1 word for the  structure return value. */
-  argp += sizeof(int);
-
-#ifdef USING_PURIFY
-  /* Purify will probably complain in our assembly routine, unless we
-     zero out this memory. */
-
-  ((int*)argp)[0] = 0;
-  ((int*)argp)[1] = 0;
-  ((int*)argp)[2] = 0;
-  ((int*)argp)[3] = 0;
-  ((int*)argp)[4] = 0;
-  ((int*)argp)[5] = 0;
-#endif
-
-  p_argv = ecif->avalue;
-
-  for (i = ecif->cif->nargs, p_arg = ecif->cif->arg_types; i; i--, p_arg++)
-    {
-      size_t z;
-
-	  if ((*p_arg)->type == FFI_TYPE_STRUCT
 #if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-	      || (*p_arg)->type == FFI_TYPE_LONGDOUBLE
+# if FFI_TYPE_LONGDOUBLE != 4
+#  error FFI_TYPE_LONGDOUBLE out of date
 # endif
-	      )
+#else
-	    {
+# undef FFI_TYPE_LONGDOUBLE
-	      *(unsigned int *) argp = (unsigned long)(* p_argv);
+# define FFI_TYPE_LONGDOUBLE 4
-	      z = sizeof(int);
-	    }
-	  else
-	    {
-	      z = (*p_arg)->size;
-	      if (z < sizeof(int))
-		{
-		  z = sizeof(int);
-		  switch ((*p_arg)->type)
-		    {
-		    case FFI_TYPE_SINT8:
-		      *(signed int *) argp = *(SINT8 *)(* p_argv);
-		      break;
-		      
-		    case FFI_TYPE_UINT8:
-		      *(unsigned int *) argp = *(UINT8 *)(* p_argv);
-		      break;
-		      
-		    case FFI_TYPE_SINT16:
-		      *(signed int *) argp = *(SINT16 *)(* p_argv);
-		      break;
-		      
-		    case FFI_TYPE_UINT16:
-		      *(unsigned int *) argp = *(UINT16 *)(* p_argv);
-		      break;
-
-		    default:
-		      FFI_ASSERT(0);
-		    }
-		}
-	      else
-		{
-		  memcpy(argp, *p_argv, z);
-		}
-	    }
-	  p_argv++;
-	  argp += z;
-    }
-  
-  return;
-}
-
-int ffi_prep_args_v9(char *stack, extended_cif *ecif)
-{
-  int i, ret = 0;
-  int tmp;
-  void **p_argv;
-  char *argp;
-  ffi_type **p_arg;
-
-  tmp = 0;
-
-  /* Skip 16 words for the window save area */
-  argp = stack + 16*sizeof(long long);
-
-#ifdef USING_PURIFY
-  /* Purify will probably complain in our assembly routine, unless we
-     zero out this memory. */
-
-  ((long long*)argp)[0] = 0;
-  ((long long*)argp)[1] = 0;
-  ((long long*)argp)[2] = 0;
-  ((long long*)argp)[3] = 0;
-  ((long long*)argp)[4] = 0;
-  ((long long*)argp)[5] = 0;
 #endif
 
-  p_argv = ecif->avalue;
-
-  if (ecif->cif->rtype->type == FFI_TYPE_STRUCT &&
-      ecif->cif->rtype->size > 32)
-    {
-      *(unsigned long long *) argp = (unsigned long)ecif->rvalue;
-      argp += sizeof(long long);
-      tmp = 1;
-    }
-
-  for (i = 0, p_arg = ecif->cif->arg_types; i < ecif->cif->nargs;
-       i++, p_arg++)
-    {
-      size_t z;
-
-      z = (*p_arg)->size;
-      switch ((*p_arg)->type)
-	{
-	case FFI_TYPE_STRUCT:
-	  if (z > 16)
-	    {
-	      /* For structures larger than 16 bytes we pass reference.  */
-	      *(unsigned long long *) argp = (unsigned long)* p_argv;
-	      argp += sizeof(long long);
-	      tmp++;
-	      p_argv++;
-	      continue;
-	    }
-	  /* FALLTHROUGH */
-	case FFI_TYPE_FLOAT:
-	case FFI_TYPE_DOUBLE:
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-	case FFI_TYPE_LONGDOUBLE:
-#endif
-	  ret = 1; /* We should promote into FP regs as well as integer.  */
-	  break;
-	}
-      if (z < sizeof(long long))
-	{
-	  switch ((*p_arg)->type)
-	    {
-	    case FFI_TYPE_SINT8:
-	      *(signed long long *) argp = *(SINT8 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_UINT8:
-	      *(unsigned long long *) argp = *(UINT8 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_SINT16:
-	      *(signed long long *) argp = *(SINT16 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_UINT16:
-	      *(unsigned long long *) argp = *(UINT16 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_SINT32:
-	      *(signed long long *) argp = *(SINT32 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_UINT32:
-	      *(unsigned long long *) argp = *(UINT32 *)(* p_argv);
-	      break;
-
-	    case FFI_TYPE_FLOAT:
-	      *(float *) (argp + 4) = *(FLOAT32 *)(* p_argv); /* Right justify */
-	      break;
-
-	    case FFI_TYPE_STRUCT:
-	      memcpy(argp, *p_argv, z);
-	      break;
-
-	    default:
-	      FFI_ASSERT(0);
-	    }
-	  z = sizeof(long long);
-	  tmp++;
-	}
-      else if (z == sizeof(long long))
-	{
-	  memcpy(argp, *p_argv, z);
-	  z = sizeof(long long);
-	  tmp++;
-	}
-      else
-	{
-	  if ((tmp & 1) && (*p_arg)->alignment > 8)
-	    {
-	      tmp++;
-	      argp += sizeof(long long);
-	    }
-	  memcpy(argp, *p_argv, z);
-	  z = 2 * sizeof(long long);
-	  tmp += 2;
-	}
-      p_argv++;
-      argp += z;
-    }
-
-  return ret;
-}
-
 /* Perform machine dependent cif processing */
-ffi_status ffi_prep_cif_machdep(ffi_cif *cif)
+ffi_status FFI_HIDDEN
+ffi_prep_cif_machdep(ffi_cif *cif)
 {
-  int wordsize;
+  ffi_type *rtype = cif->rtype;
-
+  int rtt = rtype->type;
-  if (cif->abi != FFI_V9)
+  size_t bytes;
-    {
+  int i, n, flags;
-      wordsize = 4;
-
-      /* If we are returning a struct, this will already have been added.
-	 Otherwise we need to add it because it's always got to be there! */
-
-      if (cif->rtype->type != FFI_TYPE_STRUCT)
-	cif->bytes += wordsize;
-
-      /* sparc call frames require that space is allocated for 6 args,
-	 even if they aren't used. Make that space if necessary. */
-  
-      if (cif->bytes < 4*6+4)
-	cif->bytes = 4*6+4;
-    }
-  else
-    {
-      wordsize = 8;
-
-      /* sparc call frames require that space is allocated for 6 args,
-	 even if they aren't used. Make that space if necessary. */
-  
-      if (cif->bytes < 8*6)
-	cif->bytes = 8*6;
-    }
-
-  /* Adjust cif->bytes. to include 16 words for the window save area,
-     and maybe the struct/union return pointer area, */
-
-  cif->bytes += 16 * wordsize;
-
-  /* The stack must be 2 word aligned, so round bytes up
-     appropriately. */
-
-  cif->bytes = ALIGN(cif->bytes, 2 * wordsize);
 
   /* Set the return type flag */
-  switch (cif->rtype->type)
+  switch (rtt)
     {
     case FFI_TYPE_VOID:
+      flags = SPARC_RET_VOID;
+      break;
     case FFI_TYPE_FLOAT:
+      flags = SPARC_RET_F_1;
+      break;
     case FFI_TYPE_DOUBLE:
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+      flags = SPARC_RET_F_2;
+      break;
     case FFI_TYPE_LONGDOUBLE:
-#endif
-      cif->flags = cif->rtype->type;
-      break;
-
     case FFI_TYPE_STRUCT:
-      if (cif->abi == FFI_V9 && cif->rtype->size > 32)
+      flags = (rtype->size & 0xfff) << SPARC_SIZEMASK_SHIFT;
-	cif->flags = FFI_TYPE_VOID;
+      flags |= SPARC_RET_STRUCT;
-      else
-	cif->flags = FFI_TYPE_STRUCT;
       break;
-
     case FFI_TYPE_SINT8:
+      flags = SPARC_RET_SINT8;
+      break;
     case FFI_TYPE_UINT8:
+      flags = SPARC_RET_UINT8;
+      break;
     case FFI_TYPE_SINT16:
+      flags = SPARC_RET_SINT16;
+      break;
     case FFI_TYPE_UINT16:
-      if (cif->abi == FFI_V9)
+      flags = SPARC_RET_UINT16;
-	cif->flags = FFI_TYPE_INT;
+      break;
-      else
+    case FFI_TYPE_INT:
-	cif->flags = cif->rtype->type;
+    case FFI_TYPE_SINT32:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_POINTER:
+      flags = SPARC_RET_UINT32;
       break;
-
     case FFI_TYPE_SINT64:
     case FFI_TYPE_UINT64:
-      if (cif->abi == FFI_V9)
+      flags = SPARC_RET_INT64;
-	cif->flags = FFI_TYPE_INT;
+      break;
-      else
+    case FFI_TYPE_COMPLEX:
-	cif->flags = FFI_TYPE_SINT64;
+      rtt = rtype->elements[0]->type;
+      switch (rtt)
+	{
+	case FFI_TYPE_FLOAT:
+	  flags = SPARC_RET_F_2;
+	  break;
+	case FFI_TYPE_DOUBLE:
+	  flags = SPARC_RET_F_4;
+	  break;
+	case FFI_TYPE_LONGDOUBLE:
+	  flags = SPARC_RET_F_8;
+	  break;
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	  flags = SPARC_RET_INT128;
+	  break;
+	case FFI_TYPE_INT:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT32:
+	  flags = SPARC_RET_INT64;
+	  break;
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_UINT16:
+	  flags = SP_V8_RET_CPLX16;
+	  break;
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT8:
+	  flags = SP_V8_RET_CPLX8;
+	  break;
+	default:
+	  abort();
+	}
+      break;
+    default:
+      abort();
+    }
+  cif->flags = flags;
+
+  bytes = 0;
+  for (i = 0, n = cif->nargs; i < n; ++i)
+    {
+      ffi_type *ty = cif->arg_types[i];
+      size_t z = ty->size;
+      int tt = ty->type;
+
+      switch (tt)
+	{
+	case FFI_TYPE_STRUCT:
+	case FFI_TYPE_LONGDOUBLE:
+	by_reference:
+	  /* Passed by reference.  */
+	  z = 4;
 	  break;
 
+	case FFI_TYPE_COMPLEX:
+	  tt = ty->elements[0]->type;
+	  if (tt == FFI_TYPE_FLOAT || z > 8)
+	    goto by_reference;
+	  /* FALLTHRU */
+
 	default:
-      cif->flags = FFI_TYPE_INT;
+	  z = ALIGN(z, 4);
-      break;
 	}
+      bytes += z;
+    }
+
+  /* Sparc call frames require that space is allocated for 6 args,
+     even if they aren't used. Make that space if necessary.  */
+  if (bytes < 6 * 4)
+    bytes = 6 * 4;
+
+  /* The ABI always requires space for the struct return pointer.  */
+  bytes += 4;
+
+  /* The stack must be 2 word aligned, so round bytes up appropriately. */
+  bytes = ALIGN(bytes, 2 * 4);
+
+  /* Include the call frame to prep_args.  */
+  bytes += 4*16 + 4*8;
+  cif->bytes = bytes;
+
   return FFI_OK;
 }
 
-int ffi_v9_layout_struct(ffi_type *arg, int off, char *ret, char *intg, char *flt)
+extern void ffi_call_v8(ffi_cif *cif, void (*fn)(void), void *rvalue,
-{
+			void **avalue, size_t bytes, void *closure) FFI_HIDDEN;
-  ffi_type **ptr = &arg->elements[0];
 
-  while (*ptr != NULL)
+int FFI_HIDDEN
+ffi_prep_args_v8(ffi_cif *cif, unsigned long *argp, void *rvalue, void **avalue)
 {
-      if (off & ((*ptr)->alignment - 1))
+  ffi_type **p_arg;
-	off = ALIGN(off, (*ptr)->alignment);
+  int flags = cif->flags;
+  int i, nargs;
 
-      switch ((*ptr)->type)
+  if (rvalue == NULL)
+    {
+      if ((flags & SPARC_FLAG_RET_MASK) == SPARC_RET_STRUCT)
+	{
+	  /* Since we pass the pointer to the callee, we need a value.
+	     We allowed for this space in ffi_call, before ffi_call_v8
+	     alloca'd the space.  */
+	  rvalue = (char *)argp + cif->bytes;
+	}
+      else
+	{
+	  /* Otherwise, we can ignore the return value.  */
+	  flags = SPARC_RET_VOID;
+	}
+    }
+
+  /* This could only really be done when we are returning a structure.
+     However, the space is reserved so we can do it unconditionally.  */
+  *argp++ = (unsigned long)rvalue;
+
+#ifdef USING_PURIFY
+  /* Purify will probably complain in our assembly routine,
+     unless we zero out this memory. */
+  memset(argp, 0, 6*4);
+#endif
+
+  p_arg = cif->arg_types;
+  for (i = 0, nargs = cif->nargs; i < nargs; i++)
+    {
+      ffi_type *ty = p_arg[i];
+      void *a = avalue[i];
+      int tt = ty->type;
+      size_t z;
+
+      switch (tt)
 	{
 	case FFI_TYPE_STRUCT:
-	  off = ffi_v9_layout_struct(*ptr, off, ret, intg, flt);
-	  off = ALIGN(off, FFI_SIZEOF_ARG);
-	  break;
-	case FFI_TYPE_FLOAT:
-	case FFI_TYPE_DOUBLE:
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
 	case FFI_TYPE_LONGDOUBLE:
-#endif
+	by_reference:
-	  memmove(ret + off, flt + off, (*ptr)->size);
+	  *argp++ = (unsigned long)a;
-	  off += (*ptr)->size;
 	  break;
+
+	case FFI_TYPE_DOUBLE:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	  memcpy(argp, a, 8);
+	  argp += 2;
+	  break;
+
+	case FFI_TYPE_INT:
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_POINTER:
+	  *argp++ = *(unsigned *)a;
+	  break;
+
+	case FFI_TYPE_UINT8:
+	  *argp++ = *(UINT8 *)a;
+	  break;
+	case FFI_TYPE_SINT8:
+	  *argp++ = *(SINT8 *)a;
+	  break;
+	case FFI_TYPE_UINT16:
+	  *argp++ = *(UINT16 *)a;
+	  break;
+	case FFI_TYPE_SINT16:
+	  *argp++ = *(SINT16 *)a;
+	  break;
+
+        case FFI_TYPE_COMPLEX:
+	  tt = ty->elements[0]->type;
+	  z = ty->size;
+	  if (tt == FFI_TYPE_FLOAT || z > 8)
+	    goto by_reference;
+	  if (z < 4)
+	    {
+	      memcpy((char *)argp + 4 - z, a, z);
+	      argp++;
+	    }
+	  else
+	    {
+	      memcpy(argp, a, z);
+	      argp += z / 4;
+	    }
+	  break;
+
 	default:
-	  memmove(ret + off, intg + off, (*ptr)->size);
+	  abort();
-	  off += (*ptr)->size;
-	  break;
-	}
-      ptr++;
-    }
-  return off;
-}
-
-
-#ifdef SPARC64
-extern int ffi_call_v9(void *, extended_cif *, unsigned, 
-		       unsigned, unsigned *, void (*fn)(void));
-#else
-extern int ffi_call_v8(void *, extended_cif *, unsigned, 
-		       unsigned, unsigned *, void (*fn)(void));
-#endif
-
-#ifndef __GNUC__
-void ffi_flush_icache (void *, size_t);
-#endif
-
-void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
-{
-  extended_cif ecif;
-  void *rval = rvalue;
-
-  ecif.cif = cif;
-  ecif.avalue = avalue;
-
-  /* If the return value is a struct and we don't have a return	*/
-  /* value address then we need to make one		        */
-
-  ecif.rvalue = rvalue;
-  if (cif->rtype->type == FFI_TYPE_STRUCT)
-    {
-      if (cif->rtype->size <= 32)
-	rval = alloca(64);
-      else
-	{
-	  rval = NULL;
-	  if (rvalue == NULL)
-	    ecif.rvalue = alloca(cif->rtype->size);
 	}
     }
 
-  switch (cif->abi) 
+  return flags;
+}
+
+static void
+ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	      void **avalue, void *closure)
 {
-    case FFI_V8:
+  size_t bytes = cif->bytes;
-#ifdef SPARC64
+
-      /* We don't yet support calling 32bit code from 64bit */
+  FFI_ASSERT (cif->abi == FFI_V8);
-      FFI_ASSERT(0);
+
-#else
+  /* If we've not got a return value, we need to create one if we've
-      if (rvalue && (cif->rtype->type == FFI_TYPE_STRUCT
+     got to pass the return value to the callee.  Otherwise ignore it.  */
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+  if (rvalue == NULL
-	  || cif->flags == FFI_TYPE_LONGDOUBLE
+      && (cif->flags & SPARC_FLAG_RET_MASK) == SPARC_RET_STRUCT)
-#endif
+    bytes += ALIGN (cif->rtype->size, 8);
-	  ))
+
+  ffi_call_v8(cif, fn, rvalue, avalue, -bytes, closure);
+}
+
+void
+ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 {
-	  /* For v8, we need an "unimp" with size of returning struct */
+  ffi_call_int (cif, fn, rvalue, avalue, NULL);
-	  /* behind "call", so we alloc some executable space for it. */
+}
-	  /* l7 is used, we need to make sure v8.S doesn't use %l7.   */
+
-	  unsigned int *call_struct = NULL;
+void
-	  ffi_closure_alloc(32, (void **)&call_struct);
+ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
-	  if (call_struct)
+	     void **avalue, void *closure)
 {
-	      unsigned long f = (unsigned long)fn;
+  ffi_call_int (cif, fn, rvalue, avalue, closure);
-	      call_struct[0] = 0xae10001f;		 /* mov   %i7, %l7	 */
+}
-	      call_struct[1] = 0xbe10000f;		 /* mov   %o7, %i7	 */
+
-	      call_struct[2] = 0x03000000 | f >> 10;     /* sethi %hi(fn), %g1	 */
-	      call_struct[3] = 0x9fc06000 | (f & 0x3ff); /* jmp %g1+%lo(fn), %o7 */
-	      call_struct[4] = 0x01000000;		 /* nop			 */
-	      if (cif->rtype->size < 0x7f)
-		call_struct[5] = cif->rtype->size;	 /* unimp		 */
-	      else
-		call_struct[5] = 0x01000000;	     	 /* nop			 */
-	      call_struct[6] = 0x81c7e008;		 /* ret			 */
-	      call_struct[7] = 0xbe100017;		 /* mov   %l7, %i7	 */
 #ifdef __GNUC__
-	      asm volatile ("iflush %0; iflush %0+8; iflush %0+16; iflush %0+24" : :
+static inline void
-			    "r" (call_struct) : "memory");
+ffi_flush_icache (void *p)
+{
   /* SPARC v8 requires 5 instructions for flush to be visible */
-	      asm volatile ("nop; nop; nop; nop; nop");
+  asm volatile ("iflush	%0; iflush %0+8; nop; nop; nop; nop; nop"
+		: : "r" (p) : "memory");
+}
 #else
-	      ffi_flush_icache (call_struct, 32);
+extern void ffi_flush_icache (void *) FFI_HIDDEN;
 #endif
-	      ffi_call_v8(ffi_prep_args_v8, &ecif, cif->bytes,
-			  cif->flags, rvalue, call_struct);
-	      ffi_closure_free(call_struct);
-	    }
-	  else
-	    {
-	      ffi_call_v8(ffi_prep_args_v8, &ecif, cif->bytes,
-			  cif->flags, rvalue, fn);
-	    }
-	}
-      else
-	{
-	  ffi_call_v8(ffi_prep_args_v8, &ecif, cif->bytes,
-		      cif->flags, rvalue, fn);
-	}
-#endif
-      break;
-    case FFI_V9:
-#ifdef SPARC64
-      ffi_call_v9(ffi_prep_args_v9, &ecif, cif->bytes,
-		  cif->flags, rval, fn);
-      if (rvalue && rval && cif->rtype->type == FFI_TYPE_STRUCT)
-	ffi_v9_layout_struct(cif->rtype, 0, (char *)rvalue, (char *)rval, ((char *)rval)+32);
-#else
-      /* And vice versa */
-      FFI_ASSERT(0);
-#endif
-      break;
-    default:
-      FFI_ASSERT(0);
-      break;
-    }
-}
 
-
+extern void ffi_closure_v8(void) FFI_HIDDEN;
-#ifdef SPARC64
+extern void ffi_go_closure_v8(void) FFI_HIDDEN;
-extern void ffi_closure_v9(void);
-#else
-extern void ffi_closure_v8(void);
-#endif
 
 ffi_status
 ffi_prep_closure_loc (ffi_closure *closure,
@@ -493,189 +334,135 @@ ffi_prep_closure_loc (ffi_closure* closure,
 		      void *codeloc)
 {
   unsigned int *tramp = (unsigned int *) &closure->tramp[0];
-  unsigned long fn;
+  unsigned long ctx = (unsigned long) closure;
-#ifdef SPARC64
+  unsigned long fn = (unsigned long) ffi_closure_v8;
-  /* Trampoline address is equal to the closure address.  We take advantage
+
-     of that to reduce the trampoline size by 8 bytes. */
-  if (cif->abi != FFI_V9)
-    return FFI_BAD_ABI;
-  fn = (unsigned long) ffi_closure_v9;
-  tramp[0] = 0x83414000;	/* rd	%pc, %g1	*/
-  tramp[1] = 0xca586010;	/* ldx	[%g1+16], %g5	*/
-  tramp[2] = 0x81c14000;	/* jmp	%g5		*/
-  tramp[3] = 0x01000000;	/* nop			*/
-  *((unsigned long *) &tramp[4]) = fn;
-#else
-  unsigned long ctx = (unsigned long) codeloc;
   if (cif->abi != FFI_V8)
     return FFI_BAD_ABI;
-  fn = (unsigned long) ffi_closure_v8;
+
   tramp[0] = 0x03000000 | fn >> 10;	/* sethi %hi(fn), %g1	*/
   tramp[1] = 0x05000000 | ctx >> 10;	/* sethi %hi(ctx), %g2	*/
   tramp[2] = 0x81c06000 | (fn & 0x3ff);	/* jmp   %g1+%lo(fn)	*/
   tramp[3] = 0x8410a000 | (ctx & 0x3ff);/* or    %g2, %lo(ctx)	*/
-#endif
 
   closure->cif = cif;
   closure->fun = fun;
   closure->user_data = user_data;
 
-  /* Flush the Icache.  closure is 8 bytes aligned.  */
+  ffi_flush_icache (closure);
-#ifdef __GNUC__
-#ifdef SPARC64
-  asm volatile ("flush	%0; flush %0+8" : : "r" (closure) : "memory");
-#else
-  asm volatile ("iflush	%0; iflush %0+8" : : "r" (closure) : "memory");
-  /* SPARC v8 requires 5 instructions for flush to be visible */
-  asm volatile ("nop; nop; nop; nop; nop");
-#endif
-#else
-  ffi_flush_icache (closure, 16);
-#endif
 
   return FFI_OK;
 }
 
-int
+ffi_status
-ffi_closure_sparc_inner_v8(ffi_closure *closure,
+ffi_prep_go_closure (ffi_go_closure *closure, ffi_cif *cif,
-  void *rvalue, unsigned long *gpr, unsigned long *scratch)
+		     void (*fun)(ffi_cif*, void*, void**, void*))
+{
+  if (cif->abi != FFI_V8)
+    return FFI_BAD_ABI;
+
+  closure->tramp = ffi_go_closure_v8;
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
+}
+
+int FFI_HIDDEN
+ffi_closure_sparc_inner_v8(ffi_cif *cif, 
+			   void (*fun)(ffi_cif*, void*, void**, void*),
+			   void *user_data, void *rvalue,
+			   unsigned long *argp)
 {
-  ffi_cif *cif;
   ffi_type **arg_types;
   void **avalue;
-  int i, argn;
+  int i, nargs, flags;
 
-  cif = closure->cif;
   arg_types = cif->arg_types;
-  avalue = alloca(cif->nargs * sizeof(void *));
+  nargs = cif->nargs;
+  flags = cif->flags;
+  avalue = alloca(nargs * sizeof(void *));
 
