I suspect the instruction-level annotations that exist (such as source location)
are the ones that people needed in order to get something done. If you want to
do something new, you will likely need a new kind of annotation.
Obviously the debug-info metadata survives through code generation, so you could
use that as a model. One caution: preserving the debug info requires a lot of
attention from the optimization passes (we have loads of bugs in this
department) so it seems quite likely that a new kind of annotation would require
similar attention.
--paulr
From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Son Tuan
VU via llvm-dev
Sent: Friday, July 13, 2018 4:24 PM
To: llvm-dev
Subject: [llvm-dev] MDNode equivalent in backend?
Hi all,
Let's say I want to annotate a C statement, let it be the simplest
assignment `int i = 3`. At the IR, this is straightforward as we can attach a
MDNode to the IR instruction representing the assignment. But I need to be able
to propagate the annotation to the binary ELF file. AFAIK, there is no mechanism
to do so in the backend. Is there any particular reason to not support some sort
of MIR annotation in the backend? Or is it purely because of manpower?
Thank you for your advice
Son Tuan Vu
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180716/a50e923c/attachment.html>