1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -S -loop-vectorize -prefer-predicate-over-epilogue=predicate-else-scalar-epilogue < %s | FileCheck %s
3 ; RUN: opt -S -loop-vectorize < %s | FileCheck %s
5 ; This tests should produce the same result as with default options, and when tail folding
6 ; is preferred, because the vectorizer can't fold the tail by masking (due to an
7 ; outside user of %incdec.ptr in %end) and should fallback to a scalar epilogue.
9 ; The first test (@basic_loop) simply relies on the command-line switches.
10 ; The second test (@metadata) specificies its tail-folding preference via metadata.
11 ; Both tests should always generate a scalar epilogue.
13 target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
15 define void @basic_loop(i8* nocapture readonly %ptr, i32 %size, i8** %pos) {
16 ; CHECK-LABEL: @basic_loop(
18 ; CHECK-NEXT: [[PTR0:%.*]] = load i8*, i8** [[POS:%.*]], align 4
19 ; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[SIZE:%.*]], 4
20 ; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
22 ; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[SIZE]], 4
23 ; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[SIZE]], [[N_MOD_VF]]
24 ; CHECK-NEXT: [[IND_END:%.*]] = sub i32 [[SIZE]], [[N_VEC]]
25 ; CHECK-NEXT: [[IND_END2:%.*]] = getelementptr i8, i8* [[PTR:%.*]], i32 [[N_VEC]]
26 ; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
28 ; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
29 ; CHECK-NEXT: [[OFFSET_IDX:%.*]] = sub i32 [[SIZE]], [[INDEX]]
30 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[OFFSET_IDX]], 0
31 ; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 0
32 ; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, i8* [[PTR]], i32 [[TMP1]]
33 ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, i8* [[NEXT_GEP]], i32 1
34 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, i8* [[TMP2]], i32 0
35 ; CHECK-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to <4 x i8>*
36 ; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP4]], align 1
37 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, i8* [[NEXT_GEP]], i32 0
38 ; CHECK-NEXT: [[TMP6:%.*]] = bitcast i8* [[TMP5]] to <4 x i8>*
39 ; CHECK-NEXT: store <4 x i8> [[WIDE_LOAD]], <4 x i8>* [[TMP6]], align 1
40 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
41 ; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
42 ; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !0
43 ; CHECK: middle.block:
44 ; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[SIZE]], [[N_VEC]]
45 ; CHECK-NEXT: br i1 [[CMP_N]], label [[END:%.*]], label [[SCALAR_PH]]
47 ; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[SIZE]], [[HEADER:%.*]] ]
48 ; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i8* [ [[IND_END2]], [[MIDDLE_BLOCK]] ], [ [[PTR]], [[HEADER]] ]
49 ; CHECK-NEXT: br label [[BODY:%.*]]
51 ; CHECK-NEXT: [[DEC66:%.*]] = phi i32 [ [[DEC:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
52 ; CHECK-NEXT: [[BUFF:%.*]] = phi i8* [ [[INCDEC_PTR:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]
53 ; CHECK-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i8, i8* [[BUFF]], i32 1
54 ; CHECK-NEXT: [[DEC]] = add nsw i32 [[DEC66]], -1
55 ; CHECK-NEXT: [[TMP8:%.*]] = load i8, i8* [[INCDEC_PTR]], align 1
56 ; CHECK-NEXT: store i8 [[TMP8]], i8* [[BUFF]], align 1
57 ; CHECK-NEXT: [[TOBOOL11:%.*]] = icmp eq i32 [[DEC]], 0
58 ; CHECK-NEXT: br i1 [[TOBOOL11]], label [[END]], label [[BODY]], !llvm.loop !2
60 ; CHECK-NEXT: [[INCDEC_PTR_LCSSA:%.*]] = phi i8* [ [[INCDEC_PTR]], [[BODY]] ], [ [[IND_END2]], [[MIDDLE_BLOCK]] ]
61 ; CHECK-NEXT: store i8* [[INCDEC_PTR_LCSSA]], i8** [[POS]], align 4
62 ; CHECK-NEXT: ret void
65 %ptr0 = load i8*, i8** %pos, align 4
69 %dec66 = phi i32 [ %dec, %body ], [ %size, %header ]
70 %buff = phi i8* [ %incdec.ptr, %body ], [ %ptr, %header ]
71 %incdec.ptr = getelementptr inbounds i8, i8* %buff, i32 1
