3 ; RUN: opt -passes=loop-vectorize -force-vector-width=1 -force-vector-interleave=2 -debug -disable-output %s 2>&1 | FileCheck --check-prefix=DBG %s
4 ; RUN: opt -passes=loop-vectorize -force-vector-width=1 -force-vector-interleave=2 -S %s | FileCheck %s
6 ; DBG-LABEL: 'test_scalarize_call'
7 ; DBG: VPlan 'Initial VPlan for VF={1},UF>=1' {
8 ; DBG-NEXT: Live-in vp<[[VFxUF:%.+]]> = VF * UF
9 ; DBG-NEXT: Live-in vp<[[VEC_TC:%.+]]> = vector-trip-count
10 ; DBG-NEXT: vp<[[TC:%.+]]> = original trip-count
13 ; DBG-NEXT: EMIT vp<[[TC]]> = EXPAND SCEV (1000 + (-1 * %start))
14 ; DBG-NEXT: No successors
16 ; DBG-NEXT: vector.ph:
17 ; DBG-NEXT: Successor(s): vector loop
19 ; DBG-NEXT: <x1> vector loop: {
20 ; DBG-NEXT: vector.body:
21 ; DBG-NEXT: EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
22 ; DBG-NEXT: vp<[[DERIVED_IV:%.+]]> = DERIVED-IV ir<%start> + vp<[[CAN_IV]]> * ir<1>
23 ; DBG-NEXT: vp<[[IV_STEPS:%.]]> = SCALAR-STEPS vp<[[DERIVED_IV]]>, ir<1>
24 ; DBG-NEXT: CLONE ir<%min> = call @llvm.smin.i32(vp<[[IV_STEPS]]>, ir<65535>)
25 ; DBG-NEXT: CLONE ir<%arrayidx> = getelementptr inbounds ir<%dst>, vp<[[IV_STEPS]]>
26 ; DBG-NEXT: CLONE store ir<%min>, ir<%arrayidx>
27 ; DBG-NEXT: EMIT vp<[[INC:%.+]]> = add nuw vp<[[CAN_IV]]>, vp<[[VFxUF]]>
28 ; DBG-NEXT: EMIT branch-on-count vp<[[INC]]>, vp<[[VEC_TC]]>
29 ; DBG-NEXT: No successors
32 define void @test_scalarize_call(i32 %start, ptr %dst) {
33 ; CHECK-LABEL: @test_scalarize_call(
35 ; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
36 ; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i32 %start, [[INDEX]]
37 ; CHECK-NEXT: [[INDUCTION:%.*]] = add i32 [[OFFSET_IDX]], 0
38 ; CHECK-NEXT: [[INDUCTION1:%.*]] = add i32 [[OFFSET_IDX]], 1
39 ; CHECK-NEXT: [[TMP1:%.*]] = tail call i32 @llvm.smin.i32(i32 [[INDUCTION]], i32 65535)
40 ; CHECK-NEXT: [[TMP2:%.*]] = tail call i32 @llvm.smin.i32(i32 [[INDUCTION1]], i32 65535)
41 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[DST:%.*]], i32 [[INDUCTION]]
42 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds i32, ptr [[DST]], i32 [[INDUCTION1]]
43 ; CHECK-NEXT: store i32 [[TMP1]], ptr [[TMP3]], align 8
44 ; CHECK-NEXT: store i32 [[TMP2]], ptr [[TMP4]], align 8
45 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
46 ; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[INDEX_NEXT]], %n.vec
47 ; CHECK-NEXT: br i1 [[TMP5]], label %middle.block, label %vector.body
48 ; CHECK: middle.block:
54 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ]
55 %min = tail call i32 @llvm.smin.i32(i32 %iv, i32 65535)
56 %arrayidx = getelementptr inbounds i32 , ptr %dst, i32 %iv
57 store i32 %min, ptr %arrayidx, align 8
58 %iv.next = add nsw i32 %iv, 1
59 %tobool.not = icmp eq i32 %iv.next, 1000
60 br i1 %tobool.not, label %exit, label %loop
66 declare i32 @llvm.smin.i32(i32, i32)
69 ; DBG-LABEL: 'test_scalarize_with_branch_cond'
71 ; DBG: Live-in vp<[[VFxUF:%.+]]> = VF * UF
