diff options
Diffstat (limited to 'sysdeps/ia64/bzero.S')
| -rw-r--r-- | sysdeps/ia64/bzero.S | 334 |
1 files changed, 277 insertions, 57 deletions
diff --git a/sysdeps/ia64/bzero.S b/sysdeps/ia64/bzero.S index 3225ed4975..bcca41d5e9 100644 --- a/sysdeps/ia64/bzero.S +++ b/sysdeps/ia64/bzero.S @@ -1,7 +1,8 @@ /* Optimized version of the standard bzero() function. This file is part of the GNU C Library. Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc. - Contributed by Dan Pop <Dan.Pop@cern.ch>. + Contributed by Dan Pop for Itanium <Dan.Pop@cern.ch>. + Rewritten for McKinley by Sverre Jarp, HP Labs/CERN <Sverre.Jarp@cern.ch> The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public @@ -25,8 +26,11 @@ in1: count The algorithm is fairly straightforward: set byte by byte until we - we get to a word aligned address, then set word by word as much as - possible; the remaining few bytes are set one by one. */ + we get to a 16B-aligned address, then loop on 128 B chunks using an + early store as prefetching, then loop on 32B chucks, then clear remaining + words, finally clear remaining bytes. + Since a stf.spill f0 can store 16B in one go, we use this instruction + to get peak speed. */ #include <sysdep.h> #undef ret @@ -34,62 +38,278 @@ #define dest in0 #define cnt in1 -#define save_pfs loc0 -#define ptr1 loc1 -#define ptr2 loc2 -#define tmp loc3 -#define loopcnt loc4 -#define save_lc loc5 +#define tmp r31 +#define save_lc r30 +#define ptr0 r29 +#define ptr1 r28 +#define ptr2 r27 +#define ptr3 r26 +#define ptr9 r24 +#define loopcnt r23 +#define linecnt r22 +#define bytecnt r21 -ENTRY(__bzero) +// This routine uses only scratch predicate registers (p6 - p15) +#define p_scr p6 // default register for same-cycle branches +#define p_unalgn p9 +#define p_y p11 +#define p_n p12 +#define p_yy p13 +#define p_nn p14 + +#define movi0 mov + +#define MIN1 15 +#define MIN1P1HALF 8 +#define LINE_SIZE 128 +#define LSIZE_SH 7 // shift amount +#define PREF_AHEAD 8 + +#define USE_FLP +#if defined(USE_INT) +#define store st8 +#define myval r0 +#elif defined(USE_FLP) +#define store stf8 +#define myval f0 +#endif + +.align 64 +ENTRY(bzero) +{ .mmi .prologue - alloc save_pfs = ar.pfs, 2, 6, 0, 0 - .save ar.lc, save_lc - mov save_lc = ar.lc + alloc tmp = ar.pfs, 2, 0, 0, 0 + lfetch.nt1 [dest] + .save ar.lc, save_lc + movi0 save_lc = ar.lc +} { .mmi .body - mov ret0 = dest - and tmp = 7, dest - cmp.eq p6, p0 = cnt, r0 -(p6) br.cond.spnt .restore_and_exit ;; + mov ret0 = dest // return value + nop.m 0 + cmp.eq p_scr, p0 = cnt, r0 +;; } +{ .mmi + and ptr2 = -(MIN1+1), dest // aligned address + and tmp = MIN1, dest // prepare to check for alignment + tbit.nz p_y, p_n = dest, 0 // Do we have an odd address? (M_B_U) +} { .mib mov ptr1 = dest - sub loopcnt = 8, tmp - cmp.gt p6, p0 = 16, cnt -(p6) br.cond.spnt .set_few;; - cmp.eq p6, p0 = tmp, r0 -(p6) br.cond.sptk .dest_aligned - sub cnt = cnt, loopcnt - adds loopcnt = -1, loopcnt;; - mov ar.lc = loopcnt;; -.l1: - st1 [ptr1] = r0, 1 - br.cloop.dptk .l1 ;; -.dest_aligned: - adds ptr2 = 8, ptr1 - shr.u loopcnt = cnt, 4 ;; // loopcnt = cnt / 16 - cmp.eq