[Extension] Optimizing const member objects.

[lvoufo]

Summary

Continued from D13606...

[lvoufo]

Updated from D14003 (D14008 + D14009), D14004, D14005, D14006 and D14007.

[lvoufo]

Update documentation of invariant_start intrinsic call.

[lvoufo]

Fix diff (wrong one submitted last time).

include/llvm/IR/Intrinsics.td

 //
 def int_lifetime_start  : Intrinsic<[],
                                     [llvm_i64_ty, llvm_ptr_ty],
                                     [IntrReadWriteArgMem, NoCapture<1>]>;
 def int_lifetime_end    : Intrinsic<[],
                                     [llvm_i64_ty, llvm_ptr_ty],
                                     [IntrReadWriteArgMem, NoCapture<1>]>;
 def int_invariant_start : Intrinsic<[llvm_descriptor_ty],
-                                    [llvm_i64_ty, llvm_ptr_ty],
+                                    [llvm_i64_ty, llvm_ptr_ty, llvm_vararg_ty],
                                     [IntrReadWriteArgMem, NoCapture<1>]>;
 def int_invariant_end   : Intrinsic<[],
                                     [llvm_descriptor_ty, llvm_i64_ty,
                                      llvm_ptr_ty],
                                     [IntrReadWriteArgMem, NoCapture<2>]>;
 
 def int_invariant_group_barrier : Intrinsic<[llvm_ptr_ty], 
                                             [llvm_ptr_ty], 

lib/Analysis/BasicAliasAnalysis.cpp

/*static*/ const Value *BasicAAResult::DecomposeGEPExpression(
   } while (--MaxLookup);
 
   // If the chain of expressions is too deep, just return early.
   MaxLookupReached = true;
   SearchLimitReached++;
   return V;
 }
 
+/// Checks that a given intrinsic_start call writeonce-covers a given
+/// memory location pointer.
+static bool checkCoverage(IntrinsicInst *II,
+                          const llvm::Value *LocPtr, const DataLayout &DL) {
+  if (!II) return false;
+
+  assert(II->getIntrinsicID() == Intrinsic::invariant_start &&
+         "Given intrinsic instruction must be invariant_start");
+
+  // If the instruction has no offset, then LocPtr is pointing to constant
+  // (writeonce) memory, and thus is either a global variable, an alloca
+  // instruction, or a GEP instruction accessing either one of those.
+  if (II->getNumArgOperands() <= 2)
+    return true;
+
+  unsigned GEPOffsetBits = 0;
+  bool GEPOffsetChanged = false;
+  APInt GEPOffset;
+
+  // Otherwise, LocPtr itself is not pointing to constant memory, but
+  // sub-locations at offsets are constant. In addition, LocPtr must be
+  // a GEP instruction.
+  if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LocPtr)) {
+    GEPOffsetBits = DL.getPointerTypeSizeInBits(GEP->getType());
+    GEPOffset = APInt(GEPOffsetBits, 0);
+    GEPOffsetChanged = GEP->accumulateConstantOffset(DL, GEPOffset);
+  } else if (const GEPOperator *GEP = dyn_cast<GEPOperator>(LocPtr)) {
+    GEPOffsetBits = DL.getPointerTypeSizeInBits(GEP->getType());
+    GEPOffset = APInt(GEPOffsetBits, 0);
+    GEPOffsetChanged = GEP->accumulateConstantOffset(DL, GEPOffset);
+  }
+  if (GEPOffsetChanged) {
+    GEPOffset *= APInt(GEPOffsetBits, 8);
+    for (unsigned i = 2; i < II->getNumArgOperands(); i += 2) {
+      llvm::ConstantInt *FESize =
+          dyn_cast_or_null<llvm::ConstantInt>(II->getArgOperand(i));
+      llvm::ConstantInt *FEOffset =
+          dyn_cast_or_null<llvm::ConstantInt>(II->getArgOperand(i+1));
+      assert(FEOffset && FESize &&
+             "Frontend must store these offsets (and corresponding sizes) "
+             "in bits as ConstantInts.");
+      const APInt &Offset = FEOffset->getValue();
+
+      if (Offset.ule(GEPOffset) &&
+          GEPOffset.ule(Offset + FESize->getZExtValue()))
+        return true;
+    }
+  }
+
+  // If the GEP offset has not changed, then LocPtr must be pointing to
+  // the begining of allocated memory, which is not writeonce.
+  return false;
+}
+
 /// Returns whether the given pointer value points to memory that is local to
 /// the function, with global constants being considered local to all
 /// functions.
 bool BasicAAResult::pointsToConstantMemory(const MemoryLocation &Loc,
                                            bool OrLocal) {
   assert(Visited.empty() && "Visited must be cleared after use!");
 
