Jun Bum Lim | 90b6b50 | 2016-12-16 20:38:39 +0000 | [diff] [blame] | 1 | ; RUN: opt -codegenprepare < %s -mtriple=aarch64-none-linux-gnu -S | FileCheck %s |
| 2 | |
| 3 | target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128" |
| 4 | target triple = "aarch64--linux-gnu" |
| 5 | |
| 6 | ; Expect to skip merging two empty blocks (sw.bb and sw.bb2) into sw.epilog |
| 7 | ; as both of them are unlikely executed. |
| 8 | define i32 @f_switch(i32 %c) { |
| 9 | ; CHECK-LABEL: @f_switch |
| 10 | ; CHECK-LABEL: entry: |
| 11 | ; CHECK: i32 10, label %sw.bb |
| 12 | ; CHECK: i32 20, label %sw.bb2 |
| 13 | entry: |
| 14 | switch i32 %c, label %sw.default [ |
| 15 | i32 10, label %sw.bb |
| 16 | i32 20, label %sw.bb2 |
| 17 | i32 30, label %sw.bb3 |
| 18 | i32 40, label %sw.bb4 |
| 19 | ], !prof !0 |
| 20 | |
| 21 | sw.bb: ; preds = %entry |
| 22 | br label %sw.epilog |
| 23 | |
| 24 | sw.bb2: ; preds = %entry |
| 25 | br label %sw.epilog |
| 26 | |
| 27 | sw.bb3: ; preds = %entry |
| 28 | call void bitcast (void (...)* @callcase3 to void ()*)() |
| 29 | br label %sw.epilog |
| 30 | |
| 31 | sw.bb4: ; preds = %entry |
| 32 | call void bitcast (void (...)* @callcase4 to void ()*)() |
| 33 | br label %sw.epilog |
| 34 | |
| 35 | sw.default: ; preds = %entry |
| 36 | call void bitcast (void (...)* @calldefault to void ()*)() |
| 37 | br label %sw.epilog |
| 38 | |
| 39 | ; CHECK-LABEL: sw.epilog: |
| 40 | ; CHECK: %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F2, %sw.bb2 ], [ @F1, %sw.bb ] |
| 41 | sw.epilog: ; preds = %sw.default, %sw.bb3, %sw.bb2, %sw.bb |
| 42 | %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F2, %sw.bb2 ], [ @F1, %sw.bb ] |
| 43 | %callee.knr.cast = bitcast void (...)* %fp.0 to void ()* |
| 44 | call void %callee.knr.cast() |
| 45 | ret i32 0 |
| 46 | } |
| 47 | |
| 48 | ; Expect not to merge sw.bb2 because of the conflict in the incoming value from |
| 49 | ; sw.bb which is already merged. |
| 50 | define i32 @f_switch2(i32 %c) { |
| 51 | ; CHECK-LABEL: @f_switch2 |
| 52 | ; CHECK-LABEL: entry: |
| 53 | ; CHECK: i32 10, label %sw.epilog |
| 54 | ; CHECK: i32 20, label %sw.bb2 |
| 55 | entry: |
| 56 | switch i32 %c, label %sw.default [ |
| 57 | i32 10, label %sw.bb |
| 58 | i32 20, label %sw.bb2 |
| 59 | i32 30, label %sw.bb3 |
| 60 | i32 40, label %sw.bb4 |
| 61 | ], !prof !1 |
| 62 | |
| 63 | sw.bb: ; preds = %entry |
| 64 | br label %sw.epilog |
| 65 | |
| 66 | sw.bb2: ; preds = %entry |
| 67 | br label %sw.epilog |
| 68 | |
| 69 | sw.bb3: ; preds = %entry |
| 70 | call void bitcast (void (...)* @callcase3 to void ()*)() |
| 71 | br label %sw.epilog |
| 72 | |
| 73 | sw.bb4: ; preds = %entry |
| 74 | call void bitcast (void (...)* @callcase4 to void ()*)() |
| 75 | br label %sw.epilog |
| 76 | |
| 77 | sw.default: ; preds = %entry |
| 78 | call void bitcast (void (...)* @calldefault to void ()*)() |
| 79 | br label %sw.epilog |
| 80 | |
| 81 | ; CHECK-LABEL: sw.epilog: |
| 82 | ; CHECK: %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F2, %sw.bb2 ], [ @F1, %entry ] |
| 83 | sw.epilog: ; preds = %sw.default, %sw.bb3, %sw.bb2, %sw.bb |
