Incorrect code generated for llvm.masked.store <2 x i1> with a variable mask

[ZviRackover]

Bugzilla Link	31305
Version	trunk
OS	Windows NT
Blocks	#30613
CC	@topperc,@RKSimon

Extended Description

define void @​unmasked(<2 x i1> %x, <2 x i1>* %ptr) nounwind {
store <2 x i1> %x, <2 x i1>* %ptr
ret void
}

define void @​masked_const(<2 x i1> %x, <2 x i1>* %ptr) nounwind {
call void @​llvm.masked.store.v2i1.p0v2i1(<2 x i1> %x, <2 x i1>* %ptr, i32 0, <2 x i1> <i1 1, i1 1>)
ret void
}

For both functions above we generate:
llc -mcpu=skx

vpsllq $63, %xmm0, %xmm0
vptestmq %xmm0, %xmm0, %k0
kmovb %k0, (%rdi)
retq

But in the case of a variable mask:

define void @​masked_var(<2 x i1> %x, <2 x i1>* %ptr) nounwind {
call void @​llvm.masked.store.v2i1.p0v2i1(<2 x i1> %x, <2 x i1>* %ptr, i32 0, <2 x i1> %x)
ret void
}

the generated code stores each vector element in a distinct byte:
vpsllq $63, %xmm0, %xmm0
vptestmq %xmm0, %xmm0, %k0
kshiftlw $15, %k0, %k1
kshiftrw $15, %k1, %k1
kmovw %k1, %eax
andl $1, %eax
testb %al, %al
je .LBB2_2

BB#1: # %cond.store

    kmovb   %k1, (%rdi)

.LBB2_2: # %else
kshiftlw $14, %k0, %k0
kshiftrw $15, %k0, %k0
kmovw %k0, %eax
andl $1, %eax
testb %al, %al
je .LBB2_4

BB#3: # %cond.store1

    kmovb   %k0, 1(%rdi)

.LBB2_4: # %else2
retq

[RKSimon]

Trunk is still as bad: https://godbolt.org/z/qNsVis

[RKSimon]

These tests currently fail with:

masked_store: alignment must be a power of 2
call void @​llvm.masked.store.v2i1.p0v2i1(<2 x i1> %x, <2 x i1>* %ptr, i32 0, <2 x i1> <i1 true, i1 true>)