-// $Id$
-//***************************************************************************
-// File:
-// Sparc.cpp
-//
-// Purpose:
-//
-// History:
-// 7/15/01 - Vikram Adve - Created
-//**************************************************************************/
-
+//===-- Sparc.cpp - General implementation file for the Sparc Target ------===//
+//
+// This file contains the code for the Sparc Target that does not fit in any of
+// the other files in this directory.
+//
+//===----------------------------------------------------------------------===//
#include "SparcInternals.h"
-#include "llvm/Target/Sparc.h"
-#include "llvm/CodeGen/InstrScheduling.h"
+#include "llvm/Target/TargetMachineImpls.h"
+#include "llvm/Function.h"
+#include "llvm/PassManager.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionInfo.h"
+#include "llvm/CodeGen/PreSelection.h"
+#include "llvm/CodeGen/StackSlots.h"
+#include "llvm/CodeGen/PeepholeOpts.h"
#include "llvm/CodeGen/InstrSelection.h"
+#include "llvm/CodeGen/InstrScheduling.h"
+#include "llvm/CodeGen/RegisterAllocation.h"
#include "llvm/CodeGen/MachineCodeForInstruction.h"
-#include "llvm/CodeGen/MachineCodeForMethod.h"
-#include "llvm/CodeGen/PhyRegAlloc.h"
-#include "llvm/Method.h"
-#include <iostream>
-using std::cerr;
+#include "llvm/Reoptimizer/Mapping/MappingInfo.h"
+#include "Support/CommandLine.h"
+#include "llvm/Assembly/PrintModulePass.h"
+static const unsigned ImplicitRegUseList[] = { 0 }; /* not used yet */
// Build the MachineInstruction Description Array...
-const MachineInstrDescriptor SparcMachineInstrDesc[] = {
+const TargetInstrDescriptor SparcMachineInstrDesc[] = {
#define I(ENUM, OPCODESTRING, NUMOPERANDS, RESULTPOS, MAXIMM, IMMSE, \
NUMDELAYSLOTS, LATENCY, SCHEDCLASS, INSTFLAGS) \
{ OPCODESTRING, NUMOPERANDS, RESULTPOS, MAXIMM, IMMSE, \
- NUMDELAYSLOTS, LATENCY, SCHEDCLASS, INSTFLAGS },
+ NUMDELAYSLOTS, LATENCY, SCHEDCLASS, INSTFLAGS, 0, \
+ ImplicitRegUseList, ImplicitRegUseList },
#include "SparcInstr.def"
};
-//----------------------------------------------------------------------------
-// allocateSparcTargetMachine - Allocate and return a subclass of TargetMachine
-// that implements the Sparc backend. (the llvm/CodeGen/Sparc.h interface)
-//----------------------------------------------------------------------------
-//
-
-TargetMachine *allocateSparcTargetMachine() { return new UltraSparc(); }
-
-
-//----------------------------------------------------------------------------
-// Entry point for register allocation for a module
-//----------------------------------------------------------------------------
-
-void AllocateRegisters(Method *M, TargetMachine &target)
-{
-
- if ( (M)->isExternal() ) // don't process prototypes
- return;
-
- if( DEBUG_RA )
- cerr << "\n******************** Method "<< M->getName()
- << " ********************\n";
-
- MethodLiveVarInfo LVI(M ); // Analyze live varaibles
- LVI.analyze();
-
-
- PhyRegAlloc PRA(M, target, &LVI); // allocate registers
- PRA.allocateRegisters();
-
-
- if( DEBUG_RA ) cerr << "\nRegister allocation complete!\n";
-
-}
-
-
//---------------------------------------------------------------------------
-// Function InsertPrologCode
-// Function InsertEpilogCode
-// Function InsertPrologEpilog
-//
-// Insert prolog code at the unique method entry point.
-// Insert epilog code at each method exit point.
-// InsertPrologEpilog invokes these only if the method is not compiled
-// with the leaf method optimization.
+// Command line options to control choice of code generation passes.
