Support frame introspection

The approach in this changeset is to attach frame args
to every call site via stackmap args.  We make sure that
every callsite is a patchpoint, and serialize the symbol
table into its arguments.

In this patch, that is just used for supporting the locals()
method.

It's currently disabled since it has a hugely negative impact
on performance (LLVM takes much longer to JIT with all the extra
function arguments).

src/asm_writing/assembler.h

@@ -18,6 +18,7 @@
 #include <unordered_set>
 
 #include "asm_writing/types.h"
+#include "codegen/stackmaps.h"
 #include "core/ast.h"
 
 namespace pyston {

src/asm_writing/rewriter.cpp

@@ -1111,6 +1111,72 @@ int RewriterVar::nvars = 0;
 
 static const int INITIAL_CALL_SIZE = 13;
 static const int DWARF_RBP_REGNUM = 6;
+bool spillFrameArgumentIfNecessary(StackMap::Record::Location& l, uint8_t*& inst_addr, uint8_t* inst_end,
+                                   int& scratch_offset, int& scratch_size, SpillMap& remapped) {
+    switch (l.type) {
+        case StackMap::Record::Location::LocationType::Direct:
+        case StackMap::Record::Location::LocationType::Indirect:
+        case StackMap::Record::Location::LocationType::Constant:
+        case StackMap::Record::Location::LocationType::ConstIndex:
+            return false;
+        case StackMap::Record::Location::LocationType::Register: {
+            assembler::GenericRegister ru = assembler::GenericRegister::fromDwarf(l.regnum);
+
+            if (!Location(ru).isClobberedByCall())
+                return false;
+
+            auto it = remapped.find(ru);
+            if (it != remapped.end()) {
+                if (VERBOSITY()) {
+                    printf("Already spilled ");
+                    ru.dump();
+                }
+                l = it->second;
+                return false;
+            }
+
+            if (VERBOSITY()) {
+                printf("Spilling reg ");
+                ru.dump();
+            }
+
+            assembler::Assembler assembler(inst_addr, inst_end - inst_addr);
+
+            int bytes_pushed;
+            if (ru.type == assembler::GenericRegister::GP) {
+                auto dest = assembler::Indirect(assembler::RBP, scratch_offset);
+                assembler.mov(ru.gp, dest);
+                bytes_pushed = 8;
+            } else if (ru.type == assembler::GenericRegister::XMM) {
+                auto dest = assembler::Indirect(assembler::RBP, scratch_offset);
+                assembler.movsd(ru.xmm, dest);
+                bytes_pushed = 8;
+            } else {
+                abort();
+            }
+
+            assert(scratch_size >= bytes_pushed);
+            assert(!assembler.hasFailed());
+
+            uint8_t* cur_addr = assembler.curInstPointer();
+            inst_addr = cur_addr;
+
+            l.type = StackMap::Record::Location::LocationType::Indirect;
+            l.regnum = DWARF_RBP_REGNUM;
+            l.offset = scratch_offset;
+
+            scratch_offset += bytes_pushed;
+            scratch_size -= bytes_pushed;
+
+            remapped[ru] = l;
+
+            return true;
+        }
+        default:
+            abort();
+    }
+}
+
 void* extractSlowpathFunc(uint8_t* pp_addr) {
 #ifndef NDEBUG
     // mov $imm, %r11:
@@ -1135,12 +1201,13 @@ void* extractSlowpathFunc(uint8_t* pp_addr) {
 
 std::pair<uint8_t*, uint8_t*> initializePatchpoint3(void* slowpath_func, uint8_t* start_addr, uint8_t* end_addr,
                                                     int scratch_offset, int scratch_size,
-                                                    const std::unordered_set<int>& live_outs) {
+                                                    const std::unordered_set<int>& live_outs, SpillMap& remapped) {
     assert(start_addr < end_addr);
 
     int est_slowpath_size = INITIAL_CALL_SIZE;
 
     std::vector<assembler::GenericRegister> regs_to_spill;
+    std::vector<assembler::Register> regs_to_reload;
 
     for (int dwarf_regnum : live_outs) {
         assembler::GenericRegister ru = assembler::GenericRegister::fromDwarf(dwarf_regnum);
@@ -1154,6 +1221,14 @@ std::pair<uint8_t*, uint8_t*> initializePatchpoint3(void* slowpath_func, uint8_t
 
         // Location(ru).dump();
 
+        if (ru.type == assembler::GenericRegister::GP && remapped.count(ru)) {
+            // printf("already spilled!\n");
+
+            regs_to_reload.push_back(ru.gp);
+            est_slowpath_size += 7; // 7 bytes for a single mov
+            continue;
+        }
+
         regs_to_spill.push_back(ru);
 
         if (ru.type == assembler::GenericRegister::GP)
@@ -1183,6 +1258,14 @@ std::pair<uint8_t*, uint8_t*> initializePatchpoint3(void* slowpath_func, uint8_t
     uint8_t* rtn = assem.emitCall(slowpath_func, assembler::R11);
     assem.emitBatchPop(scratch_offset, scratch_size, regs_to_spill);
 
+    for (assembler::Register r : regs_to_reload) {
+        StackMap::Record::Location& l = remapped[r];
+        assert(l.type == StackMap::Record::Location::LocationType::Indirect);
+        assert(l.regnum == DWARF_RBP_REGNUM);
+
+        assem.mov(assembler::Indirect(assembler::RBP, l.offset), r);
+    }
+
     assem.fillWithNops();
     assert(!assem.hasFailed());
 

src/asm_writing/rewriter.h

@@ -15,6 +15,7 @@
 #ifndef PYSTON_ASMWRITING_REWRITER_H
 #define PYSTON_ASMWRITING_REWRITER_H
 
+#include <map>
 #include <memory>
 
 #include "asm_writing/assembler.h"
@@ -315,10 +316,28 @@ public:
 
 void* extractSlowpathFunc(uint8_t* pp_addr);
 
