[release-branch.go1.5] runtime: adjust the arm64 memmove and memclr to operate by word as much as they can

Not only is this an obvious optimization:

benchmark                           old MB/s     new MB/s     speedup
BenchmarkMemmove1-4                 35.35        29.65        0.84x
BenchmarkMemmove2-4                 63.78        52.53        0.82x
BenchmarkMemmove3-4                 89.72        73.96        0.82x
BenchmarkMemmove4-4                 109.94       95.73        0.87x
BenchmarkMemmove5-4                 127.60       112.80       0.88x
BenchmarkMemmove6-4                 143.59       126.67       0.88x
BenchmarkMemmove7-4                 157.90       138.92       0.88x
BenchmarkMemmove8-4                 167.18       231.81       1.39x
BenchmarkMemmove9-4                 175.23       252.07       1.44x
BenchmarkMemmove10-4                165.68       261.10       1.58x
BenchmarkMemmove11-4                174.43       263.31       1.51x
BenchmarkMemmove12-4                180.76       267.56       1.48x
BenchmarkMemmove13-4                189.06       284.93       1.51x
BenchmarkMemmove14-4                186.31       284.72       1.53x
BenchmarkMemmove15-4                195.75       281.62       1.44x
BenchmarkMemmove16-4                202.96       439.23       2.16x
BenchmarkMemmove32-4                264.77       775.77       2.93x
BenchmarkMemmove64-4                306.81       1209.64      3.94x
BenchmarkMemmove128-4               357.03       1515.41      4.24x
BenchmarkMemmove256-4               380.77       2066.01      5.43x
BenchmarkMemmove512-4               385.05       2556.45      6.64x
BenchmarkMemmove1024-4              381.23       2804.10      7.36x
BenchmarkMemmove2048-4              379.06       2814.83      7.43x
BenchmarkMemmove4096-4              387.43       3064.96      7.91x
BenchmarkMemmoveUnaligned1-4        28.91        25.40        0.88x
BenchmarkMemmoveUnaligned2-4        56.13        47.56        0.85x
BenchmarkMemmoveUnaligned3-4        74.32        69.31        0.93x
BenchmarkMemmoveUnaligned4-4        97.02        83.58        0.86x
BenchmarkMemmoveUnaligned5-4        110.17       103.62       0.94x
BenchmarkMemmoveUnaligned6-4        124.95       113.26       0.91x
BenchmarkMemmoveUnaligned7-4        142.37       130.82       0.92x
BenchmarkMemmoveUnaligned8-4        151.20       205.64       1.36x
BenchmarkMemmoveUnaligned9-4        166.97       215.42       1.29x
BenchmarkMemmoveUnaligned10-4       148.49       221.22       1.49x
BenchmarkMemmoveUnaligned11-4       159.47       239.57       1.50x
BenchmarkMemmoveUnaligned12-4       163.52       247.32       1.51x
BenchmarkMemmoveUnaligned13-4       167.55       256.54       1.53x
BenchmarkMemmoveUnaligned14-4       175.12       251.03       1.43x
BenchmarkMemmoveUnaligned15-4       192.10       267.13       1.39x
BenchmarkMemmoveUnaligned16-4       190.76       378.87       1.99x
BenchmarkMemmoveUnaligned32-4       259.02       562.98       2.17x
BenchmarkMemmoveUnaligned64-4       317.72       842.44       2.65x
BenchmarkMemmoveUnaligned128-4      355.43       1274.49      3.59x
BenchmarkMemmoveUnaligned256-4      378.17       1815.74      4.80x
BenchmarkMemmoveUnaligned512-4      362.15       2180.81      6.02x
BenchmarkMemmoveUnaligned1024-4     376.07       2453.58      6.52x
BenchmarkMemmoveUnaligned2048-4     381.66       2568.32      6.73x
BenchmarkMemmoveUnaligned4096-4     398.51       2669.36      6.70x
BenchmarkMemclr5-4                  113.83       107.93       0.95x
BenchmarkMemclr16-4                 223.84       389.63       1.74x
BenchmarkMemclr64-4                 421.99       1209.58      2.87x
BenchmarkMemclr256-4                525.94       2411.58      4.59x
BenchmarkMemclr4096-4               581.66       4372.20      7.52x
BenchmarkMemclr65536-4              565.84       4747.48      8.39x
BenchmarkGoMemclr5-4                194.63       160.31       0.82x
BenchmarkGoMemclr16-4               295.30       630.07       2.13x
BenchmarkGoMemclr64-4               480.24       1884.03      3.92x
BenchmarkGoMemclr256-4              540.23       2926.49      5.42x

but it turns out that it's necessary to avoid the GC seeing partially written
pointers.

