[PowerPC] Support constrained vector fp/int conversion

[qiucf]

Summary

This patch makes these operations legal, and add necessary codegen patterns.

There's still some issue similar to D77033 for conversion from/to v1i128, but normal type tests synced from X86/SystemZ's vector-constrained-fp-intrinsics.ll are all okay.

[uweigand]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

8638

It would seem that this doesn't create a correct node when called with a STRICT version of Opc -- for one, in this case it will need to handle chains correctly (in and out).

It is strange that this was not detected by any tests -- is the coverage good enough?

The rest of the algorithm seems OK for the strict case, since it only introduces integer operations.

[qiucf]

Fix incorrect operand in LowerINT_TO_FPVector.

[qiucf]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

8638

Yes, it was missed here. Thanks for comments!

This method is for custom lowering v4i8/v4i16. They are not set to custom in this patch so such cases will be expanded automatically. After setting them, it hits an assert since PPC doesn't override TargetLowering::LowerOperationWrapper. This can be done in later patches.

if (SDValue Res = LowerOperation(SDValue(N, 0), DAG))
  Results.push_back(Res);

BTW, the method looks strange to me, maybe we can fill rest of operands in base class implementation.

[uweigand]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

8638

Well, if the changes to this routine are never exercised without some additional changes, maybe they should be removed from this patch and added to the patch that'll introduce the first use?

On the other hand, I see above that at least for v4i1 the method is set to Custom, so is this routine used after all? (Not sure when v4i1 is actually used here, a bool->float conversion looks rare ... Are there test cases for v4i1?)

[qiucf]

Add handling for v4i16 and tests

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

8638

Oh, v4i1 would be promoted to v4i32. So setting it to custom (both strict and non-strict) is actually meaningless.. I'll add the handling for v4i16 and others along with tests.

[uweigand]

This LGTM now.