radeon/llvm: Add R600ExpandSpecialInstrs pass

This pass expends reduction instructions into a MachineInstrBundle that
contains 4 instruction, one for each instruction slot.

6 files changed, 112 insertions, 14 deletions

diff --git a/src/gallium/drivers/radeon/AMDGPU.h b/src/gallium/drivers/radeon/AMDGPU.h
index 191f495eaa4..927e62a27dd 100644
--- a/src/gallium/drivers/radeon/AMDGPU.h
+++ b/src/gallium/drivers/radeon/AMDGPU.h
@@ -22,6 +22,7 @@ class AMDGPUTargetMachine;
 // R600 Passes
 FunctionPass* createR600KernelParametersPass(const TargetData* TD);
 FunctionPass *createR600CodeEmitterPass(formatted_raw_ostream &OS);
+FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
 
 // SI Passes
 FunctionPass *createSIAssignInterpRegsPass(TargetMachine &tm);
diff --git a/src/gallium/drivers/radeon/AMDGPUInstrInfo.h b/src/gallium/drivers/radeon/AMDGPUInstrInfo.h
index 31400a7b60f..de3c5940613 100644
--- a/src/gallium/drivers/radeon/AMDGPUInstrInfo.h
+++ b/src/gallium/drivers/radeon/AMDGPUInstrInfo.h
@@ -30,6 +30,7 @@
 #define MO_FLAG_ABS   (1 << 2)
 #define MO_FLAG_MASK  (1 << 3)
 #define MO_FLAG_PUSH  (1 << 4)
+#define MO_FLAG_LAST  (1 << 5)
 
 #define OPCODE_IS_ZERO_INT 0x00000045
 #define OPCODE_IS_NOT_ZERO_INT 0x00000042
diff --git a/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp b/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
index b97c0febfc4..6f15430bd53 100644
--- a/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
+++ b/src/gallium/drivers/radeon/AMDGPUTargetMachine.cpp
@@ -158,6 +158,12 @@ bool AMDGPUPassConfig::addPreEmitPass() {
   PM->add(createAMDGPUCFGPreparationPass(*TM));
   PM->add(createAMDGPUCFGStructurizerPass(*TM));
 
+  const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
+  if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
+    PM->add(createR600ExpandSpecialInstrsPass(*TM));
+    addPass(FinalizeMachineBundlesID);
+  }
+
   return false;
 }
 
diff --git a/src/gallium/drivers/radeon/Makefile.sources b/src/gallium/drivers/radeon/Makefile.sources
index 3a75ce96945..0e9825f0fe8 100644
--- a/src/gallium/drivers/radeon/Makefile.sources
+++ b/src/gallium/drivers/radeon/Makefile.sources
@@ -35,6 +35,7 @@ CPP_SOURCES := \
 	AMDGPUInstrInfo.cpp		\
 	AMDGPURegisterInfo.cpp		\
 	R600CodeEmitter.cpp		\
+	R600ExpandSpecialInstrs.cpp	\
 	R600ISelLowering.cpp		\
 	R600InstrInfo.cpp		\
 	R600KernelParameters.cpp	\
diff --git a/src/gallium/drivers/radeon/R600CodeEmitter.cpp b/src/gallium/drivers/radeon/R600CodeEmitter.cpp
index 02b6fdb748b..14e877b2518 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 IsReduction;
   bool IsVector;
   unsigned currentElement;
   bool IsLast;
@@ -60,7 +59,7 @@ private:
 public:
 
   R600CodeEmitter(formatted_raw_ostream &OS) : MachineFunctionPass(ID),
-      _OS(OS), TM(NULL), IsCube(false), IsReduction(false), IsVector(false),
+      _OS(OS), TM(NULL), IsCube(false), IsVector(false),
       IsLast(true) { }
 
