Tweaking configure.in

[grinvader]

Currently...

Code: Select all

if test x$debug != xyes; then
	case "$target" in
		i486-*-*)
			CFLAGS="$CFLAGS -march=i486"
			AC_MSG_RESULT(486)
			;;
		i586-*-*)
			CFLAGS="$CFLAGS -march=pentium"
			AC_MSG_RESULT(586)
			;;
		i686-*-*)
			CFLAGS="$CFLAGS -march=pentiumpro"
			dnl CFLAGS="$CFLAGS -march=pentium3 -mmmx -msse -mfpmath=sse,387"
			AC_MSG_RESULT(686)
			;;
		*)
			AC_MSG_RESULT(386)
			AC_MSG_WARN(*** This is probably not what you want use --target)
			;;

Now that I know where to modify, I'd like some pointers on how to add other cpus to help optimizations.
Of course, we can do the ./autogen.sh, then manually edit Makefile to set the CFlag to whatever we want, but might as well put it in the autogen, no ?

[Nach]

Do be aware that changing the optimizations on C code is not going to do much. Sure you can get faster DSP-1, Seta 10, some C4, but that's about it.

[Kagerato]

An excellent point, Nach. With the massive amount of x86, the C optimization settings won't make much difference.

However, I believe grinvader is probably thinking of the future. In which case, adding other processor's preferred optimizations is not a bad idea.

[grinvader]

After some tries... I now use:

Code: Select all

CFLAGS="$CFLAGS -march=pentium4 -mmmx -msse -msse2 -mfpmath=sse,387"

The size of the UPX-compressed exe went down by 20 kB. And I swear that the water effect A (my personnal favourite CPU eater) is 10 times faster.

Also worth of note: Yoshi's Island (U) [SuperFX] is now always over 52/60. @1024*768 OGL FULL with hq2x and bilinear filters.
Even after 'touching fuzzy, getting dizzy'.

[Nach]

After some tries... I now use:

Code: Select all

CFLAGS="$CFLAGS -march=pentium4 -mmmx -msse -msse2 -mfpmath=sse,387"

Do you realize that -march=pentium4 includes those other ones?

[ipher]

when you use -march=whatever, it usually uses all of the optimizations that the processor uses, since you're basically saying "this processor is required".

[grinvader]

I wasn't sure about that, so I put them explicitly.
Let me rebuild and check if the exe's the same...

Edit: not quite. pentium4 alone makes an even smaller exe, but slightly slower on some intensive parts of Yoshi's Island involving flying drug clouds.

[Kagerato]

There's been much confusion and debate about whether the march flags actually include the mmx/sse/etc optimizations (they certainly imply them; I believe that's even mentioned in the gcc manual). I'm fairly sure they are included, and that only in rare circumstances would manually specifying -msse and so on makes a positive difference.

I use "-march=pentium4 -O2" for my CFLAGS in make.conf (gentoo system). I've found that -O3 does not improve speed by any significant margin over -O2.

[jdratlif]

Currently...

Code: Select all

if test x$debug != xyes; then
	case "$target" in
		i486-*-*)
			CFLAGS="$CFLAGS -march=i486"
			AC_MSG_RESULT(486)
			;;
		i586-*-*)
			CFLAGS="$CFLAGS -march=pentium"
			AC_MSG_RESULT(586)
			;;
		i686-*-*)
			CFLAGS="$CFLAGS -march=pentiumpro"
			dnl CFLAGS="$CFLAGS -march=pentium3 -mmmx -msse -mfpmath=sse,387"
			AC_MSG_RESULT(686)
			;;
		*)
			AC_MSG_RESULT(386)
			AC_MSG_WARN(*** This is probably not what you want use --target)
			;;

Now that I know where to modify, I'd like some pointers on how to add other cpus to help optimizations.
Of course, we can do the ./autogen.sh, then manually edit Makefile to set the CFlag to whatever we want, but might as well put it in the autogen, no ?

No. autogen should be thrown away and the configure script should be distributed. Not everyone has autoconf/automake installed, and it shouldn't be required to compile. autoconf is a developer tool, not an end-user tool.

The big point of autoconf is so that you can build that platform-independent configure script that doesn't need autoconf/automake/gnu m4 or some specialized make like gmake.

The generic INSTALL doc talks about configure. You shouldn't need to generate configure to run it. It should be part and parcel of the distribution.

BTW: make install doesn't create the parent directories if you don't install to the default prefix. This should be changed...