//---------------------------------------------------------------------------
-static MachineInstr* minstrVec[MAX_INSTR_PER_VMINSTR];
+static cl::opt<bool> DisablePreOpt("disable-preopt",
+ cl::desc("Disable optimizations prior to instruction selection"));
-static void
-InsertPrologCode(Method* method, TargetMachine& target)
-{
- BasicBlock* entryBB = method->getEntryNode();
- unsigned N = GetInstructionsForProlog(entryBB, target, minstrVec);
- assert(N <= MAX_INSTR_PER_VMINSTR);
- if (N > 0)
- {
- MachineCodeForBasicBlock& bbMvec = entryBB->getMachineInstrVec();
- bbMvec.insert(bbMvec.begin(), minstrVec, minstrVec+N);
- }
-}
+static cl::opt<bool> DisableSched("disable-sched",
+ cl::desc("Disable local scheduling pass"));
+static cl::opt<bool> DisablePeephole("disable-peephole",
+ cl::desc("Disable peephole optimization pass"));
-static void
-InsertEpilogCode(Method* method, TargetMachine& target)
-{
- for (Method::iterator I=method->begin(), E=method->end(); I != E; ++I)
- if ((*I)->getTerminator()->getOpcode() == Instruction::Ret)
- {
- BasicBlock* exitBB = *I;
- unsigned N = GetInstructionsForEpilog(exitBB, target, minstrVec);
-
- MachineCodeForBasicBlock& bbMvec = exitBB->getMachineInstrVec();
- MachineCodeForInstruction &termMvec =
- MachineCodeForInstruction::get(exitBB->getTerminator());
-
- // Remove the NOPs in the delay slots of the return instruction
- const MachineInstrInfo &mii = target.getInstrInfo();
- unsigned numNOPs = 0;
- while (termMvec.back()->getOpCode() == NOP)
- {
- assert( termMvec.back() == bbMvec.back());
- termMvec.pop_back();
- bbMvec.pop_back();
- ++numNOPs;
- }
- assert(termMvec.back() == bbMvec.back());
-
- // Check that we found the right number of NOPs and have the right
- // number of instructions to replace them.
- unsigned ndelays = mii.getNumDelaySlots(termMvec.back()->getOpCode());
- assert(numNOPs == ndelays && "Missing NOPs in delay slots?");
- assert(N == ndelays && "Cannot use epilog code for delay slots?");
-
- // Append the epilog code to the end of the basic block.
- bbMvec.push_back(minstrVec[0]);
- }
-}
-
-
-// Insert SAVE/RESTORE instructions for the method
-static void
-InsertPrologEpilog(Method *method, TargetMachine &target)
-{
- MachineCodeForMethod& mcodeInfo = MachineCodeForMethod::get(method);
- if (mcodeInfo.isCompiledAsLeafMethod())
- return; // nothing to do
-
- InsertPrologCode(method, target);
- InsertEpilogCode(method, target);
-}
+static cl::opt<bool> EmitMappingInfo("emitmaps",
+ cl::desc("Emit LLVM-to-MachineCode mapping info to assembly"));
+static cl::opt<bool> DisableStrip("disable-strip",
+ cl::desc("Do not strip the LLVM bytecode included in executable"));
-//---------------------------------------------------------------------------
-// class UltraSparcSchedInfo
-//
-// Purpose:
-// Scheduling information for the UltraSPARC.
-// Primarily just initializes machine-dependent parameters in
-// class MachineSchedInfo.