   const char *getPassName() const { return "AMDGPU Machine Code Emitter"; }
@@ -159,10 +158,9 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
   for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
                                                   BB != BB_E; ++BB) {
      MachineBasicBlock &MBB = *BB;
-     for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
-                                                       I != E; ++I) {
+     for (MachineBasicBlock::instr_iterator I = MBB.instr_begin(),
+                                            E = MBB.instr_end(); I != E; ++I) {
           MachineInstr &MI = *I;
-	  IsReduction = TII->isReductionOp(MI.getOpcode());
 	  IsVector = TII->isVector(MI);
 	  IsCube = TII->isCubeOp(MI.getOpcode());
           if (MI.getNumOperands() > 1 && MI.getOperand(0).isReg() && MI.getOperand(0).isDead()) {
@@ -172,7 +170,7 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
             EmitTexInstr(MI);
           } else if (TII->isFCOp(MI.getOpcode())){
             EmitFCInstr(MI);
-          } else if (IsReduction || IsVector || IsCube) {
+          } else if (IsVector || IsCube) {
             IsLast = false;
             // XXX: On Cayman, some (all?) of the vector instructions only need
             // to fill the first three slots.
@@ -180,7 +178,6 @@ bool R600CodeEmitter::runOnMachineFunction(MachineFunction &MF) {
               IsLast = (currentElement == 3);
               EmitALUInstr(MI);
             }
-            IsReduction = false;
 	    IsVector = false;
 	    IsCube = false;
           } else if (MI.getOpcode() == AMDGPU::RETURN ||
@@ -310,8 +307,6 @@ void R600CodeEmitter::EmitSrc(const MachineOperand & MO, int chan_override)
   // Emit the source channel (1 byte)
   if (chan_override != -1) {
     EmitByte(chan_override);
-  } else if (IsReduction) {
-    EmitByte(currentElement);
   } else if (MO.isReg()) {
     EmitByte(TRI->getHWRegChan(MO.getReg()));
   } else {
@@ -353,7 +348,7 @@ void R600CodeEmitter::EmitDst(const MachineOperand & MO)
     EmitByte(getHWReg(MO.getReg()));
 
     // Emit the element of the destination register (1 byte)
-    if (IsReduction || IsCube || IsVector) {
+    if (IsCube || IsVector) {
       EmitByte(currentElement);
     } else {
       EmitByte(TRI->getHWRegChan(MO.getReg()));
@@ -367,7 +362,7 @@ void R600CodeEmitter::EmitDst(const MachineOperand & MO)
     }
 
     // Emit writemask (1 byte).
-    if (((IsReduction || IsVector) &&
+    if ((IsVector &&
           currentElement != TRI->getHWRegChan(MO.getReg()))
        || MO.getTargetFlags() & MO_FLAG_MASK) {
       EmitByte(0);
@@ -389,11 +384,14 @@ void R600CodeEmitter::EmitALU(MachineInstr &MI, unsigned numSrc)
   EmitTwoBytes(getBinaryCodeForInstr(MI));
 
   // Emit IsLast (for this instruction group) (1 byte)
-  if (IsLast) {
-    EmitByte(1);
-  } else {
+  if (!IsLast ||
+				(MI.isInsideBundle() &&
+				!(MI.getOperand(0).getTargetFlags() & MO_FLAG_LAST))) {
     EmitByte(0);
+  } else {
+    EmitByte(1);
   }
+
   // Emit isOp3 (1 byte)
   if (numSrc == 3) {
     EmitByte(1);
diff --git a/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
new file mode 100644
index 00000000000..4c67ba47568
--- /dev/null
+++ b/src/gallium/drivers/radeon/R600ExpandSpecialInstrs.cpp
@@ -0,0 +1,91 @@
+//===-- R600ExpandSpecialInstrs.cpp - Expand special instructions ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// Vector, Reduction, and Cube instructions need to fill the entire instruction
+// group to work correctly.  This pass expands these individual instructions
+// into several instructions that will completely fill the instruction group. 
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "R600InstrInfo.h"
+#include "R600RegisterInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+class R600ExpandSpecialInstrsPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const R600InstrInfo *TII;
+
+public:
+  R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID),
+    TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const {
+    return "R600 Expand special instructions pass";
+  }
+};
+
+} // End anonymous namespace
+
+char R600ExpandSpecialInstrsPass::ID = 0;
+
+FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) {
+  return new R600ExpandSpecialInstrsPass(TM);
+}
+
+bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
+
+  const R600RegisterInfo &TRI = TII->getRegisterInfo();
+
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    MachineBasicBlock::iterator I = MBB.begin();
+    while (I != MBB.end()) {
+      MachineInstr &MI = *I;
+      I = llvm::next(I);
+
+      if (!TII->isReductionOp(MI.getOpcode())) {
+        continue;
+      }
+
+      // Expand the instruction
+      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);
+        unsigned DstBase = TRI.getHWRegIndex(DstReg);
+        unsigned NewDstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
+        unsigned Flags = (Chan != TRI.getHWRegChan(DstReg) ? MO_FLAG_MASK : 0);
+        Flags |= (Chan == 3 ? MO_FLAG_LAST : 0);
+        MachineOperand NewDstOp = MachineOperand::CreateReg(NewDstReg, true);
+        NewDstOp.addTargetFlag(Flags);
+
+        BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(MI.getOpcode()))
+                .addOperand(NewDstOp)
+                .addReg(NewSrc0)
+                .addReg(NewSrc1)
+                ->setIsInsideBundle(Chan != 0);
+      }
+      MI.eraseFromParent();
+    }
+  }
+  return false;
+}