| blob | afc28cfda45aad52e10489160e3fe9ef4b9fa434 |
1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
2 ; RUN: opt -passes=loop-idiom-vectorize -mtriple=riscv64-unknown-linux-gnu -loop-idiom-vectorize-style=predicated -mattr=+v -S < %s | FileCheck %s
3 ; RUN: opt -passes=loop-idiom-vectorize -mtriple=riscv64-unknown-linux-gnu -loop-idiom-vectorize-style=predicated -loop-idiom-vectorize-bytecmp-vf=64 -mattr=+v -S < %s | FileCheck %s --check-prefix=LMUL8
4 ; RUN: opt -passes='loop(loop-idiom-vectorize),simplifycfg' -mtriple=riscv64-unknown-linux-gnu -loop-idiom-vectorize-style=predicated -mattr=+v -S < %s | FileCheck %s --check-prefix=LOOP-DEL
5 ; RUN: opt -passes=loop-idiom-vectorize -mtriple=riscv64-unknown-linux-gnu -loop-idiom-vectorize-style=masked -mattr=+v -S < %s | FileCheck %s --check-prefix=MASKED
7 define i32 @compare_bytes_simple(ptr %a, ptr %b, i32 %len, i32 %n) {
8 ; CHECK-LABEL: define i32 @compare_bytes_simple(
9 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
10 ; CHECK-NEXT: entry:
11 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
12 ; CHECK-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
13 ; CHECK: mismatch_min_it_check:
14 ; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
15 ; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
16 ; CHECK-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
17 ; CHECK-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0:![0-9]+]]
18 ; CHECK: mismatch_mem_check:
19 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
20 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
21 ; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
22 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
23 ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
24 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
25 ; CHECK-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
26 ; CHECK-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
27 ; CHECK-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
28 ; CHECK-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
29 ; CHECK-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
30 ; CHECK-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
31 ; CHECK-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
32 ; CHECK-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
33 ; CHECK-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
34 ; CHECK-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1:![0-9]+]]
35 ; CHECK: mismatch_vec_loop_preheader:
36 ; CHECK-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
37 ; CHECK: mismatch_vec_loop:
38 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
39 ; CHECK-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
40 ; CHECK-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
41 ; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
42 ; CHECK-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
43 ; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
44 ; CHECK-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
45 ; CHECK-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
46 ; CHECK-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
47 ; CHECK-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
48 ; CHECK-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
49 ; CHECK: mismatch_vec_loop_inc:
50 ; CHECK-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
51 ; CHECK-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
52 ; CHECK-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
53 ; CHECK-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
54 ; CHECK: mismatch_vec_loop_found:
55 ; CHECK-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
56 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
57 ; CHECK-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
58 ; CHECK-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
59 ; CHECK-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
60 ; CHECK-NEXT: br label [[MISMATCH_END]]
61 ; CHECK: mismatch_loop_pre:
62 ; CHECK-NEXT: br label [[MISMATCH_LOOP:%.*]]
63 ; CHECK: mismatch_loop:
64 ; CHECK-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
65 ; CHECK-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
66 ; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
67 ; CHECK-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
68 ; CHECK-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
69 ; CHECK-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
70 ; CHECK-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
71 ; CHECK-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
72 ; CHECK: mismatch_loop_inc:
73 ; CHECK-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
74 ; CHECK-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
75 ; CHECK-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
76 ; CHECK: mismatch_end:
77 ; CHECK-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
78 ; CHECK-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
79 ; CHECK: while.cond:
80 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
81 ; CHECK-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
82 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
83 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
84 ; CHECK: while.body:
85 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
86 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
87 ; CHECK-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
88 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
89 ; CHECK-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
90 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
91 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
92 ; CHECK: byte.compare:
93 ; CHECK-NEXT: br label [[WHILE_END]]
94 ; CHECK: while.end:
95 ; CHECK-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
96 ; CHECK-NEXT: ret i32 [[INC_LCSSA]]
97 ;
98 ; LMUL8-LABEL: define i32 @compare_bytes_simple(
99 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
100 ; LMUL8-NEXT: entry:
101 ; LMUL8-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
102 ; LMUL8-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
103 ; LMUL8: mismatch_min_it_check:
104 ; LMUL8-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
105 ; LMUL8-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
106 ; LMUL8-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
107 ; LMUL8-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0:![0-9]+]]
108 ; LMUL8: mismatch_mem_check:
109 ; LMUL8-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
110 ; LMUL8-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
111 ; LMUL8-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
112 ; LMUL8-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
113 ; LMUL8-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
114 ; LMUL8-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
115 ; LMUL8-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
116 ; LMUL8-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
117 ; LMUL8-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
118 ; LMUL8-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
119 ; LMUL8-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
120 ; LMUL8-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
121 ; LMUL8-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
122 ; LMUL8-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
123 ; LMUL8-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
124 ; LMUL8-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1:![0-9]+]]
125 ; LMUL8: mismatch_vec_loop_preheader:
126 ; LMUL8-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
127 ; LMUL8: mismatch_vec_loop:
128 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
129 ; LMUL8-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
130 ; LMUL8-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 64, i1 true)
131 ; LMUL8-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
132 ; LMUL8-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP20]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
133 ; LMUL8-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
134 ; LMUL8-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP21]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
135 ; LMUL8-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 64 x i1> @llvm.vp.icmp.nxv64i8(<vscale x 64 x i8> [[LHS_LOAD]], <vscale x 64 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
136 ; LMUL8-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv64i1(<vscale x 64 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
137 ; LMUL8-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
138 ; LMUL8-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
139 ; LMUL8: mismatch_vec_loop_inc:
140 ; LMUL8-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
141 ; LMUL8-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
142 ; LMUL8-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
143 ; LMUL8-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
144 ; LMUL8: mismatch_vec_loop_found:
145 ; LMUL8-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
146 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
147 ; LMUL8-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
148 ; LMUL8-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
149 ; LMUL8-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
150 ; LMUL8-NEXT: br label [[MISMATCH_END]]
151 ; LMUL8: mismatch_loop_pre:
152 ; LMUL8-NEXT: br label [[MISMATCH_LOOP:%.*]]
153 ; LMUL8: mismatch_loop:
154 ; LMUL8-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
155 ; LMUL8-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
156 ; LMUL8-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
157 ; LMUL8-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
158 ; LMUL8-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
159 ; LMUL8-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
160 ; LMUL8-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
161 ; LMUL8-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
162 ; LMUL8: mismatch_loop_inc:
163 ; LMUL8-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
164 ; LMUL8-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
165 ; LMUL8-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
166 ; LMUL8: mismatch_end:
167 ; LMUL8-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
168 ; LMUL8-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
169 ; LMUL8: while.cond:
170 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
171 ; LMUL8-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
172 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
173 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
174 ; LMUL8: while.body:
175 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
176 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
177 ; LMUL8-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
178 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
179 ; LMUL8-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
180 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
181 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
182 ; LMUL8: byte.compare:
183 ; LMUL8-NEXT: br label [[WHILE_END]]
184 ; LMUL8: while.end:
185 ; LMUL8-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
186 ; LMUL8-NEXT: ret i32 [[INC_LCSSA]]
187 ;
188 ; LOOP-DEL-LABEL: define i32 @compare_bytes_simple(
189 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
190 ; LOOP-DEL-NEXT: entry:
191 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
192 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
193 ; LOOP-DEL-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
194 ; LOOP-DEL-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
195 ; LOOP-DEL-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0:![0-9]+]]
196 ; LOOP-DEL: mismatch_mem_check:
197 ; LOOP-DEL-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
198 ; LOOP-DEL-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
199 ; LOOP-DEL-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
200 ; LOOP-DEL-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
201 ; LOOP-DEL-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
202 ; LOOP-DEL-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
203 ; LOOP-DEL-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
204 ; LOOP-DEL-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
205 ; LOOP-DEL-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
206 ; LOOP-DEL-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
207 ; LOOP-DEL-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
208 ; LOOP-DEL-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
209 ; LOOP-DEL-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
210 ; LOOP-DEL-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
211 ; LOOP-DEL-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
212 ; LOOP-DEL-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP:%.*]], !prof [[PROF1:![0-9]+]]
213 ; LOOP-DEL: mismatch_vec_loop:
214 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ], [ [[TMP1]], [[MISMATCH_MEM_CHECK]] ]
215 ; LOOP-DEL-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
216 ; LOOP-DEL-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
217 ; LOOP-DEL-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
218 ; LOOP-DEL-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
219 ; LOOP-DEL-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
220 ; LOOP-DEL-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
221 ; LOOP-DEL-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
222 ; LOOP-DEL-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
223 ; LOOP-DEL-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
224 ; LOOP-DEL-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
225 ; LOOP-DEL: mismatch_vec_loop_inc:
226 ; LOOP-DEL-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
227 ; LOOP-DEL-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
228 ; LOOP-DEL-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
229 ; LOOP-DEL-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[WHILE_END:%.*]]
230 ; LOOP-DEL: mismatch_vec_loop_found:
231 ; LOOP-DEL-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
232 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
233 ; LOOP-DEL-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
234 ; LOOP-DEL-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
235 ; LOOP-DEL-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
236 ; LOOP-DEL-NEXT: br label [[WHILE_END]]
237 ; LOOP-DEL: mismatch_loop_pre:
238 ; LOOP-DEL-NEXT: br label [[MISMATCH_LOOP:%.*]]
239 ; LOOP-DEL: mismatch_loop:
240 ; LOOP-DEL-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
241 ; LOOP-DEL-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
242 ; LOOP-DEL-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
243 ; LOOP-DEL-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
244 ; LOOP-DEL-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
245 ; LOOP-DEL-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
246 ; LOOP-DEL-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
247 ; LOOP-DEL-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[WHILE_END]]
248 ; LOOP-DEL: mismatch_loop_inc:
249 ; LOOP-DEL-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
250 ; LOOP-DEL-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
251 ; LOOP-DEL-NEXT: br i1 [[TMP36]], label [[WHILE_END]], label [[MISMATCH_LOOP]]
252 ; LOOP-DEL: while.end:
253 ; LOOP-DEL-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
254 ; LOOP-DEL-NEXT: ret i32 [[MISMATCH_RESULT]]
255 ;
256 ; MASKED-LABEL: define i32 @compare_bytes_simple(
257 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
258 ; MASKED-NEXT: entry:
259 ; MASKED-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
260 ; MASKED-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
261 ; MASKED: mismatch_min_it_check:
262 ; MASKED-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
263 ; MASKED-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
264 ; MASKED-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
265 ; MASKED-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0:![0-9]+]]
266 ; MASKED: mismatch_mem_check:
267 ; MASKED-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
268 ; MASKED-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
269 ; MASKED-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP5]] to i64
270 ; MASKED-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
271 ; MASKED-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
272 ; MASKED-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
273 ; MASKED-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP8]] to i64
274 ; MASKED-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP9]] to i64
275 ; MASKED-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP7]], 12
276 ; MASKED-NEXT: [[TMP13:%.*]] = lshr i64 [[TMP10]], 12
277 ; MASKED-NEXT: [[TMP14:%.*]] = lshr i64 [[TMP6]], 12
278 ; MASKED-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP11]], 12
279 ; MASKED-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP12]], [[TMP13]]
280 ; MASKED-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP14]], [[TMP15]]
281 ; MASKED-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
282 ; MASKED-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VEC_LOOP_PREHEADER:%.*]], !prof [[PROF1:![0-9]+]]
283 ; MASKED: mismatch_vec_loop_preheader:
284 ; MASKED-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP1]], i64 [[TMP2]])
285 ; MASKED-NEXT: [[TMP20:%.*]] = call i64 @llvm.vscale.i64()
286 ; MASKED-NEXT: [[TMP21:%.*]] = mul nuw nsw i64 [[TMP20]], 16
287 ; MASKED-NEXT: br label [[MISMATCH_VEC_LOOP:%.*]]
288 ; MASKED: mismatch_vec_loop:
289 ; MASKED-NEXT: [[MISMATCH_VEC_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP19]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP30:%.*]], [[MISMATCH_VEC_LOOP_INC:%.*]] ]
