xref: /llvm-project/llvm/test/Transforms/LoopVectorize/ARM/mve-qabs.ll (revision f7685af4a5bd188e6d548967d818d8569f10a70d) 
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -passes=loop-vectorize,instcombine,simplifycfg -simplifycfg-require-and-preserve-domtree=1 < %s -S -o - | FileCheck %s
3
4target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
5target triple = "thumbv8.1m.main-arm-none-eabi"
6
7define void @arm_abs_q7(ptr nocapture readonly %pSrc, ptr nocapture %pDst, i32 %blockSize) #0 {
8; CHECK-LABEL: @arm_abs_q7(
9; CHECK-NEXT:  entry:
10; CHECK-NEXT:    [[PSRC2:%.*]] = ptrtoint ptr [[PSRC:%.*]] to i32
11; CHECK-NEXT:    [[PDST1:%.*]] = ptrtoint ptr [[PDST:%.*]] to i32
12; CHECK-NEXT:    [[CMP_NOT19:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 0
13; CHECK-NEXT:    br i1 [[CMP_NOT19]], label [[WHILE_END:%.*]], label [[WHILE_BODY_PREHEADER:%.*]]
14; CHECK:       while.body.preheader:
15; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[BLOCKSIZE]], 16
16; CHECK-NEXT:    [[TMP0:%.*]] = sub i32 [[PDST1]], [[PSRC2]]
17; CHECK-NEXT:    [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP0]], 16
18; CHECK-NEXT:    [[OR_COND:%.*]] = select i1 [[MIN_ITERS_CHECK]], i1 true, i1 [[DIFF_CHECK]]
19; CHECK-NEXT:    br i1 [[OR_COND]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
20; CHECK:       vector.ph:
21; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[BLOCKSIZE]], -16
22; CHECK-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[N_VEC]]
23; CHECK-NEXT:    [[IND_END3:%.*]] = and i32 [[BLOCKSIZE]], 15
24; CHECK-NEXT:    [[IND_END5:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[N_VEC]]
25; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
26; CHECK:       vector.body:
27; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
28; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[INDEX]]
29; CHECK-NEXT:    [[NEXT_GEP7:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[INDEX]]
30; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <16 x i8>, ptr [[NEXT_GEP]], align 1
31; CHECK-NEXT:    [[TMP1:%.*]] = icmp sgt <16 x i8> [[WIDE_LOAD]], zeroinitializer
32; CHECK-NEXT:    [[TMP2:%.*]] = icmp eq <16 x i8> [[WIDE_LOAD]], splat (i8 -128)
33; CHECK-NEXT:    [[TMP3:%.*]] = sub <16 x i8> zeroinitializer, [[WIDE_LOAD]]
34; CHECK-NEXT:    [[TMP4:%.*]] = select <16 x i1> [[TMP2]], <16 x i8> splat (i8 127), <16 x i8> [[TMP3]]
35; CHECK-NEXT:    [[TMP5:%.*]] = select <16 x i1> [[TMP1]], <16 x i8> [[WIDE_LOAD]], <16 x i8> [[TMP4]]
36; CHECK-NEXT:    store <16 x i8> [[TMP5]], ptr [[NEXT_GEP7]], align 1
37; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
38; CHECK-NEXT:    [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
39; CHECK-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
40; CHECK:       middle.block:
41; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[BLOCKSIZE]], [[N_VEC]]
42; CHECK-NEXT:    br i1 [[CMP_N]], label [[WHILE_END]], label [[SCALAR_PH]]
43; CHECK:       scalar.ph:
44; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[PSRC]], [[WHILE_BODY_PREHEADER]] ]
