xref: /llvm-project/llvm/test/Transforms/LoopVectorize/instruction-only-used-outside-of-loop.ll (revision 38fffa630ee80163dc65e759392ad29798905679) 
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -passes=loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S %s | FileCheck %s
3
4; Tests for PR54266.
5define i32 @one_direct_branch(ptr %src) {
6; CHECK-LABEL: @one_direct_branch(
7; CHECK-NEXT:  entry:
8; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
9; CHECK:       vector.ph:
10; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
11; CHECK:       vector.body:
12; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
13; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
14; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[SRC:%.*]], i32 [[TMP0]]
15; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
16; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP2]], align 4
17; CHECK-NEXT:    [[TMP3:%.*]] = xor <4 x i32> splat (i32 25500), [[WIDE_LOAD]]
18; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
19; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
20; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
21; CHECK:       middle.block:
22; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i32> [[TMP3]], i32 3
23; CHECK-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
24; CHECK:       scalar.ph:
25; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
26; CHECK-NEXT:    br label [[LOOP:%.*]]
27; CHECK:       loop:
28; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
29; CHECK-NEXT:    [[SRC_GEP:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i32 [[IV]]
30; CHECK-NEXT:    [[LV:%.*]] = load i32, ptr [[SRC_GEP]], align 4
31; CHECK-NEXT:    [[XOR:%.*]] = xor i32 25500, [[LV]]
32; CHECK-NEXT:    br label [[LOOP_LATCH]]
33; CHECK:       loop.latch:
34; CHECK-NEXT:    [[PHI_XOR:%.*]] = phi i32 [ [[XOR]], [[LOOP]] ]
35; CHECK-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
36; CHECK-NEXT:    [[TOBOOL_NOT:%.*]] = icmp eq i32 [[IV_NEXT]], 1000
37; CHECK-NEXT:    br i1 [[TOBOOL_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
38; CHECK:       exit:
39; CHECK-NEXT:    [[XOR_LCSSA:%.*]] = phi i32 [ [[PHI_XOR]], [[LOOP_LATCH]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
40; CHECK-NEXT:    ret i32 [[XOR_LCSSA]]
41;
42entry:
43  br label %loop
44
45loop:
46  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop.latch ]
47  %src.gep = getelementptr inbounds i32, ptr %src, i32 %iv
48  %lv = load i32, ptr %src.gep
49  %xor = xor i32 25500, %lv
50  br label %loop.latch
51
52loop.latch:
53  %phi.xor = phi i32 [ %xor, %loop ]
54  %iv.next = add nsw i32 %iv, 1
55  %tobool.not = icmp eq i32 %iv.next, 1000
56  br i1 %tobool.not, label %exit, label %loop
57
58exit:
59  %xor.lcssa = phi i32 [ %phi.xor, %loop.latch ]
60  ret i32 %xor.lcssa
61}
62
63define i32 @two_direct_branch(ptr %src) {
64; CHECK-LABEL: @two_direct_branch(
65; CHECK-NEXT:  entry:
66; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
67; CHECK:       vector.ph:
68; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
69; CHECK:       vector.body:
70; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
71; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
72; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[SRC:%.*]], i32 [[TMP0]]
73; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
74; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP2]], align 4
75; CHECK-NEXT:    [[TMP3:%.*]] = xor <4 x i32> splat (i32 25500), [[WIDE_LOAD]]
76; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
77; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
78; CHECK-NEXT:    br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
79; CHECK:       middle.block:
80; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i32> [[TMP3]], i32 3
81; CHECK-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
82; CHECK:       scalar.ph:
83; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
84; CHECK-NEXT:    br label [[LOOP:%.*]]
85; CHECK:       loop:
86; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
87; CHECK-NEXT:    [[SRC_GEP:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i32 [[IV]]
88; CHECK-NEXT:    [[LV:%.*]] = load i32, ptr [[SRC_GEP]], align 4
89; CHECK-NEXT:    [[XOR:%.*]] = xor i32 25500, [[LV]]
90; CHECK-NEXT:    br label [[BB:%.*]]
91; CHECK:       bb:
92; CHECK-NEXT:    [[PHI_XOR_1:%.*]] = phi i32 [ [[XOR]], [[LOOP]] ]
93; CHECK-NEXT:    br label [[LOOP_LATCH]]
94; CHECK:       loop.latch:
95; CHECK-NEXT:    [[PHI_XOR:%.*]] = phi i32 [ [[PHI_XOR_1]], [[BB]] ]
96; CHECK-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
97; CHECK-NEXT:    [[TOBOOL_NOT:%.*]] = icmp eq i32 [[IV_NEXT]], 1000
98; CHECK-NEXT:    br i1 [[TOBOOL_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
99; CHECK:       exit:
100; CHECK-NEXT:    [[XOR_LCSSA:%.*]] = phi i32 [ [[PHI_XOR]], [[LOOP_LATCH]] ], [ [[TMP5]], [[MIDDLE_BLOCK]] ]
101; CHECK-NEXT:    ret i32 [[XOR_LCSSA]]
102;
103entry:
104  br label %loop
105
106loop:                              ; preds = %for.inc3, %entry
107  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop.latch ]
108  %src.gep = getelementptr inbounds i32, ptr %src, i32 %iv
109  %lv = load i32, ptr %src.gep
110  %xor = xor i32 25500, %lv
111  br label %bb
112
113bb:
114  %phi.xor.1 = phi i32 [ %xor, %loop ]
115  br label %loop.latch
116
117loop.latch:
118  %phi.xor = phi i32 [ %phi.xor.1, %bb ]
119  %iv.next = add nsw i32 %iv, 1
120  %tobool.not = icmp eq i32 %iv.next, 1000
121  br i1 %tobool.not, label %exit, label %loop
122
123exit:
124  %xor.lcssa = phi i32 [ %phi.xor, %loop.latch ]
125  ret i32 %xor.lcssa
126}
127
128define i32 @cond_branch(i32 %a, ptr %src) {
129; CHECK-LABEL: @cond_branch(
130; CHECK-NEXT:  entry:
131; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
132; CHECK:       vector.ph:
133; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[A:%.*]], i64 0
134; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer
135; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
136; CHECK:       vector.body:
137; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
138; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
139; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[INDEX]], 0
140; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[SRC:%.*]], i32 [[TMP0]]
141; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
142; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP2]], align 4
143; CHECK-NEXT:    [[TMP3:%.*]] = xor <4 x i32> splat (i32 25500), [[WIDE_LOAD]]
144; CHECK-NEXT:    [[TMP4:%.*]] = icmp ne <4 x i32> [[VEC_IND]], [[BROADCAST_SPLAT]]
145; CHECK-NEXT:    [[TMP5:%.*]] = xor <4 x i1> [[TMP4]], splat (i1 true)
146; CHECK-NEXT:    [[PREDPHI:%.*]] = select <4 x i1> [[TMP5]], <4 x i32> splat (i32 10), <4 x i32> [[TMP3]]
147; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
148; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
149; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000
150; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
151; CHECK:       middle.block:
152; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <4 x i32> [[PREDPHI]], i32 3
153; CHECK-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
154; CHECK:       scalar.ph:
155; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i32 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
156; CHECK-NEXT:    br label [[LOOP:%.*]]
157; CHECK:       loop:
158; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
