Tcl Source Code

I wanted to test compat/strtod.c on Linux 6.2
on my Alpha XL-300:

$ cd unix
$ setenv LIBOBJS strtod.o
$ ./configure
$ make test

but the resulting tclsh didn't use the strtod()
from compat/strtod.c.

The problem was that with optimization level "-O"
the compiler egcs-2.91.66 and the header files from
glibc-devel-2.1.3-21 automatically inlines the
function strtod() so it can't be replaced.

The attached patch fixes the problem.  It detects
the intent to link in routines from compat/* and
in that case adds to EXTRA_CFLAGS the switch
-fno-inline.

Logged In: YES 
user_id=80530
Originator: YES


This bug was apparently only
for the egcs-2.91.66 compiler,
and the world has moved on.

Logged In: YES 
user_id=72656

Feel free to recorrect, but only with checks for that old 
version.  Newer compilers do not have the bug, and it does 
improve speed to allow inlining.

Logged In: YES 
user_id=80530

looks like the new commits
went into HEAD only.

On the HEAD, when I follow
the steps in this report, I get
in the Makefile:

COMPAT_OBJS     = strtod$U.o

I have no idea what that $U is
for, but it seems to be produced
by the new autoconf-.2.57 generated
configure script.  Anyhow, it does seem
to work, putting strtod.o on the list
of compiled files that get put into libtcl.

As expected, though, due to the inlining,
the compat/strtod() routine is never called.
This report is re-opened for Tcl 8.5a0.

Logged In: YES 
user_id=80530


Just curious how the commits
of Nov. 11, 2003 will impact
on this report?

Logged In: YES 
user_id=90858

Doh! You are quite right. The -fno-builtin flag
should work but it seems to be broken in this case.
I will go ahead and use the -fno-inline flag.

Logged In: YES 
user_id=80530

I see exactly the same thing on my Linux system.
However:

$ make tclCmdMZ.o CFLAGS="-O -fno-builtin"
$ nm tclCmdMZ.o | grep strtod
         U __strtod_internal

So, what is the key to success?  Is it adding -fno-builtin,
or is it just removing -O ?

Logged In: YES 
user_id=90858

I took a look at the .h files and it does seem
like something funny is going on. Thing is,
passing -fno-builtin does seem to work for me.

% make tclCmdMZ.o CFLAGS="-O"

% nm tclCmdMZ.o | grep strtod
         U __strtod_internal

% make tclCmdMZ.o CFLAGS=""

% nm tclCmdMZ.o | grep strtod
         U strtod

% make tclCmdMZ.o CFLAGS="-fno-builtin"

% nm tclCmdMZ.o | grep strtod
         U strtod

Do you still get a ref to __strtod_internal
when passing -fno-builtin? If not, what do
you get? I am running on an x86 box with
RH 6.2 in case that matters.

Logged In: YES 
user_id=80530

Tcl also adds -D__NO_STRING_INLINES and -D__NO_MATH_INLINES
(they come from the shared tcl.m4 file) but that has no
effect on the issue in this report.

Logged In: YES 
user_id=72656

Tk adds -D__NO_STRING_INLINES and -D__NO_MATH_INLINES which 
may be the right fix.

Logged In: YES 
user_id=80530


No, -fno-builtin doesn't solve the problem.

Logged In: YES 
user_id=90858

Does passing the -fno-builtin also work for your
case? The docs seem to indicate that it will.
I would rather use -fno-builtin instead of -fno-inline
since it will only stop inlining of builtin functions.

Logged In: YES 
user_id=80530

assigning to maintainer for review.

File Added - 8370: inline.patch