45; CHECK-NEXT:    [[BC_RESUME_VAL4:%.*]] = phi i32 [ [[IND_END3]], [[MIDDLE_BLOCK]] ], [ [[BLOCKSIZE]], [[WHILE_BODY_PREHEADER]] ]
46; CHECK-NEXT:    [[BC_RESUME_VAL6:%.*]] = phi ptr [ [[IND_END5]], [[MIDDLE_BLOCK]] ], [ [[PDST]], [[WHILE_BODY_PREHEADER]] ]
47; CHECK-NEXT:    br label [[WHILE_BODY:%.*]]
48; CHECK:       while.body:
49; CHECK-NEXT:    [[PSRC_ADDR_022:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
50; CHECK-NEXT:    [[BLKCNT_021:%.*]] = phi i32 [ [[DEC:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL4]], [[SCALAR_PH]] ]
51; CHECK-NEXT:    [[PDST_ADDR_020:%.*]] = phi ptr [ [[INCDEC_PTR13:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL6]], [[SCALAR_PH]] ]
52; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds nuw i8, ptr [[PSRC_ADDR_022]], i32 1
53; CHECK-NEXT:    [[TMP7:%.*]] = load i8, ptr [[PSRC_ADDR_022]], align 1
54; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i8 [[TMP7]], 0
55; CHECK-NEXT:    [[CMP5:%.*]] = icmp eq i8 [[TMP7]], -128
56; CHECK-NEXT:    [[SUB:%.*]] = sub i8 0, [[TMP7]]
57; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP5]], i8 127, i8 [[SUB]]
58; CHECK-NEXT:    [[COND11:%.*]] = select i1 [[CMP1]], i8 [[TMP7]], i8 [[COND]]
59; CHECK-NEXT:    [[INCDEC_PTR13]] = getelementptr inbounds nuw i8, ptr [[PDST_ADDR_020]], i32 1
60; CHECK-NEXT:    store i8 [[COND11]], ptr [[PDST_ADDR_020]], align 1
61; CHECK-NEXT:    [[DEC]] = add i32 [[BLKCNT_021]], -1
62; CHECK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[DEC]], 0
63; CHECK-NEXT:    br i1 [[CMP_NOT]], label [[WHILE_END]], label [[WHILE_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
64; CHECK:       while.end:
65; CHECK-NEXT:    ret void
66;
67entry:
68  %cmp.not19 = icmp eq i32 %blockSize, 0
69  br i1 %cmp.not19, label %while.end, label %while.body.preheader
70
71while.body.preheader:                             ; preds = %entry
72  br label %while.body
73
74while.body:                                       ; preds = %while.body.preheader, %while.body
75  %pSrc.addr.022 = phi ptr [ %incdec.ptr, %while.body ], [ %pSrc, %while.body.preheader ]
76  %blkCnt.021 = phi i32 [ %dec, %while.body ], [ %blockSize, %while.body.preheader ]
77  %pDst.addr.020 = phi ptr [ %incdec.ptr13, %while.body ], [ %pDst, %while.body.preheader ]
78  %incdec.ptr = getelementptr inbounds i8, ptr %pSrc.addr.022, i32 1
79  %0 = load i8, ptr %pSrc.addr.022, align 1
80  %cmp1 = icmp sgt i8 %0, 0
81  %cmp5 = icmp eq i8 %0, -128
82  %sub = sub i8 0, %0
83  %cond = select i1 %cmp5, i8 127, i8 %sub
84  %cond11 = select i1 %cmp1, i8 %0, i8 %cond
85  %incdec.ptr13 = getelementptr inbounds i8, ptr %pDst.addr.020, i32 1
86  store i8 %cond11, ptr %pDst.addr.020, align 1
87  %dec = add i32 %blkCnt.021, -1
88  %cmp.not = icmp eq i32 %dec, 0
89  br i1 %cmp.not, label %while.end.loopexit, label %while.body
90
91while.end.loopexit:                               ; preds = %while.body
92  br label %while.end
93
94while.end:                                        ; preds = %while.end.loopexit, %entry
95  ret void
96}
97
