/* Optimized version of the standard memset() function. Copyright (c) 2002 Hewlett-Packard Co/CERN Sverre Jarp Return: dest Inputs: in0: dest in1: value in2: count The algorithm is fairly straightforward: set byte by byte until we 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 when value = 0. */ #include #undef ret #define dest in0 #define value in1 #define cnt in2 #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 #define fvalue f6 // This routine uses only scratch predicate registers (p6 - p15) #define p_scr p6 // default register for same-cycle branches #define p_nz p7 #define p_zr p8 #define p_unalgn p9 #define p_y p11 #define p_n p12 #define p_yy p13 #define p_nn p14 #define MIN1 15 #define MIN1P1HALF 8 #define LINE_SIZE 128 #define LSIZE_SH 7 // shift amount #define PREF_AHEAD 8 GLOBAL_ENTRY(memset) { .mmi .prologue alloc tmp = ar.pfs, 3, 0, 0, 0 .body lfetch.nt1 [dest] // .save ar.lc, save_lc mov.i save_lc = ar.lc } { .mmi mov ret0 = dest // return value cmp.ne p_nz, p_zr = value, r0 // use stf.spill if value is zero cmp.eq p_scr, p0 = cnt, r0 ;; } { .mmi and ptr2 = -(MIN1+1), dest // aligned address and tmp = MIN1, dest // prepare to check for correct alignment tbit.nz p_y, p_n = dest, 0 // Do we have an odd address? (M_B_U) } { .mib mov ptr1 = dest mux1 value = value, @brcst // create 8 identical bytes in word (p_scr) br.ret.dpnt.many rp // return immediately if count = 0 ;; } { .mib cmp.ne p_unalgn, p0 = tmp, r0 // } { .mib sub bytecnt = (MIN1+1), tmp // NB: # of bytes to move is 1 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] = value,-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] = value,-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 setf.sig fvalue=value // transfer value to FLP side (p_y) st2 [ptr2] = value,-1 // (p_n) add ptr2 = 1, ptr2 // ;; } { .mmi (p_yy) st1 [ptr2] = value // 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 (p_zr) br.cond.dptk.many .l1b // Jump to use stf.spill ;; } .align 32 // -------------------------- // L1A: 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 = 8, ptr1 // start of stores (beyond prefetch stores) add ptr1 = tmp, ptr1 // first address beyond total range ;; } { .mmi add tmp = -1, linecnt // next loop count mov.i ar.lc = loopcnt // ;; } .pref_l1a: { .mib stf8 [ptr9] = fvalue, 128 // Do stores one cache line apart nop.i 0 br.cloop.dptk.few .pref_l1a ;; } { .mmi add ptr0 = 16, ptr2 // Two stores in parallel mov.i ar.lc = tmp // ;; } .l1ax: { .mmi stf8 [ptr2] = fvalue, 8 stf8 [ptr0] = fvalue, 8 ;; } { .mmi stf8 [ptr2] = fvalue, 24 stf8 [ptr0] = fvalue, 24 ;; } { .mmi stf8 [ptr2] = fvalue, 8 stf8 [ptr0] = fvalue, 8 ;; } { .mmi stf8 [ptr2] = fvalue, 24 stf8 [ptr0] = fvalue, 24 ;; } { .mmi stf8 [ptr2] = fvalue, 8 stf8 [ptr0] = fvalue, 8 ;; } { .mmi stf8 [ptr2] = fvalue, 24 stf8 [ptr0] = fvalue, 24 ;; } { .mmi stf8 [ptr2] = fvalue, 8 stf8 [ptr0] = fvalue, 32 cmp.lt p_scr, p0 = ptr9, ptr1 // do we need more prefetching? ;; } { .mmb stf8 [ptr2] = fvalue, 24 (p_scr) stf8 [ptr9] = fvalue, 128 br.cloop.dptk.few .l1ax ;; } { .mbb cmp.le p_scr, p0 = 8, cnt // just a few bytes left ? (p_scr) br.cond.dpnt.many .fraction_of_line // Branch no. 2 br.cond.dpnt.many .move_bytes_from_alignment // Branch no. 3 ;; } .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 mov.i 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 mov.i 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 mov.i ar.lc = loopcnt ;; } .align 32 .l2: // ------------------------------------ // L2A: store 32B in 2 cycles { .mmb stf8 [ptr1] = fvalue, 8 stf8 [ptr2] = fvalue, 8 ;; } { .mmb stf8 [ptr1] = fvalue, 24 stf8 [ptr2] = fvalue, 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 stf8 [ptr1] = fvalue, 8 // store cmp.le p_y, p_n = 16, cnt add cnt = -8, cnt // subtract ;; } { .mmi (p_y) stf8 [ptr1] = fvalue, 8 // store (p_y) cmp.le.unc p_yy, p_nn = 16, cnt (p_y) add cnt = -8, cnt // subtract ;; } { .mmi // store (p_yy) stf8 [ptr1] = fvalue, 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] = value,4 tbit.nz.unc p_yy, p0 = cnt, 1 // should we terminate with a st2 ? ;; } { .mib (p_yy) st2 [ptr1] = value,2 tbit.nz.unc p_y, p0 = cnt, 0 // should we terminate with a st1 ? ;; } { .mib (p_y) st1 [ptr1] = value ;; } .restore_and_exit: { .mib nop.m 0 mov.i 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] = value, 1 // fill 1 (odd-aligned) byte [15, 14 (or less) left] (p_y) add cnt = -1, cnt ;; } { .mmi (p_yy) cmp.le.unc p_y, p0 = 8, cnt add ptr3 = ptr1, cnt // prepare last store mov.i ar.lc = save_lc } { .mmi (p_yy) st2 [ptr1] = value, 4 // fill 2 (aligned) bytes (p_yy) st2 [ptr2] = value, 4 // fill 2 (aligned) bytes [11, 10 (o less) left] (p_yy) add cnt = -4, cnt ;; } { .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] = value, 4 // fill 2 (aligned) bytes (p_y) st2 [ptr2] = value, 4 // fill 2 (aligned) bytes [7, 6 (or less) left] (p_y) add cnt = -4, cnt ;; } { .mmi (p_yy) st2 [ptr1] = value, 4 // fill 2 (aligned) bytes (p_yy) st2 [ptr2] = value, 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] = value // fill 2 (aligned) bytes (p_y) st1 [ptr3] = value // fill last byte (using ptr3) br.ret.sptk.many rp } END(memset)