[CIR][ThroughMLIR] Lower pointers to memref<?xType>

[TopiLeppanen]

Hello, I attempted to resolve the #1729, as I have been compiling some simple OpenCL kernels similar to @mahmood82.

According to the @felixdaas's suggestion of using the 'polygeist way', I have changed the default type conversion of cir.ptr to memref<?xType>.

To make the intermediate legalization work, this also changes the lowering of AllocaOp and GlobalOp to generate memref<1xType> followed by a CastOp to memref<?xType>. CastOp is later eliminated in Load/Store-lowering. Load/Store-lowering's findBaseAndIndices was then adjusted to generate op with zero indices in case a non-indexed op is performed.

I added a testcase ptr-arg.cir, with example code similar to one in #1729

I also had to touch the loop lowering to SCF, since the alloca-change affected the 'canonical-loop recognition'.

I also had to change all the ThroughMLIR tests to work with the alloca of memref<1xType>.

Additionally, the ptr.dialect based lowering by @terapines-osc-cir is now basically dead code, as the ptrstride-ptr.cir-test they added is now lowered directly to memref-ops. So probably a new, more complex testcase would be needed to also test it, and to change the ptrstrideop-lowering to still take that path in the more complex cases.

[TopiLeppanen]

CastOp is later eliminated in Load/Store-lowering.

Forgot to mention, but this was an optional step I decided to add in order to generate slightly less code, but we could just as well just leave the CastOp there, to simplify the implementation. I noticed that canonicalization is able to remove it as well. Leaving it in there would probably also reduce the changes I had to make in SCF-loop lowering as there's no longer this issue of referencing an op that has already been bypassed. As a tradeoff, all the tests that use alloca would then need to be fixed to have this memref.cast-op after the alloca, but that wouldn't be too difficult, just a bit noisy.

Let me know if that would make more sense, and I can do it pretty easily.

[bcardosolopes]

@TopiLeppanen thanks for your first CIR contribution! I like the plan, the 'polygeist way' makes us closer to things we already know to work, way to go. Before we land this I just want to make sure that @terapines-osc-cir is happy with the path, can you give us a stamp?

[terapines-osc-cir]

Before we land this I just want to make sure that @terapines-osc-cir is happy with the path

Cool! Like it 👍

[bcardosolopes]

LGTM once conflicts are fixed