aarch64: thunderx2 memmove performance improvements

The performance improvement is about 20%-30% for
larger cases and about 1%-5% for smaller cases.

Used SIMD load/store instead of GPR for large
overlapping forward moves.

Reused existing memcpy implementation for smaller
or overlapping backward moves.

Fixed the existing memcpy implementation to allow it
to deal with the overlapping case.

Simplified loop tails in the memcpy implementation -
use branchless overlapping sequence of fixed length
load/stores instead of branching depending on the
size.

A cleanup/optimization converting str's to stp's.

Added __memmove_thunderx2 to the list of the
available implementations.

ChangeLog

@@ -1,3 +1,14 @@
+2019-05-03  Anton Youdkevitch  <anton.youdkevitch@bell-sw.com>
+
+	* sysdeps/aarch64/multiarch/ifunc-impl-list.c: Added
+	__memmove_thunderx2 to the list of implementations
+	* sysdeps/aarch64/multiarch/memmove.c: Likewise
+	* sysdeps/aarch64/multiarch/memcpy_thunderx2.S:
+	(__memmove_thunderx2): Rewritten using SIMD ld/st
+	(__memcpy_thunderx2): Fixed handling overlapping cases.
+	Used ldp/stp instead of ldr/str if possible. Made loops
+	tails branchless.
+
 2019-05-03  Florian Weimer  <fweimer@redhat.com>
 
 	* misc/tst-tsearch.c (walk_tree): Add more error checking.

sysdeps/aarch64/multiarch/ifunc-impl-list.c

@@ -45,6 +45,7 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_generic))
   IFUNC_IMPL (i, name, memmove,
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx)
+	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx2)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_falkor)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic))
   IFUNC_IMPL (i, name, memset,

sysdeps/aarch64/multiarch/memcpy_thunderx2.S

@@ -85,16 +85,14 @@
 #if IS_IN (libc)
 
 #undef MEMCPY
-#undef MEMMOVE
 #define MEMCPY __memcpy_thunderx2
+#undef MEMMOVE
 #define MEMMOVE __memmove_thunderx2
 
 
-/* Moves are split into 3 main cases: small copies of up to 16 bytes,
+/* Overlapping large forward memmoves use a loop that copies backwards.
-   medium copies of 17..96 bytes which are fully unrolled. Large copies
+   Otherwise memcpy is used. Small moves branch to memcopy16 directly.
-   of more than 96 bytes align the destination and use an unrolled loop
+   The longer memcpy cases fall through to the memcpy head.
-   processing 64 bytes per iteration.
-   Overlapping large forward memmoves use a loop that copies backwards.
 */
 
 ENTRY_ALIGN (MEMMOVE, 6)
@@ -103,188 +101,14 @@ ENTRY_ALIGN (MEMMOVE, 6)
 	DELOUSE (1)
 	DELOUSE (2)
 
+	add	srcend, src, count
+	cmp	count, 16
+	b.ls	L(memcopy16)
 	sub	tmp1, dstin, src
 	cmp	count, 96
 	ccmp	tmp1, count, 2, hi
 	b.lo	L(move_long)
 
