172 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --filter-out-after "scalar.ph:" --version 62; RUN: opt -passes=loop-vectorize -enable-mem-access-versioning -force-vector-width=2 -force-vector-interleave=1 -S %s | FileCheck %s3 4; Check that we version this loop with speculating the value 1 for symbolic5; strides. This also checks that the symbolic stride information is correctly6; propagated to the memcheck generation. Without this the loop wouldn't7; vectorize because we couldn't determine the array bounds for the required8; memchecks.9 10define void @test(ptr noalias %A, i64 %AStride, ptr noalias %B, i32 %BStride, ptr noalias %C, i64 %CStride) {11; CHECK-LABEL: define void @test(12; CHECK-SAME: ptr noalias [[A:%.*]], i64 [[ASTRIDE:%.*]], ptr noalias [[B:%.*]], i32 [[BSTRIDE:%.*]], ptr noalias [[C:%.*]], i64 [[CSTRIDE:%.*]]) {13; CHECK-NEXT: [[ENTRY:.*:]]14; CHECK-NEXT: br label %[[VECTOR_SCEVCHECK:.*]]15; CHECK: [[VECTOR_SCEVCHECK]]:16; CHECK-NEXT: [[IDENT_CHECK1:%.*]] = icmp ne i32 [[BSTRIDE]], 117; CHECK-NEXT: [[IDENT_CHECK2:%.*]] = icmp ne i64 [[CSTRIDE]], 118; CHECK-NEXT: [[DIFF_CHECK6:%.*]] = icmp ne i64 [[ASTRIDE]], 119; CHECK-NEXT: [[DIFF_CHECK:%.*]] = or i1 [[IDENT_CHECK1]], [[IDENT_CHECK2]]20; CHECK-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[DIFF_CHECK]], [[DIFF_CHECK6]]21; CHECK-NEXT: br i1 [[CONFLICT_RDX]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]22; CHECK: [[VECTOR_PH]]:23; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]24; CHECK: [[VECTOR_BODY]]:25; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]26; CHECK-NEXT: [[TMP4:%.*]] = trunc i64 [[INDEX]] to i3227; CHECK-NEXT: [[TMP5:%.*]] = zext i32 [[TMP4]] to i6428; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[TMP5]]29; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP6]], align 430; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDEX]]31; CHECK-NEXT: [[WIDE_LOAD7:%.*]] = load <2 x i32>, ptr [[TMP7]], align 432; CHECK-NEXT: [[TMP8:%.*]] = mul nsw <2 x i32> [[WIDE_LOAD7]], [[WIDE_LOAD]]33; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDEX]]34; CHECK-NEXT: store <2 x i32> [[TMP8]], ptr [[TMP9]], align 435; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 236; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10037; CHECK-NEXT: br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]38; CHECK: [[MIDDLE_BLOCK]]:39; CHECK-NEXT: br [[EXIT:label %.*]]40; CHECK: [[SCALAR_PH]]:41;42entry:43 br label %loop44 45loop:46 %iv = phi i64 [ %iv.next, %loop ], [ 0, %entry ]47 %iv.trunc = trunc i64 %iv to i3248 %mul = mul i32 %iv.trunc, %BStride49 %mul64 = zext i32 %mul to i6450 %gep.x = getelementptr inbounds i32, ptr %B, i64 %mul6451 %0 = load i32, ptr %gep.x, align 452 %mul2 = mul nsw i64 %iv, %CStride53 %gep.c = getelementptr inbounds i32, ptr %C, i64 %mul254 %1 = load i32, ptr %gep.c, align 455 %mul4 = mul nsw i32 %1, %056 %mul3 = mul nsw i64 %iv, %AStride57 %gep.a = getelementptr inbounds i32, ptr %A, i64 %mul358 store i32 %mul4, ptr %gep.a, align 459 %iv.next = add nuw nsw i64 %iv, 160 %ec = icmp eq i64 %iv.next, 10061 br i1 %ec, label %exit, label %loop62 63exit:64 ret void65}66 67; We used to crash on this function because we removed the fptosi cast when68; replacing the symbolic stride '%conv'.69; PR1848070 71define void @fn1(ptr noalias %x, ptr noalias %c, double %a) {72; CHECK-LABEL: define void @fn1(73; CHECK-SAME: ptr noalias [[X:%.*]], ptr noalias [[C:%.*]], double [[A:%.*]]) {74; CHECK-NEXT: [[ENTRY:.*:]]75; CHECK-NEXT: [[CONV:%.*]] = fptosi double [[A]] to i3276; CHECK-NEXT: [[CONV2:%.*]] = add i32 [[CONV]], 477; CHECK-NEXT: [[CMP8:%.*]] = icmp sgt i32 [[CONV2]], 078; CHECK-NEXT: br i1 [[CMP8]], label %[[LOOP_PREHEADER:.*]], [[EXIT:label %.*]]79; CHECK: [[LOOP_PREHEADER]]:80; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[CONV2]] to i6481; CHECK-NEXT: br label %[[VECTOR_SCEVCHECK:.