Aarch64: Add new memset for Qualcomm's oryon-1 core

Qualcom's new core, oryon-1, has a different characteristics for
memset than the current versions of memset. For non-zero, larger
sizes, using GPRs rather than the SIMD stores is ~30% faster.
For even larger sizes, using the nontemporal stores is needed
not to polute the L1/L2 caches.

For zero values, using `dc zva` should be used. Since we
know the size will always be 64 bytes, we don't need to figure
out the size there.

I started with the emag memset and added back the `dc zva` code.

Changes since v1:
* v3: Fix comment formating

Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
Reviewed-by: Adhemerval Zanella  <adhemerval.zanella@linaro.org>

sysdeps/aarch64/multiarch/Makefile

@@ -15,6 +15,7 @@ sysdep_routines += \
   memset_generic \
   memset_kunpeng \
   memset_mops \
+  memset_oryon1 \
   memset_zva64 \
   strlen_asimd \
   strlen_generic \

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

@@ -56,6 +56,7 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic))
   IFUNC_IMPL (i, name, memset,
 	      IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_zva64)
+	      IFUNC_IMPL_ADD (array, i, memset, (zva_size == 64), __memset_oryon1)
 	      IFUNC_IMPL_ADD (array, i, memset, 1, __memset_emag)
 	      IFUNC_IMPL_ADD (array, i, memset, 1, __memset_kunpeng)
 #if HAVE_AARCH64_SVE_ASM

sysdeps/aarch64/multiarch/memset.c

@@ -1,5 +1,6 @@
 /* Multiple versions of memset. AARCH64 version.
    Copyright (C) 2017-2024 Free Software Foundation, Inc.
+   Copyright The GNU Toolchain Authors.
    This file is part of the GNU C Library.
 
    The GNU C Library is free software; you can redistribute it and/or
@@ -34,6 +35,7 @@ extern __typeof (__redirect_memset) __memset_kunpeng attribute_hidden;
 extern __typeof (__redirect_memset) __memset_a64fx attribute_hidden;
 extern __typeof (__redirect_memset) __memset_generic attribute_hidden;
 extern __typeof (__redirect_memset) __memset_mops attribute_hidden;
+extern __typeof (__redirect_memset) __memset_oryon1 attribute_hidden;
 
 static inline __typeof (__redirect_memset) *
 select_memset_ifunc (void)
@@ -49,6 +51,9 @@ select_memset_ifunc (void)
 	return __memset_a64fx;
     }
 
+  if (IS_ORYON1 (midr) && zva_size == 64)
+    return __memset_oryon1;
+
   if (IS_KUNPENG920 (midr))
     return __memset_kunpeng;
 