-	prfm	PLDL1KEEP, [src]
-	add	srcend, src, count
-	add	dstend, dstin, count
-	cmp	count, 16
-	b.ls	L(copy16)
-	cmp	count, 96
-	b.hi	L(copy_long)
-
-	/* Medium copies: 17..96 bytes.  */
-	sub	tmp1, count, 1
-	ldp	A_l, A_h, [src]
-	tbnz	tmp1, 6, L(copy96)
-	ldp	D_l, D_h, [srcend, -16]
-	tbz	tmp1, 5, 1f
-	ldp	B_l, B_h, [src, 16]
-	ldp	C_l, C_h, [srcend, -32]
-	stp	B_l, B_h, [dstin, 16]
-	stp	C_l, C_h, [dstend, -32]
-1:
-	stp	A_l, A_h, [dstin]
-	stp	D_l, D_h, [dstend, -16]
-	ret
-
-	.p2align 4
-	/* Small copies: 0..16 bytes.  */
-L(copy16):
-	cmp	count, 8
-	b.lo	1f
-	ldr	A_l, [src]
-	ldr	A_h, [srcend, -8]
-	str	A_l, [dstin]
-	str	A_h, [dstend, -8]
-	ret
-	.p2align 4
-1:
-	tbz	count, 2, 1f
-	ldr	A_lw, [src]
-	ldr	A_hw, [srcend, -4]
-	str	A_lw, [dstin]
-	str	A_hw, [dstend, -4]
-	ret
-
-	/* Copy 0..3 bytes.  Use a branchless sequence that copies the same
-	   byte 3 times if count==1, or the 2nd byte twice if count==2.  */
-1:
-	cbz	count, 2f
-	lsr	tmp1, count, 1
-	ldrb	A_lw, [src]
-	ldrb	A_hw, [srcend, -1]
-	ldrb	B_lw, [src, tmp1]
-	strb	A_lw, [dstin]
-	strb	B_lw, [dstin, tmp1]
-	strb	A_hw, [dstend, -1]
-2:	ret
-
-	.p2align 4
-	/* Copy 64..96 bytes.  Copy 64 bytes from the start and
-	   32 bytes from the end.  */
-L(copy96):
-	ldp	B_l, B_h, [src, 16]
-	ldp	C_l, C_h, [src, 32]
-	ldp	D_l, D_h, [src, 48]
-	ldp	E_l, E_h, [srcend, -32]
-	ldp	F_l, F_h, [srcend, -16]
-	stp	A_l, A_h, [dstin]
-	stp	B_l, B_h, [dstin, 16]
-	stp	C_l, C_h, [dstin, 32]
-	stp	D_l, D_h, [dstin, 48]
-	stp	E_l, E_h, [dstend, -32]
-	stp	F_l, F_h, [dstend, -16]
-	ret
-
-	/* Align DST to 16 byte alignment so that we don't cross cache line
-	   boundaries on both loads and stores.  There are at least 96 bytes
-	   to copy, so copy 16 bytes unaligned and then align.  The loop
-	   copies 64 bytes per iteration and prefetches one iteration ahead.  */
-
-	.p2align 4
-L(copy_long):
-	and	tmp1, dstin, 15
-	bic	dst, dstin, 15
-	ldp	D_l, D_h, [src]
-	sub	src, src, tmp1
-	add	count, count, tmp1	/* Count is now 16 too large.  */
-	ldp	A_l, A_h, [src, 16]
-	stp	D_l, D_h, [dstin]
-	ldp	B_l, B_h, [src, 32]
-	ldp	C_l, C_h, [src, 48]
-	ldp	D_l, D_h, [src, 64]!
-	subs	count, count, 128 + 16	/* Test and readjust count.  */
-	b.ls	L(last64)
-L(loop64):
-	stp	A_l, A_h, [dst, 16]
-	ldp	A_l, A_h, [src, 16]
-	stp	B_l, B_h, [dst, 32]