290 ; MASKED-NEXT: [[MISMATCH_VEC_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP29:%.*]], [[MISMATCH_VEC_LOOP_INC]] ]
291 ; MASKED-NEXT: [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VEC_INDEX]]
292 ; MASKED-NEXT: [[TMP23:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP22]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
293 ; MASKED-NEXT: [[TMP24:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VEC_INDEX]]
294 ; MASKED-NEXT: [[TMP25:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP24]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
295 ; MASKED-NEXT: [[TMP26:%.*]] = icmp ne <vscale x 16 x i8> [[TMP23]], [[TMP25]]
296 ; MASKED-NEXT: [[TMP27:%.*]] = select <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i1> [[TMP26]], <vscale x 16 x i1> zeroinitializer
297 ; MASKED-NEXT: [[TMP28:%.*]] = call i1 @llvm.vector.reduce.or.nxv16i1(<vscale x 16 x i1> [[TMP27]])
298 ; MASKED-NEXT: br i1 [[TMP28]], label [[MISMATCH_VEC_LOOP_FOUND:%.*]], label [[MISMATCH_VEC_LOOP_INC]]
299 ; MASKED: mismatch_vec_loop_inc:
300 ; MASKED-NEXT: [[TMP29]] = add nuw nsw i64 [[MISMATCH_VEC_INDEX]], [[TMP21]]
301 ; MASKED-NEXT: [[TMP30]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP29]], i64 [[TMP2]])
302 ; MASKED-NEXT: [[TMP31:%.*]] = extractelement <vscale x 16 x i1> [[TMP30]], i64 0
303 ; MASKED-NEXT: br i1 [[TMP31]], label [[MISMATCH_VEC_LOOP]], label [[MISMATCH_END:%.*]]
304 ; MASKED: mismatch_vec_loop_found:
305 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP27]], [[MISMATCH_VEC_LOOP]] ]
306 ; MASKED-NEXT: [[MISMATCH_VEC_LAST_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[MISMATCH_VEC_LOOP_PRED]], [[MISMATCH_VEC_LOOP]] ]
307 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_INDEX:%.*]] = phi i64 [ [[MISMATCH_VEC_INDEX]], [[MISMATCH_VEC_LOOP]] ]
308 ; MASKED-NEXT: [[TMP32:%.*]] = and <vscale x 16 x i1> [[MISMATCH_VEC_LAST_LOOP_PRED]], [[MISMATCH_VEC_FOUND_PRED]]
309 ; MASKED-NEXT: [[TMP33:%.*]] = call i32 @llvm.experimental.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[TMP32]], i1 true)
310 ; MASKED-NEXT: [[TMP34:%.*]] = zext i32 [[TMP33]] to i64
311 ; MASKED-NEXT: [[TMP35:%.*]] = add nuw nsw i64 [[MISMATCH_VEC_FOUND_INDEX]], [[TMP34]]
312 ; MASKED-NEXT: [[TMP36:%.*]] = trunc i64 [[TMP35]] to i32
313 ; MASKED-NEXT: br label [[MISMATCH_END]]
314 ; MASKED: mismatch_loop_pre:
315 ; MASKED-NEXT: br label [[MISMATCH_LOOP:%.*]]
316 ; MASKED: mismatch_loop:
317 ; MASKED-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP43:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
318 ; MASKED-NEXT: [[TMP37:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
319 ; MASKED-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP37]]
320 ; MASKED-NEXT: [[TMP39:%.*]] = load i8, ptr [[TMP38]], align 1
321 ; MASKED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP37]]
322 ; MASKED-NEXT: [[TMP41:%.*]] = load i8, ptr [[TMP40]], align 1
323 ; MASKED-NEXT: [[TMP42:%.*]] = icmp eq i8 [[TMP39]], [[TMP41]]
324 ; MASKED-NEXT: br i1 [[TMP42]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
325 ; MASKED: mismatch_loop_inc:
326 ; MASKED-NEXT: [[TMP43]] = add i32 [[MISMATCH_INDEX]], 1
327 ; MASKED-NEXT: [[TMP44:%.*]] = icmp eq i32 [[TMP43]], [[N]]
328 ; MASKED-NEXT: br i1 [[TMP44]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
329 ; MASKED: mismatch_end:
330 ; MASKED-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VEC_LOOP_INC]] ], [ [[TMP36]], [[MISMATCH_VEC_LOOP_FOUND]] ]
331 ; MASKED-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
332 ; MASKED: while.cond:
333 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
334 ; MASKED-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
335 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
336 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
337 ; MASKED: while.body:
338 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
339 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
340 ; MASKED-NEXT: [[TMP45:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
341 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
342 ; MASKED-NEXT: [[TMP46:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
343 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP45]], [[TMP46]]
344 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
345 ; MASKED: byte.compare:
346 ; MASKED-NEXT: br label [[WHILE_END]]
347 ; MASKED: while.end:
348 ; MASKED-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
349 ; MASKED-NEXT: ret i32 [[INC_LCSSA]]
350 ;
351 entry:
352 br label %while.cond
354 while.cond:
355 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
356 %inc = add i32 %len.addr, 1
357 %cmp.not = icmp eq i32 %inc, %n
358 br i1 %cmp.not, label %while.end, label %while.body
360 while.body:
361 %idxprom = zext i32 %inc to i64
362 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
363 %0 = load i8, ptr %arrayidx
364 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
365 %1 = load i8, ptr %arrayidx2
366 %cmp.not2 = icmp eq i8 %0, %1
367 br i1 %cmp.not2, label %while.cond, label %while.end
369 while.end:
370 %inc.lcssa = phi i32 [ %inc, %while.body ], [ %inc, %while.cond ]
371 ret i32 %inc.lcssa
372 }
374 define i32 @compare_bytes_signed_wrap(ptr %a, ptr %b, i32 %len, i32 %n) {
375 ; CHECK-LABEL: define i32 @compare_bytes_signed_wrap(
376 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
377 ; CHECK-NEXT: entry:
378 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
379 ; CHECK-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
380 ; CHECK: mismatch_min_it_check:
381 ; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
382 ; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
383 ; CHECK-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
384 ; CHECK-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
385 ; CHECK: mismatch_mem_check:
386 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
387 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
388 ; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
389 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
390 ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
391 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
392 ; CHECK-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
393 ; CHECK-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
394 ; CHECK-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
395 ; CHECK-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
396 ; CHECK-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
397 ; CHECK-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
398 ; CHECK-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
399 ; CHECK-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
400 ; CHECK-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
401 ; CHECK-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
402 ; CHECK: mismatch_vec_loop_preheader:
403 ; CHECK-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
404 ; CHECK: mismatch_vec_loop:
405 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
406 ; CHECK-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
407 ; CHECK-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
408 ; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
409 ; CHECK-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
410 ; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
411 ; CHECK-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
412 ; CHECK-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
413 ; CHECK-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
414 ; CHECK-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
415 ; CHECK-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
416 ; CHECK: mismatch_vec_loop_inc:
417 ; CHECK-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
418 ; CHECK-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
419 ; CHECK-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
420 ; CHECK-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
421 ; CHECK: mismatch_vec_loop_found:
422 ; CHECK-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
423 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
424 ; CHECK-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
425 ; CHECK-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
426 ; CHECK-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
427 ; CHECK-NEXT: br label [[MISMATCH_END]]
428 ; CHECK: mismatch_loop_pre:
429 ; CHECK-NEXT: br label [[MISMATCH_LOOP:%.*]]
430 ; CHECK: mismatch_loop:
431 ; CHECK-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
432 ; CHECK-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
433 ; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
434 ; CHECK-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
435 ; CHECK-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
436 ; CHECK-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
437 ; CHECK-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
438 ; CHECK-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
439 ; CHECK: mismatch_loop_inc:
440 ; CHECK-NEXT: [[TMP35]] = add nsw i32 [[MISMATCH_INDEX]], 1
441 ; CHECK-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
442 ; CHECK-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
443 ; CHECK: mismatch_end:
444 ; CHECK-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
445 ; CHECK-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
446 ; CHECK: while.cond:
447 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
448 ; CHECK-NEXT: [[INC:%.*]] = add nsw i32 [[LEN_ADDR]], 1
449 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
450 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
451 ; CHECK: while.body:
452 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
453 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
454 ; CHECK-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
455 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
456 ; CHECK-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
457 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
458 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
459 ; CHECK: byte.compare:
460 ; CHECK-NEXT: br label [[WHILE_END]]
461 ; CHECK: while.end:
462 ; CHECK-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
463 ; CHECK-NEXT: ret i32 [[INC_LCSSA]]
464 ;
465 ; LMUL8-LABEL: define i32 @compare_bytes_signed_wrap(
466 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
467 ; LMUL8-NEXT: entry:
468 ; LMUL8-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
469 ; LMUL8-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
470 ; LMUL8: mismatch_min_it_check:
471 ; LMUL8-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
472 ; LMUL8-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
473 ; LMUL8-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
474 ; LMUL8-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
475 ; LMUL8: mismatch_mem_check:
476 ; LMUL8-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
477 ; LMUL8-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
478 ; LMUL8-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
479 ; LMUL8-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
480 ; LMUL8-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
481 ; LMUL8-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
482 ; LMUL8-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
483 ; LMUL8-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
484 ; LMUL8-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
485 ; LMUL8-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
486 ; LMUL8-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
487 ; LMUL8-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
488 ; LMUL8-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
489 ; LMUL8-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
490 ; LMUL8-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
491 ; LMUL8-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
492 ; LMUL8: mismatch_vec_loop_preheader:
493 ; LMUL8-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
494 ; LMUL8: mismatch_vec_loop:
495 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
496 ; LMUL8-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
497 ; LMUL8-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 64, i1 true)
498 ; LMUL8-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
499 ; LMUL8-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP20]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
500 ; LMUL8-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
501 ; LMUL8-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP21]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
502 ; LMUL8-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 64 x i1> @llvm.vp.icmp.nxv64i8(<vscale x 64 x i8> [[LHS_LOAD]], <vscale x 64 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
503 ; LMUL8-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv64i1(<vscale x 64 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
504 ; LMUL8-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
505 ; LMUL8-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
506 ; LMUL8: mismatch_vec_loop_inc:
507 ; LMUL8-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
508 ; LMUL8-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
509 ; LMUL8-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
510 ; LMUL8-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
511 ; LMUL8: mismatch_vec_loop_found:
512 ; LMUL8-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
513 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
514 ; LMUL8-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
515 ; LMUL8-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
516 ; LMUL8-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
517 ; LMUL8-NEXT: br label [[MISMATCH_END]]
518 ; LMUL8: mismatch_loop_pre:
519 ; LMUL8-NEXT: br label [[MISMATCH_LOOP:%.*]]
520 ; LMUL8: mismatch_loop:
521 ; LMUL8-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
522 ; LMUL8-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
523 ; LMUL8-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
524 ; LMUL8-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
525 ; LMUL8-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
526 ; LMUL8-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
527 ; LMUL8-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
528 ; LMUL8-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
529 ; LMUL8: mismatch_loop_inc:
530 ; LMUL8-NEXT: [[TMP35]] = add nsw i32 [[MISMATCH_INDEX]], 1
531 ; LMUL8-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
532 ; LMUL8-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
533 ; LMUL8: mismatch_end:
534 ; LMUL8-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
535 ; LMUL8-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
536 ; LMUL8: while.cond:
537 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
538 ; LMUL8-NEXT: [[INC:%.*]] = add nsw i32 [[LEN_ADDR]], 1
539 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
540 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
541 ; LMUL8: while.body:
542 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
543 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
544 ; LMUL8-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
545 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
546 ; LMUL8-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
547 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
548 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
549 ; LMUL8: byte.compare:
550 ; LMUL8-NEXT: br label [[WHILE_END]]
551 ; LMUL8: while.end:
552 ; LMUL8-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
553 ; LMUL8-NEXT: ret i32 [[INC_LCSSA]]
554 ;
555 ; LOOP-DEL-LABEL: define i32 @compare_bytes_signed_wrap(
556 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
557 ; LOOP-DEL-NEXT: entry:
558 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
559 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
560 ; LOOP-DEL-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
561 ; LOOP-DEL-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
562 ; LOOP-DEL-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
563 ; LOOP-DEL: mismatch_mem_check:
564 ; LOOP-DEL-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
565 ; LOOP-DEL-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
566 ; LOOP-DEL-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
567 ; LOOP-DEL-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
568 ; LOOP-DEL-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
569 ; LOOP-DEL-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
570 ; LOOP-DEL-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
571 ; LOOP-DEL-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
572 ; LOOP-DEL-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
573 ; LOOP-DEL-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
574 ; LOOP-DEL-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
575 ; LOOP-DEL-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
576 ; LOOP-DEL-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
577 ; LOOP-DEL-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
578 ; LOOP-DEL-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
579 ; LOOP-DEL-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP:%.*]], !prof [[PROF1]]
580 ; LOOP-DEL: mismatch_vec_loop:
581 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ], [ [[TMP1]], [[MISMATCH_MEM_CHECK]] ]
582 ; LOOP-DEL-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
583 ; LOOP-DEL-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
584 ; LOOP-DEL-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
585 ; LOOP-DEL-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
586 ; LOOP-DEL-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
587 ; LOOP-DEL-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
588 ; LOOP-DEL-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