   unsigned MaxLookup = 8;
   SmallVector<const Value *, 16> Worklist;
   Worklist.push_back(Loc.Ptr);
   do {
-    const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
+    //const Value *V = GetUnderlyingObject(Worklist.pop_back_val(), DL);
+    auto LocPtr = Worklist.pop_back_val();
+    const Value *V = GetUnderlyingObject(LocPtr, DL);
     if (!Visited.insert(V).second) {
       Visited.clear();
       return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
     }
 
     // An alloca instruction defines local memory.
     if (OrLocal && isa<AllocaInst>(V))
       continue;
 
     if (const AllocaInst *AI = dyn_cast<AllocaInst>(V))
-      if (AI->getInvariantStartInstruction()) continue;
+      if (checkCoverage(AI->getInvariantStartInstruction(), LocPtr, DL))
+        continue;
 
     // A global constant counts as local memory for our purposes.
     if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
       // Note: this doesn't require GV to be "ODR" because it isn't legal for a
       // global to be marked constant in some modules and non-constant in
       // others.  GV may even be a declaration, not a definition.
-      if (!GV->isConstant() && !GV->getInvariantStartInstruction()) {
+      if (!GV->isConstant()  &&
+          !checkCoverage(GV->getInvariantStartInstruction(), LocPtr, DL)) {
         Visited.clear();
         return AAResultBase::pointsToConstantMemory(Loc, OrLocal);
       }
       continue;
     }
 
     // If both select values point to local memory, then so does the select.
     if (const SelectInst *SI = dyn_cast<SelectInst>(V)) {

lib/Analysis/MemoryDependenceAnalysis.cpp

MemDepResult MemoryDependenceAnalysis::getSimplePointerDependencyFrom(
   // position between two instructions in a BB and can be used by
   // AliasAnalysis::callCapturesBefore.
   OrderedBasicBlock OBB(BB);
 
   // Walk backwards through the basic block, looking for dependencies.
   while (ScanIt != BB->begin()) {
     Instruction *Inst = --ScanIt;
 
-    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
+    if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
       // Debug intrinsics don't (and can't) cause dependencies.
       if (isa<DbgInfoIntrinsic>(II)) continue;
 
     // Same for invariant intrinsics.
     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
       if (II->getIntrinsicID() == Intrinsic::invariant_start) {
         if (II == Preserved.II)
           // We did not skip any instruction earlier. So, we must express that
           // the given load is no longer pointing to constant memory.
           llvm::setInvariantStartInstruction(Preserved.LoadedAddr, nullptr);
         continue;
       } else if (II->getIntrinsicID() == Intrinsic::invariant_end)
         continue;
     }
 
+    // Same for invariant intrinsics.
+      if (II->getIntrinsicID() == Intrinsic::invariant_start) {
+        if (II == Preserved.II)
+          // We did not skip any instruction earlier. So, we must express that
+          // the given load is no longer pointing to constant memory.
+          llvm::setInvariantStartInstruction(Preserved.LoadedAddr, nullptr);
+        continue;
+      } else if (II->getIntrinsicID() == Intrinsic::invariant_end)
+        continue;
+    }
+
     // Limit the amount of scanning we do so we don't end up with quadratic
     // running time on extreme testcases.
     --Limit;
     if (!Limit)
       return MemDepResult::getUnknown();
 
     if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
       // If we reach a lifetime begin or end marker, then the query ends here
#endif
     // a subsequent bitcast of the malloc call result.  There can be stores to
     // the malloced memory between the malloc call and its bitcast uses, and we
     // need to continue scanning until the malloc call.
     if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
       const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
 
