Merge pull request #205 from vext01/o1

O1

llvm/include/llvm/CodeGen/StackMaps.h

@@ -29,6 +29,8 @@ class MCStreamer;
 class raw_ostream;
 class TargetRegisterInfo;
 
+unsigned getDwarfRegNum(unsigned Reg, const TargetRegisterInfo *TRI);
+
 /// MI-level stackmap operands.
 ///
 /// MI stackmap operations take the form:

llvm/include/llvm/Support/Yk.h

@@ -8,5 +8,6 @@ void initYkOptions(void);
 // YKFIXME: all of our command-line arguments should be collected here instead
 // of us randomly introducing `extern bool`s all over the place.
 extern bool YkOptNoneAfterIRPasses;
+extern bool YkDontOptFuncABI;
 
 #endif

llvm/include/llvm/Transforms/Yk/ControlPoint.h

@@ -14,6 +14,9 @@
 // right instruction in AOT from where to continue.
 #define YK_RECONSTRUCT_FRAMES "__ykrt_reconstruct_frames"
 
+// The name of the patchpoint intrinsic we use for the control point.
+#define CP_PPNAME "llvm.experimental.patchpoint.void"
+
 namespace llvm {
 ModulePass *createYkControlPointPass();
 } // namespace llvm

llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp

@@ -1982,10 +1982,18 @@ void AsmPrinter::emitFunctionBody() {
         break;
       default:
 
+        // So that the ykpt decoder can work without disasembling instructions
+        // to find call-sites and sucessor blocks, we encode that info
+        // statically into the blockmap at AOT compile time.
+        //
+        // Some things that look like calls in IR don't actually emit a
+        // call into the binary. Namely a stackmap intrinsic.
+        //
+        // Note that patchpoint and statepoint intrinsics, although similar to
+        // the stackmap intrinsic, do actually emit a call in the binary, so we
+        // DO need to include those callsites.
         if (YkExtendedLLVMBBAddrMapSection && MI.isCall() &&
-            (MI.getOpcode() != TargetOpcode::STACKMAP) &&
-            (MI.getOpcode() != TargetOpcode::PATCHPOINT) &&
-            (MI.getOpcode() != TargetOpcode::STATEPOINT)) {
+            (MI.getOpcode() != TargetOpcode::STACKMAP)) {
           // Record the address of the call instruction itself.
           MCSymbol *YkPreCallSym =
               MF->getContext().createTempSymbol("yk_precall", true);

llvm/lib/CodeGen/StackMaps.cpp

@@ -24,7 +24,6 @@
 #include "llvm/MC/MCContext.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCObjectFileInfo.h"
-#include "llvm/MC/MCRegisterInfo.h"
 #include "llvm/MC/MCStreamer.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
@@ -211,7 +210,7 @@ unsigned StackMaps::getNextMetaArgIdx(const MachineInstr *MI, unsigned CurIdx) {
 }
 
 /// Go up the super-register chain until we hit a valid dwarf register number.
-static unsigned getDwarfRegNum(unsigned Reg, const TargetRegisterInfo *TRI) {
+unsigned llvm::getDwarfRegNum(unsigned Reg, const TargetRegisterInfo *TRI) {
   int RegNum;
   for (MCPhysReg SR : TRI->superregs_inclusive(Reg)) {
     RegNum = TRI->getDwarfRegNum(SR, false);

llvm/lib/CodeGen/Yk/FixStackmapsSpillReloads.cpp

@@ -89,8 +89,10 @@ INITIALIZE_PASS_BEGIN(FixStackmapsSpillReloads, DEBUG_TYPE, "Fixup Stackmap Spil
 INITIALIZE_PASS_END(FixStackmapsSpillReloads, DEBUG_TYPE, "Fixup Stackmap Spills",
                     false, false)
 
+const TargetRegisterInfo *TRI;
 
