Path: blob/main/contrib/bionic-x86_64-string/sse4-memcmp-slm.S
39475 views
/*1Copyright (c) 2014, Intel Corporation2All rights reserved.34Redistribution and use in source and binary forms, with or without5modification, are permitted provided that the following conditions are met:67* Redistributions of source code must retain the above copyright notice,8* this list of conditions and the following disclaimer.910* Redistributions in binary form must reproduce the above copyright notice,11* this list of conditions and the following disclaimer in the documentation12* and/or other materials provided with the distribution.1314* Neither the name of Intel Corporation nor the names of its contributors15* may be used to endorse or promote products derived from this software16* without specific prior written permission.1718THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND19ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED20WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE21DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR22ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES23(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;24LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON25ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT26(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS27SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.28*/2930#include "cache.h"3132#ifndef MEMCMP33# define MEMCMP memcmp34#endif3536#ifndef L37# define L(label) .L##label38#endif3940#ifndef ALIGN41# define ALIGN(n) .p2align n42#endif4344#ifndef cfi_startproc45# define cfi_startproc .cfi_startproc46#endif4748#ifndef cfi_endproc49# define cfi_endproc .cfi_endproc50#endif5152#ifndef ENTRY53# define ENTRY(name) \54.type name, @function; \55.globl name; \56.p2align 4; \57name: \58cfi_startproc59#endif6061#ifndef END62# define END(name) \63cfi_endproc; \64.size name, .-name65#endif6667#ifndef ALIGN68# define ALIGN(n) .p2align n69#endif7071#define JMPTBL(I, B) (I - B)7273#define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \74lea TABLE(%rip), %r11; \75movslq (%r11, INDEX, SCALE), %rcx; \76add %r11, %rcx; \77jmp *%rcx; \78ud27980.section .text.sse4.1,"ax",@progbits81ENTRY (MEMCMP)82#ifdef USE_AS_WMEMCMP83shl $2, %rdx84#endif85pxor %xmm0, %xmm086cmp $79, %rdx87ja L(79bytesormore)88#ifndef USE_AS_WMEMCMP89cmp $1, %rdx90je L(firstbyte)91#endif92add %rdx, %rsi93add %rdx, %rdi94BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)9596#ifndef USE_AS_WMEMCMP97ALIGN (4)98L(firstbyte):99movzbl (%rdi), %eax100movzbl (%rsi), %ecx101sub %ecx, %eax102ret103#endif104105ALIGN (4)106L(79bytesormore):107movdqu (%rsi), %xmm1108movdqu (%rdi), %xmm2109pxor %xmm1, %xmm2110ptest %xmm2, %xmm0111jnc L(16bytesin256)112mov %rsi, %rcx113and $-16, %rsi114add $16, %rsi115sub %rsi, %rcx116117sub %rcx, %rdi118add %rcx, %rdx119test $0xf, %rdi120jz L(2aligned)121122cmp $128, %rdx123ja L(128bytesormore)124L(less128bytes):125sub $64, %rdx126127movdqu (%rdi), %xmm2128pxor (%rsi), %xmm2129ptest %xmm2, %xmm0130jnc L(16bytesin256)131132movdqu 16(%rdi), %xmm2133pxor 16(%rsi), %xmm2134ptest %xmm2, %xmm0135jnc L(32bytesin256)136137movdqu 32(%rdi), %xmm2138pxor 32(%rsi), %xmm2139ptest %xmm2, %xmm0140jnc L(48bytesin256)141142movdqu 48(%rdi), %xmm2143pxor 48(%rsi), %xmm2144ptest %xmm2, %xmm0145jnc L(64bytesin256)146cmp $32, %rdx147jb L(less32bytesin64)148149movdqu 64(%rdi), %xmm2150pxor 64(%rsi), %xmm2151ptest %xmm2, %xmm0152jnc L(80bytesin256)153154movdqu 80(%rdi), %xmm2155pxor 80(%rsi), %xmm2156ptest %xmm2, %xmm0157jnc L(96bytesin256)158sub $32, %rdx159add $32, %rdi160add $32, %rsi161L(less32bytesin64):162add $64, %rdi163add $64, %rsi164add %rdx, %rsi165add %rdx, %rdi166BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)167168L(128bytesormore):169cmp $512, %rdx170ja L(512bytesormore)171cmp $256, %rdx172ja L(less512bytes)173L(less256bytes):174sub $128, %rdx175176movdqu (%rdi), %xmm2177pxor (%rsi), %xmm2178ptest %xmm2, %xmm0179jnc L(16bytesin256)180181movdqu 16(%rdi), %xmm2182pxor 16(%rsi), %xmm2183ptest %xmm2, %xmm0184jnc L(32bytesin256)185186movdqu 32(%rdi), %xmm2187pxor 32(%rsi), %xmm2188ptest %xmm2, %xmm0189jnc L(48bytesin256)190191movdqu 48(%rdi), %xmm2192pxor 48(%rsi), %xmm2193ptest %xmm2, %xmm0194jnc L(64bytesin256)195196movdqu 64(%rdi), %xmm2197pxor 64(%rsi), %xmm2198ptest %xmm2, %xmm0199jnc L(80bytesin256)200201movdqu 80(%rdi), %xmm2202pxor 80(%rsi), %xmm2203ptest %xmm2, %xmm0204jnc L(96bytesin256)205206movdqu 96(%rdi), %xmm2207pxor 96(%rsi), %xmm2208ptest %xmm2, %xmm0209jnc L(112bytesin256)210211movdqu 112(%rdi), %xmm2212pxor 112(%rsi), %xmm2213ptest %xmm2, %xmm0214jnc L(128bytesin256)215216add $128, %rsi217add $128, %rdi218219cmp $64, %rdx220jae L(less128bytes)221222cmp $32, %rdx223jb L(less32bytesin128)224225movdqu (%rdi), %xmm2226pxor (%rsi), %xmm2227ptest %xmm2, %xmm0228jnc L(16bytesin256)229230movdqu 16(%rdi), %xmm2231pxor 16(%rsi), %xmm2232ptest %xmm2, %xmm0233jnc