       if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
         return MemDepResult::getDef(Inst);
+
       if (isInvariantLoad)
         continue;
       // Be conservative if the accessed pointer may alias the allocation.
       if (AA->alias(Inst, AccessPtr) != NoAlias)
         return MemDepResult::getClobber(Inst);
       // If the allocation is not aliased and does not read memory (like
       // strdup), it is safe to ignore.
       if (isa<AllocaInst>(Inst) ||

test/Feature/memorymarkers.ll

 ; RUN: llvm-as -disable-output < %s
 
 %"struct.std::pair<int,int>" = type { i32, i32 }
 
 declare void @_Z3barRKi(i32*)
 
 declare void @llvm.lifetime.start(i64, i8* nocapture) nounwind
 declare void @llvm.lifetime.end(i64, i8* nocapture) nounwind
-declare {}* @llvm.invariant.start(i64, i8* nocapture) readonly nounwind
+declare {}* @llvm.invariant.start(i64, i8* nocapture, ...) readonly nounwind
 declare void @llvm.invariant.end({}*, i64, i8* nocapture) nounwind
 
 define i32 @_Z4foo2v() nounwind {
 entry:
   %x = alloca %"struct.std::pair<int,int>"
   %y = bitcast %"struct.std::pair<int,int>"* %x to i8*
 
   ;; Constructor starts here (this isn't needed since it is immediately
   ;; preceded by an alloca, but shown for completeness).
   call void @llvm.lifetime.start(i64 8, i8* %y)
 
   %0 = getelementptr %"struct.std::pair<int,int>", %"struct.std::pair<int,int>"* %x, i32 0, i32 0
   store i32 4, i32* %0, align 8
   %1 = getelementptr %"struct.std::pair<int,int>", %"struct.std::pair<int,int>"* %x, i32 0, i32 1
   store i32 5, i32* %1, align 4
 
   ;; Constructor has finished here.
-  %inv = call {}* @llvm.invariant.start(i64 8, i8* %y)
+  %inv = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 8, i8* %y)
   call void @_Z3barRKi(i32* %0) nounwind
   %2 = load i32, i32* %0, align 8
 
   ;; Destructor is run here.
   call void @llvm.invariant.end({}* %inv, i64 8, i8* %y)
   ;; Destructor is done here.
   call void @llvm.lifetime.end(i64 8, i8* %y)
   ret i32 %2
 }

test/Transforms/GlobalOpt/invariant-nodatalayout.ll

 ; RUN: opt -globalopt -S -o - < %s | FileCheck %s
 ; The check here is that it doesn't crash.
 
-declare {}* @llvm.invariant.start(i64 %size, i8* nocapture %ptr)
+declare {}* @llvm.invariant.start(i64 %size, i8* nocapture %ptr, ...)
 
 @object1 = global { i32, i32 } zeroinitializer
 ; CHECK: @object1 = global { i32, i32 } zeroinitializer
 
 define void @ctor1() {
   %ptr = bitcast {i32, i32}* @object1 to i8*
-  call {}* @llvm.invariant.start(i64 4, i8* %ptr)
+  call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %ptr)
   ret void
 }
 
 @llvm.global_ctors = appending constant
   [1 x { i32, void ()* }]
   [ { i32, void ()* } { i32 65535, void ()* @ctor1 } ]

test/Transforms/GlobalOpt/invariant.ll

 ; RUN: opt -globalopt -S -o - < %s | FileCheck %s
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
-declare {}* @llvm.invariant.start(i64 %size, i8* nocapture %ptr)
+declare {}* @llvm.invariant.start(i64 %size, i8* nocapture %ptr, ...)
 
 define void @test1(i8* %ptr) {
-  call {}* @llvm.invariant.start(i64 4, i8* %ptr)
+  call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %ptr)
   ret void
 }
 
 @object1 = global i32 0
 ; CHECK: @object1 = constant i32 -1
 define void @ctor1() {
   store i32 -1, i32* @object1
   %A = bitcast i32* @object1 to i8*
   call void @test1(i8* %A)
   ret void
 }
 
