[LLVMdev] Handling segmented instruction space in backend for custom target

Anoushe Jamshidi ajamshidi at gmail.com
Fri Nov 16 12:09:53 PST 2012

Hi all,

I'm building a backend for a custom target, and I'm trying to figure out
how to handle global and external target address with my architecture's
call instructions.

This arch. has 16-bit addresses into a segmented address space, and to do a
direct call I need to set both an instruction counter (IC, 10 bits wide)
and an instruction segment register (ISR, 6 bits wide). My CALL instruction
can set the IC, and a LISR instruction can load the ISR.  The ISR doesn't
need to be set if the callee lives within the same instruction segment,
i.e. it's not a far call, no LISR instruction needed.  If it is a far call,
the LISR instruction precedes the CALL.

1) I'm not sure exactly how to transform the global & external target
addresses, nor how to take the transformation and match them to my LISR &
CALL instructions in my XXXInstrInfo.td TableGen file.  I've been looking
at how the Mips target has hi/lo relocations to handle 16-bits of an
address at a time, but I don't see how the relocations are inserted(?) into
the patterns for the JAL/JALR nodes.  Does anyone have any advice on the
best way to do this, and/or how the Mips addresses are handled?

2) I'm also not sure how (or at what stage of codegen) to check if the
callee lives in the same instruction segment.  How can I compare a call
instruction's address with that of its target?  If I can do this, I could
remove (or maybe not select) LISR instructions when they are unnecessary.

Maybe this is all elementary and I just haven't been searching for the
right things, but I'm new at this and would greatly appreciate any help.

Thank you,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20121116/13c736e6/attachment.html>

More information about the llvm-dev mailing list