fooling strict aliasing?

[Andrew Haley]

On 09/29/2010 07:39 AM, André Bjärby wrote:

> The following cobe generates a warning "dereferencing type-punned
> pointer will break strict-aliasing rules" in function foo(). When I
> add an intermediate  variable (as in function bar()) the warning
> disappears.
> 
> #include <stdint.h>
> 
> typedef struct blob_St {
> 	uint32_t length;
> 	char data[0] __attribute__ ((aligned (sizeof (uint32_t))));;
> } blob_t;
> 
> 
> int
> foo (blob_t *arg)
> {
> 	return *(uint32_t*)arg->data == 10;
> }
> 
> int
> bar (blob_t *arg)
> {
> 	uint32_t *tmp;
> 
> 	tmp = (uint32_t*)arg->data;
> 	return *tmp == 10;
> }
> 
> 
> This code was compiled using 'gcc -c -Wall test.c -O2'. With gcc
> version 4.4 I get the error, with versions 4.3 and 4.1 I don't.
> 
> My questions are:
> 1. Why does the error disappear when I use an intermediate variable?

Because gcc doesn't do full data-flow analysis.

> 2. Do I run the risk of generating incorrent code if using an
> intermediate variable and compiling with strict aliasing?

Yes.

Andrew.