72 %dec = add nsw i32 %dec66, -1
73 %0 = load i8, i8* %incdec.ptr, align 1
74 store i8 %0, i8* %buff, align 1
75 %tobool11 = icmp eq i32 %dec, 0
76 br i1 %tobool11, label %end, label %body
79 store i8* %incdec.ptr, i8** %pos, align 4
83 define void @metadata(i8* nocapture readonly %ptr, i32 %size, i8** %pos) {
84 ; CHECK-LABEL: @metadata(
86 ; CHECK-NEXT: [[PTR0:%.*]] = load i8*, i8** [[POS:%.*]], align 4
87 ; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[SIZE:%.*]], 4
88 ; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
90 ; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[SIZE]], 4
91 ; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[SIZE]], [[N_MOD_VF]]
92 ; CHECK-NEXT: [[IND_END:%.*]] = sub i32 [[SIZE]], [[N_VEC]]
93 ; CHECK-NEXT: [[IND_END2:%.*]] = getelementptr i8, i8* [[PTR:%.*]], i32 [[N_VEC]]
94 ; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
96 ; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
97 ; CHECK-NEXT: [[OFFSET_IDX:%.*]] = sub i32 [[SIZE]], [[INDEX]]
98 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[OFFSET_IDX]], 0
99 ; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[INDEX]], 0
100 ; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, i8* [[PTR]], i32 [[TMP1]]
101 ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, i8* [[NEXT_GEP]], i32 1
102 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, i8* [[TMP2]], i32 0
103 ; CHECK-NEXT: [[TMP4:%.*]] = bitcast i8* [[TMP3]] to <4 x i8>*
104 ; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i8>, <4 x i8>* [[TMP4]], align 1
105 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, i8* [[NEXT_GEP]], i32 0
106 ; CHECK-NEXT: [[TMP6:%.*]] = bitcast i8* [[TMP5]] to <4 x i8>*
107 ; CHECK-NEXT: store <4 x i8> [[WIDE_LOAD]], <4 x i8>* [[TMP6]], align 1
108 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
109 ; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
110 ; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop !4
111 ; CHECK: middle.block:
112 ; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[SIZE]], [[N_VEC]]
113 ; CHECK-NEXT: br i1 [[CMP_N]], label [[END:%.*]], label [[SCALAR_PH]]
115 ; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[SIZE]], [[HEADER:%.*]] ]
116 ; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi i8* [ [[IND_END2]], [[MIDDLE_BLOCK]] ], [ [[PTR]], [[HEADER]] ]
117 ; CHECK-NEXT: br label [[BODY:%.*]]
119 ; CHECK-NEXT: [[DEC66:%.*]] = phi i32 [ [[DEC:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
120 ; CHECK-NEXT: [[BUFF:%.*]] = phi i8* [ [[INCDEC_PTR:%.*]], [[BODY]] ], [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ]
121 ; CHECK-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i8, i8* [[BUFF]], i32 1
122 ; CHECK-NEXT: [[DEC]] = add nsw i32 [[DEC66]], -1
123 ; CHECK-NEXT: [[TMP8:%.*]] = load i8, i8* [[INCDEC_PTR]], align 1
124 ; CHECK-NEXT: store i8 [[TMP8]], i8* [[BUFF]], align 1
125 ; CHECK-NEXT: [[TOBOOL11:%.*]] = icmp eq i32 [[DEC]], 0
126 ; CHECK-NEXT: br i1 [[TOBOOL11]], label [[END]], label [[BODY]], !llvm.loop !5
128 ; CHECK-NEXT: [[INCDEC_PTR_LCSSA:%.*]] = phi i8* [ [[INCDEC_PTR]], [[BODY]] ], [ [[IND_END2]], [[MIDDLE_BLOCK]] ]
129 ; CHECK-NEXT: store i8* [[INCDEC_PTR_LCSSA]], i8** [[POS]], align 4
130 ; CHECK-NEXT: ret void
133 %ptr0 = load i8*, i8** %pos, align 4
137 %dec66 = phi i32 [ %dec, %body ], [ %size, %header ]
138 %buff = phi i8* [ %incdec.ptr, %body ], [ %ptr, %header ]
139 %incdec.ptr = getelementptr inbounds i8, i8* %buff, i32 1
140 %dec = add nsw i32 %dec66, -1
141 %0 = load i8, i8* %incdec.ptr, align 1
142 store i8 %0, i8* %buff, align 1
143 %tobool11 = icmp eq i32 %dec, 0
144 br i1 %tobool11, label %end, label %body, !llvm.loop !1
147 store i8* %incdec.ptr, i8** %pos, align 4
151 !1 = distinct !{!1, !2, !3}
152 !2 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
153 !3 = !{!"llvm.loop.vectorize.enable", i1 true}