   /* Copy the caller's structure return address so that the closure
-     returns the data directly to the caller.  */
+     returns the data directly to the caller.  Also install it so we
-  if (cif->flags == FFI_TYPE_STRUCT
+     can return the address in %o0.  */
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE  
+  if ((flags & SPARC_FLAG_RET_MASK) == SPARC_RET_STRUCT)
-      || cif->flags == FFI_TYPE_LONGDOUBLE
+    {
-#endif
+      void *new_rvalue = (void *)*argp;
-     )
+      *(void **)rvalue = new_rvalue;
-    rvalue = (void *) gpr[0];
+      rvalue = new_rvalue;
+    }
 
   /* Always skip the structure return address.  */
-  argn = 1;
+  argp++;
 
   /* Grab the addresses of the arguments from the stack frame.  */
-  for (i = 0; i < cif->nargs; i++)
+  for (i = 0; i < nargs; i++)
     {
-      if (arg_types[i]->type == FFI_TYPE_STRUCT
+      ffi_type *ty = arg_types[i];
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+      int tt = ty->type;
-	  || arg_types[i]->type == FFI_TYPE_LONGDOUBLE
+      void *a = argp;
-#endif
+      size_t z;
-         )
+
+      switch (tt)
 	{
+	case FFI_TYPE_STRUCT:
+	case FFI_TYPE_LONGDOUBLE:
+	by_reference:
 	  /* Straight copy of invisible reference.  */
-	  avalue[i] = (void *)gpr[argn++];
+	  a = (void *)*argp;
-	}
+	  break;
-      else if ((arg_types[i]->type == FFI_TYPE_DOUBLE
+
-	       || arg_types[i]->type == FFI_TYPE_SINT64
+	case FFI_TYPE_DOUBLE:
-	       || arg_types[i]->type == FFI_TYPE_UINT64)
+	case FFI_TYPE_SINT64:
-	       /* gpr is 8-byte aligned.  */
+	case FFI_TYPE_UINT64:
-	       && (argn % 2) != 0)
+	  if ((unsigned long)a & 7)
 	    {
 	      /* Align on a 8-byte boundary.  */
-	  scratch[0] = gpr[argn];
+	      UINT64 *tmp = alloca(8);
-	  scratch[1] = gpr[argn+1];
+	      *tmp = ((UINT64)argp[0] << 32) | argp[1];
-	  avalue[i] = scratch;
+	      a = tmp;
-	  scratch -= 2;
-	  argn += 2;
 	    }
-      else
+	  argp++;
-	{
+	  break;
-	  /* Always right-justify.  */
+
-	  argn += ALIGN(arg_types[i]->size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
+	case FFI_TYPE_INT:
-	  avalue[i] = ((char *) &gpr[argn]) - arg_types[i]->size;
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_POINTER:
+	  break;
+        case FFI_TYPE_UINT16:
+        case FFI_TYPE_SINT16:
+	  a += 2;
+	  break;
+        case FFI_TYPE_UINT8:
+        case FFI_TYPE_SINT8:
+	  a += 3;
+	  break;
+
+        case FFI_TYPE_COMPLEX:
+	  tt = ty->elements[0]->type;
+	  z = ty->size;
+	  if (tt == FFI_TYPE_FLOAT || z > 8)
+	    goto by_reference;
+	  if (z < 4)
+	    a += 4 - z;
+	  else if (z > 4)
+	    argp++;
+	  break;
+
+	default:
+	  abort();
 	}
+      argp++;
+      avalue[i] = a;
     }
 
   /* Invoke the closure.  */
-  (closure->fun) (cif, rvalue, avalue, closure->user_data);
+  fun (cif, rvalue, avalue, user_data);
 
   /* Tell ffi_closure_sparc how to perform return type promotions.  */
-  return cif->rtype->type;
+  return flags;
-}
-
-int
-ffi_closure_sparc_inner_v9(ffi_closure *closure,
-  void *rvalue, unsigned long *gpr, double *fpr)
-{
-  ffi_cif *cif;
-  ffi_type **arg_types;
-  void **avalue;
-  int i, argn, fp_slot_max;
-
-  cif = closure->cif;
-  arg_types = cif->arg_types;
-  avalue = alloca(cif->nargs * sizeof(void *));
-
-  /* Copy the caller's structure return address so that the closure
-     returns the data directly to the caller.  */
-  if (cif->flags == FFI_TYPE_VOID
-      && cif->rtype->type == FFI_TYPE_STRUCT)
-    {
-      rvalue = (void *) gpr[0];
-      /* Skip the structure return address.  */
-      argn = 1;
-    }
-  else
-    argn = 0;
-
-  fp_slot_max = 16 - argn;
-
-  /* Grab the addresses of the arguments from the stack frame.  */
-  for (i = 0; i < cif->nargs; i++)
-    {
-      if (arg_types[i]->type == FFI_TYPE_STRUCT)
-	{
-	  if (arg_types[i]->size > 16)
-	    {
-	      /* Straight copy of invisible reference.  */
-	      avalue[i] = (void *)gpr[argn++];
-	    }
-	  else
-	    {
-	      /* Left-justify.  */
-	      ffi_v9_layout_struct(arg_types[i],
-				   0,
-				   (char *) &gpr[argn],
-				   (char *) &gpr[argn],
-				   (char *) &fpr[argn]);
-	      avalue[i] = &gpr[argn];
-	      argn += ALIGN(arg_types[i]->size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
-	    }
-	}
-      else
-	{
-	  /* Right-justify.  */
-	  argn += ALIGN(arg_types[i]->size, FFI_SIZEOF_ARG) / FFI_SIZEOF_ARG;
-
-	  /* Align on a 16-byte boundary.  */
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-	  if (arg_types[i]->type == FFI_TYPE_LONGDOUBLE && (argn % 2) != 0)
-	    argn++;
-#endif
-	  if (i < fp_slot_max
-	      && (arg_types[i]->type == FFI_TYPE_FLOAT
-		  || arg_types[i]->type == FFI_TYPE_DOUBLE
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-		  || arg_types[i]->type == FFI_TYPE_LONGDOUBLE
-#endif
-		  ))
-	    avalue[i] = ((char *) &fpr[argn]) - arg_types[i]->size;
-	  else
-	    avalue[i] = ((char *) &gpr[argn]) - arg_types[i]->size;
-	}
-    }
-
-  /* Invoke the closure.  */
-  (closure->fun) (cif, rvalue, avalue, closure->user_data);
-
-  /* Tell ffi_closure_sparc how to perform return type promotions.  */
-  return cif->rtype->type;
 }
+#endif /* !SPARC64 */

src/sparc/ffi64.c

@@ -0,0 +1,592 @@
+/* -----------------------------------------------------------------------
+   ffi.c - Copyright (c) 2011, 2013 Anthony Green
+           Copyright (c) 1996, 2003-2004, 2007-2008 Red Hat, Inc.
+
+   SPARC Foreign Function Interface
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   ``Software''), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be included
+   in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+   ----------------------------------------------------------------------- */
+
+#include <ffi.h>
+#include <ffi_common.h>
+#include <stdlib.h>
+#include "internal.h"
+
+/* Force FFI_TYPE_LONGDOUBLE to be different than FFI_TYPE_DOUBLE;
+   all further uses in this file will refer to the 128-bit type.  */
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+# if FFI_TYPE_LONGDOUBLE != 4
+#  error FFI_TYPE_LONGDOUBLE out of date
+# endif
+#else
+# undef FFI_TYPE_LONGDOUBLE
+# define FFI_TYPE_LONGDOUBLE 4
+#endif
+
+#ifdef SPARC64
+
+/* Flatten the contents of a structure to the parts that are passed in
+   floating point registers.  The return is a bit mask wherein bit N
+   set means bytes [4*n, 4*n+3] are passed in %fN.
+
+   We encode both the (running) size (maximum 32) and mask (maxumum 255)
+   into one integer.  The size is placed in the low byte, so that align
+   and addition work correctly.  The mask is placed in the second byte.  */
+
+static int
+ffi_struct_float_mask (ffi_type *outer_type, int size_mask)
+{
+  ffi_type **elts;
+  ffi_type *t;
+
+  if (outer_type->type == FFI_TYPE_COMPLEX)
+    {
+      int m = 0, tt = outer_type->elements[0]->type;
+      size_t z = outer_type->size;
+
+      if (tt == FFI_TYPE_FLOAT
+	  || tt == FFI_TYPE_DOUBLE
+	  || tt == FFI_TYPE_LONGDOUBLE)
+        m = (1 << (z / 4)) - 1;
+      return (m << 8) | z;
+    }
+  FFI_ASSERT (outer_type->type == FFI_TYPE_STRUCT);
+
+  for (elts = outer_type->elements; (t = *elts) != NULL; elts++)
+    {
+      size_t z = t->size;
+      int o, m, tt;
+
+      size_mask = ALIGN(size_mask, t->alignment);
+      switch (t->type)
+	{
+	case FFI_TYPE_STRUCT:
+	  size_mask = ffi_struct_float_mask (t, size_mask);
+	  continue;
+	case FFI_TYPE_COMPLEX:
+	  tt = t->elements[0]->type;
+	  if (tt != FFI_TYPE_FLOAT
+	      && tt != FFI_TYPE_DOUBLE
+	      && tt != FFI_TYPE_LONGDOUBLE)
+	    break;
+	  /* FALLTHRU */
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_DOUBLE:
+	case FFI_TYPE_LONGDOUBLE:
+	  m = (1 << (z / 4)) - 1;	/* compute mask for type */
+	  o = (size_mask >> 2) & 0x3f;	/* extract word offset */
+	  size_mask |= m << (o + 8);	/* insert mask into place */
+	  break;
+	}
+      size_mask += z;
+    }
+
+  size_mask = ALIGN(size_mask, outer_type->alignment);
+  FFI_ASSERT ((size_mask & 0xff) == outer_type->size);
+
+  return size_mask;
+}
+
+/* Merge floating point data into integer data.  If the structure is
+   entirely floating point, simply return a pointer to the fp data.  */
+
+static void *
+ffi_struct_float_merge (int size_mask, void *vi, void *vf)
+{
+  int size = size_mask & 0xff;
+  int mask = size_mask >> 8;
+  int n = size >> 2;
+
+  if (mask == 0)
+    return vi;
+  else if (mask == (1 << n) - 1)
+    return vf;
+  else
+    {
+      unsigned int *wi = vi, *wf = vf;
+      int i;
+
+      for (i = 0; i < n; ++i)
+	if ((mask >> i) & 1)
+	  wi[i] = wf[i];
+
+      return vi;
+    }
+}
+
+/* Similar, but place the data into VD in the end.  */
+
+void FFI_HIDDEN
+ffi_struct_float_copy (int size_mask, void *vd, void *vi, void *vf)
+{
+  int size = size_mask & 0xff;
+  int mask = size_mask >> 8;
+  int n = size >> 2;
+
+  if (mask == 0)
+    ;
+  else if (mask == (1 << n) - 1)
+    vi = vf;
+  else
+    {
+      unsigned int *wd = vd, *wi = vi, *wf = vf;
+      int i;
+
+      for (i = 0; i < n; ++i)
+	wd[i] = ((mask >> i) & 1 ? wf : wi)[i];
+      return;
+    }
+  memcpy (vd, vi, size);
+}
+
+/* Perform machine dependent cif processing */
+
+ffi_status FFI_HIDDEN
+ffi_prep_cif_machdep(ffi_cif *cif)
+{
+  ffi_type *rtype = cif->rtype;
+  int rtt = rtype->type;
+  size_t bytes = 0;
+  int i, n, flags;
+
+  /* Set the return type flag */
+  switch (rtt)
+    {
+    case FFI_TYPE_VOID:
+      flags = SPARC_RET_VOID;
+      break;
+    case FFI_TYPE_FLOAT:
+      flags = SPARC_RET_F_1;
+      break;
+    case FFI_TYPE_DOUBLE:
+      flags = SPARC_RET_F_2;
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+      flags = SPARC_RET_F_4;
+      break;
+
+    case FFI_TYPE_COMPLEX:
+    case FFI_TYPE_STRUCT:
+      if (rtype->size > 32)
+	{
+	  flags = SPARC_RET_VOID | SPARC_FLAG_RET_IN_MEM;
+	  bytes = 8;
+	}
+      else
+	{
+	  int size_mask = ffi_struct_float_mask (rtype, 0);
+	  int word_size = (size_mask >> 2) & 0x3f;
+	  int all_mask = (1 << word_size) - 1;
+	  int fp_mask = size_mask >> 8;
+
+	  flags = (size_mask << SPARC_SIZEMASK_SHIFT) | SPARC_RET_STRUCT;
+
+	  /* For special cases of all-int or all-fp, we can return
+	     the value directly without popping through a struct copy.  */
+	  if (fp_mask == 0)
+	    {
+	      if (rtype->alignment >= 8)
+		{
+		  if (rtype->size == 8)
+		    flags = SPARC_RET_INT64;
+		  else if (rtype->size == 16)
+		    flags = SPARC_RET_INT128;
+		}
+	    }
+	  else if (fp_mask == all_mask)
+	    switch (word_size)
+	      {
+	      case 1: flags = SPARC_RET_F_1; break;
+	      case 2: flags = SPARC_RET_F_2; break;
+	      case 3: flags = SP_V9_RET_F_3; break;
+	      case 4: flags = SPARC_RET_F_4; break;
+	      /* 5 word structures skipped; handled via RET_STRUCT.  */
+	      case 6: flags = SPARC_RET_F_6; break;
+	      /* 7 word structures skipped; handled via RET_STRUCT.  */
+	      case 8: flags = SPARC_RET_F_8; break;
+	      }
+	}
+      break;
+
+    case FFI_TYPE_SINT8:
+      flags = SPARC_RET_SINT8;
+      break;
+    case FFI_TYPE_UINT8:
+      flags = SPARC_RET_UINT8;
+      break;
+    case FFI_TYPE_SINT16:
+      flags = SPARC_RET_SINT16;
+      break;
+    case FFI_TYPE_UINT16:
+      flags = SPARC_RET_UINT16;
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+      flags = SP_V9_RET_SINT32;
+      break;
+    case FFI_TYPE_UINT32:
+      flags = SPARC_RET_UINT32;
+      break;
+    case FFI_TYPE_SINT64:
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_POINTER:
+      flags = SPARC_RET_INT64;
+      break;
+
+    default:
+      abort();
+    }
+
+  bytes = 0;
+  for (i = 0, n = cif->nargs; i < n; ++i)
+    {
+      ffi_type *ty = cif->arg_types[i];
+      size_t z = ty->size;
+      size_t a = ty->alignment;
+
+      switch (ty->type)
+	{
+	case FFI_TYPE_COMPLEX:
+	case FFI_TYPE_STRUCT:
+	  /* Large structs passed by reference.  */
+	  if (z > 16)
+	    {
+	      a = z = 8;
+	      break;
+	    }
+	  /* Small structs may be passed in integer or fp regs or both.  */
+	  if (bytes >= 16*8)
+	    break;
+	  if ((ffi_struct_float_mask (ty, 0) & 0xff00) == 0)
+	    break;
+	  /* FALLTHRU */
+	case FFI_TYPE_FLOAT:
+	case FFI_TYPE_DOUBLE:
+	case FFI_TYPE_LONGDOUBLE:
+	  flags |= SPARC_FLAG_FP_ARGS;
+	  break;
+	}
+      bytes = ALIGN(bytes, a);
+      bytes += ALIGN(z, 8);
+    }
+
+  /* Sparc call frames require that space is allocated for 6 args,
+     even if they aren't used. Make that space if necessary. */
+  if (bytes < 6 * 8)
+    bytes = 6 * 8;
+
+  /* The stack must be 2 word aligned, so round bytes up appropriately. */
+  bytes = ALIGN(bytes, 16);
+
+  /* Include the call frame to prep_args.  */
+  bytes += 8*16 + 8*8;
+
+  cif->bytes = bytes;
+  cif->flags = flags;
+  return FFI_OK;
+}
+
+extern void ffi_call_v9(ffi_cif *cif, void (*fn)(void), void *rvalue,
+			void **avalue, size_t bytes, void *closure) FFI_HIDDEN;
+
+/* ffi_prep_args is called by the assembly routine once stack space
+   has been allocated for the function's arguments */
+
+int FFI_HIDDEN
+ffi_prep_args_v9(ffi_cif *cif, unsigned long *argp, void *rvalue, void **avalue)
+{
+  ffi_type **p_arg;
+  int flags = cif->flags;
+  int i, nargs;
+
+  if (rvalue == NULL)
+    {
+      if (flags & SPARC_FLAG_RET_IN_MEM)
+	{
+	  /* Since we pass the pointer to the callee, we need a value.
+	     We allowed for this space in ffi_call, before ffi_call_v8
+	     alloca'd the space.  */
+	  rvalue = (char *)argp + cif->bytes;
+	}
+      else
+	{
+	  /* Otherwise, we can ignore the return value.  */
+	  flags = SPARC_RET_VOID;
+	}
+    }
+
+#ifdef USING_PURIFY
+  /* Purify will probably complain in our assembly routine,
+     unless we zero out this memory. */
+  memset(argp, 0, 6*8);
+#endif
+
+  if (flags & SPARC_FLAG_RET_IN_MEM)
+    *argp++ = (unsigned long)rvalue;
+
+  p_arg = cif->arg_types;
+  for (i = 0, nargs = cif->nargs; i < nargs; i++)
+    {
+      ffi_type *ty = p_arg[i];
+      void *a = avalue[i];
+      size_t z;
+
+      switch (ty->type)
+	{
+	case FFI_TYPE_SINT8:
+	  *argp++ = *(SINT8 *)a;
+	  break;
+	case FFI_TYPE_UINT8:
+	  *argp++ = *(UINT8 *)a;
+	  break;
+	case FFI_TYPE_SINT16:
+	  *argp++ = *(SINT16 *)a;
+	  break;
+	case FFI_TYPE_UINT16:
+	  *argp++ = *(UINT16 *)a;
+	  break;
+	case FFI_TYPE_INT:
+	case FFI_TYPE_SINT32:
+	  *argp++ = *(SINT32 *)a;
+	  break;
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_FLOAT:
+	  *argp++ = *(UINT32 *)a;
+	  break;
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_POINTER:
+	case FFI_TYPE_DOUBLE:
+	  *argp++ = *(UINT64 *)a;
+	  break;
+
+	case FFI_TYPE_LONGDOUBLE:
+	case FFI_TYPE_COMPLEX:
+	case FFI_TYPE_STRUCT:
+	  z = ty->size;
+	  if (z > 16)
+	    {
+	      /* For structures larger than 16 bytes we pass reference.  */
+	      *argp++ = (unsigned long)a;
+	      break;
+	    }
+	  if (((unsigned long)argp & 15) && ty->alignment > 8)
+	    argp++;
+	  memcpy(argp, a, z);
+	  argp += ALIGN(z, 8) / 8;
+	  break;
+
+	default:
+	  abort();
+	}
+    }
+
+  return flags;
+}
+
+static void
+ffi_call_int(ffi_cif *cif, void (*fn)(void), void *rvalue,
+	     void **avalue, void *closure)
+{
+  size_t bytes = cif->bytes;
+
+  FFI_ASSERT (cif->abi == FFI_V9);
+
+  if (rvalue == NULL && (cif->flags & SPARC_FLAG_RET_IN_MEM))
+    bytes += ALIGN (cif->rtype->size, 16);
+
+  ffi_call_v9(cif, fn, rvalue, avalue, -bytes, closure);
+}
+
+void
+ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, NULL);
+}
+
+void
+ffi_call_go(ffi_cif *cif, void (*fn)(void), void *rvalue,
+	    void **avalue, void *closure)
+{
+  ffi_call_int(cif, fn, rvalue, avalue, closure);
+}
+
+#ifdef __GNUC__
+static inline void
+ffi_flush_icache (void *p)
+{
+  asm volatile ("flush	%0; flush %0+8" : : "r" (p) : "memory");
+}
+#else
+extern void ffi_flush_icache (void *) FFI_HIDDEN;
+#endif
+
+extern void ffi_closure_v9(void) FFI_HIDDEN;
+extern void ffi_go_closure_v9(void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure* closure,
+		      ffi_cif* cif,
+		      void (*fun)(ffi_cif*, void*, void**, void*),
+		      void *user_data,
+		      void *codeloc)
+{
+  unsigned int *tramp = (unsigned int *) &closure->tramp[0];
+  unsigned long fn;
+
+  if (cif->abi != FFI_V9)
+    return FFI_BAD_ABI;
+
+  /* Trampoline address is equal to the closure address.  We take advantage
+     of that to reduce the trampoline size by 8 bytes. */
+  fn = (unsigned long) ffi_closure_v9;
+  tramp[0] = 0x83414000;	/* rd	%pc, %g1	*/
+  tramp[1] = 0xca586010;	/* ldx	[%g1+16], %g5	*/
+  tramp[2] = 0x81c14000;	/* jmp	%g5		*/
+  tramp[3] = 0x01000000;	/* nop			*/
+  *((unsigned long *) &tramp[4]) = fn;
+
+  closure->cif = cif;
+  closure->fun = fun;
+  closure->user_data = user_data;
+
+  ffi_flush_icache (closure);
+
+  return FFI_OK;
+}
+
+ffi_status
+ffi_prep_go_closure (ffi_go_closure* closure, ffi_cif* cif,
+		     void (*fun)(ffi_cif*, void*, void**, void*))
+{
+  if (cif->abi != FFI_V9)
+    return FFI_BAD_ABI;
+
+  closure->tramp = ffi_go_closure_v9;
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
+}
+
+int FFI_HIDDEN
+ffi_closure_sparc_inner_v9(ffi_cif *cif,
+			   void (*fun)(ffi_cif*, void*, void**, void*),
+			   void *user_data, void *rvalue,
+			   unsigned long *gpr, unsigned long *fpr)
+{
+  ffi_type **arg_types;
+  void **avalue;
+  int i, argn, argx, nargs, flags;
+
+  arg_types = cif->arg_types;
+  nargs = cif->nargs;
+  flags = cif->flags;
+
+  avalue = alloca(nargs * sizeof(void *));
+
+  /* Copy the caller's structure return address so that the closure
+     returns the data directly to the caller.  */
+  if (flags & SPARC_FLAG_RET_IN_MEM)
+    {
+      rvalue = (void *) gpr[0];
+      /* Skip the structure return address.  */
+      argn = 1;
+    }
+  else
+    argn = 0;
+
+  /* Grab the addresses of the arguments from the stack frame.  */
+  for (i = 0; i < nargs; i++, argn = argx)
+    {
+      ffi_type *ty = arg_types[i];
+      void *a = &gpr[argn];
+      size_t z;
+
+      argx = argn + 1;
+      switch (ty->type)
+	{
+	case FFI_TYPE_COMPLEX:
+	case FFI_TYPE_STRUCT:
+	  z = ty->size;
+	  if (z > 16)
+	    a = *(void **)a;
+	  else
+	    {
+	      argx = argn + ALIGN (z, 8) / 8;
+	      if (argn < 16)
+		{
+		  int size_mask = ffi_struct_float_mask (ty, 0);
+		  int argn_mask = (0xffff00 >> argn) & 0xff00;
+
+		  /* Eliminate fp registers off the end.  */
+		  size_mask = (size_mask & 0xff) | (size_mask & argn_mask);
+		  a = ffi_struct_float_merge (size_mask, gpr+argn, fpr+argn);
+		}
+	    }
+	  break;
+
+	case FFI_TYPE_LONGDOUBLE:
+	  argn = ALIGN (argn, 2);
+	  a = (argn < 16 ? fpr : gpr) + argn;
+	  argx = argn + 2;
+	  break;
+	case FFI_TYPE_DOUBLE:
+	  if (argn <= 16)
+	    a = fpr + argn;
+	  break;
+	case FFI_TYPE_FLOAT:
+	  if (argn <= 16)
+	    a = fpr + argn;
+	  a += 4;
+	  break;
+
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	case FFI_TYPE_POINTER:
+	  break;
+	case FFI_TYPE_INT:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	  a += 4;
+	  break;
+        case FFI_TYPE_UINT16:
+        case FFI_TYPE_SINT16:
+	  a += 6;
+	  break;
+        case FFI_TYPE_UINT8:
+        case FFI_TYPE_SINT8:
+	  a += 7;
+	  break;
+
+	default:
+	  abort();
+	}
+      avalue[i] = a;
+    }
+
+  /* Invoke the closure.  */
+  fun (cif, rvalue, avalue, user_data);
+
+  /* Tell ffi_closure_sparc how to perform return type promotions.  */
+  return flags;
+}
+#endif /* SPARC64 */

src/sparc/ffitarget.h

@@ -46,21 +46,24 @@ typedef signed long            ffi_sarg;
 
 typedef enum ffi_abi {
   FFI_FIRST_ABI = 0,
-  FFI_V8,
-  FFI_V8PLUS,
-  FFI_V9,
-  FFI_LAST_ABI,
 #ifdef SPARC64
-  FFI_DEFAULT_ABI = FFI_V9
+  FFI_V9,
+  FFI_DEFAULT_ABI = FFI_V9,
 #else
-  FFI_DEFAULT_ABI = FFI_V8
+  FFI_V8,
+  FFI_DEFAULT_ABI = FFI_V8,
 #endif
+  FFI_LAST_ABI
 } ffi_abi;
 #endif
 