p6, p0 = loopcnt, r0 -(p6) br.cond.spnt .one_more - and cnt = 0xf, cnt // compute the remaining cnt - adds loopcnt = -1, loopcnt;; - mov ar.lc = loopcnt;; -.l2: - st8 [ptr1] = r0, 16 - st8 [ptr2] = r0, 16 - br.cloop.dptk .l2 - cmp.le p6, p0 = 8, cnt ;; -.one_more: -(p6) st8 [ptr1] = r0, 8 -(p6) adds cnt = -8, cnt ;; - cmp.eq p6, p0 = cnt, r0 -(p6) br.cond.spnt .restore_and_exit -.set_few: - adds loopcnt = -1, cnt;; - mov ar.lc = loopcnt;; -.l3: - st1 [ptr1] = r0, 1 - br.cloop.dptk .l3 ;; + nop.i 0 +(p_scr) br.ret.dpnt.many rp // return immediately if count = 0 +;; } +{ .mib + cmp.ne p_unalgn, p0 = tmp, r0 +} { .mib // NB: # of bytes to move is 1 + sub bytecnt = (MIN1+1), tmp // higher than loopcnt + cmp.gt p_scr, p0 = 16, cnt // is it a minimalistic task? +(p_scr) br.cond.dptk.many .move_bytes_unaligned // go move just a few (M_B_U) +;; } +{ .mmi +(p_unalgn) add ptr1 = (MIN1+1), ptr2 // after alignment +(p_unalgn) add ptr2 = MIN1P1HALF, ptr2 // after alignment +(p_unalgn) tbit.nz.unc p_y, p_n = bytecnt, 3 // should we do a st8 ? +;; } +{ .mib +(p_y) add cnt = -8, cnt +(p_unalgn) tbit.nz.unc p_yy, p_nn = bytecnt, 2 // should we do a st4 ? +} { .mib +(p_y) st8 [ptr2] = r0,-4 +(p_n) add ptr2 = 4, ptr2 +;; } +{ .mib +(p_yy) add cnt = -4, cnt +(p_unalgn) tbit.nz.unc p_y, p_n = bytecnt, 1 // should we do a st2 ? +} { .mib +(p_yy) st4 [ptr2] = r0,-2 +(p_nn) add ptr2 = 2, ptr2 +;; } +{ .mmi + mov tmp = LINE_SIZE+1 // for compare +(p_y) add cnt = -2, cnt +(p_unalgn) tbit.nz.unc p_yy, p_nn = bytecnt, 0 // should we do a st1 ? +} { .mmi + nop.m 0 +(p_y) st2 [ptr2] = r0,-1 +(p_n) add ptr2 = 1, ptr2 +;; } + +{ .mmi +(p_yy) st1 [ptr2] = r0 + cmp.gt p_scr, p0 = tmp, cnt // is it a minimalistic task? +} { .mbb +(p_yy) add cnt = -1, cnt +(p_scr) br.cond.dpnt.many .fraction_of_line // go move just a few +;; } +{ .mib + nop.m 0 + shr.u linecnt = cnt, LSIZE_SH + nop.b 0 +;; } + + .align 32 +.l1b: // ------------------// L1B: store ahead into cache lines; fill later +{ .mmi + and tmp = -(LINE_SIZE), cnt // compute end of range + mov ptr9 = ptr1 // used for prefetching + and cnt = (LINE_SIZE-1), cnt // remainder +} { .mmi + mov loopcnt = PREF_AHEAD-1 // default prefetch loop + cmp.gt p_scr, p0 = PREF_AHEAD, linecnt // check against actual value +;; } +{ .mmi +(p_scr) add loopcnt = -1, linecnt + add ptr2 = 16, ptr1 // start of stores (beyond prefetch stores) + add ptr1 = tmp, ptr1 // first address beyond total range +;; } +{ .mmi + add tmp = -1, linecnt // next loop count + movi0 ar.lc = loopcnt +;; } +.pref_l1b: +{ .mib + stf.spill [ptr9] = f0, 128 // Do stores one cache line apart + nop.i 0 + br.cloop.dptk.few .pref_l1b +;; } +{ .mmi + add ptr0 = 16, ptr2 // Two stores in parallel + movi0 ar.lc = tmp +;; } +.l1bx: + { .mmi + stf.spill [ptr2] = f0, 32 + stf.spill [ptr0] = f0, 32 + ;; } + { .mmi + stf.spill [ptr2] = f0, 32 + stf.spill [ptr0] = f0, 32 + ;; } + { .mmi + stf.spill [ptr2] = f0, 32 + stf.spill [ptr0] = f0, 64 + cmp.lt p_scr, p0 = ptr9, ptr1 // do we need more prefetching? + ;; } +{ .mmb + stf.spill [ptr2] = f0, 32 +(p_scr) stf.spill [ptr9] = f0, 128 + br.cloop.dptk.few .l1bx +;; } +{ .mib + cmp.gt p_scr, p0 = 8, cnt // just a few bytes left ? +(p_scr) br.cond.dpnt.many .move_bytes_from_alignment +;; } + +.fraction_of_line: +{ .mib + add ptr2 = 16, ptr1 + shr.u loopcnt = cnt, 5 // loopcnt = cnt / 32 +;; } +{ .mib + cmp.eq p_scr, p0 = loopcnt, r0 + add loopcnt = -1, loopcnt +(p_scr) br.cond.dpnt.many .store_words +;; } +{ .mib + and cnt = 0x1f, cnt // compute the remaining cnt + movi0 ar.lc = loopcnt +;; } + .align 32 +.l2: // -----------------------------// L2A: store 32B in 2 cycles +{ .mmb + store [ptr1] = myval, 8 + store [ptr2] = myval, 8 +;; } { .mmb + store [ptr1] = myval, 24 + store [ptr2] = myval, 24 + br.cloop.dptk.many .l2 +;; } +.store_words: +{ .mib + cmp.gt p_scr, p0 = 8, cnt // just a few bytes left ? +(p_scr) br.cond.dpnt.many .move_bytes_from_alignment // Branch +;; } + +{ .mmi + store [ptr1] = myval, 8 // store + cmp.le p_y, p_n = 16, cnt // + add cnt = -8, cnt // subtract +;; } +{ .mmi +(p_y) store [ptr1] = myval, 8 // store +(p_y) cmp.le.unc p_yy, p_nn = 16, cnt +(p_y) add cnt = -8, cnt // subtract +;; } +{ .mmi // store +(p_yy) store [ptr1] = myval, 8 +(p_yy) add cnt = -8, cnt // subtract +;; } + +.move_bytes_from_alignment: +{ .mib + cmp.eq p_scr, p0 = cnt, r0 + tbit.nz.unc p_y, p0 = cnt, 2 // should we terminate with a st4 ? +(p_scr) br.cond.dpnt.few .restore_and_exit +;; } +{ .mib +(p_y) st4 [ptr1] = r0,4 + tbit.nz.unc p_yy, p0 = cnt, 1 // should we terminate with a st2 ? +;; } +{ .mib +(p_yy) st2 [ptr1] = r0,2 + tbit.nz.unc p_y, p0 = cnt, 0 // should we terminate with a st1 ? +;; } + +{ .mib +(p_y) st1 [ptr1] = r0 +;; } .restore_and_exit: - mov ar.lc = save_lc - mov ar.pfs = save_pfs - br.ret.sptk.many b0 -END(__bzero) -weak_alias (__bzero, bzero) +{ .mib + nop.m 0 + movi0 ar.lc = save_lc + br.ret.sptk.many rp +;; } + +.move_bytes_unaligned: +{ .mmi + .pred.rel "mutex",p_y, p_n + .pred.rel "mutex",p_yy, p_nn +(p_n) cmp.le p_yy, p_nn = 4, cnt +(p_y) cmp.le p_yy, p_nn = 5, cnt +(p_n) add ptr2 = 2, ptr1 +} { .mmi +(p_y) add ptr2 = 3, ptr1 +(p_y) st1 [ptr1] = r0, 1 // fill 1 (odd-aligned) byte +(p_y) add cnt = -1, cnt // [15, 14 (or less) left] +;; } +{ .mmi +(p_yy) cmp.le.unc p_y, p0 = 8, cnt + add ptr3 = ptr1, cnt // prepare last store + movi0 ar.lc = save_lc +} { .mmi +(p_yy) st2 [ptr1] = r0, 4 // fill 2 (aligned) bytes +(p_yy) st2 [ptr2] = r0, 4 // fill 2 (aligned) bytes +(p_yy) add cnt = -4, cnt // [11, 10 (o less) left] +;; } +{ .mmi +(p_y) cmp.le.unc p_yy, p0 = 8, cnt + add ptr3 = -1, ptr3 // last store + tbit.nz p_scr, p0 = cnt, 1 // will there be a st2 at the end ? +} { .mmi +(p_y) st2 [ptr1] = r0, 4 // fill 2 (aligned) bytes +(p_y) st2 [ptr2] = r0, 4 // fill 2 (aligned) bytes +(p_y) add cnt = -4, cnt // [7, 6 (or less) left] +;; } +{ .mmi +(p_yy) st2 [ptr1] = r0, 4 // fill 2 (aligned) bytes +(p_yy) st2 [ptr2] = r0, 4 // fill 2 (aligned) bytes + // [3, 2 (or less) left] + tbit.nz p_y, p0 = cnt, 0 // will there be a st1 at the end ? +} { .mmi +(p_yy) add cnt = -4, cnt +;; } +{ .mmb +(p_scr) st2 [ptr1] = r0 // fill 2 (aligned) bytes +(p_y) st1 [ptr3] = r0 // fill last byte (using ptr3) + br.ret.sptk.many rp +;; } +END(bzero) |