porffor 0.16.0-ab08df866 → 0.16.0-b099006b8

package/compiler/havoc.js

@@ -0,0 +1,93 @@
+// havoc: wasm rewrite library (it wreaks havoc upon wasm bytecode hence "havoc")
+import { Opcodes, Valtype } from './wasmSpec.js';
+
+export const localsToConsts = (func, targets, consts, { localKeys }) => {
+  const { wasm, locals } = func;
+
+  // update locals object
+  localKeys ??= Object.keys(locals).sort((a, b) => a.idx - b.idx);
+  for (const x of targets) {
+    delete locals[localKeys[x]];
+  }
+
+  const idxLut = {};
+  for (const x in locals) {
+    let i = locals[x].idx;
+    const o = i;
+    for (let j = 0; j < targets.length; j++) {
+      if (o > targets[j]) i--;
+    }
+
+    locals[x].idx = i;
+    idxLut[o] = i;
+  }
+
+  // update wasm
+  for (let i = 0; i < wasm.length; i++) {
+    const op = wasm[i];
+    const opcode = op[0];
+    const idx = op[1];
+
+    if (opcode >= 0x20 && opcode <= 0x22) { // local.* op
+      if (targets.includes(idx)) {
+        if (opcode === Opcodes.local_get || opcode === Opcodes.local_tee) {
+          // get/tee -> const
+          const c = consts[targets.indexOf(idx)];
+          wasm.splice(i, 1, c);
+          continue;
+        }
+
+        // set -> drop
+        wasm.splice(i, 1, [ Opcodes.drop ]);
+        continue;
+      }
+
+      // adjust index to compensate for removed
+      wasm[i] = [ opcode, idxLut[idx] ];
+    }
+  }
+};
+
+export const f64ToI32s = (func, targets) => {
+  const { wasm, locals } = func;
+
+  // update locals object
+  for (const x in locals) {
+    let i = locals[x].idx;
+    if (targets.includes(i)) {
+      locals[x].type = Valtype.i32;
+    }
+  }
+
+  // update wasm
+  for (let i = 0; i < wasm.length; i++) {
+    const op = wasm[i];
+    const opcode = op[0];
+    const idx = op[1];
+
+    if (opcode >= 0x20 && opcode <= 0x22 && targets.includes(idx)) { // local.* op
+      if (opcode === Opcodes.local_get) {
+        // add i32 -> f64 after
+        wasm.splice(i + 1, 0, Opcodes.i32_from);
+        continue;
+      }
+
+      if (opcode === Opcodes.local_set) {
+        // add f64 -> i32 before
+        wasm.splice(i, 0, Opcodes.i32_to);
+        i++;
+        continue;
+      }
+
+      if (opcode === Opcodes.local_tee) {
+        // add f64 -> i32 before
+        wasm.splice(i, 0, Opcodes.i32_to);
+        i++;
+
+        // add i32 -> f64 after
+        wasm.splice(i + 1, 0, Opcodes.i32_from);
+        continue;
+      }
+    }
+  }
+};

package/compiler/index.js

@@ -1,10 +1,13 @@
 import { underline, bold, log } from './log.js';
+import { Valtype, PageSize } from './wasmSpec.js';
 import parse from './parse.js';
 import codegen from './codegen.js';
 import opt from './opt.js';
 import assemble from './assemble.js';
 import decompile from './decompile.js';
 import toc from './2c.js';
+import * as pgo from './pgo.js';
+import cyclone from './cyclone.js';
 import Prefs from './prefs.js';
 
 globalThis.decompile = decompile;
@@ -13,6 +16,7 @@ const logFuncs = (funcs, globals, exceptions) => {
   console.log('\n' + underline(bold('funcs')));
 
   for (const f of funcs) {
+    if (f.internal) continue;
     console.log(decompile(f.wasm, f.name, f.index, f.locals, f.params, f.returns, funcs, globals, exceptions));
   }
 
