rng: Add AVX2 + NEON asm.

Merge after minio#228 to include arm64 checks.

Before/after...

```
pkg: github.com/minio/pkg/v3/rng
cpu: AMD Ryzen 9 9950X 16-Core Processor
BenchmarkReader
BenchmarkReader/1000-32         	46546988	        25.88 ns/op	38635.30 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/1024-32         	70920727	        17.14 ns/op	59755.33 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/16384-32        	 5805674	       204.9 ns/op	79950.02 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/1048576-32      	   92539	     14080 ns/op	74470.24 MB/s	       0 B/op	       0 allocs/op

BenchmarkReader/1000-32         	52974752	        22.57 ns/op	44300.70 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/1024-32         	100000000	        11.37 ns/op	90096.95 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/16384-32        	14598060	        81.69 ns/op	200552.58 MB/s	       0 B/op	       0 allocs/op
BenchmarkReader/1048576-32      	  174301	      6384 ns/op	164256.53 MB/s	       0 B/op	       0 allocs/op
```

Author: klauspost

rng/reader_test.go

@@ -19,6 +19,7 @@ package rng
 
 import (
 	"bytes"
+	"encoding/binary"
 	"io"
 	"math/rand"
 	"strconv"
@@ -188,3 +189,114 @@ func TestXor(t *testing.T) {
 		}
 	}
 }
+
+func TestXorZeroKey(t *testing.T) {
+	var keys [4]uint64
+	for _, size := range []int{0, 32, 64, 96, 128, 1024} {
+		in := make([]byte, size)
+		for i := range in {
+			in[i] = byte(i)
+		}
+		out := make([]byte, size)
+		xorSlice(in, out, &keys)
+		if !bytes.Equal(in, out) {
+			t.Fatalf("size %d: zero-key xor should copy input\nexpected %x\ngot      %x", size, in, out)
+		}
+		out2 := make([]byte, size)
+		xor32Go(in, out2, &keys)
+		if !bytes.Equal(in, out2) {
+			t.Fatalf("size %d: zero-key xor32Go should copy input", size)
+		}
+	}
+}
+
+func TestXorDoubleApply(t *testing.T) {
+	rng := rand.New(rand.NewSource(42))
+	var keys [4]uint64
+	for i := range keys {
+		keys[i] = rng.Uint64()
+	}
+	for _, size := range []int{32, 64, 96, 128, 256, 1024, 4096} {
+		in := make([]byte, size)
+		_, _ = io.ReadFull(rng, in)
+		orig := make([]byte, size)
+		copy(orig, in)
+
+		tmp := make([]byte, size)
+		out := make([]byte, size)
+		xorSlice(in, tmp, &keys)
+		xorSlice(tmp, out, &keys)
+		if !bytes.Equal(orig, out) {
+			t.Fatalf("size %d: double xor should return original\nexpected %x\ngot      %x", size, orig[:32], out[:32])
+		}
+	}
+}
+
+func TestXorAllSizes(t *testing.T) {
+	rng := rand.New(rand.NewSource(99))
+	var keys [4]uint64
+	for i := range keys {
+		keys[i] = rng.Uint64()
+	}
+	in := make([]byte, 8192)
+	_, _ = io.ReadFull(rng, in)
+
+	// Every multiple of 32 up to 8192 exercises different loop paths:
+	// 0 (empty), 32 (single block), 64 (two blocks / 64-byte loop only),
+	// 96 (64-byte loop + 32-byte tail), etc.
+	for size := 0; size <= len(in); size += 32 {
+		outAsm := make([]byte, size)
+		outGo := make([]byte, size)
+		xorSlice(in[:size], outAsm, &keys)
+		xor32Go(in[:size], outGo, &keys)
+		if !bytes.Equal(outAsm, outGo) {
+			t.Fatalf("size %d: asm and Go disagree\nasm %x\ngo  %x", size, outAsm[:min(64, size)], outGo[:min(64, size)])
+		}
+	}
+}
+
+func TestXorDistinctKeys(t *testing.T) {
+	in := make([]byte, 256)
+	for i := range in {
+		in[i] = byte(i)
+	}
+	keys1 := [4]uint64{1, 2, 3, 4}
+	keys2 := [4]uint64{5, 6, 7, 8}
+	out1 := make([]byte, 256)
+	out2 := make([]byte, 256)
+	xorSlice(in, out1, &keys1)
+	xorSlice(in, out2, &keys2)
+	if bytes.Equal(out1, out2) {
+		t.Fatal("different keys should produce different output")
+	}
+}
+
+func TestXorKnownValues(t *testing.T) {
+	in := make([]byte, 32)
+	for i := range in {
+		in[i] = byte(i)
+	}
+	keys := [4]uint64{0x0807060504030201, 0x100f0e0d0c0b0a09, 0x1817161514131211, 0x201f1e1d1c1b1a19}
+	out := make([]byte, 32)
+	xor32Go(in, out, &keys)
+
+	// Each 8-byte block: out[i..i+7] = in[i..i+7] XOR keys[i/8]
+	// Block 0: in[0..7] = 00 01 02 03 04 05 06 07, key = 01 02 03 04 05 06 07 08 (LE)
+	//   XOR => 01 03 01 07 01 03 01 0f
+	expected := make([]byte, 32)
+	for i := 0; i < 32; i++ {
+		keyBytes := make([]byte, 8)
+		binary.LittleEndian.PutUint64(keyBytes, keys[i/8])
+		expected[i] = in[i] ^ keyBytes[i%8]
+	}
+	if !bytes.Equal(out, expected) {
+		t.Fatalf("known values mismatch\nexpected %x\ngot      %x", expected, out)
+	}
+
+	// Now verify xorSlice matches
+	outAsm := make([]byte, 32)
+	xorSlice(in, outAsm, &keys)
+	if !bytes.Equal(outAsm, expected) {
+		t.Fatalf("xorSlice known values mismatch\nexpected %x\ngot      %x", expected, outAsm)
+	}
+}

