127 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --filter-out-after "^scalar.ph:" --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=1 -max-dependences=2 -debug-only=loop-vectorize,loop-accesses -S 2>&1 | FileCheck %s3; RUN: opt < %s -mattr=+sve2 -passes=loop-vectorize,instcombine -enable-histogram-loop-vectorization -sve-gather-overhead=2 -sve-scatter-overhead=2 -force-vector-interleave=1 -debug-only=loop-vectorize,loop-accesses -S 2>&1 | FileCheck %s --check-prefix=NORMAL_DEP_LIMIT4; REQUIRES: asserts5 6target triple = "aarch64-unknown-linux-gnu"7 8;; Check that we don't crash if LAA gives up on recording dependences and9;; returns a null pointer.10 11; CHECK-LABEL: LAA: Checking a loop in 'many_deps'12; CHECK: Too many dependences, stopped recording13; CHECK: LV: Can't vectorize due to memory conflicts14; CHECK: LV: Not vectorizing: Cannot prove legality.15 16define void @many_deps(ptr noalias %buckets, ptr %array, ptr %indices, ptr %other, i64 %N) {17; CHECK-LABEL: define void @many_deps(18; CHECK-SAME: ptr noalias [[BUCKETS:%.*]], ptr [[ARRAY:%.*]], ptr [[INDICES:%.*]], ptr [[OTHER:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {19; CHECK-NEXT: entry:20; CHECK-NEXT: br label [[FOR_BODY:%.*]]21; CHECK: for.body:22; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]23; CHECK-NEXT: [[GEP_INDICES:%.*]] = getelementptr inbounds nuw i32, ptr [[INDICES]], i64 [[IV]]24; CHECK-NEXT: [[L_IDX:%.*]] = load i32, ptr [[GEP_INDICES]], align 425; CHECK-NEXT: [[IDXPROM1:%.*]] = zext i32 [[L_IDX]] to i6426; CHECK-NEXT: [[GEP_BUCKET:%.*]] = getelementptr inbounds nuw i32, ptr [[BUCKETS]], i64 [[IDXPROM1]]27; CHECK-NEXT: [[L_BUCKET:%.*]] = load i32, ptr [[GEP_BUCKET]], align 428; CHECK-NEXT: [[INC:%.*]] = add nsw i32 [[L_BUCKET]], 129; CHECK-NEXT: store i32 [[INC]], ptr [[GEP_BUCKET]], align 430; CHECK-NEXT: [[IDX_ADDR:%.*]] = getelementptr inbounds nuw i32, ptr [[ARRAY]], i64 [[IV]]31; CHECK-NEXT: [[IV_TRUNC:%.*]] = trunc i64 [[IV]] to i3232; CHECK-NEXT: store i32 [[IV_TRUNC]], ptr [[IDX_ADDR]], align 433; CHECK-NEXT: [[GEP_OTHER:%.*]] = getelementptr inbounds nuw i32, ptr [[OTHER]], i64 [[IV]]34; CHECK-NEXT: [[L_OTHER:%.*]] = load i32, ptr [[GEP_OTHER]], align 435; CHECK-NEXT: [[ADD_OTHER:%.*]] = add i32 [[L_OTHER]], [[IV_TRUNC]]36; CHECK-NEXT: store i32 [[ADD_OTHER]], ptr [[GEP_OTHER]], align 437; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 138; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]39; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_EXIT:%.*]], label [[FOR_BODY]]40; CHECK: for.exit:41; CHECK-NEXT: ret void42;43; NORMAL_DEP_LIMIT-LABEL: define void @many_deps(44; NORMAL_DEP_LIMIT-SAME: ptr noalias [[BUCKETS:%.*]], ptr [[ARRAY:%.*]], ptr [[INDICES:%.