98define void @arm_abs_q15(ptr nocapture readonly %pSrc, ptr nocapture %pDst, i32 %blockSize) #0 {
99; CHECK-LABEL: @arm_abs_q15(
100; CHECK-NEXT:  entry:
101; CHECK-NEXT:    [[PSRC2:%.*]] = ptrtoint ptr [[PSRC:%.*]] to i32
102; CHECK-NEXT:    [[PDST1:%.*]] = ptrtoint ptr [[PDST:%.*]] to i32
103; CHECK-NEXT:    [[CMP_NOT20:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 0
104; CHECK-NEXT:    br i1 [[CMP_NOT20]], label [[WHILE_END:%.*]], label [[WHILE_BODY_PREHEADER:%.*]]
105; CHECK:       while.body.preheader:
106; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[BLOCKSIZE]], 8
107; CHECK-NEXT:    [[TMP0:%.*]] = sub i32 [[PDST1]], [[PSRC2]]
108; CHECK-NEXT:    [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP0]], 16
109; CHECK-NEXT:    [[OR_COND:%.*]] = select i1 [[MIN_ITERS_CHECK]], i1 true, i1 [[DIFF_CHECK]]
110; CHECK-NEXT:    br i1 [[OR_COND]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
111; CHECK:       vector.ph:
112; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[BLOCKSIZE]], -8
113; CHECK-NEXT:    [[TMP1:%.*]] = shl i32 [[N_VEC]], 1
114; CHECK-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[TMP1]]
115; CHECK-NEXT:    [[IND_END3:%.*]] = and i32 [[BLOCKSIZE]], 7
116; CHECK-NEXT:    [[TMP2:%.*]] = shl i32 [[N_VEC]], 1
117; CHECK-NEXT:    [[IND_END5:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[TMP2]]
118; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
119; CHECK:       vector.body:
120; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
121; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i32 [[INDEX]], 1
122; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[OFFSET_IDX]]
123; CHECK-NEXT:    [[OFFSET_IDX7:%.*]] = shl i32 [[INDEX]], 1
124; CHECK-NEXT:    [[NEXT_GEP8:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[OFFSET_IDX7]]
125; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x i16>, ptr [[NEXT_GEP]], align 2
126; CHECK-NEXT:    [[TMP3:%.*]] = icmp sgt <8 x i16> [[WIDE_LOAD]], zeroinitializer
127; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq <8 x i16> [[WIDE_LOAD]], splat (i16 -32768)
128; CHECK-NEXT:    [[TMP5:%.*]] = sub <8 x i16> zeroinitializer, [[WIDE_LOAD]]
129; CHECK-NEXT:    [[TMP6:%.*]] = select <8 x i1> [[TMP4]], <8 x i16> splat (i16 32767), <8 x i16> [[TMP5]]
130; CHECK-NEXT:    [[TMP7:%.*]] = select <8 x i1> [[TMP3]], <8 x i16> [[WIDE_LOAD]], <8 x i16> [[TMP6]]
131; CHECK-NEXT:    store <8 x i16> [[TMP7]], ptr [[NEXT_GEP8]], align 2
132; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 8
133; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
134; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
135; CHECK:       middle.block:
136; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[BLOCKSIZE]], [[N_VEC]]
137; CHECK-NEXT:    br i1 [[CMP_N]], label [[WHILE_END]], label [[SCALAR_PH]]
138; CHECK:       scalar.ph:
139; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[PSRC]], [[WHILE_BODY_PREHEADER]] ]
140; CHECK-NEXT:    [[BC_RESUME_VAL4:%.*]] = phi i32 [ [[IND_END3]], [[MIDDLE_BLOCK]] ], [ [[BLOCKSIZE]], [[WHILE_BODY_PREHEADER]] ]
