1; RUN: opt -S -instcombine < %s | FileCheck %s
2
3declare void @v4float_user(<4 x float>) #0
4
5define float @extract_one_select(<4 x float> %a, <4 x float> %b, i32 %c) #0 {
6; CHECK-LABEL: @extract_one_select(
7; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 %c, 0
8; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], <4 x float> %b, <4 x float> %a
9; CHECK-NEXT:    [[EXTRACT:%.*]] = extractelement <4 x float> [[SEL]], i32 2
10; CHECK-NEXT:    ret float [[EXTRACT]]
11;
12  %cmp = icmp ne i32 %c, 0
13  %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b
14  %extract = extractelement <4 x float> %sel, i32 2
15  ret float %extract
16}
17
18; Multiple extractelements
19define <2 x float> @extract_two_select(<4 x float> %a, <4 x float> %b, i32 %c) #0 {
20; CHECK-LABEL: @extract_two_select(
21; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 %c, 0
22; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], <4 x float> %b, <4 x float> %a
23; CHECK-NEXT:    [[BUILD2:%.*]] = shufflevector <4 x float> [[SEL]], <4 x float> undef, <2 x i32> <i32 1, i32 2>
24; CHECK-NEXT:    ret <2 x float> [[BUILD2]]
25;
26  %cmp = icmp ne i32 %c, 0
27  %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b
28  %extract1 = extractelement <4 x float> %sel, i32 1
29  %extract2 = extractelement <4 x float> %sel, i32 2
30  %build1 = insertelement <2 x float> undef, float %extract1, i32 0
31  %build2 = insertelement <2 x float> %build1, float %extract2, i32 1
32  ret <2 x float> %build2
33}
34
35; Select has an extra non-extractelement user, don't change it
36define float @extract_one_select_user(<4 x float> %a, <4 x float> %b, i32 %c) #0 {
37; CHECK-LABEL: @extract_one_select_user(
38; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 %c, 0
39; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], <4 x float> %b, <4 x float> %a
40; CHECK-NEXT:    [[EXTRACT:%.*]] = extractelement <4 x float> [[SEL]], i32 2
41; CHECK-NEXT:    call void @v4float_user(<4 x float> [[SEL]])
42; CHECK-NEXT:    ret float [[EXTRACT]]
43;
44  %cmp = icmp ne i32 %c, 0
45  %sel = select i1 %cmp, <4 x float> %a, <4 x float> %b
46  %extract = extractelement <4 x float> %sel, i32 2
47  call void @v4float_user(<4 x float> %sel)
48  ret float %extract
49}
50
51define float @extract_one_vselect_user(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 {
52; CHECK-LABEL: @extract_one_vselect_user(
53; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <4 x i32> %c, zeroinitializer
54; CHECK-NEXT:    [[SEL:%.*]] = select <4 x i1> [[CMP]], <4 x float> %b, <4 x float> %a
55; CHECK-NEXT:    [[EXTRACT:%.*]] = extractelement <4 x float> [[SEL]], i32 2
56; CHECK-NEXT:    call void @v4float_user(<4 x float> [[SEL]])
57; CHECK-NEXT:    ret float [[EXTRACT]]
58;
59  %cmp = icmp ne <4 x i32> %c, zeroinitializer
60  %sel = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b
61  %extract = extractelement <4 x float> %sel, i32 2
62  call void @v4float_user(<4 x float> %sel)
63  ret float %extract
64}
65
66; Do not convert the vector select into a scalar select. That would increase
67; the instruction count and potentially obfuscate a vector min/max idiom.
68
69define float @extract_one_vselect(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 {
70; CHECK-LABEL: @extract_one_vselect(
71; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <4 x i32> %c, zeroinitializer
72; CHECK-NEXT:    [[SELECT:%.*]] = select <4 x i1> [[CMP]], <4 x float> %b, <4 x float> %a
73; CHECK-NEXT:    [[EXTRACT:%.*]] = extractelement <4 x float> [[SELECT]], i32 0
74; CHECK-NEXT:    ret float [[EXTRACT]]
75;
76  %cmp = icmp ne <4 x i32> %c, zeroinitializer
77  %select = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b
78  %extract = extractelement <4 x float> %select, i32 0
79  ret float %extract
80}
81
82; Multiple extractelements from a vector select
83define <2 x float> @extract_two_vselect(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 {
84; CHECK-LABEL: @extract_two_vselect(
85; CHECK-NEXT:    [[CMP:%.*]] = icmp eq <4 x i32> %c, zeroinitializer
86; CHECK-NEXT:    [[SEL:%.*]] = select <4 x i1> [[CMP]], <4 x float> %b, <4 x float> %a
87; CHECK-NEXT:    [[BUILD2:%.*]] = shufflevector <4 x float> [[SEL]], <4 x float> undef, <2 x i32> <i32 1, i32 2>
88; CHECK-NEXT:    ret <2 x float> [[BUILD2]]
89;
90  %cmp = icmp ne <4 x i32> %c, zeroinitializer
91  %sel = select <4 x i1> %cmp, <4 x float> %a, <4 x float> %b
92  %extract1 = extractelement <4 x float> %sel, i32 1
93  %extract2 = extractelement <4 x float> %sel, i32 2
94  %build1 = insertelement <2 x float> undef, float %extract1, i32 0
95  %build2 = insertelement <2 x float> %build1, float %extract2, i32 1
96  ret <2 x float> %build2
97}
98
99; The vector selects are not decomposed into scalar selects because that would increase
100; the instruction count. Extract+insert is converted to non-lane-crossing shuffles.
