Many years ago (in the early 90's given I was OSF at the time), I was working on a MIPS based machine that there were no user callable instructions to flush the cache (these instructions were added in MIPS-IV, but weren't available at the time), and at least one of the OS's had no system call to make the stack executable (DEC's Ultrix). I attempted to fix all of the places where there were hidden assumptions about the trampolines being on the stack, and after a month or so of frustration, finally gave up and just made it a restriction for that particular port. Since then, I've done two other ports where trampolines were not possible (one where the code and data were in completely different address spaces, and it was only by use that the hardware would go to the I or D space, and another machine with cache problems). -----Original Message----- From: gcc-help-owner@xxxxxxxxxxx [mailto:gcc-help-owner@xxxxxxxxxxx] On Behalf Of John Yates Sent: Thursday, December 29, 2005 10:33 AM To: gcc-help@xxxxxxx Cc: Marco Gerards Subject: RE: Nested functions Wednesday, December 28, 2005 7:51 AM mgerards@xxxxxxxxx wrote: > The second problem is caused because some OS'es have no executable > stack. Redhat even removes the nested functions from our code so it > builds for them. But of course it would be better for all of our > users if it just works. Is it not the case that building the trampoline on the stack is simply an efficiency issue, akin to alloca versus malloc? Surely, at least in C++ with support for calling destructors, trampolines could be build on the heap. /john