+#define FFI_TARGET_SPECIFIC_STACK_SPACE_ALLOCATION
+#define FFI_TARGET_HAS_COMPLEX_TYPE
+
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
+#define FFI_GO_CLOSURES 1
 #define FFI_NATIVE_RAW_API 0
 
 #ifdef SPARC64

src/sparc/internal.h

@@ -0,0 +1,26 @@
+#define SPARC_RET_VOID		0
+#define SPARC_RET_STRUCT	1
+#define SPARC_RET_UINT8		2
+#define SPARC_RET_SINT8		3
+#define SPARC_RET_UINT16	4
+#define SPARC_RET_SINT16	5
+#define SPARC_RET_UINT32	6
+#define SP_V9_RET_SINT32	7	/* v9 only */
+#define SP_V8_RET_CPLX16	7	/* v8 only */
+#define SPARC_RET_INT64		8
+#define SPARC_RET_INT128	9
+
+/* Note that F_7 is missing, and is handled by SPARC_RET_STRUCT.  */
+#define SPARC_RET_F_8		10
+#define SPARC_RET_F_6		11
+#define SPARC_RET_F_4		12
+#define SPARC_RET_F_2		13
+#define SP_V9_RET_F_3		14	/* v9 only */
+#define SP_V8_RET_CPLX8		14	/* v8 only */
+#define SPARC_RET_F_1		15
+
+#define SPARC_FLAG_RET_MASK	15
+#define SPARC_FLAG_RET_IN_MEM	32
+#define SPARC_FLAG_FP_ARGS	64
+
+#define SPARC_SIZEMASK_SHIFT	8

src/sparc/v8.S

@@ -28,27 +28,33 @@
 #define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
+#include "internal.h"
 
-#define STACKFRAME 96		/* Minimum stack framesize for SPARC */
+#ifndef SPARC64
-#define ARGS (64+4)		/* Offset of register area in frame */
+
+#define C2(X, Y)  X ## Y
+#define C1(X, Y)  C2(X, Y)
+
+#ifdef __USER_LABEL_PREFIX__
+# define C(Y)	C1(__USER_LABEL_PREFIX__, Y)
+#else
+# define C(Y)	Y
+#endif
+#define L(Y)	C1(.L, Y)
+
+	.text
 
 #ifndef __GNUC__
-	.text
         .align 8
-.globl ffi_flush_icache
+	.globl	C(ffi_flush_icache)
-.globl _ffi_flush_icache
+	.type	C(ffi_flush_icache),@function
+	FFI_HIDDEN(C(ffi_flush_icache))
 
-ffi_flush_icache:
+C(ffi_flush_icache):
-_ffi_flush_icache:	
+	cfi_startproc
-        add %o0, %o1, %o2
-#ifdef SPARC64	
-1:	flush %o0
-#else	
 1:	iflush %o0
-#endif
+	iflush %o+8
-	add %o0, 8, %o0
-	cmp %o0, %o2
-	blt 1b
 	nop
 	nop
 	nop
@@ -56,148 +62,206 @@ _ffi_flush_icache:
 	nop
 	retl
 	 nop
-.ffi_flush_icache_end:
+	cfi_endproc
-	.size	ffi_flush_icache,.ffi_flush_icache_end-ffi_flush_icache
+	.size	C(ffi_flush_icache), . - C(ffi_flush_icache)
 #endif
 
-	.text
+.macro E index
+	.align	16
+	.org	2b + \index * 16
+.endm
+
         .align 8
-.globl ffi_call_v8
+	.globl	C(ffi_call_v8)
-.globl _ffi_call_v8
+	.type	C(ffi_call_v8),@function
+	FFI_HIDDEN(C(ffi_call_v8))
 
-ffi_call_v8:
+C(ffi_call_v8):
-_ffi_call_v8:
+	cfi_startproc
-.LLFB1:
+	! Allocate a stack frame sized by ffi_call.
-	save	%sp, -STACKFRAME, %sp
+	save	%sp, %o4, %sp
-.LLCFI0:
+	cfi_def_cfa_register(%fp)
+	cfi_window_save
 
-	sub	%sp, %i2, %sp	! alloca() space in stack for frame to set up
+	mov	%i0, %o0		! copy cif
-	add	%sp, STACKFRAME, %l0	! %l0 has start of 
+	add	%sp, 64+32, %o1		! load args area
-					! frame to set up
+	mov	%i2, %o2		! copy rvalue
+	call	C(ffi_prep_args_v8)
+	 mov	%i3, %o3		! copy avalue
 
-	mov	%l0, %o0	! call routine to set up frame
+	add	%sp, 32, %sp		! deallocate prep frame
-	call	%i0
+	and	%o0, SPARC_FLAG_RET_MASK, %l0	! save return type
-	mov	%i1, %o1	! (delay)
+	srl	%o0, SPARC_SIZEMASK_SHIFT, %l1	! save return size
+	ld	[%sp+64+4], %o0		! load all argument registers
+	ld	[%sp+64+8], %o1
+	ld	[%sp+64+12], %o2
+	ld	[%sp+64+16], %o3
+	cmp	%l0, SPARC_RET_STRUCT	! struct return needs an unimp 4
+	ld	[%sp+64+20], %o4
+	be	8f
+	 ld	[%sp+64+24], %o5
 
-	ld	[%l0+ARGS], %o0	! call foreign function
+	! Call foreign function
-	ld	[%l0+ARGS+4], %o1
+	call	%i1
-	ld	[%l0+ARGS+8], %o2
+	 mov	%i5, %g2		! load static chain
-	ld	[%l0+ARGS+12], %o3
-	ld	[%l0+ARGS+16], %o4
-	ld	[%l0+ARGS+20], %o5
-	call	%i5
-	mov	%l0, %sp	! (delay) switch to frame
-	nop			! STRUCT returning functions skip 12 instead of 8 bytes
 
-	! If the return value pointer is NULL, assume no return value.
+0:	call	1f		! load pc in %o7
-	tst	%i4
+	 sll	%l0, 4, %l0
-	bz	done
+1:	add	%o7, %l0, %o7	! o7 = 0b + ret_type*16
+	jmp	%o7+(2f-0b)
 	 nop
 
-	cmp	%i3, FFI_TYPE_INT
+	! Note that each entry is 4 insns, enforced by the E macro.
-	be,a	done
+	.align	16
-	st	%o0, [%i4]	! (delay)
+2:
-
+E SPARC_RET_VOID
-	cmp	%i3, FFI_TYPE_FLOAT
-	be,a	done
-	st	%f0, [%i4+0]	! (delay)
-
-	cmp	%i3, FFI_TYPE_DOUBLE
-	be,a	double
-	st	%f0, [%i4+0]	! (delay)
-
-	cmp	%i3, FFI_TYPE_SINT8
-	be,a	sint8
-	sll	%o0, 24, %o0	! (delay)
-
-	cmp	%i3, FFI_TYPE_UINT8
-	be,a	uint8
-	sll	%o0, 24, %o0	! (delay)
-
-	cmp	%i3, FFI_TYPE_SINT16
-	be,a	sint16
-	sll	%o0, 16, %o0	! (delay)
-
-	cmp	%i3, FFI_TYPE_UINT16
-	be,a	uint16
-	sll	%o0, 16, %o0	! (delay)
-
-	cmp	%i3, FFI_TYPE_SINT64
-	be,a	longlong
-	st	%o0, [%i4+0]	! (delay)
-done:
 	ret
 	 restore
-
+E SPARC_RET_STRUCT
-double:
+	unimp
-	st	%f1, [%i4+4]
+E SPARC_RET_UINT8
+	and	%o0, 0xff, %o0
+	st	%o0, [%i2]
 	ret
 	 restore
-
+E SPARC_RET_SINT8
-sint8:
+	sll	%o0, 24, %o0
+	b	7f
 	 sra	%o0, 24, %o0
-	st	%o0, [%i4+0]
+E SPARC_RET_UINT16
-	ret
+	sll	%o0, 16, %o0
-	restore
+	b	7f
-
-uint8:
-	srl	%o0, 24, %o0
-	st	%o0, [%i4+0]
-	ret
-	restore
-
-sint16:
-	sra	%o0, 16, %o0
-	st	%o0, [%i4+0]
-	ret
-	restore
-
-uint16:
 	 srl	%o0, 16, %o0
-	st	%o0, [%i4+0]
+E SPARC_RET_SINT16
+	sll	%o0, 16, %o0
+	b	7f
+	 sra	%o0, 16, %o0
+E SPARC_RET_UINT32
+7:	st	%o0, [%i2]
+	ret
+	 restore
+E SP_V8_RET_CPLX16
+	sth	%o0, [%i2+2]
+	b	9f
+	 srl	%o0, 16, %o0
+E SPARC_RET_INT64
+	st	%o0, [%i2]
+	st	%o1, [%i2+4]
+	ret
+	 restore
+E SPARC_RET_INT128
+	std	%o0, [%i2]
+	std	%o2, [%i2+8]
+	ret
+	 restore
+E SPARC_RET_F_8
+	st	%f7, [%i2+7*4]
+	nop
+	st	%f6, [%i2+6*4]
+	nop
+E SPARC_RET_F_6
+	st	%f5, [%i2+5*4]
+	nop
+	st	%f4, [%i2+4*4]
+	nop
+E SPARC_RET_F_4
+	st	%f3, [%i2+3*4]
+	nop
+	st	%f2, [%i2+2*4]
+	nop
+E SPARC_RET_F_2
+	st	%f1, [%i2+4]
+	st	%f0, [%i2]
+	ret
+	 restore
+E SP_V8_RET_CPLX8
+	stb	%o0, [%i2+1]
+	b	10f
+	 srl	%o0, 8, %o0
+E SPARC_RET_F_1
+	st	%f0, [%i2]
 	ret
 	 restore
 
-longlong:
+	.align	8
-	st	%o1, [%i4+4]
+9:	sth	%o0, [%i2]
+	ret
+	 restore
+	.align	8
+10:	stb	%o0, [%i2]
 	ret
 	 restore
-.LLFE1:
 
-.ffi_call_v8_end:
+	! Struct returning functions expect and skip the unimp here.
-	.size	ffi_call_v8,.ffi_call_v8_end-ffi_call_v8
+	! To make it worse, conforming callees examine the unimp and
+	! make sure the low 12 bits of the unimp match the size of
+	! the struct being returned.
+	.align	8
+8:	call	1f				! load pc in %o7
+	 sll	%l1, 2, %l0			! size * 4
+1:	sll	%l1, 4, %l1			! size * 16
+	add	%l0, %l1, %l0			! size * 20
+	add	%o7, %l0, %o7			! o7 = 0b + size*20
+	jmp	%o7+(2f-8b)
+	 mov	%i5, %g2			! load static chain
+2:
+.rept	0x1000
+	call	%i1
+	 nop
+	unimp	(. - 2b) / 20
+	ret
+	 restore
+.endr
+
+	cfi_endproc
+	.size	C(ffi_call_v8),. - C(ffi_call_v8)
 
 
-#undef STACKFRAME
+/* 16*4 register window + 1*4 struct return + 6*4 args backing store
-#define	STACKFRAME	104	/* 16*4 register window +
+   + 8*4 return storage + 1*4 alignment.  */
-				   1*4 struct return +	
+#define	STACKFRAME	(16*4 + 4 + 6*4 + 8*4 + 4)
-				   6*4 args backing store +
-				   3*4 locals */
 
 /* ffi_closure_v8(...)
 
    Receives the closure argument in %g2.   */
 
-	.text
-	.align 8
-	.globl ffi_closure_v8
-
-ffi_closure_v8:
 #ifdef HAVE_AS_REGISTER_PSEUDO_OP
 	.register	%g2, #scratch
 #endif
-.LLFB2:
-	! Reserve frame space for all arguments in case
-	! we need to align them on a 8-byte boundary.
-	ld	[%g2+FFI_TRAMPOLINE_SIZE], %g1
-	ld	[%g1+4], %g1
-	sll	%g1, 3, %g1
-	add	%g1, STACKFRAME, %g1
-	! %g1 == STACKFRAME + 8*nargs
-	neg	%g1
-	save	%sp, %g1, %sp
-.LLCFI1:
 
+	.align 8
+	.globl	C(ffi_go_closure_v8)
+	.type	C(ffi_go_closure_v8),@function
+	FFI_HIDDEN(C(ffi_go_closure_v8))
+
+C(ffi_go_closure_v8):
+	cfi_startproc
+	save	%sp, -STACKFRAME, %sp
+	cfi_def_cfa_register(%fp)
+	cfi_window_save
+
+	ld	[%g2+4], %o0			! load cif
+	ld	[%g2+8], %o1			! load fun
+	b	0f
+	 mov	%g2, %o2			! load user_data
+	cfi_endproc
+	.size	C(ffi_go_closure_v8), . - C(ffi_go_closure_v8)
+
+	.align 8
+	.globl	C(ffi_closure_v8)
+	.type	C(ffi_closure_v8),@function
+	FFI_HIDDEN(C(ffi_closure_v8))
+
+C(ffi_closure_v8):
+	cfi_startproc
+	save	%sp, -STACKFRAME, %sp
+	cfi_def_cfa_register(%fp)
+	cfi_window_save
+
+	ld	[%g2+FFI_TRAMPOLINE_SIZE], %o0		! load cif
+	ld	[%g2+FFI_TRAMPOLINE_SIZE+4], %o1	! load fun
+	ld	[%g2+FFI_TRAMPOLINE_SIZE+8], %o2	! load user_data
+0:
 	! Store all of the potential argument registers in va_list format.
 	st	%i0, [%fp+68+0]
 	st	%i1, [%fp+68+4]
@@ -207,140 +271,91 @@ ffi_closure_v8:
 	st	%i5, [%fp+68+20]
 
 	! Call ffi_closure_sparc_inner to do the bulk of the work.
-	mov	%g2, %o0
+	add	%fp, -8*4, %o3
-	add	%fp, -8, %o1
-	add	%fp,  64, %o2
 	call	ffi_closure_sparc_inner_v8
-	 add	%fp, -16, %o3
+	 add	%fp,  64, %o4
 
-	! Load up the return value in the proper type.
+0:	call	1f
-	! See ffi_prep_cif_machdep for the list of cases.
+	 and	%o0, SPARC_FLAG_RET_MASK, %o0
-	cmp	%o0, FFI_TYPE_VOID
+1:	sll	%o0, 4, %o0	! o0 = o0 * 16
-	be	done1
+	add	%o7, %o0, %o7	! o7 = 0b + o0*16
+	jmp	%o7+(2f-0b)
+	 add	%fp, -8*4, %i2
 	 
-	cmp	%o0, FFI_TYPE_INT
+	! Note that each entry is 4 insns, enforced by the E macro.
-	be	done1
+	.align	16
-	 ld	[%fp-8], %i0
+2:
-
+E SPARC_RET_VOID
-	cmp	%o0, FFI_TYPE_FLOAT
+	ret
-	be,a	done1
-	 ld	[%fp-8], %f0
-
-	cmp	%o0, FFI_TYPE_DOUBLE
-	be,a	done1
-	 ldd	[%fp-8], %f0
-
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-	cmp	%o0, FFI_TYPE_LONGDOUBLE
-	be	done2
-#endif
-
-	cmp	%o0, FFI_TYPE_STRUCT
-	be	done2
-
-	cmp	%o0, FFI_TYPE_SINT64
-	be,a	done1
-	 ldd	[%fp-8], %i0
-
-	cmp	%o0, FFI_TYPE_UINT64
-	be,a	done1
-	 ldd	[%fp-8], %i0
-
-	ld	[%fp-8], %i0
-done1:
-	jmp	%i7+8
 	 restore
-done2:
+E SPARC_RET_STRUCT
-	! Skip 'unimp'.
+	ld	[%i2], %i0
 	jmp	%i7+12
 	 restore
-.LLFE2:
+E SPARC_RET_UINT8
-
+	ldub	[%i2+3], %i0
-.ffi_closure_v8_end:
+	ret
-	.size	ffi_closure_v8,.ffi_closure_v8_end-ffi_closure_v8
+	 restore
-
+E SPARC_RET_SINT8
-#ifdef SPARC64
+	ldsb	[%i2+3], %i0
-#define WS 8
+	ret
-#define nword	xword
+	 restore
-#define uanword	uaxword
+E SPARC_RET_UINT16
-#else
+	lduh	[%i2+2], %i0
-#define WS 4
+	ret
-#define nword	long
+	 restore
-#define uanword	uaword
+E SPARC_RET_SINT16
-#endif
+	ldsh	[%i2+2], %i0
-
+	ret
-#ifdef HAVE_RO_EH_FRAME
+	 restore
-	.section	".eh_frame",#alloc
+E SPARC_RET_UINT32
-#else
+	ld	[%i2], %i0
-	.section	".eh_frame",#alloc,#write
+	ret
-#endif
+	 restore
-.LLframe1:
+E SP_V8_RET_CPLX16
-	.uaword	.LLECIE1-.LLSCIE1	! Length of Common Information Entry
+	ld	[%i2], %i0
-.LLSCIE1:
+	ret
-	.uaword	0x0	! CIE Identifier Tag
+	 restore
-	.byte	0x1	! CIE Version
+E SPARC_RET_INT64
-	.ascii "zR\0"	! CIE Augmentation
+	ldd	[%i2], %i0
-	.byte	0x1	! uleb128 0x1; CIE Code Alignment Factor
+	ret
-	.byte	0x80-WS	! sleb128 -WS; CIE Data Alignment Factor
+	 restore
-	.byte	0xf	! CIE RA Column
+E SPARC_RET_INT128
-	.byte	0x1	! uleb128 0x1; Augmentation size
+	ldd	[%i2], %i0
-#ifdef HAVE_AS_SPARC_UA_PCREL
+	ldd	[%i2+8], %i2
-	.byte	0x1b	! FDE Encoding (pcrel sdata4)
+	ret
-#else
+	 restore
-	.byte	0x50	! FDE Encoding (aligned absolute)
+E SPARC_RET_F_8
-#endif
+	ld	[%i2+7*4], %f7
-	.byte	0xc	! DW_CFA_def_cfa
+	nop
-	.byte	0xe	! uleb128 0xe
+	ld	[%i2+6*4], %f6
-	.byte	0x0	! uleb128 0x0
+	nop
-	.align	WS
+E SPARC_RET_F_6
-.LLECIE1:
+	ld	[%i2+5*4], %f5
-.LLSFDE1:
+	nop
-	.uaword	.LLEFDE1-.LLASFDE1	! FDE Length
+	ld	[%i2+4*4], %f4
-.LLASFDE1:
+	nop
-	.uaword	.LLASFDE1-.LLframe1	! FDE CIE offset
+E SPARC_RET_F_4
-#ifdef HAVE_AS_SPARC_UA_PCREL
+	ld	[%i2+3*4], %f3
-	.uaword	%r_disp32(.LLFB1)
+	nop
-	.uaword	.LLFE1-.LLFB1	! FDE address range
+	ld	[%i2+2*4], %f2
-#else
+	nop
-	.align	WS
+E SPARC_RET_F_2
-	.nword	.LLFB1
+	ldd	[%i2], %f0
-	.uanword .LLFE1-.LLFB1	! FDE address range
+	ret
-#endif
+	 restore
-	.byte	0x0	! uleb128 0x0; Augmentation size
+E SP_V8_RET_CPLX8
-	.byte	0x4	! DW_CFA_advance_loc4
+	lduh	[%i2], %i0
-	.uaword	.LLCFI0-.LLFB1
+	ret
-	.byte	0xd	! DW_CFA_def_cfa_register
+	 restore
-	.byte	0x1e	! uleb128 0x1e
+E SPARC_RET_F_1
-	.byte	0x2d	! DW_CFA_GNU_window_save
+	ld	[%i2], %f0
-	.byte	0x9	! DW_CFA_register
+	ret
-	.byte	0xf	! uleb128 0xf
+	 restore
-	.byte	0x1f	! uleb128 0x1f
-	.align	WS
-.LLEFDE1:
-.LLSFDE2:
-	.uaword	.LLEFDE2-.LLASFDE2	! FDE Length
-.LLASFDE2:
-	.uaword	.LLASFDE2-.LLframe1	! FDE CIE offset
-#ifdef HAVE_AS_SPARC_UA_PCREL
-	.uaword	%r_disp32(.LLFB2)
-	.uaword	.LLFE2-.LLFB2	! FDE address range
-#else
-	.align	WS
-	.nword	.LLFB2
-	.uanword .LLFE2-.LLFB2	! FDE address range
-#endif
-	.byte	0x0	! uleb128 0x0; Augmentation size
-	.byte	0x4	! DW_CFA_advance_loc4
-	.uaword	.LLCFI1-.LLFB2
-	.byte	0xd	! DW_CFA_def_cfa_register
-	.byte	0x1e	! uleb128 0x1e
-	.byte	0x2d	! DW_CFA_GNU_window_save
-	.byte	0x9	! DW_CFA_register
-	.byte	0xf	! uleb128 0xf
-	.byte	0x1f	! uleb128 0x1f
-	.align	WS
-.LLEFDE2:
 
+	cfi_endproc
+	.size	C(ffi_closure_v8), . - C(ffi_closure_v8)
+#endif /* !SPARC64 */
 #if defined __ELF__ && defined __linux__
 	.section	.note.GNU-stack,"",@progbits
 #endif

src/sparc/v9.S

@@ -27,105 +27,178 @@
 #define LIBFFI_ASM	
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
+#include "internal.h"
 
 #ifdef SPARC64
-/* Only compile this in for 64bit builds, because otherwise the object file
-   will have inproper architecture due to used instructions.  */
 
-#define STACKFRAME 176		/* Minimum stack framesize for SPARC 64-bit */
+#define C2(X, Y)  X ## Y
+#define C1(X, Y)  C2(X, Y)
+
+#ifdef __USER_LABEL_PREFIX__
+# define C(Y)	C1(__USER_LABEL_PREFIX__, Y)
+#else
+# define C(Y)	Y
+#endif
+#define L(Y)	C1(.L, Y)
+
+.macro E index
+	.align	16
+	.org	2b + \index * 16
+.endm
+
 #define STACK_BIAS 2047
-#define ARGS (128)		/* Offset of register area in frame */
 
 	.text
         .align 8
-.globl ffi_call_v9
+	.globl	C(ffi_call_v9)
-.globl _ffi_call_v9
+	.type	C(ffi_call_v9),@function
+	FFI_HIDDEN(C(ffi_call_v9))
 
-ffi_call_v9:
+C(ffi_call_v9):
-_ffi_call_v9:
+	cfi_startproc
-.LLFB1:
+	save	%sp, %o4, %sp
-	save	%sp, -STACKFRAME, %sp
+	cfi_def_cfa_register(%fp)
-.LLCFI0:
+	cfi_window_save
 	
-	sub	%sp, %i2, %sp	! alloca() space in stack for frame to set up
+	mov	%i0, %o0			! copy cif
-	add	%sp, STACKFRAME+STACK_BIAS, %l0	! %l0 has start of 
+	add	%sp, STACK_BIAS+128+48, %o1	! load args area
-						! frame to set up
+	mov	%i2, %o2			! copy rvalue
+	call	C(ffi_prep_args_v9)
+	 mov	%i3, %o3			! copy avalue
 
-	mov	%l0, %o0	! call routine to set up frame
+	andcc	%o0, SPARC_FLAG_FP_ARGS, %g0	! need fp regs?
-	call	%i0
+	add	%sp, 48, %sp			! deallocate prep frame
-	 mov	%i1, %o1	! (delay)
+	be,pt	%xcc, 1f
-	brz,pt	%o0, 1f
+	 mov	%o0, %l0			! save flags
-	 ldx	[%l0+ARGS], %o0	! call foreign function
 
