It's not necessary to do rounding for alloca operations when the requested
alignment is equal to the stack alignment.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@40004 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/IA64/IA64Bundling.cpp b/lib/Target/IA64/IA64Bundling.cpp
new file mode 100644
index 0000000..6c9fa29
--- /dev/null
+++ b/lib/Target/IA64/IA64Bundling.cpp
@@ -0,0 +1,118 @@
+//===-- IA64Bundling.cpp - IA-64 instruction bundling pass. ------------ --===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file was developed by Duraid Madina and is distributed under the
+// University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Add stops where required to prevent read-after-write and write-after-write
+// dependencies, for both registers and memory addresses. There are exceptions:
+//
+//    - Compare instructions (cmp*, tbit, tnat, fcmp, frcpa) are OK with
+//      WAW dependencies so long as they all target p0, or are of parallel
+//      type (.and*/.or*)
+//
+// FIXME: bundling, for now, is left to the assembler.
+// FIXME: this might be an appropriate place to translate between different
+//        instructions that do the same thing, if this helps bundling.
+// 
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "ia64-codegen"
+#include "IA64.h"
+#include "IA64InstrInfo.h"
+#include "IA64TargetMachine.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/ADT/SetOperations.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/Debug.h"
+#include <set>
+using namespace llvm;
+
+STATISTIC(StopBitsAdded, "Number of stop bits added");
+
+namespace {
+  struct IA64BundlingPass : public MachineFunctionPass {
+    static char ID;
+    /// Target machine description which we query for reg. names, data
+    /// layout, etc.
+    ///
+    IA64TargetMachine &TM;
+
+    IA64BundlingPass(IA64TargetMachine &tm) 
+      : MachineFunctionPass((intptr_t)&ID), TM(tm) { }
+
+    virtual const char *getPassName() const {
+      return "IA64 (Itanium) Bundling Pass";
+    }
+
+    bool runOnMachineBasicBlock(MachineBasicBlock &MBB);
+    bool runOnMachineFunction(MachineFunction &F) {
+      bool Changed = false;
+      for (MachineFunction::iterator FI = F.begin(), FE = F.end();
+           FI != FE; ++FI)
+        Changed |= runOnMachineBasicBlock(*FI);
+      return Changed;
+    }
+
+    // XXX: ugly global, but pending writes can cross basic blocks. Note that
+    // taken branches end instruction groups. So we only need to worry about
+    // 'fallthrough' code
+    std::set<unsigned> PendingRegWrites;
+  };
+  char IA64BundlingPass::ID = 0;
+} // end of anonymous namespace
+
+/// createIA64BundlingPass - Returns a pass that adds STOP (;;) instructions
+/// and arranges the result into bundles.
+///
+FunctionPass *llvm::createIA64BundlingPass(IA64TargetMachine &tm) {
+  return new IA64BundlingPass(tm);
+}
+
+/// runOnMachineBasicBlock - add stops and bundle this MBB.
+///
+bool IA64BundlingPass::runOnMachineBasicBlock(MachineBasicBlock &MBB) {
+  bool Changed = false;
+
+  for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ) {
+    MachineInstr *CurrentInsn = I++;
+    std::set<unsigned> CurrentReads, CurrentWrites, OrigWrites;
+
+    for(unsigned i=0; i < CurrentInsn->getNumOperands(); i++) {
+      MachineOperand &MO=CurrentInsn->getOperand(i);
+      if(MO.isRegister()) {
+        if(MO.isUse()) { // TODO: exclude p0
+          CurrentReads.insert(MO.getReg());
+        }
+        if(MO.isDef()) { // TODO: exclude p0
+          CurrentWrites.insert(MO.getReg());
+          OrigWrites.insert(MO.getReg()); // FIXME: use a nondestructive
+                                          // set_intersect instead?
+        }
+      }
+    }
+    
+    // CurrentReads/CurrentWrites contain info for the current instruction.
+    // Does it read or write any registers that are pending a write?
+    // (i.e. not separated by a stop)
+    set_intersect(CurrentReads, PendingRegWrites);
+    set_intersect(CurrentWrites, PendingRegWrites);
+    
+    if(! (CurrentReads.empty() && CurrentWrites.empty()) ) {
+      // there is a conflict, insert a stop and reset PendingRegWrites
+      CurrentInsn = BuildMI(MBB, CurrentInsn,
+                            TM.getInstrInfo()->get(IA64::STOP), 0);
+      PendingRegWrites=OrigWrites; // carry over current writes to next insn
+      Changed=true; StopBitsAdded++; // update stats      
+    } else { // otherwise, track additional pending writes
+      set_union(PendingRegWrites, OrigWrites);
+    }
+  } // onto the next insn in the MBB
+
+  return Changed;
+}
+