72 ; DBG-NEXT: Live-in vp<[[VEC_TC:%.+]]> = vector-trip-count
73 ; DBG-NEXT: Live-in ir<1000> = original trip-count
75 ; DBG-NEXT: vector.ph:
76 ; DBG-NEXT: Successor(s): vector loop
78 ; DBG-NEXT: <x1> vector loop: {
79 ; DBG-NEXT: vector.body:
80 ; DBG-NEXT: EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
81 ; DBG-NEXT: vp<[[DERIVED_IV:%.+]]> = DERIVED-IV ir<false> + vp<[[CAN_IV]]> * ir<true>
82 ; DBG-NEXT: vp<[[STEPS1:%.+]]> = SCALAR-STEPS vp<[[DERIVED_IV]]>, ir<true>
83 ; DBG-NEXT: Successor(s): pred.store
85 ; DBG-NEXT: <xVFxUF> pred.store: {
86 ; DBG-NEXT: pred.store.entry:
87 ; DBG-NEXT: BRANCH-ON-MASK vp<[[STEPS1]]>
88 ; DBG-NEXT: Successor(s): pred.store.if, pred.store.continue
90 ; DBG-NEXT: pred.store.if:
91 ; DBG-NEXT: vp<[[STEPS2:%.+]]> = SCALAR-STEPS vp<[[CAN_IV]]>, ir<1>
92 ; DBG-NEXT: CLONE ir<%gep.src> = getelementptr inbounds ir<%src>, vp<[[STEPS2]]>
93 ; DBG-NEXT: CLONE ir<%l> = load ir<%gep.src>
94 ; DBG-NEXT: CLONE ir<%gep.dst> = getelementptr inbounds ir<%dst>, vp<[[STEPS2]]>
95 ; DBG-NEXT: CLONE store ir<%l>, ir<%gep.dst>
96 ; DBG-NEXT: Successor(s): pred.store.continue
98 ; DBG-NEXT: pred.store.continue:
99 ; DBG-NEXT: PHI-PREDICATED-INSTRUCTION vp<{{.+}}> = ir<%l>
100 ; DBG-NEXT: No successors
102 ; DBG-NEXT: Successor(s): cond.false.1
104 ; DBG-NEXT: cond.false.1:
105 ; DBG-NEXT: EMIT vp<[[CAN_IV_INC:%.+]]> = add nuw vp<[[CAN_IV]]>, vp<[[VFxUF]]>
106 ; DBG-NEXT: EMIT branch-on-count vp<[[CAN_IV_INC]]>, vp<[[VEC_TC]]>
107 ; DBG-NEXT: No successors
109 ; DBG-NEXT: Successor(s): middle.block
111 ; DBG-NEXT: middle.block:
112 ; DBG-NEXT: No successors
115 define void @test_scalarize_with_branch_cond(ptr %src, ptr %dst) {
116 ; CHECK-LABEL: @test_scalarize_with_branch_cond(
117 ; CHECK: vector.body:
118 ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %pred.store.continue5 ]
119 ; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[INDEX]] to i1
120 ; CHECK-NEXT: [[OFFSET_IDX:%.*]] = sub i1 false, [[TMP0]]
121 ; CHECK-NEXT: [[INDUCTION:%.*]] = add i1 [[OFFSET_IDX]], false
122 ; CHECK-NEXT: [[INDUCTION3:%.*]] = add i1 [[OFFSET_IDX]], true
123 ; CHECK-NEXT: br i1 [[INDUCTION]], label %pred.store.if, label %pred.store.continue
124 ; CHECK: pred.store.if:
125 ; CHECK-NEXT: [[INDUCTION4:%.*]] = add i64 [[INDEX]], 0
126 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr %src, i64 [[INDUCTION4]]
127 ; CHECK-NEXT: [[TMP4:%.*]] = load i32, ptr [[TMP3]], align 4
128 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i32, ptr %dst, i64 [[INDUCTION4]]
129 ; CHECK-NEXT: store i32 [[TMP4]], ptr [[TMP1]], align 4
130 ; CHECK-NEXT: br label %pred.store.continue
131 ; CHECK: pred.store.continue:
132 ; CHECK-NEXT: [[TMP5:%.*]] = phi i32 [ poison, %vector.body ], [ [[TMP4]], %pred.store.if ]
133 ; CHECK-NEXT: br i1 [[INDUCTION3]], label %pred.store.if4, label %pred.store.continue5
134 ; CHECK: pred.store.if4:
135 ; CHECK-NEXT: [[INDUCTION5:%.*]] = add i64 [[INDEX]], 1