-	ldd	[%l0+ARGS], %f0
+	ldd	[%sp+STACK_BIAS+128], %f0	! load all fp arg regs
-	ldd	[%l0+ARGS+8], %f2
+	ldd	[%sp+STACK_BIAS+128+8], %f2
-	ldd	[%l0+ARGS+16], %f4
+	ldd	[%sp+STACK_BIAS+128+16], %f4
-	ldd	[%l0+ARGS+24], %f6
+	ldd	[%sp+STACK_BIAS+128+24], %f6
-	ldd	[%l0+ARGS+32], %f8
+	ldd	[%sp+STACK_BIAS+128+32], %f8
-	ldd	[%l0+ARGS+40], %f10
+	ldd	[%sp+STACK_BIAS+128+40], %f10
-	ldd	[%l0+ARGS+48], %f12
+	ldd	[%sp+STACK_BIAS+128+48], %f12
-	ldd	[%l0+ARGS+56], %f14
+	ldd	[%sp+STACK_BIAS+128+56], %f14
-	ldd	[%l0+ARGS+64], %f16
+	ldd	[%sp+STACK_BIAS+128+64], %f16
-	ldd	[%l0+ARGS+72], %f18
+	ldd	[%sp+STACK_BIAS+128+72], %f18
-	ldd	[%l0+ARGS+80], %f20
+	ldd	[%sp+STACK_BIAS+128+80], %f20
-	ldd	[%l0+ARGS+88], %f22
+	ldd	[%sp+STACK_BIAS+128+88], %f22
-	ldd	[%l0+ARGS+96], %f24
+	ldd	[%sp+STACK_BIAS+128+96], %f24
-	ldd	[%l0+ARGS+104], %f26
+	ldd	[%sp+STACK_BIAS+128+104], %f26
-	ldd	[%l0+ARGS+112], %f28
+	ldd	[%sp+STACK_BIAS+128+112], %f28
-	ldd	[%l0+ARGS+120], %f30
+	ldd	[%sp+STACK_BIAS+128+120], %f30
 
-1:	ldx	[%l0+ARGS+8], %o1
+1:	ldx	[%sp+STACK_BIAS+128], %o0	! load all int arg regs
-	ldx	[%l0+ARGS+16], %o2
+	ldx	[%sp+STACK_BIAS+128+8], %o1
-	ldx	[%l0+ARGS+24], %o3
+	ldx	[%sp+STACK_BIAS+128+16], %o2
-	ldx	[%l0+ARGS+32], %o4
+	ldx	[%sp+STACK_BIAS+128+24], %o3
-	ldx	[%l0+ARGS+40], %o5
+	ldx	[%sp+STACK_BIAS+128+32], %o4
-	call	%i5
+	ldx	[%sp+STACK_BIAS+128+40], %o5
-	 sub	%l0, STACK_BIAS, %sp	! (delay) switch to frame
+	call	%i1
+	 mov	%i5, %g5			! load static chain
 
-	! If the return value pointer is NULL, assume no return value.
+0:	call	1f		! load pc in %o7
-	brz,pn	%i4, done
+	 and	%l0, SPARC_FLAG_RET_MASK, %l1
+1:	sll	%l1, 4, %l1
+	add	%o7, %l1, %o7	! o7 = 0b + ret_type*16
+	jmp	%o7+(2f-0b)
 	 nop
 
-	cmp	%i3, FFI_TYPE_INT
+	.align	16
-	be,a,pt	%icc, done
+2:
-	 stx	%o0, [%i4+0]	! (delay)
+E SPARC_RET_VOID
-
+	return	%i7+8
-	cmp	%i3, FFI_TYPE_FLOAT
-	be,a,pn	%icc, done
-	 st	%f0, [%i4+0]	! (delay)
-
-	cmp	%i3, FFI_TYPE_DOUBLE
-	be,a,pn	%icc, done
-	 std	%f0, [%i4+0]	! (delay)
-
-	cmp	%i3, FFI_TYPE_STRUCT
-	be,pn	%icc, dostruct
-
-	cmp	%i3, FFI_TYPE_LONGDOUBLE
-	bne,pt	%icc, done
 	 nop
-	std	%f0, [%i4+0]
+E SPARC_RET_STRUCT
-	std	%f2, [%i4+8]
+	add	%sp, STACK_BIAS-64+128+48, %l2
+	sub	%sp, 64, %sp
+	b	8f
+	 stx	%o0, [%l2]
+E SPARC_RET_UINT8
+	and	%o0, 0xff, %i0
+	return	%i7+8
+	  stx	%o0, [%o2]
+E SPARC_RET_SINT8
+	sll	%o0, 24, %o0
+	sra	%o0, 24, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E SPARC_RET_UINT16
+	sll	%o0, 16, %o0
+	srl	%o0, 16, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E SPARC_RET_SINT16
+	sll	%o0, 16, %o0
+	sra	%o0, 16, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E SPARC_RET_UINT32
+	srl	%o0, 0, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E SP_V9_RET_SINT32
+	sra	%o0, 0, %i0
+	return	%i7+8
+	 stx	%o0, [%o2]
+E SPARC_RET_INT64
+	stx	%o0, [%i2]
+	return	%i7+8
+	 nop
+E SPARC_RET_INT128
+	stx	%o0, [%i2]
+	stx	%o1, [%i2+8]
+	return	%i7+8
+	 nop
+E SPARC_RET_F_8
+	st	%f7, [%i2+7*4]
+	nop
+	st	%f6, [%i2+6*4]
+	nop
+E SPARC_RET_F_6
+	st	%f5, [%i2+5*4]
+	nop
+	st	%f4, [%i2+4*4]
+	nop
+E SPARC_RET_F_4
+	std	%f2, [%i2+2*4]
+	return	%i7+8
+	 std	%f0, [%o2]
+E SPARC_RET_F_2
+	return	%i7+8
+	 std	%f0, [%o2]
+E SP_V9_RET_F_3
+	st	%f2, [%i2+2*4]
+	nop
+	st	%f1, [%i2+1*4]
+	nop
+E SPARC_RET_F_1
+	return	%i7+8
+	 st	%f0, [%o2]
 
-done:	ret
+	! Finish the SPARC_RET_STRUCT sequence.
-	 restore
+	.align	8
+8:	stx	%o1, [%l2+8]
+	stx	%o2, [%l2+16]
+	stx	%o3, [%l2+24]
+	std	%f0, [%l2+32]
+	std	%f2, [%l2+40]
+	std	%f4, [%l2+48]
+	std	%f6, [%l2+56]
 
-dostruct:
+	! Copy the structure into place.
-	/* This will not work correctly for unions. */
+	srl	%l0, SPARC_SIZEMASK_SHIFT, %o0	! load size_mask
-	stx	%o0, [%i4+0]
+	mov	%i2, %o1			! load dst
-	stx	%o1, [%i4+8]
+	mov	%l2, %o2			! load src_gp
-	stx	%o2, [%i4+16]
+	call	C(ffi_struct_float_copy)
-	stx	%o3, [%i4+24]
+	 add	%l2, 32, %o3			! load src_fp
-	std	%f0, [%i4+32]
-	std	%f2, [%i4+40]
-	std	%f4, [%i4+48]
-	std	%f6, [%i4+56]
-	ret
-	 restore
-.LLFE1:
 
-.ffi_call_v9_end:
+	return	%i7+8
-	.size	ffi_call_v9,.ffi_call_v9_end-ffi_call_v9
+	 nop
+
+	cfi_endproc
+	.size	C(ffi_call_v9), . - C(ffi_call_v9)
 
 
 #undef STACKFRAME
@@ -138,15 +211,40 @@ dostruct:
 
    Receives the closure argument in %g1.   */
 
-	.text
 	.align 8
-	.globl ffi_closure_v9
+	.globl	C(ffi_go_closure_v9)
+	.type	C(ffi_go_closure_v9),@function
+	FFI_HIDDEN(C(ffi_go_closure_v9))
 
-ffi_closure_v9:
+C(ffi_go_closure_v9):
-.LLFB2:
+	cfi_startproc
 	save	%sp, -STACKFRAME, %sp
-.LLCFI1:
+	cfi_def_cfa_register(%fp)
+	cfi_window_save
 
+	ldx	[%g5+8], %o0
+	ldx	[%g5+16], %o1
+	b	0f
+	 mov	%g5, %o2
+
+	cfi_endproc
+	.size	C(ffi_go_closure_v9), . - C(ffi_go_closure_v9)
+
+	.align 8
+	.globl	C(ffi_closure_v9)
+	.type	C(ffi_closure_v9),@function
+	FFI_HIDDEN(C(ffi_closure_v9))
+
+C(ffi_closure_v9):
+	cfi_startproc
+	save	%sp, -STACKFRAME, %sp
+	cfi_def_cfa_register(%fp)
+	cfi_window_save
+
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE], %o0
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE+8], %o1
+	ldx	[%g1+FFI_TRAMPOLINE_SIZE+16], %o2
+0:
 	! Store all of the potential argument registers in va_list format.
 	stx	%i0, [FP+128+0]
 	stx	%i1, [FP+128+8]
@@ -174,134 +272,106 @@ ffi_closure_v9:
 	std     %f30, [FP-8]
 
 	! Call ffi_closure_sparc_inner to do the bulk of the work.
-	mov	%g1, %o0
+	add	%fp, STACK_BIAS-160, %o3
-	add	%fp, STACK_BIAS-160, %o1
+	add	%fp, STACK_BIAS+128, %o4
-	add	%fp, STACK_BIAS+128, %o2
+	call	C(ffi_closure_sparc_inner_v9)
-	call	ffi_closure_sparc_inner_v9
+	 add	%fp, STACK_BIAS-128, %o5
-	 add	%fp, STACK_BIAS-128, %o3
 
-	! Load up the return value in the proper type.
+0:	call	1f		! load pc in %o7
-	! See ffi_prep_cif_machdep for the list of cases.
+	 and	%o0, SPARC_FLAG_RET_MASK, %o0
-	cmp	%o0, FFI_TYPE_VOID
+1:	sll	%o0, 4, %o0	! o2 = i2 * 16
-	be,pn	%icc, done1
+	add	%o7, %o0, %o7	! o7 = 0b + i2*16
+	jmp	%o7+(2f-0b)
+	 nop
 
-	cmp	%o0, FFI_TYPE_INT
+	! Note that we cannot load the data in the delay slot of
-	be,pn	%icc, integer
+	! the return insn because the data is in the stack frame
-
+	! that is deallocated by the return.
-	cmp	%o0, FFI_TYPE_FLOAT
+	.align	16
-	be,a,pn	%icc, done1
+2:
-	 ld	[FP-160], %f0
+E SPARC_RET_VOID
-
+	return	%i7+8
-	cmp	%o0, FFI_TYPE_DOUBLE
+	 nop
-	be,a,pn	%icc, done1
+E SPARC_RET_STRUCT
-	 ldd	[FP-160], %f0
-
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
-	cmp	%o0, FFI_TYPE_LONGDOUBLE
-	be,a,pn	%icc, longdouble1
-	 ldd	[FP-160], %f0
-#endif
-
-	! FFI_TYPE_STRUCT
-	ldx	[FP-152], %i1
-	ldx	[FP-144], %i2
-	ldx	[FP-136], %i3
-	ldd	[FP-160], %f0
-	ldd	[FP-152], %f2
-	ldd	[FP-144], %f4
-	ldd	[FP-136], %f6
-
-integer:
 	ldx	[FP-160], %i0
+	ldd	[FP-160], %f0
+	b	8f
+	 ldx	[FP-152], %i1
+E SPARC_RET_UINT8
+	ldub	[FP-160+7], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_SINT8
+	ldsb	[FP-160+7], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_UINT16
+	lduh	[FP-160+6], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_SINT16
+	ldsh	[FP-160+6], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_UINT32
+	lduw	[FP-160+4], %i0
+	return	%i7+8
+	 nop
+E SP_V9_RET_SINT32
+	ldsw	[FP-160+4], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_INT64
+	ldx	[FP-160], %i0
+	return	%i7+8
+	 nop
+E SPARC_RET_INT128
+	ldx	[FP-160], %i0
+	ldx	[FP-160+8], %i1
+	return	%i7+8
+	 nop
+E SPARC_RET_F_8
+	ld	[FP-160+7*4], %f7
+	nop
+	ld	[FP-160+6*4], %f6
+	nop
+E SPARC_RET_F_6
+	ld	[FP-160+5*4], %f5
+	nop
+	ld	[FP-160+4*4], %f4
+	nop
+E SPARC_RET_F_4
+	ldd	[FP-160], %f0
+	ldd	[FP-160+8], %f2
+	return	%i7+8
+	 nop
+E SPARC_RET_F_2
+	ldd	[FP-160], %f0
+	return	%i7+8
+	 nop
+E SP_V9_RET_F_3
+	ld	[FP-160+2*4], %f2
+	nop
+	ld	[FP-160+1*4], %f1
+	nop
+E SPARC_RET_F_1
+	ld	[FP-160], %f0
+	return	%i7+8
+	 nop
 
-done1:
+	! Finish the SPARC_RET_STRUCT sequence.
-	ret
+	.align	8
-	 restore
+8:	ldd	[FP-152], %f2
-
+	ldx	[FP-144], %i2
-#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+	ldd	[FP-144], %f4
-longdouble1:
+	ldx	[FP-136], %i3
-	ldd	[FP-152], %f2
+	ldd	[FP-136], %f6
-	ret
+	return	%i7+8
-	 restore
+	 nop
-#endif
-.LLFE2:
-
-.ffi_closure_v9_end:
-	.size	ffi_closure_v9,.ffi_closure_v9_end-ffi_closure_v9
-
-#ifdef HAVE_RO_EH_FRAME
-	.section	".eh_frame",#alloc
-#else
-	.section	".eh_frame",#alloc,#write
-#endif
-.LLframe1:
-	.uaword	.LLECIE1-.LLSCIE1	! Length of Common Information Entry
-.LLSCIE1:
-	.uaword	0x0	! CIE Identifier Tag
-	.byte	0x1	! CIE Version
-	.ascii "zR\0"	! CIE Augmentation
-	.byte	0x1	! uleb128 0x1; CIE Code Alignment Factor
-	.byte	0x78	! sleb128 -8; CIE Data Alignment Factor
-	.byte	0xf	! CIE RA Column
-	.byte	0x1	! uleb128 0x1; Augmentation size
-#ifdef HAVE_AS_SPARC_UA_PCREL
-	.byte	0x1b	! FDE Encoding (pcrel sdata4)
-#else
-	.byte	0x50	! FDE Encoding (aligned absolute)
-#endif
-	.byte	0xc	! DW_CFA_def_cfa
-	.byte	0xe	! uleb128 0xe
-	.byte	0xff,0xf	! uleb128 0x7ff
-	.align 8
-.LLECIE1:
-.LLSFDE1:
-	.uaword	.LLEFDE1-.LLASFDE1	! FDE Length
-.LLASFDE1:
-	.uaword	.LLASFDE1-.LLframe1	! FDE CIE offset
-#ifdef HAVE_AS_SPARC_UA_PCREL
-	.uaword	%r_disp32(.LLFB1)
-	.uaword	.LLFE1-.LLFB1		! FDE address range
-#else
-	.align 8
-	.xword	.LLFB1
-	.uaxword	.LLFE1-.LLFB1	! FDE address range
-#endif
-	.byte	0x0	! uleb128 0x0; Augmentation size
-	.byte	0x4	! DW_CFA_advance_loc4
-	.uaword	.LLCFI0-.LLFB1
-	.byte	0xd	! DW_CFA_def_cfa_register
-	.byte	0x1e	! uleb128 0x1e
-	.byte	0x2d	! DW_CFA_GNU_window_save
-	.byte	0x9	! DW_CFA_register
-	.byte	0xf	! uleb128 0xf
-	.byte	0x1f	! uleb128 0x1f
-	.align 8
-.LLEFDE1:
-.LLSFDE2:
-	.uaword	.LLEFDE2-.LLASFDE2	! FDE Length
-.LLASFDE2:
-	.uaword	.LLASFDE2-.LLframe1	! FDE CIE offset
-#ifdef HAVE_AS_SPARC_UA_PCREL
-	.uaword	%r_disp32(.LLFB2)
-	.uaword	.LLFE2-.LLFB2		! FDE address range
-#else
-	.align 8
-	.xword	.LLFB2
-	.uaxword	.LLFE2-.LLFB2	! FDE address range
-#endif
-	.byte	0x0	! uleb128 0x0; Augmentation size
-	.byte	0x4	! DW_CFA_advance_loc4
-	.uaword	.LLCFI1-.LLFB2
-	.byte	0xd	! DW_CFA_def_cfa_register
-	.byte	0x1e	! uleb128 0x1e
-	.byte	0x2d	! DW_CFA_GNU_window_save
-	.byte	0x9	! DW_CFA_register
-	.byte	0xf	! uleb128 0xf
-	.byte	0x1f	! uleb128 0x1f
-	.align 8
-.LLEFDE2:
-#endif
 
+	cfi_endproc
+	.size	C(ffi_closure_v9), . - C(ffi_closure_v9)
+#endif /* SPARC64 */
 #ifdef __linux__
 	.section	.note.GNU-stack,"",@progbits
 #endif

src/x86/ffi64.c

@@ -32,6 +32,8 @@
 
 #include <stdlib.h>
 #include <stdarg.h>
+#include <stdint.h>
+#include "internal64.h"
 
 #ifdef __x86_64__
 
@@ -62,10 +64,12 @@ struct register_args
   /* Registers for argument passing.  */
   UINT64 gpr[MAX_GPR_REGS];
   union big_int_union sse[MAX_SSE_REGS];
+  UINT64 rax;	/* ssecount */
+  UINT64 r10;	/* static chain */
 };
 
 extern void ffi_call_unix64 (void *args, unsigned long bytes, unsigned flags,
-			     void *raddr, void (*fnaddr)(void), unsigned ssecount);
+			     void *raddr, void (*fnaddr)(void)) FFI_HIDDEN;
 
 /* All reference to register classes here is identical to the code in
    gcc/config/i386/i386.c. Do *not* change one without the other.  */
@@ -167,6 +171,7 @@ classify_argument (ffi_type *type, enum x86_64_reg_class classes[],
     case FFI_TYPE_UINT64:
     case FFI_TYPE_SINT64:
     case FFI_TYPE_POINTER:
+    do_integer:
       {
 	size_t size = byte_offset + type->size;
 
@@ -188,7 +193,7 @@ classify_argument (ffi_type *type, enum x86_64_reg_class classes[],
 	  }
 	else if (size <= 16)
 	  {
-	    classes[0] = classes[1] = X86_64_INTEGERSI_CLASS;
+	    classes[0] = classes[1] = X86_64_INTEGER_CLASS;
 	    return 2;
 	  }
 	else
@@ -297,11 +302,42 @@ classify_argument (ffi_type *type, enum x86_64_reg_class classes[],
 	  }
 	return words;
       }
+    case FFI_TYPE_COMPLEX:
+      {
+	ffi_type *inner = type->elements[0];
+	switch (inner->type)
+	  {
+	  case FFI_TYPE_INT:
+	  case FFI_TYPE_UINT8:
+	  case FFI_TYPE_SINT8:
+	  case FFI_TYPE_UINT16:
+	  case FFI_TYPE_SINT16:
+	  case FFI_TYPE_UINT32:
+	  case FFI_TYPE_SINT32:
+	  case FFI_TYPE_UINT64:
+	  case FFI_TYPE_SINT64:
+	    goto do_integer;
 
-    default:
+	  case FFI_TYPE_FLOAT:
-      FFI_ASSERT(0);
+	    classes[0] = X86_64_SSE_CLASS;
+	    if (byte_offset % 8)
+	      {
+		classes[1] = X86_64_SSESF_CLASS;
+		return 2;
 	      }
-  return 0; /* Never reached.  */
+	    return 1;
+	  case FFI_TYPE_DOUBLE:
+	    classes[0] = classes[1] = X86_64_SSEDF_CLASS;
+	    return 2;
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+	  case FFI_TYPE_LONGDOUBLE:
+	    classes[0] = X86_64_COMPLEX_X87_CLASS;
+	    return 1;
+#endif
+	  }
+      }
+    }
+  abort();
 }
 
 /* Examine the argument and return set number of register required in each
@@ -356,13 +392,57 @@ ffi_prep_cif_machdep (ffi_cif *cif)
 {
   int gprcount, ssecount, i, avn, ngpr, nsse, flags;
   enum x86_64_reg_class classes[MAX_CLASSES];
-  size_t bytes, n;
+  size_t bytes, n, rtype_size;
+  ffi_type *rtype;
+
+  if (cif->abi != FFI_UNIX64)
+    return FFI_BAD_ABI;
 
   gprcount = ssecount = 0;
 
-  flags = cif->rtype->type;
+  rtype = cif->rtype;
-  if (flags != FFI_TYPE_VOID)
+  rtype_size = rtype->size;
+  switch (rtype->type)
     {
+    case FFI_TYPE_VOID:
+      flags = UNIX64_RET_VOID;
+      break;
+    case FFI_TYPE_UINT8:
+      flags = UNIX64_RET_UINT8;
+      break;
+    case FFI_TYPE_SINT8:
+      flags = UNIX64_RET_SINT8;
+      break;
+    case FFI_TYPE_UINT16:
+      flags = UNIX64_RET_UINT16;
+      break;
+    case FFI_TYPE_SINT16:
+      flags = UNIX64_RET_SINT16;
+      break;
+    case FFI_TYPE_UINT32:
+      flags = UNIX64_RET_UINT32;
+      break;
+    case FFI_TYPE_INT:
+    case FFI_TYPE_SINT32:
+      flags = UNIX64_RET_SINT32;
+      break;
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+      flags = UNIX64_RET_INT64;
+      break;
+    case FFI_TYPE_POINTER:
+      flags = (sizeof(void *) == 4 ? UNIX64_RET_UINT32 : UNIX64_RET_INT64);
+      break;
+    case FFI_TYPE_FLOAT:
+      flags = UNIX64_RET_XMM32;
+      break;
+    case FFI_TYPE_DOUBLE:
+      flags = UNIX64_RET_XMM64;
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+      flags = UNIX64_RET_X87;
+      break;
+    case FFI_TYPE_STRUCT:
       n = examine_argument (cif->rtype, classes, 1, &ngpr, &nsse);
       if (n == 0)
 	{
@@ -370,23 +450,63 @@ ffi_prep_cif_machdep (ffi_cif *cif)
 	     memory is the first argument.  Allocate a register for it.  */
 	  gprcount++;
 	  /* We don't have to do anything in asm for the return.  */
-	  flags = FFI_TYPE_VOID;
+	  flags = UNIX64_RET_VOID | UNIX64_FLAG_RET_IN_MEM;
 	}
-      else if (flags == FFI_TYPE_STRUCT)
+      else
 	{
-	  /* Mark which registers the result appears in.  */
 	  _Bool sse0 = SSE_CLASS_P (classes[0]);
+
+	  if (rtype_size == 4 && sse0)
+	    flags = UNIX64_RET_XMM32;
+	  else if (rtype_size == 8)
+	    flags = sse0 ? UNIX64_RET_XMM64 : UNIX64_RET_INT64;
+	  else
+	    {
 	      _Bool sse1 = n == 2 && SSE_CLASS_P (classes[1]);
-	  if (sse0 && !sse1)
+	      if (sse0 && sse1)
-	    flags |= 1 << 8;
+		flags = UNIX64_RET_ST_XMM0_XMM1;
-	  else if (!sse0 && sse1)
+	      else if (sse0)
-	    flags |= 1 << 9;
+		flags = UNIX64_RET_ST_XMM0_RAX;
-	  else if (sse0 && sse1)
+	      else if (sse1)
-	    flags |= 1 << 10;
+		flags = UNIX64_RET_ST_RAX_XMM0;
-	  /* Mark the true size of the structure.  */
+	      else
-	  flags |= cif->rtype->size << 12;
+		flags = UNIX64_RET_ST_RAX_RDX;
+	      flags |= rtype_size << UNIX64_SIZE_SHIFT;
 	    }
 	}
+      break;
+    case FFI_TYPE_COMPLEX:
+      switch (rtype->elements[0]->type)
+	{
+	case FFI_TYPE_UINT8:
+	case FFI_TYPE_SINT8:
+	case FFI_TYPE_UINT16:
+	case FFI_TYPE_SINT16:
+	case FFI_TYPE_INT:
+	case FFI_TYPE_UINT32:
+	case FFI_TYPE_SINT32:
+	case FFI_TYPE_UINT64:
+	case FFI_TYPE_SINT64:
+	  flags = UNIX64_RET_ST_RAX_RDX | (rtype_size << UNIX64_SIZE_SHIFT);
+	  break;
+	case FFI_TYPE_FLOAT:
+	  flags = UNIX64_RET_XMM64;
+	  break;
+	case FFI_TYPE_DOUBLE:
+	  flags = UNIX64_RET_ST_XMM0_XMM1 | (16 << UNIX64_SIZE_SHIFT);
+	  break;
+#if FFI_TYPE_LONGDOUBLE != FFI_TYPE_DOUBLE
+	case FFI_TYPE_LONGDOUBLE:
+	  flags = UNIX64_RET_X87_2;
+	  break;
+#endif
+	default:
+	  return FFI_BAD_TYPEDEF;
+	}
+      break;
+    default:
+      return FFI_BAD_TYPEDEF;
+    }
 
