Initial support for x64-mingw32

git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@65868 b2dd03c8-39d4-4d8f-98ff-823fe69b080e

configure.ac

@@ -2325,8 +2325,14 @@ AS_IF([test "${universal_binary-no}" = yes ], [
     AC_DEFINE_UNQUOTED(STACK_GROW_DIRECTION, $dir)
 ])
 
-AC_MSG_CHECKING(native fiber implementation)
+AC_MSG_CHECKING(native coroutine implementation for $target_cpu-$target_os)
 AS_CASE(["$target_cpu-$target_os"],
+  [x*64-darwin], [
+    AC_MSG_RESULT(amd64)
+    COROUTINE_H=coroutine/amd64/Context.h
+    AC_DEFINE_UNQUOTED(FIBER_USE_COROUTINE, ["$COROUTINE_H"])
+    AC_LIBOBJ([coroutine/amd64/Context])
+  ],
   [x*64-linux], [
     AC_MSG_RESULT(amd64)
     COROUTINE_H=coroutine/amd64/Context.h
@@ -2339,6 +2345,12 @@ AS_CASE(["$target_cpu-$target_os"],
     AC_DEFINE_UNQUOTED(FIBER_USE_COROUTINE, ["$COROUTINE_H"])
     AC_LIBOBJ([coroutine/x86/Context])
   ]
+  [x64-mingw32], [
+    AC_MSG_RESULT(x86)
+    COROUTINE_H=coroutine/win64/Context.h
+    AC_DEFINE_UNQUOTED(FIBER_USE_COROUTINE, ["$COROUTINE_H"])
+    AC_LIBOBJ([coroutine/win64/Context])
+  ]
   [*], [
     AC_MSG_RESULT(no)
   ]

coroutine/win32/Context.asm

@@ -9,6 +9,8 @@
 
 .code
 
+assume fs:nothing
+
 ; Using fastcall is a big win (and it's the same has how x64 works).
 ; In coroutine transfer, the arguments are passed in ecx and edx. We don't need
 ; to touch these in order to pass them to the destination coroutine.

coroutine/win32/Context.h

@@ -24,7 +24,7 @@ struct coroutine_context
 	void **stack_pointer;
 };
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef void(__fastcall * coroutine_start)(coroutine_context *from, coroutine_context *self);
 
 static inline void coroutine_initialize(
 	coroutine_context *context,
@@ -40,12 +40,12 @@ static inline void coroutine_initialize(
 		return;
 	}
 
+	*--context->stack_pointer = (void*)start;
+
 	/* Windows Thread Information Block */
 	*--context->stack_pointer = 0; /* fs:[0] */
-	*--context->stack_pointer = stack_pointer + stack_size; /* fs:[4] */
+	*--context->stack_pointer = (void*)stack_pointer; /* fs:[4] */
-	*--context->stack_pointer = (void*)stack_pointer;  /* fs:[8] */
+	*--context->stack_pointer = (void*)((char *)stack_pointer - stack_size);  /* fs:[8] */
-
-	*--context->stack_pointer = (void*)start;
 
 	context->stack_pointer -= COROUTINE_REGISTERS;
 	memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);

coroutine/win64/Context.asm

@@ -9,9 +9,8 @@
 
 coroutine_transfer proc
 	; Save the thread information block:
-	push gs:[0x00]
+	push qword ptr gs:[8]
-	push gs:[0x08]
+	push qword ptr gs:[16]
-	push gs:[0x10]
 
 	; Save caller registers:
 	push rbp
@@ -23,12 +22,34 @@ coroutine_transfer proc
 	push r14
 	push r15
 
+	movaps [rsp - 24], xmm6
+	movaps [rsp - 40], xmm7
+	movaps [rsp - 56], xmm8
+	movaps [rsp - 72], xmm9
+	movaps [rsp - 88], xmm10
+	movaps [rsp - 104], xmm11
+	movaps [rsp - 120], xmm12
+	movaps [rsp - 136], xmm13
+	movaps [rsp - 152], xmm14
+	movaps [rsp - 168], xmm15
+
 	; Save caller stack pointer:
 	mov [rcx], rsp
 