136 ; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr %src, i64 [[INDUCTION5]]
137 ; CHECK-NEXT: [[TMP7:%.*]] = load i32, ptr [[TMP6]], align 4
138 ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i32, ptr %dst, i64 [[INDUCTION5]]
139 ; CHECK-NEXT: store i32 [[TMP7]], ptr [[TMP2]], align 4
140 ; CHECK-NEXT: br label %pred.store.continue5
141 ; CHECK: pred.store.continue5:
142 ; CHECK-NEXT: [[TMP8:%.*]] = phi i32 [ poison, %pred.store.continue ], [ [[TMP7]], %pred.store.if4 ]
143 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
144 ; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1000
145 ; CHECK-NEXT: br i1 [[TMP9]], label %middle.block, label %vector.body
146 ; CHECK: middle.block:
149 br label %loop.header
152 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]
153 %d = phi i1 [ false, %entry ], [ %d.next, %loop.latch ]
154 %d.next = xor i1 %d, true
155 br i1 %d, label %cond.false, label %loop.latch
158 %gep.src = getelementptr inbounds i32, ptr %src, i64 %iv
159 %gep.dst = getelementptr inbounds i32, ptr %dst, i64 %iv
160 %l = load i32, ptr %gep.src, align 4
161 store i32 %l, ptr %gep.dst
165 %iv.next = add nsw i64 %iv, 1
166 %ec = icmp eq i64 %iv.next, 1000
167 br i1 %ec, label %exit, label %loop.header
173 ; Make sure the widened induction gets replaced by scalar-steps for plans
174 ; including the scalar VF, if it is used in first-order recurrences.
176 ; DBG-LABEL: 'first_order_recurrence_using_induction'
177 ; DBG: VPlan 'Initial VPlan for VF={1},UF>=1' {
178 ; DBG-NEXT: Live-in vp<[[VFxUF:%.+]]> = VF * UF
179 ; DBG-NEXT: Live-in vp<[[VTC:%.+]]> = vector-trip-count
180 ; DBG-NEXT: vp<[[TC:%.+]]> = original trip-count
183 ; DBG-NEXT: EMIT vp<[[TC]]> = EXPAND SCEV (zext i32 (1 smax %n) to i64)
184 ; DBG-NEXT: No successors
186 ; DBG-NEXT: vector.ph:
187 ; DBG-NEXT: Successor(s): vector loop
189 ; DBG-NEXT: <x1> vector loop: {
190 ; DBG-NEXT: vector.body:
191 ; DBG-NEXT: EMIT vp<[[CAN_IV:%.+]]> = CANONICAL-INDUCTION
192 ; DBG-NEXT: FIRST-ORDER-RECURRENCE-PHI ir<%for> = phi ir<0>, vp<[[SCALAR_STEPS:.+]]>
193 ; DBG-NEXT: vp<[[DERIVED_IV:%.+]]> = DERIVED-IV ir<0> + vp<[[CAN_IV]]> * ir<1> (truncated to i32)
194 ; DBG-NEXT: vp<[[SCALAR_STEPS]]> = SCALAR-STEPS vp<[[DERIVED_IV]]>, ir<1>
195 ; DBG-NEXT: EMIT vp<[[SPLICE:%.+]]> = first-order splice ir<%for>, vp<[[SCALAR_STEPS]]>
196 ; DBG-NEXT: CLONE store vp<[[SPLICE]]>, ir<%dst>
197 ; DBG-NEXT: EMIT vp<[[IV_INC:%.+]]> = add nuw vp<[[CAN_IV]]>, vp<[[VFxUF]]>
198 ; DBG-NEXT: EMIT branch-on-count vp<[[IV_INC]]>, vp<[[VTC]]>
199 ; DBG-NEXT: No successors
201 ; DBG-NEXT: Successor(s): middle.block
203 ; DBG-NEXT: middle.block:
204 ; DBG-NEXT: No successors
207 define void @first_order_recurrence_using_induction(i32 %n, ptr %dst) {
208 ; CHECK-LABEL: @first_order_recurrence_using_induction(
209 ; CHECK: vector.body:
210 ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
211 ; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi i32 [ 0, %vector.ph ], [ [[INDUCTION1:%.*]], %vector.body ]
