Modify configure.host to detect compilation with Microsoft Visual C++ and use assembly with Intel syntax in that case

1 files changed, 237 insertions, 0 deletions

diff --git a/src/x86/win64_intel.S b/src/x86/win64_intel.S
new file mode 100644
index 0000000..7df78b3
--- /dev/null
+++ b/src/x86/win64_intel.S
@@ -0,0 +1,237 @@
+#define LIBFFI_ASM
+#include <fficonfig.h>
+#include <ffi.h>
+#include <ffi_cfi.h>
+#include "asmnames.h"
+
+#if defined(HAVE_AS_CFI_PSEUDO_OP)
+        .cfi_sections   .debug_frame
+#endif
+
+#ifdef X86_WIN64
+#define SEH(...) __VA_ARGS__
+#define arg0	rcx
+#define arg1	rdx
+#define arg2	r8
+#define arg3	r9
+#else
+#define SEH(...)
+#define arg0	rdi
+#define arg1	rsi
+#define arg2	rdx
+#define arg3	rcx
+#endif
+
+/* 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)	ALIGN 8
+#else
+# define E(BASE, X)	ALIGN 8; ORG BASE + X * 8
+#endif
+
+	.CODE
+	extern PLT(C(abort)):near
+	extern C(ffi_closure_win64_inner):near
+
+/* ffi_call_win64 (void *stack, struct win64_call_frame *frame, void *r10)
+
+   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	8
+	PUBLIC	C(ffi_call_win64)
+
+	; SEH(.safesh ffi_call_win64)
+C(ffi_call_win64) proc SEH(frame)
+	cfi_startproc
+	/* Set up the local stack frame and install it in rbp/rsp.  */
+	mov	RAX, [RSP] ; 	movq	(%rsp), %rax
+	mov [arg1], RBP ; movq	%rbp, (arg1)
+	mov [arg1 + 8], RAX;	movq	%rax, 8(arg1)
+	mov	 RBP, arg1; movq	arg1, %rbp
+	cfi_def_cfa(rbp, 16)
+	cfi_rel_offset(rbp, 0)
+	SEH(.pushreg rbp)
+	SEH(.setframe rbp, 0)
+	SEH(.endprolog)
+	mov	RSP, arg0 ;	movq	arg0, %rsp
+
+	mov	R10, arg2 ; movq	arg2, %r10
+
+	/* Load all slots into both general and xmm registers.  */
+	mov	RCX, [RSP] ;	movq	(%rsp), %rcx
+	movsd XMM0, qword ptr [RSP] ; movsd	(%rsp), %xmm0
+	mov	RDX, [RSP + 8] ;movq	8(%rsp), %rdx
+	movsd XMM1, qword ptr [RSP + 8];	movsd	8(%rsp), %xmm1
+	mov R8, [RSP + 16] ; movq	16(%rsp), %r8
+	movsd	XMM2, qword ptr [RSP + 16] ; movsd	16(%rsp), %xmm2
+	mov	R9, [RSP + 24] ; movq	24(%rsp), %r9
+	movsd	XMM3, qword ptr [RSP + 24] ;movsd	24(%rsp), %xmm3
+
+	CALL qword ptr [RBP + 16] ; call	*16(%rbp)
+
+	mov	 ECX, [RBP + 24] ; movl	24(%rbp), %ecx
+	mov	R8, [RBP + 32] ; movq	32(%rbp), %r8
+	LEA	R10, ffi_call_win64_tab ; leaq	0f(%rip), %r10
+	CMP	ECX, FFI_TYPE_SMALL_STRUCT_4B ; cmpl	$FFI_TYPE_SMALL_STRUCT_4B, %ecx
+	LEA	R10, [R10 + RCX*8] ; leaq	(%r10, %rcx, 8), %r10
+	JA	L99 ; ja	99f
+	JMP	R10 ; jmp	*%r10
+
+/* Below, we're space constrained most of the time.  Thus we eschew the
+   modern "mov, pop, ret" sequence (5 bytes) for "leave, ret" (2 bytes).  */
+epilogue macro
+	LEAVE
+	cfi_remember_state
+	cfi_def_cfa(rsp, 8)
+	cfi_restore(rbp)
+	RET
+	cfi_restore_state
+endm
+
+	ALIGN 8
+ffi_call_win64_tab LABEL NEAR
+E(0b, FFI_TYPE_VOID)
+	epilogue
+E(0b, FFI_TYPE_INT)
+	movsxd rax, eax ; movslq	%eax, %rax
+	mov qword ptr [r8], rax; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_FLOAT)
+	movss dword ptr [r8], xmm0 ; movss	%xmm0, (%r8)
+	epilogue
+E(0b, FFI_TYPE_DOUBLE)
+	movsd qword ptr[r8], xmm0; movsd	%xmm0, (%r8)
+	epilogue
+E(0b, FFI_TYPE_LONGDOUBLE)
+	call	PLT(C(abort))
+E(0b, FFI_TYPE_UINT8)
+	movzx eax, al ;movzbl	%al, %eax
+	mov qword ptr[r8], rax; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SINT8)
+	movsx rax, al ; movsbq	%al, %rax
+	jmp	L98
+E(0b, FFI_TYPE_UINT16)
+	movzx eax, ax ; movzwl	%ax, %eax
+	mov qword ptr[r8], rax; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SINT16)
+	movsx rax, ax; movswq	%ax, %rax
+	jmp	L98
+E(0b, FFI_TYPE_UINT32)