-	ldp	B_l, B_h, [src, 32]
-	stp	C_l, C_h, [dst, 48]
-	ldp	C_l, C_h, [src, 48]
-	stp	D_l, D_h, [dst, 64]!
-	ldp	D_l, D_h, [src, 64]!
-	subs	count, count, 64
-	b.hi	L(loop64)
-
-	/* Write the last full set of 64 bytes.  The remainder is at most 64
-	   bytes, so it is safe to always copy 64 bytes from the end even if
-	   there is just 1 byte left.  */
-L(last64):
-	ldp	E_l, E_h, [srcend, -64]
-	stp	A_l, A_h, [dst, 16]
-	ldp	A_l, A_h, [srcend, -48]
-	stp	B_l, B_h, [dst, 32]
-	ldp	B_l, B_h, [srcend, -32]
-	stp	C_l, C_h, [dst, 48]
-	ldp	C_l, C_h, [srcend, -16]
-	stp	D_l, D_h, [dst, 64]
-	stp	E_l, E_h, [dstend, -64]
-	stp	A_l, A_h, [dstend, -48]
-	stp	B_l, B_h, [dstend, -32]
-	stp	C_l, C_h, [dstend, -16]
-	ret
-
-	.p2align 4
-L(move_long):
-	cbz	tmp1, 3f
-
-	add	srcend, src, count
-	add	dstend, dstin, count
-
-	/* Align dstend to 16 byte alignment so that we don't cross cache line
-	   boundaries on both loads and stores.  There are at least 96 bytes
-	   to copy, so copy 16 bytes unaligned and then align.  The loop
-	   copies 64 bytes per iteration and prefetches one iteration ahead.  */
-
-	and	tmp1, dstend, 15
-	ldp	D_l, D_h, [srcend, -16]
-	sub	srcend, srcend, tmp1
-	sub	count, count, tmp1
-	ldp	A_l, A_h, [srcend, -16]
-	stp	D_l, D_h, [dstend, -16]
-	ldp	B_l, B_h, [srcend, -32]
-	ldp	C_l, C_h, [srcend, -48]
-	ldp	D_l, D_h, [srcend, -64]!
-	sub	dstend, dstend, tmp1
-	subs	count, count, 128
-	b.ls	2f
-
-	nop
-1:
-	stp	A_l, A_h, [dstend, -16]
-	ldp	A_l, A_h, [srcend, -16]
-	stp	B_l, B_h, [dstend, -32]
-	ldp	B_l, B_h, [srcend, -32]
-	stp	C_l, C_h, [dstend, -48]
-	ldp	C_l, C_h, [srcend, -48]
-	stp	D_l, D_h, [dstend, -64]!
-	ldp	D_l, D_h, [srcend, -64]!
-	subs	count, count, 64
-	b.hi	1b
-
-	/* Write the last full set of 64 bytes.  The remainder is at most 64
-	   bytes, so it is safe to always copy 64 bytes from the start even if
-	   there is just 1 byte left.  */
-2:
-	ldp	G_l, G_h, [src, 48]
-	stp	A_l, A_h, [dstend, -16]
-	ldp	A_l, A_h, [src, 32]
-	stp	B_l, B_h, [dstend, -32]
-	ldp	B_l, B_h, [src, 16]
-	stp	C_l, C_h, [dstend, -48]
-	ldp	C_l, C_h, [src]
-	stp	D_l, D_h, [dstend, -64]
-	stp	G_l, G_h, [dstin, 48]
-	stp	A_l, A_h, [dstin, 32]
-	stp	B_l, B_h, [dstin, 16]
-	stp	C_l, C_h, [dstin]
-3:	ret
-
 END (MEMMOVE)
 libc_hidden_builtin_def (MEMMOVE)
 