*]], ptr [[OTHER:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {45; NORMAL_DEP_LIMIT-NEXT: entry:46; NORMAL_DEP_LIMIT-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()47; NORMAL_DEP_LIMIT-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 248; NORMAL_DEP_LIMIT-NEXT: [[TMP2:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP1]], i64 8)49; NORMAL_DEP_LIMIT-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], [[TMP2]]50; NORMAL_DEP_LIMIT-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]51; NORMAL_DEP_LIMIT: vector.memcheck:52; NORMAL_DEP_LIMIT-NEXT: [[TMP3:%.*]] = shl i64 [[N]], 253; NORMAL_DEP_LIMIT-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[ARRAY]], i64 [[TMP3]]54; NORMAL_DEP_LIMIT-NEXT: [[SCEVGEP1:%.*]] = getelementptr i8, ptr [[OTHER]], i64 [[TMP3]]55; NORMAL_DEP_LIMIT-NEXT: [[SCEVGEP2:%.*]] = getelementptr i8, ptr [[INDICES]], i64 [[TMP3]]56; NORMAL_DEP_LIMIT-NEXT: [[BOUND0:%.*]] = icmp ult ptr [[ARRAY]], [[SCEVGEP1]]57; NORMAL_DEP_LIMIT-NEXT: [[BOUND1:%.*]] = icmp ult ptr [[OTHER]], [[SCEVGEP]]58; NORMAL_DEP_LIMIT-NEXT: [[FOUND_CONFLICT:%.*]] = and i1 [[BOUND0]], [[BOUND1]]59; NORMAL_DEP_LIMIT-NEXT: [[BOUND03:%.*]] = icmp ult ptr [[ARRAY]], [[SCEVGEP2]]60; NORMAL_DEP_LIMIT-NEXT: [[BOUND14:%.*]] = icmp ult ptr [[INDICES]], [[SCEVGEP]]61; NORMAL_DEP_LIMIT-NEXT: [[FOUND_CONFLICT5:%.*]] = and i1 [[BOUND03]], [[BOUND14]]62; NORMAL_DEP_LIMIT-NEXT: [[CONFLICT_RDX:%.*]] = or i1 [[FOUND_CONFLICT]], [[FOUND_CONFLICT5]]63; NORMAL_DEP_LIMIT-NEXT: [[BOUND06:%.*]] = icmp ult ptr [[OTHER]], [[SCEVGEP2]]64; NORMAL_DEP_LIMIT-NEXT: [[BOUND17:%.*]] = icmp ult ptr [[INDICES]], [[SCEVGEP1]]65; NORMAL_DEP_LIMIT-NEXT: [[FOUND_CONFLICT8:%.*]] = and i1 [[BOUND06]], [[BOUND17]]66; NORMAL_DEP_LIMIT-NEXT: [[CONFLICT_RDX9:%.*]] = or i1 [[CONFLICT_RDX]], [[FOUND_CONFLICT8]]67; NORMAL_DEP_LIMIT-NEXT: br i1 [[CONFLICT_RDX9]], label [[SCALAR_PH]], label [[ENTRY:%.*]]68; NORMAL_DEP_LIMIT: vector.ph:69; NORMAL_DEP_LIMIT-NEXT: [[TMP4:%.*]] = call i64 @llvm.vscale.i64()70; NORMAL_DEP_LIMIT-NEXT: [[TMP8:%.*]] = shl nuw i64 [[TMP4]], 271; NORMAL_DEP_LIMIT-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP8]]72; NORMAL_DEP_LIMIT-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]73; NORMAL_DEP_LIMIT-NEXT: [[TMP7:%.*]] = call <vscale x 4 x i32> @llvm.stepvector.nxv4i32()74; NORMAL_DEP_LIMIT-NEXT: [[TMP9:%.*]] = trunc i64 [[TMP8]] to i3275; NORMAL_DEP_LIMIT-NEXT: [[DOTSPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[TMP9]], i64 076; NORMAL_DEP_LIMIT-NEXT: [[DOTSPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[DOTSPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer77; NORMAL_DEP_LIMIT-NEXT: br label [[FOR_BODY:%.*]]78; NORMAL_DEP_LIMIT: vector.body:79; NORMAL_DEP_LIMIT-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY]] ], [ [[IV_NEXT:%.*]], [[FOR_BODY]] ]80; NORMAL_DEP_LIMIT-NEXT: [[VEC_IND:%.*]] = phi <vscale x 4 x i32> [ [[TMP7]], [[ENTRY]] ], [ [[VEC_IND_NEXT:%.*]], [[FOR_BODY]] ]81; NORMAL_DEP_LIMIT-NEXT: [[GEP_INDICES:%.