 
 @object2 = global i32 0
 ; CHECK: @object2 = global i32 0
 define void @ctor2() {
   store i32 -1, i32* @object2
   %A = bitcast i32* @object2 to i8*
-  %B = call {}* @llvm.invariant.start(i64 4, i8* %A)
+  %B = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %A)
   %C = bitcast {}* %B to i8*
   ret void
 }
 
 
 @object3 = global i32 0
 ; CHECK: @object3 = global i32 -1
 define void @ctor3() {
   store i32 -1, i32* @object3
   %A = bitcast i32* @object3 to i8*
-  call {}* @llvm.invariant.start(i64 3, i8* %A)
+  call {}* (i64, i8*, ...) @llvm.invariant.start(i64 3, i8* %A)
   ret void
 }
 
 
 @object4 = global i32 0
 ; CHECK: @object4 = global i32 -1
 define void @ctor4() {
   store i32 -1, i32* @object4
   %A = bitcast i32* @object4 to i8*
-  call {}* @llvm.invariant.start(i64 -1, i8* %A)
+  call {}* (i64, i8*, ...) @llvm.invariant.start(i64 -1, i8* %A)
   ret void
 }
 
 
 @llvm.global_ctors = appending constant
   [4 x { i32, void ()* }]
   [ { i32, void ()* } { i32 65535, void ()* @ctor1 },
     { i32, void ()* } { i32 65535, void ()* @ctor2 },
     { i32, void ()* } { i32 65535, void ()* @ctor3 },
     { i32, void ()* } { i32 65535, void ()* @ctor4 } ]

test/Transforms/InstCombine/invariant.ll

 ; Test to make sure unused llvm.invariant.start calls are not trivially eliminated
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 declare void @g(i8*)
 
-declare {}* @llvm.invariant.start(i64, i8* nocapture) nounwind readonly
+declare {}* @llvm.invariant.start(i64, i8* nocapture, ...) nounwind readonly
 
 define i8 @f() {
   %a = alloca i8                                  ; <i8*> [#uses=4]
   store i8 0, i8* %a
-  %i = call {}* @llvm.invariant.start(i64 1, i8* %a) ; <{}*> [#uses=0]
-  ; CHECK: call {}* @llvm.invariant.start
+  %i = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 1, i8* %a) ; <{}*> [#uses=0]
+  ; CHECK: call {}* (i64, i8*, ...) @llvm.invariant.start
   call void @g(i8* %a)
   %r = load i8, i8* %a                                ; <i8> [#uses=1]
   ret i8 %r
 }

test/Transforms/LoadElim/global-local-vars.ll

entry:
   %j = alloca %struct.A
   %agg.tmp = alloca %struct.A
   %agg.tmp1 = alloca %struct.A
   %0 = bitcast %struct.A* %j to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0)
   %1 = bitcast %struct.A* %j to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)
   %2 = bitcast %struct.A* %j to i8*
-  %3 = call {}* @llvm.invariant.start(i64 4, i8* %2)
+  %3 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %2)
   ; CHECK-NOT: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %4 = bitcast %struct.A* %agg.tmp to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)
   %coerce.dive = getelementptr inbounds %struct.A, %struct.A* %agg.tmp, i32 0, i32 0
   %5 = load i32, i32* %coerce.dive
   ; CHECK: load i32, i32*
   call void @_Z3bar1A(i32 %5)
   call void @_Z3fooPK1A(%struct.A* @_ZL1i)
entry:
   %0 = bitcast %struct.A* %j to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0)
   %1 = bitcast %struct.A* %j to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %1, i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)
   %2 = bitcast %struct.A* %j to i8*
   ; CHECK-4-5A1-5B: load i32, i32*
   ; CHECK-5A2-5B: load i32, i32*
   ; CHECK-ALL: load i32, i32*
-  %3 = call {}* @llvm.invariant.start(i64 4, i8* %2)
+  %3 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %2)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %4 = bitcast %struct.A* %agg.tmp to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)
   %coerce.dive = getelementptr inbounds %struct.A, %struct.A* %agg.tmp, i32 0, i32 0
   %5 = load i32, i32* %coerce.dive
   ; CHECK-5B: load i32, i32*
   call void @_Z3bar1A(i32 %5)
   %6 = bitcast %struct.A* %agg.tmp1 to i8*
entry:
   ret void
 }
 
