97 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 62; RUN: opt -passes='default<O3>' -S %s | FileCheck %s3 4target triple = "arm64-apple-macosx"5 6%"class.dealii::VectorizedArray" = type { [4 x double] }7 8define void @hoist_invariant_load(ptr %invariant_ptr, i64 %num_elements, ptr %array) {9; CHECK-LABEL: define void @hoist_invariant_load(10; CHECK-SAME: ptr readonly captures(none) [[INVARIANT_PTR:%.*]], i64 [[NUM_ELEMENTS:%.*]], ptr captures(none) [[ARRAY:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {11; CHECK-NEXT: [[ENTRY:.*:]]12; CHECK-NEXT: [[CMP1_NOT:%.*]] = icmp eq i64 [[NUM_ELEMENTS]], 013; CHECK-NEXT: br i1 [[CMP1_NOT]], label %[[EXIT:.*]], label %[[LOOP_LATCH_PREHEADER:.*]]14; CHECK: [[LOOP_LATCH_PREHEADER]]:15; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[NUM_ELEMENTS]], 1116; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[LOOP_LATCH_PREHEADER6:.*]], label %[[VECTOR_MEMCHECK:.*]]17; CHECK: [[VECTOR_MEMCHECK]]:18; CHECK-NEXT: [[TMP0:%.*]] = shl i64 [[NUM_ELEMENTS]], 519; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[ARRAY]], i64 [[TMP0]]20; CHECK-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[TMP1]], i64 -2421; CHECK-NEXT: [[SCEVGEP3:%.*]] = getelementptr i8, ptr [[INVARIANT_PTR]], i64 822; CHECK-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[ARRAY]], [[SCEVGEP3]]23; CHECK-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[INVARIANT_PTR]], [[SCEVGEP]]24; CHECK-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]25; CHECK-NEXT: br i1 [[FOUND_CONFLICT]], label %[[LOOP_LATCH_PREHEADER6]], label %[[VECTOR_PH:.*]]26; CHECK: [[VECTOR_PH]]:27; CHECK-NEXT: [[N_MOD_VF:%.*]] = and i64 [[NUM_ELEMENTS]], 328; CHECK-NEXT: [[TMP2:%.*]] = icmp eq i64 [[N_MOD_VF]], 029; CHECK-NEXT: [[TMP3:%.*]] = select i1 [[TMP2]], i64 4, i64 [[N_MOD_VF]]30; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[NUM_ELEMENTS]], [[TMP3]]31; CHECK-NEXT: [[TMP4:%.*]] = load double, ptr [[INVARIANT_PTR]], align 8, !alias.scope [[META0:![0-9]+]]32; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x double> poison, double [[TMP4]], i64 033; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x double> [[BROADCAST_SPLATINSERT]], <2 x double> poison, <2 x i32> zeroinitializer34; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]35; CHECK: [[VECTOR_BODY]]:36; CHECK-NEXT: [[I2:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]37; CHECK-NEXT: [[GEP:%.*]] = getelementptr nusw %"class.dealii::VectorizedArray", ptr [[ARRAY]], i64 [[I2]]38; CHECK-NEXT: [[TMP6:%.*]] = getelementptr %"class.dealii::VectorizedArray", ptr [[ARRAY]], i64 [[I2]]39; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[TMP6]], i64 3240; CHECK-NEXT: [[TMP8:%.*]] = getelementptr %"class.dealii::VectorizedArray", ptr [[ARRAY]], i64 [[I2]]41; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[TMP8]], i64 6442; CHECK-NEXT: [[TMP10:%.*]] = getelementptr %"class.dealii::VectorizedArray", ptr [[ARRAY]], i64 [[I2]]43; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i8, ptr [[TMP10]], i64 9644; CHECK-NEXT: [[TMP12:%.*]] = load <5 x double>, ptr [[GEP]], align 8, !alias.scope [[META3:![0-9]+]], !noalias [[META0]]45; CHECK-NEXT: [[STRIDED_VEC:%.