runtime: improve arm64 memclr implementation

Improve runtime memclr_arm64.s using ZVA feature to zero out memory when n
is at least 64 bytes.

Also add DCZID_EL0 system register to use in MRS instruction.

    Benchmark results of runtime/Memclr on Amberwing:
name          old time/op    new time/op    delta
Memclr/5        12.7ns ± 0%    12.7ns ± 0%      ~     (all equal)
Memclr/16       12.7ns ± 0%    12.2ns ± 1%    -4.13%  (p=0.000 n=7+8)
Memclr/64       14.0ns ± 0%    14.6ns ± 1%    +4.29%  (p=0.000 n=7+8)
Memclr/256      23.7ns ± 0%    25.7ns ± 0%    +8.44%  (p=0.000 n=8+7)
Memclr/4096      204ns ± 0%      74ns ± 0%   -63.71%  (p=0.000 n=8+8)
Memclr/65536    2.89µs ± 0%    0.84µs ± 0%   -70.91%  (p=0.000 n=8+8)
Memclr/1M       45.9µs ± 0%    17.0µs ± 0%   -62.88%  (p=0.000 n=8+8)
Memclr/4M        184µs ± 0%      77µs ± 4%   -57.94%  (p=0.001 n=6+8)
Memclr/8M        367µs ± 0%     144µs ± 1%   -60.72%  (p=0.000 n=7+8)
Memclr/16M       734µs ± 0%     293µs ± 1%   -60.09%  (p=0.000 n=8+8)
Memclr/64M      2.94ms ± 0%    1.23ms ± 0%   -58.06%  (p=0.000 n=7+8)
GoMemclr/5      8.00ns ± 0%    8.79ns ± 0%    +9.83%  (p=0.000 n=8+8)
GoMemclr/16     8.00ns ± 0%    7.60ns ± 0%    -5.00%  (p=0.000 n=8+8)
GoMemclr/64     10.8ns ± 0%    10.4ns ± 0%    -3.70%  (p=0.000 n=8+8)
GoMemclr/256    20.4ns ± 0%    21.2ns ± 0%    +3.92%  (p=0.000 n=8+8)

name          old speed      new speed      delta
Memclr/5       394MB/s ± 0%   393MB/s ± 0%    -0.28%  (p=0.006 n=8+8)
Memclr/16     1.26GB/s ± 0%  1.31GB/s ± 1%    +4.07%  (p=0.000 n=7+8)
Memclr/64     4.57GB/s ± 0%  4.39GB/s ± 2%    -3.91%  (p=0.000 n=7+8)
Memclr/256    10.8GB/s ± 0%  10.0GB/s ± 0%    -7.95%  (p=0.001 n=7+6)
Memclr/4096   20.1GB/s ± 0%  55.3GB/s ± 0%  +175.46%  (p=0.000 n=8+8)
Memclr/65536  22.6GB/s ± 0%  77.8GB/s ± 0%  +243.63%  (p=0.000 n=7+8)
Memclr/1M     22.8GB/s ± 0%  61.5GB/s ± 0%  +169.38%  (p=0.000 n=8+8)
Memclr/4M     22.8GB/s ± 0%  54.3GB/s ± 4%  +137.85%  (p=0.001 n=6+8)
Memclr/8M     22.8GB/s ± 0%  58.1GB/s ± 1%  +154.56%  (p=0.000 n=7+8)
Memclr/16M    22.8GB/s ± 0%  57.2GB/s ± 1%  +150.54%  (p=0.000 n=8+8)
Memclr/64M    22.8GB/s ± 0%  54.4GB/s ± 0%  +138.42%  (p=0.000 n=7+8)
GoMemclr/5     625MB/s ± 0%   569MB/s ± 0%    -8.90%  (p=0.000 n=7+8)
GoMemclr/16   2.00GB/s ± 0%  2.10GB/s ± 0%    +5.26%  (p=0.000 n=8+8)
GoMemclr/64   5.92GB/s ± 0%  6.15GB/s ± 0%    +3.83%  (p=0.000 n=7+8)
GoMemclr/256  12.5GB/s ± 0%  12.1GB/s ± 0%    -3.77%  (p=0.000 n=8+7)