L(32bytesin256)234sub $32, %rdx235add $32, %rdi236add $32, %rsi237L(less32bytesin128):238add %rdx, %rsi239add %rdx, %rdi240BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)241242L(less512bytes):243sub $256, %rdx244movdqu (%rdi), %xmm2245pxor (%rsi), %xmm2246ptest %xmm2, %xmm0247jnc L(16bytesin256)248249movdqu 16(%rdi), %xmm2250pxor 16(%rsi), %xmm2251ptest %xmm2, %xmm0252jnc L(32bytesin256)253254movdqu 32(%rdi), %xmm2255pxor 32(%rsi), %xmm2256ptest %xmm2, %xmm0257jnc L(48bytesin256)258259movdqu 48(%rdi), %xmm2260pxor 48(%rsi), %xmm2261ptest %xmm2, %xmm0262jnc L(64bytesin256)263264movdqu 64(%rdi), %xmm2265pxor 64(%rsi), %xmm2266ptest %xmm2, %xmm0267jnc L(80bytesin256)268269movdqu 80(%rdi), %xmm2270pxor 80(%rsi), %xmm2271ptest %xmm2, %xmm0272jnc L(96bytesin256)273274movdqu 96(%rdi), %xmm2275pxor 96(%rsi), %xmm2276ptest %xmm2, %xmm0277jnc L(112bytesin256)278279movdqu 112(%rdi), %xmm2280pxor 112(%rsi), %xmm2281ptest %xmm2, %xmm0282jnc L(128bytesin256)283284movdqu 128(%rdi), %xmm2285pxor 128(%rsi), %xmm2286ptest %xmm2, %xmm0287jnc L(144bytesin256)288289movdqu 144(%rdi), %xmm2290pxor 144(%rsi), %xmm2291ptest %xmm2, %xmm0292jnc L(160bytesin256)293294movdqu 160(%rdi), %xmm2295pxor 160(%rsi), %xmm2296ptest %xmm2, %xmm0297jnc L(176bytesin256)298299movdqu 176(%rdi), %xmm2300pxor 176(%rsi), %xmm2301ptest %xmm2, %xmm0302jnc L(192bytesin256)303304movdqu 192(%rdi), %xmm2305pxor 192(%rsi), %xmm2306ptest %xmm2, %xmm0307jnc L(208bytesin256)308309movdqu 208(%rdi), %xmm2310pxor 208(%rsi), %xmm2311ptest %xmm2, %xmm0312jnc L(224bytesin256)313314movdqu 224(%rdi), %xmm2315pxor 224(%rsi), %xmm2316ptest %xmm2, %xmm0317jnc L(240bytesin256)318319movdqu 240(%rdi), %xmm2320pxor 240(%rsi), %xmm2321ptest %xmm2, %xmm0322jnc L(256bytesin256)323324add $256, %rsi325add $256, %rdi326327cmp $128, %rdx328jae L(less256bytes)329330cmp $64, %rdx331jae L(less128bytes)332333cmp $32, %rdx334jb L(less32bytesin256)335336movdqu (%rdi), %xmm2337pxor (%rsi), %xmm2338ptest %xmm2, %xmm0339jnc L(16bytesin256)340341movdqu 16(%rdi), %xmm2342pxor 16(%rsi), %xmm2343ptest %xmm2, %xmm0344jnc L(32bytesin256)345sub $32, %rdx346add $32, %rdi347add $32, %rsi348L(less32bytesin256):349add %rdx, %rsi350add %rdx, %rdi351BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)352353ALIGN (4)354L(512bytesormore):355#ifdef DATA_CACHE_SIZE_HALF356mov $DATA_CACHE_SIZE_HALF, %r8357#else358mov __x86_64_data_cache_size_half(%rip), %r8359#endif360mov %r8, %r9361shr $1, %r8362add %r9, %r8363cmp %r8, %rdx364ja L(L2_L3_cache_unaglined)365sub $64, %rdx366ALIGN (4)367L(64bytesormore_loop):368movdqu (%rdi), %xmm2369pxor (%rsi), %xmm2370movdqa %xmm2, %xmm1371372movdqu 16(%rdi), %xmm3373pxor 16(%rsi), %xmm3374por %xmm3, %xmm1375376movdqu 32(%rdi), %xmm4377pxor 32(%rsi), %xmm4378por %xmm4, %xmm1379380movdqu 48(%rdi), %xmm5381pxor 48(%rsi), %xmm5382por %xmm5, %xmm1383384ptest %xmm1, %xmm0385jnc L(64bytesormore_loop_end)386add $64, %rsi387add $64, %rdi388sub $64, %rdx389jae L(64bytesormore_loop)390391add $64, %rdx392add %rdx, %rsi393add %rdx, %rdi394BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)395396L(L2_L3_cache_unaglined):397sub $64, %rdx398ALIGN (4)399L(L2_L3_unaligned_128bytes_loop):400prefetchnta 0x1c0(%rdi)401prefetchnta 0x1c0(%rsi)402movdqu (%rdi), %xmm2403pxor (%rsi), %xmm2404movdqa %xmm2, %xmm1405406movdqu 16(%rdi), %xmm3407pxor 16(%rsi), %xmm3408por %xmm3, %xmm1409410movdqu 32(%rdi), %xmm4411pxor 32(%rsi), %xmm4412por %xmm4, %xmm1413414movdqu 48(%rdi), %xmm5415pxor 48(%rsi), %xmm5416por %xmm5, %xmm1417418ptest %xmm1, %xmm0419jnc L(64bytesormore_loop_end)420add $64, %rsi421add $64, %rdi422sub $64, %rdx423jae L(L2_L3_unaligned_128bytes_loop)424425add $64, %rdx426add %rdx, %rsi427add %rdx, %rdi428BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)429430/*431* This case is for machines which are sensitive for unaligned instructions.432*/433ALIGN (4)434L(2aligned):435cmp $128, %rdx436ja L(128bytesormorein2aligned)437L(less128bytesin2aligned):438sub $64, %rdx439440movdqa (%rdi), %xmm2441pxor (%rsi), %xmm2442ptest %xmm2, %xmm0443jnc L(16bytesin256)444445movdqa 16(%rdi), %xmm2446pxor 16(%rsi), %xmm2447ptest %xmm2, %xmm0448jnc L(32bytesin256)449450movdqa 32(%rdi), %xmm2451pxor 32(%rsi), %xmm2452ptest %xmm2, %xmm0453jnc L(48bytesin256)454455movdqa 48(%rdi), %xmm2456pxor 48(%rsi), %xmm2457ptest %xmm2, %xmm0458jnc L(64bytesin256)459cmp $32, %rdx460jb L(less32bytesin64in2alinged)461462movdqa 64(%rdi), %xmm2463pxor 64(%rsi), %xmm2464ptest %xmm2, %xmm0465jnc L(80bytesin256)466467movdqa 80(%rdi), %xmm2468pxor 80(%rsi), %xmm2469ptest %xmm2, %xmm0470jnc L(96bytesin256)471sub $32, %rdx472add $32, %rdi473add $32, %rsi474L(less32bytesin64in2alinged):475add $64, %rdi476add $64, %rsi477add %rdx, %rsi478add %rdx, %rdi479BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)480481ALIGN (4)482L(128bytesormorein2aligned):483cmp $512, %rdx484ja L(512bytesormorein2aligned)485cmp $256, %rdx486ja L(256bytesormorein2aligned)487L(less256bytesin2alinged):488sub $128, %rdx489490movdqa (%rdi), %xmm2491pxor (%rsi), %xmm2492ptest %xmm2, %xmm0493jnc L(16bytesin256)494495movdqa 16(%rdi), %xmm2496pxor 16(%rsi), %xmm2497ptest %xmm2, %xmm0498jnc L(32bytesin256)499500movdqa 32(%rdi), %xmm2501pxor 32(%rsi), %xmm2502ptest %xmm2, %xmm0503jnc L(48bytesin256)504505movdqa 48(%rdi), %xmm2506pxor 48(%rsi), %xmm2507ptest %xmm2, %xmm0508jnc L(64bytesin256)509510movdqa 64(%rdi), %xmm2511pxor 64(%rsi), %xmm2512ptest %xmm2, %xmm0513jnc L(80bytesin256)514515movdqa 80(%rdi), %xmm2516pxor 80(%rsi), %xmm2517ptest %xmm2, %xmm0518jnc L(96bytesin256)519520movdqa 96(%rdi), %xmm2521pxor 96(%rsi), %xmm2522ptest %xmm2, %xmm0523jnc L(112bytesin256)524525movdqa 112(%rdi), %xmm2526pxor 112(%rsi), %xmm2527ptest %xmm2, %xmm0528jnc L(128bytesin256)529530add $128, %rsi531add $128, %rdi532533cmp $64, %rdx534jae L(less128bytesin2aligned)535536cmp $32, %rdx537jb L(less32bytesin128in2aligned)538539movdqu (%rdi), %xmm2540pxor (%rsi), %xmm2541ptest %xmm2, %xmm0542jnc L(16bytesin256)543544movdqu 16(%rdi), %xmm2545pxor 16(%rsi), %xmm2546ptest %xmm2, %xmm0547jnc L(32bytesin256)548sub $32, %rdx549add $32, %rdi550add $32, %rsi551L(less32bytesin128in2aligned):552add %rdx, %rsi553add %rdx, %rdi554BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)555556ALIGN (4)557L(256bytesormorein2aligned):558559sub $256, %rdx560movdqa (%rdi), %xmm2561pxor (%rsi), %xmm2562ptest %xmm2, %xmm0563jnc L(16bytesin256)564565movdqa 16(%rdi), %xmm2566pxor 16(%rsi), %xmm2567ptest %xmm2, %xmm0568jnc L(32bytesin256)569570movdqa 32(%rdi), %xmm2571pxor 32(%rsi), %xmm2572ptest %xmm2, %xmm0573jnc L(48bytesin256)574575movdqa 48(%rdi), %xmm2576pxor 48(%rsi), %xmm2577ptest %xmm2, %xmm0578jnc L(64bytesin256)579580movdqa 64(%rdi), %xmm2581pxor 64(%rsi), %xmm2582ptest %xmm2, %xmm0583jnc L(80bytesin256)584585movdqa 80(%rdi), %xmm2586pxor 80(%rsi), %xmm2587ptest %xmm2, %xmm0588jnc L(96bytesin256)589590movdqa 96(%rdi), %xmm2591pxor 96(%rsi), %xmm2592ptest %xmm2, %xmm0593jnc L(112bytesin256)594595movdqa 112(%rdi), %xmm2596pxor 112(%rsi), %xmm2597ptest %xmm2, %xmm0598jnc L(128bytesin256)599600movdqa 128(%rdi), %xmm2601pxor 128(%rsi), %xmm2602ptest %xmm2, %xmm0603jnc L(144bytesin256)604605movdqa 144(%rdi), %xmm2606pxor 144(%rsi), %xmm2607ptest %xmm2, %xmm0608jnc L(160bytesin256)609610movdqa 160(%rdi), %xmm2611pxor 160(%rsi), %xmm2612ptest %xmm2, %xmm0613jnc L(176bytesin256)614615movdqa 176(%rdi), %xmm2616pxor 176(%rsi), %xmm2617ptest %xmm2, %xmm0618jnc L(192bytesin256)619620movdqa 192(%rdi), %xmm2621pxor 192(%rsi), %xmm2622ptest %xmm2, %xmm0623jnc L(208bytesin256)624625movdqa 208(%rdi), %xmm2626pxor 208(%rsi), %xmm2627ptest %xmm2, %xmm0628jnc L(224bytesin256)629630movdqa 224(%rdi), %xmm2631pxor 224(%rsi), %xmm2632ptest %xmm2, %xmm0633jnc L(240bytesin256)634635movdqa 240(%rdi), %xmm2636pxor 240(%rsi), %xmm2637ptest %xmm2, %xmm0638jnc L(256bytesin256)639640add $256, %rsi641add $256, %rdi642643cmp $128, %rdx644jae L(less256bytesin2alinged)645646cmp $64, %rdx647jae L(less128bytesin2aligned)648649cmp $32, %rdx650jb L(less32bytesin256in2alinged)651652movdqa (%rdi), %xmm2653pxor (%rsi), %xmm2654ptest %xmm2, %xmm0655jnc L(16bytesin256)656657movdqa 16(%rdi), %xmm2658pxor 16(%rsi), %xmm2659ptest %xmm2, %xmm0660jnc L(32bytesin256)661sub $32, %rdx662add $32, %rdi663add $32, %rsi664L(less32bytesin256in2alinged):665add %rdx, %rsi666add %rdx, %rdi667BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)668669ALIGN (4)670L(512bytesormorein2aligned):671#ifdef DATA_CACHE_SIZE_HALF672mov $DATA_CACHE_SIZE_HALF, %r8673#else674mov __x86_64_data_cache_size_half(%rip), %r8675#endif676mov %r8, %r9677shr $1, %r8678add %r9, %r8679cmp %r8, %rdx680ja L(L2_L3_cache_aglined)681682sub $64, %rdx683ALIGN (4)684L(64bytesormore_loopin2aligned):685movdqa (%rdi), %xmm2686pxor (%rsi), %xmm2687movdqa %xmm2, %xmm1688689movdqa 16(%rdi), %xmm3690pxor 16(%rsi), %xmm3691por %xmm3, %xmm1692693movdqa 32(%rdi), %xmm4694pxor 32(%rsi), %xmm4695por %xmm4, %xmm1696697movdqa 48(%rdi), %xmm5698pxor 48(%rsi), %xmm5699por %xmm5, %xmm1700701ptest %xmm1, %xmm0702jnc L(64bytesormore_loop_end)703add $64, %rsi704add $64, %rdi705sub $64, %rdx706jae L(64bytesormore_loopin2aligned)707708add $64, %rdx709add %rdx, %rsi710add %rdx, %rdi711BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)712L(L2_L3_cache_aglined):713sub $64, %rdx714ALIGN (4)715L(L2_L3_aligned_128bytes_loop):716prefetchnta 0x1c0(%rdi)717prefetchnta 0x1c0(%rsi)718movdqa (%rdi), %xmm2719pxor (%rsi), %xmm2720movdqa %xmm2, %xmm1721722movdqa 16(%rdi), %xmm3723pxor 16(%rsi), %xmm3724por %xmm3, %xmm1725726movdqa 32(%rdi), %xmm4727pxor 32(%rsi), %xmm4728por %xmm4, %xmm1729730movdqa 48(%rdi), %xmm5731pxor 48(%rsi), %xmm5732por %xmm5, %xmm1733734ptest %xmm1, %xmm0735jnc L(64bytesormore_loop_end)736add $64, %rsi737add $64, %rdi738sub $64, %rdx739jae L(L2_L3_aligned_128bytes_loop)740741add $64, %rdx742add %rdx, %rsi743add %rdx, %rdi744BRANCH_TO_JMPTBL_ENTRY(L(table_64bytes), %rdx, 