1; RUN: llc < %s -o - -mcpu=generic -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s
2; RUN: llc < %s -o - -mcpu=generic -mtriple=x86_64-apple-darwin -mattr=+sse4.2 | FileCheck %s
3
4; For a setult against a constant, turn it into a setule and lower via psubusw.
5
6define void @loop_no_const_reload(<2 x i64>*  %in, <2 x i64>* %out, i32 %n) {
7; CHECK: .short 25
8; CHECK-NEXT: .short 25
9; CHECK-NEXT: .short 25
10; CHECK-NEXT: .short 25
11; CHECK-NEXT: .short 25
12; CHECK-NEXT: .short 25
13; CHECK-NEXT: .short 25
14; CHECK-NEXT: .short 25
15; CHECK-LABEL: loop_no_const_reload:
16; CHECK: psubusw
17
18; Constant is no longer clobbered so no need to reload it in the loop.
19
20; CHECK-NOT: movdqa {{%xmm[0-9]+}}, {{%xmm[0-9]+}}
21
22entry:
23  %cmp9 = icmp eq i32 %n, 0
24  br i1 %cmp9, label %for.end, label %for.body
25
26for.body:                                         ; preds = %for.body, %entry
27  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
28  %arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
29  %arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
30  %0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
31  %cmp.i.i = icmp ult <8 x i16> %0, <i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26>
32  %sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
33  %1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
34  %arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
35  store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
36  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
37  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
38  %exitcond = icmp eq i32 %lftr.wideiv, %n
39  br i1 %exitcond, label %for.end, label %for.body
40
41for.end:                                          ; preds = %for.body, %entry
42  ret void
43}
44
45; Be careful if decrementing the constant would undeflow.
46
47define void @loop_const_folding_underflow(<2 x i64>*  %in, <2 x i64>* %out, i32 %n) {
48; CHECK-NOT: .short 25
49; CHECK-LABEL: loop_const_folding_underflow:
50; CHECK-NOT: psubusw
51entry:
52  %cmp9 = icmp eq i32 %n, 0
53  br i1 %cmp9, label %for.end, label %for.body
54
55for.body:                                         ; preds = %for.body, %entry
56  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
57  %arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
58  %arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
59  %0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
60  %cmp.i.i = icmp ult <8 x i16> %0, <i16 0, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26, i16 26>
61  %sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
62  %1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
63  %arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
64  store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
65  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
66  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
67  %exitcond = icmp eq i32 %lftr.wideiv, %n
68  br i1 %exitcond, label %for.end, label %for.body
69
70for.end:                                          ; preds = %for.body, %entry
71  ret void
72}
73
74; Test for PSUBUSB
75
76define <16 x i8> @test_ult_byte(<16 x i8> %a) {
77; CHECK: .space 16,10
78; CHECK-LABEL: test_ult_byte:
79; CHECK: psubus
80entry:
81  %icmp = icmp ult <16 x i8> %a, <i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11, i8 11>
82  %sext = sext <16 x i1> %icmp to <16 x i8>
83  ret <16 x i8> %sext
84}
85
86; Only do this when we can turn the comparison into a setule.  I.e. not for
87; register operands.
88
89define <8 x i16> @test_ult_register(<8 x i16> %a, <8 x i16> %b) {
90; CHECK-LABEL: test_ult_register:
91; CHECK-NOT: psubus
92entry:
93  %icmp = icmp ult <8 x i16> %a, %b
94  %sext = sext <8 x i1> %icmp to <8 x i16>
95  ret <8 x i16> %sext
96}
97