I can't get used to the lack of reply-to field for gcc-help.
On Tue, Aug 24, 2010 at 00:33, Segher Boessenkool
<segher@xxxxxxxxxxxxxxxxxxx> wrote:
>> inline uint64x64_t &operator += (uint64x64_t &lhs, const uint64x64_t &rhs)
>> {
>> lhs._v.hi += rhs._v.hi;
>> lhs._v.lo += rhs._v.lo;
>> if (lhs._v.hi < rhs._v.lo)
>
> if (lhs._v.lo < rhs._v.lo)
gah :/
>
>> {
>> lhs._v.hi++;
>> }
>> return lhs;
>> }
>
> Does the generated code look any better with that correction? If not, you
> want to tell us the exact command line and GCC version you used.
[mlacage@diese simulator]$ gcc --version
gcc (GCC) 4.3.2 20081105 (Red Hat 4.3.2-7)
Yes, it looks much better, I get:
804b850: 8b 55 e4 mov -0x1c(%ebp),%edx
804b853: 8b 7d 08 mov 0x8(%ebp),%edi
804b856: 8b 44 0a 08 mov 0x8(%edx,%ecx,1),%eax
804b85a: 8b 54 0a 0c mov 0xc(%edx,%ecx,1),%edx
804b85e: 01 47 08 add %eax,0x8(%edi)
804b861: 8b 45 e4 mov -0x1c(%ebp),%eax
804b864: 11 57 0c adc %edx,0xc(%edi)
804b867: 03 c 08 add (%eax,%ecx,1),%ebx
804b86a: 13 74 08 04 adc 0x4(%eax,%ecx,1),%esi
804b86e: 89 f mov %ebx,(%edi)
804b870: 89 77 04 mov %esi,0x4(%edi)
804b873: 3b 74 08 04 cmp 0x4(%eax,%ecx,1),%esi
804b877: 77 12 ja 804b88b
<ns3::uint64x64_t run_add<ns3::uint64x64_t>(ns3::uint64x64_t,
ns3::uint64x64_t, long)+0xfb>
But the above is not as good as the following quick hack that is most
likely not correct but that should be close to the minimal code:
asm ("mov 0(%0),%%eax\n\t"
"add 0(%1),%%eax\n\t"
"mov %%eax,0(%0)\n\t"
"mov 4(%0),%%eax\n\t"
"adc 4(%1),%%eax\n\t"
"mov %%eax,4(%0)\n\t"
"mov 8(%0),%%eax\n\t"
"adc 8(%1),%%eax\n\t"
"mov %%eax,8(%0)\n\t"
"mov 12(%0),%%eax\n\t"
"adc 12(%1),%%eax\n\t"
"mov %%eax,12(%0)\n\t"
:
: "r" (&lhs._v), "r" (&rhs._v)
: "%eax", "cc");
I get around 3.5ns for the handcoded assembly while I get 5.1ns for
the compiler-generated one.
Mathieu
--
Mathieu Lacage <mathieu.lacage@xxxxxxxxx>
--
Mathieu Lacage <mathieu.lacage@xxxxxxxxx>
