1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -O3 -rotation-max-header-size=0 -S -enable-new-pm=0 < %s | FileCheck %s --check-prefix=HOIST 3; RUN: opt -passes='default<O3>' -rotation-max-header-size=0 -S < %s | FileCheck %s --check-prefix=HOIST 4 5; RUN: opt -O3 -rotation-max-header-size=1 -S -enable-new-pm=0 < %s | FileCheck %s --check-prefix=HOIST 6; RUN: opt -passes='default<O3>' -rotation-max-header-size=1 -S < %s | FileCheck %s --check-prefix=HOIST 7 8; RUN: opt -O3 -rotation-max-header-size=2 -S -enable-new-pm=0 < %s | FileCheck %s --check-prefix=ROTATED_LATER_OLDPM 9; RUN: opt -passes='default<O3>' -rotation-max-header-size=2 -S < %s | FileCheck %s --check-prefix=ROTATED_LATER_NEWPM 10 11; RUN: opt -O3 -rotation-max-header-size=3 -S -enable-new-pm=0 < %s | FileCheck %s --check-prefix=ROTATE_OLDPM 12; RUN: opt -passes='default<O3>' -rotation-max-header-size=3 -S < %s | FileCheck %s --check-prefix=ROTATE_NEWPM 13 14; This example is produced from a very basic C code: 15; 16; void f0(); 17; void f1(); 18; void f2(); 19; 20; void loop(int width) { 21; if(width < 1) 22; return; 23; for(int i = 0; i < width - 1; ++i) { 24; f0(); 25; f1(); 26; } 27; f0(); 28; f2(); 29; } 30 31; We have a choice here. We can either 32; * hoist the f0() call into loop header, 33; * which potentially makes loop rotation unprofitable since loop header might 34; have grown above certain threshold, and such unrotated loops will be 35; ignored by LoopVectorizer, preventing vectorization 36; * or loop rotation will succeed, resulting in some weird PHIs that will also 37; harm vectorization 38; * or not hoist f0() call before performing loop rotation, 39; at the cost of potential code bloat and/or potentially successfully rotating 40; the loops, vectorizing them at the cost of compile time. 41 42target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128" 43 44declare void @f0() 45declare void @f1() 46declare void @f2() 47 48declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture) 49declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture) 50 51define void @_Z4loopi(i32 %width) { 52; HOIST-LABEL: @_Z4loopi( 53; HOIST-NEXT: entry: 54; HOIST-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 55; HOIST-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 56; HOIST: for.cond.preheader: 57; HOIST-NEXT: [[TMP0:%.*]] = add nsw i32 [[WIDTH]], -1 58; HOIST-NEXT: br label [[FOR_COND:%.*]] 59; HOIST: for.cond: 60; HOIST-NEXT: [[I_0:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY:%.*]] ], [ 0, [[FOR_COND_PREHEADER]] ] 61; HOIST-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[I_0]], [[TMP0]] 62; HOIST-NEXT: tail call void @f0() 63; HOIST-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]] 64; HOIST: for.cond.cleanup: 65; HOIST-NEXT: tail call void @f2() 66; HOIST-NEXT: br label [[RETURN]] 67; HOIST: for.body: 68; HOIST-NEXT: tail call void @f1() 69; HOIST-NEXT: [[INC]] = add nuw i32 [[I_0]], 1 70; HOIST-NEXT: br label [[FOR_COND]] 71; HOIST: return: 72; HOIST-NEXT: ret void 73; 74; ROTATED_LATER_OLDPM-LABEL: @_Z4loopi( 75; ROTATED_LATER_OLDPM-NEXT: entry: 76; ROTATED_LATER_OLDPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 77; ROTATED_LATER_OLDPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 78; ROTATED_LATER_OLDPM: for.cond.preheader: 79; ROTATED_LATER_OLDPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 80; ROTATED_LATER_OLDPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY_PREHEADER:%.*]] 81; ROTATED_LATER_OLDPM: for.body.preheader: 82; ROTATED_LATER_OLDPM-NEXT: [[TMP0:%.*]] = add nsw i32 [[WIDTH]], -1 83; ROTATED_LATER_OLDPM-NEXT: br label [[FOR_BODY:%.*]] 84; ROTATED_LATER_OLDPM: for.cond.cleanup: 85; ROTATED_LATER_OLDPM-NEXT: tail call void @f0() 86; ROTATED_LATER_OLDPM-NEXT: tail call void @f2() 87; ROTATED_LATER_OLDPM-NEXT: br label [[RETURN]] 88; ROTATED_LATER_OLDPM: for.body: 89; ROTATED_LATER_OLDPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ] 90; ROTATED_LATER_OLDPM-NEXT: tail call void @f0() 91; ROTATED_LATER_OLDPM-NEXT: tail call void @f1() 92; ROTATED_LATER_OLDPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 93; ROTATED_LATER_OLDPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[TMP0]] 94; ROTATED_LATER_OLDPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 95; ROTATED_LATER_OLDPM: return: 96; ROTATED_LATER_OLDPM-NEXT: ret void 97; 98; ROTATED_LATER_NEWPM-LABEL: @_Z4loopi( 99; ROTATED_LATER_NEWPM-NEXT: entry: 100; ROTATED_LATER_NEWPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 101; ROTATED_LATER_NEWPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 102; ROTATED_LATER_NEWPM: for.cond.preheader: 103; ROTATED_LATER_NEWPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 104; ROTATED_LATER_NEWPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY_PREHEADER:%.