radeon/llvm: ExpandSpecialInstrs - Add support for cube instructions

4 files changed, 100 insertions, 63 deletions

diff --git a/src/gallium/drivers/radeon/R600CodeEmitter.cpp b/src/gallium/drivers/radeon/R600CodeEmitter.cpp
index 396ae6f5054..efe194b5206 100644
--- a/src/gallium/drivers/radeon/R600CodeEmitter.cpp
+++ b/src/gallium/drivers/radeon/R600CodeEmitter.cpp
@@ -49,17 +49,14 @@ private:
   const R600RegisterInfo * TRI;
   const R600InstrInfo * TII;
 
-  bool IsCube;
   unsigned currentElement;
-  bool IsLast;
 
   unsigned section_start;
 
 public:
 
   R600CodeEmitter(formatted_raw_ostream &OS) : MachineFunctionPass(ID),
-      _OS(OS), TM(NULL), IsCube(false),
-      IsLast(true) { }
+      _OS(OS), TM(NULL) { }
 
   const char *getPassName() const { return "AMDGPU Machine Code Emitter"; }
 
@@ -70,7 +67,7 @@ public:
 private:
 
   void EmitALUInstr(MachineInstr  &MI);
-  void EmitSrc(const MachineOperand & MO, int chan_override  = -1);
+  void EmitSrc(const MachineOperand & MO);
   void EmitDst(const MachineOperand & MO);
   void EmitALU(MachineInstr &MI, unsigned numSrc);
   void EmitTexInstr(MachineInstr &MI);
@@ -160,7 +157,6 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
      for (MachineBasicBlock::instr_iterator I = MBB.instr_begin(),
                                             E = MBB.instr_end(); I != E; ++I) {
           MachineInstr &MI = *I;
-	  IsCube = TII->isCubeOp(MI.getOpcode());
           if (MI.getNumOperands() > 1 && MI.getOperand(0).isReg() && MI.getOperand(0).isDead()) {
             continue;
           }
@@ -168,15 +164,6 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
             EmitTexInstr(MI);
           } else if (TII->isFCOp(MI.getOpcode())){
             EmitFCInstr(MI);
-          } else if (IsCube) {
-            IsLast = false;
-            // XXX: On Cayman, some (all?) of the vector instructions only need
-            // to fill the first three slots.
-            for (currentElement = 0; currentElement < 4; currentElement++) {
-              IsLast = (currentElement == 3);
-              EmitALUInstr(MI);
-            }
-	    IsCube = false;
           } else if (MI.getOpcode() == AMDGPU::RETURN ||
                      MI.getOpcode() == AMDGPU::BUNDLE ||
                      MI.getOpcode() == AMDGPU::KILL) {
@@ -250,25 +237,18 @@ void R600CodeEmitter::EmitALUInstr(MachineInstr &MI)
   // Emit instruction type
   EmitByte(0);
 
-  if (IsCube) {
-    static const int cube_src_swz[] = {2, 2, 0, 1};
-    EmitSrc(MI.getOperand(1), cube_src_swz[currentElement]);
-    EmitSrc(MI.getOperand(1), cube_src_swz[3-currentElement]);
-    EmitNullBytes(SRC_BYTE_COUNT);
-  } else {
-    unsigned int opIndex;
-    for (opIndex = 1; opIndex < numOperands; opIndex++) {
-      // Literal constants are always stored as the last operand.
-      if (MI.getOperand(opIndex).isImm() || MI.getOperand(opIndex).isFPImm()) {
-        break;
-      }
-      EmitSrc(MI.getOperand(opIndex));
+  unsigned int opIndex;
+  for (opIndex = 1; opIndex < numOperands; opIndex++) {
+    // Literal constants are always stored as the last operand.
+    if (MI.getOperand(opIndex).isImm() || MI.getOperand(opIndex).isFPImm()) {
+      break;
     }
+    EmitSrc(MI.getOperand(opIndex));
+  }
 
-    // Emit zeros for unused sources
-    for ( ; opIndex < 4; opIndex++) {
-      EmitNullBytes(SRC_BYTE_COUNT);
-    }
+  // Emit zeros for unused sources
+  for ( ; opIndex < 4; opIndex++) {
+    EmitNullBytes(SRC_BYTE_COUNT);
   }
 
   EmitDst(dstOp);
@@ -276,7 +256,7 @@ void R600CodeEmitter::EmitALUInstr(MachineInstr &MI)
   EmitALU(MI, numOperands - 1);
 }
 
