2017-10-10 Siddhesh Poyarekar * sysdeps/aarch64/multiarch/Makefile (sysdep_routines): Add memmove_falkor. * sysdeps/aarch64/multiarch/ifunc-impl-list.c (__libc_ifunc_impl_list): Likewise. * sysdeps/aarch64/multiarch/memmove.c: Likewise. * sysdeps/aarch64/multiarch/memmove_falkor.S: New file. * benchtests/bench-memmove-walk.c: New file. * benchtests/Makefile (string-benchset): Add it. * benchtests/bench-memset-walk.c: New file. * benchtests/Makefile (string-benchset): Add it. * benchtests/bench-memcpy-walk.c: New file. * benchtests/Makefile (string-benchset): Add it. 2017-10-10 Siddhesh Poyarekar * manual/tunables.texi (Tunable glibc.tune.cpu): Add falkor. * sysdeps/aarch64/multiarch/Makefile (sysdep_routines): Add memcpy_falkor. * sysdeps/aarch64/multiarch/ifunc-impl-list.c (MAX_IFUNC): Bump. (__libc_ifunc_impl_list): Add __memcpy_falkor. * sysdeps/aarch64/multiarch/memcpy.c: Likewise. * sysdeps/aarch64/multiarch/memcpy_falkor.S: New file. * sysdeps/unix/sysv/linux/aarch64/cpu-features.c (cpu_list): Add falkor. * sysdeps/unix/sysv/linux/aarch64/cpu-features.h (IS_FALKOR): New macro. Index: glibc-2.26/manual/tunables.texi =================================================================== --- glibc-2.26.orig/manual/tunables.texi +++ glibc-2.26/manual/tunables.texi @@ -267,7 +267,7 @@ This tunable is specific to i386 and x86 @deftp Tunable glibc.tune.cpu The @code{glibc.tune.cpu=xxx} tunable allows the user to tell @theglibc{} to assume that the CPU is @code{xxx} where xxx may have one of these values: -@code{generic}, @code{thunderxt88}. +@code{generic}, @code{falkor}, @code{thunderxt88}. This tunable is specific to aarch64. @end deftp Index: glibc-2.26/sysdeps/aarch64/multiarch/Makefile =================================================================== --- glibc-2.26.orig/sysdeps/aarch64/multiarch/Makefile +++ glibc-2.26/sysdeps/aarch64/multiarch/Makefile @@ -1,3 +1,4 @@ ifeq ($(subdir),string) -sysdep_routines += memcpy_generic memcpy_thunderx +sysdep_routines += memcpy_generic memcpy_thunderx memcpy_falkor \ + memmove_falkor endif Index: glibc-2.26/sysdeps/aarch64/multiarch/ifunc-impl-list.c =================================================================== --- glibc-2.26.orig/sysdeps/aarch64/multiarch/ifunc-impl-list.c +++ glibc-2.26/sysdeps/aarch64/multiarch/ifunc-impl-list.c @@ -25,7 +25,7 @@ #include /* Maximum number of IFUNC implementations. */ -#define MAX_IFUNC 2 +#define MAX_IFUNC 3 size_t __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, @@ -40,9 +40,11 @@ __libc_ifunc_impl_list (const char *name /* Support sysdeps/aarch64/multiarch/memcpy.c and memmove.c. */ IFUNC_IMPL (i, name, memcpy, IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_thunderx) + IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_falkor) 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_falkor) IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic)) return i; Index: glibc-2.26/sysdeps/aarch64/multiarch/memcpy.c =================================================================== --- glibc-2.26.orig/sysdeps/aarch64/multiarch/memcpy.c +++ glibc-2.26/sysdeps/aarch64/multiarch/memcpy.c @@ -30,9 +30,14 @@ extern __typeof (__redirect_memcpy) __li extern __typeof (__redirect_memcpy) __memcpy_generic attribute_hidden; extern __typeof (__redirect_memcpy) __memcpy_thunderx attribute_hidden; +extern __typeof (__redirect_memcpy) __memcpy_falkor attribute_hidden; libc_ifunc (__libc_memcpy, - IS_THUNDERX (midr) ? __memcpy_thunderx : __memcpy_generic); + (IS_THUNDERX (midr) + ? __memcpy_thunderx + : (IS_FALKOR (midr) + ? __memcpy_falkor + : __memcpy_generic))); # undef memcpy strong_alias (__libc_memcpy, memcpy); Index: glibc-2.26/sysdeps/aarch64/multiarch/memcpy_falkor.S =================================================================== --- /dev/null +++ glibc-2.26/sysdeps/aarch64/multiarch/memcpy_falkor.S @@ -0,0 +1,184 @@ +/* Optimized memcpy for Qualcomm Falkor processor. + Copyright (C) 2017 Free Software Foundation, Inc. + + 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 + . */ + +#include + +/* Assumptions: + + ARMv8-a, AArch64, falkor, unaligned accesses. */ + +#define dstin x0 +#define src x1 +#define count x2 +#define dst x3 +#define srcend x4 +#define dstend x5 +#define A_l x6 +#define A_lw w6 +#define A_h x7 +#define A_hw w7 +#define tmp1 x14 + +/* Copies are split into 3 main cases: + + 1. Small copies of up to 32 bytes + 2. Medium copies of 33..128 bytes which are fully unrolled + 3. Large copies of more than 128 bytes. + + Large copies align the sourceto a quad word and use an unrolled loop + processing 64 bytes per iteration. + + FALKOR-SPECIFIC DESIGN: + + The smallest copies (32 bytes or less) focus on optimal pipeline usage, + which is why the redundant copies of 0-3 bytes have been replaced with + conditionals, since the former would unnecessarily break across multiple + issue groups. The medium copy group has been enlarged to 128 bytes since + bumping up the small copies up to 32 bytes allows us to do that without + cost and also allows us to reduce the size of the prep code before loop64. + + All copies are done only via two registers r6 and r7. This is to ensure + that all loads hit a single hardware prefetcher which can get correctly + trained to prefetch a single stream. + + The non-temporal stores help optimize cache utilization. */ + +#if IS_IN (libc) +ENTRY_ALIGN (__memcpy_falkor, 6) + + cmp count, 32 + add srcend, src, count + add dstend, dstin, count + b.ls L(copy32) + ldp A_l, A_h, [src] + cmp count, 128 + stp A_l, A_h, [dstin] + b.hi L(copy_long) + + /* Medium copies: 33..128 bytes. */ + sub tmp1, count, 1 + ldp A_l, A_h, [src, 16] + stp A_l, A_h, [dstin, 16] + tbz tmp1, 6, 1f + ldp A_l, A_h, [src, 32] + stp A_l, A_h, [dstin, 32] + ldp A_l, A_h, [src, 48] + stp A_l, A_h, [dstin, 48] + ldp A_l, A_h, [srcend, -64] + stp A_l, A_h, [dstend, -64] + ldp A_l, A_h, [srcend, -48] + stp A_l, A_h, [dstend, -48] +1: + ldp A_l, A_h, [srcend, -32] + stp A_l, A_h, [dstend, -32] + ldp A_l, A_h, [srcend, -16] + stp A_l, A_h, [dstend, -16] + ret + + .p2align 4 + /* Small copies: 0..32 bytes. */ +L(copy32): + /* 16-32 */ + cmp count, 16 + b.lo 1f + ldp A_l, A_h, [src] + stp A_l, A_h, [dstin] + ldp A_l, A_h, [srcend, -16] + stp A_l, A_h, [dstend, -16] + ret + .p2align 4 +1: + /* 8-15 */ + tbz count, 3, 1f + ldr A_l, [src] + str A_l, [dstin] + ldr A_l, [srcend, -8] + str A_l, [dstend, -8] + ret + .p2align 4 +1: + /* 4-7 */ + tbz count, 2, 1f + ldr A_lw, [src] + str A_lw, [dstin] + ldr A_lw, [srcend, -4] + str A_lw, [dstend, -4] + ret + .p2align 4 +1: + /* 2-3 */ + tbz count, 1, 1f + ldrh A_lw, [src] + strh A_lw, [dstin] + ldrh A_lw, [srcend, -2] + strh A_lw, [dstend, -2] + ret + .p2align 4 +1: + /* 0-1 */ + tbz count, 0, 1f + ldrb A_lw, [src] + strb A_lw, [dstin] +1: + ret + + /* Align SRC to 16 bytes and copy; that