11; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2- ; RUN: opt -passes='loop-vectorize,loop-unroll ' -force-vector-width=4 -S < %s | FileCheck %s
2+ ; RUN: opt -passes='loop-vectorize' -force-vector-width=4 -S < %s | FileCheck %s
33
44; Test if the follow-up metadata for loops works fine. The original code is
5- ; something like below. In this case, unrolling should be applied after
6- ; vectorization .
5+ ; something like below. In this case, the unrolling metadata from the original
6+ ; loop should be applied to the vector loop as well .
77;
88; void f(float *a, float x) {
99; #pragma clang loop vectorize(enable) unroll_count(8)
1212; }
1313; }
1414;
15- define void @f (ptr noundef captures(none) %a , float noundef %x ) {
16- ; CHECK-LABEL: define void @f (
15+ define void @scalar_loop_dead (ptr noundef captures(none) %a , float noundef %x ) {
16+ ; CHECK-LABEL: define void @scalar_loop_dead (
1717; CHECK-SAME: ptr noundef captures(none) [[A:%.*]], float noundef [[X:%.*]]) {
1818; CHECK-NEXT: [[ENTRY:.*:]]
1919; CHECK-NEXT: br label %[[VECTOR_PH:.*]]
@@ -22,93 +22,115 @@ define void @f(ptr noundef captures(none) %a, float noundef %x) {
2222; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT]], <4 x float> poison, <4 x i32> zeroinitializer
2323; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
2424; CHECK: [[VECTOR_BODY]]:
25- ; CHECK-NEXT: [[INDEX_NEXT_6:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
26- ; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_6]]
27- ; CHECK-NEXT: [[WIDE_LOAD_7:%.*]] = load <4 x float>, ptr [[TMP14]], align 4
28- ; CHECK-NEXT: [[TMP15:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_7]]
29- ; CHECK-NEXT: store <4 x float> [[TMP15]], ptr [[TMP14]], align 4
30- ; CHECK-NEXT: [[INDEX_NEXT1:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 4
31- ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT1]]
32- ; CHECK-NEXT: [[WIDE_LOAD_1:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
33- ; CHECK-NEXT: [[TMP3:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_1]]
34- ; CHECK-NEXT: store <4 x float> [[TMP3]], ptr [[TMP2]], align 4
35- ; CHECK-NEXT: [[INDEX_NEXT_1:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 8
36- ; CHECK-NEXT: [[TMP16:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_1]]
37- ; CHECK-NEXT: [[WIDE_LOAD_2:%.*]] = load <4 x float>, ptr [[TMP16]], align 4
38- ; CHECK-NEXT: [[TMP5:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_2]]
39- ; CHECK-NEXT: store <4 x float> [[TMP5]], ptr [[TMP16]], align 4
40- ; CHECK-NEXT: [[INDEX_NEXT_2:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 12
41- ; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_2]]
42- ; CHECK-NEXT: [[WIDE_LOAD_3:%.*]] = load <4 x float>, ptr [[TMP6]], align 4
43- ; CHECK-NEXT: [[TMP7:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_3]]
44- ; CHECK-NEXT: store <4 x float> [[TMP7]], ptr [[TMP6]], align 4
45- ; CHECK-NEXT: [[INDEX_NEXT_3:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 16
46- ; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_3]]
47- ; CHECK-NEXT: [[WIDE_LOAD_4:%.*]] = load <4 x float>, ptr [[TMP8]], align 4
48- ; CHECK-NEXT: [[TMP9:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_4]]
49- ; CHECK-NEXT: store <4 x float> [[TMP9]], ptr [[TMP8]], align 4
50- ; CHECK-NEXT: [[INDEX_NEXT_4:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 20
51- ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_4]]
52- ; CHECK-NEXT: [[WIDE_LOAD_5:%.*]] = load <4 x float>, ptr [[TMP10]], align 4
53- ; CHECK-NEXT: [[TMP11:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_5]]
54- ; CHECK-NEXT: store <4 x float> [[TMP11]], ptr [[TMP10]], align 4
