Lower PtrToInst instructions.

[ptersilie]

No description provided.

[vext01]

Just reading the LLVM reference, ptrtoint has more semantics than we first thought:
https://llvm.org/docs/LangRef.html#ptrtoint-to-instruction

• We should reject the vector variant (encode an unimplemented instruction). You can copy/adapt existing code from other opcodes that have vector variants.
• ptrtoint can truncate to a smaller integer size, so zero extension isn't generally correct. If you don't want to handle those now, you can make an unimplemented instruction if the output type isn't pointer sized.

[vext01]

(when rejecting certain versions of instructions, I've also been adding a case to tests/ir_lowering/unsupported_variants.ll to check they are properly rejected)

[ptersilie]

My first instinct after your comment was to just add a aot_ir::PtrToIntInst. But reading the langref, it does say they either truncate/zeroextend/noop anyway. So we can probably do without such an instruction and just lower each case to different existing aot instructions.

[vext01]

that would be fine by me.

[ptersilie]

I've updated the PR.

[vext01]

Bitwidth check looks good to me.

We just need to reject the vector variant of inttoptr now. e.g.

%Z = inttoptr <4 x i32> %G to <4 x ptr>

You can adapt this kind of thing:

ykllvm/llvm/lib/YkIR/YkIRWriter.cpp

Line 380 in c645c79

I->getType()->isVectorTy()) {

[ptersilie]

Here you go.

[vext01]

Perfect. Please squash.

[ptersilie]

Done.

[ptersilie]

Ok to squash?

[vext01]

Please squash

[ptersilie]

Done.