Re: 1.5.18: suppressing non-C language checks / pass flags to cc in link mode

[Paul Jakma]

Hi Ralf,

On Sat, 26 Nov 2005, Ralf Wildenhues wrote:

> Well, in a way libtool *cannot* allow all flags through, and thus 
> *has* to know about them.  At least it has to know whether a 
> certain flag eats an argument or not.

Well, one simple trick at least would be if:

        -XCClink "x y z"

would result in cc being invoked with x, y and z as seperate 
arguments. Unfortunately it does not, cc is then invoked with 'x y z' 
as one arg ;).

> And it cannot allow flags through without knowing, when those flags 
> alter linking behavior which libtool assumes differently.  Also 
> note that on some systems $LD is used for linking rather than $CC 
> or $CXX.

Right. But libtool knows that, no point passing CFLAGS to ld.

I tried going through libtool.m4 to find an easy fix, but it's not 
very obvious.

> > The only other alternative then, it seems, is to write your 
> > configure script to include some kind of extra 'LINK_MODE_CFLAGS' 
> > variable, then tack that on to CFLAGS and also generate an 
> > additional 'LI5BBTOOL_CCLINK_FLAGS' variable for each word in 
> > LINK_MODE_CFLAGS combined with -XCClinker and then use 
> > LIBTOOL_CCLINK_FLAGS in prog_LDADD in your Makefile.am's.

> something like this, yes.  It would be nice to adjust Autoconf/Automake
> to cope better with this situation.

Well, a libtool autoconf macro to specify the CClink flags to use 
might be nice.

> Which flags do you need passed through?

For gcc:
        -fpic
        -fPIC
        -fpie
        -fPIE

(I think you have some support for -fpic already?)

For Sun Studio, the ones i'm interested in:

        -xarch=....
        -xtarget=....
        -xchip=....
        -xcode=....
        -fast
        -xOn    (n is an integer)
        -xipo
        -xipo=n

For more see A.1.2 in:

        http://docs.sun.com/source/819-0494/cc_ops.app.html

And A.1.11 obviously should be passed through too.

Note that -xOn and -xtarget are not listed in A.1.2, because they are 
macro arguments which simply are equivalent to several other base 
arguments, however they must still be given.

> We should deal with all those correctly.  Please list the ones we're
> currently missing, so that can be fixed.

See above. -xtarget=.... is probably the most annoying one which is 
missing, because it is required in order to be compile for 64bit. ;)

> This patch is a decent workaround.
>
> I'm inclined to accept a fix for branch-1-5 if somebody has the 
> patience to dig out the patch that fixed this in CVS HEAD (plus 
> eventual subsequent fixes of the patch) and it can be applied to 
> branch-1-5 without adding new interfaces or changing existing ones, 
> and is not too disruptive in general.

Sounds good, I've not followed libtool development at all though, so 
I've no idea what/where that patch might be.

> If merely for the reason that I hate to read yet another bug report 
> about this issue, or see yet another dysfunctional workaround for 
> this (unlike above!).

;)

> Other than that, I can only reiterate that we are working on the 
> last few known issues in CVS HEAD, hope to have them fixed in the 
> not too distant future, and will then release another 2.0 alpha, 
> and 2.0 hopefully soon afterwards.  Yes, I know I sound like a 
> street organ. :-/

:)

Thanks very much!

regards,
--
Paul Jakma      address@hidden  address@hidden  Key ID: 64A2FF6A
Fortune:
"Here's something to think about:  How come you never see a headline like
 `Psychic Wins Lottery.'"
-- Comedian Jay Leno