diff --git a/src/common/hdl/hdlruntime.ts b/src/common/hdl/hdlruntime.ts
index e512e3b3..4d448b39 100644
--- a/src/common/hdl/hdlruntime.ts
+++ b/src/common/hdl/hdlruntime.ts
@@ -24,6 +24,7 @@ export class HDLModuleJS implements HDLModuleRunner {
     curconsts: {};
     constused: number;
     specfuncs: VerilatorUnit[] = [];
+    getFileData = null;
     
     constructor(mod: HDLModuleDef, constpool: HDLModuleDef) {
         this.mod = mod;
@@ -377,7 +378,7 @@ export class HDLModuleJS implements HDLModuleRunner {
         barr.reverse(); // reverse it
         var strfn = byteArrayToString(barr); // convert to string
         // parse hex/binary file
-        var strdata = this.getFile(strfn) as string;
+        var strdata = this.getFileData(strfn) as string;
         if (strdata == null) throw Error("Could not $readmem '" + strfn + "'");
         var data = strdata.split('\n').filter(s => s !== '').map(s => parseInt(s, ishex ? 16 : 2));
         console.log('$readmem', ishex, strfn, data.length);
@@ -388,11 +389,6 @@ export class HDLModuleJS implements HDLModuleRunner {
             memp[i] = data[i];
     }
 
-    getFile(path: string) : string {
-        // TODO: override
-        return null;
-    }
-
     isStopped() { return this.stopped; }
     isFinished() { return this.finished; }
     
diff --git a/src/common/hdl/hdltypes.ts b/src/common/hdl/hdltypes.ts
index 05a1faa7..597dd9aa 100644
--- a/src/common/hdl/hdltypes.ts
+++ b/src/common/hdl/hdltypes.ts
@@ -11,6 +11,7 @@ export interface HDLModuleRunner {
     saveState() : {};
     loadState(state: {}) : void;
     dispose() : void;
+    getFileData : (filename : string) => string|Uint8Array;
 }
 
 export interface HDLModuleTrace {
diff --git a/src/common/hdl/hdlwasm.ts b/src/common/hdl/hdlwasm.ts
index 39b230a7..10fbfb72 100644
--- a/src/common/hdl/hdlwasm.ts
+++ b/src/common/hdl/hdlwasm.ts
@@ -29,10 +29,12 @@ const TRACEBUF = "$$tbuf";
 
 export class HDLError extends Error {
     obj: any;
+    $loc: HDLSourceLocation;
     constructor(obj: any, msg: string) {
         super(msg);
         Object.setPrototypeOf(this, HDLError.prototype);
         this.obj = obj;
+        if (obj && obj.$loc) this.$loc = obj.$loc;
         if (obj) console.log(obj);
     }
 }
@@ -57,6 +59,10 @@ function getDataTypeSize(dt: HDLDataType) : number {
     }
 }
 
+function isReferenceType(dt: HDLDataType) : boolean {
+    return getDataTypeSize(dt) > 8;
+}
+
 function getArrayElementSizeFromType(dtype: HDLDataType) : number {
     if (isArrayType(dtype)) {
         return getArrayElementSizeFromType(dtype.subtype);
@@ -188,6 +194,7 @@ export class HDLModuleWASM implements HDLModuleRunner {
     traceStartOffset: number;
     traceEndOffset: number;
     trace: any;
+    getFileData = null;
 
     constructor(moddef: HDLModuleDef, constpool: HDLModuleDef) {
         this.hdlmod = moddef;
@@ -345,7 +352,12 @@ export class HDLModuleWASM implements HDLModuleRunner {
              this.pushScope(fscope);
              block.exprs.forEach((e) => {
                  if (e && isVarDecl(e)) {
-                     fscope.addVar(e);
+                     // TODO: make local reference types, instead of promoting local arrays to global
+                     if (isReferenceType(e.dtype)) {
+                         this.globals.addVar(e);
+                     } else {
+                         fscope.addVar(e);
+                     }
                  }
              })
              // create function body
@@ -464,26 +476,44 @@ export class HDLModuleWASM implements HDLModuleRunner {
     }
 