 	; Restore callee stack pointer:
 	mov rsp, [rdx]
 
+	movaps xmm15, [rsp - 168]
+	movaps xmm14, [rsp - 152]
+	movaps xmm13, [rsp - 136]
+	movaps xmm12, [rsp - 120]
+	movaps xmm11, [rsp - 104]
+	movaps xmm10, [rsp - 88]
+	movaps xmm9, [rsp - 72]
+	movaps xmm8, [rsp - 56]
+	movaps xmm7, [rsp - 40]
+	movaps xmm6, [rsp - 24]
+
 	; Restore callee stack:
 	pop r15
 	pop r14
@@ -40,9 +61,8 @@ coroutine_transfer proc
 	pop rbp
 
 	; Restore the thread information block:
-	pop gs:[0x10]
+	pop qword ptr gs:[16]
-	pop gs:[0x08]
+	pop qword ptr gs:[8]
-	pop gs:[0x00]
 
 	; Put the first argument into the return value:
 	mov rax, rcx
@@ -51,4 +71,9 @@ coroutine_transfer proc
 	ret
 coroutine_transfer endp
 
+coroutine_trampoline proc
+	; Do not remove this. This forces 16-byte alignment when entering the coroutine.
+	ret
+coroutine_trampoline endp
+
 end

coroutine/win64/Context.h

@@ -17,13 +17,16 @@ extern "C" {
 #define COROUTINE __declspec(noreturn) void
 
 const size_t COROUTINE_REGISTERS = 8;
+const size_t COROUTINE_XMM_REGISTERS = 1+10*2;
 
 struct coroutine_context
 {
 	void **stack_pointer;
 };
 
-typedef COROUTINE(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef void(* coroutine_start)(coroutine_context *from, coroutine_context *self);
+
+void coroutine_trampoline();
 
 static inline void coroutine_initialize(
 	coroutine_context *context,
@@ -31,7 +34,8 @@ static inline void coroutine_initialize(
 	void *stack_pointer,
 	size_t stack_size
 ) {
-	context->stack_pointer = (void**)stack_pointer;
+	/* Force 16-byte alignment */
+	context->stack_pointer = (void**)((uintptr_t)stack_pointer & ~0xF);
 
 	if (!start) {
 		assert(!context->stack_pointer);
@@ -39,16 +43,22 @@ static inline void coroutine_initialize(
 		return;
 	}
 
-	/* Windows Thread Information Block */
+	/* Win64 ABI requires space for arguments */
-	*--context->stack_pointer = 0; /* gs:[0x00] */
+	context->stack_pointer -= 4;
-	*--context->stack_pointer = stack_pointer + stack_size; /* gs:[0x08] */
-	*--context->stack_pointer = (void*)stack_pointer;  /* gs:[0x10] */
-
 
+	/* Return address */
+	*--context->stack_pointer = 0;
 	*--context->stack_pointer = (void*)start;
+	*--context->stack_pointer = (void*)coroutine_trampoline;
+
+	/* Windows Thread Information Block */
+	/* *--context->stack_pointer = 0; */ /* gs:[0x00] is not used */
+	*--context->stack_pointer = (void*)stack_pointer; /* gs:[0x08] */
+	*--context->stack_pointer = (void*)((char *)stack_pointer - stack_size);  /* gs:[0x10] */
 
 	context->stack_pointer -= COROUTINE_REGISTERS;
 	memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
+	memset(context->stack_pointer - COROUTINE_XMM_REGISTERS, 0, sizeof(void*) * COROUTINE_XMM_REGISTERS);
 }
 
 coroutine_context * coroutine_transfer(coroutine_context * current, coroutine_context * target);