[llvm] ba84cfb - [BasicAA] Use nuw attribute of GEPs (#98608)

via llvm-commits llvm-commits at lists.llvm.org
Tue Aug 20 03:00:08 PDT 2024


Author: Hari Limaye
Date: 2024-08-20T11:00:03+01:00
New Revision: ba84cfbe0ce2c697702c377b32e4dbf50dd1e114

URL: https://github.com/llvm/llvm-project/commit/ba84cfbe0ce2c697702c377b32e4dbf50dd1e114
DIFF: https://github.com/llvm/llvm-project/commit/ba84cfbe0ce2c697702c377b32e4dbf50dd1e114.diff

LOG: [BasicAA] Use nuw attribute of GEPs (#98608)

Use the nuw attribute of GEPs to prove that pointers do not alias, in
cases matching the following:

   +                +                     +
   | BaseOffset     |   +<nuw> Indices    |
   ---------------->|-------------------->|
   |-->V2Size       |                     |-------> V1Size
  LHS                                    RHS

If the difference between pointers is Offset +<nuw> Indices then we know
that the addition does not wrap the pointer index type (add nuw) and the
constant Offset is a lower bound on the distance between the pointers. We
can then prove NoAlias via Offset u>= V2Size.

Added: 
    llvm/test/Analysis/BasicAA/gep-nuw-alias.ll

Modified: 
    llvm/lib/Analysis/BasicAliasAnalysis.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Analysis/BasicAliasAnalysis.cpp b/llvm/lib/Analysis/BasicAliasAnalysis.cpp
index 6eac30f19d7851..72db28929c0c37 100644
--- a/llvm/lib/Analysis/BasicAliasAnalysis.cpp
+++ b/llvm/lib/Analysis/BasicAliasAnalysis.cpp
@@ -555,17 +555,17 @@ struct BasicAAResult::DecomposedGEP {
   APInt Offset;
   // Scaled variable (non-constant) indices.
   SmallVector<VariableGEPIndex, 4> VarIndices;
-  // Are all operations inbounds GEPs or non-indexing operations?
-  // (std::nullopt iff expression doesn't involve any geps)
-  std::optional<bool> InBounds;
+  // Nowrap flags common to all GEP operations involved in expression.
+  GEPNoWrapFlags NWFlags = GEPNoWrapFlags::all();
 
   void dump() const {
     print(dbgs());
     dbgs() << "\n";
   }
   void print(raw_ostream &OS) const {
-    OS << "(DecomposedGEP Base=" << Base->getName()
-       << ", Offset=" << Offset
+    OS << ", inbounds=" << (NWFlags.isInBounds() ? "1" : "0")
+       << ", nuw=" << (NWFlags.hasNoUnsignedWrap() ? "1" : "0")
+       << "(DecomposedGEP Base=" << Base->getName() << ", Offset=" << Offset
        << ", VarIndices=[";
     for (size_t i = 0; i < VarIndices.size(); i++) {
       if (i != 0)
@@ -644,12 +644,8 @@ BasicAAResult::DecomposeGEPExpression(const Value *V, const DataLayout &DL,
       return Decomposed;
     }
 
-    // Track whether we've seen at least one in bounds gep, and if so, whether
-    // all geps parsed were in bounds.
-    if (Decomposed.InBounds == std::nullopt)
-      Decomposed.InBounds = GEPOp->isInBounds();
-    else if (!GEPOp->isInBounds())
-      Decomposed.InBounds = false;
+    // Track the common nowrap flags for all GEPs we see.
+    Decomposed.NWFlags &= GEPOp->getNoWrapFlags();
 
     assert(GEPOp->getSourceElementType()->isSized() && "GEP must be sized");
 
@@ -1112,6 +1108,13 @@ AliasResult BasicAAResult::aliasGEP(
   if (DecompGEP1.Base == GEP1 && DecompGEP2.Base == V2)
     return AliasResult::MayAlias;
 