    Benchmark results of runtime/Memclr on Amberwing without ZVA:
name          old time/op    new time/op    delta
Memclr/5        12.7ns ± 0%    12.8ns ± 0%   +0.79%  (p=0.008 n=5+5)
Memclr/16       12.7ns ± 0%    12.7ns ± 0%     ~     (p=0.444 n=5+5)
Memclr/64       14.0ns ± 0%    14.4ns ± 0%   +2.86%  (p=0.008 n=5+5)
Memclr/256      23.7ns ± 1%    19.2ns ± 0%  -19.06%  (p=0.008 n=5+5)
Memclr/4096      203ns ± 0%     119ns ± 0%  -41.38%  (p=0.008 n=5+5)
Memclr/65536    2.89µs ± 0%    1.66µs ± 0%  -42.76%  (p=0.008 n=5+5)
Memclr/1M       45.9µs ± 0%    26.2µs ± 0%  -42.82%  (p=0.008 n=5+5)
Memclr/4M        184µs ± 0%     105µs ± 0%  -42.81%  (p=0.008 n=5+5)
Memclr/8M        367µs ± 0%     210µs ± 0%  -42.76%  (p=0.008 n=5+5)
Memclr/16M       734µs ± 0%     420µs ± 0%  -42.74%  (p=0.008 n=5+5)
Memclr/64M      2.94ms ± 0%    1.69ms ± 0%  -42.46%  (p=0.008 n=5+5)
GoMemclr/5      8.00ns ± 0%    8.40ns ± 0%   +5.00%  (p=0.008 n=5+5)
GoMemclr/16     8.00ns ± 0%    8.40ns ± 0%   +5.00%  (p=0.008 n=5+5)
GoMemclr/64     10.8ns ± 0%     9.6ns ± 0%  -11.02%  (p=0.008 n=5+5)
GoMemclr/256    20.4ns ± 0%    17.2ns ± 0%  -15.69%  (p=0.008 n=5+5)

name          old speed      new speed      delta
Memclr/5       393MB/s ± 0%   391MB/s ± 0%   -0.64%  (p=0.008 n=5+5)
Memclr/16     1.26GB/s ± 0%  1.26GB/s ± 0%   -0.55%  (p=0.008 n=5+5)
Memclr/64     4.57GB/s ± 0%  4.44GB/s ± 0%   -2.79%  (p=0.008 n=5+5)
Memclr/256    10.8GB/s ± 0%  13.3GB/s ± 0%  +23.07%  (p=0.016 n=4+5)
Memclr/4096   20.1GB/s ± 0%  34.3GB/s ± 0%  +70.91%  (p=0.008 n=5+5)
Memclr/65536  22.7GB/s ± 0%  39.6GB/s ± 0%  +74.65%  (p=0.008 n=5+5)
Memclr/1M     22.8GB/s ± 0%  40.0GB/s ± 0%  +74.88%  (p=0.008 n=5+5)
Memclr/4M     22.8GB/s ± 0%  39.9GB/s ± 0%  +74.84%  (p=0.008 n=5+5)
Memclr/8M     22.9GB/s ± 0%  39.9GB/s ± 0%  +74.71%  (p=0.008 n=5+5)
Memclr/16M    22.9GB/s ± 0%  39.9GB/s ± 0%  +74.64%  (p=0.008 n=5+5)
Memclr/64M    22.8GB/s ± 0%  39.7GB/s ± 0%  +73.79%  (p=0.008 n=5+5)
GoMemclr/5     625MB/s ± 0%   595MB/s ± 0%   -4.77%  (p=0.000 n=4+5)
GoMemclr/16   2.00GB/s ± 0%  1.90GB/s ± 0%   -4.77%  (p=0.008 n=5+5)
GoMemclr/64   5.92GB/s ± 0%  6.66GB/s ± 0%  +12.48%  (p=0.016 n=4+5)
GoMemclr/256  12.5GB/s ± 0%  14.9GB/s ± 0%  +18.95%  (p=0.008 n=5+5)