@@ -23,6 +27,24 @@ const fs = (typeof process?.version !== 'undefined' ? (await import('node:fs'))
 const execSync = (typeof process?.version !== 'undefined' ? (await import('node:child_process')).execSync : undefined);
 
 export default (code, flags) => {
+  let target = Prefs.target ?? 'wasm';
+  if (Prefs.native) target = 'native';
+
+  let outFile = Prefs.o;
+
+  globalThis.valtype = 'f64';
+  const valtypeOpt = process.argv.find(x => x.startsWith('--valtype='));
+  if (valtypeOpt) valtype = valtypeOpt.split('=')[1];
+  globalThis.valtypeBinary = Valtype[valtype];
+
+  globalThis.pageSize = PageSize;
+  const pageSizeOpt = process.argv.find(x => x.startsWith('--page-size='));
+  if (pageSizeOpt) pageSize = parseInt(pageSizeOpt.split('=')[1]) * 1024;
+
+  // enable pgo by default for c/native
+  if (target !== 'wasm') Prefs.pgo = Prefs.pgo === false ? false : true;
+  if (Prefs.pgo) pgo.setup();
+
   const t0 = performance.now();
   const program = parse(code, flags);
   if (Prefs.profileCompiler) console.log(`1. parsed in ${(performance.now() - t0).toFixed(2)}ms`);
@@ -35,9 +57,49 @@ export default (code, flags) => {
 
   const t2 = performance.now();
   opt(funcs, globals, pages, tags, exceptions);
-  if (Prefs.profileCompiler) console.log(`3. optimized in ${(performance.now() - t2).toFixed(2)}ms`);
 
-  // if (Prefs.optFuncs) logFuncs(funcs, globals, exceptions);
+  if (Prefs.pgo) {
+    if (Prefs.pgoLog) {
+      const oldSize = assemble(funcs, globals, tags, pages, data, flags, true).byteLength;
+      const t = performance.now();
+
+      pgo.run({ funcs, globals, tags, exceptions, pages, data });
+      opt(funcs, globals, pages, tags, exceptions);
+
+      console.log(`PGO total time: ${(performance.now() - t).toFixed(2)}ms`);
+
+      const newSize = assemble(funcs, globals, tags, pages, data, flags, true).byteLength;
+      console.log(`PGO size diff: ${oldSize - newSize} bytes (${oldSize} -> ${newSize})\n`);
+    } else {
+      pgo.run({ funcs, globals, tags, exceptions, pages, data });
+      opt(funcs, globals, pages, tags, exceptions);
+    }
+  }
+
+  if (Prefs.cyclone) {
+    if (Prefs.cycloneLog) {
+      const oldSize = assemble(funcs, globals, tags, pages, data, flags, true).byteLength;
+      const t = performance.now();
+
+      for (const x of funcs) {
+        const preOps = x.wasm.length;
+        cyclone(x.wasm);
+
+        console.log(`${x.name}: ${preOps} -> ${x.wasm.length} ops`);
+      }
+
+      console.log(`cyclone total time: ${(performance.now() - t).toFixed(2)}ms`);
+
+      const newSize = assemble(funcs, globals, tags, pages, data, flags, true).byteLength;
+      console.log(`cyclone size diff: ${oldSize - newSize} bytes (${oldSize} -> ${newSize})\n`);
+    } else {
+      for (const x of funcs) {
+        cyclone(x.wasm);
+      }
+    }
+  }
+
+  if (Prefs.profileCompiler) console.log(`3. optimized in ${(performance.now() - t2).toFixed(2)}ms`);
 
   const t3 = performance.now();
   const wasm = assemble(funcs, globals, tags, pages, data, flags);
@@ -54,9 +116,6 @@ export default (code, flags) => {
 
   const out = { wasm, funcs, globals, tags, exceptions, pages, data };
 
-  const target = Prefs.target ?? 'wasm';
-  const outFile = Prefs.o;
-
   if (target === 'wasm' && outFile) {
     fs.writeFileSync(outFile, Buffer.from(wasm));
 
@@ -77,6 +136,8 @@ export default (code, flags) => {
   }
 
   if (target === 'native') {
+    outFile ??= Prefs.native ? './porffor_tmp' : file.split('/').at(-1).split('.').at(0, -1).join('.');
+
     let compiler = Prefs.compiler ?? 'clang';
     const cO = Prefs._cO ?? 'Ofast';
 
@@ -95,7 +156,29 @@ export default (code, flags) => {
 
     fs.unlinkSync(tmpfile);
 
