[AArch64] Suspect code near ICE on memset call.

[llvmbot]

Bugzilla Link	20147
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@jmolloy,@TNorthover

Extended Description

In r211960, I fixed an ICE in getMemsetValue.

The segment of code just below the fix looks suspect as well, but I was unable to generate a test case in my quick attempt.

Specifically, line 3638 seems to assume the VT of a vector element is either f32 or f64.

3636 else if (VT.isVector()) {
3637 unsigned NumElts = VT.getVectorNumElements();
3638 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64;
3639 return DAG.getNode(ISD::BITCAST, dl, VT,
3640 DAG.getConstant(0, EVT::getVectorVT(*DAG.getContext(),
3641 EltVT, NumElts)));
3642 } else

Hoping someone wouldn't mind taking a look.

[TNorthover]

The "isInteger" check above captures vectors of integers too.

[TNorthover]

Actually, I suppose vNf16 could exist.

[llvmbot]

Actually, I suppose vNf16 could exist.

Perhaps, we should just add an assertion?