porffor 0.36.7 → 0.37.0

package/compiler/builtins/__internal_string.ts

@@ -6,8 +6,8 @@ export const __Porffor_strcmp = (a: any, b: any): boolean => {
   // fast path: check if pointers are equal
   if (Porffor.wasm`local.get ${a}` == Porffor.wasm`local.get ${b}`) return true;
 
-  const al: i32 = Porffor.wasm.i32.load(a, 0, 0);
-  const bl: i32 = Porffor.wasm.i32.load(b, 0, 0);
+  let al: i32 = Porffor.wasm.i32.load(a, 0, 0);
+  let bl: i32 = Porffor.wasm.i32.load(b, 0, 0);
 
   // fast path: check if lengths are inequal
   if (al != bl) return false;
@@ -15,9 +15,81 @@ export const __Porffor_strcmp = (a: any, b: any): boolean => {
   if (Porffor.wasm`local.get ${a+1}` == Porffor.TYPES.bytestring) {
     if (Porffor.wasm`local.get ${b+1}` == Porffor.TYPES.bytestring) {
       // bytestring, bytestring
-      for (let i: i32 = 0; i < al; i++) {
-        if (Porffor.wasm.i32.load8_u(Porffor.wasm`local.get ${a}` + i, 0, 4) !=
-            Porffor.wasm.i32.load8_u(Porffor.wasm`local.get ${b}` + i, 0, 4)) return false;
+      // this path is hyper-optimized as it is by far the most common and (perf) important
+
+      let ap: i32 = a - 4;
+      let bp: i32 = b - 4;
+      Porffor.wasm`
+;; load in i64 chunks while length >= 8
+local.get ${al}
+i32.const 8
+i32.ge_s
+if 64
+  loop 64
+    local.get ${ap}
+    local.get ${al}
+    i32.add
+    i64.load 0 0
+
+    local.get ${bp}
+    local.get ${al}
+    i32.add
+    i64.load 0 0
+
+    i64.ne
+    if 64
+      i32.const 0
+      i32.const 2
+      return
+    end
+
+    local.get ${al}
+    i32.const 8
+    i32.sub
+    local.tee ${al}
+    i32.const 8
+    i32.ge_s
+    br_if 0
+  end
+end
+
+;; load in u16 chunks while length >= 2
+local.get ${al}
+i32.const 2
+i32.ge_s
+if 64
+  loop 64
+    local.get ${ap}
+    local.get ${al}
+    i32.add
+    i32.load16_u 0 6
+
+    local.get ${bp}
+    local.get ${al}
+    i32.add
+    i32.load16_u 0 6
+
+    i32.ne
+    if 64
+      i32.const 0
+      i32.const 2
+      return
+    end
+
+    local.get ${al}
+    i32.const 2
+    i32.sub
+    local.tee ${al}
+    i32.const 2
+    i32.ge_s
+    br_if 0
+  end
+end`;
+
+      // check bonus char if exists
+      if (al == 1) {
+        if (Porffor.wasm.i32.load8_u(Porffor.wasm`local.get ${a}`, 0, 4) !=
+            Porffor.wasm.i32.load8_u(Porffor.wasm`local.get ${b}`, 0, 4)) return false;
       }
       return true;
     } else {
@@ -38,6 +110,40 @@ export const __Porffor_strcmp = (a: any, b: any): boolean => {
       return true;
     } else {
       // string, string
+      let ap: i32 = a - 4;
+      let bp: i32 = b - 4;
+      Porffor.wasm`
+loop 64
+  local.get ${ap}
+  local.get ${al}
+  i32.const 2
+  i32.mul
+  i32.add
+  i64.load 0 0
+
+  local.get ${bp}
+  local.get ${al}
+  i32.const 2
+  i32.mul
+  i32.add
+  i64.load 0 0
+
+  i64.ne
+  if 64
+    i32.const 0
+    i32.const 2
+    return
+  end
+
+  local.get ${al}
+  i32.const 4
+  i32.sub
+  local.tee ${al}
+  i32.const 4
+  i32.ge_s
+  br_if 0
+end`;
+
       for (let i: i32 = 0; i < al; i++) {
         if (Porffor.wasm.i32.load16_u(Porffor.wasm`local.get ${a}` + i*2, 0, 4) !=
             Porffor.wasm.i32.load16_u(Porffor.wasm`local.get ${b}` + i*2, 0, 4)) return false;