212 ; CHECK-NEXT: [[TMP3:%.*]] = trunc i64 [[INDEX]] to i32
213 ; CHECK-NEXT: [[INDUCTION:%.*]] = add i32 [[TMP3]], 0
214 ; CHECK-NEXT: [[INDUCTION1]] = add i32 [[TMP3]], 1
215 ; CHECK-NEXT: store i32 [[VECTOR_RECUR]], ptr [[DST:%.*]], align 4
216 ; CHECK-NEXT: store i32 [[INDUCTION]], ptr [[DST]], align 4
217 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
218 ; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], %n.vec
219 ; CHECK-NEXT: br i1 [[TMP4]], label %middle.block, label %vector.body
220 ; CHECK: middle.block:
226 %iv = phi i64 [ 0, %entry ],[ %iv.next, %loop ]
227 %for = phi i32 [ 0, %entry ], [ %iv.trunc, %loop ]
228 %iv.trunc = trunc i64 %iv to i32
229 store i32 %for, ptr %dst
230 %iv.next = add nuw nsw i64 %iv, 1
231 %iv.next.trunc = trunc i64 %iv.next to i32
232 %ec = icmp slt i32 %iv.next.trunc, %n
233 br i1 %ec, label %loop, label %exit
239 define i16 @reduction_with_casts() {
240 ; CHECK-LABEL: define i16 @reduction_with_casts() {
241 ; CHECK: vector.body:
242 ; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH:%.+]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY:%.+]] ]
243 ; CHECK-NEXT: [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP2:%.*]], [[VECTOR_BODY]] ]
244 ; CHECK-NEXT: [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
245 ; CHECK-NEXT: [[TMP0:%.*]] = and i32 [[VEC_PHI]], 65535
246 ; CHECK-NEXT: [[TMP1:%.*]] = and i32 [[VEC_PHI1]], 65535
247 ; CHECK-NEXT: [[TMP2]] = add i32 [[TMP0]], 1
248 ; CHECK-NEXT: [[TMP3]] = add i32 [[TMP1]], 1
249 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 2
250 ; CHECK-NEXT: [[TMP4:%.*]] = icmp eq i32 [[INDEX_NEXT]], 9998
251 ; CHECK-NEXT: br i1 [[TMP4]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]]
252 ; CHECK: middle.block:
253 ; CHECK-NEXT: [[BIN_RDX:%.*]] = add i32 [[TMP3]], [[TMP2]]
254 ; CHECK-NEXT: br i1 false, label [[EXIT:%.*]], label %scalar.ph
260 %count.0.in1 = phi i32 [ 0, %entry ], [ %add, %loop ]
261 %iv = phi i16 [ 1, %entry ], [ %iv.next, %loop ]
262 %conv1 = and i32 %count.0.in1, 65535
263 %add = add nuw nsw i32 %conv1, 1
264 %iv.next = add i16 %iv, 1
265 %cmp = icmp eq i16 %iv.next, 10000
266 br i1 %cmp, label %exit, label %loop
269 %add.lcssa = phi i32 [ %add, %loop ]
270 %count.0 = trunc i32 %add.lcssa to i16
274 define void @scalarize_ptrtoint(ptr %src, ptr %dst) {
275 ; CHECK: vector.body:
276 ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
277 ; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
278 ; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
279 ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr ptr, ptr %src, i64 [[TMP0]]
280 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr ptr, ptr %src, i64 [[TMP1]]
281 ; CHECK-NEXT: [[TMP4:%.*]] = load ptr, ptr [[TMP2]], align 8
282 ; CHECK-NEXT: [[TMP5:%.*]] = load ptr, ptr [[TMP3]], align 8
283 ; CHECK-NEXT: [[TMP6:%.*]] = ptrtoint ptr [[TMP4]] to i64
