ffi 1.9.18-x64-mingw32 → 1.9.21-x64-mingw32

data/ext/ffi_c/libffi/src/x86/unix64.S

@@ -1,6 +1,7 @@
 /* -----------------------------------------------------------------------
-   unix64.S - Copyright (c) 2002  Bo Thorsen <bo@suse.de>
-	      Copyright (c) 2008  Red Hat, Inc
+   unix64.S - Copyright (c) 2013  The Written Word, Inc.
+	    - Copyright (c) 2008  Red Hat, Inc
+	    - Copyright (c) 2002  Bo Thorsen <bo@suse.de>
 
    x86-64 Foreign Function Interface 
 
@@ -29,8 +30,20 @@
 #define LIBFFI_ASM	
 #include <fficonfig.h>
 #include <ffi.h>
+#include "internal64.h"
+#include "asmnames.h"
 
-.text
+	.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.  */
+#if defined(__clang__) || defined(__APPLE__) || (defined (__sun__) && defined(__svr4__))
+# define E(BASE, X)	.balign 8
+#else
+# define E(BASE, X)	.balign 8; .org BASE + X * 8
+#endif
 
 /* ffi_call_unix64 (void *args, unsigned long bytes, unsigned flags,
 	            void *raddr, void (*fnaddr)(void));
@@ -39,12 +52,12 @@
    for this function.  This has been allocated by ffi_call.  We also
    deallocate some of the stack that has been alloca'd.  */
 
-	.align	2
-	.globl	ffi_call_unix64
-	.type	ffi_call_unix64,@function
+	.balign	8
+	.globl	C(ffi_call_unix64)
+	FFI_HIDDEN(C(ffi_call_unix64))
 
-ffi_call_unix64:
-.LUW0:
+C(ffi_call_unix64):
+L(UW0):
 	movq	(%rsp), %r10		/* Load return address.  */
 	leaq	(%rdi, %rsi), %rax	/* Find local stack base.  */
 	movq	%rdx, (%rax)		/* Save flags.  */
@@ -52,24 +65,37 @@ 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.  */
+L(UW1):
+	/* 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	16(%r10), %rdx
-	movq	24(%r10), %rcx
-	movq	32(%r10), %r8
-	movq	40(%r10), %r9
+	movq	0x08(%r10), %rsi
+	movq	0x10(%r10), %rdx
+	movq	0x18(%r10), %rcx
+	movq	0x20(%r10), %r8
+	movq	0x28(%r10), %r9
+	movl	0xb0(%r10), %eax
 	testl	%eax, %eax
-	jnz	.Lload_sse
-.Lret_from_load_sse:
+	jnz	L(load_sse)
+L(ret_from_load_sse):
 
-	/* Deallocate the reg arg area.  */
-	leaq	176(%r10), %rsp
+	/* Deallocate the reg arg area, except for r10, then load via pop.  */
+	leaq	0xb8(%r10), %rsp
+	popq	%r10
 
 	/* Call the user function.  */
 	call	*%r11
@@ -80,347 +106,420 @@ ffi_call_unix64:
 	movq	0(%rbp), %rcx		/* Reload flags.  */
 	movq	8(%rbp), %rdi		/* Reload raddr.  */
 	movq	16(%rbp), %rbp		/* Reload old frame pointer.  */
-.LUW2:
+L(UW2):
+	/* 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
-	movslq	(%r11, %r10, 4), %r10
-	addq	%r11, %r10
+	leaq	L(store_table)(%rip), %r11
+	ja	L(sa)
+	leaq	(%r11, %r10, 8), %r10
+
+	/* Prep for the structure cases: scratch area in redzone.  */
+	leaq	-20(%rsp), %rsi
 	jmp	*%r10
 