 define internal void @__cxx_global_var_init() {
 entry:
   %call = call i32 @_Z3onev()
   call void @_ZN1AC1Ei(%struct.A* @_ZL1i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
-  %0 = call {}* @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1i to i8*))
+  %0 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1i to i8*))
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   ret void
 }
 
 declare i32 @_Z3onev()
 declare void @_Z3bar1A(i32)
 declare void @_Z3fooPK1A(%struct.A*)
 declare void @_Z4foo2PK1AS1_(%struct.A*, %struct.A*)
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32, i1)
-declare {}* @llvm.invariant.start(i64, i8* nocapture)
+declare {}* @llvm.invariant.start(i64, i8* nocapture, ...)
 declare void @llvm.invariant.end({}*, i64, i8* nocapture)
 declare void @llvm.lifetime.start(i64, i8* nocapture)
 declare void @llvm.lifetime.end(i64, i8* nocapture)
 
 define internal void @_GLOBAL__sub_I_global_local() {
 entry:
   call void @__cxx_global_var_init()
   ret void

test/Transforms/LoadElim/global-vars.ll

 
 ;; * When the load scan limit is 8,
 ;;   '-functionattrs -tailcallelim -instcombine' may be as good as '-gvn'.
 ;;   But the same can't be said when the limit is 6.
 ; RUN: opt < %s -globalopt -available-load-scan-limit=8 -instcombine -functionattrs -tailcallelim -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-ALL
 ; RUN: opt < %s -globalopt -available-load-scan-limit=6 -instcombine -functionattrs -tailcallelim -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=CHECK-4-5A-5B1 --check-prefix=CHECKLOAD-5B2
 
 %struct.A = type { i32 }
+%struct.B = type { i32, %struct.P }
+%struct.P = type { i32 }
 
 @_ZL1i = internal global %struct.A zeroinitializer
+@i = global %struct.B zeroinitializer
 @k = global %struct.A zeroinitializer
 @_ZL1j = internal global %struct.A zeroinitializer
 @llvm.global_ctors = appending global [1 x { i32, void ()*, i8* }] [{ i32, void ()*, i8* } { i32 65535, void ()* @_GLOBAL__sub_I_global, i8* null }]
 
 ;; Example 1: Duplicate loads.
 ;; void ex1() {
 ;;   const Type i(one());
 ;;   bar(i);            // First load.
entry:
   %7 = load i32, i32* %coerce.dive6
   ; CHECKLOAD-5B2: load i32, i32*
   ; CHECK-5A2-5B-NOT: load i32, i32*
   ; CHECK-ALL-NOT: load i32, i32*
   call void @_Z3bar1A(i32 %7)
   ret void
 }
 
+;; Example 1 with member objects.
+;; void ex1m() {
+;;  A i(one());
+;;  bar(i.b);          // First load.
+;;  foo(&i.b);
+;;  bar(i.b);          // No load.
+define void @_Z4ex1mv() {
+entry:
+  %agg.tmp = alloca %struct.P
+  %agg.tmp1 = alloca %struct.P
+  %0 = bitcast %struct.P* %agg.tmp to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %0, i8* bitcast (%struct.P* getelementptr inbounds (%struct.B, %struct.B* @i, i32 0, i32 1) to i8*), i64 4, i32 4, i1 false)
+  %coerce.dive = getelementptr inbounds %struct.P, %struct.P* %agg.tmp, i32 0, i32 0
+  %1 = load i32, i32* %coerce.dive
+  ; CHECK: load i32, i32*
+  call void @_Z3bar1P(i32 %1)
+  call void @_Z3fooPK1P(%struct.P* getelementptr inbounds (%struct.B, %struct.B* @i, i32 0, i32 1))
+  %2 = bitcast %struct.P* %agg.tmp1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %2, i8* bitcast (%struct.P* getelementptr inbounds (%struct.B, %struct.B* @i, i32 0, i32 1) to i8*), i64 4, i32 4, i1 false)
+  %coerce.dive2 = getelementptr inbounds %struct.P, %struct.P* %agg.tmp1, i32 0, i32 0
+  %3 = load i32, i32* %coerce.dive2
+  ; CHECK-NOT: load i32, i32*
+  call void @_Z3bar1P(i32 %3)
+  ret void
+}
+
+
 define internal void @__cxx_global_var_init() {
 entry:
   %call = call i32 @_Z3onev()
   call void @_ZN1AC1Ei(%struct.A* @_ZL1i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
-  %0 = call {}* @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1i to i8*))
+  %0 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1i to i8*))
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   ret void
 }
 