way at least one of the + accesses is aligned throughout the copy sequence. + + The count is off by 0 to 15 bytes, but this is OK because we trim + off the last 64 bytes to copy off from the end. Due to this the + loop never runs out of bounds. */ + .p2align 6 +L(copy_long): + sub count, count, 64 + 16 + and tmp1, src, 15 + bic src, src, 15 + sub dst, dstin, tmp1 + add count, count, tmp1 + +L(loop64): + ldp A_l, A_h, [src, 16]! + stnp A_l, A_h, [dst, 16] + ldp A_l, A_h, [src, 16]! + subs count, count, 64 + stnp A_l, A_h, [dst, 32] + ldp A_l, A_h, [src, 16]! + stnp A_l, A_h, [dst, 48] + ldp A_l, A_h, [src, 16]! + stnp A_l, A_h, [dst, 64] + add dst, dst, 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 A_l, A_h, [srcend, -64] + stnp A_l, A_h, [dstend, -64] + ldp A_l, A_h, [srcend, -48] + stnp A_l, A_h, [dstend, -48] + ldp A_l, A_h, [srcend, -32] + stnp A_l, A_h, [dstend, -32] + ldp A_l, A_h, [srcend, -16] + stnp A_l, A_h, [dstend, -16] + ret + +END (__memcpy_falkor) +libc_hidden_builtin_def (__memcpy_falkor) +#endif Index: glibc-2.26/sysdeps/aarch64/multiarch/memmove.c =================================================================== --- glibc-2.26.orig/sysdeps/aarch64/multiarch/memmove.c +++ glibc-2.26/sysdeps/aarch64/multiarch/memmove.c @@ -30,9 +30,14 @@ extern __typeof (__redirect_memmove) __l extern __typeof (__redirect_memmove) __memmove_generic attribute_hidden; extern __typeof (__redirect_memmove) __memmove_thunderx attribute_hidden; +extern __typeof (__redirect_memmove) __memmove_falkor attribute_hidden; libc_ifunc (__libc_memmove, - IS_THUNDERX (midr) ? __memmove_thunderx : __memmove_generic); + (IS_THUNDERX (midr) + ? __memmove_thunderx + : (IS_FALKOR (midr) + ? __memmove_falkor + : __memmove_generic))); # undef memmove strong_alias (__libc_memmove, memmove); Index: glibc-2.26/sysdeps/aarch64/multiarch/memmove_falkor.S =================================================================== --- /dev/null +++ glibc-2.26/sysdeps/aarch64/multiarch/memmove_falkor.S @@ -0,0 +1,232 @@ +/* Copyright (C) 2017 Free Software Foundation, Inc. + + 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 + . */ + +#include + +/* Assumptions: ARMv8-a, AArch64, falkor, unaligned accesses. */ + +#define dstin x0 +#define src x1 +#define count x2 +#define dstlen x3 +#define dst x3 +#define srcend x4 +#define dstend x5 +#define A_l x6 +#define A_lw w6 +#define A_h x7 +#define A_hw w7 +#define B_l x8 +#define B_lw w8 +#define B_h x9 +#define C_l x10 +#define C_h x11 +#define D_l x12 +#define D_h x13 +#define E_l src +#define E_h count +#define F_l srcend +#define F_h dst +#define tmp1 x14 + +/* Alias with A_l and A_h to train the prefetcher. */ +#define Q_l x22 +#define Q_h x23 + +/* RATIONALE: + + The copy has 4 distinct parts: + * Small copies of 16 bytes and under + * Medium sized copies of 17-96 bytes + * Large copies where the source address is higher than the destination + (forward copies) + * Large copies where the destination address is higher than the source + (copy backward, or move). + + We use only two registerpairs x6,x7 and x22,x23 for the copies and copy 32 + bytes at a time to correctly train the hardware prefetcher for better + throughput. */ +ENTRY_ALIGN (__memmove_falkor, 6) + + sub tmp1, dstin, src + add srcend, src, count + add dstend, dstin, count + cmp count, 96 + ccmp tmp1, count, 2, hi + b.lo L(move_long) + + 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: + /* 4-7 */ + tbz count, 2, 1f + ldr A_lw, [src] + ldr A_hw, [srcend, -4] + str A_lw, [dstin] + str A_hw, [dstend, -4] + ret + .p2align 4 +1: + /* 2-3 */ + tbz count, 1, 1f + ldrh A_lw, [src] + ldrh A_hw, [srcend, -2] + strh A_lw, [dstin] + strh A_hw, [dstend, -2] + ret + .p2align 4 +1: + /* 0-1 */ + tbz count, 0, 1f + ldrb A_lw, [src] + strb A_lw, [dstin] +1: 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 SRC 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 32 bytes per iteration and prefetches one iteration ahead. */ + + .p2align 4 +L(copy_long): + sub count, count, 64 + 16 /* Test and readjust count. */ + mov B_l, Q_l + mov B_h, Q_h + ldp A_l, A_h, [src] + and tmp1, src, 15 + bic src, src, 15 + sub dst, dstin, tmp1 + add count, count, tmp1 /* Count is now 16 too large. */ + ldp Q_l, Q_h, [src, 16]! + stp A_l, A_h, [dstin] + ldp A_l, A_h, [src, 16]! + +L(loop64): + subs count, count, 32 + stp Q_l, Q_h, [dst, 16] + ldp Q_l, Q_h, [src, 16]! + stp A_l, A_h, [dst, 32]! + ldp A_l, A_h, [src, 16]! + b.hi L(loop64) + + /* Write the last full set of 32 bytes. The remainder is at most 32 + bytes, so it is safe to always copy 32 bytes from the end even if + there is just 1 byte left. */ +L(last64): + ldp C_l, C_h, [srcend, -32] + stp Q_l, Q_h, [dst, 16] + ldp Q_l, Q_h, [srcend, -16] + stp A_l, A_h, [dst, 32] + stp C_l, C_h, [dstend, -32] + stp Q_l, Q_h, [dstend, -16] + mov Q_l, B_l + mov Q_h, B_h + ret + + .p2align 4 +L(move_long): + cbz tmp1, 3f + + mov B_l, Q_l + mov B_h, Q_h + + /* Align SRCEND 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 32 bytes per iteration and prefetches one iteration ahead. */ + + ldp A_l, A_h, [srcend, -16] + and tmp1, srcend, 15 + sub srcend, srcend, tmp1 + ldp Q_l, Q_h, [srcend, -16]! + stp A_l, A_h, [dstend, -16] + sub count, count, tmp1 + ldp A_l, A_h, [srcend, -16]! + sub dstend, dstend, tmp1 + sub count, count, 64 + +1: + subs count, count, 32 + stp Q_l, Q_h, [dstend, -16] + ldp Q_l, Q_h, [srcend, -16]! + stp A_l, A_h, [dstend, -32]! + ldp A_l, A_h, [srcend, -16]! + b.hi 1b + + /* Write the last full set of 32 bytes. The remainder is at most 32 + bytes, so it is safe to always copy 32 bytes from the start even if + there is just 1 byte left. */ +2: + ldp C_l, C_h, [src, 16] + stp Q_l, Q_h, [dstend, -16] + ldp Q_l, Q_h, [src] + stp A_l, A_h, [dstend, -32] + stp C_l, C_h, [dstin, 16] + stp Q_l, Q_h, [dstin] + mov Q_l, B_l + mov Q_h, B_h +3: ret + +END (__memmove_falkor) +libc_hidden_builtin_def (__memmove_falkor) Index: glibc-2.26/sysdeps/unix/sysv/linux/aarch64/cpu-features.c =================================================================== --- glibc-2.26.orig/sysdeps/unix/sysv/linux/aarch64/cpu-features.c +++ glibc-2.26/sysdeps/unix/sysv/linux/aarch64/cpu-features.c @@ -28,6 +28,7 @@ struct cpu_list }; static struct cpu_list cpu_list[] = { + {"falkor", 0x510FC000}, {"thunderxt88", 0x430F0A10}, {"generic", 0x0} }; Index: glibc-2.26/sysdeps/unix/sysv/linux/aarch64/cpu-features.h =================================================================== --- glibc-2.26.orig/sysdeps/unix/sysv/linux/aarch64/cpu-features.h +++ glibc-2.26/sysdeps/unix/sysv/linux/aarch64/cpu-features.h @@ -41,6 +41,9 @@ #define IS_THUNDERX(midr) (MIDR_IMPLEMENTOR(midr) == 'C' \ && MIDR_PARTNUM(midr) == 0x0a1) +#define IS_FALKOR(midr) (MIDR_IMPLEMENTOR(midr) == 'Q' \ + && MIDR_PARTNUM(midr) == 0xc00) + struct cpu_features { uint64_t midr_el1;