llvm/test/CodeGen/X86/code_placement_ext_tsp_large.ll

   1 ; REQUIRES: asserts
   2 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -enable-ext-tsp-block-placement=1 -ext-tsp-chain-split-threshold=128 -debug-only=block-placement < %s 2>&1 | FileCheck %s
   3 ; RUN: llc -mcpu=corei7 -mtriple=x86_64-linux -enable-ext-tsp-block-placement=1 -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 -enable-ext-tsp-block-placement=0 -debug-only=block-placement < %s 2>&1 | FileCheck %s -check-prefix=CHECK3
   5
   6 @yydebug = dso_local global i32 0, align 4
   7
   8 define void @func_large() !prof !0 {
   9 ; A large CFG instance where chain splitting helps to
  10 ; compute a better basic block ordering. The test verifies that with chain
  11 ; splitting, the resulting layout is improved (e.g., the score is increased).
  12 ;
  13 ;                                     +----------------+
  14 ;                                     | b0 [76 bytes]  | -------------------+
  15 ;                                     +----------------+                    |
  16 ;                                       |                                   |
  17 ;                                       | 3,065,981,778                     |
  18 ;                                       v                                   |
  19 ; +----------------+  766,495,444     +----------------+                    |
  20 ; | b8 [244 bytes] | <--------------- |  b2 [4 bytes]  |                    |
  21 ; +----------------+                  +----------------+                    |
  22 ;   |        ^                          |                                   |
  23 ;   |        |                          | 2,299,486,333                     |
  24 ;   |        | 766,495,444              v                                   |
  25 ;   |        |                        +----------------+                    |
  26 ;   |        +----------------------- | b3 [12 bytes]  |                    |
  27 ;   |                                 +----------------+                    |
  28 ;   |                                   |                                   |
  29 ;   |                                   | 1,532,990,888                     |
  30 ;   |                                   v                                   |
  31 ;   |                                 +----------------+                    | 574,869,946
  32 ;   |                 +-------------- | b4 [12 bytes]  |                    |
  33 ;   |                 |               +----------------+                    |
  34 ;   |                 |                 |                                   |
  35 ;   |                 |                 | 574,871,583                       |
  36 ;   |                 |                 v                                   |
  37 ;   |                 |               +----------------+                    |
  38 ;   |                 |               | b5 [116 bytes] | -+                 |
  39 ;   |                 |               +----------------+  |                 |
  40 ;   |                 |                 |                 |                 |
  41 ;   |                 |                 | 1,636           |                 |
  42 ;   |                 |                 v                 |                 |
  43 ;   |                 |               +----------------+  |                 |
  44 ;   |                 |       +------ | b6 [32 bytes]  |  |                 |
  45 ;   |                 |       |       +----------------+  |                 |
  46 ;   |                 |       |         |                 |                 |
  47 ;   |                 |       |         | 7               | 3,065,981,778   |
  48 ;   |                 |       |         v                 |                 |
  49 ;   |                 |       |       +----------------+  |                 |
  50 ;   |                 |       | 1,628 | b9 [16 bytes]  |  |                 |
  51 ;   |                 |       |       +----------------+  |                 |
  52 ;   |                 |       |         |                 |                 |
  53 ;   |                 |       |         | 7               |                 |
  54 ;   |                 |       |         v                 |                 |
  55 ;   |                 |       |       +----------------+  |                 |
  56 ;   |                 |       +-----> | b7 [12 bytes]  |  |                 |
  57 ;   |                 |               +----------------+  |                 |
  58 ;   |                 |                 |                 |                 |
  59 ;   |                 | 958,119,305     | 1,636           |                 |
  60 ;   |                 |                 v                 v                 v
