vector miscompile from AArch64 backend

[regehr]

please bear with me here, this is the best I could get out of llvm-reduce:

define i16 @f(<8 x i16> %0) {
  %2 = shufflevector <8 x i16> %0, <8 x i16> zeroinitializer, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 9, i32 7, i32 5, i32 3>
  %3 = zext <8 x i16> %2 to <8 x i64>
  %new0 = add <8 x i64> %3, %3
  %last = trunc <8 x i64> %new0 to <8 x i16>
  %4 = extractelement <8 x i16> %last, i32 0
  ret i16 %4
}

Alive thinks (and I agree) that when this is called like this:

call i16 @f(<8 x i16> <i16 0, i16 3, i16 3, i16 3, i16 1, i16 3, i16 3, i16 3>)

the result should be 2.

but when I compile to AArch64 using current llc I get:

_f:
	movi.2d	v1, #0x00ffff0000ffff
	uzp1.4s	v0, v0, v0
	and.16b	v0, v0, v1
	add.4s	v0, v0, v0
	xtn.4h	v0, v0
	umov.h	w0, v0[0]
	ret

and this -- for that same input -- computes 0

cc @Hatsunespica @nunoplopes

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: John Regehr (regehr)

please bear with me here, this is the best I could get out of llvm-reduce: ```llvm define i16 @f(<8 x i16> %0) { %2 = shufflevector <8 x i16> %0, <8 x i16> zeroinitializer, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 9, i32 7, i32 5, i32 3> %3 = zext <8 x i16> %2 to <8 x i64> %new0 = add <8 x i64> %3, %3 %last = trunc <8 x i64> %new0 to <8 x i16> %4 = extractelement <8 x i16> %last, i32 0 ret i16 %4 } ``` Alive thinks (and I agree) that when this is called like this: ``` call i16 @f(<8 x i16> <i16 0, i16 3, i16 3, i16 3, i16 1, i16 3, i16 3, i16 3>) ``` the result should be 2.

but when I compile to AArch64 using current llc I get:

_f:
	movi.2d	v1, #<!-- -->0x00ffff0000ffff
	uzp1.4s	v0, v0, v0
	and.16b	v0, v0, v1
	add.4s	v0, v0, v0
	xtn.4h	v0, v0
	umov.h	w0, v0[0]
	ret

and this -- for that same input -- computes 0

cc @hatsunespica @nunoplopes

[davemgreen]

Thanks for the report. I think performZExtDeinterleaveShuffleCombine might be missing an extra check for undef elements.

[davemgreen]

/cherry-pick 2c43479

[llvmbot]

/pull-request #126263