+struct GRCompare {
+    bool operator()(assembler::GenericRegister gr1, assembler::GenericRegister gr2) const {
+        if (gr1.type != gr2.type)
+            return gr1.type < gr2.type;
+
+        if (gr1.type == assembler::GenericRegister::GP)
+            return gr1.gp.regnum < gr2.gp.regnum;
+        if (gr1.type == assembler::GenericRegister::XMM)
+            return gr1.xmm.regnum < gr2.xmm.regnum;
+        abort();
+    }
+};
+typedef std::map<assembler::GenericRegister, StackMap::Record::Location, GRCompare> SpillMap;
+// Spills the stackmap argument and guarantees that it will be readable by the unwinder.
+// Updates the arguments if it did any spilling, and returns whether spilling happened.
+bool spillFrameArgumentIfNecessary(StackMap::Record::Location& l, uint8_t*& inst_addr, uint8_t* inst_end,
+                                   int& scratch_offset, int& scratch_size, SpillMap& remapped);
+
 // returns (start_of_slowpath, return_addr_of_slowpath_call)
 std::pair<uint8_t*, uint8_t*> initializePatchpoint3(void* slowpath_func, uint8_t* start_addr, uint8_t* end_addr,
                                                     int scratch_offset, int scratch_size,
-                                                    const std::unordered_set<int>& live_outs);
+                                                    const std::unordered_set<int>& live_outs, SpillMap& remapped);
 }
 
 #endif

src/codegen/compvars.cpp

@@ -34,6 +34,16 @@
 
 namespace pyston {
 
+void ConcreteCompilerType::serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) {
+#ifndef NDEBUG
+    if (llvmType() == g.i1) {
+        var->getValue()->dump();
+        ASSERT(llvmType() != g.i1, "due to an llvm limitation cannot add these to stackmaps yet");
+    }
+#endif
+    stackmap_args.push_back(var->getValue());
+}
+
 std::string ValuedCompilerType<llvm::Value*>::debugName() {
     std::string rtn;
     llvm::raw_string_ostream os(rtn);
@@ -150,6 +160,18 @@ public:
         }
         return rtn;
     }
+
+    void serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) override {
+        var->getValue()->obj->serializeToFrame(stackmap_args);
+        var->getValue()->func->serializeToFrame(stackmap_args);
+    }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+        abort();
+    }
+
+    int numFrameArgs() override { return obj_type->numFrameArgs() + function_type->numFrameArgs(); }
 };
 std::unordered_map<std::pair<CompilerType*, CompilerType*>, InstanceMethodType*> InstanceMethodType::made;
 
@@ -353,6 +375,11 @@ public:
 
         return new ConcreteCompilerVariable(UNKNOWN, rtn, true);
     }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+        return reinterpret_cast<Box*>(vals[0]);
+    }
 };
 
 ConcreteCompilerType* UNKNOWN = new UnknownType();
@@ -702,6 +729,13 @@ public:
     }
 
     static CompilerType* get(const std::vector<Sig*>& sigs) { return new AbstractFunctionType(sigs); }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+        abort();
+    }
+
+    int numFrameArgs() override { abort(); }
 };
 
 class IntType : public ConcreteCompilerType {
@@ -908,6 +942,12 @@ public:
     }
 
     virtual ConcreteCompilerType* getBoxType() { return BOXED_INT; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+
+        return boxInt(vals[0]);
+    }
 } _INT;
 ConcreteCompilerType* INT = &_INT;
 
@@ -1122,6 +1162,13 @@ public:
     }
 
     virtual ConcreteCompilerType* getBoxType() { return BOXED_FLOAT; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+
+        double d = *reinterpret_cast<const double*>(&vals[0]);
+        return boxFloat(d);
+    }
 } _FLOAT;
 ConcreteCompilerType* FLOAT = &_FLOAT;
 
@@ -1171,6 +1218,15 @@ public:
         assert(is_well_defined);
         return typeFromClass(cls);
     }
+
+    void serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) override { abort(); }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+        abort();
+    }
+
+    int numFrameArgs() override { return 0; }
 };
 std::unordered_map<BoxedClass*, KnownClassobjType*> KnownClassobjType::made;
 
@@ -1490,6 +1546,11 @@ public:
     virtual BoxedClass* guaranteedClass() { return cls; }
 
     virtual ConcreteCompilerType* getBoxType() { return this; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+        return reinterpret_cast<Box*>(vals[0]);
+    }
 };
 std::unordered_map<BoxedClass*, NormalObjectType*> NormalObjectType::made;
 ConcreteCompilerType* STR, *BOXED_INT, *BOXED_FLOAT, *BOXED_BOOL, *NONE;
@@ -1513,12 +1574,15 @@ public:
     }
 
     virtual ConcreteCompilerType* getConcreteType() { return this; }
-    // Shouldn't call this:
-    virtual ConcreteCompilerType* getBoxType() { RELEASE_ASSERT(0, ""); }
+    virtual ConcreteCompilerType* getBoxType() { return this; }
 
     void drop(IREmitter& emitter, VAR* var) override {}
     void grab(IREmitter& emitter, VAR* var) override {}
 
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+        abort();
+    }
 } _CLOSURE;
 ConcreteCompilerType* CLOSURE = &_CLOSURE;
 
@@ -1536,6 +1600,11 @@ public:
     virtual void grab(IREmitter& emitter, VAR* var) {
         // pass
     }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+        abort();
+    }
 } _GENERATOR;
 ConcreteCompilerType* GENERATOR = &_GENERATOR;
 
@@ -1612,6 +1681,18 @@ public:
         }
         return rtn;
     }
+
+    void serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) override {
+        stackmap_args.push_back(embedConstantPtr(var->getValue(), g.i8_ptr));
+    }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+
+        return boxStringPtr(reinterpret_cast<std::string*>(vals[0]));
+    }
+
+    int numFrameArgs() override { return 1; }
 };
 static ValuedCompilerType<const std::string*>* STR_CONSTANT = new StrConstantType();
 
