diff --git a/platforms/app/main.c b/platforms/app/main.c
index 7f9f264..172ede8 100644
--- a/platforms/app/main.c
+++ b/platforms/app/main.c
@@ -92,10 +92,19 @@ M3Result link_all  (IM3Module module)
     return res;
 }
 
+const char* modname_from_fn(const char* fn)
+{
+	const char* off = strrchr(fn, '/');
+	if (off) return off+1;
+	off = strrchr(fn, '\\');
+	if (off) return off+1;
+	return fn;
+}
 
 M3Result repl_load  (const char* fn)
 {
     M3Result result = m3Err_none;
+    IM3Module module = NULL;
 
     u8* wasm = NULL;
     u32 fsize = 0;
@@ -129,13 +138,14 @@ M3Result repl_load  (const char* fn)
     fclose (f);
     f = NULL;
 
-    IM3Module module;
     result = m3_ParseModule (env, &module, wasm, fsize);
     if (result) goto on_error;
 
     result = m3_LoadModule (runtime, module);
     if (result) goto on_error;
 
+    m3_SetModuleName(module, modname_from_fn(fn));
+
     result = link_all (module);
     if (result) goto on_error;
 
@@ -144,7 +154,9 @@ M3Result repl_load  (const char* fn)
     }
 
     return result;
+
 on_error:
+    m3_FreeModule(module);
     if (wasm) free(wasm);
     if (f) fclose(f);
 
diff --git a/source/m3_api_defs.h b/source/m3_api_defs.h
index 9d52bac..221d962 100644
--- a/source/m3_api_defs.h
+++ b/source/m3_api_defs.h
@@ -32,16 +32,20 @@
 #define m3ApiSuccess()             { return m3Err_none; }
 
 # if defined(M3_BIG_ENDIAN)
+#  define m3ApiReadMem8(ptr)         (* (u8 *)(ptr))
 #  define m3ApiReadMem16(ptr)        __builtin_bswap16((* (u16 *)(ptr)))
 #  define m3ApiReadMem32(ptr)        __builtin_bswap32((* (u32 *)(ptr)))
 #  define m3ApiReadMem64(ptr)        __builtin_bswap64((* (u64 *)(ptr)))
+#  define m3ApiWriteMem8(ptr, val)   { * (u8  *)(ptr)  = (val); }
 #  define m3ApiWriteMem16(ptr, val)  { * (u16 *)(ptr) = __builtin_bswap16((val)); }
 #  define m3ApiWriteMem32(ptr, val)  { * (u32 *)(ptr) = __builtin_bswap32((val)); }
 #  define m3ApiWriteMem64(ptr, val)  { * (u64 *)(ptr) = __builtin_bswap64((val)); }
 # else
+#  define m3ApiReadMem8(ptr)         (* (u8 *)(ptr))
 #  define m3ApiReadMem16(ptr)        (* (u16 *)(ptr))
 #  define m3ApiReadMem32(ptr)        (* (u32 *)(ptr))
 #  define m3ApiReadMem64(ptr)        (* (u64 *)(ptr))
+#  define m3ApiWriteMem8(ptr, val)   { * (u8  *)(ptr) = (val); }
 #  define m3ApiWriteMem16(ptr, val)  { * (u16 *)(ptr) = (val); }
 #  define m3ApiWriteMem32(ptr, val)  { * (u32 *)(ptr) = (val); }
 #  define m3ApiWriteMem64(ptr, val)  { * (u64 *)(ptr) = (val); }
diff --git a/source/m3_compile.c b/source/m3_compile.c
index 2f7f9e7..6912fb7 100644
--- a/source/m3_compile.c
+++ b/source/m3_compile.c
@@ -362,7 +362,7 @@ u16  GetMaxUsedSlotPlusOne  (IM3Compilation o)
 
         o->slotMaxAllocatedIndexPlusOne--;
     }
-    
+
 #   ifdef DEBUG
         u16 maxSlot = o->slotMaxAllocatedIndexPlusOne;
         while (maxSlot < d_m3MaxFunctionSlots)
@@ -465,8 +465,8 @@ M3Result  Push  (IM3Compilation o, u8 i_type, u16 i_slot)
             u32 regSelect = IsFpRegisterSlotAlias (i_slot);
             AllocateRegister (o, regSelect, stackIndex);
         }
-        
-        if (d_m3LogWasmStack) dump_type_stack (o);
+
+		if (d_m3LogWasmStack) dump_type_stack (o);
     }
     else result = m3Err_functionStackOverflow;
 