-void R600CodeEmitter::EmitSrc(const MachineOperand & MO, int chan_override)
+void R600CodeEmitter::EmitSrc(const MachineOperand & MO)
 {
   uint32_t value = 0;
   // Emit the source select (2 bytes).  For GPRs, this is the register index.
@@ -302,9 +282,7 @@ void R600CodeEmitter::EmitSrc(const MachineOperand & MO, int chan_override)
   }
 
   // Emit the source channel (1 byte)
-  if (chan_override != -1) {
-    EmitByte(chan_override);
-  } else if (MO.isReg()) {
+  if (MO.isReg()) {
     EmitByte(TRI->getHWRegChan(MO.getReg()));
   } else {
     EmitByte(0);
@@ -345,11 +323,7 @@ void R600CodeEmitter::EmitDst(const MachineOperand & MO)
     EmitByte(getHWReg(MO.getReg()));
 
     // Emit the element of the destination register (1 byte)
-    if (IsCube) {
-      EmitByte(currentElement);
-    } else {
-      EmitByte(TRI->getHWRegChan(MO.getReg()));
-    }
+    EmitByte(TRI->getHWRegChan(MO.getReg()));
 
     // Emit isClamped (1 byte)
     if (MO.getTargetFlags() & MO_FLAG_CLAMP) {
@@ -379,9 +353,8 @@ void R600CodeEmitter::EmitALU(MachineInstr &MI, unsigned numSrc)
   EmitTwoBytes(getBinaryCodeForInstr(MI));
 
   // Emit IsLast (for this instruction group) (1 byte)
-  if (!IsLast ||
-				(MI.isInsideBundle() &&
-				!(MI.getOperand(0).getTargetFlags() & MO_FLAG_LAST))) {
+  if (MI.isInsideBundle() &&
+      !(MI.getOperand(0).getTargetFlags() & MO_FLAG_LAST)) {
     EmitByte(0);
   } else {
     EmitByte(1);
diff --git a/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
index ba336a37467..9f1b8168a65 100644
--- a/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
+++ b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
@@ -61,7 +61,8 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
 
       bool IsReduction = TII->isReductionOp(MI.getOpcode());
       bool IsVector = TII->isVector(MI);
-      if (!IsReduction && !IsVector) {
+	    bool IsCube = TII->isCubeOp(MI.getOpcode());
+      if (!IsReduction && !IsVector && !IsCube) {
         continue;
       }
 
@@ -82,23 +83,73 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
       // T0_Y (write masked) = MULLO_INT T1_X, T2_X
       // T0_Z (write masked) = MULLO_INT T1_X, T2_X
       // T0_W (write masked) = MULLO_INT T1_X, T2_X
+      //
+      // Cube instructions:
+      // T0_XYZW = CUBE T1_XYZW
+      // becomes:
+      // TO_X = CUBE T1_Z, T1_Y
+      // T0_Y = CUBE T1_Z, T1_X
+      // T0_Z = CUBE T1_X, T1_Z
+      // T0_W = CUBE T1_Y, T1_Z
       for (unsigned Chan = 0; Chan < 4; Chan++) {
         unsigned DstReg = MI.getOperand(0).getReg();
         unsigned Src0 = MI.getOperand(1).getReg();
-        unsigned Src1 = MI.getOperand(2).getReg();
+        unsigned Src1 = 0;
+
+        // Determine the correct source registers
+        if (!IsCube) {
+          Src1 = MI.getOperand(2).getReg();
+        }
         if (IsReduction) {
           unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
           Src0 = TRI.getSubReg(Src0, SubRegIndex);
           Src1 = TRI.getSubReg(Src1, SubRegIndex);
+        } else if (IsCube) {
+          static const int CubeSrcSwz[] = {2, 2, 0, 1};
+          unsigned SubRegIndex0 = TRI.getSubRegFromChannel(CubeSrcSwz[Chan]);
+          unsigned SubRegIndex1 = TRI.getSubRegFromChannel(CubeSrcSwz[3 - Chan]);
+          Src1 = TRI.getSubReg(Src0, SubRegIndex1);
+          Src0 = TRI.getSubReg(Src0, SubRegIndex0);
+        }
+
+        // Determine the correct destination registers;
+        unsigned Flags = 0;
+        if (IsCube) {
+          unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
+          DstReg = TRI.getSubReg(DstReg, SubRegIndex);
+        } else {
+          // Mask the write if the original instruction does not write to
+          // the current Channel.
+          Flags |= (Chan != TRI.getHWRegChan(DstReg) ? MO_FLAG_MASK : 0);
+          unsigned DstBase = TRI.getHWRegIndex(DstReg);
+          DstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
         }
-        unsigned DstBase = TRI.getHWRegIndex(DstReg);
-        unsigned NewDstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
-        unsigned Flags = (Chan != TRI.getHWRegChan(DstReg) ? MO_FLAG_MASK : 0);
+
+        // Set the IsLast bit
         Flags |= (Chan == 3 ? MO_FLAG_LAST : 0);
-        MachineOperand NewDstOp = MachineOperand::CreateReg(NewDstReg, true);
+
+        // Add the new instruction
+        unsigned Opcode;
+        if (IsCube) {
+          switch (MI.getOpcode()) {
+          case AMDGPU::CUBE_r600_pseudo:
+            Opcode = AMDGPU::CUBE_r600_real;
+            break;
+          case AMDGPU::CUBE_eg_pseudo:
+            Opcode = AMDGPU::CUBE_eg_real;
+            break;
+          default:
+            assert(!"Unknown CUBE instruction");
+            Opcode = 0;
+            break;
+          }
+        } else {
+          Opcode = MI.getOpcode();
+        }
+        MachineOperand NewDstOp = MachineOperand::CreateReg(DstReg, true);
         NewDstOp.addTargetFlag(Flags);
 