@@ -1622,6 +1703,8 @@ CompilerVariable* makeStr(const std::string* s) {
 class VoidType : public ConcreteCompilerType {
 public:
     llvm::Type* llvmType() { return g.void_; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override { abort(); }
 };
 ConcreteCompilerType* VOID = new VoidType();
 
@@ -1630,17 +1713,14 @@ ConcreteCompilerType* typeFromClass(BoxedClass* c) {
     return NormalObjectType::fromClass(c);
 }
 
-// Due to a temporary LLVM limitation, can represent bools as i64's instead of i1's.
-// #define BOOLS_AS_I64
 class BoolType : public ConcreteCompilerType {
 public:
     std::string debugName() { return "bool"; }
     llvm::Type* llvmType() {
-#ifdef BOOLS_AS_I64
-        return g.i64;
-#else
-        return g.i1;
-#endif
+        if (BOOLS_AS_I64)
+            return g.i64;
+        else
+            return g.i1;
     }
 
     virtual bool isFitBy(BoxedClass* c) { return false; }
@@ -1704,6 +1784,13 @@ public:
     }
 
     virtual ConcreteCompilerType* getBoxType() { return BOXED_BOOL; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+        assert(llvmType() == g.i64);
+        bool b = (bool)vals[0];
+        return boxBool(b);
+    }
 };
 ConcreteCompilerType* BOOL = new BoolType();
 ConcreteCompilerVariable* makeBool(bool b) {
@@ -1858,6 +1945,38 @@ public:
         return makeConverted(emitter, var, getConcreteType())
             ->callattr(emitter, info, attr, clsonly, argspec, args, keyword_names);
     }
+
+    void serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) override {
+        for (auto v : *var->getValue()) {
+            v->serializeToFrame(stackmap_args);
+        }
+    }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == numFrameArgs());
+
+        BoxedTuple::GCVector elts;
+        int cur_idx = 0;
+        for (auto e : elt_types) {
+            int num_args = e->numFrameArgs();
+            // TODO: inefficient to make these copies
+            FrameVals sub_vals(vals.begin() + cur_idx, vals.begin() + cur_idx + num_args);
+
+            elts.push_back(e->deserializeFromFrame(sub_vals));
+
+            cur_idx += num_args;
+        }
+        assert(cur_idx == vals.size());
+
+        return new BoxedTuple(std::move(elts));
+    }
+
+    int numFrameArgs() override {
+        int rtn = 0;
+        for (auto e : elt_types)
+            rtn += e->numFrameArgs();
+        return rtn;
+    }
 };
 
 CompilerType* makeTupleType(const std::vector<CompilerType*>& elt_types) {
@@ -1941,6 +2060,11 @@ public:
     virtual bool canConvertTo(ConcreteCompilerType* other_type) { return true; }
 
     virtual BoxedClass* guaranteedClass() { return NULL; }
+
+    Box* deserializeFromFrame(const FrameVals& vals) override {
+        assert(vals.size() == 1);
+        abort();
+    }
 } _UNDEF;
 CompilerType* UNDEF = &_UNDEF;
 
@@ -1949,25 +2073,25 @@ ConcreteCompilerVariable* undefVariable() {
 }
 
 ConcreteCompilerVariable* boolFromI1(IREmitter& emitter, llvm::Value* v) {
-#ifdef BOOLS_AS_I64
-    assert(v->getType() == g.i1);
-    assert(BOOL->llvmType() == g.i64);
-    llvm::Value* v2 = emitter.getBuilder()->CreateZExt(v, BOOL->llvmType());
-    return new ConcreteCompilerVariable(BOOL, v2, true);
-#else
-    return new ConcreteCompilerVariable(BOOL, v, true);
-#endif
+    if (BOOLS_AS_I64) {
+        assert(v->getType() == g.i1);
+        assert(BOOL->llvmType() == g.i64);
+        llvm::Value* v2 = emitter.getBuilder()->CreateZExt(v, BOOL->llvmType());
+        return new ConcreteCompilerVariable(BOOL, v2, true);
+    } else {
+        return new ConcreteCompilerVariable(BOOL, v, true);
+    }
 }
 
 llvm::Value* i1FromBool(IREmitter& emitter, ConcreteCompilerVariable* v) {
-#ifdef BOOLS_AS_I64
-    assert(v->getType() == BOOL);
-    assert(BOOL->llvmType() == g.i64);
-    llvm::Value* v2 = emitter.getBuilder()->CreateTrunc(v->getValue(), g.i1);
-    return v2;
-#else
-    return v->getValue();
-#endif
+    if (BOOLS_AS_I64) {
+        assert(v->getType() == BOOL);
+        assert(BOOL->llvmType() == g.i64);
+        llvm::Value* v2 = emitter.getBuilder()->CreateTrunc(v->getValue(), g.i1);
+        return v2;
+    } else {
+        return v->getValue();
+    }
 }
 
 

src/codegen/compvars.h

@@ -18,6 +18,7 @@
 #include <stdint.h>
 #include <vector>
 
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/IR/Type.h"
 #include "llvm/IR/Value.h"
 
@@ -31,6 +32,8 @@ class OpInfo;
 class CompilerType;
 class IREmitter;
 
