diff options
Diffstat (limited to 'contrib/arm-optimized-routines/string/aarch64/strcpy.S')
-rw-r--r-- | contrib/arm-optimized-routines/string/aarch64/strcpy.S | 395 |
1 files changed, 120 insertions, 275 deletions
diff --git a/contrib/arm-optimized-routines/string/aarch64/strcpy.S b/contrib/arm-optimized-routines/string/aarch64/strcpy.S index 6e9ed424b693..97ae37ea4229 100644 --- a/contrib/arm-optimized-routines/string/aarch64/strcpy.S +++ b/contrib/arm-optimized-routines/string/aarch64/strcpy.S @@ -1,311 +1,156 @@ /* * strcpy/stpcpy - copy a string returning pointer to start/end. * - * Copyright (c) 2013-2020, Arm Limited. - * SPDX-License-Identifier: MIT + * Copyright (c) 2020-2023, Arm Limited. + * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception */ /* Assumptions: * - * ARMv8-a, AArch64, unaligned accesses, min page size 4k. + * ARMv8-a, AArch64, Advanced SIMD. + * MTE compatible. */ -#include "../asmdefs.h" +#include "asmdefs.h" -/* To build as stpcpy, define BUILD_STPCPY before compiling this file. - - To test the page crossing code path more thoroughly, compile with - -DSTRCPY_TEST_PAGE_CROSS - this will force all copies through the slower - entry path. This option is not intended for production use. */ - -/* Arguments and results. */ #define dstin x0 #define srcin x1 +#define result x0 -/* Locals and temporaries. */ #define src x2 #define dst x3 -#define data1 x4 -#define data1w w4 -#define data2 x5 -#define data2w w5 -#define has_nul1 x6 -#define has_nul2 x7 -#define tmp1 x8 -#define tmp2 x9 -#define tmp3 x10 -#define tmp4 x11 -#define zeroones x12 -#define data1a x13 -#define data2a x14 -#define pos x15 -#define len x16 -#define to_align x17 +#define len x4 +#define synd x4 +#define tmp x5 +#define shift x5 +#define data1 x6 +#define dataw1 w6 +#define data2 x7 +#define dataw2 w7 + +#define dataq q0 +#define vdata v0 +#define vhas_nul v1 +#define vend v2 +#define dend d2 +#define dataq2 q1 #ifdef BUILD_STPCPY -#define STRCPY __stpcpy_aarch64 -#else -#define STRCPY __strcpy_aarch64 -#endif - - /* NUL detection works on the principle that (X - 1) & (~X) & 0x80 - (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and - can be done in parallel across the entire word. */ - -#define REP8_01 0x0101010101010101 -#define REP8_7f 0x7f7f7f7f7f7f7f7f -#define REP8_80 0x8080808080808080 - - /* AArch64 systems have a minimum page size of 4k. We can do a quick - page size check for crossing this boundary on entry and if we - do not, then we can short-circuit much of the entry code. We - expect early page-crossing strings to be rare (probability of - 16/MIN_PAGE_SIZE ~= 0.4%), so the branch should be quite - predictable, even with random strings. - - We don't bother checking for larger page sizes, the cost of setting - up the correct page size is just not worth the extra gain from - a small reduction in the cases taking the slow path. Note that - we only care about whether the first fetch, which may be - misaligned, crosses a page boundary - after that we move to aligned - fetches for the remainder of the string. */ - -#ifdef