From: Agniva De Sarker Date: Sat, 22 Jun 2019 07:37:57 +0000 (+0530) Subject: runtime,syscall/js: reuse wasm memory DataView X-Git-Tag: go1.14beta1~1323 X-Git-Url: http://www.git.cypherpunks.su/?a=commitdiff_plain;h=07f04607370bf46c6f3a12ac3e182068da403b86;p=gostls13.git runtime,syscall/js: reuse wasm memory DataView Currently, every call to mem() incurs a new DataView object. This was necessary because the wasm linear memory could grow at any time. Now, whenever the memory grows, we make a call to the front-end. This allows us to reuse the existing DataView object and create a new one only when the memory actually grows. This gives us a boost in performance during DOM operations, while incurring an extra trip to front-end when memory grows. However, since the GrowMemory calls are meant to decrease over the runtime of an application, this is a good tradeoff in the long run. The benchmarks have been tested inside a browser (Google Chrome 75.0.3770.90 (Official Build) (64-bit)). It is hard to get stable nos. for DOM operations since the jumps make the timing very unreliable. But overall, it shows a clear gain. name old time/op new time/op delta DOM 135µs ±26% 84µs ±10% -37.22% (p=0.000 n=10+9) Go1 benchmarks do not show any noticeable degradation: name old time/op new time/op delta BinaryTree17 22.5s ± 0% 22.5s ± 0% ~ (p=0.743 n=8+9) Fannkuch11 15.1s ± 0% 15.1s ± 0% +0.17% (p=0.000 n=9+9) FmtFprintfEmpty 324ns ± 1% 303ns ± 0% -6.64% (p=0.000 n=9+10) FmtFprintfString 535ns ± 1% 515ns ± 0% -3.85% (p=0.000 n=10+10) FmtFprintfInt 609ns ± 0% 589ns ± 0% -3.28% (p=0.000 n=10+10) FmtFprintfIntInt 938ns ± 0% 920ns ± 0% -1.92% (p=0.000 n=9+10) FmtFprintfPrefixedInt 950ns ± 0% 924ns ± 0% -2.72% (p=0.000 n=10+9) FmtFprintfFloat 1.41µs ± 1% 1.43µs ± 0% +1.01% (p=0.000 n=10+10) FmtManyArgs 3.66µs ± 1% 3.46µs ± 0% -5.43% (p=0.000 n=9+10) GobDecode 38.8ms ± 1% 37.8ms ± 0% -2.50% (p=0.000 n=10+8) GobEncode 26.3ms ± 1% 26.3ms ± 0% ~ (p=0.853 n=10+10) Gzip 1.16s ± 1% 1.16s ± 0% -0.37% (p=0.008 n=10+9) Gunzip 210ms ± 0% 208ms ± 1% -1.01% (p=0.000 n=10+10) JSONEncode 48.0ms ± 0% 48.1ms ± 1% +0.29% (p=0.019 n=9+9) JSONDecode 348ms ± 1% 326ms ± 1% -6.34% (p=0.000 n=10+10) Mandelbrot200 6.62ms ± 0% 6.64ms ± 0% +0.37% (p=0.000 n=7+9) GoParse 23.9ms ± 1% 24.7ms ± 1% +2.98% (p=0.000 n=9+9) RegexpMatchEasy0_32 555ns ± 0% 561ns ± 0% +1.10% (p=0.000 n=8+10) RegexpMatchEasy0_1K 3.94µs ± 1% 3.94µs ± 0% ~ (p=0.906 n=9+8) RegexpMatchEasy1_32 516ns ± 0% 524ns ± 0% +1.51% (p=0.000 n=9+10) RegexpMatchEasy1_1K 4.39µs ± 1% 4.40µs ± 1% ~ (p=0.171 n=10+10) RegexpMatchMedium_32 25.1ns ± 0% 25.5ns ± 0% +1.51% (p=0.000 n=9+8) RegexpMatchMedium_1K 196µs ± 0% 203µs ± 1% +3.23% (p=0.000 n=9+10) RegexpMatchHard_32 11.2µs ± 1% 11.6µs ± 1% +3.62% (p=0.000 n=10+10) RegexpMatchHard_1K 334µs ± 1% 348µs ± 1% +4.21% (p=0.000 n=9+10) Revcomp 2.39s ± 0% 2.41s ± 0% +0.78% (p=0.000 n=8+9) Template 385ms ± 1% 336ms ± 0% -12.61% (p=0.000 n=10+9) TimeParse 2.18µs ± 1% 2.18µs ± 1% ~ (p=0.424 n=10+10) TimeFormat 2.28µs ± 1% 2.22µs ± 1% -2.30% (p=0.000 n=10+10) name old speed new speed delta GobDecode 19.8MB/s ± 1% 20.3MB/s ± 0% +2.56% (p=0.000 n=10+8) GobEncode 29.1MB/s ± 1% 29.2MB/s ± 0% ~ (p=0.810 n=10+10) Gzip 16.7MB/s ± 1% 16.8MB/s ± 0% +0.37% (p=0.007 n=10+9) Gunzip 92.2MB/s ± 0% 93.2MB/s ± 1% +1.03% (p=0.000 n=10+10) JSONEncode 40.4MB/s ± 0% 40.3MB/s ± 1% -0.28% (p=0.025 n=9+9) JSONDecode 5.58MB/s ± 1% 5.96MB/s ± 1% +6.80% (p=0.000 