Micro optimizations of interpreter loop (#47)

- make RemainingStackSize inlinable - mark an out-of-bounds check as unlikely - make ErrorExit and Quit non-inlinable This seem to give a surprising amount of speed: median -3% (min. -25%, max. 10%) https://rebench.dev/SOMpp/compare/4ba47cfe11c5e5b4d5379c36de923faf601ffc93..8866a58999526c54a50fc5d4e4c948ee94176a89
SOM-st · Aug 9, 2024 · e1feaf1 · e1feaf1
2 parents 4ba47cf + 8866a58
commit e1feaf1
Show file tree

Hide file tree

Showing 27 changed files with 110 additions and 87 deletions.
diff --git a/src/Main.cpp b/src/Main.cpp
@@ -67,5 +67,5 @@ int main(int argc, char** argv) {
 
     Universe::Start(argc, argv);
 
-    Universe::Quit(ERR_SUCCESS);
+    Quit(ERR_SUCCESS);
 }
diff --git a/src/compiler/Lexer.cpp b/src/compiler/Lexer.cpp
@@ -33,7 +33,7 @@
 #include <string>
 
 #include "../misc/defs.h"
-#include "../vm/Universe.h"
+#include "../vm/Print.h"
 
 Lexer::Lexer(istream& file) : infile(file), peekDone(false) {}
 
@@ -322,7 +322,7 @@ Symbol Lexer::Peek() {
     const LexerState old = state;
 
     if (peekDone) {
-        Universe::ErrorExit("Cannot Peek twice!\n");
+        ErrorExit("Cannot Peek twice!\n");
     }
     GetSym();
     const Symbol nextSym = state.sym;

diff --git a/src/compiler/MethodGenerationContext.cpp b/src/compiler/MethodGenerationContext.cpp
@@ -38,6 +38,7 @@
 #include "../interpreter/bytecodes.h"
 #include "../misc/VectorUtil.h"
 #include "../vm/Globals.h"
+#include "../vm/Print.h"
 #include "../vm/Universe.h"
 #include "../vmobjects/ObjectFormats.h"
 #include "../vmobjects/VMMethod.h"
@@ -147,7 +148,7 @@ VMTrivialMethod* MethodGenerationContext::assembleLiteralReturn(
         }
     }
 
-    GetUniverse()->ErrorExit(
+    ErrorExit(
         "Unexpected situation when trying to create trivial method that "
         "returns a literal");
 }
@@ -159,7 +160,7 @@ VMTrivialMethod* MethodGenerationContext::assembleGlobalReturn() {
     }
 
     if (literals.size() != 1) {
-        GetUniverse()->ErrorExit(
+        ErrorExit(
             "Unexpected situation when trying to create trivial method that "
             "reads a global. New Bytecode?");
     }
@@ -228,7 +229,7 @@ VMTrivialMethod* MethodGenerationContext::assembleFieldSetter() {
             break;
         }
         default: {
-            GetUniverse()->ErrorExit("Unexpected bytecode");
+            ErrorExit("Unexpected bytecode");
         }
     }
 
@@ -246,7 +247,7 @@ VMTrivialMethod* MethodGenerationContext::assembleFieldSetter() {
             break;
         }
         default: {
-            GetUniverse()->ErrorExit("Unexpected bytecode");
+            ErrorExit("Unexpected bytecode");
         }
     }
 
@@ -708,7 +709,7 @@ uint8_t MethodGenerationContext::GetInlinedLocalIdx(const Variable* var) const {
         "Unexpected issue trying to find an inlined variable. %s could not "
         "be found.",
         qualifiedName.data());
-    Universe::ErrorExit(msg);
+    ErrorExit(msg);
 }
 
 void MethodGenerationContext::checkJumpOffset(size_t jumpOffset,
@@ -718,7 +719,7 @@ void MethodGenerationContext::checkJumpOffset(size_t jumpOffset,
         snprintf(msg, 100,
                  "The jumpOffset for the %s bytecode is out of range: %zu\n",
                  Bytecode::GetBytecodeName(bytecode), jumpOffset);
-        Universe::ErrorExit(msg);
+        ErrorExit(msg);
     }
 }
 
