On 29/11/11 16:44, Joseph S. Myers wrote:
On Tue, 29 Nov 2011, Nadezhda Ivan�vna Vyukova wrote:
I've explained the customer that by default char is treated
as signed char on our platform and therefore this program
does not conform ISO C90, as it causes the integer overflow
(undefined behavior). But he was not satisfied.
Actually this program does not involve undefined behavior, because the
increment happens in type int and then implementation-defined behavior
converting the int to char (which GCC defines as modulo) applies. See PR
35634 which has various discussion of possible approaches for fixing
this, and patches that introduce vectorizer optimization regressions.
With this interpretation, then I can see that the program does not
involve undefined behaviour. But I think it should behave differently.
When e is 127 before the loop, the "while (++e)" will promote e from a
(signed) char to a signed int - giving 127. This is incremented to 128
(valid for an int). The value is stored back to e by modulo truncation,
which gives -128 for a signed char. And the value 128 is tested for the
while loop.
So after printing 127, it should print -128.
This will be followed by -127, -126, ..., -1 when e is -1.
When executing "while (++e)", ++e is now 0 and the loop will exit.
But when compiled with -O2 with "-Wstrict-overflow", the compiler warns
"assuming signed overflow does not occur when simplifying conditional to
constant" - and the program runs prints a never-ending sequence of
positive integers (well, until I control-C'd the program).
Does this mean that there is a bug somewhere, or is it the conversion
from int value 128 to a signed char that is undefined?
mvh.,
David
