Hi all,
Seems like it that bug 18702 has nothing to do with my changes to the
stack-balancing code, and that nested function proc's never functioned
well on Windows.
The problem is that pop_parasize() is used to pop the stack-frame
pointer form the stack. But on 32-bit Windows, pop_parasize() has some
code to ensure that the funcret pointer is removed from the stack by the
caller. This is not applicable to the situation where only the
stack-frame pointer has to be popped from the stack.
A possible solution is attached, but I didn't commit it yet because I'm
nut sure if it is 'allowed' to use a 'target_info.system =
system_i386_win32' condition in ncgcal.pas. Other solutions could be to
add a parameter to pop_parasize or to add something like
pop_funcret_parasize().
Thoughts?
Joost.
Index: i386/n386cal.pas
===================================================================
--- i386/n386cal.pas (revision 19881)
+++ i386/n386cal.pas (working copy)
@@ -81,16 +81,6 @@
exit;
end;
- { on win32, the caller is responsible for removing the funcret }
- { pointer from the stack, unlike on Linux. Don't know about }
- { elsewhere (except Darwin, handled above), but since the default }
- { was "callee removes funcret pointer from stack" until now, we'll }
- { keep that default for everyone else (ncgcal decreases popsize by }
- { sizeof(aint) in case of ret_in_param()) }
- if (target_info.system = system_i386_win32) and
- paramanager.ret_in_param(procdefinition.returndef,procdefinition.proccalloption) then
- inc(pop_size,sizeof(aint));
-
{ better than an add on all processors }
if pop_size=4 then
begin
Index: ncgcal.pas
===================================================================
--- ncgcal.pas (revision 19881)
+++ ncgcal.pas (working copy)
@@ -875,6 +875,18 @@
(tf_safecall_exceptions in target_info.flags)) and
paramanager.ret_in_param(procdefinition.returndef,procdefinition.proccalloption) then
dec(pop_size,sizeof(pint));
+
+ { on win32, the caller is responsible for removing the funcret }
+ { pointer from the stack, unlike on Linux. Don't know about }
+ { elsewhere (except Darwin, handled above), but since the default }
+ { was "callee removes funcret pointer from stack" until now, we'll }
+ { keep that default for everyone else (ncgcal decreases popsize by }
+ { sizeof(aint) in case of ret_in_param()) }
+ if (not paramanager.use_fixed_Stack) and
+ (target_info.system = system_i386_win32) and
+ paramanager.ret_in_param(procdefinition.returndef,procdefinition.proccalloption) then
+ inc(pop_size,sizeof(aint));
+
{ Remove parameters/alignment from the stack }
pop_parasize(pop_size);
end
_______________________________________________
fpc-devel maillist - [email protected]
http://lists.freepascal.org/mailman/listinfo/fpc-devel
