Seperating LLVM representation for processing with different backends

Hi,

I am writing a pass which identifies the parts(basic
blocks/functions) of
the input algorithm with more of data flow or control
flow. These parts are to
be separated executed by different backends.
My idea is to insert replacement basic blocks with
intrinsic instructions
within the basic block. These instructions should
satisfy the data
dependencies within the LLVM data structure and should
do the parameter
marshalling for switching to the other architecture to
and fro.

Dependencies of the different passes:

analysis part
          >
       filter
      / \
ppc custom backend

Switching between the ppc and custom backend would be
by using different
filter passes in two compiler runs.

I have a couple of questions regarding the problem
outlined above:
Is the approach using intrinsic instructions
worthwile?
What has to be done to the existing backends to make
these new instrinsic
instructions known?
Is there an way to eliminate two compiler runs, one
for each backend?

Best regards
Aditya

Hi,

I am writing a pass which identifies the parts(basic
blocks/functions) of
the input algorithm with more of data flow or control
flow. These parts are to
be separated executed by different backends.
My idea is to insert replacement basic blocks with
intrinsic instructions
within the basic block. These instructions should
satisfy the data
dependencies within the LLVM data structure and should
do the parameter
marshalling for switching to the other architecture to
and fro.

Dependencies of the different passes:

analysis part
          >
       filter
      / \
ppc custom backend

Ok.

Switching between the ppc and custom backend would be
by using different
filter passes in two compiler runs.

I have a couple of questions regarding the problem
outlined above:
Is the approach using intrinsic instructions
worthwile?

Probably. If what you are trying to represent is truly target specific then using intrinsics is good solution.

What has to be done to the existing backends to make
these new instrinsic
instructions known?

You need to add Intrinsics??.td file to include/llvm as well as instruction selection patterns / code to translate them to target specific instructions.

Is there an way to eliminate two compiler runs, one
for each backend?

Right now, the only way I see that can be possible is if you have a single target capable of handling codegen for both parts. So for example, if you can make the target into a variant of ppc, then you can potentially dynamically switch between the two for different BB's. But I would suggest you put the idea away for now. It would seem to me the problem you described present plenty of challenges for now. :slight_smile:

Evan