155 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt -passes=loop-vectorize -force-vector-width=4 -S %s | FileCheck %s3 4define void @test(ptr %A, i32 %x) {5; CHECK-LABEL: define void @test(6; CHECK-SAME: ptr [[A:%.*]], i32 [[X:%.*]]) {7; CHECK-NEXT: [[ENTRY:.*:]]8; CHECK-NEXT: br label %[[VECTOR_SCEVCHECK:.*]]9; CHECK: [[VECTOR_SCEVCHECK]]:10; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[X]], 111; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]12; CHECK: [[VECTOR_PH]]:13; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]14; CHECK: [[VECTOR_BODY]]:15; CHECK-NEXT: [[OFFSET_IDX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]16; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[OFFSET_IDX]] to i3217; CHECK-NEXT: [[TMP3:%.*]] = add nuw nsw i64 [[OFFSET_IDX]], 118; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 [[TMP3]] to i3219; CHECK-NEXT: [[TMP5:%.*]] = zext i32 [[TMP4]] to i6420; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP5]]21; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[TMP6]], align 422; CHECK-NEXT: [[TMP8:%.*]] = zext i32 [[TMP0]] to i6423; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP8]]24; CHECK-NEXT: store <4 x float> [[WIDE_LOAD]], ptr [[TMP9]], align 425; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[OFFSET_IDX]], 426; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100027; CHECK-NEXT: br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]28; CHECK: [[MIDDLE_BLOCK]]:29; CHECK-NEXT: br label %[[EXIT:.*]]30; CHECK: [[SCALAR_PH]]:31; CHECK-NEXT: br label %[[LOOP:.*]]32; CHECK: [[LOOP]]:33; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[IV_NEXT:%.*]], %[[LOOP]] ], [ 0, %[[SCALAR_PH]] ]34; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 135; CHECK-NEXT: [[T_IV_NEXT:%.*]] = trunc i64 [[IV_NEXT]] to i3236; CHECK-NEXT: [[MUL_IV_NEXT:%.*]] = mul i32 [[T_IV_NEXT]], [[X]]37; CHECK-NEXT: [[IDX_1:%.*]] = zext i32 [[MUL_IV_NEXT]] to i6438; CHECK-NEXT: [[ARRAYIDX1215:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[IDX_1]]39; CHECK-NEXT: [[LV:%.*]] = load float, ptr [[ARRAYIDX1215]], align 440; CHECK-NEXT: [[T_IV:%.*]] = trunc i64 [[IV]] to i3241; CHECK-NEXT: [[MUL_IV:%.*]] = mul i32 [[T_IV]], [[X]]42; CHECK-NEXT: [[IDX_2:%.*]] = zext i32 [[MUL_IV]] to i6443; CHECK-NEXT: [[ARRAYIDX1209:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[IDX_2]]44; CHECK-NEXT: store float [[LV]], ptr [[ARRAYIDX1209]], align 445; CHECK-NEXT: [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], 100046; CHECK-NEXT: br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]47; CHECK: [[EXIT]]:48; CHECK-NEXT: ret void49;50entry:51 br label %loop52 53loop:54 %iv = phi i64 [ %iv.next, %loop ], [ 0, %entry ]55 %iv.next = add nuw nsw i64 %iv, 156 %t.iv.next = trunc i64 %iv.next to i3257 %mul.iv.next = mul i32 %t.iv.next, %x58 %idx.1 = zext i32 %mul.iv.next to i6459 %arrayidx1215 = getelementptr inbounds float, ptr %A, i64 %idx.160 %lv = load float, ptr %arrayidx1215, align 461 %t.iv = trunc i64 %iv to i3262 %mul.iv = mul i32 %t.iv, %x63 %idx.2 = zext i32 %mul.iv to i6464 %arrayidx1209 = getelementptr inbounds float, ptr %A, i64 %idx.265 store float %lv, ptr %arrayidx1209, align 466 %ec = icmp eq i64 %iv.next, 100067 br i1 %ec, label %exit, label %loop68 69exit:70 ret void71}72 73; !llvm.loop.unroll.runtime.disable metadata should be added to the74; scalar loop, as there are no runtime checks needed (or they can be proven75; false).76define void @diff_memcheck_known_false_for_vf_4(ptr %B, ptr %A, ptr %end) {77; CHECK-LABEL: define void @diff_memcheck_known_false_for_vf_4(78; CHECK-SAME: ptr [[B:%.*]], ptr [[A:%.*]], ptr [[END:%.*]]) {79; CHECK-NEXT: [[ENTRY:.*]]:80; CHECK-NEXT: [[END1:%.*]] = ptrtoint ptr [[END]] to i6481; CHECK-NEXT: [[A_INT:%.*]] = ptrtoint ptr [[A]] to i6482; CHECK-NEXT: [[B_CAST:%.