Basic (new) deopt support

Old deopt worked by compiling two copies of every BB, one with
speculations and one without, and stitching the two together.
This has a number of issues:
- doubles the amount of code LLVM has to jit
- can't ever get back on the optimized path
- doesn't support 'deopt if branch taken'
- horrifically complex
- doesn't support deopt from within try blocks

We actually ran into that last issue (see test from previous commit).  So
rather than wade in and try to fix old-deopt, just start switching to new-deopt.

(new) deopt works by using the frame introspection features, gathering up all
the locals, and passing them to the interpreter.

src/codegen/ast_interpreter.cpp

@@ -65,7 +65,7 @@ public:
 
     void initArguments(int nargs, BoxedClosure* closure, BoxedGenerator* generator, Box* arg1, Box* arg2, Box* arg3,
                        Box** args);
-    static Value execute(ASTInterpreter& interpreter, AST_stmt* start_at = NULL);
+    static Value execute(ASTInterpreter& interpreter, CFGBlock* start_block = NULL, AST_stmt* start_at = NULL);
 
 private:
     Box* createFunction(AST* node, AST_arguments* args, const std::vector<AST_stmt*>& body);
@@ -138,6 +138,7 @@ public:
     CompiledFunction* getCF() { return compiled_func; }
     FrameInfo* getFrameInfo() { return &frame_info; }
     const SymMap& getSymbolTable() { return sym_table; }
+    void addSymbol(const std::string& name, Box* value, bool allow_duplicates);
     void gcVisit(GCVisitor* visitor);
 };
 
@@ -169,12 +170,76 @@ Box* astInterpretFunction(CompiledFunction* cf, int nargs, Box* closure, Box* ge
     return v.o ? v.o : None;
 }
 
-Box* astInterpretFrom(CompiledFunction* cf, AST_stmt* start_at, BoxedDict* locals) {
-    assert(locals->d.size() == 0);
+void ASTInterpreter::addSymbol(const std::string& name, Box* value, bool allow_duplicates) {
+    if (!allow_duplicates)
+        assert(sym_table.count(name) == 0);
+    sym_table[name] = value;
+}
+
+Box* astInterpretFrom(CompiledFunction* cf, AST_expr* after_expr, AST_stmt* enclosing_stmt, Box* expr_val,
+                      BoxedDict* locals) {
+    assert(cf);
+    assert(enclosing_stmt);
+    assert(locals);
+    assert(after_expr);
+    assert(expr_val);
 
     ASTInterpreter interpreter(cf);
 
-    Value v = ASTInterpreter::execute(interpreter, start_at);
+    for (const auto& p : locals->d) {
+        assert(p.first->cls == str_cls);
+        interpreter.addSymbol(static_cast<BoxedString*>(p.first)->s, p.second, false);
+    }
+
+    CFGBlock* start_block = NULL;
+    AST_stmt* starting_statement = NULL;
+    while (true) {
+        if (enclosing_stmt->type == AST_TYPE::Assign) {
+            auto asgn = ast_cast<AST_Assign>(enclosing_stmt);
+            assert(asgn->value == after_expr);
+            assert(asgn->targets.size() == 1);
+            assert(asgn->targets[0]->type == AST_TYPE::Name);
+            auto name = ast_cast<AST_Name>(asgn->targets[0]);
+            assert(name->id[0] == '#');
+            interpreter.addSymbol(name->id, expr_val, true);
+            break;
+        } else if (enclosing_stmt->type == AST_TYPE::Expr) {
+            auto expr = ast_cast<AST_Expr>(enclosing_stmt);
+            assert(expr->value == after_expr);
+            break;
+        } else if (enclosing_stmt->type == AST_TYPE::Invoke) {
+            auto invoke = ast_cast<AST_Invoke>(enclosing_stmt);
+            start_block = invoke->normal_dest;
+            starting_statement = start_block->body[0];
+            enclosing_stmt = invoke->stmt;
+        } else {
+            RELEASE_ASSERT(0, "should not be able to reach here with anything other than an Assign (got %d)",
+                           enclosing_stmt->type);
+        }
+    }
+
+    if (start_block == NULL) {
+        // TODO innefficient
+        for (auto block : cf->clfunc->source->cfg->blocks) {
+            int n = block->body.size();
+            for (int i = 0; i < n; i++) {
+                if (block->body[i] == enclosing_stmt) {
+                    ASSERT(i + 1 < n, "how could we deopt from a non-invoke terminator?");
+                    start_block = block;
+                    starting_statement = block->body[i + 1];
+                    break;
+                }
+            }
+
+            if (start_block)
+                break;
+        }
+
+        ASSERT(start_block, "was unable to find the starting block??");
+        assert(starting_statement);
+    }
+
+    Value v = ASTInterpreter::execute(interpreter, start_block, starting_statement);
 
     return v.o ? v.o : None;
 }
@@ -288,16 +353,33 @@ public:
 };
 }
 
-Value ASTInterpreter::execute(ASTInterpreter& interpreter, AST_stmt* start_at) {
+Value ASTInterpreter::execute(ASTInterpreter& interpreter, CFGBlock* start_block, AST_stmt* start_at) {
     threading::allowGLReadPreemption();
 
