Complete static field accesses; clear TODO's

This adds the fast path for static field accesses.  Miscellaneous
changes include adding dummy helper for filled array allocation,
elimination of redunant arg1/r1 load, deleting TODO's that I'm not
going to do, and altered compiler_test to use System.out.println
instead of System.out.printf.

Change-Id: I560aa6093c98f1d288d64f346c8fd9ab7f180994

6 files changed, 256 insertions, 228 deletions

diff --git a/src/compiler/codegen/arm/MethodCodegenDriver.cc b/src/compiler/codegen/arm/MethodCodegenDriver.cc
index eb9296bb05..81eb24cbcb 100644
--- a/src/compiler/codegen/arm/MethodCodegenDriver.cc
+++ b/src/compiler/codegen/arm/MethodCodegenDriver.cc
@@ -60,10 +60,8 @@ static void genFilledNewArray(CompilationUnit* cUnit, MIR* mir, bool isRange)
         typeId = dInsn->vC;
     }
     oatFlushAllRegs(cUnit);    /* Everything to home location */
-    // TODO: Alloc variant that checks types (see header comment) */
-    UNIMPLEMENTED(WARNING) << "Need AllocFromCode variant w/ extra checks";
     loadWordDisp(cUnit, rSELF,
-                 OFFSETOF_MEMBER(Thread, pAllocFromCode), rLR);
+                 OFFSETOF_MEMBER(Thread, pCheckAndAllocFromCode), rLR);
     loadCurrMethodDirect(cUnit, r1);              // arg1 <- Method*
     loadConstant(cUnit, r0, typeId);              // arg0 <- type_id
     loadConstant(cUnit, r2, elems);               // arg2 <- count
@@ -139,17 +137,14 @@ static void genFilledNewArray(CompilationUnit* cUnit, MIR* mir, bool isRange)
 
 static void genSput(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc)
 {
-    bool slow_path = true;
     bool isObject = ((mir->dalvikInsn.opcode == OP_SPUT_OBJECT) ||
                      (mir->dalvikInsn.opcode == OP_SPUT_OBJECT_VOLATILE));
-    UNIMPLEMENTED(WARNING) << "Implement sput fast path";
-    int funcOffset;
-    if (slow_path) {
-        if (isObject) {
-            funcOffset = OFFSETOF_MEMBER(Thread, pSetObjStatic);
-        } else {
-            funcOffset = OFFSETOF_MEMBER(Thread, pSet32Static);
-        }
+    int fieldIdx = mir->dalvikInsn.vB;
+    Field* field = cUnit->method->GetDexCacheResolvedFields()->Get(fieldIdx);
+    if (field == NULL) {
+        // Slow path
+        int funcOffset = isObject ? OFFSETOF_MEMBER(Thread, pSetObjStatic)
+                                  : OFFSETOF_MEMBER(Thread, pSet32Static);
         oatFlushAllRegs(cUnit);
         loadWordDisp(cUnit, rSELF, funcOffset, rLR);
         loadConstant(cUnit, r0, mir->dalvikInsn.vB);
@@ -158,153 +153,171 @@ static void genSput(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc)
         opReg(cUnit, kOpBlx, rLR);
         oatClobberCallRegs(cUnit);
     } else {
-        UNIMPLEMENTED(FATAL) << "Must update for new world";
-#if 0
-    int valOffset = OFFSETOF_MEMBER(StaticField, value);
-    int tReg = oatAllocTemp(cUnit);
-    int objHead;
-    bool isVolatile;
-    bool isSputObject;
-    const Method *method = (mir->OptimizationFlags & MIR_CALLEE) ?
-        mir->meta.calleeMethod : cUnit->method;
-    void* fieldPtr = (void*)
-      (method->clazz->pDvmDex->pResFields[mir->dalvikInsn.vB]);
-    Opcode opcode = mir->dalvikInsn.opcode;
-
-    if (fieldPtr == NULL) {
-        // FIXME: need to handle this case for oat();
-        UNIMPLEMENTED(FATAL);
-    }
-
+        // fast path
+        int fieldOffset = field->GetOffset().Int32Value();
+        art::ClassLinker* class_linker = art::Runtime::Current()->
+            GetClassLinker();
+        const art::DexFile& dex_file = class_linker->
+            FindDexFile(field->GetDeclaringClass()->GetDexCache());
+        const art::DexFile::FieldId& field_id = dex_file.GetFieldId(fieldIdx);
+        int typeIdx = field_id.class_idx_;
+        // Using fixed register to sync with slow path
+        int rMethod = r1;
+        oatLockTemp(cUnit, rMethod);
+        loadCurrMethodDirect(cUnit, rMethod);
+        int rBase = r0;
+        oatLockTemp(cUnit, rBase);
+        loadWordDisp(cUnit, rMethod,
+            Method::DexCacheInitializedStaticStorageOffset().Int32Value(),
+            rBase);
+        loadWordDisp(cUnit, rBase, art::Array::DataOffset().Int32Value() +
+                      sizeof(int32_t*)* typeIdx, rBase);
+        // TUNING: fast path should fall through
+        ArmLIR* branchOver = genCmpImmBranch(cUnit, kArmCondNe, rBase, 0);
+        loadWordDisp(cUnit, rSELF,
+                     OFFSETOF_MEMBER(Thread, pInitializeStaticStorage), rLR);
+        loadConstant(cUnit, r0, typeIdx);
+        opReg(cUnit, kOpBlx, rLR);
+        ArmLIR* skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
+        skipTarget->defMask = ENCODE_ALL;
+        branchOver->generic.target = (LIR*)skipTarget;
+        rlSrc = oatGetSrc(cUnit, mir, 0);
+        rlSrc = loadValue(cUnit, rlSrc, kAnyReg);
+        storeWordDisp(cUnit, rBase, fieldOffset, rlSrc.lowReg);
 #if ANDROID_SMP != 0
-    isVolatile = (opcode == OP_SPUT_VOLATILE) ||
-                 (opcode == OP_SPUT_VOLATILE_JUMBO) ||
-                 (opcode == OP_SPUT_OBJECT_VOLATILE) ||
-                 (opcode == OP_SPUT_OBJECT_VOLATILE_JUMBO);
-    assert(isVolatile == artIsVolatileField((Field *) fieldPtr));
-#else
-    isVolatile = artIsVolatileField((Field *) fieldPtr);
-#endif
-
-    isSputObject = (opcode == OP_SPUT_OBJECT) ||
-                   (opcode == OP_SPUT_OBJECT_VOLATILE);
-
-    rlSrc = oatGetSrc(cUnit, mir, 0);
-    rlSrc = loadValue(cUnit, rlSrc, kAnyReg);
-    loadConstant(cUnit, tReg,  (int) fieldPtr);
-    if (isSputObject) {
-        objHead = oatAllocTemp(cUnit);
-        loadWordDisp(cUnit, tReg, OFFSETOF_MEMBER(Field, clazz), objHead);
-    }
-    storeWordDisp(cUnit, tReg, valOffset ,rlSrc.lowReg);
-    oatFreeTemp(cUnit, tReg);
-    if (isVolatile) {
-        oatGenMemBarrier(cUnit, kSY);
-    }
-    if (isSputObject) {
-        /* NOTE: marking card based sfield->clazz */
-        markGCCard(cUnit, rlSrc.lowReg, objHead);
-        oatFreeTemp(cUnit, objHead);
-    }
+        if (field->IsVolatile()) {
+            oatGenMemBarrier(cUnit, kSY);
+        }
 #endif
+        if (isObject) {
+            markGCCard(cUnit, rlSrc.lowReg, rBase);
+        }
+        oatFreeTemp(cUnit, rBase);
     }
 }
 
