/* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 2012, Oracle and/or its affiliates. All rights reserved. */ .file "memset.s" /* * This version of memset is tuned for SPARC-T4 and other processors * which share the S3 SPARC architecture core. * * char *memset(sp, c, n) * * Set an array of n chars starting at sp to the character c. * Return sp. * * Fast assembler language version of the following C-program for memset * which represents the `standard' for the C-library. * * void * * memset(void *sp1, int c, size_t n) * { * if (n != 0) { * char *sp = sp1; * do { * *sp++ = (char)c; * } while (--n != 0); * } * return (sp1); * } * * Flow : * * For small 6 or fewer bytes stores, bytes will be stored. * * For less than 32 bytes stores, align the address on 4 byte boundary. * Then store as many 4-byte chunks, followed by trailing bytes. * * For sizes greater than 32 bytes, align the address on 8 byte boundary. * if (count > 64) { * store 8-bytes chunks to align the address on 64 byte boundary * if (count > 1664 ) { * if(value to be set is zero) { * Using BIS stores, set the first long word of each 32-byte * cache line to zero which will also clear the other * three long words of the cache line. * } else { * Using BIS stores, set the first long word of each of * 44 cache lines (32 bytes each) before the main loop * is entered. * In the main loop, continue pre-setting the first long * word of each cache line 44 lines in advance while setting * the other three long words (24 bytes)of each cache line * until fewer than 1536 bytes remain. Then set the remaining * three long words of each cache line that has already had * it's first long word set. * } * store remaining data in 64-byte chunks until less than * 64 bytes remain. * } * Store as many 8-byte chunks, followed by trialing bytes. * * BIS = Block Init Store * Doing the advance store of the first element of the cache line * initiates the displacement of a cache line while only using a single * instruction in the pipeline. That avoids various pipeline delays, * such as filling the miss buffer. The performance effect is * similar to prefetching for normal stores. * The special case for zero fills runs faster and uses fewer instruction * cycles than the normal memset loop. * * We only use BIS for memset of greater than 1664 bytes because a sequence * BIS stores must be followed by a membar #StoreStore. The benefit of * the BIS store must be balanced against the cost of the membar operation. */ #include #include #include #define ASI_PNF 0x82 /* primary no fault */ #define ASI_BLK_P 0xF0 /* block primary */ #define ASI_BLK_INIT_ST_QUAD_LDD_P 0xE2 /* documented name for primary block initializing store */ #define ASI_STBI_P ASI_BLK_INIT_ST_QUAD_LDD_P #define ASI_STBIMRU_P 0xF2 /* * ASI_STBI_P marks the cache line as "least recently used" * which means if many threads are active, it has a high chance * of being pushed out of the cache between the first initializing * store and the final stores. * Thus, we use ASI_STBIMRU_P which marks the cache line as * "most recently used" for all but the last cache line */ ANSI_PRAGMA_WEAK(memset,function) ENTRY(memset) mov %o0, %o5 ! copy sp1 before using it cmp %o2, 7 ! if small counts, just write bytes blu,pn %ncc, .wrchar and %o1, 0xff, %o1 ! o1 is (char)c sll %o1, 8, %o3 or %o1, %o3, %o1 ! now o1 has 2 bytes of c sll %o1, 16, %o3 cmp %o2, 32 blu,pn %ncc, .wdalign or %o1, %o3, %o1 ! now o1 has 4 bytes of c sllx %o1, 32, %o3 or %o1, %o3, %o1 ! now o1 has 8 bytes of c .dbalign: andcc %o5, 7, %o3 ! is sp1 aligned on a 8 byte bound bz,pt %ncc, .blkalign ! already long word aligned sub %o3, 8, %o3 ! -(bytes till long word aligned) add %o2, %o3, %o2 ! update o2 with new count ! Set -(%o3) bytes till sp1 long word aligned 1: stb %o1, [%o5] ! there is at least 1 byte to set inccc %o3 ! byte clearing loop bl,pt %ncc, 1b inc %o5 ! Now sp1 is long word aligned (sp1 is found in %o5) .blkalign: cmp %o2, 64 ! check if there are 64 bytes to set blu,pn %ncc, .wrshort mov %o2, %o3 andcc %o5, 63, %o3 ! is sp1 block aligned? bz,pt %ncc, .blkwr ! now block aligned sub %o3, 64, %o3 ! o3 is -(bytes till block aligned) add %o2, %o3, %o2 ! o2 is the remainder ! Store -(%o3) bytes till dst is block (64 byte) aligned. ! Use long word word stores. ! Recall that dst is already long word aligned 1: addcc %o3, 8, %o3 stx %o1, [%o5] bl,pt %ncc, 1b add %o5, 8, %o5 ! Now sp1 is block aligned .blkwr: andn %o2, 63, %o4 ! calculate size of blocks in bytes cmp %o4, 1664 ! check there are enough bytes to set blu,pn %ncc, 3f ! to justify cost of membar ! must be > pre-cleared lines and %o2, 63, %o3 ! %o3 = bytes left after blk stores. brz,pn %o1, .wrzero ! special case if c1 == 0 mov ASI_STBIMRU_P, %asi ! initial cache-clearing stores ! get store pipeline moving stxa %o1, [%o5+0x00+0x000]%asi stxa %o1, [%o5+0x40+0x000]%asi stxa %o1, [%o5+0x20+0x000]%asi stxa %o1, [%o5+0x60+0x000]%asi stxa %o1, [%o5+0x00+0x080]%asi stxa %o1, [%o5+0x40+0x080]%asi stxa %o1, [%o5+0x20+0x080]%asi stxa %o1, [%o5+0x60+0x080]%asi stxa %o1, [%o5+0x00+0x100]%asi stxa %o1, [%o5+0x40+0x100]%asi stxa %o1, [%o5+0x20+0x100]%asi stxa %o1, [%o5+0x60+0x100]%asi stxa %o1, [%o5+0x00+0x180]%asi stxa %o1, [%o5+0x40+0x180]%asi stxa %o1, [%o5+0x20+0x180]%asi stxa %o1, [%o5+0x60+0x180]%asi stxa %o1, [%o5+0x00+0x200]%asi stxa %o1, [%o5+0x40+0x200]%asi stxa %o1, [%o5+0x20+0x200]%asi stxa %o1, [%o5+0x60+0x200]%asi stxa %o1, [%o5+0x00+0x280]%asi stxa %o1, [%o5+0x40+0x280]%asi stxa %o1, [%o5+0x20+0x280]%asi stxa %o1, [%o5+0x60+0x280]%asi stxa %o1, [%o5+0x00+0x300]%asi stxa %o1, [%o5+0x40+0x300]%asi stxa %o1, [%o5+0x20+0x300]%asi stxa %o1, [%o5+0x60+0x300]%asi stxa %o1, [%o5+0x00+0x380]%asi stxa %o1, [%o5+0x40+0x380]%asi stxa %o1, [%o5+0x20+0x380]%asi stxa %o1, [%o5+0x60+0x380]%asi stxa %o1, [%o5+0x00+0x400]%asi stxa %o1, [%o5+0x40+0x400]%asi stxa %o1, [%o5+0x20+0x400]%asi stxa %o1, [%o5+0x60+0x400]%asi stxa %o1, [%o5+0x00+0x480]%asi stxa %o1, [%o5+0x40+0x480]%asi stxa %o1, [%o5+0x20+0x480]%asi stxa %o1, [%o5+0x60+0x480]%asi stxa %o1, [%o5+0x00+0x500]%asi stxa %o1, [%o5+0x40+0x500]%asi stxa %o1, [%o5+0x20+0x500]%asi stxa %o1, [%o5+0x60+0x500]%asi ! Primary memset loop for large memsets ! Continue with advance stores to keep store pipeline moving sub %o5, 8, %o5 ! adjust %o5 for immediate %asi value 2: stxa %o1, [%o5+0x08+0x580]%asi stxa %o1, [%o5+0x48+0x580]%asi stxa %o1, [%o5+0x28+0x580]%asi stxa %o1, [%o5+0x68+0x580]%asi sub %o4, 128, %o4 stxa %o1, [%o5+0x10]%asi stxa %o1, [%o5+0x50]%asi stxa %o1, [%o5+0x30]%asi stxa %o1, [%o5+0x70]%asi cmp %o4, 1536 ! 