589 ; LOOP-DEL-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
590 ; LOOP-DEL-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
591 ; LOOP-DEL-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
592 ; LOOP-DEL: mismatch_vec_loop_inc:
593 ; LOOP-DEL-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
594 ; LOOP-DEL-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
595 ; LOOP-DEL-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
596 ; LOOP-DEL-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[WHILE_END:%.*]]
597 ; LOOP-DEL: mismatch_vec_loop_found:
598 ; LOOP-DEL-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
599 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
600 ; LOOP-DEL-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
601 ; LOOP-DEL-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
602 ; LOOP-DEL-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
603 ; LOOP-DEL-NEXT: br label [[WHILE_END]]
604 ; LOOP-DEL: mismatch_loop_pre:
605 ; LOOP-DEL-NEXT: br label [[MISMATCH_LOOP:%.*]]
606 ; LOOP-DEL: mismatch_loop:
607 ; LOOP-DEL-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
608 ; LOOP-DEL-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
609 ; LOOP-DEL-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
610 ; LOOP-DEL-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
611 ; LOOP-DEL-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
612 ; LOOP-DEL-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
613 ; LOOP-DEL-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
614 ; LOOP-DEL-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[WHILE_END]]
615 ; LOOP-DEL: mismatch_loop_inc:
616 ; LOOP-DEL-NEXT: [[TMP35]] = add nsw i32 [[MISMATCH_INDEX]], 1
617 ; LOOP-DEL-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
618 ; LOOP-DEL-NEXT: br i1 [[TMP36]], label [[WHILE_END]], label [[MISMATCH_LOOP]]
619 ; LOOP-DEL: while.end:
620 ; LOOP-DEL-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
621 ; LOOP-DEL-NEXT: ret i32 [[MISMATCH_RESULT]]
622 ;
623 ; MASKED-LABEL: define i32 @compare_bytes_signed_wrap(
624 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
625 ; MASKED-NEXT: entry:
626 ; MASKED-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
627 ; MASKED-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
628 ; MASKED: mismatch_min_it_check:
629 ; MASKED-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
630 ; MASKED-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
631 ; MASKED-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
632 ; MASKED-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
633 ; MASKED: mismatch_mem_check:
634 ; MASKED-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
635 ; MASKED-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
636 ; MASKED-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP5]] to i64
637 ; MASKED-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
638 ; MASKED-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
639 ; MASKED-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
640 ; MASKED-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP8]] to i64
641 ; MASKED-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP9]] to i64
642 ; MASKED-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP7]], 12
643 ; MASKED-NEXT: [[TMP13:%.*]] = lshr i64 [[TMP10]], 12
644 ; MASKED-NEXT: [[TMP14:%.*]] = lshr i64 [[TMP6]], 12
645 ; MASKED-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP11]], 12
646 ; MASKED-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP12]], [[TMP13]]
647 ; MASKED-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP14]], [[TMP15]]
648 ; MASKED-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
649 ; MASKED-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VEC_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
650 ; MASKED: mismatch_vec_loop_preheader:
651 ; MASKED-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP1]], i64 [[TMP2]])
652 ; MASKED-NEXT: [[TMP20:%.*]] = call i64 @llvm.vscale.i64()
653 ; MASKED-NEXT: [[TMP21:%.*]] = mul nuw nsw i64 [[TMP20]], 16
654 ; MASKED-NEXT: br label [[MISMATCH_VEC_LOOP:%.*]]
655 ; MASKED: mismatch_vec_loop:
656 ; MASKED-NEXT: [[MISMATCH_VEC_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP19]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP30:%.*]], [[MISMATCH_VEC_LOOP_INC:%.*]] ]
657 ; MASKED-NEXT: [[MISMATCH_VEC_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP29:%.*]], [[MISMATCH_VEC_LOOP_INC]] ]
658 ; MASKED-NEXT: [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VEC_INDEX]]
659 ; MASKED-NEXT: [[TMP23:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP22]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
660 ; MASKED-NEXT: [[TMP24:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VEC_INDEX]]
661 ; MASKED-NEXT: [[TMP25:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP24]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
662 ; MASKED-NEXT: [[TMP26:%.*]] = icmp ne <vscale x 16 x i8> [[TMP23]], [[TMP25]]
663 ; MASKED-NEXT: [[TMP27:%.*]] = select <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i1> [[TMP26]], <vscale x 16 x i1> zeroinitializer
664 ; MASKED-NEXT: [[TMP28:%.*]] = call i1 @llvm.vector.reduce.or.nxv16i1(<vscale x 16 x i1> [[TMP27]])
665 ; MASKED-NEXT: br i1 [[TMP28]], label [[MISMATCH_VEC_LOOP_FOUND:%.*]], label [[MISMATCH_VEC_LOOP_INC]]
666 ; MASKED: mismatch_vec_loop_inc:
667 ; MASKED-NEXT: [[TMP29]] = add nuw nsw i64 [[MISMATCH_VEC_INDEX]], [[TMP21]]
668 ; MASKED-NEXT: [[TMP30]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP29]], i64 [[TMP2]])
669 ; MASKED-NEXT: [[TMP31:%.*]] = extractelement <vscale x 16 x i1> [[TMP30]], i64 0
670 ; MASKED-NEXT: br i1 [[TMP31]], label [[MISMATCH_VEC_LOOP]], label [[MISMATCH_END:%.*]]
671 ; MASKED: mismatch_vec_loop_found:
672 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP27]], [[MISMATCH_VEC_LOOP]] ]
673 ; MASKED-NEXT: [[MISMATCH_VEC_LAST_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[MISMATCH_VEC_LOOP_PRED]], [[MISMATCH_VEC_LOOP]] ]
674 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_INDEX:%.*]] = phi i64 [ [[MISMATCH_VEC_INDEX]], [[MISMATCH_VEC_LOOP]] ]
675 ; MASKED-NEXT: [[TMP32:%.*]] = and <vscale x 16 x i1> [[MISMATCH_VEC_LAST_LOOP_PRED]], [[MISMATCH_VEC_FOUND_PRED]]
676 ; MASKED-NEXT: [[TMP33:%.*]] = call i32 @llvm.experimental.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[TMP32]], i1 true)
677 ; MASKED-NEXT: [[TMP34:%.*]] = zext i32 [[TMP33]] to i64
678 ; MASKED-NEXT: [[TMP35:%.*]] = add nuw nsw i64 [[MISMATCH_VEC_FOUND_INDEX]], [[TMP34]]
679 ; MASKED-NEXT: [[TMP36:%.*]] = trunc i64 [[TMP35]] to i32
680 ; MASKED-NEXT: br label [[MISMATCH_END]]
681 ; MASKED: mismatch_loop_pre:
682 ; MASKED-NEXT: br label [[MISMATCH_LOOP:%.*]]
683 ; MASKED: mismatch_loop:
684 ; MASKED-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP43:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
685 ; MASKED-NEXT: [[TMP37:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
686 ; MASKED-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP37]]
687 ; MASKED-NEXT: [[TMP39:%.*]] = load i8, ptr [[TMP38]], align 1
688 ; MASKED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP37]]
689 ; MASKED-NEXT: [[TMP41:%.*]] = load i8, ptr [[TMP40]], align 1
690 ; MASKED-NEXT: [[TMP42:%.*]] = icmp eq i8 [[TMP39]], [[TMP41]]
691 ; MASKED-NEXT: br i1 [[TMP42]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
692 ; MASKED: mismatch_loop_inc:
693 ; MASKED-NEXT: [[TMP43]] = add nsw i32 [[MISMATCH_INDEX]], 1
694 ; MASKED-NEXT: [[TMP44:%.*]] = icmp eq i32 [[TMP43]], [[N]]
695 ; MASKED-NEXT: br i1 [[TMP44]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
696 ; MASKED: mismatch_end:
697 ; MASKED-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VEC_LOOP_INC]] ], [ [[TMP36]], [[MISMATCH_VEC_LOOP_FOUND]] ]
698 ; MASKED-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
699 ; MASKED: while.cond:
700 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
701 ; MASKED-NEXT: [[INC:%.*]] = add nsw i32 [[LEN_ADDR]], 1
702 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
703 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
704 ; MASKED: while.body:
705 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
706 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
707 ; MASKED-NEXT: [[TMP45:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
708 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
709 ; MASKED-NEXT: [[TMP46:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
710 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP45]], [[TMP46]]
711 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
712 ; MASKED: byte.compare:
713 ; MASKED-NEXT: br label [[WHILE_END]]
714 ; MASKED: while.end:
715 ; MASKED-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
716 ; MASKED-NEXT: ret i32 [[INC_LCSSA]]
717 ;
718 ; NO-TRANSFORM-LABEL: define i32 @compare_bytes_signed_wrap(
719 ; NO-TRANSFORM-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) {
720 ; NO-TRANSFORM-NEXT: entry:
721 ; NO-TRANSFORM-NEXT: br label [[WHILE_COND:%.*]]
722 ; NO-TRANSFORM: while.cond:
723 ; NO-TRANSFORM-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
724 ; NO-TRANSFORM-NEXT: [[INC]] = add nsw i32 [[LEN_ADDR]], 1
725 ; NO-TRANSFORM-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
726 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
727 ; NO-TRANSFORM: while.body:
728 ; NO-TRANSFORM-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
729 ; NO-TRANSFORM-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
730 ; NO-TRANSFORM-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
731 ; NO-TRANSFORM-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
732 ; NO-TRANSFORM-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
733 ; NO-TRANSFORM-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
734 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
735 ; NO-TRANSFORM: while.end:
736 ; NO-TRANSFORM-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
737 ; NO-TRANSFORM-NEXT: ret i32 [[INC_LCSSA]]
738 entry:
739 br label %while.cond
741 while.cond:
742 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
743 %inc = add nsw i32 %len.addr, 1
744 %cmp.not = icmp eq i32 %inc, %n
745 br i1 %cmp.not, label %while.end, label %while.body
747 while.body:
748 %idxprom = zext i32 %inc to i64
749 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
750 %0 = load i8, ptr %arrayidx
751 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
752 %1 = load i8, ptr %arrayidx2
753 %cmp.not2 = icmp eq i8 %0, %1
754 br i1 %cmp.not2, label %while.cond, label %while.end
756 while.end:
757 %inc.lcssa = phi i32 [ %inc, %while.body ], [ %inc, %while.cond ]
758 ret i32 %inc.lcssa
759 }
762 define i32 @compare_bytes_simple_end_ne_found(ptr %a, ptr %b, ptr %c, ptr %d, i32 %len, i32 %n) {
763 ; CHECK-LABEL: define i32 @compare_bytes_simple_end_ne_found(
764 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
765 ; CHECK-NEXT: entry:
766 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
767 ; CHECK-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
768 ; CHECK: mismatch_min_it_check:
769 ; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
770 ; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
771 ; CHECK-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
772 ; CHECK-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
773 ; CHECK: mismatch_mem_check:
774 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
775 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
776 ; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
777 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
778 ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
779 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
780 ; CHECK-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
781 ; CHECK-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
782 ; CHECK-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
783 ; CHECK-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
784 ; CHECK-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
785 ; CHECK-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
786 ; CHECK-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
787 ; CHECK-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
788 ; CHECK-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
789 ; CHECK-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
790 ; CHECK: mismatch_vec_loop_preheader:
791 ; CHECK-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
792 ; CHECK: mismatch_vec_loop:
793 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
794 ; CHECK-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
795 ; CHECK-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
796 ; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
797 ; CHECK-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
798 ; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
799 ; CHECK-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
800 ; CHECK-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
801 ; CHECK-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
802 ; CHECK-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
803 ; CHECK-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
804 ; CHECK: mismatch_vec_loop_inc:
805 ; CHECK-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
806 ; CHECK-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
807 ; CHECK-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
808 ; CHECK-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
809 ; CHECK: mismatch_vec_loop_found:
810 ; CHECK-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
811 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
812 ; CHECK-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
813 ; CHECK-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
814 ; CHECK-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
815 ; CHECK-NEXT: br label [[MISMATCH_END]]
816 ; CHECK: mismatch_loop_pre:
817 ; CHECK-NEXT: br label [[MISMATCH_LOOP:%.*]]
818 ; CHECK: mismatch_loop:
819 ; CHECK-NEXT: [[MISMATCH_INDEX3:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
820 ; CHECK-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX3]] to i64
821 ; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
822 ; CHECK-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
823 ; CHECK-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
824 ; CHECK-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
825 ; CHECK-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
826 ; CHECK-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
827 ; CHECK: mismatch_loop_inc:
828 ; CHECK-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX3]], 1
829 ; CHECK-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
830 ; CHECK-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
831 ; CHECK: mismatch_end:
832 ; CHECK-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX3]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
833 ; CHECK-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
834 ; CHECK: while.cond:
835 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
836 ; CHECK-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
837 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
838 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
839 ; CHECK: while.body:
840 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
841 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
842 ; CHECK-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
843 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
844 ; CHECK-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
845 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
846 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_FOUND:%.*]]
847 ; CHECK: while.found:
848 ; CHECK-NEXT: [[MISMATCH_INDEX1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
849 ; CHECK-NEXT: [[FOUND_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[C]], [[BYTE_COMPARE]] ]
850 ; CHECK-NEXT: br label [[END:%.*]]
851 ; CHECK: byte.compare:
852 ; CHECK-NEXT: [[TMP39:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
853 ; CHECK-NEXT: br i1 [[TMP39]], label [[WHILE_END]], label [[WHILE_FOUND]]
854 ; CHECK: while.end:
855 ; CHECK-NEXT: [[MISMATCH_INDEX2:%.*]] = phi i32 [ [[N]], [[WHILE_COND]] ], [ [[N]], [[BYTE_COMPARE]] ]
856 ; CHECK-NEXT: [[END_PTR:%.*]] = phi ptr [ [[D]], [[WHILE_COND]] ], [ [[D]], [[BYTE_COMPARE]] ]
857 ; CHECK-NEXT: br label [[END]]
858 ; CHECK: end:
859 ; CHECK-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[MISMATCH_INDEX1]], [[WHILE_FOUND]] ], [ [[MISMATCH_INDEX2]], [[WHILE_END]] ]
860 ; CHECK-NEXT: [[STORE_PTR:%.*]] = phi ptr [ [[END_PTR]], [[WHILE_END]] ], [ [[FOUND_PTR]], [[WHILE_FOUND]] ]
861 ; CHECK-NEXT: store i32 [[MISMATCH_INDEX]], ptr [[STORE_PTR]], align 4
862 ; CHECK-NEXT: ret i32 [[MISMATCH_INDEX]]
863 ;
864 ; LMUL8-LABEL: define i32 @compare_bytes_simple_end_ne_found(
865 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
866 ; LMUL8-NEXT: entry:
867 ; LMUL8-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
868 ; LMUL8-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
869 ; LMUL8: mismatch_min_it_check:
870 ; LMUL8-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
871 ; LMUL8-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
872 ; LMUL8-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
873 ; LMUL8-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
874 ; LMUL8: mismatch_mem_check:
875 ; LMUL8-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
876 ; LMUL8-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
877 ; LMUL8-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
878 ; LMUL8-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
879 ; LMUL8-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
880 ; LMUL8-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
881 ; LMUL8-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