 static void genSputWide(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc)
 {
-    bool slow_path = true;
-    UNIMPLEMENTED(WARNING) << "Implement sput-wide fast path";
-    int funcOffset;
-    if (slow_path) {
-        funcOffset = OFFSETOF_MEMBER(Thread, pSet64Static);
+    int fieldIdx = mir->dalvikInsn.vB;
+    Field* field = cUnit->method->GetDexCacheResolvedFields()->Get(fieldIdx);
+    if (field == NULL) {
         oatFlushAllRegs(cUnit);
-        loadWordDisp(cUnit, rSELF, funcOffset, rLR);
+        loadWordDisp(cUnit, rSELF, OFFSETOF_MEMBER(Thread, pSet64Static), rLR);
         loadConstant(cUnit, r0, mir->dalvikInsn.vB);
         loadCurrMethodDirect(cUnit, r1);
         loadValueDirectWideFixed(cUnit, rlSrc, r2, r3);
         opReg(cUnit, kOpBlx, rLR);
         oatClobberCallRegs(cUnit);
     } else {
-        UNIMPLEMENTED(FATAL) << "Must update for new world";
-#if 0
-    int tReg = oatAllocTemp(cUnit);
-    int valOffset = OFFSETOF_MEMBER(StaticField, value);
-    const Method *method = (mir->OptimizationFlags & MIR_CALLEE) ?
-        mir->meta.calleeMethod : cUnit->method;
-    void* fieldPtr = (void*)
-      (method->clazz->pDvmDex->pResFields[mir->dalvikInsn.vB]);
-
-    if (fieldPtr == NULL) {
-        // FIXME: need to handle this case for oat();
-        UNIMPLEMENTED(FATAL);
-    }
-
-    rlSrc = oatGetSrcWide(cUnit, mir, 0, 1);
-    rlSrc = loadValueWide(cUnit, rlSrc, kAnyReg);
-    loadConstant(cUnit, tReg,  (int) fieldPtr + valOffset);
-
-    storePair(cUnit, tReg, rlSrc.lowReg, rlSrc.highReg);
+        // fast path
+        int fieldOffset = field->GetOffset().Int32Value();
+        art::ClassLinker* class_linker = art::Runtime::Current()->
+            GetClassLinker();
+        const art::DexFile& dex_file = class_linker->
+            FindDexFile(field->GetDeclaringClass()->GetDexCache());
+        const art::DexFile::FieldId& field_id = dex_file.GetFieldId(fieldIdx);
+        int typeIdx = field_id.class_idx_;
+        // Using fixed register to sync with slow path
+        int rMethod = r1;
+        oatLockTemp(cUnit, rMethod);
+        loadCurrMethodDirect(cUnit, r1);
+        int rBase = r0;
+        oatLockTemp(cUnit, rBase);
+        loadWordDisp(cUnit, rMethod,
+            Method::DexCacheInitializedStaticStorageOffset().Int32Value(),
+            rBase);
+        loadWordDisp(cUnit, rBase, art::Array::DataOffset().Int32Value() +
+                      sizeof(int32_t*)* typeIdx, rBase);
+        // TUNING: fast path should fall through
+        ArmLIR* branchOver = genCmpImmBranch(cUnit, kArmCondNe, rBase, 0);
+        loadWordDisp(cUnit, rSELF,
+                     OFFSETOF_MEMBER(Thread, pInitializeStaticStorage), rLR);
+        loadConstant(cUnit, r0, typeIdx);
+        opReg(cUnit, kOpBlx, rLR);
+        ArmLIR* skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
+        skipTarget->defMask = ENCODE_ALL;
+        branchOver->generic.target = (LIR*)skipTarget;
+        rlSrc = oatGetSrcWide(cUnit, mir, 0, 1);
+        rlSrc = loadValueWide(cUnit, rlSrc, kAnyReg);
+        storeBaseDispWide(cUnit, rBase, fieldOffset, rlSrc.lowReg,
+                          rlSrc.highReg);
+#if ANDROID_SMP != 0
+        if (field->IsVolatile()) {
+            oatGenMemBarrier(cUnit, kSY);
+        }
 #endif
+        oatFreeTemp(cUnit, rBase);
     }
 }
 
