i386/186848: CLANG/LLVM code generation bug with optimisation on i386

[Bruce Evans]

The following reply was made to PR i386/186848; it has been noted by GNATS.

From: Bruce Evans <brde at optusnet.com.au>
To: David Hines <freebsd at dph.fluff.org>
Cc: freebsd-gnats-submit at freebsd.org, freebsd-i386 at freebsd.org
Subject: Re: i386/186848: CLANG/LLVM code generation bug with optimisation
 on i386
Date: Thu, 20 Feb 2014 14:06:55 +1100 (EST)

 On Mon, 17 Feb 2014, David Hines wrote:
 
 >> Description:
 > "cc -O -o clang_bug-O clang_bug.c" generates incorrect code on i386, with the attached sample code. Without the "-O", or on an amd64 install the problem does not occur.
 
 This is a bug in clang_bug.c.  It's behaviour is undefined.
 
 > Xunion
 > X{
 > X    int i;
 > X} u;
 > X
 > X
 > Xint
 > Xmain(int argc, char *argv[])
 > X{
 > X    int j = 1;
 > X
 > X    for (u.i = 1;  u.i += u.i;  ++j)
 > X	;
 > X    printf("An int has %d bits\n", j);
 > X
 > X    return 0;
 > X}
 
 Undefined behaviour occurs when the addition overflows.  clang somehow
 notices this (u.i starts as 1, and repeated doublings of it cannot
 make it 0 unless overflow occurs).  The addition does in fact overflow.
 The result can be anything.
 
 Some people (not me) don't like clang not giving the "normal" behaviour
 on overflow.
 
 Bruce