@@ -293,18 +117,16 @@ libc_hidden_builtin_def (MEMMOVE)
    medium copies of 17..96 bytes which are fully unrolled. Large copies
    of more than 96 bytes align the destination and use load-and-merge
    approach in the case src and dst addresses are unaligned not evenly,
-   so that, loads and stores are always aligned.
+   so that, actual loads and stores are always aligned.
-   Large copies use an unrolled loop processing 64 bytes per iteration.
+   Large copies use the loops processing 64 bytes per iteration for
-   The current optimized memcpy implementation is not compatible with
+   unaligned case and 128 bytes per iteration for aligned ones.
-   memmove and is separated from it completely.
+*/
-
-   memcpy implementation below is not compatible with memmove
-   because of pipelined loads/stores, which are faster, but they
-   can't be used in the case of overlapping memmove arrays */
 
 #define MEMCPY_PREFETCH_LDR 640
 
+	.p2align 4
 ENTRY (MEMCPY)
+
 	DELOUSE (0)
 	DELOUSE (1)
 	DELOUSE (2)
@@ -373,29 +195,26 @@ L(bytes_0_to_8):
 	/* Copy 0..3 bytes.  Use a branchless sequence that copies the same
 	   byte 3 times if count==1, or the 2nd byte twice if count==2.  */
 L(bytes_0_to_3):
-	cbz     count, L(end)
+	cbz	count, 1f
 	lsr	tmp1, count, 1
 	ldrb	A_lw, [src]
 	ldrb	A_hw, [srcend, -1]
 	add	dstend, dstin, count
 	ldrb	B_lw, [src, tmp1]
-	strb    A_lw, [dstin]
 	strb	B_lw, [dstin, tmp1]
 	strb	A_hw, [dstend, -1]
-L(end):
+	strb	A_lw, [dstin]
+1:
 	ret
 
 	.p2align 4
 
 L(memcpy_copy96):
 	/* Copying 65..96 bytes. A_q (first 16 bytes) and
-	   E_q(last 16 bytes) are already loaded.
+	   E_q(last 16 bytes) are already loaded. The size
-
+	   is large enough to benefit from aligned loads */
-	   The size is large enough to benefit from aligned
-	   loads */
 	bic	src, src, 15
 	ldp	B_q, C_q, [src]
-	str     A_q, [dstin]
 	/* Loaded 64 bytes, second 16-bytes chunk can be
 	   overlapping with the first chunk by tmp1 bytes.
 	   Stored 16 bytes. */
@@ -408,44 +227,46 @@ L(memcpy_copy96):
 	b.gt	L(copy96_medium)
 	ldr	D_q, [src, 32]
 	stp	B_q, C_q, [dst, 16]
-	str     E_q, [dstend, -16]
 	str	D_q, [dst, 48]
+	str	A_q, [dstin]
+	str	E_q, [dstend, -16]
 	ret
 
 	.p2align 4
 L(copy96_medium):
-	ldp     D_q, A_q, [src, 32]
+	ldp	D_q, G_q, [src, 32]
-	str     B_q, [dst, 16]
 	cmp	count, 96
 	b.gt	L(copy96_large)
+	stp	B_q, C_q, [dst, 16]
+	stp	D_q, G_q, [dst, 48]
+	str	A_q, [dstin]
 	str	E_q, [dstend, -16]
-	stp     C_q, D_q, [dst, 32]
-	str     A_q, [dst, 64]
 	ret
 
 L(copy96_large):
 	ldr	F_q, [src, 64]
+	str	B_q, [dst, 16]
 	stp	C_q, D_q, [dst, 32]
+	stp	G_q, F_q, [dst, 64]
+	str	A_q, [dstin]
 	str	E_q, [dstend, -16]
-	stp     A_q, F_q, [dst, 64]
 	ret
 
 	.p2align 4
 L(memcopy_long):
 	bic	src, src, 15
 	ldp	B_q, C_q, [src], #32
-	str     A_q, [dstin]
 	sub	dst, dstin, tmp1
 	add	count, count, tmp1
 	add	dst, dst, 16
 	and	tmp1, dst, 15
 	ldp	D_q, E_q, [src], #32
-	str     B_q, [dst], #16
+	str	A_q, [dstin]
 
 	/* Already loaded 64+16 bytes. Check if at
 	   least 64 more bytes left */
 	subs	count, count, 64+64+16
-	b.lt    L(loop128_exit2)
+	b.lt	L(loop128_exit0)
 	cmp	count, MEMCPY_PREFETCH_LDR + 64 + 32
 	b.lt	L(loop128)
 	cbnz	tmp1, L(dst_unaligned)
@@ -454,66 +275,59 @@ L(memcopy_long):
 	.p2align 4
 
