Messages in this thread | | | From | David Howells <> | Subject | Re: [PATCH] Reduce the number of expensive division instructions done by _parse_integer() | Date | Thu, 09 Feb 2012 17:46:07 +0000 |
| |
Linus Torvalds <torvalds@linux-foundation.org> wrote:
> Looking at the code generated, the "val >> 60" thing actually does > generate a shift, and at least on x86-64, the attached patch generates > better code.
On fixed-size instruction arches, the runtime shift is probably the better option, as simply loading 64-bit large constant would take likely take at least four instructions - and might involve a shift anyway. On the other hand, it seems the compiler can optimise your suggestion fairly well. In both cases, the 64-bit arithmetic can be reduced to 32-bit arithmetic on the MSW only on 32-bit arches.
On x86_64 we have:
400649: 48 89 d8 mov %rbx,%rax 40064c: 48 c1 e8 3c shr $0x3c,%rax 400650: 48 85 c0 test %rax,%rax 400653: 75 52 jne 4006a7 <_parse_integer+0xa7>
And on i386 we have:
8048532: 8b 54 24 14 mov 0x14(%esp),%edx ... 8048538: 89 c7 mov %eax,%edi 804853a: c1 ea 1c shr $0x1c,%edx 804853d: 85 d2 test %edx,%edx 804853f: 75 79 jne 80485ba <_parse_integer+0xda>
With your code, we have on x86_64:
40062d: 49 bf 00 00 00 00 00 movabs $0xf000000000000000,%r15 400634: 00 00 f0 ... 400659: 4c 85 fb test %r15,%rbx 40065c: 75 59 jne 4006b7 <_parse_integer+0xb7>
And on i386:
804853c: 89 c7 mov %eax,%edi 804853e: f7 44 24 1c 00 00 00 testl $0xf0000000,0x1c(%esp) 8048545: f0 8048546: 75 79 jne 80485c1 <_parse_integer+0xe1>
But it will work too. And I like the pointer indirection removal as well.
I'm not sure there's a lot to choose between them, though I prefer mine as I think it produces slightly smaller code.
Want me to wrap these changes up with my patch description?
David
| |