49 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 32; RUN: opt < %s -mattr=+sve2 -passes=loop-vectorize,instcombine -enable-histogram-loop-vectorization -sve-gather-overhead=2 -sve-scatter-overhead=2 -force-vector-interleave=2 -force-vector-width=1 -debug-only=loop-vectorize -S 2>&1 | FileCheck %s3; REQUIRES: asserts4 5;; Make sure we don't interleave a histogram when vectorization is disabled.6 7; CHECK-LABEL: LV: Checking a loop in 'simple_histogram_forced_scalar_interleave'8; CHECK: LV: Not interleaving without vectorization due to histogram operations.9 10define void @simple_histogram_forced_scalar_interleave(ptr noalias %buckets, ptr readonly %indices, i64 %N) {11; CHECK-LABEL: define void @simple_histogram_forced_scalar_interleave(12; CHECK-SAME: ptr noalias [[BUCKETS:%.*]], ptr readonly [[INDICES:%.*]], i64 [[N:%.*]]) {13; CHECK-NEXT: entry:14; CHECK-NEXT: br label [[FOR_BODY:%.*]]15; CHECK: for.body:16; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]17; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds nuw i32, ptr [[INDICES]], i64 [[IV]]18; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 419; CHECK-NEXT: [[IDXPROM1:%.*]] = zext i32 [[TMP0]] to i6420; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds nuw i32, ptr [[BUCKETS]], i64 [[IDXPROM1]]21; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 422; CHECK-NEXT: [[INC:%.*]] = add nsw i32 [[TMP1]], 123; CHECK-NEXT: store i32 [[INC]], ptr [[ARRAYIDX2]], align 424; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 125; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]26; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_EXIT:%.*]], label [[FOR_BODY]]27; CHECK: for.exit:28; CHECK-NEXT: ret void29;30entry:31 br label %for.body32 33for.body:34 %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]35 %gep.indices = getelementptr inbounds i32, ptr %indices, i64 %iv36 %l.idx = load i32, ptr %gep.indices, align 437 %idxprom1 = zext i32 %l.idx to i6438 %gep.bucket = getelementptr inbounds i32, ptr %buckets, i64 %idxprom139 %l.bucket = load i32, ptr %gep.bucket, align 440 %inc = add nsw i32 %l.bucket, 141 store i32 %inc, ptr %gep.bucket, align 442 %iv.next = add nuw nsw i64 %iv, 143 %exitcond = icmp eq i64 %iv.next, %N44 br i1 %exitcond, label %for.exit, label %for.body45 46for.exit:47 ret void48}49