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: [PATCH] PR middle-end/18293: Fast-path expand_mult for 2^N

On Tuesday, December 7, 2004, at 05:51 PM, Roger Sayle wrote:
[But perhaps my understanding of GCC's middle-end is unduly influencing
my preferred coding style].

Yes, I think so. The code should be targeted first, at humans. When the profile proves that it is important to not do that, then we deviate. I haven't seen the data for this one case, but, I can't imagine that it blows up by more than 20%.

If you are worried about better turn around for developers, the first think I would like to see is better dependancy tracking for tm.h files, so that you can edit a single comment in a port file, and recompile, and have nothing recompile, instead of half of the entire world.

I was thinking about doing something radical like saving every -E during compilation, and doing a move if changed on them all to produce the .o files in much the same spirit as one of the existing stamp idioms.

I'd expect the benefit would be better compile times for all port maintainers, and the improvement would be large.

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