+	mov eax, eax; movl	%eax, %eax
+	mov qword ptr[r8], rax ; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SINT32)
+	movsxd rax, eax; movslq	%eax, %rax
+	mov qword ptr [r8], rax; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_UINT64)
+L98 LABEL near
+	mov qword ptr [r8], rax ; movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SINT64)
+	mov qword ptr [r8], rax;movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_STRUCT)
+	epilogue
+E(0b, FFI_TYPE_POINTER)
+	mov qword ptr [r8], rax ;movq	%rax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_COMPLEX)
+	call	PLT(C(abort))
+E(0b, FFI_TYPE_SMALL_STRUCT_1B)
+	mov byte ptr [r8], al ; movb	%al, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SMALL_STRUCT_2B)
+	mov word ptr [r8], ax ; movw	%ax, (%r8)
+	epilogue
+E(0b, FFI_TYPE_SMALL_STRUCT_4B)
+	mov dword ptr [r8], eax ; movl	%eax, (%r8)
+	epilogue
+
+	align	8
+L99 LABEL near
+	call	PLT(C(abort))
+
+	epilogue
+
+	cfi_endproc
+	C(ffi_call_win64) endp
+
+
+/* 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
+	PUBLIC	C(ffi_go_closure_win64)
+
+C(ffi_go_closure_win64) proc
+	cfi_startproc
+	/* Save all integer arguments into the incoming reg stack space.  */
+	mov qword ptr [rsp + 8], rcx; movq	%rcx, 8(%rsp)
+	mov qword ptr [rsp + 16], rdx; movq	%rdx, 16(%rsp)
+	mov qword ptr [rsp + 24], r8; movq	%r8, 24(%rsp)
+	mov qword ptr [rsp + 32], r9 ;movq	%r9, 32(%rsp)
+
+	mov rcx, qword ptr [r10 + 8]; movq	8(%r10), %rcx			/* load cif */
+	mov rdx, qword ptr [r10 + 16];  movq	16(%r10), %rdx			/* load fun */
+	mov r8, r10 ; movq	%r10, %r8			/* closure is user_data */
+	jmp	ffi_closure_win64_2
+	cfi_endproc
+	C(ffi_go_closure_win64) endp
+
+	align	8
+	
+PUBLIC C(ffi_closure_win64)
+C(ffi_closure_win64) PROC FRAME
+	cfi_startproc
+	/* Save all integer arguments into the incoming reg stack space.  */
+	mov qword ptr [rsp + 8], rcx; movq	%rcx, 8(%rsp)
+	mov qword ptr [rsp + 16], rdx;	movq	%rdx, 16(%rsp)
+	mov qword ptr [rsp + 24], r8; 	movq	%r8, 24(%rsp)
+	mov qword ptr [rsp + 32], r9;	movq	%r9, 32(%rsp)
+
+	mov rcx, qword ptr [FFI_TRAMPOLINE_SIZE + r10]	;movq	FFI_TRAMPOLINE_SIZE(%r10), %rcx		/* load cif */
+	mov rdx, qword ptr [FFI_TRAMPOLINE_SIZE + 8 + r10] ;	movq	FFI_TRAMPOLINE_SIZE+8(%r10), %rdx	/* load fun */
+	mov r8, qword ptr [FFI_TRAMPOLINE_SIZE+16+r10] ;movq	FFI_TRAMPOLINE_SIZE+16(%r10), %r8	/* load user_data */
+ffi_closure_win64_2 LABEL near
+	sub rsp, ffi_clo_FS ;subq	$ffi_clo_FS, %rsp
+	cfi_adjust_cfa_offset(ffi_clo_FS)
+	SEH(.allocstack ffi_clo_FS)
+	SEH(.endprolog)
+
+	/* Save all sse arguments into the stack frame.  */
+	movsd qword ptr [ffi_clo_OFF_X + rsp], xmm0	; movsd	%xmm0, ffi_clo_OFF_X(%rsp)
+	movsd qword ptr [ffi_clo_OFF_X+8+rsp], xmm1 ; movsd	%xmm1, ffi_clo_OFF_X+8(%rsp)
+	movsd qword ptr [ffi_clo_OFF_X+16+rsp], xmm2 ; movsd %xmm2, ffi_clo_OFF_X+16(%rsp)
+	movsd qword ptr [ffi_clo_OFF_X+24+rsp], xmm3 ; movsd %xmm3, ffi_clo_OFF_X+24(%rsp)
+
+	lea	r9, [ffi_clo_OFF_R + rsp] ; leaq	ffi_clo_OFF_R(%rsp), %r9
+	call C(ffi_closure_win64_inner)
+
+	/* Load the result into both possible result registers.  */
+	
+	mov rax, qword ptr [ffi_clo_OFF_R + rsp] ;movq    ffi_clo_OFF_R(%rsp), %rax
+	movsd xmm0, qword ptr [rsp + ffi_clo_OFF_R] ;movsd   ffi_clo_OFF_R(%rsp), %xmm0
+
+	add rsp, ffi_clo_FS ;addq	$ffi_clo_FS, %rsp
+	cfi_adjust_cfa_offset(-ffi_clo_FS)
+	ret
+
+	cfi_endproc
+	C(ffi_closure_win64) endp
+
+#if defined __ELF__ && defined __linux__
+	.section	.note.GNU-stack,"",@progbits
+#endif
+_text ends
+end
\ No newline at end of file