blob: 5ea4c1005fd2361e9f1be6aec6ce68f9cb63ee7f [file] [log] [blame]
; RUN: llc -mtriple=arm -mattr=+neon %s -o - | FileCheck %s
; Check codegen for 64-bit icmp operations, which don't directly map to any
; instruction.
define <2 x i64> @vne(ptr %A, ptr %B) nounwind {
;CHECK-LABEL: vne:
;CHECK: vceq.i32
;CHECK-NEXT: vrev64.32
;CHECK-NEXT: vand
;CHECK-NEXT: vmvn
;CHECK-NEXT: vmov
;CHECK-NEXT: vmov
;CHECK-NEXT: mov pc, lr
%tmp1 = load <2 x i64>, ptr %A
%tmp2 = load <2 x i64>, ptr %B
%tmp3 = icmp ne <2 x i64> %tmp1, %tmp2
%tmp4 = sext <2 x i1> %tmp3 to <2 x i64>
ret <2 x i64> %tmp4
}
define <2 x i64> @veq(ptr %A, ptr %B) nounwind {
;CHECK-LABEL: veq:
;CHECK: vceq.i32
;CHECK-NEXT: vrev64.32
;CHECK-NEXT: vand
;CHECK-NEXT: vmov
;CHECK-NEXT: vmov
;CHECK-NEXT: mov pc, lr
%tmp1 = load <2 x i64>, ptr %A
%tmp2 = load <2 x i64>, ptr %B
%tmp3 = icmp eq <2 x i64> %tmp1, %tmp2
%tmp4 = sext <2 x i1> %tmp3 to <2 x i64>
ret <2 x i64> %tmp4
}
; FIXME: We currently generate terrible code for this.
; (Atop < Btop) | ((ATop == BTop) & (ABottom < BBottom))
; would come out to roughly 6 instructions, but we currently
; scalarize it.
define <2 x i64> @vult(ptr %A, ptr %B) nounwind {
;CHECK-LABEL: vult:
;CHECK: subs
;CHECK: sbcs
;CHECK: subs
;CHECK: sbcs
%tmp1 = load <2 x i64>, ptr %A
%tmp2 = load <2 x i64>, ptr %B
%tmp3 = icmp ult <2 x i64> %tmp1, %tmp2
%tmp4 = sext <2 x i1> %tmp3 to <2 x i64>
ret <2 x i64> %tmp4
}