Re: [avr-chat] Missed Optimisation ?

[Graham Davies]

bob wrote:

> In the construct <load value> < OR > <save value>, it just doesn't make
> sense to me to cheat on the OR'ing but not on the load/save part.

OK, that's fine, but the problem is in what makes sense to you, not in what 
the compiler is doing.  Please read over more carefully what I wrote in my 
previous reply.  Again, when loading and storing, the compiler is 
constrained by the use of the volatile storage qualifier to do exactly what 
the programmer writes, i.e. fetch a 32-bit value from a variable and later 
write back a 32-bit variable.  In between, we tell the compiler to modify 
the value in a fixed way (known at compile time).  The *value* is not 
volatile, it's just a pattern of bits.  The compiler can use any tricks it 
wants to perform the modification as long as the result is always the same 
as doing it the stupid and obvious way.  Modifying the value (performing the 
OR operation) has no effect on the variable until it is written back.  So 
the volatile storage qualifier has no effect on how the compiler contrives 
to perform the modification, only on how it gets written back.

> Your thread argument is true but would be broken by the current compiler
> action ...

The compiler does not know about threads or any form of concurrent access to 
a variable.  The compiler only knows what you tell it in a single 
translation unit (C file and included header files).  The compiler does not 
know what is memory and what is a special function register.  You have to 
look after all this with a) the volatile storage qualifier (which disables 
certain optimizations) and b) tricks to ensure that multi-access variables 
are changed in an atomic fashion.  These two are separate.  The volatile 
declaration will not make the compiler ensure atomic access.

Graham.

P.S.  The use and effect of the volatile storage qualifier is one of the 
most misunderstood aspects of the C language.  So, with due respect to all 
concerned, those who know (or suspect) that they are *not experts* would do 
better to stay out of this discussion thread.