 
-
 static void genSgetWide(CompilationUnit* cUnit, MIR* mir,
                  RegLocation rlResult, RegLocation rlDest)
 {
-    bool slow_path = true;
-    UNIMPLEMENTED(WARNING) << "Implement sget-wide fast path";
-    int funcOffset;
-    if (slow_path) {
-        funcOffset = OFFSETOF_MEMBER(Thread, pGet64Static);
+    int fieldIdx = mir->dalvikInsn.vB;
+    Field* field = cUnit->method->GetDexCacheResolvedFields()->Get(fieldIdx);
+    if (field == NULL) {
         oatFlushAllRegs(cUnit);
-        loadWordDisp(cUnit, rSELF, funcOffset, rLR);
+        loadWordDisp(cUnit, rSELF, OFFSETOF_MEMBER(Thread, pGet64Static), rLR);
         loadConstant(cUnit, r0, mir->dalvikInsn.vB);
         loadCurrMethodDirect(cUnit, r1);
         opReg(cUnit, kOpBlx, rLR);
         RegLocation rlResult = oatGetReturnWide(cUnit);
         storeValueWide(cUnit, rlDest, rlResult);
     } else {
-        UNIMPLEMENTED(FATAL) << "Must update for new world";
-#if 0
-    int valOffset = OFFSETOF_MEMBER(StaticField, value);
-    const Method *method = (mir->OptimizationFlags & MIR_CALLEE) ?
-        mir->meta.calleeMethod : cUnit->method;
-    void* fieldPtr = (void*)
-      (method->clazz->pDvmDex->pResFields[mir->dalvikInsn.vB]);
-
-    if (fieldPtr == NULL) {
-        // FIXME: need to handle this case for oat();
-        UNIMPLEMENTED(FATAL);
-    }
-
-    int tReg = oatAllocTemp(cUnit);
-    rlDest = oatGetDestWide(cUnit, mir, 0, 1);
-    rlResult = oatEvalLoc(cUnit, rlDest, kAnyReg, true);
-    loadConstant(cUnit, tReg,  (int) fieldPtr + valOffset);
-
-    loadPair(cUnit, tReg, rlResult.lowReg, rlResult.highReg);
-
-    storeValueWide(cUnit, rlDest, rlResult);
+        // Fast path
+        int fieldOffset = field->GetOffset().Int32Value();
+        art::ClassLinker* class_linker = art::Runtime::Current()->
+            GetClassLinker();
+        const art::DexFile& dex_file = class_linker->
+            FindDexFile(field->GetDeclaringClass()->GetDexCache());
+        const art::DexFile::FieldId& field_id = dex_file.GetFieldId(fieldIdx);
+        int typeIdx = field_id.class_idx_;
+        // Using fixed register to sync with slow path
+        int rMethod = r1;
+        oatLockTemp(cUnit, rMethod);
+        loadCurrMethodDirect(cUnit, rMethod);
+        int rBase = r0;
+        oatLockTemp(cUnit, rBase);
+        loadWordDisp(cUnit, rMethod,
+            Method::DexCacheInitializedStaticStorageOffset().Int32Value(),
+            rBase);
+        loadWordDisp(cUnit, rBase, art::Array::DataOffset().Int32Value() +
+                      sizeof(int32_t*)* typeIdx, rBase);
+        // TUNING: fast path should fall through
+        ArmLIR* branchOver = genCmpImmBranch(cUnit, kArmCondNe, rBase, 0);
+        loadWordDisp(cUnit, rSELF,
+                     OFFSETOF_MEMBER(Thread, pInitializeStaticStorage), rLR);
+        loadConstant(cUnit, r0, typeIdx);
+        opReg(cUnit, kOpBlx, rLR);
+        ArmLIR* skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
+        skipTarget->defMask = ENCODE_ALL;
+        branchOver->generic.target = (LIR*)skipTarget;
+        rlDest = oatGetDestWide(cUnit, mir, 0, 1);
+        RegLocation rlResult = oatEvalLoc(cUnit, rlDest, kAnyReg, true);
+#if ANDROID_SMP != 0
+        if (isVolatile) {
+            oatGenMemBarrier(cUnit, kSY);
+        }
 #endif
+        loadBaseDispWide(cUnit, NULL, rBase, fieldOffset, rlResult.lowReg,
+                         rlResult.highReg, INVALID_SREG);
+        oatFreeTemp(cUnit, rBase);
+        storeValueWide(cUnit, rlDest, rlResult);
     }
 }
 