+typedef llvm::SmallVector<uint64_t, 1> FrameVals;
+
 class CompilerType {
 public:
     virtual ~CompilerType() {}
@@ -42,6 +45,8 @@ public:
     virtual CompilerType* callType(ArgPassSpec argspec, const std::vector<CompilerType*>& arg_types,
                                    const std::vector<const std::string*>* keyword_names) = 0;
     virtual BoxedClass* guaranteedClass() = 0;
+    virtual Box* deserializeFromFrame(const FrameVals& vals) = 0;
+    virtual int numFrameArgs() = 0;
 };
 
 typedef std::unordered_map<CompilerVariable*, CompilerVariable*> DupCache;
@@ -152,6 +157,7 @@ public:
         ASSERT((CompilerType*)getConcreteType() != this, "%s", debugName().c_str());
         return getConcreteType()->guaranteedClass();
     }
+    virtual void serializeToFrame(VAR* v, std::vector<llvm::Value*>& stackmap_args) = 0;
 };
 
 template <class V> class ValuedCompilerType : public _ValuedCompilerType<V> { public: };
@@ -180,6 +186,8 @@ public:
     virtual bool canConvertTo(ConcreteCompilerType* other_type) { return other_type == this || other_type == UNKNOWN; }
     virtual ConcreteCompilerVariable* makeConverted(IREmitter& emitter, ConcreteCompilerVariable* var,
                                                     ConcreteCompilerType* other_type);
+    void serializeToFrame(VAR* var, std::vector<llvm::Value*>& stackmap_args) override final;
+    int numFrameArgs() override final { return 1; }
 };
 
 class CompilerVariable {
@@ -254,6 +262,8 @@ public:
     virtual CompilerVariable* getitem(IREmitter& emitter, const OpInfo& info, CompilerVariable*) = 0;
     virtual CompilerVariable* binexp(IREmitter& emitter, const OpInfo& info, CompilerVariable* rhs,
                                      AST_TYPE::AST_TYPE op_type, BinExpType exp_type) = 0;
+
+    virtual void serializeToFrame(std::vector<llvm::Value*>& stackmap_args) = 0;
 };
 
 template <class V> class ValuedCompilerVariable : public CompilerVariable {
@@ -343,6 +353,10 @@ public:
     }
 
     BoxedClass* guaranteedClass() override { return type->guaranteedClass(); }
+
+    void serializeToFrame(std::vector<llvm::Value*>& stackmap_args) override {
+        type->serializeToFrame(this, stackmap_args);
+    }
 };
 
 // template <>

src/codegen/irgen.cpp

@@ -650,7 +650,37 @@ static void emitBBs(IRGenState* irstate, const char* bb_type, GuardList& out_gua
                 // Assert that the phi_st is empty, and just create the symbol table from the non-phi st:
                 ASSERT(phi_ending_symbol_tables[pred]->size() == 0, "%d %d", block->idx, pred->idx);
                 assert(ending_symbol_tables.count(pred));
-                generator->copySymbolsFrom(ending_symbol_tables[pred]);
+
+                // Filter out any names set by an invoke statement at the end
+                // of the previous block, if we're in the unwind path.
+                // This definitely doesn't seem like the most elegant way to do this,
+                // but the rest of the analysis frameworks can't (yet) support the idea of
+                // a block flowing differently to its different predecessors.
+                auto pred = block->predecessors[0];
+                auto last_inst = pred->body.back();
+
+                SymbolTable* sym_table = ending_symbol_tables[pred];
+                bool created_new_sym_table = false;
+                if (last_inst->type == AST_TYPE::Invoke) {
+                    auto invoke = ast_cast<AST_Invoke>(last_inst);
+                    if (invoke->exc_dest == block && invoke->stmt->type == AST_TYPE::Assign) {
+                        auto asgn = ast_cast<AST_Assign>(invoke->stmt);
+                        assert(asgn->targets.size() == 1);
+                        if (asgn->targets[0]->type == AST_TYPE::Name) {
+                            auto name = ast_cast<AST_Name>(asgn->targets[0]);
+
+                            // TODO: inneficient
+                            sym_table = new SymbolTable(*sym_table);
+                            assert(sym_table->count(name->id));
+                            sym_table->erase(name->id);
+                            created_new_sym_table = true;
+                        }
+                    }
+                }
+
+                generator->copySymbolsFrom(sym_table);
+                if (created_new_sym_table)
+                    delete sym_table;
             } else {
                 // With multiple predecessors, the symbol tables at the end of each predecessor should be *exactly* the
                 // same.

src/codegen/irgen/irgenerator.cpp

@@ -142,6 +142,8 @@ private:
 
         pp_args.insert(pp_args.end(), ic_stackmap_args.begin(), ic_stackmap_args.end());
 
+        irgenerator->addFrameStackmapArgs(info, pp_args);
+
         llvm::Intrinsic::ID intrinsic_id;
         if (return_type->isIntegerTy() || return_type->isPointerTy()) {
             intrinsic_id = llvm::Intrinsic::experimental_patchpoint_i64;
@@ -181,8 +183,27 @@ public:
     CompiledFunction* currentFunction() override { return irstate->getCurFunction(); }
 
     llvm::Value* createCall(ExcInfo exc_info, llvm::Value* callee, const std::vector<llvm::Value*>& args) override {
-        llvm::CallSite cs = this->emitCall(exc_info, callee, args);
-        return cs.getInstruction();
+        if (ENABLE_FRAME_INTROSPECTION) {
+            llvm::Type* rtn_type = llvm::cast<llvm::FunctionType>(llvm::cast<llvm::PointerType>(callee->getType())
+                                                                      ->getElementType())->getReturnType();
+
+            llvm::Value* bitcasted = getBuilder()->CreateBitCast(callee, g.i8->getPointerTo());
+            llvm::CallSite cs = emitPatchpoint(rtn_type, NULL, bitcasted, args, {}, exc_info);
+
+            if (rtn_type == cs->getType()) {
+                return cs.getInstruction();
+            } else if (rtn_type == g.i1) {
+                return getBuilder()->CreateTrunc(cs.getInstruction(), rtn_type);
+            } else if (llvm::isa<llvm::PointerType>(rtn_type)) {
+                return getBuilder()->CreateIntToPtr(cs.getInstruction(), rtn_type);
+            } else {
+                cs.getInstruction()->getType()->dump();
+                rtn_type->dump();
+                RELEASE_ASSERT(0, "don't know how to convert those");
+            }
+        } else {
+            return emitCall(exc_info, callee, args).getInstruction();
+        }
     }
 
