Use AA in LoadCombine

LoadCombine can be smarter about aborting when a writing instruction is
encountered, instead of aborting upon encountering any writing instruction, use
an AliasSetTracker, and only abort when encountering some write that might
alias with the loads that could potentially be combined.

This was originally motivated by comments made (and a test case provided) by
David Majnemer in response to PR21448. It turned out that LoadCombine was not
responsible for that PR, but LoadCombine should also be improved so that
unrelated stores (and @llvm.assume) don't interrupt load combining.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@221203 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/Analysis/AliasSetTracker.h b/include/llvm/Analysis/AliasSetTracker.h
index 6dcd4a0..403a2b5 100644 (file)
--- a/include/llvm/Analysis/AliasSetTracker.h
+++ b/include/llvm/Analysis/AliasSetTracker.h
@@ -370,6 +370,10 @@ public:
   /// alias sets.
   bool containsPointer(Value *P, uint64_t Size, const AAMDNodes &AAInfo) const;
 
+  /// Return true if the specified instruction "may" (or must) alias one of the
+  /// members in any of the sets.
+  bool containsUnknown(Instruction *I) const;
+
   /// getAliasAnalysis - Return the underlying alias analysis object used by
   /// this tracker.
   AliasAnalysis &getAliasAnalysis() const { return AA; }

diff --git a/lib/Analysis/AliasSetTracker.cpp b/lib/Analysis/AliasSetTracker.cpp
index 843a541..9d941e5 100644 (file)
--- a/lib/Analysis/AliasSetTracker.cpp
+++ b/lib/Analysis/AliasSetTracker.cpp
@@ -242,7 +242,12 @@ bool AliasSetTracker::containsPointer(Value *Ptr, uint64_t Size,
   return false;
 }
 
-
+bool AliasSetTracker::containsUnknown(Instruction *Inst) const {
+  for (const_iterator I = begin(), E = end(); I != E; ++I)
+    if (!I->Forward && I->aliasesUnknownInst(Inst, AA))
+      return true;
+  return false;
+}
 
 AliasSet *AliasSetTracker::findAliasSetForUnknownInst(Instruction *Inst) {
   AliasSet *FoundSet = nullptr;

diff --git a/lib/Transforms/Scalar/LoadCombine.cpp b/lib/Transforms/Scalar/LoadCombine.cpp
index 648626a..11e4d76 100644 (file)
--- a/lib/Transforms/Scalar/LoadCombine.cpp
+++ b/lib/Transforms/Scalar/LoadCombine.cpp
@@ -15,6 +15,8 @@
 
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AliasSetTracker.h"
 #include "llvm/Analysis/TargetFolder.h"
 #include "llvm/Pass.h"
 #include "llvm/IR/DataLayout.h"
@@ -51,11 +53,12 @@ struct LoadPOPPair {
 class LoadCombine : public BasicBlockPass {
   LLVMContext *C;
   const DataLayout *DL;
+  AliasAnalysis *AA;
 
 public:
   LoadCombine()
       : BasicBlockPass(ID),
-        C(nullptr), DL(nullptr) {
+        C(nullptr), DL(nullptr), AA(nullptr) {
     initializeSROAPass(*PassRegistry::getPassRegistry());
   }
   
@@ -225,19 +228,23 @@ bool LoadCombine::runOnBasicBlock(BasicBlock &BB) {
   if (skipOptnoneFunction(BB) || !DL)
     return false;
 
+  AA = &getAnalysis<AliasAnalysis>();
+
   IRBuilder<true, TargetFolder>
   TheBuilder(BB.getContext(), TargetFolder(DL));
   Builder = &TheBuilder;
 
   DenseMap<const Value *, SmallVector<LoadPOPPair, 8>> LoadMap;
+  AliasSetTracker AST(*AA);
 
   bool Combined = false;
   unsigned Index = 0;
   for (auto &I : BB) {
-    if (I.mayWriteToMemory() || I.mayThrow()) {
+    if (I.mayThrow() || (I.mayWriteToMemory() && AST.containsUnknown(&I))) {
       if (combineLoads(LoadMap))
         Combined = true;
       LoadMap.clear();
+      AST.clear();
       continue;
     }
     LoadInst *LI = dyn_cast<LoadInst>(&I);
@@ -250,6 +257,7 @@ bool LoadCombine::runOnBasicBlock(BasicBlock &BB) {
     if (!POP.Pointer)
       continue;
     LoadMap[POP.Pointer].push_back(LoadPOPPair(LI, POP, Index++));
+    AST.add(LI);
   }
   if (combineLoads(LoadMap))
     Combined = true;
@@ -258,6 +266,9 @@ bool LoadCombine::runOnBasicBlock(BasicBlock &BB) {
 
 void LoadCombine::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.setPreservesCFG();
+
+  AU.addRequired<AliasAnalysis>();
+  AU.addPreserved<AliasAnalysis>();
 }
 
 char LoadCombine::ID = 0;
@@ -266,5 +277,9 @@ BasicBlockPass *llvm::createLoadCombinePass() {
   return new LoadCombine();
 }
 