284 ; CHECK-NEXT: [[TMP7:%.*]] = ptrtoint ptr [[TMP5]] to i64
285 ; CHECK-NEXT: [[TMP8:%.*]] = add i64 [[TMP6]], 10
286 ; CHECK-NEXT: [[TMP9:%.*]] = add i64 [[TMP7]], 10
287 ; CHECK-NEXT: [[TMP10:%.*]] = inttoptr i64 [[TMP8]] to ptr
288 ; CHECK-NEXT: [[TMP11:%.*]] = inttoptr i64 [[TMP9]] to ptr
289 ; CHECK-NEXT: store ptr [[TMP10]], ptr %dst, align 8
290 ; CHECK-NEXT: store ptr [[TMP11]], ptr %dst, align 8
291 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
292 ; CHECK-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 0
293 ; CHECK-NEXT: br i1 [[TMP12]], label %middle.block, label %vector.body
299 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
300 %gep = getelementptr ptr, ptr %src, i64 %iv
301 %l = load ptr, ptr %gep, align 8
302 %cast = ptrtoint ptr %l to i64
303 %add = add i64 %cast, 10
304 %cast.2 = inttoptr i64 %add to ptr
305 store ptr %cast.2, ptr %dst, align 8
306 %iv.next = add i64 %iv, 1
307 %ec = icmp eq i64 %iv.next, 0
308 br i1 %ec, label %exit, label %loop
314 define void @pr76986_trunc_sext_interleaving_only(i16 %arg, ptr noalias %src, ptr noalias %dst) {
315 ; CHECK-LABEL: define void @pr76986_trunc_sext_interleaving_only(
316 ; CHECK: vector.body:
317 ; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
318 ; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[INDEX]], 0
319 ; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 1
320 ; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i8, ptr %src, i64 [[TMP0]]
321 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr %src, i64 [[TMP1]]
322 ; CHECK-NEXT: [[TMP4:%.*]] = load i8, ptr [[TMP2]], align 1
323 ; CHECK-NEXT: [[TMP5:%.*]] = load i8, ptr [[TMP3]], align 1
324 ; CHECK-NEXT: [[TMP6:%.*]] = sext i8 [[TMP4]] to i32
325 ; CHECK-NEXT: [[TMP7:%.*]] = sext i8 [[TMP5]] to i32
326 ; CHECK-NEXT: [[TMP8:%.*]] = trunc i32 [[TMP6]] to i16
327 ; CHECK-NEXT: [[TMP9:%.*]] = trunc i32 [[TMP7]] to i16
328 ; CHECK-NEXT: [[TMP10:%.*]] = sdiv i16 [[TMP8]], %arg
329 ; CHECK-NEXT: [[TMP11:%.*]] = sdiv i16 [[TMP9]], %arg
330 ; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i16, ptr %dst, i64 [[TMP0]]
331 ; CHECK-NEXT: [[TMP13:%.*]] = getelementptr inbounds i16, ptr %dst, i64 [[TMP1]]
332 ; CHECK-NEXT: store i16 [[TMP10]], ptr [[TMP12]], align 2
333 ; CHECK-NEXT: store i16 [[TMP11]], ptr [[TMP13]], align 2
334 ; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
335 ; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], 14934
336 ; CHECK-NEXT: br i1 [[TMP14]], label %middle.block, label %vector.body
342 %iv = phi i64 [ 0, %bb ], [ %iv.next, %loop ]
343 %gep.src = getelementptr inbounds i8, ptr %src, i64 %iv
344 %l = load i8, ptr %gep.src
345 %sext = sext i8 %l to i32
346 %trunc = trunc i32 %sext to i16
347 %sdiv = sdiv i16 %trunc, %arg
348 %gep.dst = getelementptr inbounds i16, ptr %dst, i64 %iv
349 store i16 %sdiv, ptr %gep.dst
350 %iv.next = add i64 %iv, 1
351 %icmp = icmp ult i64 %iv, 14933
352 br i1 %icmp, label %loop, label %exit