rt: Make 64-bit __morestack grow and shrink properly

src/rt/arch/x86_64/morestack.S

@@ -8,11 +8,15 @@
 #if defined(__APPLE__) || defined(_WIN32)
 #define RUST_NEW_STACK2     _rust_new_stack2
 #define RUST_DEL_STACK      _rust_del_stack
+#define RUST_DEL_STACK      _rust_del_stack
+#define RUST_GET_PREV_STACK _rust_get_prev_stack
 #define UPCALL_CALL_C       _upcall_call_shim_on_c_stack
 #define MORESTACK           ___morestack
 #else
 #define RUST_NEW_STACK2     rust_new_stack2
 #define RUST_DEL_STACK      rust_del_stack
+#define RUST_DEL_STACK      rust_del_stack
+#define RUST_GET_PREV_STACK rust_get_prev_stack
 #define UPCALL_CALL_C       upcall_call_shim_on_c_stack
 #define MORESTACK           __morestack
 #endif
@@ -59,6 +63,11 @@ MORESTACK:
 	movq %rsp, %rbp
 	.cfi_def_cfa_register %rbp
 
+	// Alignment
+	pushq $0
+
+	// FIXME: libgcc also saves rax. not sure if we need to
+
 	// Save argument registers
 	pushq	%rdi
 	pushq	%rsi
@@ -77,15 +86,46 @@ MORESTACK:
 	pushq %rcx // Address of stack arguments
 	pushq %r10 // The amount of stack needed
 
-	lea RUST_NEW_STACK2@PLT(%rip), %rsi
-	mov %rsp, %rdi
+	leaq RUST_NEW_STACK2@PLT(%rip), %rsi
+	movq %rsp, %rdi
 	call UPCALL_CALL_C@PLT
-	
-        mov (%rsp),%rdx        // Grab the return pointer.
-        inc %rdx               // Skip past the `ret`.
-        mov %rax,%rsp          // Switch to the new stack.
-        call *%rdx             // Enter the new function.
 
+	// Pop the new_stack_args struct
+	addq $32, %rsp
+
+	// Pop the saved arguments
+	popq %r9
+	popq %r8
+	popq %rcx
+	popq %rdx
+	popq %rsi
+	popq %rdi
+	
+        movq 8(%rbp),%r10       // Grab the return pointer.
+        incq %r10               // Skip past the `ret` in our parent frame
+        movq %rax,%rsp          // Switch to the new stack.
+
+        call *%r10              // Reenter the caller function
+
+	leaq RUST_GET_PREV_STACK@PLT(%rip), %rsi
+	movq $0, %rdi
+	call UPCALL_CALL_C@PLT
+
+	// Switch back to the rust stack, positioned
+	// where we pushed %ebp
+	movq %rax, %rsp
+
+	// Align the stack again
+	pushq $0
+	
+	leaq RUST_DEL_STACK@PLT(%rip), %rsi
+	movq $0, %rdi
+	call UPCALL_CALL_C@PLT
+
+	addq $8, %rsp
+	popq %rbp
+	ret
+	
 	.cfi_endproc
 #else
 MORESTACK:

src/rt/arch/x86_64/record_sp.S

@@ -11,6 +11,7 @@
 #if defined(__linux__)
 RECORD_SP:
 	movq %rdi, %fs:112
+	ret
 #else
 RECORD_SP:
 	ret

src/rt/rust_internal.h

@@ -95,7 +95,13 @@ static size_t const BUF_BYTES = 2048;
 #define PROC_FAIL_CODE 101;
 
 // FIXME: We want this to be 128 but need to slim the red zone calls down
-#define RED_ZONE_SIZE   256
+#ifdef __i386__
+#define RED_ZONE_SIZE 256
+#endif
+
+#ifdef __x86_64__
+#define RED_ZONE_SIZE 2048
+#endif
 
 // Every reference counted object should use this macro and initialize
 // ref_count.