[MemoryBuiltins][FIX] Adjust index type size properly wrt. AS casts

[jdoerfert]

Summary

Use existing functionality to strip constant offsets that works well
with AS casts and avoids the code duplication.

Since we strip AS casts during the computation of the offset we also
need to adjust the APInt properly to avoid mismatches in the bit width.
This code ensures the caller of compute sees APInts that match the
index type size of the value passed to compute, not the value result
of the strip pointer cast.

[jdoerfert]

Fix comments

[jdoerfert]

llvm/lib/Analysis/MemoryBuiltins.cpp

619

Here, and elsewhere, we could probably give up only on the problematic part (offset vs size).

[arsenm]

llvm/lib/Analysis/MemoryBuiltins.cpp

596–601

There are several flavors of stripPointerCasts. Why isn't this using stripAndAccumulateConstantOffsets?

[jdoerfert]

Use existing strip and accumulate

[jdoerfert]

Remove assertion that does not hold (non-constant GEPs can exist)

[nikic]

llvm/lib/Analysis/MemoryBuiltins.cpp

648

Constant GEPs can have constant expression offsets, which will not be stripped.

[jdoerfert]

llvm/lib/Analysis/MemoryBuiltins.cpp

648

Good point. I'll remove the assertion, nothing we can do with those otherwise.

[jdoerfert]

remove constant expr case and assert

[jdoerfert]

Re-introduce the former bit width adjustment. This passes all unit tests that
the strip change alone failed.

[arsenm]

Missing tests?

[jdoerfert]

Add test

[jdoerfert]

Add check lines

[jdoerfert]

Add test to correct (=latest) version of the patch

[arsenm]

llvm/test/Transforms/InstCombine/builtin-dynamic-object-size.ll

175

doesn't cover the special case constantexpr case?

[jdoerfert]

Add second test with constexpr gep & as cast

llvm/include/llvm/Analysis/MemoryBuiltins.h

public:
   // These are "private", except they can't actually be made private. Only
   // compute() should be used by external users.
   SizeOffsetType visitAllocaInst(AllocaInst &I);
   SizeOffsetType visitArgument(Argument &A);
   SizeOffsetType visitCallBase(CallBase &CB);
   SizeOffsetType visitConstantPointerNull(ConstantPointerNull&);
   SizeOffsetType visitExtractElementInst(ExtractElementInst &I);
   SizeOffsetType visitExtractValueInst(ExtractValueInst &I);
-  SizeOffsetType visitGEPOperator(GEPOperator &GEP);
   SizeOffsetType visitGlobalAlias(GlobalAlias &GA);
   SizeOffsetType visitGlobalVariable(GlobalVariable &GV);
   SizeOffsetType visitIntToPtrInst(IntToPtrInst&);
   SizeOffsetType visitLoadInst(LoadInst &I);
   SizeOffsetType visitPHINode(PHINode&);
   SizeOffsetType visitSelectInst(SelectInst &I);
   SizeOffsetType visitUndefValue(UndefValue&);
   SizeOffsetType visitInstruction(Instruction &I);
 
 private:
+  SizeOffsetType computeImpl(Value *V);
   bool CheckedZextOrTrunc(APInt &I);
 };
 
 using SizeOffsetEvalType = std::pair<Value *, Value *>;
 
 /// Evaluate the size and offset of an object pointed to by a Value*.
 /// May create code to compute the result at run-time.
 class ObjectSizeOffsetEvaluator

llvm/lib/Analysis/MemoryBuiltins.cpp

ObjectSizeOffsetVisitor::ObjectSizeOffsetVisitor(const DataLayout &DL,
                                                  LLVMContext &Context,
                                                  ObjectSizeOpts Options)
     : DL(DL), TLI(TLI), Options(Options) {
   // Pointer size must be rechecked for each object visited since it could have
   // a different address space.
 }
 
