From cae6801b7df7c006d2956d5b6012b902cf6eb14d Mon Sep 17 00:00:00 2001
From: Vincent Lejeune <vljn@ovi.com>
Date: Tue, 5 Mar 2013 15:04:49 +0000
Subject: [PATCH] R600: Turn BUILD_VECTOR into Reg_Sequence

Reviewed-by: Tom Stellard <thomas.stellard at amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@176487 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/R600/AMDILISelDAGToDAG.cpp | 29 +++++++++++++++++++++++++++
 lib/Target/R600/CMakeLists.txt        |  1 -
 test/CodeGen/R600/fdiv.v4f32.ll       |  6 +++---
 3 files changed, 32 insertions(+), 4 deletions(-)

diff --git a/lib/Target/R600/AMDILISelDAGToDAG.cpp b/lib/Target/R600/AMDILISelDAGToDAG.cpp
index e77b9dca49f..0c7880d232b 100644
--- a/lib/Target/R600/AMDILISelDAGToDAG.cpp
+++ b/lib/Target/R600/AMDILISelDAGToDAG.cpp
@@ -162,6 +162,35 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
   }
   switch (Opc) {
   default: break;
+  case ISD::BUILD_VECTOR: {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
+      break;
+    }
+    // BUILD_VECTOR is usually lowered into an IMPLICIT_DEF + 4 INSERT_SUBREG
+    // that adds a 128 bits reg copy when going through TwoAddressInstructions
+    // pass. We want to avoid 128 bits copies as much as possible because they
+    // can't be bundled by our scheduler.
+    SDValue RegSeqArgs[9] = {
+      CurDAG->getTargetConstant(AMDGPU::R600_Reg128RegClassID, MVT::i32),
+      SDValue(), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
+      SDValue(), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32),
+      SDValue(), CurDAG->getTargetConstant(AMDGPU::sub2, MVT::i32),
+      SDValue(), CurDAG->getTargetConstant(AMDGPU::sub3, MVT::i32)
+    };
+    bool IsRegSeq = true;
+    for (unsigned i = 0; i < N->getNumOperands(); i++) {
+      if (dyn_cast<RegisterSDNode>(N->getOperand(i))) {
+        IsRegSeq = false;
+        break;
+      }
+      RegSeqArgs[2 * i + 1] = N->getOperand(i);
+    }
+    if (!IsRegSeq)
+      break;
+    return CurDAG->SelectNodeTo(N, AMDGPU::REG_SEQUENCE, N->getVTList(),
+        RegSeqArgs, 2 * N->getNumOperands() + 1);
+  }
   case ISD::ConstantFP:
   case ISD::Constant: {
     const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
diff --git a/lib/Target/R600/CMakeLists.txt b/lib/Target/R600/CMakeLists.txt
index 00f8b106c73..6c695a08d9a 100644
--- a/lib/Target/R600/CMakeLists.txt
+++ b/lib/Target/R600/CMakeLists.txt
@@ -37,7 +37,6 @@ add_llvm_target(R600CodeGen
   R600ExpandSpecialInstrs.cpp
   R600InstrInfo.cpp
   R600ISelLowering.cpp
-  R600LowerConstCopy.cpp
   R600MachineFunctionInfo.cpp
   R600RegisterInfo.cpp
   SIAnnotateControlFlow.cpp
diff --git a/test/CodeGen/R600/fdiv.v4f32.ll b/test/CodeGen/R600/fdiv.v4f32.ll
index 459fd119d20..79e677f541f 100644
--- a/test/CodeGen/R600/fdiv.v4f32.ll
+++ b/test/CodeGen/R600/fdiv.v4f32.ll
@@ -1,13 +1,13 @@
 ;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
 
 ;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
-;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
-;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
-;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 ;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
 
 define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
   %b_ptr = getelementptr <4 x float> addrspace(1)* %in, i32 1
-- 
2.34.1