+define internal void @__cxx_global_var_init.3() {
+entry:
+  %call = call i32 @_Z3onev()
+  call void @_ZN1BC1Ei(%struct.B* @i, i32 %call)
+  %0 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 8, i8* bitcast (%struct.B* @i to i8*), i64 8, i64 32)
+  ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8* {{.*}},
+  ret void
+}
+
 define internal void @__cxx_global_var_init.2() {
 entry:
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* bitcast (%struct.A* @_ZL1j to i8*), i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)
-  %0 = call {}* @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1j to i8*))
+  %0 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* bitcast (%struct.A* @_ZL1j to i8*))
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   ret void
 }
 
+
 declare i32 @_Z3onev()
 declare void @_Z3bar1A(i32)
 declare void @_Z3fooPK1A(%struct.A*)
 declare void @_Z4foo2PK1AS1_(%struct.A*, %struct.A*)
+declare void @_Z3bar1P(i32)
+declare void @_Z3fooPK1P(%struct.P*)
+declare void @_ZN1BC1Ei(%struct.B*, i32)
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32, i1)
-declare {}* @llvm.invariant.start(i64, i8* nocapture)
+declare {}* @llvm.invariant.start(i64, i8* nocapture, ...)
 declare void @llvm.invariant.end({}*, i64, i8* nocapture)
 declare void @llvm.lifetime.start(i64, i8* nocapture)
 declare void @llvm.lifetime.end(i64, i8* nocapture)
 
 define internal void @_GLOBAL__sub_I_global() {
 entry:
   call void @__cxx_global_var_init()
+  call void @__cxx_global_var_init.3()
   call void @__cxx_global_var_init.2()
   call void @__cxx_global_var_init.1()
   ret void
 }
 
 define internal void @__cxx_global_var_init.1() {
 entry:
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* bitcast (%struct.A* @k to i8*), i8* bitcast (%struct.A* @_ZL1i to i8*), i64 4, i32 4, i1 false)

test/Transforms/LoadElim/local-vars.ll

 ;; * When the load scan limit is 8,
 ;;   '-functionattrs -tailcallelim -instcombine' may be as good as '-gvn'.
 ;;   But the same can't be said when the limit is 6.
 ; RUN: opt < %s -available-load-scan-limit=8 -instcombine -functionattrs -tailcallelim -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=CHECKNOINL --check-prefix=CHECK-ALL
 ; RUN: opt < %s -available-load-scan-limit=6 -instcombine -functionattrs -tailcallelim -instcombine -S | FileCheck %s --check-prefix=CHECK --check-prefix=CHECKNOINL --check-prefix=CHECK-5A-5B1 --check-prefix=CHECKLOAD-4-5B2
 
 
 %struct.A = type { i32 }
+%struct.B = type { i32, %struct.P }
+%struct.P = type { i32 }
 
