1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
|
; RUN: opt -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK
; RUN: opt -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -S < %s | FileCheck %s --check-prefix=CHECK
; RUN: opt -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=1 -S < %s | FileCheck %s --check-prefix=CHECK
; This test can theoretically be vectorized without a runtime-check, by
; pattern-matching on the constructs that are introduced by IndVarSimplify.
; We can check two things:
; %1 = trunc i64 %iv to i32
; This indicates that the %iv is truncated to i32. We can then check the loop
; guard is a signed i32:
; %cmp.sgt = icmp sgt i32 %n, 0
; and successfully vectorize the case without a runtime-check.
define i32 @select_icmp_const_truncated_iv_widened_exit(ptr %a, i32 %n) {
; CHECK-LABEL: define i32 @select_icmp_const_truncated_iv_widened_exit
; CHECK-NOT: vector.body:
;
entry:
%cmp.sgt = icmp sgt i32 %n, 0
br i1 %cmp.sgt, label %for.body.preheader, label %exit
for.body.preheader: ; preds = %entry
%wide.trip.count = zext i32 %n to i64
br label %for.body
for.body: ; preds = %for.body.preheader, %for.body
%iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
%rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
%0 = load i64, ptr %arrayidx, align 8
%cmp = icmp sgt i64 %0, 3
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, %wide.trip.count
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body, %entry
%rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
ret i32 %rdx.lcssa
}
; This test can theoretically be vectorized without a runtime-check, by
; pattern-matching on the constructs that are introduced by IndVarSimplify.
; We can check two things:
; %1 = trunc i64 %iv to i32
; This indicates that the %iv is truncated to i32. We can then check the loop
; exit condition, which compares to a constant that fits within i32:
; %exitcond.not = icmp eq i64 %inc, 20000
; and successfully vectorize the case without a runtime-check.
define i32 @select_icmp_const_truncated_iv_const_exit(ptr %a) {
; CHECK-LABEL: define i32 @select_icmp_const_truncated_iv_const_exit
; CHECK-NOT: vector.body:
;
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds i64, ptr %a, i64 %iv
%0 = load i64, ptr %arrayidx, align 8
%cmp = icmp sgt i64 %0, 3
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, 20000
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body
ret i32 %spec.select
}
; Without loop guard, the maximum constant trip count that can be vectorized is
; the signed maximum value of reduction type.
define i32 @select_fcmp_max_valid_const_ub(ptr %a) {
; CHECK-LABEL: define i32 @select_fcmp_max_valid_const_ub
; CHECK-NOT: vector.body:
;
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ -1, %entry ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds float, ptr %a, i64 %iv
%0 = load float, ptr %arrayidx, align 4
%cmp = fcmp fast olt float %0, 0.000000e+00
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, 2147483648
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body
ret i32 %spec.select
}
; Negative tests
; This test can theoretically be vectorized, but only with a runtime-check.
; The construct that are introduced by IndVarSimplify is:
; %1 = trunc i64 %iv to i32
; However, the loop guard is an i64:
; %cmp.sgt = icmp sgt i64 %n, 0
; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
; sentinel value), and need a runtime-check to vectorize this case.
define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit(ptr %a, i64 %n) {
; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unwidened_exit
; CHECK-NOT: vector.body:
;
entry:
%cmp.sgt = icmp sgt i64 %n, 0
br i1 %cmp.sgt, label %for.body, label %exit
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%cmp = icmp sgt i32 %0, 3
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, %n
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body, %entry
%rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
ret i32 %rdx.lcssa
}
; This test can theoretically be vectorized, but only with a runtime-check.
; The construct that are introduced by IndVarSimplify is:
; %1 = trunc i64 %iv to i32
; However, the loop guard is unsigned:
; %cmp.not = icmp eq i32 %n, 0
; We cannot guarantee that %iv won't overflow an i32 value (and hence hit the
; sentinel value), and need a runtime-check to vectorize this case.
define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard(ptr %a, i32 %n) {
; CHECK-LABEL: define i32 @not_vectorized_select_icmp_const_truncated_iv_unsigned_loop_guard
; CHECK-NOT: vector.body:
;
entry:
%cmp.not = icmp eq i32 %n, 0
br i1 %cmp.not, label %exit, label %for.body.preheader
for.body.preheader: ; preds = %entry
%wide.trip.count = zext i32 %n to i64
br label %for.body
for.body: ; preds = %for.body.preheader, %for.body
%iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
%rdx = phi i32 [ 331, %for.body.preheader ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%cmp1 = icmp sgt i32 %0, 3
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp1, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, %wide.trip.count
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body, %entry
%rdx.lcssa = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
ret i32 %rdx.lcssa
}
; This test cannot be vectorized, even with a runtime check.