 bool FixStackmapsSpillReloads::runOnMachineFunction(MachineFunction &MF) {
+  TRI = MF.getSubtarget().getRegisterInfo();
   bool Changed = false;
   const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
   for (MachineBasicBlock &MBB : MF) {
@@ -144,7 +146,7 @@ bool FixStackmapsSpillReloads::runOnMachineFunction(MachineFunction &MF) {
         MOI++;
         while (MOI != MI.operands_end()) {
           if (MOI->isReg()) {
-            Register Reg = MOI->getReg();
+            unsigned int Reg = getDwarfRegNum(MOI->getReg(), TRI);
             // Check if the register operand in the stackmap is a restored
             // spill.
             // Since implicit operands are ignored by stackmaps (they are not
@@ -240,13 +242,15 @@ bool FixStackmapsSpillReloads::runOnMachineFunction(MachineFunction &MF) {
         if (TII->isCopyInstr(MI) || TII->isLoadFromStackSlotPostFE(MI, FI)) {
           // FIXME: Can there be multiple spill reloads here? Then this would
           // need to be a loop.
-          if (TII->isCopyInstr(MI) && Spills.count(MI.getOperand(1).getReg())) {
+          unsigned int Op0 = getDwarfRegNum(MI.getOperand(0).getReg(), TRI);
+          unsigned int Op1 = getDwarfRegNum(MI.getOperand(1).getReg(), TRI);
+          if (TII->isCopyInstr(MI) && Spills.count(Op1)) {
             // The source for this copy instruction is itself a spill reload.
             // So we need to lookup the spill for the source and apply this
             // instead.
-            Spills[MI.getOperand(0).getReg()] = Spills[MI.getOperand(1).getReg()];
+            Spills[Op0] = Spills[Op1];
           } else {
-            Spills[MI.getOperand(0).getReg()] = &MI;
+            Spills[Op0] = &MI;
           }
         }
       }

llvm/lib/Support/Yk.cpp

@@ -96,11 +96,26 @@ struct CreateYkEmbedIRParser {
 } // namespace
 static ManagedStatic<cl::opt<bool, true>, CreateYkEmbedIRParser> YkEmbedIRParser;
 
+bool YkDontOptFuncABI;
+namespace {
+struct CreateYkDontOptFuncABIParser {
+  static void *call() {
+    return new cl::opt<bool, true>(
+        "yk-dont-opt-func-abi",
+        cl::desc(
+            "Don't change the ABIs of functions during optimisation"),
+        cl::NotHidden, cl::location(YkDontOptFuncABI));
+  }
+};
+} // namespace
+static ManagedStatic<cl::opt<bool, true>, CreateYkDontOptFuncABIParser> YkDontOptFuncABIParser;
+
 void llvm::initYkOptions() {
   *YkExtendedLLVMBBAddrMapSectionParser;
   *YkStackMapOffsetFixParser;
   *YkStackMapAdditionalLocsParser;
   *YkStackmapsSpillFixParser;
   *YkOptNoneAfterIRPassesParser;
   *YkEmbedIRParser;
+  *YkDontOptFuncABIParser;
 }

llvm/lib/Target/X86/X86AsmPrinter.cpp

@@ -24,6 +24,7 @@
 #include "llvm/CodeGen/MachineConstantPool.h"
 #include "llvm/CodeGen/MachineModuleInfoImpls.h"
 #include "llvm/CodeGen/MachineValueType.h"
+#include "llvm/CodeGen/StackMaps.h"
 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/InlineAsm.h"
@@ -62,22 +63,6 @@ X86AsmPrinter::X86AsmPrinter(TargetMachine &TM,
 const TargetInstrInfo *TII;
 const TargetRegisterInfo *TRI;
 
-/// Go up the super-register chain until we hit a valid dwarf register number.
-///
-/// (duplicated/adapted from StackMaps.cpp so as to not introduce extra link
-/// dependencies)
-static unsigned getDwarfRegNum(unsigned Reg) {
-  int RegNum;
-  for (MCPhysReg SR : TRI->superregs_inclusive(Reg)) {
-    RegNum = TRI->getDwarfRegNum(SR, false);
-    if (RegNum >= 0)
-      break;
-  }
-
-  assert(RegNum >= 0 && "Invalid Dwarf register number.");
-  return (unsigned)RegNum;
-}
-
 /// Clear any mappings that map to the given register.
 void clearRhs(Register Reg, std::map<Register, std::set<int64_t>> &SpillMap) {
   auto I = SpillMap.begin();
@@ -116,6 +101,13 @@ void processInstructions(
       continue;
     }
 