141; CHECK-NEXT:    [[BC_RESUME_VAL6:%.*]] = phi ptr [ [[IND_END5]], [[MIDDLE_BLOCK]] ], [ [[PDST]], [[WHILE_BODY_PREHEADER]] ]
142; CHECK-NEXT:    br label [[WHILE_BODY:%.*]]
143; CHECK:       while.body:
144; CHECK-NEXT:    [[PSRC_ADDR_023:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
145; CHECK-NEXT:    [[BLKCNT_022:%.*]] = phi i32 [ [[DEC:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL4]], [[SCALAR_PH]] ]
146; CHECK-NEXT:    [[PDST_ADDR_021:%.*]] = phi ptr [ [[INCDEC_PTR13:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL6]], [[SCALAR_PH]] ]
147; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds nuw i8, ptr [[PSRC_ADDR_023]], i32 2
148; CHECK-NEXT:    [[TMP9:%.*]] = load i16, ptr [[PSRC_ADDR_023]], align 2
149; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i16 [[TMP9]], 0
150; CHECK-NEXT:    [[CMP5:%.*]] = icmp eq i16 [[TMP9]], -32768
151; CHECK-NEXT:    [[SUB:%.*]] = sub i16 0, [[TMP9]]
152; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP5]], i16 32767, i16 [[SUB]]
153; CHECK-NEXT:    [[COND11:%.*]] = select i1 [[CMP1]], i16 [[TMP9]], i16 [[COND]]
154; CHECK-NEXT:    [[INCDEC_PTR13]] = getelementptr inbounds nuw i8, ptr [[PDST_ADDR_021]], i32 2
155; CHECK-NEXT:    store i16 [[COND11]], ptr [[PDST_ADDR_021]], align 2
156; CHECK-NEXT:    [[DEC]] = add i32 [[BLKCNT_022]], -1
157; CHECK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[DEC]], 0
158; CHECK-NEXT:    br i1 [[CMP_NOT]], label [[WHILE_END]], label [[WHILE_BODY]], !llvm.loop [[LOOP5:![0-9]+]]
159; CHECK:       while.end:
160; CHECK-NEXT:    ret void
161;
162entry:
163  %cmp.not20 = icmp eq i32 %blockSize, 0
164  br i1 %cmp.not20, label %while.end, label %while.body.preheader
165
166while.body.preheader:                             ; preds = %entry
167  br label %while.body
168
169while.body:                                       ; preds = %while.body.preheader, %while.body
170  %pSrc.addr.023 = phi ptr [ %incdec.ptr, %while.body ], [ %pSrc, %while.body.preheader ]
171  %blkCnt.022 = phi i32 [ %dec, %while.body ], [ %blockSize, %while.body.preheader ]
172  %pDst.addr.021 = phi ptr [ %incdec.ptr13, %while.body ], [ %pDst, %while.body.preheader ]
173  %incdec.ptr = getelementptr inbounds i16, ptr %pSrc.addr.023, i32 1
174  %0 = load i16, ptr %pSrc.addr.023, align 2
175  %cmp1 = icmp sgt i16 %0, 0
176  %cmp5 = icmp eq i16 %0, -32768
177  %sub = sub i16 0, %0
178  %cond = select i1 %cmp5, i16 32767, i16 %sub
179  %cond11 = select i1 %cmp1, i16 %0, i16 %cond
180  %incdec.ptr13 = getelementptr inbounds i16, ptr %pDst.addr.021, i32 1
181  store i16 %cond11, ptr %pDst.addr.021, align 2
182  %dec = add i32 %blkCnt.022, -1
183  %cmp.not = icmp eq i32 %dec, 0
184  br i1 %cmp.not, label %while.end.loopexit, label %while.body
185
186while.end.loopexit:                               ; preds = %while.body
187  br label %while.end
188
189while.end:                                        ; preds = %while.end.loopexit, %entry
190  ret void
191}
192
193define void @arm_abs_q31(ptr nocapture readonly %pSrc, ptr nocapture %pDst, i32 %blockSize) #0 {