1536 = 128+1408 (1408=0x580) stxa %o1, [%o5+0x18]%asi stxa %o1, [%o5+0x58]%asi stxa %o1, [%o5+0x38]%asi stxa %o1, [%o5+0x78]%asi stxa %o1, [%o5+0x20]%asi stxa %o1, [%o5+0x60]%asi ! Adjust %o5 to allow zero offset, allowing immediate %asi value add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBI_P add %o5, 64, %o5 bgu,pt %ncc, 2b stxa %o1, [%o5]ASI_STBI_P ! Clean up loop ! All cache lines have had initial 8 byte store ! Now just fill in remaining 24 bytes .cleanup: stxa %o1, [%o5+0x10]%asi sub %o4, 128, %o4 stxa %o1, [%o5+0x50]%asi stxa %o1, [%o5+0x30]%asi stxa %o1, [%o5+0x70]%asi cmp %o4, 128 stxa %o1, [%o5+0x18]%asi stxa %o1, [%o5+0x58]%asi stxa %o1, [%o5+0x38]%asi stxa %o1, [%o5+0x78]%asi stxa %o1, [%o5+0x20]%asi stxa %o1, [%o5+0x60]%asi ! Adjust %o5 to allow zero offset, allowing immediate %asi value add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBI_P add %o5, 64, %o5 bgu,pt %ncc, .cleanup stxa %o1, [%o5]ASI_STBI_P ba .bsi_done add %o5, 8, %o5 ! restore %o5 offset ! Special case loop for zero fill memsets ! blk store to first 8 bytes of 32 byte cache line clears all 32 bytes. ! For each 64 byte cache line, first 32 bytes marked MRU ! second 32 bytes marked LRU to allow L3 cache to replace it ! .wrzero: mov ASI_STBI_P, %asi .wrzero_loop: stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5]ASI_STBIMRU_P add %o5, 64, %o5 stxa %o1, [%o5-0x1e0]%asi stxa %o1, [%o5-0x1a0]%asi sub %o4, 512, %o4 stxa %o1, [%o5-0x160]%asi stxa %o1, [%o5-0x120]%asi cmp %o4, 512 stxa %o1, [%o5-0xe0]%asi stxa %o1, [%o5-0xa0]%asi stxa %o1, [%o5-0x60]%asi bgu,pt %ncc, .wrzero_loop stxa %o1, [%o5-0x20]%asi .bsi_done: membar #StoreStore ! required by use of Block Store Init mov ASI_PNF, %asi ! restore %asi to default ! ASI_PRIMARY_NOFAULT value 3: cmp %o4, 64 ! check if 64 bytes to set blu %ncc, 5f nop 4: ! set final blocks of 64 bytes stx %o1, [%o5+0x00] stx %o1, [%o5+0x20] stx %o1, [%o5+0x08] stx %o1, [%o5+0x10] stx %o1, [%o5+0x18] subcc %o4, 64, %o4 stx %o1, [%o5+0x28] add %o5, 64, %o5 stx %o1, [%o5-0x10] bgu,pt %ncc, 4b stx %o1, [%o5-0x08] 5: ! Set the remaining long words .wrshort: subcc %o3, 8, %o3 ! Can we store any long words? blu,pn %ncc, .wrchars and %o2, 7, %o2 ! calc bytes left after long words 6: subcc %o3, 8, %o3 stx %o1, [%o5] ! store the long words bgeu,pt %ncc, 6b add %o5, 8, %o5 .wrchars: ! check for extra chars cmp %o2, 0 bne %ncc, 7f nop retl nop .wdalign: andcc %o5, 3, %o3 ! is sp1 aligned on a word boundary bz,pn %ncc, .wrword andn %o2, 3, %o3 ! create word sized count in %o3 dec %o2 ! decrement count stb %o1, [%o5] ! clear a byte b .wdalign inc %o5 ! next byte .wrword: subcc %o3, 4, %o3 st %o1, [%o5] ! 4-byte writing loop bnz,pt %ncc, .wrword add %o5, 4, %o5 and %o2, 3, %o2 ! leftover count, if any .wrchar: ! Set the remaining bytes, if any cmp %o2, 0 be %ncc, .exit nop 7: deccc %o2 stb %o1, [%o5] bgu,pt %ncc, 7b inc %o5 .exit: retl ! %o0 was preserved nop SET_SIZE(memset)