*]] = getelementptr inbounds i32, ptr [[INDICES]], i64 [[IV]]82; NORMAL_DEP_LIMIT-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[GEP_INDICES]], align 4, !alias.scope [[META0:![0-9]+]]83; NORMAL_DEP_LIMIT-NEXT: [[TMP11:%.*]] = zext <vscale x 4 x i32> [[WIDE_LOAD]] to <vscale x 4 x i64>84; NORMAL_DEP_LIMIT-NEXT: [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[BUCKETS]], <vscale x 4 x i64> [[TMP11]]85; NORMAL_DEP_LIMIT-NEXT: call void @llvm.experimental.vector.histogram.add.nxv4p0.i32(<vscale x 4 x ptr> [[TMP12]], i32 1, <vscale x 4 x i1> splat (i1 true))86; NORMAL_DEP_LIMIT-NEXT: [[TMP13:%.*]] = getelementptr inbounds i32, ptr [[ARRAY]], i64 [[IV]]87; NORMAL_DEP_LIMIT-NEXT: store <vscale x 4 x i32> [[VEC_IND]], ptr [[TMP13]], align 4, !alias.scope [[META3:![0-9]+]], !noalias [[META5:![0-9]+]]88; NORMAL_DEP_LIMIT-NEXT: [[TMP14:%.*]] = getelementptr inbounds i32, ptr [[OTHER]], i64 [[IV]]89; NORMAL_DEP_LIMIT-NEXT: [[WIDE_LOAD10:%.*]] = load <vscale x 4 x i32>, ptr [[TMP14]], align 4, !alias.scope [[META7:![0-9]+]], !noalias [[META0]]90; NORMAL_DEP_LIMIT-NEXT: [[TMP15:%.*]] = add <vscale x 4 x i32> [[WIDE_LOAD10]], [[VEC_IND]]91; NORMAL_DEP_LIMIT-NEXT: store <vscale x 4 x i32> [[TMP15]], ptr [[TMP14]], align 4, !alias.scope [[META7]], !noalias [[META0]]92; NORMAL_DEP_LIMIT-NEXT: [[IV_NEXT]] = add nuw i64 [[IV]], [[TMP8]]93; NORMAL_DEP_LIMIT-NEXT: [[VEC_IND_NEXT]] = add <vscale x 4 x i32> [[VEC_IND]], [[DOTSPLAT]]94; NORMAL_DEP_LIMIT-NEXT: [[TMP16:%.*]] = icmp eq i64 [[IV_NEXT]], [[N_VEC]]95; NORMAL_DEP_LIMIT-NEXT: br i1 [[TMP16]], label [[MIDDLE_BLOCK:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]96; NORMAL_DEP_LIMIT: middle.block:97; NORMAL_DEP_LIMIT-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_MOD_VF]], 098; NORMAL_DEP_LIMIT-NEXT: br i1 [[CMP_N]], label [[FOR_EXIT:%.*]], label [[SCALAR_PH]]99; NORMAL_DEP_LIMIT: scalar.ph:100;101entry:102 br label %for.body103 104for.body:105 %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]106 %gep.indices = getelementptr inbounds i32, ptr %indices, i64 %iv107 %l.idx = load i32, ptr %gep.indices, align 4108 %idxprom1 = zext i32 %l.idx to i64109 %gep.bucket = getelementptr inbounds i32, ptr %buckets, i64 %idxprom1110 %l.bucket = load i32, ptr %gep.bucket, align 4111 %inc = add nsw i32 %l.bucket, 1112 store i32 %inc, ptr %gep.bucket, align 4113 %idx.addr = getelementptr inbounds i32, ptr %array, i64 %iv114 %iv.trunc = trunc i64 %iv to i32115 store i32 %iv.trunc, ptr %idx.addr, align 4116 %gep.other = getelementptr inbounds i32, ptr %other, i64 %iv117 %l.other = load i32, ptr %gep.other, align 4118 %add.other = add i32 %l.other, %iv.trunc119 store i32 %add.other, ptr %gep.other, align 4120 %iv.next = add nuw nsw i64 %iv, 1121 %exitcond = icmp eq i64 %iv.next, %N122 br i1 %exitcond, label %for.exit, label %for.body123 124for.exit:125 ret void126}127