   /* Go over all arguments and determine the way they should be passed.
      If it's in a register and there is space for it, let that be so. If
@@ -412,44 +532,50 @@ ffi_prep_cif_machdep (ffi_cif *cif)
 	}
     }
   if (ssecount)
-    flags |= 1 << 11;
+    flags |= UNIX64_FLAG_XMM_ARGS;
+
   cif->flags = flags;
-  cif->bytes = (unsigned)ALIGN (bytes, 8);
+  cif->bytes = ALIGN (bytes, 8);
 
   return FFI_OK;
 }
 
-void
+static void
-ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	      void **avalue, void *closure)
 {
   enum x86_64_reg_class classes[MAX_CLASSES];
   char *stack, *argp;
   ffi_type **arg_types;
-  int gprcount, ssecount, ngpr, nsse, i, avn;
+  int gprcount, ssecount, ngpr, nsse, i, avn, flags;
-  _Bool ret_in_memory;
   struct register_args *reg_args;
 
   /* Can't call 32-bit mode from 64-bit mode.  */
   FFI_ASSERT (cif->abi == FFI_UNIX64);
 
   /* If the return value is a struct and we don't have a return value
-     address then we need to make one.  Note the setting of flags to
+     address then we need to make one.  Otherwise we can ignore it.  */
-     VOID above in ffi_prep_cif_machdep.  */
+  flags = cif->flags;
-  ret_in_memory = (cif->rtype->type == FFI_TYPE_STRUCT
+  if (rvalue == NULL)
-		   && (cif->flags & 0xff) == FFI_TYPE_VOID);
+    {
-  if (rvalue == NULL && ret_in_memory)
+      if (flags & UNIX64_FLAG_RET_IN_MEM)
 	rvalue = alloca (cif->rtype->size);
+      else
+	flags = UNIX64_RET_VOID;
+    }
 
   /* Allocate the space for the arguments, plus 4 words of temp space.  */
   stack = alloca (sizeof (struct register_args) + cif->bytes + 4*8);
   reg_args = (struct register_args *) stack;
   argp = stack + sizeof (struct register_args);
 
+  reg_args->r10 = (uintptr_t) closure;
+
   gprcount = ssecount = 0;
 
   /* If the return value is passed in memory, add the pointer as the
      first integer argument.  */
-  if (ret_in_memory)
+  if (flags & UNIX64_FLAG_RET_IN_MEM)
     reg_args->gpr[gprcount++] = (unsigned long) rvalue;
 
   avn = cif->nargs;
@@ -494,17 +620,17 @@ ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 		  switch (arg_types[i]->type)
 		    {
 		    case FFI_TYPE_SINT8:
-		      *(SINT64 *)&reg_args->gpr[gprcount] = (SINT64) *((SINT8 *) a);
+		      reg_args->gpr[gprcount] = (SINT64) *((SINT8 *) a);
 		      break;
 		    case FFI_TYPE_SINT16:
-		      *(SINT64 *)&reg_args->gpr[gprcount] = (SINT64) *((SINT16 *) a);
+		      reg_args->gpr[gprcount] = (SINT64) *((SINT16 *) a);
 		      break;
 		    case FFI_TYPE_SINT32:
-		      *(SINT64 *)&reg_args->gpr[gprcount] = (SINT64) *((SINT32 *) a);
+		      reg_args->gpr[gprcount] = (SINT64) *((SINT32 *) a);
 		      break;
 		    default:
 		      reg_args->gpr[gprcount] = 0;
-		      memcpy (&reg_args->gpr[gprcount], a, size < 8 ? size : 8);
+		      memcpy (&reg_args->gpr[gprcount], a, size);
 		    }
 		  gprcount++;
 		  break;
@@ -521,13 +647,27 @@ ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 	    }
 	}
     }
+  reg_args->rax = ssecount;
 
   ffi_call_unix64 (stack, cif->bytes + sizeof (struct register_args),
-		   cif->flags, rvalue, fn, ssecount);
+		   flags, rvalue, fn);
 }
 
+void
+ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, NULL);
+}
 
-extern void ffi_closure_unix64(void);
+void
+ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	     void **avalue, void *closure)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, closure);
+}
+
+extern void ffi_closure_unix64(void) FFI_HIDDEN;
+extern void ffi_closure_unix64_sse(void) FFI_HIDDEN;
 
 ffi_status
 ffi_prep_closure_loc (ffi_closure* closure,
@@ -536,29 +676,27 @@ ffi_prep_closure_loc (ffi_closure* closure,
 		      void *user_data,
 		      void *codeloc)
 {
-  volatile unsigned short *tramp;
+  static const unsigned char trampoline[16] = {
+    /* leaq  -0x7(%rip),%r10   # 0x0  */
+    0x4c, 0x8d, 0x15, 0xf9, 0xff, 0xff, 0xff,
+    /* jmpq  *0x3(%rip)        # 0x10 */
+    0xff, 0x25, 0x03, 0x00, 0x00, 0x00,
+    /* nopl  (%rax) */
+    0x0f, 0x1f, 0x00
+  };
+  void (*dest)(void);
+  char *tramp = closure->tramp;
 
-  /* Sanity check on the cif ABI.  */
+  if (cif->abi != FFI_UNIX64)
-  {
-    int abi = cif->abi;
-    if (UNLIKELY (! (abi > FFI_FIRST_ABI && abi < FFI_LAST_ABI)))
     return FFI_BAD_ABI;
-  }
 
-  tramp = (volatile unsigned short *) &closure->tramp[0];
+  if (cif->flags & UNIX64_FLAG_XMM_ARGS)
+    dest = ffi_closure_unix64_sse;
+  else
+    dest = ffi_closure_unix64;
 
-  tramp[0] = 0xbb49;		/* mov <code>, %r11	*/
+  memcpy (tramp, trampoline, sizeof(trampoline));
-  *((unsigned long long * volatile) &tramp[1])
+  *(UINT64 *)(tramp + 16) = (uintptr_t)dest;
-    = (unsigned long) ffi_closure_unix64;
-  tramp[5] = 0xba49;		/* mov <data>, %r10	*/
-  *((unsigned long long * volatile) &tramp[6])
-    = (unsigned long) codeloc;
-
-  /* Set the carry bit iff the function uses any sse registers.
-     This is clc or stc, together with the first byte of the jmp.  */
-  tramp[10] = cif->flags & (1 << 11) ? 0x49f9 : 0x49f8;
-
-  tramp[11] = 0xe3ff;			/* jmp *%r11    */
 
   closure->cif = cif;
   closure->fun = fun;
@@ -567,49 +705,29 @@ ffi_prep_closure_loc (ffi_closure* closure,
   return FFI_OK;
 }
 
-int
+int FFI_HIDDEN
-ffi_closure_unix64_inner(ffi_closure *closure, void *rvalue,
+ffi_closure_unix64_inner(ffi_cif *cif,
-			 struct register_args *reg_args, char *argp)
+			 void (*fun)(ffi_cif*, void*, void**, void*),
+			 void *user_data,
+			 void *rvalue,
+			 struct register_args *reg_args,
+			 char *argp)
 {
-  ffi_cif *cif;
   void **avalue;
   ffi_type **arg_types;
   long i, avn;
   int gprcount, ssecount, ngpr, nsse;
-  int ret;
+  int flags;
-
-  cif = closure->cif;
-  avalue = alloca(cif->nargs * sizeof(void *));
-  gprcount = ssecount = 0;
-
-  ret = cif->rtype->type;
-  if (ret != FFI_TYPE_VOID)
-    {
-      enum x86_64_reg_class classes[MAX_CLASSES];
-      size_t n = examine_argument (cif->rtype, classes, 1, &ngpr, &nsse);
-      if (n == 0)
-	{
-	  /* The return value goes in memory.  Arrange for the closure
-	     return value to go directly back to the original caller.  */
-	  rvalue = (void *) (unsigned long) reg_args->gpr[gprcount++];
-	  /* We don't have to do anything in asm for the return.  */
-	  ret = FFI_TYPE_VOID;
-	}
-      else if (ret == FFI_TYPE_STRUCT && n == 2)
-	{
-	  /* Mark which register the second word of the structure goes in.  */
-	  _Bool sse0 = SSE_CLASS_P (classes[0]);
-	  _Bool sse1 = SSE_CLASS_P (classes[1]);
-	  if (!sse0 && sse1)
-	    ret |= 1 << 8;
-	  else if (sse0 && !sse1)
-	    ret |= 1 << 9;
-	}
-    }
 
   avn = cif->nargs;
-  arg_types = cif->arg_types;
+  flags = cif->flags;
+  avalue = alloca(avn * sizeof(void *));
+  gprcount = ssecount = 0;
 
+  if (flags & UNIX64_FLAG_RET_IN_MEM)
+    rvalue = (void *)(uintptr_t)reg_args->gpr[gprcount++];
+
+  arg_types = cif->arg_types;
   for (i = 0; i < avn; ++i)
     {
       enum x86_64_reg_class classes[MAX_CLASSES];
@@ -667,10 +785,29 @@ ffi_closure_unix64_inner(ffi_closure *closure, void *rvalue,
     }
 
   /* Invoke the closure.  */
-  closure->fun (cif, rvalue, avalue, closure->user_data);
+  fun (cif, rvalue, avalue, user_data);
 
   /* Tell assembly how to perform return type promotions.  */
-  return ret;
+  return flags;
+}
+
+extern void ffi_go_closure_unix64(void) FFI_HIDDEN;
+extern void ffi_go_closure_unix64_sse(void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_go_closure (ffi_go_closure* closure, ffi_cif* cif,
+		     void (*fun)(ffi_cif*, void*, void**, void*))
+{
+  if (cif->abi != FFI_UNIX64)
+    return FFI_BAD_ABI;
+
+  closure->tramp = (cif->flags & UNIX64_FLAG_XMM_ARGS
+		    ? ffi_go_closure_unix64_sse
+		    : ffi_go_closure_unix64);
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
 }
 
 #endif /* __x86_64__ */

src/x86/ffitarget.h

@@ -76,44 +76,40 @@ typedef signed long            ffi_sarg;
 #endif
 
 typedef enum ffi_abi {
+#if defined(X86_WIN64)
   FFI_FIRST_ABI = 0,
-
-  /* ---- Intel x86 Win32 ---------- */
-#ifdef X86_WIN32
-  FFI_SYSV,
-  FFI_STDCALL,
-  FFI_THISCALL,
-  FFI_FASTCALL,
-  FFI_MS_CDECL,
-  FFI_PASCAL,
-  FFI_REGISTER,
-  FFI_LAST_ABI,
-#ifdef _MSC_VER
-  FFI_DEFAULT_ABI = FFI_MS_CDECL
-#else
-  FFI_DEFAULT_ABI = FFI_SYSV
-#endif
-
-#elif defined(X86_WIN64)
   FFI_WIN64,
   FFI_LAST_ABI,
   FFI_DEFAULT_ABI = FFI_WIN64
 
-#else
+#elif defined(X86_64) || (defined (__x86_64__) && defined (X86_DARWIN))
-  /* ---- Intel x86 and AMD x86-64 - */
+  FFI_FIRST_ABI = 1,
-  FFI_SYSV,
+  FFI_UNIX64,
-  FFI_UNIX64,   /* Unix variants all use the same ABI for x86-64  */
-  FFI_THISCALL,
-  FFI_FASTCALL,
-  FFI_STDCALL,
-  FFI_PASCAL,
-  FFI_REGISTER,
   FFI_LAST_ABI,
-#if defined(__i386__) || defined(__i386)
-  FFI_DEFAULT_ABI = FFI_SYSV
-#else
   FFI_DEFAULT_ABI = FFI_UNIX64
-#endif
+
+#elif defined(X86_WIN32)
+  FFI_FIRST_ABI = 0,
+  FFI_SYSV      = 1,
+  FFI_STDCALL   = 2,
+  FFI_THISCALL  = 3,
+  FFI_FASTCALL  = 4,
+  FFI_MS_CDECL  = 5,
+  FFI_PASCAL    = 6,
+  FFI_REGISTER  = 7,
+  FFI_LAST_ABI,
+  FFI_DEFAULT_ABI = FFI_MS_CDECL
+#else
+  FFI_FIRST_ABI = 0,
+  FFI_SYSV      = 1,
+  FFI_THISCALL  = 3,
+  FFI_FASTCALL  = 4,
+  FFI_STDCALL   = 5,
+  FFI_PASCAL    = 6,
+  FFI_REGISTER  = 7,
+  FFI_MS_CDECL  = 8,
+  FFI_LAST_ABI,
+  FFI_DEFAULT_ABI = FFI_SYSV
 #endif
 } ffi_abi;
 #endif
@@ -121,30 +117,21 @@ typedef enum ffi_abi {
 /* ---- Definitions for closures ----------------------------------------- */
 
 #define FFI_CLOSURES 1
+#define FFI_GO_CLOSURES 1
+
 #define FFI_TYPE_SMALL_STRUCT_1B (FFI_TYPE_LAST + 1)
 #define FFI_TYPE_SMALL_STRUCT_2B (FFI_TYPE_LAST + 2)
 #define FFI_TYPE_SMALL_STRUCT_4B (FFI_TYPE_LAST + 3)
 #define FFI_TYPE_MS_STRUCT       (FFI_TYPE_LAST + 4)
 
-#if defined (X86_64) || (defined (__x86_64__) && defined (X86_DARWIN))
+#if defined (X86_64) || defined(X86_WIN64) \
+    || (defined (__x86_64__) && defined (X86_DARWIN))
 # define FFI_TRAMPOLINE_SIZE 24
 # define FFI_NATIVE_RAW_API 0
 #else
-#ifdef X86_WIN32
+# define FFI_TRAMPOLINE_SIZE 12
-#define FFI_TRAMPOLINE_SIZE 52
-#else
-#ifdef X86_WIN64
-#define FFI_TRAMPOLINE_SIZE 29
-#define FFI_NATIVE_RAW_API 0
-#define FFI_NO_RAW_API 1
-#else
-#define FFI_TRAMPOLINE_SIZE 10
-#endif
-#endif
-#ifndef X86_WIN64
 # define FFI_NATIVE_RAW_API 1  /* x86 has native raw api support */
 #endif
-#endif
 
 #endif
 

src/x86/ffiw64.c

@@ -0,0 +1,281 @@
+/* -----------------------------------------------------------------------
+   ffiw64.c - Copyright (c) 2014 Red Hat, Inc.
+
+   x86 win64 Foreign Function Interface
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   ``Software''), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be included
+   in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+   ----------------------------------------------------------------------- */
+
+#include <ffi.h>
+#include <ffi_common.h>
+#include <stdlib.h>
+#include <stdint.h>
+
+#ifdef X86_WIN64
+
+struct win64_call_frame
+{
+  UINT64 rbp;		/* 0 */
+  UINT64 retaddr;	/* 8 */
+  UINT64 fn;		/* 16 */
+  UINT64 flags;		/* 24 */
+  UINT64 rvalue;	/* 32 */
+};
+
+extern void ffi_call_win64 (void *stack, struct win64_call_frame *,
+			    void *closure) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_cif_machdep (ffi_cif *cif)
+{
+  int flags, n;
+
+  if (cif->abi != FFI_WIN64)
+    return FFI_BAD_ABI;
+
+  flags = cif->rtype->type;
+  switch (flags)
+    {
+    default:
+      break;
+    case FFI_TYPE_LONGDOUBLE:
+      flags = FFI_TYPE_STRUCT;
+      break;
+    case FFI_TYPE_COMPLEX:
+      flags = FFI_TYPE_STRUCT;
+      /* FALLTHRU */
+    case FFI_TYPE_STRUCT:
+      switch (cif->rtype->size)
+	{
+	case 8:
+	  flags = FFI_TYPE_UINT64;
+	  break;
+	case 4:
+	  flags = FFI_TYPE_SMALL_STRUCT_4B;
+	  break;
+	case 2:
+	  flags = FFI_TYPE_SMALL_STRUCT_2B;
+	  break;
+	case 1:
+	  flags = FFI_TYPE_SMALL_STRUCT_1B;
+	  break;
+	}
+      break;
+    }
+  cif->flags = flags;
+
+  /* Each argument either fits in a register, an 8 byte slot, or is
+     passed by reference with the pointer in the 8 byte slot.  */
+  n = cif->nargs;
+  n += (flags == FFI_TYPE_STRUCT);
+  if (n < 4)
+    n = 4;
+  cif->bytes = n * 8;
+
+  return FFI_OK;
+}
+
+static void
+ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	      void **avalue, void *closure)
+{
+  int i, j, n, flags;
+  UINT64 *stack;
+  size_t rsize;
+  struct win64_call_frame *frame;
+
+  FFI_ASSERT(cif->abi == FFI_WIN64);
+
+  flags = cif->flags;
+  rsize = 0;
+
+  /* If we have no return value for a structure, we need to create one.
+     Otherwise we can ignore the return type entirely.  */
+  if (rvalue == NULL)
+    {
+      if (flags == FFI_TYPE_STRUCT)
+	rsize = cif->rtype->size;
+      else
+	flags = FFI_TYPE_VOID;
+    }
+
+  stack = alloca(cif->bytes + sizeof(struct win64_call_frame) + rsize);
+  frame = (struct win64_call_frame *)((char *)stack + cif->bytes);
+  if (rsize)
+    rvalue = frame + 1;
+
+  frame->fn = (uintptr_t)fn;
+  frame->flags = flags;
+  frame->rvalue = (uintptr_t)rvalue;
+
+  j = 0;
+  if (flags == FFI_TYPE_STRUCT)
+    {
+      stack[0] = (uintptr_t)rvalue;
+      j = 1;
+    }
+
+  for (i = 0, n = cif->nargs; i < n; ++i, ++j)
+    {
+      switch (cif->arg_types[i]->size)
+	{
+	case 8:
+	  stack[j] = *(UINT64 *)avalue[i];
+	  break;
+	case 4:
+	  stack[j] = *(UINT32 *)avalue[i];
+	  break;
+	case 2:
+	  stack[j] = *(UINT16 *)avalue[i];
+	  break;
+	case 1:
+	  stack[j] = *(UINT8 *)avalue[i];
+	  break;
+	default:
+	  stack[j] = (uintptr_t)avalue[i];
+	  break;
+	}
+    }
+
+  ffi_call_win64 (stack, frame, closure);
+}
+
+void
+ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, NULL);
+}
+
+void
+ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
+	     void **avalue, void *closure)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, closure);
+}
+
+
+extern void ffi_closure_win64(void) FFI_HIDDEN;
+extern void ffi_go_closure_win64(void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure* closure,
+		      ffi_cif* cif,
+		      void (*fun)(ffi_cif*, void*, void**, void*),
+		      void *user_data,
+		      void *codeloc)
+{
+  static const unsigned char trampoline[16] = {
+    /* leaq  -0x7(%rip),%r10   # 0x0  */
+    0x4c, 0x8d, 0x15, 0xf9, 0xff, 0xff, 0xff,
+    /* jmpq  *0x3(%rip)        # 0x10 */
+    0xff, 0x25, 0x03, 0x00, 0x00, 0x00,
+    /* nopl  (%rax) */
+    0x0f, 0x1f, 0x00
+  };
+  void *tramp = closure->tramp;
+
+  if (cif->abi != FFI_WIN64)
+    return FFI_BAD_ABI;
+
+  memcpy (tramp, trampoline, sizeof(trampoline));
+  *(UINT64 *)(tramp + 16) = (uintptr_t)ffi_closure_win64;
+
+  closure->cif = cif;
+  closure->fun = fun;
+  closure->user_data = user_data;
+
+  return FFI_OK;
+}
+
+ffi_status
+ffi_prep_go_closure (ffi_go_closure* closure, ffi_cif* cif,
+		     void (*fun)(ffi_cif*, void*, void**, void*))
+{
+  if (cif->abi != FFI_WIN64)
+    return FFI_BAD_ABI;
+
+  closure->tramp = ffi_go_closure_win64;
+  closure->cif = cif;
+  closure->fun = fun;
+
+  return FFI_OK;
+}
+
+struct win64_closure_frame
+{
+  UINT64 rvalue[2];
+  UINT64 fargs[4];
+  UINT64 retaddr;
+  UINT64 args[];
+};
+
+int FFI_HIDDEN
+ffi_closure_win64_inner(ffi_cif *cif,
+			void (*fun)(ffi_cif*, void*, void**, void*),
+			void *user_data,
+			struct win64_closure_frame *frame)
+{
+  void **avalue;
+  void *rvalue;
+  int i, n, nreg, flags;
+
+  avalue = alloca(cif->nargs * sizeof(void *));
+  rvalue = frame->rvalue;
+  nreg = 0;
+
+  /* When returning a structure, the address is in the first argument.
+     We must also be prepared to return the same address in eax, so
+     install that address in the frame and pretend we return a pointer.  */
+  flags = cif->flags;
+  if (flags == FFI_TYPE_STRUCT)
+    {
+      rvalue = (void *)(uintptr_t)frame->args[0];
+      frame->rvalue[0] = frame->args[0];
+      nreg = 1;
+    }
+
+  for (i = 0, n = cif->nargs; i < n; ++i, ++nreg)
+    {
+      size_t size = cif->arg_types[i]->size;
+      size_t type = cif->arg_types[i]->type;
+      void *a;
+
+      if (type == FFI_TYPE_DOUBLE || type == FFI_TYPE_FLOAT)
+	{
+	  if (nreg < 4)
+	    a = &frame->fargs[nreg];
+	  else
+	    a = &frame->args[nreg];
+	}
+      else if (size == 1 || size == 2 || size == 4 || size == 8)
+	a = &frame->args[nreg];
+      else
+	a = (void *)(uintptr_t)frame->args[nreg];
+
+      avalue[i] = a;
+    }
+
+  /* Invoke the closure.  */
+  fun (cif, rvalue, avalue, user_data);
+  return flags;
+}
+
+#endif /* X86_WIN64 */

