Compiler: keep Int32 and Nativeint in the IR for the js backend

[hhugo]

Previously, primitives taking float arguments
could receive Int/Int32/NativeInt/Float arg

[hhugo]

Experimentation based on this comment #1649 (comment)

[hhugo]

With the current PR, compiling ocamlc result in the following diff. Two 4 become 4.

$ diff master.js pr.js 
35587c35587
<          Math.floor(v14535 * 4 / v14534 * 1000. + 0.5) / 1000.),
---
>          Math.floor(v14535 * 4. / v14534 * 1000. + 0.5) / 1000.),
35595c35595
<     var v14543 = [0, v14542.length - 2 | 0], v14545 = v14544 * 4;
---
>     var v14543 = [0, v14542.length - 2 | 0], v14545 = v14544 * 4.;

[hhugo]

The issue now is that wasmoo uses information from the js runtime but some //Alias: annotations are not wasmoo compatible. @vouillon, do you have any plan for another mechanism to keep "purity" information for the wasm runtime ?

[vouillon]

The issue now is that wasmoo uses information from the js runtime but some //Alias: annotations are not wasmoo compatible. @vouillon, do you have any plan for another mechanism to keep "purity" information for the wasm runtime ?

I did not plan to do that. One issue is that if we start duplicating this information, we will have to make sure to keep things in sync. The other issue is that I don't know where else to put this information.

[hhugo]

This PR is now includes

• part of Optimization of primitives: some improvements #1912 (for more static evaluation)
• Effects: introduce primitive caml_cps_trampoline #1922

With this PR, wasmoo no longer resolves aliases during codegen. Aliases are still used to compute purity though.

I think there is still an issue with #1922, can you take a look ?