//===-- TargetMachine.cpp - General Target Information ---------------------==//
//
// This file describes the general parts of a Target machine.
+// This file also implements MachineInstrInfo and MachineCacheInfo.
//
//===----------------------------------------------------------------------===//
-#include "llvm/Target/Machine.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/MachineInstrInfo.h"
+#include "llvm/Target/MachineCacheInfo.h"
+#include "llvm/CodeGen/PreSelection.h"
+#include "llvm/CodeGen/StackSlots.h"
+#include "llvm/CodeGen/InstrSelection.h"
+#include "llvm/CodeGen/InstrScheduling.h"
+#include "llvm/CodeGen/RegisterAllocation.h"
+#include "llvm/CodeGen/PeepholeOpts.h"
+#include "llvm/CodeGen/MachineCodeForMethod.h"
+#include "llvm/CodeGen/MachineCodeForInstruction.h"
+#include "llvm/Reoptimizer/Mapping/MappingInfo.h"
+#include "llvm/Reoptimizer/Mapping/FInfo.h"
+#include "llvm/Transforms/Scalar.h"
+#include "Support/CommandLine.h"
+#include "llvm/PassManager.h"
+#include "llvm/Function.h"
#include "llvm/DerivedTypes.h"
-// External object describing the machine instructions
-// Initialized only when the TargetMachine class is created
-// and reset when that class is destroyed.
-//
-const MachineInstrDescriptor* TargetInstrDescriptors = NULL;
+//---------------------------------------------------------------------------
+// Command line options to control choice of code generation passes.
+//---------------------------------------------------------------------------
-resourceId_t MachineResource::nextId = 0;
+static cl::opt<bool> DisablePreSelect("nopreselect",
+ cl::desc("Disable preselection pass"));
-static cycles_t ComputeMinGap (const InstrRUsage& fromRU,
- const InstrRUsage& toRU);
+static cl::opt<bool> DisableSched("nosched",
+ cl::desc("Disable local scheduling pass"));
-static bool RUConflict (const vector<resourceId_t>& fromRVec,
- const vector<resourceId_t>& fromRVec);
+static cl::opt<bool> DisablePeephole("nopeephole",
+ cl::desc("Disable peephole optimization pass"));
//---------------------------------------------------------------------------
// class TargetMachine
// space equal to optSizeForSubWordData, and all other primitive data
// items use space according to the type.
//
-unsigned int TargetMachine::findOptimalStorageSize(const Type* ty) const {
- switch(ty->getPrimitiveID()) {
- case Type::BoolTyID:
- case Type::UByteTyID:
- case Type::SByteTyID:
- case Type::UShortTyID:
- case Type::ShortTyID:
- return optSizeForSubWordData;
+unsigned int
+TargetMachine::findOptimalStorageSize(const Type* ty) const
+{
+ switch(ty->getPrimitiveID())
+ {
+ case Type::BoolTyID:
+ case Type::UByteTyID:
+ case Type::SByteTyID:
+ case Type::UShortTyID:
+ case Type::ShortTyID:
+ return optSizeForSubWordData;
- default:
- return DataLayout.getTypeSize(ty);
+ default:
+ return DataLayout.getTypeSize(ty);
+ }
+}
+
+
+//===---------------------------------------------------------------------===//
+// Default code generation passes.
+//
+// Native code generation for a specified target.
+//===---------------------------------------------------------------------===//
+
+class ConstructMachineCodeForFunction : public FunctionPass {
+ TargetMachine &Target;
+public:
+ inline ConstructMachineCodeForFunction(TargetMachine &T) : Target(T) {}
+
+ const char *getPassName() const {
+ return "ConstructMachineCodeForFunction";
}
+
+ bool runOnFunction(Function &F) {
+ MachineCodeForMethod::construct(&F, Target);
+ return false;
+ }
+};
+
+struct FreeMachineCodeForFunction : public FunctionPass {
+ const char *getPassName() const { return "FreeMachineCodeForFunction"; }
+
+ static void freeMachineCode(Instruction &I) {
+ MachineCodeForInstruction::destroy(&I);
+ }
+
+ bool runOnFunction(Function &F) {
+ for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
+ for (BasicBlock::iterator I = FI->begin(), E = FI->end(); I != E; ++I)
+ MachineCodeForInstruction::get(I).dropAllReferences();
+
+ for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
+ for_each(FI->begin(), FI->end(), freeMachineCode);
+
+ return false;
+ }
+};
+
+// addPassesToEmitAssembly - This method controls the entire code generation
+// process for the ultra sparc.
+//
+void
+TargetMachine::addPassesToEmitAssembly(PassManager &PM, std::ostream &Out)
+{
+ // Construct and initialize the MachineCodeForMethod object for this fn.
+ PM.add(new ConstructMachineCodeForFunction(*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));
+
+ // Specialize LLVM code for this target machine and then
+ // run basic dataflow optimizations on LLVM code.
