157 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=loop-vectorize < %s -S -o - | FileCheck %s3 4target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"5target triple = "thumbv8.1m.main-none-eabi"6 7; Test cases to make sure LV & loop versioning can handle loops with8; multiple exiting branches.9 10; Multiple branches exiting the loop to a unique exit block.11define void @multiple_exits_unique_exit_block(ptr %A, ptr %B, i32 %N) #0 {12; CHECK-LABEL: @multiple_exits_unique_exit_block(13; CHECK-NEXT: entry:14; CHECK-NEXT: [[A2:%.*]] = ptrtoint ptr [[A:%.*]] to i3215; CHECK-NEXT: [[B1:%.*]] = ptrtoint ptr [[B:%.*]] to i3216; CHECK-NEXT: [[UMIN:%.*]] = call i32 @llvm.umin.i32(i32 [[N:%.*]], i32 999)17; CHECK-NEXT: [[TMP0:%.*]] = add nuw nsw i32 [[UMIN]], 118; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 [[TMP0]], 419; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]20; CHECK: vector.memcheck:21; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[B1]], [[A2]]22; CHECK-NEXT: [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP1]], 1623; CHECK-NEXT: br i1 [[DIFF_CHECK]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]24; CHECK: vector.ph:25; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 426; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i32 [[N_MOD_VF]], 027; CHECK-NEXT: [[TMP3:%.*]] = select i1 [[TMP2]], i32 4, i32 [[N_MOD_VF]]28; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[TMP3]]29; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]30; CHECK: vector.body:31; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]32; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[INDEX]]33; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP5]], align 434; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[B]], i32 [[INDEX]]35; CHECK-NEXT: store <4 x i32> [[WIDE_LOAD]], ptr [[TMP7]], align 436; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 437; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]38; CHECK-NEXT: br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]39; CHECK: middle.block:40; CHECK-NEXT: br label [[SCALAR_PH]]41; CHECK: scalar.ph:42; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_MEMCHECK]] ]43; CHECK-NEXT: br label [[LOOP_HEADER:%.*]]44; CHECK: loop.header:45; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY:%.*]] ]46; CHECK-NEXT: [[COND_0:%.*]] = icmp eq i32 [[IV]], [[N]]47; CHECK-NEXT: br i1 [[COND_0]], label [[EXIT:%.*]], label [[FOR_BODY]]48; CHECK: for.body:49; CHECK-NEXT: [[A_GEP:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[IV]]50; CHECK-NEXT: [[LV:%.*]] = load i32, ptr [[A_GEP]], align 451; CHECK-NEXT: [[B_GEP:%.*]] = getelementptr inbounds i32, ptr [[B]], i32 [[IV]]52; CHECK-NEXT: store i32 [[LV]], ptr [[B_GEP]], align 453; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 154; CHECK-NEXT: [[COND_1:%.*]] = icmp ult i32 [[IV_NEXT]], 100055; CHECK-NEXT: br i1 [[COND_1]], label [[LOOP_HEADER]], label [[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]56; CHECK: exit:57; CHECK-NEXT: ret void58;59entry:60 br label %loop.header61 62loop.header:63 %iv = phi i32 [ 0, %entry ], [ %iv.next, %for.body ]64 %cond.0 = icmp eq i32 %iv, %N65 br i1 %cond.0, label %exit, label %for.body66 67for.body:68 %A.gep = getelementptr inbounds i32, ptr %A, i32 %iv69 %lv = load i32, ptr %A.gep, align 470 %B.gep = getelementptr inbounds i32, ptr %B, i32 %iv71 store i32 %lv, ptr %B.gep, align 472 %iv.next = add nuw i32 %iv, 173 %cond.1 = icmp ult i32 %iv.next, 100074 br i1 %cond.1, label %loop.header, label %exit75 76exit:77 ret void78}79 80 81; Multiple branches exiting the loop to different blocks.82define i32 @multiple_exits_multiple_exit_blocks(ptr %A, ptr %B, i32 %N) #0 {83; CHECK-LABEL: @multiple_exits_multiple_exit_blocks(84; CHECK-NEXT: entry:85; CHECK-NEXT: [[A2:%.