AESKEYGENASSIST $0x36, X0, X1
	CALL _expand_key_128<>(SB)
Lexp_dec:
	// dec
	SUBQ $16, BX
	MOVUPS (BX), X1
	MOVUPS X1, (DX)
	DECQ CX
Lexp_dec_loop:
	MOVUPS -16(BX), X1
	AESIMC X1, X0
	MOVUPS X0, 16(DX)
	SUBQ $16, BX
	ADDQ $16, DX
	DECQ CX
	JNZ Lexp_dec_loop
	MOVUPS -16(BX), X0
	MOVUPS X0, 16(DX)
	RET

TEXT _expand_key_128<>(SB),NOSPLIT,$0
	PSHUFD $0xff, X1, X1

TEXT ·gcmAesInit(SB),NOSPLIT,$0
#define dst DI
#define KS SI
#define NR DX

	MOVQ productTable+0(FP), dst
	MOVQ ks_base+8(FP), KS
	MOVQ ks_len+16(FP), NR

	SHRQ $2, NR
	DECQ NR

	MOVOU bswapMask<>(SB), BSWAP
	MOVOU gcmPoly<>(SB), POLY

	// Encrypt block 0, with the AES key to generate the hash key H
	MOVOU (16*0)(KS), B0
	MOVOU (16*1)(KS), T0
	AESENC T0, B0

	CMOVQCS acc5, acc1
	CMOVQCS y_ptr, acc2
	CMOVQCS t1, acc3

	MOVQ acc0, (8*0)(res_ptr)
	MOVQ acc1, (8*1)(res_ptr)
	MOVQ acc2, (8*2)(res_ptr)
	MOVQ acc3, (8*3)(res_ptr)
	MOVQ res_ptr, x_ptr
	DECQ BX
	JNE  sqrLoop

	RET
/* ---------------------------------------*/
// func p256Mul(res, in1, in2 *p256Element)
TEXT ·p256Mul(SB),NOSPLIT,$0
	MOVQ res+0(FP), res_ptr
	MOVQ in1+8(FP), x_ptr

	atomic.AddInt64(&v.nowCount, 1)
	r.qCache.bucketStats[bucket] = v
	atomic.AddInt64(&r.qCache.srQueueStats.nowBytes, sz)
	atomic.AddInt64(&r.qCache.srQueueStats.nowCount, 1)
}

func (r *ReplicationStats) decQ(bucket string, sz int64, isDelMarker bool, opType replication.Type) {
	r.qCache.Lock()
	defer r.qCache.Unlock()
	v, ok := r.qCache.bucketStats[bucket]
	if !ok {
		v = newInQueueStats(r.registry, bucket)
	}

	ef -= arr[b]
	// 386:`SUBL\s[A-Z]+,\s\([A-Z]+\)\([A-Z]+\*4\)`
	arr[c] -= b
	// 386:`ADDL\s[$]-15,\s\([A-Z]+\)\([A-Z]+\*4\)`
	arr[d] -= 15
	// 386:`DECL\s\([A-Z]+\)\([A-Z]+\*4\)`
	arr[b]--
	// amd64:`DECQ\s64\([A-Z]+\)`
	arr[8]--
	// 386:"SUBL\t4"
	// amd64:"SUBQ\t8"
	return arr[0] - arr[1]
}

func SubFromConst(a int) int {
	// ppc64x: `SUBC\tR[0-9]+,\s[$]40,\sR`
	b := 40 - a
	return b
}

				atomic.AddInt32(&p.activeMRFWorkers, 1)
				replicateObject(p.ctx, v, p.objLayer)
				atomic.AddInt32(&p.activeMRFWorkers, -1)
				globalReplicationStats.decQ(v.Bucket, v.Size, v.DeleteMarker, v.OpType)

			default:
				bugLogIf(p.ctx, fmt.Errorf("unknown mrf replication type: %T", oi), "unknown-mrf-replicate-type")
			}
		case <-p.mrfWorkerKillCh:
			return

	// xor in seed
	PXOR	X0, X8
	PXOR	X1, X9
	PXOR	X2, X10
	PXOR	X3, X11
	PXOR	X4, X12
	PXOR	X5, X13
	PXOR	X6, X14
	PXOR	X7, X15

	// compute number of remaining 128-byte blocks
	DECQ	CX
	SHRQ	$7, CX

	PCALIGN $16
aesloop:
	// scramble state
	AESENC	X8, X8
	AESENC	X9, X9
	AESENC	X10, X10
	AESENC	X11, X11
	AESENC	X12, X12
	AESENC	X13, X13
	AESENC	X14, X14