 L(loop128_prefetch):
-	str     C_q, [dst], #16
 	prfm	pldl1strm, [src, MEMCPY_PREFETCH_LDR]
-	str     D_q, [dst], #16
 	ldp	F_q, G_q, [src], #32
-	str	E_q, [dst], #16
+	stp	B_q, C_q, [dst], #32
-	ldp     H_q, A_q, [src], #32
+	ldp	H_q, I_q, [src], #32
-	str     F_q, [dst], #16
 	prfm	pldl1strm, [src, MEMCPY_PREFETCH_LDR]
-	str     G_q, [dst], #16
 	ldp	B_q, C_q, [src], #32
-	str	H_q, [dst], #16
+	stp	D_q, E_q, [dst], #32
 	ldp	D_q, E_q, [src], #32
-	stp	A_q, B_q, [dst], #32
+	stp	F_q, G_q, [dst], #32
+	stp	H_q, I_q, [dst], #32
 	subs	count, count, 128
 	b.ge	L(loop128_prefetch)
 
-L(preloop128):
 	add	count, count, MEMCPY_PREFETCH_LDR + 64 + 32
 	.p2align 4
 L(loop128):
 	ldp	F_q, G_q, [src], #32
-	str     C_q, [dst], #16
+	ldp	H_q, I_q, [src], #32
-	ldp     B_q, A_q, [src], #32
+	stp	B_q, C_q, [dst], #32
-	str     D_q, [dst], #16
+	stp	D_q, E_q, [dst], #32
-	stp     E_q, F_q, [dst], #32
-	stp     G_q, B_q, [dst], #32
 	subs	count, count, 64
 	b.lt	L(loop128_exit1)
-L(loop128_proceed):
 	ldp	B_q, C_q, [src], #32
-	str     A_q, [dst], #16
 	ldp	D_q, E_q, [src], #32
-	str     B_q, [dst], #16
+	stp	F_q, G_q, [dst], #32
+	stp	H_q, I_q, [dst], #32
 	subs	count, count, 64
 	b.ge	L(loop128)
-
+L(loop128_exit0):
-	.p2align 4
+	ldp	F_q, G_q, [srcend, -64]
-L(loop128_exit2):
+	ldp	H_q, I_q, [srcend, -32]
-	stp     C_q, D_q, [dst], #32
+	stp	B_q, C_q, [dst], #32
-	str     E_q, [dst], #16
+	stp	D_q, E_q, [dst]
-	b       L(copy_long_check32);
+	stp	F_q, G_q, [dstend, -64]
-
+	stp	H_q, I_q, [dstend, -32]
+	ret
 L(loop128_exit1):
-	/* A_q is still not stored and 0..63 bytes left,
+	ldp	B_q, C_q, [srcend, -64]
-	   so, count is -64..-1.
+	ldp	D_q, E_q, [srcend, -32]
-	   Check if less than 32 bytes left (count < -32) */
+	stp	F_q, G_q, [dst], #32
-	str     A_q, [dst], #16
+	stp	H_q, I_q, [dst]
-L(copy_long_check32):
+	stp	B_q, C_q, [dstend, -64]
-	cmn     count, 64
+	stp	D_q, E_q, [dstend, -32]
-	b.eq    L(copy_long_done)
+	ret
-	cmn     count, 32
-	b.le    L(copy_long_last32)
-	ldp     B_q, C_q, [src]
-	stp     B_q, C_q, [dst]
 
-L(copy_long_last32):
+L(dst_unaligned_tail):
-	ldp     F_q, G_q, [srcend, -32]
+	ldp	C_q, D_q, [srcend, -64]
-	stp     F_q, G_q, [dstend, -32]
+	ldp	E_q, F_q, [srcend, -32]
-
+	stp	A_q, B_q, [dst], #32
-L(copy_long_done):
+	stp	H_q, I_q, [dst], #16
+	str	G_q, [dst, tmp1]
+	stp	C_q, D_q, [dstend, -64]
+	stp	E_q, F_q, [dstend, -32]
 	ret
 