 ;; Example 1: Duplicate loads.
 ;; void ex1() {
 ;;   const Type i(one());
 ;;   bar(i);            // First load.
 ;;   foo(&i);           // Does not change i.
 ;;   bar(i);            // No load.
 ;; }
 define void @_Z3ex1v() {
 entry:
   %i = alloca %struct.A
   %agg.tmp = alloca %struct.A
   %agg.tmp1 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %agg.tmp to i8*
   %4 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %4, i64 4, i32 4, i1 false)
   %coerce.dive = getelementptr inbounds %struct.A, %struct.A* %agg.tmp, i32 0, i32 0
   %5 = load i32, i32* %coerce.dive
   ; CHECKNOINL: load i32, i32*
   ; CHECKINL-NOT: load i32, i32*
entry:
   %agg.tmp = alloca %struct.A
   %agg.tmp1 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %j to i8*
   call void @llvm.lifetime.start(i64 4, i8* %3) 
   %4 = bitcast %struct.A* %j to i8*
   %5 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %5, i64 4, i32 4, i1 false)
   %6 = bitcast %struct.A* %j to i8*
-  %7 = call {}* @llvm.invariant.start(i64 4, i8* %6)
+  %7 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %6)
   ; CHECK-NOT: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %8 = bitcast %struct.A* %agg.tmp to i8*
   %9 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %8, i8* %9, i64 4, i32 4, i1 false)
   %coerce.dive = getelementptr inbounds %struct.A, %struct.A* %agg.tmp, i32 0, i32 0
   %10 = load i32, i32* %coerce.dive
   ; CHECKNOINL: load i32, i32*
   ; CHECKINL-NOT: load i32, i32*
entry:
   %agg.tmp = alloca %struct.A
   %agg.tmp1 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECKNOINL: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   ; CHECKINV3: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   ; CHECKINV-NOT: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %k to i8*
   call void @llvm.lifetime.start(i64 4, i8* %3) 
   %4 = bitcast %struct.A* %k to i8*
   %5 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %5, i64 4, i32 4, i1 false)
entry:
   %agg.tmp1 = alloca %struct.A
   %agg.tmp3 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %k to i8*
   call void @llvm.lifetime.start(i64 4, i8* %3) 
   %4 = bitcast %struct.A* %k to i8*
   %5 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %5, i64 4, i32 4, i1 false)
   %6 = bitcast %struct.A* %agg.tmp to i8*
   %7 = bitcast %struct.A* %i to i8*
entry:
   %agg.tmp3 = alloca %struct.A
   %agg.tmp5 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %k to i8*
   call void @llvm.lifetime.start(i64 4, i8* %3) 
   %4 = bitcast %struct.A* %k to i8*
   %5 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %5, i64 4, i32 4, i1 false)
   %6 = bitcast %struct.A* %agg.tmp to i8*
   %7 = bitcast %struct.A* %i to i8*
entry:
   %agg.tmp3 = alloca %struct.A
   %agg.tmp5 = alloca %struct.A
   %0 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.start(i64 4, i8* %0) 
   %call = call i32 @_Z3onev()
   call void @_ZN1AC2Ei(%struct.A* %i, i32 %call)
   ; CHECKINL: store i32 {{.*}}, i32*
   %1 = bitcast %struct.A* %i to i8*
-  %2 = call {}* @llvm.invariant.start(i64 4, i8* %1)
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %3 = bitcast %struct.A* %j to i8*
   call void @llvm.lifetime.start(i64 4, i8* %3) 
   %4 = bitcast %struct.A* %j to i8*
   %5 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %5, i64 4, i32 4, i1 false)
   %6 = bitcast %struct.A* %j to i8*
-  %7 = call {}* @llvm.invariant.start(i64 4, i8* %6)
+  %7 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 4, i8* %6)
   ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8*
   %8 = bitcast %struct.A* %agg.tmp to i8*
   %9 = bitcast %struct.A* %i to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %8, i8* %9, i64 4, i32 4, i1 false)
   %coerce.dive = getelementptr inbounds %struct.A, %struct.A* %agg.tmp, i32 0, i32 0
   %10 = load i32, i32* %coerce.dive
   ; CHECKNOINL-ALL: load i32, i32*
   ; CHECKINL-NOT: load i32, i32*
entry:
   call void @llvm.lifetime.end(i64 4, i8* %20) 
   call void @llvm.invariant.end({}* %2, i64 4, i8* %1)
   ; CHECK: call {{.*}}@llvm.invariant.end({{.*}}, i64 {{[0-9]+}}, i8*
   %21 = bitcast %struct.A* %i to i8*
   call void @llvm.lifetime.end(i64 4, i8* %21) 
   ret void
 }
 