     private addImportedFunctions() {
-        this.bmod.addFunctionImport("$display", "builtins", "$display", binaryen.createType([binaryen.i32]), binaryen.none);
-        this.bmod.addFunctionImport("$finish", "builtins", "$finish", binaryen.createType([binaryen.i32]), binaryen.none);
-        this.bmod.addFunctionImport("$stop", "builtins", "$stop", binaryen.createType([binaryen.i32]), binaryen.none);
-        this.bmod.addFunctionImport("$time", "builtins", "$time", binaryen.createType([binaryen.i32]), binaryen.i64);
-        this.bmod.addFunctionImport("$rand", "builtins", "$rand", binaryen.createType([binaryen.i32]), binaryen.i32);
+        this.bmod.addFunctionImport("$finish_0", "builtins", "$finish", binaryen.createType([binaryen.i32]), binaryen.none);
+        this.bmod.addFunctionImport("$stop_0", "builtins", "$stop", binaryen.createType([binaryen.i32]), binaryen.none);
+        this.bmod.addFunctionImport("$time_0", "builtins", "$time", binaryen.createType([binaryen.i32]), binaryen.i64);
+        this.bmod.addFunctionImport("$rand_0", "builtins", "$rand", binaryen.createType([binaryen.i32]), binaryen.i32);
+        this.bmod.addFunctionImport("$readmem_2", "builtins", "$readmem", binaryen.createType([binaryen.i32, binaryen.i32, binaryen.i32]), binaryen.none);
     }
 
     private getImportObject() : {} {
         var n = 0;
         return {
             builtins: {
-                $display: (o) => { if (++n < 100) console.log('...',o); }, // TODO
                 $finish: (o) => { console.log('... Finished @', o); this.finished = true; },
                 $stop: (o) => { console.log('... Stopped @', o); this.stopped = true; },
                 $time: (o) => BigInt(new Date().getTime()), // TODO: timescale
                 $rand: (o) => (Math.random() * (65536 * 65536)) | 0,
+                $readmem: (o,a,b) => this.$readmem(a, b)
             }
         }
     }
 
+    private $readmem(p_filename, p_rom) {
+        var fn = '';
+        for (var i=0; i<255; i++) {
+            var charCode = this.data8[p_filename + i];
+            if (charCode == 0) break;
+            fn = String.fromCharCode(charCode) + fn;
+        }
+        var filedata = this.getFileData && this.getFileData(fn);
+        if (filedata == null) throw new HDLError(fn, `no file "${fn}" for $readmem`);
+        if (typeof filedata !== 'string') throw new HDLError(fn, `file "${fn}" must be lines of hex or binary values`);
+        var ishex = !fn.endsWith('.binary'); // TODO: hex should be attribute in xml
+        var data = filedata.split('\n').filter(s => s !== '').map(s => parseInt(s, ishex ? 16 : 2));
+        for (var i=0; i<data.length; i++) {
+            this.data8[p_rom + i] = data[i];
+        }
+        return 0;
+    }
+
     private addCopyTraceRecFunction() {
         const m = this.bmod;
         const o_TRACERECLEN = this.globals.lookup(TRACERECLEN).offset;
@@ -673,13 +703,18 @@ export class HDLModuleWASM implements HDLModuleRunner {
 
     funccall2wasm(e: HDLFuncCall, opts?:Options) : number {
         var args = [this.dataptr()];
+        for (var arg of e.args) {
+            args.push(this.e2w(arg));
+        }
+        var internal = e.funcname;
         if (e.funcname.startsWith('$')) {
             if ((e.funcname == '$stop' || e.funcname == '$finish') && e.$loc) {
                 args = [this.bmod.i32.const(e.$loc.line)]; // line # of source code
             }
+            internal += '_' + (args.length - 1);
         }
         var ret = this.funcResult(e.funcname);
-        return this.bmod.call(e.funcname, args, ret);
+        return this.bmod.call(internal, args, ret);
     }
 
