| blob | b8cc896b61d4126091bd94d01b02c06efbb8c59f |
1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -select-optimize -mtriple=aarch64-linux-gnu -mcpu=neoverse-v2 -S < %s | FileCheck %s
3 ; RUN: opt -passes='require<profile-summary>,function(select-optimize)' -mtriple=aarch64-linux-gnu -mcpu=neoverse-v2 -S < %s | FileCheck %s
5 define i32 @test(ptr nocapture noundef readnone %x, i32 noundef %iters) {
6 ; CHECK-LABEL: @test(
7 ; CHECK-NEXT: entry:
8 ; CHECK-NEXT: [[CMP3:%.*]] = icmp sgt i32 [[ITERS:%.*]], 0
9 ; CHECK-NEXT: br i1 [[CMP3]], label [[FOR_BODY_I_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
10 ; CHECK: for.body.i.preheader:
11 ; CHECK-NEXT: [[I_05:%.*]] = phi i32 [ [[INC:%.*]], [[INNER_LOOP_010_EXIT:%.*]] ], [ 0, [[ENTRY:%.*]] ]
12 ; CHECK-NEXT: [[RES_04:%.*]] = phi i32 [ [[OR:%.*]], [[INNER_LOOP_010_EXIT]] ], [ 0, [[ENTRY]] ]
13 ; CHECK-NEXT: br label [[FOR_BODY_I:%.*]]
14 ; CHECK: for.cond.cleanup:
15 ; CHECK-NEXT: [[RES_0_LCSSA:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[OR]], [[INNER_LOOP_010_EXIT]] ]
16 ; CHECK-NEXT: ret i32 [[RES_0_LCSSA]]
17 ; CHECK: for.body.i:
18 ; CHECK-NEXT: [[INDVARS_IV_I:%.*]] = phi i64 [ 0, [[FOR_BODY_I_PREHEADER]] ], [ [[INDVARS_IV_NEXT_I_4:%.*]], [[FOR_BODY_I]] ]
19 ; CHECK-NEXT: [[ALL_0_OFF010_I:%.*]] = phi i1 [ true, [[FOR_BODY_I_PREHEADER]] ], [ [[ALL_0_OFF0__I_4:%.*]], [[FOR_BODY_I]] ]
20 ; CHECK-NEXT: [[ANY_0_OFF09_I:%.*]] = phi i1 [ false, [[FOR_BODY_I_PREHEADER]] ], [ [[DOTANY_0_OFF0_I_4:%.*]], [[FOR_BODY_I]] ]
21 ; CHECK-NEXT: [[ARRAYIDX_I:%.*]] = getelementptr inbounds float, ptr [[X:%.*]], i64 [[INDVARS_IV_I]]
22 ; CHECK-NEXT: [[TMP0:%.*]] = load float, ptr [[ARRAYIDX_I]], align 4
23 ; CHECK-NEXT: [[CMP1_I:%.*]] = fcmp fast olt float [[TMP0]], 0.000000e+00
24 ; CHECK-NEXT: [[INDVARS_IV_NEXT_I:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 1
25 ; CHECK-NEXT: [[ARRAYIDX_I_1:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I]]
26 ; CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[ARRAYIDX_I_1]], align 4
27 ; CHECK-NEXT: [[CMP1_I_1:%.*]] = fcmp fast olt float [[TMP1]], 0.000000e+00
28 ; CHECK-NEXT: [[INDVARS_IV_NEXT_I_1:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 2
29 ; CHECK-NEXT: [[ARRAYIDX_I_2:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_1]]
30 ; CHECK-NEXT: [[TMP2:%.*]] = load float, ptr [[ARRAYIDX_I_2]], align 4
31 ; CHECK-NEXT: [[CMP1_I_2:%.*]] = fcmp fast olt float [[TMP2]], 0.000000e+00
32 ; CHECK-NEXT: [[INDVARS_IV_NEXT_I_2:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 3
33 ; CHECK-NEXT: [[ARRAYIDX_I_3:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_2]]
34 ; CHECK-NEXT: [[TMP3:%.*]] = load float, ptr [[ARRAYIDX_I_3]], align 4
35 ; CHECK-NEXT: [[CMP1_I_3:%.*]] = fcmp fast olt float [[TMP3]], 0.000000e+00
36 ; CHECK-NEXT: [[INDVARS_IV_NEXT_I_3:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 4
37 ; CHECK-NEXT: [[ARRAYIDX_I_4:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_3]]
38 ; CHECK-NEXT: [[TMP4:%.*]] = load float, ptr [[ARRAYIDX_I_4]], align 4
39 ; CHECK-NEXT: [[CMP1_I_4:%.*]] = fcmp fast olt float [[TMP4]], 0.000000e+00
40 ; CHECK-NEXT: [[TMP5:%.*]] = select i1 [[CMP1_I_4]], i1 true, i1 [[CMP1_I_3]]
41 ; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i1 true, i1 [[CMP1_I_2]]
42 ; CHECK-NEXT: [[TMP7:%.*]] = select i1 [[TMP6]], i1 true, i1 [[CMP1_I_1]]
43 ; CHECK-NEXT: [[TMP8:%.*]] = select i1 [[TMP7]], i1 true, i1 [[CMP1_I]]
44 ; CHECK-NEXT: [[DOTANY_0_OFF0_I_4]] = select i1 [[TMP8]], i1 true, i1 [[ANY_0_OFF09_I]]
45 ; CHECK-NEXT: [[TMP9:%.*]] = select i1 [[CMP1_I_4]], i1 [[CMP1_I_3]], i1 false
46 ; CHECK-NEXT: [[TMP10:%.*]] = select i1 [[TMP9]], i1 [[CMP1_I_2]], i1 false
47 ; CHECK-NEXT: [[TMP11:%.*]] = select i1 [[TMP10]], i1 [[CMP1_I_1]], i1 false
48 ; CHECK-NEXT: [[TMP12:%.*]] = select i1 [[TMP11]], i1 [[CMP1_I]], i1 false