rng/xor_amd64.go

@@ -19,5 +19,18 @@
 
 package rng
 
+import "github.com/klauspost/cpuid/v2"
+
+func xorSlice(in, out []byte, v *[4]uint64) {
+	if cpuid.CPU.Has(cpuid.AVX2) {
+		xorSliceAvx2(in, out, v)
+	} else {
+		xorSliceSSE2(in, out, v)
+	}
+}
+
+//go:noescape
+func xorSliceSSE2(in, out []byte, v *[4]uint64)
+
 //go:noescape
-func xorSlice(in, out []byte, v *[4]uint64)
+func xorSliceAvx2(in, out []byte, v *[4]uint64)

rng/xor_amd64.s

@@ -19,15 +19,15 @@
 //+build !appengine
 //+build !gccgo
 
-// func xorSlice(in, out []byte, v *[4]uint64)
-TEXT ·xorSlice(SB), 7, $0
-	MOVQ  v+48(FP), AX         // AX: v
-	MOVQ  in+0(FP), SI         // SI: &in
-	MOVQ  out+24(FP), DX       // DX: &out
-	MOVQ  out_len+32(FP), R9   // R9: len(out)
-	MOVOU (AX), X0             // v[x]
-	MOVOU 16(AX), X1           // v[x+2]
-	SHRQ  $5, R9               // len(in) / 32
+// func xorSliceSSE2(in, out []byte, v *[4]uint64)
+TEXT ·xorSliceSSE2(SB), 7, $0
+	MOVQ  v+48(FP), AX       // AX: v
+	MOVQ  in+0(FP), SI       // SI: &in
+	MOVQ  out+24(FP), DX     // DX: &out
+	MOVQ  out_len+32(FP), R9 // R9: len(out)
+	MOVOU (AX), X0           // v[x]
+	MOVOU 16(AX), X1         // v[x+2]
+	SHRQ  $5, R9             // len(in) / 32
 	JZ    done_xor_sse2_32
 
 loopback_xor_sse2_32:
@@ -44,3 +44,40 @@ loopback_xor_sse2_32:
 