     llvm::Value* createCall(ExcInfo exc_info, llvm::Value* callee, llvm::Value* arg1) override {
@@ -2215,6 +2236,23 @@ private:
     }
 
 public:
+    void addFrameStackmapArgs(PatchpointInfo* pp, std::vector<llvm::Value*>& stackmap_args) {
+        int initial_args = stackmap_args.size();
+        if (ENABLE_FRAME_INTROSPECTION) {
+            // TODO: don't need to use a sorted symbol table if we're explicitly recording the names!
+            // nice for debugging though.
+            SortedSymbolTable sorted_symbol_table(symbol_table.begin(), symbol_table.end());
+
+            for (const auto& p : sorted_symbol_table) {
+                CompilerVariable* v = p.second;
+
+                v->serializeToFrame(stackmap_args);
+                pp->addFrameVar(p.first, v->getType());
+            }
+        }
+        pp->setNumFrameArgs(stackmap_args.size() - initial_args);
+    }
+
     EndingState getEndingSymbolTable() override {
         assert(state == FINISHED || state == DEAD);
 

src/codegen/irgen/irgenerator.h

@@ -34,6 +34,7 @@ namespace pyston {
 
 class CFGBlock;
 class GCBuilder;
+class PatchpointInfo;
 class ScopeInfo;
 class TypeAnalysis;
 
@@ -196,6 +197,7 @@ public:
     virtual void run(const CFGBlock* block) = 0;
     virtual EndingState getEndingSymbolTable() = 0;
     virtual void doSafePoint() = 0;
+    virtual void addFrameStackmapArgs(PatchpointInfo* pp, std::vector<llvm::Value*>& stackmap_args) = 0;
 };
 
 class IREmitter;

src/codegen/llvm_interpreter.cpp

@@ -25,6 +25,7 @@
 #include "llvm/IR/Module.h"
 
 #include "codegen/codegen.h"
+#include "codegen/compvars.h"
 #include "codegen/irgen/hooks.h"
 #include "codegen/irgen/util.h"
 #include "codegen/patchpoints.h"
@@ -32,6 +33,7 @@
 #include "core/stats.h"
 #include "core/thread_utils.h"
 #include "core/util.h"
+#include "runtime/types.h"
 
 //#undef VERBOSITY
 //#define VERBOSITY(x) 2
@@ -230,6 +232,60 @@ void gatherInterpreterRoots(GCVisitor* visitor) {
                                 visitor);
 }
 
+BoxedDict* localsForInterpretedFrame(void* frame_ptr, bool only_user_visible) {
+    llvm::Instruction* inst = cur_instruction_map[frame_ptr];
+    assert(inst);
+
+    const SymMap* syms = root_stack_set.get()->back();
+    assert(syms);
+
+    ASSERT(llvm::isa<llvm::CallInst>(inst) || llvm::isa<llvm::InvokeInst>(inst),
+           "trying to unwind from not within a patchpoint!");
+
+    llvm::CallSite CS(inst);
+    llvm::Function* f = CS.getCalledFunction();
+    assert(startswith(f->getName(), "llvm.experimental.patchpoint."));
+
+    llvm::Value* pp_arg = CS.getArgument(0);
+    int64_t pp_id = llvm::cast<llvm::ConstantInt>(pp_arg)->getSExtValue();
+    PatchpointInfo* pp = reinterpret_cast<PatchpointInfo*>(pp_id);
+
+    llvm::DataLayout dl(inst->getParent()->getParent()->getParent());
+
+    BoxedDict* rtn = new BoxedDict();
+
+    int stackmap_args_start = 4 + llvm::cast<llvm::ConstantInt>(CS.getArgument(3))->getZExtValue();
+    assert(CS.arg_size() == stackmap_args_start + pp->totalStackmapArgs());
+
+    // TODO: too much duplication here with other code
+
+    int cur_arg = pp->frameStackmapArgsStart();
+    for (const PatchpointInfo::FrameVarInfo& frame_var : pp->getFrameVars()) {
+        int num_args = frame_var.type->numFrameArgs();
+
+        if (only_user_visible && (frame_var.name[0] == '!' || frame_var.name[0] == '#')) {
+            cur_arg += num_args;
+            continue;
+        }
+
+        llvm::SmallVector<uint64_t, 1> vals;
+
+        for (int i = cur_arg, e = cur_arg + num_args; i < e; i++) {
+            Val r = fetch(CS.getArgument(stackmap_args_start + i), dl, *syms);
+            vals.push_back(r.n);
+        }
+
+        Box* b = frame_var.type->deserializeFromFrame(vals);
+        ASSERT(gc::isValidGCObject(b), "%p", b);
+        rtn->d[boxString(frame_var.name)] = b;
+
+        cur_arg += num_args;
+    }
+    assert(cur_arg - pp->frameStackmapArgsStart() == pp->numFrameStackmapArgs());
+
+    return rtn;
+}
+
 class UnregisterHelper {
 private:
     void* frame_ptr;

src/codegen/llvm_interpreter.h

@@ -22,6 +22,7 @@ class Function;
 namespace pyston {
 
 class Box;
+class BoxedDict;
 class GCVisitor;
 class LineInfo;
 
@@ -30,6 +31,7 @@ Box* interpretFunction(llvm::Function* f, int nargs, Box* closure, Box* generato
 
 void gatherInterpreterRoots(GCVisitor* visitor);
 const LineInfo* getLineInfoForInterpretedFrame(void* frame_ptr);
+BoxedDict* localsForInterpretedFrame(void* frame_ptr, bool only_user_visible);
 }
 