882 ; LMUL8-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
883 ; LMUL8-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
884 ; LMUL8-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
885 ; LMUL8-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
886 ; LMUL8-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
887 ; LMUL8-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
888 ; LMUL8-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
889 ; LMUL8-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
890 ; LMUL8-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
891 ; LMUL8: mismatch_vec_loop_preheader:
892 ; LMUL8-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
893 ; LMUL8: mismatch_vec_loop:
894 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
895 ; LMUL8-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
896 ; LMUL8-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 64, i1 true)
897 ; LMUL8-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
898 ; LMUL8-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP20]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
899 ; LMUL8-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
900 ; LMUL8-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP21]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
901 ; LMUL8-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 64 x i1> @llvm.vp.icmp.nxv64i8(<vscale x 64 x i8> [[LHS_LOAD]], <vscale x 64 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
902 ; LMUL8-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv64i1(<vscale x 64 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
903 ; LMUL8-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
904 ; LMUL8-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
905 ; LMUL8: mismatch_vec_loop_inc:
906 ; LMUL8-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
907 ; LMUL8-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
908 ; LMUL8-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
909 ; LMUL8-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
910 ; LMUL8: mismatch_vec_loop_found:
911 ; LMUL8-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
912 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
913 ; LMUL8-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
914 ; LMUL8-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
915 ; LMUL8-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
916 ; LMUL8-NEXT: br label [[MISMATCH_END]]
917 ; LMUL8: mismatch_loop_pre:
918 ; LMUL8-NEXT: br label [[MISMATCH_LOOP:%.*]]
919 ; LMUL8: mismatch_loop:
920 ; LMUL8-NEXT: [[MISMATCH_INDEX3:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
921 ; LMUL8-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX3]] to i64
922 ; LMUL8-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
923 ; LMUL8-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
924 ; LMUL8-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
925 ; LMUL8-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
926 ; LMUL8-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
927 ; LMUL8-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
928 ; LMUL8: mismatch_loop_inc:
929 ; LMUL8-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX3]], 1
930 ; LMUL8-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
931 ; LMUL8-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
932 ; LMUL8: mismatch_end:
933 ; LMUL8-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX3]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
934 ; LMUL8-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
935 ; LMUL8: while.cond:
936 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
937 ; LMUL8-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
938 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
939 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
940 ; LMUL8: while.body:
941 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
942 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
943 ; LMUL8-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
944 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
945 ; LMUL8-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
946 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
947 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_FOUND:%.*]]
948 ; LMUL8: while.found:
949 ; LMUL8-NEXT: [[MISMATCH_INDEX1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
950 ; LMUL8-NEXT: [[FOUND_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[C]], [[BYTE_COMPARE]] ]
951 ; LMUL8-NEXT: br label [[END:%.*]]
952 ; LMUL8: byte.compare:
953 ; LMUL8-NEXT: [[TMP39:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
954 ; LMUL8-NEXT: br i1 [[TMP39]], label [[WHILE_END]], label [[WHILE_FOUND]]
955 ; LMUL8: while.end:
956 ; LMUL8-NEXT: [[MISMATCH_INDEX2:%.*]] = phi i32 [ [[N]], [[WHILE_COND]] ], [ [[N]], [[BYTE_COMPARE]] ]
957 ; LMUL8-NEXT: [[END_PTR:%.*]] = phi ptr [ [[D]], [[WHILE_COND]] ], [ [[D]], [[BYTE_COMPARE]] ]
958 ; LMUL8-NEXT: br label [[END]]
959 ; LMUL8: end:
960 ; LMUL8-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[MISMATCH_INDEX1]], [[WHILE_FOUND]] ], [ [[MISMATCH_INDEX2]], [[WHILE_END]] ]
961 ; LMUL8-NEXT: [[STORE_PTR:%.*]] = phi ptr [ [[END_PTR]], [[WHILE_END]] ], [ [[FOUND_PTR]], [[WHILE_FOUND]] ]
962 ; LMUL8-NEXT: store i32 [[MISMATCH_INDEX]], ptr [[STORE_PTR]], align 4
963 ; LMUL8-NEXT: ret i32 [[MISMATCH_INDEX]]
964 ;
965 ; LOOP-DEL-LABEL: define i32 @compare_bytes_simple_end_ne_found(
966 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
967 ; LOOP-DEL-NEXT: entry:
968 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
969 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
970 ; LOOP-DEL-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
971 ; LOOP-DEL-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
972 ; LOOP-DEL-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
973 ; LOOP-DEL: mismatch_mem_check:
974 ; LOOP-DEL-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
975 ; LOOP-DEL-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
976 ; LOOP-DEL-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
977 ; LOOP-DEL-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
978 ; LOOP-DEL-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
979 ; LOOP-DEL-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
980 ; LOOP-DEL-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
981 ; LOOP-DEL-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
982 ; LOOP-DEL-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
983 ; LOOP-DEL-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
984 ; LOOP-DEL-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
985 ; LOOP-DEL-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
986 ; LOOP-DEL-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
987 ; LOOP-DEL-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
988 ; LOOP-DEL-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
989 ; LOOP-DEL-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP:%.*]], !prof [[PROF1]]
990 ; LOOP-DEL: mismatch_vec_loop:
991 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ], [ [[TMP1]], [[MISMATCH_MEM_CHECK]] ]
992 ; LOOP-DEL-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
993 ; LOOP-DEL-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
994 ; LOOP-DEL-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
995 ; LOOP-DEL-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
996 ; LOOP-DEL-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
997 ; LOOP-DEL-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
998 ; LOOP-DEL-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
999 ; LOOP-DEL-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1000 ; LOOP-DEL-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
1001 ; LOOP-DEL-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1002 ; LOOP-DEL: mismatch_vec_loop_inc:
1003 ; LOOP-DEL-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
1004 ; LOOP-DEL-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
1005 ; LOOP-DEL-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
1006 ; LOOP-DEL-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[BYTE_COMPARE:%.*]]
1007 ; LOOP-DEL: mismatch_vec_loop_found:
1008 ; LOOP-DEL-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1009 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1010 ; LOOP-DEL-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST1]] to i64
1011 ; LOOP-DEL-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP26]]
1012 ; LOOP-DEL-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
1013 ; LOOP-DEL-NEXT: br label [[BYTE_COMPARE]]
1014 ; LOOP-DEL: mismatch_loop_pre:
1015 ; LOOP-DEL-NEXT: br label [[MISMATCH_LOOP:%.*]]
1016 ; LOOP-DEL: mismatch_loop:
1017 ; LOOP-DEL-NEXT: [[MISMATCH_INDEX3:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1018 ; LOOP-DEL-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX3]] to i64
1019 ; LOOP-DEL-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
1020 ; LOOP-DEL-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
1021 ; LOOP-DEL-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
1022 ; LOOP-DEL-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
1023 ; LOOP-DEL-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
1024 ; LOOP-DEL-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[BYTE_COMPARE]]
1025 ; LOOP-DEL: mismatch_loop_inc:
1026 ; LOOP-DEL-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX3]], 1
1027 ; LOOP-DEL-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
1028 ; LOOP-DEL-NEXT: br i1 [[TMP36]], label [[BYTE_COMPARE]], label [[MISMATCH_LOOP]]
1029 ; LOOP-DEL: byte.compare:
1030 ; LOOP-DEL-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX3]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1031 ; LOOP-DEL-NEXT: [[TMP37:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1032 ; LOOP-DEL-NEXT: [[SPEC_SELECT:%.*]] = select i1 [[TMP37]], i32 [[N]], i32 [[MISMATCH_RESULT]]
1033 ; LOOP-DEL-NEXT: [[SPEC_SELECT4:%.*]] = select i1 [[TMP37]], ptr [[D]], ptr [[C]]
1034 ; LOOP-DEL-NEXT: store i32 [[SPEC_SELECT]], ptr [[SPEC_SELECT4]], align 4
1035 ; LOOP-DEL-NEXT: ret i32 [[SPEC_SELECT]]
1036 ;
1037 ; MASKED-LABEL: define i32 @compare_bytes_simple_end_ne_found(
1038 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
1039 ; MASKED-NEXT: entry:
1040 ; MASKED-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
1041 ; MASKED-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1042 ; MASKED: mismatch_min_it_check:
1043 ; MASKED-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1044 ; MASKED-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
1045 ; MASKED-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
1046 ; MASKED-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1047 ; MASKED: mismatch_mem_check:
1048 ; MASKED-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
1049 ; MASKED-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1050 ; MASKED-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP5]] to i64
1051 ; MASKED-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
1052 ; MASKED-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
1053 ; MASKED-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1054 ; MASKED-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP8]] to i64
1055 ; MASKED-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP9]] to i64
1056 ; MASKED-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP7]], 12
1057 ; MASKED-NEXT: [[TMP13:%.*]] = lshr i64 [[TMP10]], 12
1058 ; MASKED-NEXT: [[TMP14:%.*]] = lshr i64 [[TMP6]], 12
1059 ; MASKED-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP11]], 12
1060 ; MASKED-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP12]], [[TMP13]]
1061 ; MASKED-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP14]], [[TMP15]]
1062 ; MASKED-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
1063 ; MASKED-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VEC_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1064 ; MASKED: mismatch_vec_loop_preheader:
1065 ; MASKED-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP1]], i64 [[TMP2]])
1066 ; MASKED-NEXT: [[TMP20:%.*]] = call i64 @llvm.vscale.i64()
1067 ; MASKED-NEXT: [[TMP21:%.*]] = mul nuw nsw i64 [[TMP20]], 16
1068 ; MASKED-NEXT: br label [[MISMATCH_VEC_LOOP:%.*]]
1069 ; MASKED: mismatch_vec_loop:
1070 ; MASKED-NEXT: [[MISMATCH_VEC_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP19]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP30:%.*]], [[MISMATCH_VEC_LOOP_INC:%.*]] ]
1071 ; MASKED-NEXT: [[MISMATCH_VEC_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP29:%.*]], [[MISMATCH_VEC_LOOP_INC]] ]
1072 ; MASKED-NEXT: [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VEC_INDEX]]
1073 ; MASKED-NEXT: [[TMP23:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP22]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1074 ; MASKED-NEXT: [[TMP24:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VEC_INDEX]]
1075 ; MASKED-NEXT: [[TMP25:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP24]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1076 ; MASKED-NEXT: [[TMP26:%.*]] = icmp ne <vscale x 16 x i8> [[TMP23]], [[TMP25]]
1077 ; MASKED-NEXT: [[TMP27:%.*]] = select <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i1> [[TMP26]], <vscale x 16 x i1> zeroinitializer
1078 ; MASKED-NEXT: [[TMP28:%.*]] = call i1 @llvm.vector.reduce.or.nxv16i1(<vscale x 16 x i1> [[TMP27]])
1079 ; MASKED-NEXT: br i1 [[TMP28]], label [[MISMATCH_VEC_LOOP_FOUND:%.*]], label [[MISMATCH_VEC_LOOP_INC]]
1080 ; MASKED: mismatch_vec_loop_inc:
1081 ; MASKED-NEXT: [[TMP29]] = add nuw nsw i64 [[MISMATCH_VEC_INDEX]], [[TMP21]]
1082 ; MASKED-NEXT: [[TMP30]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP29]], i64 [[TMP2]])
1083 ; MASKED-NEXT: [[TMP31:%.*]] = extractelement <vscale x 16 x i1> [[TMP30]], i64 0
1084 ; MASKED-NEXT: br i1 [[TMP31]], label [[MISMATCH_VEC_LOOP]], label [[MISMATCH_END:%.*]]
1085 ; MASKED: mismatch_vec_loop_found:
1086 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP27]], [[MISMATCH_VEC_LOOP]] ]
1087 ; MASKED-NEXT: [[MISMATCH_VEC_LAST_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[MISMATCH_VEC_LOOP_PRED]], [[MISMATCH_VEC_LOOP]] ]
1088 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_INDEX:%.*]] = phi i64 [ [[MISMATCH_VEC_INDEX]], [[MISMATCH_VEC_LOOP]] ]
1089 ; MASKED-NEXT: [[TMP32:%.*]] = and <vscale x 16 x i1> [[MISMATCH_VEC_LAST_LOOP_PRED]], [[MISMATCH_VEC_FOUND_PRED]]
1090 ; MASKED-NEXT: [[TMP33:%.*]] = call i32 @llvm.experimental.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[TMP32]], i1 true)
1091 ; MASKED-NEXT: [[TMP34:%.*]] = zext i32 [[TMP33]] to i64
1092 ; MASKED-NEXT: [[TMP35:%.*]] = add nuw nsw i64 [[MISMATCH_VEC_FOUND_INDEX]], [[TMP34]]
1093 ; MASKED-NEXT: [[TMP36:%.*]] = trunc i64 [[TMP35]] to i32
1094 ; MASKED-NEXT: br label [[MISMATCH_END]]
1095 ; MASKED: mismatch_loop_pre:
1096 ; MASKED-NEXT: br label [[MISMATCH_LOOP:%.*]]
1097 ; MASKED: mismatch_loop:
1098 ; MASKED-NEXT: [[MISMATCH_INDEX3:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP43:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1099 ; MASKED-NEXT: [[TMP37:%.*]] = zext i32 [[MISMATCH_INDEX3]] to i64
1100 ; MASKED-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP37]]
1101 ; MASKED-NEXT: [[TMP39:%.*]] = load i8, ptr [[TMP38]], align 1
1102 ; MASKED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP37]]
1103 ; MASKED-NEXT: [[TMP41:%.*]] = load i8, ptr [[TMP40]], align 1
1104 ; MASKED-NEXT: [[TMP42:%.*]] = icmp eq i8 [[TMP39]], [[TMP41]]
1105 ; MASKED-NEXT: br i1 [[TMP42]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1106 ; MASKED: mismatch_loop_inc:
1107 ; MASKED-NEXT: [[TMP43]] = add i32 [[MISMATCH_INDEX3]], 1
1108 ; MASKED-NEXT: [[TMP44:%.*]] = icmp eq i32 [[TMP43]], [[N]]
1109 ; MASKED-NEXT: br i1 [[TMP44]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1110 ; MASKED: mismatch_end:
1111 ; MASKED-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX3]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VEC_LOOP_INC]] ], [ [[TMP36]], [[MISMATCH_VEC_LOOP_FOUND]] ]
1112 ; MASKED-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1113 ; MASKED: while.cond:
1114 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
1115 ; MASKED-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
1116 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1117 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
1118 ; MASKED: while.body:
1119 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1120 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1121 ; MASKED-NEXT: [[TMP45:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1122 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1123 ; MASKED-NEXT: [[TMP46:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1124 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP45]], [[TMP46]]
1125 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_FOUND:%.*]]
1126 ; MASKED: while.found:
1127 ; MASKED-NEXT: [[MISMATCH_INDEX1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1128 ; MASKED-NEXT: [[FOUND_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[C]], [[BYTE_COMPARE]] ]
1129 ; MASKED-NEXT: br label [[END:%.*]]
1130 ; MASKED: byte.compare:
1131 ; MASKED-NEXT: [[TMP47:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1132 ; MASKED-NEXT: br i1 [[TMP47]], label [[WHILE_END]], label [[WHILE_FOUND]]
1133 ; MASKED: while.end:
1134 ; MASKED-NEXT: [[MISMATCH_INDEX2:%.*]] = phi i32 [ [[N]], [[WHILE_COND]] ], [ [[N]], [[BYTE_COMPARE]] ]
1135 ; MASKED-NEXT: [[END_PTR:%.*]] = phi ptr [ [[D]], [[WHILE_COND]] ], [ [[D]], [[BYTE_COMPARE]] ]
1136 ; MASKED-NEXT: br label [[END]]
1137 ; MASKED: end:
1138 ; MASKED-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[MISMATCH_INDEX1]], [[WHILE_FOUND]] ], [ [[MISMATCH_INDEX2]], [[WHILE_END]] ]
1139 ; MASKED-NEXT: [[STORE_PTR:%.*]] = phi ptr [ [[END_PTR]], [[WHILE_END]] ], [ [[FOUND_PTR]], [[WHILE_FOUND]] ]