n=10+10) GoParse 2.42MB/s ± 0% 2.35MB/s ± 1% -2.83% (p=0.000 n=8+9) RegexpMatchEasy0_32 57.7MB/s ± 0% 57.0MB/s ± 0% -1.09% (p=0.000 n=8+10) RegexpMatchEasy0_1K 260MB/s ± 1% 260MB/s ± 0% ~ (p=0.963 n=9+8) RegexpMatchEasy1_32 62.1MB/s ± 0% 61.1MB/s ± 0% -1.53% (p=0.000 n=10+10) RegexpMatchEasy1_1K 233MB/s ± 1% 233MB/s ± 1% ~ (p=0.190 n=10+10) RegexpMatchMedium_32 39.8MB/s ± 0% 39.1MB/s ± 1% -1.74% (p=0.000 n=9+10) RegexpMatchMedium_1K 5.21MB/s ± 0% 5.05MB/s ± 1% -3.09% (p=0.000 n=9+10) RegexpMatchHard_32 2.86MB/s ± 1% 2.76MB/s ± 1% -3.43% (p=0.000 n=10+10) RegexpMatchHard_1K 3.06MB/s ± 1% 2.94MB/s ± 1% -4.06% (p=0.000 n=9+10) Revcomp 106MB/s ± 0% 105MB/s ± 0% -0.77% (p=0.000 n=8+9) Template 5.04MB/s ± 1% 5.77MB/s ± 0% +14.48% (p=0.000 n=10+9) Updates #32591 Change-Id: Id567e14a788e359248b2129ef1cf0adc8cc4ab7f Reviewed-on: https://go-review.googlesource.com/c/go/+/183457 Run-TryBot: Agniva De Sarker TryBot-Result: Gobot Gobot Reviewed-by: Richard Musiol --- diff --git a/misc/wasm/wasm_exec.js b/misc/wasm/wasm_exec.js index a54bb9a95d..7341e755e7 100644 --- a/misc/wasm/wasm_exec.js +++ b/misc/wasm/wasm_exec.js @@ -113,24 +113,19 @@ this._scheduledTimeouts = new Map(); this._nextCallbackTimeoutID = 1; - const mem = () => { - // The buffer may change when requesting more memory. - return new DataView(this._inst.exports.mem.buffer); - } - const setInt64 = (addr, v) => { - mem().setUint32(addr + 0, v, true); - mem().setUint32(addr + 4, Math.floor(v / 4294967296), true); + this.mem.setUint32(addr + 0, v, true); + this.mem.setUint32(addr + 4, Math.floor(v / 4294967296), true); } const getInt64 = (addr) => { - const low = mem().getUint32(addr + 0, true); - const high = mem().getInt32(addr + 4, true); + const low = this.mem.getUint32(addr + 0, true); + const high = this.mem.getInt32(addr + 4, true); return low + high * 4294967296; } const loadValue = (addr) => { - const f = mem().getFloat64(addr, true); + const f = this.mem.getFloat64(addr, true); if (f === 0) { return undefined; } @@ -138,7 +133,7 @@ return f; } - const id = mem().getUint32(addr, true); + const id = this.mem.getUint32(addr, true); return this._values[id]; } @@ -147,34 +142,34 @@ if (typeof v === "number") { if (isNaN(v)) { - mem().setUint32(addr + 4, nanHead, true); - mem().setUint32(addr, 0, true); + this.mem.setUint32(addr + 4, nanHead, true); + this.mem.setUint32(addr, 0, true); return; } if (v === 0) { - mem().setUint32(addr + 4, nanHead, true); - mem().setUint32(addr, 1, true); + this.mem.setUint32(addr + 4, nanHead, true); + this.mem.setUint32(addr, 1, true); return; } - mem().setFloat64(addr, v, true); + this.mem.setFloat64(addr, v, true); return; } switch (v) { case undefined: - mem().setFloat64(addr, 0, true); + this.mem.setFloat64(addr, 0, true); return; case null: - mem().setUint32(addr + 4, nanHead, true); - mem().setUint32(addr, 2, true); + this.mem.setUint32(addr + 4, nanHead, true); + this.mem.setUint32(addr, 2, true); return; case true: - mem().setUint32(addr + 4, nanHead, true); - mem().setUint32(addr, 3, true); + this.mem.setUint32(addr + 4, nanHead, true); + this.mem.setUint32(addr, 3, true); return; case false: - mem().setUint32(addr + 4, nanHead, true); - mem().setUint32(addr, 4, true); + this.mem.setUint32(addr + 4, nanHead, true); + this.mem.setUint32(addr, 4, true); return; } @@ -196,8 +191,8 @@ typeFlag = 3; break; } - mem().setUint32(addr + 4, nanHead | typeFlag, true); - mem().setUint32(addr, ref, true); + this.mem.setUint32(addr + 4, nanHead | typeFlag, true); + this.mem.setUint32(addr, ref, true); } const loadSlice = (addr) => { @@ -232,7 +227,7 @@ // func wasmExit(code int32) "runtime.wasmExit": (sp) => { - const code = mem().getInt32(sp + 8, true); + const code = this.mem.getInt32(sp + 8, true); this.exited = true; delete this._inst; delete this._values; @@ -244,10 +239,15 @@ "runtime.wasmWrite": (sp) => { const fd = getInt64(sp + 8); const p = getInt64(sp + 16); - const n = mem().getInt32(sp + 24, true); + const n = this.mem.getInt32(sp + 24, true); fs.writeSync(fd, new Uint8Array(this._inst.exports.mem.buffer, p, n)); }, + // func resetMemoryDataView() + "runtime.resetMemoryDataView": (sp) => { + this.mem = new DataView(this._inst.exports.mem.buffer); + }, + // func nanotime() int64 "runtime.nanotime": (sp) => { setInt64(sp + 8, (timeOrigin + performance.now()) * 1000000); @@ -257,7 +257,7 @@ "runtime.walltime": (sp) => { const msec = (new Date).getTime(); setInt64(sp + 8, msec / 1000); - mem().setInt32(sp + 16, (msec % 1000) * 1000000, true); + this.mem.setInt32(sp + 16, (msec % 1000) * 1000000, true); }, // func scheduleTimeoutEvent(delay int64) int32 @@ -276,12 +276,12 @@ }, getInt64(sp + 8) + 1, // setTimeout has been seen to fire up to 1 millisecond early )); - mem().setInt32(sp + 16, id, true); + this.mem.setInt32(sp + 16, id, true); }, // func clearTimeoutEvent(id int32) "runtime.clearTimeoutEvent": (sp) => { - const id = mem().getInt32(sp + 8, true); + const id = this.mem.getInt32(sp + 8, true); clearTimeout(this._scheduledTimeouts.get(id)); this._scheduledTimeouts.delete(id); }, @@ -327,10 +327,10 @@ const result = Reflect.apply(m, v, args); sp = this._inst.exports.getsp(); // see comment above storeValue(sp + 56, result); - mem().setUint8(sp + 64, 1); + this.mem.setUint8(sp + 64, 1); } catch (err) { storeValue(sp + 56, err); - mem().setUint8(sp + 64, 0); + this.mem.setUint8(sp + 64, 0); } }, @@ -342,10 +342,10 @@ const result = Reflect.apply(v, undefined, args); sp = this._inst.exports.getsp(); // see comment above storeValue(sp + 40, result); - mem().setUint8(sp + 48, 1); + this.mem.setUint8(sp + 48, 1); } catch (err) { storeValue(sp + 40, err); - mem().setUint8(sp + 48, 0); + this.mem.setUint8(sp + 48, 0); } }, @@ -357,10 +357,10 @@ const result = Reflect.construct(v, args); sp = this._inst.exports.getsp(); // see comment above storeValue(sp + 40, result); - mem().setUint8(sp + 48, 1); + this.mem.setUint8(sp + 48, 1); } catch (err) { storeValue(sp + 40, err); - mem().setUint8(sp + 48, 0); + this.mem.setUint8(sp + 48, 0); } }, @@ -384,7 +384,7 @@ // func valueInstanceOf(v ref, t ref) bool "syscall/js.valueInstanceOf": (sp) => { - mem().setUint8(sp + 24, loadValue(sp + 8) instanceof loadValue(sp + 16)); + this.mem.setUint8(sp + 24, loadValue(sp + 8) instanceof loadValue(sp + 16)); }, // func copyBytesToGo(dst []byte, src ref) (int, bool) @@ -392,13 +392,13 @@ const dst = loadSlice(sp + 8); const src = loadValue(sp + 32); if (!(src instanceof Uint8Array)) { - mem().setUint8(sp + 48, 0); + this.mem.setUint8(sp + 48, 0); return; } const toCopy = src.subarray(0, dst.length); dst.set(toCopy); setInt64(sp + 40, toCopy.length); - mem().setUint8(sp + 48, 1); + this.mem.setUint8(sp + 48, 1); }, // func copyBytesToJS(dst ref, src []byte) (int, bool) @@ -406,13 +406,13 @@ const dst = loadValue(sp + 8); const src = loadSlice(sp + 16); if (!