| 84 | %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F2, %sw.bb2 ], [ @F1, %sw.bb ] |
| 85 | %callee.knr.cast = bitcast void (...)* %fp.0 to void ()* |
| 86 | call void %callee.knr.cast() |
| 87 | ret i32 0 |
| 88 | } |
| 89 | |
| 90 | ; Multiple empty blocks should be considered together if all incoming values |
| 91 | ; from them are same. We expect to merge both empty blocks (sw.bb and sw.bb2) |
| 92 | ; because the sum of frequencies are higer than the threshold. |
| 93 | define i32 @f_switch3(i32 %c) { |
| 94 | ; CHECK-LABEL: @f_switch3 |
| 95 | ; CHECK-LABEL: entry: |
| 96 | ; CHECK: i32 10, label %sw.epilog |
| 97 | ; CHECK: i32 20, label %sw.epilog |
| 98 | entry: |
| 99 | switch i32 %c, label %sw.default [ |
| 100 | i32 10, label %sw.bb |
| 101 | i32 20, label %sw.bb2 |
| 102 | i32 30, label %sw.bb3 |
| 103 | i32 40, label %sw.bb4 |
| 104 | ], !prof !2 |
| 105 | |
| 106 | sw.bb: ; preds = %entry |
| 107 | br label %sw.epilog |
| 108 | |
| 109 | sw.bb2: ; preds = %entry |
| 110 | br label %sw.epilog |
| 111 | |
| 112 | sw.bb3: ; preds = %entry |
| 113 | call void bitcast (void (...)* @callcase3 to void ()*)() |
| 114 | br label %sw.epilog |
| 115 | |
| 116 | sw.bb4: ; preds = %entry |
| 117 | call void bitcast (void (...)* @callcase4 to void ()*)() |
| 118 | br label %sw.epilog |
| 119 | |
| 120 | sw.default: ; preds = %entry |
| 121 | call void bitcast (void (...)* @calldefault to void ()*)() |
| 122 | br label %sw.epilog |
| 123 | |
| 124 | ; CHECK-LABEL: sw.epilog: |
| 125 | ; CHECK: %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F1, %entry ], [ @F1, %entry ] |
| 126 | sw.epilog: ; preds = %sw.default, %sw.bb3, %sw.bb2, %sw.bb |
| 127 | %fp.0 = phi void (...)* [ @FD, %sw.default ], [ @F4, %sw.bb4 ], [ @F3, %sw.bb3 ], [ @F1, %sw.bb2 ], [ @F1, %sw.bb ] |
| 128 | %callee.knr.cast = bitcast void (...)* %fp.0 to void ()* |
| 129 | call void %callee.knr.cast() |
| 130 | ret i32 0 |
| 131 | } |
| 132 | |
| 133 | declare void @F1(...) local_unnamed_addr |
| 134 | declare void @F2(...) local_unnamed_addr |
| 135 | declare void @F3(...) local_unnamed_addr |
| 136 | declare void @F4(...) local_unnamed_addr |
| 137 | declare void @FD(...) local_unnamed_addr |
| 138 | declare void @callcase3(...) local_unnamed_addr |
| 139 | declare void @callcase4(...) local_unnamed_addr |
| 140 | declare void @calldefault(...) local_unnamed_addr |
| 141 | |
| 142 | !0 = !{!"branch_weights", i32 5, i32 1, i32 1,i32 5, i32 5} |
| 143 | !1 = !{!"branch_weights", i32 1 , i32 5, i32 1,i32 1, i32 1} |
| 144 | !2 = !{!"branch_weights", i32 1 , i32 4, i32 1,i32 1, i32 1} |
| 145 | |
Balaram Makam | cddf3c5 | 2017-10-26 22:34:01 +0000 | [diff] [blame] | 146 | ; while.cond does not dominate return, expect to skip merging empty block |
| 147 | ; return.loopexit into return. |
| 148 | @b = external global i32, align 4 |
| 149 | @a = external global i32*, align 8 |
| 150 | |
| 151 | define void @f_switch4(i32 %i) local_unnamed_addr personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { |
| 152 | ; CHECK-LABEL: @f_switch4 |
| 153 | entry: |
