crypto/md5: Optimize amd64 assembly

* Use two ADDL instead of LEAL
* Keep ones in R11
* Use XORL with lower latency instead of NOTL
* Remove loads and load the correct value in the previous round

```
name                    old time/op    new time/op    delta
Hash8Bytes-32              106ns ± 0%     103ns ± 0%  -2.37%  (p=0.000 n=10+10)
Hash1K-32                 1.33µs ± 0%    1.26µs ± 0%  -4.78%   (p=0.000 n=8+10)
Hash8K-32                 9.95µs ± 0%    9.46µs ± 0%  -4.90%  (p=0.000 n=10+10)
Hash8BytesUnaligned-32     106ns ± 0%     103ns ± 0%  -2.37%  (p=0.000 n=10+10)
Hash1KUnaligned-32        1.33µs ± 0%    1.26µs ± 0%  -4.76%   (p=0.000 n=10+9)
Hash8KUnaligned-32        10.0µs ± 0%     9.5µs ± 0%  -4.88%  (p=0.000 n=10+10)

name                    old speed      new speed      delta
Hash8Bytes-32           75.8MB/s ± 0%  77.8MB/s ± 0%  +2.70%  (p=0.000 n=10+10)
Hash1K-32                772MB/s ± 0%   810MB/s ± 0%  +4.99%   (p=0.000 n=9+10)
Hash8K-32                823MB/s ± 0%   866MB/s ± 0%  +5.15%  (p=0.000 n=10+10)
Hash8BytesUnaligned-32  75.8MB/s ± 0%  77.8MB/s ± 0%  +2.64%  (p=0.000 n=10+10)
Hash1KUnaligned-32       771MB/s ± 0%   810MB/s ± 0%  +4.96%  (p=0.000 n=10+10)
Hash8KUnaligned-32       823MB/s ± 0%   866MB/s ± 0%  +5.13%  (p=0.000 n=10+10)
```

Change-Id: I02b31229b857e9257dc9d36538883eb3af4ad993

Author: klauspost

src/crypto/md5/md5block_amd64.s

@@ -25,6 +25,7 @@ TEXT	·block(SB),NOSPLIT,$8-32
 	MOVL	(1*4)(BP),	BX
 	MOVL	(2*4)(BP),	CX
 	MOVL	(3*4)(BP),	DX
+	MOVL	$0xffffffff,	R11
 
 	CMPQ	SI,		DI
 	JEQ	end
@@ -40,7 +41,8 @@ loop:
 
 #define ROUND1(a, b, c, d, index, const, shift) \
 	XORL	c, R9; \
-	LEAL	const(a)(R8*1), a; \
+	ADDL	$const, a; \
+	ADDL	R8, a; \
 	ANDL	b, R9; \
 	XORL d, R9; \
 	MOVL (index*4)(SI), R8; \
@@ -64,15 +66,15 @@ loop:
 	ROUND1(AX,BX,CX,DX,13,0x6b901122, 7);
 	ROUND1(DX,AX,BX,CX,14,0xfd987193,12);
 	ROUND1(CX,DX,AX,BX,15,0xa679438e,17);
-	ROUND1(BX,CX,DX,AX, 0,0x49b40821,22);
+	ROUND1(BX,CX,DX,AX, 1,0x49b40821,22);
 
-	MOVL	(1*4)(SI),	R8
 	MOVL	DX,		R9
 	MOVL	DX,		R10
 
 #define ROUND2(a, b, c, d, index, const, shift) \
-	NOTL	R9; \
-	LEAL	const(a)(R8*1),a; \
+	XORL	R11, R9; \
+	ADDL	$const, a; \
+	ADDL	R8, a; \
 	ANDL	b,		R10; \
 	ANDL	c,		R9; \
 	MOVL	(index*4)(SI),R8; \
@@ -98,13 +100,13 @@ loop:
 	ROUND2(AX,BX,CX,DX, 2,0xa9e3e905, 5);
 	ROUND2(DX,AX,BX,CX, 7,0xfcefa3f8, 9);
 	ROUND2(CX,DX,AX,BX,12,0x676f02d9,14);
-	ROUND2(BX,CX,DX,AX, 0,0x8d2a4c8a,20);
+	ROUND2(BX,CX,DX,AX, 5,0x8d2a4c8a,20);
 
-	MOVL	(5*4)(SI),	R8
 	MOVL	CX,		R9
 
 #define ROUND3(a, b, c, d, index, const, shift) \
-	LEAL	const(a)(R8*1),a; \
+	ADDL	$const, a; \
+	ADDL	R8, a; \
 	MOVL	(index*4)(SI),R8; \
 	XORL	d,		R9; \
 	XORL	b,		R9; \
@@ -130,12 +132,12 @@ loop:
 	ROUND3(CX,DX,AX,BX, 2,0x1fa27cf8,16);
 	ROUND3(BX,CX,DX,AX, 0,0xc4ac5665,23);
 
-	MOVL	(0*4)(SI),	R8
-	MOVL	$0xffffffff,	R9
+	MOVL	R11,	R9
 	XORL	DX,		R9
 
 #define ROUND4(a, b, c, d, index, const, shift) \
-	LEAL	const(a)(R8*1),a; \
+	ADDL	$const, a; \
+	ADDL	R8, a; \
 	ORL	b,		R9; \
 	XORL	c,		R9; \
 	ADDL	R9,		a; \