-INITIALIZE_PASS(LoadCombine, "load-combine", "Combine Adjacent Loads", false,
-                false)
+INITIALIZE_PASS_BEGIN(LoadCombine, "load-combine", "Combine Adjacent Loads",
+                      false, false)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(LoadCombine, "load-combine", "Combine Adjacent Loads",
+                    false, false)
+

diff --git a/test/Transforms/LoadCombine/load-combine-aa.ll b/test/Transforms/LoadCombine/load-combine-aa.ll
new file mode 100644 (file)
index 0000000..3542dce
--- /dev/null
+++ b/test/Transforms/LoadCombine/load-combine-aa.ll
@@ -0,0 +1,39 @@
+; RUN: opt -basicaa -load-combine -instcombine -S < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i64 @test1(i32* nocapture readonly noalias %a, i32* nocapture readonly noalias %b) {
+; CHECK-LABEL: @test1
+
+; CHECK: load i64*
+; CHECK: ret i64
+
+  %load1 = load i32* %a, align 4
+  %conv = zext i32 %load1 to i64
+  %arrayidx1 = getelementptr inbounds i32* %a, i64 1
+  store i32 %load1, i32* %b, align 4
+  %load2 = load i32* %arrayidx1, align 4
+  %conv2 = zext i32 %load2 to i64
+  %shl = shl nuw i64 %conv2, 32
+  %add = or i64 %shl, %conv
+  ret i64 %add
+}
+
+define i64 @test2(i32* nocapture readonly %a, i32* nocapture readonly %b) {
+; CHECK-LABEL: @test2
+
+; CHECK: load i32*
+; CHECK: load i32*
+; CHECK: ret i64
+
+  %load1 = load i32* %a, align 4
+  %conv = zext i32 %load1 to i64
+  %arrayidx1 = getelementptr inbounds i32* %a, i64 1
+  store i32 %load1, i32* %b, align 4
+  %load2 = load i32* %arrayidx1, align 4
+  %conv2 = zext i32 %load2 to i64
+  %shl = shl nuw i64 %conv2, 32
+  %add = or i64 %shl, %conv
+  ret i64 %add
+}
+

diff --git a/test/Transforms/LoadCombine/load-combine-assume.ll b/test/Transforms/LoadCombine/load-combine-assume.ll
new file mode 100644 (file)
index 0000000..94f6300
--- /dev/null
+++ b/test/Transforms/LoadCombine/load-combine-assume.ll
@@ -0,0 +1,44 @@
+; RUN: opt -basicaa -load-combine -instcombine -S < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @llvm.assume(i1) nounwind
+
+; 'load' before the 'call' gets optimized:
+define i64 @test1(i32* nocapture readonly %a, i1 %b) {
+; CHECK-LABEL: @test1
+
+; CHECK-DAG: load i64* %1, align 4
+; CHECK-DAG: tail call void @llvm.assume(i1 %b)
+; CHECK: ret i64
+
+  %load1 = load i32* %a, align 4
+  %conv = zext i32 %load1 to i64
+  %arrayidx1 = getelementptr inbounds i32* %a, i64 1
+  %load2 = load i32* %arrayidx1, align 4
+  tail call void @llvm.assume(i1 %b)
+  %conv2 = zext i32 %load2 to i64
+  %shl = shl nuw i64 %conv2, 32
+  %add = or i64 %shl, %conv
+  ret i64 %add
+}
+
+; 'call' before the 'load' doesn't get optimized:
+define i64 @test2(i32* nocapture readonly %a, i1 %b) {
+; CHECK-LABEL: @test2
+
+; CHECK-DAG: load i64* %1, align 4
+; CHECK-DAG: tail call void @llvm.assume(i1 %b)
+; CHECK: ret i64
+
+  %load1 = load i32* %a, align 4
+  %conv = zext i32 %load1 to i64
+  %arrayidx1 = getelementptr inbounds i32* %a, i64 1
+  tail call void @llvm.assume(i1 %b)
+  %load2 = load i32* %arrayidx1, align 4
+  %conv2 = zext i32 %load2 to i64
+  %shl = shl nuw i64 %conv2, 32
+  %add = or i64 %shl, %conv
+  ret i64 %add
+}
+