R600/SI: adjust writemask to only the used components

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179165 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 156 insertions, 33 deletions

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index ef0df9e..d214135 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -569,8 +569,9 @@ void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
   Operand = SDValue(Node, 0);
 }
 
-SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
-                                          SelectionDAG &DAG) const {
+/// \brief Try to fold the Nodes operands into the Node
+SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
+                                       SelectionDAG &DAG) const {
 
   // Original encoding (either e32 or e64)
   int Opcode = Node->getMachineOpcode();
@@ -703,3 +704,79 @@ SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
   return DAG.getMachineNode(Desc->Opcode, Node->getDebugLoc(),
                             Node->getVTList(), Ops.data(), Ops.size());
 }
+
+/// \brief Helper function for adjustWritemask
+unsigned SubIdx2Lane(unsigned Idx) {
+  switch (Idx) {
+  default: return 0;
+  case AMDGPU::sub0: return 0;
+  case AMDGPU::sub1: return 1;
+  case AMDGPU::sub2: return 2;
+  case AMDGPU::sub3: return 3;
+  }
+}
+
+/// \brief Adjust the writemask of MIMG instructions
+void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
+                                       SelectionDAG &DAG) const {
+  SDNode *Users[4] = { };
+  unsigned Writemask = 0;
+
+  // Try to figure out the used register components
+  for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
+       I != E; ++I) {
+
+    // Abort if we can't understand the usage
+    if (!I->isMachineOpcode() ||
+        I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
+      return;
+
+    unsigned Lane = SubIdx2Lane(I->getConstantOperandVal(1));
+
+    // Abort if we have more than one user per component
+    if (Users[Lane])
+      return;
+
+    Users[Lane] = *I;
+    Writemask |= 1 << Lane;
+  }
+
+  // Abort if all components are used
+  if (Writemask == 0xf)
+    return;
+
+  // Adjust the writemask in the node
+  std::vector<SDValue> Ops;
+  Ops.push_back(DAG.getTargetConstant(Writemask, MVT::i32));
+  for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i)
+    Ops.push_back(Node->getOperand(i));
+  Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());
+
+  // Update the users of the node with the new indices
+  for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
+
+    SDNode *User = Users[i];
+    if (!User)
+      continue;
+
+    SDValue Op = DAG.getTargetConstant(Idx, MVT::i32);
+    DAG.UpdateNodeOperands(User, User->getOperand(0), Op);
+
+    switch (Idx) {
+    default: break;
+    case AMDGPU::sub0: Idx = AMDGPU::sub1; break;
+    case AMDGPU::sub1: Idx = AMDGPU::sub2; break;
+    case AMDGPU::sub2: Idx = AMDGPU::sub3; break;
+    }
+  }
+}
+
+/// \brief Fold the instructions after slecting them
+SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
+                                          SelectionDAG &DAG) const {
+
+  if (AMDGPU::isMIMG(Node->getMachineOpcode()) != -1)
+    adjustWritemask(Node, DAG);
+
+  return foldOperands(Node, DAG);
+}
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
index ace32f8..e8a8ee3 100644
--- a/lib/Target/R600/SIISelLowering.h
+++ b/lib/Target/R600/SIISelLowering.h
@@ -34,6 +34,9 @@ class SITargetLowering : public AMDGPUTargetLowering {
   void ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand, 
                        unsigned RegClass, bool &ScalarSlotUsed) const;
 
+  SDNode *foldOperands(MachineSDNode *N, SelectionDAG &DAG) const;
+  void adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
+
 public:
   SITargetLowering(TargetMachine &tm);
 
diff --git a/lib/Target/R600/SIInstrInfo.h b/lib/Target/R600/SIInstrInfo.h
index d4e60e5..87eff4d 100644
--- a/lib/Target/R600/SIInstrInfo.h
+++ b/lib/Target/R600/SIInstrInfo.h
@@ -80,6 +80,7 @@ namespace AMDGPU {
   int getVOPe64(uint16_t Opcode);
   int getCommuteRev(uint16_t Opcode);
   int getCommuteOrig(uint16_t Opcode);
+  int isMIMG(uint16_t Opcode);
 
 } // End namespace AMDGPU
 
diff --git a/lib/Target/R600/SIInstrInfo.td b/lib/Target/R600/SIInstrInfo.td
index 6ed83db..7e2e4dc 100644
--- a/lib/Target/R600/SIInstrInfo.td
+++ b/lib/Target/R600/SIInstrInfo.td
@@ -379,4 +379,12 @@ def getCommuteOrig : InstrMapping {
   let ValueCols = [["1"]];
 }
 
