LLVM bootstrap fails due to HAVE_MMAP_FILE missing

[lattner]

Bugzilla Link	443
Resolution	FIXED
Resolved on	Mar 06, 2010 13:59
Version	trunk
OS	Linux

Extended Description

For an unknown reason, bootstrapping LLVM does not pick up HAVE_MMAP_FILE
correctly, leading to this #error:

Compiling FileUtilities.cpp
FileUtilities.cpp:299:2: #error Unimplemented ReadFileIntoAddressSpace - need to
use read/write.

To reproduce, configure with CC=llvmgcc CXX=llvmg++, and build LLVM.

When manually hacking HAVE_MMAP_FILE into config.h, things work as expected.

-Chris

[llvmbot]

There was a problem with this. The implementation provided in the
llvm/autoconf/m4/func_mmap_file.m4 file was updated by me early yesterday
morning. If you haven't updated or reconfigured since then, I'd try again.

Also, this works fine on Linux so if you have updated/reconfigured recently, I
doubt this is an "All" OS problem. What platform were you testing this on?

[lattner]

This occurs with a fresh checkout of the entire tree, circa 15 mins ago. This
was on a RH7.1 linux box. I assume that this is some funky interaction between
configure and using llvmgcc for the test or something...

-Chris

[llvmbot]

Does it work without the CC= and CXX= lines?

I tried your configure command but it failed right away because it couldn't find
llvmg++. What produces that, anyway? the CFE always just produces gcc or g++ for me.

[lattner]

Does it work without the CC= and CXX= lines?

Yup.

I tried your configure command but it failed right away because it couldn't
find llvmg++. What produces that, anyway? the CFE always just produces gcc or
g++ for me.

llvmg++ is just my local alias for /path/to/llvmgcc/bin/g++

-Chris

[llvmbot]

Could you please attach your config.log (possibly compressed) that was generated?

Thanks.

[lattner]

LLVM-llvm config.log
here ya go

[llvmbot]

The func_mmap_file test is generating a bad test program. I'm not sure why it
does this on RedHat 7.1 but not on RedHat 9.0. Continuing to investigate.

[llvmbot]

The problem is the use of AC_LANG_PROGRAM generating an incorrect program. It
had a "main" inside a "main" and the compiler complained about nested functions.
This resulted from a hasty conversion of the func_mmap_file.m4 script from "old"
autoconf style to "new" autoconf style.

[lattner]

The scary thing is that it "works" with normal GCC's, due to the nested function
extension. Bork. :)

-Chris

[llvmbot]

Problem Resolved.

Patches here:

http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20040920/018473.html
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20040920/018475.html

Reid.

[lattner]

Great, thanks Reid!!

-Chris