Lower function invocations to Verilog module instantiations

[joachimschmidt557]

Hi! I was wondering whether it would be possible to lower DSLX function invocations to module instantiations in the codegen pass. A small example:

fn foo(a: u32, b:u32) -> u32 {
  a + b
}

fn test(a: u32, b:u32) -> u32 {
  foo(a, b)
}

Converted to XLS IR, this produces

package test

fn __test__foo(a: bits[32], b: bits[32]) -> bits[32] {
  ret add.3: bits[32] = add(a, b, id=3, pos=0,1,4)
}

fn __test__test(a: bits[32], b: bits[32]) -> bits[32] {
  ret invoke.6: bits[32] = invoke(a, b, to_apply=__test__foo, id=6, pos=0,5,5)
}

Currently, when the invoke IR instruction is encountered by codegen_main, an error message is emitted. Would it be possible to convert that invoke instruction into a Verilog module instantiation and thus "preserve the function/module hierarchy"? Like this:

module __test__foo(
  input wire [31:0] a,
  input wire [31:0] b,
  output wire [31:0] out
);
  assign out = a + b;
endmodule

module __test__test(
  input wire [31:0] a,
  input wire [31:0] b,
  output wire [31:0] out
);
  foo foo_0(a, b, out);
endmodule

[meheff]

This wouldn't be too hard to implement. We already have the IR nodes and construct for module instantiation (they are not used at the moment). Codegen would have to walk the call graph and emit all the dependent modules first. To prevent inlining of the call sites, one could disable inlining entirely or add an attribute at the on functions DSLX level (and reflected in the IR) which disables inlining.

Currently we don't do any global (cross-function) optimizations so not inlining functions could potentially have some qor ramifications.

[joachimschmidt557]

Hi! Thanks for the pointers on how the implementation could look like. I implemented lowering function invocations to module instantiations in a local checkout (only for combinational modules so far, as I'm mostly interested in those). I modified GenerateCombinationalModule to walk the call graph in post-order (similar to the inlining pass). Furthermore, I introduced a new codegen pass InvocationToInstantiation which converts every invoke node in the block to an instantiation together with the InstantiationInput and InstantiationOutput nodes.

Let me know if you would be interested in upstreaming the implementation, otherwise, I'll just maintain it in a fork 👍

[proppy]

otherwise, I'll just maintain it in a fork

@joachimschmidt557 is it this fork main...joachimschmidt557:xls:spfe-framework ?

[joachimschmidt557]

Yes, I have based the changes on a previous version of XLS. Since then, multiple changes to the base files have happened, so that the changes can't be merged at the moment. I might rebase it if there is interest...

[proppy]

@meheff @grebe should we transform this in an enhancement proposal?