clang miscompiles emacs because it uses horribly undefined behavior

[efriedma-quic]

Bugzilla Link	4041
Resolution	INVALID
Resolved on	Feb 22, 2010 12:49
Version	unspecified
OS	Linux

Extended Description

A build of emacs on Linux with clang breaks in the middle; tail of the make output:
LC_ALL=C ./temacs -batch -l loadup dump
Segmentation fault
make[1]: *** [emacs] Error 139
make[1]: Leaving directory `/home/eli/emacs-22.3/src'
make: *** [src] Error 2

This is a build configured with just "CC=clang ./configure".

I'm going to try and track this down at some point, but I figured I should file it so we don't lose track of it.

[efriedma-quic]

Quote from the emacs intel386.h:

/* Define NO_ARG_ARRAY if you cannot take the address of the first of a

• group of arguments and treat it as an array of the arguments. */

/* #define NO_ARG_ARRAY */

Doesn't anyone understand "undefined behavior"? Ugh.

[llvmbot]

Wow that is awesome. Does gcc actually have extra code to support this: pin variables to a contiguous region of the stack if any of their address are taken?

[efriedma-quic]

No, what it's doing is that it's assuming &arg1 is the actual address of the first argument, which is what gcc returns... it then can assume &arg1+1 is the address of the second argument, and so on.

This is the same sort of issue as bug 2826.