159; CHECK-NEXT:    [[SRC_GEP:%.*]] = getelementptr inbounds i32, ptr [[SRC]], i32 [[IV]]
160; CHECK-NEXT:    [[LV:%.*]] = load i32, ptr [[SRC_GEP]], align 4
161; CHECK-NEXT:    [[XOR:%.*]] = xor i32 25500, [[LV]]
162; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[IV]], [[A]]
163; CHECK-NEXT:    br i1 [[CMP]], label [[LOOP_LATCH]], label [[THEN:%.*]]
164; CHECK:       then:
165; CHECK-NEXT:    br label [[LOOP_LATCH]]
166; CHECK:       loop.latch:
167; CHECK-NEXT:    [[PHI_XOR:%.*]] = phi i32 [ [[XOR]], [[LOOP]] ], [ 10, [[THEN]] ]
168; CHECK-NEXT:    [[IV_NEXT]] = add nsw i32 [[IV]], 1
169; CHECK-NEXT:    [[TOBOOL_NOT:%.*]] = icmp eq i32 [[IV_NEXT]], 1000
170; CHECK-NEXT:    br i1 [[TOBOOL_NOT]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP7:![0-9]+]]
171; CHECK:       exit:
172; CHECK-NEXT:    [[XOR_LCSSA:%.*]] = phi i32 [ [[PHI_XOR]], [[LOOP_LATCH]] ], [ [[TMP7]], [[MIDDLE_BLOCK]] ]
173; CHECK-NEXT:    ret i32 [[XOR_LCSSA]]
174;
175entry:
176  br label %loop
177
178loop:
179  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop.latch ]
180  %src.gep = getelementptr inbounds i32, ptr %src, i32 %iv
181  %lv = load i32, ptr %src.gep
182  %xor = xor i32 25500, %lv
183  %cmp = icmp ne i32 %iv, %a
184  br i1 %cmp, label %loop.latch, label %then
185
186then:
187  br label %loop.latch
188
189loop.latch:
190  %phi.xor = phi i32 [ %xor, %loop ], [ 10, %then ]
191  %iv.next = add nsw i32 %iv, 1
192  %tobool.not = icmp eq i32 %iv.next, 1000
193  br i1 %tobool.not, label %exit, label %loop
194
195exit:
196  %xor.lcssa = phi i32 [ %phi.xor, %loop.latch ]
197  ret i32 %xor.lcssa
198}
199
200; Test case for PR54370.
201; TODO: Should either compute the final value of the truncated IV independent
202; of loop or scalarize the vector IV.
203define i32 @optimizable_trunc_used_outside() {
204; CHECK-LABEL: @optimizable_trunc_used_outside(
205; CHECK-NEXT:  entry:
206; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
207; CHECK:       vector.ph:
208; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
209; CHECK:       vector.body:
210; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
211; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]
212; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
213; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)
214; CHECK-NEXT:    [[TMP0:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
215; CHECK-NEXT:    br i1 [[TMP0]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
216; CHECK:       middle.block:
217; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <4 x i32> [[VEC_IND]], i32 3
218; CHECK-NEXT:    br i1 true, label [[EXIT:%.*]], label [[SCALAR_PH]]
219; CHECK:       scalar.ph:
220; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 1000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
221; CHECK-NEXT:    br label [[LOOP:%.*]]
222; CHECK:       loop:
223; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
224; CHECK-NEXT:    [[IV_TRUNC:%.*]] = trunc i64 [[IV]] to i32
225; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
226; CHECK-NEXT:    [[EXITCOND_NOT_I_I:%.*]] = icmp eq i64 [[IV_NEXT]], 1000
227; CHECK-NEXT:    br i1 [[EXITCOND_NOT_I_I]], label [[EXIT]], label [[LOOP]], !llvm.loop [[LOOP9:![0-9]+]]
228; CHECK:       exit:
229; CHECK-NEXT:    [[IV_TRUNC_LCSSA:%.*]] = phi i32 [ [[IV_TRUNC]], [[LOOP]] ], [ [[TMP1]], [[MIDDLE_BLOCK]] ]
230; CHECK-NEXT:    ret i32 [[IV_TRUNC_LCSSA]]
231;
232entry:
233  br label %loop
234
235loop:
236  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
237  %iv.trunc = trunc i64 %iv to i32
238  %iv.next = add nuw nsw i64 %iv, 1
239  %exitcond.not.i.i = icmp eq i64 %iv.next, 1000
240  br i1 %exitcond.not.i.i, label %exit, label %loop
241
242exit:
243  %iv.trunc.lcssa = phi i32 [ %iv.trunc, %loop ]
244  ret i32 %iv.trunc.lcssa
245}
246