+    // Because a patchpoint already captures the live values at the exact
+    // moment we desire, there's no need to compute a spillmap for them nor do
+    // we have to "patch them up". We can just skip them.
+    if (Instr.getOpcode() == TargetOpcode::PATCHPOINT) {
+      continue;
+    }
+
     // Copying a value from one register B to another A, creates a mapping from
     // A to B. If A is tracked by the stackmap, then B will also be tracked and
     // assigned the same value during deoptimisation.
@@ -124,8 +116,8 @@ void processInstructions(
       const MachineOperand Rhs = Instr.getOperand(1);
       assert(Lhs.isReg() && "Is register.");
       assert(Rhs.isReg() && "Is register.");
-      auto LhsDwReg = getDwarfRegNum(Lhs.getReg());
-      auto RhsDwReg = getDwarfRegNum(Rhs.getReg());
+      auto LhsDwReg = getDwarfRegNum(Lhs.getReg(), TRI);
+      auto RhsDwReg = getDwarfRegNum(Rhs.getReg(), TRI);
       if (LhsDwReg == RhsDwReg) {
         // Moves like `mov rax, rax` are effectively a NOP for this analysis.
         continue;
@@ -149,7 +141,7 @@ void processInstructions(
       const MachineOperand OffsetOp = Instr.getOperand(3);
       const MachineOperand MO = Instr.getOperand(5);
       assert(MO.isReg() && "Is register.");
-      const Register DwReg = getDwarfRegNum(MO.getReg());
+      const Register DwReg = getDwarfRegNum(MO.getReg(), TRI);
       if (OffsetOp.isImm()) {
         const int64_t Offset = OffsetOp.getImm();
         // We don't need to do `clearRhs(DwReg)` and reset the `SpillMap` entry
@@ -168,7 +160,7 @@ void processInstructions(
       const MachineOperand OffsetOp = Instr.getOperand(4);
       const MachineOperand Lhs = Instr.getOperand(0);
       assert(Lhs.isReg() && "Is register.");
-      const Register DwReg = getDwarfRegNum(Lhs.getReg());
+      const Register DwReg = getDwarfRegNum(Lhs.getReg(), TRI);
       if (OffsetOp.isImm()) {
         const int64_t Offset = OffsetOp.getImm();
         clearRhs(DwReg, SpillMap);
@@ -180,7 +172,7 @@ void processInstructions(
     // Any other assignments to tracked registers removes their mapping.
     for (const MachineOperand MO : Instr.defs()) {
       assert(MO.isReg() && "Is register.");
-      auto DwReg = getDwarfRegNum(MO.getReg());
+      auto DwReg = getDwarfRegNum(MO.getReg(), TRI);
       SpillMap.erase(DwReg);
       clearRhs(DwReg, SpillMap);
     }
@@ -190,7 +182,7 @@ void processInstructions(
     // be used.
     for (const MachineOperand MO : Instr.uses()) {
       if (MO.isReg() && MO.isKill()) {
-        auto DwReg = getDwarfRegNum(MO.getReg());
+        auto DwReg = getDwarfRegNum(MO.getReg(), TRI);
         SpillMap.erase(DwReg);
         clearRhs(DwReg, SpillMap);
       }

llvm/lib/Transforms/IPO/GlobalOpt.cpp

@@ -58,6 +58,7 @@
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/Yk.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/Utils/CtorUtils.h"
@@ -1986,7 +1987,9 @@ OptimizeFunctions(Module &M,
       }
     }
 
-    if (hasChangeableCC(&F) && !F.isVarArg() && !F.hasAddressTaken()) {
+    if (!YkDontOptFuncABI && hasChangeableCC(&F) && !F.isVarArg()
+        && !F.hasAddressTaken())
+    {
       // If this function has a calling convention worth changing, is not a
       // varargs function, and is only called directly, promote it to use the
       // Fast calling convention.

llvm/lib/Transforms/Yk/ControlPoint.cpp

@@ -1,4 +1,4 @@
-//===- ControlPoint.cpp - Synthesise the yk control point -----------------===//
+//===- ControlPoint.cpp - Patch the yk control point -----------------===//
 //
 // This pass finds the user's call to the dummy control point and replaces it
 // with a call to a new control point that implements the necessary logic to
@@ -19,30 +19,22 @@
 // }
 // ```
 //
-// Into one that looks like this (note that this transformation happens at the
-// IR level):
+// Into one that looks like this:
 //
 // ```
-// // The YkCtrlPointStruct contains one member for each live LLVM variable
-// // just before the call to the control point.
-// struct YkCtrlPointStruct {
-//     size_t pc;
-// }
-//
-// struct YkCtrlPointStruct cp_vars;
 // pc = 0;
 // while (...) {
-//     // Now we call the patched control point.
-//     cp_vars.pc = pc;
-//     __ykrt__control_point(mt, loc, &cp_vars);
-//     pc = cp_vars.pc;
+//     llvm.experimental.patchpoint.void(..., __ykrt_control_point, ...)
 //     bc = program[pc];
 //     switch (bc) {
 //         // bytecode handlers here.
 //     }
 // }
 // ```
 //
+// A patchpoint is used to capture the locations of live variables immediately
+// before a call to __ykrt_control_point.
+//
 // Note that this transformation occurs at the LLVM IR level. The above example
 // is shown as C code for easy comprehension.
 