 SizeOffsetType ObjectSizeOffsetVisitor::compute(Value *V) {
+  unsigned InitialIntTyBits = DL.getIndexTypeSizeInBits(V->getType());
+
+  // Stripping pointer casts can strip address space casts which can change the
+  // index type size. The invariant is that we use the value type to determine
+  // the index type size and if we stripped address space casts we have to
+  // readjust the APInt as we pass it upwards in order for the APInt to match
+  // the type the caller passed in.
+  APInt Offset(InitialIntTyBits, 0);
+  V = V->stripAndAccumulateConstantOffsets(
+      DL, Offset, /* AllowNonInbounds */ true, /* AllowInvariantGroup */ true);
+
+  // Later we use the index type size and zero but it will match the type of the
+  // value that is passed to computeImpl.
   IntTyBits = DL.getIndexTypeSizeInBits(V->getType());
   Zero = APInt::getZero(IntTyBits);
 
-  V = V->stripPointerCasts();
+  bool IndexTypeSizeChanged = InitialIntTyBits != IntTyBits;
+  if (!IndexTypeSizeChanged && Offset.isZero())
+    return computeImpl(V);
+
+  // We stripped an address space cast that changed the index type size or we
+  // accumulated some constant offset (or both). Readjust the bit width to match
+  // the argument index type size and apply the offset, as required.
+  SizeOffsetType SOT = computeImpl(V);
+  if (IndexTypeSizeChanged) {
+    if (knownSize(SOT) && !::CheckedZextOrTrunc(SOT.first, InitialIntTyBits))
+      SOT.first = APInt();
+    if (knownOffset(SOT) && !::CheckedZextOrTrunc(SOT.second, InitialIntTyBits))
+      SOT.second = APInt();
+  }
+  // If the computed offset is "unknown" we cannot add the stripped offset.
+  return {SOT.first,
+          SOT.second.getBitWidth() > 1 ? SOT.second + Offset : SOT.second};
+}
+
+SizeOffsetType ObjectSizeOffsetVisitor::computeImpl(Value *V) {
   if (Instruction *I = dyn_cast<Instruction>(V)) {
     // If we have already seen this instruction, bail out. Cycles can happen in
     // unreachable code after constant propagation.
     if (!SeenInsts.insert(I).second)
       return unknown();
 
-    if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
-      return visitGEPOperator(*GEP);
     return visit(*I);
   }
   if (Argument *A = dyn_cast<Argument>(V))
     return visitArgument(*A);
   if (ConstantPointerNull *P = dyn_cast<ConstantPointerNull>(V))
     return visitConstantPointerNull(*P);
   if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V))
     return visitGlobalAlias(*GA);
   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
     return visitGlobalVariable(*GV);
   if (UndefValue *UV = dyn_cast<UndefValue>(V))
     return visitUndefValue(*UV);
-  if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
-    if (CE->getOpcode() == Instruction::IntToPtr)
-      return unknown(); // clueless
-    if (CE->getOpcode() == Instruction::GetElementPtr)
-      return visitGEPOperator(cast<GEPOperator>(*CE));
-  }
 
   LLVM_DEBUG(dbgs() << "ObjectSizeOffsetVisitor::compute() unhandled value: "
                     << *V << '\n');
   return unknown();
 }
 
 bool ObjectSizeOffsetVisitor::CheckedZextOrTrunc(APInt &I) {
   return ::CheckedZextOrTrunc(I, IntTyBits);
 }

 }
 
 SizeOffsetType
 ObjectSizeOffsetVisitor::visitExtractValueInst(ExtractValueInst&) {
   // Easy cases were already folded by previous passes.
   return unknown();
 }
 
-SizeOffsetType ObjectSizeOffsetVisitor::visitGEPOperator(GEPOperator &GEP) {
-  SizeOffsetType PtrData = compute(GEP.getPointerOperand());
-  APInt Offset(DL.getIndexTypeSizeInBits(GEP.getPointerOperand()->getType()), 0);
-  if (!bothKnown(PtrData) || !GEP.accumulateConstantOffset(DL, Offset))
-    return unknown();
-
-  return std::make_pair(PtrData.first, PtrData.second + Offset);
-}
-
 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalAlias(GlobalAlias &GA) {
   if (GA.isInterposable())
     return unknown();
   return compute(GA.getAliasee());
 }
 
