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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
|
; Several inner-loop reduction patterns are not yet supported.
; RUN: opt < %s -passes="loop-interchange" -loop-interchange-reduction-to-mem -pass-remarks-missed='loop-interchange' \
; RUN: -pass-remarks-output=%t -S | FileCheck -check-prefix=IR %s
; RUN: FileCheck --input-file=%t %s
; 1. The initial value of the reduction is not a constant.
; for (int i = 0; i < n; i++) {
; r = s[i];
; for (int j = 0; j < n; j++)
; r = r + a[j][i] * b[j][i];
; s[i] = r;
; }
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedInnerReduction
; CHECK-NEXT: Function: reduction_01
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only supported for the reduction with a constant initial value.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIInner
; CHECK-NEXT: Function: reduction_01
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only inner loops with induction or reduction PHI nodes can be interchange currently.
; IR-LABEL: @reduction_01(
; IR-NOT: split
define void @reduction_01(ptr noalias readonly %a, ptr noalias readonly %b, ptr noalias writeonly %s, i64 %n) {
entry:
%cmp = icmp sgt i64 %n, 0
br i1 %cmp, label %outerloop_header, label %exit
outerloop_header:
%index_i = phi i64 [ 0, %entry ], [ %index_i.next, %outerloop_latch ]
%addr_s = getelementptr inbounds nuw double, ptr %s, i64 %index_i
%invariant.gep.us = getelementptr inbounds nuw double, ptr %a, i64 %index_i
%invariant.gep32.us = getelementptr inbounds nuw double, ptr %b, i64 %index_i
%s_init = load double, ptr %addr_s, align 8
br label %innerloop
innerloop:
%index_j = phi i64 [ 0, %outerloop_header ], [ %index_j.next, %innerloop ]
%reduction = phi double [ %s_init, %outerloop_header ], [ %add, %innerloop ]
%addr_a_j_i = getelementptr inbounds nuw double, ptr %invariant.gep.us, i64 %index_j
%0 = load double, ptr %addr_a_j_i, align 8
%addr_b_j_i = getelementptr inbounds nuw double, ptr %invariant.gep32.us, i64 %index_j
%1 = load double, ptr %addr_b_j_i, align 8
%mul = fmul fast double %1, %0
%add = fadd fast double %mul, %reduction
%index_j.next = add nuw nsw i64 %index_j, 1
%cond1 = icmp eq i64 %index_j.next, %n
br i1 %cond1, label %outerloop_latch, label %innerloop
outerloop_latch:
%lcssa = phi double [ %add, %innerloop ]
store double %lcssa, ptr %addr_s, align 8
%index_i.next = add nuw nsw i64 %index_i, 1
%cond2 = icmp eq i64 %index_i.next, %n
br i1 %cond2, label %exit, label %outerloop_header
exit:
ret void
}
; 2. There are two or more reductions
; for (int i = 0; i < n; i++) {
; r1 = 0;
; r2 = 0;
; for (int j = 0; j < n; j++){
; r1 = r1 + a[j][i] * b[j][i];
; r2 = r2 + a[j][i];
; }
; s[i] = r1;
; s2[i] = r2;
; }
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedInnerReduction
; CHECK-NEXT: Function: reduction_02
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only supports at most one reduction.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIInner
; CHECK-NEXT: Function: reduction_02
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only inner loops with induction or reduction PHI nodes can be interchange currently.
; IR-LABEL: @reduction_02(
; IR-NOT: split
define void @reduction_02(ptr noalias readonly %a, ptr noalias readonly %b, ptr noalias writeonly %s, ptr noalias writeonly %s2, i64 %n) {
entry:
%cmp = icmp sgt i64 %n, 0
br i1 %cmp, label %outerloop_header, label %exit
outerloop_header:
%index_i = phi i64 [ 0, %entry ], [ %index_i.next, %outerloop_latch ]
%addr_s = getelementptr inbounds nuw double, ptr %s, i64 %index_i
%addr_s2 = getelementptr inbounds nuw double, ptr %s2, i64 %index_i
%invariant.gep.us = getelementptr inbounds nuw double, ptr %a, i64 %index_i
%invariant.gep32.us = getelementptr inbounds nuw double, ptr %b, i64 %index_i
br label %innerloop
innerloop:
%index_j = phi i64 [ 0, %outerloop_header ], [ %index_j.next, %innerloop ]
%reduction = phi double [ 0.000000e+00, %outerloop_header ], [ %add, %innerloop ]
%reduction2 = phi double [ 0.000000e+00, %outerloop_header ], [ %add2, %innerloop ]
%addr_a_j_i = getelementptr inbounds nuw double, ptr %invariant.gep.us, i64 %index_j
%0 = load double, ptr %addr_a_j_i, align 8
%addr_b_j_i = getelementptr inbounds nuw double, ptr %invariant.gep32.us, i64 %index_j
%1 = load double, ptr %addr_b_j_i, align 8
%mul = fmul fast double %1, %0
%add = fadd fast double %mul, %reduction
%add2 = fadd fast double %reduction2, %0
%index_j.next = add nuw nsw i64 %index_j, 1
%cond1 = icmp eq i64 %index_j.next, %n
br i1 %cond1, label %outerloop_latch, label %innerloop
outerloop_latch:
%lcssa = phi double [ %add, %innerloop ]
%lcssa2 = phi double [%add2, %innerloop]
store double %lcssa, ptr %addr_s, align 8
store double %lcssa2, ptr %addr_s2, align 8
%index_i.next = add nuw nsw i64 %index_i, 1
%cond2 = icmp eq i64 %index_i.next, %n
br i1 %cond2, label %exit, label %outerloop_header
exit:
ret void
}
; 3. The reduction is used more than once in the outer loop.