     const2wasm(e: HDLConstant, opts: Options) : number {
@@ -707,14 +742,17 @@ export class HDLModuleWASM implements HDLModuleRunner {
         if (local != null) {
             return this.bmod.local.get(local.index, local.itype);
         } else if (global != null) {
-            return this.loadmem(this.dataptr(), global.offset, global.size);
+            if (global.size <= 8)
+                return this.loadmem(this.dataptr(), global.offset, global.size);
+            else
+                return this.address2wasm(e);
         }
         throw new HDLError(e, `cannot lookup variable ${e.refname}`)
     }
 
     local2wasm(e: HDLVariableDef, opts:Options) : number {
         var local = this.locals.lookup(e.name);
-        if (local == null) throw Error(`no local for ${e.name}`)
+        //if (local == null) throw Error(`no local for ${e.name}`)
         return this.bmod.nop(); // TODO
     }
 
@@ -829,7 +867,11 @@ export class HDLModuleWASM implements HDLModuleRunner {
             block.push(this.e2w(e.precond));
         }
         if (e.loopcond) {
-            block.push(this.bmod.br_if("@block", this.e2w(e.loopcond, {resulttype:binaryen.i32})));
+            block.push(this.bmod.if(
+                    this.e2w(e.loopcond, {resulttype:binaryen.i32}),
+                    this.bmod.nop(),
+                    this.bmod.br("@block")  // exit loop
+            ));
         }
         if (e.body) {
             block.push(this.e2w(e.body));
@@ -972,6 +1014,15 @@ export class HDLModuleWASM implements HDLModuleRunner {
     _sub2wasm(e: HDLBinop) {
         return this.binop(e, this.i3264(e.dtype).sub);
     }
+    _mul2wasm(e: HDLBinop) {
+        return this.binop(e, this.i3264(e.dtype).mul);
+    }
+    _moddiv2wasm(e: HDLBinop) {
+        return this.binop(e, this.i3264(e.dtype).rem_u);
+    }
+    _div2wasm(e: HDLBinop) {
+        return this.binop(e, this.i3264(e.dtype).div_u);
+    }
     _moddivs2wasm(e: HDLBinop) {
         return this.binop(e, this.i3264(e.dtype).rem_s);
     }
diff --git a/src/common/hdl/vxmlparser.ts b/src/common/hdl/vxmlparser.ts
index 1e0999c5..3eaa1f11 100644
--- a/src/common/hdl/vxmlparser.ts
+++ b/src/common/hdl/vxmlparser.ts
@@ -165,6 +165,7 @@ export class VerilogXMLParser implements HDLUnit {
             instances: [],
             vardefs: {},
         }
+        if (this.cur_module) throw new Error(`nested modules not supported`);
         this.cur_module = module;
         return module;
     }
@@ -187,7 +188,7 @@ export class VerilogXMLParser implements HDLUnit {
         var m = re_const.exec(s);
         if (m) {
             var numstr = m[3];
-            if (numstr.length < 8)
+            if (numstr.length <= 8)
                 return parseInt(numstr, 16);
             else
                 return BigInt('0x' + numstr);
@@ -213,6 +214,9 @@ export class VerilogXMLParser implements HDLUnit {
     visit_verilator_xml(node: XMLNode) {
     }
 
+    visit_package(node: XMLNode) { // TODO?
+    }
+
     visit_module(node: XMLNode) {
         this.findChildren(node, 'var', false).forEach((n) => {
             if (isVarDecl(n.obj)) {
@@ -329,6 +333,10 @@ export class VerilogXMLParser implements HDLUnit {
     }
 
     visit_cfunc(node: XMLNode) : HDLBlock {
+        if (this.cur_module == null) { // TODO?
+            console.log('no module open, skipping', node);
+            return;
+        }
         var block = this.visit_begin(node);
         block.exprs = [];
         node.children.forEach((n) => block.exprs.push(n.obj));
@@ -648,8 +656,8 @@ export class VerilogXMLParser implements HDLUnit {
     visit_rand(node: XMLNode) { return this.__visit_func(node); }
     visit_time(node: XMLNode) { return this.__visit_func(node); }
 
