| blob | 24c52f1e88656ed7496a5310b83d41af4aa26e3d |
1 ; REQUIRES: asserts
2 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -verify-machineinstrs -enable-ext-tsp-block-placement -ext-tsp-chain-split-threshold=128 -debug-only=block-placement < %s 2>&1 | FileCheck %s
3 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -verify-machineinstrs -enable-ext-tsp-block-placement -ext-tsp-chain-split-threshold=1 -debug-only=block-placement < %s 2>&1 | FileCheck %s -check-prefix=CHECK2
4 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -verify-machineinstrs -debug-only=block-placement < %s 2>&1 | FileCheck %s -check-prefix=CHECK3
5 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -verify-machineinstrs -enable-ext-tsp-block-placement -ext-tsp-block-placement-max-blocks=8 -debug-only=block-placement < %s 2>&1 | FileCheck %s -check-prefix=CHECK4
7 @yydebug = dso_local global i32 0, align 4
9 define void @func_large() !prof !0 {
10 ; A large CFG instance where chain splitting helps to
11 ; compute a better basic block ordering. The test verifies that with chain
12 ; splitting, the resulting layout is improved (e.g., the score is increased).
13 ;
14 ; +----------------+
15 ; | b0 [76 bytes] | -------------------+
16 ; +----------------+ |
17 ; | |
18 ; | 3,065,981,778 |
19 ; v |
20 ; +----------------+ 766,495,444 +----------------+ |
21 ; | b8 [244 bytes] | <--------------- | b2 [4 bytes] | |
22 ; +----------------+ +----------------+ |
23 ; | ^ | |
24 ; | | | 2,299,486,333 |
25 ; | | 766,495,444 v |
26 ; | | +----------------+ |
27 ; | +----------------------- | b3 [12 bytes] | |
28 ; | +----------------+ |
29 ; | | |
30 ; | | 1,532,990,888 |
31 ; | v |
32 ; | +----------------+ | 574,869,946
33 ; | +-------------- | b4 [12 bytes] | |
34 ; | | +----------------+ |
35 ; | | | |
36 ; | | | 574,871,583 |
37 ; | | v |
38 ; | | +----------------+ |
39 ; | | | b5 [116 bytes] | -+ |
40 ; | | +----------------+ | |
41 ; | | | | |
42 ; | | | 1,636 | |
43 ; | | v | |
44 ; | | +----------------+ | |
45 ; | | +------ | b6 [32 bytes] | | |
46 ; | | | +----------------+ | |
47 ; | | | | | |
48 ; | | | | 7 | 3,065,981,778 |
49 ; | | | v | |
50 ; | | | +----------------+ | |
51 ; | | | 1,628 | b9 [16 bytes] | | |
52 ; | | | +----------------+ | |
53 ; | | | | | |
54 ; | | | | 7 | |
55 ; | | | v | |
56 ; | | | +----------------+ | |
57 ; | | +-----> | b7 [12 bytes] | | |
58 ; | | +----------------+ | |
59 ; | | | | |
60 ; | | 958,119,305 | 1,636 | |
61 ; | | v v v
62 ; | | +------------------------------------------+
63 ; | +-------------> | |
64 ; | 1,532,990,889 | b1 [36 bytes] |
65 ; +-------------------------------> | |
66 ; +------------------------------------------+
67 ;
68 ; An expected output with a large chain-split-threshold -- the layout score is
69 ; increased by ~17%
70 ;
71 ; CHECK-LABEL: Applying ext-tsp layout
72 ; CHECK: original layout score: 23587612604815436.00
73 ; CHECK: optimized layout score: 27891096739311172.00
74 ; CHECK: b0
75 ; CHECK: b2
76 ; CHECK: b3
77 ; CHECK: b4
78 ; CHECK: b5
79 ; CHECK: b8
80 ; CHECK: b1
81 ; CHECK: b6
82 ; CHECK: b7
83 ; CHECK: b9
84 ;
85 ; An expected output with chain-split-threshold=1 (disabling split point enumeration)
86 ;
87 ; CHECK2-LABEL: Applying ext-tsp layout
88 ; CHECK2: original layout score: 23587612604815436.00
89 ; CHECK2: optimized layout score: 27891096739311172.00
90 ; CHECK2: b0
91 ; CHECK2: b2
92 ; CHECK2: b3
93 ; CHECK2: b4
94 ; CHECK2: b5
95 ; CHECK2: b8
96 ; CHECK2: b1
97 ; CHECK2: b6
98 ; CHECK2: b7
99 ; CHECK2: b9
100 ;
101 ; An expected output with ext-tsp disabled -- the layout is not modified:
102 ;
103 ; CHECK3-LABEL: func_large:
104 ; CHECK3: b0
105 ; CHECK3: b1
106 ; CHECK3: b2
107 ; CHECK3: b3
108 ; CHECK3: b4
109 ; CHECK3: b5
110 ; CHECK3: b6
111 ; CHECK3: b7
112 ; CHECK3: b8
113 ; CHECK3: b9
114 ;
115 ; An expected output with function size larger than the threshold -- the layout is not modified:
116 ;
117 ; CHECK4-LABEL: func_large:
