Revise alignment checking/calculation on 256-bit unaligned memory access

- It's still considered aligned when the specified alignment is larger
  than the natural alignment;
- The new alignment for the high 128-bit vector should be min(16,
  alignment) as the pointer is advanced by 16, a power-of-2 offset.

llvm-svn: 177947

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -16639,11 +16639,10 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   unsigned RegSz = RegVT.getSizeInBits();
 
+  // On Sandybridge unaligned 256bit loads are inefficient.
   ISD::LoadExtType Ext = Ld->getExtensionType();
   unsigned Alignment = Ld->getAlignment();
-  bool IsAligned = Alignment == 0 || Alignment == MemVT.getSizeInBits()/8;
+  bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
-
-  // On Sandybridge unaligned 256bit loads are inefficient.
   if (RegVT.is256BitVector() && !Subtarget->hasInt256() &&
       !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
     unsigned NumElems = RegVT.getVectorNumElements();
@@ -16663,7 +16662,7 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
     SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
                                 Ld->getPointerInfo(), Ld->isVolatile(),
                                 Ld->isNonTemporal(), Ld->isInvariant(),
-                                std::max(Alignment/2U, 1U));
+                                std::min(16U, Alignment));
     SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
                              Load1.getValue(1),
                              Load2.getValue(1));
@@ -16834,13 +16833,13 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
   DebugLoc dl = St->getDebugLoc();
   SDValue StoredVal = St->getOperand(1);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-  unsigned Alignment = St->getAlignment();
-  bool IsAligned = Alignment == 0 || Alignment == VT.getSizeInBits()/8;
 
   // If we are saving a concatenation of two XMM registers, perform two stores.
   // On Sandy Bridge, 256-bit memory operations are executed by two
   // 128-bit ports. However, on Haswell it is better to issue a single 256-bit
   // memory  operation.
+  unsigned Alignment = St->getAlignment();
+  bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
   if (VT.is256BitVector() && !Subtarget->hasInt256() &&
       StVT == VT && !IsAligned) {
     unsigned NumElems = VT.getVectorNumElements();
@@ -16860,7 +16859,7 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
     SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
                                 St->getPointerInfo(), St->isVolatile(),
                                 St->isNonTemporal(),
-                                std::max(Alignment/2U, 1U));
+                                std::min(16U, Alignment));
     return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
   }
 

llvm/test/CodeGen/X86/avx-load-store.ll

@@ -81,7 +81,7 @@ define void @storev32i8_01(<32 x i8> %a) nounwind {
 ; CHECK: _double_save
 ; CHECK-NOT: vinsertf128 $1
 ; CHECK-NOT: vinsertf128 $0
-; CHECK: vmovups %xmm
+; CHECK: vmovaps %xmm
 ; CHECK: vmovaps %xmm
 define void @double_save(<4 x i32> %A, <4 x i32> %B, <8 x i32>* %P) nounwind ssp {
 entry:
@@ -127,3 +127,25 @@ define void @add8i32(<8 x i32>* %ret, <8 x i32>* %bp) nounwind {
   store <8 x i32> %x, <8 x i32>* %ret, align 1
   ret void
 }
+
+; CHECK: add4i64a64
+; CHECK: vmovaps ({{.*}}), %ymm{{.*}}
+; CHECK: vmovaps %ymm{{.*}}, ({{.*}})
+define void @add4i64a64(<4 x i64>* %ret, <4 x i64>* %bp) nounwind {
+  %b = load <4 x i64>* %bp, align 64
+  %x = add <4 x i64> zeroinitializer, %b
+  store <4 x i64> %x, <4 x i64>* %ret, align 64
+  ret void
+}
+
+; CHECK: add4i64a16
+; CHECK: vmovaps {{.*}}({{.*}}), %xmm{{.*}}
+; CHECK: vmovaps {{.*}}({{.*}}), %xmm{{.*}}
+; CHECK: vmovaps %xmm{{.*}}, {{.*}}({{.*}})
+; CHECK: vmovaps %xmm{{.*}}, {{.*}}({{.*}})
+define void @add4i64a16(<4 x i64>* %ret, <4 x i64>* %bp) nounwind {
+  %b = load <4 x i64>* %bp, align 16
+  %x = add <4 x i64> zeroinitializer, %b
+  store <4 x i64> %x, <4 x i64>* %ret, align 16
+  ret void
+}