-.Lstore_table:
-	.long	.Lst_void-.Lstore_table		/* FFI_TYPE_VOID */
-	.long	.Lst_sint32-.Lstore_table	/* FFI_TYPE_INT */
-	.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:
+	.balign	8
+L(store_table):
+E(L(store_table), UNIX64_RET_VOID)
 	ret
-	.align 2
-
-.Lst_uint8:
-	movzbq	%al, %rax
+E(L(store_table), UNIX64_RET_UINT8)
+	movzbl	%al, %eax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
-.Lst_sint8:
-	movsbq	%al, %rax
+E(L(store_table), UNIX64_RET_UINT16)
+	movzwl	%ax, %eax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
-.Lst_uint16:
-	movzwq	%ax, %rax
+E(L(store_table), UNIX64_RET_UINT32)
+	movl	%eax, %eax
 	movq	%rax, (%rdi)
-	.align 2
-.Lst_sint16:
-	movswq	%ax, %rax
+	ret
+E(L(store_table), UNIX64_RET_SINT8)
+	movsbq	%al, %rax
 	movq	%rax, (%rdi)
 	ret
-	.align 2
-.Lst_uint32:
-	movl	%eax, %eax
+E(L(store_table), UNIX64_RET_SINT16)
+	movswq	%ax, %rax
 	movq	%rax, (%rdi)
-	.align 2
-.Lst_sint32:
+	ret
+E(L(store_table), UNIX64_RET_SINT32)
 	cltq
 	movq	%rax, (%rdi)
 	ret
-	.align 2
-.Lst_int64:
+E(L(store_table), UNIX64_RET_INT64)
 	movq	%rax, (%rdi)
 	ret
-
-	.align 2
-.Lst_float:
-	movss	%xmm0, (%rdi)
+E(L(store_table), UNIX64_RET_XMM32)
+	movd	%xmm0, (%rdi)
 	ret
-	.align 2
-.Lst_double:
-	movsd	%xmm0, (%rdi)
+E(L(store_table), UNIX64_RET_XMM64)
+	movq	%xmm0, (%rdi)
 	ret
-.Lst_ldouble:
+E(L(store_table), UNIX64_RET_X87)
 	fstpt	(%rdi)
 	ret
-
-	.align 2
-.Lst_struct:
-	leaq	-20(%rsp), %rsi		/* Scratch area in redzone.  */
-
-	/* We have to locate the values now, and since we don't want to
-	   write too much data into the user's return value, we spill the
-	   value to a 16 byte scratch area first.  Bits 8, 9, and 10
-	   control where the values are located.  Only one of the three
-	   bits will be set; see ffi_prep_cif_machdep for the pattern.  */
-	movd	%xmm0, %r10
-	movd	%xmm1, %r11
-	testl	$0x100, %ecx
-	cmovnz	%rax, %rdx
-	cmovnz	%r10, %rax
-	testl	$0x200, %ecx
-	cmovnz	%r10, %rdx
-	testl	$0x400, %ecx
-	cmovnz	%r10, %rax
-	cmovnz	%r11, %rdx
-	movq	%rax, (%rsi)
+E(L(store_table), UNIX64_RET_X87_2)
+	fstpt	(%rdi)
+	fstpt	16(%rdi)
+	ret
+E(L(store_table), UNIX64_RET_ST_XMM0_RAX)
+	movq	%rax, 8(%rsi)
+	jmp	L(s3)
+E(L(store_table), UNIX64_RET_ST_RAX_XMM0)
+	movq	%xmm0, 8(%rsi)
+	jmp	L(s2)
+E(L(store_table), UNIX64_RET_ST_XMM0_XMM1)
+	movq	%xmm1, 8(%rsi)
+	jmp	L(s3)
+E(L(store_table), UNIX64_RET_ST_RAX_RDX)
 	movq	%rdx, 8(%rsi)
-
-	/* Bits 12-31 contain the true size of the structure.  Copy from
-	   the scratch area to the true destination.  */
-	shrl	$12, %ecx
+L(s2):
+	movq	%rax, (%rsi)
+	shrl	$UNIX64_SIZE_SHIFT, %ecx
 	rep movsb
 	ret
+	.balign 8
+L(s3):
+	movq	%xmm0, (%rsi)
+	shrl	$UNIX64_SIZE_SHIFT, %ecx
+	rep movsb
+	ret
+
+L(sa):	call	PLT(C(abort))
 