@@ -786,7 +787,7 @@ size_t MethodGenerationContext::getOffsetOfLastBytecode(size_t indexFromEnd) {
     for (size_t i = 0; i < indexFromEnd + 1; i += 1) {
         uint8_t actual = last4Bytecodes.at(NUM_LAST_BYTECODES - 1 - i);
         if (actual == BC_INVALID) {
-            GetUniverse()->ErrorExit("The requested bytecode is invalid");
+            ErrorExit("The requested bytecode is invalid");
         }
 
         bcOffset -= Bytecode::GetBytecodeLength(actual);

diff --git a/src/compiler/Parser.cpp b/src/compiler/Parser.cpp
@@ -908,7 +908,7 @@ __attribute__((noreturn)) void Parser::parseError(const char* msg,
     ReplacePattern(msgWithMeta, "%(found)s", foundStr);
 
     ErrorPrint(msgWithMeta);
-    GetUniverse()->Quit(ERR_FAIL);
+    Quit(ERR_FAIL);
 }
 
 __attribute__((noreturn)) void Parser::parseError(const char* msg,

diff --git a/src/interpreter/Interpreter.cpp b/src/interpreter/Interpreter.cpp
@@ -36,6 +36,7 @@
 #include "../misc/defs.h"
 #include "../vm/Globals.h"
 #include "../vm/IsValidObject.h"
+#include "../vm/Print.h"
 #include "../vm/Universe.h"
 #include "../vmobjects/IntegerBox.h"
 #include "../vmobjects/ObjectFormats.h"
@@ -237,7 +238,7 @@ void Interpreter::doPushFieldWithIndex(uint8_t fieldIndex) {
 
     if (unlikely(IS_TAGGED(self))) {
         o = nullptr;
-        Universe()->ErrorExit("Integers do not have fields!");
+        ErrorExit("Integers do not have fields!");
     } else {
         o = ((VMObject*)self)->GetField(fieldIndex);
     }
@@ -320,7 +321,7 @@ void Interpreter::doPopFieldWithIndex(uint8_t fieldIndex) {
     vm_oop_t o = GetFrame()->Pop();
 
     if (unlikely(IS_TAGGED(self))) {
-        GetUniverse()->ErrorExit("Integers do not have fields that can be set");
+        ErrorExit("Integers do not have fields that can be set");
     } else {
         ((VMObject*)self)->SetField(fieldIndex, o);
     }
@@ -436,7 +437,7 @@ void Interpreter::doInc() {
         double d = static_cast<VMDouble*>(val)->GetEmbeddedDouble();
         val = GetUniverse()->NewDouble(d + 1.0);
     } else {
-        GetUniverse()->ErrorExit("unsupported");
+        ErrorExit("unsupported");
     }
 
     GetFrame()->SetTop(store_root(val));

diff --git a/src/memory/CopyingHeap.cpp b/src/memory/CopyingHeap.cpp
@@ -7,7 +7,6 @@
 
 #include "../misc/defs.h"
 #include "../vm/Print.h"
-#include "../vm/Universe.h"
 #include "../vmobjects/AbstractObject.h"
 #include "CopyingCollector.h"
 #include "Heap.h"
@@ -52,7 +51,7 @@ void CopyingHeap::switchBuffers(bool increaseMemory) {
         currentBuffer = malloc(newSize);
 
         if (currentBuffer == nullptr) {
-            Universe::ErrorExit("unable to allocate heap memory");
+            ErrorExit("unable to allocate heap memory");
         }
 
         currentBufferEnd = (void*)((size_t)currentBuffer + newSize);
@@ -91,7 +90,7 @@ void CopyingHeap::invalidateOldBuffer() {
         oldBuffer = malloc(currentBufSize);
 
         if (oldBuffer == nullptr) {
-            Universe::ErrorExit("unable to allocate heap memory");
+            ErrorExit("unable to allocate heap memory");
         }
 
         oldBufferEnd = (void*)((size_t)oldBuffer + currentBufSize);
@@ -103,7 +102,7 @@ AbstractVMObject* CopyingHeap::AllocateObject(size_t size) {
     nextFreePosition = (void*)((size_t)nextFreePosition + size);
     if (nextFreePosition > currentBufferEnd) {
         ErrorPrint("\nFailed to allocate " + to_string(size) + " Bytes.\n");
-        Universe::Quit(-1);
+        Quit(-1);
     }
 
     // let's see if we have to trigger the GC

diff --git a/src/memory/DebugCopyingHeap.cpp b/src/memory/DebugCopyingHeap.cpp
@@ -7,7 +7,6 @@
 #include <string>
 
