1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=s390x-linux-gnu -mcpu=z13 < %s | FileCheck %s
4 ; Test that DAGCombiner gets helped by computeKnownBitsForTargetNode().
6 ; SystemZISD::REPLICATE
7 define i32 @f0(<4 x i32> %a0) {
10 ; CHECK-NEXT: vgbm %v0, 0
11 ; CHECK-NEXT: vceqf %v0, %v24, %v0
12 ; CHECK-NEXT: vrepif %v1, 1
13 ; CHECK-NEXT: vnc %v0, %v1, %v0
14 ; CHECK-NEXT: vlgvf %r2, %v0, 3
15 ; CHECK-NEXT: # kill: def $r2l killed $r2l killed $r2d
17 %cmp0 = icmp ne <4 x i32> %a0, zeroinitializer
18 %zxt0 = zext <4 x i1> %cmp0 to <4 x i32>
19 %ext0 = extractelement <4 x i32> %zxt0, i32 3
23 ; The vector icmp+zext involves a REPLICATE of 1's. If KnownBits reflects
24 ; this, DAGCombiner can see that the i32 icmp and zext here are not needed.
25 %cmp1 = icmp ne i32 %ext0, 0
26 %zxt1 = zext i1 %cmp1 to i32
30 ; SystemZISD::JOIN_DWORDS (and REPLICATE)
31 ; The DAG XOR has JOIN_DWORDS and REPLICATE operands. With KnownBits properly set
32 ; for both these nodes, ICMP is used instead of TM during lowering because
33 ; adjustForRedundantAnd() succeeds.
37 ; CHECK-NEXT: clhhsi 0, 0
38 ; CHECK-NEXT: lhi %r0, 0
39 ; CHECK-NEXT: lochie %r0, 1
40 ; CHECK-NEXT: lghi %r1, 1
41 ; CHECK-NEXT: vlvgp %v0, %r0, %r1
42 ; CHECK-NEXT: vrepig %v1, 1
43 ; CHECK-NEXT: vx %v0, %v0, %v1
44 ; CHECK-NEXT: vlgvf %r0, %v0, 1
45 ; CHECK-NEXT: cijlh %r0, 0, .LBB1_3
46 ; CHECK-NEXT: # %bb.1:
47 ; CHECK-NEXT: vlgvf %r0, %v0, 3
48 ; CHECK-NEXT: cijlh %r0, 0, .LBB1_3
49 ; CHECK-NEXT: # %bb.2:
50 ; CHECK-NEXT: .LBB1_3:
51 %1 = load i16, i16* null, align 2
52 %2 = icmp eq i16 %1, 0
53 %3 = insertelement <2 x i1> undef, i1 %2, i32 0
54 %4 = insertelement <2 x i1> %3, i1 true, i32 1
55 %5 = xor <2 x i1> %4, <i1 true, i1 true>
56 %6 = extractelement <2 x i1> %5, i32 0
57 %7 = extractelement <2 x i1> %5, i32 1
59 br i1 %8, label %10, label %9
61 ; <label>:8: ; preds = %0
64 ; <label>:9: ; preds = %0