3 files changed, 106 insertions, 41 deletions

diff --git a/coroutine/win32/Context.S b/coroutine/win32/Context.S
new file mode 100644
index 0000000000..d14bf435e8
--- /dev/null
+++ b/coroutine/win32/Context.S
@@ -0,0 +1,47 @@
+##
+##  This file is part of the "Coroutine" project and released under the MIT License.
+##
+##  Created by Lars Kanis on 06/06/2019.
+##
+
+.text
+
+# Using fastcall is a big win (and it's the same as 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.
+
+.globl @coroutine_transfer@8
+@coroutine_transfer@8:
+	# Save the thread information block:
+	pushl %fs:0
+	pushl %fs:4
+	pushl %fs:8
+
+	# Save caller registers:
+	pushl %ebp
+	pushl %ebx
+	pushl %edi
+	pushl %esi
+
+	# Save caller stack pointer:
+	movl %esp, (%ecx)
+
+	# Restore callee stack pointer:
+	movl (%edx), %esp
+
+	# Restore callee stack:
+	popl %esi
+	popl %edi
+	popl %ebx
+	popl %ebp
+
+	# Restore the thread information block:
+	popl %fs:8
+	popl %fs:4
+	popl %fs:0
+
+	# Save the first argument as the return value:
+	movl %ecx, %eax
+
+	# Jump to the address on the stack:
+	ret
diff --git a/coroutine/win32/Context.asm b/coroutine/win32/Context.asm
index 8fa13831bc..f8f431239b 100644
--- a/coroutine/win32/Context.asm
+++ b/coroutine/win32/Context.asm
@@ -2,40 +2,53 @@
 ;;  This file is part of the "Coroutine" project and released under the MIT License.
 ;;
 ;;  Created by Samuel Williams on 10/5/2018.
-;;  Copyright, 2018, by Samuel Williams. All rights reserved.
+;;  Copyright, 2018, by Samuel Williams.
 ;;
 
+.386
 .model flat
 
 .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_transfer@8 proc
-	; Save caller registers
+	; Save the thread information block:
+	push fs:[0]
+	push fs:[4]
+	push fs:[8]
+
+	; Save caller registers:
 	push ebp
 	push ebx
 	push edi
 	push esi
-	
-	; Save caller stack pointer
+
+	; Save caller stack pointer:
 	mov dword ptr [ecx], esp
-	
-	; Restore callee stack pointer
+
+	; Restore callee stack pointer:
 	mov esp, dword ptr [edx]
 
-	; Restore callee stack
+	; Restore callee stack:
 	pop esi
 	pop edi
 	pop ebx
 	pop ebp
 
-	; Save the first argument as the return value
+	; Restore the thread information block:
+	pop fs:[8]
+	pop fs:[4]
+	pop fs:[0]
+
+	; Save the first argument as the return value:
 	mov eax, dword ptr ecx
 
-	; Jump to the address on the stack
+	; Jump to the address on the stack:
 	ret
 @coroutine_transfer@8 endp
 
diff --git a/coroutine/win32/Context.h b/coroutine/win32/Context.h
index 30cb5a982d..c1c8fbcd13 100644
--- a/coroutine/win32/Context.h
+++ b/coroutine/win32/Context.h
@@ -1,61 +1,66 @@
+#ifndef COROUTINE_WIN32_CONTEXT_H
+#define COROUTINE_WIN32_CONTEXT_H 1
+
 /*
  *  This file is part of the "Coroutine" project and released under the MIT License.
  *
  *  Created by Samuel Williams on 10/5/2018.
- *  Copyright, 2018, by Samuel Williams. All rights reserved.
+ *  Copyright, 2018, by Samuel Williams.
 */
 
 #pragma once
 
 #include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
 #include <string.h>
 
-#if __cplusplus
-extern "C" {
-#endif
-
-#define COROUTINE void __fastcall
+#define COROUTINE __declspec(noreturn) void __fastcall
+#define COROUTINE_DECL void __fastcall
+#define COROUTINE_LIMITED_ADDRESS_SPACE
 
-const size_t COROUTINE_REGISTERS = 4;
+/* This doesn't include thread information block */
+enum {COROUTINE_REGISTERS = 4};
 
 struct coroutine_context
 {
-	void **stack_pointer;
+    void **stack_pointer;
+    void *argument;
 };
 
-typedef void(__fastcall * coroutine_start)(coroutine_context *from, coroutine_context *self);
+typedef void(__fastcall * coroutine_start)(struct coroutine_context *from, struct coroutine_context *self);
 
-inline void coroutine_initialize(
-	coroutine_context *context,
-	coroutine_start start,
-	void *stack_pointer,
-	size_t stack_size
+static inline void coroutine_initialize_main(struct coroutine_context * context) {
+    context->stack_pointer = NULL;
+}
+
+static inline void coroutine_initialize(
+    struct coroutine_context *context,
+    coroutine_start start,
+    void *stack,
+    size_t size
 ) {
-	context->stack_pointer = (void**)stack_pointer;
+    assert(start && stack && size >= 1024);
 
-	if (!start) {
-		assert(!context->stack_pointer);
-		/* We are main coroutine for this thread */
-		return;
-	}
+    // Stack grows down. Force 16-byte alignment.
+    char * top = (char*)stack + size;
+    context->stack_pointer = (void**)((uintptr_t)top & ~0xF);
 
-	/* Windows Thread Information Block */
-	*--context->stack_pointer = 0;
-	*--context->stack_pointer = stack_pointer;
-	*--context->stack_pointer = (void*)stack_size;
+    *--context->stack_pointer = (void*)(uintptr_t)start;
 
-	*--context->stack_pointer = (void*)start;
+    /* Windows Thread Information Block */
+    *--context->stack_pointer = (void*)0xFFFFFFFF; /* fs:[0] */
+    *--context->stack_pointer = (void*)top; /* fs:[4] */
+    *--context->stack_pointer = (void*)stack;  /* fs:[8] */
 
-	context->stack_pointer -= COROUTINE_REGISTERS;
-	memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
+    context->stack_pointer -= COROUTINE_REGISTERS;
+    memset(context->stack_pointer, 0, sizeof(void*) * COROUTINE_REGISTERS);
 }
 
-coroutine_context * __fastcall coroutine_transfer(coroutine_context * current, coroutine_context * target);
+struct coroutine_context * __fastcall coroutine_transfer(struct coroutine_context * current, struct coroutine_context * target);
 
-inline void coroutine_destroy(coroutine_context * context)
+static inline void coroutine_destroy(struct coroutine_context * context)
 {
 }
 
-#if __cplusplus
-}
-#endif
\ No newline at end of file
+#endif /* COROUTINE_WIN32_CONTEXT_H */