4)745746747ALIGN (4)748L(64bytesormore_loop_end):749add $16, %rdi750add $16, %rsi751ptest %xmm2, %xmm0752jnc L(16bytes)753754add $16, %rdi755add $16, %rsi756ptest %xmm3, %xmm0757jnc L(16bytes)758759add $16, %rdi760add $16, %rsi761ptest %xmm4, %xmm0762jnc L(16bytes)763764add $16, %rdi765add $16, %rsi766jmp L(16bytes)767768L(256bytesin256):769add $256, %rdi770add $256, %rsi771jmp L(16bytes)772L(240bytesin256):773add $240, %rdi774add $240, %rsi775jmp L(16bytes)776L(224bytesin256):777add $224, %rdi778add $224, %rsi779jmp L(16bytes)780L(208bytesin256):781add $208, %rdi782add $208, %rsi783jmp L(16bytes)784L(192bytesin256):785add $192, %rdi786add $192, %rsi787jmp L(16bytes)788L(176bytesin256):789add $176, %rdi790add $176, %rsi791jmp L(16bytes)792L(160bytesin256):793add $160, %rdi794add $160, %rsi795jmp L(16bytes)796L(144bytesin256):797add $144, %rdi798add $144, %rsi799jmp L(16bytes)800L(128bytesin256):801add $128, %rdi802add $128, %rsi803jmp L(16bytes)804L(112bytesin256):805add $112, %rdi806add $112, %rsi807jmp L(16bytes)808L(96bytesin256):809add $96, %rdi810add $96, %rsi811jmp L(16bytes)812L(80bytesin256):813add $80, %rdi814add $80, %rsi815jmp L(16bytes)816L(64bytesin256):817add $64, %rdi818add $64, %rsi819jmp L(16bytes)820L(48bytesin256):821add $16, %rdi822add $16, %rsi823L(32bytesin256):824add $16, %rdi825add $16, %rsi826L(16bytesin256):827add $16, %rdi828add $16, %rsi829L(16bytes):830mov -16(%rdi), %rax831mov -16(%rsi), %rcx832cmp %rax, %rcx833jne L(diffin8bytes)834L(8bytes):835mov -8(%rdi), %rax836mov -8(%rsi), %rcx837cmp %rax, %rcx838jne L(diffin8bytes)839xor %eax, %eax840ret841842ALIGN (4)843L(12bytes):844mov -12(%rdi), %rax845mov -12(%rsi), %rcx846cmp %rax, %rcx847jne L(diffin8bytes)848L(4bytes):849mov -4(%rsi), %ecx850mov -4(%rdi), %eax851cmp %eax, %ecx852jne L(diffin4bytes)853L(0bytes):854xor %eax, %eax855ret856857#ifndef USE_AS_WMEMCMP858/* unreal case for wmemcmp */859ALIGN (4)860L(65bytes):861movdqu -65(%rdi), %xmm1862movdqu -65(%rsi), %xmm2863mov $-65, %dl864pxor %xmm1, %xmm2865ptest %xmm2, %xmm0866jnc L(less16bytes)867L(49bytes):868movdqu -49(%rdi), %xmm1869movdqu -49(%rsi), %xmm2870mov $-49, %dl871pxor %xmm1, %xmm2872ptest %xmm2, %xmm0873jnc L(less16bytes)874L(33bytes):875movdqu -33(%rdi), %xmm1876movdqu -33(%rsi), %xmm2877mov $-33, %dl878pxor %xmm1, %xmm2879ptest %xmm2, %xmm0880jnc L(less16bytes)881L(17bytes):882mov -17(%rdi), %rax883mov -17(%rsi), %rcx884cmp %rax, %rcx885jne L(diffin8bytes)886L(9bytes):887mov -9(%rdi), %rax888mov -9(%rsi), %rcx889cmp %rax, %rcx890jne L(diffin8bytes)891movzbl -1(%rdi), %eax892movzbl -1(%rsi), %edx893sub %edx, %eax894ret895896ALIGN (4)897L(13bytes):898mov -13(%rdi), %rax899mov -13(%rsi), %rcx900cmp %rax, %rcx901jne L(diffin8bytes)902mov -8(%rdi), %rax903mov -8(%rsi), %rcx904cmp %rax, %rcx905jne L(diffin8bytes)906xor %eax, %eax907ret908909ALIGN (4)910L(5bytes):911mov -5(%rdi), %eax912mov -5(%rsi), %ecx913cmp %eax, %ecx914jne L(diffin4bytes)915movzbl -1(%rdi), %eax916movzbl -1(%rsi), %edx917sub %edx, %eax918ret919920ALIGN (4)921L(66bytes):922movdqu -66(%rdi), %xmm1923movdqu -66(%rsi), %xmm2924mov $-66, %dl925pxor %xmm1, %xmm2926ptest %xmm2, %xmm0927jnc L(less16bytes)928L(50bytes):929movdqu -50(%rdi), %xmm1930movdqu -50(%rsi), %xmm2931mov $-50, %dl932pxor %xmm1, %xmm2933ptest %xmm2, %xmm0934jnc L(less16bytes)935L(34bytes):936movdqu -34(%rdi), %xmm1937movdqu -34(%rsi), %xmm2938mov $-34, %dl939pxor %xmm1, %xmm2940ptest %xmm2, %xmm0941jnc L(less16bytes)942L(18bytes):943mov -18(%rdi), %rax944mov -18(%rsi), %rcx945cmp %rax, %rcx946jne L(diffin8bytes)947L(10bytes):948mov -10(%rdi), %rax949mov -10(%rsi), %rcx950cmp %rax, %rcx951jne L(diffin8bytes)952movzwl -2(%rdi), %eax953movzwl -2(%rsi), %ecx954cmp %cl, %al955jne L(end)956and $0xffff, %eax957and $0xffff, %ecx958sub %ecx, %eax959ret960961ALIGN (4)962L(14bytes):963mov -14(%rdi), %rax964mov -14(%rsi), %rcx965cmp %rax, %rcx966jne L(diffin8bytes)967mov -8(%rdi), %rax968mov -8(%rsi), %rcx969cmp %rax, %rcx970jne L(diffin8bytes)971xor %eax, %eax972ret973974ALIGN (4)975L(6bytes):976mov -6(%rdi), %eax977mov -6(%rsi), %ecx978cmp %eax, %ecx979jne L(diffin4bytes)980L(2bytes):981movzwl -2(%rsi), %ecx982movzwl -2(%rdi), %eax983cmp %cl, %al984jne L(end)985and $0xffff, %eax986and $0xffff, %ecx987sub %ecx, %eax988ret989990ALIGN (4)991L(67bytes):992movdqu -67(%rdi), %xmm2993movdqu -67(%rsi), %xmm1994mov $-67, %dl995pxor %xmm1, %xmm2996ptest %xmm2, %xmm0997jnc L(less16bytes)998L(51bytes):999movdqu -51(%rdi), %xmm21000movdqu -51(%rsi), %xmm11001mov $-51, %dl1002pxor %xmm1, %xmm21003ptest %xmm2, %xmm01004jnc L(less16bytes)1005L(35bytes):1006movdqu -35(%rsi), %xmm11007movdqu -35(%rdi), %xmm21008mov $-35, %dl1009pxor %xmm1, %xmm21010ptest %xmm2, %xmm01011jnc L(less16bytes)1012L(19bytes):1013mov -19(%rdi), %rax1014mov -19(%rsi), %rcx1015cmp %rax, %rcx1016jne L(diffin8bytes)1017L(11bytes):1018mov -11(%rdi), %rax1019mov -11(%rsi), %rcx1020cmp %rax, %rcx1021jne L(diffin8bytes)1022mov -4(%rdi), %eax1023mov -4(%rsi), %ecx1024cmp %eax, %ecx1025jne L(diffin4bytes)1026xor %eax, %eax1027ret10281029ALIGN (4)1030L(15bytes):1031mov -15(%rdi), %rax1032mov -15(%rsi), %rcx1033cmp %rax, %rcx1034jne L(diffin8bytes)1035mov -8(%rdi), %rax1036mov -8(%rsi), %rcx1037cmp %rax, %rcx1038jne L(diffin8bytes)1039xor %eax, %eax1040ret10411042ALIGN (4)1043L(7bytes):1044mov -7(%rdi), %eax1045mov -7(%rsi), %ecx1046cmp %eax, %ecx1047jne L(diffin4bytes)1048mov -4(%rdi), %eax1049mov -4(%rsi), %ecx1050cmp %eax, %ecx1051jne L(diffin4bytes)1052xor %eax, %eax1053ret10541055ALIGN (4)1056L(3bytes):1057movzwl -3(%rdi), %eax1058movzwl -3(%rsi), %ecx1059cmp %eax, %ecx1060jne L(diffin2bytes)1061L(1bytes):1062movzbl -1(%rdi), %eax1063movzbl -1(%rsi), %ecx1064sub %ecx, %eax1065ret1066#endif10671068ALIGN (4)1069L(68bytes):1070movdqu -68(%rdi), %xmm21071movdqu -68(%rsi), %xmm11072mov $-68, %dl1073pxor %xmm1, %xmm21074ptest %xmm2, %xmm01075jnc L(less16bytes)1076L(52bytes):1077movdqu -52(%rdi), %xmm21078movdqu -52(%rsi), %xmm11079mov $-52, %dl1080pxor %xmm1, %xmm21081ptest %xmm2, %xmm01082jnc L(less16bytes)1083L(36bytes):1084movdqu -36(%rdi), %xmm21085movdqu -36(%rsi), %xmm11086mov $-36, %dl1087pxor %xmm1, %xmm21088ptest %xmm2, %xmm01089jnc L(less16bytes)1090L(20bytes):1091movdqu -20(%rdi), %xmm21092movdqu -20(%rsi), %xmm11093mov $-20, %dl1094pxor %xmm1, %xmm21095ptest %xmm2, %xmm01096jnc L(less16bytes)1097mov -4(%rdi), %eax1098mov -4(%rsi), %ecx1099cmp %eax, %ecx1100jne L(diffin4bytes)1101xor %eax, %eax1102ret11031104#ifndef USE_AS_WMEMCMP1105/* unreal cases for wmemcmp */1106ALIGN (4)1107L(69bytes):1108movdqu -69(%rsi), %xmm11109movdqu -69(%rdi), %xmm21110mov $-69, %dl1111pxor %xmm1, %xmm21112ptest %xmm2, %xmm01113jnc L(less16bytes)1114L(53bytes):1115movdqu -53(%rsi), %xmm11116movdqu -53(%rdi), %xmm21117mov $-53, %dl1118pxor %xmm1, %xmm21119ptest %xmm2, %xmm01120jnc L(less16bytes)1121L(37bytes):1122movdqu -37(%rsi), %xmm11123movdqu -37(%rdi), %xmm21124mov $-37, %dl1125pxor %xmm1, %xmm21126ptest %xmm2, %xmm01127jnc L(less16bytes)1128L(21bytes):1129movdqu -21(%rsi), %xmm11130movdqu -21(%rdi), %xmm21131mov $-21, %dl1132pxor %xmm1, %xmm21133ptest %xmm2, %xmm01134jnc L(less16bytes)1135mov -8(%rdi), %rax1136mov -8(%rsi), %rcx1137cmp %rax, %rcx1138jne L(diffin8bytes)1139xor %eax, %eax1140ret11411142ALIGN (4)1143L(70bytes):1144movdqu -70(%rsi), %xmm11145movdqu -70(%rdi), %xmm21146mov $-70, %dl1147pxor %xmm1, %xmm21148ptest %xmm2, %xmm01149jnc L(less16bytes)1150L(54bytes):1151movdqu -54(%rsi), %xmm11152movdqu -54(%rdi), %xmm21153mov $-54, %dl1154pxor %xmm1, %xmm21155ptest %xmm2, %xmm01156jnc L(less16bytes)1157L(38bytes):1158movdqu -38(%rsi), %xmm11159movdqu -38(%rdi), %xmm21160mov $-38, %dl1161pxor %xmm1, %xmm21162ptest %xmm2, %xmm01163jnc L(less16bytes)1164L(22bytes):1165movdqu -22(%rsi), %xmm11166movdqu -22(%rdi), %xmm21167mov $-22, %dl1168pxor %xmm1, %xmm21169ptest %xmm2, %xmm01170jnc L(less16bytes)1171mov -8(%rdi), %rax1172mov -8(%rsi), %rcx1173cmp %rax, %rcx1174jne L(diffin8bytes)1175xor %eax, %eax1176ret11771178ALIGN (4)1179L(71bytes):1180movdqu -71(%rsi), %xmm11181movdqu -71(%rdi), %xmm21182mov $-71, %dl1183pxor %xmm1, %xmm21184ptest %xmm2, %xmm01185jnc L(less16bytes)1186L(55bytes):1187movdqu -55(%rdi), %xmm21188movdqu -55(%rsi), %xmm11189mov $-55, %dl1190pxor %xmm1, %xmm21191ptest %xmm2, %xmm01192jnc L(less16bytes)1193L(39bytes):1194movdqu -39(%rdi), %xmm21195movdqu -39(%rsi), %xmm11196mov $-39, %dl1197pxor %xmm1, %xmm21198ptest %xmm2, %xmm01199jnc L(less16bytes)1200L(23bytes):1201movdqu -23(%rdi), %xmm21202movdqu -23(%rsi), %xmm11203mov $-23, %dl1204pxor %xmm1, %xmm21205ptest %xmm2, %xmm01206jnc L(less16bytes)1207mov -8(%rdi), %rax1208mov -8(%rsi), %rcx1209cmp %rax, %rcx1210jne L(diffin8bytes)1211xor %eax, %eax1212ret1213#endif12141215ALIGN (4)1216L(72bytes):1217movdqu -72(%rsi), %xmm11218movdqu -72(%rdi), %xmm21219mov $-72, %dl1220pxor %xmm1, %xmm21221ptest %xmm2, %xmm01222jnc L(less16bytes)1223L(56bytes):1224movdqu -56(%rdi), %xmm21225movdqu -56(%rsi), %xmm11226mov $-56, %dl1227pxor %xmm1, %xmm21228ptest %xmm2, %xmm01229jnc L(less16bytes)1230L(40bytes):1231movdqu -40(%rdi), %xmm21232movdqu -40(%rsi), %xmm11233mov $-40, %dl1234pxor %xmm1, %xmm21235ptest %xmm2, %xmm01236jnc L(less16bytes)1237L(24bytes):1238movdqu -24(%rdi), %xmm21239movdqu -24(%rsi), %xmm11240mov $-24, %dl1241pxor %xmm1, %xmm21242ptest %xmm2, %xmm01243jnc L(less16bytes)1244mov -8(%rdi), %rax1245mov -8(%rsi), %rcx1246cmp %rax, %rcx1247jne L(diffin8bytes)1248xor %eax, %eax1249ret12501251#ifndef USE_AS_WMEMCMP1252/* unreal cases for wmemcmp */1253ALIGN (4)1254L(73bytes):1255movdqu -73(%rsi), %xmm11256movdqu -73(%rdi), %xmm21257mov $-73, %dl1258pxor %xmm1, %xmm21259ptest %xmm2, %xmm01260jnc L(less16bytes)1261L(57bytes):1262movdqu -57(%rdi), %xmm21263movdqu -57(%rsi), %xmm11264mov $-57, %dl1265pxor %xmm1, %xmm21266ptest %xmm2, %xmm01267jnc L(less16bytes)1268L(41bytes):1269movdqu -41(%rdi), %xmm21270movdqu -41(%rsi), %xmm11271mov $-41, %dl1272pxor %xmm1, %xmm21273ptest %xmm2, %xmm01274jnc L(less16bytes)1275L(25bytes):1276movdqu -25(%rdi), %xmm21277movdqu -25(%rsi), %xmm11278mov $-25, %dl1279pxor %xmm1, %xmm21280ptest %xmm2, %xmm01281jnc L(less16bytes)1282mov -9(%rdi), %rax1283mov -9(%rsi), %rcx1284cmp %rax, %rcx1285jne L(diffin8bytes)1286movzbl -1(%rdi), %eax1287movzbl -1(%rsi), %ecx1288sub %ecx, %eax1289ret12901291ALIGN (4)1292L(74bytes):1293movdqu -74(%rsi), %xmm11294movdqu -74(%rdi), %xmm21295mov $-74, %dl1296pxor %xmm1, %xmm21297ptest %xmm2, %xmm01298jnc L(less16bytes)1299L(58bytes):1300movdqu -58(%rdi), %xmm21301movdqu -58(%rsi), %xmm11302mov $-58, %dl1303pxor %xmm1, %xmm21304ptest %xmm2, %xmm01305jnc L(less16bytes)1306L(42bytes):1307movdqu -42(%rdi), %xmm21308movdqu -42(%rsi), %xmm11309mov $-42, %dl1310pxor %xmm1, %xmm21311ptest %xmm2, %xmm01312jnc L(less16bytes)1313L(26bytes):1314movdqu -26(%rdi), %xmm21315movdqu -26(%rsi), %xmm11316mov $-26, %dl1317pxor %xmm1, %xmm21318ptest %xmm2, %xmm01319jnc L(less16bytes)1320mov -10(%rdi), %rax1321mov -10(%rsi), %rcx1322cmp %rax, %rcx1323jne L(diffin8bytes)1324movzwl -2(%rdi), %eax1325movzwl -2(%rsi), %ecx1326jmp L(diffin2bytes)13271328ALIGN (4)1329L(75bytes):1330movdqu -75(%rsi), %xmm11331movdqu -75(%rdi), %xmm21332mov $-75, %dl1333pxor %xmm1, %xmm21334ptest %xmm2, %xmm01335jnc L(less16bytes)1336L(59bytes):1337movdqu -59(%rdi), %xmm21338movdqu -59(%rsi), %xmm11339mov $-59, %dl1340pxor %xmm1, %xmm21341ptest %xmm2, %xmm01342jnc L(less16bytes)1343L(43bytes):1344movdqu -43(%rdi), %xmm21345movdqu -43(%rsi), %xmm11346mov $-43, %dl1347pxor %xmm1, %xmm21348ptest %xmm2, %xmm01349jnc L(less16bytes)1350L(27bytes):1351movdqu -27(%rdi), %xmm21352movdqu -27(%rsi), %xmm11353mov $-27, %dl1354pxor %xmm1, %xmm21355ptest %xmm2, %xmm01356jnc L(less16bytes)1357mov -11(%rdi), %rax1358mov -11(%rsi), %rcx1359cmp %rax, %rcx1360jne L(diffin8bytes)1361mov -4(%rdi), %eax1362mov -4(%rsi), %ecx1363cmp %eax, %ecx1364jne L(diffin4bytes)1365xor %eax, %eax1366ret1367#endif1368ALIGN (4)1369L(76bytes):1370movdqu -76(%rsi), %xmm11371movdqu -76(%rdi), %xmm21372mov $-76, %dl1373pxor %xmm1, %xmm21374ptest %xmm2, %xmm01375jnc L(less16bytes)1376L(60bytes):1377movdqu -60(%rdi), %xmm21378movdqu -60(%rsi), %xmm11379mov $-60, %dl1380pxor %xmm1, %xmm21381ptest %xmm2, %xmm01382jnc L(less16bytes)1383L(44bytes):1384movdqu -44(%rdi), %xmm21385movdqu -44(%rsi), %xmm11386mov $-44, %dl1387pxor %xmm1, %xmm21388ptest %xmm2, %xmm01389jnc L(less16bytes)1390L(28bytes):1391movdqu -28(%rdi), %xmm21392movdqu -28(%rsi), %xmm11393mov $-28, %dl1394pxor %xmm1, %xmm21395ptest %xmm2, %xmm01396jnc L(less16bytes)1397mov -12(%rdi), %rax1398mov -12(%rsi), %rcx1399cmp %rax, %rcx1400jne L(diffin8bytes)1401mov -4(%rdi), %eax1402mov -4(%rsi), %ecx1403cmp %eax, %ecx1404jne L(diffin4bytes)1405xor %eax, %eax1406ret14071408#ifndef USE_AS_WMEMCMP1409/* unreal cases for wmemcmp */1410ALIGN (4)1411L(77bytes):1412movdqu -77(%rsi), %xmm11413movdqu -77(%rdi), %xmm21414mov $-77, %dl1415pxor %xmm1, %xmm21416ptest %xmm2, %xmm01417jnc L(less16bytes)1418L(61bytes):1419movdqu -61(%rdi), %xmm21420movdqu -61(%rsi), %xmm11421mov $-61, %dl1422pxor %xmm1, %xmm21423ptest %xmm2, %xmm01424jnc L(less16bytes)1425L(45bytes):1426movdqu -45(%rdi), %xmm21427movdqu -45(%rsi), %xmm11428mov $-45, %dl1429pxor %xmm1, %xmm21430ptest %xmm2, %xmm01431jnc L(less16bytes)1432L(29bytes):1433movdqu -29(%rdi), %xmm21434movdqu -29(%rsi), %xmm11435mov $-29, %dl1436pxor %xmm1, %xmm21437ptest %xmm2, %xmm01438jnc L(less16bytes)14391440mov -13(%rdi), %rax1441mov -13(%rsi), %rcx1442cmp %rax, %rcx1443jne L(diffin8bytes)14441445mov -8(%rdi), %rax1446mov -8(%rsi), %rcx1447cmp %rax, %rcx1448jne L(diffin8bytes)1449xor %eax, %eax1450ret14511452ALIGN (4)1453L(78bytes):1454movdqu -78(%rsi), %xmm11455movdqu -78(%rdi), %xmm21456mov $-78, %dl1457pxor %xmm1, %xmm21458ptest %xmm2, %xmm01459jnc L(less16bytes)1460L(62bytes):1461movdqu -62(%rdi), %xmm21462movdqu -62(%rsi), %xmm11463mov $-62, %dl1464pxor %xmm1, %xmm21465ptest %xmm2, %xmm01466jnc L(less16bytes)1467L(46bytes):1468movdqu -46(%rdi), %xmm21469movdqu -46(%rsi), %xmm11470mov $-46, %dl1471pxor %xmm1, %xmm21472ptest %xmm2, %xmm01473jnc L(less16bytes)1474L(30bytes):1475movdqu -30(%rdi), %xmm21476movdqu -30(%rsi), %xmm11477mov $-30, %dl1478pxor %xmm1, %xmm21479ptest %xmm2, %xmm01480jnc L(less16bytes)1481mov -14(%rdi), %rax1482mov -14(%rsi), %rcx1483cmp %rax, %rcx1484jne L(diffin8bytes)1485mov -8(%rdi), %rax1486mov -8(%rsi), %rcx1487cmp %rax, %rcx1488jne L(diffin8bytes)1489xor %eax, %eax1490ret14911492ALIGN (4)1493L(79bytes):1494movdqu -79(%rsi), %xmm11495movdqu -79(%rdi), %xmm21496mov $-79, %dl1497pxor %xmm1, %xmm21498ptest %xmm2, %xmm01499jnc L(less16bytes)1500L(63bytes):1501movdqu -63(%rdi), %xmm21502movdqu -63(%rsi), %xmm11503mov $-63, %dl1504pxor %xmm1, %xmm21505ptest %xmm2, %xmm01506jnc L(less16bytes)1507L(47bytes):1508movdqu -47(%rdi), %xmm21509movdqu -47(%rsi), %xmm11510mov $-47, %dl1511pxor %xmm1, %xmm21512ptest %xmm2, %xmm01513jnc L(less16bytes)1514L(31bytes):1515movdqu -31(%rdi), %xmm21516movdqu -31(%rsi), %xmm11517mov $-31, %dl1518pxor %xmm1, %xmm21519ptest %xmm2, %xmm01520jnc L(less16bytes)1521mov -15(%rdi), %rax1522mov -15(%rsi), %rcx1523cmp %rax, %rcx1524jne L(diffin8bytes)1525mov -8(%rdi), %rax1526mov -8(%rsi), %rcx1527cmp %rax, %rcx1528jne L(diffin8bytes)1529xor %eax, %eax1530ret1531#endif1532ALIGN (4)1533L(64bytes):1534movdqu -64(%rdi), %xmm21535movdqu -64(%rsi), %xmm11536mov $-64, %dl1537pxor %xmm1, %xmm21538ptest %xmm2, %xmm01539jnc L(less16bytes)1540L(48bytes):1541movdqu -48(%rdi), %xmm21542movdqu -48(%rsi), %xmm11543mov $-48, %dl1544pxor %xmm1, %xmm21545ptest %xmm2, %xmm01546jnc L(less16bytes)1547L(32bytes):1548movdqu -32(%rdi), %xmm21549movdqu -32(%rsi), %xmm11550mov $-32, %dl1551pxor %xmm1, %xmm21552ptest %xmm2, %xmm01553jnc L(less16bytes)15541555mov -16(%rdi), %rax1556mov -16(%rsi), %rcx1557cmp %rax, %rcx1558jne L(diffin8bytes)15591560mov -8(%rdi), %rax1561mov -8(%rsi), %rcx1562cmp %rax, %rcx1563jne L(diffin8bytes)1564xor %eax, %eax1565ret15661567/*1568* Aligned 8 bytes to avoid 2 branch "taken" in one 16 alinged code block.1569*/1570ALIGN (3)1571L(less16bytes):1572movsbq %dl, %rdx1573mov (%rsi, %rdx), %rcx1574mov (%rdi, %rdx), %rax1575cmp %rax, %rcx1576jne L(diffin8bytes)1577mov 8(%rsi, %rdx), %rcx1578mov 8(%rdi, %rdx), %rax1579L(diffin8bytes):1580cmp %eax, %ecx1581jne L(diffin4bytes)1582shr $32, %rcx1583shr $32, %rax15841585#ifdef USE_AS_WMEMCMP1586/* for wmemcmp */1587cmp %eax, %ecx1588jne L(diffin4bytes)1589xor %eax, %eax1590ret1591#endif15921593L(diffin4bytes):1594#ifndef USE_AS_WMEMCMP1595cmp %cx, %ax1596jne L(diffin2bytes)1597shr $16, %ecx1598shr $16, %eax1599L(diffin2bytes):1600cmp %cl, %al1601jne L(end)1602and $0xffff, %eax1603and $0xffff, %ecx1604sub %ecx, %eax1605ret1606#else16071608/* for wmemcmp */1609mov $1, %eax1610jl L(nequal_bigger)1611neg %eax1612ret16131614ALIGN (4)1615L(nequal_bigger):1616ret16171618L(unreal_case):1619xor %eax, %eax1620ret1621#endif16221623ALIGN (4)1624L(end):1625and $0xff, %eax1626and $0xff, %ecx1627sub %ecx, %eax1628ret16291630END (MEMCMP)16311632.section .rodata.sse4.1,"a",@progbits1633ALIGN (3)1634#ifndef USE_AS_WMEMCMP1635L(table_64bytes):1636.int JMPTBL (L(0bytes), L(table_64bytes))1637.int JMPTBL (L(1bytes), L(table_64bytes))1638.int JMPTBL (L(2bytes), L(table_64bytes))1639.int JMPTBL (L(3bytes), L(table_64bytes))1640.int JMPTBL (L(4bytes), L(table_64bytes))1641.int JMPTBL (L(5bytes), L(table_64bytes))1642.int JMPTBL (L(6bytes), L(table_64bytes))1643.int JMPTBL (L(7bytes), L(table_64bytes))1644.int JMPTBL (L(8bytes), L(table_64bytes))1645.int JMPTBL (L(9bytes), L(table_64bytes))1646.int JMPTBL (L(10bytes), L(table_64bytes))1647.int JMPTBL (L(11bytes), L(table_64bytes))1648.int JMPTBL (L(12bytes), L(table_64bytes))1649.int JMPTBL (L(13bytes), L(table_64bytes))1650.int JMPTBL (L(14bytes), L(table_64bytes))1651.int JMPTBL (L(15bytes), L(table_64bytes))1652.int JMPTBL (L(16bytes), L(table_64bytes))1653.int JMPTBL (L(17bytes), L(table_64bytes))1654.int JMPTBL (L(18bytes), L(table_64bytes))1655.int JMPTBL (L(19bytes), L(table_64bytes))1656.int JMPTBL (L(20bytes), L(table_64bytes))1657.int JMPTBL (L(21bytes), L(table_64bytes))1658.int JMPTBL (L(22bytes), L(table_64bytes))1659.int JMPTBL (L(23bytes), L(table_64bytes))1660.int JMPTBL (L(24bytes), L(table_64bytes))1661.int JMPTBL (L(25bytes), L(table_64bytes))1662.int JMPTBL (L(26bytes), L(table_64bytes))1663.int JMPTBL (L(27bytes), L(table_64bytes))1664.int JMPTBL (L(28bytes), L(table_64bytes))1665.int JMPTBL (L(29bytes), L(table_64bytes))1666.int JMPTBL (L(30bytes), L(table_64bytes))1667.int JMPTBL (L(31bytes), L(table_64bytes))1668.int JMPTBL (L(32bytes), L(table_64bytes))1669.int JMPTBL (L(33bytes), L(table_64bytes))1670.int JMPTBL (L(34bytes), L(table_64bytes))1671.int JMPTBL (L(35bytes), L(table_64bytes))1672.int JMPTBL (L(36bytes), L(table_64bytes))1673.int JMPTBL (L(37bytes), L(table_64bytes))1674.int JMPTBL (L(38bytes), L(table_64bytes))1675.int JMPTBL (L(39bytes), L(table_64bytes))1676.int JMPTBL (L(40bytes), L(table_64bytes))1677.int JMPTBL (L(41bytes), L(table_64bytes))1678.int JMPTBL (L(42bytes), L(table_64bytes))1679.int JMPTBL (L(43bytes), L(table_64bytes))1680.int JMPTBL (L(44bytes), L(table_64bytes))1681.int JMPTBL (L(45bytes), L(table_64bytes))1682.int JMPTBL (L(46bytes), L(table_64bytes))1683.int JMPTBL (L(47bytes), L(table_64bytes))1684.int JMPTBL (L(48bytes), L(table_64bytes))1685.int JMPTBL (L(49bytes), L(table_64bytes))1686.int JMPTBL (L(50bytes), L(table_64bytes))1687.int JMPTBL (L(51bytes), L(table_64bytes))1688.int JMPTBL (L(52bytes), L(table_64bytes))1689.int JMPTBL (L(53bytes), L(table_64bytes))1690.int JMPTBL (L(54bytes), L(table_64bytes))1691.int JMPTBL (L(55bytes), L(table_64bytes))1692.int JMPTBL (L(56bytes), L(table_64bytes))1693.int JMPTBL (L(57bytes), L(table_64bytes))1694.int JMPTBL (L(58bytes), L(table_64bytes))1695.int JMPTBL (L(59bytes), L(table_64bytes))1696.int JMPTBL (L(60bytes), L(table_64bytes))1697.int JMPTBL (L(61bytes), L(table_64bytes))1698.int JMPTBL (L(62bytes), L(table_64bytes))1699.int JMPTBL (L(63bytes), L(table_64bytes))1700.int JMPTBL (L(64bytes), L(table_64bytes))1701.int JMPTBL (L(65bytes), L(table_64bytes))1702.int JMPTBL (L(66bytes), L(table_64bytes))1703.int JMPTBL (L(67bytes), L(table_64bytes))1704.int JMPTBL (L(68bytes), L(table_64bytes))1705.int JMPTBL (L(69bytes), L(table_64bytes))1706.int JMPTBL (L(70bytes), L(table_64bytes))1707.int JMPTBL (L(71bytes), L(table_64bytes))1708.int JMPTBL (L(72bytes), L(table_64bytes))1709.int JMPTBL (L(73bytes), L(table_64bytes))1710.int JMPTBL (L(74bytes), L(table_64bytes))1711.int JMPTBL (L(75bytes), L(table_64bytes))1712.int JMPTBL (L(76bytes), L(table_64bytes))1713.int JMPTBL (L(77bytes), L(table_64bytes))1714.int JMPTBL (L(78bytes), L(table_64bytes))1715.int JMPTBL (L(79bytes), L(table_64bytes))1716#else1717L(table_64bytes):1718.int JMPTBL (L(0bytes), L(table_64bytes))1719.int JMPTBL (L(unreal_case), L(table_64bytes))1720.int JMPTBL (L(unreal_case), L(table_64bytes))1721.int JMPTBL (L(unreal_case), L(table_64bytes))1722.int JMPTBL (L(4bytes), L(table_64bytes))1723.int JMPTBL (L(unreal_case), L(table_64bytes))1724.int JMPTBL (L(unreal_case), L(table_64bytes))1725.int JMPTBL (L(unreal_case), L(table_64bytes))1726.int JMPTBL (L(8bytes), L(table_64bytes))1727.int JMPTBL (L(unreal_case), L(table_64bytes))1728.int JMPTBL (L(unreal_case), L(table_64bytes))1729.int JMPTBL (L(unreal_case), L(table_64bytes))1730.int JMPTBL (L(12bytes), L(table_64bytes))1731.int JMPTBL (L(unreal_case), L(table_64bytes))1732.int JMPTBL (L(unreal_case), L(table_64bytes))1733.int JMPTBL (L(unreal_case), L(table_64bytes))1734.int JMPTBL (L(16bytes), L(table_64bytes))1735.int JMPTBL (L(unreal_case), L(table_64bytes))1736.int JMPTBL (L(unreal_case), L(table_64bytes))1737.int JMPTBL (L(unreal_case), L(table_64bytes))1738.int JMPTBL (L(20bytes), L(table_64bytes))1739.int JMPTBL (L(unreal_case), L(table_64bytes))1740.int JMPTBL (L(unreal_case), L(table_64bytes))1741.int JMPTBL (L(unreal_case), L(table_64bytes))1742.int JMPTBL (L(24bytes), L(table_64bytes))1743.int JMPTBL (L(unreal_case), L(table_64bytes))1744.int JMPTBL (L(unreal_case), L(table_64bytes))1745.int JMPTBL (L(unreal_case), L(table_64bytes))1746.int JMPTBL (L(28bytes), L(table_64bytes))1747.int JMPTBL (L(unreal_case), L(table_64bytes))1748.int JMPTBL (L(unreal_case), L(table_64bytes))1749.int JMPTBL (L(unreal_case), L(table_64bytes))1750.int JMPTBL (L(32bytes), L(table_64bytes))1751.int JMPTBL (L(unreal_case), L(table_64bytes))1752.int JMPTBL (L(unreal_case), L(table_64bytes))1753.int JMPTBL (L(unreal_case), L(table_64bytes))1754.int JMPTBL (L(36bytes), L(table_64bytes))1755.int JMPTBL (L(unreal_case), L(table_64bytes))1756.int JMPTBL (L(unreal_case), L(table_64bytes))1757.int JMPTBL (L(unreal_case), L(table_64bytes))1758.int JMPTBL (L(40bytes), L(table_64bytes))1759.int JMPTBL (L(unreal_case), L(table_64bytes))1760.int JMPTBL (L(unreal_case), L(table_64bytes))1761.int JMPTBL (L(unreal_case), L(table_64bytes))1762.int JMPTBL (L(44bytes), L(table_64bytes))1763.int JMPTBL (L(unreal_case), L(table_64bytes))1764.int JMPTBL (L(unreal_case), L(table_64bytes))1765.int JMPTBL (L(unreal_case), L(table_64bytes))1766.int JMPTBL (L(48bytes), L(table_64bytes))1767.int JMPTBL (L(unreal_case), L(table_64bytes))1768.int JMPTBL (L(unreal_case), L(table_64bytes))1769.int JMPTBL (L(unreal_case), L(table_64bytes))1770.int JMPTBL (L(52bytes), L(table_64bytes))1771.int JMPTBL (L(unreal_case), L(table_64bytes))1772.int JMPTBL (L(unreal_case), L(table_64bytes))1773.int JMPTBL (L(unreal_case), L(table_64bytes))1774.int JMPTBL (L(56bytes), L(table_64bytes))1775.int JMPTBL (L(unreal_case), L(table_64bytes))1776.int JMPTBL (L(unreal_case), L(table_64bytes))1777.int JMPTBL (L(unreal_case), L(table_64bytes))1778.int JMPTBL (L(60bytes), L(table_64bytes))1779.int JMPTBL (L(unreal_case), L(table_64bytes))1780.int JMPTBL (L(unreal_case), L(table_64bytes))1781.int JMPTBL (L(unreal_case), L(table_64bytes))1782.int JMPTBL (L(64bytes), L(table_64bytes))1783.int JMPTBL (L(unreal_case), L(table_64bytes))1784.int JMPTBL (L(unreal_case), L(table_64bytes))1785.int JMPTBL (L(unreal_case), L(table_64bytes))1786.int JMPTBL (L(68bytes), L(table_64bytes))1787.int JMPTBL (L(unreal_case), L(table_64bytes))1788.int JMPTBL (L(unreal_case), L(table_64bytes))1789.int JMPTBL (L(unreal_case), L(table_64bytes))1790.int JMPTBL (L(72bytes), L(table_64bytes))1791.int JMPTBL (L(unreal_case), L(table_64bytes))1792.int JMPTBL (L(unreal_case), L(table_64bytes))1793.int JMPTBL (L(unreal_case), L(table_64bytes))1794.int JMPTBL (L(76bytes), L(table_64bytes))1795.int JMPTBL (L(unreal_case), L(table_64bytes))1796.int JMPTBL (L(unreal_case), L(table_64bytes))1797.int JMPTBL (L(unreal_case), L(table_64bytes))1798#endif179918001801