Re: performance issue with cgywin make

[Christopher Faylor]

On Sat, Dec 09, 2006 at 08:27:29AM -0500, Bill Hoffman wrote:
>Eli Zaretskii wrote:
>>>Date: Fri, 08 Dec 2006 21:46:18 -0500
>>>From: Bill Hoffman <address@hidden>
>>>
>>>So, the windows native stat is about 3 X faster.
>>
>>This is not really a valid conclusion: `stat' does much more than just
>>get the file's attributes via a single Win32 API call.  And Make does
>>use the result of the other API calls, such as st_ino, st_dev, and
>>st_mode, for example.
>  
>Since nmake is performing about 2X faster than gmake, the likely
>culprit is stat.  This all started because a cmake user was trying to
>switch from nmake to gmake, and found it so much slower, then I
>remembered the change we made in cmake.  stat does do more, but I would
>assume that make is mostly interested in comparing file modification
>times, and ignores the rest of those values.
>
>I will poke around in the gmake sources and try to figure out if I can
>replace some of the stat calls.  It looks like I can just change
>name_mtime in remake.c it uses st_mode in one place but only to check
>for symlinks.  I could maybe have a fast mtime if check_symlink_flag is
>not set, in that case only the file time is used.  One issue is that
>FILE_TIMESTAMP is a uintmax_t and the windows call uses __int64 for the
>time stamp.

I'm not sure if you are understanding this or not but there will never
be a version of gnu make in the Cygwin release which replaces stat()
calls with direct calls to Windows functions and I will actively lobby
not to have any such patches added to any cygwin-specific parts of make.

cgf