*]]82; CHECK: [[VECTOR_SCEVCHECK]]:83; CHECK-NEXT: [[IDENT_CHECK:%.*]] = icmp ne i32 [[CONV]], 184; CHECK-NEXT: br i1 [[IDENT_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]85; CHECK: [[VECTOR_PH]]:86; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]87; CHECK: [[VECTOR_BODY]]:88; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]89; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[INDEX]] to i3290; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[TMP1]] to i6491; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds double, ptr [[X]], i64 [[TMP2]]92; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP3]], align 893; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds double, ptr [[C]], i64 [[INDEX]]94; CHECK-NEXT: store <2 x double> [[WIDE_LOAD]], ptr [[TMP4]], align 895; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 296; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], 497; CHECK-NEXT: br i1 [[TMP5]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]98; CHECK: [[MIDDLE_BLOCK]]:99; CHECK-NEXT: br label %[[SCALAR_PH]]100; CHECK: [[SCALAR_PH]]:101;102entry:103 %conv = fptosi double %a to i32104 %conv2 = add i32 %conv, 4105 %cmp8 = icmp sgt i32 %conv2, 0106 br i1 %cmp8, label %loop, label %exit107 108loop:109 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]110 %0 = trunc i64 %iv to i32111 %mul = mul nsw i32 %0, %conv112 %mul.ext = sext i32 %mul to i64113 %gep.x = getelementptr inbounds double, ptr %x, i64 %mul.ext114 %1 = load double, ptr %gep.x, align 8115 %gep.c = getelementptr inbounds double, ptr %c, i64 %iv116 store double %1, ptr %gep.c, align 8117 %iv.next = add nuw nsw i64 %iv, 1118 %iv.trunc = trunc i64 %iv.next to i32119 %ec = icmp eq i32 %iv.trunc, %conv2120 br i1 %ec, label %exit, label %loop121 122exit:123 ret void124}125 126; Make sure we do not crash when the stride is poison.127; Test for https://github.com/llvm/llvm-project/issues/162922.128define void @stride_poison(ptr %dst) mustprogress {129; CHECK-LABEL: define void @stride_poison(130; CHECK-SAME: ptr [[DST:%.*]]) #[[ATTR0:[0-9]+]] {131; CHECK-NEXT: [[ENTRY:.*:]]132; CHECK-NEXT: [[UMAX:%.*]] = call i64 @llvm.umax.i64(i64 poison, i64 1)133; CHECK-NEXT: [[TMP0:%.*]] = udiv i64 99, [[UMAX]]134; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], 2135; CHECK-NEXT: br label %[[VECTOR_PH:.*]]136; CHECK: [[VECTOR_PH]]:137; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP1]], 2138; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP1]], [[N_MOD_VF]]139; CHECK-NEXT: [[TMP2:%.*]] = mul i64 [[N_VEC]], poison140; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]141; CHECK: [[VECTOR_BODY]]:142; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]143; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], poison144; CHECK-NEXT: [[TMP3:%.*]] = add i64 [[OFFSET_IDX]], poison145; CHECK-NEXT: [[TMP4:%.*]] = add i64 [[OFFSET_IDX]], poison146; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[DST]], i64 [[TMP3]]147; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[DST]], i64 [[TMP4]]148; CHECK-NEXT: store i8 0, ptr [[TMP5]], align 1149; CHECK-NEXT: store i8 0, ptr [[TMP6]], align 1150; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2151; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]152; CHECK-NEXT: br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]153; CHECK: [[MIDDLE_BLOCK]]:154; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP1]], [[N_VEC]]155; CHECK-NEXT: br i1 [[CMP_N]], [[EXIT:label %.*]], label %[[SCALAR_PH:.*]]156; CHECK: [[SCALAR_PH]]:157;158entry:159 br label %loop160 161loop:162 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]163 %gep.dst = getelementptr i8, ptr %dst, i64 %iv164 store i8 0, ptr %gep.dst, align 1165 %iv.next = add nuw nsw i64 %iv, poison166 %ec = icmp samesign ult i64 %iv, 100167 br i1 %ec, label %loop, label %exit168 169exit:170 ret void171}172