55- ; CHECK-NEXT: [[INDEX_NEXT_5:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 24
56- ; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_5]]
57- ; CHECK-NEXT: [[WIDE_LOAD_6:%.*]] = load <4 x float>, ptr [[TMP12]], align 4
58- ; CHECK-NEXT: [[TMP13:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_6]]
59- ; CHECK-NEXT: store <4 x float> [[TMP13]], ptr [[TMP12]], align 4
60- ; CHECK-NEXT: [[INDEX_NEXT_7:%.*]] = add nuw nsw i64 [[INDEX_NEXT_6]], 28
61- ; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX_NEXT_7]]
62- ; CHECK-NEXT: [[WIDE_LOAD_8:%.*]] = load <4 x float>, ptr [[TMP17]], align 4
63- ; CHECK-NEXT: [[TMP18:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD_8]]
64- ; CHECK-NEXT: store <4 x float> [[TMP18]], ptr [[TMP17]], align 4
65- ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw nsw i64 [[INDEX_NEXT_6]], 32
66- ; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
67- ; CHECK-NEXT: br i1 [[TMP4]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
25+ ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
26+ ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX]]
27+ ; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP0]], align 4
28+ ; CHECK-NEXT: [[TMP1:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]
29+ ; CHECK-NEXT: store <4 x float> [[TMP1]], ptr [[TMP0]], align 4
30+ ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
31+ ; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
32+ ; CHECK-NEXT: br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
6833; CHECK: [[MIDDLE_BLOCK]]:
6934; CHECK-NEXT: br label %[[EXIT:.*]]
7035; CHECK: [[SCALAR_PH:.*]]:
71- ; CHECK-NEXT: br label %[[FOR_BODY :.*]]
72- ; CHECK: [[FOR_BODY ]]:
73- ; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[FOR_BODY ]] ]
36+ ; CHECK-NEXT: br label %[[LOOP :.*]]
37+ ; CHECK: [[LOOP ]]:
38+ ; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP ]] ]
7439; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV]]
7540; CHECK-NEXT: [[LOAD:%.*]] = load float, ptr [[ARRAYIDX]], align 4
7641; CHECK-NEXT: [[MUL:%.*]] = fmul float [[X]], [[LOAD]]
7742; CHECK-NEXT: store float [[MUL]], ptr [[ARRAYIDX]], align 4
7843; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
7944; CHECK-NEXT: [[COMP:%.*]] = icmp eq i64 [[IV_NEXT]], 1024
80- ; CHECK-NEXT: br i1 [[COMP]], label %[[EXIT]], label %[[FOR_BODY ]], !llvm.loop [[LOOP3 :![0-9]+]]
45+ ; CHECK-NEXT: br i1 [[COMP]], label %[[EXIT]], label %[[LOOP ]], !llvm.loop [[LOOP4 :![0-9]+]]
8146; CHECK: [[EXIT]]:
8247; CHECK-NEXT: ret void
8348;
8449entry:
85- br label %for.body
50+ br label %loop
8651
87- for.body :
88- %iv = phi i64 [ 0 , %entry ], [ %iv.next , %for.body ]
52+ loop :
53+ %iv = phi i64 [ 0 , %entry ], [ %iv.next , %loop ]
8954 %arrayidx = getelementptr inbounds nuw float , ptr %a , i64 %iv
9055 %load = load float , ptr %arrayidx , align 4
9156 %mul = fmul float %x , %load
9257 store float %mul , ptr %arrayidx , align 4
9358 %iv.next = add nuw nsw i64 %iv , 1
9459 %comp = icmp eq i64 %iv.next , 1024
95- br i1 %comp , label %exit , label %for.body , !llvm.loop !0
60+ br i1 %comp , label %exit , label %loop , !llvm.loop !0
9661
9762exit:
9863 ret void
9964}
10065
66+ define void @scalar_loop_live (ptr noundef captures(none) %a , float noundef %x , i64 %n ) {
67+ ; CHECK-LABEL: define void @scalar_loop_live(
68+ ; CHECK-SAME: ptr noundef captures(none) [[A:%.*]], float noundef [[X:%.*]], i64 [[N:%.*]]) {
69+ ; CHECK-NEXT: [[ENTRY:.*]]:
70+ ; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 4
71+ ; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
72+ ; CHECK: [[VECTOR_PH]]:
73+ ; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
