/* $Id: VIScsum.S,v 1.6 2000/02/20 23:21:39 davem Exp $ * VIScsum.S: High bandwidth IP checksumming utilizing the UltraSparc * Visual Instruction Set. * * Copyright (C) 1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz) * Copyright (C) 2000 David S. Miller (davem@redhat.com) * * Based on older sparc32/sparc64 checksum.S, which is: * * Copyright(C) 1995 Linus Torvalds * Copyright(C) 1995 Miguel de Icaza * Copyright(C) 1996, 1997 David S. Miller * derived from: * Linux/Alpha checksum c-code * Linux/ix86 inline checksum assembly * RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code) * David Mosberger-Tang for optimized reference c-code * BSD4.4 portable checksum routine */ #ifdef __sparc_v9__ #define STACKOFF 2175 #else #define STACKOFF 64 #endif #ifdef __KERNEL__ #include #include #include #include #else #define ASI_BLK_P 0xf0 #define FRPS_FEF 0x04 #endif /* Dobrou noc, SunSoft engineers. Spete sladce. * This has a couple of tricks in and those * tricks are UltraLinux trade secrets :)) */ #define START_THE_TRICK(fz,f0,f2,f4,f6,f8,f10) \ fcmpgt32 %fz, %f0, %g1 /* FPM Group */; \ fcmpgt32 %fz, %f2, %g2 /* FPM Group */; \ fcmpgt32 %fz, %f4, %g3 /* FPM Group */; \ inc %g1 /* IEU0 Group */; \ fcmpgt32 %fz, %f6, %g5 /* FPM */; \ srl %g1, 1, %g1 /* IEU0 Group */; \ fcmpgt32 %fz, %f8, %g7 /* FPM */; \ inc %g2 /* IEU0 Group */; \ fcmpgt32 %fz, %f10, %o3 /* FPM */; \ srl %g2, 1, %g2 /* IEU0 Group */; \ inc %g3 /* IEU1 */; \ srl %g3, 1, %g3 /* IEU0 Group */; \ add %o2, %g1, %o2 /* IEU1 */; \ add %o2, %g2, %o2 /* IEU0 Group */; \ inc %g5 /* IEU1 */; \ add %o2, %g3, %o2 /* IEU0 Group */; #define DO_THE_TRICK(O12,O14,f0,f2,f4,f6,f8,f10,f12,f14,F0,F2,F4,F6,F8,F10,F12,F14) \ srl %g5, 1, %g5 /* IEU0 Group */; \ fpadd32 %F0, %f0, %F0 /* FPA */; \ fcmpgt32 %O12, %f12, %o4 /* FPM */; \ inc %g7 /* IEU0 Group */; \ fpadd32 %F2, %f2, %F2 /* FPA */; \ fcmpgt32 %O14, %f14, %o5 /* FPM */; \ add %o2, %g5, %o2 /* IEU1 Group */; \ fpadd32 %F4, %f4, %F4 /* FPA */; \ fcmpgt32 %f0, %F0, %g1 /* FPM */; \ srl %g7, 1, %g7 /* IEU0 Group */; \ fpadd32 %F6, %f6, %F6 /* FPA */; \ fcmpgt32 %f2, %F2, %g2 /* FPM */; \ add %o2, %g7, %o2 /* IEU0 Group */; \ fpadd32 %F8, %f8, %F8 /* FPA */; \ fcmpgt32 %f4, %F4, %g3 /* FPM */; \ inc %o3 /* IEU0 Group */; \ fpadd32 %F10, %f10, %F10 /* FPA */; \ fcmpgt32 %f6, %F6, %g5 /* FPM */; \ srl %o3, 1, %o3 /* IEU0 Group */; \ fpadd32 %F12, %f12, %F12 /* FPA */; \ fcmpgt32 %f8, %F8, %g7 /* FPM */; \ add %o2, %o3, %o2 /* IEU0 Group */; \ fpadd32 %F14, %f14, %F14 /* FPA */; \ fcmpgt32 %f10, %F10, %o3 /* FPM */; \ inc %o4 /* IEU0 Group */; \ inc %o5 /* IEU1 */; \ srl %o4, 1, %o4 /* IEU0 Group */; \ inc %g1 /* IEU1 */; \ srl %o5, 1, %o5 /* IEU0 Group */; \ add %o2, %o4, %o2 /* IEU1 */; \ srl %g1, 1, %g1 /* IEU0 Group */; \ add %o2, %o5, %o2 /* IEU1 */; \ inc %g2 /* IEU0 Group */; \ add %o2, %g1, %o2 /* IEU1 */; \ srl %g2, 1, %g2 /* IEU0 Group */; \ inc %g3 /* IEU1 */; \ srl %g3, 1, %g3 /* IEU0 Group */; \ add %o2, %g2, %o2 /* IEU1 */; \ inc %g5 /* IEU0 Group */; \ add %o2, %g3, %o2 /* IEU0 */; #define END_THE_TRICK(O12,O14,f0,f2,f4,f6,f8,f10,f12,f14,S0,S1,S2,S3,T0,T1,U0,fz) \ srl %g5, 1, %g5 /* IEU0 Group */; \ fpadd32 %f2, %f0, %S0 /* FPA */; \ fcmpgt32 %O12, %f12, %o4 /* FPM */; \ inc %g7 /* IEU0 Group */; \ fpadd32 %f6, %f4, %S1 /* FPA */; \ fcmpgt32 %O14, %f14, %o5 /* FPM */; \ srl %g7, 1, %g7 /* IEU0 Group */; \ fpadd32 %f10, %f8, %S2 /* FPA */; \ fcmpgt32 %f0, %S0, %g1 /* FPM */; \ inc %o3 /* IEU0 Group */; \ fpadd32 %f14, %f12, %S3 /* FPA */; \ fcmpgt32 %f4, %S1, %g2 /* FPM */; \ add %o2, %g5, %o2 /* IEU0 Group */; \ fpadd32 %S0, %S1, %T0 /* FPA */; \ fcmpgt32 %f8, %S2, %g3 /* FPM */; \ add %o2, %g7, %o2 /* IEU0 Group */; \ fzero %fz /* FPA */; \ fcmpgt32 %f12, %S3, %g5 /* FPM */; \ srl %o3, 1, %o3 /* IEU0 Group */; \ fpadd32 %S2, %S3, %T1 /* FPA */; \ fcmpgt32 %S0, %T0, %g7 /* FPM */; \ add %o2, %o3, %o2 /* IEU0 Group */; \ fpadd32 %T0, %T1, %U0 /* FPA */; \ fcmpgt32 %S2, %T1, %o3 /* FPM */; \ inc %o4 /* IEU0 Group */; \ inc %o5 /* IEU1 */; \ srl %o4, 1, %o4 /* IEU0 Group */; \ inc %g1 /* IEU1 */; \ add %o2, %o4, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %f2, %o4 /* FPM */; \ srl %o5, 1, %o5 /* IEU0 Group */; \ inc %g2 /* IEU1 */; \ add %o2, %o5, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %f6, %o5 /* FPM */; \ srl %g1, 1, %g1 /* IEU0 Group */; \ inc %g3 /* IEU1 */; \ add %o2, %g1, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %f10, %g1 /* FPM */; \ srl %g2, 1, %g2 /* IEU0 Group */; \ inc %g5 /* IEU1 */; \ add %o2, %g2, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %f14, %g2 /* FPM */; \ srl %g3, 1, %g3 /* IEU0 Group */; \ inc %g7 /* IEU1 */; \ add %o2, %g3, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %S1, %g3 /* FPM */; \ srl %g5, 1, %g5 /* IEU0 Group */; \ inc %o3 /* IEU1 */; \ add %o2, %g5, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %S3, %g5 /* FPM */; \ srl %g7, 1, %g7 /* IEU0 Group */; \ inc %o4 /* IEU1 */; \ add %o2, %g7, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %T1, %g7 /* FPM */; \ srl %o3, 1, %o3 /* IEU0 Group */; \ inc %o5 /* IEU1 */; \ add %o2, %o3, %o2 /* IEU0 Group */; \ fcmpgt32 %T0, %U0, %o3 /* FPM */; \ srl %o4, 1, %o4 /* IEU0 Group */; \ inc %g1 /* IEU1 */; \ sub %o2, %o4, %o2 /* IEU0 Group */; \ fcmpgt32 %fz, %U0, %o4 /* FPM */; \ srl %o5, 1, %o5 /* IEU0 Group */; \ inc %g2 /* IEU1 */; \ srl %g1, 1, %g1 /* IEU0 Group */; \ sub %o2, %o5, %o2 /* IEU1 */; \ std %U0, [%sp + STACKOFF] /* Store */; \ srl %g2, 1, %g2 /* IEU0 Group */; \ sub %o2, %g1, %o2 /* IEU1 */; \ inc %g3 /* IEU0 Group */; \ sub %o2, %g2, %o2 /* IEU1 */; \ srl %g3, 1, %g3 /* IEU0 Group */; \ inc %g5 /* IEU1 */; \ srl %g5, 1, %g5 /* IEU0 Group */; \ sub %o2, %g3, %o2 /* IEU1 */; \ ldx [%sp + STACKOFF], %o5 /* Load Group */; \ inc %g7 /* IEU0 */; \ sub %o2, %g5, %o2 /* IEU1 */; \ srl %g7, 1, %g7 /* IEU0 Group */; \ inc %o3 /* IEU1 */; \ srl %o3, 1, %o3 /* IEU0 Group */; \ sub %o2, %g7, %o2 /* IEU1 */; \ inc %o4 /* IEU0 Group */; \ add %o2, %o3, %o2 /* IEU1 */; \ srl %o4, 1, %o4 /* IEU0 Group */; \ sub %o2, %o4, %o2 /* IEU0 Group */; \ addcc %o2, %o5, %o2 /* IEU1 Group */; \ bcs,a,pn %xcc, 33f /* CTI */; \ add %o2, 1, %o2 /* IEU0 */; \ 33: /* That's it */; #define CSUM_LASTCHUNK(offset) \ ldx [%o0 - offset - 0x10], %g2; \ ldx [%o0 - offset - 0x08], %g3; \ addcc %g2, %o2, %o2; \ bcs,a,pn %xcc, 31f; \ add %o2, 1, %o2; \ 31: addcc %g3, %o2, %o2; \ bcs,a,pn %xcc, 32f; \ add %o2, 1, %o2; \ 32: .text .globl csum_partial .align 32 csum_partial: andcc %o0, 7, %g0 /* IEU1 Group */ be,pt %icc, 4f /* CTI */ andcc %o0, 0x38, %g3 /* IEU1 */ mov 1, %g5 /* IEU0 Group */ cmp %o1, 6 /* IEU1 */ bl,pn %icc, 21f /* CTI */ andcc %o0, 2, %g0 /* IEU1 Group */ be,pt %icc, 1f /* CTI */ and %o0, 4, %g7 /* IEU0 */ lduh [%o0], %g2 /* Load */ sub %o1, 2, %o1 /* IEU0 Group */ add %o0, 2, %o0 /* IEU1 */ andcc %o0, 4, %g7 /* IEU1 Group */ sll %g5, 16, %g5 /* IEU0 */ sll %g2, 16, %g2 /* IEU0 Group */ addcc %g2, %o2, %o2 /* IEU1 Group (regdep) */ bcs,a,pn %icc, 1f /* CTI */ add %o2, %g5, %o2 /* IEU0 */ 1: ld [%o0], %g2 /* Load */ brz,a,pn %g7, 4f /* CTI+IEU1 Group */ and %o0, 0x38, %g3 /* IEU0 */ add %o0, 4, %o0 /* IEU0 Group */ sub %o1, 4, %o1 /* IEU1 */ addcc %g2, %o2, %o2 /* IEU1 Group */ bcs,a,pn %icc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: and %o0, 0x38, %g3 /* IEU1 Group */ 4: srl %o2, 0, %o2 /* IEU0 Group */ mov 0x40, %g1 /* IEU1 */ brz,pn %g3, 3f /* CTI+IEU1 Group */ sub %g1, %g3, %g1 /* IEU0 */ cmp %o1, 56 /* IEU1 Group */ blu,pn %icc, 20f /* CTI */ andcc %o0, 8, %g0 /* IEU1 Group */ be,pn %icc, 1f /* CTI */ ldx [%o0], %g2 /* Load */ add %o0, 8, %o0 /* IEU0 Group */ sub %o1, 8, %o1 /* IEU1 */ addcc %g2, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: andcc %g1, 0x10, %g0 /* IEU1 Group */ be,pn %icc, 2f /* CTI */ and %g1, 0x20, %g1 /* IEU0 */ ldx [%o0], %g2 /* Load */ ldx [%o0+8], %g3 /* Load Group */ add %o0, 16, %o0 /* IEU0 */ sub %o1, 16, %o1 /* IEU1 */ addcc %g2, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: addcc %g3, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 2f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 2: brz,pn %g1, 3f /* CTI+IEU1 Group */ ldx [%o0], %g2 /* Load */ ldx [%o0+8], %g3 /* Load Group */ ldx [%o0+16], %g5 /* Load Group */ ldx [%o0+24], %g7 /* Load Group */ add %o0, 32, %o0 /* IEU0 */ sub %o1, 32, %o1 /* IEU1 */ addcc %g2, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: addcc %g3, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: addcc %g5, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: addcc %g7, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 3f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 3: cmp %o1, 0xc0 /* IEU1 Group */ blu,pn %icc, 20f /* CTI */ sllx %o2, 32, %g5 /* IEU0 */ #ifdef __KERNEL__ VISEntry #endif addcc %o2, %g5, %o2 /* IEU1 Group */ sub %o1, 0xc0, %o1 /* IEU0 */ wr %g0, ASI_BLK_P, %asi /* LSU Group */ membar #StoreLoad /* LSU Group */ srlx %o2, 32, %o2 /* IEU0 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU1 */ 1: andcc %o1, 0x80, %g0 /* IEU1 Group */ bne,pn %icc, 7f /* CTI */ andcc %o1, 0x40, %g0 /* IEU1 Group */ be,pn %icc, 6f /* CTI */ fzero %f12 /* FPA */ fzero %f14 /* FPA Group */ ldda [%o0 + 0x000] %asi, %f16 ldda [%o0 + 0x040] %asi, %f32 ldda [%o0 + 0x080] %asi, %f48 START_THE_TRICK(f12,f16,f18,f20,f22,f24,f26) ba,a,pt %xcc, 3f 6: sub %o0, 0x40, %o0 /* IEU0 Group */ fzero %f28 /* FPA */ fzero %f30 /* FPA Group */ ldda [%o0 + 0x040] %asi, %f32 ldda [%o0 + 0x080] %asi, %f48 ldda [%o0 + 0x0c0] %asi, %f0 START_THE_TRICK(f28,f32,f34,f36,f38,f40,f42) ba,a,pt %xcc, 4f 7: bne,pt %icc, 8f /* CTI */ fzero %f44 /* FPA */ add %o0, 0x40, %o0 /* IEU0 Group */ fzero %f60 /* FPA */ fzero %f62 /* FPA Group */ ldda [%o0 - 0x040] %asi, %f0 ldda [%o0 + 0x000] %asi, %f16 ldda [%o0 + 0x040] %asi, %f32 START_THE_TRICK(f60,f0,f2,f4,f6,f8,f10) ba,a,pt %xcc, 2f 8: add %o0, 0x80, %o0 /* IEU0 Group */ fzero %f46 /* FPA */ ldda [%o0 - 0x080] %asi, %f48 ldda [%o0 - 0x040] %asi, %f0 ldda [%o0 + 0x000] %asi, %f16 START_THE_TRICK(f44,f48,f50,f52,f54,f56,f58) 1: DO_THE_TRICK(f44,f46,f48,f50,f52,f54,f56,f58,f60,f62,f0,f2,f4,f6,f8,f10,f12,f14) ldda [%o0 + 0x040] %asi, %f32 2: DO_THE_TRICK(f60,f62,f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30) ldda [%o0 + 0x080] %asi, %f48 3: DO_THE_TRICK(f12,f14,f16,f18,f20,f22,f24,f26,f28,f30,f32,f34,f36,f38,f40,f42,f44,f46) ldda [%o0 + 0x0c0] %asi, %f0 4: DO_THE_TRICK(f28,f30,f32,f34,f36,f38,f40,f42,f44,f46,f48,f50,f52,f54,f56,f58,f60,f62) add %o0, 0x100, %o0 /* IEU0 Group */ subcc %o1, 0x100, %o1 /* IEU1 */ bgeu,a,pt %icc, 1b /* CTI */ ldda [%o0 + 0x000] %asi, %f16 membar #Sync /* LSU Group */ DO_THE_TRICK(f44,f46,f48,f50,f52,f54,f56,f58,f60,f62,f0,f2,f4,f6,f8,f10,f12,f14) END_THE_TRICK(f60,f62,f0,f2,f4,f6,f8,f10,f12,f14,f16,f18,f20,f22,f24,f26,f28,f30) #ifdef __KERNEL__ ldub [%g6 + AOFF_task_thread + AOFF_thread_current_ds], %g7 #endif and %o1, 0x3f, %o1 /* IEU0 Group */ #ifdef __KERNEL__ VISExit wr %g7, %g0, %asi #endif 20: andcc %o1, 0xf0, %g1 /* IEU1 Group */ be,pn %icc, 23f /* CTI */ and %o1, 0xf, %o3 /* IEU0 */ #ifdef __KERNEL__ 22: sll %g1, 1, %o4 /* IEU0 Group */ sethi %hi(23f), %g7 /* IEU1 */ sub %g7, %o4, %g7 /* IEU0 Group */ jmpl %g7 + %lo(23f), %g0 /* CTI Group brk forced*/ add %o0, %g1, %o0 /* IEU0 */ #else 22: rd %pc, %g7 /* LSU Group+4bubbles */ sll %g1, 1, %o4 /* IEU0 Group */ sub %g7, %o4, %g7 /* IEU0 Group (regdep) */ jmpl %g7 + (23f - 22b), %g0 /* CTI Group brk forced*/ add %o0, %g1, %o0 /* IEU0 */ #endif CSUM_LASTCHUNK(0xe0) CSUM_LASTCHUNK(0xd0) CSUM_LASTCHUNK(0xc0) CSUM_LASTCHUNK(0xb0) CSUM_LASTCHUNK(0xa0) CSUM_LASTCHUNK(0x90) CSUM_LASTCHUNK(0x80) CSUM_LASTCHUNK(0x70) CSUM_LASTCHUNK(0x60) CSUM_LASTCHUNK(0x50) CSUM_LASTCHUNK(0x40) CSUM_LASTCHUNK(0x30) CSUM_LASTCHUNK(0x20) CSUM_LASTCHUNK(0x10) CSUM_LASTCHUNK(0x00) 23: brnz,pn %o3, 26f /* CTI+IEU1 Group */ 24: sllx %o2, 32, %g1 /* IEU0 */ 25: addcc %o2, %g1, %o0 /* IEU1 Group */ srlx %o0, 32, %o0 /* IEU0 Group (regdep) */ bcs,a,pn %xcc, 1f /* CTI */ add %o0, 1, %o0 /* IEU1 */ 1: retl /* CTI Group brk forced*/ srl %o0, 0, %o0 /* IEU0 */ 26: andcc %o1, 8, %g0 /* IEU1 Group */ be,pn %icc, 1f /* CTI */ ldx [%o0], %g3 /* Load */ add %o0, 8, %o0 /* IEU0 Group */ addcc %g3, %o2, %o2 /* IEU1 Group */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: andcc %o1, 4, %g0 /* IEU1 Group */ be,a,pn %icc, 1f /* CTI */ clr %g2 /* IEU0 */ ld [%o0], %g2 /* Load */ add %o0, 4, %o0 /* IEU0 Group */ sllx %g2, 32, %g2 /* IEU0 Group */ 1: andcc %o1, 2, %g0 /* IEU1 */ be,a,pn %icc, 1f /* CTI */ clr %o4 /* IEU0 Group */ lduh [%o0], %o4 /* Load */ add %o0, 2, %o0 /* IEU1 */ sll %o4, 16, %o4 /* IEU0 Group */ 1: andcc %o1, 1, %g0 /* IEU1 */ be,a,pn %icc, 1f /* CTI */ clr %o5 /* IEU0 Group */ ldub [%o0], %o5 /* Load */ sll %o5, 8, %o5 /* IEU0 Group */ 1: or %g2, %o4, %o4 /* IEU1 */ or %o5, %o4, %o4 /* IEU0 Group (regdep) */ addcc %o4, %o2, %o2 /* IEU1 Group (regdep) */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: ba,pt %xcc, 25b /* CTI Group */ sllx %o2, 32, %g1 /* IEU0 */ 21: srl %o2, 0, %o2 /* IEU0 Group */ cmp %o1, 0 /* IEU1 */ be,pn %icc, 24b /* CTI */ andcc %o1, 4, %g0 /* IEU1 Group */ be,a,pn %icc, 1f /* CTI */ clr %g2 /* IEU0 */ lduh [%o0], %g3 /* Load */ lduh [%o0+2], %g2 /* Load Group */ add %o0, 4, %o0 /* IEU0 Group */ sllx %g3, 48, %g3 /* IEU0 Group */ sllx %g2, 32, %g2 /* IEU0 Group */ or %g3, %g2, %g2 /* IEU0 Group */ 1: andcc %o1, 2, %g0 /* IEU1 */ be,a,pn %icc, 1f /* CTI */ clr %o4 /* IEU0 Group */ lduh [%o0], %o4 /* Load */ add %o0, 2, %o0 /* IEU1 */ sll %o4, 16, %o4 /* IEU0 Group */ 1: andcc %o1, 1, %g0 /* IEU1 */ be,a,pn %icc, 1f /* CTI */ clr %o5 /* IEU0 Group */ ldub [%o0], %o5 /* Load */ sll %o5, 8, %o5 /* IEU0 Group */ 1: or %g2, %o4, %o4 /* IEU1 */ or %o5, %o4, %o4 /* IEU0 Group (regdep) */ addcc %o4, %o2, %o2 /* IEU1 Group (regdep) */ bcs,a,pn %xcc, 1f /* CTI */ add %o2, 1, %o2 /* IEU0 */ 1: ba,pt %xcc, 25b /* CTI Group */ sllx %o2, 32, %g1 /* IEU0 */