Fixes #22948

Change-Id: Iaae4e22391e25b54d299821bb7f8a81ac3986b93
Reviewed-on: https://go-review.googlesource.com/82055
Run-TryBot: Brad Fitzpatrick <bradfitz@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

1 files changed, 161 insertions, 38 deletions

diff --git a/src/runtime/memclr_arm64.s b/src/runtime/memclr_arm64.s
index bf954e047f..159cac2486 100644
--- a/src/runtime/memclr_arm64.s
+++ b/src/runtime/memclr_arm64.s
@@ -8,52 +8,175 @@
 TEXT runtime·memclrNoHeapPointers(SB),NOSPLIT,$0-16
 	MOVD	ptr+0(FP), R0
 	MOVD	n+8(FP), R1
-	// If size is less than 16 bytes, use tail_zero to zero what remains
+
 	CMP	$16, R1
-	BLT	tail_zero
-	// Get buffer offset into 16 byte aligned address for better performance
-	ANDS	$15, R0, ZR
-	BNE	unaligned_to_16
-aligned_to_16:
-	LSR	$4, R1, R2
+	// If n is equal to 16 bytes, use zero_exact_16 to zero
+	BEQ	zero_exact_16
+
+	// If n is greater than 16 bytes, use zero_by_16 to zero
+	BHI	zero_by_16
+
+	// n is less than 16 bytes
+	ADD	R1, R0, R7
+	TBZ	$3, R1, less_than_8
+	MOVD	ZR, (R0)
+	MOVD	ZR, -8(R7)
+	RET
+
+less_than_8:
+	TBZ	$2, R1, less_than_4
+	MOVW	ZR, (R0)
+	MOVW	ZR, -4(R7)
+	RET
+
+less_than_4:
+	CBZ	R1, ending
+	MOVB	ZR, (R0)
+	TBZ	$1, R1, ending
+	MOVH	ZR, -2(R7)
+
+ending:
+	RET
+
+zero_exact_16:
+	// n is exactly 16 bytes
+	STP	(ZR, ZR), (R0)
+	RET
+
 zero_by_16:
-	STP.P	(ZR, ZR), 16(R0)
-	SUBS	$1, R2, R2
-	BNE	zero_by_16
+	// n greater than 16 bytes, check if the start address is aligned
+	NEG	R0, R4
+	ANDS	$15, R4, R4
+	// Try zeroing using zva if the start address is aligned with 16
+	BEQ	try_zva
+
+	// Non-aligned store
+	STP	(ZR, ZR), (R0)
+	// Make the destination aligned
+	SUB	R4, R1, R1
+	ADD	R4, R0, R0
+	B	try_zva
+
+tail_maybe_long:
+	CMP	$64, R1
+	BHS	no_zva
 
+tail63:
+	ANDS	$48, R1, R3
+	BEQ	last16
+	CMPW	$32, R3
+	BEQ	last48
+	BLT	last32
+	STP.P	(ZR, ZR), 16(R0)
+last48:
+	STP.P	(ZR, ZR), 16(R0)
+last32:
+	STP.P	(ZR, ZR), 16(R0)
+	// The last store length is at most 16, so it is safe to use
+	// stp to write last 16 bytes
+last16:
 	ANDS	$15, R1, R1
-	BEQ	ending
+	CBZ	R1, last_end
+	ADD	R1, R0, R0
+	STP	(ZR, ZR), -16(R0)
+last_end:
+	RET
 
-	// Zero buffer with size=R1 < 16
-tail_zero:
-	TBZ	$3, R1, tail_zero_4
-	MOVD.P	ZR, 8(R0)
+no_zva:
+	SUB	$16, R0, R0
+	SUB	$64, R1, R1
 
