diff --git a/llvm/include/llvm/Support/GenericDomTree.h b/llvm/include/llvm/Support/GenericDomTree.h index 7e2b68e6faea2..45ef38b965b75 100644 --- a/llvm/include/llvm/Support/GenericDomTree.h +++ b/llvm/include/llvm/Support/GenericDomTree.h @@ -397,6 +397,8 @@ class DominatorTreeBase { /// may (but is not required to) be null for a forward (backwards) /// statically unreachable block. DomTreeNodeBase *getNode(const NodeT *BB) const { + assert((!BB || Parent == NodeTrait::getParent(const_cast(BB))) && + "cannot get DomTreeNode of block with different parent"); if (auto Idx = getNodeIndex(BB); Idx && *Idx < DomTreeNodes.size()) return DomTreeNodes[*Idx].get(); return nullptr; diff --git a/llvm/lib/Analysis/TypeMetadataUtils.cpp b/llvm/lib/Analysis/TypeMetadataUtils.cpp index 67ce1540112bb..9ec0785eb5034 100644 --- a/llvm/lib/Analysis/TypeMetadataUtils.cpp +++ b/llvm/lib/Analysis/TypeMetadataUtils.cpp @@ -33,6 +33,8 @@ findCallsAtConstantOffset(SmallVectorImpl &DevirtCalls, // after indirect call promotion and inlining, where we may have uses // of the vtable pointer guarded by a function pointer check, and a fallback // indirect call. + if (CI->getFunction() != User->getFunction()) + continue; if (!DT.dominates(CI, User)) continue; if (isa(User)) { diff --git a/llvm/lib/Transforms/Scalar/AlignmentFromAssumptions.cpp b/llvm/lib/Transforms/Scalar/AlignmentFromAssumptions.cpp index f3422a705dca7..8555ef5c22f82 100644 --- a/llvm/lib/Transforms/Scalar/AlignmentFromAssumptions.cpp +++ b/llvm/lib/Transforms/Scalar/AlignmentFromAssumptions.cpp @@ -208,6 +208,7 @@ bool AlignmentFromAssumptionsPass::processAssumption(CallInst *ACall, continue; if (Instruction *K = dyn_cast(J)) + if (K->getFunction() == ACall->getFunction()) WorkList.push_back(K); } diff --git a/llvm/lib/Transforms/Scalar/LoopFuse.cpp b/llvm/lib/Transforms/Scalar/LoopFuse.cpp index 8512b2accbe7c..fe0e30d1965e0 100644 --- a/llvm/lib/Transforms/Scalar/LoopFuse.cpp +++ b/llvm/lib/Transforms/Scalar/LoopFuse.cpp @@ -1729,7 +1729,9 @@ struct LoopFuser { // mergeLatch may remove the only block in FC1. SE.forgetLoop(FC1.L); SE.forgetLoop(FC0.L); - SE.forgetLoopDispositions(); + // Forget block dispositions as well, so that there are no dangling + // pointers to erased/free'ed blocks. + SE.forgetBlockAndLoopDispositions(); // Move instructions from FC0.Latch to FC1.Latch. // Note: mergeLatch requires an updated DT. @@ -2023,7 +2025,9 @@ struct LoopFuser { // mergeLatch may remove the only block in FC1. SE.forgetLoop(FC1.L); SE.forgetLoop(FC0.L); - SE.forgetLoopDispositions(); + // Forget block dispositions as well, so that there are no dangling + // pointers to erased/free'ed blocks. + SE.forgetBlockAndLoopDispositions(); // Move instructions from FC0.Latch to FC1.Latch. // Note: mergeLatch requires an updated DT. diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp index 186b382addd71..edacb2fb33540 100644 --- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp +++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp @@ -5991,6 +5991,9 @@ BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const { DenseMap> PtrToStoresMap; for (unsigned Lane : seq(0, TE->Scalars.size())) { Value *V = TE->Scalars[Lane]; + // Don't iterate over the users of constant data. + if (isa(V)) + continue; // To save compilation time we don't visit if we have too many users. if (V->hasNUsesOrMore(UsesLimit)) break; @@ -5998,7 +6001,9 @@ BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const { // Collect stores per pointer object. for (User *U : V->users()) { auto *SI = dyn_cast(U); - if (SI == nullptr || !SI->isSimple() || + // Test whether we can handle the store. V might be a global, which could + // be used in a different function. + if (SI == nullptr || !SI->isSimple() || SI->getFunction() != F || !isValidElementType(SI->getValueOperand()->getType())) continue; // Skip entry if already diff --git a/llvm/test/Transforms/AlignmentFromAssumptions/domtree-crash.ll b/llvm/test/Transforms/AlignmentFromAssumptions/domtree-crash.ll new file mode 100644 index 0000000000000..c7fc1dc699671 --- /dev/null +++ b/llvm/test/Transforms/AlignmentFromAssumptions/domtree-crash.ll @@ -0,0 +1,33 @@ +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 +; RUN: opt -passes=alignment-from-assumptions -S < %s | FileCheck %s + +; The alignment assumption is a global, which has users in a different +; function. Test that in this case the dominator tree is only queried with +; blocks from the same function. + +@global = external constant [192 x i8] + +define void @fn1() { +; CHECK-LABEL: define void @fn1() { +; CHECK-NEXT: call void @llvm.assume(i1 false) [ "align"(ptr @global, i64 1) ] +; CHECK-NEXT: ret void +; + call void @llvm.assume(i1 false) [ "align"(ptr @global, i64 1) ] + ret void +} + +define void @fn2() { +; CHECK-LABEL: define void @fn2() { +; CHECK-NEXT: ret void +; CHECK: [[LOOP:.*]]: +; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i8, ptr @global, i64 0 +; CHECK-NEXT: [[LOAD:%.*]] = load i64, ptr [[GEP]], align 1 +; CHECK-NEXT: br label %[[LOOP]] +; + ret void + +loop: + %gep = getelementptr inbounds i8, ptr @global, i64 0 + %load = load i64, ptr %gep, align 1 + br label %loop +} diff --git a/llvm/test/Transforms/SLPVectorizer/X86/const-in-different-functions.ll b/llvm/test/Transforms/SLPVectorizer/X86/const-in-different-functions.ll new file mode 100644 index 0000000000000..2e473f4f2c213 --- /dev/null +++ b/llvm/test/Transforms/SLPVectorizer/X86/const-in-different-functions.ll @@ -0,0 +1,48 @@ +; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5 +; RUN: opt -S -mtriple=x86_64 -passes=slp-vectorizer < %s | FileCheck %s + +; Test that SLP vectorize doesn't crash if a stored constant is used in multiple +; functions. + +@p = external global [64 x float] + +define void @_Z1hPfl() { +; CHECK-LABEL: define void @_Z1hPfl() { +; CHECK-NEXT: [[ENTRY:.*:]] +; CHECK-NEXT: [[TMP0:%.*]] = getelementptr i8, ptr @p, i64 28 +; CHECK-NEXT: store <2 x float> , ptr [[TMP0]], align 4 +; CHECK-NEXT: ret void +; +entry: + %0 = getelementptr i8, ptr @p, i64 28 + store float 0.000000e+00, ptr %0, align 4 + %1 = getelementptr i8, ptr @p, i64 32 + store float 1.000000e+00, ptr %1, align 16 + ret void +} + +define void @_Z1mv(i64 %arrayidx4.i.2.idx) { +; CHECK-LABEL: define void @_Z1mv( +; CHECK-SAME: i64 [[ARRAYIDX4_I_2_IDX:%.*]]) { +; CHECK-NEXT: [[ENTRY:.*:]] +; CHECK-NEXT: ret void +; CHECK: [[FOR_COND1_PREHEADER_LR_PH_I:.*:]] +; CHECK-NEXT: br label %[[FOR_COND1_PREHEADER_I:.*]] +; CHECK: [[FOR_COND1_PREHEADER_I]]: +; CHECK-NEXT: store float 1.000000e+00, ptr @p, align 4 +; CHECK-NEXT: [[ARRAYIDX4_I_2:%.*]] = getelementptr i8, ptr @p, i64 [[ARRAYIDX4_I_2_IDX]] +; CHECK-NEXT: store float 0.000000e+00, ptr [[ARRAYIDX4_I_2]], align 4 +; CHECK-NEXT: br label %[[FOR_COND1_PREHEADER_I]] +; +entry: + ret void + +for.cond1.preheader.lr.ph.i: ; No predecessors! + br label %for.cond1.preheader.i + +for.cond1.preheader.i: ; preds = %for.cond1.preheader.i, %for.cond1.preheader.lr.ph.i + store float 1.000000e+00, ptr @p, align 4 + %arrayidx4.i.2 = getelementptr i8, ptr @p, i64 %arrayidx4.i.2.idx + store float 0.000000e+00, ptr %arrayidx4.i.2, align 4 + br label %for.cond1.preheader.i +}