+def isMIMG : InstrMapping {
+  let FilterClass = "MIMG_Load_Helper";
+  let RowFields = ["Inst"];
+  let ColFields = ["Size"];
+  let KeyCol = ["8"];
+  let ValueCols = [["8"]];
+}
+
 include "SIInstructions.td"
diff --git a/test/CodeGen/R600/llvm.SI.sample.ll b/test/CodeGen/R600/llvm.SI.sample.ll
index 7c7aa5e..d8031c3 100644
--- a/test/CodeGen/R600/llvm.SI.sample.ll
+++ b/test/CodeGen/R600/llvm.SI.sample.ll
@@ -1,21 +1,21 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE_C
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
-;CHECK: IMAGE_SAMPLE
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 15
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 3
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 2
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 1
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 4
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 8
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 5
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 9
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 6
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 10
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 12
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 7
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 11
+;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 13
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 14
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 8
 
 define void @test(i32 %a1, i32 %a2, i32 %a3, i32 %a4) {
    %v1 = insertelement <4 x i32> undef, i32 %a1, i32 0
@@ -67,21 +67,55 @@ define void @test(i32 %a1, i32 %a2, i32 %a3, i32 %a4) {
    %res16 = call <4 x float> @llvm.SI.sample.(<4 x i32> %v16,
       <8 x i32> undef, <4 x i32> undef, i32 16)
    %e1 = extractelement <4 x float> %res1, i32 0
-   %e2 = extractelement <4 x float> %res2, i32 0
-   %e3 = extractelement <4 x float> %res3, i32 0
-   %e4 = extractelement <4 x float> %res4, i32 0
-   %e5 = extractelement <4 x float> %res5, i32 0
-   %e6 = extractelement <4 x float> %res6, i32 0
-   %e7 = extractelement <4 x float> %res7, i32 0
-   %e8 = extractelement <4 x float> %res8, i32 0
-   %e9 = extractelement <4 x float> %res9, i32 0
-   %e10 = extractelement <4 x float> %res10, i32 0
-   %e11 = extractelement <4 x float> %res11, i32 0
-   %e12 = extractelement <4 x float> %res12, i32 0
-   %e13 = extractelement <4 x float> %res13, i32 0
-   %e14 = extractelement <4 x float> %res14, i32 0
-   %e15 = extractelement <4 x float> %res15, i32 0
-   %e16 = extractelement <4 x float> %res16, i32 0
+   %e2 = extractelement <4 x float> %res2, i32 1
+   %e3 = extractelement <4 x float> %res3, i32 2
+   %e4 = extractelement <4 x float> %res4, i32 3
+   %t0 = extractelement <4 x float> %res5, i32 0
+   %t1 = extractelement <4 x float> %res5, i32 1
+   %e5 = fadd float %t0, %t1
+   %t2 = extractelement <4 x float> %res6, i32 0
+   %t3 = extractelement <4 x float> %res6, i32 2
+   %e6 = fadd float %t2, %t3
+   %t4 = extractelement <4 x float> %res7, i32 0
+   %t5 = extractelement <4 x float> %res7, i32 3
+   %e7 = fadd float %t4, %t5
+   %t6 = extractelement <4 x float> %res8, i32 1
+   %t7 = extractelement <4 x float> %res8, i32 2
+   %e8 = fadd float %t6, %t7
+   %t8 = extractelement <4 x float> %res9, i32 1
+   %t9 = extractelement <4 x float> %res9, i32 3
+   %e9 = fadd float %t8, %t9
+   %t10 = extractelement <4 x float> %res10, i32 2
+   %t11 = extractelement <4 x float> %res10, i32 3
+   %e10 = fadd float %t10, %t11
+   %t12 = extractelement <4 x float> %res11, i32 0
+   %t13 = extractelement <4 x float> %res11, i32 1
+   %t14 = extractelement <4 x float> %res11, i32 2
+   %t15 = fadd float %t12, %t13
+   %e11 = fadd float %t14, %t15
+   %t16 = extractelement <4 x float> %res12, i32 0
+   %t17 = extractelement <4 x float> %res12, i32 1
+   %t18 = extractelement <4 x float> %res12, i32 3
+   %t19 = fadd float %t16, %t17
+   %e12 = fadd float %t18, %t19
+   %t20 = extractelement <4 x float> %res13, i32 0
+   %t21 = extractelement <4 x float> %res13, i32 2
+   %t22 = extractelement <4 x float> %res13, i32 3
+   %t23 = fadd float %t20, %t21
+   %e13 = fadd float %t22, %t23
+   %t24 = extractelement <4 x float> %res14, i32 1
+   %t25 = extractelement <4 x float> %res14, i32 2
+   %t26 = extractelement <4 x float> %res14, i32 3
+   %t27 = fadd float %t24, %t25
+   %e14 = fadd float %t26, %t27
+   %t28 = extractelement <4 x float> %res15, i32 0
+   %t29 = extractelement <4 x float> %res15, i32 1
+   %t30 = extractelement <4 x float> %res15, i32 2
+   %t31 = extractelement <4 x float> %res15, i32 3
+   %t32 = fadd float %t28, %t29
+   %t33 = fadd float %t30, %t31
+   %e15 = fadd float %t32, %t33
+   %e16 = extractelement <4 x float> %res16, i32 3
    %s1 = fadd float %e1, %e2
    %s2 = fadd float %s1, %e3
    %s3 = fadd float %s2, %e4