 static void genSget(CompilationUnit* cUnit, MIR* mir,
              RegLocation rlResult, RegLocation rlDest)
 {
-    bool slow_path = true;
+    int fieldIdx = mir->dalvikInsn.vB;
+    Field* field = cUnit->method->GetDexCacheResolvedFields()->Get(fieldIdx);
     bool isObject = ((mir->dalvikInsn.opcode == OP_SGET_OBJECT) ||
                      (mir->dalvikInsn.opcode == OP_SGET_OBJECT_VOLATILE));
-    UNIMPLEMENTED(WARNING) << "Implement sget fast path";
-    int funcOffset;
-    if (slow_path) {
-        if (isObject) {
-            funcOffset = OFFSETOF_MEMBER(Thread, pGetObjStatic);
-        } else {
-            funcOffset = OFFSETOF_MEMBER(Thread, pGet32Static);
-        }
+    if (field == NULL) {
+        // Slow path
+        int funcOffset = isObject ? OFFSETOF_MEMBER(Thread, pGetObjStatic)
+                                  : OFFSETOF_MEMBER(Thread, pGet32Static);
         oatFlushAllRegs(cUnit);
         loadWordDisp(cUnit, rSELF, funcOffset, rLR);
         loadConstant(cUnit, r0, mir->dalvikInsn.vB);
@@ -313,48 +326,44 @@ static void genSget(CompilationUnit* cUnit, MIR* mir,
         RegLocation rlResult = oatGetReturn(cUnit);
         storeValue(cUnit, rlDest, rlResult);
     } else {
-        UNIMPLEMENTED(FATAL) << "Must update for new world";
-#if 0
-    int valOffset = OFFSETOF_MEMBER(StaticField, value);
-    int tReg = oatAllocTemp(cUnit);
-    bool isVolatile;
-    const Method *method = cUnit->method;
-    void* fieldPtr = (void*)
-      (method->clazz->pDvmDex->pResFields[mir->dalvikInsn.vB]);
-
-    if (fieldPtr == NULL) {
-        // FIXME: need to handle this case for oat();
-        UNIMPLEMENTED(FATAL);
-    }
-
-    /*
-     * On SMP systems, Dalvik opcodes found to be referencing
-     * volatile fields are rewritten to their _VOLATILE variant.
-     * However, this does not happen on non-SMP systems. The compiler
-     * still needs to know about volatility to avoid unsafe
-     * optimizations so we determine volatility based on either
-     * the opcode or the field access flags.
-     */
+        // Fast path
+        int fieldOffset = field->GetOffset().Int32Value();
+        art::ClassLinker* class_linker = art::Runtime::Current()->
+            GetClassLinker();
+        const art::DexFile& dex_file = class_linker->
+            FindDexFile(field->GetDeclaringClass()->GetDexCache());
+        const art::DexFile::FieldId& field_id = dex_file.GetFieldId(fieldIdx);
+        int typeIdx = field_id.class_idx_;
+        // Using fixed register to sync with slow path
+        int rMethod = r1;
+        oatLockTemp(cUnit, rMethod);
+        loadCurrMethodDirect(cUnit, rMethod);
+        int rBase = r0;
+        oatLockTemp(cUnit, rBase);
+        loadWordDisp(cUnit, rMethod,
+            Method::DexCacheInitializedStaticStorageOffset().Int32Value(),
+            rBase);
+        loadWordDisp(cUnit, rBase, art::Array::DataOffset().Int32Value() +
+                      sizeof(int32_t*)* typeIdx, rBase);
+        // TUNING: fast path should fall through
+        ArmLIR* branchOver = genCmpImmBranch(cUnit, kArmCondNe, rBase, 0);
+        loadWordDisp(cUnit, rSELF,
+                     OFFSETOF_MEMBER(Thread, pInitializeStaticStorage), rLR);
+        loadConstant(cUnit, r0, typeIdx);
+        opReg(cUnit, kOpBlx, rLR);
+        ArmLIR* skipTarget = newLIR0(cUnit, kArmPseudoTargetLabel);
+        skipTarget->defMask = ENCODE_ALL;
+        branchOver->generic.target = (LIR*)skipTarget;
+        rlDest = oatGetDest(cUnit, mir, 0);
+        rlResult = oatEvalLoc(cUnit, rlDest, kAnyReg, true);
 #if ANDROID_SMP != 0
-    Opcode opcode = mir->dalvikInsn.opcode;
-    isVolatile = (opcode == OP_SGET_VOLATILE) ||
-                 (opcode == OP_SGET_OBJECT_VOLATILE);
-    assert(isVolatile == artIsVolatileField((Field *) fieldPtr));
-#else
-    isVolatile = artIsVolatileField((Field *) fieldPtr);
-#endif
-
-    rlDest = oatGetDest(cUnit, mir, 0);
-    rlResult = oatEvalLoc(cUnit, rlDest, kAnyReg, true);
-    loadConstant(cUnit, tReg,  (int) fieldPtr + valOffset);
-
-    if (isVolatile) {
-        oatGenMemBarrier(cUnit, kSY);
-    }
-    loadWordDisp(cUnit, tReg, 0, rlResult.lowReg);
-
-    storeValue(cUnit, rlDest, rlResult);
+        if (isVolatile) {
+            oatGenMemBarrier(cUnit, kSY);
+        }
 #endif
+        loadWordDisp(cUnit, rBase, fieldOffset, rlResult.lowReg);
+        oatFreeTemp(cUnit, rBase);
+        storeValue(cUnit, rlDest, rlResult);
     }
 }
 