; for (int i = 0; i < n; i++) {
; r = 0;
; for (int j = 0; j < n; j++)
; r = r + a[j][i] * b[j][i];
; r += 1;
; s[i] = r;
; }
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedInnerReduction
; CHECK-NEXT: Function: reduction_03
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only supported when the reduction is used once in the outer loop.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIInner
; CHECK-NEXT: Function: reduction_03
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only inner loops with induction or reduction PHI nodes can be interchange currently.
; IR-LABEL: @reduction_03(
; IR-NOT: split
define void @reduction_03(ptr noalias readonly %a, ptr noalias readonly %b, ptr noalias writeonly %s, i64 %n) {
entry:
%cmp = icmp sgt i64 %n, 0
br i1 %cmp, label %outerloop_header, label %exit
outerloop_header:
%index_i = phi i64 [ 0, %entry ], [ %index_i.next, %outerloop_latch ]
%addr_s = getelementptr inbounds nuw double, ptr %s, i64 %index_i
%invariant.gep.us = getelementptr inbounds nuw double, ptr %a, i64 %index_i
%invariant.gep32.us = getelementptr inbounds nuw double, ptr %b, i64 %index_i
br label %innerloop
innerloop:
%index_j = phi i64 [ 0, %outerloop_header ], [ %index_j.next, %innerloop ]
%reduction = phi double [ 0.000000e+00, %outerloop_header ], [ %add, %innerloop ]
%addr_a_j_i = getelementptr inbounds nuw double, ptr %invariant.gep.us, i64 %index_j
%0 = load double, ptr %addr_a_j_i, align 8
%addr_b_j_i = getelementptr inbounds nuw double, ptr %invariant.gep32.us, i64 %index_j
%1 = load double, ptr %addr_b_j_i, align 8
%mul = fmul fast double %1, %0
%add = fadd fast double %mul, %reduction
%index_j.next = add nuw nsw i64 %index_j, 1
%cond1 = icmp eq i64 %index_j.next, %n
br i1 %cond1, label %outerloop_latch, label %innerloop
outerloop_latch:
%lcssa = phi double [ %add, %innerloop ]
store double %lcssa, ptr %addr_s, align 8
%add17.us = fadd fast double %lcssa, 1.000000e+00
store double %add17.us, ptr %addr_s, align 8
%index_i.next = add nuw nsw i64 %index_i, 1
%cond2 = icmp eq i64 %index_i.next, %n
br i1 %cond2, label %exit, label %outerloop_header
exit:
ret void
}
; 4. The reduction is not in the innermost loop.
; for (int i = 0; i < n; i++) {
; r = 0;
; for (int j = 0; j < n; j++) {
; r = r + a[j][i] * b[j][i]; // reduction
; for (int k = 0; k < n; k++)
; c[k] = 1;
; }
; s[i] = r;
; }
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIOuter
; CHECK-NEXT: Function: reduction_04
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only outer loops with induction or reduction PHI nodes can be interchanged currently.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedInnerReduction
; CHECK-NEXT: Function: reduction_04
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only supported when the loop is the innermost.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIInner
; CHECK-NEXT: Function: reduction_04
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only inner loops with induction or reduction PHI nodes can be interchange currently.