*]] = ptrtoint ptr [[A:%.*]] to i3286; CHECK-NEXT: [[B1:%.*]] = ptrtoint ptr [[B:%.*]] to i3287; CHECK-NEXT: [[UMIN:%.*]] = call i32 @llvm.umin.i32(i32 [[N:%.*]], i32 999)88; CHECK-NEXT: [[TMP0:%.*]] = add nuw nsw i32 [[UMIN]], 189; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 [[TMP0]], 490; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]91; CHECK: vector.memcheck:92; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[B1]], [[A2]]93; CHECK-NEXT: [[DIFF_CHECK:%.*]] = icmp ult i32 [[TMP1]], 1694; CHECK-NEXT: br i1 [[DIFF_CHECK]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]95; CHECK: vector.ph:96; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 497; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i32 [[N_MOD_VF]], 098; CHECK-NEXT: [[TMP3:%.*]] = select i1 [[TMP2]], i32 4, i32 [[N_MOD_VF]]99; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[TMP3]]100; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]101; CHECK: vector.body:102; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]103; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[INDEX]]104; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP5]], align 4105; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[B]], i32 [[INDEX]]106; CHECK-NEXT: store <4 x i32> [[WIDE_LOAD]], ptr [[TMP7]], align 4107; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4108; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]109; CHECK-NEXT: br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]110; CHECK: middle.block:111; CHECK-NEXT: br label [[SCALAR_PH]]112; CHECK: scalar.ph:113; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_MEMCHECK]] ]114; CHECK-NEXT: br label [[LOOP_HEADER:%.*]]115; CHECK: loop.header:116; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY:%.*]] ]117; CHECK-NEXT: [[COND_0:%.*]] = icmp eq i32 [[IV]], [[N]]118; CHECK-NEXT: br i1 [[COND_0]], label [[EXIT_0:%.*]], label [[FOR_BODY]]119; CHECK: for.body:120; CHECK-NEXT: [[A_GEP:%.*]] = getelementptr inbounds i32, ptr [[A]], i32 [[IV]]121; CHECK-NEXT: [[LV:%.*]] = load i32, ptr [[A_GEP]], align 4122; CHECK-NEXT: [[B_GEP:%.*]] = getelementptr inbounds i32, ptr [[B]], i32 [[IV]]123; CHECK-NEXT: store i32 [[LV]], ptr [[B_GEP]], align 4124; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1125; CHECK-NEXT: [[COND_1:%.*]] = icmp ult i32 [[IV_NEXT]], 1000126; CHECK-NEXT: br i1 [[COND_1]], label [[LOOP_HEADER]], label [[EXIT_1:%.*]], !llvm.loop [[LOOP5:![0-9]+]]127; CHECK: exit.0:128; CHECK-NEXT: ret i32 1129; CHECK: exit.1:130; CHECK-NEXT: ret i32 2131;132entry:133 br label %loop.header134 135loop.header:136 %iv = phi i32 [ 0, %entry ], [ %iv.next, %for.body ]137 %cond.0 = icmp eq i32 %iv, %N138 br i1 %cond.0, label %exit.0, label %for.body139 140for.body:141 %A.gep = getelementptr inbounds i32, ptr %A, i32 %iv142 %lv = load i32, ptr %A.gep, align 4143 %B.gep = getelementptr inbounds i32, ptr %B, i32 %iv144 store i32 %lv, ptr %B.gep, align 4145 %iv.next = add nuw i32 %iv, 1146 %cond.1 = icmp ult i32 %iv.next, 1000147 br i1 %cond.1, label %loop.header, label %exit.1148 149exit.0:150 ret i32 1151 152exit.1:153 ret i32 2154}155 156attributes #0 = { "target-features"="+mve" }157