The tale of "-mrtd" in GCC and Clang

https://m680x0.github.io/blog/2023/05/the-tale-of-mrtd.html

by cbmuser • 3 years ago

85 12 3 years ago

m680x0.github.io

cbmuser• 3 years ago

This story was uncovered while working on the M68k backend in LLVM.

If you think that LLVM having an M68k is a great idea, please consider supporting the development via OpenCollective or Patreon:

- https://opencollective.com/m68k-llvm-dev

- https://www.patreon.com/m68k_llvm

jwilk• 3 years ago

GCC 1.21 (released in May 1988) had -mrtd for m68k, but it didn't support i386 yet:

https://gcc.gnu.org/pub/gcc/old-releases/gcc-1/gcc-1.21.tar....

mshockwave• 3 years ago

I didn’t know they put pre-VCS GCC on their own server. Thanks for pointing out!

Y_Y• 3 years ago

Alternatively I can directly ask some of the early GCC contributors. Unfortunately I’m not sure whether their emails are still reachable or even worse, whether they are still with us.

Last I checked, rms himself is still alive and answering emails. He might be a good place to start.

unwind• 3 years ago

As someone growing up writing assembler on the M68k-based Amiga, reading about the rtd instruction was mildly shocking. Learning it was 68010 and later calmed me down, phew.

Edit: spelling.

ksherlock• 3 years ago

I re-used the 68k engine from an amiga emulator once. I only encountered one bug: the RTD instruction wasn't implemented correctly.

tempodox• 3 years ago

I remember the RTD instruction. In classic macOS 7 it's employed by code that uses the Pascal calling convention.

pjmlp• 3 years ago

That is how it worked on MS-DOS as well, the enter/leave instruction pairs for x86 had a similar purpose.

microtherion2• 3 years ago

The reason why it was generally not used in C was probably that in pre-ANSI C, prototypes were optional, calling conventions were enforced loosely, and even in functions that were not officially variadic, sometimes trailing arguments were omitted deliberately or accidentally.

fweimer• 3 years ago

Prototypes did not exist in original K&R. Even function declarations (without prototypes/parameter lists) were still optional in C89 (“ANSI C”), only C99 disallowed calling undeclared functions. To this day, GCC accepts call to undeclared functions by default. Generally, the ABIs GCC supports are carefully constructed to support such usage in many historically important cases (such as calling functions declared in <unistd.h>, many of which use the default int return type, too).

The K&R legacy is likely the reason for the caller-pops-arguments convention, but there was also a time where it was deemed more efficient because you could combine multiple stack adjustments.

There's also a neat trick I learned from LuaJIT: if you store the stack pointer in a callee-saved register before the call and restore it afterwards, it works with both conventions.

gumby• 3 years ago

Even though rcs and sccs were in every Unix distro and some other competitors existed their use was not common back then.

Rich Pixley deserves all the credit for getting gnu sources into source control, specifically CVS which at the time was the latest hotness (it was not yet client server; that was added later by IIRC Ian Taylor after some unfortunate incidents).

John and I placed some absurd hoops that needed to be jumped through in order for source control to be adopted, all of which Rich wisely just ignored.

The tale of "-mrtd" in GCC and Clang

Related Stories