From 5e742a3e1b29104fcadc26286f6daa6c450f87d9 Mon Sep 17 00:00:00 2001
From: Nadav Rotem <nadav.rotem@intel.com>
Date: Thu, 11 Aug 2011 16:41:21 +0000
Subject: [PATCH] [AVX] If the data which is going to be saved is already in
 two XMM registers (for example, after integer operation), do not pack the
 registers into a YMM before saving. Its better to save as two XMM registers.

Before:
                vinsertf128         $1, %xmm3, %ymm0, %ymm3
                vinsertf128         $0, %xmm1, %ymm3, %ymm1
                vmovaps              %ymm1, 416(%rsp)

After:
                vmovaps              %xmm3, 416+16(%rsp)
                vmovaps              %xmm1, 416(%rsp)



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@137308 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/X86/X86ISelLowering.cpp | 23 +++++++++++++++++++++++
 test/CodeGen/X86/avx-insert.ll     | 17 +++++++++++++++++
 2 files changed, 40 insertions(+)
 create mode 100644 test/CodeGen/X86/avx-insert.ll

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index be3ecd7e531..94faae5b556 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -12589,6 +12589,29 @@ static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
   EVT VT = St->getValue().getValueType();
   EVT StVT = St->getMemoryVT();
   DebugLoc dl = St->getDebugLoc();
+  SDValue StoredVal = St->getOperand(1);
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+  // If we are saving a concatination of two XMM registers, perform two stores.
+  if (VT.getSizeInBits() == 256 &&
+    StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
+    StoredVal.getNumOperands() == 2) {
+
+    SDValue Value0 = StoredVal.getOperand(0);
+    SDValue Value1 = StoredVal.getOperand(1);
+
+    SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
+    SDValue Ptr0 = St->getBasePtr();
+    SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
+
+    SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
+                                St->getPointerInfo(), St->isVolatile(),
+                                St->isNonTemporal(), St->getAlignment());
+    SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
+                                St->getPointerInfo(), St->isVolatile(),
+                                St->isNonTemporal(), St->getAlignment());
+    return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
+  }
 
   // Optimize trunc store (of multiple scalars) to shuffle and store.
   // First, pack all of the elements in one place. Next, store to memory
diff --git a/test/CodeGen/X86/avx-insert.ll b/test/CodeGen/X86/avx-insert.ll
new file mode 100644
index 00000000000..d9eae03eee7
--- /dev/null
+++ b/test/CodeGen/X86/avx-insert.ll
@@ -0,0 +1,17 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=corei7-avx -mattr=+avx | FileCheck %s
+
+; It is faster to make two saves, if the data is already in XMM registers. For
+; example, after making an integer operation.
+define void @double_save(<4 x i32>* %Ap, <4 x i32>* %Bp, <8 x i32>* %P) nounwind ssp {
+entry:
+  ; CHECK: movaps
+  ; CHECK: movaps
+  ; CHECK: movaps
+  ; CHECK: movaps
+  %A = load <4 x i32>* %Ap
+  %B = load <4 x i32>* %Bp
+  %Z = shufflevector <4 x i32>%A, <4 x i32>%B, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+  store <8 x i32> %Z, <8 x i32>* %P, align 16
+  ret void
+}
+