@@ -694,7 +703,8 @@ static int nextSuperCallInsnSP(CompilationUnit* cUnit, MIR* mir,
 static int genDalvikArgsNoRange(CompilationUnit* cUnit, MIR* mir,
                                 DecodedInstruction* dInsn, int callState,
                                 ArmLIR** pcrLabel, bool isRange,
-                                NextCallInsn nextCallInsn, ArmLIR* rollback)
+                                NextCallInsn nextCallInsn, ArmLIR* rollback,
+                                bool skipThis)
 {
     RegLocation rlArg;
     int registerArgs[3];
@@ -726,7 +736,7 @@ static int genDalvikArgsNoRange(CompilationUnit* cUnit, MIR* mir,
     }
 
     /* Load register arguments r1..r3 */
-    for (unsigned int i = 0; i < 3; i++) {
+    for (unsigned int i = skipThis ? 1 : 0; i < 3; i++) {
         if (i < dInsn->vA)
             registerArgs[i] = (isRange) ? dInsn->vC + i : i;
         else
@@ -761,7 +771,7 @@ static int genDalvikArgsNoRange(CompilationUnit* cUnit, MIR* mir,
 static int genDalvikArgsRange(CompilationUnit* cUnit, MIR* mir,
                               DecodedInstruction* dInsn, int callState,
                               ArmLIR** pcrLabel, NextCallInsn nextCallInsn,
-                              ArmLIR* rollback)
+                              ArmLIR* rollback, bool skipThis)
 {
     int firstArg = dInsn->vC;
     int numArgs = dInsn->vA;
@@ -769,7 +779,7 @@ static int genDalvikArgsRange(CompilationUnit* cUnit, MIR* mir,
     // If we can treat it as non-range (Jumbo ops will use range form)
     if (numArgs <= 5)
         return genDalvikArgsNoRange(cUnit, mir, dInsn, callState, pcrLabel,
-                                    true, nextCallInsn, rollback);
+                                    true, nextCallInsn, rollback, skipThis);
     /*
      * Make sure range list doesn't span the break between in normal
      * Dalvik vRegs and the ins.
@@ -789,7 +799,6 @@ static int genDalvikArgsRange(CompilationUnit* cUnit, MIR* mir,
     // Scan the rest of the args - if in physReg flush to memory
     for (int i = 4; i < numArgs; i++) {
         RegLocation loc = oatGetSrc(cUnit, mir, i);
-        //TODO: generic loc flushing routine
         if (loc.wide) {
             loc = oatUpdateLocWide(cUnit, loc);
             if (loc.location == kLocPhysReg) {  // TUNING: if dirty?
@@ -831,7 +840,7 @@ static int genDalvikArgsRange(CompilationUnit* cUnit, MIR* mir,
     }
 
     // Handle the 1st 3 in r1, r2 & r3
-    for (unsigned int i = 0; i < dInsn->vA && i < 3; i++) {
+    for (unsigned int i = skipThis? 1 : 0; i < dInsn->vA && i < 3; i++) {
         RegLocation loc = oatGetSrc(cUnit, mir, firstArg + i);
         loadValueDirectFixed(cUnit, loc, r1 + i);
         callState = nextCallInsn(cUnit, mir, dInsn, callState, rollback);
@@ -856,10 +865,10 @@ static void genInvokeStaticDirect(CompilationUnit* cUnit, MIR* mir,
 
     if (range) {
         callState = genDalvikArgsRange(cUnit, mir, dInsn, callState, pNullCk,
-                                       nextCallInsn, NULL);
+                                       nextCallInsn, NULL, false);
     } else {
         callState = genDalvikArgsNoRange(cUnit, mir, dInsn, callState, pNullCk,
-                                         false, nextCallInsn, NULL);
+                                         false, nextCallInsn, NULL, false);
     }
     // Finish up any of the call sequence not interleaved in arg loading
     while (callState >= 0) {
@@ -881,10 +890,11 @@ static void genInvokeInterface(CompilationUnit* cUnit, MIR* mir)
     callState = nextInterfaceCallInsn(cUnit, mir, dInsn, callState, NULL);
     if (mir->dalvikInsn.opcode == OP_INVOKE_INTERFACE)
         callState = genDalvikArgsNoRange(cUnit, mir, dInsn, callState, &nullCk,
-                                         false, nextInterfaceCallInsn, NULL);
+                                         false, nextInterfaceCallInsn, NULL,
+                                         true);
     else
         callState = genDalvikArgsRange(cUnit, mir, dInsn, callState, &nullCk,
-                                       nextInterfaceCallInsn, NULL);
+                                       nextInterfaceCallInsn, NULL, true);
     // Finish up any of the call sequence not interleaved in arg loading
     while (callState >= 0) {
         callState = nextInterfaceCallInsn(cUnit, mir, dInsn, callState, NULL);
@@ -925,13 +935,12 @@ static void genInvokeSuper(CompilationUnit* cUnit, MIR* mir)
         rollback = newLIR0(cUnit, kArmPseudoTargetLabel);
         rollback->defMask = -1;
     }
-// TODO - redundantly loading arg0/r1 ("this")
     if (mir->dalvikInsn.opcode == OP_INVOKE_SUPER)
         callState = genDalvikArgsNoRange(cUnit, mir, dInsn, callState, &nullCk,
-                                         false, nextCallInsn, rollback);
+                                         false, nextCallInsn, rollback, true);
     else
         callState = genDalvikArgsRange(cUnit, mir, dInsn, callState, &nullCk,
-                                       nextCallInsn, rollback);
+                                       nextCallInsn, rollback, true);
     // Finish up any of the call sequence not interleaved in arg loading
     while (callState >= 0) {
         callState = nextSuperCallInsn(cUnit, mir, dInsn, callState, NULL);
@@ -960,13 +969,12 @@ static void genInvokeVirtual(CompilationUnit* cUnit, MIR* mir)
         nextCallInsn = nextVCallInsn;
         rollback = NULL;
     }
-    // TODO - redundantly loading arg0/r1 ("this")
     if (mir->dalvikInsn.opcode == OP_INVOKE_VIRTUAL)
         callState = genDalvikArgsNoRange(cUnit, mir, dInsn, callState, &nullCk,
-                                         false, nextCallInsn, rollback);
+                                         false, nextCallInsn, rollback, true);
     else
         callState = genDalvikArgsRange(cUnit, mir, dInsn, callState, &nullCk,
-                                       nextCallInsn, rollback);
+                                       nextCallInsn, rollback, true);
     // Finish up any of the call sequence not interleaved in arg loading
     while (callState >= 0) {
         callState = nextCallInsn(cUnit, mir, dInsn, callState, rollback);
@@ -974,7 +982,6 @@ static void genInvokeVirtual(CompilationUnit* cUnit, MIR* mir)
     newLIR1(cUnit, kThumbBlxR, rLR);
 }
 
