bug or feature? CSE optimizes away global, fixed register

[Georg-Johann Lay]

Hi,

in the test cases

register unsigned char reg2 asm ("r2");

void foo (void)
{
     while (1)
     {
         while (x);
         reg2 = 0;
         while (!x);
     }
}

void bar (void)
{
     while (1)
         reg2 = 0;
}

gcc 4.3.2 optimizes away the accesses to the global variable reg2 that 
lives in general purpose reg r2.

Compiled with
    avr-gcc -Os -ffixed-2

As gcc 3.4.6 behaves as expected (e.g. in bar, the access to reg2 
remains in the loop), I wonder if this aggressive optimization is a bug?

The test case is a simplified excerpt of an application that uses global 
register variables for communication with an interrupt service routine 
(ISR).

This helps to improve the performance of a time critical, real-time 
embedded application and works fine with gcc 3.4.6, but with 4.3.2 CSE 
decides that reg2 is dead. The evil happens in rtl-optimize pass cse1.

IMO that is a bug because reg2 is a fixed register and should therefore 
be treated just like volatile. Unfortunately, gcc has nothing like 
REG_VOLATILE_P to tag a reg being volatile, but as I said global fixed 
registers should be treated that way.

Without that feature gcc 4 is unsuitable for that application and global 
registers are pretty much useless in asynchronous applications. A pity.

I did not try what avr-gcc 4.5 does, because it failed to compile (some 
trunk revision).

Georg-Johann