+  // Swap GEP1 and GEP2 if GEP2 has more variable indices.
+  if (DecompGEP1.VarIndices.size() < DecompGEP2.VarIndices.size()) {
+    std::swap(DecompGEP1, DecompGEP2);
+    std::swap(V1Size, V2Size);
+    std::swap(UnderlyingV1, UnderlyingV2);
+  }
+
   // Subtract the GEP2 pointer from the GEP1 pointer to find out their
   // symbolic 
diff erence.
   subtractDecomposedGEPs(DecompGEP1, DecompGEP2, AAQI);
@@ -1120,20 +1123,19 @@ AliasResult BasicAAResult::aliasGEP(
   // for the two to alias, then we can assume noalias.
   // TODO: Remove !isScalable() once BasicAA fully support scalable location
   // size
-  if (*DecompGEP1.InBounds && DecompGEP1.VarIndices.empty() &&
+
+  if (DecompGEP1.NWFlags.isInBounds() && DecompGEP1.VarIndices.empty() &&
       V2Size.hasValue() && !V2Size.isScalable() &&
       DecompGEP1.Offset.sge(V2Size.getValue()) &&
       isBaseOfObject(DecompGEP2.Base))
     return AliasResult::NoAlias;
 
-  if (isa<GEPOperator>(V2)) {
-    // Symmetric case to above.
-    if (*DecompGEP2.InBounds && DecompGEP1.VarIndices.empty() &&
-        V1Size.hasValue() && !V1Size.isScalable() &&
-        DecompGEP1.Offset.sle(-V1Size.getValue()) &&
-        isBaseOfObject(DecompGEP1.Base))
-      return AliasResult::NoAlias;
-  }
+  // Symmetric case to above.
+  if (DecompGEP2.NWFlags.isInBounds() && DecompGEP1.VarIndices.empty() &&
+      V1Size.hasValue() && !V1Size.isScalable() &&
+      DecompGEP1.Offset.sle(-V1Size.getValue()) &&
+      isBaseOfObject(DecompGEP1.Base))
+    return AliasResult::NoAlias;
 
   // For GEPs with identical offsets, we can preserve the size and AAInfo
   // when performing the alias check on the underlying objects.
@@ -1239,6 +1241,20 @@ AliasResult BasicAAResult::aliasGEP(
     }
   }
 
+  // If the 
diff erence between pointers is Offset +<nuw> Indices then we know
+  // that the addition does not wrap the pointer index type (add nuw) and the
+  // constant Offset is a lower bound on the distance between the pointers. We
+  // can then prove NoAlias via Offset u>= VLeftSize.
+  //    +                +                     +
+  //    | BaseOffset     |   +<nuw> Indices    |
+  //    ---------------->|-------------------->|
+  //    |-->V2Size       |                     |-------> V1Size
+  //   LHS                                    RHS
+  if (!DecompGEP1.VarIndices.empty() &&
+      DecompGEP1.NWFlags.hasNoUnsignedWrap() && V2Size.hasValue() &&
+      !V2Size.isScalable() && DecompGEP1.Offset.uge(V2Size.getValue()))
+    return AliasResult::NoAlias;
+
   // Bail on analysing scalable LocationSize
   if (V1Size.isScalable() || V2Size.isScalable())
     return AliasResult::MayAlias;
@@ -1843,6 +1859,11 @@ bool BasicAAResult::isValueEqualInPotentialCycles(const Value *V,
 void BasicAAResult::subtractDecomposedGEPs(DecomposedGEP &DestGEP,
                                            const DecomposedGEP &SrcGEP,
                                            const AAQueryInfo &AAQI) {
+  // Drop nuw flag from GEP if subtraction of constant offsets overflows in an
+  // unsigned sense.
+  if (DestGEP.Offset.ult(SrcGEP.Offset))
+    DestGEP.NWFlags = DestGEP.NWFlags.withoutNoUnsignedWrap();
+
   DestGEP.Offset -= SrcGEP.Offset;
   for (const VariableGEPIndex &Src : SrcGEP.VarIndices) {
     // Find V in Dest.  This is N^2, but pointer indices almost never have more
@@ -1865,6 +1886,11 @@ void BasicAAResult::subtractDecomposedGEPs(DecomposedGEP &DestGEP,
       // If we found it, subtract off Scale V's from the entry in Dest.  If it
       // goes to zero, remove the entry.
       if (Dest.Scale != Src.Scale) {
+        // Drop nuw flag from GEP if subtraction of V's Scale overflows in an
+        // unsigned sense.
+        if (Dest.Scale.ult(Src.Scale))
+          DestGEP.NWFlags = DestGEP.NWFlags.withoutNoUnsignedWrap();
+
         Dest.Scale -= Src.Scale;
         Dest.IsNSW = false;
       } else {
@@ -1879,6 +1905,9 @@ void BasicAAResult::subtractDecomposedGEPs(DecomposedGEP &DestGEP,
       VariableGEPIndex Entry = {Src.Val, Src.Scale, Src.CxtI, Src.IsNSW,
                                 /* IsNegated */ true};
       DestGEP.VarIndices.push_back(Entry);
+
+      // Drop nuw flag when we have unconsumed variable indices from SrcGEP.
+      DestGEP.NWFlags = DestGEP.NWFlags.withoutNoUnsignedWrap();
     }
   }
 }