-tail_zero_4:
-	TBZ	$2, R1, tail_zero_2
-	MOVW.P	ZR, 4(R0)
+loop_64:
+	STP	(ZR, ZR), 16(R0)
+	STP	(ZR, ZR), 32(R0)
+	STP	(ZR, ZR), 48(R0)
+	STP.W	(ZR, ZR), 64(R0)
+	SUBS	$64, R1, R1
+	BGE	loop_64
+	ANDS	$63, R1, ZR
+	ADD	$16, R0, R0
+	BNE	tail63
+	RET
 
-tail_zero_2:
-	TBZ	$1, R1, tail_zero_1
-	MOVH.P	ZR, 2(R0)
+try_zva:
+	// Try using the ZVA feature to zero entire cache lines
+	// It is not meaningful to use ZVA if the block size is less than 64,
+	// so make sure that n is greater than or equal to 64
+	CMP	$63, R1
+	BLE	tail63
 
-tail_zero_1:
-	TBZ	$0, R1, ending
-	MOVB	ZR, (R0)
+	CMP	$128, R1
+	// Ensure n is at least 128 bytes, so that there is enough to copy after
+	// alignment.
+	BLT	no_zva
+	// Check if ZVA is allowed from user code, and if so get the block size
+	MOVW	block_size<>(SB), R5
+	TBNZ	$31, R5, no_zva
+	CBNZ	R5, zero_by_line
+        // DCZID_EL0 bit assignments
+        // [63:5] Reserved
+        // [4]    DZP, if bit set DC ZVA instruction is prohibited, else permitted
+        // [3:0]  log2 of the block size in words, eg. if it returns 0x4 then block size is 16 words
+	MRS	DCZID_EL0, R3
+	TBZ	$4, R3, init
+	// ZVA not available
+	MOVW	$~0, R5
+	MOVW	R5, block_size<>(SB)
+	B	no_zva
 
-ending:
+init:
+	MOVW	$4, R9
+	ANDW	$15, R3, R5
+	LSLW	R5, R9, R5
+	MOVW	R5, block_size<>(SB)
+
+	ANDS	$63, R5, R9
+	// Block size is less than 64.
+	BNE	no_zva
+
+zero_by_line:
+	CMP	R5, R1
+	// Not enough memory to reach alignment
+	BLO	no_zva
+	SUB	$1, R5, R6
+	NEG	R0, R4
+	ANDS	R6, R4, R4
+	// Already aligned
+	BEQ	aligned
+
+	// check there is enough to copy after alignment
+	SUB	R4, R1, R3
+
+	// Check that the remaining length to ZVA after alignment
+	// is greater than 64.
+	CMP	$64, R3
+	CCMP	GE, R3, R5, $10  // condition code GE, NZCV=0b1010
+	BLT	no_zva
+
+	// We now have at least 64 bytes to zero, update n
+	MOVD	R3, R1
+
+loop_zva_prolog:
+	STP	(ZR, ZR), (R0)
+	STP	(ZR, ZR), 16(R0)
+	STP	(ZR, ZR), 32(R0)
+	SUBS	$64, R4, R4
+	STP	(ZR, ZR), 48(R0)
+	ADD	$64, R0, R0
+	BGE	loop_zva_prolog
+
+	ADD	R4, R0, R0
+
+aligned:
+	SUB	R5, R1, R1
+
+loop_zva:
+	WORD	$0xd50b7420 // DC ZVA, R0
+	ADD	R5, R0, R0
+	SUBS	R5, R1, R1
+	BHS	loop_zva
+	ANDS	R6, R1, R1
+	BNE	tail_maybe_long
 	RET
 
-unaligned_to_16:
-	MOVD	R0, R2
-head_loop:
-	MOVBU.P	ZR, 1(R0)
-	ANDS	$15, R0, ZR
-	BNE	head_loop
-	// Adjust length for what remains
-	SUB	R2, R0, R3
-	SUB	R3, R1
-	// If size is less than 16 bytes, use tail_zero to zero what remains
-	CMP	$16, R1
-	BLT	tail_zero
-	B	aligned_to_16
+GLOBL block_size<>(SB), NOPTR, $8