D51335-alignment-issue

===================================================================

Gbp-Pq: Name D51335-alignment-issue.diff

diff --git a/lib/Transforms/Scalar/SROA.cpp b/lib/Transforms/Scalar/SROA.cpp
index bfe3754f076962361b5b2ee6469faf7a8b0b067e..3874b177f7f7e9154b2d9b9e6210a88806a10e57 100644 (file)
--- a/lib/Transforms/Scalar/SROA.cpp
+++ b/lib/Transforms/Scalar/SROA.cpp
@@ -2987,6 +2987,42 @@ private:
     return true;
   }
 
+  void fixLoadStoreAlign(Instruction &Root) {
+    // This algorithm implements the same visitor loop as
+    // hasUnsafePHIOrSelectUse, and fixes the alignment of each load
+    // or store found.
+    SmallPtrSet<Instruction *, 4> Visited;
+    SmallVector<Instruction *, 4> Uses;
+    Visited.insert(&Root);
+    Uses.push_back(&Root);
+    do {
+      Instruction *I = Uses.pop_back_val();
+
+      if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+        unsigned LoadAlign = LI->getAlignment();
+        if (!LoadAlign)
+          LoadAlign = DL.getABITypeAlignment(LI->getType());
+        LI->setAlignment(std::min(LoadAlign, getSliceAlign()));
+        continue;
+      }
+      if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
+        unsigned StoreAlign = SI->getAlignment();
+        if (!StoreAlign) {
+          Value *Op = SI->getOperand(0);
+          StoreAlign = DL.getABITypeAlignment(Op->getType());
+        }
+        SI->setAlignment(std::min(StoreAlign, getSliceAlign()));
+        continue;
+      }
+
+      assert(isa<BitCastInst>(I) || isa<PHINode>(I) ||
+             isa<SelectInst>(I) || isa<GetElementPtrInst>(I));
+      for (User *U : I->users())
+        if (Visited.insert(cast<Instruction>(U)).second)
+          Uses.push_back(cast<Instruction>(U));
+    } while (!Uses.empty());
+  }
+
   bool visitPHINode(PHINode &PN) {
     DEBUG(dbgs() << "    original: " << PN << "\n");
     assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
@@ -3010,6 +3046,9 @@ private:
     DEBUG(dbgs() << "          to: " << PN << "\n");
     deleteIfTriviallyDead(OldPtr);
 
+    // Fix the alignment of any loads or stores using this PHI node.
+    fixLoadStoreAlign(PN);
+
     // PHIs can't be promoted on their own, but often can be speculated. We
     // check the speculation outside of the rewriter so that we see the
     // fully-rewritten alloca.
@@ -3034,6 +3073,9 @@ private:
     DEBUG(dbgs() << "          to: " << SI << "\n");
     deleteIfTriviallyDead(OldPtr);
 
+    // Fix the alignment of any loads or stores using this select.
+    fixLoadStoreAlign(SI);
+
     // Selects can't be promoted on their own, but often can be speculated. We
     // check the speculation outside of the rewriter so that we see the
     // fully-rewritten alloca.

diff --git a/test/Transforms/SROA/phi-and-select.ll b/test/Transforms/SROA/phi-and-select.ll
index fb76548b1d18787f262236d830b4c5a761ae8e4c..e7ba2e89d7956d5356ca845db9e938bd11cec881 100644 (file)
--- a/test/Transforms/SROA/phi-and-select.ll
+++ b/test/Transforms/SROA/phi-and-select.ll
@@ -600,3 +600,35 @@ if.then5:                                         ; preds = %if.then2, %if.end
   store %struct.S undef, %struct.S* %f1, align 4
   ret void
 }
+
+define i32 @phi_align(i32* %z) {
+; CHECK-LABEL: @phi_align(
+entry:
+  %a = alloca [8 x i8], align 8
+; CHECK: alloca [7 x i8]
+
+  %a0x = getelementptr [8 x i8], [8 x i8]* %a, i64 0, i32 1
+  %a0 = bitcast i8* %a0x to i32*
+  %a1x = getelementptr [8 x i8], [8 x i8]* %a, i64 0, i32 4
+  %a1 = bitcast i8* %a1x to i32*
+; CHECK: store i32 0, {{.*}}, align 1
+  store i32 0, i32* %a0, align 1
+; CHECK: store i32 1, {{.*}}, align 1
+  store i32 1, i32* %a1, align 4
+; CHECK: load {{.*}}, align 1
+  %v0 = load i32, i32* %a0, align 1
+; CHECK: load {{.*}}, align 1
+  %v1 = load i32, i32* %a1, align 4
+  %cond = icmp sle i32 %v0, %v1
+  br i1 %cond, label %then, label %exit
+
+then:
+  br label %exit
+
+exit:
+; CHECK: %phi = phi i32* [ {{.*}}, %then ], [ %z, %entry ]
+; CHECK-NEXT: %result = load i32, i32* %phi, align 1
+  %phi = phi i32* [ %a1, %then ], [ %z, %entry ]
+  %result = load i32, i32* %phi, align 4
+  ret i32 %result
+}