*]] 105; ROTATED_LATER_NEWPM: for.body.preheader: 106; ROTATED_LATER_NEWPM-NEXT: [[TMP0:%.*]] = add nsw i32 [[WIDTH]], -1 107; ROTATED_LATER_NEWPM-NEXT: br label [[FOR_BODY:%.*]] 108; ROTATED_LATER_NEWPM: for.cond.cleanup: 109; ROTATED_LATER_NEWPM-NEXT: tail call void @f0() 110; ROTATED_LATER_NEWPM-NEXT: tail call void @f2() 111; ROTATED_LATER_NEWPM-NEXT: br label [[RETURN]] 112; ROTATED_LATER_NEWPM: for.body: 113; ROTATED_LATER_NEWPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ] 114; ROTATED_LATER_NEWPM-NEXT: tail call void @f0() 115; ROTATED_LATER_NEWPM-NEXT: tail call void @f1() 116; ROTATED_LATER_NEWPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 117; ROTATED_LATER_NEWPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[TMP0]] 118; ROTATED_LATER_NEWPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 119; ROTATED_LATER_NEWPM: return: 120; ROTATED_LATER_NEWPM-NEXT: ret void 121; 122; ROTATE_OLDPM-LABEL: @_Z4loopi( 123; ROTATE_OLDPM-NEXT: entry: 124; ROTATE_OLDPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 125; ROTATE_OLDPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 126; ROTATE_OLDPM: for.cond.preheader: 127; ROTATE_OLDPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 128; ROTATE_OLDPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY_PREHEADER:%.*]] 129; ROTATE_OLDPM: for.body.preheader: 130; ROTATE_OLDPM-NEXT: [[TMP0:%.*]] = add nsw i32 [[WIDTH]], -1 131; ROTATE_OLDPM-NEXT: br label [[FOR_BODY:%.*]] 132; ROTATE_OLDPM: for.cond.cleanup: 133; ROTATE_OLDPM-NEXT: tail call void @f0() 134; ROTATE_OLDPM-NEXT: tail call void @f2() 135; ROTATE_OLDPM-NEXT: br label [[RETURN]] 136; ROTATE_OLDPM: for.body: 137; ROTATE_OLDPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ] 138; ROTATE_OLDPM-NEXT: tail call void @f0() 139; ROTATE_OLDPM-NEXT: tail call void @f1() 140; ROTATE_OLDPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 141; ROTATE_OLDPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[TMP0]] 142; ROTATE_OLDPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 143; ROTATE_OLDPM: return: 144; ROTATE_OLDPM-NEXT: ret void 145; 146; ROTATE_NEWPM-LABEL: @_Z4loopi( 147; ROTATE_NEWPM-NEXT: entry: 148; ROTATE_NEWPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 149; ROTATE_NEWPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 150; ROTATE_NEWPM: for.cond.preheader: 151; ROTATE_NEWPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 152; ROTATE_NEWPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY_PREHEADER:%.*]] 153; ROTATE_NEWPM: for.body.preheader: 154; ROTATE_NEWPM-NEXT: [[TMP0:%.*]] = add nsw i32 [[WIDTH]], -1 155; ROTATE_NEWPM-NEXT: br label [[FOR_BODY:%.*]] 156; ROTATE_NEWPM: for.cond.cleanup: 157; ROTATE_NEWPM-NEXT: tail call void @f0() 158; ROTATE_NEWPM-NEXT: tail call void @f2() 159; ROTATE_NEWPM-NEXT: br label [[RETURN]] 160; ROTATE_NEWPM: for.body: 161; ROTATE_NEWPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_BODY_PREHEADER]] ] 162; ROTATE_NEWPM-NEXT: tail call void @f0() 163; ROTATE_NEWPM-NEXT: tail call void @f1() 164; ROTATE_NEWPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 165; ROTATE_NEWPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[TMP0]] 166; ROTATE_NEWPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 167; ROTATE_NEWPM: return: 168; ROTATE_NEWPM-NEXT: ret void 169; 170entry: 171 %width.addr = alloca i32, align 4 172 %i = alloca i32, align 4 173 store i32 %width, i32* %width.addr, align 4 174 %i1 = load i32, i32* %width.addr, align 4 175 %cmp = icmp slt i32 %i1, 1 176 br i1 %cmp, label %if.then, label %if.end 177 178if.then: 179 br label %return 180 181if.end: 182 %i2 = bitcast i32* %i to i8* 183 call void @llvm.lifetime.start.p0i8(i64 4, i8* %i2) 184 store i32 0, i32* %i, align 4 185 br label %for.cond 186 187for.cond: 188 %i3 = load i32, i32* %i, align 4 189 %i4 = load i32, i32* %width.addr, align 4 190 %sub = sub nsw i32 %i4, 1 191 %cmp1 = icmp slt i32 %i3, %sub 192 br i1 %cmp1, label %for.body, label %for.cond.cleanup 193 194for.cond.cleanup: 195 %i5 = bitcast i32* %i to i8* 196 call void @llvm.lifetime.end.p0i8(i64 4, i8* %i5) 197 br label %for.end 198 199for.body: 200 call void @f0() 201 call void @f1() 202 br label %for.inc 203 204for.inc: 205 %i6 = load i32, i32* %i, align 4 206 %inc = add nsw i32 %i6, 1 207 store i32 %inc, i32* %i, align 4 208 br label %for.cond 209 210for.end: 211 call void @f0() 212 call void @f2() 213 br label %return 214 215return: 216 ret void 217} 218