-//---------------------------------------------------------------------------
+static cl::opt<bool> DumpInput("dump-input",
+ cl::desc("Print bytecode before native code generation"),
+ cl::Hidden);
-/*ctor*/
-UltraSparcSchedInfo::UltraSparcSchedInfo(const TargetMachine& tgt)
- : MachineSchedInfo(tgt,
- (unsigned int) SPARC_NUM_SCHED_CLASSES,
- SparcRUsageDesc,
- SparcInstrUsageDeltas,
- SparcInstrIssueDeltas,
- sizeof(SparcInstrUsageDeltas)/sizeof(InstrRUsageDelta),
- sizeof(SparcInstrIssueDeltas)/sizeof(InstrIssueDelta))
-{
- maxNumIssueTotal = 4;
- longestIssueConflict = 0; // computed from issuesGaps[]
-
- branchMispredictPenalty = 4; // 4 for SPARC IIi
- branchTargetUnknownPenalty = 2; // 2 for SPARC IIi
- l1DCacheMissPenalty = 8; // 7 or 9 for SPARC IIi
- l1ICacheMissPenalty = 8; // ? for SPARC IIi
-
- inOrderLoads = true; // true for SPARC IIi
- inOrderIssue = true; // true for SPARC IIi
- inOrderExec = false; // false for most architectures
- inOrderRetire= true; // true for most architectures
-
- // must be called after above parameters are initialized.
- this->initializeResources();
-}
+//----------------------------------------------------------------------------
+// allocateSparcTargetMachine - Allocate and return a subclass of TargetMachine
+// that implements the Sparc backend. (the llvm/CodeGen/Sparc.h interface)
+//----------------------------------------------------------------------------
-void
-UltraSparcSchedInfo::initializeResources()
-{
- // Compute MachineSchedInfo::instrRUsages and MachineSchedInfo::issueGaps
- MachineSchedInfo::initializeResources();
-
- // Machine-dependent fixups go here. None for now.
+TargetMachine *allocateSparcTargetMachine(unsigned Configuration) {
+ return new UltraSparc();
}
-
//---------------------------------------------------------------------------
// class UltraSparcFrameInfo
//
-// Purpose:
// Interface to stack frame layout info for the UltraSPARC.
// Starting offsets for each area of the stack frame are aligned at
// a multiple of getStackFrameSizeAlignment().
//---------------------------------------------------------------------------
int
-UltraSparcFrameInfo::getFirstAutomaticVarOffset(MachineCodeForMethod& ,
+UltraSparcFrameInfo::getFirstAutomaticVarOffset(MachineFunction& ,
bool& pos) const
{
pos = false; // static stack area grows downwards
}
int
-UltraSparcFrameInfo::getRegSpillAreaOffset(MachineCodeForMethod& mcInfo,
+UltraSparcFrameInfo::getRegSpillAreaOffset(MachineFunction& mcInfo,
bool& pos) const
{
+ // ensure no more auto vars are added
+ mcInfo.getInfo()->freezeAutomaticVarsArea();
+
pos = false; // static stack area grows downwards
- unsigned int autoVarsSize = mcInfo.getAutomaticVarsSize();
- if (int mod = autoVarsSize % getStackFrameSizeAlignment())
- autoVarsSize += (getStackFrameSizeAlignment() - mod);
+ unsigned autoVarsSize = mcInfo.getInfo()->getAutomaticVarsSize();
return StaticAreaOffsetFromFP - autoVarsSize;
}
int
-UltraSparcFrameInfo::getTmpAreaOffset(MachineCodeForMethod& mcInfo,
+UltraSparcFrameInfo::getTmpAreaOffset(MachineFunction& mcInfo,
bool& pos) const
{
+ MachineFunctionInfo *MFI = mcInfo.getInfo();
+ MFI->freezeAutomaticVarsArea(); // ensure no more auto vars are added
+ MFI->freezeSpillsArea(); // ensure no more spill slots are added
+
pos = false; // static stack area grows downwards
- unsigned int autoVarsSize = mcInfo.getAutomaticVarsSize();
- unsigned int spillAreaSize = mcInfo.getRegSpillsSize();
+ unsigned autoVarsSize = MFI->getAutomaticVarsSize();
+ unsigned spillAreaSize = MFI->getRegSpillsSize();
int offset = autoVarsSize + spillAreaSize;
- if (int mod = offset % getStackFrameSizeAlignment())
- offset += (getStackFrameSizeAlignment() - mod);
return StaticAreaOffsetFromFP - offset;
}
int
-UltraSparcFrameInfo::getDynamicAreaOffset(MachineCodeForMethod& mcInfo,
+UltraSparcFrameInfo::getDynamicAreaOffset(MachineFunction& mcInfo,
bool& pos) const
{
- // dynamic stack area grows downwards starting at top of opt-args area
- unsigned int optArgsSize = mcInfo.getMaxOptionalArgsSize();
+ // Dynamic stack area grows downwards starting at top of opt-args area.
