hack around the problem that x86_64 expects first few args in regs.

call on c-stack expects all data to be delivered on the stack.

src/rt/arch/x86_64/ccall.S

@@ -1,4 +1,9 @@
-    .text
+#include "regs.h"
+
+#define ARG0 RUSTRT_ARG0_S
+#define ARG1 RUSTRT_ARG1_S
+        
+        .text
 
 // upcall_call_c_stack(void (*fn)(), void *new_esp)
 //
@@ -21,8 +26,32 @@ upcall_call_c_stack_float:
 #endif
     push %rbp
     mov %rsp,%rbp          // save rsp
-    mov %rsi,%rsp          // switch stack
-    call *%rdi
+    mov ARG1,%rsp          // switch stack
+
+    // Hack: the arguments to the function are sitting
+    // on the stack right now, as in i386 calling
+    // convention.  We need them in registers.
+    // For now, we just load them into registers.
+    //
+    // This is a total hack because it does not consider
+    // the actual arguments of the target function.
+    // It fails if there are non-INTEGER class arguments,
+    // which would get pushed on the stack, or if there are
+    // additional arguments beyond those that will get
+    // passed in registers.
+    mov ARG0,%r11     // Remember target address
+    mov  0(%rsp),RUSTRT_ARG0_S
+    mov  8(%rsp),RUSTRT_ARG1_S
+    mov 16(%rsp),RUSTRT_ARG2_S
+    mov 24(%rsp),RUSTRT_ARG3_S
+#   ifdef RUSTRT_ARG4_S        
+    mov 32(%rsp),RUSTRT_ARG4_S
+#   endif
+#   ifdef RUSTRT_ARG5_S        
+    mov 40(%rsp),RUSTRT_ARG5_S
+#   endif
+        
+    call *%r11
     mov %rbp,%rsp          // would like to use "leave" but it's slower
     pop %rbp
     ret

src/rt/arch/x86_64/regs.h

@@ -21,9 +21,15 @@
 #if defined(__MINGW32__) || defined(_WINDOWS)
 #   define RUSTRT_ARG0_S %rcx
 #   define RUSTRT_ARG1_S %rdx
+#   define RUSTRT_ARG2_S %r8
+#   define RUSTRT_ARG3_S %r9
 #else
 #   define RUSTRT_ARG0_S %rdi
 #   define RUSTRT_ARG1_S %rsi
+#   define RUSTRT_ARG2_S %rdx
+#   define RUSTRT_ARG3_S %rcx
+#   define RUSTRT_ARG4_S %r8
+#   define RUSTRT_ARG5_S %r9
 #endif