 #include "../vm/Print.h"
-#include "../vm/Universe.h"
 #include "../vmobjects/AbstractObject.h"
 
 void DebugCopyingHeap::switchBuffers(bool increaseMemory) {
@@ -57,7 +56,7 @@ AbstractVMObject* DebugCopyingHeap::AllocateObject(size_t size) {
 
     if (currentHeapUsage > currentHeapSize) {
         ErrorPrint("\nFailed to allocate " + to_string(size) + " Bytes.\n");
-        Universe::Quit(-1);
+        Quit(-1);
     }
 
     // let's see if we have to trigger the GC

diff --git a/src/memory/GenerationalHeap.cpp b/src/memory/GenerationalHeap.cpp
@@ -6,7 +6,6 @@
 #include <vector>
 
 #include "../vm/Print.h"
-#include "../vm/Universe.h"
 #include "../vmobjects/AbstractObject.h"
 #include "../vmobjects/ObjectFormats.h"
 #include "../vmobjects/VMObjectBase.h"
@@ -39,7 +38,7 @@ AbstractVMObject* GenerationalHeap::AllocateNurseryObject(size_t size) {
     if ((size_t)nextFreePosition > nursery_end) {
         ErrorPrint("\nFailed to allocate " + to_string(size) +
                    " Bytes in nursery.\n");
-        GetUniverse()->Quit(-1);
+        Quit(-1);
     }
     // let's see if we have to trigger the GC
     if (nextFreePosition > collectionLimit) {
@@ -52,7 +51,7 @@ AbstractVMObject* GenerationalHeap::AllocateMatureObject(size_t size) {
     AbstractVMObject* newObject = (AbstractVMObject*)malloc(size);
     if (newObject == nullptr) {
         ErrorPrint("\nFailed to allocate " + to_string(size) + " Bytes.\n");
-        GetUniverse()->Quit(-1);
+        Quit(-1);
     }
     allocatedObjects->push_back(newObject);
     matureObjectsSize += size;

diff --git a/src/memory/MarkSweepHeap.cpp b/src/memory/MarkSweepHeap.cpp
@@ -7,7 +7,6 @@
 
 #include "../memory/Heap.h"
 #include "../vm/Print.h"
-#include "../vm/Universe.h"
 #include "../vmobjects/AbstractObject.h"
 #include "MarkSweepCollector.h"
 
@@ -23,7 +22,7 @@ AbstractVMObject* MarkSweepHeap::AllocateObject(size_t size) {
     AbstractVMObject* newObject = (AbstractVMObject*)malloc(size);
     if (newObject == nullptr) {
         ErrorPrint("\nFailed to allocate " + to_string(size) + " Bytes.\n");
-        Universe::Quit(-1);
+        Quit(-1);
     }
     spcAlloc += size;
     memset((void*)newObject, 0, size);

diff --git a/src/primitives/Double.cpp b/src/primitives/Double.cpp
@@ -34,6 +34,7 @@
 
 #include "../primitivesCore/PrimitiveContainer.h"
 #include "../vm/Globals.h"
+#include "../vm/Print.h"
 #include "../vm/Universe.h"
 #include "../vmobjects/ObjectFormats.h"
 #include "../vmobjects/VMDouble.h"
@@ -56,8 +57,7 @@ double coerceDouble(vm_oop_t x) {
     } else if (cl == load_ptr(integerClass)) {
         return (double)static_cast<VMInteger*>(x)->GetEmbeddedInteger();
     } else {
-        GetUniverse()->ErrorExit(
-            "Attempt to apply Double operation to non-number.");
+        ErrorExit("Attempt to apply Double operation to non-number.");
     }
 
     return 0.0f;

diff --git a/src/primitives/System.cpp b/src/primitives/System.cpp
@@ -91,7 +91,7 @@ static vm_oop_t sysLoad_(vm_oop_t, vm_oop_t rightObj) {
 
 static vm_oop_t sysExit_(vm_oop_t, vm_oop_t err) {
     long err_no = INT_VAL(err);
-    GetUniverse()->Quit(err_no);
+    Quit(err_no);
 }
 
 static vm_oop_t sysPrintString_(vm_oop_t leftObj, vm_oop_t rightObj) {

diff --git a/src/vm/Print.cpp b/src/vm/Print.cpp
@@ -1,20 +1,50 @@
 #include "Print.h"
 