 SizeOffsetType ObjectSizeOffsetVisitor::visitGlobalVariable(GlobalVariable &GV){
   if (!GV.hasDefinitiveInitializer())

llvm/test/Transforms/InstCombine/builtin-dynamic-object-size.ll

 ; RUN: opt -passes=instcombine -S < %s | FileCheck %s
 
-target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128-p7:32:32"
 target triple = "x86_64-apple-macosx10.14.0"
 
 ; Function Attrs: nounwind ssp uwtable
 define i64 @weird_identity_but_ok(i64 %sz) {
 entry:
   %call = tail call i8* @malloc(i64 %sz)
   %calc_size = tail call i64 @llvm.objectsize.i64.p0i8(i8* %call, i1 false, i1 true, i1 true)
   tail call void @free(i8* %call)

 }
 
 ; CHECK:  define void @bdos_cmpm1_expr(
 ; CHECK:    [[TMP:%.*]] = icmp ne i64 [[SZ:%.*]], -1
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[TMP]])
 ; CHECK-NEXT:    br i1 false, label %if.else, label %if.then
 ; CHECK:    call void @fortified_chk(i8* %obj, i64 [[SZ]])
 
+@p7 = internal addrspace(7) global i8 0
+
+; Gracefully handle AS cast when the address spaces have different pointer widths.
+define i64 @as_cast(i1 %c) {
+; CHECK:  [[TMP0:%.*]] = select i1 %c, i64 64, i64 1
+; CHECK:  [[NOT:%.*]] = xor i1 %c, true
+; CHECK:  [[NEG:%.*]] = sext i1 [[NOT]] to i64
+; CHECK:  [[TMP1:%.*]] = add nsw i64 [[TMP0]], [[NEG]]
+; CHECK:  [[TMP2:%.*]] = icmp ne i64 [[TMP1]], -1
+; CHECK:  call void @llvm.assume(i1 [[TMP2]])
+; CHECK:  ret i64 [[TMP1]]
+;
+entry:
+  %p0 = tail call i8* @malloc(i64 64)
+  %gep = getelementptr i8, i8 addrspace(7)* @p7, i32 1
+  %as = addrspacecast i8 addrspace(7)* %gep to i8*
+  %select = select i1 %c, i8* %p0, i8* %as
+  %calc_size = tail call i64 @llvm.objectsize.i64.p0i8(i8* %select, i1 false, i1 true, i1 true)
+  ret i64 %calc_size
+}
+
+define i64 @constexpr_as_cast(i1 %c) {
+; CHECK:  [[TMP0:%.*]] = select i1 %c, i64 64, i64 1
+; CHECK:  [[NOT:%.*]] = xor i1 %c, true
+; CHECK:  [[NEG:%.*]] = sext i1 [[NOT]] to i64
+; CHECK:  [[TMP1:%.*]] = add nsw i64 [[TMP0]], [[NEG]]
+; CHECK:  [[TMP2:%.*]] = icmp ne i64 [[TMP1]], -1
+; CHECK:  call void @llvm.assume(i1 [[TMP2]])
+; CHECK:  ret i64 [[TMP1]]
+;
+entry:
+  %p0 = tail call i8* @malloc(i64 64)
+  %select = select i1 %c, i8* %p0, i8* addrspacecast (i8 addrspace(7)* getelementptr (i8, i8 addrspace(7)* @p7, i32 1) to i8*)
+  %calc_size = tail call i64 @llvm.objectsize.i64.p0i8(i8* %select, i1 false, i1 true, i1 true)
+  ret i64 %calc_size
+}
+
 declare void @bury(i32) local_unnamed_addr #2
 
 ; Function Attrs: nounwind allocsize(0)
 declare i8* @malloc(i64)
 
 declare i8* @get_unknown_buffer()
 
 ; Function Attrs: nounwind
 declare void @free(i8* nocapture)
 
 ; Function Attrs: nounwind readnone speculatable
 declare i64 @llvm.objectsize.i64.p0i8(i8*, i1, i1, i1)
 
 declare void @fortified_chk(i8*, i64)
 
 declare void @unfortified(i8*, i64)