; ; strlen.asm ; ; Copyright (c) Microsoft Corporation. All rights reserved. ; ; Optimized strlen and strnlen implementations for ARM64. ; #include "ksarm64.h" ; size_t strlen(const char *str); // AV's when str == NULL ; size_t strnlen(const char *str, size_t numberOfElements); // AV's when str == NULL ; This file could also define strnlen_s. strnlen_s is currently defined in the header string.h in C ; using a check for null and a call to strnlen. This avoids making the call in the case where the string is null, ; which should be infrequent. However it makes code larger by inlining that check everywhere strnlen_s is called. ; A better alternative would be to modify the standard headers and define strnlen_s here. It would be just one ; instruction because the required return value in x0 is already 0 when null is passed for the first parameter. ; ; EXPORT strnlen_s [FUNC] ; LEAF_ENTRY strnlen_s ; cbz x0, AnyRet ; add the label AnyRet in front of any ret instruction you want ; ; fallthrough into strnlen code ; ALTERNATE_ENTRY strnlen ; change LEAF_ENTRY for strnlen to ALTERNATE_ENTRY ; ... ; current body of strnlen ; LEAF_END ; change strnlen leaf_end to strnlen_s. #if !defined(_M_ARM64EC) EXPORT A64NAME(strlen) [FUNC] EXPORT A64NAME(strnlen) [FUNC] #endif SET_COMDAT_ALIGNMENT 5 ; With strlen we will usually read some bytes past the end of the string. To avoid getting an AV ; when a byte-by-byte implementation would not, we must ensure that we never cross a page boundary with a ; vector load, so we must align the vector loads to 16-byte-aligned boundaries. ; ; For strnlen we know the buffer length and so we won't read any bytes beyond the end of the buffer. This means ; we have a choice whether to arrange our vector loads to be 16-byte aligned. (Note that on arm64 a vector load ; only produces an alignment fault when the vector *elements* are misaligned, so a "16B" vector load will never ; give an alignment fault for user memory). Aligning the vector loads on 16-byte boundaries saves one cycle ; per vector load instruction. The cost of forcing 16-byte aligned loads is the 10 instructions preceding the ; 'NoNeedToAlign' label below. On Cortex-A57, the execution latency of those 10 instructions is 27 cycles, ; assuming no branch mispredict on the 'beq'. To account for the cost of an occasional mispredict we guess a ; mispredict rate of 2% and a mispredict cost of 50 cycles, or 1 cycle per call amortized, 28 total. 28 * 16 = 448. ; In this analysis we are ignoring the chance of extra cache misses due to loads crossing cache lines when ; they are not 16-byte aligned. When the vector loads span cache line boundaries each cache line is referenced ; one more time than it is when the loads are aligned. But we assume that the cache line stays loaded for the ; short time we need to do all the references to it, and so one extra reference won't matter. ; It is expected that the number of cycles (28) will stay about the same for future processor models. If it ; changes radically, it will be worth converting the EQU to a global, using ldr to load it instead of a ; mov-immediate, and dynamically setting the global during CRT startup based on processor model. __strnlen_forceAlignThreshold EQU 448 ; code below assumes must be >= 32 ; If a strlen is performed on an unterminated buffer, strlen may try to access an invalid address ; and generate an access violation. The prime imperative is for us not to generate an AV by loading ; characters beyond the end of a valid string. But also, in the case of an invalid string that should ; generate an AV, we need to have the AV report the proper bad address. If we only perform 16-byte aligned ; vector loads then the first time we touch an invalid page we will be loading from offset 0 in that ; page, which is the correct address for the AV to report. Even though we will usually load some bytes ; from beyond the end of the string, we won't load bytes beyond the end of the page unless the string ; extends into the next page. This is the primary purpose for forcing the vector loads to be ; 16-byte aligned. ; There is a slight performance boost for using aligned vector loads vs. unaligned ones but ; it is only worth the cost of aligning for longer strings (at least 512 chars). ; For strings less than 16 