Merging r341094:

------------------------------------------------------------------------
r341094 | efriedma | 2018-08-30 20:59:24 +0200 (Thu, 30 Aug 2018) | 11 lines

[SROA] Fix alignment for uses of PHI nodes.

Splitting an alloca can decrease the alignment of GEPs into the
partition.  Normally, rewriting accounts for this, but the code was
missing for uses of PHI nodes and select instructions.

Fixes https://bugs.llvm.org/show_bug.cgi?id=38707 .

Differential Revision: https://reviews.llvm.org/D51335

------------------------------------------------------------------------

git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_70@341220 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/SROA.cpp b/lib/Transforms/Scalar/SROA.cpp
index de16b60..bf482bf 100644 (file)
--- a/lib/Transforms/Scalar/SROA.cpp
+++ b/lib/Transforms/Scalar/SROA.cpp
@@ -3046,6 +3046,42 @@ 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) {
     LLVM_DEBUG(dbgs() << "    original: " << PN << "\n");
     assert(BeginOffset >= NewAllocaBeginOffset && "PHIs are unsplittable");
@@ -3069,6 +3105,9 @@ private:
     LLVM_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.
@@ -3093,6 +3132,9 @@ private:
     LLVM_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 fb76548..e7ba2e8 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
+}