package/compiler/builtins_precompiled.js

@@ -15,9 +15,9 @@ locals:[124,124,124,127,127,124,124,124,127,124,124,124,124,127,127,127],localNa
 globalInits:{underlyingKeys:(_,{allocPage,glbl,loc})=>[...number(allocPage(_,'array: __internal_object.ts/underlyingKeys','f64'),124),...glbl(36,'underlyingKeys',124),...glbl(35,'underlyingKeys',124),[252,3],[33,loc('#makearray_pointer_tmp',127)],[32,loc('#makearray_pointer_tmp',127)],[65,0],[54,1,0],[32,loc('#makearray_pointer_tmp',127)],[68,0],[252,3],[54,1,0],[32,loc('#makearray_pointer_tmp',127)],[26]],underlyingVals:(_,{allocPage,glbl,loc})=>[...number(allocPage(_,'array: __internal_object.ts/underlyingVals','f64'),124),...glbl(36,'underlyingVals',124),...glbl(35,'underlyingVals',124),[252,3],[33,loc('#makearray_pointer_tmp',127)],[32,loc('#makearray_pointer_tmp',127)],[65,0],[54,1,0],[32,loc('#makearray_pointer_tmp',127)],[68,0],[252,3],[54,1,0],[32,loc('#makearray_pointer_tmp',127)],[26]]},data:{"bytestring: __Porffor_object_makeObject/key1":[9,0,0,0,112,114,111,116,111,116,121,112,101],"bytestring: __Porffor_object_makeObject/key2":[11,0,0,0,99,111,110,115,116,114,117,99,116,111,114]},
 };
 this.__Porffor_strcmp = {
-wasm:()=>[[32,0],[32,2],[70],[4,64],[65,1],[65,2],[15],[11],[32,0],[40,0,0],[33,4],[32,2],[40,0,0],[33,5],[32,4],[32,5],[71],[4,64],[65,0],[65,2],[15],[11],[32,1],[65,195,1],[70],[4,64],[32,3],[65,195,1],[70],[4,64],[65,0],[33,6],[3,64],[32,6],[32,4],[72],[4,64],[32,0],[32,6],[106],[45,0,4],[32,2],[32,6],[106],[45,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,6],[65,1],[106],[33,6],[12,1],[11],[11],[65,1],[65,2],[15],[5],[65,0],[33,6],[3,64],[32,6],[32,4],[72],[4,64],[32,0],[32,6],[106],[45,0,4],[32,2],[32,6],[65,2],[108],[106],[47,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,6],[65,1],[106],[33,6],[12,1],[11],[11],[65,1],[65,2],[15],[11],[5],[32,3],[65,195,1],[70],[4,64],[65,0],[33,6],[3,64],[32,6],[32,4],[72],[4,64],[32,0],[32,6],[65,2],[108],[106],[47,0,4],[32,2],[32,6],[106],[45,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,6],[65,1],[106],[33,6],[12,1],[11],[11],[65,1],[65,2],[15],[5],[65,0],[33,6],[3,64],[32,6],[32,4],[72],[4,64],[32,0],[32,6],[65,2],[108],[106],[47,0,4],[32,2],[32,6],[65,2],[108],[106],[47,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,6],[65,1],[106],[33,6],[12,1],[11],[11],[65,1],[65,2],[15],[11],[11],[65,0],[65,128,1],[15]],
+wasm:()=>[[32,0],[32,2],[70],[4,64],[65,1],[65,2],[15],[11],[32,0],[40,0,0],[33,4],[32,2],[40,0,0],[33,5],[32,4],[32,5],[71],[4,64],[65,0],[65,2],[15],[11],[32,1],[65,195,1],[70],[4,64],[32,3],[65,195,1],[70],[4,64],[32,0],[65,4],[107],[33,6],[32,2],[65,4],[107],[33,7],[32,4],[65,8],[78],[4,64],[3,64],[32,6],[32,4],[106],[41,0,0],[32,7],[32,4],[106],[41,0,0],[82],[4,64],[65,0],[65,2],[15],[11],[32,4],[65,8],[107],[34,4],[65,8],[78],[13,0],[11],[11],[32,4],[65,2],[78],[4,64],[3,64],[32,6],[32,4],[106],[47,0,6],[32,7],[32,4],[106],[47,0,6],[71],[4,64],[65,0],[65,2],[15],[11],[32,4],[65,2],[107],[34,4],[65,2],[78],[13,0],[11],[11],[32,4],[65,1],[70],[4,64],[32,0],[45,0,4],[32,2],[45,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[11],[65,1],[65,2],[15],[5],[65,0],[33,8],[3,64],[32,8],[32,4],[72],[4,64],[32,0],[32,8],[106],[45,0,4],[32,2],[32,8],[65,2],[108],[106],[47,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,8],[65,1],[106],[33,8],[12,1],[11],[11],[65,1],[65,2],[15],[11],[5],[32,3],[65,195,1],[70],[4,64],[65,0],[33,8],[3,64],[32,8],[32,4],[72],[4,64],[32,0],[32,8],[65,2],[108],[106],[47,0,4],[32,2],[32,8],[106],[45,