124 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -force-vector-width=4 -force-vector-interleave=1 -passes='loop-vectorize,loop-load-elim' -S %s | FileCheck %s3 4@glob.1 = external global [100 x double]5@glob.2 = external global [100 x double]6 7define void @g(ptr %dst.1, ptr %start, i64 %N) {8; CHECK-LABEL: @g(9; CHECK-NEXT: loop.1.lver.check:10; CHECK-NEXT: [[TMP0:%.*]] = shl i64 [[N:%.*]], 311; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[TMP0]], 1612; CHECK-NEXT: [[UGLYGEP:%.*]] = getelementptr i8, ptr @glob.2, i64 [[TMP1]]13; CHECK-NEXT: [[UGLYGEP2:%.*]] = getelementptr i8, ptr [[DST_1:%.*]], i64 814; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr @glob.2, [[UGLYGEP2]]15; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[DST_1]], [[UGLYGEP]]16; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]17; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label [[LOOP_1_PH_LVER_ORIG:%.*]], label [[LOOP_1_PH:%.*]]18; CHECK: loop.1.ph.lver.orig:19; CHECK-NEXT: br label [[LOOP_1_LVER_ORIG:%.*]]20; CHECK: loop.1.lver.orig:21; CHECK-NEXT: [[IV_LVER_ORIG:%.*]] = phi i64 [ 0, [[LOOP_1_PH_LVER_ORIG]] ], [ [[IV_NEXT_LVER_ORIG:%.*]], [[LOOP_1_LVER_ORIG]] ]22; CHECK-NEXT: [[PTR_IV_1_LVER_ORIG:%.*]] = phi ptr [ @glob.1, [[LOOP_1_PH_LVER_ORIG]] ], [ [[PTR_IV_1_NEXT_LVER_ORIG:%.*]], [[LOOP_1_LVER_ORIG]] ]23; CHECK-NEXT: [[GEP_IV_LVER_ORIG:%.*]] = getelementptr inbounds double, ptr @glob.2, i64 [[IV_LVER_ORIG]]24; CHECK-NEXT: [[L_1_LVER_ORIG:%.*]] = load double, ptr [[GEP_IV_LVER_ORIG]], align 825; CHECK-NEXT: [[GEP_IV_1_LVER_ORIG:%.*]] = getelementptr inbounds double, ptr getelementptr inbounds (double, ptr @glob.2, i64 1), i64 [[IV_LVER_ORIG]]26; CHECK-NEXT: store double 0.000000e+00, ptr [[GEP_IV_1_LVER_ORIG]], align 827; CHECK-NEXT: store double 0.000000e+00, ptr [[DST_1]], align 828; CHECK-NEXT: [[PTR_IV_1_NEXT_LVER_ORIG]] = getelementptr inbounds double, ptr [[PTR_IV_1_LVER_ORIG]], i64 129; CHECK-NEXT: [[IV_NEXT_LVER_ORIG]] = add nuw nsw i64 [[IV_LVER_ORIG]], 130; CHECK-NEXT: [[EXITCOND_NOT_LVER_ORIG:%.*]] = icmp eq i64 [[IV_LVER_ORIG]], [[N]]31; CHECK-NEXT: br i1 [[EXITCOND_NOT_LVER_ORIG]], label [[LOOP_2_PH_LOOPEXIT:%.*]], label [[LOOP_1_LVER_ORIG]]32; CHECK: loop.1.ph:33; CHECK-NEXT: [[LOAD_INITIAL:%.*]] = load double, ptr @glob.2, align 834; CHECK-NEXT: br label [[LOOP_1:%.*]]35; CHECK: loop.1:36; CHECK-NEXT: [[STORE_FORWARDED:%.*]] = phi double [ [[LOAD_INITIAL]], [[LOOP_1_PH]] ], [ 0.000000e+00, [[LOOP_1]] ]37; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[LOOP_1_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP_1]] ]38; CHECK-NEXT: [[PTR_IV_1:%.*]] = phi ptr [ @glob.1, [[LOOP_1_PH]] ], [ [[PTR_IV_1_NEXT:%.*]], [[LOOP_1]] ]39; CHECK-NEXT: [[GEP_IV:%.*]] = getelementptr inbounds double, ptr @glob.2, i64 [[IV]]40; CHECK-NEXT: [[L_1:%.*]] = load double, ptr [[GEP_IV]], align 841; CHECK-NEXT: [[GEP_IV_1:%.*]] = getelementptr inbounds double, ptr getelementptr inbounds (double, ptr @glob.2, i64 1), i64 [[IV]]42; CHECK-NEXT: store double 0.000000e+00, ptr [[GEP_IV_1]], align 843; CHECK-NEXT: store double 0.000000e+00, ptr [[DST_1]], align 844; CHECK-NEXT: [[PTR_IV_1_NEXT]] = getelementptr inbounds double, ptr [[PTR_IV_1]], i64 145; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 146; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV]], [[N]]47; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[LOOP_2_PH_LOOPEXIT3:%.*]], label [[LOOP_1]]48; CHECK: loop.2.ph.loopexit:49; CHECK-NEXT: [[LCSSA_PTR_IV_1_PH:%.*]] = phi ptr [ [[PTR_IV_1_LVER_ORIG]], [[LOOP_1_LVER_ORIG]] ]50; CHECK-NEXT: br label [[LOOP_2_PH:%.