187 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 52; RUN: opt -p loop-vectorize -force-vector-width=2 -S %s | FileCheck %s3 4define void @loop_guard_on_assume_needed_to_prove_dereferenceable_ptr_arg_noundef(i64 %x, ptr noalias noundef %A, ptr noalias %B, ptr noalias %C) nofree nosync {5; CHECK-LABEL: define void @loop_guard_on_assume_needed_to_prove_dereferenceable_ptr_arg_noundef(6; CHECK-SAME: i64 [[X:%.*]], ptr noalias noundef [[A:%.*]], ptr noalias [[B:%.*]], ptr noalias [[C:%.*]]) #[[ATTR0:[0-9]+]] {7; CHECK-NEXT: [[ENTRY:.*:]]8; CHECK-NEXT: call void @llvm.assume(i1 true) [ "align"(ptr [[A]], i64 4), "dereferenceable"(ptr [[A]], i64 [[X]]) ]9; CHECK-NEXT: [[C_X:%.*]] = icmp uge i64 [[X]], 12810; CHECK-NEXT: br i1 [[C_X]], label %[[LOOP_HEADER_PREHEADER:.*]], [[EXIT:label %.*]]11; CHECK: [[LOOP_HEADER_PREHEADER]]:12; CHECK-NEXT: br label %[[VECTOR_PH:.*]]13; CHECK: [[VECTOR_PH]]:14; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]15; CHECK: [[VECTOR_BODY]]:16; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]17; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]18; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP0]], align 419; CHECK-NEXT: [[TMP1:%.*]] = icmp eq <2 x i32> [[WIDE_LOAD]], zeroinitializer20; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]21; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP2]], align 422; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP1]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]23; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDEX]]24; CHECK-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP14]], align 425; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 226; CHECK-NEXT: [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], 3227; CHECK-NEXT: br i1 [[TMP15]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]28; CHECK: [[MIDDLE_BLOCK]]:29; CHECK-NEXT: br [[EXIT_LOOPEXIT:label %.*]]30; CHECK: [[SCALAR_PH:.*:]]31;32entry:33 call void @llvm.assume(i1 true) [ "align"(ptr %A, i64 4), "dereferenceable"(ptr %A, i64 %x) ]34 %c.x = icmp uge i64 %x, 12835 br i1 %c.x, label %loop.header, label %exit36 37loop.header:38 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]39 %gep.b = getelementptr i32, ptr %B, i64 %iv40 %l.b = load i32, ptr %gep.b, align 441 %c.1 = icmp eq i32 %l.b, 042 br i1 %c.1, label %loop.latch, label %loop.then43 44loop.then:45 %gep.a = getelementptr i32, ptr %A, i64 %iv46 %l.a = load i32, ptr %gep.a, align 447 br label %loop.latch48 49loop.latch:50 %merge = phi i32 [ %l.a, %loop.then ], [ %l.b, %loop.header ]51 %gep.c = getelementptr inbounds i32, ptr %C, i64 %iv52 store i32 %merge, ptr %gep.c, align 453 %iv.next = add nuw nsw i64 %iv, 154 %ec = icmp eq i64 %iv.next, 3255 br i1 %ec, label %exit, label %loop.header56 57exit:58 ret void59}60 61define void @loop_guard_on_assume_needed_to_prove_dereferenceable(i64 %x, ptr noalias %A, ptr noalias %B, ptr noalias %C) nofree nosync {62; CHECK-LABEL: define void @loop_guard_on_assume_needed_to_prove_dereferenceable(63; CHECK-SAME: i64 [[X:%.*]], ptr noalias [[A:%.*]], ptr noalias [[B:%.*]], ptr noalias [[C:%.*]]) #[[ATTR0]] {64; CHECK-NEXT: [[ENTRY:.*:]]65; CHECK-NEXT: call void @llvm.assume(i1 true) [ "noundef"(ptr [[A]]), "align"(ptr [[A]], i64 4), "dereferenceable"(ptr [[A]], i64 [[X]]) ]66; CHECK-NEXT: [[C_X:%.*]] = icmp uge i64 [[X]], 12867; CHECK-NEXT: br i1 [[C_X]], label %[[LOOP_HEADER_PREHEADER:.*]], [[EXIT:label %.*]]68; CHECK: [[LOOP_HEADER_PREHEADER]]:69; CHECK-NEXT: br label %[[VECTOR_PH:.*]]70; CHECK: [[VECTOR_PH]]:71; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]72; CHECK: [[VECTOR_BODY]]:73; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]74; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]75; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP0]], align 476; CHECK-NEXT: [[TMP1:%.*]] = icmp eq <2 x i32> [[WIDE_LOAD]], zeroinitializer77; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]78; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP2]], align 479; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP1]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]80; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDEX]]81; CHECK-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP14]], align 482; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 283; CHECK-NEXT: [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], 3284; CHECK-NEXT: br i1 [[TMP15]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]85; CHECK: [[MIDDLE_BLOCK]]:86; CHECK-NEXT: br [[EXIT_LOOPEXIT:label %.