Re: fix builtins/fputs.c

[Roger Sayle <roger at eyesopen dot com>]

On Fri, 9 Jul 2004, Richard Henderson wrote:
> This patch
>
> 2004-07-04  Roger Sayle  <roger@eyesopen.com>
>
>         * tree-ssa-ccp.c (set_rhs): Change function to return a bool.
>         Ensure the replacement rhs is valid gimple before performing
>         the substitution.  Return false if these sanity checks fail.
>
> caused this test to begin failing on Alpha, because
>
> 	fputs (x, f)
> =>
> 	(int) fwrite (x, 1, 1, f)
>
> We did not strip the nop on Alpha because fwrite returns a long,
> and STRIP_NOP preserves mode conversions.

Thanks for fixing this.

I've been wondering whether this should be split out into an even
more aggressive function, something like "ignore_tree_value (tree)"?
This could be used both here, and within "fold" when converting an
expression to VOID_TYPE, and within omit_one_operand/COMPOUND_EXPR
when we evaluate an expression purely for its side-effects, etc.

Currently, the middle-end's fold_convert just wraps a CONVERT_EXPR
around the expression when casting it to void(!), but clearly not only
can we strip type conversions, but also non-trapping, non-side-effecting
unary and binary operators.

For example, (void)((int)foo() + 3) is really just (void)foo().
Expressions without side-effects could be reduced enitrely to just
integer_zero_node.

This would be a useful functionality to complement tree-ssa's implicit
"cast-to-void" (a.k.a. ignore result) semantics for non-MODIFY_EXPR
statements.

Thoughts?

Roger
--