Re: Optimizing in assembly language

> So here's my question: why can't we write an "optimizing assembler"
> that lets the programmer specify machine sequences and the assembler
> does the same kinds of "book keeping" that a HLL compiler would do.


IIRC DEC/Mips used at least a bit of this approach. There were some
problems, the most obvious of which were related to use of memory as
semaphores. Code motion to improve memory bandwith removed full
protection of critical sections.
        ...
        while (lock != 0)
                ...
                ;
        lock = 1
        ...update data with mutiple instruction...
        lock = 0
        ...
Dead store elimination, if implemented, can remove locks on
resources above entirely, leaving code like:
        ...
        while (lock != 0)
                ...
                ;
        ...update data with multiple instruction...
        lock = 0
        ...
or even (on discovering the protect loop always exits with lock==0):
        ...
        while (lock != 0)
                ...
                ;
        update data with more than one indivisible instruction
        ...


A "clever" assembler may notice the "lock=1" code is useless because
lock is not examined anywhere in the update, so it may simply
eliminate the lock = 1 altogether. Often this is right, but if we are
concerned with multiple threads viewing the data, the optimization is
wrong. It is tough to determine which bits of assembly are
intentional and which are artifact, and I wouldn't trust an assembler
to do it myself.


Scott David Daniels

Return to the comp.compilers page.
Search the comp.compilers archives again.