From 6c99f2101fbd3edb7d5899c44ca9d984a3c0f8b6 Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard@amd.com>
Date: Tue, 21 Aug 2012 14:33:04 +0000
Subject: [PATCH] radeon/llvm: ExpandSpecialInstrs - Add support for vector
 instructions

---
 src/gallium/drivers/radeon/R600CodeEmitter.cpp     | 13 +++------
 .../drivers/radeon/R600ExpandSpecialInstrs.cpp     | 32 ++++++++++++++++++----
 2 files changed, 30 insertions(+), 15 deletions(-)

diff --git a/src/gallium/drivers/radeon/R600CodeEmitter.cpp b/src/gallium/drivers/radeon/R600CodeEmitter.cpp
index 14e877b..396ae6f 100644
--- a/src/gallium/drivers/radeon/R600CodeEmitter.cpp
+++ b/src/gallium/drivers/radeon/R600CodeEmitter.cpp
@@ -50,7 +50,6 @@ private:
   const R600InstrInfo * TII;
 
   bool IsCube;
-  bool IsVector;
   unsigned currentElement;
   bool IsLast;
 
@@ -59,7 +58,7 @@ private:
 public:
 
   R600CodeEmitter(formatted_raw_ostream &OS) : MachineFunctionPass(ID),
-      _OS(OS), TM(NULL), IsCube(false), IsVector(false),
+      _OS(OS), TM(NULL), IsCube(false),
       IsLast(true) { }
 
   const char *getPassName() const { return "AMDGPU Machine Code Emitter"; }
@@ -161,7 +160,6 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
      for (MachineBasicBlock::instr_iterator I = MBB.instr_begin(),
                                             E = MBB.instr_end(); I != E; ++I) {
           MachineInstr &MI = *I;
-	  IsVector = TII->isVector(MI);
 	  IsCube = TII->isCubeOp(MI.getOpcode());
           if (MI.getNumOperands() > 1 && MI.getOperand(0).isReg() && MI.getOperand(0).isDead()) {
             continue;
@@ -170,7 +168,7 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
             EmitTexInstr(MI);
           } else if (TII->isFCOp(MI.getOpcode())){
             EmitFCInstr(MI);
-          } else if (IsVector || IsCube) {
+          } else if (IsCube) {
             IsLast = false;
             // XXX: On Cayman, some (all?) of the vector instructions only need
             // to fill the first three slots.
@@ -178,7 +176,6 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
               IsLast = (currentElement == 3);
               EmitALUInstr(MI);
             }
-	    IsVector = false;
 	    IsCube = false;
           } else if (MI.getOpcode() == AMDGPU::RETURN ||
                      MI.getOpcode() == AMDGPU::BUNDLE ||
@@ -348,7 +345,7 @@ void R600CodeEmitter::EmitDst(const MachineOperand & MO)
     EmitByte(getHWReg(MO.getReg()));
 
     // Emit the element of the destination register (1 byte)
-    if (IsCube || IsVector) {
+    if (IsCube) {
       EmitByte(currentElement);
     } else {
       EmitByte(TRI->getHWRegChan(MO.getReg()));
@@ -362,9 +359,7 @@ void R600CodeEmitter::EmitDst(const MachineOperand & MO)
     }
 
     // Emit writemask (1 byte).
-    if ((IsVector &&
-          currentElement != TRI->getHWRegChan(MO.getReg()))
-       || MO.getTargetFlags() & MO_FLAG_MASK) {
+    if (MO.getTargetFlags() & MO_FLAG_MASK) {
       EmitByte(0);
     } else {
       EmitByte(1);
diff --git a/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
index 4c67ba4..ba336a3 100644
--- a/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
+++ b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
@@ -59,18 +59,38 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
       MachineInstr &MI = *I;
       I = llvm::next(I);
 
-      if (!TII->isReductionOp(MI.getOpcode())) {
+      bool IsReduction = TII->isReductionOp(MI.getOpcode());
+      bool IsVector = TII->isVector(MI);
+      if (!IsReduction && !IsVector) {
         continue;
       }
 
       // Expand the instruction
+      //
+      // Reduction instructions:
+      // T0_X = DP4 T1_XYZW, T2_XYZW
+      // becomes:
+      // TO_X = DP4 T1_X, T2_X
+      // TO_Y (write masked) = DP4 T1_Y, T2_Y
+      // TO_Z (write masked) = DP4 T1_Z, T2_Z
+      // TO_W (write masked) = DP4 T1_W, T2_W
+      //
+      // Vector instructions:
+      // T0_X = MULLO_INT T1_X, T2_X
+      // becomes:
+      // T0_X = MULLO_INT T1_X, T2_X
+      // 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
       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 SubRegIndex = TRI.getSubRegFromChannel(Chan);
-        unsigned NewSrc0 = TRI.getSubReg(Src0, SubRegIndex);
-        unsigned NewSrc1 = TRI.getSubReg(Src1, SubRegIndex);
+        if (IsReduction) {
+          unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
+          Src0 = TRI.getSubReg(Src0, SubRegIndex);
+          Src1 = TRI.getSubReg(Src1, SubRegIndex);
+        }
         unsigned DstBase = TRI.getHWRegIndex(DstReg);
         unsigned NewDstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
         unsigned Flags = (Chan != TRI.getHWRegChan(DstReg) ? MO_FLAG_MASK : 0);
@@ -80,8 +100,8 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
 
         BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(MI.getOpcode()))
                 .addOperand(NewDstOp)
-                .addReg(NewSrc0)
-                .addReg(NewSrc1)
+                .addReg(Src0)
+                .addReg(Src1)
                 ->setIsInsideBundle(Chan != 0);
       }
       MI.eraseFromParent();
-- 
2.7.4