1140 ; MASKED-NEXT: store i32 [[MISMATCH_INDEX]], ptr [[STORE_PTR]], align 4
1141 ; MASKED-NEXT: ret i32 [[MISMATCH_INDEX]]
1142 ;
1143 ; NO-TRANSFORM-LABEL: define i32 @compare_bytes_simple_end_ne_found(
1144 ; NO-TRANSFORM-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) {
1145 ; NO-TRANSFORM-NEXT: entry:
1146 ; NO-TRANSFORM-NEXT: br label [[WHILE_COND:%.*]]
1147 ; NO-TRANSFORM: while.cond:
1148 ; NO-TRANSFORM-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
1149 ; NO-TRANSFORM-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
1150 ; NO-TRANSFORM-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
1151 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
1152 ; NO-TRANSFORM: while.body:
1153 ; NO-TRANSFORM-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
1154 ; NO-TRANSFORM-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1155 ; NO-TRANSFORM-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1156 ; NO-TRANSFORM-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1157 ; NO-TRANSFORM-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1158 ; NO-TRANSFORM-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
1159 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_FOUND:%.*]]
1160 ; NO-TRANSFORM: while.found:
1161 ; NO-TRANSFORM-NEXT: [[MISMATCH_INDEX1:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ]
1162 ; NO-TRANSFORM-NEXT: [[FOUND_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ]
1163 ; NO-TRANSFORM-NEXT: br label [[END:%.*]]
1164 ; NO-TRANSFORM: while.end:
1165 ; NO-TRANSFORM-NEXT: [[MISMATCH_INDEX2:%.*]] = phi i32 [ [[N]], [[WHILE_COND]] ]
1166 ; NO-TRANSFORM-NEXT: [[END_PTR:%.*]] = phi ptr [ [[D]], [[WHILE_COND]] ]
1167 ; NO-TRANSFORM-NEXT: br label [[END]]
1168 ; NO-TRANSFORM: end:
1169 ; NO-TRANSFORM-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[MISMATCH_INDEX1]], [[WHILE_FOUND]] ], [ [[MISMATCH_INDEX2]], [[WHILE_END]] ]
1170 ; NO-TRANSFORM-NEXT: [[STORE_PTR:%.*]] = phi ptr [ [[END_PTR]], [[WHILE_END]] ], [ [[FOUND_PTR]], [[WHILE_FOUND]] ]
1171 ; NO-TRANSFORM-NEXT: store i32 [[MISMATCH_INDEX]], ptr [[STORE_PTR]], align 4
1172 ; NO-TRANSFORM-NEXT: ret i32 [[MISMATCH_INDEX]]
1173 entry:
1174 br label %while.cond
1176 while.cond:
1177 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
1178 %inc = add i32 %len.addr, 1
1179 %cmp.not = icmp eq i32 %inc, %n
1180 br i1 %cmp.not, label %while.end, label %while.body
1182 while.body:
1183 %idxprom = zext i32 %inc to i64
1184 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
1185 %0 = load i8, ptr %arrayidx
1186 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
1187 %1 = load i8, ptr %arrayidx2
1188 %cmp.not2 = icmp eq i8 %0, %1
1189 br i1 %cmp.not2, label %while.cond, label %while.found
1191 while.found:
1192 %mismatch_index1 = phi i32 [ %inc, %while.body ]
1193 %found_ptr = phi ptr [ %c, %while.body ]
1194 br label %end
1196 while.end:
1197 %mismatch_index2 = phi i32 [ %n, %while.cond ]
1198 %end_ptr = phi ptr [ %d, %while.cond ]
1199 br label %end
1201 end:
1202 %mismatch_index = phi i32 [ %mismatch_index1, %while.found ], [ %mismatch_index2, %while.end ]
1203 %store_ptr = phi ptr [ %end_ptr, %while.end ], [ %found_ptr, %while.found ]
1204 store i32 %mismatch_index, ptr %store_ptr
1205 ret i32 %mismatch_index
1206 }
1210 define i32 @compare_bytes_extra_cmp(ptr %a, ptr %b, i32 %len, i32 %n, i32 %x) {
1211 ; CHECK-LABEL: define i32 @compare_bytes_extra_cmp(
1212 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) #[[ATTR0]] {
1213 ; CHECK-NEXT: entry:
1214 ; CHECK-NEXT: [[CMP_X:%.*]] = icmp ult i32 [[N]], [[X]]
1215 ; CHECK-NEXT: br i1 [[CMP_X]], label [[PH:%.*]], label [[WHILE_END:%.*]]
1216 ; CHECK: ph:
1217 ; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
1218 ; CHECK-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1219 ; CHECK: mismatch_min_it_check:
1220 ; CHECK-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1221 ; CHECK-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
1222 ; CHECK-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
1223 ; CHECK-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1224 ; CHECK: mismatch_mem_check:
1225 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
1226 ; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1227 ; CHECK-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
1228 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
1229 ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
1230 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1231 ; CHECK-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
1232 ; CHECK-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
1233 ; CHECK-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
1234 ; CHECK-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
1235 ; CHECK-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
1236 ; CHECK-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
1237 ; CHECK-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
1238 ; CHECK-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
1239 ; CHECK-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
1240 ; CHECK-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1241 ; CHECK: mismatch_vec_loop_preheader:
1242 ; CHECK-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
1243 ; CHECK: mismatch_vec_loop:
1244 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
1245 ; CHECK-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
1246 ; CHECK-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
1247 ; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
1248 ; CHECK-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1249 ; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
1250 ; CHECK-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1251 ; CHECK-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1252 ; CHECK-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1253 ; CHECK-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
1254 ; CHECK-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1255 ; CHECK: mismatch_vec_loop_inc:
1256 ; CHECK-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
1257 ; CHECK-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
1258 ; CHECK-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
1259 ; CHECK-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
1260 ; CHECK: mismatch_vec_loop_found:
1261 ; CHECK-NEXT: [[FIRST2:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1262 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX3:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1263 ; CHECK-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST2]] to i64
1264 ; CHECK-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX3]], [[TMP26]]
1265 ; CHECK-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
1266 ; CHECK-NEXT: br label [[MISMATCH_END]]
1267 ; CHECK: mismatch_loop_pre:
1268 ; CHECK-NEXT: br label [[MISMATCH_LOOP:%.*]]
1269 ; CHECK: mismatch_loop:
1270 ; CHECK-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1271 ; CHECK-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1272 ; CHECK-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
1273 ; CHECK-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
1274 ; CHECK-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
1275 ; CHECK-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
1276 ; CHECK-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
1277 ; CHECK-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1278 ; CHECK: mismatch_loop_inc:
1279 ; CHECK-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
1280 ; CHECK-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
1281 ; CHECK-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1282 ; CHECK: mismatch_end:
1283 ; CHECK-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1284 ; CHECK-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1285 ; CHECK: while.cond:
1286 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
1287 ; CHECK-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
1288 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1289 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END_LOOPEXIT:%.*]], label [[WHILE_BODY]]
1290 ; CHECK: while.body:
1291 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1292 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1293 ; CHECK-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1294 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1295 ; CHECK-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1296 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
1297 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END_LOOPEXIT]]
1298 ; CHECK: byte.compare:
1299 ; CHECK-NEXT: br label [[WHILE_END_LOOPEXIT]]
1300 ; CHECK: while.end.loopexit:
1301 ; CHECK-NEXT: [[INC_LCSSA1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1302 ; CHECK-NEXT: br label [[WHILE_END]]
1303 ; CHECK: while.end:
1304 ; CHECK-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[X]], [[ENTRY:%.*]] ], [ [[INC_LCSSA1]], [[WHILE_END_LOOPEXIT]] ]
1305 ; CHECK-NEXT: ret i32 [[INC_LCSSA]]
1306 ;
1307 ; LMUL8-LABEL: define i32 @compare_bytes_extra_cmp(
1308 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) #[[ATTR0]] {
1309 ; LMUL8-NEXT: entry:
1310 ; LMUL8-NEXT: [[CMP_X:%.*]] = icmp ult i32 [[N]], [[X]]
1311 ; LMUL8-NEXT: br i1 [[CMP_X]], label [[PH:%.*]], label [[WHILE_END:%.*]]
1312 ; LMUL8: ph:
1313 ; LMUL8-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
1314 ; LMUL8-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1315 ; LMUL8: mismatch_min_it_check:
1316 ; LMUL8-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1317 ; LMUL8-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
1318 ; LMUL8-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
1319 ; LMUL8-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1320 ; LMUL8: mismatch_mem_check:
1321 ; LMUL8-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
1322 ; LMUL8-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1323 ; LMUL8-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
1324 ; LMUL8-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
1325 ; LMUL8-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
1326 ; LMUL8-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1327 ; LMUL8-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
1328 ; LMUL8-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
1329 ; LMUL8-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
1330 ; LMUL8-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
1331 ; LMUL8-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
1332 ; LMUL8-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
1333 ; LMUL8-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
1334 ; LMUL8-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
1335 ; LMUL8-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
1336 ; LMUL8-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1337 ; LMUL8: mismatch_vec_loop_preheader:
1338 ; LMUL8-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
1339 ; LMUL8: mismatch_vec_loop:
1340 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
1341 ; LMUL8-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
1342 ; LMUL8-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 64, i1 true)
1343 ; LMUL8-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
1344 ; LMUL8-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP20]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
1345 ; LMUL8-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
1346 ; LMUL8-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP21]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
1347 ; LMUL8-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 64 x i1> @llvm.vp.icmp.nxv64i8(<vscale x 64 x i8> [[LHS_LOAD]], <vscale x 64 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
1348 ; LMUL8-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv64i1(<vscale x 64 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 64 x i1> splat (i1 true), i32 [[TMP19]])
1349 ; LMUL8-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
1350 ; LMUL8-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1351 ; LMUL8: mismatch_vec_loop_inc:
1352 ; LMUL8-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
1353 ; LMUL8-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
1354 ; LMUL8-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
1355 ; LMUL8-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
1356 ; LMUL8: mismatch_vec_loop_found:
1357 ; LMUL8-NEXT: [[FIRST2:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1358 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX3:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1359 ; LMUL8-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST2]] to i64
1360 ; LMUL8-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX3]], [[TMP26]]
1361 ; LMUL8-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
1362 ; LMUL8-NEXT: br label [[MISMATCH_END]]
1363 ; LMUL8: mismatch_loop_pre:
1364 ; LMUL8-NEXT: br label [[MISMATCH_LOOP:%.*]]
1365 ; LMUL8: mismatch_loop:
1366 ; LMUL8-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1367 ; LMUL8-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1368 ; LMUL8-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
1369 ; LMUL8-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
1370 ; LMUL8-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
1371 ; LMUL8-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
1372 ; LMUL8-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
1373 ; LMUL8-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1374 ; LMUL8: mismatch_loop_inc:
1375 ; LMUL8-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
1376 ; LMUL8-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
1377 ; LMUL8-NEXT: br i1 [[TMP36]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1378 ; LMUL8: mismatch_end:
1379 ; LMUL8-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1380 ; LMUL8-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1381 ; LMUL8: while.cond:
1382 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
1383 ; LMUL8-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
1384 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1385 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END_LOOPEXIT:%.*]], label [[WHILE_BODY]]
1386 ; LMUL8: while.body:
1387 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1388 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1389 ; LMUL8-NEXT: [[TMP37:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1390 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1391 ; LMUL8-NEXT: [[TMP38:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1392 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP37]], [[TMP38]]
1393 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END_LOOPEXIT]]
1394 ; LMUL8: byte.compare:
1395 ; LMUL8-NEXT: br label [[WHILE_END_LOOPEXIT]]
1396 ; LMUL8: while.end.loopexit:
1397 ; LMUL8-NEXT: [[INC_LCSSA1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1398 ; LMUL8-NEXT: br label [[WHILE_END]]
1399 ; LMUL8: while.end:
1400 ; LMUL8-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[X]], [[ENTRY:%.*]] ], [ [[INC_LCSSA1]], [[WHILE_END_LOOPEXIT]] ]
1401 ; LMUL8-NEXT: ret i32 [[INC_LCSSA]]
1402 ;
1403 ; LOOP-DEL-LABEL: define i32 @compare_bytes_extra_cmp(
1404 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) #[[ATTR0]] {
1405 ; LOOP-DEL-NEXT: entry:
1406 ; LOOP-DEL-NEXT: [[CMP_X:%.*]] = icmp ult i32 [[N]], [[X]]
1407 ; LOOP-DEL-NEXT: br i1 [[CMP_X]], label [[PH:%.*]], label [[WHILE_END:%.*]]
1408 ; LOOP-DEL: ph:
1409 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
1410 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1411 ; LOOP-DEL-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
1412 ; LOOP-DEL-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
1413 ; LOOP-DEL-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1414 ; LOOP-DEL: mismatch_mem_check:
1415 ; LOOP-DEL-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
1416 ; LOOP-DEL-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1417 ; LOOP-DEL-NEXT: [[TMP8:%.*]] = ptrtoint ptr [[TMP5]] to i64
1418 ; LOOP-DEL-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
1419 ; LOOP-DEL-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
1420 ; LOOP-DEL-NEXT: [[TMP13:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1421 ; LOOP-DEL-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP10]] to i64
1422 ; LOOP-DEL-NEXT: [[TMP14:%.*]] = ptrtoint ptr [[TMP13]] to i64
1423 ; LOOP-DEL-NEXT: [[TMP6:%.*]] = lshr i64 [[TMP7]], 12
1424 ; LOOP-DEL-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP11]], 12
1425 ; LOOP-DEL-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP8]], 12
1426 ; LOOP-DEL-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP14]], 12
1427 ; LOOP-DEL-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP6]], [[TMP9]]
1428 ; LOOP-DEL-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP12]], [[TMP15]]
1429 ; LOOP-DEL-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
1430 ; LOOP-DEL-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP:%.*]], !prof [[PROF1]]
1431 ; LOOP-DEL: mismatch_vec_loop:
1432 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ [[TMP24:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ], [ [[TMP1]], [[MISMATCH_MEM_CHECK]] ]
1433 ; LOOP-DEL-NEXT: [[AVL:%.*]] = sub nuw nsw i64 [[TMP2]], [[MISMATCH_VECTOR_INDEX]]
1434 ; LOOP-DEL-NEXT: [[TMP19:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
1435 ; LOOP-DEL-NEXT: [[TMP20:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VECTOR_INDEX]]
1436 ; LOOP-DEL-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP20]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1437 ; LOOP-DEL-NEXT: [[TMP21:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VECTOR_INDEX]]
1438 ; LOOP-DEL-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP21]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1439 ; LOOP-DEL-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1440 ; LOOP-DEL-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP19]])