-// TODO: break out the case handlers.  Might make it easier to support x86
 static bool compileDalvikInstruction(CompilationUnit* cUnit, MIR* mir,
                                      BasicBlock* bb, ArmLIR* labelList)
 {
@@ -1058,7 +1065,7 @@ static bool compileDalvikInstruction(CompilationUnit* cUnit, MIR* mir,
              * formation in storeValueWide.
              */
             assert(retLocWide.lowReg == r0);
-            assert(retLocWide.lowReg == r1);
+            assert(retLocWide.highReg == r1);
             oatLockTemp(cUnit, retLocWide.lowReg);
             oatLockTemp(cUnit, retLocWide.highReg);
             storeValueWide(cUnit, rlDest, retLocWide);
diff --git a/src/compiler/codegen/arm/Thumb2/Gen.cc b/src/compiler/codegen/arm/Thumb2/Gen.cc
index c4a1fed931..d0b0683a5b 100644
--- a/src/compiler/codegen/arm/Thumb2/Gen.cc
+++ b/src/compiler/codegen/arm/Thumb2/Gen.cc
@@ -596,8 +596,7 @@ static void genInstanceof(CompilationUnit* cUnit, MIR* mir, RegLocation rlDest,
     Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()->
         GetResolvedType(mir->dalvikInsn.vC);
     if (classPtr == NULL) {
-        /* Shouldn't happen */
-        LOG(FATAL) << "Unexpected null class pointer";
+        UNIMPLEMENTED(FATAL) << "Handle null class pointer";
     }
     oatFlushAllRegs(cUnit);   /* Everything to home location */
     loadValueDirectFixed(cUnit, rlSrc, r0);  /* Ref */
@@ -631,8 +630,7 @@ static void genCheckCast(CompilationUnit* cUnit, MIR* mir, RegLocation rlSrc)
     Class* classPtr = cUnit->method->GetDeclaringClass()->GetDexCache()->
         GetResolvedType(mir->dalvikInsn.vB);
     if (classPtr == NULL) {
-        /* Shouldn't happen with our current model */
-        LOG(FATAL) << "Unexpected null class pointer";
+        UNIMPLEMENTED(FATAL) << "Unimplemented null class pointer";
     }
     oatFlushAllRegs(cUnit);   /* Everything to home location */
     loadConstant(cUnit, r1, (int) classPtr );
diff --git a/src/object.h b/src/object.h
index ca84dac39a..4edae14a2a 100644
--- a/src/object.h
+++ b/src/object.h
@@ -563,11 +563,11 @@ class Field : public AccessibleObject {
   static void SetClass(Class* java_lang_reflect_Field);
   static void ResetClass();
 
- private:
   bool IsVolatile() const {
     return (GetAccessFlags() & kAccVolatile) != 0;
   }
 
+ private:
   // private implementation of field access using raw data
   uint32_t Get32(const Object* object) const;
   void Set32(Object* object, uint32_t new_value) const;
@@ -758,6 +758,11 @@ class Method : public AccessibleObject {
     return OFFSET_OF_OBJECT_MEMBER(Method, dex_cache_strings_);
   }
 
+  static MemberOffset DexCacheInitializedStaticStorageOffset() {
+    return OFFSET_OF_OBJECT_MEMBER(Method,
+        dex_cache_initialized_static_storage_);
+  }
+
   ObjectArray<Class>* GetDexCacheResolvedTypes() const;
   void SetDexCacheResolvedTypes(ObjectArray<Class>* new_dex_cache_types);
 
diff --git a/src/thread.cc b/src/thread.cc
index 31786c9970..a617fb48ec 100644
--- a/src/thread.cc
+++ b/src/thread.cc
@@ -106,6 +106,18 @@ static void ResolveMethodFromCode(Method* method, uint32_t method_idx) {
      */
 }
 
