start going back to an i386 build

src/etc/gyp-uv

@@ -40,14 +40,17 @@ do
 
     # Comment out the gyp auto regeneration
     for os in mac unix win; do
-        sed -i ".save" \
+        sed -i "" \
             -e 's/^\(Makefile: $(srcdir)\/src\/rt\/libuv\/uv\.gyp\)/#\1/' \
             mk/libuv/$ARCH/$os/Makefile
 
-        sed -i ".save" \
+        sed -i "" \
             -e 's/^\(	$(call do_cmd,regen_makefile)\)/#\1/' \
             mk/libuv/$ARCH/$os/Makefile
-
-        rm mk/libuv/$ARCH/$os/Makefile.save
     done
 done
+
+# On Mac, GYP hardcodes a -arch i386 into the output. Fix that.
+sed -i "" \
+    -e 's/-arch i386/-arch x86_64/' \
+    mk/libuv/x86_64/mac/src/rt/libuv/*.mk

src/rt/arch/x86_64/_context.S

@@ -54,44 +54,44 @@ swap_registers:
         // jump back into the old code.
         
         // Save instruction pointer:
-        popl %rax
-        movl %rax, 104(%rdi)
+        pop %rax
+        mov %rax, 104(%rdi)
 
         // Save non-volatile integer registers:
         //   (including RSP)
-        movl %rbx,   0(%rdi)
-        movl %rsp,   8(%rdi)
-        movl %rbp,  16(%rdi)
-        movl %r12,  24(%rdi)
-        movl %r13,  32(%rdi)
-        movl %r14,  40(%rdi)
-        movl %r15,  48(%rdi)
+        mov %rbx,   0(%rdi)
+        mov %rsp,   8(%rdi)
+        mov %rbp,  16(%rdi)
+        mov %r12,  24(%rdi)
+        mov %r13,  32(%rdi)
+        mov %r14,  40(%rdi)
+        mov %r15,  48(%rdi)
 
         // Save non-volatile XMM registers:
-        movl %xmm0, 56(%rdi)
-        movl %xmm1, 64(%rdi)
-        movl %xmm2, 72(%rdi)
-        movl %xmm3, 80(%rdi)
-        movl %xmm4, 88(%rdi)
-        movl %xmm5, 96(%rdi)
+        movapd %xmm0, 56(%rdi)
+        movapd %xmm1, 64(%rdi)
+        movapd %xmm2, 72(%rdi)
+        movapd %xmm3, 80(%rdi)
+        movapd %xmm4, 88(%rdi)
+        movapd %xmm5, 96(%rdi)
 
         // Restore non-volatile integer registers:
         //   (including RSP)
-        movl  0(%rsi), %rbx
-        movl  8(%rsi), %rsp
-        movl 16(%rsi), %rbp
-        movl 24(%rsi), %r12
-        movl 32(%rsi), %r13
-        movl 40(%rsi), %r14
-        movl 48(%rsi), %r15
+        mov  0(%rsi), %rbx
+        mov  8(%rsi), %rsp
+        mov 16(%rsi), %rbp
+        mov 24(%rsi), %r12
+        mov 32(%rsi), %r13
+        mov 40(%rsi), %r14
+        mov 48(%rsi), %r15
 
         // Restore non-volatile XMM registers:
-        movl 56(%rsi), %xmm0
-        movl 64(%rsi), %xmm1
-        movl 72(%rsi), %xmm2
-        movl 80(%rsi), %xmm3
-        movl 88(%rsi), %xmm4
-        movl 96(%rsi), %xmm5
+        movapd 56(%rsi), %xmm0
+        movapd 64(%rsi), %xmm1
+        movapd 72(%rsi), %xmm2
+        movapd 80(%rsi), %xmm3
+        movapd 88(%rsi), %xmm4
+        movapd 96(%rsi), %xmm5
 
         // Jump to the instruction pointer
         // found in regs:

src/rt/arch/x86_64/ccall.S

@@ -11,11 +11,11 @@ _upcall_call_c_stack:
 .globl upcall_call_c_stack
 upcall_call_c_stack:
 #endif
-    pushl %rbp
-    movl %rsp,%rbp          // save rsp
-    movl %rsi,%rsp          // switch stack
-    calll *%rdi
-    movl %rbp,%rsp          // would like to use "leave" but it's slower
-    popl %rbp
+    push %rbp
+    mov %rsp,%rbp          // save rsp
+    mov %rsi,%rsp          // switch stack
+    call *%rdi
+    mov %rbp,%rsp          // would like to use "leave" but it's slower
+    pop %rbp
     ret
 

src/rt/arch/x86_64/context.cpp

@@ -33,6 +33,6 @@ void context::call(void *f, void *arg, void *stack) {
   *--sp = (uint64_t)arg;
   *--sp = 0xdeadbeef;
 
-  regs.regs[1] = (uint64_t)sp;
+  regs.regs[RSP] = (uint64_t)sp;
   regs.ip = (uint64_t)f;
 }

src/rt/arch/x86_64/context.h

@@ -27,6 +27,8 @@ struct registers_t {
 
 class context {
 public:
+  static const int RSP = 1;
+
   registers_t regs;
 
   context();
@@ -42,7 +44,7 @@ public:
   // function being called causes the task to fail, then we have to avoid
   // leaking space on the C stack.
   inline void *alloc_stack(size_t nbytes) {
-    uint32_t bot = regs.esp;
+    uint32_t bot = regs.regs[RSP];
     uint32_t top = align_down(bot - nbytes);
 
 #ifdef HAVE_VALGRIND