characters long the byte-by-byte loop will be about as fast as the ; compiler could produce, so we're not losing any performance vs. compiled C in any case. ARM64EC_ENTRY_THUNK A64NAME(strlen),1,0 LEAF_ENTRY_COMDAT A64NAME(strlen) ; check for empty string to avoid huge perf degradation in this case. ldrb w2, [x0], #0 cbz w2, EmptyStr mov x5, x0 ; keep original x0 value for the final 'sub' ; calculate number of bytes until first 16-byte alignment point ands x1, x5, #15 ; x1 = (addr mod 16) beq StrlenMainLoop ; no need to force alignment if already aligned ; we need to align, check whether we are within 16 bytes of the end of the page ands x2, x5, #4095 cmp x2, #4080 bgt AlignByteByByte ; too close to end of page, must align byte-by-byte ; safe to do one unaligned 16-byte vector load to force alignment ld1 v0.16b, [x5] ; don't post-increment x5 uminv b1, v0.16b fmov w2, s1 ; fmov is sometimes 1 cycle faster than 'umov w2, v1.b[0]' cbz w2, FindNullInVector ; jump when string <= 15 bytes long & not near end of page add x5, x5, #16 ; move x5 forward only to aligned address and x5, x5, 0xFFFFFFFFFFFFFFF0 ; first iter of StrlenMainLoop will retest some bytes we already tested StrlenMainLoop ; test 16 bytes at a time until we find it ld1 v0.16b, [x5], #16 uminv b1, v0.16b ; use unsigned min to look for a zero byte; too bad it doesn't set CC fmov w2, s1 ; need to move min byte into gpr to test it cbnz w2, StrlenMainLoop ; fall through when any one of the bytes in v0 is zero sub x5, x5, #16 ; undo the last #16 post-increment of x5 FindNullInVector ; this label is also the target of a jump from strnlen ldr q1, ReverseBytePos ; load the position indicator mask cmeq v0.16b, v0.16b, #0 ; +---- and v0.16b, v0.16b, v1.16b ; | umaxv b0, v0.16b ; | see big comment below fmov w2, s0 ; | eor w2, w2, #15 ; +---- add x5, x5, x2 ; which is the offset we need to add to x5 to point at the null byte sub x0, x5, x0 ; subtract ptr to null char from ptr to first char to get the strlen ret ByteByByteFoundIt ; this label is also the target of a jump from strnlen sub x5, x5, #1 ; Undo the final post-increment that happened on the load of the null char. sub x0, x5, x0 ; With x5 pointing at the null char, x5-x0 is the strlen ret AlignByteByByte sub x1, x1, #16 ; x1 = (addr mod 16) - 16 neg x1, x1 ; x1 = 16 - (addr mod 16) = count for byte-by-byte loop ByteByByteLoop ; test one byte at a time until we are 16-byte aligned ldrb w2, [x5], #1 cbz w2, ByteByByteFoundIt ; branch if byte-at-a-time testing finds the null subs x1, x1, #1 bgt ByteByByteLoop ; fall through when not found and 16-byte aligned b StrlenMainLoop EmptyStr mov x0, 0 ret ; The challenge is to find a way to efficiently determine which of the 16 bytes we loaded is the end of the string. ; The trick is to load a position indicator mask and generate the position of the rightmost null from that. ; Little-endian order means when we load the mask below v1.16b[0] has 0x0F, and v0.16b[0] is the byte of the string ; that comes first of the 16 we loaded. We do a cmeq, mapping all the characters we loaded to either 0xFF (for nulls) ; or 0x00 for non-nulls. Then we and with the mask below. SIMD lanes corresponding to a non-null character will be 0, ; and SIMD lanes corresponding to null bytes will have a byte from the mask. We take the max across the bytes of the ; vector to find the highest position that corresponds to a null character. The numbering order means we find the ; rightmost null in the vector, which is the null that occurred first in memory due to little endian loading. ; Exclusive oring the position indicator byte with 15 inverts the order, which gives us the offset of the null ; counting from the first character we loaded into the v0 SIMD reg. ReverseBytePos \ dcb 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 LEAF_END ARM64EC_ENTRY_THUNK A64NAME(strnlen),1,0 LEAF_ENTRY_COMDAT A64NAME(strnlen) ; start here for strnlen mov x5, x0 cmp x1, #16 ; x1 has length. When < 16 we have to go byte-by-byte blo ShortStrNLen ; only do byte-by-byte for 0 to 15 bytes. ; Do an aligned strnlen ands x3, x5, #15 ; x3 = start address mod 16 beq NoNeedToAlign ; branch on x3 == 0 because its already aligned ands x2, x5, #4095 ; x2 = start address mod (PAGE_SIZE - 1) cmp x2, #4080 bgt AlignByteByByte_Strnlen ; too close to end of page, must align byte-by-byte sub x3, x3, #16 ; x3 = (start address mod 16) - 16 neg x3, x3 ; x3 = 16 - (start address mod 16) = number of bytes to advance to get aligned ld1 v0.16b, [x5] ; don't post-increment uminv b1, v0.16b fmov w2, s1 ; fmov is sometimes 1 cycle faster than 'umov w2, v1.b[0]' cbz w2, FindNullInVector_Strnlen ; jump out when null found in first 16 bytes sub x1, x1, x3 ; reduce length remaining by number of bytes needed to get aligned add x5, x5, x3 ; move x5 forward only to aligned address ResumeAfterAlignByteByByte cmp x1, #16 ; check for size < 16 after alignment adjustment blo ShortStrNLen NoNeedToAlign asr x3, x1, #4 ; set up number of interations remaining after alignment point reached ; no need to check here for x3 == 0 because: ; - if we didn't align it, it is at least 16 bytes long ; - if we did align it, we checked for <16 before coming here StrNlenMainLoop ; test 16 bytes at a time until we find it ld1 v0.16b, [x5], #16 uminv b1, v0.16b ; use unsigned min to look for a zero byte; too bad it doesn't set CC fmov w2, s1 ; need to move min byte into gpr to test it cbz w2, UndoPI_FindNullInVector ; jump out to when any one of the bytes in v0 is zero subs x3, x3, #1 bne StrNlenMainLoop ands x1, x1, #15 ; check for remainder beq StrNLenOverrun ; orig buffer size was multiple of 16 bytes so no remainder; goto overrun case ; We're within 16 bytes of the end of the buffer and haven't found a '\0' yet. We know we were originally longer than ; 16 bytes so we can do an unaligned vector compare of the last 16 bytes of the buffer, overlapping with some bytes ; we already know are non-zero, without fear of underrunning the original front of the buffer. This avoids a more costly ; byte-by-byte comparison for the remainder (which would average 32 instructions executed and two branch mispredicts). ; At this point: ; x5 points at one of the last 15 chars of the buffer ; x1 has the number of chars remaining in the buffer. 1 <= x1 <= 15 ; 16 - x1 is the number of characters we have to 'back up' FastRemainderHandling sub x1, x1, #16 neg x1, x1 sub x5, x5, x1 ld1 v0.16b, [x5], #16 uminv b1, v0.16b fmov w2, s1 ; fmov is sometimes 1 cycle faster than 'umov w2, v1.b[0]' cbz w2, UndoPI_FindNullInVector ; found a '\0' b StrNLenOverrun ; x5 points one past end of buffer, we're all set for the overrun exit. ShortStrNLen cbz x1, StrNLenOverrun ; if original length was zero, we must return 0 without touching the buffer ShortStrNLenLoop ldrb w2, [x5], #1 cbz w2, ByteByByteFoundIt_Strnlen ; jump into other function to avoid code duplication subs x1, x1, #1 bhi ShortStrNLenLoop StrNLenOverrun sub x0, x5, x0 ; x5 points one past the end of the buffer, x5-x0 is original numberOfElements ret AlignByteByByte_Strnlen sub x3, x3, #16 ; x3 = (addr mod 16) - 16 neg x3, x3 ; x3 = 16 - (addr mod 16) = count for byte-by-byte loop ByteByByteLoop_Strnlen ; test one byte at a time until we are 16-byte aligned ldrb w2, [x5], #1 cbz w2, ByteByByteFoundIt_Strnlen ; branch if byte-at-a-time testing finds the null subs x1, x1, #1 ; check remaining length = 0 beq ByteByByteReachedMax_Strnlen ; branch if byte-at-a-time testing reached end of buffer count subs x3, x3, #1 bgt ByteByByteLoop_Strnlen ; fall through when not found and 16-byte aligned b ResumeAfterAlignByteByByte UndoPI_FindNullInVector ; this label is the target of a jump from strnlen sub x5, x5, #16 ; undo the last #16 post-increment of x5 FindNullInVector_Strnlen ; this label is also the target of a jump from strnlen ldr q1, ReverseBytePos_Strnlen ; load the position indicator mask cmeq v0.16b, v0.16b, #0 ; +---- and v0.16b, v0.16b, v1.16b ; | umaxv b0, v0.16b ; | see big comment below fmov w2, s0 ; | eor w2, w2, #15 ; +---- add x5, x5, x2 ; which is the offset we need to add to x5 to point at the null byte sub x0, x5, x0 ; subtract ptr to null char from ptr to first char to get the strlen ret ByteByByteFoundIt_Strnlen ; this label is also the target of a jump from strnlen sub x5, x5, #1 ; Undo the final post-increment that happened on the load of the null char. ByteByByteReachedMax_Strnlen sub x0, x5, x0 ; With x5 pointing at the null char, x5-x0 is the strlen ret ReverseBytePos_Strnlen \ dcb 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0 LEAF_END END