Also remove all (limited) type inference from the sysir. Type
inference is better done in frontend, limited inference in backend
just covers compilers issues.
Simple hello world with nasm working.
We need the ability to represent multiple calling conventions in IR.
All backends need to support a :default CC, but can also support more
for interop with system libraries, code from other compilers, syscalls, etc.
Also allow void returns.
Needs changes to IR to allow encoding immediates in all
instructions where possible. This makes the IR denser, means
we don't need `constant` and `callk`, and allows certain optimizations
like comparing to zero, using `inc` and `dec`, etc which are
specializations of more general instructions with constants.
This doesn't seem to reintroduce the original issue. There was
definitely some interplay with #1431
Doing git bisect landed me at commit
2f0c789ea1 as the first bad commit for
issue #1452.
Setup frontend.janet to show the basics of what is going on. Currently
emitting "fake" instructions just to hash out the idea.
One apparent issue is how we handle register spilling during variable
argument IR instructions (function calls). Arguments should come
_before_ the function call not after.
Disallow jumping to arbitrary instructions - instead, only allow jumps
to label ids. This will make various transformations and validations
easier since adding or remove instructions does not break jumps.
Basic frontend being prototyped in examples/sysir/frontend.janet. Still
a lot of work needs to be done here, and some of this code will
eventually move to C most likely, but this is a good way to better
exercise our backend.
Type inference - at the very least _forward_ inference, is the most
needed change here. While one could do this in the compiler
frontend, doing so in sysir/asm is not so much of an issue. "Inference"
here means inserting "bind" instructions when there is only a single
type that will work correctly.