@@ -1062,7 +1062,8 @@ _               (CopyStackIndexToSlot (o, returnSlot, stackTop--));
 _               (Pop (o));
         }
     }
-    _catch: return result;
+
+	_catch: return result;
 }
 
 
@@ -1482,7 +1483,7 @@ _       (GetBlockScope (o, & scope, target));
 
         // create a ContinueLoop operation on a fresh page
 _       (AcquireCompilationCodePage (o, & continueOpPage));
-        
+
         pc_t startPC = GetPagePC (continueOpPage);
         IM3CodePage savedPage = o->page;
         o->page = continueOpPage;
@@ -1514,7 +1515,7 @@ _                   (EmitPatchingBranch (o, scope));
                 }
             }
         }
-        
+
         ReleaseCompilationCodePage (o);     // FIX: continueOpPage can get lost if thrown
         o->page = savedPage;
 
@@ -2367,7 +2368,7 @@ const M3OpInfo c_operations [] =
 
     d_m3DebugOp (CopySlot_32),      d_m3DebugOp (PreserveCopySlot_32), d_m3DebugOp (If_s),          d_m3DebugOp (BranchIfPrologue_s),
     d_m3DebugOp (CopySlot_64),      d_m3DebugOp (PreserveCopySlot_64), d_m3DebugOp (If_r),          d_m3DebugOp (BranchIfPrologue_r),
-    
+
     d_m3DebugOp (Select_i32_rss),   d_m3DebugOp (Select_i32_srs),   d_m3DebugOp (Select_i32_ssr),   d_m3DebugOp (Select_i32_sss),
     d_m3DebugOp (Select_i64_rss),   d_m3DebugOp (Select_i64_srs),   d_m3DebugOp (Select_i64_ssr),   d_m3DebugOp (Select_i64_sss),
 
@@ -2744,7 +2745,7 @@ _   (CompileLocals (o));
 _   (ReserveConstants (o));
 
     // start tracking the max stack used (Push() also updates this value) so that op_Entry can precisely detect stack overflow
-    o->function->maxStackSlots = o->slotMaxAllocatedIndexPlusOne = o->slotFirstDynamicIndex;
+    o->maxStackSlots = o->slotMaxAllocatedIndexPlusOne = o->slotFirstDynamicIndex;
 
     o->block.blockStackIndex = o->stackFirstDynamicIndex = o->stackIndex;                           m3log (compile, "start stack index: %d",
                                                                                                           (u32) o->stackFirstDynamicIndex);
@@ -2757,6 +2758,7 @@ _   (CompileBlockStatements (o));
     _throwif(m3Err_wasmMalformed, o->previousOpcode != c_waOp_end);
 
     io_function->compiled = pc;
+    io_function->maxStackSlots = o->maxStackSlots;
 
     u16 numConstantSlots = o->slotMaxConstIndex - o->slotFirstConstIndex;                           m3log (compile, "unique constant slots: %d; unused slots: %d",
                                                                                                            numConstantSlots, o->slotFirstDynamicIndex - o->slotMaxConstIndex);
diff --git a/source/m3_compile.h b/source/m3_compile.h
index c97cda4..cca423f 100644
--- a/source/m3_compile.h
+++ b/source/m3_compile.h
@@ -91,6 +91,8 @@ typedef struct
     u16                 slotFirstLocalIndex;
     u16                 slotFirstDynamicIndex;      // numArgs + numLocals + numReservedConstants. the first mutable slot available to the compiler.
 
+    u16                 maxStackSlots;
+
     m3slot_t            constants                   [d_m3MaxConstantTableSize];
 
     // 'wasmStack' holds slot locations
diff --git a/source/m3_core.c b/source/m3_core.c
index b1e3cea..0d9ffed 100644
--- a/source/m3_core.c
+++ b/source/m3_core.c
@@ -117,7 +117,7 @@ void *  m3_Malloc  (size_t i_size)
 void  m3_FreeImpl  (void * io_ptr)
 {
 //    if (io_ptr) printf("== free %p\n", io_ptr);
-    free ((void*)io_ptr);
+    free (io_ptr);
 }
 
 void *  m3_Realloc  (void * i_ptr, size_t i_newSize, size_t i_oldSize)
diff --git a/source/m3_env.c b/source/m3_env.c
index d707265..ea1d565 100644
--- a/source/m3_env.c
+++ b/source/m3_env.c
@@ -286,6 +286,10 @@ M3Result  EvaluateExpression  (IM3Module i_module, void * o_expressed, u8 i_type
         pc_t m3code = GetPagePC (o->page);
         result = CompileBlock (o, ftype, c_waOp_block);
 