 	/* 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:
-.Lload_sse:
-	movdqa	48(%r10), %xmm0
-	movdqa	64(%r10), %xmm1
-	movdqa	80(%r10), %xmm2
-	movdqa	96(%r10), %xmm3
-	movdqa	112(%r10), %xmm4
-	movdqa	128(%r10), %xmm5
-	movdqa	144(%r10), %xmm6
-	movdqa	160(%r10), %xmm7
-	jmp	.Lret_from_load_sse
-
-.LUW4:
-	.size    ffi_call_unix64,.-ffi_call_unix64
-
-	.align	2
-	.globl ffi_closure_unix64
-	.type	ffi_closure_unix64,@function
-
-ffi_closure_unix64:
-.LUW5:
-	/* The carry flag is set by the trampoline iff SSE registers
-	   are used.  Don't clobber it before the branch instruction.  */
-	leaq    -200(%rsp), %rsp
-.LUW6:
-	movq	%rdi, (%rsp)
-	movq    %rsi, 8(%rsp)
-	movq    %rdx, 16(%rsp)
-	movq    %rcx, 24(%rsp)
-	movq    %r8, 32(%rsp)
-	movq    %r9, 40(%rsp)
-	jc      .Lsave_sse
-.Lret_from_save_sse:
-
-	movq	%r10, %rdi
-	leaq	176(%rsp), %rsi
-	movq	%rsp, %rdx
-	leaq	208(%rsp), %rcx
-	call	ffi_closure_unix64_inner@PLT
+	.balign 2
+L(UW3):
+	/* cfi_restore_state */
+L(load_sse):
+	movdqa	0x30(%r10), %xmm0
+	movdqa	0x40(%r10), %xmm1
+	movdqa	0x50(%r10), %xmm2
+	movdqa	0x60(%r10), %xmm3
+	movdqa	0x70(%r10), %xmm4
+	movdqa	0x80(%r10), %xmm5
+	movdqa	0x90(%r10), %xmm6
+	movdqa	0xa0(%r10), %xmm7
+	jmp	L(ret_from_load_sse)
+
+L(UW4):
+ENDF(C(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)
+
+	.balign	2
+	.globl	C(ffi_closure_unix64_sse)
+	FFI_HIDDEN(C(ffi_closure_unix64_sse))
+
+C(ffi_closure_unix64_sse):
+L(UW5):
+	subq	$ffi_closure_FS, %rsp
+L(UW6):
+	/* cfi_adjust_cfa_offset(ffi_closure_FS) */
+
+	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	L(sse_entry1)
+
+L(UW7):
+ENDF(C(ffi_closure_unix64_sse))
+
+	.balign	2
+	.globl	C(ffi_closure_unix64)
+	FFI_HIDDEN(C(ffi_closure_unix64))
+
+C(ffi_closure_unix64):
+L(UW8):
+	subq	$ffi_closure_FS, %rsp
+L(UW9):
+	/* cfi_adjust_cfa_offset(ffi_closure_FS) */
+L(sse_entry1):
+	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	FFI_TRAMPOLINE_SIZE(%r10), %edi		/* Load cif */
+	movl	FFI_TRAMPOLINE_SIZE+4(%r10), %esi	/* Load fun */
+	movl	FFI_TRAMPOLINE_SIZE+8(%r10), %edx	/* Load user_data */
+#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
+L(do_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	C(ffi_closure_unix64_inner)
 
 	/* Deallocate stack frame early; return value is now in redzone.  */
-	addq	$200, %rsp
-.LUW7:
+	addq	$ffi_closure_FS, %rsp
+L(UW10):
+	/* 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
-	movslq	(%r11, %r10, 4), %r10
-	addq	%r11, %r10
+	leaq	L(load_table)(%rip), %r11
+	ja	L(la)
+	leaq	(%r11, %r10, 8), %r10
+	leaq	ffi_closure_RED_RVALUE(%rsp), %rsi
 	jmp	*%r10
 