1441 ; LOOP-DEL-NEXT: [[TMP22:%.*]] = icmp ne i32 [[FIRST]], [[TMP19]]
1442 ; LOOP-DEL-NEXT: br i1 [[TMP22]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1443 ; LOOP-DEL: mismatch_vec_loop_inc:
1444 ; LOOP-DEL-NEXT: [[TMP23:%.*]] = zext i32 [[TMP19]] to i64
1445 ; LOOP-DEL-NEXT: [[TMP24]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP23]]
1446 ; LOOP-DEL-NEXT: [[TMP25:%.*]] = icmp ne i64 [[TMP24]], [[TMP2]]
1447 ; LOOP-DEL-NEXT: br i1 [[TMP25]], label [[MISMATCH_VECTOR_LOOP]], label [[WHILE_END]]
1448 ; LOOP-DEL: mismatch_vec_loop_found:
1449 ; LOOP-DEL-NEXT: [[FIRST2:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1450 ; LOOP-DEL-NEXT: [[MISMATCH_VECTOR_INDEX3:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1451 ; LOOP-DEL-NEXT: [[TMP26:%.*]] = zext i32 [[FIRST2]] to i64
1452 ; LOOP-DEL-NEXT: [[TMP27:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX3]], [[TMP26]]
1453 ; LOOP-DEL-NEXT: [[TMP28:%.*]] = trunc i64 [[TMP27]] to i32
1454 ; LOOP-DEL-NEXT: br label [[WHILE_END]]
1455 ; LOOP-DEL: mismatch_loop_pre:
1456 ; LOOP-DEL-NEXT: br label [[MISMATCH_LOOP:%.*]]
1457 ; LOOP-DEL: mismatch_loop:
1458 ; LOOP-DEL-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP35:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1459 ; LOOP-DEL-NEXT: [[TMP29:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1460 ; LOOP-DEL-NEXT: [[TMP30:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP29]]
1461 ; LOOP-DEL-NEXT: [[TMP31:%.*]] = load i8, ptr [[TMP30]], align 1
1462 ; LOOP-DEL-NEXT: [[TMP32:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP29]]
1463 ; LOOP-DEL-NEXT: [[TMP33:%.*]] = load i8, ptr [[TMP32]], align 1
1464 ; LOOP-DEL-NEXT: [[TMP34:%.*]] = icmp eq i8 [[TMP31]], [[TMP33]]
1465 ; LOOP-DEL-NEXT: br i1 [[TMP34]], label [[MISMATCH_LOOP_INC]], label [[WHILE_END]]
1466 ; LOOP-DEL: mismatch_loop_inc:
1467 ; LOOP-DEL-NEXT: [[TMP35]] = add i32 [[MISMATCH_INDEX]], 1
1468 ; LOOP-DEL-NEXT: [[TMP36:%.*]] = icmp eq i32 [[TMP35]], [[N]]
1469 ; LOOP-DEL-NEXT: br i1 [[TMP36]], label [[WHILE_END]], label [[MISMATCH_LOOP]]
1470 ; LOOP-DEL: while.end:
1471 ; LOOP-DEL-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[X]], [[ENTRY:%.*]] ], [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP28]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1472 ; LOOP-DEL-NEXT: ret i32 [[INC_LCSSA]]
1473 ;
1474 ; MASKED-LABEL: define i32 @compare_bytes_extra_cmp(
1475 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) #[[ATTR0]] {
1476 ; MASKED-NEXT: entry:
1477 ; MASKED-NEXT: [[CMP_X:%.*]] = icmp ult i32 [[N]], [[X]]
1478 ; MASKED-NEXT: br i1 [[CMP_X]], label [[PH:%.*]], label [[WHILE_END:%.*]]
1479 ; MASKED: ph:
1480 ; MASKED-NEXT: [[TMP0:%.*]] = add i32 [[LEN]], 1
1481 ; MASKED-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1482 ; MASKED: mismatch_min_it_check:
1483 ; MASKED-NEXT: [[TMP1:%.*]] = zext i32 [[TMP0]] to i64
1484 ; MASKED-NEXT: [[TMP2:%.*]] = zext i32 [[N]] to i64
1485 ; MASKED-NEXT: [[TMP3:%.*]] = icmp ule i32 [[TMP0]], [[N]]
1486 ; MASKED-NEXT: br i1 [[TMP3]], label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1487 ; MASKED: mismatch_mem_check:
1488 ; MASKED-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP1]]
1489 ; MASKED-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP1]]
1490 ; MASKED-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP5]] to i64
1491 ; MASKED-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP4]] to i64
1492 ; MASKED-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP2]]
1493 ; MASKED-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[B]], i64 [[TMP2]]
1494 ; MASKED-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP8]] to i64
1495 ; MASKED-NEXT: [[TMP11:%.*]] = ptrtoint ptr [[TMP9]] to i64
1496 ; MASKED-NEXT: [[TMP12:%.*]] = lshr i64 [[TMP7]], 12
1497 ; MASKED-NEXT: [[TMP13:%.*]] = lshr i64 [[TMP10]], 12
1498 ; MASKED-NEXT: [[TMP14:%.*]] = lshr i64 [[TMP6]], 12
1499 ; MASKED-NEXT: [[TMP15:%.*]] = lshr i64 [[TMP11]], 12
1500 ; MASKED-NEXT: [[TMP16:%.*]] = icmp ne i64 [[TMP12]], [[TMP13]]
1501 ; MASKED-NEXT: [[TMP17:%.*]] = icmp ne i64 [[TMP14]], [[TMP15]]
1502 ; MASKED-NEXT: [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
1503 ; MASKED-NEXT: br i1 [[TMP18]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VEC_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1504 ; MASKED: mismatch_vec_loop_preheader:
1505 ; MASKED-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP1]], i64 [[TMP2]])
1506 ; MASKED-NEXT: [[TMP20:%.*]] = call i64 @llvm.vscale.i64()
1507 ; MASKED-NEXT: [[TMP21:%.*]] = mul nuw nsw i64 [[TMP20]], 16
1508 ; MASKED-NEXT: br label [[MISMATCH_VEC_LOOP:%.*]]
1509 ; MASKED: mismatch_vec_loop:
1510 ; MASKED-NEXT: [[MISMATCH_VEC_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP19]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP30:%.*]], [[MISMATCH_VEC_LOOP_INC:%.*]] ]
1511 ; MASKED-NEXT: [[MISMATCH_VEC_INDEX:%.*]] = phi i64 [ [[TMP1]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP29:%.*]], [[MISMATCH_VEC_LOOP_INC]] ]
1512 ; MASKED-NEXT: [[TMP22:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[MISMATCH_VEC_INDEX]]
1513 ; MASKED-NEXT: [[TMP23:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP22]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1514 ; MASKED-NEXT: [[TMP24:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[MISMATCH_VEC_INDEX]]
1515 ; MASKED-NEXT: [[TMP25:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP24]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1516 ; MASKED-NEXT: [[TMP26:%.*]] = icmp ne <vscale x 16 x i8> [[TMP23]], [[TMP25]]
1517 ; MASKED-NEXT: [[TMP27:%.*]] = select <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i1> [[TMP26]], <vscale x 16 x i1> zeroinitializer
1518 ; MASKED-NEXT: [[TMP28:%.*]] = call i1 @llvm.vector.reduce.or.nxv16i1(<vscale x 16 x i1> [[TMP27]])
1519 ; MASKED-NEXT: br i1 [[TMP28]], label [[MISMATCH_VEC_LOOP_FOUND:%.*]], label [[MISMATCH_VEC_LOOP_INC]]
1520 ; MASKED: mismatch_vec_loop_inc:
1521 ; MASKED-NEXT: [[TMP29]] = add nuw nsw i64 [[MISMATCH_VEC_INDEX]], [[TMP21]]
1522 ; MASKED-NEXT: [[TMP30]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP29]], i64 [[TMP2]])
1523 ; MASKED-NEXT: [[TMP31:%.*]] = extractelement <vscale x 16 x i1> [[TMP30]], i64 0
1524 ; MASKED-NEXT: br i1 [[TMP31]], label [[MISMATCH_VEC_LOOP]], label [[MISMATCH_END:%.*]]
1525 ; MASKED: mismatch_vec_loop_found:
1526 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP27]], [[MISMATCH_VEC_LOOP]] ]
1527 ; MASKED-NEXT: [[MISMATCH_VEC_LAST_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[MISMATCH_VEC_LOOP_PRED]], [[MISMATCH_VEC_LOOP]] ]
1528 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_INDEX:%.*]] = phi i64 [ [[MISMATCH_VEC_INDEX]], [[MISMATCH_VEC_LOOP]] ]
1529 ; MASKED-NEXT: [[TMP32:%.*]] = and <vscale x 16 x i1> [[MISMATCH_VEC_LAST_LOOP_PRED]], [[MISMATCH_VEC_FOUND_PRED]]
1530 ; MASKED-NEXT: [[TMP33:%.*]] = call i32 @llvm.experimental.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[TMP32]], i1 true)
1531 ; MASKED-NEXT: [[TMP34:%.*]] = zext i32 [[TMP33]] to i64
1532 ; MASKED-NEXT: [[TMP35:%.*]] = add nuw nsw i64 [[MISMATCH_VEC_FOUND_INDEX]], [[TMP34]]
1533 ; MASKED-NEXT: [[TMP36:%.*]] = trunc i64 [[TMP35]] to i32
1534 ; MASKED-NEXT: br label [[MISMATCH_END]]
1535 ; MASKED: mismatch_loop_pre:
1536 ; MASKED-NEXT: br label [[MISMATCH_LOOP:%.*]]
1537 ; MASKED: mismatch_loop:
1538 ; MASKED-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ [[TMP0]], [[MISMATCH_LOOP_PRE]] ], [ [[TMP43:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1539 ; MASKED-NEXT: [[TMP37:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1540 ; MASKED-NEXT: [[TMP38:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[TMP37]]
1541 ; MASKED-NEXT: [[TMP39:%.*]] = load i8, ptr [[TMP38]], align 1
1542 ; MASKED-NEXT: [[TMP40:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[TMP37]]
1543 ; MASKED-NEXT: [[TMP41:%.*]] = load i8, ptr [[TMP40]], align 1
1544 ; MASKED-NEXT: [[TMP42:%.*]] = icmp eq i8 [[TMP39]], [[TMP41]]
1545 ; MASKED-NEXT: br i1 [[TMP42]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1546 ; MASKED: mismatch_loop_inc:
1547 ; MASKED-NEXT: [[TMP43]] = add i32 [[MISMATCH_INDEX]], 1
1548 ; MASKED-NEXT: [[TMP44:%.*]] = icmp eq i32 [[TMP43]], [[N]]
1549 ; MASKED-NEXT: br i1 [[TMP44]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1550 ; MASKED: mismatch_end:
1551 ; MASKED-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ [[N]], [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ [[N]], [[MISMATCH_VEC_LOOP_INC]] ], [ [[TMP36]], [[MISMATCH_VEC_LOOP_FOUND]] ]
1552 ; MASKED-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1553 ; MASKED: while.cond:
1554 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[MISMATCH_END]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ]
1555 ; MASKED-NEXT: [[INC:%.*]] = add i32 [[LEN_ADDR]], 1
1556 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], [[N]]
1557 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END_LOOPEXIT:%.*]], label [[WHILE_BODY]]
1558 ; MASKED: while.body:
1559 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1560 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1561 ; MASKED-NEXT: [[TMP45:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1562 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1563 ; MASKED-NEXT: [[TMP46:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1564 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP45]], [[TMP46]]
1565 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END_LOOPEXIT]]
1566 ; MASKED: byte.compare:
1567 ; MASKED-NEXT: br label [[WHILE_END_LOOPEXIT]]
1568 ; MASKED: while.end.loopexit:
1569 ; MASKED-NEXT: [[INC_LCSSA1:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_COND]] ], [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1570 ; MASKED-NEXT: br label [[WHILE_END]]
1571 ; MASKED: while.end:
1572 ; MASKED-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[X]], [[ENTRY:%.*]] ], [ [[INC_LCSSA1]], [[WHILE_END_LOOPEXIT]] ]
1573 ; MASKED-NEXT: ret i32 [[INC_LCSSA]]
1574 ;
1575 ; NO-TRANSFORM-LABEL: define i32 @compare_bytes_extra_cmp(
1576 ; NO-TRANSFORM-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) {
1577 ; NO-TRANSFORM-NEXT: entry:
1578 ; NO-TRANSFORM-NEXT: [[CMP_X:%.*]] = icmp ult i32 [[N]], [[X]]
1579 ; NO-TRANSFORM-NEXT: br i1 [[CMP_X]], label [[PH:%.*]], label [[WHILE_END:%.*]]
1580 ; NO-TRANSFORM: ph:
1581 ; NO-TRANSFORM-NEXT: br label [[WHILE_COND:%.*]]
1582 ; NO-TRANSFORM: while.cond:
1583 ; NO-TRANSFORM-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[PH]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
1584 ; NO-TRANSFORM-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
1585 ; NO-TRANSFORM-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
1586 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END]], label [[WHILE_BODY]]
1587 ; NO-TRANSFORM: while.body:
1588 ; NO-TRANSFORM-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
1589 ; NO-TRANSFORM-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1590 ; NO-TRANSFORM-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1591 ; NO-TRANSFORM-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1592 ; NO-TRANSFORM-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1593 ; NO-TRANSFORM-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
1594 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
1595 ; NO-TRANSFORM: while.end:
1596 ; NO-TRANSFORM-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ], [ [[X]], [[ENTRY:%.*]] ]
1597 ; NO-TRANSFORM-NEXT: ret i32 [[INC_LCSSA]]
1598 entry:
1599 %cmp.x = icmp ult i32 %n, %x
1600 br i1 %cmp.x, label %ph, label %while.end
1602 ph:
1603 br label %while.cond
1605 while.cond:
1606 %len.addr = phi i32 [ %len, %ph ], [ %inc, %while.body ]
1607 %inc = add i32 %len.addr, 1
1608 %cmp.not = icmp eq i32 %inc, %n
1609 br i1 %cmp.not, label %while.end, label %while.body
1611 while.body:
1612 %idxprom = zext i32 %inc to i64
1613 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
1614 %0 = load i8, ptr %arrayidx
1615 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
1616 %1 = load i8, ptr %arrayidx2
1617 %cmp.not2 = icmp eq i8 %0, %1
1618 br i1 %cmp.not2, label %while.cond, label %while.end
1620 while.end:
1621 %inc.lcssa = phi i32 [ %inc, %while.body ], [ %inc, %while.cond ], [ %x, %entry ]
1622 ret i32 %inc.lcssa
1623 }
1625 define void @compare_bytes_cleanup_block(ptr %src1, ptr %src2) {
1626 ; CHECK-LABEL: define void @compare_bytes_cleanup_block(
1627 ; CHECK-SAME: ptr [[SRC1:%.*]], ptr [[SRC2:%.*]]) #[[ATTR0]] {
1628 ; CHECK-NEXT: entry:
1629 ; CHECK-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1630 ; CHECK: mismatch_min_it_check:
1631 ; CHECK-NEXT: br i1 false, label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1632 ; CHECK: mismatch_mem_check:
1633 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i8, ptr [[SRC1]], i64 1
1634 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[SRC2]], i64 1
1635 ; CHECK-NEXT: [[TMP9:%.*]] = ptrtoint ptr [[TMP1]] to i64
1636 ; CHECK-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP0]] to i64
1637 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[SRC1]], i64 0
1638 ; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[SRC2]], i64 0
1639 ; CHECK-NEXT: [[TMP4:%.*]] = ptrtoint ptr [[TMP3]] to i64
1640 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP6]] to i64
1641 ; CHECK-NEXT: [[TMP2:%.*]] = lshr i64 [[TMP10]], 12
1642 ; CHECK-NEXT: [[TMP5:%.*]] = lshr i64 [[TMP4]], 12
1643 ; CHECK-NEXT: [[TMP8:%.*]] = lshr i64 [[TMP9]], 12
1644 ; CHECK-NEXT: [[TMP11:%.*]] = lshr i64 [[TMP7]], 12
1645 ; CHECK-NEXT: [[TMP12:%.*]] = icmp ne i64 [[TMP2]], [[TMP5]]
1646 ; CHECK-NEXT: [[TMP13:%.*]] = icmp ne i64 [[TMP8]], [[TMP11]]
1647 ; CHECK-NEXT: [[TMP14:%.*]] = or i1 [[TMP12]], [[TMP13]]
1648 ; CHECK-NEXT: br i1 [[TMP14]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1649 ; CHECK: mismatch_vec_loop_preheader:
1650 ; CHECK-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
1651 ; CHECK: mismatch_vec_loop:
1652 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ 1, [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP20:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
1653 ; CHECK-NEXT: [[AVL:%.*]] = sub nuw nsw i64 0, [[MISMATCH_VECTOR_INDEX]]
1654 ; CHECK-NEXT: [[TMP15:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 16, i1 true)
1655 ; CHECK-NEXT: [[TMP16:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[MISMATCH_VECTOR_INDEX]]
1656 ; CHECK-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP16]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP15]])
1657 ; CHECK-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[MISMATCH_VECTOR_INDEX]]
1658 ; CHECK-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 16 x i8> @llvm.vp.load.nxv16i8.p0(ptr [[TMP17]], <vscale x 16 x i1> splat (i1 true), i32 [[TMP15]])
1659 ; CHECK-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 16 x i1> @llvm.vp.icmp.nxv16i8(<vscale x 16 x i8> [[LHS_LOAD]], <vscale x 16 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 16 x i1> splat (i1 true), i32 [[TMP15]])
1660 ; CHECK-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 16 x i1> splat (i1 true), i32 [[TMP15]])
1661 ; CHECK-NEXT: [[TMP18:%.*]] = icmp ne i32 [[FIRST]], [[TMP15]]
1662 ; CHECK-NEXT: br i1 [[TMP18]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1663 ; CHECK: mismatch_vec_loop_inc:
1664 ; CHECK-NEXT: [[TMP19:%.*]] = zext i32 [[TMP15]] to i64
1665 ; CHECK-NEXT: [[TMP20]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP19]]
1666 ; CHECK-NEXT: [[TMP21:%.*]] = icmp ne i64 [[TMP20]], 0
1667 ; CHECK-NEXT: br i1 [[TMP21]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
1668 ; CHECK: mismatch_vec_loop_found:
1669 ; CHECK-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1670 ; CHECK-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1671 ; CHECK-NEXT: [[TMP22:%.*]] = zext i32 [[FIRST1]] to i64
1672 ; CHECK-NEXT: [[TMP23:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP22]]
1673 ; CHECK-NEXT: [[TMP24:%.*]] = trunc i64 [[TMP23]] to i32
1674 ; CHECK-NEXT: br label [[MISMATCH_END]]
1675 ; CHECK: mismatch_loop_pre:
1676 ; CHECK-NEXT: br label [[MISMATCH_LOOP:%.*]]
1677 ; CHECK: mismatch_loop:
1678 ; CHECK-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ 1, [[MISMATCH_LOOP_PRE]] ], [ [[TMP31:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1679 ; CHECK-NEXT: [[TMP25:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1680 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[TMP25]]
1681 ; CHECK-NEXT: [[TMP27:%.*]] = load i8, ptr [[TMP26]], align 1
1682 ; CHECK-NEXT: [[TMP28:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[TMP25]]
1683 ; CHECK-NEXT: [[TMP29:%.*]] = load i8, ptr [[TMP28]], align 1
1684 ; CHECK-NEXT: [[TMP30:%.*]] = icmp eq i8 [[TMP27]], [[TMP29]]
1685 ; CHECK-NEXT: br i1 [[TMP30]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1686 ; CHECK: mismatch_loop_inc:
1687 ; CHECK-NEXT: [[TMP31]] = add i32 [[MISMATCH_INDEX]], 1
1688 ; CHECK-NEXT: [[TMP32:%.*]] = icmp eq i32 [[TMP31]], 0
1689 ; CHECK-NEXT: br i1 [[TMP32]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1690 ; CHECK: mismatch_end:
1691 ; CHECK-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ 0, [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ 0, [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP24]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1692 ; CHECK-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1693 ; CHECK: while.cond:
1694 ; CHECK-NEXT: [[LEN:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ], [ 0, [[MISMATCH_END]] ]
1695 ; CHECK-NEXT: [[INC:%.*]] = add i32 [[LEN]], 1
1696 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1697 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[CLEANUP_THREAD:%.*]], label [[WHILE_BODY]]
1698 ; CHECK: while.body:
1699 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1700 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[IDXPROM]]
