Lower mem-ops to unaligned i32/i16 load/stores on ARM where supported.

Add support for trimming constants to GetDemandedBits. This fixes some funky
constant generation that occurs when stores are expanded for targets that don't
support unaligned stores natively.

llvm-svn: 144102

3 files changed, 26 insertions, 6 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index e67016ca0c7..8b28ea997f4 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4564,6 +4564,16 @@ SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
   switch (V.getOpcode()) {
   default: break;
+  case ISD::Constant: {
+    const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
+    assert(CV != 0 && "Const value should be ConstSDNode.");
+    const APInt &CVal = CV->getAPIntValue();
+    APInt NewVal = CVal & Mask;
+    if (NewVal != CVal) {
+      return DAG.getConstant(NewVal, V.getValueType());
+    }
+    break;
+  }
   case ISD::OR:
   case ISD::XOR:
     // If the LHS or RHS don't contribute bits to the or, drop them.
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 84a34d587bb..c51e7ae5359 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -8171,6 +8171,13 @@ EVT ARMTargetLowering::getOptimalMemOpType(uint64_t Size,
     }
   }
 
+  // Lowering to i32/i16 if the size permits.
+  if (Size >= 4) {
+    return MVT::i32;
+  } else if (Size >= 2) {
+    return MVT::i16;
+  }
+
   // Let the target-independent logic figure it out.
   return MVT::Other;
 }
diff --git a/llvm/test/CodeGen/ARM/2011-10-26-memset-inline.ll b/llvm/test/CodeGen/ARM/2011-10-26-memset-inline.ll
index a236ffe475c..ff049c89860 100644
--- a/llvm/test/CodeGen/ARM/2011-10-26-memset-inline.ll
+++ b/llvm/test/CodeGen/ARM/2011-10-26-memset-inline.ll
@@ -1,14 +1,17 @@
 ; Make sure short memsets on ARM lower to stores, even when optimizing for size.
-; RUN: llc -march=arm < %s | FileCheck %s
+; RUN: llc -march=arm < %s | FileCheck %s -check-prefix=CHECK-GENERIC
+; RUN: llc -march=arm -mcpu=cortex-a8 < %s | FileCheck %s -check-prefix=CHECK-UNALIGNED
 
 target datalayout = "e-p:32:32:32-i1:8:32-i8:8:32-i16:16:32-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:32:64-v128:32:128-a0:0:32-n32-S32"
 target triple = "thumbv7-apple-ios5.0.0"
 
-; CHECK:      strb
-; CHECK-NEXT: strb
-; CHECK-NEXT: strb
-; CHECK-NEXT: strb
-; CHECK-NEXT: strb 
+; CHECK-GENERIC:      strb
+; CHECK-GENERIT-NEXT: strb
+; CHECK-GENERIT-NEXT: strb
+; CHECK-GENERIT-NEXT: strb
+; CHECK-GENERIT-NEXT: strb
+; CHECK-UNALIGNED:      strb
+; CHECK-UNALIGNED-NEXT: str 
 define void @foo(i8* nocapture %c) nounwind optsize {
 entry:
   call void @llvm.memset.p0i8.i64(i8* %c, i8 -1, i64 5, i32 1, i1 false)