src/x86/freebsd.S

@@ -1,463 +0,0 @@
-/* -----------------------------------------------------------------------
-   freebsd.S - Copyright (c) 1996, 1998, 2001, 2002, 2003, 2005  Red Hat, Inc.
-	       Copyright (c) 2008  Björn König
-	
-   X86 Foreign Function Interface for FreeBSD
-
-   Permission is hereby granted, free of charge, to any person obtaining
-   a copy of this software and associated documentation files (the
-   ``Software''), to deal in the Software without restriction, including
-   without limitation the rights to use, copy, modify, merge, publish,
-   distribute, sublicense, and/or sell copies of the Software, and to
-   permit persons to whom the Software is furnished to do so, subject to
-   the following conditions:
-
-   The above copyright notice and this permission notice shall be included
-   in all copies or substantial portions of the Software.
-
-   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
-   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
-   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
-   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
-   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
-   DEALINGS IN THE SOFTWARE.
------------------------------------------------------------------------ */
-
-#ifndef __x86_64__
-
-#define LIBFFI_ASM	
-#include <fficonfig.h>
-#include <ffi.h>
-
-.text
-
-.globl ffi_prep_args
-
-	.align 4
-.globl ffi_call_SYSV
-        .type    ffi_call_SYSV,@function
-
-ffi_call_SYSV:
-.LFB1:
-        pushl %ebp
-.LCFI0:
-        movl  %esp,%ebp
-.LCFI1:
-	/* Make room for all of the new args.  */
-	movl  16(%ebp),%ecx
-	subl  %ecx,%esp
-
-	/* Align the stack pointer to 16-bytes */
-	andl  $0xfffffff0, %esp
-
-	movl  %esp,%eax
-
-	/* Place all of the ffi_prep_args in position  */
-	pushl 12(%ebp)
-	pushl %eax
-	call  *8(%ebp)
-
-	/* Return stack to previous state and call the function  */
-	addl  $8,%esp	
-
-	call  *28(%ebp)
-
-	/* Load %ecx with the return type code  */
-	movl  20(%ebp),%ecx	
-
-	/* Protect %esi.  We're going to pop it in the epilogue.  */
-	pushl %esi
-
-	/* If the return value pointer is NULL, assume no return value.  */
-	cmpl  $0,24(%ebp)
-	jne  0f
-
-	/* Even if there is no space for the return value, we are 
-	   obliged to handle floating-point values.  */
-	cmpl  $FFI_TYPE_FLOAT,%ecx
-	jne   noretval
-	fstp  %st(0)
-
-        jmp   epilogue
-
-0:
-	call  1f
-
-.Lstore_table:
-	.long	noretval-.Lstore_table	/* FFI_TYPE_VOID */
-	.long	retint-.Lstore_table	/* FFI_TYPE_INT */
-	.long	retfloat-.Lstore_table	/* FFI_TYPE_FLOAT */
-	.long	retdouble-.Lstore_table	/* FFI_TYPE_DOUBLE */
-	.long	retlongdouble-.Lstore_table	/* FFI_TYPE_LONGDOUBLE */
-	.long	retuint8-.Lstore_table	/* FFI_TYPE_UINT8 */
-	.long	retsint8-.Lstore_table	/* FFI_TYPE_SINT8 */
-	.long	retuint16-.Lstore_table	/* FFI_TYPE_UINT16 */
-	.long	retsint16-.Lstore_table	/* FFI_TYPE_SINT16 */
-	.long	retint-.Lstore_table	/* FFI_TYPE_UINT32 */
-	.long	retint-.Lstore_table	/* FFI_TYPE_SINT32 */
-	.long	retint64-.Lstore_table	/* FFI_TYPE_UINT64 */
-	.long	retint64-.Lstore_table	/* FFI_TYPE_SINT64 */
-	.long	retstruct-.Lstore_table	/* FFI_TYPE_STRUCT */
-	.long	retint-.Lstore_table	/* FFI_TYPE_POINTER */
-	.long   retstruct1b-.Lstore_table	/* FFI_TYPE_SMALL_STRUCT_1B */
-	.long   retstruct2b-.Lstore_table	/* FFI_TYPE_SMALL_STRUCT_2B */
-
-1:
-	pop  %esi
-	add  (%esi, %ecx, 4), %esi
-	jmp  *%esi
-
-	/* Sign/zero extend as appropriate.  */
-retsint8:
-	movsbl  %al, %eax
-	jmp  retint
-
-retsint16:
-	movswl  %ax, %eax
-	jmp  retint
-
-retuint8:
-	movzbl  %al, %eax
-	jmp  retint
-
-retuint16:
-	movzwl  %ax, %eax
-	jmp  retint
-
-retfloat:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx	
-	fstps (%ecx)
-	jmp   epilogue
-
-retdouble:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx	
-	fstpl (%ecx)
-	jmp   epilogue
-
-retlongdouble:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx	
-	fstpt (%ecx)
-	jmp   epilogue
-	
-retint64:	
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx	
-	movl  %eax,0(%ecx)
-	movl  %edx,4(%ecx)
-	jmp   epilogue
-	
-retstruct1b:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx
-	movb  %al,0(%ecx)
-	jmp   epilogue
-
-retstruct2b:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx
-	movw  %ax,0(%ecx)
-	jmp   epilogue
-
-retint:
-	/* Load %ecx with the pointer to storage for the return value  */
-	movl  24(%ebp),%ecx	
-	movl  %eax,0(%ecx)
-
-retstruct:
-	/* Nothing to do!  */
-
-noretval:
-epilogue:
-        popl %esi
-        movl %ebp,%esp
-        popl %ebp
-        ret
-.LFE1:
-.ffi_call_SYSV_end:
-        .size    ffi_call_SYSV,.ffi_call_SYSV_end-ffi_call_SYSV
-
-	.align	4
-FFI_HIDDEN (ffi_closure_SYSV)
-.globl ffi_closure_SYSV
-	.type	ffi_closure_SYSV, @function
-
-ffi_closure_SYSV:
-.LFB2:
-	pushl	%ebp
-.LCFI2:
-	movl	%esp, %ebp
-.LCFI3:
-	subl	$40, %esp
-	leal	-24(%ebp), %edx
-	movl	%edx, -12(%ebp)	/* resp */
-	leal	8(%ebp), %edx
-	movl	%edx, 4(%esp)	/* args = __builtin_dwarf_cfa () */
-	leal	-12(%ebp), %edx
-	movl	%edx, (%esp)	/* &resp */
-#if defined HAVE_HIDDEN_VISIBILITY_ATTRIBUTE || !defined __PIC__
-	call	ffi_closure_SYSV_inner
-#else
-	movl	%ebx, 8(%esp)
-.LCFI7:
-	call	1f
-1:	popl	%ebx
-	addl	$_GLOBAL_OFFSET_TABLE_+[.-1b], %ebx
-	call	ffi_closure_SYSV_inner@PLT
-	movl	8(%esp), %ebx
-#endif
-	movl	-12(%ebp), %ecx
-	cmpl	$FFI_TYPE_INT, %eax
-	je	.Lcls_retint
-
-	/* Handle FFI_TYPE_UINT8, FFI_TYPE_SINT8, FFI_TYPE_UINT16,
-	   FFI_TYPE_SINT16, FFI_TYPE_UINT32, FFI_TYPE_SINT32.  */
-	cmpl	$FFI_TYPE_UINT64, %eax
-	jge	0f
-	cmpl	$FFI_TYPE_UINT8, %eax
-	jge	.Lcls_retint
-	
-0:	cmpl	$FFI_TYPE_FLOAT, %eax
-	je	.Lcls_retfloat
-	cmpl	$FFI_TYPE_DOUBLE, %eax
-	je	.Lcls_retdouble
-	cmpl	$FFI_TYPE_LONGDOUBLE, %eax
-	je	.Lcls_retldouble
-	cmpl	$FFI_TYPE_SINT64, %eax
-	je	.Lcls_retllong
-	cmpl	$FFI_TYPE_SMALL_STRUCT_1B, %eax
-	je	.Lcls_retstruct1b
-	cmpl	$FFI_TYPE_SMALL_STRUCT_2B, %eax
-	je	.Lcls_retstruct2b
-	cmpl	$FFI_TYPE_STRUCT, %eax
-	je	.Lcls_retstruct
-.Lcls_epilogue:
-	movl	%ebp, %esp
-	popl	%ebp
-	ret
-.Lcls_retint:
-	movl	(%ecx), %eax
-	jmp	.Lcls_epilogue
-.Lcls_retfloat:
-	flds	(%ecx)
-	jmp	.Lcls_epilogue
-.Lcls_retdouble:
-	fldl	(%ecx)
-	jmp	.Lcls_epilogue
-.Lcls_retldouble:
-	fldt	(%ecx)
-	jmp	.Lcls_epilogue
-.Lcls_retllong:
-	movl	(%ecx), %eax
-	movl	4(%ecx), %edx
-	jmp	.Lcls_epilogue
-.Lcls_retstruct1b:
-	movsbl	(%ecx), %eax
-	jmp	.Lcls_epilogue
-.Lcls_retstruct2b:
-	movswl	(%ecx), %eax
-	jmp	.Lcls_epilogue
-.Lcls_retstruct:
-	movl	%ebp, %esp
-	popl	%ebp
-	ret	$4
-.LFE2:
-	.size	ffi_closure_SYSV, .-ffi_closure_SYSV
-
-#if !FFI_NO_RAW_API
-
-#define RAW_CLOSURE_CIF_OFFSET ((FFI_TRAMPOLINE_SIZE + 3) & ~3)
-#define RAW_CLOSURE_FUN_OFFSET (RAW_CLOSURE_CIF_OFFSET + 4)
-#define RAW_CLOSURE_USER_DATA_OFFSET (RAW_CLOSURE_FUN_OFFSET + 4)
-#define CIF_FLAGS_OFFSET 20
-
-	.align	4
-FFI_HIDDEN (ffi_closure_raw_SYSV)
-.globl ffi_closure_raw_SYSV
-	.type	ffi_closure_raw_SYSV, @function
-
-ffi_closure_raw_SYSV:
-.LFB3:
-	pushl	%ebp
-.LCFI4:
-	movl	%esp, %ebp
-.LCFI5:
-	pushl	%esi
-.LCFI6:
-	subl	$36, %esp
-	movl	RAW_CLOSURE_CIF_OFFSET(%eax), %esi	 /* closure->cif */
-	movl	RAW_CLOSURE_USER_DATA_OFFSET(%eax), %edx /* closure->user_data */
-	movl	%edx, 12(%esp)	/* user_data */
-	leal	8(%ebp), %edx	/* __builtin_dwarf_cfa () */
-	movl	%edx, 8(%esp)	/* raw_args */
-	leal	-24(%ebp), %edx
-	movl	%edx, 4(%esp)	/* &res */
-	movl	%esi, (%esp)	/* cif */
-	call	*RAW_CLOSURE_FUN_OFFSET(%eax)		 /* closure->fun */
-	movl	CIF_FLAGS_OFFSET(%esi), %eax		 /* rtype */
-	cmpl	$FFI_TYPE_INT, %eax
-	je	.Lrcls_retint
-
-	/* Handle FFI_TYPE_UINT8, FFI_TYPE_SINT8, FFI_TYPE_UINT16,
-	   FFI_TYPE_SINT16, FFI_TYPE_UINT32, FFI_TYPE_SINT32.  */
-	cmpl	$FFI_TYPE_UINT64, %eax
-	jge	0f
-	cmpl	$FFI_TYPE_UINT8, %eax
-	jge	.Lrcls_retint
-0:
-	cmpl	$FFI_TYPE_FLOAT, %eax
-	je	.Lrcls_retfloat
-	cmpl	$FFI_TYPE_DOUBLE, %eax
-	je	.Lrcls_retdouble
-	cmpl	$FFI_TYPE_LONGDOUBLE, %eax
-	je	.Lrcls_retldouble
-	cmpl	$FFI_TYPE_SINT64, %eax
-	je	.Lrcls_retllong
-.Lrcls_epilogue:
-	addl	$36, %esp
-	popl	%esi
-	popl	%ebp
-	ret
-.Lrcls_retint:
-	movl	-24(%ebp), %eax
-	jmp	.Lrcls_epilogue
-.Lrcls_retfloat:
-	flds	-24(%ebp)
-	jmp	.Lrcls_epilogue
-.Lrcls_retdouble:
-	fldl	-24(%ebp)
-	jmp	.Lrcls_epilogue
-.Lrcls_retldouble:
-	fldt	-24(%ebp)
-	jmp	.Lrcls_epilogue
-.Lrcls_retllong:
-	movl	-24(%ebp), %eax
-	movl	-20(%ebp), %edx
-	jmp	.Lrcls_epilogue
-.LFE3:
-	.size	ffi_closure_raw_SYSV, .-ffi_closure_raw_SYSV
-#endif
-
-	.section	.eh_frame,EH_FRAME_FLAGS,@progbits
-.Lframe1:
-	.long	.LECIE1-.LSCIE1	/* Length of Common Information Entry */
-.LSCIE1:
-	.long	0x0	/* CIE Identifier Tag */
-	.byte	0x1	/* CIE Version */
-#ifdef __PIC__
-	.ascii "zR\0"	/* CIE Augmentation */
-#else
-	.ascii "\0"	/* CIE Augmentation */
-#endif
-	.byte	0x1	/* .uleb128 0x1; CIE Code Alignment Factor */
-	.byte	0x7c	/* .sleb128 -4; CIE Data Alignment Factor */
-	.byte	0x8	/* CIE RA Column */
-#ifdef __PIC__
-	.byte	0x1	/* .uleb128 0x1; Augmentation size */
-	.byte	0x1b	/* FDE Encoding (pcrel sdata4) */
-#endif
-	.byte	0xc	/* DW_CFA_def_cfa */
-	.byte	0x4	/* .uleb128 0x4 */
-	.byte	0x4	/* .uleb128 0x4 */
-	.byte	0x88	/* DW_CFA_offset, column 0x8 */
-	.byte	0x1	/* .uleb128 0x1 */
-	.align 4
-.LECIE1:
-.LSFDE1:
-	.long	.LEFDE1-.LASFDE1	/* FDE Length */
-.LASFDE1:
-	.long	.LASFDE1-.Lframe1	/* FDE CIE offset */
-#ifdef __PIC__
-	.long	.LFB1-.	/* FDE initial location */
-#else
-	.long	.LFB1	/* FDE initial location */
-#endif
-	.long	.LFE1-.LFB1	/* FDE address range */
-#ifdef __PIC__
-	.byte	0x0	/* .uleb128 0x0; Augmentation size */
-#endif
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI0-.LFB1
-	.byte	0xe	/* DW_CFA_def_cfa_offset */
-	.byte	0x8	/* .uleb128 0x8 */
-	.byte	0x85	/* DW_CFA_offset, column 0x5 */
-	.byte	0x2	/* .uleb128 0x2 */
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI1-.LCFI0
-	.byte	0xd	/* DW_CFA_def_cfa_register */
-	.byte	0x5	/* .uleb128 0x5 */
-	.align 4
-.LEFDE1:
-.LSFDE2:
-	.long	.LEFDE2-.LASFDE2	/* FDE Length */
-.LASFDE2:
-	.long	.LASFDE2-.Lframe1	/* FDE CIE offset */
-#ifdef __PIC__
-	.long	.LFB2-.	/* FDE initial location */
-#else
-	.long	.LFB2
-#endif
-	.long	.LFE2-.LFB2	/* FDE address range */
-#ifdef __PIC__
-	.byte	0x0	/* .uleb128 0x0; Augmentation size */
-#endif
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI2-.LFB2
-	.byte	0xe	/* DW_CFA_def_cfa_offset */
-	.byte	0x8	/* .uleb128 0x8 */
-	.byte	0x85	/* DW_CFA_offset, column 0x5 */
-	.byte	0x2	/* .uleb128 0x2 */
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI3-.LCFI2
-	.byte	0xd	/* DW_CFA_def_cfa_register */
-	.byte	0x5	/* .uleb128 0x5 */
-#if !defined HAVE_HIDDEN_VISIBILITY_ATTRIBUTE && defined __PIC__
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI7-.LCFI3
-	.byte	0x83	/* DW_CFA_offset, column 0x3 */
-	.byte	0xa	/* .uleb128 0xa */
-#endif
-	.align 4
-.LEFDE2:
-
-#if !FFI_NO_RAW_API
-
-.LSFDE3:
-	.long	.LEFDE3-.LASFDE3	/* FDE Length */
-.LASFDE3:
-	.long	.LASFDE3-.Lframe1	/* FDE CIE offset */
-#ifdef __PIC__
-	.long	.LFB3-.	/* FDE initial location */
-#else
-	.long	.LFB3
-#endif
-	.long	.LFE3-.LFB3	/* FDE address range */
-#ifdef __PIC__
-	.byte	0x0	/* .uleb128 0x0; Augmentation size */
-#endif
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI4-.LFB3
-	.byte	0xe	/* DW_CFA_def_cfa_offset */
-	.byte	0x8	/* .uleb128 0x8 */
-	.byte	0x85	/* DW_CFA_offset, column 0x5 */
-	.byte	0x2	/* .uleb128 0x2 */
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI5-.LCFI4
-	.byte	0xd	/* DW_CFA_def_cfa_register */
-	.byte	0x5	/* .uleb128 0x5 */
-	.byte	0x4	/* DW_CFA_advance_loc4 */
-	.long	.LCFI6-.LCFI5
-	.byte	0x86	/* DW_CFA_offset, column 0x6 */
-	.byte	0x3	/* .uleb128 0x3 */
-	.align 4
-.LEFDE3:
-
-#endif
-
-#endif /* ifndef __x86_64__ */
-
-	.section .note.GNU-stack,"",%progbits

src/x86/internal.h

@@ -0,0 +1,23 @@
+#define X86_RET_FLOAT		0
+#define X86_RET_DOUBLE		1
+#define X86_RET_LDOUBLE		2
+#define X86_RET_SINT8		3
+#define X86_RET_SINT16		4
+#define X86_RET_UINT8		5
+#define X86_RET_UINT16		6
+#define X86_RET_INT64		7
+#define X86_RET_INT32		8
+#define X86_RET_VOID		9
+#define X86_RET_STRUCTPOP	10
+#define X86_RET_STRUCTARG       11
+#define X86_RET_STRUCT_1B	12
+#define X86_RET_STRUCT_2B	13
+#define X86_RET_UNUSED14	14
+#define X86_RET_UNUSED15	15
+
+#define X86_RET_TYPE_MASK	15
+#define X86_RET_POP_SHIFT	4
+
+#define R_EAX	0
+#define R_EDX	1
+#define R_ECX	2

src/x86/internal64.h

@@ -0,0 +1,22 @@
+#define UNIX64_RET_VOID		0
+#define UNIX64_RET_UINT8	1
+#define UNIX64_RET_UINT16	2
+#define UNIX64_RET_UINT32	3
+#define UNIX64_RET_SINT8	4
+#define UNIX64_RET_SINT16	5
+#define UNIX64_RET_SINT32	6
+#define UNIX64_RET_INT64	7
+#define UNIX64_RET_XMM32	8
+#define UNIX64_RET_XMM64	9
+#define UNIX64_RET_X87		10
+#define UNIX64_RET_X87_2	11
+#define UNIX64_RET_ST_XMM0_RAX	12
+#define UNIX64_RET_ST_RAX_XMM0	13
+#define UNIX64_RET_ST_XMM0_XMM1	14
+#define UNIX64_RET_ST_RAX_RDX	15
+
+#define UNIX64_RET_LAST		15
+
+#define UNIX64_FLAG_RET_IN_MEM	(1 << 10)
+#define UNIX64_FLAG_XMM_ARGS	(1 << 11)
+#define UNIX64_SIZE_SHIFT	12

src/x86/unix64.S

@@ -30,9 +30,22 @@
 #define LIBFFI_ASM	
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
+#include "internal64.h"
 
 	.text
 
+/* This macro allows the safe creation of jump tables without an
+   actual table.  The entry points into the table are all 8 bytes.
+   The use of ORG asserts that we're at the correct location.  */
+/* ??? The clang assembler doesn't handle .org with symbolic expressions.  */
+.macro E index
+	.align	8
+#ifndef __clang__
+	.org	0b + \index * 8, 0x90
+#endif
+.endm
+
 /* ffi_call_unix64 (void *args, unsigned long bytes, unsigned flags,
 	            void *raddr, void (*fnaddr)(void));
 
@@ -40,12 +53,13 @@
    for this function.  This has been allocated by ffi_call.  We also
    deallocate some of the stack that has been alloca'd.  */
 
-	.align	2
+	.align	8
 	.globl	ffi_call_unix64
 	.type	ffi_call_unix64,@function
+	FFI_HIDDEN(ffi_call_unix64)
 
 ffi_call_unix64:
-.LUW0:
+	cfi_startproc
 	movq	(%rsp), %r10		/* Load return address.  */
 	leaq	(%rdi, %rsi), %rax	/* Find local stack base.  */
 	movq	%rdx, (%rax)		/* Save flags.  */
@@ -53,24 +67,36 @@ ffi_call_unix64:
 	movq	%rbp, 16(%rax)		/* Save old frame pointer.  */
 	movq	%r10, 24(%rax)		/* Relocate return address.  */
 	movq	%rax, %rbp		/* Finalize local stack frame.  */
-.LUW1:
+
+	/* New stack frame based off rbp.  This is a itty bit of unwind
+	   trickery in that the CFA *has* changed.  There is no easy way
+	   to describe it correctly on entry to the function.  Fortunately,
+	   it doesn't matter too much since at all points we can correctly
+	   unwind back to ffi_call.  Note that the location to which we
+	   moved the return address is (the new) CFA-8, so from the
+	   perspective of the unwind info, it hasn't moved.  */
+	cfi_def_cfa(%rbp, 32)
+	cfi_rel_offset(%rbp, 16)
+
 	movq	%rdi, %r10		/* Save a copy of the register area. */
 	movq	%r8, %r11		/* Save a copy of the target fn.  */
 	movl	%r9d, %eax		/* Set number of SSE registers.  */
 
 	/* Load up all argument registers.  */
 	movq	(%r10), %rdi
-	movq	8(%r10), %rsi
+	movq	0x08(%r10), %rsi
-	movq	16(%r10), %rdx
+	movq	0x10(%r10), %rdx
-	movq	24(%r10), %rcx
+	movq	0x18(%r10), %rcx
-	movq	32(%r10), %r8
+	movq	0x20(%r10), %r8
-	movq	40(%r10), %r9
+	movq	0x28(%r10), %r9
+	movl	0xb0(%r10), %eax
 	testl	%eax, %eax
 	jnz	.Lload_sse
 .Lret_from_load_sse:
 
-	/* Deallocate the reg arg area.  */
+	/* Deallocate the reg arg area, except for r10, then load via pop.  */
-	leaq	176(%r10), %rsp
+	leaq	0xb8(%r10), %rsp
+	popq	%r10
 
 	/* Call the user function.  */
 	call	*%r11
@@ -81,352 +107,304 @@ ffi_call_unix64:
 	movq	0(%rbp), %rcx		/* Reload flags.  */
 	movq	8(%rbp), %rdi		/* Reload raddr.  */
 	movq	16(%rbp), %rbp		/* Reload old frame pointer.  */
-.LUW2:
+	cfi_remember_state
+	cfi_def_cfa(%rsp, 8)
+	cfi_restore(%rbp)
 
 	/* The first byte of the flags contains the FFI_TYPE.  */
+	cmpb	$UNIX64_RET_LAST, %cl
 	movzbl	%cl, %r10d
-	leaq	.Lstore_table(%rip), %r11
+	leaq	0f(%rip), %r11
-	movslq	(%r11, %r10, 4), %r10
+	ja	9f
-	addq	%r11, %r10
+	leaq	(%r11, %r10, 8), %r10
+
+	/* Prep for the structure cases: scratch area in redzone.  */
+	leaq	-20(%rsp), %rsi
 	jmp	*%r10
 
-.Lstore_table:
+	.align	8
-	.long	.Lst_void-.Lstore_table		/* FFI_TYPE_VOID */
+0:
-	.long	.Lst_sint32-.Lstore_table	/* FFI_TYPE_INT */
+E UNIX64_RET_VOID
-	.long	.Lst_float-.Lstore_table	/* FFI_TYPE_FLOAT */
-	.long	.Lst_double-.Lstore_table	/* FFI_TYPE_DOUBLE */
-	.long	.Lst_ldouble-.Lstore_table	/* FFI_TYPE_LONGDOUBLE */
-	.long	.Lst_uint8-.Lstore_table	/* FFI_TYPE_UINT8 */
-	.long	.Lst_sint8-.Lstore_table	/* FFI_TYPE_SINT8 */
-	.long	.Lst_uint16-.Lstore_table	/* FFI_TYPE_UINT16 */
-	.long	.Lst_sint16-.Lstore_table	/* FFI_TYPE_SINT16 */
-	.long	.Lst_uint32-.Lstore_table	/* FFI_TYPE_UINT32 */
-	.long	.Lst_sint32-.Lstore_table	/* FFI_TYPE_SINT32 */
-	.long	.Lst_int64-.Lstore_table	/* FFI_TYPE_UINT64 */
-	.long	.Lst_int64-.Lstore_table	/* FFI_TYPE_SINT64 */
-	.long	.Lst_struct-.Lstore_table	/* FFI_TYPE_STRUCT */
-	.long	.Lst_int64-.Lstore_table	/* FFI_TYPE_POINTER */
-
-	.align 2
-.Lst_void:
 	ret
-	.align 2
+E UNIX64_RET_UINT8
-
+	movzbl	%al, %eax
-.Lst_uint8:
-	movzbq	%al, %rax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
+E UNIX64_RET_UINT16
-.Lst_sint8:
+	movzwl	%ax, %eax
+	movq	%rax, (%rdi)
+	ret
+E UNIX64_RET_UINT32
+	movl	%eax, %eax
+	movq	%rax, (%rdi)
+	ret
+E UNIX64_RET_SINT8
 	movsbq	%al, %rax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
+E UNIX64_RET_SINT16
-.Lst_uint16:
-	movzwq	%ax, %rax
-	movq	%rax, (%rdi)
-	.align 2
-.Lst_sint16:
 	movswq	%ax, %rax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
+E UNIX64_RET_SINT32
-.Lst_uint32:
-	movl	%eax, %eax
-	movq	%rax, (%rdi)
-	.align 2
-.Lst_sint32:
 	cltq
 	movq	%rax, (%rdi)
 	ret
-	.align 2
+E UNIX64_RET_INT64
-.Lst_int64:
 	movq	%rax, (%rdi)
 	ret
-
+E UNIX64_RET_XMM32
-	.align 2
+	movd	%xmm0, (%rdi)
-.Lst_float:
-	movss	%xmm0, (%rdi)
 	ret
-	.align 2
+E UNIX64_RET_XMM64
-.Lst_double:
+	movq	%xmm0, (%rdi)
-	movsd	%xmm0, (%rdi)
 	ret
-.Lst_ldouble:
+E UNIX64_RET_X87
 	fstpt	(%rdi)
 	ret
-
+E UNIX64_RET_X87_2
-	.align 2
+	fstpt	(%rdi)
-.Lst_struct:
+	fstpt	16(%rdi)
-	leaq	-20(%rsp), %rsi		/* Scratch area in redzone.  */
+	ret
-
+E UNIX64_RET_ST_XMM0_RAX
-	/* We have to locate the values now, and since we don't want to
+	movq	%rax, 8(%rsi)
-	   write too much data into the user's return value, we spill the
+	jmp	3f
-	   value to a 16 byte scratch area first.  Bits 8, 9, and 10
+E UNIX64_RET_ST_RAX_XMM0
-	   control where the values are located.  Only one of the three
+	movq	%xmm0, 8(%rsi)
-	   bits will be set; see ffi_prep_cif_machdep for the pattern.  */
+	jmp	2f
-	movd	%xmm0, %r10
+E UNIX64_RET_ST_XMM0_XMM1
-	movd	%xmm1, %r11
+	movq	%xmm1, 8(%rsi)
-	testl	$0x100, %ecx
+	jmp	3f
-	cmovnz	%rax, %rdx
+E UNIX64_RET_ST_RAX_RDX
-	cmovnz	%r10, %rax
-	testl	$0x200, %ecx
-	cmovnz	%r10, %rdx
-	testl	$0x400, %ecx
-	cmovnz	%r10, %rax
-	cmovnz	%r11, %rdx
-	movq	%rax, (%rsi)
 	movq	%rdx, 8(%rsi)
-
+2:	movq	%rax, (%rsi)
-	/* Bits 12-31 contain the true size of the structure.  Copy from
+	shrl	$UNIX64_SIZE_SHIFT, %ecx
-	   the scratch area to the true destination.  */
-	shrl	$12, %ecx
 	rep movsb
 	ret
+	.align 8
+3:	movq	%xmm0, (%rsi)
+	shrl	$UNIX64_SIZE_SHIFT, %ecx
+	rep movsb
+	ret
+
+9:	call	abort@PLT
 