 #endif

src/codegen/patchpoints.cpp

@@ -28,11 +28,15 @@
 
 namespace pyston {
 
+void PatchpointInfo::addFrameVar(const std::string& name, CompilerType* type) {
+    frame_vars.push_back(FrameVarInfo({ .name = name, .type = type }));
+}
+
 int ICSetupInfo::totalSize() const {
     int call_size = CALL_ONLY_SIZE;
     if (getCallingConvention() != llvm::CallingConv::C) {
         // 14 bytes per reg that needs to be spilled
-        call_size += 14 * 6;
+        call_size += 14 * 4;
     }
     return num_slots * slot_size + call_size;
 }
@@ -59,6 +63,33 @@ int PatchpointInfo::patchpointSize() {
     return CALL_ONLY_SIZE;
 }
 
+void PatchpointInfo::parseLocationMap(StackMap::Record* r, LocationMap* map) {
+    assert(r->locations.size() == totalStackmapArgs());
+
+    int cur_arg = frameStackmapArgsStart();
+
+    // printf("parsing pp %ld:\n", reinterpret_cast<int64_t>(this));
+
+    for (FrameVarInfo& frame_var : frame_vars) {
+        int num_args = frame_var.type->numFrameArgs();
+
+        llvm::SmallVector<StackMap::Record::Location, 1> locations;
+        locations.append(&r->locations[cur_arg], &r->locations[cur_arg + num_args]);
+
+        // printf("%s %d %d\n", frame_var.name.c_str(), r->locations[cur_arg].type, r->locations[cur_arg].regnum);
+
+        map->names[frame_var.name].locations.push_back(
+            LocationMap::LocationTable::LocationEntry({ ._debug_pp_id = (uint64_t) this,
+                                                        .offset = r->offset,
+                                                        .length = patchpointSize(),
+                                                        .type = frame_var.type,
+                                                        .locations = std::move(locations) }));
+
+        cur_arg += num_args;
+    }
+    assert(cur_arg - frameStackmapArgsStart() == numFrameStackmapArgs());
+}
+
 static int extractScratchOffset(PatchpointInfo* pp, StackMap::Record* r) {
     StackMap::Record::Location l = r->locations[pp->scratchStackmapArg()];
 
@@ -95,8 +126,22 @@ static std::unordered_set<int> extractLiveOuts(StackMap::Record* r, llvm::Callin
 }
 
 void processStackmap(CompiledFunction* cf, StackMap* stackmap) {
+    // FIXME: this is pretty hacky, that we don't delete the patchpoints here.
+    // We need them currently for the llvm interpreter.
+    // Eventually we'll get rid of that and use an AST interpreter, and then we won't need this hack.
+    if (cf->effort == EffortLevel::INTERPRETED) {
+        assert(!stackmap);
+        new_patchpoints.clear();
+        return;
+    }
+
     int nrecords = stackmap ? stackmap->records.size() : 0;
 
+    assert(cf->location_map == NULL);
+    cf->location_map = new LocationMap();
+    if (stackmap)
+        cf->location_map->constants = stackmap->constants;
+
     for (int i = 0; i < nrecords; i++) {
         StackMap::Record* r = stackmap->records[i];
 
@@ -126,12 +171,36 @@ void processStackmap(CompiledFunction* cf, StackMap* stackmap) {
         //*start_addr = 0xcc;
         // start_addr++;
 
+        int nspills = 0;
+        SpillMap frame_remapped;
+        // TODO: if we pass the same llvm::Value as the stackmap args, we will get the same reg.
+        // we shouldn't then spill that multiple times.
+        // we could either deal with it here, or not put the same Value into the patchpoint
+        for (int j = pp->frameStackmapArgsStart(), e = j + pp->numFrameStackmapArgs(); j < e; j++) {
+            StackMap::Record::Location& l = r->locations[j];
+
+            // updates l, start_addr, scratch_rbp_offset, scratch_size:
+            bool spilled = spillFrameArgumentIfNecessary(l, start_addr, end_addr, scratch_rbp_offset, scratch_size,
+                                                         frame_remapped);
+            if (spilled)
+                nspills++;
+        }
+        ASSERT(nspills <= MAX_FRAME_SPILLS, "did %d spills but expected only %d!", nspills, MAX_FRAME_SPILLS);
+
+        assert(scratch_size % sizeof(void*) == 0);
+        assert(scratch_rbp_offset % sizeof(void*) == 0);
+
+        // TODO: is something like this necessary?
+        // llvm::sys::Memory::InvalidateInstructionCache(start, getSlotSize());
+
+        pp->parseLocationMap(r, cf->location_map);
+
         const ICSetupInfo* ic = pp->getICInfo();
         if (ic == NULL) {
             // We have to be using the C calling convention here, so we don't need to check the live outs
             // or save them across the call.
             initializePatchpoint3(slowpath_func, start_addr, end_addr, scratch_rbp_offset, scratch_size,
-                                  std::unordered_set<int>());
+                                  std::unordered_set<int>(), frame_remapped);
             continue;
         }
 
@@ -151,8 +220,8 @@ void processStackmap(CompiledFunction* cf, StackMap* stackmap) {
 
 
 
-        auto _p
-            = initializePatchpoint3(slowpath_func, start_addr, end_addr, scratch_rbp_offset, scratch_size, live_outs);
+        auto _p = initializePatchpoint3(slowpath_func, start_addr, end_addr, scratch_rbp_offset, scratch_size,
+                                        live_outs, frame_remapped);
         uint8_t* slowpath_start = _p.first;
         uint8_t* slowpath_rtn_addr = _p.second;
 

src/codegen/patchpoints.h

@@ -31,18 +31,32 @@ struct StackMap;
 class TypeRecorder;
 class ICSetupInfo;
 
