Merge pull request #104 from joshtriplett/efi64

Support the Windows/EFI calling convention on all x86-64 targets

configure.host

@@ -246,7 +246,7 @@ case "${TARGET}" in
 	SOURCES="ffi.c sysv.S"
 	;;
   X86_64)
-	SOURCES="ffi64.c unix64.S"
+	SOURCES="ffi64.c unix64.S ffiw64.c win64.S"
 	;;
   X86_WIN64)
 	SOURCES="ffiw64.c win64.S"

src/x86/asmnames.h

@@ -0,0 +1,30 @@
+#ifndef ASMNAMES_H
+#define ASMNAMES_H
+
+#define C2(X, Y)  X ## Y
+#define C1(X, Y)  C2(X, Y)
+#ifdef __USER_LABEL_PREFIX__
+# define C(X)     C1(__USER_LABEL_PREFIX__, X)
+#else
+# define C(X)     X
+#endif
+
+#ifdef __APPLE__
+# define L(X)     C1(L, X)
+#else
+# define L(X)     C1(.L, X)
+#endif
+
+#if defined(__ELF__) && defined(__PIC__)
+# define PLT(X)	  X@PLT
+#else
+# define PLT(X)	  X
+#endif
+
+#ifdef __ELF__
+# define ENDF(X)  .type	X,@function; .size X, . - X
+#else
+# define ENDF(X)
+#endif
+
+#endif /* ASMNAMES_H */

src/x86/ffi64.c

@@ -388,6 +388,9 @@ examine_argument (ffi_type *type, enum x86_64_reg_class classes[MAX_CLASSES],
 
 /* Perform machine dependent cif processing.  */
 
+extern ffi_status
+ffi_prep_cif_machdep_efi64(ffi_cif *cif);
+
 ffi_status
 ffi_prep_cif_machdep (ffi_cif *cif)
 {
@@ -396,6 +399,8 @@ ffi_prep_cif_machdep (ffi_cif *cif)
   size_t bytes, n, rtype_size;
   ffi_type *rtype;
 
+  if (cif->abi == FFI_EFI64)
+    return ffi_prep_cif_machdep_efi64(cif);
   if (cif->abi != FFI_UNIX64)
     return FFI_BAD_ABI;
 
@@ -657,22 +662,41 @@ ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
 		   flags, rvalue, fn);
 }
 
+extern void
+ffi_call_efi64(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue);
+
 void
 ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 {
+  if (cif->abi == FFI_EFI64)
+    return ffi_call_efi64(cif, fn, rvalue, avalue);
   ffi_call_int (cif, fn, rvalue, avalue, NULL);
 }
 
+extern void
+ffi_call_go_efi64(ffi_cif *cif, void (*fn)(void), void *rvalue,
+		  void **avalue, void *closure);
+
 void
 ffi_call_go (ffi_cif *cif, void (*fn)(void), void *rvalue,
 	     void **avalue, void *closure)
 {
+  if (cif->abi == FFI_EFI64)
+    ffi_call_go_efi64(cif, fn, rvalue, avalue, 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;
 
+extern ffi_status
+ffi_prep_closure_loc_efi64(ffi_closure* closure,
+			   ffi_cif* cif,
+			   void (*fun)(ffi_cif*, void*, void**, void*),
+			   void *user_data,
+			   void *codeloc);
+
 ffi_status
 ffi_prep_closure_loc (ffi_closure* closure,
 		      ffi_cif* cif,
@@ -691,6 +715,8 @@ ffi_prep_closure_loc (ffi_closure* closure,
   void (*dest)(void);
   char *tramp = closure->tramp;
 
+  if (cif->abi == FFI_EFI64)
+    return ffi_prep_closure_loc_efi64(closure, cif, fun, user_data, codeloc);
   if (cif->abi != FFI_UNIX64)
     return FFI_BAD_ABI;
 
@@ -805,10 +831,16 @@ ffi_closure_unix64_inner(ffi_cif *cif,
 extern void ffi_go_closure_unix64(void) FFI_HIDDEN;
 extern void ffi_go_closure_unix64_sse(void) FFI_HIDDEN;
 