+ if (!DisablePreSelect)
+ {
+ PM.add(createPreSelectionPass(*this));
+ /* PM.add(createReassociatePass()); */
+ PM.add(createGCSEPass());
+ PM.add(createLICMPass());
+ }
+
+ PM.add(createInstructionSelectionPass(*this));
+
+ if (!DisableSched)
+ PM.add(createInstructionSchedulingWithSSAPass(*this));
+
+ PM.add(getRegisterAllocator(*this));
+
+ PM.add(getPrologEpilogInsertionPass());
+
+ if (!DisablePeephole)
+ PM.add(createPeepholeOptsPass(*this));
+
+ PM.add(MappingInfoForFunction(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(new FreeMachineCodeForFunction()); // 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
+ PM.add(getEmitBytecodeToAsmPass(Out));
+ PM.add(getFunctionInfo(Out));
}
/*ctor*/
-MachineInstrInfo::MachineInstrInfo(const MachineInstrDescriptor* _desc,
+MachineInstrInfo::MachineInstrInfo(const TargetMachine& tgt,
+ const MachineInstrDescriptor* _desc,
unsigned int _descSize,
unsigned int _numRealOpCodes)
- : desc(_desc), descSize(_descSize), numRealOpCodes(_numRealOpCodes)
+ : target(tgt),
+ desc(_desc), descSize(_descSize), numRealOpCodes(_numRealOpCodes)
{
+ // FIXME: TargetInstrDescriptors should not be global
assert(TargetInstrDescriptors == NULL && desc != NULL);
TargetInstrDescriptors = desc; // initialize global variable
}
-/*dtor*/
MachineInstrInfo::~MachineInstrInfo()
{
TargetInstrDescriptors = NULL; // reset global variable
{
// First, check if opCode has an immed field.
bool isSignExtended;
- uint64_t maxImmedValue = this->maxImmedConstant(opCode, isSignExtended);
+ uint64_t maxImmedValue = maxImmedConstant(opCode, isSignExtended);
if (maxImmedValue != 0)
{
+ // NEED TO HANDLE UNSIGNED VALUES SINCE THEY MAY BECOME MUCH
+ // SMALLER AFTER CASTING TO SIGN-EXTENDED int, short, or char.
+ // See CreateUIntSetInstruction in SparcInstrInfo.cpp.
+
// Now check if the constant fits
if (intValue <= (int64_t) maxImmedValue &&
intValue >= -((int64_t) maxImmedValue+1))
//---------------------------------------------------------------------------
-// class MachineSchedInfo
-// Interface to machine description for instruction scheduling
+// class MachineCacheInfo
+//
+// Purpose:
+// Describes properties of the target cache architecture.
//---------------------------------------------------------------------------
/*ctor*/
-MachineSchedInfo::MachineSchedInfo(int _numSchedClasses,
- const MachineInstrInfo* _mii,
- const InstrClassRUsage* _classRUsages,
- const InstrRUsageDelta* _usageDeltas,
- const InstrIssueDelta* _issueDeltas,
- unsigned int _numUsageDeltas,
- unsigned int _numIssueDeltas)
- : numSchedClasses(_numSchedClasses),
- mii(_mii),
- classRUsages(_classRUsages),
- usageDeltas(_usageDeltas),
- issueDeltas(_issueDeltas),
- numUsageDeltas(_numUsageDeltas),
- numIssueDeltas(_numIssueDeltas)
-{
-}
-
-void
-MachineSchedInfo::initializeResources()
+MachineCacheInfo::MachineCacheInfo(const TargetMachine& tgt)
+ : target(tgt)
{
- assert(MAX_NUM_SLOTS >= (int) getMaxNumIssueTotal()
- && "Insufficient slots for static data! Increase MAX_NUM_SLOTS");
-
- // First, compute common resource usage info for each class because
- // most instructions will probably behave the same as their class.
- // Cannot allocate a vector of InstrRUsage so new each one.
- //
- vector<InstrRUsage> instrRUForClasses;
- instrRUForClasses.resize(numSchedClasses);
- for (InstrSchedClass sc=0; sc < numSchedClasses; sc++)
- {
- // instrRUForClasses.push_back(new InstrRUsage);
- instrRUForClasses[sc].setMaxSlots(getMaxNumIssueTotal());
- instrRUForClasses[sc] = classRUsages[sc];
- }
-
- computeInstrResources(instrRUForClasses);
-
- computeIssueGaps(instrRUForClasses);
+ Initialize();
}
-
void
-MachineSchedInfo::computeInstrResources(const vector<InstrRUsage>& instrRUForClasses)
+MachineCacheInfo::Initialize()
{
- int numOpCodes = mii->getNumRealOpCodes();
- instrRUsages.resize(numOpCodes);
-
- // First get the resource usage information from the class resource usages.
