Which stage will promote the slot in Mem2Reg?

IR & Optimizations
1

I would like to use Mem2Reg to eliminate the local variables, like this:
It’s a SystemVerilog source code.

nit: You can regard the following Verilog/SystemVerilog code as C++.

Before

module Foo;
  int x, y;
  always_comb begin
    int a;
    a = x + 1;
    y = a;
  end
endmodule

moore dialect IR

moore.module @Foo {
    %x = moore.variable  : !moore.i32
    %y = moore.variable  : !moore.i32
    moore.procedure always_comb {
      %a = moore.variable  : !moore.i32
      %0 = moore.constant 1 : !moore.i32
      %1 = moore.add %x, %0 : !moore.i32
      moore.blocking_assign %a, %1 : !moore.i32
      moore.blocking_assign %y, %a : !moore.i32
    }
  }

After

module Foo;
  int x, y;
  always_comb begin
    y = x + 1;
  end
endmodule

moore dialect IR

moore.module @Foo {
    %x = moore.variable  : !moore.i32
    %y = moore.variable  : !moore.i32
    moore.procedure always_comb {
      %0 = moore.constant 1 : !moore.i32
      %1 = moore.add %x, %0 : !moore.i32
      moore.blocking_assign %y, %1 : !moore.i32
    }
  }

But this pass doesn’t work as I expected.
There is the related implementation of Op Interface Method.

image
image690×422 85.7 KB

image
image657×500 87 KB

I’m not sure at which stage the local variable(a) will be substituted with x+1. And why it doesn’t work. Please click (https://github.com/llvm/circt/pull/7082) to view a more detailed example.
Thanks for any help in advance.

1 Like
2

I have solved this problem. I think the change/replace stage occurs at the dedicated load operation. In other words, we need not only ops like llvm.alloca and llvm.store, but also need an op like llvm.load. If anyone met a similar problem. Please click [Moore] Introduce Mem2Reg to eliminate local variables. by hailongSun2000 · Pull Request #7082 · llvm/circt · GitHub and [Moore] Introduce RefType and tweak the related ops. by hailongSun2000 · Pull Request #7095 · llvm/circt · GitHub to view the details. The similar ops in moore dialect are moore.variable(alloca), moore.read(load), and moore.blocking_assign(store).