(dst instanceof Uint8Array)) { - mem().setUint8(sp + 48, 0); + this.mem.setUint8(sp + 48, 0); return; } const toCopy = src.subarray(0, dst.length); dst.set(toCopy); setInt64(sp + 40, toCopy.length); - mem().setUint8(sp + 48, 1); + this.mem.setUint8(sp + 48, 1); }, "debug": (value) => { @@ -424,6 +424,7 @@ async run(instance) { this._inst = instance; + this.mem = new DataView(this._inst.exports.mem.buffer); this._values = [ // TODO: garbage collection NaN, 0, @@ -436,15 +437,13 @@ this._refs = new Map(); this.exited = false; - const mem = new DataView(this._inst.exports.mem.buffer) - // Pass command line arguments and environment variables to WebAssembly by writing them to the linear memory. let offset = 4096; const strPtr = (str) => { const ptr = offset; const bytes = encoder.encode(str + "\0"); - new Uint8Array(mem.buffer, offset, bytes.length).set(bytes); + new Uint8Array(this.mem.buffer, offset, bytes.length).set(bytes); offset += bytes.length; if (offset % 8 !== 0) { offset += 8 - (offset % 8); @@ -467,8 +466,8 @@ const argv = offset; argvPtrs.forEach((ptr) => { - mem.setUint32(offset, ptr, true); - mem.setUint32(offset + 4, 0, true); + this.mem.setUint32(offset, ptr, true); + this.mem.setUint32(offset + 4, 0, true); offset += 8; }); diff --git a/src/runtime/mem_js.go b/src/runtime/mem_js.go index de90f5305f..699a80db26 100644 --- a/src/runtime/mem_js.go +++ b/src/runtime/mem_js.go @@ -64,6 +64,7 @@ func sysReserve(v unsafe.Pointer, n uintptr) unsafe.Pointer { if growMemory(needed-current) == -1 { return nil } + resetMemoryDataView() } return v @@ -72,6 +73,10 @@ func sysReserve(v unsafe.Pointer, n uintptr) unsafe.Pointer { func currentMemory() int32 func growMemory(pages int32) int32 +// resetMemoryDataView signals the JS front-end that WebAssembly's memory.grow instruction has been used. +// This allows the front-end to replace the old DataView object with a new one. +func resetMemoryDataView() + func sysMap(v unsafe.Pointer, n uintptr, sysStat *uint64) { mSysStatInc(sysStat, n) } diff --git a/src/runtime/sys_wasm.s b/src/runtime/sys_wasm.s index d7bab926dc..823757814c 100644 --- a/src/runtime/sys_wasm.s +++ b/src/runtime/sys_wasm.s @@ -171,6 +171,10 @@ TEXT runtime·growMemory(SB), NOSPLIT, $0 I32Store ret+8(FP) RET +TEXT ·resetMemoryDataView(SB), NOSPLIT, $0 + CallImport + RET + TEXT ·wasmExit(SB), NOSPLIT, $0 CallImport RET diff --git a/src/syscall/js/js_test.go b/src/syscall/js/js_test.go index 7a1e346f55..753c2c3a0d 100644 --- a/src/syscall/js/js_test.go +++ b/src/syscall/js/js_test.go @@ -484,3 +484,26 @@ func TestCopyBytesToJS(t *testing.T) { }) } } + +// BenchmarkDOM is a simple benchmark which emulates a webapp making DOM operations. +// It creates a div, and sets its id. Then searches by that id and sets some data. +// Finally it removes that div. +func BenchmarkDOM(b *testing.B) { + document := js.Global().Get("document") + if document == js.Undefined() { + b.Skip("Not a browser environment. Skipping.") + } + const data = "someString" + for i := 0; i < b.N; i++ { + div := document.Call("createElement", "div") + div.Call("setAttribute", "id", "myDiv") + document.Get("body").Call("appendChild", div) + myDiv := document.Call("getElementById", "myDiv") + myDiv.Set("innerHTML", data) + + if got, want := myDiv.Get("innerHTML").String(), data; got != want { + b.Errorf("got %s, want %s", got, want) + } + document.Get("body").Call("removeChild", div) + } +}