From 98dee91f2568522bc129f8cffcaa2879b43646c3 Mon Sep 17 00:00:00 2001
From: Eric Holk <eholk@mozilla.com>
Date: Thu, 23 Jun 2011 18:49:29 -0700
Subject: [PATCH] Wrote swap_context in assembly. Closes #548. This also allows
 us to re-enable optimization on OSX for both gcc and clang. Also, re-enamed
 task-comm-1.rs, which was previously broken by optimization.

---
 mk/platform.mk                   |  4 ++--
 src/rt/arch/i386/_context.s      | 36 +++++++++++++++++++++++++-------
 src/rt/arch/i386/context.cpp     | 27 +++++++-----------------
 src/rt/arch/i386/context.h       |  2 --
 src/test/run-pass/task-comm-1.rs | 11 ++++++----
 5 files changed, 45 insertions(+), 35 deletions(-)

diff --git a/mk/platform.mk b/mk/platform.mk
index fb50285bbd2..e56b45ee6fc 100644
--- a/mk/platform.mk
+++ b/mk/platform.mk
@@ -48,8 +48,8 @@ ifeq ($(CFG_OSTYPE), Darwin)
   # approaches welcome!
   #
   # NB: Currently GCC's optimizer breaks rustrt (task-comm-1 hangs) on Darwin.
-  CFG_GCC_CFLAGS += -m32 -O0
-  CFG_CLANG_CFLAGS += -m32 -O0
+  CFG_GCC_CFLAGS += -m32 -O2
+  CFG_CLANG_CFLAGS += -m32 -O2
   ifeq ($(CFG_CPUTYPE), x86_64)
     CFG_GCCISH_CFLAGS += -arch i386
     CFG_GCCISH_LINK_FLAGS += -arch i386
diff --git a/src/rt/arch/i386/_context.s b/src/rt/arch/i386/_context.s
index 8b49d35de1b..2b45aef3c4c 100644
--- a/src/rt/arch/i386/_context.s
+++ b/src/rt/arch/i386/_context.s
@@ -25,7 +25,6 @@ get_registers:
 	movl %ebp, 16(%eax)
 	movl %esi, 20(%eax)
 	movl %edi, 24(%eax)
-	movl %esp, 28(%eax)
 	movw %cs, 32(%eax)
 	movw %ds, 34(%eax)
 	movw %ss, 36(%eax)
@@ -39,12 +38,14 @@ get_registers:
 	movl %ecx, 44(%eax)
 
 	// save the return address as the instruction pointer
-	movl 0(%esp), %ecx
+    // and save the stack pointer of the caller
+    popl %ecx
+    movl %esp, 28(%eax)
 	movl %ecx, 48(%eax)
 
 	// return 0
 	xor %eax, %eax
-	ret
+	jmp *%ecx
 
 .globl set_registers
 set_registers:
@@ -70,13 +71,9 @@ set_registers:
 	push %ecx
 	popf
 
-	// get ready to return back to the old eip
-	// We could write this directly to 0(%esp), but Valgrind on OS X
-	// complains.
-	pop %ecx
+    // get ready to return.
 	mov 48(%eax), %ecx
 	push %ecx	
-	//movl %ecx, 0(%esp)
 	
 	// okay, now we can restore ecx.
 	movl 8(%eax), %ecx
@@ -84,3 +81,26 @@ set_registers:
 	// return 1 to the saved eip
 	movl $1, %eax
 	ret
+
+// swap_registers(registers_t *oregs, registers_t *regs)
+.globl swap_registers
+swap_registers:
+    // %eax = get_registers(oregs);
+    movl 4(%esp), %eax
+    push %eax
+    call get_registers
+        
+    // if(!%eax) goto call_set
+    test %eax, %eax
+    jz call_set
+
+    // else
+    addl $4, %esp
+    ret
+        
+call_set:
+    // set_registers(regs)
+    movl 12(%esp), %eax
+    movl %eax, 0(%esp)
+    call set_registers
+    // set_registers never returns
\ No newline at end of file
diff --git a/src/rt/arch/i386/context.cpp b/src/rt/arch/i386/context.cpp
index c1c8ba2167b..0d33376935c 100644
--- a/src/rt/arch/i386/context.cpp
+++ b/src/rt/arch/i386/context.cpp
@@ -5,32 +5,22 @@
 #include <stdio.h>
 #include <stdlib.h>
 
-extern "C" uint32_t CDECL get_registers(registers_t *regs) 
-  asm ("get_registers");
-extern "C" uint32_t CDECL set_registers(registers_t *regs)
-  asm ("set_registers");
+//extern "C" uint32_t CDECL get_registers(registers_t *regs) 
+//  asm ("get_registers");
+extern "C" uint32_t CDECL swap_registers(registers_t *oregs, 
+                                         registers_t *regs)
+  asm ("swap_registers");
 
 context::context()
   : next(NULL)
 {
-  get_registers(&regs);
-}
-
-void context::set()
-{
-  //printf("Activating %p...\n", this);
-  set_registers(&regs);
+  //get_registers(&regs);
+  swap_registers(&regs, &regs);
 }
 
 void context::swap(context &out)
 {
-  //printf("Swapping to %p and saving in %p\n", this, &out);
-  uint32_t r = get_registers(&out.regs);
-  //printf("get_registers = %d, sp = 0x%x\n", r, out.regs.esp);
-  if(!r) {
-    set();
-  }
-  //printf("Resumed %p...\n", &out);
+  swap_registers(&out.regs, &regs);
 }
 
 void context::call(void *f, void *arg, void *stack) {
@@ -43,7 +33,6 @@ void context::call(void *f, void *arg, void *stack) {
   *--sp = (uint32_t)this;
   *--sp = (uint32_t)arg;
   *--sp = 0xdeadbeef;
-  *--sp = 0xca11ab1e;
 
   regs.esp = (uint32_t)sp;
   regs.eip = (uint32_t)f;
diff --git a/src/rt/arch/i386/context.h b/src/rt/arch/i386/context.h
index 5329016450a..cf14ffe0753 100644
--- a/src/rt/arch/i386/context.h
+++ b/src/rt/arch/i386/context.h
@@ -25,8 +25,6 @@ public:
   
   context *next;
 
-  void set();
-  
   void swap(context &out);
 
   void call(void *f, void *arg, void *sp);
diff --git a/src/test/run-pass/task-comm-1.rs b/src/test/run-pass/task-comm-1.rs
index 8adab6e9f0f..0c321022fa9 100644
--- a/src/test/run-pass/task-comm-1.rs
+++ b/src/test/run-pass/task-comm-1.rs
@@ -1,6 +1,9 @@
 // xfail-stage0
-// xfail-stage1
-// xfail-stage2
+
+use std;
+
+import std::task::join;
+
 fn main() -> () {
    test00(); 
 }
@@ -10,7 +13,7 @@ fn start() {
 }
 
 fn test00() {
-    let task t = spawn thread start();
-    join t;
+    let task t = spawn start();
+    join(t);
     log "Completing.";
 }
\ No newline at end of file