+        if (not result && o->maxStackSlots >= runtime.numStackSlots) {
+            result = m3Err_trapStackOverflow;
+        }
+
         if (not result)
         {
             m3ret_t r = Call (m3code, stack, NULL, d_m3OpDefaultArgs);
@@ -542,28 +546,32 @@ M3Result  m3_LoadModule  (IM3Runtime io_runtime, IM3Module io_module)
 {
     M3Result result = m3Err_none;
 
-    if (not io_module->runtime)
-    {
-        io_module->runtime = io_runtime;
-        M3Memory * memory = & io_runtime->memory;
+    if (UNLIKELY(io_module->runtime)) {
+        return m3Err_moduleAlreadyLinked;
+    }
 
-_       (InitMemory (io_runtime, io_module));
-_       (InitGlobals (io_module));
-_       (InitDataSegments (memory, io_module));
-_       (InitElements (io_module));
+    io_module->runtime = io_runtime;
+    M3Memory * memory = & io_runtime->memory;
 
-        io_module->next = io_runtime->modules;
-        io_runtime->modules = io_module;
+_   (InitMemory (io_runtime, io_module));
+_   (InitGlobals (io_module));
+_   (InitDataSegments (memory, io_module));
+_   (InitElements (io_module));
 
-        // Start func might use imported functions, which are not liked here yet,
-        // so it will be called before a function call is attempted (in m3_FindFuSnction)
-    }
-    else result = m3Err_moduleAlreadyLinked;
+    // Start func might use imported functions, which are not liked here yet,
+    // so it will be called before a function call is attempted (in m3_FindFunction)
 
-    if (result)
-        io_module->runtime = NULL;
+#ifdef DEBUG
+    Module_GenerateNames(io_module);
+#endif
 
-    _catch: return result;
+    io_module->next = io_runtime->modules;
+    io_runtime->modules = io_module;
+    return result; // ok
+
+_catch:
+    io_module->runtime = NULL;
+    return result;
 }
 
 IM3Global  m3_FindGlobal  (IM3Module               io_module,
diff --git a/source/m3_env.h b/source/m3_env.h
index 5f5c482..33d0e00 100644
--- a/source/m3_env.h
+++ b/source/m3_env.h
@@ -123,6 +123,8 @@ M3Result                    Module_AddGlobal            (IM3Module io_module, IM
 M3Result                    Module_AddFunction          (IM3Module io_module, u32 i_typeIndex, IM3ImportInfo i_importInfo /* can be null */);
 IM3Function                 Module_GetFunction          (IM3Module i_module, u32 i_functionIndex);
 
+void                        Module_GenerateNames        (IM3Module i_module);
+
 void                        FreeImportInfo              (M3ImportInfo * i_info);
 
 //---------------------------------------------------------------------------------------------------------------------------------
diff --git a/source/m3_exec.h b/source/m3_exec.h
index 2df684b..54b5a91 100644
--- a/source/m3_exec.h
+++ b/source/m3_exec.h
@@ -744,13 +744,7 @@ d_m3Op  (Entry)
         }
 
 #if d_m3EnableStrace >= 2
-        u16 numNames = 0;
-        cstr_t *names = GetFunctionNames(function, &numNames);
-        if (numNames) {
-            d_m3TracePrint("%s %s {", names[0], SPrintFunctionArgList (function, _sp));
-        } else {
-            d_m3TracePrint("$%d %s {", function->index, SPrintFunctionArgList (function, _sp));
-        }
+        d_m3TracePrint("%s %s {", m3_GetFunctionName(function), SPrintFunctionArgList (function, _sp));
         trace_rt->callDepth++;
 #endif
 
diff --git a/source/m3_function.h b/source/m3_function.h
index d038af9..7173b8b 100644
--- a/source/m3_function.h
+++ b/source/m3_function.h
@@ -63,7 +63,7 @@ typedef struct M3Function
     u32                     hits;
 #endif
 
-# if d_m3EnableStrace >= 2 || d_m3LogCompile
+# if d_m3LogCompile
     u32                     index;
 # endif
 
diff --git a/source/m3_info.c b/source/m3_info.c
index 4174172..4805c83 100644
--- a/source/m3_info.c
+++ b/source/m3_info.c
@@ -250,24 +250,20 @@ d_m3Decoder  (BranchTable)
 {
     u32 slot = fetch (u32);
 