-    if (process.version) process.exit();
+    if (process.version) {
+      if (Prefs.native) {
+        const cleanup = () => {
+          try {
+            fs.unlinkSync(outFile);
+          } catch {}
+        };
+
+        process.on('exit', cleanup);
+        process.on('beforeExit', cleanup);
+        process.on('SIGINT', () => {
+          cleanup();
+          process.exit();
+        });
+
+        const runArgs = process.argv.slice(2).filter(x => !x.startsWith('-'));
+        try {
+          execSync([ outFile, ...runArgs.slice(1) ].join(' '), { stdio: 'inherit' });
+        } catch {}
+      }
+
+      process.exit();
+    }
   }
 
   return out;

package/compiler/opt.js

@@ -354,7 +354,7 @@ export default (funcs, globals, pages, tags, exceptions) => {
           continue;
         }
 
-        if (lastInst[0] === Opcodes.const && (inst === Opcodes.i32_to || inst === Opcodes.i32_to_u)) {
+        if (lastInst[0] === Opcodes.const && inst[0] === Opcodes.i32_to[0] && (inst[1] === Opcodes.i32_to[1] || inst[1] === Opcodes.i32_to_u[1])) {
           // change const and immediate i32 convert to i32 const
           // f64.const 0
           // i32.trunc_sat_f64_s || i32.trunc_sat_f64_u
@@ -504,42 +504,6 @@ export default (funcs, globals, pages, tags, exceptions) => {
       const useCount = {};
       for (const x in f.locals) useCount[f.locals[x].idx] = 0;
 
-      // final pass
-      depth = [];
-      for (let i = 0; i < wasm.length; i++) {
-        let inst = wasm[i];
-        if (inst[0] === Opcodes.local_get || inst[0] === Opcodes.local_set || inst[0] === Opcodes.local_tee) useCount[inst[1]]++;
-
-        if (inst[0] === Opcodes.if || inst[0] === Opcodes.loop || inst[0] === Opcodes.block) depth.push(inst[0]);
-        if (inst[0] === Opcodes.end) depth.pop();
-
-        if (i < 2) continue;
-        const lastInst = wasm[i - 1];
-        const lastLastInst = wasm[i - 2];
-
-        // todo: add more math ops
-        if (optLevel >= 3 && (inst[0] === Opcodes.add || inst[0] === Opcodes.sub || inst[0] === Opcodes.mul) && lastLastInst[0] === Opcodes.const && lastInst[0] === Opcodes.const) {
-          // inline const math ops
-          // i32.const a
-          // i32.const b
-          // i32.add
-          // -->
-          // i32.const a + b
-
-          // does not work with leb encoded
-          if (lastInst.length > 2 || lastLastInst.length > 2) continue;
-
-          let a = lastLastInst[1];
-          let b = lastInst[1];
-
-          const val = performWasmOp(inst[0], a, b);
-          if (Prefs.optLog) log('opt', `inlined math op (${a} ${inst[0].toString(16)} ${b} -> ${val})`);
-
-          wasm.splice(i - 2, 3, ...number(val)); // remove consts, math op and add new const
-          i -= 2;
-        }
-      }
-
       const localIdxs = Object.values(f.locals).map(x => x.idx);
       // remove unused locals (cleanup)
       for (const x in useCount) {

