Cranelift: look into removing redundant SSA parameters

[lars-t-hansen]

@sunfishcode, we were discussing redundant SSA parameters. Here's a simple example. It's possible that the return block is a special case and that this is not what you were thinking about when you said redundant parameters should not occur.

Using the SpiderMonkey embedding here. Wat source:

(module
  (func (export "f") (param i32 i32 i32 i32 i32 i32 i32 i32 i32 i32) (result i32)
    (i32.div_s (local.get 5) (local.get 1))))

IR code:

$ RUST_LOG=debug dist/bin/js --wasm-compiler=cranelift ~/moz/manyparam.js
...
function u0:0(i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i64 vmctx) -> i32 uext baldrdash {
                                ebb0(v0: i32, v1: i32, v2: i32, v3: i32, v4: i32, v5: i32, v6: i32, v7: i32, v8: i32, v9: i32, v10: i64):
@0041                               v12 = sdiv v5, v1
@0042                               jump ebb1(v12)

                                ebb1(v11: i32):
@0042                               fallthrough_return v11
}

(Will look for additional examples.)

[sunfishcode]

Ah, yes, that one's the return block, which is created by the wasm translation code, rather than by SSA construction.

We might be able to fix this by special-casing end of the function in the opcode::End logic.

[cfallin]

Renaming this to a slightly more general topic (remove redundant SSA parameters); this might be an interesting pass to have eventually. It looks like right now the specific case above -- the return arg -- is still generated. In general we don't have an invariant that we keep minimal SSA, so this is a potential enhancement rather than bug (to be explicitly clear about it); labeling as such.

[fitzgen]

We do have a redundant phi (block param) removal pass.

It isn't complete, but in #8565 we found that making it complete wasn't a clear win, at least without better remat/regalloc. Will close this in favor of further explorations like that one in the future.