It's of course possible to be more sophisticated (using ldp/stp to move 16
bytes at a time in the core loop and unrolling the tail copying loops being
the obvious ideas) but I wanted something simple and (reasonably) obviously
correct.

Fixes #12552

Change-Id: Iaeaf8a812cd06f4747ba2f792de1ded738890735
Reviewed-on: https://go-review.googlesource.com/14813
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-on: https://go-review.googlesource.com/16909
Reviewed-by: Russ Cox <rsc@golang.org>

3 files changed, 127 insertions, 18 deletions

diff --git a/src/runtime/memclr_arm64.s b/src/runtime/memclr_arm64.s
index c44c1239f5..47c6b73c84 100644
--- a/src/runtime/memclr_arm64.s
+++ b/src/runtime/memclr_arm64.s
@@ -8,11 +8,30 @@
 TEXT runtime·memclr(SB),NOSPLIT,$0-16
 	MOVD	ptr+0(FP), R3
 	MOVD	n+8(FP), R4
-	CMP	$0, R4
-	BEQ	done
-	ADD	R3, R4, R4
+	// TODO(mwhudson): this is written this way to avoid tickling
+	// warnings from addpool when written as AND $7, R4, R6 (see
+	// https://golang.org/issue/12708)
+	AND	$~7, R4, R5	// R5 is N&~7
+	SUB	R5, R4, R6	// R6 is N&7
+
+	CMP	$0, R5
+	BEQ	nowords
+
+	ADD	R3, R5, R5
+
+wordloop: // TODO: Optimize for unaligned ptr.
+	MOVD.P	$0, 8(R3)
+	CMP	R3, R5
+	BNE	wordloop
+nowords:
+        CMP	$0, R6
+        BEQ	done
+
+	ADD	R3, R6, R6
+
+byteloop:
 	MOVBU.P	$0, 1(R3)
-	CMP	R3, R4
-	BNE	-2(PC)
+	CMP	R3, R6
+	BNE	byteloop
 done:
 	RET
diff --git a/src/runtime/memmove_arm64.s b/src/runtime/memmove_arm64.s
index 66059a75de..00813d4ef9 100644
--- a/src/runtime/memmove_arm64.s
+++ b/src/runtime/memmove_arm64.s
@@ -14,23 +14,78 @@ TEXT runtime·memmove(SB), NOSPLIT, $-8-24
 	RET
 
 check:
+	AND	$~7, R5, R7	// R7 is N&~7
+	// TODO(mwhudson): this is written this way to avoid tickling
+	// warnings from addpool when written as AND $7, R5, R6 (see
+	// https://golang.org/issue/12708)
+	SUB	R7, R5, R6	// R6 is N&7
+
 	CMP	R3, R4
 	BLT	backward
 
-	ADD	R3, R5
-loop:
-	MOVBU.P	1(R4), R6
-	MOVBU.P	R6, 1(R3)
-	CMP	R3, R5
-	BNE	loop
+	// Copying forward proceeds by copying R7/8 words then copying R6 bytes.
+	// R3 and R4 are advanced as we copy.
+
+        // (There may be implementations of armv8 where copying by bytes until
+        // at least one of source or dest is word aligned is a worthwhile
+        // optimization, but the on the one tested so far (xgene) it did not
+        // make a significance difference.)
+
+	CMP	$0, R7		// Do we need to do any word-by-word copying?
+	BEQ	noforwardlarge
+
+	ADD	R3, R7, R9	// R9 points just past where we copy by word
+
+forwardlargeloop:
+	MOVD.P	8(R4), R8	// R8 is just a scratch register
+	MOVD.P	R8, 8(R3)
+	CMP	R3, R9
+	BNE	forwardlargeloop
+
+noforwardlarge:
+	CMP	$0, R6		// Do we need to do any byte-by-byte copying?
+	BNE	forwardtail
+	RET
+
+forwardtail:
+	ADD	R3, R6, R9	// R9 points just past the destination memory
+
+forwardtailloop:
+	MOVBU.P 1(R4), R8
+	MOVBU.P	R8, 1(R3)
+	CMP	R3, R9
+	BNE	forwardtailloop
 	RET
 