@@ -58,6 +50,7 @@
 #include "llvm/IR/Verifier.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
+#include "llvm/Transforms/Yk/LivenessAnalysis.h"
 
 #define DEBUG_TYPE "yk-control-point"
 
@@ -68,6 +61,29 @@
 #define YK_CONTROL_POINT_ARG_VARS_IDX 2
 #define YK_CONTROL_POINT_NUM_ARGS 3
 
+// Stackmap ID zero is reserved for the control point.
+//
+// This will need to change when we support >1 control point.
+const unsigned CPStackMapID = 0;
+
+// The number of shadow bytes required for the control point's patchpoint.
+//
+// This must be large enough to accommodate the call to patchpoint target
+// function and if you use a too-big value LLVM will pad the space with NOP
+// bytes.
+//
+// This early in the pipeline we have no idea how the backend will choose the
+// encode this call, so for now we use the exact size of the observed
+// instruction at the time of writing, as determined by disassembling the binary
+// and eyeballing it.
+//
+// The good news is that LLVM will assert fail if you use a too small value.
+#if defined(__x86_64__) || defined(_M_X64)
+const unsigned CPShadow = 13;
+#else
+#error "unknown control point shadow size for this arch"
+#endif
+
 using namespace llvm;
 
 /// Find the call to the dummy control point that we want to patch.
@@ -128,34 +144,55 @@ class YkControlPoint : public ModulePass {
     }
 
     // The old control point should be of the form:
-    //    control_point(YkMT*, YkLocation*)
+    //    yk_mt_control_point(YkMT*, YkLocation*)
     assert(OldCtrlPointCall->arg_size() == YK_OLD_CONTROL_POINT_NUM_ARGS);
     Type *YkMTTy =
         OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_MT_IDX)->getType();
     Type *YkLocTy =
         OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_LOC_IDX)
             ->getType();
 
-    // Create the new control point, which is of the form:
-    //   void new_control_point(YkMT*, YkLocation*, i64)
+    // Create a call to the "new" (patched) control point, but do so via a
+    // patchpoint so that we can capture the live variables at exactly the
+    // moment before the call.
     Type *Int64Ty = Type::getInt64Ty(Context);
     FunctionType *FType = FunctionType::get(Type::getVoidTy(Context),
                                             {YkMTTy, YkLocTy, Int64Ty}, false);
     Function *NF = Function::Create(FType, GlobalVariable::ExternalLinkage,
                                     YK_NEW_CONTROL_POINT, M);
 
-    // At the top of the function, instantiate a `YkCtrlPointStruct` to pass in
-    // to the control point. We do so on the stack, so that we can pass the
-    // struct by pointer.
     IRBuilder<> Builder(OldCtrlPointCall);
 
-    // Insert call to the new control point. The last argument is the stackmap
-    // id belonging to the control point. This is temporarily set to INT_MAX
-    // and overwritten by the stackmap pass.
-    Builder.CreateCall(
-        NF, {OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_MT_IDX),
-             OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_LOC_IDX),
-             Builder.getInt64(UINT64_MAX)});
+    const Intrinsic::ID SMFuncID = Function::lookupIntrinsicID(CP_PPNAME);
+    if (SMFuncID == Intrinsic::not_intrinsic) {
+      Context.emitError("can't find stackmap()");
+      return false;
+    }
+    Function *SMFunc = Intrinsic::getDeclaration(&M, SMFuncID);
+    assert(SMFunc != nullptr);
+
+    // Get live variables.
+    LivenessAnalysis LA(Caller);
+    auto Lives = LA.getLiveVarsBefore(OldCtrlPointCall);
+
+    Value *SMID = ConstantInt::get(Type::getInt64Ty(Context), CPStackMapID);
+    Value *Shadow = ConstantInt::get(Type::getInt32Ty(Context), CPShadow);
+    std::vector<Value *> Args = {
+        SMID,
+        Shadow,
+        NF,
+        ConstantInt::get(Type::getInt32Ty(Context), 3),
+        OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_MT_IDX),
+        OldCtrlPointCall->getArgOperand(YK_CONTROL_POINT_ARG_LOC_IDX),
+        SMID,
+    };
+
+    for (auto *Live : Lives) {
+      Args.push_back(Live);
+    }
+
+    Builder.CreateCall(SMFunc->getFunctionType(), SMFunc,
+                       ArrayRef<Value *>(Args));
 
     // Replace the call to the dummy control point.
     OldCtrlPointCall->eraseFromParent();