-    visit_display(node: XMLNode) { return this.__visit_func(node); }
-    visit_sformatf(node: XMLNode) { return this.visit_begin(node); }
+    visit_display(node: XMLNode) { return null; }
+    visit_sformatf(node: XMLNode) { return null; }
 
     visit_readmem(node: XMLNode) { return this.__visit_func(node); }
 
diff --git a/src/platform/verilog.ts b/src/platform/verilog.ts
index efb33da4..0145b117 100644
--- a/src/platform/verilog.ts
+++ b/src/platform/verilog.ts
@@ -4,7 +4,7 @@ import { PLATFORMS, setKeyboardFromMap, AnimationTimer, RasterVideo, Keys, makeK
 import { SampleAudio } from "../common/audio";
 import { safe_extend } from "../common/util";
 import { WaveformView, WaveformProvider, WaveformMeta } from "../ide/waveform";
-import { setFrameRateUI, loadScript } from "../ide/ui";
+import { setFrameRateUI, loadScript, current_project } from "../ide/ui";
 import { HDLModuleRunner, HDLModuleTrace, HDLUnit, isLogicType } from "../common/hdl/hdltypes";
 import { HDLModuleJS } from "../common/hdl/hdlruntime";
 import { HDLModuleWASM } from "../common/hdl/hdlwasm";
@@ -142,13 +142,17 @@ var VerilogPlatform = function(mainElement, options) {
     if (useAudio) {
       audio.feedSample(top.state.spkr, 1);
     }
+    resetKbdStrobe();
+    if (debugCond && debugCond()) {
+      debugCond = null;
+    }
+  }
+
+  function resetKbdStrobe() {
     if (keycode && keycode >= 128 && top.state.keystrobe) { // keystrobe = clear hi bit of key buffer
       keycode = keycode & 0x7f;
       top.state.keycode = keycode;
     }
-    if (debugCond && debugCond()) {
-      debugCond = null;
-    }
   }
 
   function doreset() {
@@ -367,7 +371,7 @@ var VerilogPlatform = function(mainElement, options) {
     ncycles |= 0;
     var inspect = inspect_obj && inspect_sym;
     // use fast trace buffer-based update?
-    if (sync && !trace && top['trace'] != null) {
+    if (sync && !trace && top['trace'] != null && scanlineCycles > 0) {
       this.updateVideoFrameFast((top as any) as HDLModuleTrace);
       this.updateRecorder();
       return;
@@ -389,14 +393,17 @@ var VerilogPlatform = function(mainElement, options) {
           idata[frameidx] = rgb & 0x80000000 ? rgb : RGBLOOKUP[rgb & 15];
           frameidx++;
         }
+        scanlineCycles++;
       } else if (!framehsync && top.state.hsync) {
         framehsync = true;
+        scanlineCycles++;
       } else if ((framehsync && !top.state.hsync) || framex > videoWidth*2) {
         framehsync = false;
         framex = 0;
         framey++;
         top.state.hpaddle = framey > video.paddle_x ? 1 : 0;
         top.state.vpaddle = framey > video.paddle_y ? 1 : 0;
+        scanlineCycles = 0;
       }
       if (framey > maxVideoLines || top.state.vsync) {
         framevsync = true;
@@ -418,9 +425,9 @@ var VerilogPlatform = function(mainElement, options) {
 
   // use trace buffer to update video
   updateVideoFrameFast(tmod: HDLModuleTrace) {
-    var maxLineCycles = videoWidth < 300 ? 521 : 1009; // prime number so we eventually sync up
+    var maxLineCycles = 1009; // prime number so we eventually sync up
     if (!scanlineCycles) scanlineCycles = maxLineCycles;
-    var nextlineCycles = scanlineCycles;
+    var nextlineCycles = scanlineCycles + 1;
     // TODO: we can go faster if no paddle/sound
     frameidx = 0;
     var wasvsync = false;
@@ -434,6 +441,8 @@ var VerilogPlatform = function(mainElement, options) {
       }
       // generate frames in trace buffer
       top.tick2(nextlineCycles);
+      // TODO: this has to be done more quickly
+      resetKbdStrobe();
       // convert trace buffer to video/audio
       var n = 0;
       if (framey < videoHeight) {
@@ -446,24 +455,25 @@ var VerilogPlatform = function(mainElement, options) {
         n += videoWidth;
       }
       // find hsync
-      while (n < maxLineCycles && !tmod.trace.hsync) { spkr(); tmod.nextTrace(); n++; }
-      while (n < maxLineCycles && tmod.trace.hsync) { spkr(); tmod.nextTrace(); n++; }
+      while (n < nextlineCycles && !tmod.trace.hsync) { spkr(); tmod.nextTrace(); n++; }
+      while (n < nextlineCycles && tmod.trace.hsync) { spkr(); tmod.nextTrace(); n++; }
       // see if our scanline cycle count is stable
-      if (n == scanlineCycles) {
+      if (n == scanlineCycles + 1) {
         // scanline cycle count licked in, reset buffer to improve cache locality
         nextlineCycles = n;
-        tmod.resetTrace(); 
+        tmod.resetTrace();
       } else {
-        // not in sync, set to prime # and we'll eventually sync
-        nextlineCycles = maxLineCycles;
+        // not in sync, don't reset buffer (TODO: take some of the cycles back)
+        //console.log('scanline', scanlineCycles, nextlineCycles, n);
         scanlineCycles = n;
+        nextlineCycles = n;
       }
       // exit when vsync starts and then stops
       if (tmod.trace.vsync) {
         wasvsync = true;
-      } else if (wasvsync) {
         top.state.hpaddle = 0;
         top.state.vpaddle = 0;
+      } else if (wasvsync) {
         break;
       }
     }
@@ -541,6 +551,7 @@ var VerilogPlatform = function(mainElement, options) {
         // initialize top module and constant pool
         var useWASM = true;
         var _top = new (useWASM ? HDLModuleWASM : HDLModuleJS)(topmod, unit.modules['@CONST-POOL@']);
+        _top.getFileData = (path) => current_project.filedata[path]; // external file provider
         await _top.init();
         _top.powercycle();
         if (top) top.dispose();
diff --git a/src/worker/workermain.ts b/src/worker/workermain.ts
index f0793497..af646d92 100644
--- a/src/worker/workermain.ts
+++ b/src/worker/workermain.ts
@@ -1806,20 +1806,20 @@ function compileVerilator(step:BuildStep) {
       return {errors:errors};
     }
     var xmlParser = new emglobal['VerilogXMLParser']();
+    var listings : CodeListingMap = {};
     try {
       var xmlContent = FS.readFile(xmlPath, {encoding:'utf8'});
+      listings[step.prefix + '.lst'] = {lines:[],text:xmlContent};
       putWorkFile(xmlPath, xmlContent);
       if (!anyTargetChanged(step, [xmlPath]))
         return;
       xmlParser.parse(xmlContent);
     } catch(e) {
-      console.log(e);
+      console.log(e, e.stack);
       errors.push({line:0,msg:""+e});
-      return {errors:errors};
+      return {errors:errors, listings:listings};
     }
     //rtn.intermediate = {listing:h_file + cpp_file}; // TODO
-    var listings : CodeListingMap = {};
-    listings[step.prefix + '.lst'] = {lines:[],text:xmlContent};
     // TODO: what if found in non-top-module?
     if (asmlines.length)
       listings[step.path] = {lines:asmlines};