[InstCombine] Retain TBAA when narrowing memory accesses

Summary:
As discussed on the mailing list it is legal to propagate TBAA to loads/stores
from/to smaller regions of a larger load tagged with TBAA. Do so for
(load->extractvalue)=>(gep->load) and similar foldings.

Reviewed By: sanjoy
Differential Revision: https://reviews.llvm.org/D31954

llvm-svn: 306615

Author: Keno

llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

@@ -661,6 +661,9 @@ static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
     if (NumElements == 1) {
       LoadInst *NewLoad = combineLoadToNewType(IC, LI, ST->getTypeAtIndex(0U),
                                                ".unpack");
+      AAMDNodes AAMD;
+      LI.getAAMetadata(AAMD);
+      NewLoad->setAAMetadata(AAMD);
       return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
         UndefValue::get(T), NewLoad, 0, Name));
     }
@@ -690,6 +693,10 @@ static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
                                                 Name + ".elt");
       auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
       auto *L = IC.Builder->CreateAlignedLoad(Ptr, EltAlign, Name + ".unpack");
+      // Propagate AA metadata. It'll still be valid on the narrowed load.
+      AAMDNodes AAMD;
+      LI.getAAMetadata(AAMD);
+      L->setAAMetadata(AAMD);
       V = IC.Builder->CreateInsertValue(V, L, i);
     }
 
@@ -702,6 +709,9 @@ static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
     auto NumElements = AT->getNumElements();
     if (NumElements == 1) {
       LoadInst *NewLoad = combineLoadToNewType(IC, LI, ET, ".unpack");
+      AAMDNodes AAMD;
+      LI.getAAMetadata(AAMD);
+      NewLoad->setAAMetadata(AAMD);
       return IC.replaceInstUsesWith(LI, IC.Builder->CreateInsertValue(
         UndefValue::get(T), NewLoad, 0, Name));
     }
@@ -734,6 +744,9 @@ static Instruction *unpackLoadToAggregate(InstCombiner &IC, LoadInst &LI) {
                                                 Name + ".elt");
       auto *L = IC.Builder->CreateAlignedLoad(Ptr, MinAlign(Align, Offset),
                                               Name + ".unpack");
+      AAMDNodes AAMD;
+      LI.getAAMetadata(AAMD);
+      L->setAAMetadata(AAMD);
       V = IC.Builder->CreateInsertValue(V, L, i);
       Offset += EltSize;
     }
@@ -1192,7 +1205,11 @@ static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
                                                 AddrName);
       auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
       auto EltAlign = MinAlign(Align, SL->getElementOffset(i));
-      IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
+      llvm::Instruction *NS =
+          IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
+      AAMDNodes AAMD;
+      SI.getAAMetadata(AAMD);
+      NS->setAAMetadata(AAMD);
     }
 
     return true;
@@ -1239,7 +1256,10 @@ static bool unpackStoreToAggregate(InstCombiner &IC, StoreInst &SI) {
                                                 AddrName);
       auto *Val = IC.Builder->CreateExtractValue(V, i, EltName);
       auto EltAlign = MinAlign(Align, Offset);
-      IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
+      Instruction *NS = IC.Builder->CreateAlignedStore(Val, Ptr, EltAlign);
+      AAMDNodes AAMD;
+      SI.getAAMetadata(AAMD);
+      NS->setAAMetadata(AAMD);
       Offset += EltSize;
     }
 

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -2425,9 +2425,15 @@ Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
       Builder->SetInsertPoint(L);
       Value *GEP = Builder->CreateInBoundsGEP(L->getType(),
                                               L->getPointerOperand(), Indices);
+      Instruction *NL = Builder->CreateLoad(GEP);
+      // Whatever aliasing information we had for the orignal load must also
+      // hold for the smaller load, so propagate the annotations.
+      AAMDNodes Nodes;
+      L->getAAMetadata(Nodes);
+      NL->setAAMetadata(Nodes);
       // Returning the load directly will cause the main loop to insert it in
       // the wrong spot, so use replaceInstUsesWith().
-      return replaceInstUsesWith(EV, Builder->CreateLoad(GEP));
+      return replaceInstUsesWith(EV, NL);
     }
   // We could simplify extracts from other values. Note that nested extracts may
   // already be simplified implicitly by the above: extract (extract (insert) )

llvm/test/Transforms/InstCombine/extractinsert-tbaa.ll

@@ -0,0 +1,45 @@
+; RUN: opt -S -instcombine %s -o - | FileCheck %s
+
+%Complex = type { double, double }
+
+; Check that instcombine preserves TBAA when narrowing loads
+define double @teststructextract(%Complex *%val) {
+; CHECK: load double, {{.*}}, !tbaa
+; CHECK-NOT: load %Complex
+    %loaded = load %Complex, %Complex *%val, !tbaa !1
+    %real = extractvalue %Complex %loaded, 0
+    ret double %real
+}
+
+define double @testarrayextract([2 x double] *%val) {
+; CHECK: load double, {{.*}}, !tbaa
+; CHECK-NOT: load [2 x double]
+    %loaded = load [2 x double], [2 x double] *%val, !tbaa !1
+    %real = extractvalue [2 x double] %loaded, 0
+    ret double %real
+}
+
+; Check that inscombine preserves TBAA when breaking up stores
+define void @teststructinsert(%Complex *%loc, double %a, double %b) {
+; CHECK: store double %a, {{.*}}, !tbaa
+; CHECK: store double %b, {{.*}}, !tbaa
+; CHECK-NOT: store %Complex
+    %inserted  = insertvalue %Complex undef,      double %a, 0
+    %inserted2 = insertvalue %Complex %inserted,  double %b, 1
+    store %Complex %inserted2, %Complex *%loc, !tbaa !1
+    ret void
+}
+
+define void @testarrayinsert([2 x double] *%loc, double %a, double %b) {
+; CHECK: store double %a, {{.*}}, !tbaa
+; CHECK: store double %b, {{.*}}, !tbaa
+; CHECK-NOT: store [2 x double]
+    %inserted  = insertvalue [2 x double] undef,      double %a, 0
+    %inserted2 = insertvalue [2 x double] %inserted,  double %b, 1
+    store [2 x double] %inserted2, [2 x double] *%loc, !tbaa !1
+    ret void
+}
+
+!0 = !{!"tbaa_root"}
+!1 = !{!2, !2, i64 0}
+!2 = !{!"Complex", !0, i64 0}