R600/SI: dynamical figure out the reg class of MIMG

Depending on the number of bits set in the writemask.

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@179166 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPUAsmPrinter.cpp

@@ -107,6 +107,9 @@ void AMDGPUAsmPrinter::EmitProgramInfo(MachineFunction &MF) {
         } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
           isSGPR = false;
           width = 2;
+        } else if (AMDGPU::VReg_96RegClass.contains(reg)) {
+          isSGPR = false;
+          width = 3;
         } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
           isSGPR = true;
           width = 4;

lib/Target/R600/SIISelLowering.cpp

@@ -720,7 +720,7 @@ unsigned SubIdx2Lane(unsigned Idx) {
 void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
                                        SelectionDAG &DAG) const {
   SDNode *Users[4] = { };
-  unsigned Writemask = 0;
+  unsigned Writemask = 0, Lane = 0;
 
   // Try to figure out the used register components
   for (SDNode::use_iterator I = Node->use_begin(), E = Node->use_end();
@@ -731,7 +731,7 @@ void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
         I->getMachineOpcode() != TargetOpcode::EXTRACT_SUBREG)
       return;
 
-    unsigned Lane = SubIdx2Lane(I->getConstantOperandVal(1));
+    Lane = SubIdx2Lane(I->getConstantOperandVal(1));
 
     // Abort if we have more than one user per component
     if (Users[Lane])
@@ -752,6 +752,16 @@ void SITargetLowering::adjustWritemask(MachineSDNode *&Node,
     Ops.push_back(Node->getOperand(i));
   Node = (MachineSDNode*)DAG.UpdateNodeOperands(Node, Ops.data(), Ops.size());
 
+  // If we only got one lane, replace it with a copy
+  if (Writemask == (1U << Lane)) {
+    SDValue RC = DAG.getTargetConstant(AMDGPU::VReg_32RegClassID, MVT::i32);
+    SDNode *Copy = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+                                      DebugLoc(), MVT::f32,
+                                      SDValue(Node, 0), RC);
+    DAG.ReplaceAllUsesWith(Users[Lane], Copy);
+    return;
+  }
+
   // Update the users of the node with the new indices
   for (unsigned i = 0, Idx = AMDGPU::sub0; i < 4; ++i) {
 
@@ -780,3 +790,28 @@ SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
 
   return foldOperands(Node, DAG);
 }
+
+/// \brief Assign the register class depending on the number of
+/// bits set in the writemask
+void SITargetLowering::AdjustInstrPostInstrSelection(MachineInstr *MI,
+                                                     SDNode *Node) const {
+  if (AMDGPU::isMIMG(MI->getOpcode()) == -1)
+    return;
+
+  unsigned VReg = MI->getOperand(0).getReg();
+  unsigned Writemask = MI->getOperand(1).getImm();
+  unsigned BitsSet = 0;
+  for (unsigned i = 0; i < 4; ++i)
+    BitsSet += Writemask & (1 << i) ? 1 : 0;
+
+  const TargetRegisterClass *RC;
+  switch (BitsSet) {
+  default: return;
+  case 1:  RC = &AMDGPU::VReg_32RegClass; break;
+  case 2:  RC = &AMDGPU::VReg_64RegClass; break;
+  case 3:  RC = &AMDGPU::VReg_96RegClass; break;
+  }
+
+  MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
+  MRI.setRegClass(VReg, RC);
+}

lib/Target/R600/SIISelLowering.h

@@ -53,6 +53,8 @@ public:
   virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
   virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   virtual SDNode *PostISelFolding(MachineSDNode *N, SelectionDAG &DAG) const;
+  virtual void AdjustInstrPostInstrSelection(MachineInstr *MI,
+                                             SDNode *Node) const;
 
   int32_t analyzeImmediate(const SDNode *N) const;
 };

lib/Target/R600/SIInstrInfo.cpp

@@ -58,6 +58,10 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3, 0
   };
 