*]]87; CHECK: [[SCALAR_PH:.*:]]88;89entry:90 call void @llvm.assume(i1 true) [ "noundef"(ptr %A), "align"(ptr %A, i64 4), "dereferenceable"(ptr %A, i64 %x) ]91 %c.x = icmp uge i64 %x, 12892 br i1 %c.x, label %loop.header, label %exit93 94loop.header:95 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop.latch ]96 %gep.b = getelementptr i32, ptr %B, i64 %iv97 %l.b = load i32, ptr %gep.b, align 498 %c.1 = icmp eq i32 %l.b, 099 br i1 %c.1, label %loop.latch, label %loop.then100 101loop.then:102 %gep.a = getelementptr i32, ptr %A, i64 %iv103 %l.a = load i32, ptr %gep.a, align 4104 br label %loop.latch105 106loop.latch:107 %merge = phi i32 [ %l.a, %loop.then ], [ %l.b, %loop.header ]108 %gep.c = getelementptr inbounds i32, ptr %C, i64 %iv109 store i32 %merge, ptr %gep.c, align 4110 %iv.next = add nuw nsw i64 %iv, 1111 %ec = icmp eq i64 %iv.next, 32112 br i1 %ec, label %exit, label %loop.header113 114exit:115 ret void116}117 118define void @loop_guard_on_trip_count_needed_to_prove_dereferenceable(i32 %x, ptr noalias dereferenceable(128) align 4 %A, ptr noalias %B, ptr noalias %C) {119; CHECK-LABEL: define void @loop_guard_on_trip_count_needed_to_prove_dereferenceable(120; CHECK-SAME: i32 [[X:%.*]], ptr noalias align 4 dereferenceable(128) [[A:%.*]], ptr noalias [[B:%.*]], ptr noalias [[C:%.*]]) {121; CHECK-NEXT: [[ENTRY:.*:]]122; CHECK-NEXT: [[C_X:%.*]] = icmp sgt i32 [[X]], 0123; CHECK-NEXT: br i1 [[C_X]], label %[[PH:.*]], [[EXIT:label %.*]]124; CHECK: [[PH]]:125; CHECK-NEXT: [[N:%.*]] = tail call i32 @llvm.smin.i32(i32 [[X]], i32 31)126; CHECK-NEXT: [[N_EXT:%.*]] = zext i32 [[N]] to i64127; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N_EXT]], 2128; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]129; CHECK: [[VECTOR_PH]]:130; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N_EXT]], 2131; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[N_EXT]], [[N_MOD_VF]]132; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]133; CHECK: [[VECTOR_BODY]]:134; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]135; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]136; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP0]], align 4137; CHECK-NEXT: [[TMP1:%.*]] = icmp eq <2 x i32> [[WIDE_LOAD]], zeroinitializer138; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]139; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <2 x i32>, ptr [[TMP2]], align 4140; CHECK-NEXT: [[PREDPHI:%.*]] = select <2 x i1> [[TMP1]], <2 x i32> [[WIDE_LOAD]], <2 x i32> [[WIDE_LOAD1]]141; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[C]], i64 [[INDEX]]142; CHECK-NEXT: store <2 x i32> [[PREDPHI]], ptr [[TMP14]], align 4143; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2144; CHECK-NEXT: [[TMP15:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]145; CHECK-NEXT: br i1 [[TMP15]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]146; CHECK: [[MIDDLE_BLOCK]]:147; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_EXT]], [[N_VEC]]148; CHECK-NEXT: br i1 [[CMP_N]], [[EXIT_LOOPEXIT:label %.*]], label %[[SCALAR_PH]]149; CHECK: [[SCALAR_PH]]:150;151entry:152 %c.x = icmp sgt i32 %x, 0153 br i1 %c.x, label %ph, label %exit154 155ph:156 %n = tail call i32 @llvm.smin.i32(i32 %x, i32 31)157 %n.ext = zext i32 %n to i64158 br label %loop.header159 160loop.header:161 %iv = phi i64 [ 0, %ph ], [ %iv.next, %loop.latch ]162 %gep.b = getelementptr i32, ptr %B, i64 %iv163 %l.b = load i32, ptr %gep.b, align 4164 %c.1 = icmp eq i32 %l.b, 0165 br i1 %c.1, label %loop.latch, label %loop.then166 167loop.then:168 %gep.a = getelementptr i32, ptr %A, i64 %iv169 %l.a = load i32, ptr %gep.a, align 4170 br label %loop.latch171 172loop.latch:173 %merge = phi i32 [ %l.a, %loop.then ], [ %l.b, %loop.header ]174 %gep.c = getelementptr inbounds i32, ptr %C, i64 %iv175 store i32 %merge, ptr %gep.c, align 4176 %iv.next = add nuw nsw i64 %iv, 1177 %ec = icmp eq i64 %iv.next, %n.ext178 br i1 %ec, label %exit, label %loop.header179 180exit:181 ret void182}183 184 185 186declare i32 @llvm.smin.i32(i32, i32)187