- for (MachineOpCode op=0; op < numOpCodes; op++)
- {
- InstrSchedClass sc = getSchedClass(op);
- assert(sc >= 0 && sc < numSchedClasses);
- instrRUsages[op] = instrRUForClasses[sc];
- }
-
- // Now, modify the resource usages as specified in the deltas.
- for (unsigned i=0; i < numUsageDeltas; i++)
- {
- MachineOpCode op = usageDeltas[i].opCode;
- assert(op < numOpCodes);
- instrRUsages[op].addUsageDelta(usageDeltas[i]);
- }
-
- // Then modify the issue restrictions as specified in the deltas.
- for (unsigned i=0; i < numIssueDeltas; i++)
- {
- MachineOpCode op = issueDeltas[i].opCode;
- assert(op < numOpCodes);
- instrRUsages[issueDeltas[i].opCode].addIssueDelta(issueDeltas[i]);
- }
+ numLevels = 2;
+ cacheLineSizes.push_back(16); cacheLineSizes.push_back(32);
+ cacheSizes.push_back(1 << 15); cacheSizes.push_back(1 << 20);
+ cacheAssoc.push_back(1); cacheAssoc.push_back(4);
}
-
-
-void
-MachineSchedInfo::computeIssueGaps(const vector<InstrRUsage>& instrRUForClasses)
-{
- int numOpCodes = mii->getNumRealOpCodes();
- instrRUsages.resize(numOpCodes);
-
- assert(numOpCodes < (1 << MAX_OPCODE_SIZE) - 1
- && "numOpCodes invalid for implementation of class OpCodePair!");
-
- // First, compute issue gaps between pairs of classes based on common
- // resources usages for each class, because most instruction pairs will
- // usually behave the same as their class.
- //
- int classPairGaps[numSchedClasses][numSchedClasses];
- for (InstrSchedClass fromSC=0; fromSC < numSchedClasses; fromSC++)
- for (InstrSchedClass toSC=0; toSC < numSchedClasses; toSC++)
- {
- int classPairGap = ComputeMinGap(instrRUForClasses[fromSC],
- instrRUForClasses[toSC]);
- classPairGaps[fromSC][toSC] = classPairGap;
- }
-
- // Now, for each pair of instructions, use the class pair gap if both
- // instructions have identical resource usage as their respective classes.
- // If not, recompute the gap for the pair from scratch.
-
- longestIssueConflict = 0;
-
- for (MachineOpCode fromOp=0; fromOp < numOpCodes; fromOp++)
- for (MachineOpCode toOp=0; toOp < numOpCodes; toOp++)
- {
- int instrPairGap =
- (instrRUsages[fromOp].sameAsClass && instrRUsages[toOp].sameAsClass)
- ? classPairGaps[getSchedClass(fromOp)][getSchedClass(toOp)]
- : ComputeMinGap(instrRUsages[fromOp], instrRUsages[toOp]);
-
- if (instrPairGap > 0)
- {
- issueGaps[OpCodePair(fromOp,toOp)] = instrPairGap;
- conflictLists[fromOp].push_back(toOp);
- longestIssueConflict = max(longestIssueConflict, instrPairGap);
- }
- }
-}
-
-
-// Check if fromRVec and toRVec have *any* common entries.
-// Assume the vectors are sorted in increasing order.
-// Algorithm copied from function set_intersection() for sorted ranges (stl_algo.h).
-inline static bool
-RUConflict(const vector<resourceId_t>& fromRVec,
- const vector<resourceId_t>& toRVec)
-{
- bool commonElementFound = false;
-
- unsigned fN = fromRVec.size(), tN = toRVec.size();
- unsigned fi = 0, ti = 0;
- while (fi < fN && ti < tN)
- if (fromRVec[fi] < toRVec[ti])
- ++fi;
- else if (toRVec[ti] < fromRVec[fi])
- ++ti;
- else
- {
- commonElementFound = true;
- break;
- }
-
- return commonElementFound;
-}
-
-
-static cycles_t
-ComputeMinGap(const InstrRUsage& fromRU, const InstrRUsage& toRU)
-{
- cycles_t minGap = 0;
-
- if (fromRU.numBubbles > 0)
- minGap = fromRU.numBubbles;
-
- if (minGap < fromRU.numCycles)
- {
- // only need to check from cycle `minGap' onwards
- for (cycles_t gap=minGap; gap <= fromRU.numCycles-1; gap++)
- {
- // check if instr. #2 can start executing `gap' cycles after #1
- // by checking for resource conflicts in each overlapping cycle
- cycles_t numOverlap = min(fromRU.numCycles - gap, toRU.numCycles);
- for (cycles_t c = 0; c <= numOverlap-1; c++)
- if (RUConflict(fromRU.resourcesByCycle[gap + c],
- toRU.resourcesByCycle[c]))
- {// conflict found so minGap must be more than `gap'
- minGap = gap+1;
- break;
- }
- }
- }
-
- return minGap;
-}
-
-//---------------------------------------------------------------------------
projects / oota-llvm.git / blobdiff