| 154 | %0 = load i32, i32* @b, align 4 |
| 155 | %cond = icmp eq i32 %0, 6 |
| 156 | br i1 %cond, label %return, label %if.end |
| 157 | |
| 158 | if.end: ; preds = %entry |
| 159 | %add = add i32 %i, 2 |
| 160 | %1 = load i32*, i32** @a, align 8 |
| 161 | %magicptr = ptrtoint i32* %1 to i32 |
| 162 | br label %while.cond |
| 163 | |
| 164 | ; CHECK-LABEL: while.cond: |
| 165 | ; CHECK: i32 0, label %return.loopexit |
| 166 | ; CHECK: i32 47, label %return.loopexit |
| 167 | while.cond: ; preds = %while.cond.backedge, %if.end |
| 168 | switch i32 %magicptr, label %while.cond.if.end10_crit_edge [ |
| 169 | i32 0, label %return.loopexit |
| 170 | i32 47, label %return.loopexit |
| 171 | ] |
| 172 | |
| 173 | while.cond.if.end10_crit_edge: ; preds = %while.cond |
| 174 | br label %while.cond.backedge |
| 175 | |
| 176 | while.cond.backedge: ; preds = %while.cond.if.end10_crit_edge, %if.then9 |
| 177 | br label %while.cond |
| 178 | |
| 179 | return.loopexit: ; preds = %while.cond |
| 180 | br label %return |
| 181 | |
| 182 | ; CHECK_LABEL: return: |
| 183 | ; CHECK: %{{.*}} = phi i32 [ 0, %entry ], [ %add, %return.loopexit ] |
| 184 | return: ; preds = %return.loopexit, %entry |
| 185 | %retval.4 = phi i32 [ 0, %entry ], [ %add, %return.loopexit ] |
| 186 | ret void |
| 187 | } |
| 188 | declare i32 @__gxx_personality_v0(...) |
| 189 | |
| 190 | ; Expect to merge empty block while.cond2.loopexit into while.cond2 |
| 191 | define i32 @f_switch5(i32 %i) local_unnamed_addr { |
| 192 | ; CHECK-LABEL: @f_switch5 |
| 193 | entry: |
| 194 | %0 = load i32, i32* @b, align 4 |
| 195 | %cond = icmp eq i32 %0, 6 |
| 196 | br i1 %cond, label %while.cond.preheader, label %sw.epilog |
| 197 | |
| 198 | while.cond.preheader: ; preds = %entry |
| 199 | %1 = load i32*, i32** @a, align 8 |
| 200 | %magicptr = ptrtoint i32* %1 to i64 |
| 201 | %arrayidx = getelementptr inbounds i32, i32* %1, i64 1 |
| 202 | br label %while.cond |
| 203 | |
| 204 | ; CHECK-LABEL: while.cond: |
| 205 | ; CHECK: i64 32, label %while.cond2 |
| 206 | ; CHECK: i64 0, label %while.cond2 |
| 207 | while.cond: ; preds = %land.rhs, %while.cond.preheader |
| 208 | switch i64 %magicptr, label %land.rhs [ |
| 209 | i64 32, label %while.cond2.loopexit |
| 210 | i64 0, label %while.cond2.loopexit |
| 211 | ] |
| 212 | |
| 213 | land.rhs: ; preds = %while.cond |
| 214 | %2 = load i32, i32* %arrayidx, align 4 |
| 215 | %tobool1 = icmp eq i32 %2, 0 |
| 216 | br i1 %tobool1, label %while.cond2thread-pre-split.loopexit, label %while.cond |
| 217 | |
| 218 | while.cond2thread-pre-split.loopexit: ; preds = %land.rhs |
| 219 | br label %while.cond2thread-pre-split |
| 220 | |
| 221 | while.cond2thread-pre-split: ; preds = %while.body4, %while.cond2thread-pre-split.loopexit |
| 222 | %.pr = phi i32* [ %.pr.pre, %while.body4 ], [ %1, %while.cond2thread-pre-split.loopexit ] |
| 223 | br label %while.cond2 |
| 224 | |
| 225 | while.cond2.loopexit: ; preds = %while.cond, %while.cond |
| 226 | br label %while.cond2 |
| 227 | |
| 228 | ; CHECK-LABEL: while.cond2: |