-.Lload_table:
-	.long	.Lld_void-.Lload_table		/* FFI_TYPE_VOID */
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_INT */
-	.long	.Lld_float-.Lload_table		/* FFI_TYPE_FLOAT */
-	.long	.Lld_double-.Lload_table	/* FFI_TYPE_DOUBLE */
-	.long	.Lld_ldouble-.Lload_table	/* FFI_TYPE_LONGDOUBLE */
-	.long	.Lld_int8-.Lload_table		/* FFI_TYPE_UINT8 */
-	.long	.Lld_int8-.Lload_table		/* FFI_TYPE_SINT8 */
-	.long	.Lld_int16-.Lload_table		/* FFI_TYPE_UINT16 */
-	.long	.Lld_int16-.Lload_table		/* FFI_TYPE_SINT16 */
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_UINT32 */
-	.long	.Lld_int32-.Lload_table		/* FFI_TYPE_SINT32 */
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_UINT64 */
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_SINT64 */
-	.long	.Lld_struct-.Lload_table	/* FFI_TYPE_STRUCT */
-	.long	.Lld_int64-.Lload_table		/* FFI_TYPE_POINTER */
-
-	.align 2
-.Lld_void:
+	.balign	8
+L(load_table):
+E(L(load_table), UNIX64_RET_VOID)
 	ret
-
-	.align 2
-.Lld_int8:
-	movzbl	-24(%rsp), %eax
+E(L(load_table), UNIX64_RET_UINT8)
+	movzbl	(%rsi), %eax
 	ret
-	.align 2
-.Lld_int16:
-	movzwl	-24(%rsp), %eax
+E(L(load_table), UNIX64_RET_UINT16)
+	movzwl	(%rsi), %eax
 	ret
-	.align 2
-.Lld_int32:
-	movl	-24(%rsp), %eax
+E(L(load_table), UNIX64_RET_UINT32)
+	movl	(%rsi), %eax
 	ret
-	.align 2
-.Lld_int64:
-	movq	-24(%rsp), %rax
+E(L(load_table), UNIX64_RET_SINT8)
+	movsbl	(%rsi), %eax
 	ret
-
-	.align 2
-.Lld_float:
-	movss	-24(%rsp), %xmm0
+E(L(load_table), UNIX64_RET_SINT16)
+	movswl	(%rsi), %eax
 	ret
-	.align 2
-.Lld_double:
-	movsd	-24(%rsp), %xmm0
+E(L(load_table), UNIX64_RET_SINT32)
+	movl	(%rsi), %eax
 	ret
-	.align 2
-.Lld_ldouble:
-	fldt	-24(%rsp)
+E(L(load_table), UNIX64_RET_INT64)
+	movq	(%rsi), %rax
 	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
+E(L(load_table), UNIX64_RET_XMM32)
+	movd	(%rsi), %xmm0
+	ret
+E(L(load_table), UNIX64_RET_XMM64)
+	movq	(%rsi), %xmm0
+	ret
+E(L(load_table), UNIX64_RET_X87)
+	fldt	(%rsi)
+	ret
+E(L(load_table), UNIX64_RET_X87_2)
+	fldt	16(%rsi)
+	fldt	(%rsi)
+	ret
+E(L(load_table), UNIX64_RET_ST_XMM0_RAX)
+	movq	8(%rsi), %rax
+	jmp	L(l3)
+E(L(load_table), UNIX64_RET_ST_RAX_XMM0)
+	movq	8(%rsi), %xmm0
+	jmp	L(l2)
+E(L(load_table), UNIX64_RET_ST_XMM0_XMM1)
+	movq	8(%rsi), %xmm1
+	jmp	L(l3)
+E(L(load_table), UNIX64_RET_ST_RAX_RDX)
+	movq	8(%rsi), %rdx
+L(l2):
+	movq	(%rsi), %rax
+	ret
+	.balign	8
+L(l3):
+	movq	(%rsi), %xmm0
 	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 HAVE_AS_X86_64_UNWIND_SECTION_TYPE
-	.section	.eh_frame,"a",@unwind
-#else
-	.section	.eh_frame,"a",@progbits
-#endif
-.Lframe1:
-	.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 */
+L(la):	call	PLT(C(abort))
+
+L(UW11):
+ENDF(C(ffi_closure_unix64))
+
+	.balign	2
+	.globl	C(ffi_go_closure_unix64_sse)
+	FFI_HIDDEN(C(ffi_go_closure_unix64_sse))
+
+C(ffi_go_closure_unix64_sse):
+L(UW12):
+	subq	$ffi_closure_FS, %rsp
+L(UW13):
+	/* cfi_adjust_cfa_offset(ffi_closure_FS) */
+
+	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	L(sse_entry2)
+
+L(UW14):
+ENDF(C(ffi_go_closure_unix64_sse))
+
+	.balign	2
+	.globl	C(ffi_go_closure_unix64)
+	FFI_HIDDEN(C(ffi_go_closure_unix64))
+
+C(ffi_go_closure_unix64):
+L(UW15):
+	subq	$ffi_closure_FS, %rsp
+L(UW16):
+	/* cfi_adjust_cfa_offset(ffi_closure_FS) */
+L(sse_entry2):
+	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
-	.long	.LUW0@rel
+	movq	8(%r10), %rdi		/* Load cif */
+	movq	16(%r10), %rsi		/* Load fun */
+	movq	%r10, %rdx		/* Load closure (user_data) */
 #endif
-	.long	.LUW4-.LUW0		/* FDE address range */
-	.uleb128 0x0			/* Augmentation size */
-
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW1-.LUW0
+	jmp	L(do_closure)
 
