Add a new pre-allocation pass to assign adjacent registers for Neon instructions

that have that constraint.  This is currently just assigning a fixed set of
registers, and it only handles VLDn for n=2,3,4 with DPR registers.
I'm going to expand it to handle more operations next; we can make it smarter
once everything is working correctly.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@78256 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARM.h b/lib/Target/ARM/ARM.h
index d947678828ff134fe80f748962da8ad5dfdde9fd..582879bf025e491fe18b56dfdc1846d3413443cc 100644 (file)
--- a/lib/Target/ARM/ARM.h
+++ b/lib/Target/ARM/ARM.h
@@ -103,7 +103,7 @@ FunctionPass *createARMObjectCodeEmitterPass(ARMBaseTargetMachine &TM,
 
 FunctionPass *createARMLoadStoreOptimizationPass(bool PreAlloc = false);
 FunctionPass *createARMConstantIslandPass();
-
+FunctionPass *createNEONPreAllocPass();
 FunctionPass *createThumb2ITBlockPass();
 
 extern Target TheARMTarget, TheThumbTarget;

diff --git a/lib/Target/ARM/ARMTargetMachine.cpp b/lib/Target/ARM/ARMTargetMachine.cpp
index a6e987b086b521e97ad5f04a739048875dd1ecc8..a58089088eb17a43681e140962bb46eb5cf93204 100644 (file)
--- a/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/lib/Target/ARM/ARMTargetMachine.cpp
@@ -93,6 +93,9 @@ bool ARMBaseTargetMachine::addInstSelector(PassManagerBase &PM,
 
 bool ARMBaseTargetMachine::addPreRegAlloc(PassManagerBase &PM,
                                           CodeGenOpt::Level OptLevel) {
+  if (Subtarget.hasNEON())
+    PM.add(createNEONPreAllocPass());
+
   // FIXME: temporarily disabling load / store optimization pass for Thumb mode.
   if (OptLevel != CodeGenOpt::None && !DisableLdStOpti && !Subtarget.isThumb())
     PM.add(createARMLoadStoreOptimizationPass(true));

diff --git a/lib/Target/ARM/CMakeLists.txt b/lib/Target/ARM/CMakeLists.txt
index f40a3bc59a097dec52cbafdf6df72f59f6bd83f9..033da81a1883ca681e00dfb1bba4de02fcd39a34 100644 (file)
--- a/lib/Target/ARM/CMakeLists.txt
+++ b/lib/Target/ARM/CMakeLists.txt
@@ -26,6 +26,7 @@ add_llvm_target(ARMCodeGen
   ARMSubtarget.cpp
   ARMTargetAsmInfo.cpp
   ARMTargetMachine.cpp
+  NEONPreAllocPass.cpp
   Thumb1InstrInfo.cpp
   Thumb1RegisterInfo.cpp
   Thumb2ITBlockPass.cpp

diff --git a/lib/Target/ARM/NEONPreAllocPass.cpp b/lib/Target/ARM/NEONPreAllocPass.cpp
new file mode 100644 (file)
index 0000000..1662b76
--- /dev/null
+++ b/lib/Target/ARM/NEONPreAllocPass.cpp
@@ -0,0 +1,137 @@
+//===-- NEONPreAllocPass.cpp - Allocate adjacent NEON registers--*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "neon-prealloc"
+#include "ARM.h"
+#include "ARMInstrInfo.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+using namespace llvm;
+
+namespace {
+  class VISIBILITY_HIDDEN NEONPreAllocPass : public MachineFunctionPass {
+    const TargetInstrInfo *TII;
+
+  public:
+    static char ID;
+    NEONPreAllocPass() : MachineFunctionPass(&ID) {}
+
+    virtual bool runOnMachineFunction(MachineFunction &MF);
+
+    virtual const char *getPassName() const {
+      return "NEON register pre-allocation pass";
+    }
+
+  private:
+    bool PreAllocNEONRegisters(MachineBasicBlock &MBB);
+  };
+
+  char NEONPreAllocPass::ID = 0;
+}
+
+static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd,
+                             unsigned &NumRegs) {
+  switch (Opcode) {
+  default:
+    break;
+
+  case ARM::VLD2d8:
+  case ARM::VLD2d16:
+  case ARM::VLD2d32:
+  case ARM::VLD2d64:
+    FirstOpnd = 0;
+    NumRegs = 2;
+    return true;
+
+  case ARM::VLD3d8:
+  case ARM::VLD3d16:
+  case ARM::VLD3d32:
+  case ARM::VLD3d64:
+    FirstOpnd = 0;
+    NumRegs = 3;
+    return true;
+
+  case ARM::VLD4d8:
+  case ARM::VLD4d16:
+  case ARM::VLD4d32:
+  case ARM::VLD4d64:
+    FirstOpnd = 0;
+    NumRegs = 4;
+    return true;
+  }
+
+  return false;
+}
+
+bool NEONPreAllocPass::PreAllocNEONRegisters(MachineBasicBlock &MBB) {
+  bool Modified = false;
+
+  MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
+  for (; MBBI != E; ++MBBI) {
+    MachineInstr *MI = &*MBBI;
+    unsigned FirstOpnd, NumRegs;
+    if (!isNEONMultiRegOp(MI->getOpcode(), FirstOpnd, NumRegs))
+      continue;
+
+    MachineBasicBlock::iterator NextI = next(MBBI);
+    for (unsigned R = 0; R < NumRegs; ++R) {
+      MachineOperand &MO = MI->getOperand(FirstOpnd + R);
+      assert(MO.isReg() && MO.getSubReg() == 0 && "unexpected operand");
+      unsigned VirtReg = MO.getReg();
+      assert(TargetRegisterInfo::isVirtualRegister(VirtReg) &&
+             "expected a virtual register");
+
+      // For now, just assign a fixed set of adjacent registers.
+      // This leaves plenty of room for future improvements.
+      static const unsigned NEONDRegs[] = {
+        ARM::D0, ARM::D1, ARM::D2, ARM::D3
+      };
+      MO.setReg(NEONDRegs[R]);
+
+      if (MO.isUse()) {
+        // Insert a copy from VirtReg.
+        AddDefaultPred(BuildMI(MBB, MBBI, MI->getDebugLoc(),
+                               TII->get(ARM::FCPYD), MO.getReg())
+                       .addReg(VirtReg));
+        if (MO.isKill()) {
+          MachineInstr *CopyMI = prior(MBBI);
+          CopyMI->findRegisterUseOperand(VirtReg)->setIsKill();
+        }
+        MO.setIsKill();
+      } else if (MO.isDef() && !MO.isDead()) {
+        // Add a copy to VirtReg.
+        AddDefaultPred(BuildMI(MBB, NextI, MI->getDebugLoc(),
+                               TII->get(ARM::FCPYD), VirtReg)
+                       .addReg(MO.getReg()));
+      }
+    }
+  }
+
+  return Modified;
+}
+
+bool NEONPreAllocPass::runOnMachineFunction(MachineFunction &MF) {
+  TII = MF.getTarget().getInstrInfo();
+
+  bool Modified = false;
+  for (MachineFunction::iterator MFI = MF.begin(), E = MF.end(); MFI != E;
+       ++MFI) {
+    MachineBasicBlock &MBB = *MFI;
+    Modified |= PreAllocNEONRegisters(MBB);
+  }
+
+  return Modified;
+}
+
+/// createNEONPreAllocPass - returns an instance of the NEON register
+/// pre-allocation pass.
+FunctionPass *llvm::createNEONPreAllocPass() {
+  return new NEONPreAllocPass();
+}