sysdeps/aarch64/multiarch/memset_oryon1.S

@@ -0,0 +1,169 @@
+/* Optimized memset for Qualcomm's oyron-1 core.
+   Copyright (C) 2018-2024 Free Software Foundation, Inc.
+   Copyright The GNU Toolchain Authors.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library.  If not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+#include "memset-reg.h"
+
+/* Assumptions:
+   ARMv8-a, AArch64, unaligned accesses
+ */
+
+ENTRY (__memset_oryon1)
+
+	PTR_ARG (0)
+	SIZE_ARG (2)
+
+	bfi	valw, valw, 8, 8
+	bfi	valw, valw, 16, 16
+	bfi	val, val, 32, 32
+
+	add	dstend, dstin, count
+
+	cmp	count, 96
+	b.hi	L(set_long)
+	cmp	count, 16
+	b.hs	L(set_medium)
+
+	/* Set 0..15 bytes.  */
+	tbz	count, 3, 1f
+	str	val, [dstin]
+	str	val, [dstend, -8]
+	ret
+
+	.p2align 3
+1:	tbz	count, 2, 2f
+	str	valw, [dstin]
+	str	valw, [dstend, -4]
+	ret
+2:	cbz	count, 3f
+	strb	valw, [dstin]
+	tbz	count, 1, 3f
+	strh	valw, [dstend, -2]
+3:	ret
+
+	.p2align 3
+	/* Set 16..96 bytes.  */
+L(set_medium):
+	stp	val, val, [dstin]
+	tbnz	count, 6, L(set96)
+	stp	val, val, [dstend, -16]
+	tbz	count, 5, 1f
+	stp	val, val, [dstin, 16]
+	stp	val, val, [dstend, -32]
+1:	ret
+
+	.p2align 6
+	/* Set 64..96 bytes.  Write 64 bytes from the start and
+	   32 bytes from the end.  */
+L(set96):
+	stp	val, val, [dstin, 16]
+	stp	val, val, [dstin, 32]
+	stp	val, val, [dstin, 48]
+	stp	val, val, [dstend, -32]
+	stp	val, val, [dstend, -16]
+	ret
+
+	.p2align 6
+L(set_long):
+	stp	val, val, [dstin]
+	bic	dst, dstin, 15
+	cmp	count, 256
+	ccmp	valw, 0, 0, cs
+	b.eq	L(try_zva)
+	cmp	count, #32768
+	b.hi	L(set_long_with_nontemp)
+	/* Small-size or non-zero memset does not use DC ZVA. */
+	sub	count, dstend, dst
+
+	/* Adjust count and bias for loop. By subtracting extra 1 from count,
+	  it is easy to use tbz instruction to check whether loop tailing
+	  count is less than 33 bytes, so as to bypass 2 unnecessary stps. */
+	sub	count, count, 64+16+1
+
+1:	stp	val, val, [dst, 16]
+	stp	val, val, [dst, 32]
+	stp	val, val, [dst, 48]
+	stp	val, val, [dst, 64]!
+	subs	count, count, 64
+	b.hs	1b
+
+	tbz	count, 5, 1f	/* Remaining count is less than 33 bytes? */
+	stp	val, val, [dst, 16]
+	stp	val, val, [dst, 32]
+1:	stp	val, val, [dstend, -32]
+	stp	val, val, [dstend, -16]
+	ret
+
+L(set_long_with_nontemp):
+	/* Small-size or non-zero memset does not use DC ZVA. */
+	sub	count, dstend, dst
+
+	/* Adjust count and bias for loop. By subtracting extra 1 from count,
+	   it is easy to use tbz instruction to check whether loop tailing
+	   count is less than 33 bytes, so as to bypass 2 unnecessary stps. */
+	sub	count, count, 64+16+1
+
+1:	stnp	val, val, [dst, 16]
+	stnp	val, val, [dst, 32]
+	stnp	val, val, [dst, 48]
+	stnp	val, val, [dst, 64]
+	add	dst, dst, #64
+	subs	count, count, 64
+	b.hs	1b
+
+	tbz	count, 5, 1f	/* Remaining count is less than 33 bytes? */
+	stnp	val, val, [dst, 16]
+	stnp	val, val, [dst, 32]
+1:	stnp	val, val, [dstend, -32]
+	stnp	val, val, [dstend, -16]
+	ret
+
+L(try_zva):
+	/* Write the first and last 64 byte aligned block using stp rather
+	   than using DC ZVA as it is faster. */
+	.p2align 6
+L(zva_64):
+	stp	val, val, [dst, 16]
+	stp	val, val, [dst, 32]
+	stp	val, val, [dst, 48]
+	bic	dst, dst, 63
+	stp	val, val, [dst, 64]
+	stp	val, val, [dst, 64+16]
+	stp	val, val, [dst, 96]
+	stp	val, val, [dst, 96+16]
+	sub	count, dstend, dst	/* Count is now 128 too large.	*/
+	sub	count, count, 128+64+64	/* Adjust count and bias for loop.  */
+	add	dst, dst, 128
+1:	dc	zva, dst
+	add	dst, dst, 64
+	subs	count, count, 64
+	b.hi	1b
+	stp	val, val, [dst, 0]
+	stp	val, val, [dst, 16]
+	stp	val, val, [dst, 32]
+	stp	val, val, [dst, 48]
+
+	stp	val, val, [dstend, -64]
+	stp	val, val, [dstend, -64+16]
+	stp	val, val, [dstend, -32]
+	stp	val, val, [dstend, -16]
+	ret
+
+END (__memset_oryon1)