+// TODO: placeholder.  Helper function to alloc array for OP_FILLED_NEW_ARRAY
+static Array* CheckAndAllocFromCode(uint32_t type_index, Method* method,
+                                    int32_t component_count)
+{
+    /*
+     * Just a wrapper around Array::AllocFromCode() that additionally
+     * throws a runtime exception "bad Filled array req" for 'D' and 'J'.
+     */
+    UNIMPLEMENTED(WARNING) << "Need check that not 'D' or 'J'";
+    return Array::AllocFromCode(type_index, method, component_count);
+}
+
 void Thread::InitFunctionPointers() {
 #if defined(__arm__)
   pShlLong = art_shl_long;
@@ -137,6 +149,7 @@ void Thread::InitFunctionPointers() {
   pInvokeInterfaceTrampoline = art_invoke_interface_trampoline;
 #endif
   pAllocFromCode = Array::AllocFromCode;
+  pCheckAndAllocFromCode = CheckAndAllocFromCode;
   pAllocObjectFromCode = Class::AllocObjectFromCode;
   pMemcpy = memcpy;
   pHandleFillArrayDataFromCode = HandleFillArrayDataFromCode;
@@ -150,6 +163,7 @@ void Thread::InitFunctionPointers() {
   pThrowException = ThrowException;
   pInitializeTypeFromCode = InitializeTypeFromCode;
   pResolveMethodFromCode = ResolveMethodFromCode;
+  pInitializeStaticStorage = ClassLinker::InitializeStaticStorageFromCode;
   pDebugMe = DebugMe;
 #if 0
 bool (Thread::*pUnlockObject)(Thread*, Object*);
diff --git a/src/thread.h b/src/thread.h
index fc523a3d97..070933b389 100644
--- a/src/thread.h
+++ b/src/thread.h
@@ -29,6 +29,8 @@ class Thread;
 class ThreadList;
 class Throwable;
 class StackTraceElement;
+class StaticStorageBase;
+
 template<class T> class ObjectArray;
 template<class T> class PrimitiveArray;
 typedef PrimitiveArray<int32_t> IntArray;
@@ -204,6 +206,7 @@ class Thread {
   long long (*pLmul)(long long, long long);
   long long (*pLdivmod)(long long, long long);
   Array* (*pAllocFromCode)(uint32_t, Method*, int32_t);
+  Array* (*pCheckAndAllocFromCode)(uint32_t, Method*, int32_t);
   Object* (*pAllocObjectFromCode)(uint32_t, Method*);
   uint32_t (*pGet32Static)(uint32_t, const Method*);
   void (*pSet32Static)(uint32_t, const Method*, uint32_t);
@@ -221,6 +224,7 @@ class Thread {
   Class* (*pInitializeTypeFromCode)(uint32_t, Method*);
   void (*pResolveMethodFromCode)(Method*, uint32_t);
   void (*pInvokeInterfaceTrampoline)(void*, void*, void*, void*);
+  StaticStorageBase* (*pInitializeStaticStorage)(uint32_t, const Method*);
 
   class StackVisitor {
    public:
diff --git a/test/IntMath/IntMath.java b/test/IntMath/IntMath.java
index ba7d575927..1be8a98b0f 100644
--- a/test/IntMath/IntMath.java
+++ b/test/IntMath/IntMath.java
@@ -712,107 +712,107 @@ class IntMath extends IntMathBase {
     public static void main(String[] args) {
         int res = unopTest(38);
         if (res == 37) {
-            System.out.printf("unopTest PASSED\n");
+            System.out.println("unopTest PASSED");
         } else {
-            System.out.printf("unopTest FAILED: %d\n", res);
+            System.out.println("unopTest FAILED: " + res);
         }
         res = shiftTest1();
         if (res == 0) {
-            System.out.printf("shiftTest1 PASSED\n");
+            System.out.println("shiftTest1 PASSED");
         } else {
-            System.out.printf("shiftTest1 FAILED: %d\n", res);
+            System.out.println("shiftTest1 FAILED: " + res);
         }
         res = shiftTest2();
         if (res == 0) {
-            System.out.printf("shiftTest2 PASSED\n");
+            System.out.println("shiftTest2 PASSED");
         } else {
-            System.out.printf("shiftTest2 FAILED: %d\n", res);
+            System.out.println("shiftTest2 FAILED: " + res);
         }
         res = unsignedShiftTest();
         if (res == 0) {
-            System.out.printf("unsignedShiftTest PASSED\n");
+            System.out.println("unsignedShiftTest PASSED");
         } else {
-            System.out.printf("unsignedShiftTest FAILED: %d\n", res);
+            System.out.println("unsignedShiftTest FAILED: " + res);
         }
         res = convTest();
         if (res == 0) {
-            System.out.printf("convTest PASSED\n");
+            System.out.println("convTest PASSED");
         } else {
-            System.out.printf("convTest FAILED: %d\n", res);
+            System.out.println("convTest FAILED: " + res);
         }
         res = charSubTest();
         if (res == 0) {
-            System.out.printf("charSubTest PASSED\n");
+            System.out.println("charSubTest PASSED");
         } else {
-            System.out.printf("charSubTest FAILED: %d\n", res);
+            System.out.println("charSubTest FAILED: " + res);
         }
         res = intOperTest(70000, -3);
         if (res == 0) {
-            System.out.printf("intOperTest PASSED\n");
+            System.out.println("intOperTest PASSED");
         } else {
-            System.out.printf("intOperTest FAILED: %d\n", res);
+            System.out.println("intOperTest FAILED: " + res);
         }
         res = longOperTest(70000000000L, -3L);
         if (res == 0) {
-            System.out.printf("longOperTest PASSED\n");
+            System.out.println("longOperTest PASSED");
         } else {
-            System.out.printf("longOperTest FAILED: %d\n", res);
+            System.out.println("longOperTest FAILED: " + res);
         }
         long lres = longShiftTest(0xd5aa96deff00aa01L, 16);
         if (lres == 0x96deff00aa010000L) {
-            System.out.printf("longShiftTest PASSED\n");
+            System.out.println("longShiftTest PASSED");
         } else {
-            System.out.printf("longShiftTest FAILED: %d\n", res);
+            System.out.println("longShiftTest FAILED: " + res);
         }
 