194; CHECK-LABEL: @arm_abs_q31(
195; CHECK-NEXT:  entry:
196; CHECK-NEXT:    [[PSRC2:%.*]] = ptrtoint ptr [[PSRC:%.*]] to i32
197; CHECK-NEXT:    [[PDST1:%.*]] = ptrtoint ptr [[PDST:%.*]] to i32
198; CHECK-NEXT:    [[CMP_NOT14:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 0
199; CHECK-NEXT:    br i1 [[CMP_NOT14]], label [[WHILE_END:%.*]], label [[WHILE_BODY_PREHEADER:%.*]]
200; CHECK:       while.body.preheader:
201; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[BLOCKSIZE]], 4
202; CHECK-NEXT:    [[TMP0:%.*]] = sub i32 [[PDST1]], [[PSRC2]]
203; CHECK-NEXT:    [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP0]], 16
204; CHECK-NEXT:    [[OR_COND:%.*]] = select i1 [[MIN_ITERS_CHECK]], i1 true, i1 [[DIFF_CHECK]]
205; CHECK-NEXT:    br i1 [[OR_COND]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
206; CHECK:       vector.ph:
207; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[BLOCKSIZE]], -4
208; CHECK-NEXT:    [[TMP1:%.*]] = shl i32 [[N_VEC]], 2
209; CHECK-NEXT:    [[IND_END:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[TMP1]]
210; CHECK-NEXT:    [[IND_END3:%.*]] = and i32 [[BLOCKSIZE]], 3
211; CHECK-NEXT:    [[TMP2:%.*]] = shl i32 [[N_VEC]], 2
212; CHECK-NEXT:    [[IND_END5:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[TMP2]]
213; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
214; CHECK:       vector.body:
215; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
216; CHECK-NEXT:    [[OFFSET_IDX:%.*]] = shl i32 [[INDEX]], 2
217; CHECK-NEXT:    [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PSRC]], i32 [[OFFSET_IDX]]
218; CHECK-NEXT:    [[OFFSET_IDX7:%.*]] = shl i32 [[INDEX]], 2
219; CHECK-NEXT:    [[NEXT_GEP8:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[OFFSET_IDX7]]
220; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[NEXT_GEP]], align 4
221; CHECK-NEXT:    [[TMP3:%.*]] = icmp sgt <4 x i32> [[WIDE_LOAD]], zeroinitializer
222; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq <4 x i32> [[WIDE_LOAD]], splat (i32 -2147483648)
223; CHECK-NEXT:    [[TMP5:%.*]] = sub nsw <4 x i32> zeroinitializer, [[WIDE_LOAD]]
224; CHECK-NEXT:    [[TMP6:%.*]] = select <4 x i1> [[TMP4]], <4 x i32> splat (i32 2147483647), <4 x i32> [[TMP5]]
225; CHECK-NEXT:    [[TMP7:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[WIDE_LOAD]], <4 x i32> [[TMP6]]
226; CHECK-NEXT:    store <4 x i32> [[TMP7]], ptr [[NEXT_GEP8]], align 4
227; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4
228; CHECK-NEXT:    [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
229; CHECK-NEXT:    br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
230; CHECK:       middle.block:
231; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i32 [[BLOCKSIZE]], [[N_VEC]]
232; CHECK-NEXT:    br i1 [[CMP_N]], label [[WHILE_END]], label [[SCALAR_PH]]
233; CHECK:       scalar.ph:
234; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[PSRC]], [[WHILE_BODY_PREHEADER]] ]
235; CHECK-NEXT:    [[BC_RESUME_VAL4:%.*]] = phi i32 [ [[IND_END3]], [[MIDDLE_BLOCK]] ], [ [[BLOCKSIZE]], [[WHILE_BODY_PREHEADER]] ]
