Greetings.
Chatting on IRC today (freenode #gcc), someone brought up the
following example code:
#include <stdio.h>
template <typename T>
inline T const& max (T const& a, T const& b)
{
// if a < b then use b else use a
return a<b?b:a;
}
int main()
{
long long unsigned sum = 0;
for(int x = 1; x <= 100000000; x++)
{
sum+=max(x,x+1);
}
printf("%llu\n", sum);
}
They noticed that their earlier compiler (4.6.3 -O3) successfully
reduced the loop, while 4.8.1 didn't.
I tested it with 4.7.2 (on Fedora 18 x86-64). Without optimization,
the loop was preserved in the object code:
00000000004005ec <main>:
...
400605: 8b 45 f0 mov -0x10(%rbp),%eax
400608: 83 c0 01 add $0x1,%eax
40060b: 89 45 f4 mov %eax,-0xc(%rbp)
40060e: 48 8d 55 f4 lea -0xc(%rbp),%rdx
400612: 48 8d 45 f0 lea -0x10(%rbp),%rax
400616: 48 89 d6 mov %rdx,%rsi
400619: 48 89 c7 mov %rax,%rdi
40061c: e8 3d 00 00 00 callq 40065e <_Z3maxIiERKT_S2_S2_>
400621: 8b 00 mov (%rax),%eax
400623: 48 98 cltq
400625: 48 01 45 f8 add %rax,-0x8(%rbp)
400629: 8b 45 f0 mov -0x10(%rbp),%eax
40062c: 83 c0 01 add $0x1,%eax
40062f: 89 45 f0 mov %eax,-0x10(%rbp)
400632: 8b 45 f0 mov -0x10(%rbp),%eax
400635: 3d 00 e1 f5 05 cmp $0x5f5e100,%eax
40063a: 0f 9e c0 setle %al
40063d: 84 c0 test %al,%al
40063f: 75 c4 jne 400605 <main+0x19>
Using -O3, 4.7.2 successfully reduced the loop:
0000000000400500 <main>:
400500: 48 83 ec 08 sub $0x8,%rsp
400504: 48 be 80 51 d1 40 79 movabs $0x11c37940d15180,%rsi
40050b: c3 11 00
40050e: bf c0 06 40 00 mov $0x4006c0,%edi
400513: 31 c0 xor %eax,%eax
400515: e8 b6 ff ff ff callq 4004d0 <printf@plt>
40051a: 31 c0 xor %eax,%eax
40051c: 48 83 c4 08 add $0x8,%rsp
400520: c3 retq
400521: 0f 1f 00 nopl (%rax)
Unfortunately, 4.8.1 -O3 does *not* collapse the loop. It uses the
xmm registers, but it still does a loop (so far as I can tell -- I'm
not very good at reading assembler):
00000000004004f0 <main>:
...
400518: 66 0f 6f cc movdqa %xmm4,%xmm1
40051c: 66 0f 6f d5 movdqa %xmm5,%xmm2
400520: 83 c0 01 add $0x1,%eax
400523: 66 0f 6f e1 movdqa %xmm1,%xmm4
400527: 66 0f fe ce paddd %xmm6,%xmm1
40052b: 3d 40 78 7d 01 cmp $0x17d7840,%eax
400530: 66 0f 66 d1 pcmpgtd %xmm1,%xmm2
400534: 66 0f 6f d9 movdqa %xmm1,%xmm3
400538: 66 0f fe e7 paddd %xmm7,%xmm4
40053c: 66 0f 62 da punpckldq %xmm2,%xmm3
400540: 66 0f 6a ca punpckhdq %xmm2,%xmm1
400544: 66 0f d4 c3 paddq %xmm3,%xmm0
400548: 66 0f d4 c1 paddq %xmm1,%xmm0
40054c: 75 ca jne 400518 <main+0x28>
...
While this is noticably faster than the 4.7.2 loop, it's still an
order of magnitude slower than the fully-collapsed loop given by -O3
on 4.7.2 and 4.6.3 (times are in nanoseconds) [1]:
maxtest-4.7.2-O0: 332,635,216
maxtest-4.7.2-O3: 1,754,238
maxtest-4.8.1-O3: 28,128,557
Is this expected? Is it a side-effect of some of the other loop work
that landed in 4.8? Are there other options that can restore the
pre-4.8 behavior in this case?
Thanks,
Tony
[1] Quick and dirty script for generating those numbers:
for i in maxtest-* ; do
t0=`date +'%s%N'` ;
for j in {0..9} ; do
./$i > /dev/null ;
done ;
t1=`date +'%s%N'` ;
echo -n "$i: " ;
echo "($t1-$t0)/10" | bc ;
done