101; Test multiple extractelements
102define <4 x float> @simple_vector_select(<4 x float> %a, <4 x float> %b, <4 x i32> %c) #0 {
103; CHECK-LABEL: @simple_vector_select(
104; CHECK-NEXT:  entry:
105; CHECK-NEXT:    [[TMP0:%.*]] = extractelement <4 x i32> %c, i32 0
106; CHECK-NEXT:    [[TOBOOL:%.*]] = icmp eq i32 [[TMP0]], 0
107; CHECK-NEXT:    [[A_SINK:%.*]] = select i1 [[TOBOOL]], <4 x float> %b, <4 x float> %a
108; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <4 x i32> %c, i32 1
109; CHECK-NEXT:    [[TOBOOL1:%.*]] = icmp eq i32 [[TMP1]], 0
110; CHECK-NEXT:    [[A_SINK1:%.*]] = select i1 [[TOBOOL1]], <4 x float> %b, <4 x float> %a
111; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[A_SINK]], <4 x float> [[A_SINK1]], <4 x i32> <i32 0, i32 5, i32 undef, i32 undef>
112; CHECK-NEXT:    [[TMP3:%.*]] = extractelement <4 x i32> %c, i32 2
113; CHECK-NEXT:    [[TOBOOL6:%.*]] = icmp eq i32 [[TMP3]], 0
114; CHECK-NEXT:    [[A_SINK2:%.*]] = select i1 [[TOBOOL6]], <4 x float> %b, <4 x float> %a
115; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <4 x float> [[TMP2]], <4 x float> [[A_SINK2]], <4 x i32> <i32 0, i32 1, i32 6, i32 undef>
116; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i32> %c, i32 3
117; CHECK-NEXT:    [[TOBOOL11:%.*]] = icmp eq i32 [[TMP5]], 0
118; CHECK-NEXT:    [[A_SINK3:%.*]] = select i1 [[TOBOOL11]], <4 x float> %b, <4 x float> %a
119; CHECK-NEXT:    [[TMP6:%.*]] = shufflevector <4 x float> [[TMP4]], <4 x float> [[A_SINK3]], <4 x i32> <i32 0, i32 1, i32 2, i32 7>
120; CHECK-NEXT:    ret <4 x float> [[TMP6]]
121;
122entry:
123  %0 = extractelement <4 x i32> %c, i32 0
124  %tobool = icmp ne i32 %0, 0
125  %a.sink = select i1 %tobool, <4 x float> %a, <4 x float> %b
126  %1 = extractelement <4 x float> %a.sink, i32 0
127  %2 = insertelement <4 x float> undef, float %1, i32 0
128  %3 = extractelement <4 x i32> %c, i32 1
129  %tobool1 = icmp ne i32 %3, 0
130  %a.sink1 = select i1 %tobool1, <4 x float> %a, <4 x float> %b
131  %4 = extractelement <4 x float> %a.sink1, i32 1
132  %5 = insertelement <4 x float> %2, float %4, i32 1
133  %6 = extractelement <4 x i32> %c, i32 2
134  %tobool6 = icmp ne i32 %6, 0
135  %a.sink2 = select i1 %tobool6, <4 x float> %a, <4 x float> %b
136  %7 = extractelement <4 x float> %a.sink2, i32 2
137  %8 = insertelement <4 x float> %5, float %7, i32 2
138  %9 = extractelement <4 x i32> %c, i32 3
139  %tobool11 = icmp ne i32 %9, 0
140  %a.sink3 = select i1 %tobool11, <4 x float> %a, <4 x float> %b
141  %10 = extractelement <4 x float> %a.sink3, i32 3
142  %11 = insertelement <4 x float> %8, float %10, i32 3
143  ret <4 x float> %11
144}
145
146attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf"="true" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
147