 done_xor_sse2_32:
 	RET
+
+// func xorSliceAvx2(in, out []byte, v *[4]uint64)
+TEXT ·xorSliceAvx2(SB), 7, $0
+	MOVQ    v+48(FP), AX         // AX: v
+	MOVQ    in+0(FP), SI         // SI: &in
+	MOVQ    out+24(FP), DX       // DX: &out
+	MOVQ    out_len+32(FP), R9   // R9: len(out)
+	VMOVDQU (AX), Y0             // v[x]
+	SHRQ    $5, R9               // len(in) / 32
+	CMPQ    R9, $1
+	JEQ     loopback_xor_avx2_32
+	JB      end_xor_avx2_32
+
+loopback_xor_avx2_64:
+	SUBQ    $2, R9
+	VMOVDQU (SI), Y1
+	VMOVDQU 32(SI), Y2
+	VPXOR   Y1, Y0, Y1
+	VPXOR   Y2, Y0, Y2
+	VMOVDQU Y1, (DX)
+	VMOVDQU Y2, 32(DX)
+	ADDQ    $64, SI              // in+=64
+	ADDQ    $64, DX              // out+=64
+	CMPQ    R9, $1
+	JA      loopback_xor_avx2_64
+	JEQ     loopback_xor_avx2_32
+	JMP     end_xor_avx2_32
+
+loopback_xor_avx2_32:
+	VPXOR   (SI), Y0, Y1
+	VMOVDQU Y1, (DX)
+	ADDQ    $32, SI      // in+=32
+	ADDQ    $32, DX      // out+=32
+
+end_xor_avx2_32:
+	VZEROUPPER
+	RET

rng/xor_arm64.go

@@ -0,0 +1,27 @@
+// Copyright (c) 2015-2026 MinIO, Inc.
+//
+// This file is part of MinIO Object Storage stack
+//
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+//go:build !noasm && !appengine && !gccgo
+
+package rng
+
+func xorSlice(in, out []byte, v *[4]uint64) {
+	xorSliceNEON(in, out, v)
+}
+
+//go:noescape
+func xorSliceNEON(in, out []byte, v *[4]uint64)

rng/xor_arm64.s

@@ -0,0 +1,56 @@
+// Copyright (c) 2015-2026 MinIO, Inc.
+//
+// This file is part of MinIO Object Storage stack
+//
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program.  If not, see <http://www.gnu.org/licenses/>.
+
+//+build !noasm
+//+build !appengine
+//+build !gccgo
+
+// func xorSliceNEON(in, out []byte, v *[4]uint64)
+TEXT ·xorSliceNEON(SB), 7, $0
+	MOVD v+48(FP), R2
+	MOVD in+0(FP), R0
+	MOVD out+24(FP), R1
+	MOVD out_len+32(FP), R3
+	VLD1 (R2), [V0.B16, V1.B16]
+	LSR  $5, R3, R3
+	CBZ  R3, done_xor_neon
+	CMP  $1, R3
+	BEQ  loopback_xor_neon_32
+
+loopback_xor_neon_64:
+	SUB  $2, R3, R3
+	VLD1 (R0), [V2.B16, V3.B16, V4.B16, V5.B16]
+	VEOR V0.B16, V2.B16, V2.B16
+	VEOR V1.B16, V3.B16, V3.B16
+	VEOR V0.B16, V4.B16, V4.B16
+	VEOR V1.B16, V5.B16, V5.B16
+	VST1 [V2.B16, V3.B16, V4.B16, V5.B16], (R1)
+	ADD  $64, R0
+	ADD  $64, R1
+	CMP  $1, R3
+	BGT  loopback_xor_neon_64
+	BEQ  loopback_xor_neon_32
+	B    done_xor_neon
+
+loopback_xor_neon_32:
+	VLD1 (R0), [V2.B16, V3.B16]
+	VEOR V0.B16, V2.B16, V2.B16
+	VEOR V1.B16, V3.B16, V3.B16
+	VST1 [V2.B16, V3.B16], (R1)
+
+done_xor_neon:
+	RET

rng/xor_noasm.go

@@ -1,4 +1,4 @@
-//go:build !amd64 || noasm || appengine || gccgo
+//go:build (!amd64 && !arm64) || noasm || appengine || gccgo
 
 // Copyright (c) 2015-2021 MinIO, Inc.
 //