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,8],[65,1],[106],[33,8],[12,1],[11],[11],[65,1],[65,2],[15],[5],[32,0],[65,4],[107],[33,6],[32,2],[65,4],[107],[33,7],[3,64],[32,6],[32,4],[65,2],[108],[106],[41,0,0],[32,7],[32,4],[65,2],[108],[106],[41,0,0],[82],[4,64],[65,0],[65,2],[15],[11],[32,4],[65,4],[107],[34,4],[65,4],[78],[13,0],[11],[65,0],[33,8],[3,64],[32,8],[32,4],[72],[4,64],[32,0],[32,8],[65,2],[108],[106],[47,0,4],[32,2],[32,8],[65,2],[108],[106],[47,0,4],[71],[4,64],[65,0],[65,2],[15],[11],[32,8],[65,1],[106],[33,8],[12,1],[11],[11],[65,1],[65,2],[15],[11],[11],[65,0],[65,128,1],[15]],
 params:[127,127,127,127],typedParams:1,returns:[127,127],typedReturns:1,
-locals:[127,127,127],localNames:["a","a#type","b","b#type","al","bl","i"],
+locals:[127,127,127,127,127],localNames:["a","a#type","b","b#type","al","bl","ap","bp","i"],
 };
 this.__Porffor_strcat = {
 wasm:(_,{builtin})=>[[32,0],[40,0,0],[33,4],[32,2],[40,0,0],[33,5],[32,1],[65,195,1],[70],[4,64],[32,3],[65,195,1],[70],[4,64],[65,4],[32,4],[106],[32,5],[106],[16,builtin('__Porffor_allocateBytes')],[34,6],[32,4],[32,5],[106],[54,0,0],[32,6],[65,4],[106],[32,0],[65,4],[106],[32,4],[252,10,0,0],[32,6],[65,4],[106],[32,4],[106],[32,2],[65,4],[106],[32,5],[252,10,0,0],[32,6],[65,195,1],[15],[5],[65,4],[32,4],[32,5],[106],[65,2],[108],[106],[16,builtin('__Porffor_allocateBytes')],[34,6],[32,4],[32,5],[106],[54,0,0],[65,0],[33,7],[3,64],[32,7],[32,4],[72],[4,64],[32,6],[32,7],[65,2],[108],[106],[32,0],[32,7],[106],[45,0,4],[59,0,4],[32,7],[65,1],[106],[33,7],[12,1],[11],[11],[32,6],[65,4],[106],[32,4],[65,2],[108],[106],[32,2],[65,4],[106],[32,5],[65,2],[108],[252,10,0,0],[32,6],[65,195,0],[15],[11],[5],[32,3],[65,195,1],[70],[4,64],[65,4],[32,4],[32,5],[106],[65,2],[108],[106],[16,builtin('__Porffor_allocateBytes')],[34,6],[32,4],[32,5],[106],[54,0,0],[32,6],[65,4],[106],[32,0],[65,4],[106],[32,4],[65,2],[108],[252,10,0,0],[32,6],[32,4],[65,2],[108],[106],[33,8],[65,0],[33,7],[3,64],[32,7],[32,5],[72],[4,64],[32,8],[32,7],[65,2],[108],[106],[32,2],[32,7],[106],[45,0,4],[59,0,4],[32,7],[65,1],[106],[33,7],[12,1],[11],[11],[32,6],[65,195,0],[15],[5],[65,4],[32,4],[32,5],[106],[65,2],[108],[106],[16,builtin('__Porffor_allocateBytes')],[34,6],[32,4],[32,5],[106],[54,0,0],[32,6],[65,4],[106],[32,0],[65,4],[106],[32,4],[65,2],[108],[252,10,0,0],[32,6],[65,4],[106],[32,4],[65,2],[108],[106],[32,2],[65,4],[106],[32,5],[65,2],[108],[252,10,0,0],[32,6],[65,195,0],[15],[11],[11],[65,0],[65,128,1],[15]],

package/compiler/codegen.js

@@ -248,7 +248,8 @@ const generate = (scope, decl, global = false, name = undefined, valueUnused = f
 
             const encodeFunc = ({
               [Opcodes.f64_const]: x => x,
-              [Opcodes.if]: unsignedLEB128
+              [Opcodes.if]: unsignedLEB128,
+              [Opcodes.loop]: unsignedLEB128
             })[inst[0]] ?? signedLEB128;
             out.push([ ...inst, ...immediates.flatMap(x => encodeFunc(x)) ]);
           }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "porffor",
   "description": "a basic experimental wip aot optimizing js -> wasm engine/compiler/runtime in js",
-  "version": "0.36.7+e872827a7",
+  "version": "0.37.0+44ce75202",
   "author": "CanadaHonk",
   "license": "MIT",
   "scripts": {},

package/runner/index.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
 import fs from 'node:fs';
-globalThis.version = '0.36.7+e872827a7';
+globalThis.version = '0.37.0+44ce75202';
 
 // deno compat
 if (typeof process === 'undefined' && typeof Deno !== 'undefined') {