118 ; CHECK4: b0
119 ; CHECK4: b1
120 ; CHECK4: b2
121 ; CHECK4: b3
122 ; CHECK4: b4
123 ; CHECK4: b5
124 ; CHECK4: b6
125 ; CHECK4: b7
126 ; CHECK4: b8
127 ; CHECK4: b9
129 b0:
130 %0 = load i32, ptr @yydebug, align 4
131 %cmp = icmp ne i32 %0, 0
132 call void @a()
133 call void @a()
134 call void @a()
135 call void @a()
136 call void @a()
137 call void @a()
138 call void @a()
139 call void @a()
140 call void @a()
141 call void @a()
142 call void @a()
143 call void @a()
144 call void @a()
145 call void @a()
146 call void @a()
147 call void @a()
148 call void @a()
149 call void @a()
150 br i1 %cmp, label %b1, label %b2, !prof !1
151 b1:
152 call void @b()
153 call void @b()
154 call void @b()
155 call void @b()
156 call void @b()
157 call void @b()
158 call void @b()
159 call void @b()
160 ret void
161 b2:
162 call void @c()
163 call void @c()
164 call void @c()
165 call void @c()
166 call void @c()
167 call void @c()
168 call void @c()
169 call void @c()
170 call void @c()
171 call void @c()
172 br i1 %cmp, label %b3, label %b8, !prof !2
173 b3:
174 call void @d()
175 call void @d()
176 br i1 %cmp, label %b4, label %b8, !prof !3
177 b4:
178 call void @e()
179 call void @e()
180 br i1 %cmp, label %b5, label %b1, !prof !4
181 b5:
182 call void @f()
183 call void @f()
184 call void @f()
185 call void @f()
186 call void @f()
187 call void @f()
188 call void @f()
189 call void @f()
190 call void @f()
191 call void @f()
192 call void @f()
193 call void @f()
194 call void @f()
195 call void @f()
196 call void @f()
197 call void @f()
198 call void @f()
199 call void @f()
200 call void @f()
201 call void @f()
202 call void @f()
203 call void @f()
204 call void @f()
205 call void @f()
206 call void @f()
207 call void @f()
208 call void @f()
209 call void @f()
210 br i1 %cmp, label %b1, label %b6, !prof !5
211 b6:
212 call void @g()
213 call void @g()
214 call void @g()
215 call void @g()
216 call void @g()
217 call void @g()
218 call void @g()
219 br i1 %cmp, label %b7, label %b9, !prof !6
220 b7:
221 call void @h()
222 call void @h()
223 br label %b1
224 b8:
225 call void @i()
226 call void @i()
227 call void @i()
228 call void @i()
229 call void @i()
230 call void @i()
231 call void @i()
232 call void @i()
233 call void @i()
234 call void @i()
235 call void @i()
236 call void @i()
237 call void @i()
238 call void @i()
239 call void @i()
240 call void @i()
241 call void @i()
242 call void @i()
243 call void @i()
244 call void @i()
245 call void @i()
246 call void @i()
247 call void @i()
248 call void @i()
249 call void @i()
250 call void @i()
251 call void @i()
252 call void @i()
253 call void @i()
254 call void @i()
255 call void @i()
256 call void @i()
257 call void @i()
258 call void @i()
259 call void @i()
260 call void @i()
261 call void @i()
262 call void @i()
263 call void @i()
264 call void @i()
265 call void @i()
266 call void @i()
267 call void @i()
268 call void @i()
269 call void @i()
270 call void @i()
271 call void @i()
272 call void @i()
273 call void @i()
274 call void @i()
275 call void @i()
276 call void @i()
277 call void @i()
278 call void @i()
279 call void @i()
280 call void @i()
281 call void @i()
282 call void @i()
283 call void @i()
284 call void @i()
285 br label %b1
286 b9:
287 call void @j()
288 call void @j()
289 call void @j()
290 br label %b7
291 }
294 declare void @a()
295 declare void @b()
296 declare void @c()
297 declare void @d()
298 declare void @e()
299 declare void @f()
300 declare void @g()
301 declare void @h()
302 declare void @i()
303 declare void @j()
305 !0 = !{!"function_entry_count", i64 6131963556}
306 !1 = !{!"branch_weights", i32 3065981778, i32 3065981778}
307 !2 = !{!"branch_weights", i32 2299486333, i32 766495444}
308 !3 = !{!"branch_weights", i32 1532990888, i32 766495444}
309 !4 = !{!"branch_weights", i32 574871583, i32 958119305}
310 !5 = !{!"branch_weights", i32 574869946, i32 1636}
311 !6 = !{!"branch_weights", i32 1628, i32 7}