123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166 |
- #include <sysdep.h>
- #undef ret
- #define OP_T_THRES 16
- #define OPSIZ 8
- #define MEMLAT 2
- #define start r15
- #define saved_pr r17
- #define saved_lc r18
- #define dest r19
- #define src r20
- #define len r21
- #define asrc r22
- #define tmp r23
- #define value1 r24
- #define value2 r25
- #define sh2 r28
- #define sh1 r29
- #define loopcnt r30
- ENTRY(memcmp)
- .prologue
- alloc r2 = ar.pfs, 3, 37, 0, 40
- .rotr r[MEMLAT + 2], q[MEMLAT + 5], tmp1[4], tmp2[4], val[2]
- .rotp p[MEMLAT + 4 + 1]
- mov ret0 = r0
- .save pr, saved_pr
- mov saved_pr = pr
- .save ar.lc, saved_lc
- mov saved_lc = ar.lc
- .body
- mov dest = in0
- mov src = in1
- mov len = in2
- sub tmp = r0, in0
-
- and loopcnt = 7, tmp
- cmp.ge p6, p0 = OP_T_THRES, len
- (p6) br.cond.spnt .cmpfew
-
- cmp.eq p6, p0 = loopcnt, r0
- (p6) br.cond.sptk .dest_aligned
- sub len = len, loopcnt
- adds loopcnt = -1, loopcnt
-
- mov ar.lc = loopcnt
- .l1:
- ld1 value1 = [src], 1
- ld1 value2 = [dest], 1
-
- cmp.ne p6, p0 = value1, value2
- (p6) br.cond.spnt .done
- br.cloop.dptk .l1
- .dest_aligned:
- and sh1 = 7, src
- and tmp = -8, len
- and asrc = -8, src
- shr.u loopcnt = len, 3
- and len = 7, len
- shl sh1 = sh1, 3
- adds loopcnt = -1, loopcnt
- mov pr.rot = 1 << 16
- sub sh2 = 64, sh1
- mov ar.lc = loopcnt
- cmp.eq p6, p0 = sh1, r0
- (p6) br.cond.sptk .src_aligned
- add src = src, tmp
- mov ar.ec = MEMLAT + 4 + 1
- ld8 r[1] = [asrc], 8
- .align 32
- .l2:
- (p[0]) ld8 r[0] = [asrc], 8
- (p[0]) ld8 q[0] = [dest], 8
- (p[MEMLAT]) shr.u tmp1[0] = r[1 + MEMLAT], sh1
- (p[MEMLAT]) shl tmp2[0] = r[0 + MEMLAT], sh2
- (p[MEMLAT+4]) cmp.ne p6, p0 = q[MEMLAT + 4], val[1]
- (p[MEMLAT+3]) or val[0] = tmp1[3], tmp2[3]
- (p6) br.cond.spnt .l2exit
- br.ctop.sptk .l2
- br.cond.sptk .cmpfew
- .l3exit:
- mux1 value1 = r[MEMLAT],
- mux1 value2 = q[MEMLAT],
- cmp.ne p6, p0 = r0, r0
- .l2exit:
- (p6) mux1 value1 = val[1],
- (p6) mux1 value2 = q[MEMLAT + 4],
- cmp.ltu p6, p7 = value2, value1
- (p6) mov ret0 = -1
- (p7) mov ret0 = 1
- mov pr = saved_pr, -1
- mov ar.lc = saved_lc
- br.ret.sptk.many b0
- .src_aligned:
- cmp.ne p6, p0 = r0, r0
- mov ar.ec = MEMLAT + 1
- .l3:
- (p[0]) ld8 r[0] = [src], 8
- (p[0]) ld8 q[0] = [dest], 8
- (p[MEMLAT]) cmp.ne p6, p0 = r[MEMLAT], q[MEMLAT]
- (p6) br.cond.spnt .l3exit
- br.ctop.dptk .l3
- .cmpfew:
- cmp.eq p6, p0 = len, r0
- adds len = -1, len
- (p6) br.cond.spnt .restore_and_exit
- mov ar.lc = len
- .l4:
- ld1 value1 = [src], 1
- ld1 value2 = [dest], 1
-
- cmp.ne p6, p0 = value1, value2
- (p6) br.cond.spnt .done
- br.cloop.dptk .l4
- .done:
- (p6) sub ret0 = value2, value1
- .restore_and_exit:
- mov pr = saved_pr, -1
- mov ar.lc = saved_lc
- br.ret.sptk.many b0
- END(memcmp)
- libc_hidden_def (memcmp)
- #ifdef __UCLIBC_SUSV3_LEGACY__
- strong_alias (memcmp, bcmp)
- #endif
|