*]]51; CHECK: loop.2.ph.loopexit3:52; CHECK-NEXT: [[LCSSA_PTR_IV_1_PH4:%.*]] = phi ptr [ [[PTR_IV_1]], [[LOOP_1]] ]53; CHECK-NEXT: br label [[LOOP_2_PH]]54; CHECK: loop.2.ph:55; CHECK-NEXT: [[LCSSA_PTR_IV_1:%.*]] = phi ptr [ [[LCSSA_PTR_IV_1_PH]], [[LOOP_2_PH_LOOPEXIT]] ], [ [[LCSSA_PTR_IV_1_PH4]], [[LOOP_2_PH_LOOPEXIT3]] ]56; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 457; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]58; CHECK: vector.ph:59; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], 460; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]61; CHECK-NEXT: [[TMP2:%.*]] = mul i64 [[N_VEC]], 862; CHECK-NEXT: [[IND_END:%.*]] = getelementptr i8, ptr [[LCSSA_PTR_IV_1]], i64 [[TMP2]]63; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]64; CHECK: vector.body:65; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]66; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 867; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[LCSSA_PTR_IV_1]], i64 [[OFFSET_IDX]]68; CHECK-NEXT: store <4 x double> zeroinitializer, ptr [[NEXT_GEP]], align 869; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 470; CHECK-NEXT: [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]71; CHECK-NEXT: br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]72; CHECK: middle.block:73; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]74; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]75; CHECK: scalar.ph:76; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[LOOP_2_PH]] ]77; CHECK-NEXT: [[BC_RESUME_VAL1:%.*]] = phi ptr [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ [[LCSSA_PTR_IV_1]], [[LOOP_2_PH]] ]78; CHECK-NEXT: br label [[LOOP_2:%.*]]79; CHECK: loop.2:80; CHECK-NEXT: [[IV_2:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_2_NEXT:%.*]], [[LOOP_2]] ]81; CHECK-NEXT: [[PTR_IV_2:%.*]] = phi ptr [ [[BC_RESUME_VAL1]], [[SCALAR_PH]] ], [ [[PTR_IV_2_NEXT:%.*]], [[LOOP_2]] ]82; CHECK-NEXT: store double 0.000000e+00, ptr [[PTR_IV_2]], align 883; CHECK-NEXT: [[PTR_IV_2_NEXT]] = getelementptr inbounds double, ptr [[PTR_IV_2]], i64 184; CHECK-NEXT: [[IV_2_NEXT]] = add nuw nsw i64 [[IV_2]], 185; CHECK-NEXT: [[EXITCOND_1_NOT:%.*]] = icmp eq i64 [[IV_2_NEXT]], [[N]]86; CHECK-NEXT: br i1 [[EXITCOND_1_NOT]], label [[EXIT_LOOPEXIT:%.*]], label [[LOOP_2]], !llvm.loop [[LOOP3:![0-9]+]]87; CHECK: exit.loopexit:88; CHECK-NEXT: br label [[EXIT]]89; CHECK: exit:90; CHECK-NEXT: ret void91;92entry:93 br label %loop.194 95loop.1:96 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.1 ]97 %ptr.iv.1 = phi ptr [ @glob.1, %entry ], [ %ptr.iv.1.next, %loop.1 ]98 %gep.iv = getelementptr inbounds double, ptr @glob.2, i64 %iv99 %l.1 = load double, ptr %gep.iv, align 8100 %gep.iv.1 = getelementptr inbounds double, ptr getelementptr inbounds (double, ptr @glob.2, i64 1), i64 %iv101 store double 0.000000e+00, ptr %gep.iv.1, align 8102 store double 0.000000e+00, ptr %dst.1, align 8103 %ptr.iv.1.next = getelementptr inbounds double, ptr %ptr.iv.1, i64 1104 %iv.next = add nuw nsw i64 %iv, 1105 %exitcond.not = icmp eq i64 %iv, %N106 br i1 %exitcond.not, label %loop.2.ph, label %loop.1107 108loop.2.ph:109 %lcssa.ptr.iv.1 = phi ptr [ %ptr.iv.1, %loop.1 ]110 br label %loop.2111 112loop.2:113 %iv.2 = phi i64 [ 0, %loop.2.ph ] , [ %iv.2.next, %loop.2 ]114 %ptr.iv.2 = phi ptr [ %lcssa.ptr.iv.1, %loop.2.ph ], [ %ptr.iv.2.next, %loop.2 ]115 store double 0.000000e+00, ptr %ptr.iv.2, align 8116 %ptr.iv.2.next = getelementptr inbounds double, ptr %ptr.iv.2, i64 1117 %iv.2.next = add nuw nsw i64 %iv.2, 1118 %exitcond.1.not = icmp eq i64 %iv.2.next, %N119 br i1 %exitcond.1.not, label %exit, label %loop.2120 121exit:122 ret void123}124