74+ ; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
75+ ; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x float> poison, float [[X]], i64 0
76+ ; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT]], <4 x float> poison, <4 x i32> zeroinitializer
77+ ; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
78+ ; CHECK: [[VECTOR_BODY]]:
79+ ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
80+ ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[INDEX]]
81+ ; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP0]], align 4
82+ ; CHECK-NEXT: [[TMP1:%.*]] = fmul <4 x float> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]
83+ ; CHECK-NEXT: store <4 x float> [[TMP1]], ptr [[TMP0]], align 4
84+ ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
85+ ; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
86+ ; CHECK-NEXT: br i1 [[TMP2]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
87+ ; CHECK: [[MIDDLE_BLOCK]]:
88+ ; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
89+ ; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
90+ ; CHECK: [[SCALAR_PH]]:
91+ ; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
92+ ; CHECK-NEXT: br label %[[LOOP:.*]]
93+ ; CHECK: [[LOOP]]:
94+ ; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
95+ ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw float, ptr [[A]], i64 [[IV]]
96+ ; CHECK-NEXT: [[LOAD:%.*]] = load float, ptr [[ARRAYIDX]], align 4
97+ ; CHECK-NEXT: [[MUL:%.*]] = fmul float [[X]], [[LOAD]]
98+ ; CHECK-NEXT: store float [[MUL]], ptr [[ARRAYIDX]], align 4
99+ ; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
100+ ; CHECK-NEXT: [[COMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
101+ ; CHECK-NEXT: br i1 [[COMP]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP8:![0-9]+]]
102+ ; CHECK: [[EXIT]]:
103+ ; CHECK-NEXT: ret void
104+ ;
105+ entry:
106+ br label %loop
107+
108+ loop:
109+ %iv = phi i64 [ 0 , %entry ], [ %iv.next , %loop ]
110+ %arrayidx = getelementptr inbounds nuw float , ptr %a , i64 %iv
111+ %load = load float , ptr %arrayidx , align 4
112+ %mul = fmul float %x , %load
113+ store float %mul , ptr %arrayidx , align 4
114+ %iv.next = add nuw nsw i64 %iv , 1
115+ %comp = icmp eq i64 %iv.next , %n
116+ br i1 %comp , label %exit , label %loop , !llvm.loop !0
117+
118+ exit:
119+ ret void
120+ }
101121!0 = distinct !{!0 , !1 , !2 }
102122!1 = !{!"llvm.loop.vectorize.enable" , i1 true }
103123!2 = !{!"llvm.loop.vectorize.followup_all" , !3 , !4 }
104124!3 = !{!"llvm.loop.isvectorized" }
105125!4 = !{!"llvm.loop.unroll.count" , i32 8 }
106126;.
107- ; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
127+ ; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]], [[META3:![0-9]+]] }
108128; CHECK: [[META1]] = !{!"llvm.loop.isvectorized"}
109- ; CHECK: [[META2]] = !{!"llvm.loop.unroll.disable"}
110- ; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]], [[META5:![0-9]+]]}
111- ; CHECK: [[META4]] = !{!"llvm.loop.vectorize.enable", i1 true}
112- ; CHECK: [[META5]] = !{!"llvm.loop.vectorize.followup_all", [[META1]], [[META6:![0-9]+]]}
113- ; CHECK: [[META6]] = !{!"llvm.loop.unroll.count", i32 8}
129+ ; CHECK: [[META2]] = !{!"llvm.loop.unroll.count", i32 8}
130+ ; CHECK: [[META3]] = !{!"llvm.loop.unroll.runtime.disable"}
131+ ; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]], [[META6:![0-9]+]]}
132+ ; CHECK: [[META5]] = !{!"llvm.loop.vectorize.enable", i1 true}
133+ ; CHECK: [[META6]] = !{!"llvm.loop.vectorize.followup_all", [[META1]], [[META2]]}
134+ ; CHECK: [[LOOP7]] = distinct !{[[LOOP7]], [[META1]], [[META2]], [[META3]]}
135+ ; CHECK: [[LOOP8]] = distinct !{[[LOOP8]], [[META1]], [[META2]]}
114136;.
0 commit comments