 L(dst_unaligned):
@@ -542,17 +356,20 @@ L(dst_unaligned):
 	/* Store the 16 bytes to dst and align dst for further
 	   operations, several bytes will be stored at this
 	   address once more */
-	str     C_q, [dst], #16
+
 	ldp	F_q, G_q, [src], #32
+	stp	B_q, C_q, [dst], #32
 	bic	dst, dst, 15
-	subs    count, count, 32
+	sub	count, count, 32
 	adrp	tmp2, L(ext_table)
 	add	tmp2, tmp2, :lo12:L(ext_table)
 	add	tmp2, tmp2, tmp1, LSL #2
 	ldr	tmp3w, [tmp2]
 	add	tmp2, tmp2, tmp3w, SXTW
 	br	tmp2
-.p2align 4 ;\
+
+.p2align 4
+	/* to make the loop in each chunk 16-bytes aligned */
 	nop
 #define EXT_CHUNK(shft) \
 L(ext_size_ ## shft):;\
@@ -573,7 +390,7 @@ L(ext_size_ ## shft):;\
 	b.ge    1b;\
 2:;\
 	ext     I_v.16b, F_v.16b, G_v.16b, 16-shft;\
-	b	L(ext_tail);
+	b	L(dst_unaligned_tail);
 
 EXT_CHUNK(1)
 EXT_CHUNK(2)
@@ -591,12 +408,45 @@ EXT_CHUNK(13)
 EXT_CHUNK(14)
 EXT_CHUNK(15)
 
-L(ext_tail):
+L(move_long):
-	stp     A_q, B_q, [dst], #32
+	.p2align 4
-	stp     H_q, I_q, [dst], #16
+1:
-	add     dst, dst, tmp1
+	cbz	tmp1, 3f
-	str     G_q, [dst], #16
+
-	b       L(copy_long_check32)
+	add	srcend, src, count
+	add	dstend, dstin, count
+
+	and	tmp1, srcend, 15
+	ldr	D_q, [srcend, -16]
+	sub	srcend, srcend, tmp1
+	sub	count, count, tmp1
+	ldp	A_q, B_q, [srcend, -32]
+	str	D_q, [dstend, -16]
+	ldp	C_q, D_q, [srcend, -64]!
+	sub	dstend, dstend, tmp1
+	subs	count, count, 128
+	b.ls	2f
+
+	.p2align 4
+1:
+	subs	count, count, 64
+	stp	A_q, B_q, [dstend, -32]
+	ldp	A_q, B_q, [srcend, -32]
+	stp	C_q, D_q, [dstend, -64]!
+	ldp	C_q, D_q, [srcend, -64]!
+	b.hi	1b
+
+	/* Write the last full set of 64 bytes.  The remainder is at most 64
+	   bytes, so it is safe to always copy 64 bytes from the start even if
+	   there is just 1 byte left.  */
+2:
+	ldp	E_q, F_q, [src, 32]
+	ldp	G_q, H_q, [src]
+	stp	A_q, B_q, [dstend, -32]
+	stp	C_q, D_q, [dstend, -64]
+	stp	E_q, F_q, [dstin, 32]
+	stp	G_q, H_q, [dstin]
+3:	ret
 
 
 END (MEMCPY)

sysdeps/aarch64/multiarch/memmove.c

@@ -30,6 +30,7 @@ extern __typeof (__redirect_memmove) __libc_memmove;
 
 extern __typeof (__redirect_memmove) __memmove_generic attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_thunderx attribute_hidden;
+extern __typeof (__redirect_memmove) __memmove_thunderx2 attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_falkor attribute_hidden;
 
 libc_ifunc (__libc_memmove,
@@ -37,7 +38,9 @@ libc_ifunc (__libc_memmove,
 	     ? __memmove_thunderx
 	     : (IS_FALKOR (midr) || IS_PHECDA (midr)
 		? __memmove_falkor
-		: __memmove_generic)));
+		: (IS_THUNDERX2 (midr) || IS_THUNDERX2PA (midr)
+		  ? __memmove_thunderx2
+		  : __memmove_generic))));
 
 # undef memmove
 strong_alias (__libc_memmove, memmove);