perf: small tweaks targeting cache prediction

Author: JairusSW

assembly/deserialize/simd/string.ts

@@ -71,15 +71,17 @@ export function deserializeString_SIMD(srcStart: usize, srcEnd: usize): string {
     const block = load<v128>(srcStart);
     store<v128>(bs.offset, block);
 
-    const eq5C = i16x8.eq(load<v128>(srcStart), SPLAT_5C);
-    let mask = i16x8.bitmask(eq5C);
+    const eq5C = i16x8.eq(block, SPLAT_5C);
+
     // Early exit
-    if (mask === 0) {
+    if (!v128.any_true(eq5C)) {
       srcStart += 16;
       bs.offset += 16;
       continue;
     }
 
+    let mask = i16x8.bitmask(eq5C);
+
     let srcChg: usize = 0;
     let lastLane: usize = 0;
     do {
@@ -157,40 +159,3 @@ export function deserializeString_SIMD(srcStart: usize, srcEnd: usize): string {
 
   return bs.out<string>();
 }
-
-/**
- * Computes a per-lane mask identifying UTF-16 code units whose **low byte**
- * is the ASCII backslash (`'\\'`, 0x5C).
- *
- * The mask is produced in two stages:
- * 1. Detects bytes equal to 0x5C using a SWAR equality test.
- * 2. Clears matches where 0x5C appears in the **high byte** of a UTF-16 code unit,
- *    ensuring only valid low-byte backslashes are reported.
- *
- * Each matching lane sets itself to 0x80.
- */
-// @ts-ignore: decorator
-@inline function backslash_mask(block: u64): u64 {
-  const b = block ^ 0x005C_005C_005C_005C;
-  const backslash_mask = (b - 0x0001_0001_0001_0001) & ~b & 0x0080_0080_0080_0080;
-  const high_byte_mask =
-    ~(((block - 0x0100_0100_0100_0100) & ~block & 0x8000_8000_8000_8000)
-      ^ 0x8000_8000_8000_8000) >> 8;
-  return backslash_mask & high_byte_mask;
-}
-
-/**
- * Computes a per-lane mask identifying UTF-16 code units whose **low byte**
- * is the ASCII backslash (`'\\'`, 0x5C).
- *
- * Each matching lane sets itself to 0x80.
- * 
- * WARNING: The low byte of a code unit *may* be a backslash, thus triggering false positives!
- * This is useful for a hot path where it is possible to detect the false positive scalarly.
- */
-// @ts-ignore: decorator
-@inline function backslash_mask_unsafe(block: u64): u64 {
-  const b = block ^ 0x005C_005C_005C_005C;
-  const backslash_mask = (b - 0x0001_0001_0001_0001) & ~b & 0x0080_0080_0080_0080;
-  return backslash_mask;
-}

assembly/deserialize/swar/string.ts

@@ -30,7 +30,7 @@ import { hex4_to_u16_swar } from "../../util/swar";
 // -\n-       2 -*-_       - 2
 // --\n       4 --*_       - 2
 // ---\n      6 ---*       - 0
-// Formula: overflow = 
+// Formula: overflow =
 
 /**
  * Deserializes strings back into into their original form using SIMD operations
@@ -48,7 +48,8 @@ export function deserializeString_SWAR(srcStart: usize, srcEnd: usize): string {
   while (srcStart < srcEnd8) {
     const block = load<u64>(srcStart);
     store<u64>(bs.offset, block);
-    let mask = backslash_mask_unsafe(block);
+
+    let mask = inline.always(backslash_mask_unsafe(block));
 
     // Early exit
     if (mask === 0) {
@@ -153,7 +154,7 @@ export function deserializeString_SWAR(srcStart: usize, srcEnd: usize): string {
  * is the ASCII backslash (`'\\'`, 0x5C).
  *
  * Each matching lane sets itself to 0x80.
- * 
+ *
  * WARNING: The low byte of a code unit *may* be a backslash, thus triggering false positives!
  * This is useful for a hot path where it is possible to detect the false positive scalarly.
  */

assembly/index.ts

@@ -183,7 +183,6 @@ export namespace JSON {
         // @ts-ignore: type
         return deserializeString_SIMD(dataPtr, dataPtr + dataSize);
       }
-      // }
     } else if (isArray<T>()) {
       // @ts-ignore
       return inline.always(deserializeArray<nonnull<T>>(dataPtr, dataPtr + dataSize, changetype<usize>(instantiate<T>())));

assembly/serialize/simd/string.ts

@@ -57,15 +57,13 @@ export function serializeString_SIMD(src: string): void {
     const sieve = v128.or(eq22, v128.or(eq5C, v128.or(lt20, gteD8)));
     // console.log("sieve  : " + mask_to_string_v128(sieve));
 
-    if (v128.any_true(sieve)) {
-
-      let mask = i8x16.bitmask(sieve);
-
-      if (mask === 0) {
+    if (!v128.any_true(sieve)) {
         bs.offset += 16;
         srcStart += 16;
         continue;
-      }
+    }
+
+    let mask = i8x16.bitmask(sieve);
 
       do {
         const laneIdx = ctz(mask);
@@ -125,7 +123,6 @@ export function serializeString_SIMD(src: string): void {
         store<v128>(dstIdx, load<v128>(srcIdx, 1), 12);
         bs.offset += 10;
       } while (mask !== 0);
-    }
 
     srcStart += 16;
     bs.offset += 16;
@@ -195,4 +192,4 @@ export function serializeString_SIMD(src: string): void {
 // @ts-ignore: inline
 @inline function hexNibble(n: u16): u16 {
   return n < 10 ? (48 + n) : (87 + n);
-}
+}

assembly/serialize/swar/number.ts

@@ -189,4 +189,4 @@ export function serializeString_SWAR(src: string): void {
   );
   const hi = block & 0xFF00_FF00_FF00_FF00;
   return ascii_mask | hi;
-}
+}