*]] = ptrtoint ptr [[B]] to i6483; CHECK-NEXT: [[PTR_SUB:%.*]] = sub i64 [[A_INT]], [[B_CAST]]84; CHECK-NEXT: [[ADD_PTR11:%.*]] = getelementptr i8, ptr [[B]], i64 [[PTR_SUB]]85; CHECK-NEXT: [[TMP0:%.*]] = sub i64 [[A_INT]], [[END1]]86; CHECK-NEXT: [[TMP1:%.*]] = lshr i64 [[TMP0]], 387; CHECK-NEXT: [[TMP2:%.*]] = add nuw nsw i64 [[TMP1]], 188; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP2]], 489; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]90; CHECK: [[VECTOR_PH]]:91; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP2]], 492; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP2]], [[N_MOD_VF]]93; CHECK-NEXT: [[TMP3:%.*]] = mul i64 [[N_VEC]], -894; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP3]]95; CHECK-NEXT: [[TMP5:%.*]] = mul i64 [[N_VEC]], -896; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[ADD_PTR11]], i64 [[TMP5]]97; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]98; CHECK: [[VECTOR_BODY]]:99; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]100; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], -8101; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[A]], i64 [[OFFSET_IDX]]102; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i64, ptr [[NEXT_GEP]], i64 0103; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i64, ptr [[TMP7]], i64 -3104; CHECK-NEXT: store <4 x i64> zeroinitializer, ptr [[TMP8]], align 8105; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4106; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]107; CHECK-NEXT: br i1 [[TMP9]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]108; CHECK: [[MIDDLE_BLOCK]]:109; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP2]], [[N_VEC]]110; CHECK-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]111; CHECK: [[SCALAR_PH]]:112; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi ptr [ [[TMP4]], %[[MIDDLE_BLOCK]] ], [ [[A]], %[[ENTRY]] ]113; CHECK-NEXT: [[BC_RESUME_VAL2:%.*]] = phi ptr [ [[TMP6]], %[[MIDDLE_BLOCK]] ], [ [[ADD_PTR11]], %[[ENTRY]] ]114; CHECK-NEXT: br label %[[LOOP:.*]]115; CHECK: [[LOOP]]:116; CHECK-NEXT: [[IV_1:%.*]] = phi ptr [ [[IV_1_NEXT:%.*]], %[[LOOP]] ], [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ]117; CHECK-NEXT: [[IV_2:%.*]] = phi ptr [ [[IV_2_NEXT:%.*]], %[[LOOP]] ], [ [[BC_RESUME_VAL2]], %[[SCALAR_PH]] ]118; CHECK-NEXT: [[IV_2_NEXT]] = getelementptr nusw i8, ptr [[IV_2]], i64 -8119; CHECK-NEXT: [[IV_1_NEXT]] = getelementptr i8, ptr [[IV_1]], i64 -8120; CHECK-NEXT: [[TMP10:%.*]] = load i64, ptr [[IV_2_NEXT]], align 8121; CHECK-NEXT: store i64 0, ptr [[IV_1]], align 8122; CHECK-NEXT: [[CMP_NOT_I_I_I_I:%.*]] = icmp eq ptr [[END]], [[IV_2]]123; CHECK-NEXT: br i1 [[CMP_NOT_I_I_I_I]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]124; CHECK: [[EXIT]]:125; CHECK-NEXT: ret void126;127entry:128 %A.int = ptrtoint ptr %A to i64129 %B.cast = ptrtoint ptr %B to i64130 %ptr.sub = sub i64 %A.int, %B.cast131 %add.ptr11 = getelementptr i8, ptr %B, i64 %ptr.sub132 br label %loop133 134loop:135 %iv.1 = phi ptr [ %iv.1.next, %loop ], [ %A, %entry ]136 %iv.2 = phi ptr [ %iv.2.next, %loop ], [ %add.ptr11, %entry ]137 %iv.2.next = getelementptr nusw i8, ptr %iv.2, i64 -8138 %iv.1.next = getelementptr i8, ptr %iv.1, i64 -8139 %2 = load i64, ptr %iv.2.next, align 8140 store i64 0, ptr %iv.1, align 8141 %cmp.not.i.i.i.i = icmp eq ptr %end, %iv.2142 br i1 %cmp.not.i.i.i.i, label %exit, label %loop143 144exit:145 ret void146}147;.148; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}149; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}150; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}151; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]]}152; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}153; CHECK: [[LOOP5]] = distinct !{[[LOOP5]], [[META2]], [[META1]]}154;.155