  61 ;   |                 |               +------------------------------------------+
  62 ;   |                 +-------------> |                                          |
  63 ;   |       1,532,990,889             |                  b1 [36 bytes]           |
  64 ;   +-------------------------------> |                                          |
  65 ;                                     +------------------------------------------+
  66 ;
  67 ; An expected output with a large chain-split-threshold -- the layout score is
  68 ; increased by ~17%
  69 ;
  70 ; CHECK-LABEL: Applying ext-tsp layout
  71 ; CHECK:   original  layout score: 23587612604815436.00
  72 ; CHECK:   optimized layout score: 27891096739311172.00
  73 ; CHECK: b0
  74 ; CHECK: b2
  75 ; CHECK: b3
  76 ; CHECK: b4
  77 ; CHECK: b5
  78 ; CHECK: b8
  79 ; CHECK: b1
  80 ; CHECK: b6
  81 ; CHECK: b7
  82 ; CHECK: b9
  83 ;
  84 ; An expected output with chain-split-threshold=1 (disabling split point enumeration)
  85 ;
  86 ; CHECK2-LABEL: Applying ext-tsp layout
  87 ; CHECK2:   original  layout score: 23587612604815436.00
  88 ; CHECK2:   optimized layout score: 27891096739311172.00
  89 ; CHECK2: b0
  90 ; CHECK2: b2
  91 ; CHECK2: b3
  92 ; CHECK2: b4
  93 ; CHECK2: b5
  94 ; CHECK2: b8
  95 ; CHECK2: b1
  96 ; CHECK2: b6
  97 ; CHECK2: b7
  98 ; CHECK2: b9
  99 ;
 100 ; An expected output with ext-tsp disabled -- the layout is not modified:
 101 ;
 102 ; CHECK3-LABEL: func_large:
 103 ; CHECK3: b0
 104 ; CHECK3: b1
 105 ; CHECK3: b2
 106 ; CHECK3: b3
 107 ; CHECK3: b4
 108 ; CHECK3: b5
 109 ; CHECK3: b6
 110 ; CHECK3: b7
 111 ; CHECK3: b8
 112 ; CHECK3: b9
 113
 114 b0:
 115   %0 = load i32, ptr @yydebug, align 4
 116   %cmp = icmp ne i32 %0, 0
 117   call void @a()
 118   call void @a()
 119   call void @a()
 120   call void @a()
 121   call void @a()
 122   call void @a()
 123   call void @a()
 124   call void @a()
 125   call void @a()
 126   call void @a()
 127   call void @a()
 128   call void @a()
 129   call void @a()
 130   call void @a()
 131   call void @a()
 132   call void @a()
 133   call void @a()
 134   call void @a()
 135   br i1 %cmp, label %b1, label %b2, !prof !1
 136 b1:
 137   call void @b()
 138   call void @b()
 139   call void @b()
 140   call void @b()
 141   call void @b()
 142   call void @b()
 143   call void @b()
 144   call void @b()
 145   ret void
 146 b2:
 147   call void @c()
 148   call void @c()
 149   call void @c()
 150   call void @c()
 151   call void @c()
 152   call void @c()
 153   call void @c()
 154   call void @c()
 155   call void @c()
 156   call void @c()
 157   br i1 %cmp, label %b3, label %b8, !prof !2
 158 b3:
 159   call void @d()
 160   call void @d()
 161   br i1 %cmp, label %b4, label %b8, !prof !3
 162 b4:
 163   call void @e()
 164   call void @e()
 165   br i1 %cmp, label %b5, label %b1, !prof !4
 166 b5:
 167   call void @f()
 168   call void @f()
 169   call void @f()
 170   call void @f()
 171   call void @f()
 172   call void @f()
 173   call void @f()
 174   call void @f()
 175   call void @f()
 176   call void @f()
 177   call void @f()
 178   call void @f()
 179   call void @f()
 180   call void @f()
 181   call void @f()
 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   br i1 %cmp, label %b1, label %b6, !prof !5
 196 b6:
 197   call void @g()
 198   call void @g()
 199   call void @g()
 200   call void @g()
 201   call void @g()
 202   call void @g()
 203   call void @g()
 204   br i1 %cmp, label %b7, label %b9, !prof !6
 205 b7:
 206   call void @h()
 207   call void @h()
 208   br label %b1
 209 b8:
 210   call void @i()
 211   call void @i()
 212   call void @i()
 213   call void @i()
 214   call void @i()
 215   call void @i()
 216   call void @i()
 217   call void @i()
 218   call void @i()
 219   call void @i()
 220   call void @i()
 221   call void @i()
 222   call void @i()
 223   call void @i()
 224   call void @i()
 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   br label %b1
 271 b9:
 272   call void @j()
 273   call void @j()
 274   call void @j()
 275   br label %b7
 276 }
 277
 278
 279 declare void @a()
 280 declare void @b()
 281 declare void @c()
 282 declare void @d()
 283 declare void @e()
 284 declare void @f()
 285 declare void @g()
 286 declare void @h()
 287 declare void @i()
 288 declare void @j()
 289
 290 !0 = !{!"function_entry_count", i64 6131963556}
 291 !1 = !{!"branch_weights", i32 3065981778, i32 3065981778}
 292 !2 = !{!"branch_weights", i32 2299486333, i32 766495444}
 293 !3 = !{!"branch_weights", i32 1532990888, i32 766495444}
 294 !4 = !{!"branch_weights", i32 574871583, i32 958119305}
 295 !5 = !{!"branch_weights", i32 574869946, i32 1636}
 296 !6 = !{!"branch_weights", i32 1628, i32 7}