8
alloc r2=ar.pfs,2,0,0,0
9
add r8=31,in1 // round up to 32 byte-boundary
11
shr.u r8=r8,5 // we flush 32 bytes per iteration
15
mov r3=ar.lc // save ar.lc
21
.loop: fc in0 // issuable on M0 only
23
br.cloop.sptk.few .loop
29
mov ar.lc=r3 // restore ar.lc
33
#elif defined(__i386__)
41
# error Need flush_cache code for this architecture.