On Mon, Sep 12, 2011 at 8:13 AM, Paolo Bonzini <pbonz...@redhat.com> wrote:
> On 09/12/2011 10:01 AM, Richard Henderson wrote:
>>
>> > > After this patch set, only load and store op helpers remain in
>> > > op_helper.c. I have some patches for those but they need more
>> > > thought.
>> >
>> > Have you benchmarked it?
>>
>> Asking for a benchmark without full conversion is pointless.
>
> Agreed. But I would not push these patches without having tried them out on
> a prototype of a full conversion (i.e. with the load/store helpers
> converted, for which Blue Swirl said he has patches, and with the
> environment not pinned to AREG0 in TCG code).
The load/store helpers are tricky. Some Sparc64 helpers now need five
32/64 bit arguments, that may be a problem on some hosts. Changing
functions like tlb_fill() and do_unaligned_access() to use passed
CPUState pointer instead of AREG0 needs global changes. The template
system for generating the load/store functions is interesting. Then
there are __ldb_mmu() and friends, called from TCG generated code. It
would be highly desirable to limit the changes to only Sparc
translator but I don't think global changes can be avoided.
> So I hoped that he did have such a prototype, or alternatively that he
> benchmarked them and showed only minor degradations.
I don't see any slowdown. Maybe a real benchmark is needed. Looking at
the code, there are only minor differences.
On amd64 host, r14 is now available but does not get used for the new
code, so that doesn't help.
On i386 there are larger differences, but that is mostly because ebp
is normally used for the frame pointer. Using it for a global register
needs -fomit-frame-pointer. Disregarding the frame pointer issues, the
changes are minor.
For example i386 host, unpatched, op_helper.o:
00000dc0 <helper_udiv>:
dc0: 83 ec 1c sub $0x1c,%esp
dc3: 65 8b 0d 14 00 00 00 mov %gs:0x14,%ecx
dca: 89 4c 24 0c mov %ecx,0xc(%esp)
dce: 31 c9 xor %ecx,%ecx
dd0: 8b 44 24 20 mov 0x20(%esp),%eax
dd4: 8b 54 24 24 mov 0x24(%esp),%edx
dd8: 8b 4c 24 0c mov 0xc(%esp),%ecx
ddc: 65 33 0d 14 00 00 00 xor %gs:0x14,%ecx
de3: 75 0a jne def <helper_udiv+0x2f>
de5: 31 c9 xor %ecx,%ecx
de7: 83 c4 1c add $0x1c,%esp
dea: e9 f1 fe ff ff jmp ce0 <helper_udiv_common>
def: e8 fc ff ff ff call df0 <helper_udiv+0x30>
df4: 8d b6 00 00 00 00 lea 0x0(%esi),%esi
dfa: 8d bf 00 00 00 00 lea 0x0(%edi),%edi
Patched, function in helper.o:
000002a0 <helper_udiv>:
2a0: 55 push %ebp
2a1: 89 e5 mov %esp,%ebp
2a3: 53 push %ebx
2a4: 83 ec 14 sub $0x14,%esp
2a7: 8b 45 08 mov 0x8(%ebp),%eax
2aa: 65 8b 1d 14 00 00 00 mov %gs:0x14,%ebx
2b1: 89 5d f4 mov %ebx,-0xc(%ebp)
2b4: 31 db xor %ebx,%ebx
2b6: 8b 55 0c mov 0xc(%ebp),%edx
2b9: 8b 4d 10 mov 0x10(%ebp),%ecx
2bc: 8b 5d f4 mov -0xc(%ebp),%ebx
2bf: 65 33 1d 14 00 00 00 xor %gs:0x14,%ebx
2c6: 75 11 jne 2d9 <helper_udiv+0x39>
2c8: c7 45 08 00 00 00 00 movl $0x0,0x8(%ebp)
2cf: 83 c4 14 add $0x14,%esp
2d2: 5b pop %ebx
2d3: 5d pop %ebp
2d4: e9 e7 fe ff ff jmp 1c0 <helper_udiv_common>
2d9: e8 fc ff ff ff call 2da <helper_udiv+0x3a>
