[llvm-dev] How to correctly caller-save return address register

[Dominik Montada via llvm-dev]

Hi,

I'm working on a new calling convention for a downstream target. Our
target normally saves and restores the return address automatically with
the call and return instruction, so there is no need to to handle it in
any way. The new calling convention however must save the return address
in the caller manually. I thought that the register allocator would
spill and restore the register holding the return address around a call,
but I guess since we marked the register as reserved, it is not doing that.

I'm wondering now how I should correctly handle this. As a workaround I
inserted COPY instructions to and from virtual registers around the
call, but although that works, it looks more like a hack instead of a
solution.
Could someone tell me how handling the save and restore of a return
address register should look like?

Btw, I tried looking at ARM, but it looks like that LR is marked as
callee-saved and therefore handled during prologue/epilogue creation
(IIUC). That is not an option for us though.

Cheers,

Dominik

[Dominik Montada via llvm-dev]

I actually managed to solve my own problem. If anyone has a similar issue in the future, this is what I did:

I marked the return address register as being callee saved in both calling conventions. In determineCalleeSaves I then check whether there is a function call that uses the new calling convention. If so, I mark the return address reg as being required to be spilled. This way I basically get a caller-save behavior and can use the already existing infrastructure in TargetFrameLowering.

Cheers,

Dominik

_______________________________________________
LLVM Developers mailing list
llvm...@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev