Create assembler driver script for recording word addresses

This will allow the profiler to sample the current return stack and
instruction pointer.  It can't just use the ip and rsp globals
directly, as I had initially planned, since these cannot be sent
in between threads.

README.md

@@ -74,21 +74,20 @@ guile tests.scm | xmllint --format -
 
 | Name   | Address | Size / B | Access       |
 |--------|---------|----------|--------------|
-| TXBUF  | 000h    | 32       | write        |
+| TXBUF  | 00h     | 32       | write        |
-| RXBUF  | 080h    | 32       | read         |
+| RXBUF  | 20h     | 32       | read         |
-| TXHEAD | 100h    | 4        | atomic read  |
+| TXHEAD | 40h     | 4        | atomic read  |
-| TXTAIL | 104h    | 4        | atomic write |
+| TXTAIL | 44h     | 4        | atomic write |
-| RXHEAD | 108h    | 4        | atomic write |
+| RXHEAD | 48h     | 4        | atomic write |
-| RXTAIL | 10Ch    | 4        | atomic read  |
+| RXTAIL | 4Ch     | 4        | atomic read  |
 
 For both sending (`TX`) and receiving (`RX`), there are three
 registers: `xBUF`, `xHEAD` and `xTAIL`:
 
-- `xBUF` registers are 128-byte FIFO ring buffers used for data
+- `xBUF` registers are 32-byte FIFO ring buffers used for data
 - The `xHEAD` and `xTAIL` registers specify the start and end of data
   in the ring buffer, `xHEAD` being the offset of the first byte of
-  data, and `xTAIL` being the offset of the first byte *after* the
+  data, and `xTAIL` being the offset of the first byte *after* the data.
-  data.
 
 In order to be distinguishable from the empty state, the ring buffers
 must never be completely full -- there must always be *at least one*
@@ -98,10 +97,7 @@ unoccupied byte between the tail and the head.
 
 | Name     | Address | Size / B | Access       |
 |----------|---------|----------|--------------|
-| SYSREADY | 110h    | 4        | atomic write |
+| SYSREADY | 50h     | 1        | atomic write |
-| SYSINTER | 114h    | 4        | atomic read  |
 
 The `SYSREADY` register is used to indicate when the system has booted
-up and is ready for user input.  `SYSINTER` is set (and notified on)
+up and is ready for user input.
-once the emulator has enabled user input and the system is
-interactive.

asm.js

@@ -14,29 +14,20 @@ class Tokenizer {
 		this.buffer = [];
         this.comment = false;
         this.string = false;
-        this.cursor = 0;
 	}
 
-    find(pred) {
-        for (let i = this.cursor; i < this.buffer.length; ++i) {
-            if (pred(this.buffer[i]))
-                return i;
-        }
-        return null;
-    }
-
 	skip() {
-		const idx = this.find((cp) => !this.skips.has(cp));
+		const idx = this.buffer.findIndex((cp) => !this.skips.has(cp));
-        this.cursor = idx == null ? this.buffer.length : idx;
+        this.buffer = idx == -1 ? [] : this.buffer.slice(idx);
     }
 
     next_string() {
-        const idx = this.find((cp) => cp == this.string_quote);
+        const idx = this.buffer.findIndex((cp) => cp == this.string_quote);
-        if (idx == null) {
+        if (idx == -1) {
             this.string = true;
         } else {
-            const string = this.buffer.slice(this.cursor, idx).join("");
+            const string = this.buffer.slice(0, idx).join("");
-            this.cursor = idx + 1;
+            this.buffer = this.buffer.slice(idx + 1);
             this.string = false;
             return { string: string };
         }
@@ -47,20 +38,18 @@ class Tokenizer {
             return this.next_string();
 
         this.skip();
-        if (this.buffer[this.cursor] == LINE_END) {
+        if (this.buffer[0] == LINE_END)
-            ++this.cursor;
+            return this.buffer.shift();
-            return LINE_END;
-        }
 
-        if (this.buffer[this.cursor] == this.string_quote) {
+        if (this.buffer[0] == this.string_quote) {
-            ++this.cursor;
+            this.buffer.shift();
             return this.next_string();
         }
 
-		const idx = this.find((cp) => this.delims.has(cp));
+		const idx = this.buffer.findIndex((cp) => this.delims.has(cp));
-        if (idx != null) {
+        if (idx != -1) {
-            const token = this.buffer.slice(this.cursor, idx).join("");
+            const token = this.buffer.slice(0, idx).join("");
-            this.cursor = idx;
+            this.buffer = this.buffer.slice(idx);
             return token;
         }
 	}
@@ -155,49 +144,52 @@ const types = {
 };
 
 const opcodes = {
-    "block":         0x02,
+    "block":            0x02,
-    "loop":          0x03,
+    "loop":             0x03,
-    "if":            0x04,
+    "if":               0x04,
-    "else":          0x05,
+    "else":             0x05,
-    "end":           0x0b,
+    "end":              0x0b,
-    "br":            0x0c,
+    "br":               0x0c,
-    "br_if":         0x0d,
+    "br_if":            0x0d,
-    "call":          0x10,
+    "call":             0x10,
-    "call_indirect": 0x11,
+    "call_indirect":    0x11,
-    "drop":          0x0a,
+    "drop":             0x0a,
-    "local.get":     0x20,
+    "local.get":        0x20,
-    "local.set":     0x21,
+    "local.set":        0x21,
-    "local.tee":     0x22,
+    "local.tee":        0x22,
-    "global.get":    0x23,
+    "global.get":       0x23,
-    "global.set":    0x24,
+    "global.set":       0x24,
-    "i32.load":      0x28,
+    "i32.load":         0x28,
-    "i32.load8_u":   0x2d,
+    "i32.load8_u":      0x2d,
-    "i32.store":     0x36,
+    "i32.store":        0x36,
-    "i32.store8":    0x3a,
+    "i32.store8":       0x3a,
-    "i32.const":     0x41,
+    "i32.const":        0x41,
-    "i64.const":     0x42,
+    "i64.const":        0x42,
-    "i32.eqz":       0x45,
+    "i32.eqz":          0x45,
-    "i32.eq":        0x46,
+    "i32.eq":           0x46,
-    "i32.ne":        0x47,
+    "i32.ne":           0x47,
-    "i32.lt_s":      0x48,
+    "i32.lt_s":         0x48,
-    "i32.lt_u":      0x49,
+    "i32.lt_u":         0x49,
-    "i32.gt_s":      0x4a,
+    "i32.gt_s":         0x4a,
-    "i32.gt_u":      0x4b,
+    "i32.gt_u":         0x4b,
-    "i32.le_s":      0x4c,
+    "i32.le_s":         0x4c,
-    "i32.le_u":      0x4d,
+    "i32.le_u":         0x4d,
-    "i32.ge_s":      0x4e,
+    "i32.ge_s":         0x4e,
-    "i32.ge_u":      0x4f,
+    "i32.ge_u":         0x4f,
-    "i32.add":       0x6a,
+    "i32.add":          0x6a,
-    "i32.sub":       0x6b,
+    "i32.sub":          0x6b,
-    "i32.mul":       0x6c,
+    "i32.mul":          0x6c,
-	"i32.div_s":     0x6d,
+	"i32.div_s":        0x6d,
-	"i32.rem_s":     0x6f,
+	"i32.rem_s":        0x6f,
-    "i32.and":       0x71,
+    "i32.and":          0x71,
-    "i32.or":        0x72,
+    "i32.or":           0x72,
-    "i32.xor":       0x73,
+    "i32.xor":          0x73,
-    "i32.shl":       0x74,
+    "i32.shl":          0x74,
-    "i32.shr_s":     0x75,
+    "i32.shr_s":        0x75,
-    "i32.shr_u":     0x76,
+    "i32.shr_u":        0x76,
+    "i64.or":           0x84,
+    "i64.shl":          0x86,
+    "i64.extend_i32_u": 0xad,
 
     // Threads instructions
     "memory.atomic.notify": [ 0xfe, 0x00 ],
@@ -205,6 +197,7 @@ const opcodes = {
     "i32.atomic.load":      [ 0xfe, 0x10 ],
     "i32.atomic.load8_u":   [ 0xfe, 0x12 ],
     "i32.atomic.store":     [ 0xfe, 0x17 ],
+    "i64.atomic.store":     [ 0xfe, 0x18 ],
     "i32.atomic.store8":    [ 0xfe, 0x19 ],
 };
 
@@ -897,8 +890,9 @@ const Section = Object.freeze({
 });
 
 const Kind = Object.freeze({
-    FUNC: 0x00,
+    FUNC:   0x00,
-    MEM:  0x02,
+    MEM:    0x02,
+    GLOBAL: 0x03,
 });
 
 export class Assembler {
@@ -953,8 +947,28 @@ export class Assembler {
     }
 
     action_export(action) {
-        const index = Object.keys(this.funcs).indexOf(action.name);
+        const func_index = Object.keys(this.funcs).indexOf(action.name);
-        this.exports[action.name] = { kind: Kind.FUNC, index };
+        if (func_index != -1) {
+            this.exports[action.name] = {
+                kind: Kind.FUNC,
+                index: func_index,
+            };
+            return;
+        }
+
+        const global_index = Object.keys(this.globals).indexOf(action.name);
+        if (global_index != -1) {
+            this.exports[action.name] = {
+                kind: Kind.GLOBAL,
+                index: global_index,
+            };
+            return;
+        }
+
+        console.error(
+            `ERROR: Unable to resolve export ${action.name} `
+                + "(only functions and globals currently supported)"
+        );
     }
 
     action_func(action) {

emu.js

@@ -1,17 +1,17 @@
-const TXBUF = 0x000;
+const TXBUF = 0x00;
-const RXBUF = 0x080;
+const RXBUF = 0x20;
-const TXHEAD = 0x100;
+const TXHEAD = 0x40;
-const TXTAIL = 0x104;
+const TXTAIL = 0x44;
-const RXHEAD = 0x108;
+const RXHEAD = 0x48;
-const RXTAIL = 0x10c;
+const RXTAIL = 0x4c;
+const SYSREADY = 0x50;
 
-const SYSREADY = 0x110;
+const TXBUF_SIZE = 32;
-const SYSINTER = 0x114;
+const RXBUF_SIZE = 32;
+const PERIPHS_SIZE = 81;
 
-const PERIPHS_SIZE = 0x200;
+const POLL_INTERVAL_MS = 20;
-
+const DOT_INTERVAL_MS = 120;
-const POLL_INTERVAL_MS = 5;
-const DOT_INTERVAL_MS = 25;
 
 const COLS = 80;
 const TAB_WIDTH = 8;
@@ -51,18 +51,28 @@ class Emulator {
 		document.addEventListener('keydown', (e) => this.handle_keydown(e));
 		window.addEventListener('resize', () => this.handle_resize());
 
-		this.forth = new Worker('boot.js', { type: 'module' });
+		this.prof = new Worker("prof.js");
+		this.prof.onmessage = (e) => {
+			const blob = new Blob(
+				[JSON.stringify(e.data)],
+				{ type: "application/json" });
+			const url = URL.createObjectURL(blob);
+			const a = document.createElement("a");
+			a.href = url;
+			a.download = "wipforth-profile.json";
+			a.click();
+			URL.revokeObjectURL(url);
+		};
 
 		this.print("Assembling kernel ");
-		this.dots = setInterval(() => this.print("."), DOT_INTERVAL_MS);
+		const dots = setInterval(() => this.print("."), DOT_INTERVAL_MS);
-		this.forth.postMessage({ type: "load", mem: this.mem });
+		this.worker = new Worker('boot.js', { type: 'module' });
-		this.forth.onmessage = (e) => {
+		this.worker.postMessage({ type: "load", mem: this.mem });
-			clearInterval(this.dots);
+		this.worker.onmessage = (e) => {
+			clearInterval(dots);
 			this.print(" done\n");
-
+			this.worker.postMessage({ type: "boot" });
-			this.print("Loading prelude ");
+			this.prof.postMessage({ type: "start", mem: this.mem });
-			this.forth.postMessage({ type: "boot" });
-			this.dots = setInterval(() => this.print("."), DOT_INTERVAL_MS);
 		};
 
 		fetch('prelude.f')
@@ -89,21 +99,16 @@ class Emulator {
 		if (!this.input_enable) {
 			const sysready = Atomics.load(this.mem_u8, SYSREADY);
 			if (sysready != 0) {
-				clearInterval(this.dots);
+				this.prof.postMessage({ type: "stop" });
-				this.print(" done\n");
-
-				Atomics.store(this.mem_u8, SYSINTER, 1);
-				Atomics.notify(this.mem_i32, SYSINTER / 4);
-
 				this.input_enable = true;
-				this.blink = true;
 				this.flush_output();
+				document.getElementById('cursor').classList.add('blinking');
 			}
 		}
 	}
 
 	fifo_next(idx) {
-		return (idx + 1) & 0x7f;
+		return (idx + 1) & 0x1f;
 	}
 
 	handle_tx_data(head, tail) {
@@ -191,14 +196,12 @@ class Emulator {
 		}
 
 		e.preventDefault();
-		this.blink = false;
 		this.flush_output();
 
 		document.getElementById('cursor').classList.remove('blinking');
 		if (this.idle_timer)
 			clearTimeout(this.idle_timer);
 		this.idle_timer = setTimeout(() => {
-			this.blink = true;
 			document.getElementById('cursor').classList.add('blinking');
 		}, CURSOR_IDLE_TIME_MS);
 	}
@@ -285,12 +288,10 @@ class Emulator {
 			return row.map((c, x) => {
 				const ec = this.html_escape(c);
 				if (this.input_enable
-					&& x == this.cursor.x && y == this.cursor.y) {
+					&& x == this.cursor.x && y == this.cursor.y)
-					const cl = this.blink ? 'class="blinking"' : '';
+					return '<span id="cursor">' + ec + '</span>';
-					return `<span id="cursor" ${cl}>` + ec + '</span>';
+				else
-				} else {
 					return ec;
-				}
 			}).join('').trimEnd();
 		}).join('\n');
 		this.output.innerHTML = html;

prelude.f

@@ -1,5 +1,12 @@
+76 EMIT 111 EMIT 97 EMIT 100 EMIT 105 EMIT 110 EMIT 103 EMIT 32 EMIT
+112 EMIT 114 EMIT 101 EMIT 108 EMIT 117 EMIT 100 EMIT 101 EMIT 32 EMIT
+
 : \ KEY 10 = 0BRANCH [ -20 , ] ; IMMEDIATE \ Now we have line comments :)
 
+\ We'll periodically sprinkle these in so that it's clear to the user
+\ that things are happening.
+46 EMIT
+
 \ Conditionals
 
 : IF
@@ -20,6 +27,8 @@
   SWAP !
 ; IMMEDIATE
 
+46 EMIT
+
 \ Loops
 
 : BEGIN HERE @ ; IMMEDIATE
@@ -34,6 +43,8 @@
   HERE @ - ,
 ; IMMEDIATE
 
+46 EMIT
+
 \ Recursive calls
 
 : RECURSE LATEST @ >CFA , ; IMMEDIATE
@@ -50,6 +61,8 @@
 
 ( ( Take that, C ) )
 
+46 EMIT
+
 \ Printing utilities
 
 : CR 10 EMIT ;
@@ -67,6 +80,8 @@
   + EMIT
 ;
 
+CHAR . EMIT
+
 : .
   \ Handle negatives
   DUP 0< IF CHAR - EMIT NEGATE THEN
@@ -95,6 +110,8 @@
   2DROP
 ;
 
+CHAR . EMIT
+
 : TYPE ( addr len -- )
   BEGIN
 	DUP 0= IF 2DROP EXIT THEN
@@ -103,6 +120,8 @@
   AGAIN
 ;
 
+CHAR . EMIT
+
 : C, HERE @ C! 1 HERE +! ;
 
 : ."
@@ -132,6 +151,8 @@
   THEN
 ; IMMEDIATE
 
+CHAR . EMIT
+
 \ Misc utilities
 
 : NIP SWAP DROP ;
@@ -150,6 +171,8 @@
 
 : [COMPILE] ' , ; IMMEDIATE
 
+CHAR . EMIT
+
 \ Constants, variables and values
 
 : CONSTANT
@@ -181,21 +204,24 @@
   THEN
 ; IMMEDIATE
 
+CHAR . EMIT
+
 \ Peripheral register addresses
 
 HEX
 
-000 CONSTANT TXBUF
+00 CONSTANT TXBUF
-080 CONSTANT RXBUF
+20 CONSTANT RXBUF
-100 CONSTANT TXHEAD
+40 CONSTANT TXHEAD
-104 CONSTANT TXTAIL
+44 CONSTANT TXTAIL
-108 CONSTANT RXHEAD
+48 CONSTANT RXHEAD
-10C CONSTANT RXTAIL
+4C CONSTANT RXTAIL
-110 CONSTANT SYSREADY
+50 CONSTANT SYSREADY
-114 CONSTANT SYSINTER
 
 DECIMAL
 
+46 EMIT
+
 \ A better word-not-found handler
 
 : ANY-RX? RXHEAD AC@ RXTAIL AC@ <> ;
@@ -218,11 +244,13 @@ DECIMAL
 
 ' WNF-HANDLER TO WNFHOOK
 
+CHAR . EMIT
+
 \ Version number
 
 0 CONSTANT VERSION-MAJOR
 2 CONSTANT VERSION-MINOR
-2 CONSTANT VERSION-PATCH
+1 CONSTANT VERSION-PATCH
 
 : PRINT-VERSION
   CHAR v EMIT VERSION-MAJOR .
@@ -230,6 +258,8 @@ DECIMAL
   CHAR . EMIT VERSION-PATCH .
 ;
 
+CHAR . EMIT
+
 \ Welcome banner
 
 : BANNER
@@ -239,7 +269,7 @@ DECIMAL
   ." |__,__/_/ .__(_)___/_//_/" CR
   ."        /_/               " CR
   CR
-  ." Welcome to Wipforth " PRINT-VERSION  ." !" CR
+  ." Wipforth " PRINT-VERSION CR
   ." Copyright (c) Camden Dixie O'Brien" CR
   CR
   ." Wipforth is freely available to use, modify and distribute for personal use" CR
@@ -247,8 +277,8 @@ DECIMAL
   CR
 ;
 
-\ Set SYSREADY high and wait until interactive
+."  done" CR
-1 SYSREADY AC!
-SYSINTER WAIT DROP
-
 BANNER
+
+\ Set SYSREADY high to enable user input
+1 SYSREADY AC!

prof.js

@@ -0,0 +1,48 @@
+const INTERVAL_MS = 1;
+const RS_TOP_ADDR = 0x10000;
+const PROF_DATA_ADDR = 0x58;
+const PROF_DATA_IDX = PROF_DATA_ADDR / 8;
+
+let mem_8;
+let mem_64;
+let sampler;
+
+const samples = [];
+
+function sample() {
+	const data = Atomics.load(mem_64, PROF_DATA_IDX);
+	const ip = Number(data & 0xffffffffn);
+	const rsp = Number(data >> 32n);
+	samples.push({ ip, rs_bytes: mem_8.slice(rsp, RS_TOP_ADDR) });
+}
+
+function i32(bytes) {
+	return bytes[0]
+		| (bytes[1] << 8)
+		| (bytes[2] << 16)
+		| (bytes[3] << 24);
+}
+function postproc({ ip, rs_bytes }) {
+	const rs = [];
+	for (let i = 0; i < rs_bytes.length; i += 4)
+		rs.push(i32(rs_bytes.slice(i, i + 4)));
+	rs.reverse();
+	return { ip, rs };
+}
+
+self.onmessage = (e) => {
+	switch (e.data.type) {
+	case "start":
+		console.log("Starting profiler");
+		mem_8 = new Uint8Array(e.data.mem.buffer);
+		mem_64 = new BigUint64Array(e.data.mem.buffer);
+		ip = e.data.ip;
+		rsp = e.data.rsp;
+		sampler = setInterval(sample, INTERVAL_MS);
+		break;
+	case "stop":
+		clearInterval(sample);
+		console.log("Stopped profiler");
+		self.postMessage(samples.map(postproc));
+	}
+};

tests.scm

@@ -19,7 +19,7 @@
 (define client (client-setup))
 
 (navigate client "http://localhost:8080")
-(sleep 1)
+(sleep 5)
 
 (define-test kernel-assembles-successfully
   (let* ((display (get-display client))

wipforth.ws

@@ -2,12 +2,15 @@
 .import main "emu" "mem"
 
 ;; Peripheral registers
-.def TXBUF  000h
+.def TXBUF  00h
-.def RXBUF  080h
+.def RXBUF  20h
-.def TXHEAD 100h
+.def TXHEAD 40h
-.def TXTAIL 104h
+.def TXTAIL 44h
-.def RXHEAD 108h
+.def RXHEAD 48h
-.def RXTAIL 10Ch
+.def RXTAIL 4Ch
+
+;; Mirror of registers for profiler to sample
+.def PROF_DATA 58h
 
 .def DICT_START 0200h
 
@@ -635,15 +638,6 @@
     end
     call next
 
-.func wait
-.elem codewords wait WAIT_CODEWORD
-	call pop
-	i32.const 0
-	i64.const -1
-	memory.atomic.wait32 2 0
-	call push
-	call next
-
 ;; Core utility words
 
 .func exit
@@ -761,7 +755,7 @@
     local.get head
     i32.const 1
     i32.add
-    i32.const 7Fh
+    i32.const 1Fh
     i32.and
     i32.atomic.store8 0 0
 
@@ -777,7 +771,7 @@
       local.tee tail
       i32.const 1
       i32.add
-      i32.const 7Fh
+      i32.const 1Fh
       i32.and
       local.tee n
       i32.const TXHEAD
@@ -1235,15 +1229,6 @@ COPY:
 	.word COPY_CODEWORD
 .def PREV _COPY
 
-_WAIT:
-	.word PREV
-	.byte 4
-	.utf8 "WAIT"
-	.align
-WAIT:
-	.word WAIT_CODEWORD
-.def PREV _WAIT
-
 _EXIT:
 	.word PREV
 	.byte 4
@@ -2091,6 +2076,13 @@ KERNEL_DEFS_END:
 .func trampoline
     loop iter
       global.get fn call_indirect codeword codewords
+
+	  i32.const PROF_DATA
+	  global.get ip  i64.extend_i32_u
+	  global.get rsp i64.extend_i32_u
+	  i64.const 32 i64.shl i64.or
+	  i64.atomic.store 3 0
+
       global.get run br_if iter
     end
 

words.js

@@ -0,0 +1,16 @@
+import { Assembler } from "./asm.js";
+
+const asm = new Assembler();
+for await (const chunk of Deno.stdin.readable) {
+	asm.push(chunk);
+}
+asm.wasm();
+
+const defs = Object.entries(asm.defs);
+while (defs[0][0] != '_DUP')
+	defs.shift();
+while (defs.at(-1)[0] != 'WNF_HANDLER')
+	defs.pop();
+const words = Object.fromEntries(defs.filter(([k,v]) => !k.startsWith("_")));
+
+console.log(JSON.stringify(words));