package/compiler/pgo.js

@@ -0,0 +1,207 @@
+import { Opcodes, Valtype } from './wasmSpec.js';
+import { number } from './embedding.js';
+import { importedFuncs } from './builtins.js';
+import Prefs from './prefs.js';
+import assemble from './assemble.js';
+import wrap, { writeByteStr } from './wrap.js';
+import * as Havoc from './havoc.js';
+
+export const setup = () => {
+  importedFuncs[importedFuncs.profile2].params = [ Valtype.i32, valtypeBinary ];
+
+  // enable these prefs by default for pgo
+  Prefs.typeswitchUniqueTmp = Prefs.typeswitchUniqueTmp === false ? false : true;
+  Prefs.cyclone = Prefs.cyclone === false ? false : true;;
+};
+
+export const run = obj => {
+  const wasmFuncs = obj.funcs;
+
+  let starts = {};
+  const time = (id, msg) => {
+    if (!Prefs.pgoLog) return;
+
+    if (!starts[id]) {
+      process.stdout.write(msg);
+      starts[id] = performance.now();
+    } else {
+      process.stdout.write(`\r${' '.repeat(50)}\r[${(performance.now() - starts[id]).toFixed(2)}ms] ${msg}\n`);
+    }
+  };
+
+  time(0, `injecting PGO logging...`);
+
+  let funcs = [];
+  for (let i = 0; i < wasmFuncs.length; i++) {
+    const { name, internal, params, locals, wasm } = wasmFuncs[i];
+    if (internal) continue; // ignore internal funcs
+    wasmFuncs[i].originalWasm = structuredClone(wasm);
+
+    const invLocals = Object.keys(locals).reduce((acc, x) => { acc[locals[x].idx] = locals[x]; return acc; }, {});
+
+    const id = funcs.length;
+    funcs.push({ name, id, locals, params, invLocals });
+
+    wasm.unshift(
+      // mark active func
+      ...number(i, Valtype.i32),
+      [ Opcodes.call, importedFuncs.profile1 ],
+
+      // log args
+      ...params.flatMap((_, i) => [
+        ...number(i, Valtype.i32),
+        [ Opcodes.local_get, i ],
+        ...(invLocals[i].type !== Valtype.f64 ? [ Opcodes.i32_from ] : []),
+        [ Opcodes.call, importedFuncs.profile2 ]
+      ])
+    );
+
+    for (let j = 0; j < wasm.length; j++) {
+      const inst = wasm[j];
+      if (inst[0] === Opcodes.local_set || inst[0] === Opcodes.local_tee) {
+        wasm.splice(j + 1, 0,
+          ...number(inst[1], Valtype.i32),
+          [ Opcodes.local_get, inst[1] ],
+          ...(invLocals[inst[1]].type !== Valtype.f64 ? [ Opcodes.i32_from ] : []),
+          [ Opcodes.call, importedFuncs.profile2 ]
+        );
+      }
+    }
+  }
+
+  let localData = funcs.map(x => new Array(Object.keys(x.locals).length).fill(0).map(() => []));
+
+  time(0, `injected PGO logging`);
+  time(1, `running with PGO logging...`);
+
+  let activeFunc = null, abort = false;
+  try {
+    obj.wasm = assemble(obj.funcs, obj.globals, obj.tags, obj.pages, obj.data, obj.flags, true);
+
+    const { exports } = wrap(obj, [], {
+      y: n => {
+        activeFunc = n;
+      },
+      z: (i, n) => {
+        if (activeFunc == null) throw 'fail';
+        localData[activeFunc][i].push(n);
+      },
+      w: (ind, outPtr) => { // readArgv
+        const pgoInd = process.argv.indexOf('--pgo');
+        const args = process.argv.slice(pgoInd).filter(x => !x.startsWith('-'));
+        const str = args[ind - 1];
+        if (pgoInd === -1 || !str) {
+          if (Prefs.pgoLog) console.log('\nPGO warning: script was expecting arguments, please specify args to use for PGO after --pgo arg');
+          return -1;
+        }
+
+        writeByteStr(exports.$, outPtr, str);
+        return str.length;
+      }
+    }, () => {});
+
+    exports.main();
+  } catch (e) {
+    throw e;
+  }
+
+  for (const x of funcs) {
+    const wasmFunc = wasmFuncs.find(y => y.name === x.name);
+    wasmFunc.wasm = wasmFunc.originalWasm;
+    delete wasmFunc.originalWasm;
+  }
+
+  if (abort) {
+    console.log('aborting PGO!');
+    return false;
+  }
+
+  time(1, `ran with PGO logging`);
+  time(2, 'processing PGO data...');
+
+  // process data
+  let log = '';
+  for (let i = 0; i < localData.length; i++) {
+    const func = funcs[i];
+    const total = localData[i].length;
+
+    const localKeys = Object.keys(func.locals).sort((a, b) => a.idx - b.idx);
+    const localValues = Object.values(func.locals).sort((a, b) => a.idx - b.idx);
+    func.localKeys = localKeys;
+    func.localValues = localValues;
+
+    let counts = new Array(10).fill(0);
+    const consistents = localData[i].map(x => {
+      if (x.length === 0 || !x.every((y, i) => i < 1 ? true : y === x[i - 1])) return false;
+
+      counts[0]++;
+      return x[0];
+    });
+
+    const integerOnlyF64s = localData[i].map((x, j) => {
+      if (localValues[j].type === Valtype.i32) return false; // already i32
+      if (x.length === 0 || !x.every(y => Number.isInteger(y))) return false;
+
+      counts[1]++;
+      return true;
+    });
+
+    func.consistents = consistents;
+    func.integerOnlyF64s = integerOnlyF64s;
+
+    log += `  ${func.name}: identified ${counts[0]}/${total} locals as consistent${Prefs.verbosePgo ? ':' : ''}\n`;
+    if (Prefs.verbosePgo) {
+      for (let j = 0; j < localData[i].length; j++) {
+        log += `    ${consistents[j] !== false ? '\u001b[92m' : '\u001b[91m'}${localKeys[j]}\u001b[0m: ${new Set(localData[i][j]).size} unique values set\n`;
+      }
+    }
+
+    log += `  ${func.name}: identified ${counts[1]}/${localValues.reduce((acc, x) => acc + (x.type === Valtype.f64 ? 1 : 0), 0)} f64 locals as integer usage only${Prefs.verbosePgo ? ':' : ''}\n`;
+    if (Prefs.verbosePgo) {
+      for (let j = 0; j < localData[i].length; j++) {
+        if (localValues[j].type !== Valtype.f64) continue;
+        log += `    ${integerOnlyF64s[j] ? '\u001b[92m' : '\u001b[91m'}${localKeys[j]}\u001b[0m\n`;
+      }
+
+      log += '\n';
+    }
+  }
+
+  time(2, 'processed PGO data' + log);
+  time(3, 'optimizing using PGO data...');
+
+  log = '';
+  for (const x of funcs) {
+    const wasmFunc = wasmFuncs.find(y => y.name === x.name);
+
+    let targets = [];
+    for (let i = 0; i < x.integerOnlyF64s.length; i++) {
+      const c = x.integerOnlyF64s[i];
+      if (c === false) continue;
+      if (i < x.params.length) continue;
+
+      targets.push(i);
+    }
+
+    log += `  ${x.name}: replaced ${targets.length} f64 locals with i32s\n`;
+    if (targets.length > 0) Havoc.f64ToI32s(wasmFunc, targets);
+
+    targets = [];
+    let consts = [];
+    for (let i = 0; i < x.consistents.length; i++) {
+      const c = x.consistents[i];
+      if (c === false) continue;
+      if (i < x.params.length) continue;
+
+      targets.push(i);
+
+      const valtype = x.localValues[i].type;
+      consts.push(number(c, valtype)[0]);
+    }
+
+    log += `  ${x.name}: replaced ${targets.length} locals with consts\n`;
+    if (targets.length > 0) Havoc.localsToConsts(wasmFunc, targets, consts, { localKeys: x.localKeys });
+  }
+
+  time(3, 'optimized using PGO data' + log);
+};