 	/* Many times we can avoid loading any SSE registers at all.
 	   It's not worth an indirect jump to load the exact set of
 	   SSE registers needed; zero or all is a good compromise.  */
 	.align 2
-.LUW3:
+	cfi_restore_state
 .Lload_sse:
-	movdqa	48(%r10), %xmm0
+	movdqa	0x30(%r10), %xmm0
-	movdqa	64(%r10), %xmm1
+	movdqa	0x40(%r10), %xmm1
-	movdqa	80(%r10), %xmm2
+	movdqa	0x50(%r10), %xmm2
-	movdqa	96(%r10), %xmm3
+	movdqa	0x60(%r10), %xmm3
-	movdqa	112(%r10), %xmm4
+	movdqa	0x70(%r10), %xmm4
-	movdqa	128(%r10), %xmm5
+	movdqa	0x80(%r10), %xmm5
-	movdqa	144(%r10), %xmm6
+	movdqa	0x90(%r10), %xmm6
-	movdqa	160(%r10), %xmm7
+	movdqa	0xa0(%r10), %xmm7
 	jmp	.Lret_from_load_sse
 
-.LUW4:
+	cfi_endproc
 	.size    ffi_call_unix64,.-ffi_call_unix64
 
+/* 6 general registers, 8 vector registers,
+   32 bytes of rvalue, 8 bytes of alignment.  */
+#define ffi_closure_OFS_G	0
+#define ffi_closure_OFS_V	(6*8)
+#define ffi_closure_OFS_RVALUE	(ffi_closure_OFS_V + 8*16)
+#define ffi_closure_FS		(ffi_closure_OFS_RVALUE + 32 + 8)
+
+/* The location of rvalue within the red zone after deallocating the frame.  */
+#define ffi_closure_RED_RVALUE	(ffi_closure_OFS_RVALUE - ffi_closure_FS)
+
+	.align	2
+	.globl	ffi_closure_unix64_sse
+	.type	ffi_closure_unix64_sse,@function
+	FFI_HIDDEN(ffi_closure_unix64_sse)
+
+ffi_closure_unix64_sse:
+	cfi_startproc
+	subq	$ffi_closure_FS, %rsp
+	/* Note clang bug 21515: adjust_cfa_offset error across endproc.  */
+	cfi_def_cfa_offset(ffi_closure_FS + 8)
+
+	movdqa	%xmm0, ffi_closure_OFS_V+0x00(%rsp)
+	movdqa	%xmm1, ffi_closure_OFS_V+0x10(%rsp)
+	movdqa	%xmm2, ffi_closure_OFS_V+0x20(%rsp)
+	movdqa	%xmm3, ffi_closure_OFS_V+0x30(%rsp)
+	movdqa	%xmm4, ffi_closure_OFS_V+0x40(%rsp)
+	movdqa	%xmm5, ffi_closure_OFS_V+0x50(%rsp)
+	movdqa	%xmm6, ffi_closure_OFS_V+0x60(%rsp)
+	movdqa	%xmm7, ffi_closure_OFS_V+0x70(%rsp)
+	jmp	0f
+
+	cfi_endproc
+	.size	ffi_closure_unix64_sse,.-ffi_closure_unix64_sse
+
 	.align	2
 	.globl	ffi_closure_unix64
 	.type	ffi_closure_unix64,@function
+	FFI_HIDDEN(ffi_closure_unix64)
 
 ffi_closure_unix64:
-.LUW5:
+	cfi_startproc
-	/* The carry flag is set by the trampoline iff SSE registers
+	subq	$ffi_closure_FS, %rsp
-	   are used.  Don't clobber it before the branch instruction.  */
+	/* Note clang bug 21515: adjust_cfa_offset error across endproc.  */
-	leaq    -200(%rsp), %rsp
+	cfi_def_cfa_offset(ffi_closure_FS + 8)
-.LUW6:
+0:
-	movq	%rdi, (%rsp)
+	movq	%rdi, ffi_closure_OFS_G+0x00(%rsp)
-	movq    %rsi, 8(%rsp)
+	movq    %rsi, ffi_closure_OFS_G+0x08(%rsp)
-	movq    %rdx, 16(%rsp)
+	movq    %rdx, ffi_closure_OFS_G+0x10(%rsp)
-	movq    %rcx, 24(%rsp)
+	movq    %rcx, ffi_closure_OFS_G+0x18(%rsp)
-	movq    %r8, 32(%rsp)
+	movq    %r8,  ffi_closure_OFS_G+0x20(%rsp)
-	movq    %r9, 40(%rsp)
+	movq    %r9,  ffi_closure_OFS_G+0x28(%rsp)
-	jc      .Lsave_sse
-.Lret_from_save_sse:
 
-	movq	%r10, %rdi
+#ifdef __ILP32__
-	leaq	176(%rsp), %rsi
+	movl	FFI_TRAMPOLINE_SIZE(%r10), %edi		/* Load cif */
-	movq	%rsp, %rdx
+	movl	FFI_TRAMPOLINE_SIZE+4(%r10), %esi	/* Load fun */
-	leaq	208(%rsp), %rcx
+	movl	FFI_TRAMPOLINE_SIZE+8(%r10), %edx	/* Load user_data */
-	call	ffi_closure_unix64_inner@PLT
+#else
+	movq	FFI_TRAMPOLINE_SIZE(%r10), %rdi		/* Load cif */
+	movq	FFI_TRAMPOLINE_SIZE+8(%r10), %rsi	/* Load fun */
+	movq	FFI_TRAMPOLINE_SIZE+16(%r10), %rdx	/* Load user_data */
+#endif
+.Ldo_closure:
+	leaq	ffi_closure_OFS_RVALUE(%rsp), %rcx	/* Load rvalue */
+	movq	%rsp, %r8				/* Load reg_args */
+	leaq	ffi_closure_FS+8(%rsp), %r9		/* Load argp */
+	call	ffi_closure_unix64_inner
 
 	/* Deallocate stack frame early; return value is now in redzone.  */
-	addq	$200, %rsp
+	addq	$ffi_closure_FS, %rsp
-.LUW7:
+	cfi_adjust_cfa_offset(-ffi_closure_FS)
 
 	/* The first byte of the return value contains the FFI_TYPE.  */
+	cmpb	$UNIX64_RET_LAST, %al
 	movzbl	%al, %r10d
-	leaq	.Lload_table(%rip), %r11
+	leaq	0f(%rip), %r11
-	movslq	(%r11, %r10, 4), %r10
+	ja	9f
-	addq	%r11, %r10
+	leaq	(%r11, %r10, 8), %r10
+	leaq	ffi_closure_RED_RVALUE(%rsp), %rsi
 	jmp	*%r10
 
-.Lload_table:
+	.align	8
-	.long	.Lld_void-.Lload_table		/* FFI_TYPE_VOID */
+0:
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_INT */
+E UNIX64_RET_VOID
-	.long	.Lld_float-.Lload_table		/* FFI_TYPE_FLOAT */
+	ret
-	.long	.Lld_double-.Lload_table	/* FFI_TYPE_DOUBLE */
+E UNIX64_RET_UINT8
-	.long	.Lld_ldouble-.Lload_table	/* FFI_TYPE_LONGDOUBLE */
+	movzbl	(%rsi), %eax
-	.long	.Lld_int8-.Lload_table		/* FFI_TYPE_UINT8 */
+	ret
-	.long	.Lld_int8-.Lload_table		/* FFI_TYPE_SINT8 */
+E UNIX64_RET_UINT16
-	.long	.Lld_int16-.Lload_table		/* FFI_TYPE_UINT16 */
+	movzwl	(%rsi), %eax
-	.long	.Lld_int16-.Lload_table		/* FFI_TYPE_SINT16 */
+	ret
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_UINT32 */
+E UNIX64_RET_UINT32
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_SINT32 */
+	movl	(%rsi), %eax
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_UINT64 */
+	ret
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_SINT64 */
+E UNIX64_RET_SINT8
-	.long	.Lld_struct-.Lload_table	/* FFI_TYPE_STRUCT */
+	movsbl	(%rsi), %eax
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_POINTER */
+	ret
-
+E UNIX64_RET_SINT16
-	.align 2
+	movswl	(%rsi), %eax
-.Lld_void:
+	ret
+E UNIX64_RET_SINT32
+	movl	(%rsi), %eax
+	ret
+E UNIX64_RET_INT64
+	movq	(%rsi), %rax
+	ret
+E UNIX64_RET_XMM32
+	movd	(%rsi), %xmm0
+	ret
+E UNIX64_RET_XMM64
+	movq	(%rsi), %xmm0
+	ret
+E UNIX64_RET_X87
+	fldt	(%rsi)
+	ret
+E UNIX64_RET_X87_2
+	fldt	16(%rsi)
+	fldt	(%rsi)
+	ret
+E UNIX64_RET_ST_XMM0_RAX
+	movq	8(%rsi), %rax
+	jmp	3f
+E UNIX64_RET_ST_RAX_XMM0
+	movq	8(%rsi), %xmm0
+	jmp	2f
+E UNIX64_RET_ST_XMM0_XMM1
+	movq	8(%rsi), %xmm1
+	jmp	3f
+E UNIX64_RET_ST_RAX_RDX
+	movq	8(%rsi), %rdx
+2:	movq	(%rsi), %rax
+	ret
+	.align	8
+3:	movq	(%rsi), %xmm0
 	ret
 
-	.align 2
+9:	call	abort@PLT
-.Lld_int8:
-	movzbl	-24(%rsp), %eax
-	ret
-	.align 2
-.Lld_int16:
-	movzwl	-24(%rsp), %eax
-	ret
-	.align 2
-.Lld_int32:
-	movl	-24(%rsp), %eax
-	ret
-	.align 2
-.Lld_int64:
-	movq	-24(%rsp), %rax
-	ret
 
-	.align 2
+	cfi_endproc
-.Lld_float:
-	movss	-24(%rsp), %xmm0
-	ret
-	.align 2
-.Lld_double:
-	movsd	-24(%rsp), %xmm0
-	ret
-	.align 2
-.Lld_ldouble:
-	fldt	-24(%rsp)
-	ret
-
-	.align 2
-.Lld_struct:
-	/* There are four possibilities here, %rax/%rdx, %xmm0/%rax,
-	   %rax/%xmm0, %xmm0/%xmm1.  We collapse two by always loading
-	   both rdx and xmm1 with the second word.  For the remaining,
-	   bit 8 set means xmm0 gets the second word, and bit 9 means
-	   that rax gets the second word.  */
-	movq	-24(%rsp), %rcx
-	movq	-16(%rsp), %rdx
-	movq	-16(%rsp), %xmm1
-	testl	$0x100, %eax
-	cmovnz	%rdx, %rcx
-	movd	%rcx, %xmm0
-	testl	$0x200, %eax
-	movq	-24(%rsp), %rax
-	cmovnz	%rdx, %rax
-	ret
-
-	/* See the comment above .Lload_sse; the same logic applies here.  */
-	.align 2
-.LUW8:
-.Lsave_sse:
-	movdqa	%xmm0, 48(%rsp)
-	movdqa	%xmm1, 64(%rsp)
-	movdqa	%xmm2, 80(%rsp)
-	movdqa	%xmm3, 96(%rsp)
-	movdqa	%xmm4, 112(%rsp)
-	movdqa	%xmm5, 128(%rsp)
-	movdqa	%xmm6, 144(%rsp)
-	movdqa	%xmm7, 160(%rsp)
-	jmp	.Lret_from_save_sse
-
-.LUW9:
 	.size	ffi_closure_unix64,.-ffi_closure_unix64
 
-#ifdef __GNUC__
+	.align	2
-/* Only emit DWARF unwind info when building with the GNU toolchain.  */
+	.globl	ffi_go_closure_unix64_sse
+	.type	ffi_go_closure_unix64_sse,@function
+	FFI_HIDDEN(ffi_go_closure_unix64_sse)
 
-#ifdef HAVE_AS_X86_64_UNWIND_SECTION_TYPE
+ffi_go_closure_unix64_sse:
-	.section	.eh_frame,"a",@unwind
+	cfi_startproc
+	subq	$ffi_closure_FS, %rsp
+	/* Note clang bug 21515: adjust_cfa_offset error across endproc.  */
+	cfi_def_cfa_offset(ffi_closure_FS + 8)
+
+	movdqa	%xmm0, ffi_closure_OFS_V+0x00(%rsp)
+	movdqa	%xmm1, ffi_closure_OFS_V+0x10(%rsp)
+	movdqa	%xmm2, ffi_closure_OFS_V+0x20(%rsp)
+	movdqa	%xmm3, ffi_closure_OFS_V+0x30(%rsp)
+	movdqa	%xmm4, ffi_closure_OFS_V+0x40(%rsp)
+	movdqa	%xmm5, ffi_closure_OFS_V+0x50(%rsp)
+	movdqa	%xmm6, ffi_closure_OFS_V+0x60(%rsp)
+	movdqa	%xmm7, ffi_closure_OFS_V+0x70(%rsp)
+	jmp	0f
+
+	cfi_endproc
+	.size	ffi_go_closure_unix64_sse,.-ffi_go_closure_unix64_sse
+
+	.align	2
+	.globl	ffi_go_closure_unix64
+	.type	ffi_go_closure_unix64,@function
+	FFI_HIDDEN(ffi_go_closure_unix64)
+
+ffi_go_closure_unix64:
+	cfi_startproc
+	subq	$ffi_closure_FS, %rsp
+	/* Note clang bug 21515: adjust_cfa_offset error across endproc.  */
+	cfi_def_cfa_offset(ffi_closure_FS + 8)
+0:
+	movq	%rdi, ffi_closure_OFS_G+0x00(%rsp)
+	movq    %rsi, ffi_closure_OFS_G+0x08(%rsp)
+	movq    %rdx, ffi_closure_OFS_G+0x10(%rsp)
+	movq    %rcx, ffi_closure_OFS_G+0x18(%rsp)
+	movq    %r8,  ffi_closure_OFS_G+0x20(%rsp)
+	movq    %r9,  ffi_closure_OFS_G+0x28(%rsp)
+
+#ifdef __ILP32__
+	movl	4(%r10), %edi		/* Load cif */
+	movl	8(%r10), %esi		/* Load fun */
+	movl	%r10d, %edx		/* Load closure (user_data) */
 #else
-	.section	.eh_frame,"a",@progbits
+	movq	8(%r10), %rdi		/* Load cif */
+	movq	16(%r10), %rsi		/* Load fun */
+	movq	%r10, %rdx		/* Load closure (user_data) */
 #endif
-.Lframe1:
+	jmp	.Ldo_closure
-	.long	.LECIE1-.LSCIE1		/* CIE Length */
-.LSCIE1:
-	.long	0			/* CIE Identifier Tag */
-	.byte	1			/* CIE Version */
-	.ascii "zR\0"			/* CIE Augmentation */
-	.uleb128 1			/* CIE Code Alignment Factor */
-	.sleb128 -8			/* CIE Data Alignment Factor */
-	.byte	0x10			/* CIE RA Column */
-	.uleb128 1			/* Augmentation size */
-	.byte	0x1b			/* FDE Encoding (pcrel sdata4) */
-	.byte	0xc			/* DW_CFA_def_cfa, %rsp offset 8 */
-	.uleb128 7
-	.uleb128 8
-	.byte	0x80+16			/* DW_CFA_offset, %rip offset 1*-8 */
-	.uleb128 1
-	.align 8
-.LECIE1:
-.LSFDE1:
-	.long	.LEFDE1-.LASFDE1	/* FDE Length */
-.LASFDE1:
-	.long	.LASFDE1-.Lframe1	/* FDE CIE offset */
-#if HAVE_AS_X86_PCREL
-	.long	.LUW0-.			/* FDE initial location */
-#else
-	.long	.LUW0@rel
-#endif
-	.long	.LUW4-.LUW0		/* FDE address range */
-	.uleb128 0x0			/* Augmentation size */
 
-	.byte	0x4			/* DW_CFA_advance_loc4 */
+	cfi_endproc
-	.long	.LUW1-.LUW0
+	.size	ffi_go_closure_unix64,.-ffi_go_closure_unix64
-
-	/* New stack frame based off rbp.  This is a itty bit of unwind
-	   trickery in that the CFA *has* changed.  There is no easy way
-	   to describe it correctly on entry to the function.  Fortunately,
-	   it doesn't matter too much since at all points we can correctly
-	   unwind back to ffi_call.  Note that the location to which we
-	   moved the return address is (the new) CFA-8, so from the
-	   perspective of the unwind info, it hasn't moved.  */
-	.byte	0xc			/* DW_CFA_def_cfa, %rbp offset 32 */
-	.uleb128 6
-	.uleb128 32
-	.byte	0x80+6			/* DW_CFA_offset, %rbp offset 2*-8 */
-	.uleb128 2
-	.byte	0xa			/* DW_CFA_remember_state */
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW2-.LUW1
-	.byte	0xc			/* DW_CFA_def_cfa, %rsp offset 8 */
-	.uleb128 7
-	.uleb128 8
-	.byte	0xc0+6			/* DW_CFA_restore, %rbp */
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW3-.LUW2
-	.byte	0xb			/* DW_CFA_restore_state */
-
-	.align 8
-.LEFDE1:
-.LSFDE3:
-	.long	.LEFDE3-.LASFDE3	/* FDE Length */
-.LASFDE3:
-	.long	.LASFDE3-.Lframe1	/* FDE CIE offset */
-#if HAVE_AS_X86_PCREL
-	.long	.LUW5-.			/* FDE initial location */
-#else
-	.long	.LUW5@rel
-#endif
-	.long	.LUW9-.LUW5		/* FDE address range */
-	.uleb128 0x0			/* Augmentation size */
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW6-.LUW5
-	.byte	0xe			/* DW_CFA_def_cfa_offset */
-	.uleb128 208
-	.byte	0xa			/* DW_CFA_remember_state */
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW7-.LUW6
-	.byte	0xe			/* DW_CFA_def_cfa_offset */
-	.uleb128 8
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW8-.LUW7
-	.byte	0xb			/* DW_CFA_restore_state */
-
-	.align 8
-.LEFDE3:
-
-#endif /* __GNUC__ */
 
 #endif /* __x86_64__ */
-
 #if defined __ELF__ && defined __linux__
 	.section	.note.GNU-stack,"",@progbits
 #endif

src/x86/win64.S

@@ -1,264 +1,16 @@
 #define LIBFFI_ASM
 #include <fficonfig.h>
 #include <ffi.h>
+#include <ffi_cfi.h>
 
-/* Constants for ffi_call_win64 */
+#if defined(HAVE_AS_CFI_PSEUDO_OP)
-#define STACK 0
+        .cfi_sections   .debug_frame
-#define PREP_ARGS_FN 32
+#endif
-#define ECIF 40
-#define CIF_BYTES 48
-#define CIF_FLAGS 56
-#define RVALUE 64
-#define FN 72
 
-/* ffi_call_win64 (void (*prep_args_fn)(char *, extended_cif *),
+#define arg0	%rcx
-		   extended_cif *ecif, unsigned bytes, unsigned flags,
+#define arg1	%rdx
-		   unsigned *rvalue, void (*fn)());
+#define arg2	%r8
- */
+#define arg3	%r9
-
-#ifdef _MSC_VER
-PUBLIC	ffi_call_win64
-
-EXTRN	__chkstk:NEAR
-EXTRN	ffi_closure_win64_inner:NEAR
-
-_TEXT	SEGMENT
-
-;;; ffi_closure_win64 will be called with these registers set:
-;;;    rax points to 'closure'
-;;;    r11 contains a bit mask that specifies which of the
-;;;    first four parameters are float or double
-;;;
-;;; It must move the parameters passed in registers to their stack location,
-;;; call ffi_closure_win64_inner for the actual work, then return the result.
-;;;
-ffi_closure_win64 PROC FRAME
-	;; copy register arguments onto stack
-	test	r11, 1
-	jne	first_is_float
-	mov	QWORD PTR [rsp+8], rcx
-	jmp	second
-first_is_float:
-	movlpd	QWORD PTR [rsp+8], xmm0
-
-second:
-	test	r11, 2
-	jne	second_is_float
-	mov	QWORD PTR [rsp+16], rdx
-	jmp	third
-second_is_float:
-	movlpd	QWORD PTR [rsp+16], xmm1
-
-third:
-	test	r11, 4
-	jne	third_is_float
-	mov	QWORD PTR [rsp+24], r8
-	jmp	fourth
-third_is_float:
-	movlpd	QWORD PTR [rsp+24], xmm2
-
-fourth:
-	test	r11, 8
-	jne	fourth_is_float
-	mov	QWORD PTR [rsp+32], r9
-	jmp	done
-fourth_is_float:
-	movlpd	QWORD PTR [rsp+32], xmm3
-
-done:
-	.ALLOCSTACK 40
-	sub	rsp, 40
-	.ENDPROLOG
-	mov	rcx, rax	; context is first parameter
-	mov	rdx, rsp	; stack is second parameter
-	add	rdx, 48		; point to start of arguments
-	mov	rax, ffi_closure_win64_inner
-	call	rax		; call the real closure function
-	add	rsp, 40
-	movd	xmm0, rax	; If the closure returned a float,
-				; ffi_closure_win64_inner wrote it to rax
-	ret	0
-ffi_closure_win64 ENDP
-
-ffi_call_win64 PROC FRAME
-	;; copy registers onto stack
-	mov	QWORD PTR [rsp+32], r9
-	mov	QWORD PTR [rsp+24], r8
-	mov	QWORD PTR [rsp+16], rdx
-	mov	QWORD PTR [rsp+8], rcx
-	.PUSHREG rbp
-	push	rbp
-	.ALLOCSTACK 48
-	sub	rsp, 48					; 00000030H
-	.SETFRAME rbp, 32
-	lea	rbp, QWORD PTR [rsp+32]
-	.ENDPROLOG
-
-	mov	eax, DWORD PTR CIF_BYTES[rbp]
-	add	rax, 15
-	and	rax, -16
-	call	__chkstk
-	sub	rsp, rax
-	lea	rax, QWORD PTR [rsp+32]
-	mov	QWORD PTR STACK[rbp], rax
-
-	mov	rdx, QWORD PTR ECIF[rbp]
-	mov	rcx, QWORD PTR STACK[rbp]
-	call	QWORD PTR PREP_ARGS_FN[rbp]
-
-	mov	rsp, QWORD PTR STACK[rbp]
-
-	movlpd	xmm3, QWORD PTR [rsp+24]
-	movd	r9, xmm3
-
-	movlpd	xmm2, QWORD PTR [rsp+16]
-	movd	r8, xmm2
-
-	movlpd	xmm1, QWORD PTR [rsp+8]
-	movd	rdx, xmm1
-
-	movlpd	xmm0, QWORD PTR [rsp]
-	movd	rcx, xmm0
-
-	call	QWORD PTR FN[rbp]
-ret_struct4b$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SMALL_STRUCT_4B
- 	jne	ret_struct2b$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	mov	DWORD PTR [rcx], eax
-	jmp	ret_void$
-
-ret_struct2b$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SMALL_STRUCT_2B
- 	jne	ret_struct1b$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	mov	WORD PTR [rcx], ax
-	jmp	ret_void$
-
-ret_struct1b$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SMALL_STRUCT_1B
- 	jne	ret_uint8$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	mov	BYTE PTR [rcx], al
-	jmp	ret_void$
-
-ret_uint8$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_UINT8
- 	jne	ret_sint8$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	movzx   rax, al
-	mov	QWORD PTR [rcx], rax
-	jmp	ret_void$
-
-ret_sint8$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SINT8
- 	jne	ret_uint16$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	movsx   rax, al
-	mov	QWORD PTR [rcx], rax
-	jmp	ret_void$
-
-ret_uint16$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_UINT16
- 	jne	ret_sint16$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	movzx   rax, ax
-	mov	QWORD PTR [rcx], rax
-	jmp	SHORT ret_void$
-
-ret_sint16$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SINT16
- 	jne	ret_uint32$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	movsx   rax, ax
-	mov	QWORD PTR [rcx], rax
-	jmp	SHORT ret_void$
-
-ret_uint32$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_UINT32
- 	jne	ret_sint32$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	mov     eax, eax
-	mov	QWORD PTR [rcx], rax
-	jmp	SHORT ret_void$
-
-ret_sint32$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SINT32
- 	jne	ret_float$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	cdqe
-	mov	QWORD PTR [rcx], rax
-	jmp	SHORT ret_void$
-
-ret_float$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_FLOAT
- 	jne	SHORT ret_double$
-
- 	mov	rax, QWORD PTR RVALUE[rbp]
- 	movss	DWORD PTR [rax], xmm0
- 	jmp	SHORT ret_void$
-
-ret_double$:
- 	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_DOUBLE
- 	jne	SHORT ret_uint64$
-
- 	mov	rax, QWORD PTR RVALUE[rbp]
- 	movlpd	QWORD PTR [rax], xmm0
- 	jmp	SHORT ret_void$
-
-ret_uint64$:
-  	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_UINT64
-  	jne	SHORT ret_sint64$
-
- 	mov	rcx, QWORD PTR RVALUE[rbp]
- 	mov	QWORD PTR [rcx], rax
- 	jmp	SHORT ret_void$
-
-ret_sint64$:
-  	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_SINT64
-  	jne	SHORT ret_pointer$
-
- 	mov	rcx, QWORD PTR RVALUE[rbp]
- 	mov	QWORD PTR [rcx], rax
- 	jmp	SHORT ret_void$
-
-ret_pointer$:
-  	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_POINTER
-  	jne	SHORT ret_int$
-
- 	mov	rcx, QWORD PTR RVALUE[rbp]
- 	mov	QWORD PTR [rcx], rax
- 	jmp	SHORT ret_void$
-
-ret_int$:
-  	cmp	DWORD PTR CIF_FLAGS[rbp], FFI_TYPE_INT
-  	jne	SHORT ret_void$
-
-	mov	rcx, QWORD PTR RVALUE[rbp]
-	cdqe
-	mov	QWORD PTR [rcx], rax
- 	jmp	SHORT ret_void$
-
-ret_void$:
-	xor	rax, rax
-
-	lea	rsp, QWORD PTR [rbp+16]
-	pop	rbp
-	ret	0
-ffi_call_win64 ENDP
-_TEXT	ENDS
-END
-
-#else
 
