LGTM

Thanks for the suggestions, I cleaned up the test so it looks a bit nicer now.

I made the NextStride == 2 condition more future proof as requested by Dave's comment. This required to make the MVEMaxSupportedInterleaveFactor from ARMISelLowering non-static.

Widened the range of allowed strides to also include loop invariant expressions.

Added a test that would fail the VecTy && "No scalable vectors expected here" assert without this fix.

Changed getVectorType.

Revisited the new test to make it cleaner.

LGTM - Changed should indeed reflect possible changes by that function, too.

Hi, there was a bug concerning the location at which the new getelementptr instruction would be generated, thanks for catching this.
I have fixed the bug and derived a LoopVectorize test case from your C example.

All nitpicks have been addressed.

Hi Sam, those tests are checking whether the cost model calculates the costs for zext(gather(...)) and scatter(trunc(..)) correctly, because these can be integrated in the MVE intrinsics.
That calculation relies on I being available. I see that you did integrate I in IntrinsicCostAttributes, as an IntrinsicInst; the gather/scatter cost function was taking an Instruction because when it is called we might either be looking at a gather/scatter intrinsic or still at an LLVM load/store, which I think is no intrinsic?
I can't see where it happens, but my guess is that the instruction that ends up being handed over to getGatherScatterOpCost in BasicTTIImpl::getIntrinsicInstrCost doesn't match what was expected anymore.

Last correction: For the incrementing non-writeback gathers, removed the check that the offset only has one use.

Applied the chances suggested in the comments.

Made optimisation of offsets independent of the creation of the individual gathers and scatters, such that when looking at transforming those we're not in an environment where some instructions have been optimised but others, which will be optimised too, haven't been yet.

Restructured the code such that the differentiation between writeback and non-writeback gathers is clearer

Fixing a bug in the latest diff: We should always make sure to scale the offsets that the loop is initialised with.

Fixed calculation of the constant in getIfConst, moved isGatherScatter to ARMBaseInstrInfo so it is accessible for other passes, too.

Made some changes suggested in the comments, and also re-ordered the code such that it is easier to follow.

Implemented some changes suggested in comments (see below for details)

Changes addressing most of the comments - see below.

Added scatter tests, and some code changes (see comments).

Moved around some tests and add two more.
Also made sure that extends or truncs to 'incomplete' vector types are not considered valid candidates to be merged into the gather.

Implemented some of the changes suggested in comments.