-	/* 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 */
+L(UW17):
+ENDF(C(ffi_go_closure_unix64))
 
-	.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 */
+/* Sadly, OSX cctools-as doesn't understand .cfi directives at all.  */
 
-	.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 */
+#ifdef __APPLE__
+.section __TEXT,__eh_frame,coalesced,no_toc+strip_static_syms+live_support
+EHFrame0:
+#elif defined(HAVE_AS_X86_64_UNWIND_SECTION_TYPE)
+.section .eh_frame,"a",@unwind
 #else
-	.long	.LUW5@rel
+.section .eh_frame,"a",@progbits
 #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 */
+#ifdef HAVE_AS_X86_PCREL
+# define PCREL(X)	X - .
+#else
+# define PCREL(X)	X@rel
+#endif
 
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW7-.LUW6
-	.byte	0xe			/* DW_CFA_def_cfa_offset */
-	.uleb128 8
+/* Simplify advancing between labels.  Assume DW_CFA_advance_loc1 fits.  */
+#define ADV(N, P)	.byte 2, L(N)-L(P)
 
-	.byte	0x4			/* DW_CFA_advance_loc4 */
-	.long	.LUW8-.LUW7
+	.balign 8
+L(CIE):
+	.set	L(set0),L(ECIE)-L(SCIE)
+	.long	L(set0)			/* CIE Length */
+L(SCIE):
+	.long	0			/* CIE Identifier Tag */
+	.byte	1			/* CIE Version */
+	.ascii	"zR\0"			/* CIE Augmentation */
+	.byte	1			/* CIE Code Alignment Factor */
+	.byte	0x78			/* CIE Data Alignment Factor */
+	.byte	0x10			/* CIE RA Column */
+	.byte	1			/* Augmentation size */
+	.byte	0x1b			/* FDE Encoding (pcrel sdata4) */
+	.byte	0xc, 7, 8		/* DW_CFA_def_cfa, %rsp offset 8 */
+	.byte	0x80+16, 1		/* DW_CFA_offset, %rip offset 1*-8 */
+	.balign 8
+L(ECIE):
+
+	.set	L(set1),L(EFDE1)-L(SFDE1)
+	.long	L(set1)			/* FDE Length */
+L(SFDE1):
+	.long	L(SFDE1)-L(CIE)		/* FDE CIE offset */
+	.long	PCREL(L(UW0))		/* Initial location */
+	.long	L(UW4)-L(UW0)		/* Address range */
+	.byte	0			/* Augmentation size */
+	ADV(UW1, UW0)
+	.byte	0xc, 6, 32		/* DW_CFA_def_cfa, %rbp 32 */
+	.byte	0x80+6, 2		/* DW_CFA_offset, %rbp 2*-8 */
+	ADV(UW2, UW1)
+	.byte	0xa			/* DW_CFA_remember_state */
+	.byte	0xc, 7, 8		/* DW_CFA_def_cfa, %rsp 8 */