-        BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(MI.getOpcode()))
+        BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(Opcode))
                 .addOperand(NewDstOp)
                 .addReg(Src0)
                 .addReg(Src1)
diff --git a/src/gallium/drivers/radeon/R600InstrInfo.cpp b/src/gallium/drivers/radeon/R600InstrInfo.cpp
index 12b46654bb7..56a2cf93199 100644
--- a/src/gallium/drivers/radeon/R600InstrInfo.cpp
+++ b/src/gallium/drivers/radeon/R600InstrInfo.cpp
@@ -153,8 +153,10 @@ bool R600InstrInfo::isCubeOp(unsigned opcode) const
 {
   switch(opcode) {
     default: return false;
-    case AMDGPU::CUBE_r600:
-    case AMDGPU::CUBE_eg:
+    case AMDGPU::CUBE_r600_pseudo:
+    case AMDGPU::CUBE_r600_real:
+    case AMDGPU::CUBE_eg_pseudo:
+    case AMDGPU::CUBE_eg_real:
       return true;
   }
 }
diff --git a/src/gallium/drivers/radeon/R600Instructions.td b/src/gallium/drivers/radeon/R600Instructions.td
index f8d2bb0debe..1b02533edbb 100644
--- a/src/gallium/drivers/radeon/R600Instructions.td
+++ b/src/gallium/drivers/radeon/R600Instructions.td
@@ -593,14 +593,25 @@ class DOT4_Common <bits<32> inst> : R600_REDUCTION <
   [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
 >;
 
-class CUBE_Common <bits<32> inst> : InstR600 <
-  inst,
-  (outs R600_Reg128:$dst),
-  (ins R600_Reg128:$src),
-  "CUBE $dst $src",
-  [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
-  VecALU
->;
+multiclass CUBE_Common <bits<32> inst> {
+
+  def _pseudo : InstR600 <
+    inst,
+    (outs R600_Reg128:$dst),
+    (ins R600_Reg128:$src),
+    "CUBE $dst $src",
+    [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
+    VecALU
+  >;
+
+  def _real : InstR600 <
+    inst,
+    (outs R600_Reg32:$dst),
+    (ins R600_Reg32:$src0, R600_Reg32:$src1),
+    "CUBE $dst, $src0, $src1",
+    [], VecALU
+  >;
+}
 
 class EXP_IEEE_Common <bits<32> inst> : R600_1OP <
   inst, "EXP_IEEE",
@@ -737,7 +748,7 @@ let Predicates = [isR600] in {
   def CNDGT_r600 : CNDGT_Common<0x19>;
   def CNDGE_r600 : CNDGE_Common<0x1A>;
   def DOT4_r600 : DOT4_Common<0x50>;
-  def CUBE_r600 : CUBE_Common<0x52>;
+  defm CUBE_r600 : CUBE_Common<0x52>;
   def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
   def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
   def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
@@ -853,7 +864,7 @@ let Predicates = [isEGorCayman] in {
   def SIN_eg : SIN_Common<0x8D>;
   def COS_eg : COS_Common<0x8E>;
   def DOT4_eg : DOT4_Common<0xBE>;
-  def CUBE_eg : CUBE_Common<0xC0>;
+  defm CUBE_eg : CUBE_Common<0xC0>;
 
   def DIV_eg : DIV_Common<RECIP_IEEE_eg>;
   def POW_eg : POW_Common<LOG_IEEE_eg, EXP_IEEE_eg, MUL, GPRF32>;