STRCPY_TEST_PAGE_CROSS - /* Make everything that isn't Qword aligned look like a page cross. */ -#define MIN_PAGE_P2 4 +# define STRCPY __stpcpy_aarch64 +# define IFSTPCPY(X,...) X,__VA_ARGS__ #else -#define MIN_PAGE_P2 12 +# define STRCPY __strcpy_aarch64 +# define IFSTPCPY(X,...) #endif -#define MIN_PAGE_SIZE (1 << MIN_PAGE_P2) +/* + Core algorithm: + For each 16-byte chunk we calculate a 64-bit nibble mask value with four bits + per byte. We take 4 bits of every comparison byte with shift right and narrow + by 4 instruction. Since the bits in the nibble mask reflect the order in + which things occur in the original string, counting leading zeros identifies + exactly which byte matched. */ ENTRY (STRCPY) PTR_ARG (0) PTR_ARG (1) - /* For moderately short strings, the fastest way to do the copy is to - calculate the length of the string in the same way as strlen, then - essentially do a memcpy of the result. This avoids the need for - multiple byte copies and further means that by the time we - reach the bulk copy loop we know we can always use DWord - accesses. We expect __strcpy_aarch64 to rarely be called repeatedly - with the same source string, so branch prediction is likely to - always be difficult - we mitigate against this by preferring - conditional select operations over branches whenever this is - feasible. */ - and tmp2, srcin, #(MIN_PAGE_SIZE - 1) - mov zeroones, #REP8_01 - and to_align, srcin, #15 - cmp tmp2, #(MIN_PAGE_SIZE - 16) - neg tmp1, to_align - /* The first fetch will straddle a (possible) page boundary iff - srcin + 15 causes bit[MIN_PAGE_P2] to change value. A 16-byte - aligned string will never fail the page align check, so will - always take the fast path. */ - b.gt L(page_cross) - -L(page_cross_ok): - ldp data1, data2, [srcin] -#ifdef __AARCH64EB__ - /* Because we expect the end to be found within 16 characters - (profiling shows this is the most common case), it's worth - swapping the bytes now to save having to recalculate the - termination syndrome later. We preserve data1 and data2 - so that we can re-use the values later on. */ - rev tmp2, data1 - sub tmp1, tmp2, zeroones - orr tmp2, tmp2, #REP8_7f - bics has_nul1, tmp1, tmp2 - b.ne L(fp_le8) - rev tmp4, data2 - sub tmp3, tmp4, zeroones - orr tmp4, tmp4, #REP8_7f -#else - sub tmp1, data1, zeroones - orr tmp2, data1, #REP8_7f - bics has_nul1, tmp1, tmp2 - b.ne L(fp_le8) - sub tmp3, data2, zeroones - orr tmp4, data2, #REP8_7f + bic src, srcin, 15 + ld1 {vdata.16b}, [src] + cmeq vhas_nul.16b, vdata.16b, 0 + lsl shift, srcin, 2 + shrn vend.8b, vhas_nul.8h, 4 + fmov synd, dend + lsr synd, synd, shift + cbnz synd, L(tail) + + ldr dataq, [src, 16]! + cmeq vhas_nul.16b, vdata.16b, 0 + shrn vend.8b, vhas_nul.8h, 4 + fmov synd, dend + cbz synd, L(start_loop) + +#ifndef __AARCH64EB__ + rbit synd, synd #endif - bics has_nul2, tmp3, tmp4 - b.eq L(bulk_entry) + sub tmp, src, srcin + clz len, synd + add len, tmp, len, lsr 2 + tbz len, 4, L(less16) + sub tmp, len, 15 + ldr dataq, [srcin] + ldr dataq2, [srcin, tmp] + str dataq, [dstin] + str dataq2, [dstin, tmp] + IFSTPCPY (add result, dstin, len) + ret - /* The string is short (<=16 bytes). We don't know exactly how - short though, yet. Work out the exact length so that we can - quickly select the optimal copy strategy. */ -L(fp_gt8): - rev has_nul2, has_nul2 - clz pos, has_nul2 - mov tmp2, #56 - add dst, dstin, pos, lsr #3 /* Bits to bytes. */ - sub pos, tmp2, pos -#ifdef __AARCH64EB__ - lsr data2, data2, pos -#else - lsl data2, data2, pos -#endif - str data2, [dst, #1] +L(tail): + rbit synd, synd + clz len, synd + lsr len, len, 2 +L(less16): + tbz len, 3, L(less8) + sub tmp, len, 7 + ldr data1, [srcin] + ldr data2, [srcin, tmp] str data1, [dstin] -#ifdef BUILD_STPCPY - add dstin, dst, #8 -#endif + str data2, [dstin, tmp] + IFSTPCPY (add result, dstin, len) ret -L(fp_le8): - rev has_nul1, has_nul1 - clz pos, has_nul1 - add dst, dstin, pos, lsr #3 /* Bits to bytes. */ - subs tmp2, pos, #24 /* Pos in bits. */ - b.lt L(fp_lt4) -#ifdef __AARCH64EB__ - mov tmp2, #56 - sub pos, tmp2, pos - lsr data2, data1, pos - lsr data1, data1, #32 -#else - lsr data2, data1, tmp2 -#endif - /* 4->7 bytes to copy. */ - str data2w, [dst, #-3] - str data1w, [dstin] -#ifdef BUILD_STPCPY - mov dstin, dst -#endif + .p2align 4 +L(less8): + subs tmp, len, 3 + b.lo L(less4) + ldr dataw1, [srcin] + ldr dataw2, [srcin, tmp] + str dataw1, [dstin] + str dataw2, [dstin, tmp] + IFSTPCPY (add result, dstin, len) ret -L(fp_lt4): - cbz pos, L(fp_lt2) - /* 2->3 bytes to copy. */ -#ifdef __AARCH64EB__ - lsr data1, data1, #48 -#endif - strh data1w, [dstin] - /* Fall-through, one byte (max) to go. */ -L(fp_lt2): - /* Null-terminated string. Last character must be zero! */ - strb wzr, [dst] -#ifdef BUILD_STPCPY - mov dstin, dst -#endif - ret - - .p2align 6 - /* Aligning here ensures that the entry code and main loop all lies - within one 64-byte cache line. */ -L(bulk_entry): - sub to_align, to_align, #16 - stp data1, data2, [dstin] - sub src, srcin, to_align - sub dst, dstin, to_align - b L(entry_no_page_cross) - - /* The inner loop deals with two Dwords at a time. This has a - slightly higher start-up cost, but we should win quite quickly, - especially on cores with a high number of issue slots per - cycle, as we get much better parallelism out of the operations. */ -L(main_loop): - stp data1, data2, [dst], #16 -L(entry_no_page_cross): - ldp data1, data2, [src], #16 - sub tmp1, data1, zeroones - orr tmp2, data1, #REP8_7f - sub tmp3, data2, zeroones - orr tmp4, data2, #REP8_7f - bic has_nul1, tmp1, tmp2 - bics has_nul2, tmp3, tmp4 - ccmp has_nul1, #0, #0, eq /* NZCV = 0000 */ - b.eq L(main_loop) - /* Since we know we are copying at least 16 bytes, the fastest way - to deal with the tail is to determine the location of the - trailing NUL, then (re)copy the 16 bytes leading up to that. */ - cmp has_nul1, #0 -#ifdef __AARCH64EB__ - /* For big-endian, carry propagation (if the final byte in the - string is 0x01) means we cannot use has_nul directly. The - easiest way to get the correct byte is to byte-swap the data - and calculate the syndrome a second time. */ - csel data1, data1, data2, ne - rev