+ // The opt-args, required-args, and register-save areas are empty except
+ // during calls and traps, so they are shifted downwards on each
+ // dynamic-size alloca.
+ pos = false;
+ unsigned optArgsSize = mcInfo.getInfo()->getMaxOptionalArgsSize();
+ if (int extra = optArgsSize % getStackFrameSizeAlignment())
+ optArgsSize += (getStackFrameSizeAlignment() - extra);
int offset = optArgsSize + FirstOptionalOutgoingArgOffsetFromSP;
- assert(offset % getStackFrameSizeAlignment() == 0);
+ assert((offset - OFFSET) % getStackFrameSizeAlignment() == 0);
return offset;
}
-
//---------------------------------------------------------------------------
// class UltraSparcMachine
//
// Primary interface to machine description for the UltraSPARC.
// Primarily just initializes machine-dependent parameters in
// class TargetMachine, and creates machine-dependent subclasses
-// for classes such as MachineInstrInfo.
+// for classes such as TargetInstrInfo.
//
//---------------------------------------------------------------------------
UltraSparc::UltraSparc()
- : TargetMachine("UltraSparc-Native"),
- instrInfo(*this),
+ : TargetMachine("UltraSparc-Native", false),
schedInfo(*this),
regInfo(*this),
frameInfo(*this),
- cacheInfo(*this)
-{
- optSizeForSubWordData = 4;
- minMemOpWordSize = 8;
- maxAtomicMemOpWordSize = 8;
-}
-
-
-void
-ApplyPeepholeOptimizations(Method *method, TargetMachine &target)
-{
- return;
-
- // OptimizeLeafProcedures();
- // DeleteFallThroughBranches();
- // RemoveChainedBranches(); // should be folded with previous
- // RemoveRedundantOps(); // operations with %g0, NOP, etc.
+ cacheInfo(*this),
+ optInfo(*this) {
}
-
-bool
-UltraSparc::compileMethod(Method *method)
+// addPassesToEmitAssembly - This method controls the entire code generation
+// process for the ultra sparc.
+//
+bool UltraSparc::addPassesToEmitAssembly(PassManager &PM, std::ostream &Out)
{
- // Construct and initialize the MachineCodeForMethod object for this method.
- MachineCodeForMethod::construct(method, *this);
-
- if (SelectInstructionsForMethod(method, *this))
- {
- cerr << "Instruction selection failed for method " << method->getName()
- << "\n\n";
- return true;
- }
-
- /*
- if (ScheduleInstructionsWithSSA(method, *this))
- {
- cerr << "Instruction scheduling before allocation failed for method "
- << method->getName() << "\n\n";
- return true;
- }
- */
+ // The following 3 passes used to be inserted specially by llc.
+ // Replace malloc and free instructions with library calls.
+ PM.add(createLowerAllocationsPass());
+ // Strip all of the symbols from the bytecode so that it will be smaller...
+ if (!DisableStrip)
+ PM.add(createSymbolStrippingPass());
- AllocateRegisters(method, *this); // allocate registers
+ // FIXME: implement the switch instruction in the instruction selector.
+ PM.add(createLowerSwitchPass());
- ApplyPeepholeOptimizations(method, *this); // machine-dependent peephole opts
+ // decompose multi-dimensional array references into single-dim refs
+ PM.add(createDecomposeMultiDimRefsPass());
- InsertPrologEpilog(method, *this);
+ // Construct and initialize the MachineFunction object for this fn.