1701 ; CHECK-NEXT: [[TMP33:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1702 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[IDXPROM]]
1703 ; CHECK-NEXT: [[TMP34:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1704 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP33]], [[TMP34]]
1705 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[IF_END:%.*]]
1706 ; CHECK: byte.compare:
1707 ; CHECK-NEXT: [[TMP35:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1708 ; CHECK-NEXT: br i1 [[TMP35]], label [[CLEANUP_THREAD]], label [[IF_END]]
1709 ; CHECK: cleanup.thread:
1710 ; CHECK-NEXT: ret void
1711 ; CHECK: if.end:
1712 ; CHECK-NEXT: [[RES:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1713 ; CHECK-NEXT: ret void
1714 ;
1715 ; LMUL8-LABEL: define void @compare_bytes_cleanup_block(
1716 ; LMUL8-SAME: ptr [[SRC1:%.*]], ptr [[SRC2:%.*]]) #[[ATTR0]] {
1717 ; LMUL8-NEXT: entry:
1718 ; LMUL8-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1719 ; LMUL8: mismatch_min_it_check:
1720 ; LMUL8-NEXT: br i1 false, label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1721 ; LMUL8: mismatch_mem_check:
1722 ; LMUL8-NEXT: [[TMP0:%.*]] = getelementptr i8, ptr [[SRC1]], i64 1
1723 ; LMUL8-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[SRC2]], i64 1
1724 ; LMUL8-NEXT: [[TMP9:%.*]] = ptrtoint ptr [[TMP1]] to i64
1725 ; LMUL8-NEXT: [[TMP10:%.*]] = ptrtoint ptr [[TMP0]] to i64
1726 ; LMUL8-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[SRC1]], i64 0
1727 ; LMUL8-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[SRC2]], i64 0
1728 ; LMUL8-NEXT: [[TMP4:%.*]] = ptrtoint ptr [[TMP3]] to i64
1729 ; LMUL8-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP6]] to i64
1730 ; LMUL8-NEXT: [[TMP2:%.*]] = lshr i64 [[TMP10]], 12
1731 ; LMUL8-NEXT: [[TMP5:%.*]] = lshr i64 [[TMP4]], 12
1732 ; LMUL8-NEXT: [[TMP8:%.*]] = lshr i64 [[TMP9]], 12
1733 ; LMUL8-NEXT: [[TMP11:%.*]] = lshr i64 [[TMP7]], 12
1734 ; LMUL8-NEXT: [[TMP12:%.*]] = icmp ne i64 [[TMP2]], [[TMP5]]
1735 ; LMUL8-NEXT: [[TMP13:%.*]] = icmp ne i64 [[TMP8]], [[TMP11]]
1736 ; LMUL8-NEXT: [[TMP14:%.*]] = or i1 [[TMP12]], [[TMP13]]
1737 ; LMUL8-NEXT: br i1 [[TMP14]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VECTOR_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1738 ; LMUL8: mismatch_vec_loop_preheader:
1739 ; LMUL8-NEXT: br label [[MISMATCH_VECTOR_LOOP:%.*]]
1740 ; LMUL8: mismatch_vec_loop:
1741 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX:%.*]] = phi i64 [ 1, [[MISMATCH_VECTOR_LOOP_PREHEADER]] ], [ [[TMP20:%.*]], [[MISMATCH_VECTOR_LOOP_INC:%.*]] ]
1742 ; LMUL8-NEXT: [[AVL:%.*]] = sub nuw nsw i64 0, [[MISMATCH_VECTOR_INDEX]]
1743 ; LMUL8-NEXT: [[TMP15:%.*]] = call i32 @llvm.experimental.get.vector.length.i64(i64 [[AVL]], i32 64, i1 true)
1744 ; LMUL8-NEXT: [[TMP16:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[MISMATCH_VECTOR_INDEX]]
1745 ; LMUL8-NEXT: [[LHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP16]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP15]])
1746 ; LMUL8-NEXT: [[TMP17:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[MISMATCH_VECTOR_INDEX]]
1747 ; LMUL8-NEXT: [[RHS_LOAD:%.*]] = call <vscale x 64 x i8> @llvm.vp.load.nxv64i8.p0(ptr [[TMP17]], <vscale x 64 x i1> splat (i1 true), i32 [[TMP15]])
1748 ; LMUL8-NEXT: [[MISMATCH_CMP:%.*]] = call <vscale x 64 x i1> @llvm.vp.icmp.nxv64i8(<vscale x 64 x i8> [[LHS_LOAD]], <vscale x 64 x i8> [[RHS_LOAD]], metadata !"ne", <vscale x 64 x i1> splat (i1 true), i32 [[TMP15]])
1749 ; LMUL8-NEXT: [[FIRST:%.*]] = call i32 @llvm.vp.cttz.elts.i32.nxv64i1(<vscale x 64 x i1> [[MISMATCH_CMP]], i1 false, <vscale x 64 x i1> splat (i1 true), i32 [[TMP15]])
1750 ; LMUL8-NEXT: [[TMP18:%.*]] = icmp ne i32 [[FIRST]], [[TMP15]]
1751 ; LMUL8-NEXT: br i1 [[TMP18]], label [[MISMATCH_VECTOR_LOOP_FOUND:%.*]], label [[MISMATCH_VECTOR_LOOP_INC]]
1752 ; LMUL8: mismatch_vec_loop_inc:
1753 ; LMUL8-NEXT: [[TMP19:%.*]] = zext i32 [[TMP15]] to i64
1754 ; LMUL8-NEXT: [[TMP20]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX]], [[TMP19]]
1755 ; LMUL8-NEXT: [[TMP21:%.*]] = icmp ne i64 [[TMP20]], 0
1756 ; LMUL8-NEXT: br i1 [[TMP21]], label [[MISMATCH_VECTOR_LOOP]], label [[MISMATCH_END:%.*]]
1757 ; LMUL8: mismatch_vec_loop_found:
1758 ; LMUL8-NEXT: [[FIRST1:%.*]] = phi i32 [ [[FIRST]], [[MISMATCH_VECTOR_LOOP]] ]
1759 ; LMUL8-NEXT: [[MISMATCH_VECTOR_INDEX2:%.*]] = phi i64 [ [[MISMATCH_VECTOR_INDEX]], [[MISMATCH_VECTOR_LOOP]] ]
1760 ; LMUL8-NEXT: [[TMP22:%.*]] = zext i32 [[FIRST1]] to i64
1761 ; LMUL8-NEXT: [[TMP23:%.*]] = add nuw nsw i64 [[MISMATCH_VECTOR_INDEX2]], [[TMP22]]
1762 ; LMUL8-NEXT: [[TMP24:%.*]] = trunc i64 [[TMP23]] to i32
1763 ; LMUL8-NEXT: br label [[MISMATCH_END]]
1764 ; LMUL8: mismatch_loop_pre:
1765 ; LMUL8-NEXT: br label [[MISMATCH_LOOP:%.*]]
1766 ; LMUL8: mismatch_loop:
1767 ; LMUL8-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ 1, [[MISMATCH_LOOP_PRE]] ], [ [[TMP31:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1768 ; LMUL8-NEXT: [[TMP25:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1769 ; LMUL8-NEXT: [[TMP26:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[TMP25]]
1770 ; LMUL8-NEXT: [[TMP27:%.*]] = load i8, ptr [[TMP26]], align 1
1771 ; LMUL8-NEXT: [[TMP28:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[TMP25]]
1772 ; LMUL8-NEXT: [[TMP29:%.*]] = load i8, ptr [[TMP28]], align 1
1773 ; LMUL8-NEXT: [[TMP30:%.*]] = icmp eq i8 [[TMP27]], [[TMP29]]
1774 ; LMUL8-NEXT: br i1 [[TMP30]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1775 ; LMUL8: mismatch_loop_inc:
1776 ; LMUL8-NEXT: [[TMP31]] = add i32 [[MISMATCH_INDEX]], 1
1777 ; LMUL8-NEXT: [[TMP32:%.*]] = icmp eq i32 [[TMP31]], 0
1778 ; LMUL8-NEXT: br i1 [[TMP32]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1779 ; LMUL8: mismatch_end:
1780 ; LMUL8-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ 0, [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ 0, [[MISMATCH_VECTOR_LOOP_INC]] ], [ [[TMP24]], [[MISMATCH_VECTOR_LOOP_FOUND]] ]
1781 ; LMUL8-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1782 ; LMUL8: while.cond:
1783 ; LMUL8-NEXT: [[LEN:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ], [ 0, [[MISMATCH_END]] ]
1784 ; LMUL8-NEXT: [[INC:%.*]] = add i32 [[LEN]], 1
1785 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1786 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[CLEANUP_THREAD:%.*]], label [[WHILE_BODY]]
1787 ; LMUL8: while.body:
1788 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1789 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[IDXPROM]]
1790 ; LMUL8-NEXT: [[TMP33:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1791 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[IDXPROM]]
1792 ; LMUL8-NEXT: [[TMP34:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1793 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP33]], [[TMP34]]
1794 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[IF_END:%.*]]
1795 ; LMUL8: byte.compare:
1796 ; LMUL8-NEXT: [[TMP35:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1797 ; LMUL8-NEXT: br i1 [[TMP35]], label [[CLEANUP_THREAD]], label [[IF_END]]
1798 ; LMUL8: cleanup.thread:
1799 ; LMUL8-NEXT: ret void
1800 ; LMUL8: if.end:
1801 ; LMUL8-NEXT: [[RES:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1802 ; LMUL8-NEXT: ret void
1803 ;
1804 ; LOOP-DEL-LABEL: define void @compare_bytes_cleanup_block(
1805 ; LOOP-DEL-SAME: ptr [[SRC1:%.*]], ptr [[SRC2:%.*]]) #[[ATTR0]] {
1806 ; LOOP-DEL-NEXT: entry:
1807 ; LOOP-DEL-NEXT: br label [[MISMATCH_LOOP:%.*]]
1808 ; LOOP-DEL: mismatch_loop:
1809 ; LOOP-DEL-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[TMP6:%.*]], [[MISMATCH_LOOP]] ]
1810 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1811 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[TMP0]]
1812 ; LOOP-DEL-NEXT: [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
1813 ; LOOP-DEL-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[TMP0]]
1814 ; LOOP-DEL-NEXT: [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 1
1815 ; LOOP-DEL-NEXT: [[TMP5:%.*]] = icmp ne i8 [[TMP2]], [[TMP4]]
1816 ; LOOP-DEL-NEXT: [[TMP6]] = add i32 [[MISMATCH_INDEX]], 1
1817 ; LOOP-DEL-NEXT: [[TMP7:%.*]] = icmp eq i32 [[TMP6]], 0
1818 ; LOOP-DEL-NEXT: [[OR_COND:%.*]] = or i1 [[TMP5]], [[TMP7]]
1819 ; LOOP-DEL-NEXT: br i1 [[OR_COND]], label [[COMMON_RET:%.*]], label [[MISMATCH_LOOP]]
1820 ; LOOP-DEL: common.ret:
1821 ; LOOP-DEL-NEXT: ret void
1822 ;
1823 ; MASKED-LABEL: define void @compare_bytes_cleanup_block(
1824 ; MASKED-SAME: ptr [[SRC1:%.*]], ptr [[SRC2:%.*]]) #[[ATTR0]] {
1825 ; MASKED-NEXT: entry:
1826 ; MASKED-NEXT: br label [[MISMATCH_MIN_IT_CHECK:%.*]]
1827 ; MASKED: mismatch_min_it_check:
1828 ; MASKED-NEXT: br i1 false, label [[MISMATCH_MEM_CHECK:%.*]], label [[MISMATCH_LOOP_PRE:%.*]], !prof [[PROF0]]
1829 ; MASKED: mismatch_mem_check:
1830 ; MASKED-NEXT: [[TMP0:%.*]] = getelementptr i8, ptr [[SRC1]], i64 1
1831 ; MASKED-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[SRC2]], i64 1
1832 ; MASKED-NEXT: [[TMP2:%.*]] = ptrtoint ptr [[TMP1]] to i64
1833 ; MASKED-NEXT: [[TMP3:%.*]] = ptrtoint ptr [[TMP0]] to i64
1834 ; MASKED-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[SRC1]], i64 0
1835 ; MASKED-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[SRC2]], i64 0
1836 ; MASKED-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP4]] to i64
1837 ; MASKED-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP5]] to i64
1838 ; MASKED-NEXT: [[TMP8:%.*]] = lshr i64 [[TMP3]], 12
1839 ; MASKED-NEXT: [[TMP9:%.*]] = lshr i64 [[TMP6]], 12
1840 ; MASKED-NEXT: [[TMP10:%.*]] = lshr i64 [[TMP2]], 12
1841 ; MASKED-NEXT: [[TMP11:%.*]] = lshr i64 [[TMP7]], 12
1842 ; MASKED-NEXT: [[TMP12:%.*]] = icmp ne i64 [[TMP8]], [[TMP9]]
1843 ; MASKED-NEXT: [[TMP13:%.*]] = icmp ne i64 [[TMP10]], [[TMP11]]
1844 ; MASKED-NEXT: [[TMP14:%.*]] = or i1 [[TMP12]], [[TMP13]]
1845 ; MASKED-NEXT: br i1 [[TMP14]], label [[MISMATCH_LOOP_PRE]], label [[MISMATCH_VEC_LOOP_PREHEADER:%.*]], !prof [[PROF1]]
1846 ; MASKED: mismatch_vec_loop_preheader:
1847 ; MASKED-NEXT: [[TMP15:%.*]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 1, i64 0)
1848 ; MASKED-NEXT: [[TMP16:%.*]] = call i64 @llvm.vscale.i64()
1849 ; MASKED-NEXT: [[TMP17:%.*]] = mul nuw nsw i64 [[TMP16]], 16
1850 ; MASKED-NEXT: br label [[MISMATCH_VEC_LOOP:%.*]]
1851 ; MASKED: mismatch_vec_loop:
1852 ; MASKED-NEXT: [[MISMATCH_VEC_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP15]], [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP26:%.*]], [[MISMATCH_VEC_LOOP_INC:%.*]] ]
1853 ; MASKED-NEXT: [[MISMATCH_VEC_INDEX:%.*]] = phi i64 [ 1, [[MISMATCH_VEC_LOOP_PREHEADER]] ], [ [[TMP25:%.*]], [[MISMATCH_VEC_LOOP_INC]] ]
1854 ; MASKED-NEXT: [[TMP18:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[MISMATCH_VEC_INDEX]]
1855 ; MASKED-NEXT: [[TMP19:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP18]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1856 ; MASKED-NEXT: [[TMP20:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[MISMATCH_VEC_INDEX]]
1857 ; MASKED-NEXT: [[TMP21:%.*]] = call <vscale x 16 x i8> @llvm.masked.load.nxv16i8.p0(ptr [[TMP20]], i32 1, <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i8> zeroinitializer)
1858 ; MASKED-NEXT: [[TMP22:%.*]] = icmp ne <vscale x 16 x i8> [[TMP19]], [[TMP21]]
1859 ; MASKED-NEXT: [[TMP23:%.*]] = select <vscale x 16 x i1> [[MISMATCH_VEC_LOOP_PRED]], <vscale x 16 x i1> [[TMP22]], <vscale x 16 x i1> zeroinitializer
1860 ; MASKED-NEXT: [[TMP24:%.*]] = call i1 @llvm.vector.reduce.or.nxv16i1(<vscale x 16 x i1> [[TMP23]])
1861 ; MASKED-NEXT: br i1 [[TMP24]], label [[MISMATCH_VEC_LOOP_FOUND:%.*]], label [[MISMATCH_VEC_LOOP_INC]]
1862 ; MASKED: mismatch_vec_loop_inc:
1863 ; MASKED-NEXT: [[TMP25]] = add nuw nsw i64 [[MISMATCH_VEC_INDEX]], [[TMP17]]
1864 ; MASKED-NEXT: [[TMP26]] = call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 [[TMP25]], i64 0)
1865 ; MASKED-NEXT: [[TMP27:%.*]] = extractelement <vscale x 16 x i1> [[TMP26]], i64 0
1866 ; MASKED-NEXT: br i1 [[TMP27]], label [[MISMATCH_VEC_LOOP]], label [[MISMATCH_END:%.*]]
1867 ; MASKED: mismatch_vec_loop_found:
1868 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_PRED:%.*]] = phi <vscale x 16 x i1> [ [[TMP23]], [[MISMATCH_VEC_LOOP]] ]
1869 ; MASKED-NEXT: [[MISMATCH_VEC_LAST_LOOP_PRED:%.*]] = phi <vscale x 16 x i1> [ [[MISMATCH_VEC_LOOP_PRED]], [[MISMATCH_VEC_LOOP]] ]
1870 ; MASKED-NEXT: [[MISMATCH_VEC_FOUND_INDEX:%.*]] = phi i64 [ [[MISMATCH_VEC_INDEX]], [[MISMATCH_VEC_LOOP]] ]
1871 ; MASKED-NEXT: [[TMP28:%.*]] = and <vscale x 16 x i1> [[MISMATCH_VEC_LAST_LOOP_PRED]], [[MISMATCH_VEC_FOUND_PRED]]
1872 ; MASKED-NEXT: [[TMP29:%.*]] = call i32 @llvm.experimental.cttz.elts.i32.nxv16i1(<vscale x 16 x i1> [[TMP28]], i1 true)
1873 ; MASKED-NEXT: [[TMP30:%.*]] = zext i32 [[TMP29]] to i64
1874 ; MASKED-NEXT: [[TMP31:%.*]] = add nuw nsw i64 [[MISMATCH_VEC_FOUND_INDEX]], [[TMP30]]
1875 ; MASKED-NEXT: [[TMP32:%.*]] = trunc i64 [[TMP31]] to i32
1876 ; MASKED-NEXT: br label [[MISMATCH_END]]
1877 ; MASKED: mismatch_loop_pre:
1878 ; MASKED-NEXT: br label [[MISMATCH_LOOP:%.*]]
1879 ; MASKED: mismatch_loop:
1880 ; MASKED-NEXT: [[MISMATCH_INDEX:%.*]] = phi i32 [ 1, [[MISMATCH_LOOP_PRE]] ], [ [[TMP39:%.*]], [[MISMATCH_LOOP_INC:%.*]] ]
1881 ; MASKED-NEXT: [[TMP33:%.*]] = zext i32 [[MISMATCH_INDEX]] to i64
1882 ; MASKED-NEXT: [[TMP34:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[TMP33]]
1883 ; MASKED-NEXT: [[TMP35:%.*]] = load i8, ptr [[TMP34]], align 1
1884 ; MASKED-NEXT: [[TMP36:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[TMP33]]
1885 ; MASKED-NEXT: [[TMP37:%.*]] = load i8, ptr [[TMP36]], align 1
1886 ; MASKED-NEXT: [[TMP38:%.*]] = icmp eq i8 [[TMP35]], [[TMP37]]
1887 ; MASKED-NEXT: br i1 [[TMP38]], label [[MISMATCH_LOOP_INC]], label [[MISMATCH_END]]
1888 ; MASKED: mismatch_loop_inc:
1889 ; MASKED-NEXT: [[TMP39]] = add i32 [[MISMATCH_INDEX]], 1
1890 ; MASKED-NEXT: [[TMP40:%.*]] = icmp eq i32 [[TMP39]], 0
1891 ; MASKED-NEXT: br i1 [[TMP40]], label [[MISMATCH_END]], label [[MISMATCH_LOOP]]
1892 ; MASKED: mismatch_end:
1893 ; MASKED-NEXT: [[MISMATCH_RESULT:%.*]] = phi i32 [ 0, [[MISMATCH_LOOP_INC]] ], [ [[MISMATCH_INDEX]], [[MISMATCH_LOOP]] ], [ 0, [[MISMATCH_VEC_LOOP_INC]] ], [ [[TMP32]], [[MISMATCH_VEC_LOOP_FOUND]] ]
1894 ; MASKED-NEXT: br i1 true, label [[BYTE_COMPARE:%.*]], label [[WHILE_COND:%.*]]
1895 ; MASKED: while.cond:
1896 ; MASKED-NEXT: [[LEN:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY:%.*]] ], [ 0, [[MISMATCH_END]] ]
1897 ; MASKED-NEXT: [[INC:%.*]] = add i32 [[LEN]], 1
1898 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1899 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[CLEANUP_THREAD:%.*]], label [[WHILE_BODY]]
1900 ; MASKED: while.body:
1901 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[MISMATCH_RESULT]] to i64
1902 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[IDXPROM]]
1903 ; MASKED-NEXT: [[TMP41:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1904 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[IDXPROM]]
1905 ; MASKED-NEXT: [[TMP42:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1906 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP41]], [[TMP42]]
1907 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[IF_END:%.*]]
1908 ; MASKED: byte.compare:
1909 ; MASKED-NEXT: [[TMP43:%.*]] = icmp eq i32 [[MISMATCH_RESULT]], 0
1910 ; MASKED-NEXT: br i1 [[TMP43]], label [[CLEANUP_THREAD]], label [[IF_END]]
1911 ; MASKED: cleanup.thread:
1912 ; MASKED-NEXT: ret void
1913 ; MASKED: if.end:
1914 ; MASKED-NEXT: [[RES:%.*]] = phi i32 [ [[MISMATCH_RESULT]], [[WHILE_BODY]] ], [ [[MISMATCH_RESULT]], [[BYTE_COMPARE]] ]
1915 ; MASKED-NEXT: ret void
1916 ;
1917 ; NO-TRANSFORM-LABEL: define void @compare_bytes_cleanup_block(
1918 ; NO-TRANSFORM-SAME: ptr [[SRC1:%.*]], ptr [[SRC2:%.*]]) {
1919 ; NO-TRANSFORM-NEXT: entry:
1920 ; NO-TRANSFORM-NEXT: br label [[WHILE_COND:%.*]]
1921 ; NO-TRANSFORM: while.cond:
1922 ; NO-TRANSFORM-NEXT: [[LEN:%.*]] = phi i32 [ [[INC:%.*]], [[WHILE_BODY:%.*]] ], [ 0, [[ENTRY:%.*]] ]
1923 ; NO-TRANSFORM-NEXT: [[INC]] = add i32 [[LEN]], 1
1924 ; NO-TRANSFORM-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], 0
1925 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT]], label [[CLEANUP_THREAD:%.*]], label [[WHILE_BODY]]
1926 ; NO-TRANSFORM: while.body:
1927 ; NO-TRANSFORM-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
1928 ; NO-TRANSFORM-NEXT: [[ARRAYIDX:%.*]] = getelementptr i8, ptr [[SRC1]], i64 [[IDXPROM]]
1929 ; NO-TRANSFORM-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1930 ; NO-TRANSFORM-NEXT: [[ARRAYIDX2:%.*]] = getelementptr i8, ptr [[SRC2]], i64 [[IDXPROM]]
1931 ; NO-TRANSFORM-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1932 ; NO-TRANSFORM-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
1933 ; NO-TRANSFORM-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[IF_END:%.*]]
1934 ; NO-TRANSFORM: cleanup.thread:
1935 ; NO-TRANSFORM-NEXT: ret void
1936 ; NO-TRANSFORM: if.end:
1937 ; NO-TRANSFORM-NEXT: [[RES:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ]
1938 ; NO-TRANSFORM-NEXT: ret void
1939 entry:
1940 br label %while.cond
1942 while.cond:
1943 %len = phi i32 [ %inc, %while.body ], [ 0, %entry ]
1944 %inc = add i32 %len, 1
1945 %cmp.not = icmp eq i32 %inc, 0
1946 br i1 %cmp.not, label %cleanup.thread, label %while.body
1948 while.body:
1949 %idxprom = zext i32 %inc to i64
1950 %arrayidx = getelementptr i8, ptr %src1, i64 %idxprom
1951 %0 = load i8, ptr %arrayidx, align 1
1952 %arrayidx2 = getelementptr i8, ptr %src2, i64 %idxprom
1953 %1 = load i8, ptr %arrayidx2, align 1
1954 %cmp.not2 = icmp eq i8 %0, %1
1955 br i1 %cmp.not2, label %while.cond, label %if.end
1957 cleanup.thread:
1958 ret void
1960 if.end:
1961 %res = phi i32 [ %inc, %while.body ]
1962 ret void
1963 }
1965 ;
1966 ; NEGATIVE TESTS
1967 ;
1969 ; Similar to @compare_bytes_simple, except in the while.end block we have an extra PHI
1970 ; with unique values for each incoming block from the loop.