diff  --git a/llvm/test/Analysis/BasicAA/gep-nuw-alias.ll b/llvm/test/Analysis/BasicAA/gep-nuw-alias.ll
new file mode 100644
index 00000000000000..b80a457f85176c
--- /dev/null
+++ b/llvm/test/Analysis/BasicAA/gep-nuw-alias.ll
@@ -0,0 +1,214 @@
+; RUN: opt < %s -aa-pipeline=basic-aa -passes=aa-eval -print-all-alias-modref-info -disable-output 2>&1 | FileCheck %s
+
+; CHECK-LABEL: test_no_lower_bound
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+define void @test_no_lower_bound(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p, i64 4
+  %b = getelementptr nuw i8, ptr %p, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: test_lower_bound_lt_size
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+define void @test_lower_bound_lt_size(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p
+  %add = getelementptr nuw i8, ptr %p, i64 2
+  %b = getelementptr nuw i8, ptr %add, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: test_lower_bound_ge_size
+;
+; CHECK-DAG: NoAlias: i32* %a, i32* %b
+define void @test_lower_bound_ge_size(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %b = getelementptr nuw i8, ptr %add, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: test_not_all_nuw
+;
+; If part of the addressing is done with non-nuw GEPs, we can't use properties
+; implied by the last GEP with the whole offset. In this case, the calculation
+; of %add (%p + 4) could wrap the pointer index type, such that %add +<nuw> %i
+; could still alias with %p.
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+define void @test_not_all_nuw(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p
+  %add = getelementptr i8, ptr %p, i64 4
+  %b = getelementptr nuw i8, ptr %add, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: test_multi_step_not_all_nuw
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+define void @test_multi_step_not_all_nuw(ptr %p, i64 %i, i64 %j, i64 %k) {
+  %a = getelementptr i8, ptr %p
+  %add = getelementptr i8, ptr %p, i64 4
+  %step1 = getelementptr i8, ptr %add, i64 %i
+  %step2 = getelementptr i8, ptr %step1, i64 %j
+  %b = getelementptr nuw i8, ptr %step2, i64 %k
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: test_multi_step_all_nuw
+;
+; CHECK-DAG: NoAlias: i32* %a, i32* %b
+define void @test_multi_step_all_nuw(ptr %p, i64 %i, i64 %j, i64 %k) {
+  %a = getelementptr i8, ptr %p
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %step1 = getelementptr nuw i8, ptr %add, i64 %i
+  %step2 = getelementptr nuw i8, ptr %step1, i64 %j
+  %b = getelementptr nuw i8, ptr %step2, i64 %k
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+%struct = type { i64, [2 x i32], i64 }
+
+; CHECK-LABEL: test_struct_no_nuw
+;
+; The array access may alias with the struct elements before and after, because
+; we cannot prove that (%arr + %i) does not alias with the base pointer %p.
+;
+; CHECK-DAG: MayAlias: i32* %arrayidx, i64* %st
+; CHECK-DAG: NoAlias: i64* %after, i64* %st
+; CHECK-DAG: MayAlias: i64* %after, i32* %arrayidx
+