-    sprintf (o_string, "slot: %" PRIu32 "; targets: ", slot);
+    o_string += sprintf (o_string, "slot: %" PRIu32 "; targets: ", slot);
 
 //    IM3Function function = fetch2 (IM3Function);
 
     i32 targets = fetch (i32);
 
-    char str [1000];
-
     for (i32 i = 0; i < targets; ++i)
     {
         pc_t addr = fetch (pc_t);
-        sprintf (str, "%" PRIi32 "=%p, ", i, addr);
-        strcat (o_string, str);
+        o_string += sprintf (o_string, "%" PRIi32 "=%p, ", i, addr);
     }
 
     pc_t addr = fetch (pc_t);
-    sprintf (str, "def=%p ", addr);
-    strcat (o_string, str);
+    sprintf (o_string, "def=%p ", addr);
 }
 
 
@@ -315,7 +311,7 @@ void  dump_code_page  (IM3CodePage i_codePage, pc_t i_startPC)
 
                 if (i.info)
                 {
-                    char infoString [1000] = { 0 };
+                    char infoString [8*1024] = { 0 };
 
                     DecodeOperation (infoString, i.opcode, op, i.info, & pc);
 
@@ -424,7 +420,7 @@ void  dump_type_stack  (IM3Compilation o)
 //        d_m3Assert (regAllocated [r] == 0);         // reg allocation & stack out of sync
     
     u16 maxSlot = GetMaxUsedSlotPlusOne (o);
-    
+
     if (maxSlot > o->slotFirstDynamicIndex)
     {
         d_m3Log (stack, "                      -");
@@ -445,7 +441,7 @@ void  dump_type_stack  (IM3Compilation o)
         {
             printf ("%3d|", o->m3Slots [i]);
         }
-        
+
         printf ("\n");
     }
     d_m3Log(stack, "\n");
diff --git a/source/m3_module.c b/source/m3_module.c
index 1566fb0..0286409 100644
--- a/source/m3_module.c
+++ b/source/m3_module.c
@@ -37,9 +37,6 @@ void  m3_FreeModule  (IM3Module i_module)
         for (u32 i = 0; i < i_module->numGlobals; ++i)
         {
             m3_Free (i_module->globals[i].name);
-        }
-        for (u32 i = 0; i < i_module->numGlobals; ++i)
-        {
             FreeImportInfo(&(i_module->globals[i].import));
         }
         m3_Free (i_module->globals);
@@ -84,15 +81,15 @@ _try {
 
     IM3Function func = Module_GetFunction (io_module, index);
     func->funcType = ft;
-#if d_m3EnableStrace >= 2 || d_m3LogCompile
+#	if d_m3LogCompile
     func->index = index;
-#endif
+#	endif
 
     if (i_importInfo and func->numNames == 0)
     {
         func->import = * i_importInfo;
-        func->numNames = 1;
         func->names[0] = i_importInfo->fieldUtf8;
+        func->numNames = 1;
     }
 
     m3log (module, "   added function: %3d; sig: %d", index, i_typeIndex);
@@ -101,6 +98,32 @@ _try {
     return result;
 }
 
+void  Module_GenerateNames  (IM3Module i_module)
+{
+    for (u32 i = 0; i < i_module->numFunctions; ++i)
+    {
+        IM3Function func = & i_module->functions [i];
+
+        if (func->numNames == 0)
+        {
+            char* buff = m3_AllocArray(char, 16);
+            snprintf(buff, 16, "$func%d", i);
+            func->names[0] = buff;
+            func->numNames = 1;
+        }
+    }
+    for (u32 i = 0; i < i_module->numGlobals; ++i)
+    {
+        IM3Global global = & i_module->globals [i];
+
+        if (global->name == NULL)
+        {
+            char* buff = m3_AllocArray(char, 16);
+            snprintf(buff, 16, "$global%d", i);
+            global->name = buff;
+        }
+    }
+}
 
 IM3Function  Module_GetFunction  (IM3Module i_module, u32 i_functionIndex)
 {
@@ -119,11 +142,16 @@ IM3Function  Module_GetFunction  (IM3Module i_module, u32 i_functionIndex)
 const char*  m3_GetModuleName  (IM3Module i_module)
 {
     if (!i_module || !i_module->name)
-        return "<unknown>";
+        return ".unnamed";
 
     return i_module->name;
 }
 
+void  m3_SetModuleName  (IM3Module i_module, const char* name)
+{
+    if (i_module) i_module->name = name;
+}
+
 IM3Runtime  m3_GetModuleRuntime  (IM3Module i_module)
 {
     return i_module ? i_module->runtime : NULL;
diff --git a/source/m3_parse.c b/source/m3_parse.c
index 81f55f5..4ea7bb1 100644
--- a/source/m3_parse.c
+++ b/source/m3_parse.c
@@ -127,6 +127,8 @@ _   (ReadLEB_u32 (& numFunctions, & i_bytes, i_end));
 