*]] = shufflevector <5 x double> [[TMP12]], <5 x double> poison, <2 x i32> <i32 0, i32 4>46; CHECK-NEXT: [[TMP13:%.*]] = load <5 x double>, ptr [[TMP9]], align 8, !alias.scope [[META3]], !noalias [[META0]]47; CHECK-NEXT: [[STRIDED_VEC5:%.*]] = shufflevector <5 x double> [[TMP13]], <5 x double> poison, <2 x i32> <i32 0, i32 4>48; CHECK-NEXT: [[TMP14:%.*]] = fadd <2 x double> [[BROADCAST_SPLAT]], [[STRIDED_VEC]]49; CHECK-NEXT: [[TMP15:%.*]] = extractelement <2 x double> [[TMP14]], i64 050; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[TMP14]], i64 151; CHECK-NEXT: [[TMP17:%.*]] = fadd <2 x double> [[BROADCAST_SPLAT]], [[STRIDED_VEC5]]52; CHECK-NEXT: [[TMP18:%.*]] = extractelement <2 x double> [[TMP17]], i64 053; CHECK-NEXT: [[TMP19:%.*]] = extractelement <2 x double> [[TMP17]], i64 154; CHECK-NEXT: store double [[TMP15]], ptr [[GEP]], align 8, !alias.scope [[META3]], !noalias [[META0]]55; CHECK-NEXT: store double [[TMP16]], ptr [[TMP7]], align 8, !alias.scope [[META3]], !noalias [[META0]]56; CHECK-NEXT: store double [[TMP18]], ptr [[TMP9]], align 8, !alias.scope [[META3]], !noalias [[META0]]57; CHECK-NEXT: store double [[TMP19]], ptr [[TMP11]], align 8, !alias.scope [[META3]], !noalias [[META0]]58; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[I2]], 459; CHECK-NEXT: [[TMP20:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]60; CHECK-NEXT: br i1 [[TMP20]], label %[[LOOP_LATCH_PREHEADER6]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]]61; CHECK: [[LOOP_LATCH_PREHEADER6]]:62; CHECK-NEXT: [[I2_PH:%.*]] = phi i64 [ 0, %[[VECTOR_MEMCHECK]] ], [ 0, %[[LOOP_LATCH_PREHEADER]] ], [ [[N_VEC]], %[[VECTOR_BODY]] ]63; CHECK-NEXT: br label %[[LOOP_LATCH:.*]]64; CHECK: [[LOOP_LATCH]]:65; CHECK-NEXT: [[I3:%.*]] = phi i64 [ [[I_NEXT:%.*]], %[[LOOP_LATCH]] ], [ [[I2_PH]], %[[LOOP_LATCH_PREHEADER6]] ]66; CHECK-NEXT: [[GEP1:%.*]] = getelementptr nusw %"class.dealii::VectorizedArray", ptr [[ARRAY]], i64 [[I3]]67; CHECK-NEXT: [[INVARIANT_VAL:%.*]] = load double, ptr [[INVARIANT_PTR]], align 868; CHECK-NEXT: [[ARRAY_VAL:%.*]] = load double, ptr [[GEP1]], align 869; CHECK-NEXT: [[SUM:%.*]] = fadd double [[INVARIANT_VAL]], [[ARRAY_VAL]]70; CHECK-NEXT: store double [[SUM]], ptr [[GEP1]], align 871; CHECK-NEXT: [[I_NEXT]] = add nuw i64 [[I3]], 172; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[I_NEXT]], [[NUM_ELEMENTS]]73; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label %[[EXIT]], label %[[LOOP_LATCH]], !llvm.loop [[LOOP8:![0-9]+]]74; CHECK: [[EXIT]]:75; CHECK-NEXT: ret void76;77entry:78 br label %loop.header79 80loop.header: ; preds = %loop.latch, %entry81 %i = phi i64 [ 0, %entry ], [ %i.next, %loop.latch ]82 %cmp = icmp ult i64 %i, %num_elements83 br i1 %cmp, label %loop.latch, label %exit84 85loop.latch: ; preds = %loop.header86 %gep = getelementptr nusw %"class.dealii::VectorizedArray", ptr %array, i64 %i87 %invariant_val = load double, ptr %invariant_ptr, align 888 %array_val = load double, ptr %gep, align 889 %sum = fadd double %array_val, %invariant_val90 store double %sum, ptr %gep, align 891 %i.next = add i64 %i, 192 br label %loop.header93 94exit: ; preds = %loop.header95 ret void96}97