+define void @test_struct_no_nuw(ptr %st, i64 %i) {
+  %arr = getelementptr i8, ptr %st, i64 8
+  %arrayidx = getelementptr [2 x i32], ptr %arr, i64 0, i64 %i
+  %after = getelementptr i8, ptr %st, i64 16
+
+  load i64, ptr %st
+  load i32, ptr %arrayidx
+  load i64, ptr %after
+
+  ret void
+}
+
+; CHECK-LABEL: test_struct_nuw
+;
+; We can prove that the array access does not alias with struct element before,
+; because we can prove that (%arr +<nuw> %i) does not wrap the pointer index
+; type (add nuw). The array access may still alias with the struct element
+; after, as the add nuw property does not preclude this.
+;
+; CHECK-DAG: NoAlias: i32* %arrayidx, i64* %st
+; CHECK-DAG: NoAlias: i64* %after, i64* %st
+; CHECK-DAG: MayAlias: i64* %after, i32* %arrayidx
+
+define void @test_struct_nuw(ptr %st, i64 %i) {
+  %arr = getelementptr nuw i8, ptr %st, i64 8
+  %arrayidx = getelementptr nuw [2 x i32], ptr %arr, i64 0, i64 %i
+  %after = getelementptr nuw i8, ptr %st, i64 16
+
+  load i64, ptr %st
+  load i32, ptr %arrayidx
+  load i64, ptr %after
+
+  ret void
+}
+
+; CHECK-LABEL: constant_offset_overflow
+;
+; If subtraction of constant offsets could overflow in an unsigned sense, we
+; cannot prove the lower bound between the GEPs and so they may still alias.
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+
+define void @constant_offset_overflow(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p, i64 -8
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %b = getelementptr nuw i8, ptr %add, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: equal_var_idx_noalias
+;
+; If GEPs have equal variable indices, we can prove NoAlias when the Scale of
+; the RHS GEP is greater, as in this scenario the constant lower bound holds.
+;
+; CHECK-DAG: NoAlias: i32* %a, i32* %b
+
+define void @equal_var_idx_noalias(ptr %p, i64 %i) {
+  %a = getelementptr i8, ptr %p, i64 %i
+
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %b = getelementptr nuw i16, ptr %add, i64 %i
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}
+
+; CHECK-LABEL: equal_var_idx_alias
+;
+; If GEPs have equal variable indices, we cannot prove NoAlias when the Scale of
+; the RHS GEP is ult Scale of the LHS GEP.
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+
+define void @equal_var_idx_alias(ptr %p, i64 %i) {
+  %a = getelementptr i32, ptr %p, i64 %i
+
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %b = getelementptr nuw i16, ptr %add, i64 %i
+
+  load i32, ptr %b
+  load i32, ptr %a
+
+  ret void
+}
+
+; CHECK-LABEL: both_var_idx
+;
+; If the RHS GEP has unmatched variable indices, we cannot prove a constant
+; lower bound between GEPs.
+;
+; CHECK-DAG: MayAlias: i32* %a, i32* %b
+
+define void @both_var_idx(ptr %p, i64 %i, i64 %j) {
+  %a = getelementptr i8, ptr %p, i64 %i
+
+  %add = getelementptr nuw i8, ptr %p, i64 4
+  %b = getelementptr nuw i8, ptr %add, i64 %j
+
+  load i32, ptr %a
+  load i32, ptr %b
+
+  ret void
+}


        


More information about the llvm-commits mailing list