| 229 | ; CHECK: %{{.*}} = phi i32* [ %.pr.pre, %while.body4 ], [ %1, %land.rhs ], [ %1, %while.cond ], [ %1, %while.cond ] |
| 230 | while.cond2: ; preds = %while.cond2.loopexit, %while.cond2thread-pre-split |
| 231 | %3 = phi i32* [ %.pr, %while.cond2thread-pre-split ], [ %1, %while.cond2.loopexit ] |
| 232 | %tobool3 = icmp eq i32* %3, null |
| 233 | br i1 %tobool3, label %sw.epilog, label %while.body4 |
| 234 | |
| 235 | while.body4: ; preds = %while.cond2 |
| 236 | tail call void bitcast (void (...)* @fn2 to void ()*)() |
| 237 | %.pr.pre = load i32*, i32** @a, align 8 |
| 238 | br label %while.cond2thread-pre-split |
| 239 | |
| 240 | sw.epilog: ; preds = %while.cond2, %entry |
| 241 | ret i32 undef |
| 242 | } |
| 243 | |
Jun Bum Lim | 90b6b50 | 2016-12-16 20:38:39 +0000 | [diff] [blame] | 244 | |
| 245 | ; This test that BFI/BPI is created without any assertion in isMergingEmptyBlockProfitable() |
| 246 | ; in the case where empty blocks are removed before creating BFI/BPI. |
Jun Bum Lim | 90b6b50 | 2016-12-16 20:38:39 +0000 | [diff] [blame] | 247 | define i32 @should_not_assert(i32 %i) local_unnamed_addr { |
| 248 | entry: |
| 249 | %0 = load i32, i32* @b, align 4 |
| 250 | %cond = icmp eq i32 %0, 6 |
| 251 | br i1 %cond, label %while.cond.preheader, label %sw.epilog |
| 252 | |
| 253 | while.cond.preheader: ; preds = %entry |
| 254 | %1 = load i32*, i32** @a, align 8 |
| 255 | %magicptr = ptrtoint i32* %1 to i64 |
| 256 | %arrayidx = getelementptr inbounds i32, i32* %1, i64 1 |
| 257 | br label %while.cond |
| 258 | |
| 259 | while.cond: ; preds = %while.cond.preheader, %land.rhs |
| 260 | switch i64 %magicptr, label %land.rhs [ |
| 261 | i64 32, label %while.cond2.loopexit |
| 262 | i64 0, label %while.cond2.loopexit |
| 263 | ] |
| 264 | |
| 265 | land.rhs: ; preds = %while.cond |
| 266 | %2 = load i32, i32* %arrayidx, align 4 |
| 267 | %tobool1 = icmp eq i32 %2, 0 |
| 268 | br i1 %tobool1, label %while.cond2thread-pre-split.loopexit, label %while.cond |
| 269 | |
| 270 | while.cond2thread-pre-split.loopexit: ; preds = %land.rhs |
| 271 | br label %while.cond2thread-pre-split |
| 272 | |
| 273 | while.cond2thread-pre-split: ; preds = %while.cond2thread-pre-split.loopexit, %while.body4 |
| 274 | %.pr = phi i32* [ %.pr.pre, %while.body4 ], [ %1, %while.cond2thread-pre-split.loopexit ] |
| 275 | br label %while.cond2 |
| 276 | |
| 277 | while.cond2.loopexit: ; preds = %while.cond, %while.cond |
| 278 | br label %while.cond2 |
| 279 | |
| 280 | while.cond2: ; preds = %while.cond2.loopexit, %while.cond2thread-pre-split |
| 281 | %3 = phi i32* [ %.pr, %while.cond2thread-pre-split ], [ %1, %while.cond2.loopexit ] |
| 282 | %tobool3 = icmp eq i32* %3, null |
| 283 | br i1 %tobool3, label %sw.epilog, label %while.body4 |
| 284 | |
| 285 | while.body4: ; preds = %while.cond2 |
| 286 | tail call void bitcast (void (...)* @fn2 to void ()*)() |
| 287 | %.pr.pre = load i32*, i32** @a, align 8 |
| 288 | br label %while.cond2thread-pre-split |
| 289 | |
| 290 | sw.epilog: ; preds = %while.cond2, %entry |
| 291 | ret i32 undef |
| 292 | } |
| 293 | |
| 294 | |
| 295 | declare void @fn2(...) local_unnamed_addr |
| 296 | |