+  const int16_t Sub0_2[] = {
+    AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, 0
+  };
+
   const int16_t Sub0_1[] = {
     AMDGPU::sub0, AMDGPU::sub1, 0
   };
@@ -125,6 +129,11 @@ SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
     Opcode = AMDGPU::V_MOV_B32_e32;
     SubIndices = Sub0_1;
 
+  } else if (AMDGPU::VReg_96RegClass.contains(DestReg)) {
+    assert(AMDGPU::VReg_96RegClass.contains(SrcReg));
+    Opcode = AMDGPU::V_MOV_B32_e32;
+    SubIndices = Sub0_2;
+
   } else if (AMDGPU::VReg_128RegClass.contains(DestReg)) {
     assert(AMDGPU::VReg_128RegClass.contains(SrcReg) ||
 	   AMDGPU::SReg_128RegClass.contains(SrcReg));

lib/Target/R600/SIInstrInfo.td

@@ -346,6 +346,7 @@ class MIMG_Load_Helper <bits<7> op, string asm> : MIMG <
   []> {
   let mayLoad = 1;
   let mayStore = 0;
+  let hasPostISelHook = 1;
 }
 
 //===----------------------------------------------------------------------===//
@@ -379,6 +380,7 @@ def getCommuteOrig : InstrMapping {
   let ValueCols = [["1"]];
 }
 
+// Test if the supplied opcode is an MIMG instruction
 def isMIMG : InstrMapping {
   let FilterClass = "MIMG_Load_Helper";
   let RowFields = ["Inst"];

lib/Target/R600/SIRegisterInfo.td

@@ -94,6 +94,12 @@ def VGPR_64 : RegisterTuples<[sub0, sub1],
                              [(add (trunc VGPR_32, 255)),
                               (add (shl VGPR_32, 1))]>;
 
+// VGPR 96-bit registers
+def VGPR_96 : RegisterTuples<[sub0, sub1, sub2],
+                             [(add (trunc VGPR_32, 254)),
+                              (add (shl VGPR_32, 1)),
+                              (add (shl VGPR_32, 2))]>;
+
 // VGPR 128-bit registers
 def VGPR_128 : RegisterTuples<[sub0, sub1, sub2, sub3],
                               [(add (trunc VGPR_32, 253)),
@@ -162,6 +168,10 @@ def VReg_32 : RegisterClass<"AMDGPU", [i32, f32, v1i32], 32, (add VGPR_32)>;
 
 def VReg_64 : RegisterClass<"AMDGPU", [i64, f64, v2i32, v2f32], 64, (add VGPR_64)>;
 
+def VReg_96 : RegisterClass<"AMDGPU", [untyped], 96, (add VGPR_96)> {
+  let Size = 96;
+}
+
 def VReg_128 : RegisterClass<"AMDGPU", [v4i32, v4f32], 128, (add VGPR_128)>;
 
 def VReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 256, (add VGPR_256)>;

test/CodeGen/R600/llvm.SI.sample.ll

@@ -1,21 +1,21 @@
 ;RUN: llc < %s -march=r600 -mcpu=verde | FileCheck %s
 
 ;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]+}}, 3
-;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 2
+;CHECK: IMAGE_SAMPLE {{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]+}}, 1
-;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 4
+;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+}}, 4
-;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 8
+;CHECK: IMAGE_SAMPLE {{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]+}}, 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]+}}, 9
-;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 6
+;CHECK: IMAGE_SAMPLE_C {{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]+}}, 10
-;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 12
+;CHECK: IMAGE_SAMPLE {{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]+}}, 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]+}}, 11
-;CHECK: IMAGE_SAMPLE_C {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 13
+;CHECK: IMAGE_SAMPLE_C {{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]+}}, 14
-;CHECK: IMAGE_SAMPLE {{VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+_VGPR[0-9]+}}, 8
+;CHECK: IMAGE_SAMPLE {{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