package/compiler/prefs.js

@@ -23,9 +23,14 @@ const obj = new Proxy({}, {
     // do not cache in web demo as args are changed live
     if (!globalThis.document) cache[p] = ret;
     return ret;
+  },
+
+  set(_, p, v) {
+    cache[p] = v;
+    return true;
   }
 });
 
-obj.uncache = () => cache = {};
+export const uncache = () => cache = {};
 
 export default obj;

package/compiler/wrap.js

@@ -6,8 +6,25 @@ import { TYPES } from './types.js';
 import { log } from './log.js';
 import Prefs from './prefs.js';
 
+const fs = (typeof process?.version !== 'undefined' ? (await import('node:fs')) : undefined);
+
 const bold = x => `\u001b[1m${x}\u001b[0m`;
 
+export const readByteStr = (memory, ptr) => {
+  const length = (new Int32Array(memory.buffer, ptr, 1))[0];
+  return Array.from(new Uint8Array(memory.buffer, ptr + 4, length)).map(x => String.fromCharCode(x)).join('');
+};
+
+export const writeByteStr = (memory, ptr, str) => {
+  const length = str.length;
+  (new Int32Array(memory.buffer, ptr, 1))[0] = length;
+
+  const arr = new Uint8Array(memory.buffer, ptr + 4, length);
+  for (let i = 0; i < length; i++) {
+    arr[i] = str.charCodeAt(i);
+  }
+};
+
 const porfToJSValue = ({ memory, funcs, pages }, value, type) => {
   switch (type) {
     case TYPES.boolean: return Boolean(value);
@@ -101,28 +118,31 @@ const porfToJSValue = ({ memory, funcs, pages }, value, type) => {
   }
 };
 
