diff options
author | Krzysztof Koch <Krzysztof.Koch@arm.com> | 2020-06-08 14:06:15 +0100 |
---|---|---|
committer | Wilco Dijkstra <wdijkstr@arm.com> | 2020-06-08 14:13:05 +0100 |
commit | d1f75e964484504e4f30f4623569d5889a97ac18 (patch) | |
tree | 5b123eeec64573bb9fa690b3f0186b66bb53ae6a /sysdeps/aarch64/multiarch/memcpy_falkor.S | |
parent | f112dcc506a6ec0aac5c34891736eec3c4f5dad6 (diff) | |
download | glibc-d1f75e964484504e4f30f4623569d5889a97ac18.tar.gz glibc-d1f75e964484504e4f30f4623569d5889a97ac18.tar.xz glibc-d1f75e964484504e4f30f4623569d5889a97ac18.zip |
AArch64: Merge Falkor memcpy and memmove implementations
Falkor's memcpy and memmove share some implementation details, therefore, the two routines are moved to a single source file for code reuse. The two routines now share code for small and medium copies (up to and including 128 bytes). Large copies in memcpy do not handle overlap correctly, consequently, the loops for moving/copying more than 128 bytes stay separate for memcpy and memmove. To increase code reuse a number of small modifications were made: 1. The old implementation of memcpy copied the first 16-bytes as soon as the size of data was determined to be greater than 32 bytes. For memcpy code to also work when copying small/medium overlapping data, the first load and store was moved to the large copy case. 2. Medium memcpy case no longer assumes that 16 bytes were already copied and uses 8 registers to copy up to 128 bytes. 3. Small case for memmove was enlarged to that of memcpy, which is less than or equal to 32 bytes. 4. Medium case for memmove was enlarged to that of memcpy, which is less than or equal to 128 bytes. Other changes include: 1. Improve alignment of existing loop bodies. 2. 'Delouse' memmove and memcpy input arguments. Make sure that upper 32-bits of input registers are zeroed if unused. 3. Do one more iteration in memmove loops and reduce the number of copies made from the start/end of the buffer, depending on the direction of the memmove loop. Benchmarking: Looking at the results from bench-memcpy-random.out, we can see that now memmove_falkor is about 5% faster than memcpy_falkor_old, while memmove_falkor_old was more than 15% slower. The memcpy implementation remained largely unmodified, so there is no significant performance change. The reason for such a significant memmove performance gain is the increase of the upper bound on the small copy case to 32 bytes and the increase of the upper bound on the medium copy case to 128 bytes. Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Diffstat (limited to 'sysdeps/aarch64/multiarch/memcpy_falkor.S')
-rw-r--r-- | sysdeps/aarch64/multiarch/memcpy_falkor.S | 162 |
1 files changed, 143 insertions, 19 deletions
diff --git a/sysdeps/aarch64/multiarch/memcpy_falkor.S b/sysdeps/aarch64/multiarch/memcpy_falkor.S index 35a1fae1b9..8dfc2c759f 100644 --- a/sysdeps/aarch64/multiarch/memcpy_falkor.S +++ b/sysdeps/aarch64/multiarch/memcpy_falkor.S @@ -42,6 +42,9 @@ #define E_q q4 #define F_q q5 #define G_q q6 +#define H_q q7 +#define Q_q q6 +#define S_q q22 /* Copies are split into 3 main cases: @@ -49,7 +52,7 @@ 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 + Large copies align the source to a quad word and use an unrolled loop processing 64 bytes per iteration. FALKOR-SPECIFIC DESIGN: @@ -70,33 +73,38 @@ #if IS_IN (libc) ENTRY_ALIGN (__memcpy_falkor, 6) + DELOUSE (0) + DELOUSE (1) + DELOUSE (2) + cmp count, 32 add srcend, src, count add dstend, dstin, count b.ls L(copy32) - ldr A_q, [src] cmp count, 128 - str A_q, [dstin] b.hi L(copy_long) /* Medium copies: 33..128 bytes. */ +L(copy128): sub tmp1, count, 1 - ldr A_q, [src, 16] - ldr B_q, [srcend, -32] - ldr C_q, [srcend, -16] + ldr A_q, [src] + ldr B_q, [src, 16] + ldr C_q, [srcend, -32] + ldr D_q, [srcend, -16] tbz tmp1, 6, 1f - ldr