R600/SI: Fix assertion from copying a TargetGlobalAddress

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

diff --git a/lib/Target/R600/SIInsertWaits.cpp b/lib/Target/R600/SIInsertWaits.cpp
index 24fc9293695cfa6d7400eef1be17577031a5fdd5..23ade02927cb3ed79a2282c78d2cddcfa2236384 100644 (file)
--- a/lib/Target/R600/SIInsertWaits.cpp
+++ b/lib/Target/R600/SIInsertWaits.cpp
@@ -17,6 +17,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
 #include "SIInstrInfo.h"
 #include "SIMachineFunctionInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
@@ -47,7 +48,7 @@ class SIInsertWaits : public MachineFunctionPass {
 private:
   static char ID;
   const SIInstrInfo *TII;
-  const SIRegisterInfo &TRI;
+  const SIRegisterInfo *TRI;
   const MachineRegisterInfo *MRI;
 
   /// \brief Constant hardware limits
@@ -88,18 +89,22 @@ private:
                   MachineBasicBlock::iterator I,
                   const Counters &Counts);
 
+  /// \brief Do we need def2def checks?
+  bool unorderedDefines(MachineInstr &MI);
+
   /// \brief Resolve all operand dependencies to counter requirements
   Counters handleOperands(MachineInstr &MI);
 
 public:
   SIInsertWaits(TargetMachine &tm) :
     MachineFunctionPass(ID),
-    TII(static_cast<const SIInstrInfo*>(tm.getInstrInfo())),
-    TRI(TII->getRegisterInfo()) { }
+    TII(nullptr),
+    TRI(nullptr),
+    ExpInstrTypesSeen(0) { }
 
-  virtual bool runOnMachineFunction(MachineFunction &MF);
+  bool runOnMachineFunction(MachineFunction &MF) override;
 
-  const char *getPassName() const {
+  const char *getPassName() const override {
     return "SI insert wait  instructions";
   }
 
@@ -125,17 +130,24 @@ Counters SIInsertWaits::getHwCounts(MachineInstr &MI) {
 
   // Only consider stores or EXP for EXP_CNT
   Result.Named.EXP = !!(TSFlags & SIInstrFlags::EXP_CNT &&
-      (MI.getOpcode() == AMDGPU::EXP || !MI.getDesc().mayStore()));
+      (MI.getOpcode() == AMDGPU::EXP || MI.getDesc().mayStore()));
 
   // LGKM may uses larger values
   if (TSFlags & SIInstrFlags::LGKM_CNT) {
 
-    MachineOperand &Op = MI.getOperand(0);
-    assert(Op.isReg() && "First LGKM operand must be a register!");
+    if (TII->isSMRD(MI.getOpcode())) {
+
+      MachineOperand &Op = MI.getOperand(0);
+      assert(Op.isReg() && "First LGKM operand must be a register!");
+
+      unsigned Reg = Op.getReg();
+      unsigned Size = TRI->getMinimalPhysRegClass(Reg)->getSize();
+      Result.Named.LGKM = Size > 4 ? 2 : 1;
 
-    unsigned Reg = Op.getReg();
-    unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize();
-    Result.Named.LGKM = Size > 4 ? 2 : 1;
+    } else {
+      // DS
+      Result.Named.LGKM = 1;
+    }
 
   } else {
     Result.Named.LGKM = 0;
@@ -175,16 +187,16 @@ bool SIInsertWaits::isOpRelevant(MachineOperand &Op) {
 
 RegInterval SIInsertWaits::getRegInterval(MachineOperand &Op) {
 
-  if (!Op.isReg())
+  if (!Op.isReg() || !TRI->isInAllocatableClass(Op.getReg()))
     return std::make_pair(0, 0);
 
   unsigned Reg = Op.getReg();
-  unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize();
+  unsigned Size = TRI->getMinimalPhysRegClass(Reg)->getSize();
 
   assert(Size >= 4);
 
   RegInterval Result;
-  Result.first = TRI.getEncodingValue(Reg);
+  Result.first = TRI->getEncodingValue(Reg);
   Result.second = Result.first + Size / 4;
 
   return Result;
@@ -262,17 +274,17 @@ bool SIInsertWaits::insertWait(MachineBasicBlock &MBB,
       continue;
 
     NeedWait = true;
-    
+
     if (Ordered[i]) {
       unsigned Value = LastIssued.Array[i] - Required.Array[i];
 
-      // adjust the value to the real hardware posibilities
+      // Adjust the value to the real hardware possibilities.
       Counts.Array[i] = std::min(Value, WaitCounts.Array[i]);
 
     } else
       Counts.Array[i] = 0;
 
-    // Remember on what we have waited on
+    // Remember on what we have waited on.
     WaitedOn.Array[i] = LastIssued.Array[i] - Counts.Array[i];
   }
 
@@ -303,6 +315,12 @@ Counters SIInsertWaits::handleOperands(MachineInstr &MI) {
 
   Counters Result = ZeroCounts;
 
+  // S_SENDMSG implicitly waits for all outstanding LGKM transfers to finish,
+  // but we also want to wait for any other outstanding transfers before
+  // signalling other hardware blocks
+  if (MI.getOpcode() == AMDGPU::S_SENDMSG)
+    return LastIssued;
+
   // For each register affected by this
   // instruction increase the result sequence
   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
@@ -311,8 +329,10 @@ Counters SIInsertWaits::handleOperands(MachineInstr &MI) {
     RegInterval Interval = getRegInterval(Op);
     for (unsigned j = Interval.first; j < Interval.second; ++j) {
 
-      if (Op.isDef())
+      if (Op.isDef()) {
         increaseCounters(Result, UsedRegs[j]);
+        increaseCounters(Result, DefinedRegs[j]);
+      }
 
       if (Op.isUse())
         increaseCounters(Result, DefinedRegs[j]);
@@ -322,10 +342,15 @@ Counters SIInsertWaits::handleOperands(MachineInstr &MI) {
   return Result;
 }
 
+// FIXME: Insert waits listed in Table 4.2 "Required User-Inserted Wait States"
+// around other non-memory instructions.
 bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
-
   bool Changes = false;
 
+  TII = static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
+  TRI =
+      static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
+
   MRI = &MF.getRegInfo();
 
   WaitedOn = ZeroCounts;