+;; Example 1 with member objects.
+;; void ex1m() {
+;;  A i(one());
+;;  bar(i.b);          // First load.
+;;  foo(&i.b);
+;;  bar(i.b);          // No load.
+define void @_Z4ex1mv() {
+entry:
+  %i = alloca %struct.B
+  %agg.tmp = alloca %struct.P
+  %agg.tmp2 = alloca %struct.P
+  %0 = bitcast %struct.B* %i to i8*
+  call void @llvm.lifetime.start(i64 8, i8* %0)
+  %call = call i32 @_Z3onev()
+  call void @_ZN1BC1Ei(%struct.B* %i, i32 %call)
+  %1 = bitcast %struct.B* %i to i8*
+  %2 = call {}* (i64, i8*, ...) @llvm.invariant.start(i64 8, i8* %1, i64 8, i64 32)
+  ; CHECK: call {{.*}}@llvm.invariant.start(i64 {{[0-9]+}}, i8* {{.*}},
+  %b = getelementptr inbounds %struct.B, %struct.B* %i, i32 0, i32 1
+  %3 = bitcast %struct.P* %agg.tmp to i8*
+  %4 = bitcast %struct.P* %b to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %3, i8* %4, i64 4, i32 4, i1 false)
+  %coerce.dive = getelementptr inbounds %struct.P, %struct.P* %agg.tmp, i32 0, i32 0
+  %5 = load i32, i32* %coerce.dive
+  ; CHECKNOINL: load i32, i32*
+  ; CHECKINL: load i32, i32*
+  call void @_Z3bar1P(i32 %5)
+  %b1 = getelementptr inbounds %struct.B, %struct.B* %i, i32 0, i32 1
+  call void @_Z3fooPK1P(%struct.P* %b1)
+  %b3 = getelementptr inbounds %struct.B, %struct.B* %i, i32 0, i32 1
+  %6 = bitcast %struct.P* %agg.tmp2 to i8*
+  %7 = bitcast %struct.P* %b3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %6, i8* %7, i64 4, i32 4, i1 false)
+  %coerce.dive4 = getelementptr inbounds %struct.P, %struct.P* %agg.tmp2, i32 0, i32 0
+  %8 = load i32, i32* %coerce.dive4
+  ; CHECK-NOT: load i32, i32*
+  call void @_Z3bar1P(i32 %8)
+  call void @llvm.invariant.end({}* %2, i64 8, i8* %1)
+  ; CHECK: call {{.*}}@llvm.invariant.end({{.*}}, i64 {{[0-9]+}}, i8*
+  %9 = bitcast %struct.B* %i to i8*
+  call void @llvm.lifetime.end(i64 8, i8* %9)
+  ret void
+}
+
 declare i32 @_Z3onev()
 declare void @_Z3bar1A(i32)
 declare void @_Z3fooPK1A(%struct.A*)
 declare void @_Z4foo2PK1AS1_(%struct.A*, %struct.A*)
+declare void @_Z3bar1P(i32)
+declare void @_Z3fooPK1P(%struct.P*)
+declare void @_ZN1BC1Ei(%struct.B*, i32)
 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture readonly, i64, i32, i1)
-declare {}* @llvm.invariant.start(i64, i8* nocapture)
+declare {}* @llvm.invariant.start(i64, i8* nocapture, ...)
 declare void @llvm.invariant.end({}*, i64, i8* nocapture)
 declare void @llvm.lifetime.start(i64, i8* nocapture)
 declare void @llvm.lifetime.end(i64, i8* nocapture)
 
 define linkonce_odr void @_ZN1AC2Ei(%struct.A* %this, i32 %a) unnamed_addr {
 entry:
   %this.addr = alloca %struct.A*
   %a.addr = alloca i32
   store %struct.A* %this, %struct.A** %this.addr
   store i32 %a, i32* %a.addr
   %this1 = load %struct.A*, %struct.A** %this.addr
   %a2 = getelementptr inbounds %struct.A, %struct.A* %this1, i32 0, i32 0
   %0 = load i32, i32* %a.addr
   store i32 %0, i32* %a2
   ret void
 }