; The construct that are introduced by IndVarSimplify is:
; %1 = trunc i64 %iv to i32
; However, the loop exit condition is a constant that overflows i32:
; %exitcond.not = icmp eq i64 %inc, 4294967294
; Hence, the i32 will most certainly wrap and hit the sentinel value, and we
; cannot vectorize this case.
define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound(ptr %a) {
; CHECK-LABEL: define i32 @not_vectorized_select_icmp_truncated_iv_out_of_bound
; CHECK-NOT: vector.body:
;
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 2147483646, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%cmp = icmp sgt i32 %0, 3
%conv = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %conv, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, 4294967294
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body
ret i32 %spec.select
}
; Forbidding vectorization of the FindLastIV pattern involving a truncated
; induction variable in the absence of any loop guard.
define i32 @not_vectorized_select_iv_icmp_no_guard(ptr %a, ptr %b, i32 %start, i32 %n) {
; CHECK-LABEL: define i32 @not_vectorized_select_iv_icmp_no_guard
; CHECK-NOT: vector.body:
;
entry:
%wide.trip.count = zext i32 %n to i64
br label %for.body
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ %start, %entry ], [ %cond, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%arrayidx2 = getelementptr inbounds i32, ptr %b, i64 %iv
%1 = load i32, ptr %arrayidx2, align 4
%cmp = icmp sgt i32 %0, %1
%2 = trunc i64 %iv to i32
%cond = select i1 %cmp, i32 %2, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, %wide.trip.count
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body
ret i32 %cond
}
; Without loop guard, when the constant trip count exceeds the maximum signed
; value of the reduction type, truncation may cause overflow. Therefore,
; vectorizer is unable to guarantee that the induction variable is monotonic
; increasing.
define i32 @not_vectorized_select_fcmp_invalid_const_ub(ptr %a) {
; CHECK-LABEL: define i32 @not_vectorized_select_fcmp_invalid_const_ub
; CHECK-NOT: vector.body:
;
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%iv = phi i64 [ 0, %entry ], [ %inc, %for.body ]
%rdx = phi i32 [ -1, %entry ], [ %spec.select, %for.body ]
%arrayidx = getelementptr inbounds float, ptr %a, i64 %iv
%0 = load float, ptr %arrayidx, align 4
%cmp = fcmp fast olt float %0, 0.000000e+00
%1 = trunc i64 %iv to i32
%spec.select = select i1 %cmp, i32 %1, i32 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, 2147483649
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body
ret i32 %spec.select
}
; Even with loop guard protection, if the destination type of the truncation
; instruction is smaller than the trip count type before extension, overflow
; could still occur.
define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount(ptr %a, ptr %b, i16 %start, i32 %n) {
; CHECK-LABEL: define i16 @not_vectorized_select_iv_icmp_overflow_unwidened_tripcount
; CHECK-NOT: vector.body:
;
entry:
%cmp9 = icmp sgt i32 %n, 0
br i1 %cmp9, label %for.body.preheader, label %exit
for.body.preheader: ; preds = %entry
%wide.trip.count = zext i32 %n to i64
br label %for.body
for.body: ; preds = %for.body.preheader, %for.body
%iv = phi i64 [ 0, %for.body.preheader ], [ %inc, %for.body ]
%rdx = phi i16 [ %start, %for.body.preheader ], [ %cond, %for.body ]
%arrayidx = getelementptr inbounds i32, ptr %a, i64 %iv
%0 = load i32, ptr %arrayidx, align 4
%arrayidx2 = getelementptr inbounds i32, ptr %b, i64 %iv
%1 = load i32, ptr %arrayidx2, align 4
%cmp3 = icmp sgt i32 %0, %1
%2 = trunc i64 %iv to i16
%cond = select i1 %cmp3, i16 %2, i16 %rdx
%inc = add nuw nsw i64 %iv, 1
%exitcond.not = icmp eq i64 %inc, %wide.trip.count
br i1 %exitcond.not, label %exit, label %for.body
exit: ; preds = %for.body, %entry
%rdx.0.lcssa = phi i16 [ %start, %entry ], [ %cond, %for.body ]
ret i16 %rdx.0.lcssa
}
|