+	.byte	0xc0+6			/* DW_CFA_restore, %rbp */
+	ADV(UW3, UW2)
 	.byte	0xb			/* DW_CFA_restore_state */
-
-	.align 8
-.LEFDE3:
+	.balign	8
+L(EFDE1):
+
+	.set	L(set2),L(EFDE2)-L(SFDE2)
+	.long	L(set2)			/* FDE Length */
+L(SFDE2):
+	.long	L(SFDE2)-L(CIE)		/* FDE CIE offset */
+	.long	PCREL(L(UW5))		/* Initial location */
+	.long	L(UW7)-L(UW5)		/* Address range */
+	.byte	0			/* Augmentation size */
+	ADV(UW6, UW5)
+	.byte	0xe			/* DW_CFA_def_cfa_offset */
+	.byte	ffi_closure_FS + 8, 1	/* uleb128, assuming 128 <= FS < 255 */
+	.balign	8
+L(EFDE2):
+
+	.set	L(set3),L(EFDE3)-L(SFDE3)
+	.long	L(set3)			/* FDE Length */
+L(SFDE3):
+	.long	L(SFDE3)-L(CIE)		/* FDE CIE offset */
+	.long	PCREL(L(UW8))		/* Initial location */
+	.long	L(UW11)-L(UW8)		/* Address range */
+	.byte	0			/* Augmentation size */
+	ADV(UW9, UW8)
+	.byte	0xe			/* DW_CFA_def_cfa_offset */
+	.byte	ffi_closure_FS + 8, 1	/* uleb128, assuming 128 <= FS < 255 */
+	ADV(UW10, UW9)
+	.byte	0xe, 8			/* DW_CFA_def_cfa_offset 8 */
+L(EFDE3):
+
+	.set	L(set4),L(EFDE4)-L(SFDE4)
+	.long	L(set4)			/* FDE Length */
+L(SFDE4):
+	.long	L(SFDE4)-L(CIE)		/* FDE CIE offset */
+	.long	PCREL(L(UW12))		/* Initial location */
+	.long	L(UW14)-L(UW12)		/* Address range */
+	.byte	0			/* Augmentation size */
+	ADV(UW13, UW12)
+	.byte	0xe			/* DW_CFA_def_cfa_offset */
+	.byte	ffi_closure_FS + 8, 1	/* uleb128, assuming 128 <= FS < 255 */
+	.balign	8
+L(EFDE4):
+
+	.set	L(set5),L(EFDE5)-L(SFDE5)
+	.long	L(set5)			/* FDE Length */
+L(SFDE5):
+	.long	L(SFDE5)-L(CIE)		/* FDE CIE offset */
+	.long	PCREL(L(UW15))		/* Initial location */
+	.long	L(UW17)-L(UW15)		/* Address range */
+	.byte	0			/* Augmentation size */
+	ADV(UW16, UW15)
+	.byte	0xe			/* DW_CFA_def_cfa_offset */
+	.byte	ffi_closure_FS + 8, 1	/* uleb128, assuming 128 <= FS < 255 */
+	.balign	8
+L(EFDE5):
+#ifdef __APPLE__
+	.subsections_via_symbols
+#endif
 
 #endif /* __x86_64__ */
-
 #if defined __ELF__ && defined __linux__
 	.section	.note.GNU-stack,"",@progbits
 #endif