1971 define i32 @compare_bytes_simple2(ptr %a, ptr %b, ptr %c, ptr %d, i32 %len, i32 %n) {
1972 ; CHECK-LABEL: define i32 @compare_bytes_simple2(
1973 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
1974 ; CHECK-NEXT: entry:
1975 ; CHECK-NEXT: br label [[WHILE_COND:%.*]]
1976 ; CHECK: while.cond:
1977 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
1978 ; CHECK-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
1979 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
1980 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
1981 ; CHECK: while.body:
1982 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
1983 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
1984 ; CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
1985 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
1986 ; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
1987 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
1988 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
1989 ; CHECK: while.end:
1990 ; CHECK-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
1991 ; CHECK-NEXT: [[FINAL_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[D]], [[WHILE_COND]] ]
1992 ; CHECK-NEXT: store i32 [[INC_LCSSA]], ptr [[FINAL_PTR]], align 4
1993 ; CHECK-NEXT: ret i32 [[INC_LCSSA]]
1994 ;
1995 ; LMUL8-LABEL: define i32 @compare_bytes_simple2(
1996 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
1997 ; LMUL8-NEXT: entry:
1998 ; LMUL8-NEXT: br label [[WHILE_COND:%.*]]
1999 ; LMUL8: while.cond:
2000 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2001 ; LMUL8-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2002 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2003 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2004 ; LMUL8: while.body:
2005 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2006 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2007 ; LMUL8-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2008 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2009 ; LMUL8-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2010 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2011 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2012 ; LMUL8: while.end:
2013 ; LMUL8-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2014 ; LMUL8-NEXT: [[FINAL_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[D]], [[WHILE_COND]] ]
2015 ; LMUL8-NEXT: store i32 [[INC_LCSSA]], ptr [[FINAL_PTR]], align 4
2016 ; LMUL8-NEXT: ret i32 [[INC_LCSSA]]
2017 ;
2018 ; LOOP-DEL-LABEL: define i32 @compare_bytes_simple2(
2019 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2020 ; LOOP-DEL-NEXT: entry:
2021 ; LOOP-DEL-NEXT: br label [[WHILE_COND:%.*]]
2022 ; LOOP-DEL: while.cond:
2023 ; LOOP-DEL-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2024 ; LOOP-DEL-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2025 ; LOOP-DEL-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2026 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2027 ; LOOP-DEL: while.body:
2028 ; LOOP-DEL-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2029 ; LOOP-DEL-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2030 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2031 ; LOOP-DEL-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2032 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2033 ; LOOP-DEL-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2034 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2035 ; LOOP-DEL: while.end:
2036 ; LOOP-DEL-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2037 ; LOOP-DEL-NEXT: [[FINAL_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[D]], [[WHILE_COND]] ]
2038 ; LOOP-DEL-NEXT: store i32 [[INC_LCSSA]], ptr [[FINAL_PTR]], align 4
2039 ; LOOP-DEL-NEXT: ret i32 [[INC_LCSSA]]
2040 ;
2041 ; MASKED-LABEL: define i32 @compare_bytes_simple2(
2042 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], ptr [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2043 ; MASKED-NEXT: entry:
2044 ; MASKED-NEXT: br label [[WHILE_COND:%.*]]
2045 ; MASKED: while.cond:
2046 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2047 ; MASKED-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2048 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2049 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2050 ; MASKED: while.body:
2051 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2052 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2053 ; MASKED-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2054 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2055 ; MASKED-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2056 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2057 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2058 ; MASKED: while.end:
2059 ; MASKED-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2060 ; MASKED-NEXT: [[FINAL_PTR:%.*]] = phi ptr [ [[C]], [[WHILE_BODY]] ], [ [[D]], [[WHILE_COND]] ]
2061 ; MASKED-NEXT: store i32 [[INC_LCSSA]], ptr [[FINAL_PTR]], align 4
2062 ; MASKED-NEXT: ret i32 [[INC_LCSSA]]
2063 ;
2064 entry:
2065 br label %while.cond
2067 while.cond:
2068 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
2069 %inc = add i32 %len.addr, 1
2070 %cmp.not = icmp eq i32 %inc, %n
2071 br i1 %cmp.not, label %while.end, label %while.body
2073 while.body:
2074 %idxprom = zext i32 %inc to i64
2075 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
2076 %0 = load i8, ptr %arrayidx
2077 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
2078 %1 = load i8, ptr %arrayidx2
2079 %cmp.not2 = icmp eq i8 %0, %1
2080 br i1 %cmp.not2, label %while.cond, label %while.end
2082 while.end:
2083 %inc.lcssa = phi i32 [ %inc, %while.body ], [ %inc, %while.cond ]
2084 %final_ptr = phi ptr [ %c, %while.body ], [ %d, %while.cond ]
2085 store i32 %inc.lcssa, ptr %final_ptr
2086 ret i32 %inc.lcssa
2087 }
2089 define i32 @compare_bytes_simple3(ptr %a, ptr %b, ptr %c, i32 %d, i32 %len, i32 %n) {
2090 ; CHECK-LABEL: define i32 @compare_bytes_simple3(
2091 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2092 ; CHECK-NEXT: entry:
2093 ; CHECK-NEXT: br label [[WHILE_COND:%.*]]
2094 ; CHECK: while.cond:
2095 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2096 ; CHECK-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2097 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2098 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2099 ; CHECK: while.body:
2100 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2101 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2102 ; CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2103 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2104 ; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2105 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2106 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2107 ; CHECK: while.end:
2108 ; CHECK-NEXT: [[FINAL_VAL:%.*]] = phi i32 [ [[D]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2109 ; CHECK-NEXT: store i32 [[FINAL_VAL]], ptr [[C]], align 4
2110 ; CHECK-NEXT: ret i32 [[FINAL_VAL]]
2111 ;
2112 ; LMUL8-LABEL: define i32 @compare_bytes_simple3(
2113 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2114 ; LMUL8-NEXT: entry:
2115 ; LMUL8-NEXT: br label [[WHILE_COND:%.*]]
2116 ; LMUL8: while.cond:
2117 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2118 ; LMUL8-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2119 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2120 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2121 ; LMUL8: while.body:
2122 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2123 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2124 ; LMUL8-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2125 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2126 ; LMUL8-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2127 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2128 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2129 ; LMUL8: while.end:
2130 ; LMUL8-NEXT: [[FINAL_VAL:%.*]] = phi i32 [ [[D]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2131 ; LMUL8-NEXT: store i32 [[FINAL_VAL]], ptr [[C]], align 4
2132 ; LMUL8-NEXT: ret i32 [[FINAL_VAL]]
2133 ;
2134 ; LOOP-DEL-LABEL: define i32 @compare_bytes_simple3(
2135 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2136 ; LOOP-DEL-NEXT: entry:
2137 ; LOOP-DEL-NEXT: br label [[WHILE_COND:%.*]]
2138 ; LOOP-DEL: while.cond:
2139 ; LOOP-DEL-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2140 ; LOOP-DEL-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2141 ; LOOP-DEL-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2142 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2143 ; LOOP-DEL: while.body:
2144 ; LOOP-DEL-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2145 ; LOOP-DEL-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2146 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2147 ; LOOP-DEL-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2148 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2149 ; LOOP-DEL-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2150 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2151 ; LOOP-DEL: while.end:
2152 ; LOOP-DEL-NEXT: [[FINAL_VAL:%.*]] = phi i32 [ [[D]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2153 ; LOOP-DEL-NEXT: store i32 [[FINAL_VAL]], ptr [[C]], align 4
2154 ; LOOP-DEL-NEXT: ret i32 [[FINAL_VAL]]
2155 ;
2156 ; MASKED-LABEL: define i32 @compare_bytes_simple3(
2157 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i32 [[D:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR0]] {
2158 ; MASKED-NEXT: entry:
2159 ; MASKED-NEXT: br label [[WHILE_COND:%.*]]
2160 ; MASKED: while.cond:
2161 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2162 ; MASKED-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2163 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2164 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2165 ; MASKED: while.body:
2166 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2167 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2168 ; MASKED-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2169 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2170 ; MASKED-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2171 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2172 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2173 ; MASKED: while.end:
2174 ; MASKED-NEXT: [[FINAL_VAL:%.*]] = phi i32 [ [[D]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2175 ; MASKED-NEXT: store i32 [[FINAL_VAL]], ptr [[C]], align 4
2176 ; MASKED-NEXT: ret i32 [[FINAL_VAL]]
2177 ;
2178 entry:
2179 br label %while.cond
2181 while.cond:
2182 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
2183 %inc = add i32 %len.addr, 1
2184 %cmp.not = icmp eq i32 %inc, %n
2185 br i1 %cmp.not, label %while.end, label %while.body
2187 while.body:
2188 %idxprom = zext i32 %inc to i64
2189 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
2190 %0 = load i8, ptr %arrayidx
2191 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
2192 %1 = load i8, ptr %arrayidx2
2193 %cmp.not2 = icmp eq i8 %0, %1
2194 br i1 %cmp.not2, label %while.cond, label %while.end
2196 while.end:
2197 %final_val = phi i32 [ %d, %while.body ], [ %inc, %while.cond ]
2198 store i32 %final_val, ptr %c
2199 ret i32 %final_val
2200 }
2202 ; Disable the optimization when noimplicitfloat is present.
2203 define i32 @no_implicit_float(ptr %a, ptr %b, i32 %len, i32 %n) noimplicitfloat {
2204 ; CHECK-LABEL: define i32 @no_implicit_float(
2205 ; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR1:[0-9]+]] {
2206 ; CHECK-NEXT: entry:
2207 ; CHECK-NEXT: br label [[WHILE_COND:%.*]]
2208 ; CHECK: while.cond:
2209 ; CHECK-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2210 ; CHECK-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2211 ; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2212 ; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2213 ; CHECK: while.body:
2214 ; CHECK-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2215 ; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2216 ; CHECK-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2217 ; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2218 ; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2219 ; CHECK-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2220 ; CHECK-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2221 ; CHECK: while.end:
2222 ; CHECK-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2223 ; CHECK-NEXT: ret i32 [[INC_LCSSA]]
2224 ;
2225 ; LMUL8-LABEL: define i32 @no_implicit_float(
2226 ; LMUL8-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR1:[0-9]+]] {
2227 ; LMUL8-NEXT: entry:
2228 ; LMUL8-NEXT: br label [[WHILE_COND:%.*]]
2229 ; LMUL8: while.cond:
2230 ; LMUL8-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2231 ; LMUL8-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2232 ; LMUL8-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2233 ; LMUL8-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2234 ; LMUL8: while.body:
2235 ; LMUL8-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2236 ; LMUL8-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2237 ; LMUL8-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2238 ; LMUL8-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2239 ; LMUL8-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2240 ; LMUL8-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2241 ; LMUL8-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2242 ; LMUL8: while.end:
2243 ; LMUL8-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2244 ; LMUL8-NEXT: ret i32 [[INC_LCSSA]]
2245 ;
2246 ; LOOP-DEL-LABEL: define i32 @no_implicit_float(
2247 ; LOOP-DEL-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR1:[0-9]+]] {
2248 ; LOOP-DEL-NEXT: entry:
2249 ; LOOP-DEL-NEXT: br label [[WHILE_COND:%.*]]
2250 ; LOOP-DEL: while.cond:
2251 ; LOOP-DEL-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2252 ; LOOP-DEL-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2253 ; LOOP-DEL-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2254 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2255 ; LOOP-DEL: while.body:
2256 ; LOOP-DEL-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2257 ; LOOP-DEL-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2258 ; LOOP-DEL-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2259 ; LOOP-DEL-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2260 ; LOOP-DEL-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2261 ; LOOP-DEL-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2262 ; LOOP-DEL-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2263 ; LOOP-DEL: while.end:
2264 ; LOOP-DEL-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2265 ; LOOP-DEL-NEXT: ret i32 [[INC_LCSSA]]
2266 ;
2267 ; MASKED-LABEL: define i32 @no_implicit_float(
2268 ; MASKED-SAME: ptr [[A:%.*]], ptr [[B:%.*]], i32 [[LEN:%.*]], i32 [[N:%.*]]) #[[ATTR1:[0-9]+]] {
2269 ; MASKED-NEXT: entry:
2270 ; MASKED-NEXT: br label [[WHILE_COND:%.*]]
2271 ; MASKED: while.cond:
2272 ; MASKED-NEXT: [[LEN_ADDR:%.*]] = phi i32 [ [[LEN]], [[ENTRY:%.*]] ], [ [[INC:%.*]], [[WHILE_BODY:%.*]] ]
2273 ; MASKED-NEXT: [[INC]] = add i32 [[LEN_ADDR]], 1
2274 ; MASKED-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[INC]], [[N]]
2275 ; MASKED-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END:%.*]], label [[WHILE_BODY]]
2276 ; MASKED: while.body:
2277 ; MASKED-NEXT: [[IDXPROM:%.*]] = zext i32 [[INC]] to i64
2278 ; MASKED-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[A]], i64 [[IDXPROM]]
2279 ; MASKED-NEXT: [[TMP0:%.*]] = load i8, ptr [[ARRAYIDX]], align 1
2280 ; MASKED-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[IDXPROM]]
2281 ; MASKED-NEXT: [[TMP1:%.*]] = load i8, ptr [[ARRAYIDX2]], align 1
2282 ; MASKED-NEXT: [[CMP_NOT2:%.*]] = icmp eq i8 [[TMP0]], [[TMP1]]
2283 ; MASKED-NEXT: br i1 [[CMP_NOT2]], label [[WHILE_COND]], label [[WHILE_END]]
2284 ; MASKED: while.end:
2285 ; MASKED-NEXT: [[INC_LCSSA:%.*]] = phi i32 [ [[INC]], [[WHILE_BODY]] ], [ [[INC]], [[WHILE_COND]] ]
2286 ; MASKED-NEXT: ret i32 [[INC_LCSSA]]
2287 ;
2288 entry:
2289 br label %while.cond
2291 while.cond:
2292 %len.addr = phi i32 [ %len, %entry ], [ %inc, %while.body ]
2293 %inc = add i32 %len.addr, 1
2294 %cmp.not = icmp eq i32 %inc, %n
2295 br i1 %cmp.not, label %while.end, label %while.body
2297 while.body:
2298 %idxprom = zext i32 %inc to i64
2299 %arrayidx = getelementptr inbounds i8, ptr %a, i64 %idxprom
2300 %0 = load i8, ptr %arrayidx
2301 %arrayidx2 = getelementptr inbounds i8, ptr %b, i64 %idxprom
2302 %1 = load i8, ptr %arrayidx2
2303 %cmp.not2 = icmp eq i8 %0, %1
2304 br i1 %cmp.not2, label %while.cond, label %while.end
2306 while.end:
2307 %inc.lcssa = phi i32 [ %inc, %while.body ], [ %inc, %while.cond ]
2308 ret i32 %inc.lcssa
2309 }