Re: your change to tree-sra.c

[kenner at vlsi1 dot ultra dot nyu dot edu (Richard Kenner)]

    DECL_INITIAL being NULL has *always* meant that the variable goes
    in BSS, and thus is zero initialized.  We're talking about loading
    from static, read-only memory.  There must be *SOME* value there,
    and by definition it cannot change.

That's why I raised the question about the meaning of TREE_READONLY.
See your answer at http://gcc.gnu.org/ml/gcc/2004-07/msg00195.html
This is a case where the optimizer *does* use it.

    You've not yet explained to me what is wrong with assuming zero from
    null DECL_INITIAL.  How about you start there?

OK.

For automatic variables, we always emit the code to explicitly initialize
a VAR_DECL and then clear DECL_INITIAL.

For statics, if initializer_constant_valid_p is not set, we can't initialize
it load time.  So the Ada front end, by analogy with locals, emits the
initialization code into a elaboration procedure and clears DECL_INITIAL.

This was the issue for my query about TREE_READONLY.  If it literally means
that a decl isn't written, then we have to clear it when taking something
from DECL_INITIAL and emitting code for it.  But if it's like RTX_UNCHANGING_P
where we're allowed to set it *once*, we can leave it set.