 backward:
-	ADD	R5, R4
-	ADD	R3, R5
-loop1:
-	MOVBU.W	-1(R4), R6
-	MOVBU.W	R6, -1(R5)
-	CMP	R3, R5
-	BNE	loop1
+	// Copying backwards proceeds by copying R6 bytes then copying R7/8 words.
+	// R3 and R4 are advanced to the end of the destination/source buffers
+	// respectively and moved back as we copy.
+
+	ADD	R4, R5, R4	// R4 points just past the last source byte
+	ADD	R3, R5, R3	// R3 points just past the last destination byte
+
+	CMP	$0, R6		// Do we need to do any byte-by-byte copying?
+	BEQ	nobackwardtail
+
+	SUB	R6, R3, R9	// R9 points at the lowest destination byte that should be copied by byte.
+backwardtailloop:
+	MOVBU.W	-1(R4), R8
+	MOVBU.W	R8, -1(R3)
+	CMP	R9, R3
+	BNE	backwardtailloop
+
+nobackwardtail:
+	CMP     $0, R7		// Do we need to do any word-by-word copying?
+	BNE	backwardlarge
+	RET
+
+backwardlarge:
+        SUB	R7, R3, R9      // R9 points at the lowest destination byte
+
+backwardlargeloop:
+	MOVD.W	-8(R4), R8
+	MOVD.W	R8, -8(R3)
+	CMP	R9, R3
+	BNE	backwardlargeloop
 	RET
diff --git a/src/runtime/memmove_test.go b/src/runtime/memmove_test.go
index 857f99bc4c..d5a2ad8372 100644
--- a/src/runtime/memmove_test.go
+++ b/src/runtime/memmove_test.go
@@ -116,6 +116,41 @@ func BenchmarkMemmove1024(b *testing.B) { bmMemmove(b, 1024) }
 func BenchmarkMemmove2048(b *testing.B) { bmMemmove(b, 2048) }
 func BenchmarkMemmove4096(b *testing.B) { bmMemmove(b, 4096) }
 
+func bmMemmoveUnaligned(b *testing.B, n int) {
+	x := make([]byte, n+1)
+	y := make([]byte, n)
+	b.SetBytes(int64(n))
+	for i := 0; i < b.N; i++ {
+		copy(x[1:], y)
+	}
+}
+
+func BenchmarkMemmoveUnaligned0(b *testing.B)    { bmMemmoveUnaligned(b, 0) }
+func BenchmarkMemmoveUnaligned1(b *testing.B)    { bmMemmoveUnaligned(b, 1) }
+func BenchmarkMemmoveUnaligned2(b *testing.B)    { bmMemmoveUnaligned(b, 2) }
+func BenchmarkMemmoveUnaligned3(b *testing.B)    { bmMemmoveUnaligned(b, 3) }
+func BenchmarkMemmoveUnaligned4(b *testing.B)    { bmMemmoveUnaligned(b, 4) }
+func BenchmarkMemmoveUnaligned5(b *testing.B)    { bmMemmoveUnaligned(b, 5) }
+func BenchmarkMemmoveUnaligned6(b *testing.B)    { bmMemmoveUnaligned(b, 6) }
+func BenchmarkMemmoveUnaligned7(b *testing.B)    { bmMemmoveUnaligned(b, 7) }
+func BenchmarkMemmoveUnaligned8(b *testing.B)    { bmMemmoveUnaligned(b, 8) }
+func BenchmarkMemmoveUnaligned9(b *testing.B)    { bmMemmoveUnaligned(b, 9) }
+func BenchmarkMemmoveUnaligned10(b *testing.B)   { bmMemmoveUnaligned(b, 10) }
+func BenchmarkMemmoveUnaligned11(b *testing.B)   { bmMemmoveUnaligned(b, 11) }
+func BenchmarkMemmoveUnaligned12(b *testing.B)   { bmMemmoveUnaligned(b, 12) }
+func BenchmarkMemmoveUnaligned13(b *testing.B)   { bmMemmoveUnaligned(b, 13) }
+func BenchmarkMemmoveUnaligned14(b *testing.B)   { bmMemmoveUnaligned(b, 14) }
+func BenchmarkMemmoveUnaligned15(b *testing.B)   { bmMemmoveUnaligned(b, 15) }
+func BenchmarkMemmoveUnaligned16(b *testing.B)   { bmMemmoveUnaligned(b, 16) }
+func BenchmarkMemmoveUnaligned32(b *testing.B)   { bmMemmoveUnaligned(b, 32) }
+func BenchmarkMemmoveUnaligned64(b *testing.B)   { bmMemmoveUnaligned(b, 64) }
+func BenchmarkMemmoveUnaligned128(b *testing.B)  { bmMemmoveUnaligned(b, 128) }
+func BenchmarkMemmoveUnaligned256(b *testing.B)  { bmMemmoveUnaligned(b, 256) }
+func BenchmarkMemmoveUnaligned512(b *testing.B)  { bmMemmoveUnaligned(b, 512) }
+func BenchmarkMemmoveUnaligned1024(b *testing.B) { bmMemmoveUnaligned(b, 1024) }
+func BenchmarkMemmoveUnaligned2048(b *testing.B) { bmMemmoveUnaligned(b, 2048) }
+func BenchmarkMemmoveUnaligned4096(b *testing.B) { bmMemmoveUnaligned(b, 4096) }
+
 func TestMemclr(t *testing.T) {
 	size := 512
 	if testing.Short() {