+extern ffi_status
+ffi_prep_go_closure_efi64(ffi_go_closure* closure, ffi_cif* cif,
+			  void (*fun)(ffi_cif*, void*, void**, void*));
+
 ffi_status
 ffi_prep_go_closure (ffi_go_closure* closure, ffi_cif* cif,
 		     void (*fun)(ffi_cif*, void*, void**, void*))
 {
+  if (cif->abi == FFI_EFI64)
+    return ffi_prep_go_closure_efi64(closure, cif, fun);
   if (cif->abi != FFI_UNIX64)
     return FFI_BAD_ABI;
 

src/x86/ffitarget.h

@@ -87,6 +87,8 @@ typedef enum ffi_abi {
 #elif defined(X86_64) || (defined (__x86_64__) && defined (X86_DARWIN))
   FFI_FIRST_ABI = 1,
   FFI_UNIX64,
+  FFI_WIN64,
+  FFI_EFI64 = FFI_WIN64,
   FFI_LAST_ABI,
   FFI_DEFAULT_ABI = FFI_UNIX64
 

src/x86/ffiw64.c

@@ -30,6 +30,10 @@
 #include <stdint.h>
 
 #ifdef X86_WIN64
+#define EFI64(name) name
+#else
+#define EFI64(name) name##_efi64
+#endif
 
 struct win64_call_frame
 {
@@ -44,7 +48,7 @@ extern void ffi_call_win64 (void *stack, struct win64_call_frame *,
 			    void *closure) FFI_HIDDEN;
 
 ffi_status
-ffi_prep_cif_machdep (ffi_cif *cif)
+EFI64(ffi_prep_cif_machdep)(ffi_cif *cif)
 {
   int flags, n;
 
@@ -159,13 +163,13 @@ ffi_call_int (ffi_cif *cif, void (*fn)(void), void *rvalue,
 }
 
 void
-ffi_call (ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
+EFI64(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,
+EFI64(ffi_call_go)(ffi_cif *cif, void (*fn)(void), void *rvalue,
 	     void **avalue, void *closure)
 {
   ffi_call_int (cif, fn, rvalue, avalue, closure);
@@ -176,7 +180,7 @@ 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,
+EFI64(ffi_prep_closure_loc)(ffi_closure* closure,
 		      ffi_cif* cif,
 		      void (*fun)(ffi_cif*, void*, void**, void*),
 		      void *user_data,
@@ -190,7 +194,7 @@ ffi_prep_closure_loc (ffi_closure* closure,
     /* nopl  (%rax) */
     0x0f, 0x1f, 0x00
   };
-  unsigned char *tramp = closure->tramp;
+  char *tramp = closure->tramp;
 
   if (cif->abi != FFI_WIN64)
     return FFI_BAD_ABI;
@@ -206,7 +210,7 @@ ffi_prep_closure_loc (ffi_closure* closure,
 }
 
 ffi_status
-ffi_prep_go_closure (ffi_go_closure* closure, ffi_cif* cif,
+EFI64(ffi_prep_go_closure)(ffi_go_closure* closure, ffi_cif* cif,
 		     void (*fun)(ffi_cif*, void*, void**, void*))
 {
   if (cif->abi != FFI_WIN64)
@@ -277,5 +281,3 @@ ffi_closure_win64_inner(ffi_cif *cif,
   fun (cif, rvalue, avalue, user_data);
   return flags;
 }
-
-#endif /* X86_WIN64 */

src/x86/unix64.S

@@ -31,31 +31,10 @@
 #include <fficonfig.h>
 #include <ffi.h>
 #include "internal64.h"
+#include "asmnames.h"
 
 	.text
 
-#define C2(X, Y)  X ## Y
-#define C1(X, Y)  C2(X, Y)
-#ifdef __USER_LABEL_PREFIX__
-# define C(X)     C1(__USER_LABEL_PREFIX__, X)
-#else
-# define C(X)     X
-#endif
-
-#ifdef __APPLE__
-# define L(X)     C1(L, X)
-#else
-# define L(X)     C1(.L, X)
-#endif
-
-#ifdef __ELF__
-# define PLT(X)	  X@PLT
-# define ENDF(X)  .type	X,@function; .size X, . - X
-#else
-# define PLT(X)	  X
-# define ENDF(X)
-#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.  */

src/x86/win64.S

@@ -2,20 +2,24 @@
 #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
-
-#ifdef SYMBOL_UNDERSCORE
-#define SYMBOL_NAME(name) _##name
 #else
-#define SYMBOL_NAME(name) name
+#define SEH(...)
+#define arg0	%rdi
+#define arg1	%rsi
+#define arg2	%rdx
+#define arg3	%rcx
 #endif
 