data1, data1 - sub tmp1, data1, zeroones - orr tmp2, data1, #REP8_7f - bic has_nul1, tmp1, tmp2 -#else - csel has_nul1, has_nul1, has_nul2, ne -#endif - rev has_nul1, has_nul1 - clz pos, has_nul1 - add tmp1, pos, #72 - add pos, pos, #8 - csel pos, pos, tmp1, ne - add src, src, pos, lsr #3 - add dst, dst, pos, lsr #3 - ldp data1, data2, [src, #-32] - stp data1, data2, [dst, #-16] -#ifdef BUILD_STPCPY - sub dstin, dst, #1 -#endif +L(less4): + cbz len, L(zerobyte) + ldrh dataw1, [srcin] + strh dataw1, [dstin] +L(zerobyte): + strb wzr, [dstin, len] + IFSTPCPY (add result, dstin, len) ret -L(page_cross): - bic src, srcin, #15 - /* Start by loading two words at [srcin & ~15], then forcing the - bytes that precede srcin to 0xff. This means they never look - like termination bytes. */ - ldp data1, data2, [src] - lsl tmp1, tmp1, #3 /* Bytes beyond alignment -> bits. */ - tst to_align, #7 - csetm tmp2, ne -#ifdef __AARCH64EB__ - lsl tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */ -#else - lsr tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */ + .p2align 4 +L(start_loop): + sub tmp, srcin, dstin + ldr dataq2, [srcin] + sub dst, src, tmp + str dataq2, [dstin] +L(loop): + str dataq, [dst], 32 + ldr dataq, [src, 16] + cmeq vhas_nul.16b, vdata.16b, 0 + umaxp vend.16b, vhas_nul.16b, vhas_nul.16b + fmov synd, dend + cbnz synd, L(loopend) + str dataq, [dst, -16] + ldr dataq, [src, 32]! + cmeq vhas_nul.16b, vdata.16b, 0 + umaxp vend.16b, vhas_nul.16b, vhas_nul.16b + fmov synd, dend + cbz synd, L(loop) + add dst, dst, 16 +L(loopend): + shrn vend.8b, vhas_nul.8h, 4 /* 128->64 */ + fmov synd, dend + sub dst, dst, 31 +#ifndef __AARCH64EB__ + rbit synd, synd #endif - orr data1, data1, tmp2 - orr data2a, data2, tmp2 - cmp to_align, #8 - csinv data1, data1, xzr, lt - csel data2, data2, data2a, lt - sub tmp1, data1, zeroones - orr tmp2, data1, #REP8_7f - sub tmp3, data2, zeroones - orr tmp4, data2, #REP8_7f - bic has_nul1, tmp1, tmp2 - bics has_nul2, tmp3, tmp4 - ccmp has_nul1, #0, #0, eq /* NZCV = 0000 */ - b.eq L(page_cross_ok) - /* We now need to make data1 and data2 look like they've been - loaded directly from srcin. Do a rotate on the 128-bit value. */ - lsl tmp1, to_align, #3 /* Bytes->bits. */ - neg tmp2, to_align, lsl #3 -#ifdef __AARCH64EB__ - lsl data1a, data1, tmp1 - lsr tmp4, data2, tmp2 - lsl data2, data2, tmp1 - orr tmp4, tmp4, data1a - cmp to_align, #8 - csel data1, tmp4, data2, lt - rev tmp2, data1 - rev tmp4, data2 - sub tmp1, tmp2, zeroones - orr tmp2, tmp2, #REP8_7f - sub tmp3, tmp4, zeroones - orr tmp4, tmp4, #REP8_7f -#else - lsr data1a, data1, tmp1 - lsl tmp4, data2, tmp2 - lsr data2, data2, tmp1 - orr tmp4, tmp4, data1a - cmp to_align, #8 - csel data1, tmp4, data2, lt - sub tmp1, data1, zeroones - orr tmp2, data1, #REP8_7f - sub tmp3, data2, zeroones - orr tmp4, data2, #REP8_7f -#endif - bic has_nul1, tmp1, tmp2 - cbnz has_nul1, L(fp_le8) - bic has_nul2, tmp3, tmp4 - b L(fp_gt8) + clz len, synd + lsr len, len, 2 + add dst, dst, len + ldr dataq, [dst, tmp] + str dataq, [dst] + IFSTPCPY (add result, dst, 15) + ret END (STRCPY) - |