-static const int CALL_ONLY_SIZE = 13;
+static const int MAX_FRAME_SPILLS = 9; // TODO this shouldn't have to be larger than the set of non-callee-save args (9)
+                                       // except that will we currently spill the same reg multiple times
+static const int CALL_ONLY_SIZE
+    = 13 + (MAX_FRAME_SPILLS * 9)
+      + 1; // 13 for the call, 9 bytes per spill (7 for GP, 9 for XMM), + 1 if we want to nop/trap
 
 void processStackmap(CompiledFunction* cf, StackMap* stackmap);
 
 struct PatchpointInfo {
+public:
+    struct FrameVarInfo {
+        std::string name;
+        CompilerType* type;
+    };
+
 private:
     CompiledFunction* const parent_cf;
     const ICSetupInfo* icinfo;
     int num_ic_stackmap_args;
+    int num_frame_stackmap_args;
+
+    std::vector<FrameVarInfo> frame_vars;
 
     PatchpointInfo(CompiledFunction* parent_cf, const ICSetupInfo* icinfo, int num_ic_stackmap_args)
-        : parent_cf(parent_cf), icinfo(icinfo), num_ic_stackmap_args(num_ic_stackmap_args) {}
+        : parent_cf(parent_cf), icinfo(icinfo), num_ic_stackmap_args(num_ic_stackmap_args),
+          num_frame_stackmap_args(-1) {}
 
 public:
     const ICSetupInfo* getICInfo() { return icinfo; }
@@ -50,13 +64,29 @@ public:
     int patchpointSize();
     CompiledFunction* parentFunction() { return parent_cf; }
 
+    const std::vector<FrameVarInfo>& getFrameVars() { return frame_vars; }
+
     int scratchStackmapArg() { return 0; }
-    int scratchSize() { return 80; }
+    int scratchSize() { return 80 + MAX_FRAME_SPILLS * sizeof(void*); }
+
+    void addFrameVar(const std::string& name, CompilerType* type);
+    void setNumFrameArgs(int num_frame_args) {
+        assert(num_frame_stackmap_args == -1);
+        num_frame_stackmap_args = num_frame_args;
+    }
 
     int icStackmapArgsStart() { return 1; }
     int numICStackmapArgs() { return num_ic_stackmap_args; }
 
-    int totalStackmapArgs() { return icStackmapArgsStart() + numICStackmapArgs(); }
+    int frameStackmapArgsStart() { return icStackmapArgsStart() + numICStackmapArgs(); }
+    int numFrameStackmapArgs() {
+        assert(num_frame_stackmap_args >= 0);
+        return num_frame_stackmap_args;
+    }
+
+    void parseLocationMap(StackMap::Record* r, LocationMap* map);
+
+    int totalStackmapArgs() { return frameStackmapArgsStart() + numFrameStackmapArgs(); }
 
     static PatchpointInfo* create(CompiledFunction* parent_cf, const ICSetupInfo* icinfo, int num_ic_stackmap_args);
 };

src/codegen/stackmaps.h

@@ -16,14 +16,19 @@
 #define PYSTON_CODEGEN_STACKMAPS_H
 
 #include <cstdint>
+#include <unordered_map>
 #include <vector>
 
+#include "llvm/ADT/SmallVector.h"
+
 namespace llvm {
 class JITEventListener;
 }
 
 namespace pyston {
 
+class CompilerType;
+
 struct StackMap {
     struct __attribute__((__packed__)) StackSizeRecord {
         uint64_t offset;
@@ -64,6 +69,25 @@ struct StackMap {
     std::vector<Record*> records;
 };
 
+// TODO this belongs somewhere else?
+class LocationMap {
+public:
+    std::vector<uint64_t> constants;
+    struct LocationTable {
+        struct LocationEntry {
+            uint64_t _debug_pp_id;
+
+            unsigned offset;
+            int length;
+            CompilerType* type;
+            llvm::SmallVector<StackMap::Record::Location, 1> locations;
+        };
+        std::vector<LocationEntry> locations;
+    };
+
+    std::unordered_map<std::string, LocationTable> names;
+};
+
 StackMap* parseStackMap();
 llvm::JITEventListener* makeStackMapListener();
 }

src/codegen/unwinding.cpp

@@ -27,6 +27,8 @@
 #include "codegen/compvars.h"
 #include "codegen/irgen/hooks.h"
 #include "codegen/llvm_interpreter.h"
+#include "codegen/stackmaps.h"
+#include "runtime/types.h"
 
 
 #define UNW_LOCAL_ONLY
@@ -333,7 +335,7 @@ public:
                 return *this;
             }
 
-            const PythonFrameIterator& operator*() const {
+            PythonFrameIterator& operator*() const {
                 assert(it.get());
                 return *it.get();
             }
@@ -397,6 +399,71 @@ BoxedModule* getCurrentModule() {
     return getTopCompiledFunction()->clfunc->source->parent_module;
 }
 