         res = switchTest(1);
         if (res == 1234) {
-            System.out.printf("switchTest PASSED\n");
+            System.out.println("switchTest PASSED");
         } else {
-            System.out.printf("switchTest FAILED: %d\n", res);
+            System.out.println("switchTest FAILED: " + res);
         }
 
         res = testIntCompare(-5, 4, 4, 0);
         if (res == 1111) {
-            System.out.printf("testIntCompare PASSED\n");
+            System.out.println("testIntCompare PASSED");
         } else {
-            System.out.printf("testIntCompare FAILED: %d\n", res);
+            System.out.println("testIntCompare FAILED: " + res);
         }
 
         res = testLongCompare(-5L, -4294967287L, 4L, 8L);
         if (res == 2222) {
-            System.out.printf("testLongCompare PASSED\n");
+            System.out.println("testLongCompare PASSED");
         } else {
-            System.out.printf("testLongCompare FAILED: %d\n", res);
+            System.out.println("testLongCompare FAILED: " + res);
         }
 
         res = testFloatCompare(-5.0f, 4.0f, 4.0f, (1.0f/0.0f) / (1.0f/0.0f));
         if (res == 3333) {
-            System.out.printf("testFloatCompare PASSED\n");
+            System.out.println("testFloatCompare PASSED");
         } else {
-            System.out.printf("testFloatCompare FAILED: %d\n", res);
+            System.out.println("testFloatCompare FAILED: " + res);
         }
 
         res = testDoubleCompare(-5.0, 4.0, 4.0, (1.0/0.0) / (1.0/0.0));
         if (res == 4444) {
-            System.out.printf("testDoubleCompare PASSED\n");
+            System.out.println("testDoubleCompare PASSED");
         } else {
-            System.out.printf("testDoubleCompare FAILED: %d\n", res);
+            System.out.println("testDoubleCompare FAILED: " + res);
         }
 
         res = fibonacci(10);
         if (res == 55) {
-            System.out.printf("fibonacci PASSED\n");
+            System.out.println("fibonacci PASSED");
         } else {
-            System.out.printf("fibonacci FAILED: %d\n", res);
+            System.out.println("fibonacci FAILED: " + res);
         }
 
         /*
           res = throwAndCatch();
           if (res == 0) {
-          System.out.printf("throwAndCatch PASSED\n");
+          System.out.println("throwAndCatch PASSED");
           } else {
-          System.out.printf("throwAndCatch FAILED: %d\n", res);
+          System.out.println("throwAndCatch FAILED: " + res);
           }
         */
 
@@ -820,44 +820,44 @@ class IntMath extends IntMathBase {
                        (short)11, 12, (char)13, 14, 15, (byte)-16, true, 18,
                        19, 20L, 21L, 22, 23, 24, 25, 26);
         if (res == -1) {
-            System.out.printf("manyArgs PASSED\n");
+            System.out.println("manyArgs PASSED");
         } else {
-            System.out.printf("manyArgs FAILED: %d\n", res);
+            System.out.println("manyArgs FAILED: " + res);
         }
 
         res = staticCall(3);
         if (res == 6) {
-            System.out.printf("virtualCall PASSED\n");
+            System.out.println("virtualCall PASSED");
         } else {
-            System.out.printf("virtualCall FAILED: %d\n", res);
+            System.out.println("virtualCall FAILED: " + res);
         }
 
         res = testIGetPut(111);
         if (res == 333) {
-            System.out.printf("testGetPut PASSED\n");
+            System.out.println("testGetPut PASSED");
         } else {
-            System.out.printf("testGetPut FAILED: %d\n", res);
+            System.out.println("testGetPut FAILED: " + res);
         }
 
         res = staticFieldTest(404);
         if (res == 1404) {
-            System.out.printf("staticFieldTest PASSED\n");
+            System.out.println("staticFieldTest PASSED");
         } else {
-            System.out.printf("staticFieldTest FAILED: %d\n", res);
+            System.out.println("staticFieldTest FAILED: " + res);
         }
 
         res = catchBlock(1000);
         if (res == 1579) {
-            System.out.printf("catchBlock PASSED\n");
+            System.out.println("catchBlock PASSED");
         } else {
-            System.out.printf("catchBlock FAILED: %d\n", res);
+            System.out.println("catchBlock FAILED: " + res);
         }
 
         res = superTest(4141);
         if (res == 4175) {
-            System.out.printf("superTest PASSED\n");
+            System.out.println("superTest PASSED");
         } else {
-            System.out.printf("superTest FAILED: %d\n", res);
+            System.out.println("superTest FAILED: " + res);
         }
     }
 }