-    assert(start_at == NULL);
-
     void* frame_addr = __builtin_frame_address(0);
     RegisterHelper frame_registerer(&interpreter, frame_addr);
 
     Value v;
-    interpreter.next_block = interpreter.source_info->cfg->getStartingBlock();
+
+    assert((start_block == NULL) == (start_at == NULL));
+    if (start_block == NULL) {
+        start_block = interpreter.source_info->cfg->getStartingBlock();
+        start_at = start_block->body[0];
+    }
+
+    interpreter.current_block = start_block;
+    bool started = false;
+    for (auto s : start_block->body) {
+        if (!started) {
+            if (s != start_at)
+                continue;
+            started = true;
+        }
+
+        interpreter.current_inst = s;
+        v = interpreter.visit_stmt(s);
+    }
+
     while (interpreter.next_block) {
         interpreter.current_block = interpreter.next_block;
         interpreter.next_block = 0;

src/codegen/ast_interpreter.h

@@ -21,6 +21,7 @@ namespace gc {
 class GCVisitor;
 }
 
+class AST_expr;
 class AST_stmt;
 class Box;
 class BoxedDict;
@@ -31,7 +32,8 @@ extern const void* interpreter_instr_addr;
 
 Box* astInterpretFunction(CompiledFunction* f, int nargs, Box* closure, Box* generator, Box* arg1, Box* arg2, Box* arg3,
                           Box** args);
-Box* astInterpretFrom(CompiledFunction* cf, AST_stmt* start_at, BoxedDict* locals);
+Box* astInterpretFrom(CompiledFunction* cf, AST_expr* after_expr, AST_stmt* enclosing_stmt, Box* expr_val,
+                      BoxedDict* locals);
 
 AST_stmt* getCurrentStatementForInterpretedFrame(void* frame_ptr);
 CompiledFunction* getCFForInterpretedFrame(void* frame_ptr);

src/codegen/irgen/irgenerator.cpp

@@ -346,7 +346,8 @@ private:
 
     OpInfo getEmptyOpInfo(UnwindInfo unw_info) { return OpInfo(irstate->getEffortLevel(), NULL, unw_info); }
 
-    void createExprTypeGuard(llvm::Value* check_val, AST_expr* node, CompilerVariable* node_value) {
+    void createExprTypeGuard(llvm::Value* check_val, AST_expr* node, llvm::Value* node_value,
+                             AST_stmt* current_statement) {
         assert(check_val->getType() == g.i1);
 
         llvm::Metadata* md_vals[]
@@ -361,17 +362,18 @@ private:
             = llvm::BasicBlock::Create(g.context, "check_succeeded", irstate->getLLVMFunction());
         success_bb->moveAfter(curblock);
 
-        // Create the guard with both branches leading to the success_bb,
-        // and let the deopt path change the failure case to point to the
-        // as-yet-unknown deopt block.
-        // TODO Not the best approach since if we fail to do that patching,
-        // the guard will just silently be ignored.
-        llvm::BranchInst* guard = emitter.getBuilder()->CreateCondBr(check_val, success_bb, success_bb, branch_weights);
+        llvm::BasicBlock* deopt_bb = llvm::BasicBlock::Create(g.context, "check_failed", irstate->getLLVMFunction());
 
-        curblock = success_bb;
+        llvm::BranchInst* guard = emitter.getBuilder()->CreateCondBr(check_val, success_bb, deopt_bb, branch_weights);
+
+        curblock = deopt_bb;
         emitter.getBuilder()->SetInsertPoint(curblock);
+        llvm::Value* v = emitter.createCall2(UnwindInfo(current_statement, NULL), g.funcs.deopt,
+                                             embedConstantPtr(node, g.i8->getPointerTo()), node_value);
+        emitter.getBuilder()->CreateRet(v);
 
-        out_guards.addExprTypeGuard(myblock, guard, node, node_value, symbol_table);
+        curblock = success_bb;
+        emitter.getBuilder()->SetInsertPoint(curblock);
     }
 
     CompilerVariable* evalAttribute(AST_Attribute* node, UnwindInfo unw_info) {
@@ -1244,7 +1246,7 @@ private:
 
                 llvm::Value* guard_check = old_rtn->makeClassCheck(emitter, speculated_class);
                 assert(guard_check->getType() == g.i1);
-                createExprTypeGuard(guard_check, node, old_rtn);
+                createExprTypeGuard(guard_check, node, old_rtn->getValue(), unw_info.current_stmt);
 
                 rtn = unboxVar(speculated_type, old_rtn->getValue(), true);
             }

src/codegen/irgen/irgenerator.h

@@ -158,6 +158,7 @@ public:
 
     void addExprTypeGuard(CFGBlock* cfg_block, llvm::BranchInst* branch, AST_expr* ast_node, CompilerVariable* val,
                           const SymbolTable& st) {
+        abort();
         ExprTypeGuard*& g = expr_type_guards[ast_node];
         assert(g == NULL);
         g = new ExprTypeGuard(cfg_block, branch, ast_node, val, st);

src/codegen/runtime_hooks.cpp

@@ -248,6 +248,7 @@ void initGlobalFuncs(GlobalState& g) {
     g.funcs.__cxa_end_catch = addFunc((void*)__cxa_end_catch, g.void_);
     GET(raise0);
     GET(raise3);
+    GET(deopt);
 