+BoxedDict* getLocals(bool only_user_visible) {
+    BoxedDict* d = new BoxedDict();
+
+    for (PythonFrameIterator& frame_info : unwindPythonFrames()) {
+        if (frame_info.getId().type == PythonFrameId::COMPILED) {
+            CompiledFunction* cf = frame_info.getCF();
+            uint64_t ip = frame_info.getId().ip;
+
+            assert(ip > cf->code_start);
+            unsigned offset = ip - cf->code_start;
+
+            assert(cf->location_map);
+            for (const auto& p : cf->location_map->names) {
+                if (only_user_visible && (p.first[0] == '#' || p.first[0] == '!'))
+                    continue;
+
+                for (const LocationMap::LocationTable::LocationEntry& e : p.second.locations) {
+                    if (e.offset < offset && offset <= e.offset + e.length) {
+                        const auto& locs = e.locations;
+
+                        llvm::SmallVector<uint64_t, 1> vals;
+                        // printf("%s: %s\n", p.first.c_str(), e.type->debugName().c_str());
+
+                        for (auto& loc : locs) {
+                            uint64_t n;
+                            // printf("%d %d %d %d\n", loc.type, loc.flags, loc.regnum, loc.offset);
+                            if (loc.type == StackMap::Record::Location::LocationType::Register) {
+                                // TODO: need to make sure we deal with patchpoints appropriately
+                                n = frame_info.getReg(loc.regnum);
+                            } else if (loc.type == StackMap::Record::Location::LocationType::Indirect) {
+                                uint64_t reg_val = frame_info.getReg(loc.regnum);
+                                uint64_t addr = reg_val + loc.offset;
+                                n = *reinterpret_cast<uint64_t*>(addr);
+                            } else if (loc.type == StackMap::Record::Location::LocationType::Constant) {
+                                n = loc.offset;
+                            } else if (loc.type == StackMap::Record::Location::LocationType::ConstIndex) {
+                                int const_idx = loc.offset;
+                                assert(const_idx >= 0);
+                                assert(const_idx < cf->location_map->constants.size());
+                                n = cf->location_map->constants[const_idx];
+                            } else {
+                                printf("%d %d %d %d\n", loc.type, loc.flags, loc.regnum, loc.offset);
+                                abort();
+                            }
+
+                            vals.push_back(n);
+                        }
+
+                        Box* v = e.type->deserializeFromFrame(vals);
+                        // printf("%s: (pp id %ld) %p\n", p.first.c_str(), e._debug_pp_id, v);
+                        assert(gc::isValidGCObject(v));
+                        d->d[boxString(p.first)] = v;
+                    }
+                }
+            }
+        } else if (frame_info.getId().type == PythonFrameId::INTERPRETED) {
+            return localsForInterpretedFrame((void*)frame_info.getId().bp, only_user_visible);
+        } else {
+            abort();
+        }
+    }
+
+    return d;
+}
+
 
 llvm::JITEventListener* makeTracebacksListener() {
     return new TracebacksEventListener();

src/core/options.cpp

@@ -55,4 +55,7 @@ bool ENABLE_INLINING = 1 && _GLOBAL_ENABLE;
 bool ENABLE_REOPT = 1 && _GLOBAL_ENABLE;
 bool ENABLE_PYSTON_PASSES = 1 && _GLOBAL_ENABLE;
 bool ENABLE_TYPE_FEEDBACK = 1 && _GLOBAL_ENABLE;
+
+bool ENABLE_FRAME_INTROSPECTION = 0;
+bool BOOLS_AS_I64 = ENABLE_FRAME_INTROSPECTION;
 }

src/core/options.h

@@ -35,7 +35,10 @@ extern bool SHOW_DISASM, FORCE_OPTIMIZE, BENCH, PROFILE, DUMPJIT, TRAP, USE_STRI
 extern bool ENABLE_ICS, ENABLE_ICGENERICS, ENABLE_ICGETITEMS, ENABLE_ICSETITEMS, ENABLE_ICDELITEMS, ENABLE_ICBINEXPS,
     ENABLE_ICNONZEROS, ENABLE_ICCALLSITES, ENABLE_ICSETATTRS, ENABLE_ICGETATTRS, ENALBE_ICDELATTRS, ENABLE_ICGETGLOBALS,
     ENABLE_SPECULATION, ENABLE_OSR, ENABLE_LLVMOPTS, ENABLE_INLINING, ENABLE_REOPT, ENABLE_PYSTON_PASSES,
-    ENABLE_TYPE_FEEDBACK;
+    ENABLE_TYPE_FEEDBACK, ENABLE_FRAME_INTROSPECTION;
+
+// Due to a temporary LLVM limitation, represent bools as i64's instead of i1's.
+extern bool BOOLS_AS_I64;
 }
 }
 

src/core/types.h

@@ -159,6 +159,7 @@ class BoxedClosure;
 class BoxedGenerator;
 class LineTable;
 class ICInfo;
+class LocationMap;
 
 struct CompiledFunction {
 private:
@@ -188,13 +189,16 @@ public:
     // Unfortunately, can't make this a std::unique_ptr if we want to forward-declare LineTable:
     LineTable* line_table;
 
+    LocationMap* location_map; // only meaningful if this is a compiled frame
+
     std::vector<ICInfo*> ics;
 
     CompiledFunction(llvm::Function* func, FunctionSpecialization* spec, bool is_interpreted, void* code,
                      llvm::Value* llvm_code, EffortLevel::EffortLevel effort,
                      const OSREntryDescriptor* entry_descriptor)
         : clfunc(NULL), func(func), spec(spec), entry_descriptor(entry_descriptor), is_interpreted(is_interpreted),
-          code(code), llvm_code(llvm_code), effort(effort), times_called(0), line_table(nullptr) {}
+          code(code), llvm_code(llvm_code), effort(effort), times_called(0), line_table(nullptr),
+          location_map(nullptr) {}
 
     // TODO this will need to be implemented eventually, and delete line_table if it exists
     ~CompiledFunction();

src/runtime/builtin_modules/builtins.cpp

@@ -20,6 +20,7 @@
 
 #include "codegen/irgen/hooks.h"
 #include "codegen/parser.h"
+#include "codegen/unwinding.h"
 #include "core/ast.h"
 #include "core/types.h"
 #include "gc/collector.h"
@@ -573,6 +574,10 @@ Box* globals() {
     return makeAttrWrapper(m);
 }
 
+Box* locals() {
+    return getLocals(true /* filter */);
+}
+
 Box* divmod(Box* lhs, Box* rhs) {
     return binopInternal(lhs, rhs, AST_TYPE::DivMod, false, NULL);
 }
@@ -788,6 +793,7 @@ void setupBuiltins() {
     builtins_module->giveAttr("open", open_obj);
 
     builtins_module->giveAttr("globals", new BoxedFunction(boxRTFunction((void*)globals, UNKNOWN, 0, 0, false, false)));
+    builtins_module->giveAttr("locals", new BoxedFunction(boxRTFunction((void*)locals, UNKNOWN, 0, 0, false, false)));
 
     builtins_module->giveAttr("iter", new BoxedFunction(boxRTFunction((void*)getiter, UNKNOWN, 1, 0, false, false)));