+ PM.add(createMachineCodeConstructionPass(*this));
+
+ //Insert empty stackslots in the stack frame of each function
+ //so %fp+offset-8 and %fp+offset-16 are empty slots now!
+ PM.add(createStackSlotsPass(*this));
+
+ if (!DisablePreOpt) {
+ // Specialize LLVM code for this target machine
+ PM.add(createPreSelectionPass(*this));
+ // Run basic dataflow optimizations on LLVM code
+ PM.add(createReassociatePass());
+ PM.add(createLICMPass());
+ PM.add(createGCSEPass());
+ }
- return false;
-}
+ // If LLVM dumping after transformations is requested, add it to the pipeline
+ if (DumpInput)
+ PM.add(new PrintFunctionPass("Input code to instsr. selection:\n",
+ &std::cerr));
-static void freeMachineCode(Instruction *I) {
- MachineCodeForInstruction::destroy(I);
+ PM.add(createInstructionSelectionPass(*this));
+
+ if (!DisableSched)
+ PM.add(createInstructionSchedulingWithSSAPass(*this));
+
+ PM.add(getRegisterAllocator(*this));
+
+ PM.add(getPrologEpilogInsertionPass());
+
+ if (!DisablePeephole)
+ PM.add(createPeepholeOptsPass(*this));
+
+ if (EmitMappingInfo)
+ PM.add(getMappingInfoCollector(Out));
+
+ // Output assembly language to the .s file. Assembly emission is split into
+ // two parts: Function output and Global value output. This is because
+ // function output is pipelined with all of the rest of code generation stuff,
+ // allowing machine code representations for functions to be free'd after the
+ // function has been emitted.
+ //
+ PM.add(getFunctionAsmPrinterPass(Out));
+ PM.add(createMachineCodeDestructionPass()); // Free stuff no longer needed
+
+ // Emit Module level assembly after all of the functions have been processed.
+ PM.add(getModuleAsmPrinterPass(Out));
+
+ // Emit bytecode to the assembly file into its special section next
+ if (EmitMappingInfo) {
+ PM.add(getEmitBytecodeToAsmPass(Out));
+ PM.add(getFunctionInfo(Out));
+ }
+
+ return false;
}
+// addPassesToJITCompile - This method controls the JIT method of code
+// generation for the UltraSparc.
//
-// freeCompiledMethod - Release all memory associated with the compiled image
-// for this method.
-//
-void
-UltraSparc::freeCompiledMethod(Method *M)
-{
- for_each(M->inst_begin(), M->inst_end(), freeMachineCode);
- // Don't destruct MachineCodeForMethod - The global printer needs it
- //MachineCodeForMethod::destruct(M);
+bool UltraSparc::addPassesToJITCompile(PassManager &PM) {
+ const TargetData &TD = getTargetData();
+
+ PM.add(new TargetData("lli", TD.isLittleEndian(), TD.getPointerSize(),
+ TD.getPointerAlignment(), TD.getDoubleAlignment()));
+
+ // Replace malloc and free instructions with library calls.
+ // Do this after tracing until lli implements these lib calls.
+ // For now, it will emulate malloc and free internally.
+ PM.add(createLowerAllocationsPass());
+
+ // FIXME: implement the switch instruction in the instruction selector.
+ PM.add(createLowerSwitchPass());
+
+ // decompose multi-dimensional array references into single-dim refs
+ PM.add(createDecomposeMultiDimRefsPass());
+
+ // Construct and initialize the MachineFunction object for this fn.
+ PM.add(createMachineCodeConstructionPass(*this));
+
+ PM.add(createInstructionSelectionPass(*this));
+
+ // new pass: convert Value* in MachineOperand to an unsigned register
+ // this brings it in line with what the X86 JIT's RegisterAllocator expects
+ //PM.add(createAddRegNumToValuesPass());
+
+ PM.add(getRegisterAllocator(*this));
+ PM.add(getPrologEpilogInsertionPass());
+
+ if (!DisablePeephole)
+ PM.add(createPeepholeOptsPass(*this));
+
+ return false; // success!
}