[X86][SSE] Vector integer/float conversion memory folding

[oota-llvm.git] / lib / Target / R600 / R600ExpandSpecialInstrs.cpp

diff --git a/lib/Target/R600/R600ExpandSpecialInstrs.cpp b/lib/Target/R600/R600ExpandSpecialInstrs.cpp
index a1919305111ba625d9c1960344547e0212bed715..211d392e8fcc982cf23d0323330a4b07ac59d6d6 100644 (file)
--- a/lib/Target/R600/R600ExpandSpecialInstrs.cpp
+++ b/lib/Target/R600/R600ExpandSpecialInstrs.cpp
@@ -19,6 +19,7 @@
 #include "R600InstrInfo.h"
 #include "R600MachineFunctionInfo.h"
 #include "R600RegisterInfo.h"
+#include "AMDGPUSubtarget.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
@@ -33,16 +34,16 @@ private:
   static char ID;
   const R600InstrInfo *TII;
 
-  bool ExpandInputPerspective(MachineInstr& MI);
-  bool ExpandInputConstant(MachineInstr& MI);
+  void SetFlagInNewMI(MachineInstr *NewMI, const MachineInstr *OldMI,
+      unsigned Op);
 
 public:
   R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID),
-    TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+    TII(nullptr) { }
 
-  virtual bool runOnMachineFunction(MachineFunction &MF);
+  bool runOnMachineFunction(MachineFunction &MF) override;
 
-  const char *getPassName() const {
+  const char *getPassName() const override {
     return "R600 Expand special instructions pass";
   }
 };
@@ -55,7 +56,17 @@ FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) {
   return new R600ExpandSpecialInstrsPass(TM);
 }
 
+void R600ExpandSpecialInstrsPass::SetFlagInNewMI(MachineInstr *NewMI,
+    const MachineInstr *OldMI, unsigned Op) {
+  int OpIdx = TII->getOperandIdx(*OldMI, Op);
+  if (OpIdx > -1) {
+    uint64_t Val = OldMI->getOperand(OpIdx).getImm();
+    TII->setImmOperand(NewMI, Op, Val);
+  }
+}
+
 bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
+  TII = static_cast<const R600InstrInfo *>(MF.getSubtarget().getInstrInfo());
 
   const R600RegisterInfo &TRI = TII->getRegisterInfo();
 
@@ -65,7 +76,24 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
     MachineBasicBlock::iterator I = MBB.begin();
     while (I != MBB.end()) {
       MachineInstr &MI = *I;
-      I = llvm::next(I);
+      I = std::next(I);
+
+      // Expand LDS_*_RET instructions
+      if (TII->isLDSRetInstr(MI.getOpcode())) {
+        int DstIdx = TII->getOperandIdx(MI.getOpcode(), AMDGPU::OpName::dst);
+        assert(DstIdx != -1);
+        MachineOperand &DstOp = MI.getOperand(DstIdx);
+        MachineInstr *Mov = TII->buildMovInstr(&MBB, I,
+                                               DstOp.getReg(), AMDGPU::OQAP);
+        DstOp.setReg(AMDGPU::OQAP);
+        int LDSPredSelIdx = TII->getOperandIdx(MI.getOpcode(),
+                                           AMDGPU::OpName::pred_sel);
+        int MovPredSelIdx = TII->getOperandIdx(Mov->getOpcode(),
+                                           AMDGPU::OpName::pred_sel);
+        // Copy the pred_sel bit
+        Mov->getOperand(MovPredSelIdx).setReg(
+            MI.getOperand(LDSPredSelIdx).getReg());
+      }
 
       switch (MI.getOpcode()) {
       default: break;
@@ -81,25 +109,10 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
                                             AMDGPU::ZERO);             // src1
         TII->addFlag(PredSet, 0, MO_FLAG_MASK);
         if (Flags & MO_FLAG_PUSH) {
-          TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1);
+          TII->setImmOperand(PredSet, AMDGPU::OpName::update_exec_mask, 1);
         } else {
-          TII->setImmOperand(PredSet, R600Operands::UPDATE_PREDICATE, 1);
-        }
-        MI.eraseFromParent();
-        continue;
+          TII->setImmOperand(PredSet, AMDGPU::OpName::update_pred, 1);
         }
-      case AMDGPU::BREAK: {
-        MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
-                                          AMDGPU::PRED_SETE_INT,
-                                          AMDGPU::PREDICATE_BIT,
-                                          AMDGPU::ZERO,
-                                          AMDGPU::ZERO);
-        TII->addFlag(PredSet, 0, MO_FLAG_MASK);
-        TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1);
-
-        BuildMI(MBB, I, MBB.findDebugLoc(I),
-                TII->get(AMDGPU::PREDICATED_BREAK))
-                .addReg(AMDGPU::PREDICATE_BIT);
         MI.eraseFromParent();
         continue;
         }
@@ -207,12 +220,16 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
           // While not strictly necessary from hw point of view, we force
           // all src operands of a dot4 inst to belong to the same slot.
           unsigned Src0 = BMI->getOperand(
-              TII->getOperandIdx(Opcode, R600Operands::SRC0))
+              TII->getOperandIdx(Opcode, AMDGPU::OpName::src0))
               .getReg();
           unsigned Src1 = BMI->getOperand(
-              TII->getOperandIdx(Opcode, R600Operands::SRC1))
+              TII->getOperandIdx(Opcode, AMDGPU::OpName::src1))
               .getReg();
-          assert(TRI.getHWRegChan(Src0) == TRI.getHWRegChan(Src1));
+          (void) Src0;
+          (void) Src1;
+          if ((TRI.getEncodingValue(Src0) & 0xff) < 127 &&
+              (TRI.getEncodingValue(Src1) & 0xff) < 127)
+            assert(TRI.getHWRegChan(Src0) == TRI.getHWRegChan(Src1));
         }
         MI.eraseFromParent();
         continue;
@@ -253,14 +270,14 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
       // T0_W = CUBE T1_Y, T1_Z
       for (unsigned Chan = 0; Chan < 4; Chan++) {
         unsigned DstReg = MI.getOperand(
-                            TII->getOperandIdx(MI, R600Operands::DST)).getReg();
+                            TII->getOperandIdx(MI, AMDGPU::OpName::dst)).getReg();
         unsigned Src0 = MI.getOperand(
-                           TII->getOperandIdx(MI, R600Operands::SRC0)).getReg();
+                           TII->getOperandIdx(MI, AMDGPU::OpName::src0)).getReg();
         unsigned Src1 = 0;
 
         // Determine the correct source registers
         if (!IsCube) {
-          int Src1Idx = TII->getOperandIdx(MI, R600Operands::SRC1);
+          int Src1Idx = TII->getOperandIdx(MI, AMDGPU::OpName::src1);
           if (Src1Idx != -1) {
             Src1 = MI.getOperand(Src1Idx).getReg();
           }
@@ -318,6 +335,12 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
         if (NotLast) {
           TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
         }
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::clamp);
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::literal);
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::src0_abs);
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::src1_abs);
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::src0_neg);
+        SetFlagInNewMI(NewMI, &MI, AMDGPU::OpName::src1_neg);
       }
       MI.eraseFromParent();
     }