https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114480
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amonakov at gcc dot gnu.org
--- Comment #19 from Richard Biener <rguenth at gcc dot gnu.org> ---
Alexander - the testcase at -O1 shows curiously high
3.16% 9840 cc1plus cc1plus [.] mergesort<sort_ctx>
which is attributed (by callgrind) to
if (sizeof (size_t) == 8 && LIKELY (c->size == 8))
--> MERGE_ELTSIZE (8);
and the caller in tree-into-ssa.cc:prune_unused_phi_nodes doing
qsort (defs, adef, sizeof (struct dom_dfsnum), cmp_dfsnum);
I'm not sure why callgrind pins it this way, but perf somewhat agrees:
Samples│ │MERGE_ELTSIZE (8);
▒
1 │2d0:│ mov %r9,%rsi
▒
8 │ │ mov %r9,0x8(%rsp)
▒
528 │ │ mov %r12,%rdi
▒
31 │ │→ call *0x0(%r13)
▒
236 │ │ mov 0x8(%rsp),%r9
▒
2 │ │ sar $0x1f,%eax
▒
244 │ │ mov %r12,%rcx
▒
│ │ movslq %eax,%rdx
▒
531 │ │ and $0x8,%eax
▒
62 │ │ add $0x8,%rbx
▒
│ │ cltq
◆
725 │ │ xor %r9,%rcx
▒
914 │ │ add %rax,%r12
▒
1 │ │ and %rdx,%rcx
▒
│ │ xor %r9,%rcx
▒
3 │ │ mov (%rcx),%rcx
▒
2155 │ │ mov %rcx,-0x8(%rbx)
▒
29 │ │ cmp %r12,%rbx
▒
│ └──je 1d7
I'll note the swapping of 8 bytes is a bit odd and it seems to be
if-converted, thus always doing a write.
I'm of course questioning what prune_unused_phi_nodes does but I have no
idea if that's sensible at all, but it seems slow for this testcase, and
the sorting is the slowest part of it.