     _throwif("too many functions", numFunctions > d_m3MaxSaneFunctionsCount);
 
+    // TODO: prealloc functions
+
     for (u32 i = 0; i < numFunctions; ++i)
     {
         u32 funcTypeIndex;
@@ -236,18 +238,19 @@ _       (ReadLEB_u32 (& index, & i_bytes, i_end));
         if (exportKind == d_externalKind_function)
         {
             _throwif(m3Err_wasmMalformed, index >= io_module->numFunctions);
-            u16 numNames = io_module->functions [index].numNames;
-            if (numNames < d_m3MaxDuplicateFunctionImpl)
+            IM3Function func = &(io_module->functions [index]);
+            if (func->numNames < d_m3MaxDuplicateFunctionImpl)
             {
-                io_module->functions [index].numNames++;
-                io_module->functions [index].names[numNames] = utf8;
+                func->names[func->numNames++] = utf8;
                 utf8 = NULL; // ownership transferred to M3Function
             }
         }
         else if (exportKind == d_externalKind_global)
         {
             _throwif(m3Err_wasmMalformed, index >= io_module->numGlobals);
-            io_module->globals[index].name = utf8;
+            IM3Global global = &(io_module->globals [index]);
+            m3_Free (global->name);
+            global->name = utf8;
             utf8 = NULL; // ownership transferred to M3Global
         }
 
@@ -506,10 +509,11 @@ _               (Read_utf8 (& name, & i_bytes, i_end));
 
                 if (index < io_module->numFunctions)
                 {
-                    if (io_module->functions [index].numNames == 0)
+                    IM3Function func = &(io_module->functions [index]);
+                    if (func->numNames == 0)
                     {
-                        io_module->functions [index].numNames = 1;
-                        io_module->functions [index].names[0] = name;        m3log (parse, "    naming function%5d:  %s", index, name);
+                        func->names[0] = name;        m3log (parse, "    naming function%5d:  %s", index, name);
+                        func->numNames = 1;
                         name = NULL; // transfer ownership
                     }
 //                          else m3log (parse, "prenamed: %s", io_module->functions [index].name);
diff --git a/source/wasm3.h b/source/wasm3.h
index caf7cbd..32fb3e1 100644
--- a/source/wasm3.h
+++ b/source/wasm3.h
@@ -246,6 +246,7 @@ d_m3ErrorConst  (trapStackOverflow,             "[trap] stack overflow")
                                                      const void *           i_userdata);
 
     const char*         m3_GetModuleName            (IM3Module i_module);
+    void                m3_SetModuleName            (IM3Module i_module, const char* name);
     IM3Runtime          m3_GetModuleRuntime         (IM3Module i_module);
 
 //-------------------------------------------------------------------------------------------------------------------------------
diff --git a/test/run-wasi-test.py b/test/run-wasi-test.py
index 02cdfcf..f789e79 100755
--- a/test/run-wasi-test.py
+++ b/test/run-wasi-test.py
@@ -65,7 +65,7 @@ commands_full = [
     "args":           ["16", "64"],
     "expect_sha1":    "d85df3561eb15f6f0e6f20d5640e8e1306222c6d"
   }, {
-    "skip":           True,  # Fails on Windows-uvwasi, on CI only (CNR locally)
+    "skip":           True,  # TODO
     "name":           "mal",
     "wasm":           "./wasi/mal/mal.wasm",
     "args":           ["./wasi/mal/test-fib.mal", "16"],
@@ -127,6 +127,7 @@ commands_fast = [
     "args":           ["4", "32"],
     "expect_sha1":    "ea05d85998b2f453b588ef76a1256215bf9b851c"
   }, {
+    "skip":           True,  # TODO
     "name":           "mal",
     "wasm":           "./wasi/mal/mal.wasm",
     "args":           ["./wasi/mal/test-fib.mal", "16"],
@@ -176,9 +177,9 @@ for cmd in commands:
         stats.timeout += 1
         fail("Timeout")
         continue
-    except subprocess.CalledProcessError:
+    except subprocess.CalledProcessError as e:
         stats.crashed += 1
-        fail("Crashed")
+        fail(f"Exited with error code {e.returncode}")
         continue
 
     if "expect_sha1" in cmd: