This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Combine four insns

On Fri, Aug 6, 2010 at 10:08 PM, Bernd Schmidt <> wrote:
> On 08/06/2010 05:04 PM, Richard Guenther wrote:
>> Do you have statistics how many two, three and four insn combinations
>> a) are tried, b) can be validated, c) are used in the end, for example
>> during a GCC bootstrap?
> Here are some low-tech statistics for my collection of .i files. ?This
> includes gcc, Linux kernel (double counted here since there are two
> versions for different architectures), SPECint2000, Toshi's stress
> testsuite, a popular embedded benchmark and a few other things. ?Let me
> know if you want something else. ?I only did a) and c) because I
> realized too late you probably asked about which ones were discarded due
> to costs.


> $ grep Trying.four log |wc -l
> 307743
> $ grep Trying.three log |wc -l
> 592776
> $ grep Trying.two log |wc -l
> 1643112
> $ grep Succeeded.two log |wc -l
> 204808
> $ grep Succeeded.three.into.two log |wc -l
> 2976
> $ grep log |wc -l
> 12473
> $ grep Succeeded.four.into.two log |wc -l
> 244
> $ grep log |wc -l
> 140

No four into three?  So overall it's one order of magnitude less
three than two and two order of magnitude less four than three.

I think it's still reasonable (and I agree with that combine is a proper
place to do this).  Can you, for the list of examples you found where
it helps file enhancement bugreports so that we might do sth here
on the tree level or do better initial expansion?


>> It might make sense to restrict 4 insn combinations to
>> -fexpensive-optimizations (thus, not enable it at -O1).
> This, certainly.
> Bernd

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]