49 ; CHECK-NEXT: [[ALL_0_OFF0__I_4]] = select i1 [[TMP12]], i1 [[ALL_0_OFF010_I]], i1 false
50 ; CHECK-NEXT: [[INDVARS_IV_NEXT_I_4]] = add nuw nsw i64 [[INDVARS_IV_I]], 5
51 ; CHECK-NEXT: [[EXITCOND_NOT_I_4:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_I_4]], 10000
52 ; CHECK-NEXT: br i1 [[EXITCOND_NOT_I_4]], label [[INNER_LOOP_010_EXIT]], label [[FOR_BODY_I]]
53 ; CHECK: inner_loop_010.exit:
54 ; CHECK-NEXT: [[TMP13:%.*]] = select i1 [[DOTANY_0_OFF0_I_4]], i32 2, i32 3
55 ; CHECK-NEXT: [[TMP14:%.*]] = select i1 [[ALL_0_OFF0__I_4]], i32 1, i32 [[TMP13]]
56 ; CHECK-NEXT: [[OR]] = or i32 [[TMP14]], [[RES_04]]
57 ; CHECK-NEXT: [[INC]] = add nuw nsw i32 [[I_05]], 1
58 ; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[ITERS]]
59 ; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY_I_PREHEADER]]
60 ;
61 entry:
62 %cmp3 = icmp sgt i32 %iters, 0
63 br i1 %cmp3, label %for.body.i.preheader, label %for.cond.cleanup
65 for.body.i.preheader: ; preds = %entry, %inner_loop_010.exit
66 %i.05 = phi i32 [ %inc, %inner_loop_010.exit ], [ 0, %entry ]
67 %res.04 = phi i32 [ %or, %inner_loop_010.exit ], [ 0, %entry ]
68 br label %for.body.i
70 for.cond.cleanup: ; preds = %inner_loop_010.exit, %entry
71 %res.0.lcssa = phi i32 [ 0, %entry ], [ %or, %inner_loop_010.exit ]
72 ret i32 %res.0.lcssa
74 for.body.i: ; preds = %for.body.i, %for.body.i.preheader
75 %indvars.iv.i = phi i64 [ 0, %for.body.i.preheader ], [ %indvars.iv.next.i.4, %for.body.i ]
76 %all.0.off010.i = phi i1 [ true, %for.body.i.preheader ], [ %all.0.off0..i.4, %for.body.i ]
77 %any.0.off09.i = phi i1 [ false, %for.body.i.preheader ], [ %.any.0.off0.i.4, %for.body.i ]
78 %arrayidx.i = getelementptr inbounds float, ptr %x, i64 %indvars.iv.i
79 %0 = load float, ptr %arrayidx.i, align 4
80 %cmp1.i = fcmp fast olt float %0, 0.000000e+00
81 %indvars.iv.next.i = add nuw nsw i64 %indvars.iv.i, 1
82 %arrayidx.i.1 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i
83 %1 = load float, ptr %arrayidx.i.1, align 4
84 %cmp1.i.1 = fcmp fast olt float %1, 0.000000e+00
85 %indvars.iv.next.i.1 = add nuw nsw i64 %indvars.iv.i, 2
86 %arrayidx.i.2 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.1
87 %2 = load float, ptr %arrayidx.i.2, align 4
88 %cmp1.i.2 = fcmp fast olt float %2, 0.000000e+00
89 %indvars.iv.next.i.2 = add nuw nsw i64 %indvars.iv.i, 3
90 %arrayidx.i.3 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.2
91 %3 = load float, ptr %arrayidx.i.3, align 4
92 %cmp1.i.3 = fcmp fast olt float %3, 0.000000e+00
93 %indvars.iv.next.i.3 = add nuw nsw i64 %indvars.iv.i, 4
94 %arrayidx.i.4 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.3
95 %4 = load float, ptr %arrayidx.i.4, align 4
96 %cmp1.i.4 = fcmp fast olt float %4, 0.000000e+00
97 %5 = select i1 %cmp1.i.4, i1 true, i1 %cmp1.i.3
98 %6 = select i1 %5, i1 true, i1 %cmp1.i.2
99 %7 = select i1 %6, i1 true, i1 %cmp1.i.1
100 %8 = select i1 %7, i1 true, i1 %cmp1.i
101 %.any.0.off0.i.4 = select i1 %8, i1 true, i1 %any.0.off09.i
102 %9 = select i1 %cmp1.i.4, i1 %cmp1.i.3, i1 false
103 %10 = select i1 %9, i1 %cmp1.i.2, i1 false
104 %11 = select i1 %10, i1 %cmp1.i.1, i1 false
105 %12 = select i1 %11, i1 %cmp1.i, i1 false
106 %all.0.off0..i.4 = select i1 %12, i1 %all.0.off010.i, i1 false
107 %indvars.iv.next.i.4 = add nuw nsw i64 %indvars.iv.i, 5
108 %exitcond.not.i.4 = icmp eq i64 %indvars.iv.next.i.4, 10000
109 br i1 %exitcond.not.i.4, label %inner_loop_010.exit, label %for.body.i
111 inner_loop_010.exit: ; preds = %for.body.i
112 %13 = select i1 %.any.0.off0.i.4, i32 2, i32 3
113 %14 = select i1 %all.0.off0..i.4, i32 1, i32 %13
114 %or = or i32 %14, %res.04
115 %inc = add nuw nsw i32 %i.05, 1
116 %exitcond.not = icmp eq i32 %inc, %iters
117 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body.i.preheader
118 }