-export default async (source, flags = [ 'module' ], customImports = {}, print = str => process.stdout.write(str)) => {
+export default (source, flags = [ 'module' ], customImports = {}, print = str => process.stdout.write(str)) => {
   const times = [];
 
   const t1 = performance.now();
-  const { wasm, funcs, globals, tags, exceptions, pages, c } = compile(source, flags);
+  const { wasm, funcs, globals, tags, exceptions, pages, c } = typeof source === 'object' ? source : compile(source, flags);
 
   globalThis.porfDebugInfo = { funcs, globals };
 
-  if (source.includes('export function')) flags.push('module');
+  if (source.includes?.('export ')) flags.push('module');
 
-  // (await import('node:fs')).writeFileSync('out.wasm', Buffer.from(wasm));
+  fs.writeFileSync('out.wasm', Buffer.from(wasm));
 
   times.push(performance.now() - t1);
   if (Prefs.profileCompiler) console.log(bold(`compiled in ${times[0].toFixed(2)}ms`));
 
   const backtrace = (funcInd, blobOffset) => {
-    if (funcInd == null || blobOffset == null) return false;
+    if (funcInd == null || blobOffset == null ||
+        Number.isNaN(funcInd) || Number.isNaN(blobOffset)) return false;
 
     // convert blob offset -> function wasm offset.
     // this is not good code and is somewhat duplicated
     // I just want it to work for debugging, I don't care about perf/yes
     const func = funcs.find(x => x.index === funcInd);
+    if (!func) return false;
+
     const locals = Object.values(func.locals).sort((a, b) => a.idx - b.idx).slice(func.params.length).sort((a, b) => a.idx - b.idx);
 
     let localDecl = [], typeCount = 0, lastType;
@@ -197,13 +217,15 @@ export default async (source, flags = [ 'module' ], customImports = {}, print =
 
   let instance;
   try {
-    let wasmEngine = WebAssembly;
-    if (Prefs.asur) {
-      log.warning('wrap', 'using our !experimental! asur wasm engine instead of host to run');
-      wasmEngine = await import('../asur/index.js');
-    }
-
-    0, { instance } = await wasmEngine.instantiate(wasm, {
+    // let wasmEngine = WebAssembly;
+    // if (Prefs.asur) {
+    //   log.warning('wrap', 'using our !experimental! asur wasm engine instead of host to run');
+    //   wasmEngine = await import('../asur/index.js');
+    // }
+
+    // 0, { instance } = await wasmEngine.instantiate(wasm, {
+    const module = new WebAssembly.Module(wasm);
+    instance = new WebAssembly.Instance(module, {
       '': {
         p: valtype === 'i64' ? i => print(Number(i).toString()) : i => print(i.toString()),
         c: valtype === 'i64' ? i => print(String.fromCharCode(Number(i))) : i => print(String.fromCharCode(i)),
@@ -211,12 +233,31 @@ export default async (source, flags = [ 'module' ], customImports = {}, print =
         u: () => performance.timeOrigin,
         y: () => {},
         z: () => {},
+        w: (ind, outPtr) => { // readArgv
+          const args = process.argv.slice(2).filter(x => !x.startsWith('-'));
+          const str = args[ind];
+          if (!str) return -1;
+
+          writeByteStr(memory, outPtr, str);
+          return str.length;
+        },
+        q: (pathPtr, outPtr) => { // readFile
+          try {
+            const path = readByteStr(memory, pathPtr);
+            const contents = fs.readFileSync(path, 'utf8');
+            writeByteStr(memory, outPtr, contents);
+            return contents.length;
+          } catch {
+            return -1;
+          }
+        },
         ...customImports
       }
     });
   } catch (e) {
     // only backtrace for runner, not test262/etc
     if (!process.argv[1].includes('/runner')) throw e;
+    if (!(e instanceof WebAssembly.CompileError)) throw e;
 
     const funcInd = parseInt(e.message.match(/function #([0-9]+) /)?.[1]);
     const blobOffset = parseInt(e.message.split('@')?.[1]);