D_q, [src, 32] - ldr E_q, [src, 48] - str D_q, [dstin, 32] - str E_q, [dstin, 48] - ldr F_q, [srcend, -64] - ldr G_q, [srcend, -48] - str F_q, [dstend, -64] - str G_q, [dstend, -48] + ldr E_q, [src, 32] + ldr F_q, [src, 48] + ldr G_q, [srcend, -64] + ldr H_q, [srcend, -48] + str G_q, [dstend, -64] + str H_q, [dstend, -48] + str E_q, [dstin, 32] + str F_q, [dstin, 48] 1: - str A_q, [dstin, 16] - str B_q, [dstend, -32] - str C_q, [dstend, -16] + str A_q, [dstin] + str B_q, [dstin, 16] + str C_q, [dstend, -32] + str D_q, [dstend, -16] ret .p2align 4 @@ -152,10 +160,14 @@ L(copy32): 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 + + .p2align 4 + nop /* Align loop64 below. */ L(copy_long): + ldr A_q, [src] sub count, count, 64 + 16 and tmp1, src, 15 + str A_q, [dstin] bic src, src, 15 sub dst, dstin, tmp1 add count, count, tmp1 @@ -175,7 +187,6 @@ 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): ldr E_q, [srcend, -64] str E_q, [dstend, -64] ldr D_q, [srcend, -48] @@ -188,4 +199,117 @@ L(last64): END (__memcpy_falkor) libc_hidden_builtin_def (__memcpy_falkor) + + +/* RATIONALE: + + The move has 4 distinct parts: + * Small moves of 32 bytes and under. + * Medium sized moves of 33-128 bytes (fully unrolled). + * Large moves where the source address is higher than the destination + (forward copies) + * Large moves where the destination address is higher than the source + (copy backward, or move). + + We use only two registers q6 and q22 for the moves and move 32 bytes at a + time to correctly train the hardware prefetcher for better throughput. + + For small and medium cases memcpy is used. */ + +ENTRY_ALIGN (__memmove_falkor, 6) + + DELOUSE (0) + DELOUSE (1) + DELOUSE (2) + + cmp count, 32 + add srcend, src, count + add dstend, dstin, count + b.ls L(copy32) + cmp count, 128 + b.ls L(copy128) + sub tmp1, dstin, src + ccmp tmp1, count, 2, hi + b.lo L(move_long) + + /* CASE: Copy Forwards + + Align src to 16 byte alignment so that we don't cross cache line + boundaries on both loads and stores. There are at least 128 bytes + to copy, so copy 16 bytes unaligned and then align. The loop + copies 32 bytes per iteration and prefetches one iteration ahead. */ + + ldr S_q, [src] + and tmp1, src, 15 + bic src, src, 15 + sub dst, dstin, tmp1 + add count, count, tmp1 /* Count is now 16 too large. */ + ldr Q_q, [src, 16]! + str S_q, [dstin] + ldr S_q, [src, 16]! + sub count, count, 32 + 32 + 16 /* Test and readjust count. */ + + .p2align 4 +1: + subs count, count, 32 + str Q_q, [dst, 16] + ldr Q_q, [src, 16]! + str S_q, [dst, 32]! + ldr S_q, [src, 16]! + b.hi 1b + + /* Copy 32 bytes from the end before writing the data prefetched in the + last loop iteration. */ +2: + ldr B_q, [srcend, -32] + ldr C_q, [srcend, -16] + str Q_q, [dst, 16] + str S_q, [dst, 32] + str B_q, [dstend, -32] + str C_q, [dstend, -16] + ret + + /* CASE: Copy Backwards + + Align srcend to 16 byte alignment so that we don't cross cache line + boundaries on both loads and stores. There are at least 128 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 + nop + nop +L(move_long): + cbz tmp1, 3f /* Return early if src == dstin */ + ldr S_q, [srcend, -16] + and tmp1, srcend, 15 + sub srcend, srcend, tmp1 + ldr Q_q, [srcend, -16]! + str S_q, [dstend, -16] + sub count, count, tmp1 + ldr S_q, [srcend, -16]! + sub dstend, dstend, tmp1 + sub count, count, 32 + 32 + +1: + subs count, count, 32 + str Q_q, [dstend, -16] + ldr Q_q, [srcend, -16]! + str S_q, [dstend, -32]! + ldr S_q, [srcend, -16]! + b.hi 1b + + /* Copy 32 bytes from the start before writing the data prefetched in the + last loop iteration. */ + + ldr B_q, [src, 16] + ldr C_q, [src] + str Q_q, [dstend, -16] + str S_q, [dstend, -32] + str B_q, [dstin, 16] + str C_q, [dstin] +3: ret + +END (__memmove_falkor) +libc_hidden_builtin_def (__memmove_falkor) #endif |