236; CHECK-NEXT:    [[BC_RESUME_VAL6:%.*]] = phi ptr [ [[IND_END5]], [[MIDDLE_BLOCK]] ], [ [[PDST]], [[WHILE_BODY_PREHEADER]] ]
237; CHECK-NEXT:    br label [[WHILE_BODY:%.*]]
238; CHECK:       while.body:
239; CHECK-NEXT:    [[PSRC_ADDR_017:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
240; CHECK-NEXT:    [[BLKCNT_016:%.*]] = phi i32 [ [[DEC:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL4]], [[SCALAR_PH]] ]
241; CHECK-NEXT:    [[PDST_ADDR_015:%.*]] = phi ptr [ [[INCDEC_PTR7:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL6]], [[SCALAR_PH]] ]
242; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds nuw i8, ptr [[PSRC_ADDR_017]], i32 4
243; CHECK-NEXT:    [[TMP9:%.*]] = load i32, ptr [[PSRC_ADDR_017]], align 4
244; CHECK-NEXT:    [[CMP1:%.*]] = icmp sgt i32 [[TMP9]], 0
245; CHECK-NEXT:    [[CMP2:%.*]] = icmp eq i32 [[TMP9]], -2147483648
246; CHECK-NEXT:    [[SUB:%.*]] = sub nsw i32 0, [[TMP9]]
247; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP2]], i32 2147483647, i32 [[SUB]]
248; CHECK-NEXT:    [[COND6:%.*]] = select i1 [[CMP1]], i32 [[TMP9]], i32 [[COND]]
249; CHECK-NEXT:    [[INCDEC_PTR7]] = getelementptr inbounds nuw i8, ptr [[PDST_ADDR_015]], i32 4
250; CHECK-NEXT:    store i32 [[COND6]], ptr [[PDST_ADDR_015]], align 4
251; CHECK-NEXT:    [[DEC]] = add i32 [[BLKCNT_016]], -1
252; CHECK-NEXT:    [[CMP_NOT:%.*]] = icmp eq i32 [[DEC]], 0
253; CHECK-NEXT:    br i1 [[CMP_NOT]], label [[WHILE_END]], label [[WHILE_BODY]], !llvm.loop [[LOOP7:![0-9]+]]
254; CHECK:       while.end:
255; CHECK-NEXT:    ret void
256;
257entry:
258  %cmp.not14 = icmp eq i32 %blockSize, 0
259  br i1 %cmp.not14, label %while.end, label %while.body.preheader
260
261while.body.preheader:                             ; preds = %entry
262  br label %while.body
263
264while.body:                                       ; preds = %while.body.preheader, %while.body
265  %pSrc.addr.017 = phi ptr [ %incdec.ptr, %while.body ], [ %pSrc, %while.body.preheader ]
266  %blkCnt.016 = phi i32 [ %dec, %while.body ], [ %blockSize, %while.body.preheader ]
267  %pDst.addr.015 = phi ptr [ %incdec.ptr7, %while.body ], [ %pDst, %while.body.preheader ]
268  %incdec.ptr = getelementptr inbounds i32, ptr %pSrc.addr.017, i32 1
269  %0 = load i32, ptr %pSrc.addr.017, align 4
270  %cmp1 = icmp sgt i32 %0, 0
271  %cmp2 = icmp eq i32 %0, -2147483648
272  %sub = sub nsw i32 0, %0
273  %cond = select i1 %cmp2, i32 2147483647, i32 %sub
274  %cond6 = select i1 %cmp1, i32 %0, i32 %cond
275  %incdec.ptr7 = getelementptr inbounds i32, ptr %pDst.addr.015, i32 1
276  store i32 %cond6, ptr %pDst.addr.015, align 4
277  %dec = add i32 %blkCnt.016, -1
278  %cmp.not = icmp eq i32 %dec, 0
279  br i1 %cmp.not, label %while.end.loopexit, label %while.body
280
281while.end.loopexit:                               ; preds = %while.body
282  br label %while.end
283
284while.end:                                        ; preds = %while.end.loopexit, %entry
285  ret void
286}
287
288attributes #0 = { "target-features"="+mve" }
289