Hi,

I removed my change to statements.cc and made a change as you suggested in libgo/runtime/go-defer.c in function __go_set_defering_fn:

+#if defined(__powerpc64__) && _CALL_ELF != 2
+    g->defer->__defering_fn = *(void **)defering_fn;
+#else
+    g->defer->__defering_fn = defering_fn;
+#endif

With this change along with PATCH 1/9, the recover.go testcase passes, there are no new regressions on for ppc64 LE or BE or ppc (32 bit BE) when running the go and libgo testsuites and I agree is the right solution.

Thanks,

Lynn

On 09/25/2014 06:44 AM, Ulrich Weigand wrote:
Lynn Boger wrote:

I modified the patch for statements.cc and rebuilt and that eliminates
the regressions and fixes the original problem it was intended to fix
for both ppc64 BE & LE.  The ABIs are different between BE & LE, so that
make_func_code_reference on ppc64 BE is not returning the function's
code address but the function pointer from the .opd.  The first 8 bytes
of the entry in the .opd is the function's code address.  Here is the
change to statements.cc that made it work:
+#if defined(__powerpc64__) && _CALL_ELF != 2
+      Expression* pfn =
+        Expression::make_func_code_reference(function, location);
+      Type* pfntype =
+          Type::make_pointer_type(
+              Type::make_pointer_type(Type::make_void_type()));
+      Expression* fn = Expression::make_unsafe_cast(pfntype, pfn,
location);
+      Expression* fn_code_addr = Expression::make_unary(OPERATOR_MULT, fn,
+                                                        location);
+#else
+      Expression* fn_code_addr =
+        Expression::make_func_code_reference(function, location);
+#endif
+      Expression* call = Runtime::make_call(Runtime::SET_DEFERING_FN,
+                                            location, 1, fn_code_addr);
+      Statement* s = Statement::make_statement(call, true);
This looks wrong when using gcc-go as a cross-compiler.  The #if is
evaluated in the context of the *host*, but you'd need to check the
processor architecture and ABI of the *target*.  This seems difficult
since you'd have to take into account -mabi= options, which are not
readily available to the front end.

It seems more straightforward to keep the front end as is, i.e. generate
code to pass a plain function pointer (as defined by the target ABI) to
the runtime, and have the *runtime* do whatever target-specific fiddling
is required to get from a function pointer to a code address.

For example, you could add something like:

#if defined(__powerpc64__) && _CALL_ELF != 2
   defering_fn = *(void **)defering_fn;
#endif

to __go_set_defering_fn (or possibly __go_can_recover).

[ Since the runtime is compiled for the target with the appropriate
ABI setting, the #if works as intended when in runtime code.  ]

Bye,
Ulrich


Reply via email to