get pure wrappers approach running

src/rt/arch/i386/ccall.S

@@ -45,3 +45,20 @@ upcall_call_c_stack_shim:
     movl %ebp,%esp          // would like to use "leave" but it's slower
     popl %ebp
     ret
+
+#if defined(__APPLE__) || defined(_WIN32)
+.globl _asm_call_on_stack
+_asm_call_on_stack:
+#else
+.globl asm_call_on_stack
+asm_call_on_stack:
+#endif
+    pushl %ebp
+    movl %esp,%ebp          // save esp
+    movl 16(%ebp),%esp      // load new esp
+    subl $12,%esp           // maintain 16-byte alignment
+    pushl 8(%ebp)           // push ptr to argument block
+    calll *12(%ebp)
+    movl %ebp,%esp          // would like to use "leave" but it's slower
+    popl %ebp
+    ret

src/rt/arch/i386/context.h

@@ -29,6 +29,8 @@ struct registers_t {
   uint32_t eip;
 };
 
+extern "C" void asm_call_on_stack(void *args, void *fn_ptr, uintptr_t stack_ptr);
+
 class context {
 public:
   registers_t regs;
@@ -55,6 +57,10 @@ public:
 
     return reinterpret_cast<void *>(top);
   }
+
+  void call_shim_on_c_stack(void *args, void *fn_ptr) {
+      asm_call_on_stack(args, fn_ptr, regs.esp);
+  }
 };
 
 #endif

src/rt/arch/x86_64/ccall.S

@@ -2,6 +2,7 @@
 
 #define ARG0 RUSTRT_ARG0_S
 #define ARG1 RUSTRT_ARG1_S
+#define ARG2 RUSTRT_ARG2_S
         
         .text
 
@@ -72,3 +73,18 @@ upcall_call_c_stack_shim:
     mov %rbp,%rsp
     pop %rbp
     ret
+
+#if defined(__APPLE__) || defined(_WIN32)
+.globl _asm_call_on_stack
+_asm_call_on_stack:
+#else
+.globl asm_call_on_stack
+asm_call_on_stack:
+#endif
+    push %rbp
+    mov %rsp,%rbp          // save rsp
+    mov ARG2,%rsp          // switch stack
+    call *ARG1             // invoke target address
+    mov %rbp,%rsp
+    pop %rbp
+    ret

src/rt/arch/x86_64/context.h

@@ -29,6 +29,8 @@ struct registers_t {
     uint64_t data[RUSTRT_MAX];
 };
 
+extern "C" void asm_call_on_stack(void *args, void *fn_ptr, uintptr_t stack_ptr);
+
 class context {
 public:
     registers_t regs;
@@ -55,6 +57,10 @@ public:
 
         return reinterpret_cast<void *>(top);
     }
+
+    void call_shim_on_c_stack(void *args, void *fn_ptr) {
+        asm_call_on_stack(args, fn_ptr, regs.data[RUSTRT_RSP]);
+    }
 };
 
 #endif