 #ifdef SYMBOL_UNDERSCORE
 #define SYMBOL_NAME(name) _##name
@@ -266,255 +18,202 @@ END
 #define SYMBOL_NAME(name) name
 #endif
 
+.macro E which
+	.align	8
+	.org	0b + \which * 8
+.endm
+
 	.text
 
-.extern SYMBOL_NAME(ffi_closure_win64_inner)
+/* ffi_call_win64 (void *stack, struct win64_call_frame *frame, void *r10)
 
-# ffi_closure_win64 will be called with these registers set:
+   Bit o trickiness here -- FRAME is the base of the stack frame
-#    rax points to 'closure'
+   for this function.  This has been allocated by ffi_call.  We also
-#    r11 contains a bit mask that specifies which of the
+   deallocate some of the stack that has been alloca'd.  */
-#    first four parameters are float or double
-#
-# It must move the parameters passed in registers to their stack location,
-# call ffi_closure_win64_inner for the actual work, then return the result.
-#
-	.balign 16
-	.globl SYMBOL_NAME(ffi_closure_win64)
-	.seh_proc SYMBOL_NAME(ffi_closure_win64)
-SYMBOL_NAME(ffi_closure_win64):
-	# copy register arguments onto stack
-	test	$1,%r11
-	jne	.Lfirst_is_float
-	mov	%rcx, 8(%rsp)
-	jmp	.Lsecond
-.Lfirst_is_float:
-	movlpd	%xmm0, 8(%rsp)
 
-.Lsecond:
+	.align	8
-	test	$2, %r11
+	.globl	ffi_call_win64
-	jne	.Lsecond_is_float
-	mov	%rdx, 16(%rsp)
-	jmp	.Lthird
-.Lsecond_is_float:
-	movlpd	%xmm1, 16(%rsp)
 
-.Lthird:
+	.seh_proc ffi_call_win64
-	test	$4, %r11
+ffi_call_win64:
-	jne	.Lthird_is_float
+	cfi_startproc
-	mov	%r8,24(%rsp)
+	/* Set up the local stack frame and install it in rbp/rsp.  */
-	jmp	.Lfourth
+	movq	(%rsp), %rax
-.Lthird_is_float:
+	movq	%rbp, (arg1)
-	movlpd	%xmm2, 24(%rsp)
+	movq	%rax, 8(arg1)
-
+	movq	arg1, %rbp
-.Lfourth:
+	cfi_def_cfa(%rbp, 16)
-	test	$8, %r11
+	cfi_rel_offset(%rbp, 0)
-	jne	.Lfourth_is_float
+	.seh_pushreg %rbp
-	mov	%r9, 32(%rsp)
+	.seh_setframe %rbp, 0
-	jmp	.Ldone
-.Lfourth_is_float:
-	movlpd	%xmm3, 32(%rsp)
-
-.Ldone:
-	.seh_stackalloc 40
-	sub	$40, %rsp
 	.seh_endprologue
-	mov	%rax, %rcx	# context is first parameter
+	movq	arg0, %rsp
-	mov	%rsp, %rdx	# stack is second parameter
-	add	$48, %rdx	# point to start of arguments
-	leaq	SYMBOL_NAME(ffi_closure_win64_inner)(%rip), %rax
-	callq	*%rax		# call the real closure function
-	add	$40, %rsp
-	movq	%rax, %xmm0	# If the closure returned a float,
-				# ffi_closure_win64_inner wrote it to rax
-	retq
-	.seh_endproc
 
-	.balign 16
+	movq	arg2, %r10
-	.globl	SYMBOL_NAME(ffi_call_win64)
-	.seh_proc SYMBOL_NAME(ffi_call_win64)
-SYMBOL_NAME(ffi_call_win64):
-	# copy registers onto stack
-	mov	%r9,32(%rsp)
-	mov	%r8,24(%rsp)
-	mov	%rdx,16(%rsp)
-	mov	%rcx,8(%rsp)
-	.seh_pushreg rbp
-	push	%rbp
-	.seh_stackalloc 48
-	sub	$48,%rsp
-	.seh_setframe rbp, 32
-	lea	32(%rsp),%rbp
-	.seh_endprologue
 
-	mov	CIF_BYTES(%rbp),%eax
+	/* Load all slots into both general and xmm registers.  */
-	add	$15, %rax
+	movq	(%rsp), %rcx
-	and	$-16, %rax
+	movsd	(%rsp), %xmm0
-	cmpq	$0x1000, %rax
+	movq	8(%rsp), %rdx
-	jb	Lch_done
+	movsd	8(%rsp), %xmm1
-Lch_probe:
+	movq	16(%rsp), %r8
-	subq	$0x1000,%rsp
+	movsd	16(%rsp), %xmm2
-	orl	$0x0, (%rsp)
+	movq	24(%rsp), %r9
-	subq	$0x1000,%rax
+	movsd	24(%rsp), %xmm3
-	cmpq	$0x1000,%rax
-	ja	Lch_probe
-Lch_done:
-	subq	%rax, %rsp
-	orl	$0x0, (%rsp)
-	lea	32(%rsp), %rax
-	mov	%rax, STACK(%rbp)
 
-	mov	ECIF(%rbp), %rdx
+	call	*16(%rbp)
-	mov	STACK(%rbp), %rcx
-	callq	*PREP_ARGS_FN(%rbp)
 
-	mov	STACK(%rbp), %rsp
+	movl	24(%rbp), %ecx
+	movq	32(%rbp), %r8
+	leaq	0f(%rip), %r10
+	cmpl	$FFI_TYPE_SMALL_STRUCT_4B, %ecx
+	leaq	(%r10, %rcx, 8), %r10
+	ja	99f
+	jmp	*%r10
 
-	movlpd	24(%rsp), %xmm3
+/* Below, we're space constrained most of the time.  Thus we eschew the
-	movd	%xmm3, %r9
+   modern "mov, pop, ret" sequence (5 bytes) for "leave, ret" (2 bytes).  */
+.macro epilogue
+	leaveq
+	cfi_remember_state
+	cfi_def_cfa(%rsp, 8)
+	cfi_restore(%rbp)
+	ret
+	cfi_restore_state
+.endm
 
-	movlpd	16(%rsp), %xmm2
+	.align	8
-	movd	%xmm2, %r8
+0:
-
+E FFI_TYPE_VOID
-	movlpd	8(%rsp), %xmm1
+	epilogue
-	movd	%xmm1, %rdx
+E FFI_TYPE_INT
-
+	movslq	%eax, %rax
-	movlpd	(%rsp), %xmm0
+	movq	%rax, (%r8)
-	movd	%xmm0, %rcx
+	epilogue
-
+E FFI_TYPE_FLOAT
-	callq	*FN(%rbp)
+	movss	%xmm0, (%r8)
-.Lret_struct4b:
+	epilogue
- 	cmpl	$FFI_TYPE_SMALL_STRUCT_4B, CIF_FLAGS(%rbp)
+E FFI_TYPE_DOUBLE
- 	jne .Lret_struct2b
+	movsd	%xmm0, (%r8)
-
+	epilogue
-	mov	RVALUE(%rbp), %rcx
+E FFI_TYPE_LONGDOUBLE
-	mov	%eax, (%rcx)
+	call	abort
-	jmp	.Lret_void
+E FFI_TYPE_UINT8
-
+	movzbl	%al, %eax
-.Lret_struct2b:
+	movq	%rax, (%r8)
-	cmpl	$FFI_TYPE_SMALL_STRUCT_2B, CIF_FLAGS(%rbp)
+	epilogue
-	jne .Lret_struct1b
+E FFI_TYPE_SINT8
-
-	mov	RVALUE(%rbp), %rcx
-	mov	%ax, (%rcx)
-	jmp .Lret_void
-
-.Lret_struct1b:
-	cmpl	$FFI_TYPE_SMALL_STRUCT_1B, CIF_FLAGS(%rbp)
-	jne .Lret_uint8
-
-	mov	RVALUE(%rbp), %rcx
-	mov	%al, (%rcx)
-	jmp .Lret_void
-
-.Lret_uint8:
-	cmpl	$FFI_TYPE_UINT8, CIF_FLAGS(%rbp)
-	jne .Lret_sint8
-
-	mov     RVALUE(%rbp), %rcx
-	movzbq  %al, %rax
-	movq    %rax, (%rcx)
-	jmp .Lret_void
-
-.Lret_sint8:
-	cmpl	$FFI_TYPE_SINT8, CIF_FLAGS(%rbp)
-	jne .Lret_uint16
-
-	mov     RVALUE(%rbp), %rcx
 	movsbq	%al, %rax
-	movq    %rax, (%rcx)
+	jmp	98f
-	jmp .Lret_void
+E FFI_TYPE_UINT16
-
+	movzwl	%ax, %eax
-.Lret_uint16:
+	movq	%rax, (%r8)
-	cmpl	$FFI_TYPE_UINT16, CIF_FLAGS(%rbp)
+	epilogue
-	jne .Lret_sint16
+E FFI_TYPE_SINT16
-
-	mov     RVALUE(%rbp), %rcx
-	movzwq  %ax, %rax
-	movq    %rax, (%rcx)
-	jmp .Lret_void
-
-.Lret_sint16:
-	cmpl	$FFI_TYPE_SINT16, CIF_FLAGS(%rbp)
-	jne .Lret_uint32
-
-	mov     RVALUE(%rbp), %rcx
 	movswq	%ax, %rax
-	movq    %rax, (%rcx)
+	jmp	98f
-	jmp .Lret_void
+E FFI_TYPE_UINT32
-
-.Lret_uint32:
-	cmpl	$FFI_TYPE_UINT32, CIF_FLAGS(%rbp)
-	jne .Lret_sint32
-
-	mov     RVALUE(%rbp), %rcx
 	movl	%eax, %eax
-	movq    %rax, (%rcx)
+	movq	%rax, (%r8)
-	jmp .Lret_void
+	epilogue
+E FFI_TYPE_SINT32
+	movslq	%eax, %rax
+	movq	%rax, (%r8)
+	epilogue
+E FFI_TYPE_UINT64
+98:	movq	%rax, (%r8)
+	epilogue
+E FFI_TYPE_SINT64
+	movq	%rax, (%r8)
+	epilogue
+E FFI_TYPE_STRUCT
+	epilogue
+E FFI_TYPE_POINTER
+	movq	%rax, (%r8)
+	epilogue
+E FFI_TYPE_COMPLEX
+	call	abort
+E FFI_TYPE_SMALL_STRUCT_1B
+	movb	%al, (%r8)
+	epilogue
+E FFI_TYPE_SMALL_STRUCT_2B
+	movw	%ax, (%r8)
+	epilogue
+E FFI_TYPE_SMALL_STRUCT_4B
+	movl	%eax, (%r8)
+	epilogue
 
-.Lret_sint32:
+	.align	8
- 	cmpl	$FFI_TYPE_SINT32, CIF_FLAGS(%rbp)
+99:	call	abort
- 	jne	.Lret_float
 
-	mov	RVALUE(%rbp), %rcx
+.purgem epilogue
-	cltq
-	movq	%rax, (%rcx)
-	jmp	.Lret_void
 
-.Lret_float:
+	cfi_endproc
- 	cmpl	$FFI_TYPE_FLOAT, CIF_FLAGS(%rbp)
- 	jne	.Lret_double
-
- 	mov	RVALUE(%rbp), %rax
- 	movss	%xmm0, (%rax)
- 	jmp	.Lret_void
-
-.Lret_double:
- 	cmpl	$FFI_TYPE_DOUBLE, CIF_FLAGS(%rbp)
- 	jne	.Lret_uint64
-
- 	mov	RVALUE(%rbp), %rax
- 	movlpd	%xmm0, (%rax)
- 	jmp	.Lret_void
-
-.Lret_uint64:
-  	cmpl	$FFI_TYPE_UINT64, CIF_FLAGS(%rbp)
- 	jne	.Lret_sint64
-
- 	mov	RVALUE(%rbp), %rcx
- 	mov	%rax, (%rcx)
- 	jmp	.Lret_void
-
-.Lret_sint64:
-  	cmpl	$FFI_TYPE_SINT64, CIF_FLAGS(%rbp)
-  	jne	.Lret_pointer
-
- 	mov	RVALUE(%rbp), %rcx
- 	mov	%rax, (%rcx)
- 	jmp	.Lret_void
-
-.Lret_pointer:
-  	cmpl	$FFI_TYPE_POINTER, CIF_FLAGS(%rbp)
-  	jne	.Lret_int
-
- 	mov	RVALUE(%rbp), %rcx
- 	mov	%rax, (%rcx)
- 	jmp	.Lret_void
-
-.Lret_int:
-  	cmpl	$FFI_TYPE_INT, CIF_FLAGS(%rbp)
-  	jne	.Lret_void
-
-	mov	RVALUE(%rbp), %rcx
-	cltq
-	movq	%rax, (%rcx)
-	jmp	.Lret_void
-
-.Lret_void:
-	xor	%rax, %rax
-
-	lea	16(%rbp), %rsp
-	pop	%rbp
-	retq
 	.seh_endproc
-#endif /* !_MSC_VER */
 
+
+/* 32 bytes of outgoing register stack space, 8 bytes of alignment,
+   16 bytes of result, 32 bytes of xmm registers.  */
+#define ffi_clo_FS	(32+8+16+32)
+#define ffi_clo_OFF_R	(32+8)
+#define ffi_clo_OFF_X	(32+8+16)
+
+	.align	8
+	.globl	ffi_go_closure_win64
+
+	.seh_proc ffi_go_closure_win64
+ffi_go_closure_win64:
+	cfi_startproc
+	/* Save all integer arguments into the incoming reg stack space.  */
+	movq	arg0, 8(%rsp)
+	movq	arg1, 16(%rsp)
+	movq	arg2, 24(%rsp)
+	movq	arg3, 32(%rsp)
+
+	movq	8(%r10), arg0			/* load cif */
+	movq	16(%r10), arg1			/* load fun */
+	movq	%r10, arg2			/* closure is user_data */
+	jmp	0f
+	cfi_endproc
+	.seh_endproc
+
+	.align	8
+	.globl	ffi_closure_win64
+
+	.seh_proc ffi_closure_win64
+ffi_closure_win64:
+	cfi_startproc
+	/* Save all integer arguments into the incoming reg stack space.  */
+	movq	arg0, 8(%rsp)
+	movq	arg1, 16(%rsp)
+	movq	arg2, 24(%rsp)
+	movq	arg3, 32(%rsp)
+
+	movq	FFI_TRAMPOLINE_SIZE(%r10), arg0		/* load cif */
+	movq	FFI_TRAMPOLINE_SIZE+8(%r10), arg1	/* load fun */
+	movq	FFI_TRAMPOLINE_SIZE+16(%r10), arg2	/* load user_data */
+0:
+	subq	$ffi_clo_FS, %rsp
+	cfi_adjust_cfa_offset(ffi_clo_FS)
+	.seh_stackalloc ffi_clo_FS
+	.seh_endprologue
+
+	/* Save all sse arguments into the stack frame.  */
+	movsd	%xmm0, ffi_clo_OFF_X(%rsp)
+	movsd	%xmm1, ffi_clo_OFF_X+8(%rsp)
+	movsd	%xmm2, ffi_clo_OFF_X+16(%rsp)
+	movsd	%xmm3, ffi_clo_OFF_X+24(%rsp)
+
+	leaq	ffi_clo_OFF_R(%rsp), arg3
+	call	ffi_closure_win64_inner
+
+	/* Load the result into both possible result registers.  */
+	movq    ffi_clo_OFF_R(%rsp), %rax
+	movsd   ffi_clo_OFF_R(%rsp), %xmm0
+
+	addq	$ffi_clo_FS, %rsp
+	cfi_adjust_cfa_offset(-ffi_clo_FS)
+	ret
+
+	cfi_endproc
+	.seh_endproc

testsuite/libffi.call/call.exp

@@ -24,16 +24,20 @@ set ctlist [lsearch -inline -all -glob [lsort [glob -nocomplain -- $srcdir/$subd
 
 run-many-tests $tlist ""
 
-if { ![istarget s390*] } {
+# ??? We really should preprocess ffi.h and grep
-
+# for FFI_TARGET_HAS_COMPLEX_TYPE.
+if { [istarget aarch64*]
+     || [istarget alpha*]
+     || [istarget arm*]
+     || [istarget i?86*]
+     || [istarget s390*]
+     || [istarget sparc*]
+     || [istarget x86_64*] } {
+    run-many-tests $ctlist ""
+} else {
     foreach test $ctlist {
 	unsupported "$test"
     }
-
-} else {
-
-  run-many-tests $ctlist ""
-
 }
 
 dg-finish

testsuite/libffi.call/cls_align_longdouble_split.c

@@ -4,10 +4,8 @@
    PR:		none.
    Originator:	<hos@tamanegi.org> 20031203	 */
 
-/* { dg-excess-errors "no long double format" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 /* { dg-do run { xfail strongarm*-*-* xscale*-*-* } } */
 /* { dg-options -mlong-double-128 { target powerpc64*-*-linux* } } */
-/* { dg-output "" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 
 #include "ffitest.h"
 

testsuite/libffi.call/cls_align_longdouble_split2.c

@@ -5,10 +5,8 @@
 	Originator:		Blake Chaffin	6/18/2007
 */
 
-/* { dg-excess-errors "no long double format" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 /* { dg-do run { xfail strongarm*-*-* } } */
 /* { dg-options -mlong-double-128 { target powerpc64*-*-linux* } } */
-/* { dg-output "" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 
 #include "ffitest.h"
 

testsuite/libffi.call/cls_complex_va_float.c

@@ -6,5 +6,11 @@
 
 /* { dg-do run } */
 
+/* Alpha splits _Complex into two arguments.  It's illegal to pass
+   float through varargs, so _Complex float goes badly.  In sort of
+   gets passed as _Complex double, but the compiler doesn't agree
+   with itself on this issue.  */
+/* { dg-do run { xfail alpha*-*-* } } */
+
 #include "complex_defs_float.inc"
 #include "cls_complex_va.inc"

testsuite/libffi.call/cls_longdouble.c

@@ -4,12 +4,10 @@
    PR:			none.
    Originator:	Blake Chaffin	*/
 
-/* { dg-excess-errors "no long double format" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 /* This test is known to PASS on armv7l-unknown-linux-gnueabihf, so I have
    remove the xfail for arm*-*-* below, until we know more.  */
 /* { dg-do run { xfail strongarm*-*-* xscale*-*-* } } */
 /* { dg-options -mlong-double-128 { target powerpc64*-*-linux* } } */
-/* { dg-output "" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 
 #include "ffitest.h"
 

testsuite/libffi.call/complex_int.c

@@ -12,9 +12,9 @@
 
 _Complex int f_complex(_Complex int c, int x, int *py)
 {
-  c = -(2 * creal (c)) + (cimag (c) + 1)* I;
+  __real__ c = -2 * __real__ c;
+  __imag__ c = __imag__ c + 1;
   *py += x;
-
   return c;
 }
 

testsuite/libffi.call/float2.c

@@ -3,9 +3,7 @@
    Limitations:	none.
    PR:		none.
    Originator:	From the original ffitest.c  */
-
+/* { dg-do run } */
-/* { dg-excess-errors "fails" { target x86_64-*-mingw* x86_64-*-cygwin* } } */
-/* { dg-do run { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 
 #include "ffitest.h"
 #include "float.h"

testsuite/libffi.call/huge_struct.c

@@ -5,11 +5,9 @@
 	Originator:		Blake Chaffin	6/18/2007
 */
 
-/* { dg-excess-errors "" { target x86_64-*-mingw* x86_64-*-cygwin* } } */
 /* { dg-do run { xfail strongarm*-*-* xscale*-*-* } } */
 /* { dg-options -mlong-double-128 { target powerpc64*-*-linux* } } */
 /* { dg-options -Wformat=0 { target moxie*-*-elf } } */
-/* { dg-output "" { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 
 #include "ffitest.h"
 

testsuite/libffi.call/return_complex2.inc

@@ -2,10 +2,14 @@
 #include "ffitest.h"
 #include <complex.h>
 
-static _Complex T_C_TYPE return_c(_Complex T_C_TYPE c1, _Complex T_C_TYPE c2, unsigned int in3, _Complex T_C_TYPE c4)
+_Complex T_C_TYPE
+return_c(_Complex T_C_TYPE c1, _Complex T_C_TYPE c2,
+	 unsigned int in3, _Complex T_C_TYPE c4)
 {
-  return c1 + c2 + in3 + c4;
+  volatile _Complex T_C_TYPE r = c1 + c2 + in3 + c4;
+  return r;
 }
+
 int main (void)
 {
   ffi_cif cif;
@@ -35,6 +39,6 @@ int main (void)
   printf ("%f,%fi vs %f,%fi\n",
 	  T_CONV creal (rc), T_CONV cimag (rc),
 	  T_CONV creal (rc2), T_CONV cimag (rc2));
-  CHECK(rc ==  c1 + c2 + in3 + c4);
+  CHECK(rc == rc2);
   exit(0);
 }

testsuite/libffi.call/return_ldl.c

@@ -3,8 +3,8 @@
    Limitations:	none.
    PR:		none.
    Originator:	<andreast@gcc.gnu.org> 20071113  */
+/* { dg-do run } */
 
-/* { dg-do run { xfail x86_64-*-mingw* x86_64-*-cygwin* } } */
 #include "ffitest.h"
 
 static long double return_ldl(long double ldl)