 .macro E which
@@ -34,7 +38,7 @@
 	.align	8
 	.globl	ffi_call_win64
 
-	.seh_proc ffi_call_win64
+	SEH(.seh_proc ffi_call_win64)
 ffi_call_win64:
 	cfi_startproc
 	/* Set up the local stack frame and install it in rbp/rsp.  */
@@ -44,9 +48,9 @@ ffi_call_win64:
 	movq	arg1, %rbp
 	cfi_def_cfa(%rbp, 16)
 	cfi_rel_offset(%rbp, 0)
-	.seh_pushreg %rbp
+	SEH(.seh_pushreg %rbp)
-	.seh_setframe %rbp, 0
+	SEH(.seh_setframe %rbp, 0)
-	.seh_endprologue
+	SEH(.seh_endprologue)
 	movq	arg0, %rsp
 
 	movq	arg2, %r10
@@ -97,7 +101,7 @@ E FFI_TYPE_DOUBLE
 	movsd	%xmm0, (%r8)
 	epilogue
 E FFI_TYPE_LONGDOUBLE
-	call	abort
+	call	PLT(C(abort))
 E FFI_TYPE_UINT8
 	movzbl	%al, %eax
 	movq	%rax, (%r8)
@@ -132,7 +136,7 @@ E FFI_TYPE_POINTER
 	movq	%rax, (%r8)
 	epilogue
 E FFI_TYPE_COMPLEX
-	call	abort
+	call	PLT(C(abort))
 E FFI_TYPE_SMALL_STRUCT_1B
 	movb	%al, (%r8)
 	epilogue
@@ -144,12 +148,12 @@ E FFI_TYPE_SMALL_STRUCT_4B
 	epilogue
 
 	.align	8
-99:	call	abort
+99:	call	PLT(C(abort))
 
 .purgem epilogue
 
 	cfi_endproc
-	.seh_endproc
+	SEH(.seh_endproc)
 
 
 /* 32 bytes of outgoing register stack space, 8 bytes of alignment,
@@ -161,33 +165,33 @@ E FFI_TYPE_SMALL_STRUCT_4B
 	.align	8
 	.globl	ffi_go_closure_win64
 
-	.seh_proc ffi_go_closure_win64
+	SEH(.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	%rcx, 8(%rsp)
-	movq	arg1, 16(%rsp)
+	movq	%rdx, 16(%rsp)
-	movq	arg2, 24(%rsp)
+	movq	%r8, 24(%rsp)
-	movq	arg3, 32(%rsp)
+	movq	%r9, 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
+	SEH(.seh_endproc)
 
 	.align	8
 	.globl	ffi_closure_win64
 
-	.seh_proc ffi_closure_win64
+	SEH(.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	%rcx, 8(%rsp)
-	movq	arg1, 16(%rsp)
+	movq	%rdx, 16(%rsp)
-	movq	arg2, 24(%rsp)
+	movq	%r8, 24(%rsp)
-	movq	arg3, 32(%rsp)
+	movq	%r9, 32(%rsp)
 
 	movq	FFI_TRAMPOLINE_SIZE(%r10), arg0		/* load cif */
 	movq	FFI_TRAMPOLINE_SIZE+8(%r10), arg1	/* load fun */
@@ -195,8 +199,8 @@ ffi_closure_win64:
 0:
 	subq	$ffi_clo_FS, %rsp
 	cfi_adjust_cfa_offset(ffi_clo_FS)
-	.seh_stackalloc ffi_clo_FS
+	SEH(.seh_stackalloc ffi_clo_FS)
-	.seh_endprologue
+	SEH(.seh_endprologue)
 
 	/* Save all sse arguments into the stack frame.  */
 	movsd	%xmm0, ffi_clo_OFF_X(%rsp)
@@ -216,4 +220,4 @@ ffi_closure_win64:
 	ret
 
 	cfi_endproc
-	.seh_endproc
+	SEH(.seh_endproc)