; IR-LABEL: @reduction_04(
; IR-NOT: split
define void @reduction_04(ptr noalias readonly %a, ptr noalias readonly %b, ptr noalias writeonly %c, ptr noalias writeonly %s, i64 %n) {
entry:
%cmp = icmp sgt i64 %n, 0
br i1 %cmp, label %i_loop_header, label %exit
i_loop_header:
%index_i = phi i64 [ 0, %entry ], [ %index_i.next, %i_loop_latch ]
%addr_s = getelementptr inbounds nuw double, ptr %s, i64 %index_i
%invariant.gep.us = getelementptr inbounds nuw double, ptr %a, i64 %index_i
%invariant.gep32.us = getelementptr inbounds nuw double, ptr %b, i64 %index_i
br label %j_loop
j_loop:
%index_j = phi i64 [ 0, %i_loop_header ], [ %index_j.next, %j_loop_latch ]
%reduction = phi double [ 0.000000e+00, %i_loop_header ], [ %add, %j_loop_latch ]
%addr_a_j_i = getelementptr inbounds nuw double, ptr %invariant.gep.us, i64 %index_j
%0 = load double, ptr %addr_a_j_i, align 8
%addr_b_j_i = getelementptr inbounds nuw double, ptr %invariant.gep32.us, i64 %index_j
%1 = load double, ptr %addr_b_j_i, align 8
%mul = fmul fast double %1, %0
%add = fadd fast double %mul, %reduction
br label %k_loop
k_loop:
%index_k = phi i64 [ %index_k.next, %k_loop ], [ 0, %j_loop ]
%arrayidx22.us.us = getelementptr inbounds nuw double, ptr %c, i64 %index_k
%index_k.next = add nuw nsw i64 %index_k, 1
%exitcond.not = icmp eq i64 %index_k.next, %n
br i1 %exitcond.not, label %j_loop_latch, label %k_loop
j_loop_latch:
%index_j.next = add nuw nsw i64 %index_j, 1
%cond1 = icmp eq i64 %index_j.next, %n
br i1 %cond1, label %i_loop_latch, label %j_loop
i_loop_latch:
%lcssa = phi double [ %add, %j_loop_latch ]
store double %lcssa, ptr %addr_s, align 8
%index_i.next = add nuw nsw i64 %index_i, 1
%cond2 = icmp eq i64 %index_i.next, %n
br i1 %cond2, label %exit, label %i_loop_header
exit:
ret void
}
; 5. MemRef doesn't dominate InnerLoop's HeaderBB.
; for (int i = 0; i < n; i++) {
; r = 0;
; for (int j = 0; j < n; j++)
; r = r + a[j][i] * b[j][i];
; s[i] = r;
; }
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedInnerReduction
; CHECK-NEXT: Function: reduction_05
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only supported when memory reference dominate the inner loop.
; CHECK: --- !Missed
; CHECK-NEXT: Pass: loop-interchange
; CHECK-NEXT: Name: UnsupportedPHIInner
; CHECK-NEXT: Function: reduction_05
; CHECK-NEXT: Args:
; CHECK-NEXT: - String: Only inner loops with induction or reduction PHI nodes can be interchange currently.
; IR-LABEL: @reduction_05(
; IR-NOT: split
define void @reduction_05(ptr noalias readonly %a, ptr noalias readonly %b, ptr noalias writeonly %s, i64 %n) {
entry:
%cmp = icmp sgt i64 %n, 0
br i1 %cmp, label %outerloop_header, label %exit
outerloop_header:
%index_i = phi i64 [ 0, %entry ], [ %index_i.next, %outerloop_latch ]
%invariant.gep.us = getelementptr inbounds nuw double, ptr %a, i64 %index_i
%invariant.gep32.us = getelementptr inbounds nuw double, ptr %b, i64 %index_i
br label %innerloop
innerloop:
%index_j = phi i64 [ 0, %outerloop_header ], [ %index_j.next, %innerloop ]
%reduction = phi double [ 0.000000e+00, %outerloop_header ], [ %add, %innerloop ]
%addr_a_j_i = getelementptr inbounds nuw double, ptr %invariant.gep.us, i64 %index_j
%0 = load double, ptr %addr_a_j_i, align 8
%addr_b_j_i = getelementptr inbounds nuw double, ptr %invariant.gep32.us, i64 %index_j
%1 = load double, ptr %addr_b_j_i, align 8
%mul = fmul fast double %1, %0
%add = fadd fast double %mul, %reduction
%index_j.next = add nuw nsw i64 %index_j, 1
%cond1 = icmp eq i64 %index_j.next, %n
br i1 %cond1, label %outerloop_latch, label %innerloop
outerloop_latch:
%lcssa = phi double [ %add, %innerloop ]
%addr_s = getelementptr inbounds nuw double, ptr %s, i64 %index_i
store double %lcssa, ptr %addr_s, align 8
%index_i.next = add nuw nsw i64 %index_i, 1
%cond2 = icmp eq i64 %index_i.next, %n
br i1 %cond2, label %exit, label %outerloop_header
exit:
ret void
}
|