+#include <cstdlib>
 #include <iostream>
 #include <mutex>
+#include <string>
 
 #include "../misc/defs.h"
+#include "LogAllocation.h"
+#include "Universe.h"
 
 using namespace std;
 
 static mutex output_mutex;
 
-void Print(StdString str) {
+void Print(std::string str) {
     lock_guard<mutex> lock(output_mutex);
     cout << str << flush;
 }
 
-void ErrorPrint(StdString str) {
+void ErrorPrint(std::string str) {
     lock_guard<mutex> lock(output_mutex);
     cerr << str << flush;
 }
+
+void Print(const char* str) {
+    lock_guard<mutex> lock(output_mutex);
+    cout << str << flush;
+}
+
+void ErrorPrint(const char* str) {
+    lock_guard<mutex> lock(output_mutex);
+    cerr << str << flush;
+}
+
+__attribute__((noreturn)) __attribute__((noinline)) void ErrorExit(
+    const char* err) {
+    ErrorPrint("Runtime error: " + StdString(err) + "\n");
+    Quit(ERR_FAIL);
+}
+
+__attribute__((noreturn)) __attribute__((noinline)) void Quit(long err) {
+    ErrorPrint("Time spent in GC: [" +
+               to_string(Timer::GCTimer->GetTotalTime()) + "] msec\n");
+
+    Universe::Shutdown();
+
+    OutputAllocationLogFile();
+    exit((int)err);
+}
diff --git a/src/vm/Print.h b/src/vm/Print.h
@@ -1,6 +1,11 @@
 #pragma once
 
-#include "../misc/defs.h"
+#include <string>
 
-void Print(StdString str);
-void ErrorPrint(StdString str);
+void Print(std::string str);
+void Print(const char* str);
+void ErrorPrint(std::string str);
+void ErrorPrint(const char* str);
+
+__attribute__((noreturn)) __attribute__((noinline)) void Quit(long);
+__attribute__((noreturn)) __attribute__((noinline)) void ErrorExit(const char*);
diff --git a/src/vm/Shell.cpp b/src/vm/Shell.cpp
@@ -37,6 +37,7 @@
 #include "../vmobjects/VMInvokable.h"
 #include "../vmobjects/VMMethod.h"
 #include "Globals.h"
+#include "Print.h"
 #include "Symbols.h"
 #include "Universe.h"
 
@@ -65,7 +66,7 @@ void Shell::Start(Interpreter* interp) {
 #define QUIT_CMD_L 11 + 1
 
     if (bootstrapMethod == nullptr) {
-        GetUniverse()->ErrorExit("Shell needs bootstrap method!");
+        ErrorExit("Shell needs bootstrap method!");
     }
     // the statement to evaluate
     char inbuf[INPUT_MAX_SIZE];

diff --git a/src/vm/Universe.cpp b/src/vm/Universe.cpp
@@ -92,7 +92,7 @@ void Universe::BasicInit() {
     theUniverse = new Universe();
 }
 
-__attribute__((noreturn)) void Universe::Quit(long err) {
+void Universe::Shutdown() {
     ErrorPrint("Time spent in GC: [" +
                to_string(Timer::GCTimer->GetTotalTime()) + "] msec\n");
 #ifdef GENERATE_INTEGER_HISTOGRAM
@@ -136,18 +136,9 @@ __attribute__((noreturn)) void Universe::Quit(long err) {
     }
 #endif
 
-    OutputAllocationLogFile();
-
     if (theUniverse) {
         delete (theUniverse);
     }
-
-    exit((int)err);
-}
-
-__attribute__((noreturn)) void Universe::ErrorExit(const char* err) {
-    ErrorPrint("Runtime error: " + StdString(err) + "\n");
-    Quit(ERR_FAIL);
 }
 
 vector<StdString> Universe::handleArguments(long argc, char** argv) {
@@ -619,7 +610,7 @@ void Universe::LoadSystemClass(VMClass* systemClass) {
 
     if (!result) {
         ErrorPrint("Can't load system class: " + s + "\n");
-        Universe::Quit(ERR_FAIL);
+        Quit(ERR_FAIL);
     }
 
     if (result->HasPrimitives() || result->GetClass()->HasPrimitives()) {