29 lines
1.1 KiB
LLVM
29 lines
1.1 KiB
LLVM
|
; RUN: llc -march=hexagon < %s | FileCheck %s
|
||
|
|
|
||
|
|
; Test that the compiler generates the correct code when sign-extending a
|
||
|
|
; predicate register when it is converted from one vector predicate type
|
||
|
|
; to another. In this case, the compiler generates two v4i1 EXTRACT_SUBVECT
|
||
|
|
; from a v8i1, for the lower and upper parts.
|
||
|
|
|
||
|
|
; CHECK: r[[REGH:([0-9]+)]]:[[REGL:([0-9]+)]] = mask(p{{[0-3]}})
|
||
|
|
; CHECK-DAG: = vsxtbh(r[[REGH]])
|
||
|
|
; CHECK-DAG: = vsxtbh(r[[REGL]])
|
||
|
|
|
||
|
|
target triple = "hexagon"
|
||
|
|
|
||
|
|
define void @f0(i16* %a0, <8 x i16>* %a1) #0 {
|
||
|
|
b0:
|
||
|
|
%v0 = load i16, i16* %a0, align 2
|
||
|
|
%v1 = sext i16 %v0 to i32
|
||
|
|
%v2 = insertelement <8 x i32> undef, i32 %v1, i32 0
|
||
|
|
%v3 = shufflevector <8 x i32> %v2, <8 x i32> undef, <8 x i32> zeroinitializer
|
||
|
|
%v4 = shl <8 x i32> <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>, <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
|
||
|
|
%v5 = and <8 x i32> %v4, %v3
|
||
|
|
%v6 = icmp ne <8 x i32> %v5, zeroinitializer
|
||
|
|
%v7 = zext <8 x i1> %v6 to <8 x i16>
|
||
|
|
store <8 x i16> %v7, <8 x i16>* %a1, align 8
|
||
|
|
ret void
|
||
|
|
}
|
||
|
|
|
||
|
|
attributes #0 = { nounwind optsize "target-cpu"="hexagonv65" "target-features"="+hvx-length64b,+hvxv65,-long-calls" }
|