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: [[SUB:%.*]] = 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]], [[SUB]] 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: [[SUB:%.*]] = add nsw i32 [[WIDTH]], -1 80; ROTATED_LATER_OLDPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 81; ROTATED_LATER_OLDPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY:%.*]] 82; ROTATED_LATER_OLDPM: for.cond.cleanup: 83; ROTATED_LATER_OLDPM-NEXT: tail call void @f0() 84; ROTATED_LATER_OLDPM-NEXT: tail call void @f2() 85; ROTATED_LATER_OLDPM-NEXT: br label [[RETURN]] 86; ROTATED_LATER_OLDPM: for.body: 87; ROTATED_LATER_OLDPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_COND_PREHEADER]] ] 88; ROTATED_LATER_OLDPM-NEXT: tail call void @f0() 89; ROTATED_LATER_OLDPM-NEXT: tail call void @f1() 90; ROTATED_LATER_OLDPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 91; ROTATED_LATER_OLDPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[SUB]] 92; ROTATED_LATER_OLDPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 93; ROTATED_LATER_OLDPM: return: 94; ROTATED_LATER_OLDPM-NEXT: ret void 95; 96; ROTATED_LATER_NEWPM-LABEL: @_Z4loopi( 97; ROTATED_LATER_NEWPM-NEXT: entry: 98; ROTATED_LATER_NEWPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 99; ROTATED_LATER_NEWPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 100; ROTATED_LATER_NEWPM: for.cond.preheader: 101; ROTATED_LATER_NEWPM-NEXT: [[SUB:%.*]] = add nsw i32 [[WIDTH]], -1 102; ROTATED_LATER_NEWPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 103; ROTATED_LATER_NEWPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY:%.*]] 104; ROTATED_LATER_NEWPM: for.cond.cleanup: 105; ROTATED_LATER_NEWPM-NEXT: tail call void @f0() 106; ROTATED_LATER_NEWPM-NEXT: tail call void @f2() 107; ROTATED_LATER_NEWPM-NEXT: br label [[RETURN]] 108; ROTATED_LATER_NEWPM: for.body: 109; ROTATED_LATER_NEWPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_COND_PREHEADER]] ] 110; ROTATED_LATER_NEWPM-NEXT: tail call void @f0() 111; ROTATED_LATER_NEWPM-NEXT: tail call void @f1() 112; ROTATED_LATER_NEWPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 113; ROTATED_LATER_NEWPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[SUB]] 114; ROTATED_LATER_NEWPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 115; ROTATED_LATER_NEWPM: return: 116; ROTATED_LATER_NEWPM-NEXT: ret void 117; 118; ROTATE_OLDPM-LABEL: @_Z4loopi( 119; ROTATE_OLDPM-NEXT: entry: 120; ROTATE_OLDPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 121; ROTATE_OLDPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 122; ROTATE_OLDPM: for.cond.preheader: 123; ROTATE_OLDPM-NEXT: [[SUB:%.*]] = add nsw i32 [[WIDTH]], -1 124; ROTATE_OLDPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 125; ROTATE_OLDPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY:%.*]] 126; ROTATE_OLDPM: for.cond.cleanup: 127; ROTATE_OLDPM-NEXT: tail call void @f0() 128; ROTATE_OLDPM-NEXT: tail call void @f2() 129; ROTATE_OLDPM-NEXT: br label [[RETURN]] 130; ROTATE_OLDPM: for.body: 131; ROTATE_OLDPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_COND_PREHEADER]] ] 132; ROTATE_OLDPM-NEXT: tail call void @f0() 133; ROTATE_OLDPM-NEXT: tail call void @f1() 134; ROTATE_OLDPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 135; ROTATE_OLDPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[SUB]] 136; ROTATE_OLDPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 137; ROTATE_OLDPM: return: 138; ROTATE_OLDPM-NEXT: ret void 139; 140; ROTATE_NEWPM-LABEL: @_Z4loopi( 141; ROTATE_NEWPM-NEXT: entry: 142; ROTATE_NEWPM-NEXT: [[CMP:%.*]] = icmp slt i32 [[WIDTH:%.*]], 1 143; ROTATE_NEWPM-NEXT: br i1 [[CMP]], label [[RETURN:%.*]], label [[FOR_COND_PREHEADER:%.*]] 144; ROTATE_NEWPM: for.cond.preheader: 145; ROTATE_NEWPM-NEXT: [[SUB:%.*]] = add nsw i32 [[WIDTH]], -1 146; ROTATE_NEWPM-NEXT: [[CMP13_NOT:%.*]] = icmp eq i32 [[WIDTH]], 1 147; ROTATE_NEWPM-NEXT: br i1 [[CMP13_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY:%.*]] 148; ROTATE_NEWPM: for.cond.cleanup: 149; ROTATE_NEWPM-NEXT: tail call void @f0() 150; ROTATE_NEWPM-NEXT: tail call void @f2() 151; ROTATE_NEWPM-NEXT: br label [[RETURN]] 152; ROTATE_NEWPM: for.body: 153; ROTATE_NEWPM-NEXT: [[I_04:%.*]] = phi i32 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[FOR_COND_PREHEADER]] ] 154; ROTATE_NEWPM-NEXT: tail call void @f0() 155; ROTATE_NEWPM-NEXT: tail call void @f1() 156; ROTATE_NEWPM-NEXT: [[INC]] = add nuw nsw i32 [[I_04]], 1 157; ROTATE_NEWPM-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[SUB]] 158; ROTATE_NEWPM-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]] 159; ROTATE_NEWPM: return: 160; ROTATE_NEWPM-NEXT: ret void 161; 162entry: 163 %width.addr = alloca i32, align 4 164 %i = alloca i32, align 4 165 store i32 %width, i32* %width.addr, align 4 166 %i1 = load i32, i32* %width.addr, align 4 167 %cmp = icmp slt i32 %i1, 1 168 br i1 %cmp, label %if.then, label %if.end 169 170if.then: 171 br label %return 172 173if.end: 174 %i2 = bitcast i32* %i to i8* 175 call void @llvm.lifetime.start.p0i8(i64 4, i8* %i2) 176 store i32 0, i32* %i, align 4 177 br label %for.cond 178 179for.cond: 180 %i3 = load i32, i32* %i, align 4 181 %i4 = load i32, i32* %width.addr, align 4 182 %sub = sub nsw i32 %i4, 1 183 %cmp1 = icmp slt i32 %i3, %sub 184 br i1 %cmp1, label %for.body, label %for.cond.cleanup 185 186for.cond.cleanup: 187 %i5 = bitcast i32* %i to i8* 188 call void @llvm.lifetime.end.p0i8(i64 4, i8* %i5) 189 br label %for.end 190 191for.body: 192 call void @f0() 193 call void @f1() 194 br label %for.inc 195 196for.inc: 197 %i6 = load i32, i32* %i, align 4 198 %inc = add nsw i32 %i6, 1 199 store i32 %inc, i32* %i, align 4 200 br label %for.cond 201 202for.end: 203 call void @f0() 204 call void @f2() 205 br label %return 206 207return: 208 ret void 209} 210