[Cryptography] GCC bug 30475 (was Re: bounded pointers in C)

[Bill Frantz]

On 5/1/14 at 9:38 AM, leichter at lrw.com (Jerry Leichter) wrote:

>The compiler always knows the target architecture.  But that 
>result of an expression like x > x + 1 is undefined if x + 1 
>overflows.  Again, it might trap.

I'm sorry, but I don't have much time, and am about to go away 
from the Internet for a while, so I won't give you response the 
careful thought it deserves (and probably not see your responses 
for several weeks).


While compilers don't always know the target architecture, the 
optimizer phase in some implementations may not. I have seen 
compilers that produced a intermediate language which is then 
translated to machine code.

 From the view of safe programming, if x > x+1 traps and aborts 
the program, this is usually better than throwing out the 
statement entirely. In the very common case of twos compliment 
modular arithmetic the statement has meaning and should be executed.

While I certainly don't know all architectures and compilers, 
every one I know permits arithmetic overflow traps to be masked 
so they don't occur and the result of the operation is modulo 
the arithmetic base of the machine. Both the IBM 650 and the 
1620 behaved this way with their decimal arithmetic. The 1620 
was variable word length so the modulus depended on the word length.


But what I think I really want is a safe mode where dangerous 
optimizations are avoided or terminal compile errors are 
generated. I hate thinking that the only safe low level language 
is assembler.

[I have heard mention of optimizing assemblers which will remove 
code, but never encountered one. The only optimizing assembler I 
know is the IBM 650 SOAP assembler which arranged instruction 
and data words around the rotating drum main memory to minimize 
access time.]

Cheers - Bill

---------------------------------------------------------------------------
Bill Frantz        |"We used to quip that "password" is the most common
408-356-8506       | password. Now it's 'password1.' Who said 
users haven't
www.pwpconsult.com | learned anything about security?" -- Bruce Schneier