Optimize pack_memory by moving zero check to rust.

As suggested by copy in #1219 move the check if a memory
page is all zeros to rust.  Executes on 64bit chunks, which
is likely not as good simd could do, but is still a 5x speedup
in the included snapshot benchmark on my machine.

Co-authored-by: Adam M. Smith <[email protected]>

src/cpu.js

@@ -272,6 +272,7 @@ CPU.prototype.wasm_patch = function()
 
     this.allocate_memory = get_import("allocate_memory");
     this.zero_memory = get_import("zero_memory");
+    this.is_memory_zeroed = get_import("is_memory_zeroed");
 
     this.svga_allocate_memory = get_import("svga_allocate_memory");
     this.svga_allocate_dest_buffer = get_import("svga_allocate_dest_buffer");
@@ -625,23 +626,9 @@ CPU.prototype.pack_memory = function()
 
     const page_count = this.mem8.length >> 12;
     const nonzero_pages = [];
-
     for(let page = 0; page < page_count; page++)
     {
-        const offset = page << 12;
-        const view = this.mem32s.subarray(offset >> 2, offset + 0x1000 >> 2);
-        let is_zero = true;
-
-        for(let i = 0; i < view.length; i++)
-        {
-            if(view[i] !== 0)
-            {
-                is_zero = false;
-                break;
-            }
-        }
-
-        if(!is_zero)
+        if(!this.is_memory_zeroed(page << 12, 0x1000))
         {
             nonzero_pages.push(page);
         }

src/rust/cpu/memory.rs

@@ -310,3 +310,15 @@ pub unsafe fn mmap_write128(addr: u32, v0: u64, v1: u64) {
         )
     }
 }
+
+#[no_mangle]
+pub unsafe fn is_memory_zeroed(addr: u32, length: u32) -> bool {
+    dbg_assert!(addr % 8 == 0);
+    dbg_assert!(length % 8 == 0);
+    for i in (addr..addr + length).step_by(8) {
+        if *(mem8.offset(i as isize) as *const i64) != 0 {
+            return false;
+        }
+    }
+    return true;
+}

tests/benchmark/snapshot.js

@@ -0,0 +1,63 @@
+#!/usr/bin/env node
+"use strict";
+
+const BENCH_COLLECT_STATS = +process.env.BENCH_COLLECT_STATS;
+
+const V86 = require(`../../build/${BENCH_COLLECT_STATS ? "libv86-debug" : "libv86"}.js`).V86;
+const print_stats = require("../../build/libv86.js").print_stats;
+const fs = require("fs");
+const path = require("path");
+const V86_ROOT = path.join(__dirname, "../..");
+
+const LOG_SERIAL = true;
+
+
+var emulator = new V86({
+    bios: { url: __dirname + "/../../bios/seabios.bin" },
+    vga_bios: { url: __dirname + "/../../bios/vgabios.bin" },
+    cdrom: { url: __dirname + "/../../images/linux3.iso" },
+    autostart: true,
+    memory_size: 1024 * 1024 * 1024,
+    disable_jit: +process.env.DISABLE_JIT,
+    log_level: 0,
+});
+
+
+emulator.bus.register("emulator-started", function()
+{
+    console.log("Booting now, please stand by");
+    start_time = Date.now();
+});
+
+var serial_text = "";
+var start_time;
+
+emulator.add_listener("serial0-output-byte", function(byte)
+{
+    var chr = String.fromCharCode(byte);
+    if(chr < " " && chr !== "\n" && chr !== "\t" || chr > "~")
+    {
+        return;
+    }
+
+    if(LOG_SERIAL) process.stdout.write(chr);
+
+    serial_text += chr;
+
+    if(serial_text.endsWith("~% ") || serial_text.endsWith("root@localhost:~# "))
+    {
+        console.log("Creating snapshots");
+        const start_time = Date.now();
+        for(var i = 0; i < 10; ++i) emulator.save_state();
+        const end_time = Date.now();
+        const elapsed = end_time - start_time;
+        console.log("Done in %dms", elapsed);
+        emulator.destroy();
+
+        if(BENCH_COLLECT_STATS)
+        {
+            const cpu = emulator.v86.cpu;
+            console.log(print_stats.stats_to_string(cpu));
+        }
+    }
+});