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
4 ; For a setult against a constant, turn it into a setule and lower via psubusw.
6 define void @loop_no_const_reload(<2 x i64>* %in, <2 x i64>* %out, i32 %n) {
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:
18 ; Constant is no longer clobbered so no need to reload it in the loop.
20 ; CHECK-NOT: movdqa {{%xmm[0-9]+}}, {{%xmm[0-9]+}}
22 %cmp9 = icmp eq i32 %n, 0
23 br i1 %cmp9, label %for.end, label %for.body
25 for.body: ; preds = %for.body, %entry
26 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
27 %arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
28 %arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
29 %0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
30 %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>
31 %sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
32 %1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
33 %arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
34 store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
35 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
36 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
37 %exitcond = icmp eq i32 %lftr.wideiv, %n
38 br i1 %exitcond, label %for.end, label %for.body
40 for.end: ; preds = %for.body, %entry
44 ; Be careful if decrementing the constant would undeflow.
46 define void @loop_const_folding_underflow(<2 x i64>* %in, <2 x i64>* %out, i32 %n) {
47 ; CHECK-NOT: .short 25
48 ; CHECK-LABEL: loop_const_folding_underflow:
51 %cmp9 = icmp eq i32 %n, 0
52 br i1 %cmp9, label %for.end, label %for.body
54 for.body: ; preds = %for.body, %entry
55 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
56 %arrayidx1 = getelementptr inbounds <2 x i64>* %in, i64 %indvars.iv
57 %arrayidx1.val = load <2 x i64>* %arrayidx1, align 16
58 %0 = bitcast <2 x i64> %arrayidx1.val to <8 x i16>
59 %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>
60 %sext.i.i = sext <8 x i1> %cmp.i.i to <8 x i16>
61 %1 = bitcast <8 x i16> %sext.i.i to <2 x i64>
62 %arrayidx5 = getelementptr inbounds <2 x i64>* %out, i64 %indvars.iv
63 store <2 x i64> %1, <2 x i64>* %arrayidx5, align 16
64 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
65 %lftr.wideiv = trunc i64 %indvars.iv.next to i32
66 %exitcond = icmp eq i32 %lftr.wideiv, %n
67 br i1 %exitcond, label %for.end, label %for.body
69 for.end: ; preds = %for.body, %entry
75 define <16 x i8> @test_ult_byte(<16 x i8> %a) {
77 ; CHECK-LABEL: test_ult_byte:
80 %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>
81 %sext = sext <16 x i1> %icmp to <16 x i8>
85 ; Only do this when we can turn the comparison into a setule. I.e. not for
88 define <8 x i16> @test_ult_register(<8 x i16> %a, <8 x i16> %b) {
89 ; CHECK-LABEL: test_ult_register:
92 %icmp = icmp ult <8 x i16> %a, %b
93 %sext = sext <8 x i1> %icmp to <8 x i16>