     GET(div_float_float);
     GET(floordiv_float_float);

src/codegen/runtime_hooks.h

@@ -48,6 +48,7 @@ struct GlobalFuncs {
 
     llvm::Value* __cxa_begin_catch, *__cxa_end_catch;
     llvm::Value* raise0, *raise3;
+    llvm::Value* deopt;
 
     llvm::Value* div_float_float, *floordiv_float_float, *mod_float_float, *pow_float_float;
 

src/codegen/unwinding.cpp

@@ -586,6 +586,12 @@ BoxedDict* getLocals(bool only_user_visible) {
     RELEASE_ASSERT(0, "Internal error: unable to find any python frames");
 }
 
+ExecutionPoint getExecutionPoint() {
+    auto frame = getTopPythonFrame();
+    auto cf = frame->getCF();
+    auto current_stmt = frame->getCurrentStatement();
+    return ExecutionPoint({.cf = cf, .current_stmt = current_stmt });
+}
 
 llvm::JITEventListener* makeTracebacksListener() {
     return new TracebacksEventListener();

src/codegen/unwinding.h

@@ -32,6 +32,12 @@ BoxedDict* getLocals(bool only_user_visible);
 // Fetches a writeable pointer to the frame-local excinfo object,
 // calculating it if necessary (from previous frames).
 ExcInfo* getFrameExcInfo();
+
+struct ExecutionPoint {
+    CompiledFunction* cf;
+    AST_stmt* current_stmt;
+};
+ExecutionPoint getExecutionPoint();
 }
 
 #endif

src/runtime/inline/link_forcer.cpp

@@ -106,6 +106,7 @@ void force() {
 
     FORCE(raise0);
     FORCE(raise3);
+    FORCE(deopt);
 
     FORCE(div_i64_i64);
     FORCE(mod_i64_i64);

src/runtime/objmodel.cpp

@@ -31,6 +31,7 @@
 #include "codegen/irgen/hooks.h"
 #include "codegen/parser.h"
 #include "codegen/type_recording.h"
+#include "codegen/unwinding.h"
 #include "core/ast.h"
 #include "core/options.h"
 #include "core/stats.h"
@@ -226,6 +227,15 @@ bool PyLt::operator()(Box* lhs, Box* rhs) const {
     return cmp == True;
 }
 
+extern "C" Box* deopt(AST_expr* expr, Box* value) {
+    static StatCounter num_deopt("num_deopt");
+    num_deopt.log();
+
+    auto locals = getLocals(false /* filter */);
+    auto execution_point = getExecutionPoint();
+    return astInterpretFrom(execution_point.cf, expr, execution_point.current_stmt, value, locals);
+}
+
 extern "C" bool softspace(Box* b, bool newval) {
     assert(b);
 

src/runtime/objmodel.h

@@ -36,6 +36,8 @@ extern "C" void raise3(Box*, Box*, Box*) __attribute__((__noreturn__));
 void raiseExc(Box* exc_obj) __attribute__((__noreturn__));
 void raiseRaw(const ExcInfo& e) __attribute__((__noreturn__));
 
+extern "C" Box* deopt(AST_expr* expr, Box* value);
+
 // helper function for raising from the runtime:
 void raiseExcHelper(BoxedClass*, const char* fmt, ...) __attribute__((__noreturn__));
 

test/tests/deopt_tests.py

+# statcheck: 0 <= noninit_count('num_deopt') < 500
+
+def f(o):
+    print "starting"
+
+    try:
+        print o.a
+        if o.b:
+            raise Exception()
+    except Exception, e:
+        print o.c
+        print e
+    print o.d
+
+    print "Done"
+
+class C(object):
+    pass
+c = C()
+c.a = 1
+c.b = 0
+c.c = 3
+c.d = 4
+
+# These limits are high to try to trigger OSR.
+# TODO we should have some way to lower the OSR thresholds
+for i in xrange(20000):
+    if i == 5000:
+        c.a = []
+
+    if i == 6000:
+        c.b = 1
+
+    if i == 7000:
+        c.c = []
+
+    if i == 8000:
+        c.b = 0
+        c.d = 1.0
+
+    f(c)

test/tests/regr_test.py

@@ -16,3 +16,6 @@ c = C()
 c.pid = 1
 for i in xrange(20000):
     f(c, None)
+
+    if i == 15000:
+        c.pid = 1.0

tools/tester.py

@@ -280,7 +280,7 @@ def run_test(fn, check_stats, run_memcheck):
                         statname = m.group(1)
                         raise Exception((l, statname, stats[statname]))
 
-                    m = re.match("""noninit_count\(['"]([\w_]+)['"]\)""", l)
+                    m = re.search("""noninit_count\(['"]([\w_]+)['"]\)""", l)
                     if m:
                         statname = m.group(1)
                         raise Exception((l, statname, noninit_count(statname)))